[Bf-blender-cvs] [a92abf5] master: Cycles / Perlin Noise: Optimize noise calculation by using SIMD instructions on CPU. This makes scenes with a Noise Texture render faster, the BMW file is 12-15% faster now.
Thomas Dinges
noreply at git.blender.org
Fri Dec 27 18:50:52 CET 2013
Commit: a92abf5089e152d8d1b1fd95278b307f56b6a193
Author: Thomas Dinges
Date: Fri Dec 27 18:48:37 2013 +0100
https://developer.blender.org/rBa92abf5089e152d8d1b1fd95278b307f56b6a193
Cycles / Perlin Noise: Optimize noise calculation by using SIMD instructions on CPU.
This makes scenes with a Noise Texture render faster, the BMW file is 12-15% faster now.
Patch by Sv. Lockal, many thanks! :)
===================================================================
M intern/cycles/kernel/svm/svm_noise.h
===================================================================
diff --git a/intern/cycles/kernel/svm/svm_noise.h b/intern/cycles/kernel/svm/svm_noise.h
index 2055b0e..12b3445 100644
--- a/intern/cycles/kernel/svm/svm_noise.h
+++ b/intern/cycles/kernel/svm/svm_noise.h
@@ -32,11 +32,24 @@
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__)
+ccl_device __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
+}
+#endif
+
ccl_device float bits_to_01(uint bits)
{
return bits * (1.0f/(float)0xFFFFFFFF);
@@ -71,6 +84,32 @@ ccl_device uint hash(uint kx, uint ky, uint kz)
#undef final
}
+#if defined(__KERNEL_SSE2__)
+ccl_device __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;
+#undef rot
+#undef xor_rot
+}
+#endif
+
ccl_device int imod(int a, int b)
{
a %= b;
@@ -88,16 +127,41 @@ ccl_device float floorfrac(float x, int* i)
return x - *i;
}
+#if defined(__KERNEL_SSE2__)
+ccl_device __m128 floorfrac_sse(const __m128 *x, __m128i *i)
+{
+ *i = quick_floor_sse(x);
+ return _mm_sub_ps(*x, _mm_cvtepi32_ps(*i));
+}
+#endif
+
ccl_device float fade(float t)
{
return t * t * t * (t * (t * 6.0f - 15.0f) + 10.0f);
}
+#if defined(__KERNEL_SSE2__)
+ccl_device __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));
+}
+#endif
+
ccl_device float nerp(float t, float a, float b)
{
return (1.0f - t) * a + t * b;
}
+#if defined(__KERNEL_SSE2__)
+ccl_device __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);
+}
+#endif
+
ccl_device float grad(int hash, float x, float y, float z)
{
// use vectors pointing to the edges of the cube
@@ -107,11 +171,61 @@ ccl_device float grad(int hash, float x, float y, float z)
return ((h&1) ? -u : u) + ((h&2) ? -v : v);
}
+#if defined(__KERNEL_SSE2__)
+ccl_device __m128 grad_sse(const __m128i *hash, const __m128 *x, const __m128 *y, const __m128 *z)
+{
+ __m128i c1 = _mm_set1_epi32(1);
+ __m128i c2 = _mm_set1_epi32(2);
+
+ __m128i h = _mm_and_si128(*hash, _mm_set1_epi32(15)); // h = hash & 15
+
+ __m128i case_ux = _mm_cmplt_epi32(h, _mm_set1_epi32(8)); // 0xffffffff if h < 8 else 0
+
+ __m128 ux = _mm_and_ps(_mm_castsi128_ps(case_ux), *x); // u = case_ux & x
+ __m128 uy = _mm_andnot_ps(_mm_castsi128_ps(case_ux), *y); // + !case_ux & y
+ __m128 u = _mm_add_ps(ux, uy); // ...
+
+ __m128i case_vy = _mm_cmplt_epi32(h, _mm_set1_epi32(4)); // 0xffffffff if h < 4 else 0
+
+ __m128i case_h12_raw = _mm_cmpeq_epi32(h, _mm_set1_epi32(12)); // 0xffffffff if h == 12 else 0
+ __m128i case_h14_raw = _mm_cmpeq_epi32(h, _mm_set1_epi32(14)); // 0xffffffff if h == 14 else 0
+
+ __m128i case_vxtmp = _mm_or_si128(case_h12_raw, case_h14_raw); // 0xffffffff if h == 12 or h == 14 else 0
+ __m128i case_vx = _mm_andnot_si128(case_vy, case_vxtmp); // 0xffffffff if (h == 12 or h == 14) and not(h<4)
+
+ __m128i case_vz = _mm_or_si128(case_vy, case_vx); // 0xffffffff if case_vy or case_vx else 0
+
+ __m128 vtx = _mm_and_ps(_mm_castsi128_ps(case_vx), *x); // v = case_vx & x
+ __m128 vty = _mm_and_ps(_mm_castsi128_ps(case_vy), *y); // + case_vy & y
+ __m128 vtz = _mm_andnot_ps(_mm_castsi128_ps(case_vz), *z); // + !case_vz_inv & z
+ __m128 v = _mm_add_ps(vtz, _mm_add_ps(vtx, vty)); // ...
+
+ __m128i case_uneg = _mm_slli_epi32(_mm_and_si128(h, c1), 31); // 1<<31 if h&1 else 0
+ __m128 case_uneg_mask = _mm_castsi128_ps(case_uneg); // -0.0 if h&1 else +0.0
+ __m128 ru = _mm_xor_ps(u, case_uneg_mask); // -u if h&1 else u (copy float sign)
+
+ __m128i case_vneg = _mm_slli_epi32(_mm_and_si128(h, c2), 30); // 2<<30 if h&2 else 0
+ __m128 case_vneg_mask = _mm_castsi128_ps(case_vneg); // -0.0 if h&2 else +0.0
+ __m128 rv = _mm_xor_ps(v, case_vneg_mask); // -v if h&2 else v (copy float sign)
+
+ __m128 r = _mm_add_ps(ru, rv); // ((h&1) ? -u : u) + ((h&2) ? -v : v)
+ return r;
+}
+#endif
+
ccl_device float scale3(float result)
{
return 0.9820f * result;
}
+#if defined(__KERNEL_SSE2__)
+ccl_device __m128 scale3_sse(const __m128 *result)
+{
+ return _mm_mul_ps(_mm_set1_ps(0.9820f), *result);
+}
+#endif
+
+#if !defined(__KERNEL_SSE2__)
ccl_device_noinline float perlin(float x, float y, float z)
{
int X; float fx = floorfrac(x, &X);
@@ -137,6 +251,50 @@ ccl_device_noinline float perlin(float x, float y, float z)
/* can happen for big coordinates, things even out to 0.0 then anyway */
return (isfinite(r))? r: 0.0f;
}
+#else
+ccl_device_noinline float perlin(float x, float y, float z)
+{
+#define BROADCAST_I(vec, n) _mm_shuffle_epi32((vec), _MM_SHUFFLE((n), (n), (n), (n)))
+#define BROADCAST_F(vec, n) _mm_shuffle_ps((vec), (vec), _MM_SHUFFLE((n), (n), (n), (n)))
+ __m128 xyz = _mm_setr_ps(x, y, z, 0.0f);
+ __m128i XYZ;
+
+ __m128 fxyz = floorfrac_sse(&xyz, &XYZ);
+
+ __m128 uvw = fade_sse(&fxyz);
+ __m128 u = BROADCAST_F(uvw, 0), v = BROADCAST_F(uvw, 1), w = BROADCAST_F(uvw, 2);
+
+ __m128i ci[] = {_mm_setr_epi32(1, 1, 1, 1), _mm_setr_epi32(0, 0, 1, 1), _mm_setr_epi32(0, 1, 0, 1)};
+ __m128i vp[] = {BROADCAST_I(XYZ, 0), BROADCAST_I(XYZ, 1), BROADCAST_I(XYZ, 2)};
+ __m128i vd[] = {_mm_add_epi32(vp[0], ci[0]), _mm_add_epi32(vp[1], ci[1]), _mm_add_epi32(vp[2], ci[2])};
+
+ __m128i h1 = hash_sse(vp, vd+1, vd+2); // hash directions 000, 001, 010, 011 (vp[0] is not a typo, because vp[0]+0 == vp[0])
+ __m128i h2 = hash_sse(vd, vd+1, vd+2); // hash directions 100, 101, 110, 111
+
+ __m128 cf[] = {_mm_setr_ps(1.0f, 1.0f, 1.0f, 1.0f), _mm_setr_ps(0.0f, 0.0f, 1.0f, 1.0f), _mm_setr_ps(0.0f, 1.0f, 0.0f, 1.0f)};
+ __m128 vf[] = {BROADCAST_F(fxyz, 0), BROADCAST_F(fxyz, 1), BROADCAST_F(fxyz, 2)};
+ __m128 vfd[] = {_mm_sub_ps(vf[0], cf[0]), _mm_sub_ps(vf[1], cf[1]), _mm_sub_ps(vf[2], cf[2])};
+
+ __m128 g1 = grad_sse(&h1, vf, vfd+1, vfd+2); // vf is not a typo (same as above)
+ __m128 g2 = grad_sse(&h2, vfd, vfd+1, vfd+2);
+ __m128 n1 = nerp_sse(&u, &g1, &g2);
+
+ __m128 n1_half = _mm_movehl_ps(n1, n1); // extract 2 floats to a separate vector
+ __m128 n2 = nerp_sse(&v, &n1, &n1_half); // process nerp([a, b, _, _], [c, d, _, _]) -> [a', b', _, _]
+
+ __m128 n2_second = BROADCAST_F(n2, 1); // extract b to a separate vector
+ __m128 result = nerp_sse(&w, &n2, &n2_second); // process nerp([a', _, _, _], [b', _, _, _]) -> [a'', _, _, _]
+
+ __m128 r = scale3_sse(&result);
+
+ __m128 infmask = _mm_castsi128_ps(_mm_set1_epi32(0x7f800000));
+ __m128 rinfmask = _mm_cmpeq_ps(_mm_and_ps(r, infmask), infmask); // 0xffffffff if r is inf/-inf/nan else 0
+ __m128 rfinite = _mm_andnot_ps(rinfmask, r); // 0 if r is inf/-inf/nan else r
+ return _mm_cvtss_f32(rfinite);
+#undef BROADCAST_I
+#undef BROADCAST_F
+}
+#endif
ccl_device_noinline float perlin_periodic(float x, float y, float z, float3 pperiod)
{
More information about the Bf-blender-cvs
mailing list