GPU: Add indirect dispatch support.

This uses a StorageBuf as the source of indirect dispatch argument.
The user needs to make sure the parameters are in the right order.

There is no support for argument offset for the moment as there is no
need for it. But this might be added in the future.

Note that the indirect buffer is synchronized at the backend level. This is
done for practical reasons and because this feature is almost always used
for GPU driven pipeline.
This commit is contained in:
Clément Foucault 2022-03-16 08:58:59 +01:00
parent 7ee816e32f
commit d7df0dcccb
7 changed files with 36 additions and 8 deletions

View File

@ -9,6 +9,7 @@
#include "BLI_sys_types.h"
#include "GPU_shader.h"
#include "GPU_storage_buffer.h"
#ifdef __cplusplus
extern "C" {
@ -19,6 +20,8 @@ void GPU_compute_dispatch(GPUShader *shader,
uint groups_y_len,
uint groups_z_len);
void GPU_compute_dispatch_indirect(GPUShader *shader, GPUStorageBuf *indirect_buf);
#ifdef __cplusplus
}
#endif

View File

@ -35,6 +35,7 @@ class GPUBackend {
virtual void samplers_update() = 0;
virtual void compute_dispatch(int groups_x_len, int groups_y_len, int groups_z_len) = 0;
virtual void compute_dispatch_indirect(StorageBuf *indirect_buf) = 0;
virtual Context *context_alloc(void *ghost_window) = 0;

View File

@ -7,10 +7,7 @@
#include "GPU_compute.h"
#include "gpu_backend.hh"
#ifdef __cplusplus
extern "C" {
#endif
#include "gpu_storage_buffer_private.hh"
void GPU_compute_dispatch(GPUShader *shader,
uint groups_x_len,
@ -22,6 +19,12 @@ void GPU_compute_dispatch(GPUShader *shader,
gpu_backend.compute_dispatch(groups_x_len, groups_y_len, groups_z_len);
}
#ifdef __cplusplus
void GPU_compute_dispatch_indirect(GPUShader *shader, GPUStorageBuf *indirect_buf_)
{
blender::gpu::GPUBackend &gpu_backend = *blender::gpu::GPUBackend::get();
blender::gpu::StorageBuf *indirect_buf = reinterpret_cast<blender::gpu::StorageBuf *>(
indirect_buf_);
GPU_shader_bind(shader);
gpu_backend.compute_dispatch_indirect(indirect_buf);
}
#endif

View File

@ -123,6 +123,19 @@ class GLBackend : public GPUBackend {
GLCompute::dispatch(groups_x_len, groups_y_len, groups_z_len);
}
void compute_dispatch_indirect(StorageBuf *indirect_buf) override
{
GLContext::get()->state_manager_active_get()->apply_state();
dynamic_cast<GLStorageBuf *>(indirect_buf)->bind_as(GL_DISPATCH_INDIRECT_BUFFER);
/* This barrier needs to be here as it only work on the currently bound indirect buffer. */
glMemoryBarrier(GL_DRAW_INDIRECT_BUFFER);
glDispatchComputeIndirect((GLintptr)0);
/* Unbind. */
glBindBuffer(GL_DRAW_INDIRECT_BUFFER, 0);
}
private:
static void platform_init();
static void platform_exit();

View File

@ -16,8 +16,6 @@ void GLCompute::dispatch(int group_x_len, int group_y_len, int group_z_len)
{
GL_CHECK_RESOURCES("Compute");
GLContext::get()->state_manager->apply_state();
glDispatchCompute(group_x_len, group_y_len, group_z_len);
}

View File

@ -100,6 +100,13 @@ void GLStorageBuf::bind(int slot)
#endif
}
void GLStorageBuf::bind_as(GLenum target)
{
BLI_assert_msg(ssbo_id_ != 0,
"Trying to use storage buf as indirect buffer but buffer was never filled.");
glBindBuffer(target, ssbo_id_);
}
void GLStorageBuf::unbind()
{
#ifdef DEBUG

View File

@ -37,6 +37,9 @@ class GLStorageBuf : public StorageBuf {
void unbind() override;
void clear(eGPUTextureFormat internal_format, eGPUDataFormat data_format, void *data) override;
/* Special internal function to bind SSBOs to indirect argument targets. */
void bind_as(GLenum target);
private:
void init();