[Bf-blender-cvs] [20b046d] master: Cycles: workaround for noise performance regression in CUDA 5.5

[Sv. Lockal]

Commit: 20b046d7639c6c95d942760f8004c449fc42b081
Author: Sv. Lockal
Date:   Wed Jan 8 22:25:55 2014 +0400
https://developer.blender.org/rB20b046d7639c6c95d942760f8004c449fc42b081

Cycles: workaround for noise performance regression in CUDA 5.5

Use manual ternary operation widening in grad(). Without it nvcc 5.5 produces multiple branch splits with very big branches (because of inlining). This solves 19% performance regression for BMW1M-MikePan.blend.

Also remove one redundant instruction in perlin SSE (when h == 12 or h == 14, then h is always >= 4).

Reviewed By: brecht

Differential Revision: https://developer.blender.org/D190

===================================================================

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 add5f4c..a58dfdf 100644
--- a/intern/cycles/kernel/svm/svm_noise.h
+++ b/intern/cycles/kernel/svm/svm_noise.h
@@ -167,7 +167,8 @@ ccl_device float grad(int hash, float x, float y, float z)
 	// use vectors pointing to the edges of the cube
 	int h = hash & 15;
 	float u = h<8 ? x : y;
-	float v = h<4 ? y : h == 12 || h == 14 ? x : z;
+	float vt = ((h == 12) | (h == 14)) ? x : z;
+	float v = h < 4 ? y : vt;
 	return ((h&1) ? -u : u) + ((h&2) ? -v : v);
 }
 
@@ -185,11 +186,10 @@ ccl_device_inline __m128 grad_sse(const __m128i *hash, const __m128 *x, const __
 
 	__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_h12 = _mm_cmpeq_epi32(h, _mm_set1_epi32(12));     // 0xffffffff if h == 12 else 0
+	__m128i case_h14 = _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_vx = _mm_or_si128(case_h12, case_h14);            // 0xffffffff if h == 12 or h == 14 else 0
 
 	__m128 v = blend(_mm_castsi128_ps(case_vy), *y, blend(_mm_castsi128_ps(case_vx), *x, *z)); // v = h<4 ? y : h == 12 || h == 14 ? x : z