void thread_run(DeviceTask *task)
{
flush_texture_buffers();
if(task->type == DeviceTask::FILM_CONVERT) {
film_convert(*task, task->buffer, task->rgba_byte, task->rgba_half);
}
else if(task->type == DeviceTask::SHADER) {
shader(*task);
}
else if(task->type == DeviceTask::RENDER) {
RenderTile tile;
DenoisingTask denoising(this);
/* Allocate buffer for kernel globals */
device_only_memory<KernelGlobalsDummy> kgbuffer(this, "kernel_globals");
kgbuffer.alloc_to_device(1);
/* Keep rendering tiles until done. */
while(task->acquire_tile(this, tile)) {
if(tile.task == RenderTile::PATH_TRACE) {
assert(tile.task == RenderTile::PATH_TRACE);
scoped_timer timer(&tile.buffers->render_time);
split_kernel->path_trace(task,
tile,
kgbuffer,
*const_mem_map["__data"]);
/* Complete kernel execution before release tile. */
/* This helps in multi-device render;
* The device that reaches the critical-section function
* release_tile waits (stalling other devices from entering
* release_tile) for all kernels to complete. If device1 (a
* slow-render device) reaches release_tile first then it would
* stall device2 (a fast-render device) from proceeding to render
* next tile.
*/
clFinish(cqCommandQueue);
}
else if(tile.task == RenderTile::DENOISE) {
tile.sample = tile.start_sample + tile.num_samples;
denoise(tile, denoising, *task);
task->update_progress(&tile, tile.w*tile.h);
}
task->release_tile(tile);
}
kgbuffer.free();
}
}
virtual bool load_kernels(const DeviceRequestedFeatures& requested_features,
vector<OpenCLDeviceBase::OpenCLProgram*> &programs)
{
bool single_program = OpenCLInfo::use_single_program();
program_data_init = OpenCLDeviceBase::OpenCLProgram(this,
single_program ? "split" : "split_data_init",
single_program ? "kernel_split.cl" : "kernel_data_init.cl",
get_build_options(this, requested_features));
program_data_init.add_kernel(ustring("path_trace_data_init"));
programs.push_back(&program_data_init);
program_state_buffer_size = OpenCLDeviceBase::OpenCLProgram(this,
single_program ? "split" : "split_state_buffer_size",
single_program ? "kernel_split.cl" : "kernel_state_buffer_size.cl",
get_build_options(this, requested_features));
program_state_buffer_size.add_kernel(ustring("path_trace_state_buffer_size"));
programs.push_back(&program_state_buffer_size);
return split_kernel->load_kernels(requested_features);
}
void thread_run(DeviceTask *task)
{
if(task->type == DeviceTask::FILM_CONVERT) {
film_convert(*task, task->buffer, task->rgba_byte, task->rgba_half);
}
else if(task->type == DeviceTask::SHADER) {
shader(*task);
}
else if(task->type == DeviceTask::RENDER) {
RenderTile tile;
/* Copy dummy KernelGlobals related to OpenCL from kernel_globals.h to
* fetch its size.
*/
typedef struct KernelGlobals {
ccl_constant KernelData *data;
#define KERNEL_TEX(type, ttype, name) \
ccl_global type *name;
#include "kernel/kernel_textures.h"
#undef KERNEL_TEX
SplitData split_data;
SplitParams split_param_data;
} KernelGlobals;
/* Allocate buffer for kernel globals */
device_memory kgbuffer;
kgbuffer.resize(sizeof(KernelGlobals));
mem_alloc("kernel_globals", kgbuffer, MEM_READ_WRITE);
/* Keep rendering tiles until done. */
while(task->acquire_tile(this, tile)) {
if(tile.task == RenderTile::PATH_TRACE) {
assert(tile.task == RenderTile::PATH_TRACE);
split_kernel->path_trace(task,
tile,
kgbuffer,
*const_mem_map["__data"]);
/* Complete kernel execution before release tile. */
/* This helps in multi-device render;
* The device that reaches the critical-section function
* release_tile waits (stalling other devices from entering
* release_tile) for all kernels to complete. If device1 (a
* slow-render device) reaches release_tile first then it would
* stall device2 (a fast-render device) from proceeding to render
* next tile.
*/
clFinish(cqCommandQueue);
}
else if(tile.task == RenderTile::DENOISE) {
tile.sample = tile.start_sample + tile.num_samples;
denoise(tile, *task);
task->update_progress(&tile, tile.w*tile.h);
}
task->release_tile(tile);
}
mem_free(kgbuffer);
}
}
