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

8 files changed, 14 insertions, 14 deletions

diff --git a/intern/cycles/kernel/svm/svm_brick.h b/intern/cycles/kernel/svm/svm_brick.h
index 33a2a5c7598..fcf8f47b77e 100644
--- a/intern/cycles/kernel/svm/svm_brick.h
+++ b/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) ||
diff --git a/intern/cycles/kernel/svm/svm_closure.h b/intern/cycles/kernel/svm/svm_closure.h
index 8afdb73fb38..0d2d155f827 100644
--- a/intern/cycles/kernel/svm/svm_closure.h
+++ b/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;
 
diff --git a/intern/cycles/kernel/svm/svm_gradient.h b/intern/cycles/kernel/svm/svm_gradient.h
index a5e385faddc..53d7b4f812c 100644
--- a/intern/cycles/kernel/svm/svm_gradient.h
+++ b/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);
diff --git a/intern/cycles/kernel/svm/svm_image.h b/intern/cycles/kernel/svm/svm_image.h
index ccc1e3df730..08a6c01162c 100644
--- a/intern/cycles/kernel/svm/svm_image.h
+++ b/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 {
diff --git a/intern/cycles/kernel/svm/svm_math_util.h b/intern/cycles/kernel/svm/svm_math_util.h
index 39cc14d5e8e..ff9e662e931 100644
--- a/intern/cycles/kernel/svm/svm_math_util.h
+++ b/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;
 	
diff --git a/intern/cycles/kernel/svm/svm_mix.h b/intern/cycles/kernel/svm/svm_mix.h
index b6b1966cb3b..6111214acba 100644
--- a/intern/cycles/kernel/svm/svm_mix.h
+++ b/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.y = clamp(col.y, 0.0f, 1.0f);
-	outcol.z = clamp(col.z, 0.0f, 1.0f);
+	outcol.x = saturate(col.x);
+	outcol.y = saturate(col.y);
+	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);
diff --git a/intern/cycles/kernel/svm/svm_musgrave.h b/intern/cycles/kernel/svm/svm_musgrave.h
index 2f81ddaa74c..2a0a8e0e87d 100644
--- a/intern/cycles/kernel/svm/svm_musgrave.h
+++ b/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;
diff --git a/intern/cycles/kernel/svm/svm_ramp.h b/intern/cycles/kernel/svm/svm_ramp.h
index 998f649a571..062ab013b1f 100644
--- a/intern/cycles/kernel/svm/svm_ramp.h
+++ b/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);