Cycles: SSE optimization for Voronoi cells texture

Gives 5-6% speedup for Caterpillar_PatazStudio.blend.

Reviewed By: brecht, dingto

Differential Revision: https://developer.blender.org/D419

intern/cycles/kernel/svm/svm_noise.h

@@ -357,15 +357,13 @@ ccl_device float3 cellnoise_color(float3 p)
 	return make_float3(r, g, b);
 }
 #else
-ccl_device float3 cellnoise_color(const float3& p)
+ccl_device __m128 cellnoise_color(const __m128& p)
 {
-	__m128i v_yxz = quick_floor_sse(_mm_setr_ps(p.y, p.x, p.z, 0.0f));
-	__m128i v_xyy = shuffle<1, 0, 0, 3>(v_yxz);
-	__m128i v_zzx = shuffle<2, 2, 1, 3>(v_yxz);
-	__m128 rgb = bits_to_01_sse(hash_sse(v_xyy, v_yxz, v_zzx));
-
-	float3 result = *(float3*)&rgb;
-	return result;
+	__m128i ip = quick_floor_sse(p);
+	__m128i ip_yxz = shuffle<1, 0, 2, 3>(ip);
+	__m128i ip_xyy = shuffle<0, 1, 1, 3>(ip);
+	__m128i ip_zzx = shuffle<2, 2, 0, 3>(ip);
+	return bits_to_01_sse(hash_sse(ip_xyy, ip_yxz, ip_zzx));
 }
 #endif
 

intern/cycles/kernel/svm/svm_texture.h

@@ -18,6 +18,7 @@ CCL_NAMESPACE_BEGIN
 
 /* Voronoi Distances */
 
+#if 0
 ccl_device float voronoi_distance(NodeDistanceMetric distance_metric, float3 d, float e)
 {
 #if 0
@@ -43,8 +44,7 @@ ccl_device float voronoi_distance(NodeDistanceMetric distance_metric, float3 d,
 }
 
 /* Voronoi / Worley like */
-
-ccl_device_noinline float4 voronoi_Fn(float3 p, float e, int n1, int n2)
+ccl_device_inline float4 voronoi_Fn(float3 p, float e, int n1, int n2)
 {
 	float da[4];
 	float3 pa[4];
@@ -119,7 +119,95 @@ ccl_device_noinline float4 voronoi_Fn(float3 p, float e, int n1, int n2)
 
 	return result;
 }
+#endif
 
+ccl_device float voronoi_F1_distance(float3 p)
+{
+	/* returns squared distance in da */
+	float da = 1e10f;
+
+#ifndef __KERNEL_SSE2__
+	int ix = floor_to_int(p.x), iy = floor_to_int(p.y), iz = floor_to_int(p.z);
+
+	for (int xx = -1; xx <= 1; xx++) {
+		for (int yy = -1; yy <= 1; yy++) {
+			for (int zz = -1; zz <= 1; zz++) {
+				float3 ip = make_float3(ix + xx, iy + yy, iz + zz);
+				float3 vp = ip + cellnoise_color(ip);
+				float d = len_squared(p - vp);
+				da = min(d, da);
+			}
+		}
+	}
+#else
+	__m128 vec_p = load_m128(p);
+	__m128i xyzi = quick_floor_sse(vec_p);
+
+	for (int xx = -1; xx <= 1; xx++) {
+		for (int yy = -1; yy <= 1; yy++) {
+			for (int zz = -1; zz <= 1; zz++) {
+				__m128 ip = _mm_cvtepi32_ps(_mm_add_epi32(xyzi, _mm_setr_epi32(xx, yy, zz, 0)));
+				__m128 vp = _mm_add_ps(ip, cellnoise_color(ip));
+				float d = len_squared<1, 1, 1, 0>(_mm_sub_ps(vec_p, vp));
+				da = min(d, da);
+			}
+		}
+	}
+#endif
+
+	return da;
+}
+
+ccl_device float3 voronoi_F1_color(float3 p)
+{
+	/* returns color of the nearest point */
+	float da = 1e10f;
+
+#ifndef __KERNEL_SSE2__
+	float3 pa;
+	int ix = floor_to_int(p.x), iy = floor_to_int(p.y), iz = floor_to_int(p.z);
+
+	for (int xx = -1; xx <= 1; xx++) {
+		for (int yy = -1; yy <= 1; yy++) {
+			for (int zz = -1; zz <= 1; zz++) {
+				float3 ip = make_float3(ix + xx, iy + yy, iz + zz);
+				float3 vp = ip + cellnoise_color(ip);
+				float d = len_squared(p - vp);
+
+				if(d < da) {
+					da = d;
+					pa = vp;
+				}
+			}
+		}
+	}
+
+	return cellnoise_color(pa);
+#else
+	__m128 pa, vec_p = load_m128(p);
+	__m128i xyzi = quick_floor_sse(vec_p);
+
+	for (int xx = -1; xx <= 1; xx++) {
+		for (int yy = -1; yy <= 1; yy++) {
+			for (int zz = -1; zz <= 1; zz++) {
+				__m128 ip = _mm_cvtepi32_ps(_mm_add_epi32(xyzi, _mm_setr_epi32(xx, yy, zz, 0)));
+				__m128 vp = _mm_add_ps(ip, cellnoise_color(ip));
+				float d = len_squared<1, 1, 1, 0>(_mm_sub_ps(vec_p, vp));
+
+				if(d < da) {
+					da = d;
+					pa = vp;
+				}
+			}
+		}
+	}
+
+	__m128 color = cellnoise_color(pa);
+	return (float3 &)color;
+#endif
+}
+
+#if 0
 ccl_device float voronoi_F1(float3 p) { return voronoi_Fn(p, 0.0f, 0, -1).w; }
 ccl_device float voronoi_F2(float3 p) { return voronoi_Fn(p, 0.0f, 1, -1).w; }
 ccl_device float voronoi_F3(float3 p) { return voronoi_Fn(p, 0.0f, 2, -1).w; }
@@ -139,6 +227,7 @@ ccl_device float voronoi_F3S(float3 p) { return 2.0f*voronoi_F3(p) - 1.0f; }
 ccl_device float voronoi_F4S(float3 p) { return 2.0f*voronoi_F4(p) - 1.0f; }
 ccl_device float voronoi_F1F2S(float3 p) { return 2.0f*voronoi_F1F2(p) - 1.0f; }
 ccl_device float voronoi_CrS(float3 p) { return 2.0f*voronoi_Cr(p) - 1.0f; }
+#endif
 
 /* Noise Bases */
 

intern/cycles/kernel/svm/svm_voronoi.h

@@ -20,23 +20,16 @@ CCL_NAMESPACE_BEGIN
 
 ccl_device_noinline float4 svm_voronoi(NodeVoronoiColoring coloring, float3 p)
 {
-	/* compute distance and point coordinate of 4 nearest neighbours */
-	float4 dpa0 = voronoi_Fn(p, 1.0f, 0, -1);
-
-	/* output */
-	float fac;
-	float3 color;
-
 	if(coloring == NODE_VORONOI_INTENSITY) {
-		fac = fabsf(dpa0.w);
-		color = make_float3(fac, fac, fac);
+		/* compute squared distance to the nearest neighbour */
+		float fac = voronoi_F1_distance(p);
+		return make_float4(fac, fac, fac, fac);
 	}
 	else {
-		color = cellnoise_color(float4_to_float3(dpa0));
-		fac = average(color);
+		/* compute color of the nearest neighbour */
+		float3 color = voronoi_F1_color(p);
+		return make_float4(color.x, color.y, color.z, average(color));
 	}
-
-	return make_float4(color.x, color.y, color.z, fac);
 }
 
 ccl_device void svm_node_tex_voronoi(KernelGlobals *kg, ShaderData *sd, float *stack, uint4 node, int *offset)

intern/cycles/util/util_math.h

@@ -163,11 +163,7 @@ ccl_device_inline float clamp(float a, float mn, float mx)
 
 ccl_device_inline int float_to_int(float f)
 {
-#if defined(__KERNEL_SSE2__) && !defined(_MSC_VER)
-	return _mm_cvtt_ss2si(_mm_load_ss(&f));
-#else
 	return (int)f;
-#endif
 }
 
 ccl_device_inline int floor_to_int(float f)

intern/cycles/util/util_simd.h

@@ -220,6 +220,18 @@ ccl_device_inline const __m128 dot3_splat(const __m128& a, const __m128& b)
 #endif
 }
 
+/* squared length taking only specified axes into account */
+template<size_t X, size_t Y, size_t Z, size_t W>
+ccl_device_inline float len_squared(const __m128& a)
+{
+#ifndef __KERNEL_SSE41__
+	float4& t = (float4 &)a;
+	return (X ? t.x * t.x : 0.0f) + (Y ? t.y * t.y : 0.0f) + (Z ? t.z * t.z : 0.0f) + (W ? t.w * t.w : 0.0f);
+#else
+	return _mm_cvtss_f32(_mm_dp_ps(a, a, (X << 4) | (Y << 5) | (Z << 6) | (W << 7) | 0xf));
+#endif
+}
+
 ccl_device_inline float dot3(const __m128& a, const __m128& b)
 {
 #ifdef __KERNEL_SSE41__