Fix Cycles multi scatter GGX different render results with Clang and GCC.

The order of evaluation of function arguments is undefined, and the order
was reversed between these compilers. This was causing regressions tests
to give different results between Linux and macOS.

2 files changed, 28 insertions, 20 deletions

diff --git a/intern/cycles/kernel/closure/bsdf_microfacet_multi.h b/intern/cycles/kernel/closure/bsdf_microfacet_multi.h
index 22d0092093a..2f2c35d5d1f 100644
--- a/intern/cycles/kernel/closure/bsdf_microfacet_multi.h
+++ b/intern/cycles/kernel/closure/bsdf_microfacet_multi.h
@@ -40,20 +40,20 @@ ccl_device_forceinline float D_ggx_aniso(const float3 wm, const float2 alpha)
 }
 
 /* Sample slope distribution (based on page 14 of the supplemental implementation). */
-ccl_device_forceinline float2 mf_sampleP22_11(const float cosI, const float2 randU)
+ccl_device_forceinline float2 mf_sampleP22_11(const float cosI, const float randx, const float randy)
 {
 	if(cosI > 0.9999f || fabsf(cosI) < 1e-6f) {
-		const float r = sqrtf(randU.x / max(1.0f - randU.x, 1e-7f));
-		const float phi = M_2PI_F * randU.y;
+		const float r = sqrtf(randx / max(1.0f - randx, 1e-7f));
+		const float phi = M_2PI_F * randy;
 		return make_float2(r*cosf(phi), r*sinf(phi));
 	}
 
-	const float sinI = sqrtf(1.0f - cosI*cosI);
+	const float sinI = safe_sqrtf(1.0f - cosI*cosI);
 	const float tanI = sinI/cosI;
 	const float projA = 0.5f * (cosI + 1.0f);
 	if(projA < 0.0001f)
 		return make_float2(0.0f, 0.0f);
-	const float A = 2.0f*randU.x*projA / cosI - 1.0f;
+	const float A = 2.0f*randx*projA / cosI - 1.0f;
 	float tmp = A*A-1.0f;
 	if(fabsf(tmp) < 1e-7f)
 		return make_float2(0.0f, 0.0f);
@@ -64,24 +64,24 @@ ccl_device_forceinline float2 mf_sampleP22_11(const float cosI, const float2 ran
 	const float slopeX = (A < 0.0f || slopeX2 > 1.0f/tanI)? (tanI*tmp - D) : slopeX2;
 
 	float U2;
-	if(randU.y >= 0.5f)
-		U2 = 2.0f*(randU.y - 0.5f);
+	if(randy >= 0.5f)
+		U2 = 2.0f*(randy - 0.5f);
 	else
-		U2 = 2.0f*(0.5f - randU.y);
+		U2 = 2.0f*(0.5f - randy);
 	const float z = (U2*(U2*(U2*0.27385f-0.73369f)+0.46341f)) / (U2*(U2*(U2*0.093073f+0.309420f)-1.0f)+0.597999f);
 	const float slopeY = z * sqrtf(1.0f + slopeX*slopeX);
 
-	if(randU.y >= 0.5f)
+	if(randy >= 0.5f)
 		return make_float2(slopeX, slopeY);
 	else
 		return make_float2(slopeX, -slopeY);
 }
 
 /* Visible normal sampling for the GGX distribution (based on page 7 of the supplemental implementation). */
-ccl_device_forceinline float3 mf_sample_vndf(const float3 wi, const float2 alpha, const float2 randU)
+ccl_device_forceinline float3 mf_sample_vndf(const float3 wi, const float2 alpha, const float randx, const float randy)
 {
 	const float3 wi_11 = normalize(make_float3(alpha.x*wi.x, alpha.y*wi.y, wi.z));
-	const float2 slope_11 = mf_sampleP22_11(wi_11.z, randU);
+	const float2 slope_11 = mf_sampleP22_11(wi_11.z, randx, randy);
 
 	const float3 cossin_phi = safe_normalize(make_float3(wi_11.x, wi_11.y, 0.0f));
 	const float slope_x = alpha.x*(cossin_phi.x * slope_11.x - cossin_phi.y * slope_11.y);
@@ -474,6 +474,7 @@ ccl_device int bsdf_microfacet_multi_ggx_sample(KernelGlobals *kg, const ShaderC
 	*eval *= *pdf;
 
 	*omega_in = X*localO.x + Y*localO.y + Z*localO.z;
+
 #ifdef __RAY_DIFFERENTIALS__
 	*domega_in_dx = (2 * dot(Z, dIdx)) * Z - dIdx;
 	*domega_in_dy = (2 * dot(Z, dIdy)) * Z - dIdy;
diff --git a/intern/cycles/kernel/closure/bsdf_microfacet_multi_impl.h b/intern/cycles/kernel/closure/bsdf_microfacet_multi_impl.h
index 2eb2457c9e5..e73915dbda7 100644
--- a/intern/cycles/kernel/closure/bsdf_microfacet_multi_impl.h
+++ b/intern/cycles/kernel/closure/bsdf_microfacet_multi_impl.h
@@ -100,11 +100,14 @@ ccl_device_forceinline float3 MF_FUNCTION_FULL_NAME(mf_eval)(
 	bool outside = true;
 
 	for(int order = 0; order < 10; order++) {
-		/* Sample microfacet height and normal */
-		if(!mf_sample_height(wr, &hr, &C1_r, &G1_r, &lambda_r, lcg_step_float_addrspace(lcg_state)))
+		/* Sample microfacet height. */
+		float height_rand = lcg_step_float_addrspace(lcg_state);
+		if(!mf_sample_height(wr, &hr, &C1_r, &G1_r, &lambda_r, height_rand))
 			break;
-		float3 wm = mf_sample_vndf(-wr, alpha, make_float2(lcg_step_float_addrspace(lcg_state),
-		                                                   lcg_step_float_addrspace(lcg_state)));
+		/* Sample microfacet normal. */
+		float vndf_rand_y = lcg_step_float_addrspace(lcg_state);
+		float vndf_rand_x = lcg_step_float_addrspace(lcg_state);
+		float3 wm = mf_sample_vndf(-wr, alpha, vndf_rand_x, vndf_rand_y);
 
 #ifdef MF_MULTI_GLASS
 		if(order == 0 && use_fresnel) {
@@ -136,7 +139,8 @@ ccl_device_forceinline float3 MF_FUNCTION_FULL_NAME(mf_eval)(
 #ifdef MF_MULTI_GLASS
 			bool next_outside;
 			float3 wi_prev = -wr;
-			wr = mf_sample_phase_glass(-wr, outside? eta: 1.0f/eta, wm, lcg_step_float_addrspace(lcg_state), &next_outside);
+			float phase_rand = lcg_step_float_addrspace(lcg_state);
+			wr = mf_sample_phase_glass(-wr, outside? eta: 1.0f/eta, wm, phase_rand, &next_outside);
 			if(!next_outside) {
 				outside = !outside;
 				wr = -wr;
@@ -204,14 +208,16 @@ ccl_device_forceinline float3 MF_FUNCTION_FULL_NAME(mf_sample)(
 	int order;
 	for(order = 0; order < 10; order++) {
 		/* Sample microfacet height. */
-		if(!mf_sample_height(wr, &hr, &C1_r, &G1_r, &lambda_r, lcg_step_float_addrspace(lcg_state))) {
+		float height_rand = lcg_step_float_addrspace(lcg_state);
+		if(!mf_sample_height(wr, &hr, &C1_r, &G1_r, &lambda_r, height_rand)) {
 			/* The random walk has left the surface. */
 			*wo = outside? wr: -wr;
 			return throughput;
 		}
 		/* Sample microfacet normal. */
-		float3 wm = mf_sample_vndf(-wr, alpha, make_float2(lcg_step_float_addrspace(lcg_state),
-		                                                   lcg_step_float_addrspace(lcg_state)));
+		float vndf_rand_y = lcg_step_float_addrspace(lcg_state);
+		float vndf_rand_x = lcg_step_float_addrspace(lcg_state);
+		float3 wm = mf_sample_vndf(-wr, alpha, vndf_rand_x, vndf_rand_y);
 
 		/* First-bounce color is already accounted for in mix weight. */
 		if(!use_fresnel && order > 0)
@@ -221,7 +227,8 @@ ccl_device_forceinline float3 MF_FUNCTION_FULL_NAME(mf_sample)(
 #ifdef MF_MULTI_GLASS
 		bool next_outside;
 		float3 wi_prev = -wr;
-		wr = mf_sample_phase_glass(-wr, outside? eta: 1.0f/eta, wm, lcg_step_float_addrspace(lcg_state), &next_outside);
+		float phase_rand = lcg_step_float_addrspace(lcg_state);
+		wr = mf_sample_phase_glass(-wr, outside? eta: 1.0f/eta, wm, phase_rand, &next_outside);
 		if(!next_outside) {
 			hr = -hr;
 			wr = -wr;