Add initial support for half precision builtins

v2: fix fmax implementation
    use consistent checks for __CLC_FP_SIZE
    add missing TODOs
    fix whitespace in definitions.h
v3: undef ZERO in modf.inc

Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
reviewer: Jeroen Ketema <j.ketema@xs4all.nl>
Reviewed-by: Aaron Watry <awatry@gmail.com>
Tested-by: Aaron Watry <awatry@gmail.com>
llvm-svn: 332677

22 files changed, 169 insertions, 20 deletions

diff --git a/libclc/generic/lib/math/acos.inc b/libclc/generic/lib/math/acos.inc
index cac9499260f..f5586ea0939 100644
--- a/libclc/generic/lib/math/acos.inc
+++ b/libclc/generic/lib/math/acos.inc
@@ -11,10 +11,15 @@
  * precision of #4 may be better.
  */
 
-#if __CLC_FPSIZE == 32
-#define __CLC_CONST(x) x ## f
-#else
+// TODO: Enable half precision when atan2 is implemented
+#if __CLC_FPSIZE > 16
+
+#if __CLC_FPSIZE == 64
 #define __CLC_CONST(x) x
+#elif __CLC_FPSIZE == 32
+#define __CLC_CONST(x) x ## f
+#elif __CLC_FPSIZE == 16
+#define __CLC_CONST(x) x ## h
 #endif
 
 _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE acos(__CLC_GENTYPE x) {
@@ -27,3 +32,5 @@ _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE acos(__CLC_GENTYPE x) {
 }
 
 #undef __CLC_CONST
+
+#endif
diff --git a/libclc/generic/lib/math/asin.inc b/libclc/generic/lib/math/asin.inc
index 4643cf816de..b08c7bd29ce 100644
--- a/libclc/generic/lib/math/asin.inc
+++ b/libclc/generic/lib/math/asin.inc
@@ -1,12 +1,18 @@
+// TODO: Enable half precision when atan2 is implemented
+#if __CLC_FPSIZE > 16
 
-#if __CLC_FPSIZE == 32
-#define __CLC_CONST(x) x ## f
-#else
+#if __CLC_FPSIZE == 64
 #define __CLC_CONST(x) x
+#elif __CLC_FPSIZE == 32
+#define __CLC_CONST(x) x ## f
+#elif __CLC_FPSIZE == 16
+#define __CLC_CONST(x) x ## h
 #endif
 
 _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE asin(__CLC_GENTYPE x) {
-  return atan2(x, sqrt( (__CLC_GENTYPE)__CLC_CONST(1.0) -(x*x) ));
+  return atan2(x, sqrt( (__CLC_GENTYPE)__CLC_CONST(1.0) - (x*x) ));
 }
 
 #undef __CLC_CONST
+
+#endif
diff --git a/libclc/generic/lib/math/clc_nextafter.cl b/libclc/generic/lib/math/clc_nextafter.cl
index d5c0af040ad..d32ef7079e1 100644
--- a/libclc/generic/lib/math/clc_nextafter.cl
+++ b/libclc/generic/lib/math/clc_nextafter.cl
@@ -37,3 +37,10 @@ _CLC_BINARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, float, __clc_nextafter, float, flo
 NEXTAFTER(double, ulong, long)
 _CLC_BINARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, double, __clc_nextafter, double, double)
 #endif
+
+#ifdef cl_khr_fp16
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+NEXTAFTER(half, ushort, short)
+_CLC_BINARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, half, __clc_nextafter, half, half)
+#endif
diff --git a/libclc/generic/lib/math/clc_sqrt_impl.inc b/libclc/generic/lib/math/clc_sqrt_impl.inc
index e97b5403db0..fe724e8c143 100644
--- a/libclc/generic/lib/math/clc_sqrt_impl.inc
+++ b/libclc/generic/lib/math/clc_sqrt_impl.inc
@@ -20,14 +20,15 @@
  * THE SOFTWARE.
  */
 
-#if __CLC_FPSIZE == 32
-#define __CLC_NAN NAN
-#define ZERO 0.0f
-#elif __CLC_FPSIZE == 64
+#if __CLC_FPSIZE == 64
 #define __CLC_NAN __builtin_nan("")
 #define ZERO 0.0
-#else
-#error "Invalid value for __CLC_FPSIZE"
+#elif __CLC_FPSIZE == 32
+#define __CLC_NAN NAN
+#define ZERO 0.0f
+#elif __CLC_FPSIZE == 16
+#define __CLC_NAN (half)NAN
+#define ZERO 0.0h
 #endif
 
 _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __clc_sqrt(__CLC_GENTYPE val) {
diff --git a/libclc/generic/lib/math/clc_sw_binary.inc b/libclc/generic/lib/math/clc_sw_binary.inc
index 0fe15244904..7741475c237 100644
--- a/libclc/generic/lib/math/clc_sw_binary.inc
+++ b/libclc/generic/lib/math/clc_sw_binary.inc
@@ -2,8 +2,11 @@
 
 #define __CLC_SW_FUNC(x) __CLC_CONCAT(__clc_, x)
 
+// TODO: Enable half precision when the sw routine is implemented
+#if __CLC_FPSIZE > 16
 _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __CLC_FUNC(__CLC_GENTYPE x, __CLC_GENTYPE y) {
   return __CLC_SW_FUNC(__CLC_FUNC)(x, y);
 }
+#endif
 
 #undef __CLC_SW_FUNC
diff --git a/libclc/generic/lib/math/clc_sw_unary.inc b/libclc/generic/lib/math/clc_sw_unary.inc
index 0f8467c0a1c..cd148b07a02 100644
--- a/libclc/generic/lib/math/clc_sw_unary.inc
+++ b/libclc/generic/lib/math/clc_sw_unary.inc
@@ -2,8 +2,11 @@
 
 #define __CLC_SW_FUNC(x) __CLC_CONCAT(__clc_, x)
 
+// TODO: Enable half precision when the sw routine is implemented
+#if __CLC_FPSIZE > 16
 _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __CLC_FUNC(__CLC_GENTYPE x) {
   return __CLC_SW_FUNC(__CLC_FUNC)(x);
 }
+#endif
 
 #undef __CLC_SW_FUNC
diff --git a/libclc/generic/lib/math/copysign.cl b/libclc/generic/lib/math/copysign.cl
index 4e0c51b0937..df65e9d7fec 100644
--- a/libclc/generic/lib/math/copysign.cl
+++ b/libclc/generic/lib/math/copysign.cl
@@ -10,3 +10,18 @@ _CLC_DEFINE_BINARY_BUILTIN(float, copysign, __builtin_copysignf, float, float)
 _CLC_DEFINE_BINARY_BUILTIN(double, copysign, __builtin_copysign, double, double)
 
 #endif
+
+#ifdef cl_khr_fp16
+
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+_CLC_DEF _CLC_OVERLOAD half copysign(half x, half y)
+{
+   ushort sign_x = as_ushort(x) & 0x8000u;
+   ushort unsigned_y = as_ushort(y) & 0x7ffffu;
+
+   return as_half((ushort)(sign_x | unsigned_y));
+}
+_CLC_BINARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, half, copysign, half, half)
+
+#endif
diff --git a/libclc/generic/lib/math/fmax.cl b/libclc/generic/lib/math/fmax.cl
index 239da3d3a61..5c269ceccdd 100644
--- a/libclc/generic/lib/math/fmax.cl
+++ b/libclc/generic/lib/math/fmax.cl
@@ -12,5 +12,21 @@ _CLC_DEFINE_BINARY_BUILTIN(double, fmax, __builtin_fmax, double, double);
 
 #endif
 
+#ifdef cl_khr_fp16
+
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+_CLC_DEF _CLC_OVERLOAD half fmax(half x, half y)
+{
+   if (isnan(x))
+      return y;
+   if (isnan(y))
+      return x;
+   return (x < y) ? y : x;
+}
+_CLC_BINARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, half, fmax, half, half)
+
+#endif
+
 #define __CLC_BODY <fmax.inc>
 #include <clc/math/gentype.inc>
diff --git a/libclc/generic/lib/math/fmax.inc b/libclc/generic/lib/math/fmax.inc
index 8315c5fb6e0..a91ad6b7ba7 100644
--- a/libclc/generic/lib/math/fmax.inc
+++ b/libclc/generic/lib/math/fmax.inc
@@ -15,4 +15,14 @@ _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE fmax(__CLC_GENTYPE x, double y) {
 
 #endif // ifdef cl_khr_fp64
 
+#ifdef cl_khr_fp16
+
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE fmax(__CLC_GENTYPE x, half y) {
+  return fmax(x, (__CLC_GENTYPE)((__CLC_SCALAR_GENTYPE)y));
+}
+
+#endif // ifdef cl_khr_fp16
+
 #endif // !defined(__CLC_SCALAR)
diff --git a/libclc/generic/lib/math/fmin.cl b/libclc/generic/lib/math/fmin.cl
index 28c7d0125a9..45c112d991f 100644
--- a/libclc/generic/lib/math/fmin.cl
+++ b/libclc/generic/lib/math/fmin.cl
@@ -11,6 +11,21 @@ _CLC_DEFINE_BINARY_BUILTIN(float, fmin, __builtin_fminf, float, float);
 _CLC_DEFINE_BINARY_BUILTIN(double, fmin, __builtin_fmin, double, double);
 
 #endif
+#ifdef cl_khr_fp16
+
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+_CLC_DEF _CLC_OVERLOAD half fmin(half x, half y)
+{
+   if (isnan(x))
+      return y;
+   if (isnan(y))
+      return x;
+   return (y < x) ? y : x;
+}
+_CLC_BINARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, half, fmin, half, half)
+
+#endif
 
 #define __CLC_BODY <fmin.inc>
 #include <clc/math/gentype.inc>
diff --git a/libclc/generic/lib/math/fmin.inc b/libclc/generic/lib/math/fmin.inc
index d4b5ac2d62a..98756785196 100644
--- a/libclc/generic/lib/math/fmin.inc
+++ b/libclc/generic/lib/math/fmin.inc
@@ -15,4 +15,14 @@ _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE fmin(__CLC_GENTYPE x, double y) {
 
 #endif // ifdef cl_khr_fp64
 
+#ifdef cl_khr_fp16
+
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE fmin(__CLC_GENTYPE x, half y) {
+  return fmin(x, (__CLC_GENTYPE)((__CLC_SCALAR_GENTYPE)y));
+}
+
+#endif // ifdef cl_khr_fp16
+
 #endif // !defined(__CLC_SCALAR)
diff --git a/libclc/generic/lib/math/fract.inc b/libclc/generic/lib/math/fract.inc
index 8d2a4d707a9..00d4674bfa2 100644
--- a/libclc/generic/lib/math/fract.inc
+++ b/libclc/generic/lib/math/fract.inc
@@ -20,16 +20,21 @@
  * THE SOFTWARE.
  */
 
-#if __CLC_FPSIZE == 32
-#define MIN_CONSTANT 0x1.fffffep-1f
-#else
+#if __CLC_FPSIZE == 64
 #define MIN_CONSTANT 0x1.fffffffffffffp-1
+#define ZERO 0.0
+#elif __CLC_FPSIZE == 32
+#define MIN_CONSTANT 0x1.fffffep-1f
+#define ZERO 0.0f
+#elif __CLC_FPSIZE == 16
+#define MIN_CONSTANT 0x1.ffcp-1h
+#define ZERO 0.0h
 #endif
 
 _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE fract(__CLC_GENTYPE x, private __CLC_GENTYPE *iptr) {
   *iptr = floor(x);
   __CLC_GENTYPE r = fmin(x - *iptr, MIN_CONSTANT);
-  r = isinf(x) ? 0.0f : r;
+  r = isinf(x) ? ZERO : r;
   r = isnan(x) ? x : r;
   return r;
 }
@@ -47,3 +52,4 @@ FRACT_DEF(local);
 FRACT_DEF(global);
 
 #undef MIN_CONSTANT
+#undef ZERO
diff --git a/libclc/generic/lib/math/ldexp.cl b/libclc/generic/lib/math/ldexp.cl
index 9be3127f46b..190a4d5f5fc 100644
--- a/libclc/generic/lib/math/ldexp.cl
+++ b/libclc/generic/lib/math/ldexp.cl
@@ -33,7 +33,13 @@ _CLC_DEFINE_BINARY_BUILTIN(float, ldexp, __clc_ldexp, float, int)
 #pragma OPENCL EXTENSION cl_khr_fp64 : enable
 
 _CLC_DEFINE_BINARY_BUILTIN(double, ldexp, __clc_ldexp, double, int)
+#endif
+
+#ifdef cl_khr_fp16
+
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
 
+_CLC_DEFINE_BINARY_BUILTIN(half, ldexp, __clc_ldexp, half, int)
 #endif
 
 // This defines all the ldexp(GENTYPE, int) variants
diff --git a/libclc/generic/lib/math/ldexp.inc b/libclc/generic/lib/math/ldexp.inc
index 6e28fbb94ca..d6144d7cb6d 100644
--- a/libclc/generic/lib/math/ldexp.inc
+++ b/libclc/generic/lib/math/ldexp.inc
@@ -20,6 +20,9 @@
  * THE SOFTWARE.
  */
 
+// TODO: Enable half precision when ldexp is implemented.
+#if __CLC_FPSIZE > 16
+
 #ifndef __CLC_SCALAR
 
 _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE ldexp(__CLC_GENTYPE x, int n) {
@@ -27,3 +30,5 @@ _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE ldexp(__CLC_GENTYPE x, int n) {
 }
 
 #endif
+
+#endif
diff --git a/libclc/generic/lib/math/lgamma_r.inc b/libclc/generic/lib/math/lgamma_r.inc
index 316d4fa1539..0e19ba8fb2c 100644
--- a/libclc/generic/lib/math/lgamma_r.inc
+++ b/libclc/generic/lib/math/lgamma_r.inc
@@ -21,10 +21,12 @@
  * THE SOFTWARE.
  */
 
-
+// TODO: Enable half precision when the base version is implemented.
+#if __CLC_FPSIZE > 16
 _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE lgamma_r(__CLC_GENTYPE x, __CLC_ADDRSPACE __CLC_INTN *iptr) {
     __CLC_INTN private_iptr;
     __CLC_GENTYPE ret = lgamma_r(x, &private_iptr);
     *iptr = private_iptr;
     return ret;
 }
+#endif
diff --git a/libclc/generic/lib/math/modf.inc b/libclc/generic/lib/math/modf.inc
index 1486b765bba..1ffc6d9e851 100644
--- a/libclc/generic/lib/math/modf.inc
+++ b/libclc/generic/lib/math/modf.inc
@@ -20,9 +20,17 @@
  * THE SOFTWARE.
  */
 
+#if __CLC_FPSIZE == 64
+#define ZERO 0.0
+#elif __CLC_FPSIZE == 32
+#define ZERO 0.0f
+#elif __CLC_FPSIZE == 16
+#define ZERO 0.0h
+#endif
+
 _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE modf(__CLC_GENTYPE x, __CLC_GENTYPE *iptr) {
   *iptr = trunc(x);
-  return copysign(isinf(x) ? 0.0f : x - *iptr, x);
+  return copysign(isinf(x) ? ZERO : x - *iptr, x);
 }
 
 #define MODF_DEF(addrspace) \
@@ -35,3 +43,5 @@ _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE modf(__CLC_GENTYPE x, __CLC_GENTYPE *iptr)
 
 MODF_DEF(local);
 MODF_DEF(global);
+
+#undef ZERO
diff --git a/libclc/generic/lib/math/nan.inc b/libclc/generic/lib/math/nan.inc
index 6e4afc888d5..f6508c3ab52 100644
--- a/libclc/generic/lib/math/nan.inc
+++ b/libclc/generic/lib/math/nan.inc
@@ -7,11 +7,18 @@ _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE nan(__CLC_XCONCAT(ulong, __CLC_VECSIZE) cod
 {
 	return __CLC_AS_GENTYPE(code | 0x7ff0000000000000ul);
 }
-#else
+#elif __CLC_FPSIZE == 32
 _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE nan(__CLC_XCONCAT(uint, __CLC_VECSIZE) code)
 {
 	return __CLC_AS_GENTYPE(code | 0x7fc00000);
 }
+#elif __CLC_FPSIZE == 16
+_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE nan(__CLC_XCONCAT(ushort, __CLC_VECSIZE) code)
+{
+	const ushort mask = 0x7e00;
+	const __CLC_XCONCAT(ushort, __CLC_VECSIZE) res = code | mask;
+	return __CLC_AS_GENTYPE(res);
+}
 #endif
 
 
diff --git a/libclc/generic/lib/math/pown.inc b/libclc/generic/lib/math/pown.inc
index 84729d90a79..2add2c7459d 100644
--- a/libclc/generic/lib/math/pown.inc
+++ b/libclc/generic/lib/math/pown.inc
@@ -1,3 +1,6 @@
+// TODO: Enable half precision when the sw routine is implemented
+#if __CLC_FPSIZE > 16
 _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE pown(__CLC_GENTYPE x, __CLC_INTN y) {
   return __clc_pown(x, y);
 }
+#endif
diff --git a/libclc/generic/lib/math/remquo.inc b/libclc/generic/lib/math/remquo.inc
index c1de78a5e7f..c33b5ddab31 100644
--- a/libclc/generic/lib/math/remquo.inc
+++ b/libclc/generic/lib/math/remquo.inc
@@ -1,6 +1,9 @@
+// TODO: Enable half precision when the sw routine is implemented
+#if __CLC_FPSIZE > 16
 _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE remquo(__CLC_GENTYPE x, __CLC_GENTYPE y, __CLC_ADDRESS_SPACE __CLC_INTN *q) {
   __CLC_INTN local_q;
   __CLC_GENTYPE ret = __clc_remquo(x, y, &local_q);
   *q = local_q;
   return ret;
 }
+#endif
diff --git a/libclc/generic/lib/math/rootn.inc b/libclc/generic/lib/math/rootn.inc
index 3f5b00c082c..f788649685a 100644
--- a/libclc/generic/lib/math/rootn.inc
+++ b/libclc/generic/lib/math/rootn.inc
@@ -1,3 +1,6 @@
+// TODO: Enable half precision when the sw routine is implemented
+#if __CLC_FPSIZE > 16
 _CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE rootn(__CLC_GENTYPE x, __CLC_INTN y) {
   return __clc_rootn(x, y);
 }
+#endif
diff --git a/libclc/generic/lib/math/sincos.inc b/libclc/generic/lib/math/sincos.inc
index e97f0f9641c..2318ffb73f5 100644
--- a/libclc/generic/lib/math/sincos.inc
+++ b/libclc/generic/lib/math/sincos.inc
@@ -1,3 +1,5 @@
+// TODO: Enable half precision when sin/cos is implemented
+#if __CLC_FPSIZE > 16
 #define __CLC_DECLARE_SINCOS(ADDRSPACE, TYPE) \
   _CLC_OVERLOAD _CLC_DEF TYPE sincos (TYPE x, ADDRSPACE TYPE * cosval) { \
     *cosval = cos(x); \
@@ -9,3 +11,4 @@ __CLC_DECLARE_SINCOS(local, __CLC_GENTYPE)
 __CLC_DECLARE_SINCOS(private, __CLC_GENTYPE)
 
 #undef __CLC_DECLARE_SINCOS
+#endif
diff --git a/libclc/generic/lib/math/sqrt.cl b/libclc/generic/lib/math/sqrt.cl
index 300e2741ee0..b05d6bc0a06 100644
--- a/libclc/generic/lib/math/sqrt.cl
+++ b/libclc/generic/lib/math/sqrt.cl
@@ -33,3 +33,11 @@ _CLC_DEFINE_UNARY_BUILTIN(float, sqrt, __clc_sqrt, float)
 _CLC_DEFINE_UNARY_BUILTIN(double, sqrt, __clc_sqrt, double)
 
 #endif
+
+#ifdef cl_khr_fp16
+
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+_CLC_DEFINE_UNARY_BUILTIN(half, sqrt, __clc_sqrt, half)
+
+#endif