Commit: da3fdf0b4bcded2a9473279078602ac24369472a
Author: Sv. Lockal
Date: Sat Jan 11 22:20:03 2014 +0400
https://developer.blender.org/rBda3fdf0b4bcded2a9473279078602ac24369472a
Code Cleanup: in Cycles SSE replace macros with templates, skip unused code
with preprocessor, simplify casts
===================================================================
M intern/cycles/kernel/kernel_bvh_subsurface.h
M intern/cycles/kernel/kernel_bvh_traversal.h
M intern/cycles/kernel/svm/svm_noise.h
M intern/cycles/util/util_simd.h
===================================================================
diff --git a/intern/cycles/kernel/kernel_bvh_subsurface.h
b/intern/cycles/kernel/kernel_bvh_subsurface.h
index fb41bdc..bb51986 100644
--- a/intern/cycles/kernel/kernel_bvh_subsurface.h
+++ b/intern/cycles/kernel/kernel_bvh_subsurface.h
@@ -64,16 +64,16 @@ ccl_device uint BVH_FUNCTION_NAME(KernelGlobals *kg, const
Ray *ray, Intersectio
const shuffle_swap_t shuf_identity = shuffle_swap_identity();
const shuffle_swap_t shuf_swap = shuffle_swap_swap();
- const __m128i pn = _mm_set_epi32(0x80000000, 0x80000000, 0x00000000,
0x00000000);
+ const __m128 pn = _mm_castsi128_ps(_mm_set_epi32(0x80000000,
0x80000000, 0, 0));
__m128 Psplat[3], idirsplat[3];
Psplat[0] = _mm_set_ps1(P.x);
Psplat[1] = _mm_set_ps1(P.y);
Psplat[2] = _mm_set_ps1(P.z);
- idirsplat[0] = _mm_xor_ps(_mm_set_ps1(idir.x), _mm_castsi128_ps(pn));
- idirsplat[1] = _mm_xor_ps(_mm_set_ps1(idir.y), _mm_castsi128_ps(pn));
- idirsplat[2] = _mm_xor_ps(_mm_set_ps1(idir.z), _mm_castsi128_ps(pn));
+ idirsplat[0] = _mm_xor_ps(_mm_set_ps1(idir.x), pn);
+ idirsplat[1] = _mm_xor_ps(_mm_set_ps1(idir.y), pn);
+ idirsplat[2] = _mm_xor_ps(_mm_set_ps1(idir.z), pn);
__m128 tsplat = _mm_set_ps(-tmax, -tmax, 0.0f, 0.0f);
@@ -143,8 +143,8 @@ ccl_device uint BVH_FUNCTION_NAME(KernelGlobals *kg, const
Ray *ray, Intersectio
const __m128 tminmaxy =
_mm_mul_ps(_mm_sub_ps(shuffle_swap(bvh_nodes[1], shuffley), Psplat[1]),
idirsplat[1]);
const __m128 tminmaxz =
_mm_mul_ps(_mm_sub_ps(shuffle_swap(bvh_nodes[2], shufflez), Psplat[2]),
idirsplat[2]);
- const __m128 tminmax =
_mm_xor_ps(_mm_max_ps(_mm_max_ps(tminmaxx, tminmaxy), _mm_max_ps(tminmaxz,
tsplat)), _mm_castsi128_ps(pn));
- const __m128 lrhit = _mm_cmple_ps(tminmax,
shuffle_swap(tminmax, shuf_swap));
+ const __m128 tminmax =
_mm_xor_ps(_mm_max_ps(_mm_max_ps(tminmaxx, tminmaxy), _mm_max_ps(tminmaxz,
tsplat)), pn);
+ const __m128 lrhit = _mm_cmple_ps(tminmax,
shuffle<2, 3, 0, 1>(tminmax));
/* decide which nodes to traverse next */
#ifdef __VISIBILITY_FLAG__
@@ -242,9 +242,9 @@ ccl_device uint BVH_FUNCTION_NAME(KernelGlobals *kg, const
Ray *ray, Intersectio
Psplat[1] = _mm_set_ps1(P.y);
Psplat[2] = _mm_set_ps1(P.z);
- idirsplat[0] =
_mm_xor_ps(_mm_set_ps1(idir.x), _mm_castsi128_ps(pn));
- idirsplat[1] =
_mm_xor_ps(_mm_set_ps1(idir.y), _mm_castsi128_ps(pn));
- idirsplat[2] =
_mm_xor_ps(_mm_set_ps1(idir.z), _mm_castsi128_ps(pn));
+ idirsplat[0] =
_mm_xor_ps(_mm_set_ps1(idir.x), pn);
+ idirsplat[1] =
_mm_xor_ps(_mm_set_ps1(idir.y), pn);
+ idirsplat[2] =
_mm_xor_ps(_mm_set_ps1(idir.z), pn);
tsplat = _mm_set_ps(-tmax,
-tmax, 0.0f, 0.0f);
@@ -285,9 +285,9 @@ ccl_device uint BVH_FUNCTION_NAME(KernelGlobals *kg, const
Ray *ray, Intersectio
Psplat[1] = _mm_set_ps1(P.y);
Psplat[2] = _mm_set_ps1(P.z);
- idirsplat[0] = _mm_xor_ps(_mm_set_ps1(idir.x),
_mm_castsi128_ps(pn));
- idirsplat[1] = _mm_xor_ps(_mm_set_ps1(idir.y),
_mm_castsi128_ps(pn));
- idirsplat[2] = _mm_xor_ps(_mm_set_ps1(idir.z),
_mm_castsi128_ps(pn));
+ idirsplat[0] = _mm_xor_ps(_mm_set_ps1(idir.x), pn);
+ idirsplat[1] = _mm_xor_ps(_mm_set_ps1(idir.y), pn);
+ idirsplat[2] = _mm_xor_ps(_mm_set_ps1(idir.z), pn);
tsplat = _mm_set_ps(-tmax, -tmax, 0.0f, 0.0f);
diff --git a/intern/cycles/kernel/kernel_bvh_traversal.h
b/intern/cycles/kernel/kernel_bvh_traversal.h
index a78e72e..24f5568 100644
--- a/intern/cycles/kernel/kernel_bvh_traversal.h
+++ b/intern/cycles/kernel/kernel_bvh_traversal.h
@@ -73,16 +73,16 @@ ccl_device bool BVH_FUNCTION_NAME
const shuffle_swap_t shuf_identity = shuffle_swap_identity();
const shuffle_swap_t shuf_swap = shuffle_swap_swap();
- const __m128i pn = _mm_set_epi32(0x80000000, 0x80000000, 0x00000000,
0x00000000);
+ const __m128 pn = _mm_castsi128_ps(_mm_set_epi32(0x80000000,
0x80000000, 0, 0));
__m128 Psplat[3], idirsplat[3];
Psplat[0] = _mm_set_ps1(P.x);
Psplat[1] = _mm_set_ps1(P.y);
Psplat[2] = _mm_set_ps1(P.z);
- idirsplat[0] = _mm_xor_ps(_mm_set_ps1(idir.x), _mm_castsi128_ps(pn));
- idirsplat[1] = _mm_xor_ps(_mm_set_ps1(idir.y), _mm_castsi128_ps(pn));
- idirsplat[2] = _mm_xor_ps(_mm_set_ps1(idir.z), _mm_castsi128_ps(pn));
+ idirsplat[0] = _mm_xor_ps(_mm_set_ps1(idir.x), pn);
+ idirsplat[1] = _mm_xor_ps(_mm_set_ps1(idir.y), pn);
+ idirsplat[2] = _mm_xor_ps(_mm_set_ps1(idir.z), pn);
__m128 tsplat = _mm_set_ps(-isect->t, -isect->t, 0.0f, 0.0f);
@@ -169,7 +169,7 @@ ccl_device bool BVH_FUNCTION_NAME
/* calculate { c0min, c1min, -c0max, -c1max} */
__m128 minmax = _mm_max_ps(_mm_max_ps(tminmaxx,
tminmaxy), _mm_max_ps(tminmaxz, tsplat));
- const __m128 tminmax = _mm_xor_ps(minmax,
_mm_castsi128_ps(pn));
+ const __m128 tminmax = _mm_xor_ps(minmax, pn);
#if FEATURE(BVH_HAIR_MINIMUM_WIDTH)
if(difl != 0.0f) {
@@ -189,7 +189,7 @@ ccl_device bool BVH_FUNCTION_NAME
}
}
#endif
- const __m128 lrhit = _mm_cmple_ps(tminmax,
shuffle_swap(tminmax, shuf_swap));
+ const __m128 lrhit = _mm_cmple_ps(tminmax,
shuffle<2, 3, 0, 1>(tminmax));
/* decide which nodes to traverse next */
#ifdef __VISIBILITY_FLAG__
@@ -307,9 +307,9 @@ ccl_device bool BVH_FUNCTION_NAME
Psplat[1] = _mm_set_ps1(P.y);
Psplat[2] = _mm_set_ps1(P.z);
- idirsplat[0] =
_mm_xor_ps(_mm_set_ps1(idir.x), _mm_castsi128_ps(pn));
- idirsplat[1] =
_mm_xor_ps(_mm_set_ps1(idir.y), _mm_castsi128_ps(pn));
- idirsplat[2] =
_mm_xor_ps(_mm_set_ps1(idir.z), _mm_castsi128_ps(pn));
+ idirsplat[0] =
_mm_xor_ps(_mm_set_ps1(idir.x), pn);
+ idirsplat[1] =
_mm_xor_ps(_mm_set_ps1(idir.y), pn);
+ idirsplat[2] =
_mm_xor_ps(_mm_set_ps1(idir.z), pn);
tsplat = _mm_set_ps(-isect->t,
-isect->t, 0.0f, 0.0f);
@@ -343,9 +343,9 @@ ccl_device bool BVH_FUNCTION_NAME
Psplat[1] = _mm_set_ps1(P.y);
Psplat[2] = _mm_set_ps1(P.z);
- idirsplat[0] = _mm_xor_ps(_mm_set_ps1(idir.x),
_mm_castsi128_ps(pn));
- idirsplat[1] = _mm_xor_ps(_mm_set_ps1(idir.y),
_mm_castsi128_ps(pn));
- idirsplat[2] = _mm_xor_ps(_mm_set_ps1(idir.z),
_mm_castsi128_ps(pn));
+ idirsplat[0] = _mm_xor_ps(_mm_set_ps1(idir.x), pn);
+ idirsplat[1] = _mm_xor_ps(_mm_set_ps1(idir.y), pn);
+ idirsplat[2] = _mm_xor_ps(_mm_set_ps1(idir.z), pn);
tsplat = _mm_set_ps(-isect->t, -isect->t, 0.0f, 0.0f);
diff --git a/intern/cycles/kernel/svm/svm_noise.h
b/intern/cycles/kernel/svm/svm_noise.h
index a58dfdf..6ad10b7 100644
--- a/intern/cycles/kernel/svm/svm_noise.h
+++ b/intern/cycles/kernel/svm/svm_noise.h
@@ -32,21 +32,17 @@
CCL_NAMESPACE_BEGIN
-#if defined(__KERNEL_SSE2__)
-#define FMA(a, b, c) _mm_add_ps(_mm_mul_ps((a), (b)), (c))
-#endif
-
ccl_device int quick_floor(float x)
{
return float_to_int(x) - ((x < 0) ? 1 : 0);
}
-#if defined(__KERNEL_SSE2__)
+#ifdef __KERNEL_SSE2__
ccl_device_inline __m128i quick_floor_sse(const __m128 *x)
{
- __m128i b = _mm_cvttps_epi32(*x);
- __m128i isneg = _mm_castps_si128(_mm_cmplt_ps(*x, _mm_set1_ps(0.0f)));
- return _mm_add_epi32(b, isneg); // unsaturated add 0xffffffff is the same
as subtract -1
+ __m128i b = _mm_cvttps_epi32(*x);
+ __m128i isneg = _mm_castps_si128(_mm_cmplt_ps(*x, _mm_set1_ps(0.0f)));
+ return _mm_add_epi32(b, isneg); // unsaturated add 0xffffffff is the
same as subtract -1
}
#endif
@@ -84,32 +80,33 @@ ccl_device uint hash(uint kx, uint ky, uint kz)
#undef final
}
-#if defined(__KERNEL_SSE2__)
+#ifdef __KERNEL_SSE2__
ccl_device_inline __m128i hash_sse(const __m128i *kx, const __m128i *ky, const
__m128i *kz)
{
#define rot(x,k) _mm_or_si128(_mm_slli_epi32((x), (k)), _mm_srli_epi32((x),
32-(k)))
#define xor_rot(a, b, c) do {a = _mm_xor_si128(a, b); a = _mm_sub_epi32(a,
rot(b, c));} while(0)
- uint len = 3;
- __m128i magic = _mm_set1_epi32(0xdeadbeef + (len << 2) + 13);
- __m128i a = _mm_add_epi32(magic, *kx);
- __m128i b = _mm_add_epi32(magic, *ky);
- __m128i c = _mm_add_epi32(magic, *kz);
-
- xor_rot(c, b, 14);
- xor_rot(a, c, 11);
- xor_rot(b, a, 25);
- xor_rot(c, b, 16);
- xor_rot(a, c, 4);
- xor_rot(b, a, 14);
- xor_rot(c, b, 24);
-
- return c;
+ uint len = 3;
+ __m128i magic = _mm_set1_epi32(0xdeadbeef + (len << 2) + 13);
+ __m128i a = _mm_add_epi32(magic, *kx);
+ __m128i b = _mm_add_epi32(magic, *ky);
+ __m128i c = _mm_add_epi32(magic, *kz);
+
+ xor_rot(c, b, 14);
+ xor_rot(a, c, 11);
+ xor_rot(b, a, 25);
+ xor_rot(c, b, 16);
+ xor_rot(a, c, 4);
+ xor_rot(b, a, 14);
+ xor_rot(c, b, 24);
+
+ return c;
#undef rot
#undef xor_rot
}
#endif
+#if 0 // unused
ccl_device int imod(int a, int b)
{
a %= b;
@@ -120,48 +117,50 @@ ccl_device uint phash(int kx, int ky, int kz, int3 p)
{
return hash(imod(kx, p.x), imod(ky, p.y), imod(kz, p.z));
}
+#endif
+#ifndef __KERNEL_SSE2__
ccl_device float floorfrac(float x, int* i)
{
*i = quick_floor(x);
return x - *i;
}
-
-#if defined(__KERNEL_SSE2__)
+#else
ccl_device_inline __m128 floorfrac_sse(const __m128 *x, __m128i *i)
{
- *i = quick_floor_sse(x);
- return _mm_sub_ps(*x, _mm_cvtepi32_ps(*i));
+ *i = quick_floor_sse(x);
+ return _mm_sub_ps(*x, _mm_cvtepi32_ps(*i));
}
#endif
+#ifndef __KERNEL_SSE2__
ccl_device float fade(float t)
{
return t * t * t * (t * (t * 6.0f - 15.0f) + 10.0f);
}
-
-#if defined(__KERNEL_SSE2__)
+#else
ccl_device_inline __m128 fade_sse(const __m128 *t)
{
- __m128 a = FMA(*t, _mm_set1_ps(6.0f), _mm_set1_ps(-15.0f));
- __m128 b = FMA(*t, a, _mm_set1_ps(10.0f));
- return _mm_mul_ps(_mm_mul_ps(*t, *t), _mm_mul_ps(*t, b));
+ __m128 a = fma(*t, _mm_set1_ps(6.0f), _mm_set1_ps(-15.0f));
+ __m128 b = fma(*t, a, _mm_set1_ps(10.0f));
+ return _mm_mul_ps(_mm_mul_ps(*t, *t), _mm_mul_ps(*t, b));
}
#endif
+#ifndef __KERNEL_SSE2__
ccl_device float nerp(float t, float a, float b)
{
return (1.0f - t) * a + t * b;
}
-
-#if defined(__KERNEL_SSE2__)
+#else
ccl_device_inline __m128 nerp_sse(const __m128 *t, const __m128 *a, const
__m128 *b)
{
- __m128 x1 = _mm_mul_ps(_mm_sub_ps(_mm_set1_ps(1.0f), *t), *a);
- return FMA(*t, *b, x1);
+ __m128 x1 = _mm_mul_ps(_mm_sub_ps(_mm_set1_ps(1.0f), *t), *a);
+ return fma(*t, *b, x1);
}
#endif
+#ifndef __KERNEL_SSE2__
ccl_device float grad(int hash, float x, float y, float z)
{
@@ Diff output truncated at 10240 characters. @@
_______________________________________________
Bf-blender-cvs mailing list
[email protected]
http://lists.blender.org/mailman/listinfo/bf-blender-cvs
