[Bf-blender-cvs] [9c88fe08a00] split-kernel-faster-building: Cycles: Speed up split kernel builds by using split direct_emission functions
Mai Lavelle
noreply at git.blender.org
Sat Nov 11 12:48:32 CET 2017
Commit: 9c88fe08a00b5662eb84e1b40e06bb9ff6c4a510
Author: Mai Lavelle
Date: Sat Nov 11 06:36:07 2017 -0500
Branches: split-kernel-faster-building
https://developer.blender.org/rB9c88fe08a00b5662eb84e1b40e06bb9ff6c4a510
Cycles: Speed up split kernel builds by using split direct_emission functions
This removes the indirect call to `svm_eval_nodes` from the
`kernel_direct_lighting` kernel, and speeds up kernel build times
by 4-5 seconds. Render time is slower by ~1%. It should be possible
to apply this technique to other areas of the kernel to get even
better build times.
Note: this breaks branched path and shadow catcher in the split
kernel. Should be an easy fix, will do later.
===================================================================
M intern/cycles/device/device_split_kernel.cpp
M intern/cycles/kernel/kernel_light.h
M intern/cycles/kernel/kernel_types.h
M intern/cycles/kernel/split/kernel_direct_lighting.h
M intern/cycles/kernel/split/kernel_shadow_blocked_dl.h
M intern/cycles/kernel/split/kernel_split_data_types.h
===================================================================
diff --git a/intern/cycles/device/device_split_kernel.cpp b/intern/cycles/device/device_split_kernel.cpp
index 74135a1f3c8..bcc438c86c2 100644
--- a/intern/cycles/device/device_split_kernel.cpp
+++ b/intern/cycles/device/device_split_kernel.cpp
@@ -252,6 +252,7 @@ bool DeviceSplitKernel::path_trace(DeviceTask *task,
ENQUEUE_SPLIT_KERNEL(subsurface_scatter, global_size, local_size);
ENQUEUE_SPLIT_KERNEL(queue_enqueue, global_size, local_size);
ENQUEUE_SPLIT_KERNEL(direct_lighting, global_size, local_size);
+ ENQUEUE_SPLIT_KERNEL(shader_eval, global_size, local_size);
ENQUEUE_SPLIT_KERNEL(shadow_blocked_ao, global_size, local_size);
ENQUEUE_SPLIT_KERNEL(shadow_blocked_dl, global_size, local_size);
ENQUEUE_SPLIT_KERNEL(enqueue_inactive, global_size, local_size);
diff --git a/intern/cycles/kernel/kernel_light.h b/intern/cycles/kernel/kernel_light.h
index dfa3150dc92..1a3ae4c1f32 100644
--- a/intern/cycles/kernel/kernel_light.h
+++ b/intern/cycles/kernel/kernel_light.h
@@ -16,23 +16,6 @@
CCL_NAMESPACE_BEGIN
-/* Light Sample result */
-
-typedef struct LightSample {
- float3 P; /* position on light, or direction for distant light */
- float3 Ng; /* normal on light */
- float3 D; /* direction from shading point to light */
- float t; /* distance to light (FLT_MAX for distant light) */
- float u, v; /* parametric coordinate on primitive */
- float pdf; /* light sampling probability density function */
- float eval_fac; /* intensity multiplier */
- int object; /* object id for triangle/curve lights */
- int prim; /* primitive id for triangle/curve lights */
- int shader; /* shader id */
- int lamp; /* lamp id */
- LightType type; /* type of light */
-} LightSample;
-
/* Area light sampling */
/* Uses the following paper:
diff --git a/intern/cycles/kernel/kernel_types.h b/intern/cycles/kernel/kernel_types.h
index 6b0b86ad257..9032f847ec3 100644
--- a/intern/cycles/kernel/kernel_types.h
+++ b/intern/cycles/kernel/kernel_types.h
@@ -1513,6 +1513,23 @@ typedef struct WorkTile {
ccl_global float *buffer;
} WorkTile;
+/* Light Sample result */
+
+typedef struct LightSample {
+ float3 P; /* position on light, or direction for distant light */
+ float3 Ng; /* normal on light */
+ float3 D; /* direction from shading point to light */
+ float t; /* distance to light (FLT_MAX for distant light) */
+ float u, v; /* parametric coordinate on primitive */
+ float pdf; /* light sampling probability density function */
+ float eval_fac; /* intensity multiplier */
+ int object; /* object id for triangle/curve lights */
+ int prim; /* primitive id for triangle/curve lights */
+ int shader; /* shader id */
+ int lamp; /* lamp id */
+ LightType type; /* type of light */
+} LightSample;
+
/* Utility macro to get a pointer to an object that can be used locally, while avoiding
* address space issues of the split kernel. `name` must exist in split data entries.
*/
diff --git a/intern/cycles/kernel/split/kernel_direct_lighting.h b/intern/cycles/kernel/split/kernel_direct_lighting.h
index ca79602c565..176c0ec16ad 100644
--- a/intern/cycles/kernel/split/kernel_direct_lighting.h
+++ b/intern/cycles/kernel/split/kernel_direct_lighting.h
@@ -68,14 +68,14 @@ ccl_device void kernel_direct_lighting(KernelGlobals *kg,
# ifdef __BRANCHED_PATH__
if(flag && kernel_data.integrator.branched) {
flag = false;
- enqueue_flag = 1;
+ enqueue_flag = 1; // XXX
}
# endif /* __BRANCHED_PATH__ */
# ifdef __SHADOW_TRICKS__
if(flag && state->flag & PATH_RAY_SHADOW_CATCHER) {
flag = false;
- enqueue_flag = 1;
+ enqueue_flag = 1; // XXX
}
# endif /* __SHADOW_TRICKS__ */
@@ -83,7 +83,6 @@ ccl_device void kernel_direct_lighting(KernelGlobals *kg,
/* Sample illumination from lights to find path contribution. */
float light_u, light_v;
path_state_rng_2D(kg, state, PRNG_LIGHT_U, &light_u, &light_v);
- float terminate = path_state_rng_light_termination(kg, state);
LightSample ls;
if(light_sample(kg,
@@ -91,29 +90,17 @@ ccl_device void kernel_direct_lighting(KernelGlobals *kg,
sd->time,
sd->P,
state->bounce,
- &ls)) {
-
- Ray light_ray;
- light_ray.time = sd->time;
-
- BsdfEval L_light;
- bool is_lamp;
- if(direct_emission(kg,
- sd,
- AS_SHADER_DATA(&kernel_split_state.sd_DL_shadow[ray_index]),
- &ls,
- state,
- &light_ray,
- &L_light,
- &is_lamp,
- terminate))
- {
+ &ls))
+ {
+ ShaderData *emission_sd = AS_SHADER_DATA(&kernel_split_state.sd_DL_shadow[ray_index]);
+ ShaderEvalTask *eval_task = &kernel_split_state.shader_eval_task[ray_index];
+
+ if(direct_emission_setup(kg, sd, emission_sd, &ls, state, eval_task)) {
/* Write intermediate data to global memory to access from
* the next kernel.
*/
- kernel_split_state.light_ray[ray_index] = light_ray;
- kernel_split_state.bsdf_eval[ray_index] = L_light;
- kernel_split_state.is_lamp[ray_index] = is_lamp;
+ kernel_split_state.light_sample[ray_index] = ls;
+
/* Mark ray state for next shadow kernel. */
enqueue_flag = 1;
}
@@ -131,6 +118,10 @@ ccl_device void kernel_direct_lighting(KernelGlobals *kg,
local_queue_atomics,
kernel_split_state.queue_data,
kernel_split_params.queue_index);
+
+ if(ccl_global_id(1) * ccl_global_size(0) + ccl_global_id(0) == 0) {
+ kernel_split_params.shader_eval_queue = QUEUE_SHADOW_RAY_CAST_DL_RAYS;
+ }
#endif
#ifdef __BRANCHED_PATH__
diff --git a/intern/cycles/kernel/split/kernel_shadow_blocked_dl.h b/intern/cycles/kernel/split/kernel_shadow_blocked_dl.h
index da072fd5f1a..cde46d2250d 100644
--- a/intern/cycles/kernel/split/kernel_shadow_blocked_dl.h
+++ b/intern/cycles/kernel/split/kernel_shadow_blocked_dl.h
@@ -41,14 +41,24 @@ ccl_device void kernel_shadow_blocked_dl(KernelGlobals *kg)
return;
ccl_global PathState *state = &kernel_split_state.path_state[ray_index];
- Ray ray = kernel_split_state.light_ray[ray_index];
PathRadiance *L = &kernel_split_state.path_radiance[ray_index];
ShaderData *sd = kernel_split_sd(sd, ray_index);
float3 throughput = kernel_split_state.throughput[ray_index];
-
- BsdfEval L_light = kernel_split_state.bsdf_eval[ray_index];
ShaderData *emission_sd = AS_SHADER_DATA(&kernel_split_state.sd_DL_shadow[ray_index]);
- bool is_lamp = kernel_split_state.is_lamp[ray_index];
+ LightSample ls = kernel_split_state.light_sample[ray_index];
+ ShaderEvalTask *eval_task = &kernel_split_state.shader_eval_task[ray_index];
+
+ float terminate = path_state_rng_light_termination(kg, state);
+
+ Ray ray;
+ ray.time = sd->time;
+
+ BsdfEval L_light;
+ bool is_lamp;
+
+ if(!direct_emission_finish(kg, sd, emission_sd, &ls, state, &ray, &L_light, &is_lamp, terminate, eval_task)) {
+ return;
+ }
# if defined(__BRANCHED_PATH__) || defined(__SHADOW_TRICKS__)
bool use_branched = false;
diff --git a/intern/cycles/kernel/split/kernel_split_data_types.h b/intern/cycles/kernel/split/kernel_split_data_types.h
index ae5b7d15692..8698ae7e9d5 100644
--- a/intern/cycles/kernel/split/kernel_split_data_types.h
+++ b/intern/cycles/kernel/split/kernel_split_data_types.h
@@ -110,9 +110,7 @@ typedef ccl_global struct SplitBranchedState {
SPLIT_DATA_ENTRY(ccl_global Ray, ray, 1) \
SPLIT_DATA_ENTRY(ccl_global PathState, path_state, 1) \
SPLIT_DATA_ENTRY(ccl_global Intersection, isect, 1) \
- SPLIT_DATA_ENTRY(ccl_global BsdfEval, bsdf_eval, 1) \
- SPLIT_DATA_ENTRY(ccl_global int, is_lamp, 1) \
- SPLIT_DATA_ENTRY(ccl_global Ray, light_ray, 1) \
+ SPLIT_DATA_ENTRY(ccl_global LightSample, light_sample, 1) \
SPLIT_DATA_ENTRY(ShaderEvalTask, shader_eval_task, 1) \
SPLIT_DATA_ENTRY(ccl_global int, queue_data, (NUM_QUEUES*2)) /* TODO(mai): this is too large? */ \
SPLIT_DATA_ENTRY(ccl_global uint, buffer_offset, 1) \
@@ -129,9 +127,6 @@ typedef ccl_global struct SplitBranchedState {
SPLIT_DATA_ENTRY(ccl_global Ray, ray, 1) \
SPLIT_DATA_ENTRY(ccl_global PathState, path_state, 1) \
SPLIT_DATA_ENTRY(ccl_global Intersection, isect, 1) \
- SPLIT_DATA_ENTRY(ccl_global BsdfEval, bsdf_eval, 1) \
- SPLIT_DATA_ENTRY(ccl_global int, is_lamp, 1) \
- SPLIT_DATA_ENTRY(ccl_global Ray, light_ray, 1) \
SPLIT_DATA_ENTRY(ShaderDataTinyStorage, sd_DL_shadow, 1) \
SPLIT_DATA_SUBSURFACE_ENTRIES \
SPLIT_DATA_VOLUME_ENTRIES \
More information about the Bf-blender-cvs
mailing list