Cycles: Use native saturate function for CUDA

This more a workaround for CUDA optimizer which can't optimize clamp(x, 0, 1)
into a single instruction and uses 4 instructions instead.

Original patch by @lockal with own modification:

  Don't make changes outside of the kernel. They don't make any difference
  anyway and term saturate() has a bit different meaning outside of kernel.

This gives around 2% of speedup in Barcelona file, but in more complex shader
setups with lots of math nodes with clamping speedup could be much nicer.

Subscribers: dingto

Projects: #cycles

Differential Revision: https://developer.blender.org/D1224

intern/cycles/kernel/closure/bsdf_microfacet.h

@@ -235,7 +235,7 @@ ccl_device_inline float3 microfacet_sample_stretched(
 
 ccl_device int bsdf_microfacet_ggx_setup(ShaderClosure *sc)
 {
-	sc->data0 = clamp(sc->data0, 0.0f, 1.0f); /* alpha_x */
+	sc->data0 = saturate(sc->data0); /* alpha_x */
 	sc->data1 = sc->data0; /* alpha_y */
 	
 	sc->type = CLOSURE_BSDF_MICROFACET_GGX_ID;
@@ -245,8 +245,8 @@ ccl_device int bsdf_microfacet_ggx_setup(ShaderClosure *sc)
 
 ccl_device int bsdf_microfacet_ggx_aniso_setup(ShaderClosure *sc)
 {
-	sc->data0 = clamp(sc->data0, 0.0f, 1.0f); /* alpha_x */
+	sc->data0 = saturate(sc->data0); /* alpha_x */
-	sc->data1 = clamp(sc->data1, 0.0f, 1.0f); /* alpha_y */
+	sc->data1 = saturate(sc->data1); /* alpha_y */
 	
 	sc->type = CLOSURE_BSDF_MICROFACET_GGX_ANISO_ID;
 
@@ -255,7 +255,7 @@ ccl_device int bsdf_microfacet_ggx_aniso_setup(ShaderClosure *sc)
 
 ccl_device int bsdf_microfacet_ggx_refraction_setup(ShaderClosure *sc)
 {
-	sc->data0 = clamp(sc->data0, 0.0f, 1.0f); /* alpha_x */
+	sc->data0 = saturate(sc->data0); /* alpha_x */
 	sc->data1 = sc->data0; /* alpha_y */
 
 	sc->type = CLOSURE_BSDF_MICROFACET_GGX_REFRACTION_ID;
@@ -588,7 +588,7 @@ ccl_device int bsdf_microfacet_ggx_sample(KernelGlobals *kg, const ShaderClosure
 
 ccl_device int bsdf_microfacet_beckmann_setup(ShaderClosure *sc)
 {
-	sc->data0 = clamp(sc->data0, 0.0f, 1.0f); /* alpha_x */
+	sc->data0 = saturate(sc->data0); /* alpha_x */
 	sc->data1 = sc->data0; /* alpha_y */
 
 	sc->type = CLOSURE_BSDF_MICROFACET_BECKMANN_ID;
@@ -597,8 +597,8 @@ ccl_device int bsdf_microfacet_beckmann_setup(ShaderClosure *sc)
 
 ccl_device int bsdf_microfacet_beckmann_aniso_setup(ShaderClosure *sc)
 {
-	sc->data0 = clamp(sc->data0, 0.0f, 1.0f); /* alpha_x */
+	sc->data0 = saturate(sc->data0); /* alpha_x */
-	sc->data1 = clamp(sc->data1, 0.0f, 1.0f); /* alpha_y */
+	sc->data1 = saturate(sc->data1); /* alpha_y */
 
 	sc->type = CLOSURE_BSDF_MICROFACET_BECKMANN_ANISO_ID;
 	return SD_BSDF|SD_BSDF_HAS_EVAL;
@@ -606,7 +606,7 @@ ccl_device int bsdf_microfacet_beckmann_aniso_setup(ShaderClosure *sc)
 
 ccl_device int bsdf_microfacet_beckmann_refraction_setup(ShaderClosure *sc)
 {
-	sc->data0 = clamp(sc->data0, 0.0f, 1.0f); /* alpha_x */
+	sc->data0 = saturate(sc->data0); /* alpha_x */
 	sc->data1 = sc->data0; /* alpha_y */
 
 	sc->type = CLOSURE_BSDF_MICROFACET_BECKMANN_REFRACTION_ID;

intern/cycles/kernel/closure/bsdf_oren_nayar.h

@@ -37,7 +37,7 @@ ccl_device int bsdf_oren_nayar_setup(ShaderClosure *sc)
 
 	sc->type = CLOSURE_BSDF_OREN_NAYAR_ID;
 
-	sigma = clamp(sigma, 0.0f, 1.0f);
+	sigma = saturate(sigma);
 
 	float div = 1.0f / (M_PI_F + ((3.0f * M_PI_F - 4.0f) / 6.0f) * sigma);
 

intern/cycles/kernel/closure/bsdf_toon.h

@@ -40,8 +40,8 @@ CCL_NAMESPACE_BEGIN
 ccl_device int bsdf_diffuse_toon_setup(ShaderClosure *sc)
 {
 	sc->type = CLOSURE_BSDF_DIFFUSE_TOON_ID;
-	sc->data0 = clamp(sc->data0, 0.0f, 1.0f);
+	sc->data0 = saturate(sc->data0);
-	sc->data1 = clamp(sc->data1, 0.0f, 1.0f);
+	sc->data1 = saturate(sc->data1);
 
 	return SD_BSDF|SD_BSDF_HAS_EVAL;
 }
@@ -120,8 +120,8 @@ ccl_device int bsdf_diffuse_toon_sample(const ShaderClosure *sc, float3 Ng, floa
 ccl_device int bsdf_glossy_toon_setup(ShaderClosure *sc)
 {
 	sc->type = CLOSURE_BSDF_GLOSSY_TOON_ID;
-	sc->data0 = clamp(sc->data0, 0.0f, 1.0f);
+	sc->data0 = saturate(sc->data0);
-	sc->data1 = clamp(sc->data1, 0.0f, 1.0f);
+	sc->data1 = saturate(sc->data1);
 
 	return SD_BSDF|SD_BSDF_HAS_EVAL;
 }

intern/cycles/kernel/closure/bssrdf.h

@@ -30,8 +30,8 @@ ccl_device int bssrdf_setup(ShaderClosure *sc, ClosureType type)
 		return flag;
 	}
 	else {
-		sc->data1 = clamp(sc->data1, 0.0f, 1.0f); /* texture blur */
+		sc->data1 = saturate(sc->data1); /* texture blur */
-		sc->T.x = clamp(sc->T.x, 0.0f, 1.0f); /* sharpness */
+		sc->T.x = saturate(sc->T.x); /* sharpness */
 		sc->type = type;
 
 		return SD_BSDF|SD_BSDF_HAS_EVAL|SD_BSSRDF;
@@ -168,7 +168,7 @@ ccl_device float bssrdf_cubic_quintic_root_find(float xi)
 		if(fabsf(f) < tolerance || f_ == 0.0f)
 			break;
 
-		x = clamp(x - f/f_, 0.0f, 1.0f);
+		x = saturate(x - f/f_);
 	}
 
 	return x;

intern/cycles/kernel/geom/geom_curve.h

@@ -465,7 +465,7 @@ ccl_device_inline bool bvh_cardinal_curve_intersect(KernelGlobals *kg, Intersect
 					continue;
 				}
 				w = -(p_st.x * tg.x + p_st.y * tg.y) / w;
-				w = clamp((float)w, 0.0f, 1.0f);
+				w = saturate(w);
 
 				/* compute u on the curve segment */
 				u = i_st * (1 - w) + i_en * w;
@@ -577,7 +577,7 @@ ccl_device_inline bool bvh_cardinal_curve_intersect(KernelGlobals *kg, Intersect
 				}
 
 				float w = (zcentre + (tg.z * correction)) * invl;
-				w = clamp((float)w, 0.0f, 1.0f);
+				w = saturate(w);
 				/* compute u on the curve segment */
 				u = i_st * (1 - w) + i_en * w;
 

intern/cycles/kernel/kernel_film.h

@@ -27,7 +27,7 @@ ccl_device float4 film_map(KernelGlobals *kg, float4 irradiance, float scale)
 	result.z = color_scene_linear_to_srgb(result.z*exposure);
 
 	/* clamp since alpha might be > 1.0 due to russian roulette */
-	result.w = clamp(result.w, 0.0f, 1.0f);
+	result.w = saturate(result.w);
 
 	return result;
 }
@@ -37,10 +37,10 @@ ccl_device uchar4 film_float_to_byte(float4 color)
 	uchar4 result;
 
 	/* simple float to byte conversion */
-	result.x = (uchar)clamp(color.x*255.0f, 0.0f, 255.0f);
+	result.x = (uchar)(saturate(color.x)*255.0f);
-	result.y = (uchar)clamp(color.y*255.0f, 0.0f, 255.0f);
+	result.y = (uchar)(saturate(color.y)*255.0f);
-	result.z = (uchar)clamp(color.z*255.0f, 0.0f, 255.0f);
+	result.z = (uchar)(saturate(color.z)*255.0f);
-	result.w = (uchar)clamp(color.w*255.0f, 0.0f, 255.0f);
+	result.w = (uchar)(saturate(color.w)*255.0f);
 
 	return result;
 }

intern/cycles/kernel/kernel_globals.h

@@ -94,7 +94,7 @@ typedef struct KernelGlobals {
 
 ccl_device float lookup_table_read(KernelGlobals *kg, float x, int offset, int size)
 {
-	x = clamp(x, 0.0f, 1.0f)*(size-1);
+	x = saturate(x)*(size-1);
 
 	int index = min(float_to_int(x), size-1);
 	int nindex = min(index+1, size-1);
@@ -110,7 +110,7 @@ ccl_device float lookup_table_read(KernelGlobals *kg, float x, int offset, int s
 
 ccl_device float lookup_table_read_2D(KernelGlobals *kg, float x, float y, int offset, int xsize, int ysize)
 {
-	y = clamp(y, 0.0f, 1.0f)*(ysize-1);
+	y = saturate(y)*(ysize-1);
 
 	int index = min(float_to_int(y), ysize-1);
 	int nindex = min(index+1, ysize-1);

intern/cycles/kernel/kernel_passes.h

@@ -102,7 +102,7 @@ ccl_device_inline void kernel_write_data_passes(KernelGlobals *kg, ccl_global fl
 		float mist_inv_depth = kernel_data.film.mist_inv_depth;
 
 		float depth = camera_distance(kg, sd->P);
-		float mist = clamp((depth - mist_start)*mist_inv_depth, 0.0f, 1.0f);
+		float mist = saturate((depth - mist_start)*mist_inv_depth);
 
 		/* falloff */
 		float mist_falloff = kernel_data.film.mist_falloff;

intern/cycles/kernel/svm/svm_brick.h

@@ -47,7 +47,7 @@ ccl_device_noinline float2 svm_brick(float3 p, float mortar_size, float bias,
 	y = p.y - row_height*rownum;
 
 	return make_float2(
-		clamp((brick_noise((rownum << 16) + (bricknum & 0xFFFF)) + bias), 0.0f, 1.0f),
+		saturate((brick_noise((rownum << 16) + (bricknum & 0xFFFF)) + bias)),
 
 		(x < mortar_size || y < mortar_size ||
 		x > (brick_width - mortar_size) ||

intern/cycles/kernel/svm/svm_closure.h

@@ -347,7 +347,7 @@ ccl_device void svm_node_closure_bsdf(KernelGlobals *kg, ShaderData *sd, float *
 				sc->N = N;
 
 				/* sigma */
-				sc->data0 = clamp(param1, 0.0f, 1.0f);
+				sc->data0 = saturate(param1);
 				sc->data1 = 0.0f;
 				sc->data2 = 0.0f;
 				sd->flag |= bsdf_ashikhmin_velvet_setup(sc);
@@ -655,7 +655,7 @@ ccl_device void svm_node_mix_closure(ShaderData *sd, float *stack, uint4 node)
 	decode_node_uchar4(node.y, &weight_offset, &in_weight_offset, &weight1_offset, &weight2_offset);
 
 	float weight = stack_load_float(stack, weight_offset);
-	weight = clamp(weight, 0.0f, 1.0f);
+	weight = saturate(weight);
 
 	float in_weight = (stack_valid(in_weight_offset))? stack_load_float(stack, in_weight_offset): 1.0f;
 

intern/cycles/kernel/svm/svm_gradient.h

@@ -66,7 +66,7 @@ ccl_device void svm_node_tex_gradient(ShaderData *sd, float *stack, uint4 node)
 	float3 co = stack_load_float3(stack, co_offset);
 
 	float f = svm_gradient(co, (NodeGradientType)type);
-	f = clamp(f, 0.0f, 1.0f);
+	f = saturate(f);
 
 	if(stack_valid(fac_offset))
 		stack_store_float(stack, fac_offset, f);

intern/cycles/kernel/svm/svm_image.h

@@ -433,17 +433,17 @@ ccl_device void svm_node_tex_image_box(KernelGlobals *kg, ShaderData *sd, float
 		/* in case of blending, test for mixes between two textures */
 		if(N.z < (1.0f - limit)*(N.y + N.x)) {
 			weight.x = N.x/(N.x + N.y);
-			weight.x = clamp((weight.x - 0.5f*(1.0f - blend))/blend, 0.0f, 1.0f);
+			weight.x = saturate((weight.x - 0.5f*(1.0f - blend))/blend);
 			weight.y = 1.0f - weight.x;
 		}
 		else if(N.x < (1.0f - limit)*(N.y + N.z)) {
 			weight.y = N.y/(N.y + N.z);
-			weight.y = clamp((weight.y - 0.5f*(1.0f - blend))/blend, 0.0f, 1.0f);
+			weight.y = saturate((weight.y - 0.5f*(1.0f - blend))/blend);
 			weight.z = 1.0f - weight.y;
 		}
 		else if(N.y < (1.0f - limit)*(N.x + N.z)) {
 			weight.x = N.x/(N.x + N.z);
-			weight.x = clamp((weight.x - 0.5f*(1.0f - blend))/blend, 0.0f, 1.0f);
+			weight.x = saturate((weight.x - 0.5f*(1.0f - blend))/blend);
 			weight.z = 1.0f - weight.x;
 		}
 		else {

intern/cycles/kernel/svm/svm_math_util.h

@@ -97,7 +97,7 @@ ccl_device float svm_math(NodeMath type, float Fac1, float Fac2)
 	else if(type == NODE_MATH_ABSOLUTE)
 		Fac = fabsf(Fac1);
 	else if(type == NODE_MATH_CLAMP)
-		Fac = clamp(Fac1, 0.0f, 1.0f);
+		Fac = saturate(Fac1);
 	else
 		Fac = 0.0f;
 	

intern/cycles/kernel/svm/svm_mix.h

@@ -254,16 +254,16 @@ ccl_device float3 svm_mix_clamp(float3 col)
 {
 	float3 outcol = col;
 
-	outcol.x = clamp(col.x, 0.0f, 1.0f);
+	outcol.x = saturate(col.x);
-	outcol.y = clamp(col.y, 0.0f, 1.0f);
+	outcol.y = saturate(col.y);
-	outcol.z = clamp(col.z, 0.0f, 1.0f);
+	outcol.z = saturate(col.z);
 
 	return outcol;
 }
 
 ccl_device float3 svm_mix(NodeMix type, float fac, float3 c1, float3 c2)
 {
-	float t = clamp(fac, 0.0f, 1.0f);
+	float t = saturate(fac);
 
 	switch(type) {
 		case NODE_MIX_BLEND: return svm_mix_blend(t, c1, c2);

intern/cycles/kernel/svm/svm_musgrave.h

@@ -168,7 +168,7 @@ ccl_device_noinline float noise_musgrave_ridged_multi_fractal(float3 p, NodeNois
 
 	for(i = 1; i < float_to_int(octaves); i++) {
 		p *= lacunarity;
-		weight = clamp(signal * gain, 0.0f, 1.0f);
+		weight = saturate(signal * gain);
 		signal = offset - fabsf(snoise(p));
 		signal *= signal;
 		signal *= weight;

intern/cycles/kernel/svm/svm_ramp.h

@@ -21,7 +21,7 @@ CCL_NAMESPACE_BEGIN
 
 ccl_device float4 rgb_ramp_lookup(KernelGlobals *kg, int offset, float f, bool interpolate)
 {
-	f = clamp(f, 0.0f, 1.0f)*(RAMP_TABLE_SIZE-1);
+	f = saturate(f)*(RAMP_TABLE_SIZE-1);
 
 	/* clamp int as well in case of NaN */
 	int i = clamp(float_to_int(f), 0, RAMP_TABLE_SIZE-1);

intern/cycles/render/buffers.cpp

@@ -187,7 +187,7 @@ bool RenderBuffers::get_pass_rect(PassType type, float exposure, int sample, int
 			else if(type == PASS_MIST) {
 				for(int i = 0; i < size; i++, in += pass_stride, pixels++) {
 					float f = *in;
-					pixels[0] = clamp(f*scale_exposure, 0.0f, 1.0f);
+					pixels[0] = saturate(f*scale_exposure);
 				}
 			}
 #ifdef WITH_CYCLES_DEBUG
@@ -298,7 +298,7 @@ bool RenderBuffers::get_pass_rect(PassType type, float exposure, int sample, int
 					pixels[2] = f.z*scale_exposure;
 
 					/* clamp since alpha might be > 1.0 due to russian roulette */
-					pixels[3] = clamp(f.w*scale, 0.0f, 1.0f);
+					pixels[3] = saturate(f.w*scale);
 				}
 			}
 		}

intern/cycles/render/nodes.cpp

@@ -3752,7 +3752,7 @@ void MathNode::compile(SVMCompiler& compiler)
 		                                 value1_in->value.x,
 		                                 value2_in->value.x);
 		if(use_clamp) {
-			optimized_value = clamp(optimized_value, 0.0f, 1.0f);
+			optimized_value = saturate(optimized_value);
 		}
 		compiler.add_node(NODE_VALUE_F,
 		                  __float_as_int(optimized_value),

intern/cycles/util/util_math.h

@@ -175,6 +175,15 @@ ccl_device_inline float clamp(float a, float mn, float mx)
 
 #endif
 
+#ifndef __KERNEL_CUDA__
+
+ccl_device_inline float saturate(float a)
+{
+	return clamp(a, 0.0f, 1.0f);
+}
+
+#endif
+
 ccl_device_inline int float_to_int(float f)
 {
 	return (int)f;