diff options
Diffstat (limited to 'drivers/gpu/drm/amd/amdkfd')
41 files changed, 3014 insertions, 3363 deletions
diff --git a/drivers/gpu/drm/amd/amdkfd/Kconfig b/drivers/gpu/drm/amd/amdkfd/Kconfig index a1a35d4d594b..b3672d10ea54 100644 --- a/drivers/gpu/drm/amd/amdkfd/Kconfig +++ b/drivers/gpu/drm/amd/amdkfd/Kconfig @@ -1,11 +1,11 @@ -# SPDX-License-Identifier: GPL-2.0-only +# SPDX-License-Identifier: MIT # # Heterogenous system architecture configuration # config HSA_AMD bool "HSA kernel driver for AMD GPU devices" - depends on DRM_AMDGPU && (X86_64 || ARM64) + depends on DRM_AMDGPU && (X86_64 || ARM64 || PPC64) imply AMD_IOMMU_V2 if X86_64 select MMU_NOTIFIER help diff --git a/drivers/gpu/drm/amd/amdkfd/Makefile b/drivers/gpu/drm/amd/amdkfd/Makefile index 48155060a57c..61474627a32c 100644 --- a/drivers/gpu/drm/amd/amdkfd/Makefile +++ b/drivers/gpu/drm/amd/amdkfd/Makefile @@ -38,11 +38,9 @@ AMDKFD_FILES := $(AMDKFD_PATH)/kfd_module.o \ $(AMDKFD_PATH)/kfd_mqd_manager_v9.o \ $(AMDKFD_PATH)/kfd_mqd_manager_v10.o \ $(AMDKFD_PATH)/kfd_kernel_queue.o \ - $(AMDKFD_PATH)/kfd_kernel_queue_cik.o \ - $(AMDKFD_PATH)/kfd_kernel_queue_vi.o \ - $(AMDKFD_PATH)/kfd_kernel_queue_v9.o \ - $(AMDKFD_PATH)/kfd_kernel_queue_v10.o \ $(AMDKFD_PATH)/kfd_packet_manager.o \ + $(AMDKFD_PATH)/kfd_packet_manager_vi.o \ + $(AMDKFD_PATH)/kfd_packet_manager_v9.o \ $(AMDKFD_PATH)/kfd_process_queue_manager.o \ $(AMDKFD_PATH)/kfd_device_queue_manager.o \ $(AMDKFD_PATH)/kfd_device_queue_manager_cik.o \ diff --git a/drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c b/drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c index 177d1e5329a5..9f59ba93cfe0 100644 --- a/drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c +++ b/drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c @@ -33,7 +33,9 @@ static bool cik_event_interrupt_isr(struct kfd_dev *dev, const struct cik_ih_ring_entry *ihre = (const struct cik_ih_ring_entry *)ih_ring_entry; const struct kfd2kgd_calls *f2g = dev->kfd2kgd; - unsigned int vmid, pasid; + unsigned int vmid; + uint16_t pasid; + bool ret; /* This workaround is due to HW/FW limitation on Hawaii that * VMID and PASID are not written into ih_ring_entry @@ -48,13 +50,13 @@ static bool cik_event_interrupt_isr(struct kfd_dev *dev, *tmp_ihre = *ihre; vmid = f2g->read_vmid_from_vmfault_reg(dev->kgd); - pasid = f2g->get_atc_vmid_pasid_mapping_pasid(dev->kgd, vmid); + ret = f2g->get_atc_vmid_pasid_mapping_info(dev->kgd, vmid, &pasid); tmp_ihre->ring_id &= 0x000000ff; tmp_ihre->ring_id |= vmid << 8; tmp_ihre->ring_id |= pasid << 16; - return (pasid != 0) && + return ret && (pasid != 0) && vmid >= dev->vm_info.first_vmid_kfd && vmid <= dev->vm_info.last_vmid_kfd; } diff --git a/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler.h b/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler.h index 826913c70766..d3400da6ab64 100644 --- a/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler.h +++ b/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler.h @@ -274,154 +274,227 @@ static const uint32_t cwsr_trap_gfx8_hex[] = { static const uint32_t cwsr_trap_gfx9_hex[] = { - 0xbf820001, 0xbf82015e, + 0xbf820001, 0xbf820248, 0xb8f8f802, 0x89788678, - 0xb8fbf803, 0x866eff7b, - 0x00000400, 0xbf85003b, - 0x866eff7b, 0x00000800, - 0xbf850003, 0x866eff7b, - 0x00000100, 0xbf84000c, + 0xb8eef801, 0x866eff6e, + 0x00000800, 0xbf840003, 0x866eff78, 0x00002000, - 0xbf840005, 0xbf8e0010, - 0xb8eef803, 0x866eff6e, - 0x00000400, 0xbf84fffb, - 0x8778ff78, 0x00002000, - 0x80ec886c, 0x82ed806d, - 0xb8eef807, 0x866fff6e, - 0x001f8000, 0x8e6f8b6f, - 0x8977ff77, 0xfc000000, - 0x87776f77, 0x896eff6e, - 0x001f8000, 0xb96ef807, - 0xb8faf812, 0xb8fbf813, - 0x8efa887a, 0xc0071bbd, - 0x00000000, 0xbf8cc07f, - 0xc0071ebd, 0x00000008, - 0xbf8cc07f, 0x86ee6e6e, - 0xbf840001, 0xbe801d6e, - 0xb8fbf803, 0x867bff7b, - 0x000001ff, 0xbf850002, - 0x806c846c, 0x826d806d, + 0xbf840016, 0xb8fbf803, + 0x866eff7b, 0x00000400, + 0xbf85003b, 0x866eff7b, + 0x00000800, 0xbf850003, + 0x866eff7b, 0x00000100, + 0xbf84000c, 0x866eff78, + 0x00002000, 0xbf840005, + 0xbf8e0010, 0xb8eef803, + 0x866eff6e, 0x00000400, + 0xbf84fffb, 0x8778ff78, + 0x00002000, 0x80ec886c, + 0x82ed806d, 0xb8eef807, + 0x866fff6e, 0x001f8000, + 0x8e6f8b6f, 0x8977ff77, + 0xfc000000, 0x87776f77, + 0x896eff6e, 0x001f8000, + 0xb96ef807, 0xb8faf812, + 0xb8fbf813, 0x8efa887a, + 0xc0071bbd, 0x00000000, + 0xbf8cc07f, 0xc0071ebd, + 0x00000008, 0xbf8cc07f, + 0x86ee6e6e, 0xbf840001, + 0xbe801d6e, 0xb8fbf803, + 0x867bff7b, 0x000001ff, + 0xbf850002, 0x806c846c, + 0x826d806d, 0x866dff6d, + 0x0000ffff, 0x8f6e8b77, + 0x866eff6e, 0x001f8000, + 0xb96ef807, 0x86fe7e7e, + 0x86ea6a6a, 0x8f6e8378, + 0xb96ee0c2, 0xbf800002, + 0xb9780002, 0xbe801f6c, 0x866dff6d, 0x0000ffff, - 0x8f6e8b77, 0x866eff6e, - 0x001f8000, 0xb96ef807, - 0x86fe7e7e, 0x86ea6a6a, - 0x8f6e8378, 0xb96ee0c2, - 0xbf800002, 0xb9780002, - 0xbe801f6c, 0x866dff6d, - 0x0000ffff, 0xbefa0080, - 0xb97a0283, 0xb8fa2407, - 0x8e7a9b7a, 0x876d7a6d, - 0xb8fa03c7, 0x8e7a9a7a, - 0x876d7a6d, 0xb8faf807, - 0x867aff7a, 0x00007fff, - 0xb97af807, 0xbeee007e, - 0xbeef007f, 0xbefe0180, - 0xbf900004, 0x877a8478, - 0xb97af802, 0xbf8e0002, - 0xbf88fffe, 0xb8fa2a05, - 0x807a817a, 0x8e7a8a7a, - 0xb8fb1605, 0x807b817b, - 0x8e7b867b, 0x807a7b7a, - 0x807a7e7a, 0x827b807f, - 0x867bff7b, 0x0000ffff, - 0xc04b1c3d, 0x00000050, - 0xbf8cc07f, 0xc04b1d3d, - 0x00000060, 0xbf8cc07f, - 0xc0431e7d, 0x00000074, - 0xbf8cc07f, 0xbef4007e, - 0x8675ff7f, 0x0000ffff, - 0x8775ff75, 0x00040000, - 0xbef60080, 0xbef700ff, - 0x00807fac, 0x867aff7f, - 0x08000000, 0x8f7a837a, - 0x87777a77, 0x867aff7f, - 0x70000000, 0x8f7a817a, - 0x87777a77, 0xbef1007c, - 0xbef00080, 0xb8f02a05, - 0x80708170, 0x8e708a70, - 0xb8fa1605, 0x807a817a, - 0x8e7a867a, 0x80707a70, - 0xbef60084, 0xbef600ff, - 0x01000000, 0xbefe007c, - 0xbefc0070, 0xc0611c7a, - 0x0000007c, 0xbf8cc07f, - 0x80708470, 0xbefc007e, + 0xbefa0080, 0xb97a0283, + 0xb8fa2407, 0x8e7a9b7a, + 0x876d7a6d, 0xb8fa03c7, + 0x8e7a9a7a, 0x876d7a6d, + 0xb8faf807, 0x867aff7a, + 0x00007fff, 0xb97af807, + 0xbeee007e, 0xbeef007f, + 0xbefe0180, 0xbf900004, + 0x877a8478, 0xb97af802, + 0xbf8e0002, 0xbf88fffe, + 0xb8fa2a05, 0x807a817a, + 0x8e7a8a7a, 0xb8fb1605, + 0x807b817b, 0x8e7b867b, + 0x807a7b7a, 0x807a7e7a, + 0x827b807f, 0x867bff7b, + 0x0000ffff, 0xc04b1c3d, + 0x00000050, 0xbf8cc07f, + 0xc04b1d3d, 0x00000060, + 0xbf8cc07f, 0xc0431e7d, + 0x00000074, 0xbf8cc07f, + 0xbef4007e, 0x8675ff7f, + 0x0000ffff, 0x8775ff75, + 0x00040000, 0xbef60080, + 0xbef700ff, 0x00807fac, + 0x867aff7f, 0x08000000, + 0x8f7a837a, 0x87777a77, + 0x867aff7f, 0x70000000, + 0x8f7a817a, 0x87777a77, + 0xbef1007c, 0xbef00080, + 0xb8f02a05, 0x80708170, + 0x8e708a70, 0xb8fa1605, + 0x807a817a, 0x8e7a867a, + 0x80707a70, 0xbef60084, + 0xbef600ff, 0x01000000, 0xbefe007c, 0xbefc0070, - 0xc0611b3a, 0x0000007c, + 0xc0611c7a, 0x0000007c, 0xbf8cc07f, 0x80708470, 0xbefc007e, 0xbefe007c, - 0xbefc0070, 0xc0611b7a, + 0xbefc0070, 0xc0611b3a, 0x0000007c, 0xbf8cc07f, 0x80708470, 0xbefc007e, 0xbefe007c, 0xbefc0070, - 0xc0611bba, 0x0000007c, + 0xc0611b7a, 0x0000007c, 0xbf8cc07f, 0x80708470, 0xbefc007e, 0xbefe007c, - 0xbefc0070, 0xc0611bfa, + 0xbefc0070, 0xc0611bba, 0x0000007c, 0xbf8cc07f, 0x80708470, 0xbefc007e, 0xbefe007c, 0xbefc0070, - 0xc0611e3a, 0x0000007c, - 0xbf8cc07f, 0x80708470, - 0xbefc007e, 0xb8fbf803, - 0xbefe007c, 0xbefc0070, - 0xc0611efa, 0x0000007c, + 0xc0611bfa, 0x0000007c, 0xbf8cc07f, 0x80708470, 0xbefc007e, 0xbefe007c, - 0xbefc0070, 0xc0611a3a, + 0xbefc0070, 0xc0611e3a, + 0x0000007c, 0xbf8cc07f, + 0x80708470, 0xbefc007e, + 0xb8fbf803, 0xbefe007c, + 0xbefc0070, 0xc0611efa, 0x0000007c, 0xbf8cc07f, 0x80708470, 0xbefc007e, 0xbefe007c, 0xbefc0070, - 0xc0611a7a, 0x0000007c, - 0xbf8cc07f, 0x80708470, - 0xbefc007e, 0xb8f1f801, - 0xbefe007c, 0xbefc0070, - 0xc0611c7a, 0x0000007c, + 0xc0611a3a, 0x0000007c, 0xbf8cc07f, 0x80708470, - 0xbefc007e, 0x867aff7f, - 0x04000000, 0xbeef0080, - 0x876f6f7a, 0xb8f02a05, + 0xbefc007e, 0xbefe007c, + 0xbefc0070, 0xc0611a7a, + 0x0000007c, 0xbf8cc07f, + 0x80708470, 0xbefc007e, + 0xb8f1f801, 0xbefe007c, + 0xbefc0070, 0xc0611c7a, + 0x0000007c, 0xbf8cc07f, + 0x80708470, 0xbefc007e, + 0x867aff7f, 0x04000000, + 0xbeef0080, 0x876f6f7a, + 0xb8f02a05, 0x80708170, + 0x8e708a70, 0xb8fb1605, + 0x807b817b, 0x8e7b847b, + 0x8e76827b, 0xbef600ff, + 0x01000000, 0xbef20174, + 0x80747074, 0x82758075, + 0xbefc0080, 0xbf800000, + 0xbe802b00, 0xbe822b02, + 0xbe842b04, 0xbe862b06, + 0xbe882b08, 0xbe8a2b0a, + 0xbe8c2b0c, 0xbe8e2b0e, + 0xc06b003a, 0x00000000, + 0xbf8cc07f, 0xc06b013a, + 0x00000010, 0xbf8cc07f, + 0xc06b023a, 0x00000020, + 0xbf8cc07f, 0xc06b033a, + 0x00000030, 0xbf8cc07f, + 0x8074c074, 0x82758075, + 0x807c907c, 0xbf0a7b7c, + 0xbf85ffe7, 0xbef40172, + 0xbef00080, 0xbefe00c1, + 0xbeff00c1, 0xbee80080, + 0xbee90080, 0xbef600ff, + 0x01000000, 0x867aff78, + 0x00400000, 0xbf850003, + 0xb8faf803, 0x897a7aff, + 0x10000000, 0xbf85004d, + 0xbe840080, 0xd2890000, + 0x00000900, 0x80048104, + 0xd2890001, 0x00000900, + 0x80048104, 0xd2890002, + 0x00000900, 0x80048104, + 0xd2890003, 0x00000900, + 0x80048104, 0xc069003a, + 0x00000070, 0xbf8cc07f, + 0x80709070, 0xbf06c004, + 0xbf84ffee, 0xbe840080, + 0xd2890000, 0x00000901, + 0x80048104, 0xd2890001, + 0x00000901, 0x80048104, + 0xd2890002, 0x00000901, + 0x80048104, 0xd2890003, + 0x00000901, 0x80048104, + 0xc069003a, 0x00000070, + 0xbf8cc07f, 0x80709070, + 0xbf06c004, 0xbf84ffee, + 0xbe840080, 0xd2890000, + 0x00000902, 0x80048104, + 0xd2890001, 0x00000902, + 0x80048104, 0xd2890002, + 0x00000902, 0x80048104, + 0xd2890003, 0x00000902, + 0x80048104, 0xc069003a, + 0x00000070, 0xbf8cc07f, + 0x80709070, 0xbf06c004, + 0xbf84ffee, 0xbe840080, + 0xd2890000, 0x00000903, + 0x80048104, 0xd2890001, + 0x00000903, 0x80048104, + 0xd2890002, 0x00000903, + 0x80048104, 0xd2890003, + 0x00000903, 0x80048104, + 0xc069003a, 0x00000070, + 0xbf8cc07f, 0x80709070, + 0xbf06c004, 0xbf84ffee, + 0xbf820008, 0xe0724000, + 0x701d0000, 0xe0724100, + 0x701d0100, 0xe0724200, + 0x701d0200, 0xe0724300, + 0x701d0300, 0xbefe00c1, + 0xbeff00c1, 0xb8fb4306, + 0x867bc17b, 0xbf840063, + 0xbf8a0000, 0x867aff6f, + 0x04000000, 0xbf84005f, + 0x8e7b867b, 0x8e7b827b, + 0xbef6007b, 0xb8f02a05, 0x80708170, 0x8e708a70, - 0xb8fb1605, 0x807b817b, - 0x8e7b847b, 0x8e76827b, - 0xbef600ff, 0x01000000, - 0xbef20174, 0x80747074, - 0x82758075, 0xbefc0080, - 0xbf800000, 0xbe802b00, - 0xbe822b02, 0xbe842b04, - 0xbe862b06, 0xbe882b08, - 0xbe8a2b0a, 0xbe8c2b0c, - 0xbe8e2b0e, 0xc06b003a, - 0x00000000, 0xbf8cc07f, - 0xc06b013a, 0x00000010, - 0xbf8cc07f, 0xc06b023a, - 0x00000020, 0xbf8cc07f, - 0xc06b033a, 0x00000030, - 0xbf8cc07f, 0x8074c074, - 0x82758075, 0x807c907c, - 0xbf0a7b7c, 0xbf85ffe7, - 0xbef40172, 0xbef00080, - 0xbefe00c1, 0xbeff00c1, - 0xbee80080, 0xbee90080, + 0xb8fa1605, 0x807a817a, + 0x8e7a867a, 0x80707a70, + 0x8070ff70, 0x00000080, 0xbef600ff, 0x01000000, - 0xe0724000, 0x701d0000, - 0xe0724100, 0x701d0100, - 0xe0724200, 0x701d0200, - 0xe0724300, 0x701d0300, - 0xbefe00c1, 0xbeff00c1, - 0xb8fb4306, 0x867bc17b, - 0xbf84002c, 0xbf8a0000, - 0x867aff6f, 0x04000000, - 0xbf840028, 0x8e7b867b, - 0x8e7b827b, 0xbef6007b, - 0xb8f02a05, 0x80708170, - 0x8e708a70, 0xb8fa1605, - 0x807a817a, 0x8e7a867a, - 0x80707a70, 0x8070ff70, - 0x00000080, 0xbef600ff, - 0x01000000, 0xbefc0080, - 0xd28c0002, 0x000100c1, - 0xd28d0003, 0x000204c1, + 0xbefc0080, 0xd28c0002, + 0x000100c1, 0xd28d0003, + 0x000204c1, 0x867aff78, + 0x00400000, 0xbf850003, + 0xb8faf803, 0x897a7aff, + 0x10000000, 0xbf850030, + 0x24040682, 0xd86e4000, + 0x00000002, 0xbf8cc07f, + 0xbe840080, 0xd2890000, + 0x00000900, 0x80048104, + 0xd2890001, 0x00000900, + 0x80048104, 0xd2890002, + 0x00000900, 0x80048104, + 0xd2890003, 0x00000900, + 0x80048104, 0xc069003a, + 0x00000070, 0xbf8cc07f, + 0x80709070, 0xbf06c004, + 0xbf84ffee, 0xbe840080, + 0xd2890000, 0x00000901, + 0x80048104, 0xd2890001, + 0x00000901, 0x80048104, + 0xd2890002, 0x00000901, + 0x80048104, 0xd2890003, + 0x00000901, 0x80048104, + 0xc069003a, 0x00000070, + 0xbf8cc07f, 0x80709070, + 0xbf06c004, 0xbf84ffee, + 0x680404ff, 0x00000200, + 0xd0c9006a, 0x0000f702, + 0xbf87ffd2, 0xbf820015, 0xd1060002, 0x00011103, 0x7e0602ff, 0x00000200, 0xbefc00ff, 0x00010000, @@ -438,9 +511,53 @@ static const uint32_t cwsr_trap_gfx9_hex[] = { 0x807b817b, 0x8e7b827b, 0x8e76887b, 0xbef600ff, 0x01000000, 0xbefc0084, - 0xbf0a7b7c, 0xbf840015, + 0xbf0a7b7c, 0xbf84006d, 0xbf11017c, 0x807bff7b, - 0x00001000, 0x7e000300, + 0x00001000, 0x867aff78, + 0x00400000, 0xbf850003, + 0xb8faf803, 0x897a7aff, + 0x10000000, 0xbf850051, + 0xbe840080, 0xd2890000, + 0x00000900, 0x80048104, + 0xd2890001, 0x00000900, + 0x80048104, 0xd2890002, + 0x00000900, 0x80048104, + 0xd2890003, 0x00000900, + 0x80048104, 0xc069003a, + 0x00000070, 0xbf8cc07f, + 0x80709070, 0xbf06c004, + 0xbf84ffee, 0xbe840080, + 0xd2890000, 0x00000901, + 0x80048104, 0xd2890001, + 0x00000901, 0x80048104, + 0xd2890002, 0x00000901, + 0x80048104, 0xd2890003, + 0x00000901, 0x80048104, + 0xc069003a, 0x00000070, + 0xbf8cc07f, 0x80709070, + 0xbf06c004, 0xbf84ffee, + 0xbe840080, 0xd2890000, + 0x00000902, 0x80048104, + 0xd2890001, 0x00000902, + 0x80048104, 0xd2890002, + 0x00000902, 0x80048104, + 0xd2890003, 0x00000902, + 0x80048104, 0xc069003a, + 0x00000070, 0xbf8cc07f, + 0x80709070, 0xbf06c004, + 0xbf84ffee, 0xbe840080, + 0xd2890000, 0x00000903, + 0x80048104, 0xd2890001, + 0x00000903, 0x80048104, + 0xd2890002, 0x00000903, + 0x80048104, 0xd2890003, + 0x00000903, 0x80048104, + 0xc069003a, 0x00000070, + 0xbf8cc07f, 0x80709070, + 0xbf06c004, 0xbf84ffee, + 0x807c847c, 0xbf0a7b7c, + 0xbf85ffb1, 0xbf9c0000, + 0xbf820012, 0x7e000300, 0x7e020301, 0x7e040302, 0x7e060303, 0xe0724000, 0x701d0000, 0xe0724100, @@ -563,24 +680,47 @@ static const uint32_t cwsr_trap_gfx9_hex[] = { }; static const uint32_t cwsr_trap_gfx10_hex[] = { - 0xbf820001, 0xbf82012e, - 0xb0804004, 0xb970f802, - 0x8a708670, 0xb971f803, - 0x8771ff71, 0x00000400, - 0xbf850008, 0xb971f803, - 0x8771ff71, 0x000001ff, - 0xbf850001, 0x806c846c, + 0xbf820001, 0xbf8201c1, + 0xb0804004, 0xb978f802, + 0x8a788678, 0xb971f803, + 0x876eff71, 0x00000400, + 0xbf850033, 0x876eff71, + 0x00000100, 0xbf840002, + 0x8878ff78, 0x00002000, + 0x8a77ff77, 0xff000000, + 0xb96ef807, 0x876fff6e, + 0x02000000, 0x8f6f866f, + 0x88776f77, 0x876fff6e, + 0x003f8000, 0x8f6f896f, + 0x88776f77, 0x8a6eff6e, + 0x023f8000, 0xb9eef807, + 0xb97af812, 0xb97bf813, + 0x8ffa887a, 0xf4051bbd, + 0xfa000000, 0xbf8cc07f, + 0xf4051ebd, 0xfa000008, + 0xbf8cc07f, 0x87ee6e6e, + 0xbf840001, 0xbe80206e, + 0xb971f803, 0x8771ff71, + 0x000001ff, 0xbf850002, + 0x806c846c, 0x826d806d, + 0x876dff6d, 0x0000ffff, + 0x906e8977, 0x876fff6e, + 0x003f8000, 0x906e8677, + 0x876eff6e, 0x02000000, + 0x886e6f6e, 0xb9eef807, + 0x87fe7e7e, 0x87ea6a6a, + 0xb9f8f802, 0xbe80226c, + 0xb971f803, 0x8771ff71, + 0x00000100, 0xbf840006, + 0xbef60380, 0xb9f60203, 0x876dff6d, 0x0000ffff, - 0xbe80226c, 0xb971f803, - 0x8771ff71, 0x00000100, - 0xbf840006, 0xbef60380, - 0xb9f60203, 0x876dff6d, - 0x0000ffff, 0x80ec886c, - 0x82ed806d, 0xbef60380, - 0xb9f60283, 0xb973f816, - 0xb9762c07, 0x8f769c76, - 0x886d766d, 0xb97603c7, - 0x8f769b76, 0x886d766d, + 0x80ec886c, 0x82ed806d, + 0xbef60380, 0xb9f60283, + 0xb972f816, 0xb9762c07, + 0x8f769a76, 0x886d766d, + 0xb97603c7, 0x8f769976, + 0x886d766d, 0xb9760647, + 0x8f769876, 0x886d766d, 0xb976f807, 0x8776ff76, 0x00007fff, 0xb9f6f807, 0xbeee037e, 0xbeef037f, @@ -589,274 +729,834 @@ static const uint32_t cwsr_trap_gfx10_hex[] = { 0xbef4037e, 0x8775ff7f, 0x0000ffff, 0x8875ff75, 0x00040000, 0xbef60380, - 0xbef703ff, 0x00807fac, + 0xbef703ff, 0x10807fac, 0x8776ff7f, 0x08000000, 0x90768376, 0x88777677, 0x8776ff7f, 0x70000000, 0x90768176, 0x88777677, 0xbefb037c, 0xbefa0380, - 0xb97202dc, 0x8872727f, - 0xbefe03c1, 0x877c8172, - 0xbf06817c, 0xbf850002, - 0xbeff0380, 0xbf820001, - 0xbeff03c1, 0xb9712a05, - 0x80718171, 0x8f718271, - 0x877c8172, 0xbf06817c, - 0xbf85000d, 0x8f768771, + 0xb97302dc, 0x8f739973, + 0x8873737f, 0xb97a2a05, + 0x807a817a, 0x907c9973, + 0x877c817c, 0xbf06817c, + 0xbf850002, 0x8f7a897a, + 0xbf820001, 0x8f7a8a7a, + 0xb9761e06, 0x8f768a76, + 0x807a767a, 0x807aff7a, + 0x00000200, 0xbef603ff, + 0x01000000, 0xbefe037c, + 0xbefc037a, 0xf4611efa, + 0xf8000000, 0x807a847a, + 0xbefc037e, 0xbefe037c, + 0xbefc037a, 0xf4611b3a, + 0xf8000000, 0x807a847a, + 0xbefc037e, 0xbefe037c, + 0xbefc037a, 0xf4611b7a, + 0xf8000000, 0x807a847a, + 0xbefc037e, 0xbefe037c, + 0xbefc037a, 0xf4611bba, + 0xf8000000, 0x807a847a, + 0xbefc037e, 0xbefe037c, + 0xbefc037a, 0xf4611bfa, + 0xf8000000, 0x807a847a, + 0xbefc037e, 0xbefe037c, + 0xbefc037a, 0xf4611e3a, + 0xf8000000, 0x807a847a, + 0xbefc037e, 0xb971f803, + 0xbefe037c, 0xbefc037a, + 0xf4611c7a, 0xf8000000, + 0x807a847a, 0xbefc037e, + 0xbefe037c, 0xbefc037a, + 0xf4611cba, 0xf8000000, + 0x807a847a, 0xbefc037e, + 0xb97bf801, 0xbefe037c, + 0xbefc037a, 0xf4611efa, + 0xf8000000, 0x807a847a, + 0xbefc037e, 0xb97bf814, + 0xbefe037c, 0xbefc037a, + 0xf4611efa, 0xf8000000, + 0x807a847a, 0xbefc037e, + 0xb97bf815, 0xbefe037c, + 0xbefc037a, 0xf4611efa, + 0xf8000000, 0x807a847a, + 0xbefc037e, 0x8776ff7f, + 0x04000000, 0xbeef0380, + 0x886f6f76, 0xb97a2a05, + 0x807a817a, 0x907c9973, + 0x877c817c, 0xbf06817c, + 0xbf850002, 0x8f7a897a, + 0xbf820001, 0x8f7a8a7a, + 0xb9761e06, 0x8f768a76, + 0x807a767a, 0xbef603ff, + 0x01000000, 0xbef20374, + 0x80747a74, 0x82758075, + 0xbefc0380, 0xbf800000, + 0xbe802f00, 0xbe822f02, + 0xbe842f04, 0xbe862f06, + 0xbe882f08, 0xbe8a2f0a, + 0xbe8c2f0c, 0xbe8e2f0e, + 0xf469003a, 0xfa000000, + 0xf469013a, 0xfa000010, + 0xf469023a, 0xfa000020, + 0xf469033a, 0xfa000030, + 0x8074c074, 0x82758075, + 0x807c907c, 0xbf0aff7c, + 0x00000060, 0xbf85ffea, + 0xbe802f00, 0xbe822f02, + 0xbe842f04, 0xbe862f06, + 0xbe882f08, 0xbe8a2f0a, + 0xf469003a, 0xfa000000, + 0xf469013a, 0xfa000010, + 0xf469023a, 0xfa000020, + 0x8074b074, 0x82758075, + 0xbef40372, 0xbefa0380, + 0xbefe03c1, 0x907c9973, + 0x877c817c, 0xbf06817c, + 0xbf850002, 0xbeff0380, + 0xbf820002, 0xbeff03c1, + 0xbf82000b, 0xbef603ff, + 0x01000000, 0xe0704000, + 0x7a5d0000, 0xe0704080, + 0x7a5d0100, 0xe0704100, + 0x7a5d0200, 0xe0704180, + 0x7a5d0300, 0xbf82000a, 0xbef603ff, 0x01000000, - 0xbefc0380, 0x7e008700, - 0xe0704000, 0x7a5d0000, - 0x807c817c, 0x807aff7a, - 0x00000080, 0xbf0a717c, - 0xbf85fff8, 0xbf82001b, - 0x8f768871, 0xbef603ff, - 0x01000000, 0xbefc0380, - 0x7e008700, 0xe0704000, - 0x7a5d0000, 0x807c817c, - 0x807aff7a, 0x00000100, - 0xbf0a717c, 0xbf85fff8, - 0xb9711e06, 0x8771c171, - 0xbf84000c, 0x8f718371, - 0x80717c71, 0xbefe03c1, - 0xbeff0380, 0x7e008700, 0xe0704000, 0x7a5d0000, - 0x807c817c, 0x807aff7a, - 0x00000080, 0xbf0a717c, - 0xbf85fff8, 0xbf8a0000, - 0x8776ff72, 0x04000000, - 0xbf84002b, 0xbefe03c1, - 0x877c8172, 0xbf06817c, + 0xe0704100, 0x7a5d0100, + 0xe0704200, 0x7a5d0200, + 0xe0704300, 0x7a5d0300, + 0xbefe03c1, 0x907c9973, + 0x877c817c, 0xbf06817c, 0xbf850002, 0xbeff0380, 0xbf820001, 0xbeff03c1, 0xb9714306, 0x8771c171, - 0xbf840021, 0x8f718671, + 0xbf840046, 0xbf8a0000, + 0x8776ff6f, 0x04000000, + 0xbf840042, 0x8f718671, 0x8f718271, 0xbef60371, + 0xb97a2a05, 0x807a817a, + 0x907c9973, 0x877c817c, + 0xbf06817c, 0xbf850002, + 0x8f7a897a, 0xbf820001, + 0x8f7a8a7a, 0xb9761e06, + 0x8f768a76, 0x807a767a, + 0x807aff7a, 0x00000200, + 0x807aff7a, 0x00000080, 0xbef603ff, 0x01000000, 0xd7650000, 0x000100c1, 0xd7660000, 0x000200c1, - 0x16000084, 0x877c8172, - 0xbf06817c, 0xbefc0380, - 0xbf85000a, 0x807cff7c, - 0x00000080, 0x807aff7a, - 0x00000080, 0xd5250000, - 0x0001ff00, 0x00000080, - 0xbf0a717c, 0xbf85fff7, - 0xbf820009, 0x807cff7c, - 0x00000100, 0x807aff7a, - 0x00000100, 0xd5250000, - 0x0001ff00, 0x00000100, - 0xbf0a717c, 0xbf85fff7, - 0x877c8172, 0xbf06817c, - 0xbf850003, 0x8f7687ff, - 0x0000006a, 0xbf820002, - 0x8f7688ff, 0x0000006a, + 0x16000084, 0x907c9973, + 0x877c817c, 0xbf06817c, + 0xbefc0380, 0xbf850012, + 0xbe8303ff, 0x00000080, + 0xbf800000, 0xbf800000, + 0xbf800000, 0xd8d80000, + 0x01000000, 0xbf8c0000, + 0xe0704000, 0x7a5d0100, + 0x807c037c, 0x807a037a, + 0xd5250000, 0x0001ff00, + 0x00000080, 0xbf0a717c, + 0xbf85fff4, 0xbf820011, + 0xbe8303ff, 0x00000100, + 0xbf800000, 0xbf800000, + 0xbf800000, 0xd8d80000, + 0x01000000, 0xbf8c0000, + 0xe0704000, 0x7a5d0100, + 0x807c037c, 0x807a037a, + 0xd5250000, 0x0001ff00, + 0x00000100, 0xbf0a717c, + 0xbf85fff4, 0xbefe03c1, + 0x907c9973, 0x877c817c, + 0xbf06817c, 0xbf850004, + 0xbefa03ff, 0x00000200, + 0xbeff0380, 0xbf820003, + 0xbefa03ff, 0x00000400, + 0xbeff03c1, 0xb9712a05, + 0x80718171, 0x8f718271, + 0x907c9973, 0x877c817c, + 0xbf06817c, 0xbf850017, 0xbef603ff, 0x01000000, - 0x877c8172, 0xbf06817c, - 0xbefc0380, 0xbf800000, - 0xbf85000b, 0xbe802e00, - 0x7e000200, 0xe0704000, - 0x7a5d0000, 0x807aff7a, - 0x00000080, 0x807c817c, - 0xbf0aff7c, 0x0000006a, - 0xbf85fff6, 0xbf82000a, - 0xbe802e00, 0x7e000200, - 0xe0704000, 0x7a5d0000, - 0x807aff7a, 0x00000100, - 0x807c817c, 0xbf0aff7c, - 0x0000006a, 0xbf85fff6, - 0xbef60384, 0xbef603ff, - 0x01000000, 0x877c8172, - 0xbf06817c, 0xbf850030, - 0x7e00027b, 0xe0704000, - 0x7a5d0000, 0x807aff7a, - 0x00000080, 0x7e00026c, - 0xe0704000, 0x7a5d0000, - 0x807aff7a, 0x00000080, - 0x7e00026d, 0xe0704000, - 0x7a5d0000, 0x807aff7a, - 0x00000080, 0x7e00026e, - 0xe0704000, 0x7a5d0000, - 0x807aff7a, 0x00000080, - 0x7e00026f, 0xe0704000, - 0x7a5d0000, 0x807aff7a, - 0x00000080, 0x7e000270, - 0xe0704000, 0x7a5d0000, - 0x807aff7a, 0x00000080, - 0xb971f803, 0x7e000271, + 0xbefc0384, 0xbf0a717c, + 0xbf840037, 0x7e008700, + 0x7e028701, 0x7e048702, + 0x7e068703, 0xe0704000, + 0x7a5d0000, 0xe0704080, + 0x7a5d0100, 0xe0704100, + 0x7a5d0200, 0xe0704180, + 0x7a5d0300, 0x807c847c, + 0x807aff7a, 0x00000200, + 0xbf0a717c, 0xbf85ffef, + 0xbf820025, 0xbef603ff, + 0x01000000, 0xbefc0384, + 0xbf0a717c, 0xbf840020, + 0x7e008700, 0x7e028701, + 0x7e048702, 0x7e068703, 0xe0704000, 0x7a5d0000, + 0xe0704100, 0x7a5d0100, + 0xe0704200, 0x7a5d0200, + 0xe0704300, 0x7a5d0300, + 0x807c847c, 0x807aff7a, + 0x00000400, 0xbf0a717c, + 0xbf85ffef, 0xb9711e06, + 0x8771c171, 0xbf84000c, + 0x8f718371, 0x80717c71, + 0xbefe03c1, 0xbeff0380, + 0x7e008700, 0xe0704000, + 0x7a5d0000, 0x807c817c, 0x807aff7a, 0x00000080, - 0x7e000273, 0xe0704000, - 0x7a5d0000, 0x807aff7a, - 0x00000080, 0xb97bf801, - 0x7e00027b, 0xe0704000, - 0x7a5d0000, 0x807aff7a, - 0x00000080, 0xbf82002f, - 0x7e00027b, 0xe0704000, - 0x7a5d0000, 0x807aff7a, - 0x00000100, 0x7e00026c, - 0xe0704000, 0x7a5d0000, - 0x807aff7a, 0x00000100, - 0x7e00026d, 0xe0704000, - 0x7a5d0000, 0x807aff7a, - 0x00000100, 0x7e00026e, - 0xe0704000, 0x7a5d0000, - 0x807aff7a, 0x00000100, - 0x7e00026f, 0xe0704000, - 0x7a5d0000, 0x807aff7a, - 0x00000100, 0x7e000270, - 0xe0704000, 0x7a5d0000, - 0x807aff7a, 0x00000100, - 0xb971f803, 0x7e000271, - 0xe0704000, 0x7a5d0000, - 0x807aff7a, 0x00000100, - 0x7e000273, 0xe0704000, - 0x7a5d0000, 0x807aff7a, - 0x00000100, 0xb97bf801, - 0x7e00027b, 0xe0704000, - 0x7a5d0000, 0x807aff7a, - 0x00000100, 0xbf820119, - 0xbef4037e, 0x8775ff7f, - 0x0000ffff, 0x8875ff75, - 0x00040000, 0xbef60380, - 0xbef703ff, 0x00807fac, - 0x8772ff7f, 0x08000000, - 0x90728372, 0x88777277, - 0x8772ff7f, 0x70000000, - 0x90728172, 0x88777277, - 0xb97902dc, 0x8879797f, - 0xbef80380, 0xbefe03c1, - 0x877c8179, 0xbf06817c, + 0xbf0a717c, 0xbf85fff8, + 0xbf820142, 0xbef4037e, + 0x8775ff7f, 0x0000ffff, + 0x8875ff75, 0x00040000, + 0xbef60380, 0xbef703ff, + 0x10807fac, 0x8772ff7f, + 0x08000000, 0x90728372, + 0x88777277, 0x8772ff7f, + 0x70000000, 0x90728172, + 0x88777277, 0xb97302dc, + 0x8f739973, 0x8873737f, + 0x8772ff7f, 0x04000000, + 0xbf840036, 0xbefe03c1, + 0x907c9973, 0x877c817c, + 0xbf06817c, 0xbf850002, + 0xbeff0380, 0xbf820001, + 0xbeff03c1, 0xb96f4306, + 0x876fc16f, 0xbf84002b, + 0x8f6f866f, 0x8f6f826f, + 0xbef6036f, 0xb9782a05, + 0x80788178, 0x907c9973, + 0x877c817c, 0xbf06817c, + 0xbf850002, 0x8f788978, + 0xbf820001, 0x8f788a78, + 0xb9721e06, 0x8f728a72, + 0x80787278, 0x8078ff78, + 0x00000200, 0x8078ff78, + 0x00000080, 0xbef603ff, + 0x01000000, 0x907c9973, + 0x877c817c, 0xbf06817c, + 0xbefc0380, 0xbf850009, + 0xe0310000, 0x781d0000, + 0x807cff7c, 0x00000080, + 0x8078ff78, 0x00000080, + 0xbf0a6f7c, 0xbf85fff8, + 0xbf820008, 0xe0310000, + 0x781d0000, 0x807cff7c, + 0x00000100, 0x8078ff78, + 0x00000100, 0xbf0a6f7c, + 0xbf85fff8, 0xbef80380, + 0xbefe03c1, 0x907c9973, + 0x877c817c, 0xbf06817c, 0xbf850002, 0xbeff0380, 0xbf820001, 0xbeff03c1, 0xb96f2a05, 0x806f816f, - 0x8f6f826f, 0x877c8179, - 0xbf06817c, 0xbf850013, - 0x8f76876f, 0xbef603ff, + 0x8f6f826f, 0x907c9973, + 0x877c817c, 0xbf06817c, + 0xbf850021, 0xbef603ff, 0x01000000, 0xbef20378, - 0x8078ff78, 0x00000080, - 0xbefc0381, 0xe0304000, - 0x785d0000, 0xbf8c3f70, - 0x7e008500, 0x807c817c, - 0x8078ff78, 0x00000080, - 0xbf0a6f7c, 0xbf85fff7, - 0xe0304000, 0x725d0000, - 0xbf820023, 0x8f76886f, + 0x8078ff78, 0x00000200, + 0xbefc0384, 0xe0304000, + 0x785d0000, 0xe0304080, + 0x785d0100, 0xe0304100, + 0x785d0200, 0xe0304180, + 0x785d0300, 0xbf8c3f70, + 0x7e008500, 0x7e028501, + 0x7e048502, 0x7e068503, + 0x807c847c, 0x8078ff78, + 0x00000200, 0xbf0a6f7c, + 0xbf85ffee, 0xe0304000, + 0x725d0000, 0xe0304080, + 0x725d0100, 0xe0304100, + 0x725d0200, 0xe0304180, + 0x725d0300, 0xbf820032, 0xbef603ff, 0x01000000, 0xbef20378, 0x8078ff78, - 0x00000100, 0xbefc0381, - 0xe0304000, 0x785d0000, - 0xbf8c3f70, 0x7e008500, - 0x807c817c, 0x8078ff78, - 0x00000100, 0xbf0a6f7c, - 0xbf85fff7, 0xb96f1e06, - 0x876fc16f, 0xbf84000e, - 0x8f6f836f, 0x806f7c6f, - 0xbefe03c1, 0xbeff0380, + 0x00000400, 0xbefc0384, 0xe0304000, 0x785d0000, + 0xe0304100, 0x785d0100, + 0xe0304200, 0x785d0200, + 0xe0304300, 0x785d0300, 0xbf8c3f70, 0x7e008500, - 0x807c817c, 0x8078ff78, - 0x00000080, 0xbf0a6f7c, - 0xbf85fff7, 0xbeff03c1, - 0xe0304000, 0x725d0000, - 0x8772ff79, 0x04000000, - 0xbf840020, 0xbefe03c1, - 0x877c8179, 0xbf06817c, - 0xbf850002, 0xbeff0380, - 0xbf820001, 0xbeff03c1, - 0xb96f4306, 0x876fc16f, - 0xbf840016, 0x8f6f866f, - 0x8f6f826f, 0xbef6036f, - 0xbef603ff, 0x01000000, - 0x877c8172, 0xbf06817c, - 0xbefc0380, 0xbf850007, - 0x807cff7c, 0x00000080, - 0x8078ff78, 0x00000080, - 0xbf0a6f7c, 0xbf85fffa, - 0xbf820006, 0x807cff7c, - 0x00000100, 0x8078ff78, - 0x00000100, 0xbf0a6f7c, - 0xbf85fffa, 0x877c8179, - 0xbf06817c, 0xbf850003, - 0x8f7687ff, 0x0000006a, - 0xbf820002, 0x8f7688ff, - 0x0000006a, 0xbef603ff, - 0x01000000, 0x877c8179, - 0xbf06817c, 0xbf850012, - 0xf4211cba, 0xf0000000, + 0x7e028501, 0x7e048502, + 0x7e068503, 0x807c847c, + 0x8078ff78, 0x00000400, + 0xbf0a6f7c, 0xbf85ffee, + 0xb96f1e06, 0x876fc16f, + 0xbf84000e, 0x8f6f836f, + 0x806f7c6f, 0xbefe03c1, + 0xbeff0380, 0xe0304000, + 0x785d0000, 0xbf8c3f70, + 0x7e008500, 0x807c817c, 0x8078ff78, 0x00000080, - 0xbefc0381, 0xf421003a, - 0xf0000000, 0x8078ff78, - 0x00000080, 0xbf8cc07f, - 0xbe803000, 0xbf800000, - 0x807c817c, 0xbf0aff7c, - 0x0000006a, 0xbf85fff5, - 0xbe800372, 0xbf820011, - 0xf4211cba, 0xf0000000, - 0x8078ff78, 0x00000100, - 0xbefc0381, 0xf421003a, - 0xf0000000, 0x8078ff78, - 0x00000100, 0xbf8cc07f, - 0xbe803000, 0xbf800000, - 0x807c817c, 0xbf0aff7c, - 0x0000006a, 0xbf85fff5, - 0xbe800372, 0xbef60384, + 0xbf0a6f7c, 0xbf85fff7, + 0xbeff03c1, 0xe0304000, + 0x725d0000, 0xe0304100, + 0x725d0100, 0xe0304200, + 0x725d0200, 0xe0304300, + 0x725d0300, 0xbf8c3f70, + 0xb9782a05, 0x80788178, + 0x907c9973, 0x877c817c, + 0xbf06817c, 0xbf850002, + 0x8f788978, 0xbf820001, + 0x8f788a78, 0xb9721e06, + 0x8f728a72, 0x80787278, + 0x8078ff78, 0x00000200, + 0x80f8ff78, 0x00000050, + 0xbef603ff, 0x01000000, + 0xbefc03ff, 0x0000006c, + 0x80f89078, 0xf429003a, + 0xf0000000, 0xbf8cc07f, + 0x80fc847c, 0xbf800000, + 0xbe803100, 0xbe823102, + 0x80f8a078, 0xf42d003a, + 0xf0000000, 0xbf8cc07f, + 0x80fc887c, 0xbf800000, + 0xbe803100, 0xbe823102, + 0xbe843104, 0xbe863106, + 0x80f8c078, 0xf431003a, + 0xf0000000, 0xbf8cc07f, + 0x80fc907c, 0xbf800000, + 0xbe803100, 0xbe823102, + 0xbe843104, 0xbe863106, + 0xbe883108, 0xbe8a310a, + 0xbe8c310c, 0xbe8e310e, + 0xbf06807c, 0xbf84fff0, + 0xb9782a05, 0x80788178, + 0x907c9973, 0x877c817c, + 0xbf06817c, 0xbf850002, + 0x8f788978, 0xbf820001, + 0x8f788a78, 0xb9721e06, + 0x8f728a72, 0x80787278, + 0x8078ff78, 0x00000200, 0xbef603ff, 0x01000000, - 0x877c8179, 0xbf06817c, - 0xbf850025, 0xf4211bfa, - 0xf0000000, 0x8078ff78, - 0x00000080, 0xf4211b3a, - 0xf0000000, 0x8078ff78, - 0x00000080, 0xf4211b7a, - 0xf0000000, 0x8078ff78, - 0x00000080, 0xf4211eba, - 0xf0000000, 0x8078ff78, - 0x00000080, 0xf4211efa, - 0xf0000000, 0x8078ff78, - 0x00000080, 0xf4211c3a, - 0xf0000000, 0x8078ff78, - 0x00000080, 0xf4211c7a, - 0xf0000000, 0x8078ff78, - 0x00000080, 0xf4211cfa, - 0xf0000000, 0x8078ff78, - 0x00000080, 0xf4211e7a, - 0xf0000000, 0x8078ff78, - 0x00000080, 0xbf820024, 0xf4211bfa, 0xf0000000, - 0x8078ff78, 0x00000100, - 0xf4211b3a, 0xf0000000, - 0x8078ff78, 0x00000100, + 0x80788478, 0xf4211b3a, + 0xf0000000, 0x80788478, 0xf4211b7a, 0xf0000000, - 0x8078ff78, 0x00000100, - 0xf4211eba, 0xf0000000, - 0x8078ff78, 0x00000100, + 0x80788478, 0xf4211eba, + 0xf0000000, 0x80788478, 0xf4211efa, 0xf0000000, - 0x8078ff78, 0x00000100, - 0xf4211c3a, 0xf0000000, - 0x8078ff78, 0x00000100, + 0x80788478, 0xf4211c3a, + 0xf0000000, 0x80788478, 0xf4211c7a, 0xf0000000, - 0x8078ff78, 0x00000100, + 0x80788478, 0xf4211e7a, + 0xf0000000, 0x80788478, 0xf4211cfa, 0xf0000000, - 0x8078ff78, 0x00000100, - 0xf4211e7a, 0xf0000000, - 0x8078ff78, 0x00000100, - 0xbf8cc07f, 0x876dff6d, - 0x0000ffff, 0xbefc036f, - 0xbefe037a, 0xbeff037b, - 0x876f71ff, 0x000003ff, - 0xb9ef4803, 0xb9f3f816, - 0x876f71ff, 0xfffff800, - 0x906f8b6f, 0xb9efa2c3, - 0xb9f9f801, 0x876fff6d, - 0xf0000000, 0x906f9c6f, - 0x8f6f906f, 0xbef20380, - 0x88726f72, 0x876fff6d, - 0x08000000, 0x906f9b6f, - 0x8f6f8f6f, 0x88726f72, + 0x80788478, 0xf4211bba, + 0xf0000000, 0x80788478, + 0xbf8cc07f, 0xb9eef814, + 0xf4211bba, 0xf0000000, + 0x80788478, 0xbf8cc07f, + 0xb9eef815, 0xbef2036d, + 0x876dff72, 0x0000ffff, + 0xbefc036f, 0xbefe037a, + 0xbeff037b, 0x876f71ff, + 0x000003ff, 0xb9ef4803, + 0xb9f9f816, 0x876f71ff, + 0xfffff800, 0x906f8b6f, + 0xb9efa2c3, 0xb9f3f801, + 0x876fff72, 0xfc000000, + 0x906f9a6f, 0x8f6f906f, + 0xbef30380, 0x88736f73, + 0x876fff72, 0x02000000, + 0x906f996f, 0x8f6f8f6f, + 0x88736f73, 0x876fff72, + 0x01000000, 0x906f986f, + 0x8f6f996f, 0x88736f73, 0x876fff70, 0x00800000, - 0x906f976f, 0xb9f2f807, + 0x906f976f, 0xb9f3f807, + 0x87fe7e7e, 0x87ea6a6a, 0xb9f0f802, 0xbf8a0000, 0xbe80226c, 0xbf810000, 0xbf9f0000, 0xbf9f0000, 0xbf9f0000, 0xbf9f0000, 0xbf9f0000, 0x00000000, }; +static const uint32_t cwsr_trap_arcturus_hex[] = { + 0xbf820001, 0xbf8202c4, + 0xb8f8f802, 0x89788678, + 0xb8eef801, 0x866eff6e, + 0x00000800, 0xbf840003, + 0x866eff78, 0x00002000, + 0xbf840016, 0xb8fbf803, + 0x866eff7b, 0x00000400, + 0xbf85003b, 0x866eff7b, + 0x00000800, 0xbf850003, + 0x866eff7b, 0x00000100, + 0xbf84000c, 0x866eff78, + 0x00002000, 0xbf840005, + 0xbf8e0010, 0xb8eef803, + 0x866eff6e, 0x00000400, + 0xbf84fffb, 0x8778ff78, + 0x00002000, 0x80ec886c, + 0x82ed806d, 0xb8eef807, + 0x866fff6e, 0x001f8000, + 0x8e6f8b6f, 0x8977ff77, + 0xfc000000, 0x87776f77, + 0x896eff6e, 0x001f8000, + 0xb96ef807, 0xb8faf812, + 0xb8fbf813, 0x8efa887a, + 0xc0071bbd, 0x00000000, + 0xbf8cc07f, 0xc0071ebd, + 0x00000008, 0xbf8cc07f, + 0x86ee6e6e, 0xbf840001, + 0xbe801d6e, 0xb8fbf803, + 0x867bff7b, 0x000001ff, + 0xbf850002, 0x806c846c, + 0x826d806d, 0x866dff6d, + 0x0000ffff, 0x8f6e8b77, + 0x866eff6e, 0x001f8000, + 0xb96ef807, 0x86fe7e7e, + 0x86ea6a6a, 0x8f6e8378, + 0xb96ee0c2, 0xbf800002, + 0xb9780002, 0xbe801f6c, + 0x866dff6d, 0x0000ffff, + 0xbefa0080, 0xb97a0283, + 0xb8fa2407, 0x8e7a9b7a, + 0x876d7a6d, 0xb8fa03c7, + 0x8e7a9a7a, 0x876d7a6d, + 0xb8faf807, 0x867aff7a, + 0x00007fff, 0xb97af807, + 0xbeee007e, 0xbeef007f, + 0xbefe0180, 0xbf900004, + 0x877a8478, 0xb97af802, + 0xbf8e0002, 0xbf88fffe, + 0xb8fa2a05, 0x807a817a, + 0x8e7a8a7a, 0x8e7a817a, + 0xb8fb1605, 0x807b817b, + 0x8e7b867b, 0x807a7b7a, + 0x807a7e7a, 0x827b807f, + 0x867bff7b, 0x0000ffff, + 0xc04b1c3d, 0x00000050, + 0xbf8cc07f, 0xc04b1d3d, + 0x00000060, 0xbf8cc07f, + 0xc0431e7d, 0x00000074, + 0xbf8cc07f, 0xbef4007e, + 0x8675ff7f, 0x0000ffff, + 0x8775ff75, 0x00040000, + 0xbef60080, 0xbef700ff, + 0x00807fac, 0x867aff7f, + 0x08000000, 0x8f7a837a, + 0x87777a77, 0x867aff7f, + 0x70000000, 0x8f7a817a, + 0x87777a77, 0xbef1007c, + 0xbef00080, 0xb8f02a05, + 0x80708170, 0x8e708a70, + 0x8e708170, 0xb8fa1605, + 0x807a817a, 0x8e7a867a, + 0x80707a70, 0xbef60084, + 0xbef600ff, 0x01000000, + 0xbefe007c, 0xbefc0070, + 0xc0611c7a, 0x0000007c, + 0xbf8cc07f, 0x80708470, + 0xbefc007e, 0xbefe007c, + 0xbefc0070, 0xc0611b3a, + 0x0000007c, 0xbf8cc07f, + 0x80708470, 0xbefc007e, + 0xbefe007c, 0xbefc0070, + 0xc0611b7a, 0x0000007c, + 0xbf8cc07f, 0x80708470, + 0xbefc007e, 0xbefe007c, + 0xbefc0070, 0xc0611bba, + 0x0000007c, 0xbf8cc07f, + 0x80708470, 0xbefc007e, + 0xbefe007c, 0xbefc0070, + 0xc0611bfa, 0x0000007c, + 0xbf8cc07f, 0x80708470, + 0xbefc007e, 0xbefe007c, + 0xbefc0070, 0xc0611e3a, + 0x0000007c, 0xbf8cc07f, + 0x80708470, 0xbefc007e, + 0xb8fbf803, 0xbefe007c, + 0xbefc0070, 0xc0611efa, + 0x0000007c, 0xbf8cc07f, + 0x80708470, 0xbefc007e, + 0xbefe007c, 0xbefc0070, + 0xc0611a3a, 0x0000007c, + 0xbf8cc07f, 0x80708470, + 0xbefc007e, 0xbefe007c, + 0xbefc0070, 0xc0611a7a, + 0x0000007c, 0xbf8cc07f, + 0x80708470, 0xbefc007e, + 0xb8f1f801, 0xbefe007c, + 0xbefc0070, 0xc0611c7a, + 0x0000007c, 0xbf8cc07f, + 0x80708470, 0xbefc007e, + 0x867aff7f, 0x04000000, + 0xbeef0080, 0x876f6f7a, + 0xb8f02a05, 0x80708170, + 0x8e708a70, 0x8e708170, + 0xb8fb1605, 0x807b817b, + 0x8e7b847b, 0x8e76827b, + 0xbef600ff, 0x01000000, + 0xbef20174, 0x80747074, + 0x82758075, 0xbefc0080, + 0xbf800000, 0xbe802b00, + 0xbe822b02, 0xbe842b04, + 0xbe862b06, 0xbe882b08, + 0xbe8a2b0a, 0xbe8c2b0c, + 0xbe8e2b0e, 0xc06b003a, + 0x00000000, 0xbf8cc07f, + 0xc06b013a, 0x00000010, + 0xbf8cc07f, 0xc06b023a, + 0x00000020, 0xbf8cc07f, + 0xc06b033a, 0x00000030, + 0xbf8cc07f, 0x8074c074, + 0x82758075, 0x807c907c, + 0xbf0a7b7c, 0xbf85ffe7, + 0xbef40172, 0xbef00080, + 0xbefe00c1, 0xbeff00c1, + 0xbee80080, 0xbee90080, + 0xbef600ff, 0x01000000, + 0x867aff78, 0x00400000, + 0xbf850003, 0xb8faf803, + 0x897a7aff, 0x10000000, + 0xbf85004d, 0xbe840080, + 0xd2890000, 0x00000900, + 0x80048104, 0xd2890001, + 0x00000900, 0x80048104, + 0xd2890002, 0x00000900, + 0x80048104, 0xd2890003, + 0x00000900, 0x80048104, + 0xc069003a, 0x00000070, + 0xbf8cc07f, 0x80709070, + 0xbf06c004, 0xbf84ffee, + 0xbe840080, 0xd2890000, + 0x00000901, 0x80048104, + 0xd2890001, 0x00000901, + 0x80048104, 0xd2890002, + 0x00000901, 0x80048104, + 0xd2890003, 0x00000901, + 0x80048104, 0xc069003a, + 0x00000070, 0xbf8cc07f, + 0x80709070, 0xbf06c004, + 0xbf84ffee, 0xbe840080, + 0xd2890000, 0x00000902, + 0x80048104, 0xd2890001, + 0x00000902, 0x80048104, + 0xd2890002, 0x00000902, + 0x80048104, 0xd2890003, + 0x00000902, 0x80048104, + 0xc069003a, 0x00000070, + 0xbf8cc07f, 0x80709070, + 0xbf06c004, 0xbf84ffee, + 0xbe840080, 0xd2890000, + 0x00000903, 0x80048104, + 0xd2890001, 0x00000903, + 0x80048104, 0xd2890002, + 0x00000903, 0x80048104, + 0xd2890003, 0x00000903, + 0x80048104, 0xc069003a, + 0x00000070, 0xbf8cc07f, + 0x80709070, 0xbf06c004, + 0xbf84ffee, 0xbf820008, + 0xe0724000, 0x701d0000, + 0xe0724100, 0x701d0100, + 0xe0724200, 0x701d0200, + 0xe0724300, 0x701d0300, + 0xbefe00c1, 0xbeff00c1, + 0xb8fb4306, 0x867bc17b, + 0xbf840064, 0xbf8a0000, + 0x867aff6f, 0x04000000, + 0xbf840060, 0x8e7b867b, + 0x8e7b827b, 0xbef6007b, + 0xb8f02a05, 0x80708170, + 0x8e708a70, 0x8e708170, + 0xb8fa1605, 0x807a817a, + 0x8e7a867a, 0x80707a70, + 0x8070ff70, 0x00000080, + 0xbef600ff, 0x01000000, + 0xbefc0080, 0xd28c0002, + 0x000100c1, 0xd28d0003, + 0x000204c1, 0x867aff78, + 0x00400000, 0xbf850003, + 0xb8faf803, 0x897a7aff, + 0x10000000, 0xbf850030, + 0x24040682, 0xd86e4000, + 0x00000002, 0xbf8cc07f, + 0xbe840080, 0xd2890000, + 0x00000900, 0x80048104, + 0xd2890001, 0x00000900, + 0x80048104, 0xd2890002, + 0x00000900, 0x80048104, + 0xd2890003, 0x00000900, + 0x80048104, 0xc069003a, + 0x00000070, 0xbf8cc07f, + 0x80709070, 0xbf06c004, + 0xbf84ffee, 0xbe840080, + 0xd2890000, 0x00000901, + 0x80048104, 0xd2890001, + 0x00000901, 0x80048104, + 0xd2890002, 0x00000901, + 0x80048104, 0xd2890003, + 0x00000901, 0x80048104, + 0xc069003a, 0x00000070, + 0xbf8cc07f, 0x80709070, + 0xbf06c004, 0xbf84ffee, + 0x680404ff, 0x00000200, + 0xd0c9006a, 0x0000f702, + 0xbf87ffd2, 0xbf820015, + 0xd1060002, 0x00011103, + 0x7e0602ff, 0x00000200, + 0xbefc00ff, 0x00010000, + 0xbe800077, 0x8677ff77, + 0xff7fffff, 0x8777ff77, + 0x00058000, 0xd8ec0000, + 0x00000002, 0xbf8cc07f, + 0xe0765000, 0x701d0002, + 0x68040702, 0xd0c9006a, + 0x0000f702, 0xbf87fff7, + 0xbef70000, 0xbef000ff, + 0x00000400, 0xbefe00c1, + 0xbeff00c1, 0xb8fb2a05, + 0x807b817b, 0x8e7b827b, + 0x8e76887b, 0xbef600ff, + 0x01000000, 0xbefc0084, + 0xbf0a7b7c, 0xbf84006d, + 0xbf11017c, 0x807bff7b, + 0x00001000, 0x867aff78, + 0x00400000, 0xbf850003, + 0xb8faf803, 0x897a7aff, + 0x10000000, 0xbf850051, + 0xbe840080, 0xd2890000, + 0x00000900, 0x80048104, + 0xd2890001, 0x00000900, + 0x80048104, 0xd2890002, + 0x00000900, 0x80048104, + 0xd2890003, 0x00000900, + 0x80048104, 0xc069003a, + 0x00000070, 0xbf8cc07f, + 0x80709070, 0xbf06c004, + 0xbf84ffee, 0xbe840080, + 0xd2890000, 0x00000901, + 0x80048104, 0xd2890001, + 0x00000901, 0x80048104, + 0xd2890002, 0x00000901, + 0x80048104, 0xd2890003, + 0x00000901, 0x80048104, + 0xc069003a, 0x00000070, + 0xbf8cc07f, 0x80709070, + 0xbf06c004, 0xbf84ffee, + 0xbe840080, 0xd2890000, + 0x00000902, 0x80048104, + 0xd2890001, 0x00000902, + 0x80048104, 0xd2890002, + 0x00000902, 0x80048104, + 0xd2890003, 0x00000902, + 0x80048104, 0xc069003a, + 0x00000070, 0xbf8cc07f, + 0x80709070, 0xbf06c004, + 0xbf84ffee, 0xbe840080, + 0xd2890000, 0x00000903, + 0x80048104, 0xd2890001, + 0x00000903, 0x80048104, + 0xd2890002, 0x00000903, + 0x80048104, 0xd2890003, + 0x00000903, 0x80048104, + 0xc069003a, 0x00000070, + 0xbf8cc07f, 0x80709070, + 0xbf06c004, 0xbf84ffee, + 0x807c847c, 0xbf0a7b7c, + 0xbf85ffb1, 0xbf9c0000, + 0xbf820012, 0x7e000300, + 0x7e020301, 0x7e040302, + 0x7e060303, 0xe0724000, + 0x701d0000, 0xe0724100, + 0x701d0100, 0xe0724200, + 0x701d0200, 0xe0724300, + 0x701d0300, 0x807c847c, + 0x8070ff70, 0x00000400, + 0xbf0a7b7c, 0xbf85ffef, + 0xbf9c0000, 0xbefc0080, + 0xbf11017c, 0x867aff78, + 0x00400000, 0xbf850003, + 0xb8faf803, 0x897a7aff, + 0x10000000, 0xbf850059, + 0xd3d84000, 0x18000100, + 0xd3d84001, 0x18000101, + 0xd3d84002, 0x18000102, + 0xd3d84003, 0x18000103, + 0xbe840080, 0xd2890000, + 0x00000900, 0x80048104, + 0xd2890001, 0x00000900, + 0x80048104, 0xd2890002, + 0x00000900, 0x80048104, + 0xd2890003, 0x00000900, + 0x80048104, 0xc069003a, + 0x00000070, 0xbf8cc07f, + 0x80709070, 0xbf06c004, + 0xbf84ffee, 0xbe840080, + 0xd2890000, 0x00000901, + 0x80048104, 0xd2890001, + 0x00000901, 0x80048104, + 0xd2890002, 0x00000901, + 0x80048104, 0xd2890003, + 0x00000901, 0x80048104, + 0xc069003a, 0x00000070, + 0xbf8cc07f, 0x80709070, + 0xbf06c004, 0xbf84ffee, + 0xbe840080, 0xd2890000, + 0x00000902, 0x80048104, + 0xd2890001, 0x00000902, + 0x80048104, 0xd2890002, + 0x00000902, 0x80048104, + 0xd2890003, 0x00000902, + 0x80048104, 0xc069003a, + 0x00000070, 0xbf8cc07f, + 0x80709070, 0xbf06c004, + 0xbf84ffee, 0xbe840080, + 0xd2890000, 0x00000903, + 0x80048104, 0xd2890001, + 0x00000903, 0x80048104, + 0xd2890002, 0x00000903, + 0x80048104, 0xd2890003, + 0x00000903, 0x80048104, + 0xc069003a, 0x00000070, + 0xbf8cc07f, 0x80709070, + 0xbf06c004, 0xbf84ffee, + 0x807c847c, 0xbf0a7b7c, + 0xbf85ffa9, 0xbf9c0000, + 0xbf820016, 0xd3d84000, + 0x18000100, 0xd3d84001, + 0x18000101, 0xd3d84002, + 0x18000102, 0xd3d84003, + 0x18000103, 0xe0724000, + 0x701d0000, 0xe0724100, + 0x701d0100, 0xe0724200, + 0x701d0200, 0xe0724300, + 0x701d0300, 0x807c847c, + 0x8070ff70, 0x00000400, + 0xbf0a7b7c, 0xbf85ffeb, + 0xbf9c0000, 0xbf820106, + 0xbef4007e, 0x8675ff7f, + 0x0000ffff, 0x8775ff75, + 0x00040000, 0xbef60080, + 0xbef700ff, 0x00807fac, + 0x866eff7f, 0x08000000, + 0x8f6e836e, 0x87776e77, + 0x866eff7f, 0x70000000, + 0x8f6e816e, 0x87776e77, + 0x866eff7f, 0x04000000, + 0xbf84001f, 0xbefe00c1, + 0xbeff00c1, 0xb8ef4306, + 0x866fc16f, 0xbf84001a, + 0x8e6f866f, 0x8e6f826f, + 0xbef6006f, 0xb8f82a05, + 0x80788178, 0x8e788a78, + 0x8e788178, 0xb8ee1605, + 0x806e816e, 0x8e6e866e, + 0x80786e78, 0x8078ff78, + 0x00000080, 0xbef600ff, + 0x01000000, 0xbefc0080, + 0xe0510000, 0x781d0000, + 0xe0510100, 0x781d0000, + 0x807cff7c, 0x00000200, + 0x8078ff78, 0x00000200, + 0xbf0a6f7c, 0xbf85fff6, + 0xbef80080, 0xbefe00c1, + 0xbeff00c1, 0xb8ef2a05, + 0x806f816f, 0x8e6f826f, + 0x8e76886f, 0xbef90076, + 0xbef600ff, 0x01000000, + 0xbeee0078, 0x8078ff78, + 0x00000400, 0xbef30079, + 0x8079ff79, 0x00000400, + 0xbefc0084, 0xbf11087c, + 0x806fff6f, 0x00008000, + 0xe0524000, 0x791d0000, + 0xe0524100, 0x791d0100, + 0xe0524200, 0x791d0200, + 0xe0524300, 0x791d0300, + 0x8079ff79, 0x00000400, + 0xbf8c0f70, 0xd3d94000, + 0x18000100, 0xd3d94001, + 0x18000101, 0xd3d94002, + 0x18000102, 0xd3d94003, + 0x18000103, 0xe0524000, + 0x781d0000, 0xe0524100, + 0x781d0100, 0xe0524200, + 0x781d0200, 0xe0524300, + 0x781d0300, 0xbf8c0f70, + 0x7e000300, 0x7e020301, + 0x7e040302, 0x7e060303, + 0x807c847c, 0x8078ff78, + 0x00000400, 0xbf0a6f7c, + 0xbf85ffdb, 0xbf9c0000, + 0xe0524000, 0x731d0000, + 0xe0524100, 0x731d0100, + 0xe0524200, 0x731d0200, + 0xe0524300, 0x731d0300, + 0xbf8c0f70, 0xd3d94000, + 0x18000100, 0xd3d94001, + 0x18000101, 0xd3d94002, + 0x18000102, 0xd3d94003, + 0x18000103, 0xe0524000, + 0x6e1d0000, 0xe0524100, + 0x6e1d0100, 0xe0524200, + 0x6e1d0200, 0xe0524300, + 0x6e1d0300, 0xb8f82a05, + 0x80788178, 0x8e788a78, + 0x8e788178, 0xb8ee1605, + 0x806e816e, 0x8e6e866e, + 0x80786e78, 0x80f8c078, + 0xb8ef1605, 0x806f816f, + 0x8e6f846f, 0x8e76826f, + 0xbef600ff, 0x01000000, + 0xbefc006f, 0xc031003a, + 0x00000078, 0x80f8c078, + 0xbf8cc07f, 0x80fc907c, + 0xbf800000, 0xbe802d00, + 0xbe822d02, 0xbe842d04, + 0xbe862d06, 0xbe882d08, + 0xbe8a2d0a, 0xbe8c2d0c, + 0xbe8e2d0e, 0xbf06807c, + 0xbf84fff0, 0xb8f82a05, + 0x80788178, 0x8e788a78, + 0x8e788178, 0xb8ee1605, + 0x806e816e, 0x8e6e866e, + 0x80786e78, 0xbef60084, + 0xbef600ff, 0x01000000, + 0xc0211bfa, 0x00000078, + 0x80788478, 0xc0211b3a, + 0x00000078, 0x80788478, + 0xc0211b7a, 0x00000078, + 0x80788478, 0xc0211c3a, + 0x00000078, 0x80788478, + 0xc0211c7a, 0x00000078, + 0x80788478, 0xc0211eba, + 0x00000078, 0x80788478, + 0xc0211efa, 0x00000078, + 0x80788478, 0xc0211a3a, + 0x00000078, 0x80788478, + 0xc0211a7a, 0x00000078, + 0x80788478, 0xc0211cfa, + 0x00000078, 0x80788478, + 0xbf8cc07f, 0xbefc006f, + 0xbefe0070, 0xbeff0071, + 0x866f7bff, 0x000003ff, + 0xb96f4803, 0x866f7bff, + 0xfffff800, 0x8f6f8b6f, + 0xb96fa2c3, 0xb973f801, + 0xb8ee2a05, 0x806e816e, + 0x8e6e8a6e, 0x8e6e816e, + 0xb8ef1605, 0x806f816f, + 0x8e6f866f, 0x806e6f6e, + 0x806e746e, 0x826f8075, + 0x866fff6f, 0x0000ffff, + 0xc00b1c37, 0x00000050, + 0xc00b1d37, 0x00000060, + 0xc0031e77, 0x00000074, + 0xbf8cc07f, 0x866fff6d, + 0xf8000000, 0x8f6f9b6f, + 0x8e6f906f, 0xbeee0080, + 0x876e6f6e, 0x866fff6d, + 0x04000000, 0x8f6f9a6f, + 0x8e6f8f6f, 0x876e6f6e, + 0x866fff7a, 0x00800000, + 0x8f6f976f, 0xb96ef807, + 0x866dff6d, 0x0000ffff, + 0x86fe7e7e, 0x86ea6a6a, + 0x8f6e837a, 0xb96ee0c2, + 0xbf800002, 0xb97a0002, + 0xbf8a0000, 0x95806f6c, + 0xbf810000, 0x00000000, +}; diff --git a/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx10.asm b/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx10.asm index f20e463e748b..4433bda2ce25 100644 --- a/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx10.asm +++ b/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx10.asm @@ -20,1105 +20,948 @@ * OTHER DEALINGS IN THE SOFTWARE. */ +var SQ_WAVE_STATUS_INST_ATC_SHIFT = 23 +var SQ_WAVE_STATUS_INST_ATC_MASK = 0x00800000 +var SQ_WAVE_STATUS_SPI_PRIO_MASK = 0x00000006 +var SQ_WAVE_STATUS_HALT_MASK = 0x2000 + +var SQ_WAVE_LDS_ALLOC_LDS_SIZE_SHIFT = 12 +var SQ_WAVE_LDS_ALLOC_LDS_SIZE_SIZE = 9 +var SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SHIFT = 8 +var SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SIZE = 6 +var SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SHIFT = 24 +var SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SIZE = 4 +var SQ_WAVE_LDS_ALLOC_VGPR_SHARED_SIZE_SHIFT = 24 +var SQ_WAVE_LDS_ALLOC_VGPR_SHARED_SIZE_SIZE = 4 +var SQ_WAVE_IB_STS2_WAVE64_SHIFT = 11 +var SQ_WAVE_IB_STS2_WAVE64_SIZE = 1 + +var SQ_WAVE_TRAPSTS_SAVECTX_MASK = 0x400 +var SQ_WAVE_TRAPSTS_EXCE_MASK = 0x1FF +var SQ_WAVE_TRAPSTS_SAVECTX_SHIFT = 10 +var SQ_WAVE_TRAPSTS_MEM_VIOL_MASK = 0x100 +var SQ_WAVE_TRAPSTS_MEM_VIOL_SHIFT = 8 +var SQ_WAVE_TRAPSTS_PRE_SAVECTX_MASK = 0x3FF +var SQ_WAVE_TRAPSTS_PRE_SAVECTX_SHIFT = 0x0 +var SQ_WAVE_TRAPSTS_PRE_SAVECTX_SIZE = 10 +var SQ_WAVE_TRAPSTS_POST_SAVECTX_MASK = 0xFFFFF800 +var SQ_WAVE_TRAPSTS_POST_SAVECTX_SHIFT = 11 +var SQ_WAVE_TRAPSTS_POST_SAVECTX_SIZE = 21 +var SQ_WAVE_TRAPSTS_ILLEGAL_INST_MASK = 0x800 + +var SQ_WAVE_IB_STS_RCNT_SHIFT = 16 +var SQ_WAVE_IB_STS_FIRST_REPLAY_SHIFT = 15 +var SQ_WAVE_IB_STS_REPLAY_W64H_SHIFT = 25 +var SQ_WAVE_IB_STS_REPLAY_W64H_SIZE = 1 +var SQ_WAVE_IB_STS_REPLAY_W64H_MASK = 0x02000000 +var SQ_WAVE_IB_STS_FIRST_REPLAY_SIZE = 1 +var SQ_WAVE_IB_STS_RCNT_SIZE = 6 +var SQ_WAVE_IB_STS_RCNT_FIRST_REPLAY_MASK = 0x003F8000 +var SQ_WAVE_IB_STS_RCNT_FIRST_REPLAY_MASK_NEG = 0x00007FFF + +var SQ_BUF_RSRC_WORD1_ATC_SHIFT = 24 +var SQ_BUF_RSRC_WORD3_MTYPE_SHIFT = 27 + +// bits [31:24] unused by SPI debug data +var TTMP11_SAVE_REPLAY_W64H_SHIFT = 31 +var TTMP11_SAVE_REPLAY_W64H_MASK = 0x80000000 +var TTMP11_SAVE_RCNT_FIRST_REPLAY_SHIFT = 24 +var TTMP11_SAVE_RCNT_FIRST_REPLAY_MASK = 0x7F000000 + +// SQ_SEL_X/Y/Z/W, BUF_NUM_FORMAT_FLOAT, (0 for MUBUF stride[17:14] +// when ADD_TID_ENABLE and BUF_DATA_FORMAT_32 for MTBUF), ADD_TID_ENABLE +var S_SAVE_BUF_RSRC_WORD1_STRIDE = 0x00040000 +var S_SAVE_BUF_RSRC_WORD3_MISC = 0x10807FAC + +var S_SAVE_SPI_INIT_ATC_MASK = 0x08000000 +var S_SAVE_SPI_INIT_ATC_SHIFT = 27 +var S_SAVE_SPI_INIT_MTYPE_MASK = 0x70000000 +var S_SAVE_SPI_INIT_MTYPE_SHIFT = 28 +var S_SAVE_SPI_INIT_FIRST_WAVE_MASK = 0x04000000 +var S_SAVE_SPI_INIT_FIRST_WAVE_SHIFT = 26 + +var S_SAVE_PC_HI_RCNT_SHIFT = 26 +var S_SAVE_PC_HI_RCNT_MASK = 0xFC000000 +var S_SAVE_PC_HI_FIRST_REPLAY_SHIFT = 25 +var S_SAVE_PC_HI_FIRST_REPLAY_MASK = 0x02000000 +var S_SAVE_PC_HI_REPLAY_W64H_SHIFT = 24 +var S_SAVE_PC_HI_REPLAY_W64H_MASK = 0x01000000 + +var s_sgpr_save_num = 108 + +var s_save_spi_init_lo = exec_lo +var s_save_spi_init_hi = exec_hi +var s_save_pc_lo = ttmp0 +var s_save_pc_hi = ttmp1 +var s_save_exec_lo = ttmp2 +var s_save_exec_hi = ttmp3 +var s_save_status = ttmp12 +var s_save_trapsts = ttmp5 +var s_save_xnack_mask = ttmp6 +var s_wave_size = ttmp7 +var s_save_buf_rsrc0 = ttmp8 +var s_save_buf_rsrc1 = ttmp9 +var s_save_buf_rsrc2 = ttmp10 +var s_save_buf_rsrc3 = ttmp11 +var s_save_mem_offset = ttmp14 +var s_save_alloc_size = s_save_trapsts +var s_save_tmp = s_save_buf_rsrc2 +var s_save_m0 = ttmp15 + +var S_RESTORE_BUF_RSRC_WORD1_STRIDE = S_SAVE_BUF_RSRC_WORD1_STRIDE +var S_RESTORE_BUF_RSRC_WORD3_MISC = S_SAVE_BUF_RSRC_WORD3_MISC + +var S_RESTORE_SPI_INIT_ATC_MASK = 0x08000000 +var S_RESTORE_SPI_INIT_ATC_SHIFT = 27 +var S_RESTORE_SPI_INIT_MTYPE_MASK = 0x70000000 +var S_RESTORE_SPI_INIT_MTYPE_SHIFT = 28 +var S_RESTORE_SPI_INIT_FIRST_WAVE_MASK = 0x04000000 +var S_RESTORE_SPI_INIT_FIRST_WAVE_SHIFT = 26 +var S_WAVE_SIZE = 25 + +var S_RESTORE_PC_HI_RCNT_SHIFT = S_SAVE_PC_HI_RCNT_SHIFT +var S_RESTORE_PC_HI_RCNT_MASK = S_SAVE_PC_HI_RCNT_MASK +var S_RESTORE_PC_HI_FIRST_REPLAY_SHIFT = S_SAVE_PC_HI_FIRST_REPLAY_SHIFT +var S_RESTORE_PC_HI_FIRST_REPLAY_MASK = S_SAVE_PC_HI_FIRST_REPLAY_MASK + +var s_restore_spi_init_lo = exec_lo +var s_restore_spi_init_hi = exec_hi +var s_restore_mem_offset = ttmp12 +var s_restore_alloc_size = ttmp3 +var s_restore_tmp = ttmp6 +var s_restore_mem_offset_save = s_restore_tmp +var s_restore_m0 = s_restore_alloc_size +var s_restore_mode = ttmp7 +var s_restore_flat_scratch = ttmp2 +var s_restore_pc_lo = ttmp0 +var s_restore_pc_hi = ttmp1 +var s_restore_exec_lo = ttmp14 +var s_restore_exec_hi = ttmp15 +var s_restore_status = ttmp4 +var s_restore_trapsts = ttmp5 +var s_restore_xnack_mask = ttmp13 +var s_restore_buf_rsrc0 = ttmp8 +var s_restore_buf_rsrc1 = ttmp9 +var s_restore_buf_rsrc2 = ttmp10 +var s_restore_buf_rsrc3 = ttmp11 +var s_restore_size = ttmp7 shader main + asic(DEFAULT) + type(CS) + wave_size(32) -asic(DEFAULT) - -type(CS) - -wave_size(32) -/*************************************************************************/ -/* control on how to run the shader */ -/*************************************************************************/ -//any hack that needs to be made to run this code in EMU (either becasue various EMU code are not ready or no compute save & restore in EMU run) -var EMU_RUN_HACK = 0 -var EMU_RUN_HACK_RESTORE_NORMAL = 0 -var EMU_RUN_HACK_SAVE_NORMAL_EXIT = 0 -var EMU_RUN_HACK_SAVE_SINGLE_WAVE = 0 -var EMU_RUN_HACK_SAVE_FIRST_TIME = 0 //for interrupted restore in which the first save is through EMU_RUN_HACK -var SAVE_LDS = 0 -var WG_BASE_ADDR_LO = 0x9000a000 -var WG_BASE_ADDR_HI = 0x0 -var WAVE_SPACE = 0x9000 //memory size that each wave occupies in workgroup state mem, increase from 5000 to 9000 for more SGPR need to be saved -var CTX_SAVE_CONTROL = 0x0 -var CTX_RESTORE_CONTROL = CTX_SAVE_CONTROL -var SIM_RUN_HACK = 0 //any hack that needs to be made to run this code in SIM (either becasue various RTL code are not ready or no compute save & restore in RTL run) -var SGPR_SAVE_USE_SQC = 0 //use SQC D$ to do the write -var USE_MTBUF_INSTEAD_OF_MUBUF = 0 //need to change BUF_DATA_FORMAT in S_SAVE_BUF_RSRC_WORD3_MISC from 0 to BUF_DATA_FORMAT_32 if set to 1 (i.e. 0x00827FAC) -var SWIZZLE_EN = 0 //whether we use swizzled buffer addressing -var SAVE_RESTORE_HWID_DDID = 0 -var RESTORE_DDID_IN_SGPR18 = 0 -/**************************************************************************/ -/* variables */ -/**************************************************************************/ -var SQ_WAVE_STATUS_INST_ATC_SHIFT = 23 -var SQ_WAVE_STATUS_INST_ATC_MASK = 0x00800000 -var SQ_WAVE_STATUS_SPI_PRIO_MASK = 0x00000006 - -var SQ_WAVE_LDS_ALLOC_LDS_SIZE_SHIFT = 12 -var SQ_WAVE_LDS_ALLOC_LDS_SIZE_SIZE = 9 -var SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SHIFT = 8 -var SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SIZE = 6 -var SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SHIFT = 24 -var SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SIZE = 4 //FIXME sq.blk still has 4 bits at this time while SQ programming guide has 3 bits -var SQ_WAVE_LDS_ALLOC_VGPR_SHARED_SIZE_SHIFT = 24 -var SQ_WAVE_LDS_ALLOC_VGPR_SHARED_SIZE_SIZE = 4 -var SQ_WAVE_IB_STS2_WAVE64_SHIFT = 11 -var SQ_WAVE_IB_STS2_WAVE64_SIZE = 1 - -var SQ_WAVE_TRAPSTS_SAVECTX_MASK = 0x400 -var SQ_WAVE_TRAPSTS_EXCE_MASK = 0x1FF // Exception mask -var SQ_WAVE_TRAPSTS_SAVECTX_SHIFT = 10 -var SQ_WAVE_TRAPSTS_MEM_VIOL_MASK = 0x100 -var SQ_WAVE_TRAPSTS_MEM_VIOL_SHIFT = 8 -var SQ_WAVE_TRAPSTS_PRE_SAVECTX_MASK = 0x3FF -var SQ_WAVE_TRAPSTS_PRE_SAVECTX_SHIFT = 0x0 -var SQ_WAVE_TRAPSTS_PRE_SAVECTX_SIZE = 10 -var SQ_WAVE_TRAPSTS_POST_SAVECTX_MASK = 0xFFFFF800 -var SQ_WAVE_TRAPSTS_POST_SAVECTX_SHIFT = 11 -var SQ_WAVE_TRAPSTS_POST_SAVECTX_SIZE = 21 - -var SQ_WAVE_IB_STS_RCNT_SHIFT = 16 //FIXME -var SQ_WAVE_IB_STS_FIRST_REPLAY_SHIFT = 15 //FIXME -var SQ_WAVE_IB_STS_FIRST_REPLAY_SIZE = 1 //FIXME -var SQ_WAVE_IB_STS_RCNT_SIZE = 6 //FIXME -var SQ_WAVE_IB_STS_RCNT_FIRST_REPLAY_MASK_NEG = 0x00007FFF //FIXME - -var SQ_BUF_RSRC_WORD1_ATC_SHIFT = 24 -var SQ_BUF_RSRC_WORD3_MTYPE_SHIFT = 27 - - -/* Save */ -var S_SAVE_BUF_RSRC_WORD1_STRIDE = 0x00040000 //stride is 4 bytes -var S_SAVE_BUF_RSRC_WORD3_MISC = 0x00807FAC //SQ_SEL_X/Y/Z/W, BUF_NUM_FORMAT_FLOAT, (0 for MUBUF stride[17:14] when ADD_TID_ENABLE and BUF_DATA_FORMAT_32 for MTBUF), ADD_TID_ENABLE - -var S_SAVE_SPI_INIT_ATC_MASK = 0x08000000 //bit[27]: ATC bit -var S_SAVE_SPI_INIT_ATC_SHIFT = 27 -var S_SAVE_SPI_INIT_MTYPE_MASK = 0x70000000 //bit[30:28]: Mtype -var S_SAVE_SPI_INIT_MTYPE_SHIFT = 28 -var S_SAVE_SPI_INIT_FIRST_WAVE_MASK = 0x04000000 //bit[26]: FirstWaveInTG -var S_SAVE_SPI_INIT_FIRST_WAVE_SHIFT = 26 - -var S_SAVE_PC_HI_RCNT_SHIFT = 28 //FIXME check with Brian to ensure all fields other than PC[47:0] can be used -var S_SAVE_PC_HI_RCNT_MASK = 0xF0000000 //FIXME -var S_SAVE_PC_HI_FIRST_REPLAY_SHIFT = 27 //FIXME -var S_SAVE_PC_HI_FIRST_REPLAY_MASK = 0x08000000 //FIXME - -var s_save_spi_init_lo = exec_lo -var s_save_spi_init_hi = exec_hi - -var s_save_pc_lo = ttmp0 //{TTMP1, TTMP0} = {3ˇŻh0,pc_rewind[3:0], HT[0],trapID[7:0], PC[47:0]} -var s_save_pc_hi = ttmp1 -var s_save_exec_lo = ttmp2 -var s_save_exec_hi = ttmp3 -var s_save_status = ttmp4 -var s_save_trapsts = ttmp5 //not really used until the end of the SAVE routine -var s_wave_size = ttmp6 //ttmp6 is not needed now, since it's only 32bit xnack mask, now use it to determine wave32 or wave64 in EMU_HACK -var s_save_xnack_mask = ttmp7 -var s_save_buf_rsrc0 = ttmp8 -var s_save_buf_rsrc1 = ttmp9 -var s_save_buf_rsrc2 = ttmp10 -var s_save_buf_rsrc3 = ttmp11 - -var s_save_mem_offset = ttmp14 -var s_sgpr_save_num = 106 //in gfx10, all sgpr must be saved -var s_save_alloc_size = s_save_trapsts //conflict -var s_save_tmp = s_save_buf_rsrc2 //shared with s_save_buf_rsrc2 (conflict: should not use mem access with s_save_tmp at the same time) -var s_save_m0 = ttmp15 - -/* Restore */ -var S_RESTORE_BUF_RSRC_WORD1_STRIDE = S_SAVE_BUF_RSRC_WORD1_STRIDE -var S_RESTORE_BUF_RSRC_WORD3_MISC = S_SAVE_BUF_RSRC_WORD3_MISC - -var S_RESTORE_SPI_INIT_ATC_MASK = 0x08000000 //bit[27]: ATC bit -var S_RESTORE_SPI_INIT_ATC_SHIFT = 27 -var S_RESTORE_SPI_INIT_MTYPE_MASK = 0x70000000 //bit[30:28]: Mtype -var S_RESTORE_SPI_INIT_MTYPE_SHIFT = 28 -var S_RESTORE_SPI_INIT_FIRST_WAVE_MASK = 0x04000000 //bit[26]: FirstWaveInTG -var S_RESTORE_SPI_INIT_FIRST_WAVE_SHIFT = 26 - -var S_RESTORE_PC_HI_RCNT_SHIFT = S_SAVE_PC_HI_RCNT_SHIFT -var S_RESTORE_PC_HI_RCNT_MASK = S_SAVE_PC_HI_RCNT_MASK -var S_RESTORE_PC_HI_FIRST_REPLAY_SHIFT = S_SAVE_PC_HI_FIRST_REPLAY_SHIFT -var S_RESTORE_PC_HI_FIRST_REPLAY_MASK = S_SAVE_PC_HI_FIRST_REPLAY_MASK - -var s_restore_spi_init_lo = exec_lo -var s_restore_spi_init_hi = exec_hi - -var s_restore_mem_offset = ttmp12 -var s_restore_alloc_size = ttmp3 -var s_restore_tmp = ttmp6 -var s_restore_mem_offset_save = s_restore_tmp //no conflict - -var s_restore_m0 = s_restore_alloc_size //no conflict - -var s_restore_mode = ttmp13 -var s_restore_hwid1 = ttmp2 -var s_restore_ddid = s_restore_hwid1 -var s_restore_pc_lo = ttmp0 -var s_restore_pc_hi = ttmp1 -var s_restore_exec_lo = ttmp14 -var s_restore_exec_hi = ttmp15 -var s_restore_status = ttmp4 -var s_restore_trapsts = ttmp5 -//var s_restore_xnack_mask_lo = xnack_mask_lo -//var s_restore_xnack_mask_hi = xnack_mask_hi -var s_restore_xnack_mask = ttmp7 -var s_restore_buf_rsrc0 = ttmp8 -var s_restore_buf_rsrc1 = ttmp9 -var s_restore_buf_rsrc2 = ttmp10 -var s_restore_buf_rsrc3 = ttmp11 -var s_restore_size = ttmp13 //ttmp13 has no conflict - -/**************************************************************************/ -/* trap handler entry points */ -/**************************************************************************/ - if ((EMU_RUN_HACK) && (!EMU_RUN_HACK_RESTORE_NORMAL)) //hack to use trap_id for determining save/restore - //FIXME VCCZ un-init assertion s_getreg_b32 s_save_status, hwreg(HW_REG_STATUS) //save STATUS since we will change SCC - s_and_b32 s_save_tmp, s_save_pc_hi, 0xffff0000 //change SCC - s_cmp_eq_u32 s_save_tmp, 0x007e0000 //Save: trap_id = 0x7e. Restore: trap_id = 0x7f. - s_cbranch_scc0 L_JUMP_TO_RESTORE //do not need to recover STATUS here since we are going to RESTORE - //FIXME s_setreg_b32 hwreg(HW_REG_STATUS), s_save_status //need to recover STATUS since we are going to SAVE - s_branch L_SKIP_RESTORE //NOT restore, SAVE actually - else - s_branch L_SKIP_RESTORE //NOT restore. might be a regular trap or save - end + s_branch L_SKIP_RESTORE //NOT restore. might be a regular trap or save L_JUMP_TO_RESTORE: - s_branch L_RESTORE //restore + s_branch L_RESTORE L_SKIP_RESTORE: - - s_getreg_b32 s_save_status, hwreg(HW_REG_STATUS) //save STATUS since we will change SCC - s_andn2_b32 s_save_status, s_save_status, SQ_WAVE_STATUS_SPI_PRIO_MASK //check whether this is for save - s_getreg_b32 s_save_trapsts, hwreg(HW_REG_TRAPSTS) - s_and_b32 s_save_trapsts, s_save_trapsts, SQ_WAVE_TRAPSTS_SAVECTX_MASK //check whether this is for save - s_cbranch_scc1 L_SAVE //this is the operation for save - - // ********* Handle non-CWSR traps ******************* - if (!EMU_RUN_HACK) - s_getreg_b32 s_save_trapsts, hwreg(HW_REG_TRAPSTS) - s_and_b32 s_save_trapsts, s_save_trapsts, SQ_WAVE_TRAPSTS_EXCE_MASK // Check whether it is an exception - s_cbranch_scc1 L_EXCP_CASE // Exception, jump back to the shader program directly. - s_add_u32 ttmp0, ttmp0, 4 // S_TRAP case, add 4 to ttmp0 - - L_EXCP_CASE: - s_and_b32 ttmp1, ttmp1, 0xFFFF - s_rfe_b64 [ttmp0, ttmp1] - end - // ********* End handling of non-CWSR traps ******************* - -/**************************************************************************/ -/* save routine */ -/**************************************************************************/ - -L_SAVE: - + s_getreg_b32 s_save_status, hwreg(HW_REG_STATUS) //save STATUS since we will change SCC + s_andn2_b32 s_save_status, s_save_status, SQ_WAVE_STATUS_SPI_PRIO_MASK + s_getreg_b32 s_save_trapsts, hwreg(HW_REG_TRAPSTS) + s_and_b32 ttmp2, s_save_trapsts, SQ_WAVE_TRAPSTS_SAVECTX_MASK //check whether this is for save + s_cbranch_scc1 L_SAVE + + // If STATUS.MEM_VIOL is asserted then halt the wave to prevent + // the exception raising again and blocking context save. + s_and_b32 ttmp2, s_save_trapsts, SQ_WAVE_TRAPSTS_MEM_VIOL_MASK + s_cbranch_scc0 L_FETCH_2ND_TRAP + s_or_b32 s_save_status, s_save_status, SQ_WAVE_STATUS_HALT_MASK + +L_FETCH_2ND_TRAP: + // Preserve and clear scalar XNACK state before issuing scalar loads. + // Save IB_STS.REPLAY_W64H[25], RCNT[21:16], FIRST_REPLAY[15] into + // unused space ttmp11[31:24]. + s_andn2_b32 ttmp11, ttmp11, (TTMP11_SAVE_REPLAY_W64H_MASK | TTMP11_SAVE_RCNT_FIRST_REPLAY_MASK) + s_getreg_b32 ttmp2, hwreg(HW_REG_IB_STS) + s_and_b32 ttmp3, ttmp2, SQ_WAVE_IB_STS_REPLAY_W64H_MASK + s_lshl_b32 ttmp3, ttmp3, (TTMP11_SAVE_REPLAY_W64H_SHIFT - SQ_WAVE_IB_STS_REPLAY_W64H_SHIFT) + s_or_b32 ttmp11, ttmp11, ttmp3 + s_and_b32 ttmp3, ttmp2, SQ_WAVE_IB_STS_RCNT_FIRST_REPLAY_MASK + s_lshl_b32 ttmp3, ttmp3, (TTMP11_SAVE_RCNT_FIRST_REPLAY_SHIFT - SQ_WAVE_IB_STS_FIRST_REPLAY_SHIFT) + s_or_b32 ttmp11, ttmp11, ttmp3 + s_andn2_b32 ttmp2, ttmp2, (SQ_WAVE_IB_STS_REPLAY_W64H_MASK | SQ_WAVE_IB_STS_RCNT_FIRST_REPLAY_MASK) + s_setreg_b32 hwreg(HW_REG_IB_STS), ttmp2 + + // Read second-level TBA/TMA from first-level TMA and jump if available. + // ttmp[2:5] and ttmp12 can be used (others hold SPI-initialized debug data) + // ttmp12 holds SQ_WAVE_STATUS + s_getreg_b32 ttmp14, hwreg(HW_REG_SHADER_TMA_LO) + s_getreg_b32 ttmp15, hwreg(HW_REG_SHADER_TMA_HI) + s_lshl_b64 [ttmp14, ttmp15], [ttmp14, ttmp15], 0x8 + s_load_dwordx2 [ttmp2, ttmp3], [ttmp14, ttmp15], 0x0 glc:1 // second-level TBA + s_waitcnt lgkmcnt(0) + s_load_dwordx2 [ttmp14, ttmp15], [ttmp14, ttmp15], 0x8 glc:1 // second-level TMA + s_waitcnt lgkmcnt(0) + s_and_b64 [ttmp2, ttmp3], [ttmp2, ttmp3], [ttmp2, ttmp3] + s_cbranch_scc0 L_NO_NEXT_TRAP // second-level trap handler not been set + s_setpc_b64 [ttmp2, ttmp3] // jump to second-level trap handler + +L_NO_NEXT_TRAP: + s_getreg_b32 s_save_trapsts, hwreg(HW_REG_TRAPSTS) + s_and_b32 s_save_trapsts, s_save_trapsts, SQ_WAVE_TRAPSTS_EXCE_MASK + s_cbranch_scc1 L_EXCP_CASE // Exception, jump back to the shader program directly. + s_add_u32 ttmp0, ttmp0, 4 // S_TRAP case, add 4 to ttmp0 + s_addc_u32 ttmp1, ttmp1, 0 +L_EXCP_CASE: + s_and_b32 ttmp1, ttmp1, 0xFFFF + + // Restore SQ_WAVE_IB_STS. + s_lshr_b32 ttmp2, ttmp11, (TTMP11_SAVE_RCNT_FIRST_REPLAY_SHIFT - SQ_WAVE_IB_STS_FIRST_REPLAY_SHIFT) + s_and_b32 ttmp3, ttmp2, SQ_WAVE_IB_STS_RCNT_FIRST_REPLAY_MASK + s_lshr_b32 ttmp2, ttmp11, (TTMP11_SAVE_REPLAY_W64H_SHIFT - SQ_WAVE_IB_STS_REPLAY_W64H_SHIFT) + s_and_b32 ttmp2, ttmp2, SQ_WAVE_IB_STS_REPLAY_W64H_MASK + s_or_b32 ttmp2, ttmp2, ttmp3 + s_setreg_b32 hwreg(HW_REG_IB_STS), ttmp2 + + // Restore SQ_WAVE_STATUS. + s_and_b64 exec, exec, exec // Restore STATUS.EXECZ, not writable by s_setreg_b32 + s_and_b64 vcc, vcc, vcc // Restore STATUS.VCCZ, not writable by s_setreg_b32 + s_setreg_b32 hwreg(HW_REG_STATUS), s_save_status + + s_rfe_b64 [ttmp0, ttmp1] + +L_SAVE: //check whether there is mem_viol - s_getreg_b32 s_save_trapsts, hwreg(HW_REG_TRAPSTS) - s_and_b32 s_save_trapsts, s_save_trapsts, SQ_WAVE_TRAPSTS_MEM_VIOL_MASK + s_getreg_b32 s_save_trapsts, hwreg(HW_REG_TRAPSTS) + s_and_b32 s_save_trapsts, s_save_trapsts, SQ_WAVE_TRAPSTS_MEM_VIOL_MASK s_cbranch_scc0 L_NO_PC_REWIND - + //if so, need rewind PC assuming GDS operation gets NACKed - s_mov_b32 s_save_tmp, 0 //clear mem_viol bit - s_setreg_b32 hwreg(HW_REG_TRAPSTS, SQ_WAVE_TRAPSTS_MEM_VIOL_SHIFT, 1), s_save_tmp //clear mem_viol bit - s_and_b32 s_save_pc_hi, s_save_pc_hi, 0x0000ffff //pc[47:32] - s_sub_u32 s_save_pc_lo, s_save_pc_lo, 8 //pc[31:0]-8 - s_subb_u32 s_save_pc_hi, s_save_pc_hi, 0x0 // -scc + s_mov_b32 s_save_tmp, 0 + s_setreg_b32 hwreg(HW_REG_TRAPSTS, SQ_WAVE_TRAPSTS_MEM_VIOL_SHIFT, 1), s_save_tmp //clear mem_viol bit + s_and_b32 s_save_pc_hi, s_save_pc_hi, 0x0000ffff //pc[47:32] + s_sub_u32 s_save_pc_lo, s_save_pc_lo, 8 //pc[31:0]-8 + s_subb_u32 s_save_pc_hi, s_save_pc_hi, 0x0 L_NO_PC_REWIND: - s_mov_b32 s_save_tmp, 0 //clear saveCtx bit - s_setreg_b32 hwreg(HW_REG_TRAPSTS, SQ_WAVE_TRAPSTS_SAVECTX_SHIFT, 1), s_save_tmp //clear saveCtx bit - - //s_mov_b32 s_save_xnack_mask_lo, xnack_mask_lo //save XNACK_MASK - //s_mov_b32 s_save_xnack_mask_hi, xnack_mask_hi - s_getreg_b32 s_save_xnack_mask, hwreg(HW_REG_SHADER_XNACK_MASK) - s_getreg_b32 s_save_tmp, hwreg(HW_REG_IB_STS, SQ_WAVE_IB_STS_RCNT_SHIFT, SQ_WAVE_IB_STS_RCNT_SIZE) //save RCNT - s_lshl_b32 s_save_tmp, s_save_tmp, S_SAVE_PC_HI_RCNT_SHIFT - s_or_b32 s_save_pc_hi, s_save_pc_hi, s_save_tmp - s_getreg_b32 s_save_tmp, hwreg(HW_REG_IB_STS, SQ_WAVE_IB_STS_FIRST_REPLAY_SHIFT, SQ_WAVE_IB_STS_FIRST_REPLAY_SIZE) //save FIRST_REPLAY - s_lshl_b32 s_save_tmp, s_save_tmp, S_SAVE_PC_HI_FIRST_REPLAY_SHIFT - s_or_b32 s_save_pc_hi, s_save_pc_hi, s_save_tmp - s_getreg_b32 s_save_tmp, hwreg(HW_REG_IB_STS) //clear RCNT and FIRST_REPLAY in IB_STS - s_and_b32 s_save_tmp, s_save_tmp, SQ_WAVE_IB_STS_RCNT_FIRST_REPLAY_MASK_NEG + s_mov_b32 s_save_tmp, 0 + s_setreg_b32 hwreg(HW_REG_TRAPSTS, SQ_WAVE_TRAPSTS_SAVECTX_SHIFT, 1), s_save_tmp //clear saveCtx bit + + s_getreg_b32 s_save_xnack_mask, hwreg(HW_REG_SHADER_XNACK_MASK) + s_getreg_b32 s_save_tmp, hwreg(HW_REG_IB_STS, SQ_WAVE_IB_STS_RCNT_SHIFT, SQ_WAVE_IB_STS_RCNT_SIZE) + s_lshl_b32 s_save_tmp, s_save_tmp, S_SAVE_PC_HI_RCNT_SHIFT + s_or_b32 s_save_pc_hi, s_save_pc_hi, s_save_tmp + s_getreg_b32 s_save_tmp, hwreg(HW_REG_IB_STS, SQ_WAVE_IB_STS_FIRST_REPLAY_SHIFT, SQ_WAVE_IB_STS_FIRST_REPLAY_SIZE) + s_lshl_b32 s_save_tmp, s_save_tmp, S_SAVE_PC_HI_FIRST_REPLAY_SHIFT + s_or_b32 s_save_pc_hi, s_save_pc_hi, s_save_tmp + s_getreg_b32 s_save_tmp, hwreg(HW_REG_IB_STS, SQ_WAVE_IB_STS_REPLAY_W64H_SHIFT, SQ_WAVE_IB_STS_REPLAY_W64H_SIZE) + s_lshl_b32 s_save_tmp, s_save_tmp, S_SAVE_PC_HI_REPLAY_W64H_SHIFT + s_or_b32 s_save_pc_hi, s_save_pc_hi, s_save_tmp + s_getreg_b32 s_save_tmp, hwreg(HW_REG_IB_STS) //clear RCNT and FIRST_REPLAY and REPLAY_W64H in IB_STS + s_and_b32 s_save_tmp, s_save_tmp, SQ_WAVE_IB_STS_RCNT_FIRST_REPLAY_MASK_NEG s_setreg_b32 hwreg(HW_REG_IB_STS), s_save_tmp - - /* inform SPI the readiness and wait for SPI's go signal */ - s_mov_b32 s_save_exec_lo, exec_lo //save EXEC and use EXEC for the go signal from SPI - s_mov_b32 s_save_exec_hi, exec_hi - s_mov_b64 exec, 0x0 //clear EXEC to get ready to receive - if (EMU_RUN_HACK) - - else - s_sendmsg sendmsg(MSG_SAVEWAVE) //send SPI a message and wait for SPI's write to EXEC - end - - L_SLEEP: - s_sleep 0x2 - - if (EMU_RUN_HACK) - - else - s_cbranch_execz L_SLEEP - end - - - /* setup Resource Contants */ - if ((EMU_RUN_HACK) && (!EMU_RUN_HACK_SAVE_SINGLE_WAVE)) - //calculate wd_addr using absolute thread id - v_readlane_b32 s_save_tmp, v9, 0 - //determine it is wave32 or wave64 - s_getreg_b32 s_wave_size, hwreg(HW_REG_IB_STS2,SQ_WAVE_IB_STS2_WAVE64_SHIFT,SQ_WAVE_IB_STS2_WAVE64_SIZE) - s_cmp_eq_u32 s_wave_size, 0 - s_cbranch_scc1 L_SAVE_WAVE32 - s_lshr_b32 s_save_tmp, s_save_tmp, 6 //SAVE WAVE64 - s_branch L_SAVE_CON - L_SAVE_WAVE32: - s_lshr_b32 s_save_tmp, s_save_tmp, 5 //SAVE WAVE32 - L_SAVE_CON: - s_mul_i32 s_save_tmp, s_save_tmp, WAVE_SPACE - s_add_i32 s_save_spi_init_lo, s_save_tmp, WG_BASE_ADDR_LO - s_mov_b32 s_save_spi_init_hi, WG_BASE_ADDR_HI - s_and_b32 s_save_spi_init_hi, s_save_spi_init_hi, CTX_SAVE_CONTROL - else - end - if ((EMU_RUN_HACK) && (EMU_RUN_HACK_SAVE_SINGLE_WAVE)) - s_add_i32 s_save_spi_init_lo, s_save_tmp, WG_BASE_ADDR_LO - s_mov_b32 s_save_spi_init_hi, WG_BASE_ADDR_HI - s_and_b32 s_save_spi_init_hi, s_save_spi_init_hi, CTX_SAVE_CONTROL - else - end - - - s_mov_b32 s_save_buf_rsrc0, s_save_spi_init_lo //base_addr_lo - s_and_b32 s_save_buf_rsrc1, s_save_spi_init_hi, 0x0000FFFF //base_addr_hi - s_or_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, S_SAVE_BUF_RSRC_WORD1_STRIDE - s_mov_b32 s_save_buf_rsrc2, 0 //NUM_RECORDS initial value = 0 (in bytes) although not neccessarily inited - s_mov_b32 s_save_buf_rsrc3, S_SAVE_BUF_RSRC_WORD3_MISC - s_and_b32 s_save_tmp, s_save_spi_init_hi, S_SAVE_SPI_INIT_ATC_MASK - s_lshr_b32 s_save_tmp, s_save_tmp, (S_SAVE_SPI_INIT_ATC_SHIFT-SQ_BUF_RSRC_WORD1_ATC_SHIFT) //get ATC bit into position - s_or_b32 s_save_buf_rsrc3, s_save_buf_rsrc3, s_save_tmp //or ATC - s_and_b32 s_save_tmp, s_save_spi_init_hi, S_SAVE_SPI_INIT_MTYPE_MASK - s_lshr_b32 s_save_tmp, s_save_tmp, (S_SAVE_SPI_INIT_MTYPE_SHIFT-SQ_BUF_RSRC_WORD3_MTYPE_SHIFT) //get MTYPE bits into position - s_or_b32 s_save_buf_rsrc3, s_save_buf_rsrc3, s_save_tmp //or MTYPE - - s_mov_b32 s_save_m0, m0 //save M0 - - /* global mem offset */ - s_mov_b32 s_save_mem_offset, 0x0 //mem offset initial value = 0 - s_getreg_b32 s_wave_size, hwreg(HW_REG_IB_STS2,SQ_WAVE_IB_STS2_WAVE64_SHIFT,SQ_WAVE_IB_STS2_WAVE64_SIZE) //get wave_save_size - s_or_b32 s_wave_size, s_save_spi_init_hi, s_wave_size //share s_wave_size with exec_hi - - /* save VGPRs */ - ////////////////////////////// - L_SAVE_VGPR: - - s_mov_b32 exec_lo, 0xFFFFFFFF //need every thread from now on - s_and_b32 m0, s_wave_size, 1 - s_cmp_eq_u32 m0, 1 - s_cbranch_scc1 L_ENABLE_SAVE_VGPR_EXEC_HI - s_mov_b32 exec_hi, 0x00000000 - s_branch L_SAVE_VGPR_NORMAL - L_ENABLE_SAVE_VGPR_EXEC_HI: - s_mov_b32 exec_hi, 0xFFFFFFFF - L_SAVE_VGPR_NORMAL: - s_getreg_b32 s_save_alloc_size, hwreg(HW_REG_GPR_ALLOC,SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SHIFT,SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SIZE) //vpgr_size - //for wave32 and wave64, the num of vgpr function is the same? - s_add_u32 s_save_alloc_size, s_save_alloc_size, 1 - s_lshl_b32 s_save_alloc_size, s_save_alloc_size, 2 //Number of VGPRs = (vgpr_size + 1) * 4 (non-zero value) //FIXME for GFX, zero is possible - //determine it is wave32 or wave64 - s_and_b32 m0, s_wave_size, 1 - s_cmp_eq_u32 m0, 1 - s_cbranch_scc1 L_SAVE_VGPR_WAVE64 - - //zhenxu added it for save vgpr for wave32 - s_lshl_b32 s_save_buf_rsrc2, s_save_alloc_size, 7 //NUM_RECORDS in bytes (32 threads*4) - if (SWIZZLE_EN) - s_add_u32 s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0 //FIXME need to use swizzle to enable bounds checking? - else - s_mov_b32 s_save_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - end - - s_mov_b32 m0, 0x0 //VGPR initial index value =0 - //s_set_gpr_idx_on m0, 0x1 //M0[7:0] = M0[7:0] and M0[15:12] = 0x1 - //s_add_u32 s_save_alloc_size, s_save_alloc_size, 0x1000 //add 0x1000 since we compare m0 against it later, doesn't need this in gfx10 - - L_SAVE_VGPR_WAVE32_LOOP: - v_movrels_b32 v0, v0 //v0 = v[0+m0] - - if(USE_MTBUF_INSTEAD_OF_MUBUF) - tbuffer_store_format_x v0, v0, s_save_buf_rsrc0, s_save_mem_offset format:BUF_NUM_FORMAT_FLOAT format: BUF_DATA_FORMAT_32 slc:1 glc:1 - else - buffer_store_dword v0, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 - end - - s_add_u32 m0, m0, 1 //next vgpr index - s_add_u32 s_save_mem_offset, s_save_mem_offset, 128 //every buffer_store_dword does 128 bytes - s_cmp_lt_u32 m0, s_save_alloc_size //scc = (m0 < s_save_alloc_size) ? 1 : 0 - s_cbranch_scc1 L_SAVE_VGPR_WAVE32_LOOP //VGPR save is complete? - s_branch L_SAVE_LDS - //save vgpr for wave32 ends - - L_SAVE_VGPR_WAVE64: - s_lshl_b32 s_save_buf_rsrc2, s_save_alloc_size, 8 //NUM_RECORDS in bytes (64 threads*4) - if (SWIZZLE_EN) - s_add_u32 s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0 //FIXME need to use swizzle to enable bounds checking? - else - s_mov_b32 s_save_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - end - - s_mov_b32 m0, 0x0 //VGPR initial index value =0 - //s_set_gpr_idx_on m0, 0x1 //M0[7:0] = M0[7:0] and M0[15:12] = 0x1 - //s_add_u32 s_save_alloc_size, s_save_alloc_size, 0x1000 //add 0x1000 since we compare m0 against it later, doesn't need this in gfx10 - - L_SAVE_VGPR_WAVE64_LOOP: - v_movrels_b32 v0, v0 //v0 = v[0+m0] - - if(USE_MTBUF_INSTEAD_OF_MUBUF) - tbuffer_store_format_x v0, v0, s_save_buf_rsrc0, s_save_mem_offset format:BUF_NUM_FORMAT_FLOAT format: BUF_DATA_FORMAT_32 slc:1 glc:1 - else - buffer_store_dword v0, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 - end - - s_add_u32 m0, m0, 1 //next vgpr index - s_add_u32 s_save_mem_offset, s_save_mem_offset, 256 //every buffer_store_dword does 256 bytes - s_cmp_lt_u32 m0, s_save_alloc_size //scc = (m0 < s_save_alloc_size) ? 1 : 0 - s_cbranch_scc1 L_SAVE_VGPR_WAVE64_LOOP //VGPR save is complete? - //s_set_gpr_idx_off - // - //Below part will be the save shared vgpr part (new for gfx10) - s_getreg_b32 s_save_alloc_size, hwreg(HW_REG_LDS_ALLOC,SQ_WAVE_LDS_ALLOC_VGPR_SHARED_SIZE_SHIFT,SQ_WAVE_LDS_ALLOC_VGPR_SHARED_SIZE_SIZE) //shared_vgpr_size - s_and_b32 s_save_alloc_size, s_save_alloc_size, 0xFFFFFFFF //shared_vgpr_size is zero? - s_cbranch_scc0 L_SAVE_LDS //no shared_vgpr used? jump to L_SAVE_LDS - s_lshl_b32 s_save_alloc_size, s_save_alloc_size, 3 //Number of SHARED_VGPRs = shared_vgpr_size * 8 (non-zero value) - //m0 now has the value of normal vgpr count, just add the m0 with shared_vgpr count to get the total count. - //save shared_vgpr will start from the index of m0 - s_add_u32 s_save_alloc_size, s_save_alloc_size, m0 - s_mov_b32 exec_lo, 0xFFFFFFFF - s_mov_b32 exec_hi, 0x00000000 - L_SAVE_SHARED_VGPR_WAVE64_LOOP: - v_movrels_b32 v0, v0 //v0 = v[0+m0] - buffer_store_dword v0, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 - s_add_u32 m0, m0, 1 //next vgpr index - s_add_u32 s_save_mem_offset, s_save_mem_offset, 128 //every buffer_store_dword does 256 bytes - s_cmp_lt_u32 m0, s_save_alloc_size //scc = (m0 < s_save_alloc_size) ? 1 : 0 - s_cbranch_scc1 L_SAVE_SHARED_VGPR_WAVE64_LOOP //SHARED_VGPR save is complete? - - /* save LDS */ - ////////////////////////////// - L_SAVE_LDS: - - //Only check the first wave need LDS - /* the first wave in the threadgroup */ - s_barrier //FIXME not performance-optimal "LDS is used? wait for other waves in the same TG" - s_and_b32 s_save_tmp, s_wave_size, S_SAVE_SPI_INIT_FIRST_WAVE_MASK //exec is still used here - s_cbranch_scc0 L_SAVE_SGPR - - s_mov_b32 exec_lo, 0xFFFFFFFF //need every thread from now on - s_and_b32 m0, s_wave_size, 1 - s_cmp_eq_u32 m0, 1 - s_cbranch_scc1 L_ENABLE_SAVE_LDS_EXEC_HI - s_mov_b32 exec_hi, 0x00000000 - s_branch L_SAVE_LDS_NORMAL - L_ENABLE_SAVE_LDS_EXEC_HI: - s_mov_b32 exec_hi, 0xFFFFFFFF - L_SAVE_LDS_NORMAL: - s_getreg_b32 s_save_alloc_size, hwreg(HW_REG_LDS_ALLOC,SQ_WAVE_LDS_ALLOC_LDS_SIZE_SHIFT,SQ_WAVE_LDS_ALLOC_LDS_SIZE_SIZE) //lds_size - s_and_b32 s_save_alloc_size, s_save_alloc_size, 0xFFFFFFFF //lds_size is zero? - s_cbranch_scc0 L_SAVE_SGPR //no lds used? jump to L_SAVE_VGPR - s_lshl_b32 s_save_alloc_size, s_save_alloc_size, 6 //LDS size in dwords = lds_size * 64dw - s_lshl_b32 s_save_alloc_size, s_save_alloc_size, 2 //LDS size in bytes - s_mov_b32 s_save_buf_rsrc2, s_save_alloc_size //NUM_RECORDS in bytes - if (SWIZZLE_EN) - s_add_u32 s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0 //FIXME need to use swizzle to enable bounds checking? - else - s_mov_b32 s_save_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - end - - //load 0~63*4(byte address) to vgpr v15 - v_mbcnt_lo_u32_b32 v0, -1, 0 - v_mbcnt_hi_u32_b32 v0, -1, v0 - v_mul_u32_u24 v0, 4, v0 - - s_and_b32 m0, s_wave_size, 1 - s_cmp_eq_u32 m0, 1 - s_mov_b32 m0, 0x0 - s_cbranch_scc1 L_SAVE_LDS_LOOP_W64 - - L_SAVE_LDS_LOOP_W32: - if (SAVE_LDS) - ds_read_b32 v1, v0 - s_waitcnt 0 //ensure data ready - buffer_store_dword v1, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 - //buffer_store_lds_dword s_save_buf_rsrc0, s_save_mem_offset lds:1 //save lds to memory doesn't exist in 10 - end - s_add_u32 m0, m0, 128 //every buffer_store_lds does 128 bytes - s_add_u32 s_save_mem_offset, s_save_mem_offset, 128 //mem offset increased by 128 bytes - v_add_nc_u32 v0, v0, 128 - s_cmp_lt_u32 m0, s_save_alloc_size //scc=(m0 < s_save_alloc_size) ? 1 : 0 - s_cbranch_scc1 L_SAVE_LDS_LOOP_W32 //LDS save is complete? - s_branch L_SAVE_SGPR - - L_SAVE_LDS_LOOP_W64: - if (SAVE_LDS) - ds_read_b32 v1, v0 - s_waitcnt 0 //ensure data ready - buffer_store_dword v1, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 - //buffer_store_lds_dword s_save_buf_rsrc0, s_save_mem_offset lds:1 //save lds to memory doesn't exist in 10 - end - s_add_u32 m0, m0, 256 //every buffer_store_lds does 256 bytes - s_add_u32 s_save_mem_offset, s_save_mem_offset, 256 //mem offset increased by 256 bytes - v_add_nc_u32 v0, v0, 256 - s_cmp_lt_u32 m0, s_save_alloc_size //scc=(m0 < s_save_alloc_size) ? 1 : 0 - s_cbranch_scc1 L_SAVE_LDS_LOOP_W64 //LDS save is complete? - - - /* save SGPRs */ - ////////////////////////////// - //s_getreg_b32 s_save_alloc_size, hwreg(HW_REG_GPR_ALLOC,SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SHIFT,SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SIZE) //spgr_size - //s_add_u32 s_save_alloc_size, s_save_alloc_size, 1 - //s_lshl_b32 s_save_alloc_size, s_save_alloc_size, 4 //Number of SGPRs = (sgpr_size + 1) * 16 (non-zero value) - //s_lshl_b32 s_save_alloc_size, s_save_alloc_size, 3 //In gfx10, Number of SGPRs = (sgpr_size + 1) * 8 (non-zero value) - L_SAVE_SGPR: - //need to look at it is wave32 or wave64 - s_and_b32 m0, s_wave_size, 1 - s_cmp_eq_u32 m0, 1 - s_cbranch_scc1 L_SAVE_SGPR_VMEM_WAVE64 - if (SGPR_SAVE_USE_SQC) - s_lshl_b32 s_save_buf_rsrc2, s_sgpr_save_num, 2 //NUM_RECORDS in bytes - else - s_lshl_b32 s_save_buf_rsrc2, s_sgpr_save_num, 7 //NUM_RECORDS in bytes (32 threads) - end - s_branch L_SAVE_SGPR_CONT - L_SAVE_SGPR_VMEM_WAVE64: - if (SGPR_SAVE_USE_SQC) - s_lshl_b32 s_save_buf_rsrc2, s_sgpr_save_num, 2 //NUM_RECORDS in bytes - else - s_lshl_b32 s_save_buf_rsrc2, s_sgpr_save_num, 8 //NUM_RECORDS in bytes (64 threads) - end - L_SAVE_SGPR_CONT: - if (SWIZZLE_EN) - s_add_u32 s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0 //FIXME need to use swizzle to enable bounds checking? - else - s_mov_b32 s_save_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - end - - //s_mov_b32 m0, 0x0 //SGPR initial index value =0 - //s_nop 0x0 //Manually inserted wait states - - s_and_b32 m0, s_wave_size, 1 - s_cmp_eq_u32 m0, 1 - - s_mov_b32 m0, 0x0 //SGPR initial index value =0 - s_nop 0x0 //Manually inserted wait states - - s_cbranch_scc1 L_SAVE_SGPR_LOOP_WAVE64 - - L_SAVE_SGPR_LOOP_WAVE32: - s_movrels_b32 s0, s0 //s0 = s[0+m0] - //zhenxu, adding one more argument to save sgpr function, this is only for vmem, using sqc is not change - write_sgpr_to_mem_wave32(s0, s_save_buf_rsrc0, s_save_mem_offset, SGPR_SAVE_USE_SQC, USE_MTBUF_INSTEAD_OF_MUBUF) //PV: the best performance should be using s_buffer_store_dwordx4 - s_add_u32 m0, m0, 1 //next sgpr index - s_cmp_lt_u32 m0, s_sgpr_save_num //scc = (m0 < s_sgpr_save_num) ? 1 : 0 - s_cbranch_scc1 L_SAVE_SGPR_LOOP_WAVE32 //SGPR save is complete? - s_branch L_SAVE_HWREG - - L_SAVE_SGPR_LOOP_WAVE64: - s_movrels_b32 s0, s0 //s0 = s[0+m0] - //zhenxu, adding one more argument to save sgpr function, this is only for vmem, using sqc is not change - write_sgpr_to_mem_wave64(s0, s_save_buf_rsrc0, s_save_mem_offset, SGPR_SAVE_USE_SQC, USE_MTBUF_INSTEAD_OF_MUBUF) //PV: the best performance should be using s_buffer_store_dwordx4 - s_add_u32 m0, m0, 1 //next sgpr index - s_cmp_lt_u32 m0, s_sgpr_save_num //scc = (m0 < s_sgpr_save_num) ? 1 : 0 - s_cbranch_scc1 L_SAVE_SGPR_LOOP_WAVE64 //SGPR save is complete? - - - /* save HW registers */ - ////////////////////////////// - L_SAVE_HWREG: - s_mov_b32 s_save_buf_rsrc2, 0x4 //NUM_RECORDS in bytes - if (SWIZZLE_EN) - s_add_u32 s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0 //FIXME need to use swizzle to enable bounds checking? - else - s_mov_b32 s_save_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - end - - s_and_b32 m0, s_wave_size, 1 - s_cmp_eq_u32 m0, 1 - s_cbranch_scc1 L_SAVE_HWREG_WAVE64 - - write_sgpr_to_mem_wave32(s_save_m0, s_save_buf_rsrc0, s_save_mem_offset, SGPR_SAVE_USE_SQC, USE_MTBUF_INSTEAD_OF_MUBUF) //M0 - - if ((EMU_RUN_HACK) && (EMU_RUN_HACK_SAVE_FIRST_TIME)) - s_add_u32 s_save_pc_lo, s_save_pc_lo, 4 //pc[31:0]+4 - s_addc_u32 s_save_pc_hi, s_save_pc_hi, 0x0 //carry bit over - end - - write_sgpr_to_mem_wave32(s_save_pc_lo, s_save_buf_rsrc0, s_save_mem_offset, SGPR_SAVE_USE_SQC, USE_MTBUF_INSTEAD_OF_MUBUF) //PC - write_sgpr_to_mem_wave32(s_save_pc_hi, s_save_buf_rsrc0, s_save_mem_offset, SGPR_SAVE_USE_SQC, USE_MTBUF_INSTEAD_OF_MUBUF) - write_sgpr_to_mem_wave32(s_save_exec_lo, s_save_buf_rsrc0, s_save_mem_offset, SGPR_SAVE_USE_SQC, USE_MTBUF_INSTEAD_OF_MUBUF) //EXEC - write_sgpr_to_mem_wave32(s_save_exec_hi, s_save_buf_rsrc0, s_save_mem_offset, SGPR_SAVE_USE_SQC, USE_MTBUF_INSTEAD_OF_MUBUF) - write_sgpr_to_mem_wave32(s_save_status, s_save_buf_rsrc0, s_save_mem_offset, SGPR_SAVE_USE_SQC, USE_MTBUF_INSTEAD_OF_MUBUF) //STATUS - - //s_save_trapsts conflicts with s_save_alloc_size - s_getreg_b32 s_save_trapsts, hwreg(HW_REG_TRAPSTS) - write_sgpr_to_mem_wave32(s_save_trapsts, s_save_buf_rsrc0, s_save_mem_offset, SGPR_SAVE_USE_SQC, USE_MTBUF_INSTEAD_OF_MUBUF) //TRAPSTS - - //write_sgpr_to_mem_wave32(s_save_xnack_mask_lo, s_save_buf_rsrc0, s_save_mem_offset, SGPR_SAVE_USE_SQC, USE_MTBUF_INSTEAD_OF_MUBUF) //XNACK_MASK_LO - write_sgpr_to_mem_wave32(s_save_xnack_mask, s_save_buf_rsrc0, s_save_mem_offset, SGPR_SAVE_USE_SQC, USE_MTBUF_INSTEAD_OF_MUBUF) //XNACK_MASK_HI - - //use s_save_tmp would introduce conflict here between s_save_tmp and s_save_buf_rsrc2 - s_getreg_b32 s_save_m0, hwreg(HW_REG_MODE) //MODE - write_sgpr_to_mem_wave32(s_save_m0, s_save_buf_rsrc0, s_save_mem_offset, SGPR_SAVE_USE_SQC, USE_MTBUF_INSTEAD_OF_MUBUF) - if(SAVE_RESTORE_HWID_DDID) - s_getreg_b32 s_save_m0, hwreg(HW_REG_HW_ID1) //HW_ID1, handler records the SE/SA/WGP/SIMD/wave of the original wave - write_sgpr_to_mem_wave32(s_save_m0, s_save_buf_rsrc0, s_save_mem_offset, SGPR_SAVE_USE_SQC, USE_MTBUF_INSTEAD_OF_MUBUF) - end - s_branch L_S_PGM_END_SAVED - - L_SAVE_HWREG_WAVE64: - write_sgpr_to_mem_wave64(s_save_m0, s_save_buf_rsrc0, s_save_mem_offset, SGPR_SAVE_USE_SQC, USE_MTBUF_INSTEAD_OF_MUBUF) //M0 - - if ((EMU_RUN_HACK) && (EMU_RUN_HACK_SAVE_FIRST_TIME)) - s_add_u32 s_save_pc_lo, s_save_pc_lo, 4 //pc[31:0]+4 - s_addc_u32 s_save_pc_hi, s_save_pc_hi, 0x0 //carry bit over - end - - write_sgpr_to_mem_wave64(s_save_pc_lo, s_save_buf_rsrc0, s_save_mem_offset, SGPR_SAVE_USE_SQC, USE_MTBUF_INSTEAD_OF_MUBUF) //PC - write_sgpr_to_mem_wave64(s_save_pc_hi, s_save_buf_rsrc0, s_save_mem_offset, SGPR_SAVE_USE_SQC, USE_MTBUF_INSTEAD_OF_MUBUF) - write_sgpr_to_mem_wave64(s_save_exec_lo, s_save_buf_rsrc0, s_save_mem_offset, SGPR_SAVE_USE_SQC, USE_MTBUF_INSTEAD_OF_MUBUF) //EXEC - write_sgpr_to_mem_wave64(s_save_exec_hi, s_save_buf_rsrc0, s_save_mem_offset, SGPR_SAVE_USE_SQC, USE_MTBUF_INSTEAD_OF_MUBUF) - write_sgpr_to_mem_wave64(s_save_status, s_save_buf_rsrc0, s_save_mem_offset, SGPR_SAVE_USE_SQC, USE_MTBUF_INSTEAD_OF_MUBUF) //STATUS - - //s_save_trapsts conflicts with s_save_alloc_size - s_getreg_b32 s_save_trapsts, hwreg(HW_REG_TRAPSTS) - write_sgpr_to_mem_wave64(s_save_trapsts, s_save_buf_rsrc0, s_save_mem_offset, SGPR_SAVE_USE_SQC, USE_MTBUF_INSTEAD_OF_MUBUF) //TRAPSTS - - //write_sgpr_to_mem_wave64(s_save_xnack_mask_lo, s_save_buf_rsrc0, s_save_mem_offset, SGPR_SAVE_USE_SQC, USE_MTBUF_INSTEAD_OF_MUBUF) //XNACK_MASK_LO - write_sgpr_to_mem_wave64(s_save_xnack_mask, s_save_buf_rsrc0, s_save_mem_offset, SGPR_SAVE_USE_SQC, USE_MTBUF_INSTEAD_OF_MUBUF) //XNACK_MASK_HI - - //use s_save_tmp would introduce conflict here between s_save_tmp and s_save_buf_rsrc2 - s_getreg_b32 s_save_m0, hwreg(HW_REG_MODE) //MODE - write_sgpr_to_mem_wave64(s_save_m0, s_save_buf_rsrc0, s_save_mem_offset, SGPR_SAVE_USE_SQC, USE_MTBUF_INSTEAD_OF_MUBUF) - - - if(SAVE_RESTORE_HWID_DDID) - s_getreg_b32 s_save_m0, hwreg(HW_REG_HW_ID1) //HW_ID1, handler records the SE/SA/WGP/SIMD/wave of the original wave - write_sgpr_to_mem_wave64(s_save_m0, s_save_buf_rsrc0, s_save_mem_offset, SGPR_SAVE_USE_SQC, USE_MTBUF_INSTEAD_OF_MUBUF) - - /* save DDID */ - ////////////////////////////// - L_SAVE_DDID: - //EXEC has been saved, no vector inst following - s_mov_b32 exec_lo, 0x80000000 //Set MSB to 1. Cleared when draw index is returned - s_sendmsg sendmsg(MSG_GET_DDID) - - L_WAIT_DDID_LOOP: - s_nop 7 // sleep a bit - s_bitcmp0_b32 exec_lo, 31 // test to see if MSB is cleared, meaning done - s_cbranch_scc0 L_WAIT_DDID_LOOP - - s_mov_b32 s_save_m0, exec_lo - - - s_mov_b32 s_save_buf_rsrc2, 0x4 //NUM_RECORDS in bytes - if (SWIZZLE_EN) - s_add_u32 s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0 //FIXME need to use swizzle to enable bounds checking? - else - s_mov_b32 s_save_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - end - s_and_b32 m0, s_wave_size, 1 - s_cmp_eq_u32 m0, 1 - s_cbranch_scc1 L_SAVE_DDID_WAVE64 - - write_sgpr_to_mem_wave32(s_save_m0, s_save_buf_rsrc0, s_save_mem_offset, SGPR_SAVE_USE_SQC, USE_MTBUF_INSTEAD_OF_MUBUF) - - L_SAVE_DDID_WAVE64: - write_sgpr_to_mem_wave64(s_save_m0, s_save_buf_rsrc0, s_save_mem_offset, SGPR_SAVE_USE_SQC, USE_MTBUF_INSTEAD_OF_MUBUF) - - end - - L_S_PGM_END_SAVED: - /* S_PGM_END_SAVED */ //FIXME graphics ONLY - if ((EMU_RUN_HACK) && (!EMU_RUN_HACK_SAVE_NORMAL_EXIT)) - s_and_b32 s_save_pc_hi, s_save_pc_hi, 0x0000ffff //pc[47:32] - s_add_u32 s_save_pc_lo, s_save_pc_lo, 4 //pc[31:0]+4 - s_addc_u32 s_save_pc_hi, s_save_pc_hi, 0x0 //carry bit over - s_rfe_b64 s_save_pc_lo //Return to the main shader program - else - end - - - s_branch L_END_PGM - - - -/**************************************************************************/ -/* restore routine */ -/**************************************************************************/ + + /* inform SPI the readiness and wait for SPI's go signal */ + s_mov_b32 s_save_exec_lo, exec_lo //save EXEC and use EXEC for the go signal from SPI + s_mov_b32 s_save_exec_hi, exec_hi + s_mov_b64 exec, 0x0 //clear EXEC to get ready to receive + + s_sendmsg sendmsg(MSG_SAVEWAVE) //send SPI a message and wait for SPI's write to EXEC + +L_SLEEP: + // sleep 1 (64clk) is not enough for 8 waves per SIMD, which will cause + // SQ hang, since the 7,8th wave could not get arbit to exec inst, while + // other waves are stuck into the sleep-loop and waiting for wrexec!=0 + s_sleep 0x2 + s_cbranch_execz L_SLEEP + + /* setup Resource Contants */ + s_mov_b32 s_save_buf_rsrc0, s_save_spi_init_lo //base_addr_lo + s_and_b32 s_save_buf_rsrc1, s_save_spi_init_hi, 0x0000FFFF //base_addr_hi + s_or_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, S_SAVE_BUF_RSRC_WORD1_STRIDE + s_mov_b32 s_save_buf_rsrc2, 0 //NUM_RECORDS initial value = 0 (in bytes) although not neccessarily inited + s_mov_b32 s_save_buf_rsrc3, S_SAVE_BUF_RSRC_WORD3_MISC + s_and_b32 s_save_tmp, s_save_spi_init_hi, S_SAVE_SPI_INIT_ATC_MASK + s_lshr_b32 s_save_tmp, s_save_tmp, (S_SAVE_SPI_INIT_ATC_SHIFT-SQ_BUF_RSRC_WORD1_ATC_SHIFT) + s_or_b32 s_save_buf_rsrc3, s_save_buf_rsrc3, s_save_tmp //or ATC + s_and_b32 s_save_tmp, s_save_spi_init_hi, S_SAVE_SPI_INIT_MTYPE_MASK + s_lshr_b32 s_save_tmp, s_save_tmp, (S_SAVE_SPI_INIT_MTYPE_SHIFT-SQ_BUF_RSRC_WORD3_MTYPE_SHIFT) + s_or_b32 s_save_buf_rsrc3, s_save_buf_rsrc3, s_save_tmp //or MTYPE + + s_mov_b32 s_save_m0, m0 + + /* global mem offset */ + s_mov_b32 s_save_mem_offset, 0x0 + s_getreg_b32 s_wave_size, hwreg(HW_REG_IB_STS2,SQ_WAVE_IB_STS2_WAVE64_SHIFT,SQ_WAVE_IB_STS2_WAVE64_SIZE) + s_lshl_b32 s_wave_size, s_wave_size, S_WAVE_SIZE + s_or_b32 s_wave_size, s_save_spi_init_hi, s_wave_size //share s_wave_size with exec_hi, it's at bit25 + + /* save HW registers */ + +L_SAVE_HWREG: + // HWREG SR memory offset : size(VGPR)+size(SVGPR)+size(SGPR) + get_vgpr_size_bytes(s_save_mem_offset, s_wave_size) + get_svgpr_size_bytes(s_save_tmp) + s_add_u32 s_save_mem_offset, s_save_mem_offset, s_save_tmp + s_add_u32 s_save_mem_offset, s_save_mem_offset, get_sgpr_size_bytes() + + s_mov_b32 s_save_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes + + write_hwreg_to_mem(s_save_m0, s_save_buf_rsrc0, s_save_mem_offset) + write_hwreg_to_mem(s_save_pc_lo, s_save_buf_rsrc0, s_save_mem_offset) + write_hwreg_to_mem(s_save_pc_hi, s_save_buf_rsrc0, s_save_mem_offset) + write_hwreg_to_mem(s_save_exec_lo, s_save_buf_rsrc0, s_save_mem_offset) + write_hwreg_to_mem(s_save_exec_hi, s_save_buf_rsrc0, s_save_mem_offset) + write_hwreg_to_mem(s_save_status, s_save_buf_rsrc0, s_save_mem_offset) + + s_getreg_b32 s_save_trapsts, hwreg(HW_REG_TRAPSTS) + write_hwreg_to_mem(s_save_trapsts, s_save_buf_rsrc0, s_save_mem_offset) + write_hwreg_to_mem(s_save_xnack_mask, s_save_buf_rsrc0, s_save_mem_offset) + + s_getreg_b32 s_save_m0, hwreg(HW_REG_MODE) + write_hwreg_to_mem(s_save_m0, s_save_buf_rsrc0, s_save_mem_offset) + + s_getreg_b32 s_save_m0, hwreg(HW_REG_SHADER_FLAT_SCRATCH_LO) + write_hwreg_to_mem(s_save_m0, s_save_buf_rsrc0, s_save_mem_offset) + + s_getreg_b32 s_save_m0, hwreg(HW_REG_SHADER_FLAT_SCRATCH_HI) + write_hwreg_to_mem(s_save_m0, s_save_buf_rsrc0, s_save_mem_offset) + + /* the first wave in the threadgroup */ + s_and_b32 s_save_tmp, s_save_spi_init_hi, S_SAVE_SPI_INIT_FIRST_WAVE_MASK + s_mov_b32 s_save_exec_hi, 0x0 + s_or_b32 s_save_exec_hi, s_save_tmp, s_save_exec_hi // save first wave bit to s_save_exec_hi.bits[26] + + /* save SGPRs */ + // Save SGPR before LDS save, then the s0 to s4 can be used during LDS save... + + // SGPR SR memory offset : size(VGPR)+size(SVGPR) + get_vgpr_size_bytes(s_save_mem_offset, s_wave_size) + get_svgpr_size_bytes(s_save_tmp) + s_add_u32 s_save_mem_offset, s_save_mem_offset, s_save_tmp + s_mov_b32 s_save_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes + + // backup s_save_buf_rsrc0,1 to s_save_pc_lo/hi, since write_16sgpr_to_mem function will change the rsrc0 + s_mov_b32 s_save_xnack_mask, s_save_buf_rsrc0 + s_add_u32 s_save_buf_rsrc0, s_save_buf_rsrc0, s_save_mem_offset + s_addc_u32 s_save_buf_rsrc1, s_save_buf_rsrc1, 0 + + s_mov_b32 m0, 0x0 //SGPR initial index value =0 + s_nop 0x0 //Manually inserted wait states +L_SAVE_SGPR_LOOP: + // SGPR is allocated in 16 SGPR granularity + s_movrels_b64 s0, s0 //s0 = s[0+m0], s1 = s[1+m0] + s_movrels_b64 s2, s2 //s2 = s[2+m0], s3 = s[3+m0] + s_movrels_b64 s4, s4 //s4 = s[4+m0], s5 = s[5+m0] + s_movrels_b64 s6, s6 //s6 = s[6+m0], s7 = s[7+m0] + s_movrels_b64 s8, s8 //s8 = s[8+m0], s9 = s[9+m0] + s_movrels_b64 s10, s10 //s10 = s[10+m0], s11 = s[11+m0] + s_movrels_b64 s12, s12 //s12 = s[12+m0], s13 = s[13+m0] + s_movrels_b64 s14, s14 //s14 = s[14+m0], s15 = s[15+m0] + + write_16sgpr_to_mem(s0, s_save_buf_rsrc0, s_save_mem_offset) + s_add_u32 m0, m0, 16 //next sgpr index + s_cmp_lt_u32 m0, 96 //scc = (m0 < first 96 SGPR) ? 1 : 0 + s_cbranch_scc1 L_SAVE_SGPR_LOOP //first 96 SGPR save is complete? + + //save the rest 12 SGPR + s_movrels_b64 s0, s0 //s0 = s[0+m0], s1 = s[1+m0] + s_movrels_b64 s2, s2 //s2 = s[2+m0], s3 = s[3+m0] + s_movrels_b64 s4, s4 //s4 = s[4+m0], s5 = s[5+m0] + s_movrels_b64 s6, s6 //s6 = s[6+m0], s7 = s[7+m0] + s_movrels_b64 s8, s8 //s8 = s[8+m0], s9 = s[9+m0] + s_movrels_b64 s10, s10 //s10 = s[10+m0], s11 = s[11+m0] + write_12sgpr_to_mem(s0, s_save_buf_rsrc0, s_save_mem_offset) + + // restore s_save_buf_rsrc0,1 + s_mov_b32 s_save_buf_rsrc0, s_save_xnack_mask + + /* save first 4 VGPR, then LDS save could use */ + // each wave will alloc 4 vgprs at least... + + s_mov_b32 s_save_mem_offset, 0 + s_mov_b32 exec_lo, 0xFFFFFFFF //need every thread from now on + s_lshr_b32 m0, s_wave_size, S_WAVE_SIZE + s_and_b32 m0, m0, 1 + s_cmp_eq_u32 m0, 1 + s_cbranch_scc1 L_ENABLE_SAVE_4VGPR_EXEC_HI + s_mov_b32 exec_hi, 0x00000000 + s_branch L_SAVE_4VGPR_WAVE32 +L_ENABLE_SAVE_4VGPR_EXEC_HI: + s_mov_b32 exec_hi, 0xFFFFFFFF + s_branch L_SAVE_4VGPR_WAVE64 +L_SAVE_4VGPR_WAVE32: + s_mov_b32 s_save_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes + + // VGPR Allocated in 4-GPR granularity + + buffer_store_dword v0, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 + buffer_store_dword v1, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 offset:128 + buffer_store_dword v2, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 offset:128*2 + buffer_store_dword v3, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 offset:128*3 + s_branch L_SAVE_LDS + +L_SAVE_4VGPR_WAVE64: + s_mov_b32 s_save_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes + + // VGPR Allocated in 4-GPR granularity + + buffer_store_dword v0, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 + buffer_store_dword v1, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 offset:256 + buffer_store_dword v2, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 offset:256*2 + buffer_store_dword v3, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 offset:256*3 + + /* save LDS */ + +L_SAVE_LDS: + // Change EXEC to all threads... + s_mov_b32 exec_lo, 0xFFFFFFFF //need every thread from now on + s_lshr_b32 m0, s_wave_size, S_WAVE_SIZE + s_and_b32 m0, m0, 1 + s_cmp_eq_u32 m0, 1 + s_cbranch_scc1 L_ENABLE_SAVE_LDS_EXEC_HI + s_mov_b32 exec_hi, 0x00000000 + s_branch L_SAVE_LDS_NORMAL +L_ENABLE_SAVE_LDS_EXEC_HI: + s_mov_b32 exec_hi, 0xFFFFFFFF +L_SAVE_LDS_NORMAL: + s_getreg_b32 s_save_alloc_size, hwreg(HW_REG_LDS_ALLOC,SQ_WAVE_LDS_ALLOC_LDS_SIZE_SHIFT,SQ_WAVE_LDS_ALLOC_LDS_SIZE_SIZE) + s_and_b32 s_save_alloc_size, s_save_alloc_size, 0xFFFFFFFF //lds_size is zero? + s_cbranch_scc0 L_SAVE_LDS_DONE //no lds used? jump to L_SAVE_DONE + + s_barrier //LDS is used? wait for other waves in the same TG + s_and_b32 s_save_tmp, s_save_exec_hi, S_SAVE_SPI_INIT_FIRST_WAVE_MASK + s_cbranch_scc0 L_SAVE_LDS_DONE + + // first wave do LDS save; + + s_lshl_b32 s_save_alloc_size, s_save_alloc_size, 6 //LDS size in dwords = lds_size * 64dw + s_lshl_b32 s_save_alloc_size, s_save_alloc_size, 2 //LDS size in bytes + s_mov_b32 s_save_buf_rsrc2, s_save_alloc_size //NUM_RECORDS in bytes + + // LDS at offset: size(VGPR)+size(SVGPR)+SIZE(SGPR)+SIZE(HWREG) + // + get_vgpr_size_bytes(s_save_mem_offset, s_wave_size) + get_svgpr_size_bytes(s_save_tmp) + s_add_u32 s_save_mem_offset, s_save_mem_offset, s_save_tmp + s_add_u32 s_save_mem_offset, s_save_mem_offset, get_sgpr_size_bytes() + s_add_u32 s_save_mem_offset, s_save_mem_offset, get_hwreg_size_bytes() + + s_mov_b32 s_save_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes + + //load 0~63*4(byte address) to vgpr v0 + v_mbcnt_lo_u32_b32 v0, -1, 0 + v_mbcnt_hi_u32_b32 v0, -1, v0 + v_mul_u32_u24 v0, 4, v0 + + s_lshr_b32 m0, s_wave_size, S_WAVE_SIZE + s_and_b32 m0, m0, 1 + s_cmp_eq_u32 m0, 1 + s_mov_b32 m0, 0x0 + s_cbranch_scc1 L_SAVE_LDS_W64 + +L_SAVE_LDS_W32: + s_mov_b32 s3, 128 + s_nop 0 + s_nop 0 + s_nop 0 +L_SAVE_LDS_LOOP_W32: + ds_read_b32 v1, v0 + s_waitcnt 0 + buffer_store_dword v1, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 + + s_add_u32 m0, m0, s3 //every buffer_store_lds does 256 bytes + s_add_u32 s_save_mem_offset, s_save_mem_offset, s3 + v_add_nc_u32 v0, v0, 128 //mem offset increased by 128 bytes + s_cmp_lt_u32 m0, s_save_alloc_size //scc=(m0 < s_save_alloc_size) ? 1 : 0 + s_cbranch_scc1 L_SAVE_LDS_LOOP_W32 //LDS save is complete? + + s_branch L_SAVE_LDS_DONE + +L_SAVE_LDS_W64: + s_mov_b32 s3, 256 + s_nop 0 + s_nop 0 + s_nop 0 +L_SAVE_LDS_LOOP_W64: + ds_read_b32 v1, v0 + s_waitcnt 0 + buffer_store_dword v1, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 + + s_add_u32 m0, m0, s3 //every buffer_store_lds does 256 bytes + s_add_u32 s_save_mem_offset, s_save_mem_offset, s3 + v_add_nc_u32 v0, v0, 256 //mem offset increased by 256 bytes + s_cmp_lt_u32 m0, s_save_alloc_size //scc=(m0 < s_save_alloc_size) ? 1 : 0 + s_cbranch_scc1 L_SAVE_LDS_LOOP_W64 //LDS save is complete? + +L_SAVE_LDS_DONE: + /* save VGPRs - set the Rest VGPRs */ +L_SAVE_VGPR: + // VGPR SR memory offset: 0 + s_mov_b32 exec_lo, 0xFFFFFFFF //need every thread from now on + s_lshr_b32 m0, s_wave_size, S_WAVE_SIZE + s_and_b32 m0, m0, 1 + s_cmp_eq_u32 m0, 1 + s_cbranch_scc1 L_ENABLE_SAVE_VGPR_EXEC_HI + s_mov_b32 s_save_mem_offset, (0+128*4) // for the rest VGPRs + s_mov_b32 exec_hi, 0x00000000 + s_branch L_SAVE_VGPR_NORMAL +L_ENABLE_SAVE_VGPR_EXEC_HI: + s_mov_b32 s_save_mem_offset, (0+256*4) // for the rest VGPRs + s_mov_b32 exec_hi, 0xFFFFFFFF +L_SAVE_VGPR_NORMAL: + s_getreg_b32 s_save_alloc_size, hwreg(HW_REG_GPR_ALLOC,SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SHIFT,SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SIZE) + s_add_u32 s_save_alloc_size, s_save_alloc_size, 1 + s_lshl_b32 s_save_alloc_size, s_save_alloc_size, 2 //Number of VGPRs = (vgpr_size + 1) * 4 (non-zero value) + //determine it is wave32 or wave64 + s_lshr_b32 m0, s_wave_size, S_WAVE_SIZE + s_and_b32 m0, m0, 1 + s_cmp_eq_u32 m0, 1 + s_cbranch_scc1 L_SAVE_VGPR_WAVE64 + + s_mov_b32 s_save_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes + + // VGPR Allocated in 4-GPR granularity + + // VGPR store using dw burst + s_mov_b32 m0, 0x4 //VGPR initial index value =4 + s_cmp_lt_u32 m0, s_save_alloc_size + s_cbranch_scc0 L_SAVE_VGPR_END + +L_SAVE_VGPR_W32_LOOP: + v_movrels_b32 v0, v0 //v0 = v[0+m0] + v_movrels_b32 v1, v1 //v1 = v[1+m0] + v_movrels_b32 v2, v2 //v2 = v[2+m0] + v_movrels_b32 v3, v3 //v3 = v[3+m0] + + buffer_store_dword v0, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 + buffer_store_dword v1, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 offset:128 + buffer_store_dword v2, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 offset:128*2 + buffer_store_dword v3, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 offset:128*3 + + s_add_u32 m0, m0, 4 //next vgpr index + s_add_u32 s_save_mem_offset, s_save_mem_offset, 128*4 //every buffer_store_dword does 128 bytes + s_cmp_lt_u32 m0, s_save_alloc_size //scc = (m0 < s_save_alloc_size) ? 1 : 0 + s_cbranch_scc1 L_SAVE_VGPR_W32_LOOP //VGPR save is complete? + + s_branch L_SAVE_VGPR_END + +L_SAVE_VGPR_WAVE64: + s_mov_b32 s_save_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes + + // VGPR store using dw burst + s_mov_b32 m0, 0x4 //VGPR initial index value =4 + s_cmp_lt_u32 m0, s_save_alloc_size + s_cbranch_scc0 L_SAVE_VGPR_END + +L_SAVE_VGPR_W64_LOOP: + v_movrels_b32 v0, v0 //v0 = v[0+m0] + v_movrels_b32 v1, v1 //v1 = v[1+m0] + v_movrels_b32 v2, v2 //v2 = v[2+m0] + v_movrels_b32 v3, v3 //v3 = v[3+m0] + + buffer_store_dword v0, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 + buffer_store_dword v1, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 offset:256 + buffer_store_dword v2, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 offset:256*2 + buffer_store_dword v3, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 offset:256*3 + + s_add_u32 m0, m0, 4 //next vgpr index + s_add_u32 s_save_mem_offset, s_save_mem_offset, 256*4 //every buffer_store_dword does 256 bytes + s_cmp_lt_u32 m0, s_save_alloc_size //scc = (m0 < s_save_alloc_size) ? 1 : 0 + s_cbranch_scc1 L_SAVE_VGPR_W64_LOOP //VGPR save is complete? + + //Below part will be the save shared vgpr part (new for gfx10) + s_getreg_b32 s_save_alloc_size, hwreg(HW_REG_LDS_ALLOC,SQ_WAVE_LDS_ALLOC_VGPR_SHARED_SIZE_SHIFT,SQ_WAVE_LDS_ALLOC_VGPR_SHARED_SIZE_SIZE) + s_and_b32 s_save_alloc_size, s_save_alloc_size, 0xFFFFFFFF //shared_vgpr_size is zero? + s_cbranch_scc0 L_SAVE_VGPR_END //no shared_vgpr used? jump to L_SAVE_LDS + s_lshl_b32 s_save_alloc_size, s_save_alloc_size, 3 //Number of SHARED_VGPRs = shared_vgpr_size * 8 (non-zero value) + //m0 now has the value of normal vgpr count, just add the m0 with shared_vgpr count to get the total count. + //save shared_vgpr will start from the index of m0 + s_add_u32 s_save_alloc_size, s_save_alloc_size, m0 + s_mov_b32 exec_lo, 0xFFFFFFFF + s_mov_b32 exec_hi, 0x00000000 +L_SAVE_SHARED_VGPR_WAVE64_LOOP: + v_movrels_b32 v0, v0 //v0 = v[0+m0] + buffer_store_dword v0, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 + s_add_u32 m0, m0, 1 //next vgpr index + s_add_u32 s_save_mem_offset, s_save_mem_offset, 128 + s_cmp_lt_u32 m0, s_save_alloc_size //scc = (m0 < s_save_alloc_size) ? 1 : 0 + s_cbranch_scc1 L_SAVE_SHARED_VGPR_WAVE64_LOOP //SHARED_VGPR save is complete? + +L_SAVE_VGPR_END: + s_branch L_END_PGM L_RESTORE: - /* Setup Resource Contants */ - if ((EMU_RUN_HACK) && (!EMU_RUN_HACK_RESTORE_NORMAL)) - //calculate wd_addr using absolute thread id - v_readlane_b32 s_restore_tmp, v9, 0 - //determine it is wave32 or wave64 - s_getreg_b32 s_restore_size, hwreg(HW_REG_IB_STS2,SQ_WAVE_IB_STS2_WAVE64_SHIFT,SQ_WAVE_IB_STS2_WAVE64_SIZE) //change to ttmp13 - s_cmp_eq_u32 s_restore_size, 0 - s_cbranch_scc1 L_RESTORE_WAVE32 - s_lshr_b32 s_restore_tmp, s_restore_tmp, 6 //SAVE WAVE64 - s_branch L_RESTORE_CON - L_RESTORE_WAVE32: - s_lshr_b32 s_restore_tmp, s_restore_tmp, 5 //SAVE WAVE32 - L_RESTORE_CON: - s_mul_i32 s_restore_tmp, s_restore_tmp, WAVE_SPACE - s_add_i32 s_restore_spi_init_lo, s_restore_tmp, WG_BASE_ADDR_LO - s_mov_b32 s_restore_spi_init_hi, WG_BASE_ADDR_HI - s_and_b32 s_restore_spi_init_hi, s_restore_spi_init_hi, CTX_RESTORE_CONTROL - else - end - - s_mov_b32 s_restore_buf_rsrc0, s_restore_spi_init_lo //base_addr_lo - s_and_b32 s_restore_buf_rsrc1, s_restore_spi_init_hi, 0x0000FFFF //base_addr_hi - s_or_b32 s_restore_buf_rsrc1, s_restore_buf_rsrc1, S_RESTORE_BUF_RSRC_WORD1_STRIDE - s_mov_b32 s_restore_buf_rsrc2, 0 //NUM_RECORDS initial value = 0 (in bytes) - s_mov_b32 s_restore_buf_rsrc3, S_RESTORE_BUF_RSRC_WORD3_MISC - s_and_b32 s_restore_tmp, s_restore_spi_init_hi, S_RESTORE_SPI_INIT_ATC_MASK - s_lshr_b32 s_restore_tmp, s_restore_tmp, (S_RESTORE_SPI_INIT_ATC_SHIFT-SQ_BUF_RSRC_WORD1_ATC_SHIFT) //get ATC bit into position - s_or_b32 s_restore_buf_rsrc3, s_restore_buf_rsrc3, s_restore_tmp //or ATC - s_and_b32 s_restore_tmp, s_restore_spi_init_hi, S_RESTORE_SPI_INIT_MTYPE_MASK - s_lshr_b32 s_restore_tmp, s_restore_tmp, (S_RESTORE_SPI_INIT_MTYPE_SHIFT-SQ_BUF_RSRC_WORD3_MTYPE_SHIFT) //get MTYPE bits into position - s_or_b32 s_restore_buf_rsrc3, s_restore_buf_rsrc3, s_restore_tmp //or MTYPE - //determine it is wave32 or wave64 - s_getreg_b32 s_restore_size, hwreg(HW_REG_IB_STS2,SQ_WAVE_IB_STS2_WAVE64_SHIFT,SQ_WAVE_IB_STS2_WAVE64_SIZE) - s_or_b32 s_restore_size, s_restore_spi_init_hi, s_restore_size //share s_wave_size with exec_hi - - /* global mem offset */ - s_mov_b32 s_restore_mem_offset, 0x0 //mem offset initial value = 0 - - /* restore VGPRs */ - ////////////////////////////// - L_RESTORE_VGPR: - - s_mov_b32 exec_lo, 0xFFFFFFFF //need every thread from now on //be consistent with SAVE although can be moved ahead - s_and_b32 m0, s_restore_size, 1 - s_cmp_eq_u32 m0, 1 - s_cbranch_scc1 L_ENABLE_RESTORE_VGPR_EXEC_HI - s_mov_b32 exec_hi, 0x00000000 - s_branch L_RESTORE_VGPR_NORMAL - L_ENABLE_RESTORE_VGPR_EXEC_HI: - s_mov_b32 exec_hi, 0xFFFFFFFF - L_RESTORE_VGPR_NORMAL: - s_getreg_b32 s_restore_alloc_size, hwreg(HW_REG_GPR_ALLOC,SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SHIFT,SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SIZE) //vpgr_size - s_add_u32 s_restore_alloc_size, s_restore_alloc_size, 1 - s_lshl_b32 s_restore_alloc_size, s_restore_alloc_size, 2 //Number of VGPRs = (vgpr_size + 1) * 4 (non-zero value) - //determine it is wave32 or wave64 - s_and_b32 m0, s_restore_size, 1 - s_cmp_eq_u32 m0, 1 - s_cbranch_scc1 L_RESTORE_VGPR_WAVE64 - - s_lshl_b32 s_restore_buf_rsrc2, s_restore_alloc_size, 7 //NUM_RECORDS in bytes (32 threads*4) - if (SWIZZLE_EN) - s_add_u32 s_restore_buf_rsrc2, s_restore_buf_rsrc2, 0x0 //FIXME need to use swizzle to enable bounds checking? - else - s_mov_b32 s_restore_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - end - - s_mov_b32 s_restore_mem_offset_save, s_restore_mem_offset // restore start with v1, v0 will be the last - s_add_u32 s_restore_mem_offset, s_restore_mem_offset, 128 - s_mov_b32 m0, 1 //VGPR initial index value = 1 - //s_set_gpr_idx_on m0, 0x8 //M0[7:0] = M0[7:0] and M0[15:12] = 0x8 - //s_add_u32 s_restore_alloc_size, s_restore_alloc_size, 0x8000 //add 0x8000 since we compare m0 against it later, might not need this in gfx10 - - L_RESTORE_VGPR_WAVE32_LOOP: - if(USE_MTBUF_INSTEAD_OF_MUBUF) - tbuffer_load_format_x v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset format:BUF_NUM_FORMAT_FLOAT format: BUF_DATA_FORMAT_32 slc:1 glc:1 - else - buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 - end - s_waitcnt vmcnt(0) //ensure data ready - v_movreld_b32 v0, v0 //v[0+m0] = v0 - s_add_u32 m0, m0, 1 //next vgpr index - s_add_u32 s_restore_mem_offset, s_restore_mem_offset, 128 //every buffer_load_dword does 128 bytes - s_cmp_lt_u32 m0, s_restore_alloc_size //scc = (m0 < s_restore_alloc_size) ? 1 : 0 - s_cbranch_scc1 L_RESTORE_VGPR_WAVE32_LOOP //VGPR restore (except v0) is complete? - //s_set_gpr_idx_off - /* VGPR restore on v0 */ - if(USE_MTBUF_INSTEAD_OF_MUBUF) - tbuffer_load_format_x v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save format:BUF_NUM_FORMAT_FLOAT format: BUF_DATA_FORMAT_32 slc:1 glc:1 - else - buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save slc:1 glc:1 - end - - s_branch L_RESTORE_LDS - - L_RESTORE_VGPR_WAVE64: - s_lshl_b32 s_restore_buf_rsrc2, s_restore_alloc_size, 8 //NUM_RECORDS in bytes (64 threads*4) - if (SWIZZLE_EN) - s_add_u32 s_restore_buf_rsrc2, s_restore_buf_rsrc2, 0x0 //FIXME need to use swizzle to enable bounds checking? - else - s_mov_b32 s_restore_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - end - - s_mov_b32 s_restore_mem_offset_save, s_restore_mem_offset // restore start with v1, v0 will be the last - s_add_u32 s_restore_mem_offset, s_restore_mem_offset, 256 - s_mov_b32 m0, 1 //VGPR initial index value = 1 - L_RESTORE_VGPR_WAVE64_LOOP: - if(USE_MTBUF_INSTEAD_OF_MUBUF) - tbuffer_load_format_x v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset format:BUF_NUM_FORMAT_FLOAT format: BUF_DATA_FORMAT_32 slc:1 glc:1 - else - buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 - end - s_waitcnt vmcnt(0) //ensure data ready - v_movreld_b32 v0, v0 //v[0+m0] = v0 - s_add_u32 m0, m0, 1 //next vgpr index - s_add_u32 s_restore_mem_offset, s_restore_mem_offset, 256 //every buffer_load_dword does 256 bytes - s_cmp_lt_u32 m0, s_restore_alloc_size //scc = (m0 < s_restore_alloc_size) ? 1 : 0 - s_cbranch_scc1 L_RESTORE_VGPR_WAVE64_LOOP //VGPR restore (except v0) is complete? - //s_set_gpr_idx_off - // - //Below part will be the restore shared vgpr part (new for gfx10) - s_getreg_b32 s_restore_alloc_size, hwreg(HW_REG_LDS_ALLOC,SQ_WAVE_LDS_ALLOC_VGPR_SHARED_SIZE_SHIFT,SQ_WAVE_LDS_ALLOC_VGPR_SHARED_SIZE_SIZE) //shared_vgpr_size - s_and_b32 s_restore_alloc_size, s_restore_alloc_size, 0xFFFFFFFF //shared_vgpr_size is zero? - s_cbranch_scc0 L_RESTORE_V0 //no shared_vgpr used? jump to L_SAVE_LDS - s_lshl_b32 s_restore_alloc_size, s_restore_alloc_size, 3 //Number of SHARED_VGPRs = shared_vgpr_size * 8 (non-zero value) - //m0 now has the value of normal vgpr count, just add the m0 with shared_vgpr count to get the total count. - //restore shared_vgpr will start from the index of m0 - s_add_u32 s_restore_alloc_size, s_restore_alloc_size, m0 - s_mov_b32 exec_lo, 0xFFFFFFFF - s_mov_b32 exec_hi, 0x00000000 - L_RESTORE_SHARED_VGPR_WAVE64_LOOP: - buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 - s_waitcnt vmcnt(0) //ensure data ready - v_movreld_b32 v0, v0 //v[0+m0] = v0 - s_add_u32 m0, m0, 1 //next vgpr index - s_add_u32 s_restore_mem_offset, s_restore_mem_offset, 128 //every buffer_load_dword does 256 bytes - s_cmp_lt_u32 m0, s_restore_alloc_size //scc = (m0 < s_restore_alloc_size) ? 1 : 0 - s_cbranch_scc1 L_RESTORE_SHARED_VGPR_WAVE64_LOOP //VGPR restore (except v0) is complete? - - s_mov_b32 exec_hi, 0xFFFFFFFF //restore back exec_hi before restoring V0!! - - /* VGPR restore on v0 */ - L_RESTORE_V0: - if(USE_MTBUF_INSTEAD_OF_MUBUF) - tbuffer_load_format_x v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save format:BUF_NUM_FORMAT_FLOAT format: BUF_DATA_FORMAT_32 slc:1 glc:1 - else - buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save slc:1 glc:1 - end - - - /* restore LDS */ - ////////////////////////////// - L_RESTORE_LDS: - - //Only need to check the first wave - /* the first wave in the threadgroup */ - s_and_b32 s_restore_tmp, s_restore_size, S_RESTORE_SPI_INIT_FIRST_WAVE_MASK - s_cbranch_scc0 L_RESTORE_SGPR - - s_mov_b32 exec_lo, 0xFFFFFFFF //need every thread from now on //be consistent with SAVE although can be moved ahead - s_and_b32 m0, s_restore_size, 1 - s_cmp_eq_u32 m0, 1 - s_cbranch_scc1 L_ENABLE_RESTORE_LDS_EXEC_HI - s_mov_b32 exec_hi, 0x00000000 - s_branch L_RESTORE_LDS_NORMAL - L_ENABLE_RESTORE_LDS_EXEC_HI: - s_mov_b32 exec_hi, 0xFFFFFFFF - L_RESTORE_LDS_NORMAL: - s_getreg_b32 s_restore_alloc_size, hwreg(HW_REG_LDS_ALLOC,SQ_WAVE_LDS_ALLOC_LDS_SIZE_SHIFT,SQ_WAVE_LDS_ALLOC_LDS_SIZE_SIZE) //lds_size - s_and_b32 s_restore_alloc_size, s_restore_alloc_size, 0xFFFFFFFF //lds_size is zero? - s_cbranch_scc0 L_RESTORE_SGPR //no lds used? jump to L_RESTORE_VGPR - s_lshl_b32 s_restore_alloc_size, s_restore_alloc_size, 6 //LDS size in dwords = lds_size * 64dw - s_lshl_b32 s_restore_alloc_size, s_restore_alloc_size, 2 //LDS size in bytes - s_mov_b32 s_restore_buf_rsrc2, s_restore_alloc_size //NUM_RECORDS in bytes - if (SWIZZLE_EN) - s_add_u32 s_restore_buf_rsrc2, s_restore_buf_rsrc2, 0x0 //FIXME need to use swizzle to enable bounds checking? - else - s_mov_b32 s_restore_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - end - - s_and_b32 m0, s_wave_size, 1 - s_cmp_eq_u32 m0, 1 - s_mov_b32 m0, 0x0 - s_cbranch_scc1 L_RESTORE_LDS_LOOP_W64 - - L_RESTORE_LDS_LOOP_W32: - if (SAVE_LDS) - buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset lds:1 - s_waitcnt 0 - end - s_add_u32 m0, m0, 128 //every buffer_load_dword does 256 bytes - s_add_u32 s_restore_mem_offset, s_restore_mem_offset, 128 //mem offset increased by 256 bytes - s_cmp_lt_u32 m0, s_restore_alloc_size //scc=(m0 < s_restore_alloc_size) ? 1 : 0 - s_cbranch_scc1 L_RESTORE_LDS_LOOP_W32 //LDS restore is complete? - s_branch L_RESTORE_SGPR - - L_RESTORE_LDS_LOOP_W64: - if (SAVE_LDS) - buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset lds:1 - s_waitcnt 0 - end - s_add_u32 m0, m0, 256 //every buffer_load_dword does 256 bytes - s_add_u32 s_restore_mem_offset, s_restore_mem_offset, 256 //mem offset increased by 256 bytes - s_cmp_lt_u32 m0, s_restore_alloc_size //scc=(m0 < s_restore_alloc_size) ? 1 : 0 - s_cbranch_scc1 L_RESTORE_LDS_LOOP_W64 //LDS restore is complete? - - - /* restore SGPRs */ - ////////////////////////////// - //s_getreg_b32 s_restore_alloc_size, hwreg(HW_REG_GPR_ALLOC,SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SHIFT,SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SIZE) //spgr_size - //s_add_u32 s_restore_alloc_size, s_restore_alloc_size, 1 - //s_lshl_b32 s_restore_alloc_size, s_restore_alloc_size, 4 //Number of SGPRs = (sgpr_size + 1) * 16 (non-zero value) - //s_lshl_b32 s_restore_alloc_size, s_restore_alloc_size, 3 //Number of SGPRs = (sgpr_size + 1) * 8 (non-zero value) - L_RESTORE_SGPR: - //need to look at it is wave32 or wave64 - s_and_b32 m0, s_restore_size, 1 - s_cmp_eq_u32 m0, 1 - s_cbranch_scc1 L_RESTORE_SGPR_VMEM_WAVE64 - if (SGPR_SAVE_USE_SQC) - s_lshl_b32 s_restore_buf_rsrc2, s_sgpr_save_num, 2 //NUM_RECORDS in bytes - else - s_lshl_b32 s_restore_buf_rsrc2, s_sgpr_save_num, 7 //NUM_RECORDS in bytes (32 threads) - end - s_branch L_RESTORE_SGPR_CONT - L_RESTORE_SGPR_VMEM_WAVE64: - if (SGPR_SAVE_USE_SQC) - s_lshl_b32 s_restore_buf_rsrc2, s_sgpr_save_num, 2 //NUM_RECORDS in bytes - else - s_lshl_b32 s_restore_buf_rsrc2, s_sgpr_save_num, 8 //NUM_RECORDS in bytes (64 threads) - end - - L_RESTORE_SGPR_CONT: - if (SWIZZLE_EN) - s_add_u32 s_restore_buf_rsrc2, s_restore_buf_rsrc2, 0x0 //FIXME need to use swizzle to enable bounds checking? - else - s_mov_b32 s_restore_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - end - - s_and_b32 m0, s_restore_size, 1 - s_cmp_eq_u32 m0, 1 - s_cbranch_scc1 L_RESTORE_SGPR_WAVE64 - - read_sgpr_from_mem_wave32(s_restore_tmp, s_restore_buf_rsrc0, s_restore_mem_offset, SGPR_SAVE_USE_SQC) //save s0 to s_restore_tmp - s_mov_b32 m0, 0x1 - - L_RESTORE_SGPR_LOOP_WAVE32: - read_sgpr_from_mem_wave32(s0, s_restore_buf_rsrc0, s_restore_mem_offset, SGPR_SAVE_USE_SQC) //PV: further performance improvement can be made - s_waitcnt lgkmcnt(0) //ensure data ready - s_movreld_b32 s0, s0 //s[0+m0] = s0 - s_nop 0 // hazard SALU M0=> S_MOVREL - s_add_u32 m0, m0, 1 //next sgpr index - s_cmp_lt_u32 m0, s_sgpr_save_num //scc = (m0 < s_restore_alloc_size) ? 1 : 0 - s_cbranch_scc1 L_RESTORE_SGPR_LOOP_WAVE32 //SGPR restore (except s0) is complete? - s_mov_b32 s0, s_restore_tmp /* SGPR restore on s0 */ - s_branch L_RESTORE_HWREG - - L_RESTORE_SGPR_WAVE64: - read_sgpr_from_mem_wave64(s_restore_tmp, s_restore_buf_rsrc0, s_restore_mem_offset, SGPR_SAVE_USE_SQC) //save s0 to s_restore_tmp - s_mov_b32 m0, 0x1 //SGPR initial index value =1 //go on with with s1 - - L_RESTORE_SGPR_LOOP_WAVE64: - read_sgpr_from_mem_wave64(s0, s_restore_buf_rsrc0, s_restore_mem_offset, SGPR_SAVE_USE_SQC) //PV: further performance improvement can be made - s_waitcnt lgkmcnt(0) //ensure data ready - s_movreld_b32 s0, s0 //s[0+m0] = s0 - s_nop 0 // hazard SALU M0=> S_MOVREL - s_add_u32 m0, m0, 1 //next sgpr index - s_cmp_lt_u32 m0, s_sgpr_save_num //scc = (m0 < s_restore_alloc_size) ? 1 : 0 - s_cbranch_scc1 L_RESTORE_SGPR_LOOP_WAVE64 //SGPR restore (except s0) is complete? - s_mov_b32 s0, s_restore_tmp /* SGPR restore on s0 */ - - - /* restore HW registers */ - ////////////////////////////// - L_RESTORE_HWREG: - s_mov_b32 s_restore_buf_rsrc2, 0x4 //NUM_RECORDS in bytes - if (SWIZZLE_EN) - s_add_u32 s_restore_buf_rsrc2, s_restore_buf_rsrc2, 0x0 //FIXME need to use swizzle to enable bounds checking? - else - s_mov_b32 s_restore_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - end - - s_and_b32 m0, s_restore_size, 1 - s_cmp_eq_u32 m0, 1 - s_cbranch_scc1 L_RESTORE_HWREG_WAVE64 - - read_sgpr_from_mem_wave32(s_restore_m0, s_restore_buf_rsrc0, s_restore_mem_offset, SGPR_SAVE_USE_SQC) //M0 - read_sgpr_from_mem_wave32(s_restore_pc_lo, s_restore_buf_rsrc0, s_restore_mem_offset, SGPR_SAVE_USE_SQC) //PC - read_sgpr_from_mem_wave32(s_restore_pc_hi, s_restore_buf_rsrc0, s_restore_mem_offset, SGPR_SAVE_USE_SQC) - read_sgpr_from_mem_wave32(s_restore_exec_lo, s_restore_buf_rsrc0, s_restore_mem_offset, SGPR_SAVE_USE_SQC) //EXEC - read_sgpr_from_mem_wave32(s_restore_exec_hi, s_restore_buf_rsrc0, s_restore_mem_offset, SGPR_SAVE_USE_SQC) - read_sgpr_from_mem_wave32(s_restore_status, s_restore_buf_rsrc0, s_restore_mem_offset, SGPR_SAVE_USE_SQC) //STATUS - read_sgpr_from_mem_wave32(s_restore_trapsts, s_restore_buf_rsrc0, s_restore_mem_offset, SGPR_SAVE_USE_SQC) //TRAPSTS - //read_sgpr_from_mem_wave32(xnack_mask_lo, s_restore_buf_rsrc0, s_restore_mem_offset, SGPR_SAVE_USE_SQC) //XNACK_MASK_LO - //read_sgpr_from_mem_wave32(xnack_mask_hi, s_restore_buf_rsrc0, s_restore_mem_offset, SGPR_SAVE_USE_SQC) //XNACK_MASK_HI - read_sgpr_from_mem_wave32(s_restore_xnack_mask, s_restore_buf_rsrc0, s_restore_mem_offset, SGPR_SAVE_USE_SQC) //XNACK_MASK - read_sgpr_from_mem_wave32(s_restore_mode, s_restore_buf_rsrc0, s_restore_mem_offset, SGPR_SAVE_USE_SQC) //MODE - if(SAVE_RESTORE_HWID_DDID) - read_sgpr_from_mem_wave32(s_restore_hwid1, s_restore_buf_rsrc0, s_restore_mem_offset, SGPR_SAVE_USE_SQC) //HW_ID1 - end - s_branch L_RESTORE_HWREG_FINISH - - L_RESTORE_HWREG_WAVE64: - read_sgpr_from_mem_wave64(s_restore_m0, s_restore_buf_rsrc0, s_restore_mem_offset, SGPR_SAVE_USE_SQC) //M0 - read_sgpr_from_mem_wave64(s_restore_pc_lo, s_restore_buf_rsrc0, s_restore_mem_offset, SGPR_SAVE_USE_SQC) //PC - read_sgpr_from_mem_wave64(s_restore_pc_hi, s_restore_buf_rsrc0, s_restore_mem_offset, SGPR_SAVE_USE_SQC) - read_sgpr_from_mem_wave64(s_restore_exec_lo, s_restore_buf_rsrc0, s_restore_mem_offset, SGPR_SAVE_USE_SQC) //EXEC - read_sgpr_from_mem_wave64(s_restore_exec_hi, s_restore_buf_rsrc0, s_restore_mem_offset, SGPR_SAVE_USE_SQC) - read_sgpr_from_mem_wave64(s_restore_status, s_restore_buf_rsrc0, s_restore_mem_offset, SGPR_SAVE_USE_SQC) //STATUS - read_sgpr_from_mem_wave64(s_restore_trapsts, s_restore_buf_rsrc0, s_restore_mem_offset, SGPR_SAVE_USE_SQC) //TRAPSTS - //read_sgpr_from_mem_wave64(xnack_mask_lo, s_restore_buf_rsrc0, s_restore_mem_offset, SGPR_SAVE_USE_SQC) //XNACK_MASK_LO - //read_sgpr_from_mem_wave64(xnack_mask_hi, s_restore_buf_rsrc0, s_restore_mem_offset, SGPR_SAVE_USE_SQC) //XNACK_MASK_HI - read_sgpr_from_mem_wave64(s_restore_xnack_mask, s_restore_buf_rsrc0, s_restore_mem_offset, SGPR_SAVE_USE_SQC) //XNACK_MASK - read_sgpr_from_mem_wave64(s_restore_mode, s_restore_buf_rsrc0, s_restore_mem_offset, SGPR_SAVE_USE_SQC) //MODE - if(SAVE_RESTORE_HWID_DDID) - read_sgpr_from_mem_wave64(s_restore_hwid1, s_restore_buf_rsrc0, s_restore_mem_offset, SGPR_SAVE_USE_SQC) //HW_ID1 - end - L_RESTORE_HWREG_FINISH: - s_waitcnt lgkmcnt(0) //from now on, it is safe to restore STATUS and IB_STS - - - - if(SAVE_RESTORE_HWID_DDID) - L_RESTORE_DDID: - s_mov_b32 m0, s_restore_hwid1 //virture ttrace support: The save-context handler records the SE/SA/WGP/SIMD/wave of the original wave - s_ttracedata //and then can output it as SHADER_DATA to ttrace on restore to provide a correlation across the save-restore - - s_mov_b32 s_restore_buf_rsrc2, 0x4 //NUM_RECORDS in bytes - if (SWIZZLE_EN) - s_add_u32 s_restore_buf_rsrc2, s_restore_buf_rsrc2, 0x0 //FIXME need to use swizzle to enable bounds checking? - else - s_mov_b32 s_restore_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - end - - s_and_b32 m0, s_restore_size, 1 - s_cmp_eq_u32 m0, 1 - s_cbranch_scc1 L_RESTORE_DDID_WAVE64 - - read_sgpr_from_mem_wave32(s_restore_ddid, s_restore_buf_rsrc0, s_restore_mem_offset, SGPR_SAVE_USE_SQC) - s_branch L_RESTORE_DDID_FINISH - L_RESTORE_DDID_WAVE64: - read_sgpr_from_mem_wave64(s_restore_ddid, s_restore_buf_rsrc0, s_restore_mem_offset, SGPR_SAVE_USE_SQC) - - L_RESTORE_DDID_FINISH: - s_waitcnt lgkmcnt(0) - //s_mov_b32 m0, s_restore_ddid - //s_ttracedata - if (RESTORE_DDID_IN_SGPR18) - s_mov_b32 s18, s_restore_ddid - end - - end - - s_and_b32 s_restore_pc_hi, s_restore_pc_hi, 0x0000ffff //pc[47:32] //Do it here in order not to affect STATUS - - //for normal save & restore, the saved PC points to the next inst to execute, no adjustment needs to be made, otherwise: - if ((EMU_RUN_HACK) && (!EMU_RUN_HACK_RESTORE_NORMAL)) - s_add_u32 s_restore_pc_lo, s_restore_pc_lo, 8 //pc[31:0]+8 //two back-to-back s_trap are used (first for save and second for restore) - s_addc_u32 s_restore_pc_hi, s_restore_pc_hi, 0x0 //carry bit over - end - if ((EMU_RUN_HACK) && (EMU_RUN_HACK_RESTORE_NORMAL)) - s_add_u32 s_restore_pc_lo, s_restore_pc_lo, 4 //pc[31:0]+4 // save is hack through s_trap but restore is normal - s_addc_u32 s_restore_pc_hi, s_restore_pc_hi, 0x0 //carry bit over - end - - s_mov_b32 m0, s_restore_m0 - s_mov_b32 exec_lo, s_restore_exec_lo - s_mov_b32 exec_hi, s_restore_exec_hi - - s_and_b32 s_restore_m0, SQ_WAVE_TRAPSTS_PRE_SAVECTX_MASK, s_restore_trapsts + /* Setup Resource Contants */ + s_mov_b32 s_restore_buf_rsrc0, s_restore_spi_init_lo //base_addr_lo + s_and_b32 s_restore_buf_rsrc1, s_restore_spi_init_hi, 0x0000FFFF //base_addr_hi + s_or_b32 s_restore_buf_rsrc1, s_restore_buf_rsrc1, S_RESTORE_BUF_RSRC_WORD1_STRIDE + s_mov_b32 s_restore_buf_rsrc2, 0 //NUM_RECORDS initial value = 0 (in bytes) + s_mov_b32 s_restore_buf_rsrc3, S_RESTORE_BUF_RSRC_WORD3_MISC + s_and_b32 s_restore_tmp, s_restore_spi_init_hi, S_RESTORE_SPI_INIT_ATC_MASK + s_lshr_b32 s_restore_tmp, s_restore_tmp, (S_RESTORE_SPI_INIT_ATC_SHIFT-SQ_BUF_RSRC_WORD1_ATC_SHIFT) + s_or_b32 s_restore_buf_rsrc3, s_restore_buf_rsrc3, s_restore_tmp //or ATC + s_and_b32 s_restore_tmp, s_restore_spi_init_hi, S_RESTORE_SPI_INIT_MTYPE_MASK + s_lshr_b32 s_restore_tmp, s_restore_tmp, (S_RESTORE_SPI_INIT_MTYPE_SHIFT-SQ_BUF_RSRC_WORD3_MTYPE_SHIFT) + s_or_b32 s_restore_buf_rsrc3, s_restore_buf_rsrc3, s_restore_tmp //or MTYPE + //determine it is wave32 or wave64 + s_getreg_b32 s_restore_size, hwreg(HW_REG_IB_STS2,SQ_WAVE_IB_STS2_WAVE64_SHIFT,SQ_WAVE_IB_STS2_WAVE64_SIZE) + s_lshl_b32 s_restore_size, s_restore_size, S_WAVE_SIZE + s_or_b32 s_restore_size, s_restore_spi_init_hi, s_restore_size + + s_and_b32 s_restore_tmp, s_restore_spi_init_hi, S_RESTORE_SPI_INIT_FIRST_WAVE_MASK + s_cbranch_scc0 L_RESTORE_VGPR + + /* restore LDS */ +L_RESTORE_LDS: + s_mov_b32 exec_lo, 0xFFFFFFFF //need every thread from now on + s_lshr_b32 m0, s_restore_size, S_WAVE_SIZE + s_and_b32 m0, m0, 1 + s_cmp_eq_u32 m0, 1 + s_cbranch_scc1 L_ENABLE_RESTORE_LDS_EXEC_HI + s_mov_b32 exec_hi, 0x00000000 + s_branch L_RESTORE_LDS_NORMAL +L_ENABLE_RESTORE_LDS_EXEC_HI: + s_mov_b32 exec_hi, 0xFFFFFFFF +L_RESTORE_LDS_NORMAL: + s_getreg_b32 s_restore_alloc_size, hwreg(HW_REG_LDS_ALLOC,SQ_WAVE_LDS_ALLOC_LDS_SIZE_SHIFT,SQ_WAVE_LDS_ALLOC_LDS_SIZE_SIZE) + s_and_b32 s_restore_alloc_size, s_restore_alloc_size, 0xFFFFFFFF //lds_size is zero? + s_cbranch_scc0 L_RESTORE_VGPR //no lds used? jump to L_RESTORE_VGPR + s_lshl_b32 s_restore_alloc_size, s_restore_alloc_size, 6 //LDS size in dwords = lds_size * 64dw + s_lshl_b32 s_restore_alloc_size, s_restore_alloc_size, 2 //LDS size in bytes + s_mov_b32 s_restore_buf_rsrc2, s_restore_alloc_size //NUM_RECORDS in bytes + + // LDS at offset: size(VGPR)+size(SVGPR)+SIZE(SGPR)+SIZE(HWREG) + // + get_vgpr_size_bytes(s_restore_mem_offset, s_restore_size) + get_svgpr_size_bytes(s_restore_tmp) + s_add_u32 s_restore_mem_offset, s_restore_mem_offset, s_restore_tmp + s_add_u32 s_restore_mem_offset, s_restore_mem_offset, get_sgpr_size_bytes() + s_add_u32 s_restore_mem_offset, s_restore_mem_offset, get_hwreg_size_bytes() + + s_mov_b32 s_restore_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes + + s_lshr_b32 m0, s_wave_size, S_WAVE_SIZE + s_and_b32 m0, m0, 1 + s_cmp_eq_u32 m0, 1 + s_mov_b32 m0, 0x0 + s_cbranch_scc1 L_RESTORE_LDS_LOOP_W64 + +L_RESTORE_LDS_LOOP_W32: + buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset lds:1 // first 64DW + s_add_u32 m0, m0, 128 // 128 DW + s_add_u32 s_restore_mem_offset, s_restore_mem_offset, 128 //mem offset increased by 128DW + s_cmp_lt_u32 m0, s_restore_alloc_size //scc=(m0 < s_restore_alloc_size) ? 1 : 0 + s_cbranch_scc1 L_RESTORE_LDS_LOOP_W32 //LDS restore is complete? + s_branch L_RESTORE_VGPR + +L_RESTORE_LDS_LOOP_W64: + buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset lds:1 // first 64DW + s_add_u32 m0, m0, 256 // 256 DW + s_add_u32 s_restore_mem_offset, s_restore_mem_offset, 256 //mem offset increased by 256DW + s_cmp_lt_u32 m0, s_restore_alloc_size //scc=(m0 < s_restore_alloc_size) ? 1 : 0 + s_cbranch_scc1 L_RESTORE_LDS_LOOP_W64 //LDS restore is complete? + + /* restore VGPRs */ +L_RESTORE_VGPR: + // VGPR SR memory offset : 0 + s_mov_b32 s_restore_mem_offset, 0x0 + s_mov_b32 exec_lo, 0xFFFFFFFF //need every thread from now on + s_lshr_b32 m0, s_restore_size, S_WAVE_SIZE + s_and_b32 m0, m0, 1 + s_cmp_eq_u32 m0, 1 + s_cbranch_scc1 L_ENABLE_RESTORE_VGPR_EXEC_HI + s_mov_b32 exec_hi, 0x00000000 + s_branch L_RESTORE_VGPR_NORMAL +L_ENABLE_RESTORE_VGPR_EXEC_HI: + s_mov_b32 exec_hi, 0xFFFFFFFF +L_RESTORE_VGPR_NORMAL: + s_getreg_b32 s_restore_alloc_size, hwreg(HW_REG_GPR_ALLOC,SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SHIFT,SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SIZE) + s_add_u32 s_restore_alloc_size, s_restore_alloc_size, 1 + s_lshl_b32 s_restore_alloc_size, s_restore_alloc_size, 2 //Number of VGPRs = (vgpr_size + 1) * 4 (non-zero value) + //determine it is wave32 or wave64 + s_lshr_b32 m0, s_restore_size, S_WAVE_SIZE + s_and_b32 m0, m0, 1 + s_cmp_eq_u32 m0, 1 + s_cbranch_scc1 L_RESTORE_VGPR_WAVE64 + + s_mov_b32 s_restore_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes + + // VGPR load using dw burst + s_mov_b32 s_restore_mem_offset_save, s_restore_mem_offset // restore start with v1, v0 will be the last + s_add_u32 s_restore_mem_offset, s_restore_mem_offset, 128*4 + s_mov_b32 m0, 4 //VGPR initial index value = 4 + +L_RESTORE_VGPR_WAVE32_LOOP: + buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 + buffer_load_dword v1, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 offset:128 + buffer_load_dword v2, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 offset:128*2 + buffer_load_dword v3, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 offset:128*3 + s_waitcnt vmcnt(0) + v_movreld_b32 v0, v0 //v[0+m0] = v0 + v_movreld_b32 v1, v1 + v_movreld_b32 v2, v2 + v_movreld_b32 v3, v3 + s_add_u32 m0, m0, 4 //next vgpr index + s_add_u32 s_restore_mem_offset, s_restore_mem_offset, 128*4 //every buffer_load_dword does 128 bytes + s_cmp_lt_u32 m0, s_restore_alloc_size //scc = (m0 < s_restore_alloc_size) ? 1 : 0 + s_cbranch_scc1 L_RESTORE_VGPR_WAVE32_LOOP //VGPR restore (except v0) is complete? + + /* VGPR restore on v0 */ + buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save slc:1 glc:1 + buffer_load_dword v1, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save slc:1 glc:1 offset:128 + buffer_load_dword v2, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save slc:1 glc:1 offset:128*2 + buffer_load_dword v3, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save slc:1 glc:1 offset:128*3 + + s_branch L_RESTORE_SGPR + +L_RESTORE_VGPR_WAVE64: + s_mov_b32 s_restore_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes + + // VGPR load using dw burst + s_mov_b32 s_restore_mem_offset_save, s_restore_mem_offset // restore start with v4, v0 will be the last + s_add_u32 s_restore_mem_offset, s_restore_mem_offset, 256*4 + s_mov_b32 m0, 4 //VGPR initial index value = 4 + +L_RESTORE_VGPR_WAVE64_LOOP: + buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 + buffer_load_dword v1, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 offset:256 + buffer_load_dword v2, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 offset:256*2 + buffer_load_dword v3, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 offset:256*3 + s_waitcnt vmcnt(0) + v_movreld_b32 v0, v0 //v[0+m0] = v0 + v_movreld_b32 v1, v1 + v_movreld_b32 v2, v2 + v_movreld_b32 v3, v3 + s_add_u32 m0, m0, 4 //next vgpr index + s_add_u32 s_restore_mem_offset, s_restore_mem_offset, 256*4 //every buffer_load_dword does 256 bytes + s_cmp_lt_u32 m0, s_restore_alloc_size //scc = (m0 < s_restore_alloc_size) ? 1 : 0 + s_cbranch_scc1 L_RESTORE_VGPR_WAVE64_LOOP //VGPR restore (except v0) is complete? + + //Below part will be the restore shared vgpr part (new for gfx10) + s_getreg_b32 s_restore_alloc_size, hwreg(HW_REG_LDS_ALLOC,SQ_WAVE_LDS_ALLOC_VGPR_SHARED_SIZE_SHIFT,SQ_WAVE_LDS_ALLOC_VGPR_SHARED_SIZE_SIZE) //shared_vgpr_size + s_and_b32 s_restore_alloc_size, s_restore_alloc_size, 0xFFFFFFFF //shared_vgpr_size is zero? + s_cbranch_scc0 L_RESTORE_V0 //no shared_vgpr used? + s_lshl_b32 s_restore_alloc_size, s_restore_alloc_size, 3 //Number of SHARED_VGPRs = shared_vgpr_size * 8 (non-zero value) + //m0 now has the value of normal vgpr count, just add the m0 with shared_vgpr count to get the total count. + //restore shared_vgpr will start from the index of m0 + s_add_u32 s_restore_alloc_size, s_restore_alloc_size, m0 + s_mov_b32 exec_lo, 0xFFFFFFFF + s_mov_b32 exec_hi, 0x00000000 +L_RESTORE_SHARED_VGPR_WAVE64_LOOP: + buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 + s_waitcnt vmcnt(0) + v_movreld_b32 v0, v0 //v[0+m0] = v0 + s_add_u32 m0, m0, 1 //next vgpr index + s_add_u32 s_restore_mem_offset, s_restore_mem_offset, 128 + s_cmp_lt_u32 m0, s_restore_alloc_size //scc = (m0 < s_restore_alloc_size) ? 1 : 0 + s_cbranch_scc1 L_RESTORE_SHARED_VGPR_WAVE64_LOOP //VGPR restore (except v0) is complete? + + s_mov_b32 exec_hi, 0xFFFFFFFF //restore back exec_hi before restoring V0!! + + /* VGPR restore on v0 */ +L_RESTORE_V0: + buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save slc:1 glc:1 + buffer_load_dword v1, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save slc:1 glc:1 offset:256 + buffer_load_dword v2, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save slc:1 glc:1 offset:256*2 + buffer_load_dword v3, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save slc:1 glc:1 offset:256*3 + s_waitcnt vmcnt(0) + + /* restore SGPRs */ + //will be 2+8+16*6 + // SGPR SR memory offset : size(VGPR)+size(SVGPR) +L_RESTORE_SGPR: + get_vgpr_size_bytes(s_restore_mem_offset, s_restore_size) + get_svgpr_size_bytes(s_restore_tmp) + s_add_u32 s_restore_mem_offset, s_restore_mem_offset, s_restore_tmp + s_add_u32 s_restore_mem_offset, s_restore_mem_offset, get_sgpr_size_bytes() + s_sub_u32 s_restore_mem_offset, s_restore_mem_offset, 20*4 //s108~s127 is not saved + + s_mov_b32 s_restore_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes + + s_mov_b32 m0, s_sgpr_save_num + + read_4sgpr_from_mem(s0, s_restore_buf_rsrc0, s_restore_mem_offset) + s_waitcnt lgkmcnt(0) + + s_sub_u32 m0, m0, 4 // Restore from S[0] to S[104] + s_nop 0 // hazard SALU M0=> S_MOVREL + + s_movreld_b64 s0, s0 //s[0+m0] = s0 + s_movreld_b64 s2, s2 + + read_8sgpr_from_mem(s0, s_restore_buf_rsrc0, s_restore_mem_offset) + s_waitcnt lgkmcnt(0) + + s_sub_u32 m0, m0, 8 // Restore from S[0] to S[96] + s_nop 0 // hazard SALU M0=> S_MOVREL + + s_movreld_b64 s0, s0 //s[0+m0] = s0 + s_movreld_b64 s2, s2 + s_movreld_b64 s4, s4 + s_movreld_b64 s6, s6 + + L_RESTORE_SGPR_LOOP: + read_16sgpr_from_mem(s0, s_restore_buf_rsrc0, s_restore_mem_offset) + s_waitcnt lgkmcnt(0) + + s_sub_u32 m0, m0, 16 // Restore from S[n] to S[0] + s_nop 0 // hazard SALU M0=> S_MOVREL + + s_movreld_b64 s0, s0 //s[0+m0] = s0 + s_movreld_b64 s2, s2 + s_movreld_b64 s4, s4 + s_movreld_b64 s6, s6 + s_movreld_b64 s8, s8 + s_movreld_b64 s10, s10 + s_movreld_b64 s12, s12 + s_movreld_b64 s14, s14 + + s_cmp_eq_u32 m0, 0 //scc = (m0 < s_sgpr_save_num) ? 1 : 0 + s_cbranch_scc0 L_RESTORE_SGPR_LOOP + + /* restore HW registers */ +L_RESTORE_HWREG: + // HWREG SR memory offset : size(VGPR)+size(SVGPR)+size(SGPR) + get_vgpr_size_bytes(s_restore_mem_offset, s_restore_size) + get_svgpr_size_bytes(s_restore_tmp) + s_add_u32 s_restore_mem_offset, s_restore_mem_offset, s_restore_tmp + s_add_u32 s_restore_mem_offset, s_restore_mem_offset, get_sgpr_size_bytes() + + s_mov_b32 s_restore_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes + + read_hwreg_from_mem(s_restore_m0, s_restore_buf_rsrc0, s_restore_mem_offset) + read_hwreg_from_mem(s_restore_pc_lo, s_restore_buf_rsrc0, s_restore_mem_offset) + read_hwreg_from_mem(s_restore_pc_hi, s_restore_buf_rsrc0, s_restore_mem_offset) + read_hwreg_from_mem(s_restore_exec_lo, s_restore_buf_rsrc0, s_restore_mem_offset) + read_hwreg_from_mem(s_restore_exec_hi, s_restore_buf_rsrc0, s_restore_mem_offset) + read_hwreg_from_mem(s_restore_status, s_restore_buf_rsrc0, s_restore_mem_offset) + read_hwreg_from_mem(s_restore_trapsts, s_restore_buf_rsrc0, s_restore_mem_offset) + read_hwreg_from_mem(s_restore_xnack_mask, s_restore_buf_rsrc0, s_restore_mem_offset) + read_hwreg_from_mem(s_restore_mode, s_restore_buf_rsrc0, s_restore_mem_offset) + read_hwreg_from_mem(s_restore_flat_scratch, s_restore_buf_rsrc0, s_restore_mem_offset) + s_waitcnt lgkmcnt(0) + + s_setreg_b32 hwreg(HW_REG_SHADER_FLAT_SCRATCH_LO), s_restore_flat_scratch + + read_hwreg_from_mem(s_restore_flat_scratch, s_restore_buf_rsrc0, s_restore_mem_offset) + s_waitcnt lgkmcnt(0) //from now on, it is safe to restore STATUS and IB_STS + + s_setreg_b32 hwreg(HW_REG_SHADER_FLAT_SCRATCH_HI), s_restore_flat_scratch + + s_mov_b32 s_restore_tmp, s_restore_pc_hi + s_and_b32 s_restore_pc_hi, s_restore_tmp, 0x0000ffff //pc[47:32] //Do it here in order not to affect STATUS + + s_mov_b32 m0, s_restore_m0 + s_mov_b32 exec_lo, s_restore_exec_lo + s_mov_b32 exec_hi, s_restore_exec_hi + + s_and_b32 s_restore_m0, SQ_WAVE_TRAPSTS_PRE_SAVECTX_MASK, s_restore_trapsts s_setreg_b32 hwreg(HW_REG_TRAPSTS, SQ_WAVE_TRAPSTS_PRE_SAVECTX_SHIFT, SQ_WAVE_TRAPSTS_PRE_SAVECTX_SIZE), s_restore_m0 - s_setreg_b32 hwreg(HW_REG_SHADER_XNACK_MASK), s_restore_xnack_mask //restore xnack_mask - s_and_b32 s_restore_m0, SQ_WAVE_TRAPSTS_POST_SAVECTX_MASK, s_restore_trapsts - s_lshr_b32 s_restore_m0, s_restore_m0, SQ_WAVE_TRAPSTS_POST_SAVECTX_SHIFT + s_setreg_b32 hwreg(HW_REG_SHADER_XNACK_MASK), s_restore_xnack_mask + s_and_b32 s_restore_m0, SQ_WAVE_TRAPSTS_POST_SAVECTX_MASK, s_restore_trapsts + s_lshr_b32 s_restore_m0, s_restore_m0, SQ_WAVE_TRAPSTS_POST_SAVECTX_SHIFT s_setreg_b32 hwreg(HW_REG_TRAPSTS, SQ_WAVE_TRAPSTS_POST_SAVECTX_SHIFT, SQ_WAVE_TRAPSTS_POST_SAVECTX_SIZE), s_restore_m0 - //s_setreg_b32 hwreg(HW_REG_TRAPSTS), s_restore_trapsts //don't overwrite SAVECTX bit as it may be set through external SAVECTX during restore - s_setreg_b32 hwreg(HW_REG_MODE), s_restore_mode - //reuse s_restore_m0 as a temp register - s_and_b32 s_restore_m0, s_restore_pc_hi, S_SAVE_PC_HI_RCNT_MASK - s_lshr_b32 s_restore_m0, s_restore_m0, S_SAVE_PC_HI_RCNT_SHIFT - s_lshl_b32 s_restore_m0, s_restore_m0, SQ_WAVE_IB_STS_RCNT_SHIFT - s_mov_b32 s_restore_tmp, 0x0 //IB_STS is zero - s_or_b32 s_restore_tmp, s_restore_tmp, s_restore_m0 - s_and_b32 s_restore_m0, s_restore_pc_hi, S_SAVE_PC_HI_FIRST_REPLAY_MASK - s_lshr_b32 s_restore_m0, s_restore_m0, S_SAVE_PC_HI_FIRST_REPLAY_SHIFT - s_lshl_b32 s_restore_m0, s_restore_m0, SQ_WAVE_IB_STS_FIRST_REPLAY_SHIFT - s_or_b32 s_restore_tmp, s_restore_tmp, s_restore_m0 - s_and_b32 s_restore_m0, s_restore_status, SQ_WAVE_STATUS_INST_ATC_MASK - s_lshr_b32 s_restore_m0, s_restore_m0, SQ_WAVE_STATUS_INST_ATC_SHIFT - s_setreg_b32 hwreg(HW_REG_IB_STS), s_restore_tmp - s_setreg_b32 hwreg(HW_REG_STATUS), s_restore_status - - s_barrier //barrier to ensure the readiness of LDS before access attemps from any other wave in the same TG //FIXME not performance-optimal at this time - - -// s_rfe_b64 s_restore_pc_lo //Return to the main shader program and resume execution - s_rfe_b64 s_restore_pc_lo // s_restore_m0[0] is used to set STATUS.inst_atc - - -/**************************************************************************/ -/* the END */ -/**************************************************************************/ -L_END_PGM: + s_setreg_b32 hwreg(HW_REG_MODE), s_restore_mode + s_and_b32 s_restore_m0, s_restore_tmp, S_SAVE_PC_HI_RCNT_MASK + s_lshr_b32 s_restore_m0, s_restore_m0, S_SAVE_PC_HI_RCNT_SHIFT + s_lshl_b32 s_restore_m0, s_restore_m0, SQ_WAVE_IB_STS_RCNT_SHIFT + s_mov_b32 s_restore_mode, 0x0 + s_or_b32 s_restore_mode, s_restore_mode, s_restore_m0 + s_and_b32 s_restore_m0, s_restore_tmp, S_SAVE_PC_HI_FIRST_REPLAY_MASK + s_lshr_b32 s_restore_m0, s_restore_m0, S_SAVE_PC_HI_FIRST_REPLAY_SHIFT + s_lshl_b32 s_restore_m0, s_restore_m0, SQ_WAVE_IB_STS_FIRST_REPLAY_SHIFT + s_or_b32 s_restore_mode, s_restore_mode, s_restore_m0 + s_and_b32 s_restore_m0, s_restore_tmp, S_SAVE_PC_HI_REPLAY_W64H_MASK + s_lshr_b32 s_restore_m0, s_restore_m0, S_SAVE_PC_HI_REPLAY_W64H_SHIFT + s_lshl_b32 s_restore_m0, s_restore_m0, SQ_WAVE_IB_STS_REPLAY_W64H_SHIFT + s_or_b32 s_restore_mode, s_restore_mode, s_restore_m0 + + s_and_b32 s_restore_m0, s_restore_status, SQ_WAVE_STATUS_INST_ATC_MASK + s_lshr_b32 s_restore_m0, s_restore_m0, SQ_WAVE_STATUS_INST_ATC_SHIFT + s_setreg_b32 hwreg(HW_REG_IB_STS), s_restore_mode + + s_and_b64 exec, exec, exec // Restore STATUS.EXECZ, not writable by s_setreg_b32 + s_and_b64 vcc, vcc, vcc // Restore STATUS.VCCZ, not writable by s_setreg_b32 + s_setreg_b32 hwreg(HW_REG_STATUS), s_restore_status // SCC is included, which is changed by previous salu + + s_barrier //barrier to ensure the readiness of LDS before access attemps from any other wave in the same TG + + s_rfe_b64 s_restore_pc_lo //Return to the main shader program and resume execution + +L_END_PGM: s_endpgm - -end - - -/**************************************************************************/ -/* the helper functions */ -/**************************************************************************/ -function write_sgpr_to_mem_wave32(s, s_rsrc, s_mem_offset, use_sqc, use_mtbuf) - if (use_sqc) - s_mov_b32 exec_lo, m0 //assuming exec_lo is not needed anymore from this point on - s_mov_b32 m0, s_mem_offset - s_buffer_store_dword s, s_rsrc, m0 glc:1 - s_add_u32 s_mem_offset, s_mem_offset, 4 - s_mov_b32 m0, exec_lo - elsif (use_mtbuf) - v_mov_b32 v0, s - tbuffer_store_format_x v0, v0, s_rsrc, s_mem_offset format:BUF_NUM_FORMAT_FLOAT format: BUF_DATA_FORMAT_32 slc:1 glc:1 - s_add_u32 s_mem_offset, s_mem_offset, 128 - else - v_mov_b32 v0, s - buffer_store_dword v0, v0, s_rsrc, s_mem_offset slc:1 glc:1 - s_add_u32 s_mem_offset, s_mem_offset, 128 - end end -function write_sgpr_to_mem_wave64(s, s_rsrc, s_mem_offset, use_sqc, use_mtbuf) - if (use_sqc) - s_mov_b32 exec_lo, m0 //assuming exec_lo is not needed anymore from this point on - s_mov_b32 m0, s_mem_offset - s_buffer_store_dword s, s_rsrc, m0 glc:1 - s_add_u32 s_mem_offset, s_mem_offset, 4 - s_mov_b32 m0, exec_lo - elsif (use_mtbuf) - v_mov_b32 v0, s - tbuffer_store_format_x v0, v0, s_rsrc, s_mem_offset format:BUF_NUM_FORMAT_FLOAT format: BUF_DATA_FORMAT_32 slc:1 glc:1 - s_add_u32 s_mem_offset, s_mem_offset, 256 - else - v_mov_b32 v0, s - buffer_store_dword v0, v0, s_rsrc, s_mem_offset slc:1 glc:1 - s_add_u32 s_mem_offset, s_mem_offset, 256 - end +function write_hwreg_to_mem(s, s_rsrc, s_mem_offset) + s_mov_b32 exec_lo, m0 + s_mov_b32 m0, s_mem_offset + s_buffer_store_dword s, s_rsrc, m0 glc:1 + s_add_u32 s_mem_offset, s_mem_offset, 4 + s_mov_b32 m0, exec_lo +end + + +function write_16sgpr_to_mem(s, s_rsrc, s_mem_offset) + s_buffer_store_dwordx4 s[0], s_rsrc, 0 glc:1 + s_buffer_store_dwordx4 s[4], s_rsrc, 16 glc:1 + s_buffer_store_dwordx4 s[8], s_rsrc, 32 glc:1 + s_buffer_store_dwordx4 s[12], s_rsrc, 48 glc:1 + s_add_u32 s_rsrc[0], s_rsrc[0], 4*16 + s_addc_u32 s_rsrc[1], s_rsrc[1], 0x0 +end + +function write_12sgpr_to_mem(s, s_rsrc, s_mem_offset) + s_buffer_store_dwordx4 s[0], s_rsrc, 0 glc:1 + s_buffer_store_dwordx4 s[4], s_rsrc, 16 glc:1 + s_buffer_store_dwordx4 s[8], s_rsrc, 32 glc:1 + s_add_u32 s_rsrc[0], s_rsrc[0], 4*12 + s_addc_u32 s_rsrc[1], s_rsrc[1], 0x0 +end + + +function read_hwreg_from_mem(s, s_rsrc, s_mem_offset) + s_buffer_load_dword s, s_rsrc, s_mem_offset glc:1 + s_add_u32 s_mem_offset, s_mem_offset, 4 end -function read_sgpr_from_mem_wave32(s, s_rsrc, s_mem_offset, use_sqc) - s_buffer_load_dword s, s_rsrc, s_mem_offset glc:1 - if (use_sqc) - s_add_u32 s_mem_offset, s_mem_offset, 4 - else - s_add_u32 s_mem_offset, s_mem_offset, 128 - end +function read_16sgpr_from_mem(s, s_rsrc, s_mem_offset) + s_sub_u32 s_mem_offset, s_mem_offset, 4*16 + s_buffer_load_dwordx16 s, s_rsrc, s_mem_offset glc:1 end -function read_sgpr_from_mem_wave64(s, s_rsrc, s_mem_offset, use_sqc) - s_buffer_load_dword s, s_rsrc, s_mem_offset glc:1 - if (use_sqc) - s_add_u32 s_mem_offset, s_mem_offset, 4 - else - s_add_u32 s_mem_offset, s_mem_offset, 256 - end +function read_8sgpr_from_mem(s, s_rsrc, s_mem_offset) + s_sub_u32 s_mem_offset, s_mem_offset, 4*8 + s_buffer_load_dwordx8 s, s_rsrc, s_mem_offset glc:1 end +function read_4sgpr_from_mem(s, s_rsrc, s_mem_offset) + s_sub_u32 s_mem_offset, s_mem_offset, 4*4 + s_buffer_load_dwordx4 s, s_rsrc, s_mem_offset glc:1 +end + + +function get_lds_size_bytes(s_lds_size_byte) + s_getreg_b32 s_lds_size_byte, hwreg(HW_REG_LDS_ALLOC, SQ_WAVE_LDS_ALLOC_LDS_SIZE_SHIFT, SQ_WAVE_LDS_ALLOC_LDS_SIZE_SIZE) + s_lshl_b32 s_lds_size_byte, s_lds_size_byte, 8 //LDS size in dwords = lds_size * 64 *4Bytes // granularity 64DW +end + +function get_vgpr_size_bytes(s_vgpr_size_byte, s_size) + s_getreg_b32 s_vgpr_size_byte, hwreg(HW_REG_GPR_ALLOC,SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SHIFT,SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SIZE) + s_add_u32 s_vgpr_size_byte, s_vgpr_size_byte, 1 + s_lshr_b32 m0, s_size, S_WAVE_SIZE + s_and_b32 m0, m0, 1 + s_cmp_eq_u32 m0, 1 + s_cbranch_scc1 L_ENABLE_SHIFT_W64 + s_lshl_b32 s_vgpr_size_byte, s_vgpr_size_byte, (2+7) //Number of VGPRs = (vgpr_size + 1) * 4 * 32 * 4 (non-zero value) + s_branch L_SHIFT_DONE +L_ENABLE_SHIFT_W64: + s_lshl_b32 s_vgpr_size_byte, s_vgpr_size_byte, (2+8) //Number of VGPRs = (vgpr_size + 1) * 4 * 64 * 4 (non-zero value) +L_SHIFT_DONE: +end + +function get_svgpr_size_bytes(s_svgpr_size_byte) + s_getreg_b32 s_svgpr_size_byte, hwreg(HW_REG_LDS_ALLOC,SQ_WAVE_LDS_ALLOC_VGPR_SHARED_SIZE_SHIFT,SQ_WAVE_LDS_ALLOC_VGPR_SHARED_SIZE_SIZE) + s_lshl_b32 s_svgpr_size_byte, s_svgpr_size_byte, (3+7) +end + +function get_sgpr_size_bytes + return 512 +end + +function get_hwreg_size_bytes + return 128 +end diff --git a/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx8.asm b/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx8.asm index a47f5b933120..b195b7cd8a17 100644 --- a/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx8.asm +++ b/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx8.asm @@ -24,78 +24,6 @@ * PROJECT=vi ./sp3 cwsr_trap_handler_gfx8.asm -hex tmp.hex */ -/* HW (VI) source code for CWSR trap handler */ -/* Version 18 + multiple trap handler */ - -// this performance-optimal version was originally from Seven Xu at SRDC - -// Revison #18 --... -/* Rev History -** #1. Branch from gc dv. //gfxip/gfx8/main/src/test/suites/block/cs/sr/cs_trap_handler.sp3#1,#50, #51, #52-53(Skip, Already Fixed by PV), #54-56(merged),#57-58(mergerd, skiped-already fixed by PV) -** #4. SR Memory Layout: -** 1. VGPR-SGPR-HWREG-{LDS} -** 2. tba_hi.bits.26 - reconfigured as the first wave in tg bits, for defer Save LDS for a threadgroup.. performance concern.. -** #5. Update: 1. Accurate g8sr_ts_save_d timestamp -** #6. Update: 1. Fix s_barrier usage; 2. VGPR s/r using swizzle buffer?(NoNeed, already matched the swizzle pattern, more investigation) -** #7. Update: 1. don't barrier if noLDS -** #8. Branch: 1. Branch to ver#0, which is very similar to gc dv version -** 2. Fix SQ issue by s_sleep 2 -** #9. Update: 1. Fix scc restore failed issue, restore wave_status at last -** 2. optimize s_buffer save by burst 16sgprs... -** #10. Update 1. Optimize restore sgpr by busrt 16 sgprs. -** #11. Update 1. Add 2 more timestamp for debug version -** #12. Update 1. Add VGPR SR using DWx4, some case improve and some case drop performance -** #13. Integ 1. Always use MUBUF for PV trap shader... -** #14. Update 1. s_buffer_store soft clause... -** #15. Update 1. PERF - sclar write with glc:0/mtype0 to allow L2 combine. perf improvement a lot. -** #16. Update 1. PRRF - UNROLL LDS_DMA got 2500cycle save in IP tree -** #17. Update 1. FUNC - LDS_DMA has issues while ATC, replace with ds_read/buffer_store for save part[TODO restore part] -** 2. PERF - Save LDS before save VGPR to cover LDS save long latency... -** #18. Update 1. FUNC - Implicitly estore STATUS.VCCZ, which is not writable by s_setreg_b32 -** 2. FUNC - Handle non-CWSR traps -*/ - -var G8SR_WDMEM_HWREG_OFFSET = 0 -var G8SR_WDMEM_SGPR_OFFSET = 128 // in bytes - -// Keep definition same as the app shader, These 2 time stamps are part of the app shader... Should before any Save and after restore. - -var G8SR_DEBUG_TIMESTAMP = 0 -var G8SR_DEBUG_TS_SAVE_D_OFFSET = 40*4 // ts_save_d timestamp offset relative to SGPR_SR_memory_offset -var s_g8sr_ts_save_s = s[34:35] // save start -var s_g8sr_ts_sq_save_msg = s[36:37] // The save shader send SAVEWAVE msg to spi -var s_g8sr_ts_spi_wrexec = s[38:39] // the SPI write the sr address to SQ -var s_g8sr_ts_save_d = s[40:41] // save end -var s_g8sr_ts_restore_s = s[42:43] // restore start -var s_g8sr_ts_restore_d = s[44:45] // restore end - -var G8SR_VGPR_SR_IN_DWX4 = 0 -var G8SR_SAVE_BUF_RSRC_WORD1_STRIDE_DWx4 = 0x00100000 // DWx4 stride is 4*4Bytes -var G8SR_RESTORE_BUF_RSRC_WORD1_STRIDE_DWx4 = G8SR_SAVE_BUF_RSRC_WORD1_STRIDE_DWx4 - - -/*************************************************************************/ -/* control on how to run the shader */ -/*************************************************************************/ -//any hack that needs to be made to run this code in EMU (either because various EMU code are not ready or no compute save & restore in EMU run) -var EMU_RUN_HACK = 0 -var EMU_RUN_HACK_RESTORE_NORMAL = 0 -var EMU_RUN_HACK_SAVE_NORMAL_EXIT = 0 -var EMU_RUN_HACK_SAVE_SINGLE_WAVE = 0 -var EMU_RUN_HACK_SAVE_FIRST_TIME = 0 //for interrupted restore in which the first save is through EMU_RUN_HACK -var EMU_RUN_HACK_SAVE_FIRST_TIME_TBA_LO = 0 //for interrupted restore in which the first save is through EMU_RUN_HACK -var EMU_RUN_HACK_SAVE_FIRST_TIME_TBA_HI = 0 //for interrupted restore in which the first save is through EMU_RUN_HACK -var SAVE_LDS = 1 -var WG_BASE_ADDR_LO = 0x9000a000 -var WG_BASE_ADDR_HI = 0x0 -var WAVE_SPACE = 0x5000 //memory size that each wave occupies in workgroup state mem -var CTX_SAVE_CONTROL = 0x0 -var CTX_RESTORE_CONTROL = CTX_SAVE_CONTROL -var SIM_RUN_HACK = 0 //any hack that needs to be made to run this code in SIM (either because various RTL code are not ready or no compute save & restore in RTL run) -var SGPR_SAVE_USE_SQC = 1 //use SQC D$ to do the write -var USE_MTBUF_INSTEAD_OF_MUBUF = 0 //because TC EMU currently asserts on 0 of // overload DFMT field to carry 4 more bits of stride for MUBUF opcodes -var SWIZZLE_EN = 0 //whether we use swizzled buffer addressing - /**************************************************************************/ /* variables */ /**************************************************************************/ @@ -226,16 +154,7 @@ shader main type(CS) - if ((EMU_RUN_HACK) && (!EMU_RUN_HACK_RESTORE_NORMAL)) //hack to use trap_id for determining save/restore - //FIXME VCCZ un-init assertion s_getreg_b32 s_save_status, hwreg(HW_REG_STATUS) //save STATUS since we will change SCC - s_and_b32 s_save_tmp, s_save_pc_hi, 0xffff0000 //change SCC - s_cmp_eq_u32 s_save_tmp, 0x007e0000 //Save: trap_id = 0x7e. Restore: trap_id = 0x7f. - s_cbranch_scc0 L_JUMP_TO_RESTORE //do not need to recover STATUS here since we are going to RESTORE - //FIXME s_setreg_b32 hwreg(HW_REG_STATUS), s_save_status //need to recover STATUS since we are going to SAVE - s_branch L_SKIP_RESTORE //NOT restore, SAVE actually - else s_branch L_SKIP_RESTORE //NOT restore. might be a regular trap or save - end L_JUMP_TO_RESTORE: s_branch L_RESTORE //restore @@ -249,7 +168,7 @@ L_SKIP_RESTORE: s_cbranch_scc1 L_SAVE //this is the operation for save // ********* Handle non-CWSR traps ******************* -if (!EMU_RUN_HACK) + /* read tba and tma for next level trap handler, ttmp4 is used as s_save_status */ s_load_dwordx4 [ttmp8,ttmp9,ttmp10, ttmp11], [tma_lo,tma_hi], 0 s_waitcnt lgkmcnt(0) @@ -268,7 +187,7 @@ L_EXCP_CASE: s_and_b32 ttmp1, ttmp1, 0xFFFF set_status_without_spi_prio(s_save_status, ttmp2) //restore HW status(SCC) s_rfe_b64 [ttmp0, ttmp1] -end + // ********* End handling of non-CWSR traps ******************* /**************************************************************************/ @@ -276,12 +195,6 @@ end /**************************************************************************/ L_SAVE: - -if G8SR_DEBUG_TIMESTAMP - s_memrealtime s_g8sr_ts_save_s - s_waitcnt lgkmcnt(0) //FIXME, will cause xnack?? -end - s_mov_b32 s_save_tmp, 0 //clear saveCtx bit s_setreg_b32 hwreg(HW_REG_TRAPSTS, SQ_WAVE_TRAPSTS_SAVECTX_SHIFT, 1), s_save_tmp //clear saveCtx bit @@ -303,16 +216,7 @@ end s_mov_b32 s_save_exec_hi, exec_hi s_mov_b64 exec, 0x0 //clear EXEC to get ready to receive -if G8SR_DEBUG_TIMESTAMP - s_memrealtime s_g8sr_ts_sq_save_msg - s_waitcnt lgkmcnt(0) -end - - if (EMU_RUN_HACK) - - else s_sendmsg sendmsg(MSG_SAVEWAVE) //send SPI a message and wait for SPI's write to EXEC - end // Set SPI_PRIO=2 to avoid starving instruction fetch in the waves we're waiting for. s_or_b32 s_save_tmp, s_save_status, (2 << SQ_WAVE_STATUS_SPI_PRIO_SHIFT) @@ -321,36 +225,9 @@ end L_SLEEP: s_sleep 0x2 // sleep 1 (64clk) is not enough for 8 waves per SIMD, which will cause SQ hang, since the 7,8th wave could not get arbit to exec inst, while other waves are stuck into the sleep-loop and waiting for wrexec!=0 - if (EMU_RUN_HACK) - - else s_cbranch_execz L_SLEEP - end - -if G8SR_DEBUG_TIMESTAMP - s_memrealtime s_g8sr_ts_spi_wrexec - s_waitcnt lgkmcnt(0) -end /* setup Resource Contants */ - if ((EMU_RUN_HACK) && (!EMU_RUN_HACK_SAVE_SINGLE_WAVE)) - //calculate wd_addr using absolute thread id - v_readlane_b32 s_save_tmp, v9, 0 - s_lshr_b32 s_save_tmp, s_save_tmp, 6 - s_mul_i32 s_save_tmp, s_save_tmp, WAVE_SPACE - s_add_i32 s_save_spi_init_lo, s_save_tmp, WG_BASE_ADDR_LO - s_mov_b32 s_save_spi_init_hi, WG_BASE_ADDR_HI - s_and_b32 s_save_spi_init_hi, s_save_spi_init_hi, CTX_SAVE_CONTROL - else - end - if ((EMU_RUN_HACK) && (EMU_RUN_HACK_SAVE_SINGLE_WAVE)) - s_add_i32 s_save_spi_init_lo, s_save_tmp, WG_BASE_ADDR_LO - s_mov_b32 s_save_spi_init_hi, WG_BASE_ADDR_HI - s_and_b32 s_save_spi_init_hi, s_save_spi_init_hi, CTX_SAVE_CONTROL - else - end - - s_mov_b32 s_save_buf_rsrc0, s_save_spi_init_lo //base_addr_lo s_and_b32 s_save_buf_rsrc1, s_save_spi_init_hi, 0x0000FFFF //base_addr_hi s_or_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, S_SAVE_BUF_RSRC_WORD1_STRIDE @@ -383,22 +260,10 @@ end s_mov_b32 s_save_buf_rsrc2, 0x4 //NUM_RECORDS in bytes - if (SWIZZLE_EN) - s_add_u32 s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0 //FIXME need to use swizzle to enable bounds checking? - else s_mov_b32 s_save_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - end write_hwreg_to_mem(s_save_m0, s_save_buf_rsrc0, s_save_mem_offset) //M0 - - if ((EMU_RUN_HACK) && (EMU_RUN_HACK_SAVE_FIRST_TIME)) - s_add_u32 s_save_pc_lo, s_save_pc_lo, 4 //pc[31:0]+4 - s_addc_u32 s_save_pc_hi, s_save_pc_hi, 0x0 //carry bit over - s_mov_b32 tba_lo, EMU_RUN_HACK_SAVE_FIRST_TIME_TBA_LO - s_mov_b32 tba_hi, EMU_RUN_HACK_SAVE_FIRST_TIME_TBA_HI - end - write_hwreg_to_mem(s_save_pc_lo, s_save_buf_rsrc0, s_save_mem_offset) //PC write_hwreg_to_mem(s_save_pc_hi, s_save_buf_rsrc0, s_save_mem_offset) write_hwreg_to_mem(s_save_exec_lo, s_save_buf_rsrc0, s_save_mem_offset) //EXEC @@ -440,18 +305,8 @@ end s_add_u32 s_save_alloc_size, s_save_alloc_size, 1 s_lshl_b32 s_save_alloc_size, s_save_alloc_size, 4 //Number of SGPRs = (sgpr_size + 1) * 16 (non-zero value) - if (SGPR_SAVE_USE_SQC) s_lshl_b32 s_save_buf_rsrc2, s_save_alloc_size, 2 //NUM_RECORDS in bytes - else - s_lshl_b32 s_save_buf_rsrc2, s_save_alloc_size, 8 //NUM_RECORDS in bytes (64 threads) - end - - if (SWIZZLE_EN) - s_add_u32 s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0 //FIXME need to use swizzle to enable bounds checking? - else s_mov_b32 s_save_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - end - // backup s_save_buf_rsrc0,1 to s_save_pc_lo/hi, since write_16sgpr_to_mem function will change the rsrc0 //s_mov_b64 s_save_pc_lo, s_save_buf_rsrc0 @@ -490,30 +345,14 @@ end s_mov_b32 exec_lo, 0xFFFFFFFF //need every thread from now on s_mov_b32 exec_hi, 0xFFFFFFFF - if (SWIZZLE_EN) - s_add_u32 s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0 //FIXME need to use swizzle to enable bounds checking? - else s_mov_b32 s_save_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - end - // VGPR Allocated in 4-GPR granularity -if G8SR_VGPR_SR_IN_DWX4 - // the const stride for DWx4 is 4*4 bytes - s_and_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, 0x0000FFFF // reset const stride to 0 - s_or_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, G8SR_SAVE_BUF_RSRC_WORD1_STRIDE_DWx4 // const stride to 4*4 bytes - - buffer_store_dwordx4 v0, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 - - s_and_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, 0x0000FFFF // reset const stride to 0 - s_or_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, S_SAVE_BUF_RSRC_WORD1_STRIDE // reset const stride to 4 bytes -else buffer_store_dword v0, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 buffer_store_dword v1, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 offset:256 buffer_store_dword v2, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 offset:256*2 buffer_store_dword v3, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 offset:256*3 -end @@ -549,64 +388,10 @@ end s_add_u32 s_save_mem_offset, s_save_mem_offset, get_hwreg_size_bytes() - if (SWIZZLE_EN) - s_add_u32 s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0 //FIXME need to use swizzle to enable bounds checking? - else s_mov_b32 s_save_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - end - s_mov_b32 m0, 0x0 //lds_offset initial value = 0 -var LDS_DMA_ENABLE = 0 -var UNROLL = 0 -if UNROLL==0 && LDS_DMA_ENABLE==1 - s_mov_b32 s3, 256*2 - s_nop 0 - s_nop 0 - s_nop 0 - L_SAVE_LDS_LOOP: - //TODO: looks the 2 buffer_store/load clause for s/r will hurt performance.??? - if (SAVE_LDS) //SPI always alloc LDS space in 128DW granularity - buffer_store_lds_dword s_save_buf_rsrc0, s_save_mem_offset lds:1 // first 64DW - buffer_store_lds_dword s_save_buf_rsrc0, s_save_mem_offset lds:1 offset:256 // second 64DW - end - - s_add_u32 m0, m0, s3 //every buffer_store_lds does 256 bytes - s_add_u32 s_save_mem_offset, s_save_mem_offset, s3 //mem offset increased by 256 bytes - s_cmp_lt_u32 m0, s_save_alloc_size //scc=(m0 < s_save_alloc_size) ? 1 : 0 - s_cbranch_scc1 L_SAVE_LDS_LOOP //LDS save is complete? - -elsif LDS_DMA_ENABLE==1 && UNROLL==1 // UNROOL , has ichace miss - // store from higest LDS address to lowest - s_mov_b32 s3, 256*2 - s_sub_u32 m0, s_save_alloc_size, s3 - s_add_u32 s_save_mem_offset, s_save_mem_offset, m0 - s_lshr_b32 s_save_alloc_size, s_save_alloc_size, 9 // how many 128 trunks... - s_sub_u32 s_save_alloc_size, 128, s_save_alloc_size // store from higheset addr to lowest - s_mul_i32 s_save_alloc_size, s_save_alloc_size, 6*4 // PC offset increment, each LDS save block cost 6*4 Bytes instruction - s_add_u32 s_save_alloc_size, s_save_alloc_size, 3*4 //2is the below 2 inst...//s_addc and s_setpc - s_nop 0 - s_nop 0 - s_nop 0 //pad 3 dw to let LDS_DMA align with 64Bytes - s_getpc_b64 s[0:1] // reuse s[0:1], since s[0:1] already saved - s_add_u32 s0, s0,s_save_alloc_size - s_addc_u32 s1, s1, 0 - s_setpc_b64 s[0:1] - - - for var i =0; i< 128; i++ - // be careful to make here a 64Byte aligned address, which could improve performance... - buffer_store_lds_dword s_save_buf_rsrc0, s_save_mem_offset lds:1 offset:0 // first 64DW - buffer_store_lds_dword s_save_buf_rsrc0, s_save_mem_offset lds:1 offset:256 // second 64DW - - if i!=127 - s_sub_u32 m0, m0, s3 // use a sgpr to shrink 2DW-inst to 1DW inst to improve performance , i.e. pack more LDS_DMA inst to one Cacheline - s_sub_u32 s_save_mem_offset, s_save_mem_offset, s3 - end - end - -else // BUFFER_STORE v_mbcnt_lo_u32_b32 v2, 0xffffffff, 0x0 v_mbcnt_hi_u32_b32 v3, 0xffffffff, v2 // tid v_mul_i32_i24 v2, v3, 8 // tid*8 @@ -628,8 +413,6 @@ L_SAVE_LDS_LOOP_VECTOR: // restore rsrc3 s_mov_b32 s_save_buf_rsrc3, s0 -end - L_SAVE_LDS_DONE: @@ -647,44 +430,8 @@ L_SAVE_LDS_DONE: s_add_u32 s_save_alloc_size, s_save_alloc_size, 1 s_lshl_b32 s_save_alloc_size, s_save_alloc_size, 2 //Number of VGPRs = (vgpr_size + 1) * 4 (non-zero value) //FIXME for GFX, zero is possible s_lshl_b32 s_save_buf_rsrc2, s_save_alloc_size, 8 //NUM_RECORDS in bytes (64 threads*4) - if (SWIZZLE_EN) - s_add_u32 s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0 //FIXME need to use swizzle to enable bounds checking? - else s_mov_b32 s_save_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - end - - - // VGPR Allocated in 4-GPR granularity - -if G8SR_VGPR_SR_IN_DWX4 - // the const stride for DWx4 is 4*4 bytes - s_and_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, 0x0000FFFF // reset const stride to 0 - s_or_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, G8SR_SAVE_BUF_RSRC_WORD1_STRIDE_DWx4 // const stride to 4*4 bytes - - s_mov_b32 m0, 4 // skip first 4 VGPRs - s_cmp_lt_u32 m0, s_save_alloc_size - s_cbranch_scc0 L_SAVE_VGPR_LOOP_END // no more vgprs - s_set_gpr_idx_on m0, 0x1 // This will change M0 - s_add_u32 s_save_alloc_size, s_save_alloc_size, 0x1000 // because above inst change m0 -L_SAVE_VGPR_LOOP: - v_mov_b32 v0, v0 // v0 = v[0+m0] - v_mov_b32 v1, v1 - v_mov_b32 v2, v2 - v_mov_b32 v3, v3 - - - buffer_store_dwordx4 v0, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 - s_add_u32 m0, m0, 4 - s_add_u32 s_save_mem_offset, s_save_mem_offset, 256*4 - s_cmp_lt_u32 m0, s_save_alloc_size - s_cbranch_scc1 L_SAVE_VGPR_LOOP //VGPR save is complete? - s_set_gpr_idx_off -L_SAVE_VGPR_LOOP_END: - - s_and_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, 0x0000FFFF // reset const stride to 0 - s_or_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, S_SAVE_BUF_RSRC_WORD1_STRIDE // reset const stride to 4 bytes -else // VGPR store using dw burst s_mov_b32 m0, 0x4 //VGPR initial index value =0 s_cmp_lt_u32 m0, s_save_alloc_size @@ -700,52 +447,18 @@ else v_mov_b32 v2, v2 //v0 = v[0+m0] v_mov_b32 v3, v3 //v0 = v[0+m0] - if(USE_MTBUF_INSTEAD_OF_MUBUF) - tbuffer_store_format_x v0, v0, s_save_buf_rsrc0, s_save_mem_offset format:BUF_NUM_FORMAT_FLOAT format: BUF_DATA_FORMAT_32 slc:1 glc:1 - else buffer_store_dword v0, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 buffer_store_dword v1, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 offset:256 buffer_store_dword v2, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 offset:256*2 buffer_store_dword v3, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 offset:256*3 - end s_add_u32 m0, m0, 4 //next vgpr index s_add_u32 s_save_mem_offset, s_save_mem_offset, 256*4 //every buffer_store_dword does 256 bytes s_cmp_lt_u32 m0, s_save_alloc_size //scc = (m0 < s_save_alloc_size) ? 1 : 0 s_cbranch_scc1 L_SAVE_VGPR_LOOP //VGPR save is complete? s_set_gpr_idx_off -end L_SAVE_VGPR_END: - - - - - - - /* S_PGM_END_SAVED */ //FIXME graphics ONLY - if ((EMU_RUN_HACK) && (!EMU_RUN_HACK_SAVE_NORMAL_EXIT)) - s_and_b32 s_save_pc_hi, s_save_pc_hi, 0x0000ffff //pc[47:32] - s_add_u32 s_save_pc_lo, s_save_pc_lo, 4 //pc[31:0]+4 - s_addc_u32 s_save_pc_hi, s_save_pc_hi, 0x0 //carry bit over - s_rfe_b64 s_save_pc_lo //Return to the main shader program - else - end - -// Save Done timestamp -if G8SR_DEBUG_TIMESTAMP - s_memrealtime s_g8sr_ts_save_d - // SGPR SR memory offset : size(VGPR) - get_vgpr_size_bytes(s_save_mem_offset) - s_add_u32 s_save_mem_offset, s_save_mem_offset, G8SR_DEBUG_TS_SAVE_D_OFFSET - s_waitcnt lgkmcnt(0) //FIXME, will cause xnack?? - // Need reset rsrc2?? - s_mov_b32 m0, s_save_mem_offset - s_mov_b32 s_save_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - s_buffer_store_dwordx2 s_g8sr_ts_save_d, s_save_buf_rsrc0, m0 glc:1 -end - - s_branch L_END_PGM @@ -756,27 +469,6 @@ end L_RESTORE: /* Setup Resource Contants */ - if ((EMU_RUN_HACK) && (!EMU_RUN_HACK_RESTORE_NORMAL)) - //calculate wd_addr using absolute thread id - v_readlane_b32 s_restore_tmp, v9, 0 - s_lshr_b32 s_restore_tmp, s_restore_tmp, 6 - s_mul_i32 s_restore_tmp, s_restore_tmp, WAVE_SPACE - s_add_i32 s_restore_spi_init_lo, s_restore_tmp, WG_BASE_ADDR_LO - s_mov_b32 s_restore_spi_init_hi, WG_BASE_ADDR_HI - s_and_b32 s_restore_spi_init_hi, s_restore_spi_init_hi, CTX_RESTORE_CONTROL - else - end - -if G8SR_DEBUG_TIMESTAMP - s_memrealtime s_g8sr_ts_restore_s - s_waitcnt lgkmcnt(0) //FIXME, will cause xnack?? - // tma_lo/hi are sgpr 110, 111, which will not used for 112 SGPR allocated case... - s_mov_b32 s_restore_pc_lo, s_g8sr_ts_restore_s[0] - s_mov_b32 s_restore_pc_hi, s_g8sr_ts_restore_s[1] //backup ts to ttmp0/1, sicne exec will be finally restored.. -end - - - s_mov_b32 s_restore_buf_rsrc0, s_restore_spi_init_lo //base_addr_lo s_and_b32 s_restore_buf_rsrc1, s_restore_spi_init_hi, 0x0000FFFF //base_addr_hi s_or_b32 s_restore_buf_rsrc1, s_restore_buf_rsrc1, S_RESTORE_BUF_RSRC_WORD1_STRIDE @@ -818,18 +510,12 @@ end s_add_u32 s_restore_mem_offset, s_restore_mem_offset, get_hwreg_size_bytes() //FIXME, Check if offset overflow??? - if (SWIZZLE_EN) - s_add_u32 s_restore_buf_rsrc2, s_restore_buf_rsrc2, 0x0 //FIXME need to use swizzle to enable bounds checking? - else s_mov_b32 s_restore_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - end s_mov_b32 m0, 0x0 //lds_offset initial value = 0 L_RESTORE_LDS_LOOP: - if (SAVE_LDS) buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset lds:1 // first 64DW buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset lds:1 offset:256 // second 64DW - end s_add_u32 m0, m0, 256*2 // 128 DW s_add_u32 s_restore_mem_offset, s_restore_mem_offset, 256*2 //mem offset increased by 128DW s_cmp_lt_u32 m0, s_restore_alloc_size //scc=(m0 < s_restore_alloc_size) ? 1 : 0 @@ -848,40 +534,8 @@ end s_add_u32 s_restore_alloc_size, s_restore_alloc_size, 1 s_lshl_b32 s_restore_alloc_size, s_restore_alloc_size, 2 //Number of VGPRs = (vgpr_size + 1) * 4 (non-zero value) s_lshl_b32 s_restore_buf_rsrc2, s_restore_alloc_size, 8 //NUM_RECORDS in bytes (64 threads*4) - if (SWIZZLE_EN) - s_add_u32 s_restore_buf_rsrc2, s_restore_buf_rsrc2, 0x0 //FIXME need to use swizzle to enable bounds checking? - else s_mov_b32 s_restore_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - end - -if G8SR_VGPR_SR_IN_DWX4 - get_vgpr_size_bytes(s_restore_mem_offset) - s_sub_u32 s_restore_mem_offset, s_restore_mem_offset, 256*4 - - // the const stride for DWx4 is 4*4 bytes - s_and_b32 s_restore_buf_rsrc1, s_restore_buf_rsrc1, 0x0000FFFF // reset const stride to 0 - s_or_b32 s_restore_buf_rsrc1, s_restore_buf_rsrc1, G8SR_RESTORE_BUF_RSRC_WORD1_STRIDE_DWx4 // const stride to 4*4 bytes - - s_mov_b32 m0, s_restore_alloc_size - s_set_gpr_idx_on m0, 0x8 // Note.. This will change m0 - -L_RESTORE_VGPR_LOOP: - buffer_load_dwordx4 v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 - s_waitcnt vmcnt(0) - s_sub_u32 m0, m0, 4 - v_mov_b32 v0, v0 // v[0+m0] = v0 - v_mov_b32 v1, v1 - v_mov_b32 v2, v2 - v_mov_b32 v3, v3 - s_sub_u32 s_restore_mem_offset, s_restore_mem_offset, 256*4 - s_cmp_eq_u32 m0, 0x8000 - s_cbranch_scc0 L_RESTORE_VGPR_LOOP - s_set_gpr_idx_off - - s_and_b32 s_restore_buf_rsrc1, s_restore_buf_rsrc1, 0x0000FFFF // reset const stride to 0 - s_or_b32 s_restore_buf_rsrc1, s_restore_buf_rsrc1, S_RESTORE_BUF_RSRC_WORD1_STRIDE // const stride to 4*4 bytes - -else + // VGPR load using dw burst s_mov_b32 s_restore_mem_offset_save, s_restore_mem_offset // restore start with v1, v0 will be the last s_add_u32 s_restore_mem_offset, s_restore_mem_offset, 256*4 @@ -890,14 +544,10 @@ else s_add_u32 s_restore_alloc_size, s_restore_alloc_size, 0x8000 //add 0x8000 since we compare m0 against it later L_RESTORE_VGPR_LOOP: - if(USE_MTBUF_INSTEAD_OF_MUBUF) - tbuffer_load_format_x v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset format:BUF_NUM_FORMAT_FLOAT format: BUF_DATA_FORMAT_32 slc:1 glc:1 - else buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 buffer_load_dword v1, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 offset:256 buffer_load_dword v2, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 offset:256*2 buffer_load_dword v3, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 offset:256*3 - end s_waitcnt vmcnt(0) //ensure data ready v_mov_b32 v0, v0 //v[0+m0] = v0 v_mov_b32 v1, v1 @@ -909,16 +559,10 @@ else s_cbranch_scc1 L_RESTORE_VGPR_LOOP //VGPR restore (except v0) is complete? s_set_gpr_idx_off /* VGPR restore on v0 */ - if(USE_MTBUF_INSTEAD_OF_MUBUF) - tbuffer_load_format_x v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save format:BUF_NUM_FORMAT_FLOAT format: BUF_DATA_FORMAT_32 slc:1 glc:1 - else buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save slc:1 glc:1 buffer_load_dword v1, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save slc:1 glc:1 offset:256 buffer_load_dword v2, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save slc:1 glc:1 offset:256*2 buffer_load_dword v3, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save slc:1 glc:1 offset:256*3 - end - -end /* restore SGPRs */ ////////////////////////////// @@ -934,16 +578,8 @@ end s_add_u32 s_restore_alloc_size, s_restore_alloc_size, 1 s_lshl_b32 s_restore_alloc_size, s_restore_alloc_size, 4 //Number of SGPRs = (sgpr_size + 1) * 16 (non-zero value) - if (SGPR_SAVE_USE_SQC) s_lshl_b32 s_restore_buf_rsrc2, s_restore_alloc_size, 2 //NUM_RECORDS in bytes - else - s_lshl_b32 s_restore_buf_rsrc2, s_restore_alloc_size, 8 //NUM_RECORDS in bytes (64 threads) - end - if (SWIZZLE_EN) - s_add_u32 s_restore_buf_rsrc2, s_restore_buf_rsrc2, 0x0 //FIXME need to use swizzle to enable bounds checking? - else s_mov_b32 s_restore_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - end /* If 112 SGPRs ar allocated, 4 sgprs are not used TBA(108,109),TMA(110,111), However, we are safe to restore these 4 SGPRs anyway, since TBA,TMA will later be restored by HWREG @@ -972,12 +608,6 @@ end ////////////////////////////// L_RESTORE_HWREG: - -if G8SR_DEBUG_TIMESTAMP - s_mov_b32 s_g8sr_ts_restore_s[0], s_restore_pc_lo - s_mov_b32 s_g8sr_ts_restore_s[1], s_restore_pc_hi -end - // HWREG SR memory offset : size(VGPR)+size(SGPR) get_vgpr_size_bytes(s_restore_mem_offset) get_sgpr_size_bytes(s_restore_tmp) @@ -985,11 +615,7 @@ end s_mov_b32 s_restore_buf_rsrc2, 0x4 //NUM_RECORDS in bytes - if (SWIZZLE_EN) - s_add_u32 s_restore_buf_rsrc2, s_restore_buf_rsrc2, 0x0 //FIXME need to use swizzle to enable bounds checking? - else s_mov_b32 s_restore_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - end read_hwreg_from_mem(s_restore_m0, s_restore_buf_rsrc0, s_restore_mem_offset) //M0 read_hwreg_from_mem(s_restore_pc_lo, s_restore_buf_rsrc0, s_restore_mem_offset) //PC @@ -1006,16 +632,6 @@ end s_waitcnt lgkmcnt(0) //from now on, it is safe to restore STATUS and IB_STS - //for normal save & restore, the saved PC points to the next inst to execute, no adjustment needs to be made, otherwise: - if ((EMU_RUN_HACK) && (!EMU_RUN_HACK_RESTORE_NORMAL)) - s_add_u32 s_restore_pc_lo, s_restore_pc_lo, 8 //pc[31:0]+8 //two back-to-back s_trap are used (first for save and second for restore) - s_addc_u32 s_restore_pc_hi, s_restore_pc_hi, 0x0 //carry bit over - end - if ((EMU_RUN_HACK) && (EMU_RUN_HACK_RESTORE_NORMAL)) - s_add_u32 s_restore_pc_lo, s_restore_pc_lo, 4 //pc[31:0]+4 // save is hack through s_trap but restore is normal - s_addc_u32 s_restore_pc_hi, s_restore_pc_hi, 0x0 //carry bit over - end - s_mov_b32 m0, s_restore_m0 s_mov_b32 exec_lo, s_restore_exec_lo s_mov_b32 exec_hi, s_restore_exec_hi @@ -1048,11 +664,6 @@ end s_barrier //barrier to ensure the readiness of LDS before access attempts from any other wave in the same TG //FIXME not performance-optimal at this time -if G8SR_DEBUG_TIMESTAMP - s_memrealtime s_g8sr_ts_restore_d - s_waitcnt lgkmcnt(0) -end - // s_rfe_b64 s_restore_pc_lo //Return to the main shader program and resume execution s_rfe_restore_b64 s_restore_pc_lo, s_restore_m0 // s_restore_m0[0] is used to set STATUS.inst_atc diff --git a/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx9.asm b/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx9.asm index 6bae2e022c6e..75f29d13c90f 100644 --- a/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx9.asm +++ b/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx9.asm @@ -24,76 +24,9 @@ * PROJECT=greenland ./sp3 cwsr_trap_handler_gfx9.asm -hex tmp.hex */ -/* HW (GFX9) source code for CWSR trap handler */ -/* Version 18 + multiple trap handler */ - -// this performance-optimal version was originally from Seven Xu at SRDC - -// Revison #18 --... -/* Rev History -** #1. Branch from gc dv. //gfxip/gfx9/main/src/test/suites/block/cs/sr/cs_trap_handler.sp3#1,#50, #51, #52-53(Skip, Already Fixed by PV), #54-56(merged),#57-58(mergerd, skiped-already fixed by PV) -** #4. SR Memory Layout: -** 1. VGPR-SGPR-HWREG-{LDS} -** 2. tba_hi.bits.26 - reconfigured as the first wave in tg bits, for defer Save LDS for a threadgroup.. performance concern.. -** #5. Update: 1. Accurate g8sr_ts_save_d timestamp -** #6. Update: 1. Fix s_barrier usage; 2. VGPR s/r using swizzle buffer?(NoNeed, already matched the swizzle pattern, more investigation) -** #7. Update: 1. don't barrier if noLDS -** #8. Branch: 1. Branch to ver#0, which is very similar to gc dv version -** 2. Fix SQ issue by s_sleep 2 -** #9. Update: 1. Fix scc restore failed issue, restore wave_status at last -** 2. optimize s_buffer save by burst 16sgprs... -** #10. Update 1. Optimize restore sgpr by busrt 16 sgprs. -** #11. Update 1. Add 2 more timestamp for debug version -** #12. Update 1. Add VGPR SR using DWx4, some case improve and some case drop performance -** #13. Integ 1. Always use MUBUF for PV trap shader... -** #14. Update 1. s_buffer_store soft clause... -** #15. Update 1. PERF - sclar write with glc:0/mtype0 to allow L2 combine. perf improvement a lot. -** #16. Update 1. PRRF - UNROLL LDS_DMA got 2500cycle save in IP tree -** #17. Update 1. FUNC - LDS_DMA has issues while ATC, replace with ds_read/buffer_store for save part[TODO restore part] -** 2. PERF - Save LDS before save VGPR to cover LDS save long latency... -** #18. Update 1. FUNC - Implicitly estore STATUS.VCCZ, which is not writable by s_setreg_b32 -** 2. FUNC - Handle non-CWSR traps -*/ - -var G8SR_WDMEM_HWREG_OFFSET = 0 -var G8SR_WDMEM_SGPR_OFFSET = 128 // in bytes - -// Keep definition same as the app shader, These 2 time stamps are part of the app shader... Should before any Save and after restore. - -var G8SR_DEBUG_TIMESTAMP = 0 -var G8SR_DEBUG_TS_SAVE_D_OFFSET = 40*4 // ts_save_d timestamp offset relative to SGPR_SR_memory_offset -var s_g8sr_ts_save_s = s[34:35] // save start -var s_g8sr_ts_sq_save_msg = s[36:37] // The save shader send SAVEWAVE msg to spi -var s_g8sr_ts_spi_wrexec = s[38:39] // the SPI write the sr address to SQ -var s_g8sr_ts_save_d = s[40:41] // save end -var s_g8sr_ts_restore_s = s[42:43] // restore start -var s_g8sr_ts_restore_d = s[44:45] // restore end - -var G8SR_VGPR_SR_IN_DWX4 = 0 -var G8SR_SAVE_BUF_RSRC_WORD1_STRIDE_DWx4 = 0x00100000 // DWx4 stride is 4*4Bytes -var G8SR_RESTORE_BUF_RSRC_WORD1_STRIDE_DWx4 = G8SR_SAVE_BUF_RSRC_WORD1_STRIDE_DWx4 - - -/*************************************************************************/ -/* control on how to run the shader */ -/*************************************************************************/ -//any hack that needs to be made to run this code in EMU (either because various EMU code are not ready or no compute save & restore in EMU run) -var EMU_RUN_HACK = 0 -var EMU_RUN_HACK_RESTORE_NORMAL = 0 -var EMU_RUN_HACK_SAVE_NORMAL_EXIT = 0 -var EMU_RUN_HACK_SAVE_SINGLE_WAVE = 0 -var EMU_RUN_HACK_SAVE_FIRST_TIME = 0 //for interrupted restore in which the first save is through EMU_RUN_HACK -var SAVE_LDS = 1 -var WG_BASE_ADDR_LO = 0x9000a000 -var WG_BASE_ADDR_HI = 0x0 -var WAVE_SPACE = 0x5000 //memory size that each wave occupies in workgroup state mem -var CTX_SAVE_CONTROL = 0x0 -var CTX_RESTORE_CONTROL = CTX_SAVE_CONTROL -var SIM_RUN_HACK = 0 //any hack that needs to be made to run this code in SIM (either because various RTL code are not ready or no compute save & restore in RTL run) -var SGPR_SAVE_USE_SQC = 1 //use SQC D$ to do the write -var USE_MTBUF_INSTEAD_OF_MUBUF = 0 //because TC EMU currently asserts on 0 of // overload DFMT field to carry 4 more bits of stride for MUBUF opcodes -var SWIZZLE_EN = 0 //whether we use swizzled buffer addressing var ACK_SQC_STORE = 1 //workaround for suspected SQC store bug causing incorrect stores under concurrency +var SAVE_AFTER_XNACK_ERROR = 1 //workaround for TCP store failure after XNACK error when ALLOW_REPLAY=0, for debugger +var SINGLE_STEP_MISSED_WORKAROUND = 1 //workaround for lost MODE.DEBUG_EN exception when SAVECTX raised /**************************************************************************/ /* variables */ @@ -107,6 +40,7 @@ var SQ_WAVE_STATUS_PRE_SPI_PRIO_SHIFT = 0 var SQ_WAVE_STATUS_PRE_SPI_PRIO_SIZE = 1 var SQ_WAVE_STATUS_POST_SPI_PRIO_SHIFT = 3 var SQ_WAVE_STATUS_POST_SPI_PRIO_SIZE = 29 +var SQ_WAVE_STATUS_ALLOW_REPLAY_MASK = 0x400000 var SQ_WAVE_LDS_ALLOC_LDS_SIZE_SHIFT = 12 var SQ_WAVE_LDS_ALLOC_LDS_SIZE_SIZE = 9 @@ -127,12 +61,15 @@ var SQ_WAVE_TRAPSTS_POST_SAVECTX_MASK = 0xFFFFF800 var SQ_WAVE_TRAPSTS_POST_SAVECTX_SHIFT = 11 var SQ_WAVE_TRAPSTS_POST_SAVECTX_SIZE = 21 var SQ_WAVE_TRAPSTS_ILLEGAL_INST_MASK = 0x800 +var SQ_WAVE_TRAPSTS_XNACK_ERROR_MASK = 0x10000000 var SQ_WAVE_IB_STS_RCNT_SHIFT = 16 //FIXME var SQ_WAVE_IB_STS_FIRST_REPLAY_SHIFT = 15 //FIXME var SQ_WAVE_IB_STS_RCNT_FIRST_REPLAY_MASK = 0x1F8000 var SQ_WAVE_IB_STS_RCNT_FIRST_REPLAY_MASK_NEG = 0x00007FFF //FIXME +var SQ_WAVE_MODE_DEBUG_EN_MASK = 0x800 + var SQ_BUF_RSRC_WORD1_ATC_SHIFT = 24 var SQ_BUF_RSRC_WORD3_MTYPE_SHIFT = 27 @@ -197,13 +134,15 @@ var s_restore_spi_init_lo = exec_lo var s_restore_spi_init_hi = exec_hi var s_restore_mem_offset = ttmp12 +var s_restore_accvgpr_offset = ttmp13 var s_restore_alloc_size = ttmp3 var s_restore_tmp = ttmp2 var s_restore_mem_offset_save = s_restore_tmp //no conflict +var s_restore_accvgpr_offset_save = ttmp7 var s_restore_m0 = s_restore_alloc_size //no conflict -var s_restore_mode = ttmp7 +var s_restore_mode = s_restore_accvgpr_offset_save var s_restore_pc_lo = ttmp0 var s_restore_pc_hi = ttmp1 @@ -226,20 +165,11 @@ var s_restore_ttmps_hi = s_restore_alloc_size //no conflict /* Shader Main*/ shader main - asic(GFX9) + asic(DEFAULT) type(CS) - if ((EMU_RUN_HACK) && (!EMU_RUN_HACK_RESTORE_NORMAL)) //hack to use trap_id for determining save/restore - //FIXME VCCZ un-init assertion s_getreg_b32 s_save_status, hwreg(HW_REG_STATUS) //save STATUS since we will change SCC - s_and_b32 s_save_tmp, s_save_pc_hi, 0xffff0000 //change SCC - s_cmp_eq_u32 s_save_tmp, 0x007e0000 //Save: trap_id = 0x7e. Restore: trap_id = 0x7f. - s_cbranch_scc0 L_JUMP_TO_RESTORE //do not need to recover STATUS here since we are going to RESTORE - //FIXME s_setreg_b32 hwreg(HW_REG_STATUS), s_save_status //need to recover STATUS since we are going to SAVE - s_branch L_SKIP_RESTORE //NOT restore, SAVE actually - else s_branch L_SKIP_RESTORE //NOT restore. might be a regular trap or save - end L_JUMP_TO_RESTORE: s_branch L_RESTORE //restore @@ -248,12 +178,29 @@ L_SKIP_RESTORE: s_getreg_b32 s_save_status, hwreg(HW_REG_STATUS) //save STATUS since we will change SCC s_andn2_b32 s_save_status, s_save_status, SQ_WAVE_STATUS_SPI_PRIO_MASK //check whether this is for save + +if SINGLE_STEP_MISSED_WORKAROUND + // No single step exceptions if MODE.DEBUG_EN=0. + s_getreg_b32 ttmp2, hwreg(HW_REG_MODE) + s_and_b32 ttmp2, ttmp2, SQ_WAVE_MODE_DEBUG_EN_MASK + s_cbranch_scc0 L_NO_SINGLE_STEP_WORKAROUND + + // Second-level trap already handled exception if STATUS.HALT=1. + s_and_b32 ttmp2, s_save_status, SQ_WAVE_STATUS_HALT_MASK + + // Prioritize single step exception over context save. + // Second-level trap will halt wave and RFE, re-entering for SAVECTX. + s_cbranch_scc0 L_FETCH_2ND_TRAP + +L_NO_SINGLE_STEP_WORKAROUND: +end + s_getreg_b32 s_save_trapsts, hwreg(HW_REG_TRAPSTS) s_and_b32 ttmp2, s_save_trapsts, SQ_WAVE_TRAPSTS_SAVECTX_MASK //check whether this is for save s_cbranch_scc1 L_SAVE //this is the operation for save // ********* Handle non-CWSR traps ******************* -if (!EMU_RUN_HACK) + // Illegal instruction is a non-maskable exception which blocks context save. // Halt the wavefront and return from the trap. s_and_b32 ttmp2, s_save_trapsts, SQ_WAVE_TRAPSTS_ILLEGAL_INST_MASK @@ -330,7 +277,7 @@ L_EXCP_CASE: set_status_without_spi_prio(s_save_status, ttmp2) s_rfe_b64 [ttmp0, ttmp1] -end + // ********* End handling of non-CWSR traps ******************* /**************************************************************************/ @@ -338,12 +285,6 @@ end /**************************************************************************/ L_SAVE: - -if G8SR_DEBUG_TIMESTAMP - s_memrealtime s_g8sr_ts_save_s - s_waitcnt lgkmcnt(0) //FIXME, will cause xnack?? -end - s_and_b32 s_save_pc_hi, s_save_pc_hi, 0x0000ffff //pc[47:32] s_mov_b32 s_save_tmp, 0 //clear saveCtx bit @@ -365,16 +306,7 @@ end s_mov_b32 s_save_exec_hi, exec_hi s_mov_b64 exec, 0x0 //clear EXEC to get ready to receive -if G8SR_DEBUG_TIMESTAMP - s_memrealtime s_g8sr_ts_sq_save_msg - s_waitcnt lgkmcnt(0) -end - - if (EMU_RUN_HACK) - - else s_sendmsg sendmsg(MSG_SAVEWAVE) //send SPI a message and wait for SPI's write to EXEC - end // Set SPI_PRIO=2 to avoid starving instruction fetch in the waves we're waiting for. s_or_b32 s_save_tmp, s_save_status, (2 << SQ_WAVE_STATUS_SPI_PRIO_SHIFT) @@ -383,33 +315,7 @@ end L_SLEEP: s_sleep 0x2 // sleep 1 (64clk) is not enough for 8 waves per SIMD, which will cause SQ hang, since the 7,8th wave could not get arbit to exec inst, while other waves are stuck into the sleep-loop and waiting for wrexec!=0 - if (EMU_RUN_HACK) - - else s_cbranch_execz L_SLEEP - end - -if G8SR_DEBUG_TIMESTAMP - s_memrealtime s_g8sr_ts_spi_wrexec - s_waitcnt lgkmcnt(0) -end - - if ((EMU_RUN_HACK) && (!EMU_RUN_HACK_SAVE_SINGLE_WAVE)) - //calculate wd_addr using absolute thread id - v_readlane_b32 s_save_tmp, v9, 0 - s_lshr_b32 s_save_tmp, s_save_tmp, 6 - s_mul_i32 s_save_tmp, s_save_tmp, WAVE_SPACE - s_add_i32 s_save_spi_init_lo, s_save_tmp, WG_BASE_ADDR_LO - s_mov_b32 s_save_spi_init_hi, WG_BASE_ADDR_HI - s_and_b32 s_save_spi_init_hi, s_save_spi_init_hi, CTX_SAVE_CONTROL - else - end - if ((EMU_RUN_HACK) && (EMU_RUN_HACK_SAVE_SINGLE_WAVE)) - s_add_i32 s_save_spi_init_lo, s_save_tmp, WG_BASE_ADDR_LO - s_mov_b32 s_save_spi_init_hi, WG_BASE_ADDR_HI - s_and_b32 s_save_spi_init_hi, s_save_spi_init_hi, CTX_SAVE_CONTROL - else - end // Save trap temporaries 4-11, 13 initialized by SPI debug dispatch logic // ttmp SR memory offset : size(VGPR)+size(SGPR)+0x40 @@ -459,20 +365,10 @@ end s_mov_b32 s_save_buf_rsrc2, 0x4 //NUM_RECORDS in bytes - if (SWIZZLE_EN) - s_add_u32 s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0 //FIXME need to use swizzle to enable bounds checking? - else s_mov_b32 s_save_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - end write_hwreg_to_mem(s_save_m0, s_save_buf_rsrc0, s_save_mem_offset) //M0 - - if ((EMU_RUN_HACK) && (EMU_RUN_HACK_SAVE_FIRST_TIME)) - s_add_u32 s_save_pc_lo, s_save_pc_lo, 4 //pc[31:0]+4 - s_addc_u32 s_save_pc_hi, s_save_pc_hi, 0x0 //carry bit over - end - write_hwreg_to_mem(s_save_pc_lo, s_save_buf_rsrc0, s_save_mem_offset) //PC write_hwreg_to_mem(s_save_pc_hi, s_save_buf_rsrc0, s_save_mem_offset) write_hwreg_to_mem(s_save_exec_lo, s_save_buf_rsrc0, s_save_mem_offset) //EXEC @@ -510,17 +406,9 @@ end s_add_u32 s_save_alloc_size, s_save_alloc_size, 1 s_lshl_b32 s_save_alloc_size, s_save_alloc_size, 4 //Number of SGPRs = (sgpr_size + 1) * 16 (non-zero value) - if (SGPR_SAVE_USE_SQC) s_lshl_b32 s_save_buf_rsrc2, s_save_alloc_size, 2 //NUM_RECORDS in bytes - else - s_lshl_b32 s_save_buf_rsrc2, s_save_alloc_size, 8 //NUM_RECORDS in bytes (64 threads) - end - if (SWIZZLE_EN) - s_add_u32 s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0 //FIXME need to use swizzle to enable bounds checking? - else s_mov_b32 s_save_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - end // backup s_save_buf_rsrc0,1 to s_save_pc_lo/hi, since write_16sgpr_to_mem function will change the rsrc0 @@ -563,30 +451,25 @@ end s_mov_b32 xnack_mask_lo, 0x0 s_mov_b32 xnack_mask_hi, 0x0 - if (SWIZZLE_EN) - s_add_u32 s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0 //FIXME need to use swizzle to enable bounds checking? - else s_mov_b32 s_save_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - end // VGPR Allocated in 4-GPR granularity -if G8SR_VGPR_SR_IN_DWX4 - // the const stride for DWx4 is 4*4 bytes - s_and_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, 0x0000FFFF // reset const stride to 0 - s_or_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, G8SR_SAVE_BUF_RSRC_WORD1_STRIDE_DWx4 // const stride to 4*4 bytes +if SAVE_AFTER_XNACK_ERROR + check_if_tcp_store_ok() + s_cbranch_scc1 L_SAVE_FIRST_VGPRS_WITH_TCP + + write_vgprs_to_mem_with_sqc(v0, 4, s_save_buf_rsrc0, s_save_mem_offset) + s_branch L_SAVE_LDS - buffer_store_dwordx4 v0, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 +L_SAVE_FIRST_VGPRS_WITH_TCP: +end - s_and_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, 0x0000FFFF // reset const stride to 0 - s_or_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, S_SAVE_BUF_RSRC_WORD1_STRIDE // reset const stride to 4 bytes -else buffer_store_dword v0, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 buffer_store_dword v1, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 offset:256 buffer_store_dword v2, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 offset:256*2 buffer_store_dword v3, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 offset:256*3 -end @@ -621,66 +504,34 @@ end s_add_u32 s_save_mem_offset, s_save_mem_offset, get_hwreg_size_bytes() - if (SWIZZLE_EN) - s_add_u32 s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0 //FIXME need to use swizzle to enable bounds checking? - else s_mov_b32 s_save_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - end s_mov_b32 m0, 0x0 //lds_offset initial value = 0 -var LDS_DMA_ENABLE = 0 -var UNROLL = 0 -if UNROLL==0 && LDS_DMA_ENABLE==1 - s_mov_b32 s3, 256*2 - s_nop 0 - s_nop 0 - s_nop 0 - L_SAVE_LDS_LOOP: - //TODO: looks the 2 buffer_store/load clause for s/r will hurt performance.??? - if (SAVE_LDS) //SPI always alloc LDS space in 128DW granularity - buffer_store_lds_dword s_save_buf_rsrc0, s_save_mem_offset lds:1 // first 64DW - buffer_store_lds_dword s_save_buf_rsrc0, s_save_mem_offset lds:1 offset:256 // second 64DW - end - - s_add_u32 m0, m0, s3 //every buffer_store_lds does 256 bytes - s_add_u32 s_save_mem_offset, s_save_mem_offset, s3 //mem offset increased by 256 bytes - s_cmp_lt_u32 m0, s_save_alloc_size //scc=(m0 < s_save_alloc_size) ? 1 : 0 - s_cbranch_scc1 L_SAVE_LDS_LOOP //LDS save is complete? - -elsif LDS_DMA_ENABLE==1 && UNROLL==1 // UNROOL , has ichace miss - // store from higest LDS address to lowest - s_mov_b32 s3, 256*2 - s_sub_u32 m0, s_save_alloc_size, s3 - s_add_u32 s_save_mem_offset, s_save_mem_offset, m0 - s_lshr_b32 s_save_alloc_size, s_save_alloc_size, 9 // how many 128 trunks... - s_sub_u32 s_save_alloc_size, 128, s_save_alloc_size // store from higheset addr to lowest - s_mul_i32 s_save_alloc_size, s_save_alloc_size, 6*4 // PC offset increment, each LDS save block cost 6*4 Bytes instruction - s_add_u32 s_save_alloc_size, s_save_alloc_size, 3*4 //2is the below 2 inst...//s_addc and s_setpc - s_nop 0 - s_nop 0 - s_nop 0 //pad 3 dw to let LDS_DMA align with 64Bytes - s_getpc_b64 s[0:1] // reuse s[0:1], since s[0:1] already saved - s_add_u32 s0, s0,s_save_alloc_size - s_addc_u32 s1, s1, 0 - s_setpc_b64 s[0:1] - - - for var i =0; i< 128; i++ - // be careful to make here a 64Byte aligned address, which could improve performance... - buffer_store_lds_dword s_save_buf_rsrc0, s_save_mem_offset lds:1 offset:0 // first 64DW - buffer_store_lds_dword s_save_buf_rsrc0, s_save_mem_offset lds:1 offset:256 // second 64DW - - if i!=127 - s_sub_u32 m0, m0, s3 // use a sgpr to shrink 2DW-inst to 1DW inst to improve performance , i.e. pack more LDS_DMA inst to one Cacheline - s_sub_u32 s_save_mem_offset, s_save_mem_offset, s3 - end - end - -else // BUFFER_STORE v_mbcnt_lo_u32_b32 v2, 0xffffffff, 0x0 v_mbcnt_hi_u32_b32 v3, 0xffffffff, v2 // tid + +if SAVE_AFTER_XNACK_ERROR + check_if_tcp_store_ok() + s_cbranch_scc1 L_SAVE_LDS_WITH_TCP + + v_lshlrev_b32 v2, 2, v3 +L_SAVE_LDS_LOOP_SQC: + ds_read2_b32 v[0:1], v2 offset0:0 offset1:0x40 + s_waitcnt lgkmcnt(0) + + write_vgprs_to_mem_with_sqc(v0, 2, s_save_buf_rsrc0, s_save_mem_offset) + + v_add_u32 v2, 0x200, v2 + v_cmp_lt_u32 vcc[0:1], v2, s_save_alloc_size + s_cbranch_vccnz L_SAVE_LDS_LOOP_SQC + + s_branch L_SAVE_LDS_DONE + +L_SAVE_LDS_WITH_TCP: +end + v_mul_i32_i24 v2, v3, 8 // tid*8 v_mov_b32 v3, 256*2 s_mov_b32 m0, 0x10000 @@ -701,8 +552,6 @@ L_SAVE_LDS_LOOP_VECTOR: // restore rsrc3 s_mov_b32 s_save_buf_rsrc3, s0 -end - L_SAVE_LDS_DONE: @@ -720,44 +569,9 @@ L_SAVE_LDS_DONE: s_add_u32 s_save_alloc_size, s_save_alloc_size, 1 s_lshl_b32 s_save_alloc_size, s_save_alloc_size, 2 //Number of VGPRs = (vgpr_size + 1) * 4 (non-zero value) //FIXME for GFX, zero is possible s_lshl_b32 s_save_buf_rsrc2, s_save_alloc_size, 8 //NUM_RECORDS in bytes (64 threads*4) - if (SWIZZLE_EN) - s_add_u32 s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0 //FIXME need to use swizzle to enable bounds checking? - else s_mov_b32 s_save_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - end - - - // VGPR Allocated in 4-GPR granularity - -if G8SR_VGPR_SR_IN_DWX4 - // the const stride for DWx4 is 4*4 bytes - s_and_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, 0x0000FFFF // reset const stride to 0 - s_or_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, G8SR_SAVE_BUF_RSRC_WORD1_STRIDE_DWx4 // const stride to 4*4 bytes - - s_mov_b32 m0, 4 // skip first 4 VGPRs - s_cmp_lt_u32 m0, s_save_alloc_size - s_cbranch_scc0 L_SAVE_VGPR_LOOP_END // no more vgprs - - s_set_gpr_idx_on m0, 0x1 // This will change M0 - s_add_u32 s_save_alloc_size, s_save_alloc_size, 0x1000 // because above inst change m0 -L_SAVE_VGPR_LOOP: - v_mov_b32 v0, v0 // v0 = v[0+m0] - v_mov_b32 v1, v1 - v_mov_b32 v2, v2 - v_mov_b32 v3, v3 - - buffer_store_dwordx4 v0, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 - s_add_u32 m0, m0, 4 - s_add_u32 s_save_mem_offset, s_save_mem_offset, 256*4 - s_cmp_lt_u32 m0, s_save_alloc_size - s_cbranch_scc1 L_SAVE_VGPR_LOOP //VGPR save is complete? - s_set_gpr_idx_off -L_SAVE_VGPR_LOOP_END: - s_and_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, 0x0000FFFF // reset const stride to 0 - s_or_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, S_SAVE_BUF_RSRC_WORD1_STRIDE // reset const stride to 4 bytes -else // VGPR store using dw burst s_mov_b32 m0, 0x4 //VGPR initial index value =0 s_cmp_lt_u32 m0, s_save_alloc_size @@ -767,57 +581,82 @@ else s_set_gpr_idx_on m0, 0x1 //M0[7:0] = M0[7:0] and M0[15:12] = 0x1 s_add_u32 s_save_alloc_size, s_save_alloc_size, 0x1000 //add 0x1000 since we compare m0 against it later +if SAVE_AFTER_XNACK_ERROR + check_if_tcp_store_ok() + s_cbranch_scc1 L_SAVE_VGPR_LOOP + +L_SAVE_VGPR_LOOP_SQC: + write_vgprs_to_mem_with_sqc(v0, 4, s_save_buf_rsrc0, s_save_mem_offset) + + s_add_u32 m0, m0, 4 + s_cmp_lt_u32 m0, s_save_alloc_size + s_cbranch_scc1 L_SAVE_VGPR_LOOP_SQC + + s_set_gpr_idx_off + s_branch L_SAVE_VGPR_END +end + L_SAVE_VGPR_LOOP: v_mov_b32 v0, v0 //v0 = v[0+m0] v_mov_b32 v1, v1 //v0 = v[0+m0] v_mov_b32 v2, v2 //v0 = v[0+m0] v_mov_b32 v3, v3 //v0 = v[0+m0] - if(USE_MTBUF_INSTEAD_OF_MUBUF) - tbuffer_store_format_x v0, v0, s_save_buf_rsrc0, s_save_mem_offset format:BUF_NUM_FORMAT_FLOAT format: BUF_DATA_FORMAT_32 slc:1 glc:1 - else buffer_store_dword v0, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 buffer_store_dword v1, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 offset:256 buffer_store_dword v2, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 offset:256*2 buffer_store_dword v3, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 offset:256*3 - end s_add_u32 m0, m0, 4 //next vgpr index s_add_u32 s_save_mem_offset, s_save_mem_offset, 256*4 //every buffer_store_dword does 256 bytes s_cmp_lt_u32 m0, s_save_alloc_size //scc = (m0 < s_save_alloc_size) ? 1 : 0 s_cbranch_scc1 L_SAVE_VGPR_LOOP //VGPR save is complete? s_set_gpr_idx_off -end L_SAVE_VGPR_END: +if ASIC_TARGET_ARCTURUS + // Save ACC VGPRs + s_mov_b32 m0, 0x0 //VGPR initial index value =0 + s_set_gpr_idx_on m0, 0x1 //M0[7:0] = M0[7:0] and M0[15:12] = 0x1 + +if SAVE_AFTER_XNACK_ERROR + check_if_tcp_store_ok() + s_cbranch_scc1 L_SAVE_ACCVGPR_LOOP +L_SAVE_ACCVGPR_LOOP_SQC: + for var vgpr = 0; vgpr < 4; ++ vgpr + v_accvgpr_read v[vgpr], acc[vgpr] // v[N] = acc[N+m0] + end + write_vgprs_to_mem_with_sqc(v0, 4, s_save_buf_rsrc0, s_save_mem_offset) + s_add_u32 m0, m0, 4 + s_cmp_lt_u32 m0, s_save_alloc_size + s_cbranch_scc1 L_SAVE_ACCVGPR_LOOP_SQC + s_set_gpr_idx_off + s_branch L_SAVE_ACCVGPR_END +end - /* S_PGM_END_SAVED */ //FIXME graphics ONLY - if ((EMU_RUN_HACK) && (!EMU_RUN_HACK_SAVE_NORMAL_EXIT)) - s_and_b32 s_save_pc_hi, s_save_pc_hi, 0x0000ffff //pc[47:32] - s_add_u32 s_save_pc_lo, s_save_pc_lo, 4 //pc[31:0]+4 - s_addc_u32 s_save_pc_hi, s_save_pc_hi, 0x0 //carry bit over - s_rfe_b64 s_save_pc_lo //Return to the main shader program - else +L_SAVE_ACCVGPR_LOOP: + for var vgpr = 0; vgpr < 4; ++ vgpr + v_accvgpr_read v[vgpr], acc[vgpr] // v[N] = acc[N+m0] end -// Save Done timestamp -if G8SR_DEBUG_TIMESTAMP - s_memrealtime s_g8sr_ts_save_d - // SGPR SR memory offset : size(VGPR) - get_vgpr_size_bytes(s_save_mem_offset) - s_add_u32 s_save_mem_offset, s_save_mem_offset, G8SR_DEBUG_TS_SAVE_D_OFFSET - s_waitcnt lgkmcnt(0) //FIXME, will cause xnack?? - // Need reset rsrc2?? - s_mov_b32 m0, s_save_mem_offset - s_mov_b32 s_save_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - s_buffer_store_dwordx2 s_g8sr_ts_save_d, s_save_buf_rsrc0, m0 glc:1 -end + buffer_store_dword v0, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 + buffer_store_dword v1, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 offset:256 + buffer_store_dword v2, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 offset:256*2 + buffer_store_dword v3, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1 offset:256*3 + s_add_u32 m0, m0, 4 + s_add_u32 s_save_mem_offset, s_save_mem_offset, 256*4 + s_cmp_lt_u32 m0, s_save_alloc_size + s_cbranch_scc1 L_SAVE_ACCVGPR_LOOP + s_set_gpr_idx_off + +L_SAVE_ACCVGPR_END: +end s_branch L_END_PGM @@ -829,27 +668,6 @@ end L_RESTORE: /* Setup Resource Contants */ - if ((EMU_RUN_HACK) && (!EMU_RUN_HACK_RESTORE_NORMAL)) - //calculate wd_addr using absolute thread id - v_readlane_b32 s_restore_tmp, v9, 0 - s_lshr_b32 s_restore_tmp, s_restore_tmp, 6 - s_mul_i32 s_restore_tmp, s_restore_tmp, WAVE_SPACE - s_add_i32 s_restore_spi_init_lo, s_restore_tmp, WG_BASE_ADDR_LO - s_mov_b32 s_restore_spi_init_hi, WG_BASE_ADDR_HI - s_and_b32 s_restore_spi_init_hi, s_restore_spi_init_hi, CTX_RESTORE_CONTROL - else - end - -if G8SR_DEBUG_TIMESTAMP - s_memrealtime s_g8sr_ts_restore_s - s_waitcnt lgkmcnt(0) //FIXME, will cause xnack?? - // tma_lo/hi are sgpr 110, 111, which will not used for 112 SGPR allocated case... - s_mov_b32 s_restore_pc_lo, s_g8sr_ts_restore_s[0] - s_mov_b32 s_restore_pc_hi, s_g8sr_ts_restore_s[1] //backup ts to ttmp0/1, sicne exec will be finally restored.. -end - - - s_mov_b32 s_restore_buf_rsrc0, s_restore_spi_init_lo //base_addr_lo s_and_b32 s_restore_buf_rsrc1, s_restore_spi_init_hi, 0x0000FFFF //base_addr_hi s_or_b32 s_restore_buf_rsrc1, s_restore_buf_rsrc1, S_RESTORE_BUF_RSRC_WORD1_STRIDE @@ -891,18 +709,12 @@ end s_add_u32 s_restore_mem_offset, s_restore_mem_offset, get_hwreg_size_bytes() //FIXME, Check if offset overflow??? - if (SWIZZLE_EN) - s_add_u32 s_restore_buf_rsrc2, s_restore_buf_rsrc2, 0x0 //FIXME need to use swizzle to enable bounds checking? - else s_mov_b32 s_restore_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - end s_mov_b32 m0, 0x0 //lds_offset initial value = 0 L_RESTORE_LDS_LOOP: - if (SAVE_LDS) buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset lds:1 // first 64DW buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset lds:1 offset:256 // second 64DW - end s_add_u32 m0, m0, 256*2 // 128 DW s_add_u32 s_restore_mem_offset, s_restore_mem_offset, 256*2 //mem offset increased by 128DW s_cmp_lt_u32 m0, s_restore_alloc_size //scc=(m0 < s_restore_alloc_size) ? 1 : 0 @@ -921,56 +733,43 @@ end s_add_u32 s_restore_alloc_size, s_restore_alloc_size, 1 s_lshl_b32 s_restore_alloc_size, s_restore_alloc_size, 2 //Number of VGPRs = (vgpr_size + 1) * 4 (non-zero value) s_lshl_b32 s_restore_buf_rsrc2, s_restore_alloc_size, 8 //NUM_RECORDS in bytes (64 threads*4) - if (SWIZZLE_EN) - s_add_u32 s_restore_buf_rsrc2, s_restore_buf_rsrc2, 0x0 //FIXME need to use swizzle to enable bounds checking? - else + +if ASIC_TARGET_ARCTURUS + s_mov_b32 s_restore_accvgpr_offset, s_restore_buf_rsrc2 //ACC VGPRs at end of VGPRs +end + s_mov_b32 s_restore_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - end -if G8SR_VGPR_SR_IN_DWX4 - get_vgpr_size_bytes(s_restore_mem_offset) - s_sub_u32 s_restore_mem_offset, s_restore_mem_offset, 256*4 - - // the const stride for DWx4 is 4*4 bytes - s_and_b32 s_restore_buf_rsrc1, s_restore_buf_rsrc1, 0x0000FFFF // reset const stride to 0 - s_or_b32 s_restore_buf_rsrc1, s_restore_buf_rsrc1, G8SR_RESTORE_BUF_RSRC_WORD1_STRIDE_DWx4 // const stride to 4*4 bytes - - s_mov_b32 m0, s_restore_alloc_size - s_set_gpr_idx_on m0, 0x8 // Note.. This will change m0 - -L_RESTORE_VGPR_LOOP: - buffer_load_dwordx4 v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 - s_waitcnt vmcnt(0) - s_sub_u32 m0, m0, 4 - v_mov_b32 v0, v0 // v[0+m0] = v0 - v_mov_b32 v1, v1 - v_mov_b32 v2, v2 - v_mov_b32 v3, v3 - s_sub_u32 s_restore_mem_offset, s_restore_mem_offset, 256*4 - s_cmp_eq_u32 m0, 0x8000 - s_cbranch_scc0 L_RESTORE_VGPR_LOOP - s_set_gpr_idx_off - - s_and_b32 s_restore_buf_rsrc1, s_restore_buf_rsrc1, 0x0000FFFF // reset const stride to 0 - s_or_b32 s_restore_buf_rsrc1, s_restore_buf_rsrc1, S_RESTORE_BUF_RSRC_WORD1_STRIDE // const stride to 4*4 bytes - -else // VGPR load using dw burst s_mov_b32 s_restore_mem_offset_save, s_restore_mem_offset // restore start with v1, v0 will be the last s_add_u32 s_restore_mem_offset, s_restore_mem_offset, 256*4 +if ASIC_TARGET_ARCTURUS + s_mov_b32 s_restore_accvgpr_offset_save, s_restore_accvgpr_offset + s_add_u32 s_restore_accvgpr_offset, s_restore_accvgpr_offset, 256*4 +end s_mov_b32 m0, 4 //VGPR initial index value = 1 s_set_gpr_idx_on m0, 0x8 //M0[7:0] = M0[7:0] and M0[15:12] = 0x8 s_add_u32 s_restore_alloc_size, s_restore_alloc_size, 0x8000 //add 0x8000 since we compare m0 against it later L_RESTORE_VGPR_LOOP: - if(USE_MTBUF_INSTEAD_OF_MUBUF) - tbuffer_load_format_x v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset format:BUF_NUM_FORMAT_FLOAT format: BUF_DATA_FORMAT_32 slc:1 glc:1 - else + +if ASIC_TARGET_ARCTURUS + buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_accvgpr_offset slc:1 glc:1 + buffer_load_dword v1, v0, s_restore_buf_rsrc0, s_restore_accvgpr_offset slc:1 glc:1 offset:256 + buffer_load_dword v2, v0, s_restore_buf_rsrc0, s_restore_accvgpr_offset slc:1 glc:1 offset:256*2 + buffer_load_dword v3, v0, s_restore_buf_rsrc0, s_restore_accvgpr_offset slc:1 glc:1 offset:256*3 + s_add_u32 s_restore_accvgpr_offset, s_restore_accvgpr_offset, 256*4 + s_waitcnt vmcnt(0) + + for var vgpr = 0; vgpr < 4; ++ vgpr + v_accvgpr_write acc[vgpr], v[vgpr] + end +end + buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 buffer_load_dword v1, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 offset:256 buffer_load_dword v2, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 offset:256*2 buffer_load_dword v3, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 offset:256*3 - end s_waitcnt vmcnt(0) //ensure data ready v_mov_b32 v0, v0 //v[0+m0] = v0 v_mov_b32 v1, v1 @@ -982,16 +781,22 @@ else s_cbranch_scc1 L_RESTORE_VGPR_LOOP //VGPR restore (except v0) is complete? s_set_gpr_idx_off /* VGPR restore on v0 */ - if(USE_MTBUF_INSTEAD_OF_MUBUF) - tbuffer_load_format_x v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save format:BUF_NUM_FORMAT_FLOAT format: BUF_DATA_FORMAT_32 slc:1 glc:1 - else +if ASIC_TARGET_ARCTURUS + buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_accvgpr_offset_save slc:1 glc:1 + buffer_load_dword v1, v0, s_restore_buf_rsrc0, s_restore_accvgpr_offset_save slc:1 glc:1 offset:256 + buffer_load_dword v2, v0, s_restore_buf_rsrc0, s_restore_accvgpr_offset_save slc:1 glc:1 offset:256*2 + buffer_load_dword v3, v0, s_restore_buf_rsrc0, s_restore_accvgpr_offset_save slc:1 glc:1 offset:256*3 + s_waitcnt vmcnt(0) + + for var vgpr = 0; vgpr < 4; ++ vgpr + v_accvgpr_write acc[vgpr], v[vgpr] + end +end + buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save slc:1 glc:1 buffer_load_dword v1, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save slc:1 glc:1 offset:256 buffer_load_dword v2, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save slc:1 glc:1 offset:256*2 buffer_load_dword v3, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save slc:1 glc:1 offset:256*3 - end - -end /* restore SGPRs */ ////////////////////////////// @@ -1007,16 +812,8 @@ end s_add_u32 s_restore_alloc_size, s_restore_alloc_size, 1 s_lshl_b32 s_restore_alloc_size, s_restore_alloc_size, 4 //Number of SGPRs = (sgpr_size + 1) * 16 (non-zero value) - if (SGPR_SAVE_USE_SQC) s_lshl_b32 s_restore_buf_rsrc2, s_restore_alloc_size, 2 //NUM_RECORDS in bytes - else - s_lshl_b32 s_restore_buf_rsrc2, s_restore_alloc_size, 8 //NUM_RECORDS in bytes (64 threads) - end - if (SWIZZLE_EN) - s_add_u32 s_restore_buf_rsrc2, s_restore_buf_rsrc2, 0x0 //FIXME need to use swizzle to enable bounds checking? - else s_mov_b32 s_restore_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - end s_mov_b32 m0, s_restore_alloc_size @@ -1044,11 +841,6 @@ end L_RESTORE_HWREG: -if G8SR_DEBUG_TIMESTAMP - s_mov_b32 s_g8sr_ts_restore_s[0], s_restore_pc_lo - s_mov_b32 s_g8sr_ts_restore_s[1], s_restore_pc_hi -end - // HWREG SR memory offset : size(VGPR)+size(SGPR) get_vgpr_size_bytes(s_restore_mem_offset) get_sgpr_size_bytes(s_restore_tmp) @@ -1056,11 +848,7 @@ end s_mov_b32 s_restore_buf_rsrc2, 0x4 //NUM_RECORDS in bytes - if (SWIZZLE_EN) - s_add_u32 s_restore_buf_rsrc2, s_restore_buf_rsrc2, 0x0 //FIXME need to use swizzle to enable bounds checking? - else s_mov_b32 s_restore_buf_rsrc2, 0x1000000 //NUM_RECORDS in bytes - end read_hwreg_from_mem(s_restore_m0, s_restore_buf_rsrc0, s_restore_mem_offset) //M0 read_hwreg_from_mem(s_restore_pc_lo, s_restore_buf_rsrc0, s_restore_mem_offset) //PC @@ -1075,16 +863,6 @@ end s_waitcnt lgkmcnt(0) //from now on, it is safe to restore STATUS and IB_STS - //for normal save & restore, the saved PC points to the next inst to execute, no adjustment needs to be made, otherwise: - if ((EMU_RUN_HACK) && (!EMU_RUN_HACK_RESTORE_NORMAL)) - s_add_u32 s_restore_pc_lo, s_restore_pc_lo, 8 //pc[31:0]+8 //two back-to-back s_trap are used (first for save and second for restore) - s_addc_u32 s_restore_pc_hi, s_restore_pc_hi, 0x0 //carry bit over - end - if ((EMU_RUN_HACK) && (EMU_RUN_HACK_RESTORE_NORMAL)) - s_add_u32 s_restore_pc_lo, s_restore_pc_lo, 4 //pc[31:0]+4 // save is hack through s_trap but restore is normal - s_addc_u32 s_restore_pc_hi, s_restore_pc_hi, 0x0 //carry bit over - end - s_mov_b32 m0, s_restore_m0 s_mov_b32 exec_lo, s_restore_exec_lo s_mov_b32 exec_hi, s_restore_exec_hi @@ -1131,11 +909,6 @@ end s_barrier //barrier to ensure the readiness of LDS before access attempts from any other wave in the same TG //FIXME not performance-optimal at this time -if G8SR_DEBUG_TIMESTAMP - s_memrealtime s_g8sr_ts_restore_d - s_waitcnt lgkmcnt(0) -end - // s_rfe_b64 s_restore_pc_lo //Return to the main shader program and resume execution s_rfe_restore_b64 s_restore_pc_lo, s_restore_m0 // s_restore_m0[0] is used to set STATUS.inst_atc @@ -1190,7 +963,39 @@ function read_16sgpr_from_mem(s, s_rsrc, s_mem_offset) s_sub_u32 s_mem_offset, s_mem_offset, 4*16 end +function check_if_tcp_store_ok + // If STATUS.ALLOW_REPLAY=0 and TRAPSTS.XNACK_ERROR=1 then TCP stores will fail. + s_and_b32 s_save_tmp, s_save_status, SQ_WAVE_STATUS_ALLOW_REPLAY_MASK + s_cbranch_scc1 L_TCP_STORE_CHECK_DONE + + s_getreg_b32 s_save_tmp, hwreg(HW_REG_TRAPSTS) + s_andn2_b32 s_save_tmp, SQ_WAVE_TRAPSTS_XNACK_ERROR_MASK, s_save_tmp + +L_TCP_STORE_CHECK_DONE: +end + +function write_vgpr_to_mem_with_sqc(v, s_rsrc, s_mem_offset) + s_mov_b32 s4, 0 +L_WRITE_VGPR_LANE_LOOP: + for var lane = 0; lane < 4; ++ lane + v_readlane_b32 s[lane], v, s4 + s_add_u32 s4, s4, 1 + end + + s_buffer_store_dwordx4 s[0:3], s_rsrc, s_mem_offset glc:1 + ack_sqc_store_workaround() + + s_add_u32 s_mem_offset, s_mem_offset, 0x10 + s_cmp_eq_u32 s4, 0x40 + s_cbranch_scc0 L_WRITE_VGPR_LANE_LOOP +end + +function write_vgprs_to_mem_with_sqc(v, n_vgprs, s_rsrc, s_mem_offset) + for var vgpr = 0; vgpr < n_vgprs; ++ vgpr + write_vgpr_to_mem_with_sqc(v[vgpr], s_rsrc, s_mem_offset) + end +end function get_lds_size_bytes(s_lds_size_byte) // SQ LDS granularity is 64DW, while PGM_RSRC2.lds_size is in granularity 128DW @@ -1202,6 +1007,10 @@ function get_vgpr_size_bytes(s_vgpr_size_byte) s_getreg_b32 s_vgpr_size_byte, hwreg(HW_REG_GPR_ALLOC,SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SHIFT,SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SIZE) //vpgr_size s_add_u32 s_vgpr_size_byte, s_vgpr_size_byte, 1 s_lshl_b32 s_vgpr_size_byte, s_vgpr_size_byte, (2+8) //Number of VGPRs = (vgpr_size + 1) * 4 * 64 * 4 (non-zero value) //FIXME for GFX, zero is possible + +if ASIC_TARGET_ARCTURUS + s_lshl_b32 s_vgpr_size_byte, s_vgpr_size_byte, 1 // Double size for ACC VGPRs +end end function get_sgpr_size_bytes(s_sgpr_size_byte) diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index 1d3cd5c50d5f..3f0300e53727 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -42,6 +42,7 @@ static long kfd_ioctl(struct file *, unsigned int, unsigned long); static int kfd_open(struct inode *, struct file *); +static int kfd_release(struct inode *, struct file *); static int kfd_mmap(struct file *, struct vm_area_struct *); static const char kfd_dev_name[] = "kfd"; @@ -49,8 +50,9 @@ static const char kfd_dev_name[] = "kfd"; static const struct file_operations kfd_fops = { .owner = THIS_MODULE, .unlocked_ioctl = kfd_ioctl, - .compat_ioctl = kfd_ioctl, + .compat_ioctl = compat_ptr_ioctl, .open = kfd_open, + .release = kfd_release, .mmap = kfd_mmap, }; @@ -124,8 +126,13 @@ static int kfd_open(struct inode *inode, struct file *filep) if (IS_ERR(process)) return PTR_ERR(process); - if (kfd_is_locked()) + if (kfd_is_locked()) { + kfd_unref_process(process); return -EAGAIN; + } + + /* filep now owns the reference returned by kfd_create_process */ + filep->private_data = process; dev_dbg(kfd_device, "process %d opened, compat mode (32 bit) - %d\n", process->pasid, process->is_32bit_user_mode); @@ -133,6 +140,16 @@ static int kfd_open(struct inode *inode, struct file *filep) return 0; } +static int kfd_release(struct inode *inode, struct file *filep) +{ + struct kfd_process *process = filep->private_data; + + if (process) + kfd_unref_process(process); + + return 0; +} + static int kfd_ioctl_get_version(struct file *filep, struct kfd_process *p, void *data) { @@ -258,6 +275,7 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p, unsigned int queue_id; struct kfd_process_device *pdd; struct queue_properties q_properties; + uint32_t doorbell_offset_in_process = 0; memset(&q_properties, 0, sizeof(struct queue_properties)); @@ -282,11 +300,12 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p, goto err_bind_process; } - pr_debug("Creating queue for PASID %d on gpu 0x%x\n", + pr_debug("Creating queue for PASID 0x%x on gpu 0x%x\n", p->pasid, dev->id); - err = pqm_create_queue(&p->pqm, dev, filep, &q_properties, &queue_id); + err = pqm_create_queue(&p->pqm, dev, filep, &q_properties, &queue_id, + &doorbell_offset_in_process); if (err != 0) goto err_create_queue; @@ -296,14 +315,11 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p, /* Return gpu_id as doorbell offset for mmap usage */ args->doorbell_offset = KFD_MMAP_TYPE_DOORBELL; args->doorbell_offset |= KFD_MMAP_GPU_ID(args->gpu_id); - args->doorbell_offset <<= PAGE_SHIFT; if (KFD_IS_SOC15(dev->device_info->asic_family)) - /* On SOC15 ASICs, doorbell allocation must be - * per-device, and independent from the per-process - * queue_id. Return the doorbell offset within the - * doorbell aperture to user mode. + /* On SOC15 ASICs, include the doorbell offset within the + * process doorbell frame, which is 2 pages. */ - args->doorbell_offset |= q_properties.doorbell_off; + args->doorbell_offset |= doorbell_offset_in_process; mutex_unlock(&p->mutex); @@ -332,7 +348,7 @@ static int kfd_ioctl_destroy_queue(struct file *filp, struct kfd_process *p, int retval; struct kfd_ioctl_destroy_queue_args *args = data; - pr_debug("Destroying queue id %d for pasid %d\n", + pr_debug("Destroying queue id %d for pasid 0x%x\n", args->queue_id, p->pasid); @@ -378,7 +394,7 @@ static int kfd_ioctl_update_queue(struct file *filp, struct kfd_process *p, properties.queue_percent = args->queue_percentage; properties.priority = args->queue_priority; - pr_debug("Updating queue id %d for pasid %d\n", + pr_debug("Updating queue id %d for pasid 0x%x\n", args->queue_id, p->pasid); mutex_lock(&p->mutex); @@ -855,7 +871,7 @@ static int kfd_ioctl_get_process_apertures(struct file *filp, struct kfd_process_device_apertures *pAperture; struct kfd_process_device *pdd; - dev_dbg(kfd_device, "get apertures for PASID %d", p->pasid); + dev_dbg(kfd_device, "get apertures for PASID 0x%x", p->pasid); args->num_of_nodes = 0; @@ -913,7 +929,7 @@ static int kfd_ioctl_get_process_apertures_new(struct file *filp, uint32_t nodes = 0; int ret; - dev_dbg(kfd_device, "get apertures for PASID %d", p->pasid); + dev_dbg(kfd_device, "get apertures for PASID 0x%x", p->pasid); if (args->num_of_nodes == 0) { /* Return number of nodes, so that user space can alloacate @@ -1128,7 +1144,7 @@ static int kfd_ioctl_set_scratch_backing_va(struct file *filep, mutex_unlock(&p->mutex); if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS && - pdd->qpd.vmid != 0) + pdd->qpd.vmid != 0 && dev->kfd2kgd->set_scratch_backing_va) dev->kfd2kgd->set_scratch_backing_va( dev->kgd, args->va_addr, pdd->qpd.vmid); @@ -1312,10 +1328,9 @@ static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep, /* MMIO is mapped through kfd device * Generate a kfd mmap offset */ - if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) { - args->mmap_offset = KFD_MMAP_TYPE_MMIO | KFD_MMAP_GPU_ID(args->gpu_id); - args->mmap_offset <<= PAGE_SHIFT; - } + if (flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP) + args->mmap_offset = KFD_MMAP_TYPE_MMIO + | KFD_MMAP_GPU_ID(args->gpu_id); return 0; @@ -1801,11 +1816,16 @@ static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg) } else goto err_i1; - dev_dbg(kfd_device, "ioctl cmd 0x%x (#%d), arg 0x%lx\n", cmd, nr, arg); + dev_dbg(kfd_device, "ioctl cmd 0x%x (#0x%x), arg 0x%lx\n", cmd, nr, arg); - process = kfd_get_process(current); - if (IS_ERR(process)) { - dev_dbg(kfd_device, "no process\n"); + /* Get the process struct from the filep. Only the process + * that opened /dev/kfd can use the file descriptor. Child + * processes need to create their own KFD device context. + */ + process = filep->private_data; + if (process->lead_thread != current->group_leader) { + dev_dbg(kfd_device, "Using KFD FD in wrong process\n"); + retcode = -EBADF; goto err_i1; } @@ -1856,7 +1876,8 @@ err_i1: kfree(kdata); if (retcode) - dev_dbg(kfd_device, "ret = %d\n", retcode); + dev_dbg(kfd_device, "ioctl cmd (#0x%x), arg 0x%lx, ret = %d\n", + nr, arg, retcode); return retcode; } @@ -1877,7 +1898,7 @@ static int kfd_mmio_mmap(struct kfd_dev *dev, struct kfd_process *process, vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); - pr_debug("Process %d mapping mmio page\n" + pr_debug("pasid 0x%x mapping mmio page\n" " target user address == 0x%08llX\n" " physical address == 0x%08llX\n" " vm_flags == 0x%04lX\n" @@ -1898,20 +1919,19 @@ static int kfd_mmap(struct file *filp, struct vm_area_struct *vma) { struct kfd_process *process; struct kfd_dev *dev = NULL; - unsigned long vm_pgoff; + unsigned long mmap_offset; unsigned int gpu_id; process = kfd_get_process(current); if (IS_ERR(process)) return PTR_ERR(process); - vm_pgoff = vma->vm_pgoff; - vma->vm_pgoff = KFD_MMAP_OFFSET_VALUE_GET(vm_pgoff); - gpu_id = KFD_MMAP_GPU_ID_GET(vm_pgoff); + mmap_offset = vma->vm_pgoff << PAGE_SHIFT; + gpu_id = KFD_MMAP_GET_GPU_ID(mmap_offset); if (gpu_id) dev = kfd_device_by_id(gpu_id); - switch (vm_pgoff & KFD_MMAP_TYPE_MASK) { + switch (mmap_offset & KFD_MMAP_TYPE_MASK) { case KFD_MMAP_TYPE_DOORBELL: if (!dev) return -ENODEV; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_crat.c b/drivers/gpu/drm/amd/amdkfd/kfd_crat.c index 4e3fc284f6ac..de9f68d5c312 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_crat.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_crat.c @@ -138,6 +138,7 @@ static struct kfd_gpu_cache_info carrizo_cache_info[] = { /* TODO - check & update Vega10 cache details */ #define vega10_cache_info carrizo_cache_info #define raven_cache_info carrizo_cache_info +#define renoir_cache_info carrizo_cache_info /* TODO - check & update Navi10 cache details */ #define navi10_cache_info carrizo_cache_info @@ -662,6 +663,7 @@ static int kfd_fill_gpu_cache_info(struct kfd_dev *kdev, case CHIP_VEGA10: case CHIP_VEGA12: case CHIP_VEGA20: + case CHIP_ARCTURUS: pcache_info = vega10_cache_info; num_of_cache_types = ARRAY_SIZE(vega10_cache_info); break; @@ -669,7 +671,13 @@ static int kfd_fill_gpu_cache_info(struct kfd_dev *kdev, pcache_info = raven_cache_info; num_of_cache_types = ARRAY_SIZE(raven_cache_info); break; + case CHIP_RENOIR: + pcache_info = renoir_cache_info; + num_of_cache_types = ARRAY_SIZE(renoir_cache_info); + break; case CHIP_NAVI10: + case CHIP_NAVI12: + case CHIP_NAVI14: pcache_info = navi10_cache_info; num_of_cache_types = ARRAY_SIZE(navi10_cache_info); break; @@ -702,7 +710,7 @@ static int kfd_fill_gpu_cache_info(struct kfd_dev *kdev, pcache_info, cu_info, mem_available, - cu_info->cu_bitmap[i][j], + cu_info->cu_bitmap[i % 4][j + i / 4], ct, cu_processor_id, k); @@ -788,7 +796,7 @@ int kfd_create_crat_image_acpi(void **crat_image, size_t *size) * is put in the code to ensure we don't overwrite. */ #define VCRAT_SIZE_FOR_CPU (2 * PAGE_SIZE) -#define VCRAT_SIZE_FOR_GPU (3 * PAGE_SIZE) +#define VCRAT_SIZE_FOR_GPU (4 * PAGE_SIZE) /* kfd_fill_cu_for_cpu - Fill in Compute info for the given CPU NUMA node * diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_dbgdev.c b/drivers/gpu/drm/amd/amdkfd/kfd_dbgdev.c index a3441b0e385b..27bcc5b472f6 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_dbgdev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_dbgdev.c @@ -72,11 +72,11 @@ static int dbgdev_diq_submit_ib(struct kfd_dbgdev *dbgdev, * The receive packet buff will be sitting on the Indirect Buffer * and in the PQ we put the IB packet + sync packet(s). */ - status = kq->ops.acquire_packet_buffer(kq, + status = kq_acquire_packet_buffer(kq, pq_packets_size_in_bytes / sizeof(uint32_t), &ib_packet_buff); if (status) { - pr_err("acquire_packet_buffer failed\n"); + pr_err("kq_acquire_packet_buffer failed\n"); return status; } @@ -115,7 +115,7 @@ static int dbgdev_diq_submit_ib(struct kfd_dbgdev *dbgdev, if (status) { pr_err("Failed to allocate GART memory\n"); - kq->ops.rollback_packet(kq); + kq_rollback_packet(kq); return status; } @@ -151,7 +151,7 @@ static int dbgdev_diq_submit_ib(struct kfd_dbgdev *dbgdev, rm_packet->data_lo = QUEUESTATE__ACTIVE; - kq->ops.submit_packet(kq); + kq_submit_packet(kq); /* Wait till CP writes sync code: */ status = amdkfd_fence_wait_timeout( @@ -185,7 +185,7 @@ static int dbgdev_register_diq(struct kfd_dbgdev *dbgdev) properties.type = KFD_QUEUE_TYPE_DIQ; status = pqm_create_queue(dbgdev->pqm, dbgdev->dev, NULL, - &properties, &qid); + &properties, &qid, NULL); if (status) { pr_err("Failed to create DIQ\n"); @@ -761,6 +761,7 @@ int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p) { int status = 0; unsigned int vmid; + uint16_t queried_pasid; union SQ_CMD_BITS reg_sq_cmd; union GRBM_GFX_INDEX_BITS reg_gfx_index; struct kfd_process_device *pdd; @@ -782,19 +783,18 @@ int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p) */ for (vmid = first_vmid_to_scan; vmid <= last_vmid_to_scan; vmid++) { - if (dev->kfd2kgd->get_atc_vmid_pasid_mapping_valid - (dev->kgd, vmid)) { - if (dev->kfd2kgd->get_atc_vmid_pasid_mapping_pasid - (dev->kgd, vmid) == p->pasid) { - pr_debug("Killing wave fronts of vmid %d and pasid %d\n", - vmid, p->pasid); - break; - } + status = dev->kfd2kgd->get_atc_vmid_pasid_mapping_info + (dev->kgd, vmid, &queried_pasid); + + if (status && queried_pasid == p->pasid) { + pr_debug("Killing wave fronts of vmid %d and pasid 0x%x\n", + vmid, p->pasid); + break; } } if (vmid > last_vmid_to_scan) { - pr_err("Didn't find vmid for pasid %d\n", p->pasid); + pr_err("Didn't find vmid for pasid 0x%x\n", p->pasid); return -EFAULT; } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_dbgmgr.c b/drivers/gpu/drm/amd/amdkfd/kfd_dbgmgr.c index 9d4af961c5d1..9bfa50633654 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_dbgmgr.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_dbgmgr.c @@ -96,7 +96,7 @@ bool kfd_dbgmgr_create(struct kfd_dbgmgr **ppmgr, struct kfd_dev *pdev) long kfd_dbgmgr_register(struct kfd_dbgmgr *pmgr, struct kfd_process *p) { if (pmgr->pasid != 0) { - pr_debug("H/W debugger is already active using pasid %d\n", + pr_debug("H/W debugger is already active using pasid 0x%x\n", pmgr->pasid); return -EBUSY; } @@ -117,7 +117,7 @@ long kfd_dbgmgr_unregister(struct kfd_dbgmgr *pmgr, struct kfd_process *p) { /* Is the requests coming from the already registered process? */ if (pmgr->pasid != p->pasid) { - pr_debug("H/W debugger is not registered by calling pasid %d\n", + pr_debug("H/W debugger is not registered by calling pasid 0x%x\n", p->pasid); return -EINVAL; } @@ -134,7 +134,7 @@ long kfd_dbgmgr_wave_control(struct kfd_dbgmgr *pmgr, { /* Is the requests coming from the already registered process? */ if (pmgr->pasid != wac_info->process->pasid) { - pr_debug("H/W debugger support was not registered for requester pasid %d\n", + pr_debug("H/W debugger support was not registered for requester pasid 0x%x\n", wac_info->process->pasid); return -EINVAL; } @@ -147,7 +147,7 @@ long kfd_dbgmgr_address_watch(struct kfd_dbgmgr *pmgr, { /* Is the requests coming from the already registered process? */ if (pmgr->pasid != adw_info->process->pasid) { - pr_debug("H/W debugger support was not registered for requester pasid %d\n", + pr_debug("H/W debugger support was not registered for requester pasid 0x%x\n", adw_info->process->pasid); return -EINVAL; } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debugfs.c b/drivers/gpu/drm/amd/amdkfd/kfd_debugfs.c index 15c523027285..511712c2e382 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debugfs.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debugfs.c @@ -93,7 +93,7 @@ void kfd_debugfs_init(void) kfd_debugfs_hqds_by_device, &kfd_debugfs_fops); debugfs_create_file("rls", S_IFREG | 0444, debugfs_root, kfd_debugfs_rls_by_device, &kfd_debugfs_fops); - debugfs_create_file("hang_hws", S_IFREG | 0644, debugfs_root, + debugfs_create_file("hang_hws", S_IFREG | 0200, debugfs_root, NULL, &kfd_debugfs_hang_hws_fops); } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index 3322a443dfb2..2a9e40131735 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -39,9 +39,45 @@ */ static atomic_t kfd_locked = ATOMIC_INIT(0); +#ifdef CONFIG_DRM_AMDGPU_CIK +extern const struct kfd2kgd_calls gfx_v7_kfd2kgd; +#endif +extern const struct kfd2kgd_calls gfx_v8_kfd2kgd; +extern const struct kfd2kgd_calls gfx_v9_kfd2kgd; +extern const struct kfd2kgd_calls arcturus_kfd2kgd; +extern const struct kfd2kgd_calls gfx_v10_kfd2kgd; + +static const struct kfd2kgd_calls *kfd2kgd_funcs[] = { +#ifdef KFD_SUPPORT_IOMMU_V2 +#ifdef CONFIG_DRM_AMDGPU_CIK + [CHIP_KAVERI] = &gfx_v7_kfd2kgd, +#endif + [CHIP_CARRIZO] = &gfx_v8_kfd2kgd, + [CHIP_RAVEN] = &gfx_v9_kfd2kgd, +#endif +#ifdef CONFIG_DRM_AMDGPU_CIK + [CHIP_HAWAII] = &gfx_v7_kfd2kgd, +#endif + [CHIP_TONGA] = &gfx_v8_kfd2kgd, + [CHIP_FIJI] = &gfx_v8_kfd2kgd, + [CHIP_POLARIS10] = &gfx_v8_kfd2kgd, + [CHIP_POLARIS11] = &gfx_v8_kfd2kgd, + [CHIP_POLARIS12] = &gfx_v8_kfd2kgd, + [CHIP_VEGAM] = &gfx_v8_kfd2kgd, + [CHIP_VEGA10] = &gfx_v9_kfd2kgd, + [CHIP_VEGA12] = &gfx_v9_kfd2kgd, + [CHIP_VEGA20] = &gfx_v9_kfd2kgd, + [CHIP_RENOIR] = &gfx_v9_kfd2kgd, + [CHIP_ARCTURUS] = &arcturus_kfd2kgd, + [CHIP_NAVI10] = &gfx_v10_kfd2kgd, + [CHIP_NAVI12] = &gfx_v10_kfd2kgd, + [CHIP_NAVI14] = &gfx_v10_kfd2kgd, +}; + #ifdef KFD_SUPPORT_IOMMU_V2 static const struct kfd_device_info kaveri_device_info = { .asic_family = CHIP_KAVERI, + .asic_name = "kaveri", .max_pasid_bits = 16, /* max num of queues for KV.TODO should be a dynamic value */ .max_no_of_hqd = 24, @@ -60,6 +96,7 @@ static const struct kfd_device_info kaveri_device_info = { static const struct kfd_device_info carrizo_device_info = { .asic_family = CHIP_CARRIZO, + .asic_name = "carrizo", .max_pasid_bits = 16, /* max num of queues for CZ.TODO should be a dynamic value */ .max_no_of_hqd = 24, @@ -78,6 +115,7 @@ static const struct kfd_device_info carrizo_device_info = { static const struct kfd_device_info raven_device_info = { .asic_family = CHIP_RAVEN, + .asic_name = "raven", .max_pasid_bits = 16, .max_no_of_hqd = 24, .doorbell_size = 8, @@ -96,6 +134,7 @@ static const struct kfd_device_info raven_device_info = { static const struct kfd_device_info hawaii_device_info = { .asic_family = CHIP_HAWAII, + .asic_name = "hawaii", .max_pasid_bits = 16, /* max num of queues for KV.TODO should be a dynamic value */ .max_no_of_hqd = 24, @@ -114,6 +153,7 @@ static const struct kfd_device_info hawaii_device_info = { static const struct kfd_device_info tonga_device_info = { .asic_family = CHIP_TONGA, + .asic_name = "tonga", .max_pasid_bits = 16, .max_no_of_hqd = 24, .doorbell_size = 4, @@ -131,6 +171,7 @@ static const struct kfd_device_info tonga_device_info = { static const struct kfd_device_info fiji_device_info = { .asic_family = CHIP_FIJI, + .asic_name = "fiji", .max_pasid_bits = 16, .max_no_of_hqd = 24, .doorbell_size = 4, @@ -148,6 +189,7 @@ static const struct kfd_device_info fiji_device_info = { static const struct kfd_device_info fiji_vf_device_info = { .asic_family = CHIP_FIJI, + .asic_name = "fiji", .max_pasid_bits = 16, .max_no_of_hqd = 24, .doorbell_size = 4, @@ -166,6 +208,7 @@ static const struct kfd_device_info fiji_vf_device_info = { static const struct kfd_device_info polaris10_device_info = { .asic_family = CHIP_POLARIS10, + .asic_name = "polaris10", .max_pasid_bits = 16, .max_no_of_hqd = 24, .doorbell_size = 4, @@ -183,6 +226,7 @@ static const struct kfd_device_info polaris10_device_info = { static const struct kfd_device_info polaris10_vf_device_info = { .asic_family = CHIP_POLARIS10, + .asic_name = "polaris10", .max_pasid_bits = 16, .max_no_of_hqd = 24, .doorbell_size = 4, @@ -200,6 +244,7 @@ static const struct kfd_device_info polaris10_vf_device_info = { static const struct kfd_device_info polaris11_device_info = { .asic_family = CHIP_POLARIS11, + .asic_name = "polaris11", .max_pasid_bits = 16, .max_no_of_hqd = 24, .doorbell_size = 4, @@ -217,6 +262,7 @@ static const struct kfd_device_info polaris11_device_info = { static const struct kfd_device_info polaris12_device_info = { .asic_family = CHIP_POLARIS12, + .asic_name = "polaris12", .max_pasid_bits = 16, .max_no_of_hqd = 24, .doorbell_size = 4, @@ -234,6 +280,7 @@ static const struct kfd_device_info polaris12_device_info = { static const struct kfd_device_info vegam_device_info = { .asic_family = CHIP_VEGAM, + .asic_name = "vegam", .max_pasid_bits = 16, .max_no_of_hqd = 24, .doorbell_size = 4, @@ -251,6 +298,7 @@ static const struct kfd_device_info vegam_device_info = { static const struct kfd_device_info vega10_device_info = { .asic_family = CHIP_VEGA10, + .asic_name = "vega10", .max_pasid_bits = 16, .max_no_of_hqd = 24, .doorbell_size = 8, @@ -268,6 +316,7 @@ static const struct kfd_device_info vega10_device_info = { static const struct kfd_device_info vega10_vf_device_info = { .asic_family = CHIP_VEGA10, + .asic_name = "vega10", .max_pasid_bits = 16, .max_no_of_hqd = 24, .doorbell_size = 8, @@ -285,6 +334,7 @@ static const struct kfd_device_info vega10_vf_device_info = { static const struct kfd_device_info vega12_device_info = { .asic_family = CHIP_VEGA12, + .asic_name = "vega12", .max_pasid_bits = 16, .max_no_of_hqd = 24, .doorbell_size = 8, @@ -302,6 +352,7 @@ static const struct kfd_device_info vega12_device_info = { static const struct kfd_device_info vega20_device_info = { .asic_family = CHIP_VEGA20, + .asic_name = "vega20", .max_pasid_bits = 16, .max_no_of_hqd = 24, .doorbell_size = 8, @@ -317,8 +368,45 @@ static const struct kfd_device_info vega20_device_info = { .num_sdma_queues_per_engine = 8, }; +static const struct kfd_device_info arcturus_device_info = { + .asic_family = CHIP_ARCTURUS, + .asic_name = "arcturus", + .max_pasid_bits = 16, + .max_no_of_hqd = 24, + .doorbell_size = 8, + .ih_ring_entry_size = 8 * sizeof(uint32_t), + .event_interrupt_class = &event_interrupt_class_v9, + .num_of_watch_points = 4, + .mqd_size_aligned = MQD_SIZE_ALIGNED, + .supports_cwsr = true, + .needs_iommu_device = false, + .needs_pci_atomics = false, + .num_sdma_engines = 2, + .num_xgmi_sdma_engines = 6, + .num_sdma_queues_per_engine = 8, +}; + +static const struct kfd_device_info renoir_device_info = { + .asic_family = CHIP_RENOIR, + .asic_name = "renoir", + .max_pasid_bits = 16, + .max_no_of_hqd = 24, + .doorbell_size = 8, + .ih_ring_entry_size = 8 * sizeof(uint32_t), + .event_interrupt_class = &event_interrupt_class_v9, + .num_of_watch_points = 4, + .mqd_size_aligned = MQD_SIZE_ALIGNED, + .supports_cwsr = true, + .needs_iommu_device = false, + .needs_pci_atomics = false, + .num_sdma_engines = 1, + .num_xgmi_sdma_engines = 0, + .num_sdma_queues_per_engine = 2, +}; + static const struct kfd_device_info navi10_device_info = { .asic_family = CHIP_NAVI10, + .asic_name = "navi10", .max_pasid_bits = 16, .max_no_of_hqd = 24, .doorbell_size = 8, @@ -334,130 +422,64 @@ static const struct kfd_device_info navi10_device_info = { .num_sdma_queues_per_engine = 8, }; -struct kfd_deviceid { - unsigned short did; - const struct kfd_device_info *device_info; +static const struct kfd_device_info navi12_device_info = { + .asic_family = CHIP_NAVI12, + .asic_name = "navi12", + .max_pasid_bits = 16, + .max_no_of_hqd = 24, + .doorbell_size = 8, + .ih_ring_entry_size = 8 * sizeof(uint32_t), + .event_interrupt_class = &event_interrupt_class_v9, + .num_of_watch_points = 4, + .mqd_size_aligned = MQD_SIZE_ALIGNED, + .needs_iommu_device = false, + .supports_cwsr = true, + .needs_pci_atomics = false, + .num_sdma_engines = 2, + .num_xgmi_sdma_engines = 0, + .num_sdma_queues_per_engine = 8, +}; + +static const struct kfd_device_info navi14_device_info = { + .asic_family = CHIP_NAVI14, + .asic_name = "navi14", + .max_pasid_bits = 16, + .max_no_of_hqd = 24, + .doorbell_size = 8, + .ih_ring_entry_size = 8 * sizeof(uint32_t), + .event_interrupt_class = &event_interrupt_class_v9, + .num_of_watch_points = 4, + .mqd_size_aligned = MQD_SIZE_ALIGNED, + .needs_iommu_device = false, + .supports_cwsr = true, + .needs_pci_atomics = false, + .num_sdma_engines = 2, + .num_xgmi_sdma_engines = 0, + .num_sdma_queues_per_engine = 8, }; -static const struct kfd_deviceid supported_devices[] = { +/* For each entry, [0] is regular and [1] is virtualisation device. */ +static const struct kfd_device_info *kfd_supported_devices[][2] = { #ifdef KFD_SUPPORT_IOMMU_V2 - { 0x1304, &kaveri_device_info }, /* Kaveri */ - { 0x1305, &kaveri_device_info }, /* Kaveri */ - { 0x1306, &kaveri_device_info }, /* Kaveri */ - { 0x1307, &kaveri_device_info }, /* Kaveri */ - { 0x1309, &kaveri_device_info }, /* Kaveri */ - { 0x130A, &kaveri_device_info }, /* Kaveri */ - { 0x130B, &kaveri_device_info }, /* Kaveri */ - { 0x130C, &kaveri_device_info }, /* Kaveri */ - { 0x130D, &kaveri_device_info }, /* Kaveri */ - { 0x130E, &kaveri_device_info }, /* Kaveri */ - { 0x130F, &kaveri_device_info }, /* Kaveri */ - { 0x1310, &kaveri_device_info }, /* Kaveri */ - { 0x1311, &kaveri_device_info }, /* Kaveri */ - { 0x1312, &kaveri_device_info }, /* Kaveri */ - { 0x1313, &kaveri_device_info }, /* Kaveri */ - { 0x1315, &kaveri_device_info }, /* Kaveri */ - { 0x1316, &kaveri_device_info }, /* Kaveri */ - { 0x1317, &kaveri_device_info }, /* Kaveri */ - { 0x1318, &kaveri_device_info }, /* Kaveri */ - { 0x131B, &kaveri_device_info }, /* Kaveri */ - { 0x131C, &kaveri_device_info }, /* Kaveri */ - { 0x131D, &kaveri_device_info }, /* Kaveri */ - { 0x9870, &carrizo_device_info }, /* Carrizo */ - { 0x9874, &carrizo_device_info }, /* Carrizo */ - { 0x9875, &carrizo_device_info }, /* Carrizo */ - { 0x9876, &carrizo_device_info }, /* Carrizo */ - { 0x9877, &carrizo_device_info }, /* Carrizo */ - { 0x15DD, &raven_device_info }, /* Raven */ - { 0x15D8, &raven_device_info }, /* Raven */ + [CHIP_KAVERI] = {&kaveri_device_info, NULL}, + [CHIP_CARRIZO] = {&carrizo_device_info, NULL}, + [CHIP_RAVEN] = {&raven_device_info, NULL}, #endif - { 0x67A0, &hawaii_device_info }, /* Hawaii */ - { 0x67A1, &hawaii_device_info }, /* Hawaii */ - { 0x67A2, &hawaii_device_info }, /* Hawaii */ - { 0x67A8, &hawaii_device_info }, /* Hawaii */ - { 0x67A9, &hawaii_device_info }, /* Hawaii */ - { 0x67AA, &hawaii_device_info }, /* Hawaii */ - { 0x67B0, &hawaii_device_info }, /* Hawaii */ - { 0x67B1, &hawaii_device_info }, /* Hawaii */ - { 0x67B8, &hawaii_device_info }, /* Hawaii */ - { 0x67B9, &hawaii_device_info }, /* Hawaii */ - { 0x67BA, &hawaii_device_info }, /* Hawaii */ - { 0x67BE, &hawaii_device_info }, /* Hawaii */ - { 0x6920, &tonga_device_info }, /* Tonga */ - { 0x6921, &tonga_device_info }, /* Tonga */ - { 0x6928, &tonga_device_info }, /* Tonga */ - { 0x6929, &tonga_device_info }, /* Tonga */ - { 0x692B, &tonga_device_info }, /* Tonga */ - { 0x6938, &tonga_device_info }, /* Tonga */ - { 0x6939, &tonga_device_info }, /* Tonga */ - { 0x7300, &fiji_device_info }, /* Fiji */ - { 0x730F, &fiji_vf_device_info }, /* Fiji vf*/ - { 0x67C0, &polaris10_device_info }, /* Polaris10 */ - { 0x67C1, &polaris10_device_info }, /* Polaris10 */ - { 0x67C2, &polaris10_device_info }, /* Polaris10 */ - { 0x67C4, &polaris10_device_info }, /* Polaris10 */ - { 0x67C7, &polaris10_device_info }, /* Polaris10 */ - { 0x67C8, &polaris10_device_info }, /* Polaris10 */ - { 0x67C9, &polaris10_device_info }, /* Polaris10 */ - { 0x67CA, &polaris10_device_info }, /* Polaris10 */ - { 0x67CC, &polaris10_device_info }, /* Polaris10 */ - { 0x67CF, &polaris10_device_info }, /* Polaris10 */ - { 0x67D0, &polaris10_vf_device_info }, /* Polaris10 vf*/ - { 0x67DF, &polaris10_device_info }, /* Polaris10 */ - { 0x6FDF, &polaris10_device_info }, /* Polaris10 */ - { 0x67E0, &polaris11_device_info }, /* Polaris11 */ - { 0x67E1, &polaris11_device_info }, /* Polaris11 */ - { 0x67E3, &polaris11_device_info }, /* Polaris11 */ - { 0x67E7, &polaris11_device_info }, /* Polaris11 */ - { 0x67E8, &polaris11_device_info }, /* Polaris11 */ - { 0x67E9, &polaris11_device_info }, /* Polaris11 */ - { 0x67EB, &polaris11_device_info }, /* Polaris11 */ - { 0x67EF, &polaris11_device_info }, /* Polaris11 */ - { 0x67FF, &polaris11_device_info }, /* Polaris11 */ - { 0x6980, &polaris12_device_info }, /* Polaris12 */ - { 0x6981, &polaris12_device_info }, /* Polaris12 */ - { 0x6985, &polaris12_device_info }, /* Polaris12 */ - { 0x6986, &polaris12_device_info }, /* Polaris12 */ - { 0x6987, &polaris12_device_info }, /* Polaris12 */ - { 0x6995, &polaris12_device_info }, /* Polaris12 */ - { 0x6997, &polaris12_device_info }, /* Polaris12 */ - { 0x699F, &polaris12_device_info }, /* Polaris12 */ - { 0x694C, &vegam_device_info }, /* VegaM */ - { 0x694E, &vegam_device_info }, /* VegaM */ - { 0x694F, &vegam_device_info }, /* VegaM */ - { 0x6860, &vega10_device_info }, /* Vega10 */ - { 0x6861, &vega10_device_info }, /* Vega10 */ - { 0x6862, &vega10_device_info }, /* Vega10 */ - { 0x6863, &vega10_device_info }, /* Vega10 */ - { 0x6864, &vega10_device_info }, /* Vega10 */ - { 0x6867, &vega10_device_info }, /* Vega10 */ - { 0x6868, &vega10_device_info }, /* Vega10 */ - { 0x6869, &vega10_device_info }, /* Vega10 */ - { 0x686A, &vega10_device_info }, /* Vega10 */ - { 0x686B, &vega10_device_info }, /* Vega10 */ - { 0x686C, &vega10_vf_device_info }, /* Vega10 vf*/ - { 0x686D, &vega10_device_info }, /* Vega10 */ - { 0x686E, &vega10_device_info }, /* Vega10 */ - { 0x686F, &vega10_device_info }, /* Vega10 */ - { 0x687F, &vega10_device_info }, /* Vega10 */ - { 0x69A0, &vega12_device_info }, /* Vega12 */ - { 0x69A1, &vega12_device_info }, /* Vega12 */ - { 0x69A2, &vega12_device_info }, /* Vega12 */ - { 0x69A3, &vega12_device_info }, /* Vega12 */ - { 0x69AF, &vega12_device_info }, /* Vega12 */ - { 0x66a0, &vega20_device_info }, /* Vega20 */ - { 0x66a1, &vega20_device_info }, /* Vega20 */ - { 0x66a2, &vega20_device_info }, /* Vega20 */ - { 0x66a3, &vega20_device_info }, /* Vega20 */ - { 0x66a4, &vega20_device_info }, /* Vega20 */ - { 0x66a7, &vega20_device_info }, /* Vega20 */ - { 0x66af, &vega20_device_info }, /* Vega20 */ - /* Navi10 */ - { 0x7310, &navi10_device_info }, /* Navi10 */ - { 0x7312, &navi10_device_info }, /* Navi10 */ - { 0x7318, &navi10_device_info }, /* Navi10 */ - { 0x731a, &navi10_device_info }, /* Navi10 */ - { 0x731f, &navi10_device_info }, /* Navi10 */ + [CHIP_HAWAII] = {&hawaii_device_info, NULL}, + [CHIP_TONGA] = {&tonga_device_info, NULL}, + [CHIP_FIJI] = {&fiji_device_info, &fiji_vf_device_info}, + [CHIP_POLARIS10] = {&polaris10_device_info, &polaris10_vf_device_info}, + [CHIP_POLARIS11] = {&polaris11_device_info, NULL}, + [CHIP_POLARIS12] = {&polaris12_device_info, NULL}, + [CHIP_VEGAM] = {&vegam_device_info, NULL}, + [CHIP_VEGA10] = {&vega10_device_info, &vega10_vf_device_info}, + [CHIP_VEGA12] = {&vega12_device_info, NULL}, + [CHIP_VEGA20] = {&vega20_device_info, NULL}, + [CHIP_RENOIR] = {&renoir_device_info, NULL}, + [CHIP_ARCTURUS] = {&arcturus_device_info, &arcturus_device_info}, + [CHIP_NAVI10] = {&navi10_device_info, NULL}, + [CHIP_NAVI12] = {&navi12_device_info, &navi12_device_info}, + [CHIP_NAVI14] = {&navi14_device_info, NULL}, }; static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size, @@ -466,32 +488,25 @@ static void kfd_gtt_sa_fini(struct kfd_dev *kfd); static int kfd_resume(struct kfd_dev *kfd); -static const struct kfd_device_info *lookup_device_info(unsigned short did) +struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd, + struct pci_dev *pdev, unsigned int asic_type, bool vf) { - size_t i; + struct kfd_dev *kfd; + const struct kfd_device_info *device_info; + const struct kfd2kgd_calls *f2g; - for (i = 0; i < ARRAY_SIZE(supported_devices); i++) { - if (supported_devices[i].did == did) { - WARN_ON(!supported_devices[i].device_info); - return supported_devices[i].device_info; - } + if (asic_type >= sizeof(kfd_supported_devices) / (sizeof(void *) * 2) + || asic_type >= sizeof(kfd2kgd_funcs) / sizeof(void *)) { + dev_err(kfd_device, "asic_type %d out of range\n", asic_type); + return NULL; /* asic_type out of range */ } - dev_warn(kfd_device, "DID %04x is missing in supported_devices\n", - did); - - return NULL; -} + device_info = kfd_supported_devices[asic_type][vf]; + f2g = kfd2kgd_funcs[asic_type]; -struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd, - struct pci_dev *pdev, const struct kfd2kgd_calls *f2g) -{ - struct kfd_dev *kfd; - const struct kfd_device_info *device_info = - lookup_device_info(pdev->device); - - if (!device_info) { - dev_err(kfd_device, "kgd2kfd_probe failed\n"); + if (!device_info || !f2g) { + dev_err(kfd_device, "%s %s not supported in kfd\n", + amdgpu_asic_name[asic_type], vf ? "VF" : ""); return NULL; } @@ -536,6 +551,10 @@ static void kfd_cwsr_init(struct kfd_dev *kfd) BUILD_BUG_ON(sizeof(cwsr_trap_gfx8_hex) > PAGE_SIZE); kfd->cwsr_isa = cwsr_trap_gfx8_hex; kfd->cwsr_isa_size = sizeof(cwsr_trap_gfx8_hex); + } else if (kfd->device_info->asic_family == CHIP_ARCTURUS) { + BUILD_BUG_ON(sizeof(cwsr_trap_arcturus_hex) > PAGE_SIZE); + kfd->cwsr_isa = cwsr_trap_arcturus_hex; + kfd->cwsr_isa_size = sizeof(cwsr_trap_arcturus_hex); } else if (kfd->device_info->asic_family < CHIP_NAVI10) { BUILD_BUG_ON(sizeof(cwsr_trap_gfx9_hex) > PAGE_SIZE); kfd->cwsr_isa = cwsr_trap_gfx9_hex; @@ -551,10 +570,12 @@ static void kfd_cwsr_init(struct kfd_dev *kfd) } bool kgd2kfd_device_init(struct kfd_dev *kfd, + struct drm_device *ddev, const struct kgd2kfd_shared_resources *gpu_resources) { unsigned int size; + kfd->ddev = ddev; kfd->mec_fw_version = amdgpu_amdkfd_get_fw_version(kfd->kgd, KGD_ENGINE_MEC1); kfd->sdma_fw_version = amdgpu_amdkfd_get_fw_version(kfd->kgd, @@ -707,10 +728,10 @@ int kgd2kfd_pre_reset(struct kfd_dev *kfd) { if (!kfd->init_complete) return 0; - kgd2kfd_suspend(kfd); - /* hold dqm->lock to prevent further execution*/ - dqm_lock(kfd->dqm); + kfd->dqm->ops.pre_reset(kfd->dqm); + + kgd2kfd_suspend(kfd); kfd_signal_reset_event(kfd); return 0; @@ -724,17 +745,15 @@ int kgd2kfd_pre_reset(struct kfd_dev *kfd) int kgd2kfd_post_reset(struct kfd_dev *kfd) { - int ret, count; + int ret; if (!kfd->init_complete) return 0; - dqm_unlock(kfd->dqm); - ret = kfd_resume(kfd); if (ret) return ret; - count = atomic_dec_return(&kfd_locked); + atomic_dec(&kfd_locked); atomic_set(&kfd->sram_ecc_flag, 0); @@ -806,6 +825,21 @@ dqm_start_error: return err; } +static inline void kfd_queue_work(struct workqueue_struct *wq, + struct work_struct *work) +{ + int cpu, new_cpu; + + cpu = new_cpu = smp_processor_id(); + do { + new_cpu = cpumask_next(new_cpu, cpu_online_mask) % nr_cpu_ids; + if (cpu_to_node(new_cpu) == numa_node_id()) + break; + } while (cpu != new_cpu); + + queue_work_on(new_cpu, wq, work); +} + /* This is called directly from KGD at ISR. */ void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry) { @@ -828,7 +862,7 @@ void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry) patched_ihre, &is_patched) && enqueue_ih_ring_entry(kfd, is_patched ? patched_ihre : ih_ring_entry)) - queue_work(kfd->ih_wq, &kfd->interrupt_work); + kfd_queue_work(kfd->ih_wq, &kfd->interrupt_work); spin_unlock_irqrestore(&kfd->interrupt_lock, flags); } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c index e6a4288bfaa6..80d22bf702e8 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c @@ -170,7 +170,7 @@ static int allocate_doorbell(struct qcm_process_device *qpd, struct queue *q) } q->properties.doorbell_off = - kfd_doorbell_id_to_offset(dev, q->process, + kfd_get_doorbell_dw_offset_in_bar(dev, q->process, q->doorbell_id); return 0; @@ -195,20 +195,30 @@ static int allocate_vmid(struct device_queue_manager *dqm, struct qcm_process_device *qpd, struct queue *q) { - int bit, allocated_vmid; + int allocated_vmid = -1, i; - if (dqm->vmid_bitmap == 0) - return -ENOMEM; + for (i = dqm->dev->vm_info.first_vmid_kfd; + i <= dqm->dev->vm_info.last_vmid_kfd; i++) { + if (!dqm->vmid_pasid[i]) { + allocated_vmid = i; + break; + } + } - bit = ffs(dqm->vmid_bitmap) - 1; - dqm->vmid_bitmap &= ~(1 << bit); + if (allocated_vmid < 0) { + pr_err("no more vmid to allocate\n"); + return -ENOSPC; + } + + pr_debug("vmid allocated: %d\n", allocated_vmid); + + dqm->vmid_pasid[allocated_vmid] = q->process->pasid; + + set_pasid_vmid_mapping(dqm, q->process->pasid, allocated_vmid); - allocated_vmid = bit + dqm->dev->vm_info.first_vmid_kfd; - pr_debug("vmid allocation %d\n", allocated_vmid); qpd->vmid = allocated_vmid; q->properties.vmid = allocated_vmid; - set_pasid_vmid_mapping(dqm, q->process->pasid, q->properties.vmid); program_sh_mem_settings(dqm, qpd); /* qpd->page_table_base is set earlier when register_process() @@ -220,8 +230,9 @@ static int allocate_vmid(struct device_queue_manager *dqm, /* invalidate the VM context after pasid and vmid mapping is set up */ kfd_flush_tlb(qpd_to_pdd(qpd)); - dqm->dev->kfd2kgd->set_scratch_backing_va( - dqm->dev->kgd, qpd->sh_hidden_private_base, qpd->vmid); + if (dqm->dev->kfd2kgd->set_scratch_backing_va) + dqm->dev->kfd2kgd->set_scratch_backing_va(dqm->dev->kgd, + qpd->sh_hidden_private_base, qpd->vmid); return 0; } @@ -248,8 +259,6 @@ static void deallocate_vmid(struct device_queue_manager *dqm, struct qcm_process_device *qpd, struct queue *q) { - int bit = qpd->vmid - dqm->dev->vm_info.first_vmid_kfd; - /* On GFX v7, CP doesn't flush TC at dequeue */ if (q->device->device_info->asic_family == CHIP_HAWAII) if (flush_texture_cache_nocpsch(q->device, qpd)) @@ -259,8 +268,8 @@ static void deallocate_vmid(struct device_queue_manager *dqm, /* Release the vmid mapping */ set_pasid_vmid_mapping(dqm, 0, qpd->vmid); + dqm->vmid_pasid[qpd->vmid] = 0; - dqm->vmid_bitmap |= (1 << bit); qpd->vmid = 0; q->properties.vmid = 0; } @@ -331,6 +340,10 @@ static int create_queue_nocpsch(struct device_queue_manager *dqm, mqd_mgr->init_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj, &q->gart_mqd_addr, &q->properties); if (q->properties.is_active) { + if (!dqm->sched_running) { + WARN_ONCE(1, "Load non-HWS mqd while stopped\n"); + goto add_queue_to_list; + } if (WARN(q->process->mm != current->mm, "should only run in user thread")) @@ -342,6 +355,7 @@ static int create_queue_nocpsch(struct device_queue_manager *dqm, goto out_free_mqd; } +add_queue_to_list: list_add(&q->list, &qpd->queues_list); qpd->queue_count++; if (q->properties.is_active) @@ -449,6 +463,11 @@ static int destroy_queue_nocpsch_locked(struct device_queue_manager *dqm, deallocate_doorbell(qpd, q); + if (!dqm->sched_running) { + WARN_ONCE(1, "Destroy non-HWS queue while stopped\n"); + return 0; + } + retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd, KFD_PREEMPT_TYPE_WAVEFRONT_RESET, KFD_UNMAP_LATENCY_MS, @@ -524,6 +543,12 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q) (q->properties.type == KFD_QUEUE_TYPE_COMPUTE || q->properties.type == KFD_QUEUE_TYPE_SDMA || q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) { + + if (!dqm->sched_running) { + WARN_ONCE(1, "Update non-HWS queue while stopped\n"); + goto out_unlock; + } + retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd, KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN, KFD_UNMAP_LATENCY_MS, q->pipe, q->queue); @@ -579,7 +604,7 @@ static int evict_process_queues_nocpsch(struct device_queue_manager *dqm, goto out; pdd = qpd_to_pdd(qpd); - pr_info_ratelimited("Evicting PASID %u queues\n", + pr_info_ratelimited("Evicting PASID 0x%x queues\n", pdd->process->pasid); /* Mark all queues as evicted. Deactivate all active queues on @@ -593,6 +618,11 @@ static int evict_process_queues_nocpsch(struct device_queue_manager *dqm, mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( q->properties.type)]; q->properties.is_active = false; + dqm->queue_count--; + + if (WARN_ONCE(!dqm->sched_running, "Evict when stopped\n")) + continue; + retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd, KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN, KFD_UNMAP_LATENCY_MS, q->pipe, q->queue); @@ -601,7 +631,6 @@ static int evict_process_queues_nocpsch(struct device_queue_manager *dqm, * maintain a consistent eviction state */ ret = retval; - dqm->queue_count--; } out: @@ -621,7 +650,7 @@ static int evict_process_queues_cpsch(struct device_queue_manager *dqm, goto out; pdd = qpd_to_pdd(qpd); - pr_info_ratelimited("Evicting PASID %u queues\n", + pr_info_ratelimited("Evicting PASID 0x%x queues\n", pdd->process->pasid); /* Mark all queues as evicted. Deactivate all active queues on @@ -667,7 +696,7 @@ static int restore_process_queues_nocpsch(struct device_queue_manager *dqm, goto out; } - pr_info_ratelimited("Restoring PASID %u queues\n", + pr_info_ratelimited("Restoring PASID 0x%x queues\n", pdd->process->pasid); /* Update PD Base in QPD */ @@ -702,6 +731,11 @@ static int restore_process_queues_nocpsch(struct device_queue_manager *dqm, mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( q->properties.type)]; q->properties.is_active = true; + dqm->queue_count++; + + if (WARN_ONCE(!dqm->sched_running, "Restore when stopped\n")) + continue; + retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd, q->pipe, q->queue, &q->properties, mm); if (retval && !ret) @@ -709,7 +743,6 @@ static int restore_process_queues_nocpsch(struct device_queue_manager *dqm, * maintain a consistent eviction state */ ret = retval; - dqm->queue_count++; } qpd->evicted = 0; out: @@ -739,7 +772,7 @@ static int restore_process_queues_cpsch(struct device_queue_manager *dqm, goto out; } - pr_info_ratelimited("Restoring PASID %u queues\n", + pr_info_ratelimited("Restoring PASID 0x%x queues\n", pdd->process->pasid); /* Update PD Base in QPD */ @@ -879,9 +912,10 @@ static int initialize_nocpsch(struct device_queue_manager *dqm) dqm->allocated_queues[pipe] |= 1 << queue; } - dqm->vmid_bitmap = (1 << dqm->dev->vm_info.vmid_num_kfd) - 1; - dqm->sdma_bitmap = (1ULL << get_num_sdma_queues(dqm)) - 1; - dqm->xgmi_sdma_bitmap = (1ULL << get_num_xgmi_sdma_queues(dqm)) - 1; + memset(dqm->vmid_pasid, 0, sizeof(dqm->vmid_pasid)); + + dqm->sdma_bitmap = ~0ULL >> (64 - get_num_sdma_queues(dqm)); + dqm->xgmi_sdma_bitmap = ~0ULL >> (64 - get_num_xgmi_sdma_queues(dqm)); return 0; } @@ -896,21 +930,36 @@ static void uninitialize(struct device_queue_manager *dqm) for (i = 0 ; i < KFD_MQD_TYPE_MAX ; i++) kfree(dqm->mqd_mgrs[i]); mutex_destroy(&dqm->lock_hidden); - kfd_gtt_sa_free(dqm->dev, dqm->pipeline_mem); } static int start_nocpsch(struct device_queue_manager *dqm) { + pr_info("SW scheduler is used"); init_interrupts(dqm); - return pm_init(&dqm->packets, dqm); + + if (dqm->dev->device_info->asic_family == CHIP_HAWAII) + return pm_init(&dqm->packets, dqm); + dqm->sched_running = true; + + return 0; } static int stop_nocpsch(struct device_queue_manager *dqm) { - pm_uninit(&dqm->packets); + if (dqm->dev->device_info->asic_family == CHIP_HAWAII) + pm_uninit(&dqm->packets, false); + dqm->sched_running = false; + return 0; } +static void pre_reset(struct device_queue_manager *dqm) +{ + dqm_lock(dqm); + dqm->is_resetting = true; + dqm_unlock(dqm); +} + static int allocate_sdma_queue(struct device_queue_manager *dqm, struct queue *q) { @@ -1019,8 +1068,8 @@ static int initialize_cpsch(struct device_queue_manager *dqm) dqm->sdma_queue_count = 0; dqm->xgmi_sdma_queue_count = 0; dqm->active_runlist = false; - dqm->sdma_bitmap = (1ULL << get_num_sdma_queues(dqm)) - 1; - dqm->xgmi_sdma_bitmap = (1ULL << get_num_xgmi_sdma_queues(dqm)) - 1; + dqm->sdma_bitmap = ~0ULL >> (64 - get_num_sdma_queues(dqm)); + dqm->xgmi_sdma_bitmap = ~0ULL >> (64 - get_num_xgmi_sdma_queues(dqm)); INIT_WORK(&dqm->hw_exception_work, kfd_process_hw_exception); @@ -1058,25 +1107,32 @@ static int start_cpsch(struct device_queue_manager *dqm) dqm_lock(dqm); /* clear hang status when driver try to start the hw scheduler */ dqm->is_hws_hang = false; + dqm->is_resetting = false; + dqm->sched_running = true; execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); dqm_unlock(dqm); return 0; fail_allocate_vidmem: fail_set_sched_resources: - pm_uninit(&dqm->packets); + pm_uninit(&dqm->packets, false); fail_packet_manager_init: return retval; } static int stop_cpsch(struct device_queue_manager *dqm) { + bool hanging; + dqm_lock(dqm); - unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0); + if (!dqm->is_hws_hang) + unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0); + hanging = dqm->is_hws_hang || dqm->is_resetting; + dqm->sched_running = false; dqm_unlock(dqm); kfd_gtt_sa_free(dqm->dev, dqm->fence_mem); - pm_uninit(&dqm->packets); + pm_uninit(&dqm->packets, hanging); return 0; } @@ -1181,16 +1237,18 @@ static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q, list_add(&q->list, &qpd->queues_list); qpd->queue_count++; + + if (q->properties.type == KFD_QUEUE_TYPE_SDMA) + dqm->sdma_queue_count++; + else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) + dqm->xgmi_sdma_queue_count++; + if (q->properties.is_active) { dqm->queue_count++; retval = execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); } - if (q->properties.type == KFD_QUEUE_TYPE_SDMA) - dqm->sdma_queue_count++; - else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) - dqm->xgmi_sdma_queue_count++; /* * Unconditionally increment this counter, regardless of the queue's * type or whether the queue is active. @@ -1259,9 +1317,10 @@ static int map_queues_cpsch(struct device_queue_manager *dqm) { int retval; + if (!dqm->sched_running) + return 0; if (dqm->queue_count <= 0 || dqm->processes_count <= 0) return 0; - if (dqm->active_runlist) return 0; @@ -1283,6 +1342,8 @@ static int unmap_queues_cpsch(struct device_queue_manager *dqm, { int retval = 0; + if (!dqm->sched_running) + return 0; if (dqm->is_hws_hang) return -EIO; if (!dqm->active_runlist) @@ -1305,8 +1366,17 @@ static int unmap_queues_cpsch(struct device_queue_manager *dqm, /* should be timed out */ retval = amdkfd_fence_wait_timeout(dqm->fence_addr, KFD_FENCE_COMPLETED, queue_preemption_timeout_ms); - if (retval) + if (retval) { + pr_err("The cp might be in an unrecoverable state due to an unsuccessful queues preemption\n"); + dqm->is_hws_hang = true; + /* It's possible we're detecting a HWS hang in the + * middle of a GPU reset. No need to schedule another + * reset in this case. + */ + if (!dqm->is_resetting) + schedule_work(&dqm->hw_exception_work); return retval; + } pm_release_ib(&dqm->packets); dqm->active_runlist = false; @@ -1324,12 +1394,8 @@ static int execute_queues_cpsch(struct device_queue_manager *dqm, if (dqm->is_hws_hang) return -EIO; retval = unmap_queues_cpsch(dqm, filter, filter_param); - if (retval) { - pr_err("The cp might be in an unrecoverable state due to an unsuccessful queues preemption\n"); - dqm->is_hws_hang = true; - schedule_work(&dqm->hw_exception_work); + if (retval) return retval; - } return map_queues_cpsch(dqm); } @@ -1548,7 +1614,7 @@ static int get_wave_state(struct device_queue_manager *dqm, goto dqm_unlock; } - mqd_mgr = dqm->mqd_mgrs[KFD_MQD_TYPE_COMPUTE]; + mqd_mgr = dqm->mqd_mgrs[KFD_MQD_TYPE_CP]; if (!mqd_mgr->get_wave_state) { r = -EINVAL; @@ -1676,7 +1742,8 @@ static int allocate_hiq_sdma_mqd(struct device_queue_manager *dqm) struct kfd_dev *dev = dqm->dev; struct kfd_mem_obj *mem_obj = &dqm->hiq_sdma_mqd; uint32_t size = dqm->mqd_mgrs[KFD_MQD_TYPE_SDMA]->mqd_size * - dev->device_info->num_sdma_engines * + (dev->device_info->num_sdma_engines + + dev->device_info->num_xgmi_sdma_engines) * dev->device_info->num_sdma_queues_per_engine + dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ]->mqd_size; @@ -1722,6 +1789,7 @@ struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev) dqm->ops.initialize = initialize_cpsch; dqm->ops.start = start_cpsch; dqm->ops.stop = stop_cpsch; + dqm->ops.pre_reset = pre_reset; dqm->ops.destroy_queue = destroy_queue_cpsch; dqm->ops.update_queue = update_queue; dqm->ops.register_process = register_process; @@ -1740,6 +1808,7 @@ struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev) /* initialize dqm for no cp scheduling */ dqm->ops.start = start_nocpsch; dqm->ops.stop = stop_nocpsch; + dqm->ops.pre_reset = pre_reset; dqm->ops.create_queue = create_queue_nocpsch; dqm->ops.destroy_queue = destroy_queue_nocpsch; dqm->ops.update_queue = update_queue; @@ -1786,9 +1855,13 @@ struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev) case CHIP_VEGA12: case CHIP_VEGA20: case CHIP_RAVEN: + case CHIP_RENOIR: + case CHIP_ARCTURUS: device_queue_manager_init_v9(&dqm->asic_ops); break; case CHIP_NAVI10: + case CHIP_NAVI12: + case CHIP_NAVI14: device_queue_manager_init_v10_navi10(&dqm->asic_ops); break; default: @@ -1813,7 +1886,8 @@ out_free: return NULL; } -void deallocate_hiq_sdma_mqd(struct kfd_dev *dev, struct kfd_mem_obj *mqd) +static void deallocate_hiq_sdma_mqd(struct kfd_dev *dev, + struct kfd_mem_obj *mqd) { WARN(!mqd, "No hiq sdma mqd trunk to free"); @@ -1881,6 +1955,12 @@ int dqm_debugfs_hqds(struct seq_file *m, void *data) int pipe, queue; int r = 0; + if (!dqm->sched_running) { + seq_printf(m, " Device is stopped\n"); + + return 0; + } + r = dqm->dev->kfd2kgd->hqd_dump(dqm->dev->kgd, KFD_CIK_HIQ_PIPE, KFD_CIK_HIQ_QUEUE, &dump, &n_regs); @@ -1915,7 +1995,8 @@ int dqm_debugfs_hqds(struct seq_file *m, void *data) } } - for (pipe = 0; pipe < get_num_sdma_engines(dqm); pipe++) { + for (pipe = 0; pipe < get_num_sdma_engines(dqm) + + get_num_xgmi_sdma_engines(dqm); pipe++) { for (queue = 0; queue < dqm->dev->device_info->num_sdma_queues_per_engine; queue++) { diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h index 90db2c9275f6..871d3b628d2d 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h @@ -32,6 +32,8 @@ #include "kfd_mqd_manager.h" +#define VMID_NUM 16 + struct device_process_node { struct qcm_process_device *qpd; struct list_head list; @@ -102,6 +104,7 @@ struct device_queue_manager_ops { int (*initialize)(struct device_queue_manager *dqm); int (*start)(struct device_queue_manager *dqm); int (*stop)(struct device_queue_manager *dqm); + void (*pre_reset)(struct device_queue_manager *dqm); void (*uninitialize)(struct device_queue_manager *dqm); int (*create_kernel_queue)(struct device_queue_manager *dqm, struct kernel_queue *kq, @@ -185,9 +188,9 @@ struct device_queue_manager { unsigned int *allocated_queues; uint64_t sdma_bitmap; uint64_t xgmi_sdma_bitmap; - unsigned int vmid_bitmap; + /* the pasid mapping for each kfd vmid */ + uint16_t vmid_pasid[VMID_NUM]; uint64_t pipelines_addr; - struct kfd_mem_obj *pipeline_mem; uint64_t fence_gpu_addr; unsigned int *fence_addr; struct kfd_mem_obj *fence_mem; @@ -196,8 +199,10 @@ struct device_queue_manager { /* hw exception */ bool is_hws_hang; + bool is_resetting; struct work_struct hw_exception_work; struct kfd_mem_obj hiq_sdma_mqd; + bool sched_running; }; void device_queue_manager_init_cik( diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_doorbell.c b/drivers/gpu/drm/amd/amdkfd/kfd_doorbell.c index ebe79bf00145..8e0c00b9555e 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_doorbell.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_doorbell.c @@ -91,7 +91,7 @@ int kfd_doorbell_init(struct kfd_dev *kfd) kfd->doorbell_base = kfd->shared_resources.doorbell_physical_address + doorbell_start_offset; - kfd->doorbell_id_offset = doorbell_start_offset / sizeof(u32); + kfd->doorbell_base_dw_offset = doorbell_start_offset / sizeof(u32); kfd->doorbell_kernel_ptr = ioremap(kfd->doorbell_base, kfd_doorbell_process_slice(kfd)); @@ -103,8 +103,8 @@ int kfd_doorbell_init(struct kfd_dev *kfd) pr_debug("doorbell base == 0x%08lX\n", (uintptr_t)kfd->doorbell_base); - pr_debug("doorbell_id_offset == 0x%08lX\n", - kfd->doorbell_id_offset); + pr_debug("doorbell_base_dw_offset == 0x%08lX\n", + kfd->doorbell_base_dw_offset); pr_debug("doorbell_process_limit == 0x%08lX\n", doorbell_process_limit); @@ -185,7 +185,7 @@ void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd, * Calculating the kernel doorbell offset using the first * doorbell page. */ - *doorbell_off = kfd->doorbell_id_offset + inx; + *doorbell_off = kfd->doorbell_base_dw_offset + inx; pr_debug("Get kernel queue doorbell\n" " doorbell offset == 0x%08X\n" @@ -225,17 +225,17 @@ void write_kernel_doorbell64(void __iomem *db, u64 value) } } -unsigned int kfd_doorbell_id_to_offset(struct kfd_dev *kfd, +unsigned int kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev *kfd, struct kfd_process *process, unsigned int doorbell_id) { /* - * doorbell_id_offset accounts for doorbells taken by KGD. + * doorbell_base_dw_offset accounts for doorbells taken by KGD. * index * kfd_doorbell_process_slice/sizeof(u32) adjusts to * the process's doorbells. The offset returned is in dword * units regardless of the ASIC-dependent doorbell size. */ - return kfd->doorbell_id_offset + + return kfd->doorbell_base_dw_offset + process->doorbell_index * kfd_doorbell_process_slice(kfd) / sizeof(u32) + doorbell_id * kfd->device_info->doorbell_size / sizeof(u32); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_events.c b/drivers/gpu/drm/amd/amdkfd/kfd_events.c index d674d4b3340f..1f8365575b12 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_events.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_events.c @@ -346,7 +346,6 @@ int kfd_event_create(struct file *devkfd, struct kfd_process *p, ret = create_signal_event(devkfd, p, ev); if (!ret) { *event_page_offset = KFD_MMAP_TYPE_EVENTS; - *event_page_offset <<= PAGE_SHIFT; *event_slot_index = ev->event_id; } break; @@ -852,8 +851,8 @@ static void lookup_events_by_type_and_signal(struct kfd_process *p, if (type == KFD_EVENT_TYPE_MEMORY) { dev_warn(kfd_device, - "Sending SIGSEGV to HSA Process with PID %d ", - p->lead_thread->pid); + "Sending SIGSEGV to process %d (pasid 0x%x)", + p->lead_thread->pid, p->pasid); send_sig(SIGSEGV, p->lead_thread, 0); } @@ -861,13 +860,13 @@ static void lookup_events_by_type_and_signal(struct kfd_process *p, if (send_signal) { if (send_sigterm) { dev_warn(kfd_device, - "Sending SIGTERM to HSA Process with PID %d ", - p->lead_thread->pid); + "Sending SIGTERM to process %d (pasid 0x%x)", + p->lead_thread->pid, p->pasid); send_sig(SIGTERM, p->lead_thread, 0); } else { dev_err(kfd_device, - "HSA Process (PID %d) got unhandled exception", - p->lead_thread->pid); + "Process %d (pasid 0x%x) got unhandled exception", + p->lead_thread->pid, p->pasid); } } } @@ -936,7 +935,8 @@ void kfd_signal_iommu_event(struct kfd_dev *dev, unsigned int pasid, /* Workaround on Raven to not kill the process when memory is freed * before IOMMU is able to finish processing all the excessive PPRs */ - if (dev->device_info->asic_family != CHIP_RAVEN) { + if (dev->device_info->asic_family != CHIP_RAVEN && + dev->device_info->asic_family != CHIP_RENOIR) { mutex_lock(&p->event_mutex); /* Lookup events by type and signal them */ diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c b/drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c index 60521366dd31..bb77b8890e77 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c @@ -369,8 +369,13 @@ int kfd_init_apertures(struct kfd_process *process) /*Iterating over all devices*/ while (kfd_topology_enum_kfd_devices(id, &dev) == 0) { - if (!dev) { - id++; /* Skip non GPU devices */ + if (!dev || kfd_devcgroup_check_permission(dev)) { + /* Skip non GPU devices and devices to which the + * current process have no access to. Access can be + * limited by placing the process in a specific + * cgroup hierarchy + */ + id++; continue; } @@ -405,7 +410,11 @@ int kfd_init_apertures(struct kfd_process *process) case CHIP_VEGA12: case CHIP_VEGA20: case CHIP_RAVEN: + case CHIP_RENOIR: + case CHIP_ARCTURUS: case CHIP_NAVI10: + case CHIP_NAVI12: + case CHIP_NAVI14: kfd_init_apertures_v9(pdd, id); break; default: diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c index a85904ad0d5f..e05d75ecda21 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c @@ -54,8 +54,7 @@ static bool event_interrupt_isr_v9(struct kfd_dev *dev, memcpy(patched_ihre, ih_ring_entry, dev->device_info->ih_ring_entry_size); - pasid = dev->kfd2kgd->get_atc_vmid_pasid_mapping_pasid( - dev->kgd, vmid); + pasid = dev->dqm->vmid_pasid[vmid]; /* Patch the pasid field */ patched_ihre[3] = cpu_to_le32((le32_to_cpu(patched_ihre[3]) @@ -80,6 +79,7 @@ static bool event_interrupt_isr_v9(struct kfd_dev *dev, source_id == SOC15_INTSRC_SQ_INTERRUPT_MSG || source_id == SOC15_INTSRC_CP_BAD_OPCODE || client_id == SOC15_IH_CLIENTID_VMC || + client_id == SOC15_IH_CLIENTID_VMC1 || client_id == SOC15_IH_CLIENTID_UTCL2; } @@ -104,6 +104,7 @@ static void event_interrupt_wq_v9(struct kfd_dev *dev, else if (source_id == SOC15_INTSRC_CP_BAD_OPCODE) kfd_signal_hw_exception_event(pasid); else if (client_id == SOC15_IH_CLIENTID_VMC || + client_id == SOC15_IH_CLIENTID_VMC1 || client_id == SOC15_IH_CLIENTID_UTCL2) { struct kfd_vm_fault_info info = {0}; uint16_t ring_id = SOC15_RING_ID_FROM_IH_ENTRY(ih_ring_entry); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_interrupt.c b/drivers/gpu/drm/amd/amdkfd/kfd_interrupt.c index c56ac47cd318..bc47f6a44456 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_interrupt.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_interrupt.c @@ -62,6 +62,11 @@ int kfd_interrupt_init(struct kfd_dev *kfd) } kfd->ih_wq = alloc_workqueue("KFD IH", WQ_HIGHPRI, 1); + if (unlikely(!kfd->ih_wq)) { + kfifo_free(&kfd->ih_fifo); + dev_err(kfd_chardev(), "Failed to allocate KFD IH workqueue\n"); + return -ENOMEM; + } spin_lock_init(&kfd->interrupt_lock); INIT_WORK(&kfd->interrupt_work, interrupt_wq); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_iommu.c b/drivers/gpu/drm/amd/amdkfd/kfd_iommu.c index 5f35df23fb18..8d871514671e 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_iommu.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_iommu.c @@ -62,9 +62,6 @@ int kfd_iommu_device_init(struct kfd_dev *kfd) struct amd_iommu_device_info iommu_info; unsigned int pasid_limit; int err; - struct kfd_topology_device *top_dev; - - top_dev = kfd_topology_device_by_id(kfd->id); if (!kfd->device_info->needs_iommu_device) return 0; @@ -160,7 +157,7 @@ static void iommu_pasid_shutdown_callback(struct pci_dev *pdev, int pasid) if (!p) return; - pr_debug("Unbinding process %d from IOMMU\n", pasid); + pr_debug("Unbinding process 0x%x from IOMMU\n", pasid); mutex_lock(kfd_get_dbgmgr_mutex()); @@ -194,7 +191,7 @@ static int iommu_invalid_ppr_cb(struct pci_dev *pdev, int pasid, struct kfd_dev *dev; dev_warn_ratelimited(kfd_device, - "Invalid PPR device %x:%x.%x pasid %d address 0x%lX flags 0x%X", + "Invalid PPR device %x:%x.%x pasid 0x%x address 0x%lX flags 0x%X", PCI_BUS_NUM(pdev->devfn), PCI_SLOT(pdev->devfn), PCI_FUNC(pdev->devfn), @@ -235,7 +232,7 @@ static int kfd_bind_processes_to_device(struct kfd_dev *kfd) err = amd_iommu_bind_pasid(kfd->pdev, p->pasid, p->lead_thread); if (err < 0) { - pr_err("Unexpected pasid %d binding failure\n", + pr_err("Unexpected pasid 0x%x binding failure\n", p->pasid); mutex_unlock(&p->mutex); break; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c b/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c index 29c0bd2d7a5c..bae706462f96 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c @@ -34,7 +34,10 @@ #define PM4_COUNT_ZERO (((1 << 15) - 1) << 16) -static bool initialize(struct kernel_queue *kq, struct kfd_dev *dev, +/* Initialize a kernel queue, including allocations of GART memory + * needed for the queue. + */ +static bool kq_initialize(struct kernel_queue *kq, struct kfd_dev *dev, enum kfd_queue_type type, unsigned int queue_size) { struct queue_properties prop; @@ -87,9 +90,17 @@ static bool initialize(struct kernel_queue *kq, struct kfd_dev *dev, kq->pq_kernel_addr = kq->pq->cpu_ptr; kq->pq_gpu_addr = kq->pq->gpu_addr; - retval = kq->ops_asic_specific.initialize(kq, dev, type, queue_size); - if (!retval) - goto err_eop_allocate_vidmem; + /* For CIK family asics, kq->eop_mem is not needed */ + if (dev->device_info->asic_family > CHIP_MULLINS) { + retval = kfd_gtt_sa_allocate(dev, PAGE_SIZE, &kq->eop_mem); + if (retval != 0) + goto err_eop_allocate_vidmem; + + kq->eop_gpu_addr = kq->eop_mem->gpu_addr; + kq->eop_kernel_addr = kq->eop_mem->cpu_ptr; + + memset(kq->eop_kernel_addr, 0, PAGE_SIZE); + } retval = kfd_gtt_sa_allocate(dev, sizeof(*kq->rptr_kernel), &kq->rptr_mem); @@ -183,9 +194,10 @@ err_get_kernel_doorbell: } -static void uninitialize(struct kernel_queue *kq) +/* Uninitialize a kernel queue and free all its memory usages. */ +static void kq_uninitialize(struct kernel_queue *kq, bool hanging) { - if (kq->queue->properties.type == KFD_QUEUE_TYPE_HIQ) + if (kq->queue->properties.type == KFD_QUEUE_TYPE_HIQ && !hanging) kq->mqd_mgr->destroy_mqd(kq->mqd_mgr, kq->queue->mqd, KFD_PREEMPT_TYPE_WAVEFRONT_RESET, @@ -200,14 +212,19 @@ static void uninitialize(struct kernel_queue *kq) kfd_gtt_sa_free(kq->dev, kq->rptr_mem); kfd_gtt_sa_free(kq->dev, kq->wptr_mem); - kq->ops_asic_specific.uninitialize(kq); + + /* For CIK family asics, kq->eop_mem is Null, kfd_gtt_sa_free() + * is able to handle NULL properly. + */ + kfd_gtt_sa_free(kq->dev, kq->eop_mem); + kfd_gtt_sa_free(kq->dev, kq->pq); kfd_release_kernel_doorbell(kq->dev, kq->queue->properties.doorbell_ptr); uninit_queue(kq->queue); } -static int acquire_packet_buffer(struct kernel_queue *kq, +int kq_acquire_packet_buffer(struct kernel_queue *kq, size_t packet_size_in_dwords, unsigned int **buffer_ptr) { size_t available_size; @@ -268,7 +285,7 @@ err_no_space: return -ENOMEM; } -static void submit_packet(struct kernel_queue *kq) +void kq_submit_packet(struct kernel_queue *kq) { #ifdef DEBUG int i; @@ -280,11 +297,18 @@ static void submit_packet(struct kernel_queue *kq) } pr_debug("\n"); #endif - - kq->ops_asic_specific.submit_packet(kq); + if (kq->dev->device_info->doorbell_size == 8) { + *kq->wptr64_kernel = kq->pending_wptr64; + write_kernel_doorbell64(kq->queue->properties.doorbell_ptr, + kq->pending_wptr64); + } else { + *kq->wptr_kernel = kq->pending_wptr; + write_kernel_doorbell(kq->queue->properties.doorbell_ptr, + kq->pending_wptr); + } } -static void rollback_packet(struct kernel_queue *kq) +void kq_rollback_packet(struct kernel_queue *kq) { if (kq->dev->device_info->doorbell_size == 8) { kq->pending_wptr64 = *kq->wptr64_kernel; @@ -304,56 +328,18 @@ struct kernel_queue *kernel_queue_init(struct kfd_dev *dev, if (!kq) return NULL; - kq->ops.initialize = initialize; - kq->ops.uninitialize = uninitialize; - kq->ops.acquire_packet_buffer = acquire_packet_buffer; - kq->ops.submit_packet = submit_packet; - kq->ops.rollback_packet = rollback_packet; - - switch (dev->device_info->asic_family) { - case CHIP_CARRIZO: - case CHIP_TONGA: - case CHIP_FIJI: - case CHIP_POLARIS10: - case CHIP_POLARIS11: - case CHIP_POLARIS12: - case CHIP_VEGAM: - kernel_queue_init_vi(&kq->ops_asic_specific); - break; - - case CHIP_KAVERI: - case CHIP_HAWAII: - kernel_queue_init_cik(&kq->ops_asic_specific); - break; - - case CHIP_VEGA10: - case CHIP_VEGA12: - case CHIP_VEGA20: - case CHIP_RAVEN: - kernel_queue_init_v9(&kq->ops_asic_specific); - break; - case CHIP_NAVI10: - kernel_queue_init_v10(&kq->ops_asic_specific); - break; - default: - WARN(1, "Unexpected ASIC family %u", - dev->device_info->asic_family); - goto out_free; - } - - if (kq->ops.initialize(kq, dev, type, KFD_KERNEL_QUEUE_SIZE)) + if (kq_initialize(kq, dev, type, KFD_KERNEL_QUEUE_SIZE)) return kq; pr_err("Failed to init kernel queue\n"); -out_free: kfree(kq); return NULL; } -void kernel_queue_uninit(struct kernel_queue *kq) +void kernel_queue_uninit(struct kernel_queue *kq, bool hanging) { - kq->ops.uninitialize(kq); + kq_uninitialize(kq, hanging); kfree(kq); } @@ -373,7 +359,7 @@ static __attribute__((unused)) void test_kq(struct kfd_dev *dev) return; } - retval = kq->ops.acquire_packet_buffer(kq, 5, &buffer); + retval = kq_acquire_packet_buffer(kq, 5, &buffer); if (unlikely(retval != 0)) { pr_err(" Failed to acquire packet buffer\n"); pr_err("Kernel queue test failed\n"); @@ -381,7 +367,7 @@ static __attribute__((unused)) void test_kq(struct kfd_dev *dev) } for (i = 0; i < 5; i++) buffer[i] = kq->nop_packet; - kq->ops.submit_packet(kq); + kq_submit_packet(kq); pr_err("Ending kernel queue test\n"); } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.h b/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.h index 365fc674fea4..f4cfe9f1871c 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.h @@ -29,45 +29,28 @@ #include "kfd_priv.h" /** - * struct kernel_queue_ops - * - * @initialize: Initialize a kernel queue, including allocations of GART memory - * needed for the queue. - * - * @uninitialize: Uninitialize a kernel queue and free all its memory usages. - * - * @acquire_packet_buffer: Returns a pointer to the location in the kernel + * kq_acquire_packet_buffer: Returns a pointer to the location in the kernel * queue ring buffer where the calling function can write its packet. It is * Guaranteed that there is enough space for that packet. It also updates the * pending write pointer to that location so subsequent calls to * acquire_packet_buffer will get a correct write pointer * - * @submit_packet: Update the write pointer and doorbell of a kernel queue. - * - * @sync_with_hw: Wait until the write pointer and the read pointer of a kernel - * queue are equal, which means the CP has read all the submitted packets. + * kq_submit_packet: Update the write pointer and doorbell of a kernel queue. * - * @rollback_packet: This routine is called if we failed to build an acquired + * kq_rollback_packet: This routine is called if we failed to build an acquired * packet for some reason. It just overwrites the pending wptr with the current * one * */ -struct kernel_queue_ops { - bool (*initialize)(struct kernel_queue *kq, struct kfd_dev *dev, - enum kfd_queue_type type, unsigned int queue_size); - void (*uninitialize)(struct kernel_queue *kq); - int (*acquire_packet_buffer)(struct kernel_queue *kq, - size_t packet_size_in_dwords, - unsigned int **buffer_ptr); - void (*submit_packet)(struct kernel_queue *kq); - void (*rollback_packet)(struct kernel_queue *kq); -}; +int kq_acquire_packet_buffer(struct kernel_queue *kq, + size_t packet_size_in_dwords, + unsigned int **buffer_ptr); +void kq_submit_packet(struct kernel_queue *kq); +void kq_rollback_packet(struct kernel_queue *kq); -struct kernel_queue { - struct kernel_queue_ops ops; - struct kernel_queue_ops ops_asic_specific; +struct kernel_queue { /* data */ struct kfd_dev *dev; struct mqd_manager *mqd_mgr; @@ -99,9 +82,4 @@ struct kernel_queue { struct list_head list; }; -void kernel_queue_init_cik(struct kernel_queue_ops *ops); -void kernel_queue_init_vi(struct kernel_queue_ops *ops); -void kernel_queue_init_v9(struct kernel_queue_ops *ops); -void kernel_queue_init_v10(struct kernel_queue_ops *ops); - #endif /* KFD_KERNEL_QUEUE_H_ */ diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue_cik.c b/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue_cik.c deleted file mode 100644 index 19e54acb4125..000000000000 --- a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue_cik.c +++ /dev/null @@ -1,53 +0,0 @@ -/* - * Copyright 2014 Advanced Micro Devices, Inc. - * - * Permission is hereby granted, free of charge, to any person obtaining a - * copy of this software and associated documentation files (the "Software"), - * to deal in the Software without restriction, including without limitation - * the rights to use, copy, modify, merge, publish, distribute, sublicense, - * and/or sell copies of the Software, and to permit persons to whom the - * Software is furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in - * all copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL - * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR - * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, - * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR - * OTHER DEALINGS IN THE SOFTWARE. - * - */ - -#include "kfd_kernel_queue.h" - -static bool initialize_cik(struct kernel_queue *kq, struct kfd_dev *dev, - enum kfd_queue_type type, unsigned int queue_size); -static void uninitialize_cik(struct kernel_queue *kq); -static void submit_packet_cik(struct kernel_queue *kq); - -void kernel_queue_init_cik(struct kernel_queue_ops *ops) -{ - ops->initialize = initialize_cik; - ops->uninitialize = uninitialize_cik; - ops->submit_packet = submit_packet_cik; -} - -static bool initialize_cik(struct kernel_queue *kq, struct kfd_dev *dev, - enum kfd_queue_type type, unsigned int queue_size) -{ - return true; -} - -static void uninitialize_cik(struct kernel_queue *kq) -{ -} - -static void submit_packet_cik(struct kernel_queue *kq) -{ - *kq->wptr_kernel = kq->pending_wptr; - write_kernel_doorbell(kq->queue->properties.doorbell_ptr, - kq->pending_wptr); -} diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue_v10.c b/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue_v10.c deleted file mode 100644 index aed32ab7102e..000000000000 --- a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue_v10.c +++ /dev/null @@ -1,348 +0,0 @@ -/* - * Copyright 2018 Advanced Micro Devices, Inc. - * - * Permission is hereby granted, free of charge, to any person obtaining a - * copy of this software and associated documentation files (the "Software"), - * to deal in the Software without restriction, including without limitation - * the rights to use, copy, modify, merge, publish, distribute, sublicense, - * and/or sell copies of the Software, and to permit persons to whom the - * Software is furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in - * all copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL - * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR - * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, - * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR - * OTHER DEALINGS IN THE SOFTWARE. - * - */ - -#include "kfd_kernel_queue.h" -#include "kfd_device_queue_manager.h" -#include "kfd_pm4_headers_ai.h" -#include "kfd_pm4_opcodes.h" -#include "gc/gc_10_1_0_sh_mask.h" - -static bool initialize_v10(struct kernel_queue *kq, struct kfd_dev *dev, - enum kfd_queue_type type, unsigned int queue_size); -static void uninitialize_v10(struct kernel_queue *kq); -static void submit_packet_v10(struct kernel_queue *kq); - -void kernel_queue_init_v10(struct kernel_queue_ops *ops) -{ - ops->initialize = initialize_v10; - ops->uninitialize = uninitialize_v10; - ops->submit_packet = submit_packet_v10; -} - -static bool initialize_v10(struct kernel_queue *kq, struct kfd_dev *dev, - enum kfd_queue_type type, unsigned int queue_size) -{ - int retval; - - retval = kfd_gtt_sa_allocate(dev, PAGE_SIZE, &kq->eop_mem); - if (retval != 0) - return false; - - kq->eop_gpu_addr = kq->eop_mem->gpu_addr; - kq->eop_kernel_addr = kq->eop_mem->cpu_ptr; - - memset(kq->eop_kernel_addr, 0, PAGE_SIZE); - - return true; -} - -static void uninitialize_v10(struct kernel_queue *kq) -{ - kfd_gtt_sa_free(kq->dev, kq->eop_mem); -} - -static void submit_packet_v10(struct kernel_queue *kq) -{ - *kq->wptr64_kernel = kq->pending_wptr64; - write_kernel_doorbell64(kq->queue->properties.doorbell_ptr, - kq->pending_wptr64); -} - -static int pm_map_process_v10(struct packet_manager *pm, - uint32_t *buffer, struct qcm_process_device *qpd) -{ - struct pm4_mes_map_process *packet; - uint64_t vm_page_table_base_addr = qpd->page_table_base; - - packet = (struct pm4_mes_map_process *)buffer; - memset(buffer, 0, sizeof(struct pm4_mes_map_process)); - - packet->header.u32All = pm_build_pm4_header(IT_MAP_PROCESS, - sizeof(struct pm4_mes_map_process)); - packet->bitfields2.diq_enable = (qpd->is_debug) ? 1 : 0; - packet->bitfields2.process_quantum = 1; - packet->bitfields2.pasid = qpd->pqm->process->pasid; - packet->bitfields14.gds_size = qpd->gds_size; - packet->bitfields14.num_gws = qpd->num_gws; - packet->bitfields14.num_oac = qpd->num_oac; - packet->bitfields14.sdma_enable = 1; - - packet->bitfields14.num_queues = (qpd->is_debug) ? 0 : qpd->queue_count; - - packet->sh_mem_config = qpd->sh_mem_config; - packet->sh_mem_bases = qpd->sh_mem_bases; - if (qpd->tba_addr) { - packet->sq_shader_tba_lo = lower_32_bits(qpd->tba_addr >> 8); - packet->sq_shader_tba_hi = (1 << SQ_SHADER_TBA_HI__TRAP_EN__SHIFT) | - upper_32_bits(qpd->tba_addr >> 8); - packet->sq_shader_tma_lo = lower_32_bits(qpd->tma_addr >> 8); - packet->sq_shader_tma_hi = upper_32_bits(qpd->tma_addr >> 8); - } - - packet->gds_addr_lo = lower_32_bits(qpd->gds_context_area); - packet->gds_addr_hi = upper_32_bits(qpd->gds_context_area); - - packet->vm_context_page_table_base_addr_lo32 = - lower_32_bits(vm_page_table_base_addr); - packet->vm_context_page_table_base_addr_hi32 = - upper_32_bits(vm_page_table_base_addr); - - return 0; -} - -static int pm_runlist_v10(struct packet_manager *pm, uint32_t *buffer, - uint64_t ib, size_t ib_size_in_dwords, bool chain) -{ - struct pm4_mes_runlist *packet; - - int concurrent_proc_cnt = 0; - struct kfd_dev *kfd = pm->dqm->dev; - - /* Determine the number of processes to map together to HW: - * it can not exceed the number of VMIDs available to the - * scheduler, and it is determined by the smaller of the number - * of processes in the runlist and kfd module parameter - * hws_max_conc_proc. - * Note: the arbitration between the number of VMIDs and - * hws_max_conc_proc has been done in - * kgd2kfd_device_init(). - */ - concurrent_proc_cnt = min(pm->dqm->processes_count, - kfd->max_proc_per_quantum); - - - packet = (struct pm4_mes_runlist *)buffer; - - memset(buffer, 0, sizeof(struct pm4_mes_runlist)); - packet->header.u32All = pm_build_pm4_header(IT_RUN_LIST, - sizeof(struct pm4_mes_runlist)); - - packet->bitfields4.ib_size = ib_size_in_dwords; - packet->bitfields4.chain = chain ? 1 : 0; - packet->bitfields4.offload_polling = 0; - packet->bitfields4.valid = 1; - packet->bitfields4.process_cnt = concurrent_proc_cnt; - packet->ordinal2 = lower_32_bits(ib); - packet->ib_base_hi = upper_32_bits(ib); - - return 0; -} - -static int pm_map_queues_v10(struct packet_manager *pm, uint32_t *buffer, - struct queue *q, bool is_static) -{ - struct pm4_mes_map_queues *packet; - bool use_static = is_static; - - packet = (struct pm4_mes_map_queues *)buffer; - memset(buffer, 0, sizeof(struct pm4_mes_map_queues)); - - packet->header.u32All = pm_build_pm4_header(IT_MAP_QUEUES, - sizeof(struct pm4_mes_map_queues)); - packet->bitfields2.num_queues = 1; - packet->bitfields2.queue_sel = - queue_sel__mes_map_queues__map_to_hws_determined_queue_slots_vi; - - packet->bitfields2.engine_sel = - engine_sel__mes_map_queues__compute_vi; - packet->bitfields2.queue_type = - queue_type__mes_map_queues__normal_compute_vi; - - switch (q->properties.type) { - case KFD_QUEUE_TYPE_COMPUTE: - if (use_static) - packet->bitfields2.queue_type = - queue_type__mes_map_queues__normal_latency_static_queue_vi; - break; - case KFD_QUEUE_TYPE_DIQ: - packet->bitfields2.queue_type = - queue_type__mes_map_queues__debug_interface_queue_vi; - break; - case KFD_QUEUE_TYPE_SDMA: - case KFD_QUEUE_TYPE_SDMA_XGMI: - packet->bitfields2.engine_sel = q->properties.sdma_engine_id + - engine_sel__mes_map_queues__sdma0_vi; - use_static = false; /* no static queues under SDMA */ - break; - default: - WARN(1, "queue type %d\n", q->properties.type); - return -EINVAL; - } - packet->bitfields3.doorbell_offset = - q->properties.doorbell_off; - - packet->mqd_addr_lo = - lower_32_bits(q->gart_mqd_addr); - - packet->mqd_addr_hi = - upper_32_bits(q->gart_mqd_addr); - - packet->wptr_addr_lo = - lower_32_bits((uint64_t)q->properties.write_ptr); - - packet->wptr_addr_hi = - upper_32_bits((uint64_t)q->properties.write_ptr); - - return 0; -} - -static int pm_unmap_queues_v10(struct packet_manager *pm, uint32_t *buffer, - enum kfd_queue_type type, - enum kfd_unmap_queues_filter filter, - uint32_t filter_param, bool reset, - unsigned int sdma_engine) -{ - struct pm4_mes_unmap_queues *packet; - - packet = (struct pm4_mes_unmap_queues *)buffer; - memset(buffer, 0, sizeof(struct pm4_mes_unmap_queues)); - - packet->header.u32All = pm_build_pm4_header(IT_UNMAP_QUEUES, - sizeof(struct pm4_mes_unmap_queues)); - switch (type) { - case KFD_QUEUE_TYPE_COMPUTE: - case KFD_QUEUE_TYPE_DIQ: - packet->bitfields2.engine_sel = - engine_sel__mes_unmap_queues__compute; - break; - case KFD_QUEUE_TYPE_SDMA: - case KFD_QUEUE_TYPE_SDMA_XGMI: - packet->bitfields2.engine_sel = - engine_sel__mes_unmap_queues__sdma0 + sdma_engine; - break; - default: - WARN(1, "queue type %d\n", type); - break; - } - - if (reset) - packet->bitfields2.action = - action__mes_unmap_queues__reset_queues; - else - packet->bitfields2.action = - action__mes_unmap_queues__preempt_queues; - - switch (filter) { - case KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE: - packet->bitfields2.queue_sel = - queue_sel__mes_unmap_queues__perform_request_on_specified_queues; - packet->bitfields2.num_queues = 1; - packet->bitfields3b.doorbell_offset0 = filter_param; - break; - case KFD_UNMAP_QUEUES_FILTER_BY_PASID: - packet->bitfields2.queue_sel = - queue_sel__mes_unmap_queues__perform_request_on_pasid_queues; - packet->bitfields3a.pasid = filter_param; - break; - case KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: - packet->bitfields2.queue_sel = - queue_sel__mes_unmap_queues__unmap_all_queues; - break; - case KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES: - /* in this case, we do not preempt static queues */ - packet->bitfields2.queue_sel = - queue_sel__mes_unmap_queues__unmap_all_non_static_queues; - break; - default: - WARN(1, "filter %d\n", filter); - break; - } - - return 0; - -} - -static int pm_query_status_v10(struct packet_manager *pm, uint32_t *buffer, - uint64_t fence_address, uint32_t fence_value) -{ - struct pm4_mes_query_status *packet; - - packet = (struct pm4_mes_query_status *)buffer; - memset(buffer, 0, sizeof(struct pm4_mes_query_status)); - - - packet->header.u32All = pm_build_pm4_header(IT_QUERY_STATUS, - sizeof(struct pm4_mes_query_status)); - - packet->bitfields2.context_id = 0; - packet->bitfields2.interrupt_sel = - interrupt_sel__mes_query_status__completion_status; - packet->bitfields2.command = - command__mes_query_status__fence_only_after_write_ack; - - packet->addr_hi = upper_32_bits((uint64_t)fence_address); - packet->addr_lo = lower_32_bits((uint64_t)fence_address); - packet->data_hi = upper_32_bits((uint64_t)fence_value); - packet->data_lo = lower_32_bits((uint64_t)fence_value); - - return 0; -} - - -static int pm_release_mem_v10(uint64_t gpu_addr, uint32_t *buffer) -{ - struct pm4_mec_release_mem *packet; - - WARN_ON(!buffer); - - packet = (struct pm4_mec_release_mem *)buffer; - memset(buffer, 0, sizeof(struct pm4_mec_release_mem)); - - packet->header.u32All = pm_build_pm4_header(IT_RELEASE_MEM, - sizeof(struct pm4_mec_release_mem)); - - packet->bitfields2.event_type = CACHE_FLUSH_AND_INV_TS_EVENT; - packet->bitfields2.event_index = event_index__mec_release_mem__end_of_pipe; - packet->bitfields2.tcl1_action_ena = 1; - packet->bitfields2.tc_action_ena = 1; - packet->bitfields2.cache_policy = cache_policy__mec_release_mem__lru; - - packet->bitfields3.data_sel = data_sel__mec_release_mem__send_32_bit_low; - packet->bitfields3.int_sel = - int_sel__mec_release_mem__send_interrupt_after_write_confirm; - - packet->bitfields4.address_lo_32b = (gpu_addr & 0xffffffff) >> 2; - packet->address_hi = upper_32_bits(gpu_addr); - - packet->data_lo = 0; - - return sizeof(struct pm4_mec_release_mem) / sizeof(unsigned int); -} - -const struct packet_manager_funcs kfd_v10_pm_funcs = { - .map_process = pm_map_process_v10, - .runlist = pm_runlist_v10, - .set_resources = pm_set_resources_vi, - .map_queues = pm_map_queues_v10, - .unmap_queues = pm_unmap_queues_v10, - .query_status = pm_query_status_v10, - .release_mem = pm_release_mem_v10, - .map_process_size = sizeof(struct pm4_mes_map_process), - .runlist_size = sizeof(struct pm4_mes_runlist), - .set_resources_size = sizeof(struct pm4_mes_set_resources), - .map_queues_size = sizeof(struct pm4_mes_map_queues), - .unmap_queues_size = sizeof(struct pm4_mes_unmap_queues), - .query_status_size = sizeof(struct pm4_mes_query_status), - .release_mem_size = sizeof(struct pm4_mec_release_mem) -}; - diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_module.c b/drivers/gpu/drm/amd/amdkfd/kfd_module.c index 986ff52d5750..f4b7f7e6c40e 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_module.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_module.c @@ -82,7 +82,7 @@ static void kfd_exit(void) kfd_chardev_exit(); } -int kgd2kfd_init() +int kgd2kfd_init(void) { return kfd_init(); } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c index d6cf391da591..88813dad731f 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c @@ -98,8 +98,8 @@ void mqd_symmetrically_map_cu_mask(struct mqd_manager *mm, uint32_t *se_mask) { struct kfd_cu_info cu_info; - uint32_t cu_per_sh[4] = {0}; - int i, se, cu = 0; + uint32_t cu_per_se[KFD_MAX_NUM_SE] = {0}; + int i, se, sh, cu = 0; amdgpu_amdkfd_get_cu_info(mm->dev->kgd, &cu_info); @@ -107,8 +107,8 @@ void mqd_symmetrically_map_cu_mask(struct mqd_manager *mm, cu_mask_count = cu_info.cu_active_number; for (se = 0; se < cu_info.num_shader_engines; se++) - for (i = 0; i < 4; i++) - cu_per_sh[se] += hweight32(cu_info.cu_bitmap[se][i]); + for (sh = 0; sh < cu_info.num_shader_arrays_per_engine; sh++) + cu_per_se[se] += hweight32(cu_info.cu_bitmap[se % 4][sh + (se / 4)]); /* Symmetrically map cu_mask to all SEs: * cu_mask[0] bit0 -> se_mask[0] bit0; @@ -128,6 +128,6 @@ void mqd_symmetrically_map_cu_mask(struct mqd_manager *mm, se = 0; cu++; } - } while (cu >= cu_per_sh[se] && cu < 32); + } while (cu >= cu_per_se[se] && cu < 32); } } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h index 550b61e81015..fbdb16418847 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h @@ -26,6 +26,8 @@ #include "kfd_priv.h" +#define KFD_MAX_NUM_SE 8 + /** * struct mqd_manager * diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c index 28876aceb14b..19f0fe547c57 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c @@ -374,7 +374,6 @@ struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type, switch (type) { case KFD_MQD_TYPE_CP: - case KFD_MQD_TYPE_COMPUTE: mqd->allocate_mqd = allocate_mqd; mqd->init_mqd = init_mqd; mqd->free_mqd = free_mqd; @@ -401,7 +400,7 @@ struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type, #endif break; case KFD_MQD_TYPE_DIQ: - mqd->allocate_mqd = allocate_hiq_mqd; + mqd->allocate_mqd = allocate_mqd; mqd->init_mqd = init_mqd_hiq; mqd->free_mqd = free_mqd; mqd->load_mqd = load_mqd; @@ -442,7 +441,7 @@ struct mqd_manager *mqd_manager_init_cik_hawaii(enum KFD_MQD_TYPE type, mqd = mqd_manager_init_cik(type, dev); if (!mqd) return NULL; - if ((type == KFD_MQD_TYPE_CP) || (type == KFD_MQD_TYPE_COMPUTE)) + if (type == KFD_MQD_TYPE_CP) mqd->update_mqd = update_mqd_hawaii; return mqd; } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c index 9cd3eb2d90bd..d1d68a51bfb8 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c @@ -66,38 +66,22 @@ static void update_cu_mask(struct mqd_manager *mm, void *mqd, m->compute_static_thread_mgmt_se3); } +static void set_priority(struct v10_compute_mqd *m, struct queue_properties *q) +{ + m->cp_hqd_pipe_priority = pipe_priority_map[q->priority]; + m->cp_hqd_queue_priority = q->priority; +} + static struct kfd_mem_obj *allocate_mqd(struct kfd_dev *kfd, struct queue_properties *q) { - int retval; - struct kfd_mem_obj *mqd_mem_obj = NULL; - - /* From V9, for CWSR, the control stack is located on the next page - * boundary after the mqd, we will use the gtt allocation function - * instead of sub-allocation function. - */ - if (kfd->cwsr_enabled && (q->type == KFD_QUEUE_TYPE_COMPUTE)) { - mqd_mem_obj = kzalloc(sizeof(struct kfd_mem_obj), GFP_NOIO); - if (!mqd_mem_obj) - return NULL; - retval = amdgpu_amdkfd_alloc_gtt_mem(kfd->kgd, - ALIGN(q->ctl_stack_size, PAGE_SIZE) + - ALIGN(sizeof(struct v10_compute_mqd), PAGE_SIZE), - &(mqd_mem_obj->gtt_mem), - &(mqd_mem_obj->gpu_addr), - (void *)&(mqd_mem_obj->cpu_ptr), true); - } else { - retval = kfd_gtt_sa_allocate(kfd, sizeof(struct v10_compute_mqd), - &mqd_mem_obj); - } + struct kfd_mem_obj *mqd_mem_obj; - if (retval) { - kfree(mqd_mem_obj); + if (kfd_gtt_sa_allocate(kfd, sizeof(struct v10_compute_mqd), + &mqd_mem_obj)) return NULL; - } return mqd_mem_obj; - } static void init_mqd(struct mqd_manager *mm, void **mqd, @@ -131,9 +115,6 @@ static void init_mqd(struct mqd_manager *mm, void **mqd, 1 << CP_HQD_QUANTUM__QUANTUM_SCALE__SHIFT | 10 << CP_HQD_QUANTUM__QUANTUM_DURATION__SHIFT; - m->cp_hqd_pipe_priority = 1; - m->cp_hqd_queue_priority = 15; - if (q->format == KFD_QUEUE_FORMAT_AQL) { m->cp_hqd_aql_control = 1 << CP_HQD_AQL_CONTROL__CONTROL0__SHIFT; @@ -172,6 +153,14 @@ static int load_mqd(struct mqd_manager *mm, void *mqd, return r; } +static int hiq_load_mqd_kiq(struct mqd_manager *mm, void *mqd, + uint32_t pipe_id, uint32_t queue_id, + struct queue_properties *p, struct mm_struct *mms) +{ + return mm->dev->kfd2kgd->hiq_mqd_load(mm->dev->kgd, mqd, pipe_id, + queue_id, p->doorbell_off); +} + static void update_mqd(struct mqd_manager *mm, void *mqd, struct queue_properties *q) { @@ -230,11 +219,9 @@ static void update_mqd(struct mqd_manager *mm, void *mqd, m->cp_hqd_ctx_save_control = 0; update_cu_mask(mm, mqd, q); + set_priority(m, q); - q->is_active = (q->queue_size > 0 && - q->queue_address != 0 && - q->queue_percent > 0 && - !q->is_evicted); + q->is_active = QUEUE_IS_ACTIVE(*q); } static int destroy_mqd(struct mqd_manager *mm, void *mqd, @@ -250,14 +237,7 @@ static int destroy_mqd(struct mqd_manager *mm, void *mqd, static void free_mqd(struct mqd_manager *mm, void *mqd, struct kfd_mem_obj *mqd_mem_obj) { - struct kfd_dev *kfd = mm->dev; - - if (mqd_mem_obj->gtt_mem) { - amdgpu_amdkfd_free_gtt_mem(kfd->kgd, mqd_mem_obj->gtt_mem); - kfree(mqd_mem_obj); - } else { - kfd_gtt_sa_free(mm->dev, mqd_mem_obj); - } + kfd_gtt_sa_free(mm->dev, mqd_mem_obj); } static bool is_occupied(struct mqd_manager *mm, void *mqd, @@ -276,18 +256,22 @@ static int get_wave_state(struct mqd_manager *mm, void *mqd, { struct v10_compute_mqd *m; - /* Control stack is located one page after MQD. */ - void *mqd_ctl_stack = (void *)((uintptr_t)mqd + PAGE_SIZE); - m = get_mqd(mqd); + /* Control stack is written backwards, while workgroup context data + * is written forwards. Both starts from m->cp_hqd_cntl_stack_size. + * Current position is at m->cp_hqd_cntl_stack_offset and + * m->cp_hqd_wg_state_offset, respectively. + */ *ctl_stack_used_size = m->cp_hqd_cntl_stack_size - m->cp_hqd_cntl_stack_offset; *save_area_used_size = m->cp_hqd_wg_state_offset - m->cp_hqd_cntl_stack_size; - if (copy_to_user(ctl_stack, mqd_ctl_stack, m->cp_hqd_cntl_stack_size)) - return -EFAULT; + /* Control stack is not copied to user mode for GFXv10 because + * it's part of the context save area that is already + * accessible to user mode + */ return 0; } @@ -306,18 +290,6 @@ static void init_mqd_hiq(struct mqd_manager *mm, void **mqd, 1 << CP_HQD_PQ_CONTROL__KMD_QUEUE__SHIFT; } -static void update_mqd_hiq(struct mqd_manager *mm, void *mqd, - struct queue_properties *q) -{ - struct v10_compute_mqd *m; - - update_mqd(mm, mqd, q); - - /* TODO: what's the point? update_mqd already does this. */ - m = get_mqd(mqd); - m->cp_hqd_vmid = q->vmid; -} - static void init_mqd_sdma(struct mqd_manager *mm, void **mqd, struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr, struct queue_properties *q) @@ -369,11 +341,7 @@ static void update_mqd_sdma(struct mqd_manager *mm, void *mqd, m->sdma_queue_id = q->sdma_queue_id; m->sdmax_rlcx_dummy_reg = SDMA_RLC_DUMMY_DEFAULT; - - q->is_active = (q->queue_size > 0 && - q->queue_address != 0 && - q->queue_percent > 0 && - !q->is_evicted); + q->is_active = QUEUE_IS_ACTIVE(*q); } /* @@ -421,7 +389,7 @@ struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type, if (WARN_ON(type >= KFD_MQD_TYPE_MAX)) return NULL; - mqd = kzalloc(sizeof(*mqd), GFP_NOIO); + mqd = kzalloc(sizeof(*mqd), GFP_KERNEL); if (!mqd) return NULL; @@ -429,7 +397,6 @@ struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type, switch (type) { case KFD_MQD_TYPE_CP: - case KFD_MQD_TYPE_COMPUTE: pr_debug("%s@%i\n", __func__, __LINE__); mqd->allocate_mqd = allocate_mqd; mqd->init_mqd = init_mqd; @@ -450,8 +417,8 @@ struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type, mqd->allocate_mqd = allocate_hiq_mqd; mqd->init_mqd = init_mqd_hiq; mqd->free_mqd = free_mqd_hiq_sdma; - mqd->load_mqd = load_mqd; - mqd->update_mqd = update_mqd_hiq; + mqd->load_mqd = hiq_load_mqd_kiq; + mqd->update_mqd = update_mqd; mqd->destroy_mqd = destroy_mqd; mqd->is_occupied = is_occupied; mqd->mqd_size = sizeof(struct v10_compute_mqd); @@ -461,11 +428,11 @@ struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type, pr_debug("%s@%i\n", __func__, __LINE__); break; case KFD_MQD_TYPE_DIQ: - mqd->allocate_mqd = allocate_hiq_mqd; + mqd->allocate_mqd = allocate_mqd; mqd->init_mqd = init_mqd_hiq; mqd->free_mqd = free_mqd; mqd->load_mqd = load_mqd; - mqd->update_mqd = update_mqd_hiq; + mqd->update_mqd = update_mqd; mqd->destroy_mqd = destroy_mqd; mqd->is_occupied = is_occupied; mqd->mqd_size = sizeof(struct v10_compute_mqd); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c index 0c58f91b3ff3..436b7f518979 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c @@ -46,7 +46,7 @@ static void update_cu_mask(struct mqd_manager *mm, void *mqd, struct queue_properties *q) { struct v9_mqd *m; - uint32_t se_mask[4] = {0}; /* 4 is the max # of SEs */ + uint32_t se_mask[KFD_MAX_NUM_SE] = {0}; if (q->cu_mask_count == 0) return; @@ -59,12 +59,20 @@ static void update_cu_mask(struct mqd_manager *mm, void *mqd, m->compute_static_thread_mgmt_se1 = se_mask[1]; m->compute_static_thread_mgmt_se2 = se_mask[2]; m->compute_static_thread_mgmt_se3 = se_mask[3]; + m->compute_static_thread_mgmt_se4 = se_mask[4]; + m->compute_static_thread_mgmt_se5 = se_mask[5]; + m->compute_static_thread_mgmt_se6 = se_mask[6]; + m->compute_static_thread_mgmt_se7 = se_mask[7]; - pr_debug("update cu mask to %#x %#x %#x %#x\n", + pr_debug("update cu mask to %#x %#x %#x %#x %#x %#x %#x %#x\n", m->compute_static_thread_mgmt_se0, m->compute_static_thread_mgmt_se1, m->compute_static_thread_mgmt_se2, - m->compute_static_thread_mgmt_se3); + m->compute_static_thread_mgmt_se3, + m->compute_static_thread_mgmt_se4, + m->compute_static_thread_mgmt_se5, + m->compute_static_thread_mgmt_se6, + m->compute_static_thread_mgmt_se7); } static void set_priority(struct v9_mqd *m, struct queue_properties *q) @@ -84,7 +92,7 @@ static struct kfd_mem_obj *allocate_mqd(struct kfd_dev *kfd, * instead of sub-allocation function. */ if (kfd->cwsr_enabled && (q->type == KFD_QUEUE_TYPE_COMPUTE)) { - mqd_mem_obj = kzalloc(sizeof(struct kfd_mem_obj), GFP_NOIO); + mqd_mem_obj = kzalloc(sizeof(struct kfd_mem_obj), GFP_KERNEL); if (!mqd_mem_obj) return NULL; retval = amdgpu_amdkfd_alloc_gtt_mem(kfd->kgd, @@ -125,6 +133,10 @@ static void init_mqd(struct mqd_manager *mm, void **mqd, m->compute_static_thread_mgmt_se1 = 0xFFFFFFFF; m->compute_static_thread_mgmt_se2 = 0xFFFFFFFF; m->compute_static_thread_mgmt_se3 = 0xFFFFFFFF; + m->compute_static_thread_mgmt_se4 = 0xFFFFFFFF; + m->compute_static_thread_mgmt_se5 = 0xFFFFFFFF; + m->compute_static_thread_mgmt_se6 = 0xFFFFFFFF; + m->compute_static_thread_mgmt_se7 = 0xFFFFFFFF; m->cp_hqd_persistent_state = CP_HQD_PERSISTENT_STATE__PRELOAD_REQ_MASK | 0x53 << CP_HQD_PERSISTENT_STATE__PRELOAD_SIZE__SHIFT; @@ -179,6 +191,14 @@ static int load_mqd(struct mqd_manager *mm, void *mqd, wptr_shift, 0, mms); } +static int hiq_load_mqd_kiq(struct mqd_manager *mm, void *mqd, + uint32_t pipe_id, uint32_t queue_id, + struct queue_properties *p, struct mm_struct *mms) +{ + return mm->dev->kfd2kgd->hiq_mqd_load(mm->dev->kgd, mqd, pipe_id, + queue_id, p->doorbell_off); +} + static void update_mqd(struct mqd_manager *mm, void *mqd, struct queue_properties *q) { @@ -290,7 +310,8 @@ static int get_wave_state(struct mqd_manager *mm, void *mqd, *ctl_stack_used_size = m->cp_hqd_cntl_stack_size - m->cp_hqd_cntl_stack_offset; - *save_area_used_size = m->cp_hqd_wg_state_offset; + *save_area_used_size = m->cp_hqd_wg_state_offset - + m->cp_hqd_cntl_stack_size; if (copy_to_user(ctl_stack, mqd_ctl_stack, m->cp_hqd_cntl_stack_size)) return -EFAULT; @@ -312,18 +333,6 @@ static void init_mqd_hiq(struct mqd_manager *mm, void **mqd, 1 << CP_HQD_PQ_CONTROL__KMD_QUEUE__SHIFT; } -static void update_mqd_hiq(struct mqd_manager *mm, void *mqd, - struct queue_properties *q) -{ - struct v9_mqd *m; - - update_mqd(mm, mqd, q); - - /* TODO: what's the point? update_mqd already does this. */ - m = get_mqd(mqd); - m->cp_hqd_vmid = q->vmid; -} - static void init_mqd_sdma(struct mqd_manager *mm, void **mqd, struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr, struct queue_properties *q) @@ -431,7 +440,6 @@ struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type, switch (type) { case KFD_MQD_TYPE_CP: - case KFD_MQD_TYPE_COMPUTE: mqd->allocate_mqd = allocate_mqd; mqd->init_mqd = init_mqd; mqd->free_mqd = free_mqd; @@ -449,8 +457,8 @@ struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type, mqd->allocate_mqd = allocate_hiq_mqd; mqd->init_mqd = init_mqd_hiq; mqd->free_mqd = free_mqd_hiq_sdma; - mqd->load_mqd = load_mqd; - mqd->update_mqd = update_mqd_hiq; + mqd->load_mqd = hiq_load_mqd_kiq; + mqd->update_mqd = update_mqd; mqd->destroy_mqd = destroy_mqd; mqd->is_occupied = is_occupied; mqd->mqd_size = sizeof(struct v9_mqd); @@ -459,11 +467,11 @@ struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type, #endif break; case KFD_MQD_TYPE_DIQ: - mqd->allocate_mqd = allocate_hiq_mqd; + mqd->allocate_mqd = allocate_mqd; mqd->init_mqd = init_mqd_hiq; mqd->free_mqd = free_mqd; mqd->load_mqd = load_mqd; - mqd->update_mqd = update_mqd_hiq; + mqd->update_mqd = update_mqd; mqd->destroy_mqd = destroy_mqd; mqd->is_occupied = is_occupied; mqd->mqd_size = sizeof(struct v9_mqd); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c index 7d144f56f421..a5e8ff1e5945 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c @@ -312,11 +312,7 @@ static void init_mqd_hiq(struct mqd_manager *mm, void **mqd, static void update_mqd_hiq(struct mqd_manager *mm, void *mqd, struct queue_properties *q) { - struct vi_mqd *m; __update_mqd(mm, mqd, q, MTYPE_UC, 0); - - m = get_mqd(mqd); - m->cp_hqd_vmid = q->vmid; } static void init_mqd_sdma(struct mqd_manager *mm, void **mqd, @@ -425,7 +421,6 @@ struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type, switch (type) { case KFD_MQD_TYPE_CP: - case KFD_MQD_TYPE_COMPUTE: mqd->allocate_mqd = allocate_mqd; mqd->init_mqd = init_mqd; mqd->free_mqd = free_mqd; @@ -453,7 +448,7 @@ struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type, #endif break; case KFD_MQD_TYPE_DIQ: - mqd->allocate_mqd = allocate_hiq_mqd; + mqd->allocate_mqd = allocate_mqd; mqd->init_mqd = init_mqd_hiq; mqd->free_mqd = free_mqd; mqd->load_mqd = load_mqd; @@ -494,7 +489,7 @@ struct mqd_manager *mqd_manager_init_vi_tonga(enum KFD_MQD_TYPE type, mqd = mqd_manager_init_vi(type, dev); if (!mqd) return NULL; - if ((type == KFD_MQD_TYPE_CP) || (type == KFD_MQD_TYPE_COMPUTE)) + if (type == KFD_MQD_TYPE_CP) mqd->update_mqd = update_mqd_tonga; return mqd; } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c index ccf6b2310316..dc406e6dee23 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c @@ -239,10 +239,12 @@ int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm) case CHIP_VEGA12: case CHIP_VEGA20: case CHIP_RAVEN: - pm->pmf = &kfd_v9_pm_funcs; - break; + case CHIP_RENOIR: + case CHIP_ARCTURUS: case CHIP_NAVI10: - pm->pmf = &kfd_v10_pm_funcs; + case CHIP_NAVI12: + case CHIP_NAVI14: + pm->pmf = &kfd_v9_pm_funcs; break; default: WARN(1, "Unexpected ASIC family %u", @@ -262,10 +264,10 @@ int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm) return 0; } -void pm_uninit(struct packet_manager *pm) +void pm_uninit(struct packet_manager *pm, bool hanging) { mutex_destroy(&pm->lock); - kernel_queue_uninit(pm->priv_queue); + kernel_queue_uninit(pm->priv_queue, hanging); } int pm_send_set_resources(struct packet_manager *pm, @@ -276,7 +278,7 @@ int pm_send_set_resources(struct packet_manager *pm, size = pm->pmf->set_resources_size; mutex_lock(&pm->lock); - pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue, + kq_acquire_packet_buffer(pm->priv_queue, size / sizeof(uint32_t), (unsigned int **)&buffer); if (!buffer) { @@ -287,9 +289,9 @@ int pm_send_set_resources(struct packet_manager *pm, retval = pm->pmf->set_resources(pm, buffer, res); if (!retval) - pm->priv_queue->ops.submit_packet(pm->priv_queue); + kq_submit_packet(pm->priv_queue); else - pm->priv_queue->ops.rollback_packet(pm->priv_queue); + kq_rollback_packet(pm->priv_queue); out: mutex_unlock(&pm->lock); @@ -314,7 +316,7 @@ int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues) packet_size_dwords = pm->pmf->runlist_size / sizeof(uint32_t); mutex_lock(&pm->lock); - retval = pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue, + retval = kq_acquire_packet_buffer(pm->priv_queue, packet_size_dwords, &rl_buffer); if (retval) goto fail_acquire_packet_buffer; @@ -324,14 +326,14 @@ int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues) if (retval) goto fail_create_runlist; - pm->priv_queue->ops.submit_packet(pm->priv_queue); + kq_submit_packet(pm->priv_queue); mutex_unlock(&pm->lock); return retval; fail_create_runlist: - pm->priv_queue->ops.rollback_packet(pm->priv_queue); + kq_rollback_packet(pm->priv_queue); fail_acquire_packet_buffer: mutex_unlock(&pm->lock); fail_create_runlist_ib: @@ -350,7 +352,7 @@ int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address, size = pm->pmf->query_status_size; mutex_lock(&pm->lock); - pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue, + kq_acquire_packet_buffer(pm->priv_queue, size / sizeof(uint32_t), (unsigned int **)&buffer); if (!buffer) { pr_err("Failed to allocate buffer on kernel queue\n"); @@ -360,9 +362,9 @@ int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address, retval = pm->pmf->query_status(pm, buffer, fence_address, fence_value); if (!retval) - pm->priv_queue->ops.submit_packet(pm->priv_queue); + kq_submit_packet(pm->priv_queue); else - pm->priv_queue->ops.rollback_packet(pm->priv_queue); + kq_rollback_packet(pm->priv_queue); out: mutex_unlock(&pm->lock); @@ -379,7 +381,7 @@ int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type, size = pm->pmf->unmap_queues_size; mutex_lock(&pm->lock); - pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue, + kq_acquire_packet_buffer(pm->priv_queue, size / sizeof(uint32_t), (unsigned int **)&buffer); if (!buffer) { pr_err("Failed to allocate buffer on kernel queue\n"); @@ -390,9 +392,9 @@ int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type, retval = pm->pmf->unmap_queues(pm, buffer, type, filter, filter_param, reset, sdma_engine); if (!retval) - pm->priv_queue->ops.submit_packet(pm->priv_queue); + kq_submit_packet(pm->priv_queue); else - pm->priv_queue->ops.rollback_packet(pm->priv_queue); + kq_rollback_packet(pm->priv_queue); out: mutex_unlock(&pm->lock); @@ -437,7 +439,7 @@ int pm_debugfs_hang_hws(struct packet_manager *pm) size = pm->pmf->query_status_size; mutex_lock(&pm->lock); - pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue, + kq_acquire_packet_buffer(pm->priv_queue, size / sizeof(uint32_t), (unsigned int **)&buffer); if (!buffer) { pr_err("Failed to allocate buffer on kernel queue\n"); @@ -445,7 +447,7 @@ int pm_debugfs_hang_hws(struct packet_manager *pm) goto out; } memset(buffer, 0x55, size); - pm->priv_queue->ops.submit_packet(pm->priv_queue); + kq_submit_packet(pm->priv_queue); pr_info("Submitting %x %x %x %x %x %x %x to HIQ to hang the HWS.", buffer[0], buffer[1], buffer[2], buffer[3], diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c index 2d5ddf199bd0..2de01009f1b6 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c @@ -25,47 +25,7 @@ #include "kfd_device_queue_manager.h" #include "kfd_pm4_headers_ai.h" #include "kfd_pm4_opcodes.h" - -static bool initialize_v9(struct kernel_queue *kq, struct kfd_dev *dev, - enum kfd_queue_type type, unsigned int queue_size); -static void uninitialize_v9(struct kernel_queue *kq); -static void submit_packet_v9(struct kernel_queue *kq); - -void kernel_queue_init_v9(struct kernel_queue_ops *ops) -{ - ops->initialize = initialize_v9; - ops->uninitialize = uninitialize_v9; - ops->submit_packet = submit_packet_v9; -} - -static bool initialize_v9(struct kernel_queue *kq, struct kfd_dev *dev, - enum kfd_queue_type type, unsigned int queue_size) -{ - int retval; - - retval = kfd_gtt_sa_allocate(dev, PAGE_SIZE, &kq->eop_mem); - if (retval) - return false; - - kq->eop_gpu_addr = kq->eop_mem->gpu_addr; - kq->eop_kernel_addr = kq->eop_mem->cpu_ptr; - - memset(kq->eop_kernel_addr, 0, PAGE_SIZE); - - return true; -} - -static void uninitialize_v9(struct kernel_queue *kq) -{ - kfd_gtt_sa_free(kq->dev, kq->eop_mem); -} - -static void submit_packet_v9(struct kernel_queue *kq) -{ - *kq->wptr64_kernel = kq->pending_wptr64; - write_kernel_doorbell64(kq->queue->properties.doorbell_ptr, - kq->pending_wptr64); -} +#include "gc/gc_10_1_0_sh_mask.h" static int pm_map_process_v9(struct packet_manager *pm, uint32_t *buffer, struct qcm_process_device *qpd) @@ -81,7 +41,8 @@ static int pm_map_process_v9(struct packet_manager *pm, packet->bitfields2.diq_enable = (qpd->is_debug) ? 1 : 0; packet->bitfields2.process_quantum = 1; packet->bitfields2.pasid = qpd->pqm->process->pasid; - packet->bitfields14.gds_size = qpd->gds_size; + packet->bitfields14.gds_size = qpd->gds_size & 0x3F; + packet->bitfields14.gds_size_hi = (qpd->gds_size >> 6) & 0xF; packet->bitfields14.num_gws = qpd->num_gws; packet->bitfields14.num_oac = qpd->num_oac; packet->bitfields14.sdma_enable = 1; @@ -89,10 +50,17 @@ static int pm_map_process_v9(struct packet_manager *pm, packet->sh_mem_config = qpd->sh_mem_config; packet->sh_mem_bases = qpd->sh_mem_bases; - packet->sq_shader_tba_lo = lower_32_bits(qpd->tba_addr >> 8); - packet->sq_shader_tba_hi = upper_32_bits(qpd->tba_addr >> 8); - packet->sq_shader_tma_lo = lower_32_bits(qpd->tma_addr >> 8); - packet->sq_shader_tma_hi = upper_32_bits(qpd->tma_addr >> 8); + if (qpd->tba_addr) { + packet->sq_shader_tba_lo = lower_32_bits(qpd->tba_addr >> 8); + /* On GFX9, unlike GFX10, bit TRAP_EN of SQ_SHADER_TBA_HI is + * not defined, so setting it won't do any harm. + */ + packet->sq_shader_tba_hi = upper_32_bits(qpd->tba_addr >> 8) + | 1 << SQ_SHADER_TBA_HI__TRAP_EN__SHIFT; + + packet->sq_shader_tma_lo = lower_32_bits(qpd->tma_addr >> 8); + packet->sq_shader_tma_hi = upper_32_bits(qpd->tma_addr >> 8); + } packet->gds_addr_lo = lower_32_bits(qpd->gds_context_area); packet->gds_addr_hi = upper_32_bits(qpd->gds_context_area); @@ -143,6 +111,34 @@ static int pm_runlist_v9(struct packet_manager *pm, uint32_t *buffer, return 0; } +static int pm_set_resources_v9(struct packet_manager *pm, uint32_t *buffer, + struct scheduling_resources *res) +{ + struct pm4_mes_set_resources *packet; + + packet = (struct pm4_mes_set_resources *)buffer; + memset(buffer, 0, sizeof(struct pm4_mes_set_resources)); + + packet->header.u32All = pm_build_pm4_header(IT_SET_RESOURCES, + sizeof(struct pm4_mes_set_resources)); + + packet->bitfields2.queue_type = + queue_type__mes_set_resources__hsa_interface_queue_hiq; + packet->bitfields2.vmid_mask = res->vmid_mask; + packet->bitfields2.unmap_latency = KFD_UNMAP_LATENCY_MS / 100; + packet->bitfields7.oac_mask = res->oac_mask; + packet->bitfields8.gds_heap_base = res->gds_heap_base; + packet->bitfields8.gds_heap_size = res->gds_heap_size; + + packet->gws_mask_lo = lower_32_bits(res->gws_mask); + packet->gws_mask_hi = upper_32_bits(res->gws_mask); + + packet->queue_mask_lo = lower_32_bits(res->queue_mask); + packet->queue_mask_hi = upper_32_bits(res->queue_mask); + + return 0; +} + static int pm_map_queues_v9(struct packet_manager *pm, uint32_t *buffer, struct queue *q, bool is_static) { @@ -161,6 +157,8 @@ static int pm_map_queues_v9(struct packet_manager *pm, uint32_t *buffer, packet->bitfields2.engine_sel = engine_sel__mes_map_queues__compute_vi; packet->bitfields2.gws_control_queue = q->gws ? 1 : 0; + packet->bitfields2.extended_engine_sel = + extended_engine_sel__mes_map_queues__legacy_engine_sel; packet->bitfields2.queue_type = queue_type__mes_map_queues__normal_compute_vi; @@ -176,9 +174,15 @@ static int pm_map_queues_v9(struct packet_manager *pm, uint32_t *buffer, break; case KFD_QUEUE_TYPE_SDMA: case KFD_QUEUE_TYPE_SDMA_XGMI: - packet->bitfields2.engine_sel = q->properties.sdma_engine_id + - engine_sel__mes_map_queues__sdma0_vi; use_static = false; /* no static queues under SDMA */ + if (q->properties.sdma_engine_id < 2) + packet->bitfields2.engine_sel = q->properties.sdma_engine_id + + engine_sel__mes_map_queues__sdma0_vi; + else { + packet->bitfields2.extended_engine_sel = + extended_engine_sel__mes_map_queues__sdma0_to_7_sel; + packet->bitfields2.engine_sel = q->properties.sdma_engine_id; + } break; default: WARN(1, "queue type %d", q->properties.type); @@ -218,13 +222,23 @@ static int pm_unmap_queues_v9(struct packet_manager *pm, uint32_t *buffer, switch (type) { case KFD_QUEUE_TYPE_COMPUTE: case KFD_QUEUE_TYPE_DIQ: + packet->bitfields2.extended_engine_sel = + extended_engine_sel__mes_unmap_queues__legacy_engine_sel; packet->bitfields2.engine_sel = engine_sel__mes_unmap_queues__compute; break; case KFD_QUEUE_TYPE_SDMA: case KFD_QUEUE_TYPE_SDMA_XGMI: - packet->bitfields2.engine_sel = - engine_sel__mes_unmap_queues__sdma0 + sdma_engine; + if (sdma_engine < 2) { + packet->bitfields2.extended_engine_sel = + extended_engine_sel__mes_unmap_queues__legacy_engine_sel; + packet->bitfields2.engine_sel = + engine_sel__mes_unmap_queues__sdma0 + sdma_engine; + } else { + packet->bitfields2.extended_engine_sel = + extended_engine_sel__mes_unmap_queues__sdma0_to_7_sel; + packet->bitfields2.engine_sel = sdma_engine; + } break; default: WARN(1, "queue type %d", type); @@ -294,48 +308,19 @@ static int pm_query_status_v9(struct packet_manager *pm, uint32_t *buffer, return 0; } - -static int pm_release_mem_v9(uint64_t gpu_addr, uint32_t *buffer) -{ - struct pm4_mec_release_mem *packet; - - packet = (struct pm4_mec_release_mem *)buffer; - memset(buffer, 0, sizeof(struct pm4_mec_release_mem)); - - packet->header.u32All = pm_build_pm4_header(IT_RELEASE_MEM, - sizeof(struct pm4_mec_release_mem)); - - packet->bitfields2.event_type = CACHE_FLUSH_AND_INV_TS_EVENT; - packet->bitfields2.event_index = event_index__mec_release_mem__end_of_pipe; - packet->bitfields2.tcl1_action_ena = 1; - packet->bitfields2.tc_action_ena = 1; - packet->bitfields2.cache_policy = cache_policy__mec_release_mem__lru; - - packet->bitfields3.data_sel = data_sel__mec_release_mem__send_32_bit_low; - packet->bitfields3.int_sel = - int_sel__mec_release_mem__send_interrupt_after_write_confirm; - - packet->bitfields4.address_lo_32b = (gpu_addr & 0xffffffff) >> 2; - packet->address_hi = upper_32_bits(gpu_addr); - - packet->data_lo = 0; - - return 0; -} - const struct packet_manager_funcs kfd_v9_pm_funcs = { .map_process = pm_map_process_v9, .runlist = pm_runlist_v9, - .set_resources = pm_set_resources_vi, + .set_resources = pm_set_resources_v9, .map_queues = pm_map_queues_v9, .unmap_queues = pm_unmap_queues_v9, .query_status = pm_query_status_v9, - .release_mem = pm_release_mem_v9, + .release_mem = NULL, .map_process_size = sizeof(struct pm4_mes_map_process), .runlist_size = sizeof(struct pm4_mes_runlist), .set_resources_size = sizeof(struct pm4_mes_set_resources), .map_queues_size = sizeof(struct pm4_mes_map_queues), .unmap_queues_size = sizeof(struct pm4_mes_unmap_queues), .query_status_size = sizeof(struct pm4_mes_query_status), - .release_mem_size = sizeof(struct pm4_mec_release_mem) + .release_mem_size = 0, }; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue_vi.c b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_vi.c index 2adaf40027eb..bed4d0ccb6b1 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue_vi.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_vi.c @@ -26,47 +26,6 @@ #include "kfd_pm4_headers_vi.h" #include "kfd_pm4_opcodes.h" -static bool initialize_vi(struct kernel_queue *kq, struct kfd_dev *dev, - enum kfd_queue_type type, unsigned int queue_size); -static void uninitialize_vi(struct kernel_queue *kq); -static void submit_packet_vi(struct kernel_queue *kq); - -void kernel_queue_init_vi(struct kernel_queue_ops *ops) -{ - ops->initialize = initialize_vi; - ops->uninitialize = uninitialize_vi; - ops->submit_packet = submit_packet_vi; -} - -static bool initialize_vi(struct kernel_queue *kq, struct kfd_dev *dev, - enum kfd_queue_type type, unsigned int queue_size) -{ - int retval; - - retval = kfd_gtt_sa_allocate(dev, PAGE_SIZE, &kq->eop_mem); - if (retval != 0) - return false; - - kq->eop_gpu_addr = kq->eop_mem->gpu_addr; - kq->eop_kernel_addr = kq->eop_mem->cpu_ptr; - - memset(kq->eop_kernel_addr, 0, PAGE_SIZE); - - return true; -} - -static void uninitialize_vi(struct kernel_queue *kq) -{ - kfd_gtt_sa_free(kq->dev, kq->eop_mem); -} - -static void submit_packet_vi(struct kernel_queue *kq) -{ - *kq->wptr_kernel = kq->pending_wptr; - write_kernel_doorbell(kq->queue->properties.doorbell_ptr, - kq->pending_wptr); -} - unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size) { union PM4_MES_TYPE_3_HEADER header; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h b/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h index e3e21404cfa0..4d7add843746 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h @@ -83,10 +83,10 @@ struct pm4_mes_set_resources { union { struct { - uint32_t gds_heap_base:6; - uint32_t reserved3:5; - uint32_t gds_heap_size:6; - uint32_t reserved4:15; + uint32_t gds_heap_base:10; + uint32_t reserved3:1; + uint32_t gds_heap_size:10; + uint32_t reserved4:11; } bitfields8; uint32_t ordinal8; }; @@ -179,7 +179,7 @@ struct pm4_mes_map_process { uint32_t num_gws:7; uint32_t sdma_enable:1; uint32_t num_oac:4; - uint32_t reserved8:4; + uint32_t gds_size_hi:4; uint32_t gds_size:6; uint32_t num_queues:10; } bitfields14; @@ -260,6 +260,10 @@ enum mes_map_queues_engine_sel_enum { engine_sel__mes_map_queues__sdma1_vi = 3 }; +enum mes_map_queues_extended_engine_sel_enum { + extended_engine_sel__mes_map_queues__legacy_engine_sel = 0, + extended_engine_sel__mes_map_queues__sdma0_to_7_sel = 1 +}; struct pm4_mes_map_queues { union { @@ -269,7 +273,8 @@ struct pm4_mes_map_queues { union { struct { - uint32_t reserved1:4; + uint32_t reserved1:2; + enum mes_map_queues_extended_engine_sel_enum extended_engine_sel:2; enum mes_map_queues_queue_sel_enum queue_sel:2; uint32_t reserved5:6; uint32_t gws_control_queue:1; @@ -382,6 +387,11 @@ enum mes_unmap_queues_engine_sel_enum { engine_sel__mes_unmap_queues__sdmal = 3 }; +enum mes_unmap_queues_extended_engine_sel_enum { + extended_engine_sel__mes_unmap_queues__legacy_engine_sel = 0, + extended_engine_sel__mes_unmap_queues__sdma0_to_7_sel = 1 +}; + struct pm4_mes_unmap_queues { union { union PM4_MES_TYPE_3_HEADER header; /* header */ @@ -391,7 +401,7 @@ struct pm4_mes_unmap_queues { union { struct { enum mes_unmap_queues_action_enum action:2; - uint32_t reserved1:2; + enum mes_unmap_queues_extended_engine_sel_enum extended_engine_sel:2; enum mes_unmap_queues_queue_sel_enum queue_sel:2; uint32_t reserved2:20; enum mes_unmap_queues_engine_sel_enum engine_sel:3; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index 3933fb6a371e..6af1b5881f43 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -36,6 +36,10 @@ #include <linux/seq_file.h> #include <linux/kref.h> #include <linux/sysfs.h> +#include <linux/device_cgroup.h> +#include <drm/drm_file.h> +#include <drm/drm_drv.h> +#include <drm/drm_device.h> #include <kgd_kfd_interface.h> #include "amd_shared.h" @@ -55,24 +59,21 @@ * NOTE: struct vm_area_struct.vm_pgoff uses offset in pages. Hence, these * defines are w.r.t to PAGE_SIZE */ -#define KFD_MMAP_TYPE_SHIFT (62 - PAGE_SHIFT) +#define KFD_MMAP_TYPE_SHIFT 62 #define KFD_MMAP_TYPE_MASK (0x3ULL << KFD_MMAP_TYPE_SHIFT) #define KFD_MMAP_TYPE_DOORBELL (0x3ULL << KFD_MMAP_TYPE_SHIFT) #define KFD_MMAP_TYPE_EVENTS (0x2ULL << KFD_MMAP_TYPE_SHIFT) #define KFD_MMAP_TYPE_RESERVED_MEM (0x1ULL << KFD_MMAP_TYPE_SHIFT) #define KFD_MMAP_TYPE_MMIO (0x0ULL << KFD_MMAP_TYPE_SHIFT) -#define KFD_MMAP_GPU_ID_SHIFT (46 - PAGE_SHIFT) +#define KFD_MMAP_GPU_ID_SHIFT 46 #define KFD_MMAP_GPU_ID_MASK (((1ULL << KFD_GPU_ID_HASH_WIDTH) - 1) \ << KFD_MMAP_GPU_ID_SHIFT) #define KFD_MMAP_GPU_ID(gpu_id) ((((uint64_t)gpu_id) << KFD_MMAP_GPU_ID_SHIFT)\ & KFD_MMAP_GPU_ID_MASK) -#define KFD_MMAP_GPU_ID_GET(offset) ((offset & KFD_MMAP_GPU_ID_MASK) \ +#define KFD_MMAP_GET_GPU_ID(offset) ((offset & KFD_MMAP_GPU_ID_MASK) \ >> KFD_MMAP_GPU_ID_SHIFT) -#define KFD_MMAP_OFFSET_VALUE_MASK (0x3FFFFFFFFFFFULL >> PAGE_SHIFT) -#define KFD_MMAP_OFFSET_VALUE_GET(offset) (offset & KFD_MMAP_OFFSET_VALUE_MASK) - /* * When working with cp scheduler we should assign the HIQ manually or via * the amdgpu driver to a fixed hqd slot, here are the fixed HIQ hqd slot @@ -179,10 +180,6 @@ enum cache_policy { cache_policy_noncoherent }; -#define KFD_IS_VI(chip) ((chip) >= CHIP_CARRIZO && (chip) <= CHIP_POLARIS11) -#define KFD_IS_DGPU(chip) (((chip) >= CHIP_TONGA && \ - (chip) <= CHIP_NAVI10) || \ - (chip) == CHIP_HAWAII) #define KFD_IS_SOC15(chip) ((chip) >= CHIP_VEGA10) struct kfd_event_interrupt_class { @@ -195,6 +192,7 @@ struct kfd_event_interrupt_class { struct kfd_device_info { enum amd_asic_type asic_family; + const char *asic_name; const struct kfd_event_interrupt_class *event_interrupt_class; unsigned int max_pasid_bits; unsigned int max_no_of_hqd; @@ -229,6 +227,7 @@ struct kfd_dev { const struct kfd_device_info *device_info; struct pci_dev *pdev; + struct drm_device *ddev; unsigned int id; /* topology stub index */ @@ -236,9 +235,10 @@ struct kfd_dev { * KFD. It is aligned for mapping * into user mode */ - size_t doorbell_id_offset; /* Doorbell offset (from KFD doorbell - * to HW doorbell, GFX reserved some - * at the start) + size_t doorbell_base_dw_offset; /* Offset from the start of the PCI + * doorbell BAR to the first KFD + * doorbell in dwords. GFX reserves + * the segment before this offset. */ u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells * page used by kernel queue @@ -508,8 +508,7 @@ struct queue { * Please read the kfd_mqd_manager.h description. */ enum KFD_MQD_TYPE { - KFD_MQD_TYPE_COMPUTE = 0, /* for no cp scheduling */ - KFD_MQD_TYPE_HIQ, /* for hiq */ + KFD_MQD_TYPE_HIQ = 0, /* for hiq */ KFD_MQD_TYPE_CP, /* for cp queues and diq */ KFD_MQD_TYPE_SDMA, /* for sdma queues */ KFD_MQD_TYPE_DIQ, /* for diq */ @@ -686,10 +685,7 @@ struct kfd_process { /* We want to receive a notification when the mm_struct is destroyed */ struct mmu_notifier mmu_notifier; - /* Use for delayed freeing of kfd_process structure */ - struct rcu_head rcu; - - unsigned int pasid; + uint16_t pasid; unsigned int doorbell_index; /* @@ -819,7 +815,7 @@ void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr); u32 read_kernel_doorbell(u32 __iomem *db); void write_kernel_doorbell(void __iomem *db, u32 value); void write_kernel_doorbell64(void __iomem *db, u64 value); -unsigned int kfd_doorbell_id_to_offset(struct kfd_dev *kfd, +unsigned int kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev *kfd, struct kfd_process *process, unsigned int doorbell_id); phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev, @@ -887,7 +883,7 @@ struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev); void device_queue_manager_uninit(struct device_queue_manager *dqm); struct kernel_queue *kernel_queue_init(struct kfd_dev *dev, enum kfd_queue_type type); -void kernel_queue_uninit(struct kernel_queue *kq); +void kernel_queue_uninit(struct kernel_queue *kq, bool hanging); int kfd_process_vm_fault(struct device_queue_manager *dqm, unsigned int pasid); /* Process Queue Manager */ @@ -905,7 +901,8 @@ int pqm_create_queue(struct process_queue_manager *pqm, struct kfd_dev *dev, struct file *f, struct queue_properties *properties, - unsigned int *qid); + unsigned int *qid, + uint32_t *p_doorbell_offset_in_process); int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid); int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid, struct queue_properties *p); @@ -973,10 +970,9 @@ struct packet_manager_funcs { extern const struct packet_manager_funcs kfd_vi_pm_funcs; extern const struct packet_manager_funcs kfd_v9_pm_funcs; -extern const struct packet_manager_funcs kfd_v10_pm_funcs; int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm); -void pm_uninit(struct packet_manager *pm); +void pm_uninit(struct packet_manager *pm, bool hanging); int pm_send_set_resources(struct packet_manager *pm, struct scheduling_resources *res); int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues); @@ -992,9 +988,6 @@ void pm_release_ib(struct packet_manager *pm); /* Following PM funcs can be shared among VI and AI */ unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size); -int pm_set_resources_vi(struct packet_manager *pm, uint32_t *buffer, - struct scheduling_resources *res); - uint64_t kfd_get_number_elems(struct kfd_dev *kfd); @@ -1042,6 +1035,21 @@ bool kfd_is_locked(void); void kfd_inc_compute_active(struct kfd_dev *dev); void kfd_dec_compute_active(struct kfd_dev *dev); +/* Cgroup Support */ +/* Check with device cgroup if @kfd device is accessible */ +static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd) +{ +#if defined(CONFIG_CGROUP_DEVICE) + struct drm_device *ddev = kfd->ddev; + + return devcgroup_check_permission(DEVCG_DEV_CHAR, ddev->driver->major, + ddev->render->index, + DEVCG_ACC_WRITE | DEVCG_ACC_READ); +#else + return 0; +#endif +} + /* Debugfs */ #if defined(CONFIG_DEBUG_FS) diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c index 8f1076c0c88a..25b90f70aecd 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c @@ -32,6 +32,7 @@ #include <linux/mman.h> #include <linux/file.h> #include "amdgpu_amdkfd.h" +#include "amdgpu.h" struct mm_struct; @@ -62,8 +63,8 @@ static struct workqueue_struct *kfd_restore_wq; static struct kfd_process *find_process(const struct task_struct *thread); static void kfd_process_ref_release(struct kref *ref); -static struct kfd_process *create_process(const struct task_struct *thread, - struct file *filep); +static struct kfd_process *create_process(const struct task_struct *thread); +static int kfd_process_init_cwsr_apu(struct kfd_process *p, struct file *filep); static void evict_process_worker(struct work_struct *work); static void restore_process_worker(struct work_struct *work); @@ -289,7 +290,15 @@ struct kfd_process *kfd_create_process(struct file *filep) if (process) { pr_debug("Process already found\n"); } else { - process = create_process(thread, filep); + process = create_process(thread); + if (IS_ERR(process)) + goto out; + + ret = kfd_process_init_cwsr_apu(process, filep); + if (ret) { + process = ERR_PTR(ret); + goto out; + } if (!procfs.kobj) goto out; @@ -316,6 +325,8 @@ struct kfd_process *kfd_create_process(struct file *filep) (int)process->lead_thread->pid); } out: + if (!IS_ERR(process)) + kref_get(&process->ref); mutex_unlock(&kfd_processes_mutex); return process; @@ -408,7 +419,7 @@ static void kfd_process_destroy_pdds(struct kfd_process *p) list_for_each_entry_safe(pdd, temp, &p->per_device_data, per_device_list) { - pr_debug("Releasing pdd (topology id %d) for process (pasid %d)\n", + pr_debug("Releasing pdd (topology id %d) for process (pasid 0x%x)\n", pdd->dev->id, p->pasid); if (pdd->drm_file) { @@ -478,11 +489,9 @@ static void kfd_process_ref_release(struct kref *ref) queue_work(kfd_process_wq, &p->release_work); } -static void kfd_process_destroy_delayed(struct rcu_head *rcu) +static void kfd_process_free_notifier(struct mmu_notifier *mn) { - struct kfd_process *p = container_of(rcu, struct kfd_process, rcu); - - kfd_unref_process(p); + kfd_unref_process(container_of(mn, struct kfd_process, mmu_notifier)); } static void kfd_process_notifier_release(struct mmu_notifier *mn, @@ -534,12 +543,12 @@ static void kfd_process_notifier_release(struct mmu_notifier *mn, mutex_unlock(&p->mutex); - mmu_notifier_unregister_no_release(&p->mmu_notifier, mm); - mmu_notifier_call_srcu(&p->rcu, &kfd_process_destroy_delayed); + mmu_notifier_put(&p->mmu_notifier); } static const struct mmu_notifier_ops kfd_process_mmu_notifier_ops = { .release = kfd_process_notifier_release, + .free_notifier = kfd_process_free_notifier, }; static int kfd_process_init_cwsr_apu(struct kfd_process *p, struct file *filep) @@ -554,8 +563,7 @@ static int kfd_process_init_cwsr_apu(struct kfd_process *p, struct file *filep) if (!dev->cwsr_enabled || qpd->cwsr_kaddr || qpd->cwsr_base) continue; - offset = (KFD_MMAP_TYPE_RESERVED_MEM | KFD_MMAP_GPU_ID(dev->id)) - << PAGE_SHIFT; + offset = KFD_MMAP_TYPE_RESERVED_MEM | KFD_MMAP_GPU_ID(dev->id); qpd->tba_addr = (int64_t)vm_mmap(filep, 0, KFD_CWSR_TBA_TMA_SIZE, PROT_READ | PROT_EXEC, MAP_SHARED, offset); @@ -609,81 +617,69 @@ static int kfd_process_device_init_cwsr_dgpu(struct kfd_process_device *pdd) return 0; } -static struct kfd_process *create_process(const struct task_struct *thread, - struct file *filep) +/* + * On return the kfd_process is fully operational and will be freed when the + * mm is released + */ +static struct kfd_process *create_process(const struct task_struct *thread) { struct kfd_process *process; int err = -ENOMEM; process = kzalloc(sizeof(*process), GFP_KERNEL); - if (!process) goto err_alloc_process; - process->pasid = kfd_pasid_alloc(); - if (process->pasid == 0) - goto err_alloc_pasid; - - if (kfd_alloc_process_doorbells(process) < 0) - goto err_alloc_doorbells; - kref_init(&process->ref); - mutex_init(&process->mutex); - process->mm = thread->mm; - - /* register notifier */ - process->mmu_notifier.ops = &kfd_process_mmu_notifier_ops; - err = mmu_notifier_register(&process->mmu_notifier, process->mm); - if (err) - goto err_mmu_notifier; - - hash_add_rcu(kfd_processes_table, &process->kfd_processes, - (uintptr_t)process->mm); - process->lead_thread = thread->group_leader; - get_task_struct(process->lead_thread); - INIT_LIST_HEAD(&process->per_device_data); - + INIT_DELAYED_WORK(&process->eviction_work, evict_process_worker); + INIT_DELAYED_WORK(&process->restore_work, restore_process_worker); + process->last_restore_timestamp = get_jiffies_64(); kfd_event_init_process(process); + process->is_32bit_user_mode = in_compat_syscall(); + + process->pasid = kfd_pasid_alloc(); + if (process->pasid == 0) + goto err_alloc_pasid; + + if (kfd_alloc_process_doorbells(process) < 0) + goto err_alloc_doorbells; err = pqm_init(&process->pqm, process); if (err != 0) goto err_process_pqm_init; /* init process apertures*/ - process->is_32bit_user_mode = in_compat_syscall(); err = kfd_init_apertures(process); if (err != 0) goto err_init_apertures; - INIT_DELAYED_WORK(&process->eviction_work, evict_process_worker); - INIT_DELAYED_WORK(&process->restore_work, restore_process_worker); - process->last_restore_timestamp = get_jiffies_64(); - - err = kfd_process_init_cwsr_apu(process, filep); + /* Must be last, have to use release destruction after this */ + process->mmu_notifier.ops = &kfd_process_mmu_notifier_ops; + err = mmu_notifier_register(&process->mmu_notifier, process->mm); if (err) - goto err_init_cwsr; + goto err_register_notifier; + + get_task_struct(process->lead_thread); + hash_add_rcu(kfd_processes_table, &process->kfd_processes, + (uintptr_t)process->mm); return process; -err_init_cwsr: +err_register_notifier: kfd_process_free_outstanding_kfd_bos(process); kfd_process_destroy_pdds(process); err_init_apertures: pqm_uninit(&process->pqm); err_process_pqm_init: - hash_del_rcu(&process->kfd_processes); - synchronize_rcu(); - mmu_notifier_unregister_no_release(&process->mmu_notifier, process->mm); -err_mmu_notifier: - mutex_destroy(&process->mutex); kfd_free_process_doorbells(process); err_alloc_doorbells: kfd_pasid_free(process->pasid); err_alloc_pasid: + mutex_destroy(&process->mutex); kfree(process); err_alloc_process: return ERR_PTR(err); @@ -693,6 +689,8 @@ static int init_doorbell_bitmap(struct qcm_process_device *qpd, struct kfd_dev *dev) { unsigned int i; + int range_start = dev->shared_resources.non_cp_doorbells_start; + int range_end = dev->shared_resources.non_cp_doorbells_end; if (!KFD_IS_SOC15(dev->device_info->asic_family)) return 0; @@ -704,14 +702,16 @@ static int init_doorbell_bitmap(struct qcm_process_device *qpd, return -ENOMEM; /* Mask out doorbells reserved for SDMA, IH, and VCN on SOC15. */ + pr_debug("reserved doorbell 0x%03x - 0x%03x\n", range_start, range_end); + pr_debug("reserved doorbell 0x%03x - 0x%03x\n", + range_start + KFD_QUEUE_DOORBELL_MIRROR_OFFSET, + range_end + KFD_QUEUE_DOORBELL_MIRROR_OFFSET); + for (i = 0; i < KFD_MAX_NUM_OF_QUEUES_PER_PROCESS / 2; i++) { - if (i >= dev->shared_resources.non_cp_doorbells_start - && i <= dev->shared_resources.non_cp_doorbells_end) { + if (i >= range_start && i <= range_end) { set_bit(i, qpd->doorbell_bitmap); set_bit(i + KFD_QUEUE_DOORBELL_MIRROR_OFFSET, qpd->doorbell_bitmap); - pr_debug("reserved doorbell 0x%03x and 0x%03x\n", i, - i + KFD_QUEUE_DOORBELL_MIRROR_OFFSET); } } @@ -801,6 +801,8 @@ int kfd_process_device_init_vm(struct kfd_process_device *pdd, return ret; } + amdgpu_vm_set_task_info(pdd->vm); + ret = kfd_process_device_reserve_ib_mem(pdd); if (ret) goto err_reserve_ib_mem; @@ -1024,7 +1026,7 @@ static void evict_process_worker(struct work_struct *work) */ flush_delayed_work(&p->restore_work); - pr_debug("Started evicting pasid %d\n", p->pasid); + pr_debug("Started evicting pasid 0x%x\n", p->pasid); ret = kfd_process_evict_queues(p); if (!ret) { dma_fence_signal(p->ef); @@ -1033,16 +1035,15 @@ static void evict_process_worker(struct work_struct *work) queue_delayed_work(kfd_restore_wq, &p->restore_work, msecs_to_jiffies(PROCESS_RESTORE_TIME_MS)); - pr_debug("Finished evicting pasid %d\n", p->pasid); + pr_debug("Finished evicting pasid 0x%x\n", p->pasid); } else - pr_err("Failed to evict queues of pasid %d\n", p->pasid); + pr_err("Failed to evict queues of pasid 0x%x\n", p->pasid); } static void restore_process_worker(struct work_struct *work) { struct delayed_work *dwork; struct kfd_process *p; - struct kfd_process_device *pdd; int ret = 0; dwork = to_delayed_work(work); @@ -1051,17 +1052,7 @@ static void restore_process_worker(struct work_struct *work) * lifetime of this thread, kfd_process p will be valid */ p = container_of(dwork, struct kfd_process, restore_work); - - /* Call restore_process_bos on the first KGD device. This function - * takes care of restoring the whole process including other devices. - * Restore can fail if enough memory is not available. If so, - * reschedule again. - */ - pdd = list_first_entry(&p->per_device_data, - struct kfd_process_device, - per_device_list); - - pr_debug("Started restoring pasid %d\n", p->pasid); + pr_debug("Started restoring pasid 0x%x\n", p->pasid); /* Setting last_restore_timestamp before successful restoration. * Otherwise this would have to be set by KGD (restore_process_bos) @@ -1077,7 +1068,7 @@ static void restore_process_worker(struct work_struct *work) ret = amdgpu_amdkfd_gpuvm_restore_process_bos(p->kgd_process_info, &p->ef); if (ret) { - pr_debug("Failed to restore BOs of pasid %d, retry after %d ms\n", + pr_debug("Failed to restore BOs of pasid 0x%x, retry after %d ms\n", p->pasid, PROCESS_BACK_OFF_TIME_MS); ret = queue_delayed_work(kfd_restore_wq, &p->restore_work, msecs_to_jiffies(PROCESS_BACK_OFF_TIME_MS)); @@ -1087,9 +1078,9 @@ static void restore_process_worker(struct work_struct *work) ret = kfd_process_restore_queues(p); if (!ret) - pr_debug("Finished restoring pasid %d\n", p->pasid); + pr_debug("Finished restoring pasid 0x%x\n", p->pasid); else - pr_err("Failed to restore queues of pasid %d\n", p->pasid); + pr_err("Failed to restore queues of pasid 0x%x\n", p->pasid); } void kfd_suspend_all_processes(void) @@ -1103,7 +1094,7 @@ void kfd_suspend_all_processes(void) cancel_delayed_work_sync(&p->restore_work); if (kfd_process_evict_queues(p)) - pr_err("Failed to suspend process %d\n", p->pasid); + pr_err("Failed to suspend process 0x%x\n", p->pasid); dma_fence_signal(p->ef); dma_fence_put(p->ef); p->ef = NULL; @@ -1162,16 +1153,17 @@ int kfd_reserved_mem_mmap(struct kfd_dev *dev, struct kfd_process *process, void kfd_flush_tlb(struct kfd_process_device *pdd) { struct kfd_dev *dev = pdd->dev; - const struct kfd2kgd_calls *f2g = dev->kfd2kgd; if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) { /* Nothing to flush until a VMID is assigned, which * only happens when the first queue is created. */ if (pdd->qpd.vmid) - f2g->invalidate_tlbs_vmid(dev->kgd, pdd->qpd.vmid); + amdgpu_amdkfd_flush_gpu_tlb_vmid(dev->kgd, + pdd->qpd.vmid); } else { - f2g->invalidate_tlbs(dev->kgd, pdd->process->pasid); + amdgpu_amdkfd_flush_gpu_tlb_pasid(dev->kgd, + pdd->process->pasid); } } @@ -1186,7 +1178,7 @@ int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data) int idx = srcu_read_lock(&kfd_processes_srcu); hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) { - seq_printf(m, "Process %d PASID %d:\n", + seq_printf(m, "Process %d PASID 0x%x:\n", p->lead_thread->tgid, p->pasid); mutex_lock(&p->mutex); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c index 7e6c3ee82f5b..31fcd1b51f00 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c @@ -53,7 +53,7 @@ static int find_available_queue_slot(struct process_queue_manager *pqm, pr_debug("The new slot id %lu\n", found); if (found >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) { - pr_info("Cannot open more queues for process with pasid %d\n", + pr_info("Cannot open more queues for process with pasid 0x%x\n", pqm->process->pasid); return -ENOMEM; } @@ -162,7 +162,7 @@ void pqm_uninit(struct process_queue_manager *pqm) pqm->queue_slot_bitmap = NULL; } -static int create_cp_queue(struct process_queue_manager *pqm, +static int init_user_queue(struct process_queue_manager *pqm, struct kfd_dev *dev, struct queue **q, struct queue_properties *q_properties, struct file *f, unsigned int qid) @@ -192,7 +192,8 @@ int pqm_create_queue(struct process_queue_manager *pqm, struct kfd_dev *dev, struct file *f, struct queue_properties *properties, - unsigned int *qid) + unsigned int *qid, + uint32_t *p_doorbell_offset_in_process) { int retval; struct kfd_process_device *pdd; @@ -250,7 +251,7 @@ int pqm_create_queue(struct process_queue_manager *pqm, goto err_create_queue; } - retval = create_cp_queue(pqm, dev, &q, properties, f, *qid); + retval = init_user_queue(pqm, dev, &q, properties, f, *qid); if (retval != 0) goto err_create_queue; pqn->q = q; @@ -271,7 +272,7 @@ int pqm_create_queue(struct process_queue_manager *pqm, goto err_create_queue; } - retval = create_cp_queue(pqm, dev, &q, properties, f, *qid); + retval = init_user_queue(pqm, dev, &q, properties, f, *qid); if (retval != 0) goto err_create_queue; pqn->q = q; @@ -298,17 +299,20 @@ int pqm_create_queue(struct process_queue_manager *pqm, } if (retval != 0) { - pr_err("Pasid %d DQM create queue %d failed. ret %d\n", + pr_err("Pasid 0x%x DQM create queue %d failed. ret %d\n", pqm->process->pasid, type, retval); goto err_create_queue; } - if (q) + if (q && p_doorbell_offset_in_process) /* Return the doorbell offset within the doorbell page * to the caller so it can be passed up to user mode * (in bytes). + * There are always 1024 doorbells per process, so in case + * of 8-byte doorbells, there are two doorbell pages per + * process. */ - properties->doorbell_off = + *p_doorbell_offset_in_process = (q->properties.doorbell_off * sizeof(uint32_t)) & (kfd_doorbell_process_slice(dev) - 1); @@ -370,14 +374,14 @@ int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid) /* destroy kernel queue (DIQ) */ dqm = pqn->kq->dev->dqm; dqm->ops.destroy_kernel_queue(dqm, pqn->kq, &pdd->qpd); - kernel_queue_uninit(pqn->kq); + kernel_queue_uninit(pqn->kq, false); } if (pqn->q) { dqm = pqn->q->device->dqm; retval = dqm->ops.destroy_queue(dqm, &pdd->qpd, pqn->q); if (retval) { - pr_err("Pasid %d destroy queue %d failed, ret %d\n", + pr_err("Pasid 0x%x destroy queue %d failed, ret %d\n", pqm->process->pasid, pqn->q->properties.queue_id, retval); if (retval != -ETIME) diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c index c2e6e47abaf2..203c823d65f1 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c @@ -269,6 +269,8 @@ static ssize_t iolink_show(struct kobject *kobj, struct attribute *attr, buffer[0] = 0; iolink = container_of(attr, struct kfd_iolink_properties, attr); + if (iolink->gpu && kfd_devcgroup_check_permission(iolink->gpu)) + return -EPERM; sysfs_show_32bit_prop(buffer, "type", iolink->iolink_type); sysfs_show_32bit_prop(buffer, "version_major", iolink->ver_maj); sysfs_show_32bit_prop(buffer, "version_minor", iolink->ver_min); @@ -305,6 +307,8 @@ static ssize_t mem_show(struct kobject *kobj, struct attribute *attr, buffer[0] = 0; mem = container_of(attr, struct kfd_mem_properties, attr); + if (mem->gpu && kfd_devcgroup_check_permission(mem->gpu)) + return -EPERM; sysfs_show_32bit_prop(buffer, "heap_type", mem->heap_type); sysfs_show_64bit_prop(buffer, "size_in_bytes", mem->size_in_bytes); sysfs_show_32bit_prop(buffer, "flags", mem->flags); @@ -334,6 +338,8 @@ static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr, buffer[0] = 0; cache = container_of(attr, struct kfd_cache_properties, attr); + if (cache->gpu && kfd_devcgroup_check_permission(cache->gpu)) + return -EPERM; sysfs_show_32bit_prop(buffer, "processor_id_low", cache->processor_id_low); sysfs_show_32bit_prop(buffer, "level", cache->cache_level); @@ -406,8 +412,6 @@ static ssize_t node_show(struct kobject *kobj, struct attribute *attr, char *buffer) { struct kfd_topology_device *dev; - char public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE]; - uint32_t i; uint32_t log_max_watch_addr; /* Making sure that the buffer is an empty string */ @@ -416,24 +420,24 @@ static ssize_t node_show(struct kobject *kobj, struct attribute *attr, if (strcmp(attr->name, "gpu_id") == 0) { dev = container_of(attr, struct kfd_topology_device, attr_gpuid); + if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu)) + return -EPERM; return sysfs_show_32bit_val(buffer, dev->gpu_id); } if (strcmp(attr->name, "name") == 0) { dev = container_of(attr, struct kfd_topology_device, attr_name); - for (i = 0; i < KFD_TOPOLOGY_PUBLIC_NAME_SIZE; i++) { - public_name[i] = - (char)dev->node_props.marketing_name[i]; - if (dev->node_props.marketing_name[i] == 0) - break; - } - public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE-1] = 0x0; - return sysfs_show_str_val(buffer, public_name); + + if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu)) + return -EPERM; + return sysfs_show_str_val(buffer, dev->node_props.name); } dev = container_of(attr, struct kfd_topology_device, attr_props); + if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu)) + return -EPERM; sysfs_show_32bit_prop(buffer, "cpu_cores_count", dev->node_props.cpu_cores_count); sysfs_show_32bit_prop(buffer, "simd_count", @@ -482,6 +486,10 @@ static ssize_t node_show(struct kobject *kobj, struct attribute *attr, dev->node_props.num_sdma_engines); sysfs_show_32bit_prop(buffer, "num_sdma_xgmi_engines", dev->node_props.num_sdma_xgmi_engines); + sysfs_show_32bit_prop(buffer, "num_sdma_queues_per_engine", + dev->node_props.num_sdma_queues_per_engine); + sysfs_show_32bit_prop(buffer, "num_cp_queues", + dev->node_props.num_cp_queues); if (dev->gpu) { log_max_watch_addr = @@ -1106,6 +1114,9 @@ static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu) { struct kfd_topology_device *dev; struct kfd_topology_device *out_dev = NULL; + struct kfd_mem_properties *mem; + struct kfd_cache_properties *cache; + struct kfd_iolink_properties *iolink; down_write(&topology_lock); list_for_each_entry(dev, &topology_device_list, list) { @@ -1119,6 +1130,13 @@ static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu) if (!dev->gpu && (dev->node_props.simd_count > 0)) { dev->gpu = gpu; out_dev = dev; + + list_for_each_entry(mem, &dev->mem_props, list) + mem->gpu = dev->gpu; + list_for_each_entry(cache, &dev->cache_props, list) + cache->gpu = dev->gpu; + list_for_each_entry(iolink, &dev->io_link_props, list) + iolink->gpu = dev->gpu; break; } } @@ -1274,6 +1292,10 @@ int kfd_topology_add_device(struct kfd_dev *gpu) */ amdgpu_amdkfd_get_cu_info(dev->gpu->kgd, &cu_info); + + strncpy(dev->node_props.name, gpu->device_info->asic_name, + KFD_TOPOLOGY_PUBLIC_NAME_SIZE); + dev->node_props.simd_arrays_per_engine = cu_info.num_shader_arrays_per_engine; @@ -1291,9 +1313,12 @@ int kfd_topology_add_device(struct kfd_dev *gpu) dev->node_props.num_sdma_engines = gpu->device_info->num_sdma_engines; dev->node_props.num_sdma_xgmi_engines = gpu->device_info->num_xgmi_sdma_engines; + dev->node_props.num_sdma_queues_per_engine = + gpu->device_info->num_sdma_queues_per_engine; dev->node_props.num_gws = (hws_gws_support && dev->gpu->dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) ? amdgpu_amdkfd_get_num_gws(dev->gpu->kgd) : 0; + dev->node_props.num_cp_queues = get_queues_num(dev->gpu->dqm); kfd_fill_mem_clk_max_info(dev); kfd_fill_iolink_non_crat_info(dev); @@ -1321,7 +1346,11 @@ int kfd_topology_add_device(struct kfd_dev *gpu) case CHIP_VEGA12: case CHIP_VEGA20: case CHIP_RAVEN: + case CHIP_RENOIR: + case CHIP_ARCTURUS: case CHIP_NAVI10: + case CHIP_NAVI12: + case CHIP_NAVI14: dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_2_0 << HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) & HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.h b/drivers/gpu/drm/amd/amdkfd/kfd_topology.h index 276354aa0fcc..74e9b1682af8 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.h @@ -27,7 +27,7 @@ #include <linux/list.h> #include "kfd_crat.h" -#define KFD_TOPOLOGY_PUBLIC_NAME_SIZE 128 +#define KFD_TOPOLOGY_PUBLIC_NAME_SIZE 32 #define HSA_CAP_HOT_PLUGGABLE 0x00000001 #define HSA_CAP_ATS_PRESENT 0x00000002 @@ -81,7 +81,9 @@ struct kfd_node_properties { int32_t drm_render_minor; uint32_t num_sdma_engines; uint32_t num_sdma_xgmi_engines; - uint16_t marketing_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE]; + uint32_t num_sdma_queues_per_engine; + uint32_t num_cp_queues; + char name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE]; }; #define HSA_MEM_HEAP_TYPE_SYSTEM 0 @@ -102,6 +104,7 @@ struct kfd_mem_properties { uint32_t flags; uint32_t width; uint32_t mem_clk_max; + struct kfd_dev *gpu; struct kobject *kobj; struct attribute attr; }; @@ -123,6 +126,7 @@ struct kfd_cache_properties { uint32_t cache_latency; uint32_t cache_type; uint8_t sibling_map[CRAT_SIBLINGMAP_SIZE]; + struct kfd_dev *gpu; struct kobject *kobj; struct attribute attr; }; @@ -141,6 +145,7 @@ struct kfd_iolink_properties { uint32_t max_bandwidth; uint32_t rec_transfer_size; uint32_t flags; + struct kfd_dev *gpu; struct kobject *kobj; struct attribute attr; }; |
