tan: Port from amd_builtins

v2: fixup constant precision
Passes piglit on turks and carrizo.
Passes CTS on carrizo
Fixes half_tan to pass CTS on carrizo

Acked-By: Aaron Watry <awatry@gmail.com>
Tested-By: Aaron Watry <awatry@gmail.com>
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 322979

9 files changed, 164 insertions, 21 deletions

diff --git a/libclc/generic/include/math/clc_tan.h b/libclc/generic/include/math/clc_tan.h
new file mode 100644
index 00000000000..bc97902965d
--- /dev/null
+++ b/libclc/generic/include/math/clc_tan.h
@@ -0,0 +1,5 @@
+#define __CLC_FUNCTION __clc_tan
+#define __CLC_BODY <clc/math/unary_decl.inc>
+#include <clc/math/gentype.inc>
+#undef __CLC_BODY
+#undef __CLC_FUNCTION
diff --git a/libclc/generic/lib/SOURCES b/libclc/generic/lib/SOURCES
index 9dbd35e4f76..4546a9ef610 100644
--- a/libclc/generic/lib/SOURCES
+++ b/libclc/generic/lib/SOURCES
@@ -164,6 +164,7 @@ math/sinh.cl
 math/sinpi.cl
 math/clc_sqrt.cl
 math/sqrt.cl
+math/clc_tan.cl
 math/tan.cl
 math/tanh.cl
 math/tgamma.cl
diff --git a/libclc/generic/lib/math/clc_sw_unary.inc b/libclc/generic/lib/math/clc_sw_unary.inc
new file mode 100644
index 00000000000..0f8467c0a1c
--- /dev/null
+++ b/libclc/generic/lib/math/clc_sw_unary.inc
@@ -0,0 +1,9 @@
+#include <utils.h>
+
+#define __CLC_SW_FUNC(x) __CLC_CONCAT(__clc_, x)
+
+_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __CLC_FUNC(__CLC_GENTYPE x) {
+  return __CLC_SW_FUNC(__CLC_FUNC)(x);
+}
+
+#undef __CLC_SW_FUNC
diff --git a/libclc/generic/lib/math/clc_tan.cl b/libclc/generic/lib/math/clc_tan.cl
new file mode 100644
index 00000000000..ebba36a0d25
--- /dev/null
+++ b/libclc/generic/lib/math/clc_tan.cl
@@ -0,0 +1,71 @@
+/*
+ * Copyright (c) 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
+ * AUTHORS OR COPYRIGHT HOLDERS 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 <clc/clc.h>
+
+#include "math.h"
+#include "sincos_helpers.h"
+#include "../clcmacro.h"
+#include "tables.h"
+
+_CLC_DEF _CLC_OVERLOAD float __clc_tan(float x)
+{
+    int ix = as_int(x);
+    int ax = ix & 0x7fffffff;
+    float dx = as_float(ax);
+
+    float r0, r1;
+    int regn = __clc_argReductionS(&r0, &r1, dx);
+
+    float t = __clc_tanf_piby4(r0 + r1, regn);
+    t = as_float(as_int(t) ^ (ix ^ ax));
+
+    t = ax >= PINFBITPATT_SP32 ? as_float(QNANBITPATT_SP32) : t;
+    //Take care of subnormals
+    t = (x == 0.0f) ? x : t;
+    return t;
+}
+_CLC_UNARY_VECTORIZE(_CLC_DEF _CLC_OVERLOAD, float, __clc_tan, float);
+
+#ifdef cl_khr_fp64
+#include "sincosD_piby4.h"
+
+_CLC_DEF _CLC_OVERLOAD double __clc_tan(double x)
+{
+    double y = fabs(x);
+
+    double r, rr;
+    int regn;
+
+    if (y < 0x1.0p+30)
+        __clc_remainder_piby2_medium(y, &r, &rr, &regn);
+    else
+        __clc_remainder_piby2_large(y, &r, &rr, &regn);
+
+    double2 tt = __clc_tan_piby4(r, rr);
+
+    int2 t = as_int2(regn & 1 ? tt.y : tt.x);
+    t.hi ^= (x < 0.0) << 31;
+
+    return isnan(x) || isinf(x) ? as_double(QNANBITPATT_DP64) : as_double(t);
+}
+_CLC_UNARY_VECTORIZE(_CLC_DEF _CLC_OVERLOAD, double, __clc_tan, double);
+#endif
diff --git a/libclc/generic/lib/math/sincosD_piby4.h b/libclc/generic/lib/math/sincosD_piby4.h
index a00db854891..c98488b33ed 100644
--- a/libclc/generic/lib/math/sincosD_piby4.h
+++ b/libclc/generic/lib/math/sincosD_piby4.h
@@ -76,3 +76,58 @@ __libclc__sincos_piby4(double x, double xx)
 
     return ret;
 }
+
+_CLC_INLINE double2
+__clc_tan_piby4(double x, double xx)
+{
+    const double piby4_lead = 7.85398163397448278999e-01; // 0x3fe921fb54442d18
+    const double piby4_tail = 3.06161699786838240164e-17; // 0x3c81a62633145c06
+
+    // In order to maintain relative precision transform using the identity:
+    // tan(pi/4-x) = (1-tan(x))/(1+tan(x)) for arguments close to pi/4.
+    // Similarly use tan(x-pi/4) = (tan(x)-1)/(tan(x)+1) close to -pi/4.
+
+    int ca = x >  0.68;
+    int cb = x < -0.68;
+    double transform = ca ?  1.0 : 0.0;
+    transform = cb ? -1.0 : transform;
+
+    double tx = fma(-transform, x, piby4_lead) + fma(-transform, xx, piby4_tail);
+    int c = ca | cb;
+    x = c ? tx : x;
+    xx = c ? 0.0 : xx;
+
+    // Core Remez [2,3] approximation to tan(x+xx) on the interval [0,0.68].
+    double t1 = x;
+    double r = fma(2.0, x*xx, x*x);
+
+    double a = fma(r,
+                   fma(r, 0.224044448537022097264602535574e-3, -0.229345080057565662883358588111e-1),
+                   0.372379159759792203640806338901e0);
+
+    double b = fma(r,
+                   fma(r,
+                       fma(r, -0.232371494088563558304549252913e-3, 0.260656620398645407524064091208e-1),
+                       -0.515658515729031149329237816945e0),
+                   0.111713747927937668539901657944e1);
+
+    double t2 = fma(MATH_DIVIDE(a, b), x*r, xx);
+
+    double tp = t1 + t2;
+
+    // Compute -1.0/(t1 + t2) accurately
+    double z1 = as_double(as_long(tp) & 0xffffffff00000000L);
+    double z2 = t2 - (z1 - t1);
+    double trec = -MATH_RECIP(tp);
+    double trec_top = as_double(as_long(trec) & 0xffffffff00000000L);
+
+    double tpr = fma(fma(trec_top, z2, fma(trec_top, z1, 1.0)), trec, trec_top);
+
+    double tpt = transform * (1.0 - MATH_DIVIDE(2.0*tp, 1.0 + tp));
+    double tptr = transform * (MATH_DIVIDE(2.0*tp, tp - 1.0) - 1.0);
+
+    double2 ret;
+    ret.lo = c ? tpt : tp;
+    ret.hi = c ? tptr : tpr;
+    return ret;
+}
diff --git a/libclc/generic/lib/math/sincos_helpers.cl b/libclc/generic/lib/math/sincos_helpers.cl
index 251b7f94efd..3c466bcf9f8 100644
--- a/libclc/generic/lib/math/sincos_helpers.cl
+++ b/libclc/generic/lib/math/sincos_helpers.cl
@@ -90,6 +90,23 @@ _CLC_DEF float __clc_cosf_piby4(float x, float y) {
     return ret;
 }
 
+_CLC_DEF float __clc_tanf_piby4(float x, int regn)
+{
+    // Core Remez [1,2] approximation to tan(x) on the interval [0,pi/4].
+    float r = x * x;
+
+    float a = mad(r, -0.0172032480471481694693109f, 0.385296071263995406715129f);
+
+    float b = mad(r,
+	          mad(r, 0.01844239256901656082986661f, -0.51396505478854532132342f),
+	          1.15588821434688393452299f);
+
+    float t = mad(x*r, native_divide(a, b), x);
+    float tr = -MATH_RECIP(t);
+
+    return regn & 1 ? tr : t;
+}
+
 _CLC_DEF void __clc_fullMulS(float *hi, float *lo, float a, float b, float bh, float bt)
 {
     if (HAVE_HW_FMA32()) {
diff --git a/libclc/generic/lib/math/sincos_helpers.h b/libclc/generic/lib/math/sincos_helpers.h
index 2565d44eb02..e307abc48b2 100644
--- a/libclc/generic/lib/math/sincos_helpers.h
+++ b/libclc/generic/lib/math/sincos_helpers.h
@@ -20,8 +20,11 @@
  * THE SOFTWARE.
  */
 
+#include "clc/clcfunc.h"
+
 _CLC_DECL float __clc_sinf_piby4(float x, float y);
 _CLC_DECL float __clc_cosf_piby4(float x, float y);
+_CLC_DECL float __clc_tanf_piby4(float x, int y);
 _CLC_DECL int __clc_argReductionS(float *r, float *rr, float x);
 
 #ifdef cl_khr_fp64
diff --git a/libclc/generic/lib/math/tan.cl b/libclc/generic/lib/math/tan.cl
index a447999ea8b..380db67e364 100644
--- a/libclc/generic/lib/math/tan.cl
+++ b/libclc/generic/lib/math/tan.cl
@@ -1,8 +1,7 @@
 #include <clc/clc.h>
 
-#ifdef cl_khr_fp64
-#pragma OPENCL EXTENSION cl_khr_fp64 : enable
-#endif
+#include <math/clc_tan.h>
 
-#define __CLC_BODY <tan.inc>
+#define __CLC_FUNC tan
+#define __CLC_BODY <clc_sw_unary.inc>
 #include <clc/math/gentype.inc>
diff --git a/libclc/generic/lib/math/tan.inc b/libclc/generic/lib/math/tan.inc
deleted file mode 100644
index b9ce33ef823..00000000000
--- a/libclc/generic/lib/math/tan.inc
+++ /dev/null
@@ -1,17 +0,0 @@
-/*
- * Note: tan(x) = sin(x)/cos(x) also, but the final assembly ends up being
- *       twice as long for R600 (maybe for others as well).
- */
-
-#if __CLC_FPSIZE == 32
-#define __CLC_CONST(x) x ## f
-#else
-#define __CLC_CONST(x) x
-#endif
-
-_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE tan(__CLC_GENTYPE x) {
-  __CLC_GENTYPE sinx = sin(x);
-  return sinx / sqrt( (__CLC_GENTYPE) __CLC_CONST(1.0) - (sinx*sinx) );
-}
-
-#undef __CLC_CONST