[Bf-blender-cvs] [148795bf01] temp_cycles_split_kernel: Cycles: OpenCL split kernel refactor
Mai Lavelle
noreply at git.blender.org
Wed Feb 22 15:59:20 CET 2017
Commit: 148795bf0128a1e78f11d8683f2454d3c656d72d
Author: Mai Lavelle
Date: Wed Feb 22 08:10:02 2017 -0500
Branches: temp_cycles_split_kernel
https://developer.blender.org/rB148795bf0128a1e78f11d8683f2454d3c656d72d
Cycles: OpenCL split kernel refactor
This does a few things at once:
- Refactors host side split kernel logic into a new device
agnostic class `DeviceSplitKernel`.
- Removes tile splitting, a new work pool implementation takes its place and
allows as many threads as will fit in memory regardless of tile size, which
can give performance gains.
- Refactors split state buffers into one buffer, as well as reduces the
number of arguments passed to kernels. Means there's less code to deal
with overall.
- Moves kernel logic out of OpenCL kernel files so they can later be used by
other device types.
- Replaced OpenCL specific APIs with new generic versions
- Tiles can now be seen updating during rendering
===================================================================
M intern/cycles/device/CMakeLists.txt
A intern/cycles/device/device_split_kernel.cpp
A intern/cycles/device/device_split_kernel.h
M intern/cycles/device/opencl/opencl.h
M intern/cycles/device/opencl/opencl_base.cpp
M intern/cycles/device/opencl/opencl_mega.cpp
M intern/cycles/device/opencl/opencl_split.cpp
M intern/cycles/kernel/CMakeLists.txt
M intern/cycles/kernel/kernel_compat_opencl.h
M intern/cycles/kernel/kernel_globals.h
M intern/cycles/kernel/kernel_passes.h
M intern/cycles/kernel/kernel_queues.h
M intern/cycles/kernel/kernel_shadow.h
M intern/cycles/kernel/kernel_types.h
M intern/cycles/kernel/kernel_work_stealing.h
M intern/cycles/kernel/kernels/opencl/kernel.cl
M intern/cycles/kernel/kernels/opencl/kernel_background_buffer_update.cl
M intern/cycles/kernel/kernels/opencl/kernel_data_init.cl
M intern/cycles/kernel/kernels/opencl/kernel_direct_lighting.cl
M intern/cycles/kernel/kernels/opencl/kernel_holdout_emission_blurring_pathtermination_ao.cl
M intern/cycles/kernel/kernels/opencl/kernel_lamp_emission.cl
M intern/cycles/kernel/kernels/opencl/kernel_next_iteration_setup.cl
M intern/cycles/kernel/kernels/opencl/kernel_queue_enqueue.cl
M intern/cycles/kernel/kernels/opencl/kernel_scene_intersect.cl
M intern/cycles/kernel/kernels/opencl/kernel_shader_eval.cl
M intern/cycles/kernel/kernels/opencl/kernel_shadow_blocked.cl
M intern/cycles/kernel/kernels/opencl/kernel_sum_all_radiance.cl
M intern/cycles/kernel/split/kernel_background_buffer_update.h
M intern/cycles/kernel/split/kernel_data_init.h
M intern/cycles/kernel/split/kernel_direct_lighting.h
M intern/cycles/kernel/split/kernel_holdout_emission_blurring_pathtermination_ao.h
M intern/cycles/kernel/split/kernel_lamp_emission.h
M intern/cycles/kernel/split/kernel_next_iteration_setup.h
A intern/cycles/kernel/split/kernel_queue_enqueue.h
M intern/cycles/kernel/split/kernel_scene_intersect.h
M intern/cycles/kernel/split/kernel_shader_eval.h
M intern/cycles/kernel/split/kernel_shadow_blocked.h
M intern/cycles/kernel/split/kernel_split_common.h
A intern/cycles/kernel/split/kernel_split_data.h
M intern/cycles/kernel/split/kernel_sum_all_radiance.h
M intern/cycles/util/util_types.h
===================================================================
diff --git a/intern/cycles/device/CMakeLists.txt b/intern/cycles/device/CMakeLists.txt
index 966ff5e52b..a237345169 100644
--- a/intern/cycles/device/CMakeLists.txt
+++ b/intern/cycles/device/CMakeLists.txt
@@ -3,6 +3,7 @@ set(INC
.
../graph
../kernel
+ ../kernel/split
../kernel/svm
../kernel/osl
../util
@@ -33,6 +34,7 @@ set(SRC
device_cuda.cpp
device_multi.cpp
device_opencl.cpp
+ device_split_kernel.cpp
device_task.cpp
)
@@ -56,6 +58,7 @@ set(SRC_HEADERS
device_memory.h
device_intern.h
device_network.h
+ device_split_kernel.h
device_task.h
)
diff --git a/intern/cycles/device/device_split_kernel.cpp b/intern/cycles/device/device_split_kernel.cpp
new file mode 100644
index 0000000000..cf43e499d0
--- /dev/null
+++ b/intern/cycles/device/device_split_kernel.cpp
@@ -0,0 +1,283 @@
+/*
+ * Copyright 2011-2016 Blender Foundation
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "device_split_kernel.h"
+
+#include "kernel_types.h"
+#include "kernel_split_data.h"
+
+#include "util_time.h"
+
+CCL_NAMESPACE_BEGIN
+
+static const double alpha = 0.1; /* alpha for rolling average */
+
+DeviceSplitKernel::DeviceSplitKernel(Device *device) : device(device)
+{
+ current_max_closure = -1;
+ first_tile = true;
+
+ avg_time_per_sample = 0.0;
+}
+
+DeviceSplitKernel::~DeviceSplitKernel()
+{
+ device->mem_free(split_data);
+ device->mem_free(ray_state);
+ device->mem_free(use_queues_flag);
+ device->mem_free(queue_index);
+ device->mem_free(work_pool_wgs);
+
+ delete kernel_scene_intersect;
+ delete kernel_lamp_emission;
+ delete kernel_queue_enqueue;
+ delete kernel_background_buffer_update;
+ delete kernel_shader_eval;
+ delete kernel_holdout_emission_blurring_pathtermination_ao;
+ delete kernel_direct_lighting;
+ delete kernel_shadow_blocked;
+ delete kernel_next_iteration_setup;
+ delete kernel_sum_all_radiance;
+}
+
+bool DeviceSplitKernel::load_kernels(const DeviceRequestedFeatures& requested_features)
+{
+#define LOAD_KERNEL(name) \
+ kernel_##name = get_split_kernel_function(#name, requested_features); \
+ if(!kernel_##name) { \
+ return false; \
+ }
+
+ LOAD_KERNEL(scene_intersect);
+ LOAD_KERNEL(lamp_emission);
+ LOAD_KERNEL(queue_enqueue);
+ LOAD_KERNEL(background_buffer_update);
+ LOAD_KERNEL(shader_eval);
+ LOAD_KERNEL(holdout_emission_blurring_pathtermination_ao);
+ LOAD_KERNEL(direct_lighting);
+ LOAD_KERNEL(shadow_blocked);
+ LOAD_KERNEL(next_iteration_setup);
+ LOAD_KERNEL(sum_all_radiance);
+
+#undef LOAD_KERNEL
+
+ current_max_closure = requested_features.max_closure;
+
+ return true;
+}
+
+size_t DeviceSplitKernel::max_elements_for_max_buffer_size(size_t max_buffer_size, size_t passes_size)
+{
+ size_t size_per_element = split_data_buffer_size(1024, current_max_closure, passes_size) / 1024;
+ return max_buffer_size / size_per_element;
+}
+
+bool DeviceSplitKernel::path_trace(DeviceTask *task,
+ RenderTile& tile,
+ device_memory& kgbuffer,
+ device_memory& kernel_data)
+{
+ if(device->have_error()) {
+ return false;
+ }
+
+ /* Get local size */
+ size_t local_size[2];
+ {
+ int2 lsize = split_kernel_local_size();
+ local_size[0] = lsize[0];
+ local_size[1] = lsize[1];
+ }
+
+ /* Calculate per_thread_output_buffer_size. */
+ size_t per_thread_output_buffer_size = task->passes_size;
+
+ /* Set gloabl size */
+ size_t global_size[2];
+ {
+ int2 gsize = split_kernel_global_size(task);
+
+ /* Make sure that set work size is a multiple of local
+ * work size dimensions.
+ */
+ global_size[0] = round_up(gsize[0], local_size[0]);
+ global_size[1] = round_up(gsize[1], local_size[1]);
+ }
+
+ /* Number of elements in the global state buffer */
+ int num_global_elements = global_size[0] * global_size[1];
+
+ /* Allocate all required global memory once. */
+ if(first_tile) {
+ first_tile = false;
+
+ /* Calculate max groups */
+
+ /* Denotes the maximum work groups possible w.r.t. current requested tile size. */
+ unsigned int max_work_groups = num_global_elements / WORK_POOL_SIZE + 1;
+
+ /* Allocate work_pool_wgs memory. */
+ work_pool_wgs.resize(max_work_groups * sizeof(unsigned int));
+ device->mem_alloc(work_pool_wgs, MEM_READ_WRITE);
+
+ queue_index.resize(NUM_QUEUES * sizeof(int));
+ device->mem_alloc(queue_index, MEM_READ_WRITE);
+
+ use_queues_flag.resize(sizeof(char));
+ device->mem_alloc(use_queues_flag, MEM_READ_WRITE);
+
+ ray_state.resize(num_global_elements);
+ device->mem_alloc(ray_state, MEM_READ_WRITE);
+
+ split_data.resize(split_data_buffer_size(num_global_elements,
+ current_max_closure,
+ per_thread_output_buffer_size));
+ device->mem_alloc(split_data, MEM_READ_WRITE);
+ }
+
+#define ENQUEUE_SPLIT_KERNEL(name, global_size, local_size) \
+ if(device->have_error()) { \
+ return false; \
+ } \
+ if(!kernel_##name->enqueue(KernelDimensions(global_size, local_size), kgbuffer, kernel_data)) { \
+ return false; \
+ }
+
+ tile.sample = tile.start_sample;
+
+ /* for exponential increase between tile updates */
+ int time_multiplier = 1;
+
+ while(tile.sample < tile.start_sample + tile.num_samples) {
+ /* to keep track of how long it takes to run a number of samples */
+ double start_time = time_dt();
+
+ /* initial guess to start rolling average */
+ const int initial_num_samples = 1;
+ /* approx number of samples per second */
+ int samples_per_second = (avg_time_per_sample > 0.0) ?
+ int(double(time_multiplier) / avg_time_per_sample) + 1 : initial_num_samples;
+
+ RenderTile subtile = tile;
+ subtile.start_sample = tile.sample;
+ subtile.num_samples = min(samples_per_second, tile.start_sample + tile.num_samples - tile.sample);
+
+ if(device->have_error()) {
+ return false;
+ }
+
+ /* reset state memory here as global size for data_init
+ * kernel might not be large enough to do in kernel
+ */
+ device->mem_zero(work_pool_wgs);
+ device->mem_zero(split_data);
+
+ if(!enqueue_split_kernel_data_init(KernelDimensions(global_size, local_size),
+ subtile,
+ num_global_elements,
+ kgbuffer,
+ kernel_data,
+ split_data,
+ ray_state,
+ queue_index,
+ use_queues_flag,
+ work_pool_wgs
+ ))
+ {
+ return false;
+ }
+
+ bool activeRaysAvailable = true;
+
+ while(activeRaysAvailable) {
+ /* Twice the global work size of other kernels for
+ * ckPathTraceKernel_shadow_blocked_direct_lighting. */
+ size_t global_size_shadow_blocked[2];
+ global_size_shadow_blocked[0] = global_size[0] * 2;
+ global_size_shadow_blocked[1] = global_size[1];
+
+ /* Do path-iteration in host [Enqueue Path-iteration kernels. */
+ for(int PathIter = 0; PathIter < 16; PathIter++) {
+ ENQUEUE_SPLIT_KERNEL(scene_intersect, global_size, local_size);
+ ENQUEUE_SPLIT_KERNEL(lamp_emission, global_size, local_size);
+ ENQUEUE_SPLIT_KERNEL(queue_enqueue, global_size, local_size);
+ ENQUEUE_SPLIT_KERNEL(background_buffer_update, global_size, local_size);
+ ENQUEUE_SPLIT_KERNEL(shader_eval, global_size, local_size);
+ ENQUEUE_SPLIT_KERNEL(holdout_emission_blurring_pathtermination_ao, global_size, local_size);
+ ENQUEUE_SPLIT_KERNEL(direct_lighting, global_size, local_size);
+ ENQUEUE_SPLIT_KERNEL(shadow_blocked, global_size_shadow_blocked, local_size);
+ ENQUEUE_SPLIT_KERNEL(next_iteration_setup, global_size, local_size);
+
+ if(task->get_cancel()) {
+ return true;
+ }
+ }
+
+ /* Decide if we should exit path-iteration in host. */
+ device->mem_copy_from(ray_state, 0, global_size[0] * global_size[1] * sizeof(char), 1, 1);
+
+ activeRaysAvailable = false;
+
+ for(int rayStateIter = 0; rayStateIter < global_size[0] * global_size[1]; ++rayStateIter) {
+ if(int8_t(ray_state.get_data()[rayStateIter]) != RAY_INACTIVE) {
+ /* Not all rays are RAY_INACTIVE. */
+ activeRaysAvailable = true;
+ break;
+ }
+ }
+
+ if(task->get_cancel()) {
+ return true;
+ }
+ }
+
+ double time_per_sample = ((time_dt()-start_time) / subtile.num_samples);
+
+ if(avg_time_per_sample == 0.0) {
+ /* start rolling average */
+ avg_time_per_sample = time_per_sample;
+ }
+ else {
+ avg_time_per_sample = alpha*time_per_sample + (1.0-alpha)*avg_time_per_sample;
+ }
+
+ size_t sum_all_radiance_local_size[2] = {16, 16};
+ size_t sum_all_radiance_global_size[2];
+ sum_all_radiance_global_size[0] = round_up(tile.w, sum_all_radiance_local_size[0]);
+ sum_all_radiance_global_size[1] = round_up(tile.h, sum_all_radiance_local_size[1]);
+
+ ENQUEUE_SPLIT_KERNEL(sum_all_radiance,
+ sum_all_radiance_global_size,
+ sum_all_radiance_local_size);
+
+#undef ENQUEUE_SPLIT_KERNEL
+
+ tile.sample += subtile.num_samples;
+ task->update_progress(&tile, tile.w*tile.h*subtile.num_samples);
+
+ time_multiplier = min(time_multiplier << 1, 10);
+
+ if(task->get_cancel()) {
+ return true;
+ }
+ }
+
+ return true;
+}
+
+CCL_NAMESPACE_END
+
+
diff --git a/intern/cycles/device/device_split_kernel.h b/intern/cycles/device/device_split_kernel.h
new file mode 100644
index 0000000000..b3106fd563
--- /dev/null
+++ b/intern/cycles/device/device_split_ke
@@ Diff output truncated at 10240 characters. @@
More information about the Bf-blender-cvs
mailing list