Cycles: Fix crash with split branched path tracing

ShaderData memory was getting clobbered in the branched path code paths.

Was caused by 087331c495b04ebd37903c0dc0e46262354cf026

5 files changed, 22 insertions, 16 deletions

diff --git a/intern/cycles/kernel/split/kernel_branched.h b/intern/cycles/kernel/split/kernel_branched.h
index 6456636caaa..368a4395760 100644
--- a/intern/cycles/kernel/split/kernel_branched.h
+++ b/intern/cycles/kernel/split/kernel_branched.h
@@ -33,9 +33,9 @@ ccl_device_inline void kernel_split_branched_path_indirect_loop_init(KernelGloba
 	BRANCHED_STORE(isect);
 	BRANCHED_STORE(ray_state);
 
-	branched_state->sd = *kernel_split_sd(sd, ray_index);
-	for(int i = 0; i < branched_state->sd.num_closure; i++) {
-		branched_state->sd.closure[i] = kernel_split_sd(sd, ray_index)->closure[i];
+	*kernel_split_sd(branched_state_sd, ray_index) = *kernel_split_sd(sd, ray_index);
+	for(int i = 0; i < kernel_split_sd(branched_state_sd, ray_index)->num_closure; i++) {
+		kernel_split_sd(branched_state_sd, ray_index)->closure[i] = kernel_split_sd(sd, ray_index)->closure[i];
 	}
 
 #undef BRANCHED_STORE
@@ -60,9 +60,9 @@ ccl_device_inline void kernel_split_branched_path_indirect_loop_end(KernelGlobal
 	BRANCHED_RESTORE(isect);
 	BRANCHED_RESTORE(ray_state);
 
-	*kernel_split_sd(sd, ray_index) = branched_state->sd;
-	for(int i = 0; i < branched_state->sd.num_closure; i++) {
-		kernel_split_sd(sd, ray_index)->closure[i] = branched_state->sd.closure[i];
+	*kernel_split_sd(sd, ray_index) = *kernel_split_sd(branched_state_sd, ray_index);
+	for(int i = 0; i < kernel_split_sd(branched_state_sd, ray_index)->num_closure; i++) {
+		kernel_split_sd(sd, ray_index)->closure[i] = kernel_split_sd(branched_state_sd, ray_index)->closure[i];
 	}
 
 #undef BRANCHED_RESTORE
@@ -83,10 +83,17 @@ ccl_device_inline bool kernel_split_branched_indirect_start_shared(KernelGlobals
 	}
 
 #define SPLIT_DATA_ENTRY(type, name, num) \
-		kernel_split_state.name[inactive_ray] = kernel_split_state.name[ray_index];
+		if(num) { \
+			kernel_split_state.name[inactive_ray] = kernel_split_state.name[ray_index]; \
+		}
 	SPLIT_DATA_ENTRIES_BRANCHED_SHARED
 #undef SPLIT_DATA_ENTRY
 
+	*kernel_split_sd(sd, inactive_ray) = *kernel_split_sd(sd, ray_index);
+	for(int i = 0; i < kernel_split_sd(sd, ray_index)->num_closure; i++) {
+		kernel_split_sd(sd, inactive_ray)->closure[i] = kernel_split_sd(sd, ray_index)->closure[i];
+	}
+
 	kernel_split_state.branched_state[inactive_ray].shared_sample_count = 0;
 	kernel_split_state.branched_state[inactive_ray].original_ray = ray_index;
 	kernel_split_state.branched_state[inactive_ray].waiting_on_shared_samples = false;
diff --git a/intern/cycles/kernel/split/kernel_next_iteration_setup.h b/intern/cycles/kernel/split/kernel_next_iteration_setup.h
index bb6bf1cc7e6..75a0af7567b 100644
--- a/intern/cycles/kernel/split/kernel_next_iteration_setup.h
+++ b/intern/cycles/kernel/split/kernel_next_iteration_setup.h
@@ -145,7 +145,7 @@ ccl_device void kernel_next_iteration_setup(KernelGlobals *kg,
 			if(kernel_split_branched_path_surface_indirect_light_iter(kg,
 			                                                          ray_index,
 			                                                          1.0f,
-			                                                          &kernel_split_state.branched_state[ray_index].sd,
+			                                                          kernel_split_sd(branched_state_sd, ray_index),
 			                                                          true,
 			                                                          true))
 			{
@@ -190,7 +190,7 @@ ccl_device void kernel_next_iteration_setup(KernelGlobals *kg,
 		if(kernel_split_branched_path_surface_indirect_light_iter(kg,
 		                                                          ray_index,
 		                                                          1.0f,
-		                                                          &kernel_split_state.branched_state[ray_index].sd,
+		                                                          kernel_split_sd(branched_state_sd, ray_index),
 		                                                          true,
 		                                                          true))
 		{
diff --git a/intern/cycles/kernel/split/kernel_split_data.h b/intern/cycles/kernel/split/kernel_split_data.h
index fa2f0b20a83..9297e1e0ad5 100644
--- a/intern/cycles/kernel/split/kernel_split_data.h
+++ b/intern/cycles/kernel/split/kernel_split_data.h
@@ -34,7 +34,7 @@ ccl_device_inline uint64_t split_data_buffer_size(KernelGlobals *kg, size_t num_
 	uint64_t closure_size = sizeof(ShaderClosure) * (kernel_data.integrator.max_closures-1);
 
 #ifdef __BRANCHED_PATH__
-	size += align_up(closure_size * num_elements, 16);
+	size += align_up(num_elements * (sizeof(ShaderData) + closure_size), 16);
 #endif
 
 	size += align_up(num_elements * (sizeof(ShaderData) + closure_size), 16);
@@ -60,7 +60,8 @@ ccl_device_inline void split_data_init(KernelGlobals *kg,
 	uint64_t closure_size = sizeof(ShaderClosure) * (kernel_data.integrator.max_closures-1);
 
 #ifdef __BRANCHED_PATH__
-	p += align_up(closure_size * num_elements, 16);
+	split_data->_branched_state_sd = (ShaderData*)p;
+	p += align_up(num_elements * (sizeof(ShaderData) + closure_size), 16);
 #endif
 
 	split_data->_sd = (ShaderData*)p;
diff --git a/intern/cycles/kernel/split/kernel_split_data_types.h b/intern/cycles/kernel/split/kernel_split_data_types.h
index 9ac3f904819..5f40fdc9240 100644
--- a/intern/cycles/kernel/split/kernel_split_data_types.h
+++ b/intern/cycles/kernel/split/kernel_split_data_types.h
@@ -76,13 +76,11 @@ typedef ccl_global struct SplitBranchedState {
 	int shared_sample_count; /* number of branched samples shared with other threads */
 	int original_ray; /* index of original ray when sharing branched samples */
 	bool waiting_on_shared_samples;
-
-	/* Must be last in to allow for dynamic size of closures */
-	struct ShaderData sd;
 } SplitBranchedState;
 
 #define SPLIT_DATA_BRANCHED_ENTRIES \
-	SPLIT_DATA_ENTRY( SplitBranchedState, branched_state, 1)
+	SPLIT_DATA_ENTRY( SplitBranchedState, branched_state, 1) \
+	SPLIT_DATA_ENTRY(ShaderData, _branched_state_sd, 0)
 #else
 #define SPLIT_DATA_BRANCHED_ENTRIES
 #endif  /* __BRANCHED_PATH__ */
diff --git a/intern/cycles/kernel/split/kernel_subsurface_scatter.h b/intern/cycles/kernel/split/kernel_subsurface_scatter.h
index 887c3e313d1..5bf7483e9a2 100644
--- a/intern/cycles/kernel/split/kernel_subsurface_scatter.h
+++ b/intern/cycles/kernel/split/kernel_subsurface_scatter.h
@@ -37,7 +37,7 @@ ccl_device_noinline bool kernel_split_branched_path_subsurface_indirect_light_it
 {
 	SplitBranchedState *branched_state = &kernel_split_state.branched_state[ray_index];
 
-	ShaderData *sd = &branched_state->sd;
+	ShaderData *sd = kernel_split_sd(branched_state_sd, ray_index);
 	PathRadiance *L = &kernel_split_state.path_radiance[ray_index];
 	ShaderData *emission_sd = AS_SHADER_DATA(&kernel_split_state.sd_DL_shadow[ray_index]);