GPUCreateInfo Compile Time Descriptors #116594

New Issue

Clément Foucault · 2023-12-28T02:07:13+01:00

Clément Foucault commented

2023-12-28 02:07:13 +01:00

The aim of this task is to remove the need for string lookup for binding resources.
Another use of this would be to avoid missing resources.

While the overhead of the string lookups is debatable, avoiding missing resource bind by compilation error would be a great time saver.

This could be also the opportunity to create a system closer to what the backends use for tracking and uploading the resources but this isn't the priority.

Current system

The current way of avoiding the string lookup is to resort to macro definitions in a shared file like eevee_defines.h and use the same macros in both the shader interface and the binding code.

/* eevee_defines.h */
#define UNIFORM_BUF_SLOT 1

/* Create Info */
GPU_SHADER_CREATE_INFO(eevee_global_ubo)
    .uniform_buf(UNIFORM_BUF_SLOT, "UniformData", "uniform_buf");
    
/* Binding code */
pass->bind_ubo(UNIFORM_BUF_SLOT, &global_ubo_);

This is quite straightforward but you have to track the lifetime of each slot binding as the slots can be used differently by different shaders.
Also, not all slots indices are important. Only those reused across shaders are important and should use defines like this.

Proposed system

The **_info.hh files are quite standalone and could be preprocessed and compiled as an other executable that could generate the descriptor classes.

The following eevee_ray_tile_classify with its dependencies ...

GPU_SHADER_CREATE_INFO(eevee_gbuffer_data)
    .define("GBUFFER_LOAD")
    .sampler(8, ImageType::UINT_2D, "gbuf_header_tx")
    .sampler(9, ImageType::FLOAT_2D_ARRAY, "gbuf_closure_tx")
    .sampler(10, ImageType::FLOAT_2D_ARRAY, "gbuf_normal_tx");

GPU_SHADER_CREATE_INFO(eevee_global_ubo)
    .uniform_buf(UNIFORM_BUF_SLOT, "UniformData", "uniform_buf");

GPU_SHADER_CREATE_INFO(eevee_ray_tile_classify)
    .do_static_compilation(true)
    .local_group_size(RAYTRACE_GROUP_SIZE, RAYTRACE_GROUP_SIZE)
    .additional_info("eevee_shared", "eevee_gbuffer_data", "eevee_global_ubo")
    .typedef_source("draw_shader_shared.h")
    .image_out(0, RAYTRACE_TILEMASK_FORMAT, ImageType::UINT_2D_ARRAY, "tile_raytrace_denoise_img")
    .image_out(1, RAYTRACE_TILEMASK_FORMAT, ImageType::UINT_2D_ARRAY, "tile_raytrace_tracing_img")
    .image_out(2, RAYTRACE_TILEMASK_FORMAT, ImageType::UINT_2D_ARRAY, "tile_horizon_denoise_img")
    .image_out(3, RAYTRACE_TILEMASK_FORMAT, ImageType::UINT_2D_ARRAY, "tile_horizon_tracing_img")
    .compute_source("eevee_ray_tile_classify_comp.glsl");

... could generate the following class ...

namespace blender::gpu::shader::descriptor {

struct eevee_ray_tile_classify {
    /* From eevee_gbuffer_data */
    SamplerSlot gbuf_header_tx = {8};
    SamplerSlot gbuf_closure_tx = {9};
    SamplerSlot gbuf_normal_tx = {10};
    /* From eevee_global_ubo */
    UniformBufferSlot uniform_buf = {1};
    /* From eevee_ray_tile_classify */
    ImageSlot tile_raytrace_denoise_img = {0};
    ImageSlot tile_raytrace_tracing_img = {1};
    ImageSlot tile_horizon_denoise_img = {2};
    ImageSlot tile_horizon_tracing_img = {3};

    void bind(draw::PassBase<T> &pass) {
        if (debug) {
            BLI_assert(gbuf_header_tx.is_set)
            BLI_assert(gbuf_closure_tx.is_set)
            BLI_assert(gbuf_normal_tx.is_set)
            BLI_assert(uniform_buf.is_set)
            BLI_assert(tile_raytrace_denoise_img.is_set)
            BLI_assert(tile_raytrace_tracing_img.is_set)
            BLI_assert(tile_horizon_denoise_img.is_set)
            BLI_assert(tile_horizon_tracing_img.is_set)
        }

        gbuf_header_tx.bind(pass);
        gbuf_closure_tx.bind(pass);
        gbuf_normal_tx.bind(pass);
        uniform_buf.bind(pass);
        tile_raytrace_denoise_img.bind(pass);
        tile_raytrace_tracing_img.bind(pass);
        tile_horizon_denoise_img.bind(pass);
        tile_horizon_tracing_img.bind(pass);
    }
};

}

... with a slot structure looking like this.

struct ImageSlot {
  const int binding;
  bool is_set = false;

  bool is_reference;
  union {
    GPUTexture **ptr_ref;
    GPUTexture *ptr;
  };

  ImageSlot(int i) : binding(i){};

  void set(GPUTexture *ptr) { this->ptr = ptr; is_reference = false; };
  void set(GPUTexture **ptr_ref) { this->ptr_ref = ptr_ref; is_reference = true; };

  void bind(draw::PassBase<T> &pass)
  {
    if (is_reference) {
        pass.bind_image(binding, ptr_ref);
    }
    else {
        pass.bind_image(binding, ptr);
    }
  }
};

The user code would look like this:

gpu::shader::descriptor::eevee_ray_tile_classify desc();
desc.gbuf_header_tx.set(&inst_.gbuffer.header_tx);
desc.gbuf_closure_tx.set(&inst_.gbuffer.closure_tx);
desc.gbuf_normal_tx.set(&inst_.gbuffer.normal_tx);
desc.uniform_buf.set(&inst_.global_ubo_);
desc.tile_raytrace_denoise_img.set(&tile_raytrace_denoise_tx_);
desc.tile_raytrace_denoise_img.set(&tile_raytrace_tracing_tx_);
desc.tile_raytrace_denoise_img.set(&tile_horizon_denoise_tx_);
desc.tile_raytrace_denoise_img.set(&tile_horizon_tracing_tx_);
desc.tile_raytrace_denoise_img.set(&tile_horizon_tracing_tx_);

PassSimple &pass = tile_classify_ps_;
pass.init();
/* bind_descriptor calls desc.bind(pass).
 * Note that it can be called before the shader binding. */
pass.bind_descriptor(desc);

This have the benefit of creating a compilation error if any renaming in the create info is not reflected on the C++ side. More over, it adds missing resources check.

Another potential benefit of this is the possibility of completely phasing out the gpu::ShaderInterface. But that would require changing the C style API in the whole codebase.

The aim of this task is to remove the need for string lookup for binding resources. Another use of this would be to avoid missing resources. While the overhead of the string lookups is debatable, avoiding missing resource bind by compilation error would be a great time saver. This could be also the opportunity to create a system closer to what the backends use for tracking and uploading the resources but this isn't the priority. ### Current system The current way of avoiding the string lookup is to resort to macro definitions in a shared file like `eevee_defines.h` and use the same macros in both the shader interface and the binding code. ```cpp /* eevee_defines.h */ #define UNIFORM_BUF_SLOT 1 /* Create Info */ GPU_SHADER_CREATE_INFO(eevee_global_ubo) .uniform_buf(UNIFORM_BUF_SLOT, "UniformData", "uniform_buf"); /* Binding code */ pass->bind_ubo(UNIFORM_BUF_SLOT, &global_ubo_); ``` This is quite straightforward but you have to track the lifetime of each slot binding as the slots can be used differently by different shaders. Also, not all slots indices are important. Only those reused across shaders are important and should use defines like this. ### Proposed system The `**_info.hh` files are quite standalone and could be preprocessed and compiled as an other executable that could generate the descriptor classes. The following `eevee_ray_tile_classify` with its dependencies ... ```cpp GPU_SHADER_CREATE_INFO(eevee_gbuffer_data) .define("GBUFFER_LOAD") .sampler(8, ImageType::UINT_2D, "gbuf_header_tx") .sampler(9, ImageType::FLOAT_2D_ARRAY, "gbuf_closure_tx") .sampler(10, ImageType::FLOAT_2D_ARRAY, "gbuf_normal_tx"); GPU_SHADER_CREATE_INFO(eevee_global_ubo) .uniform_buf(UNIFORM_BUF_SLOT, "UniformData", "uniform_buf"); GPU_SHADER_CREATE_INFO(eevee_ray_tile_classify) .do_static_compilation(true) .local_group_size(RAYTRACE_GROUP_SIZE, RAYTRACE_GROUP_SIZE) .additional_info("eevee_shared", "eevee_gbuffer_data", "eevee_global_ubo") .typedef_source("draw_shader_shared.h") .image_out(0, RAYTRACE_TILEMASK_FORMAT, ImageType::UINT_2D_ARRAY, "tile_raytrace_denoise_img") .image_out(1, RAYTRACE_TILEMASK_FORMAT, ImageType::UINT_2D_ARRAY, "tile_raytrace_tracing_img") .image_out(2, RAYTRACE_TILEMASK_FORMAT, ImageType::UINT_2D_ARRAY, "tile_horizon_denoise_img") .image_out(3, RAYTRACE_TILEMASK_FORMAT, ImageType::UINT_2D_ARRAY, "tile_horizon_tracing_img") .compute_source("eevee_ray_tile_classify_comp.glsl"); ``` ... could generate the following class ... ```cpp namespace blender::gpu::shader::descriptor { struct eevee_ray_tile_classify { /* From eevee_gbuffer_data */ SamplerSlot gbuf_header_tx = {8}; SamplerSlot gbuf_closure_tx = {9}; SamplerSlot gbuf_normal_tx = {10}; /* From eevee_global_ubo */ UniformBufferSlot uniform_buf = {1}; /* From eevee_ray_tile_classify */ ImageSlot tile_raytrace_denoise_img = {0}; ImageSlot tile_raytrace_tracing_img = {1}; ImageSlot tile_horizon_denoise_img = {2}; ImageSlot tile_horizon_tracing_img = {3}; void bind(draw::PassBase<T> &pass) { if (debug) { BLI_assert(gbuf_header_tx.is_set) BLI_assert(gbuf_closure_tx.is_set) BLI_assert(gbuf_normal_tx.is_set) BLI_assert(uniform_buf.is_set) BLI_assert(tile_raytrace_denoise_img.is_set) BLI_assert(tile_raytrace_tracing_img.is_set) BLI_assert(tile_horizon_denoise_img.is_set) BLI_assert(tile_horizon_tracing_img.is_set) } gbuf_header_tx.bind(pass); gbuf_closure_tx.bind(pass); gbuf_normal_tx.bind(pass); uniform_buf.bind(pass); tile_raytrace_denoise_img.bind(pass); tile_raytrace_tracing_img.bind(pass); tile_horizon_denoise_img.bind(pass); tile_horizon_tracing_img.bind(pass); } }; } ``` ... with a slot structure looking like this. ```cpp struct ImageSlot { const int binding; bool is_set = false; bool is_reference; union { GPUTexture **ptr_ref; GPUTexture *ptr; }; ImageSlot(int i) : binding(i){}; void set(GPUTexture *ptr) { this->ptr = ptr; is_reference = false; }; void set(GPUTexture **ptr_ref) { this->ptr_ref = ptr_ref; is_reference = true; }; void bind(draw::PassBase<T> &pass) { if (is_reference) { pass.bind_image(binding, ptr_ref); } else { pass.bind_image(binding, ptr); } } }; ``` The user code would look like this: ```cpp gpu::shader::descriptor::eevee_ray_tile_classify desc(); desc.gbuf_header_tx.set(&inst_.gbuffer.header_tx); desc.gbuf_closure_tx.set(&inst_.gbuffer.closure_tx); desc.gbuf_normal_tx.set(&inst_.gbuffer.normal_tx); desc.uniform_buf.set(&inst_.global_ubo_); desc.tile_raytrace_denoise_img.set(&tile_raytrace_denoise_tx_); desc.tile_raytrace_denoise_img.set(&tile_raytrace_tracing_tx_); desc.tile_raytrace_denoise_img.set(&tile_horizon_denoise_tx_); desc.tile_raytrace_denoise_img.set(&tile_horizon_tracing_tx_); desc.tile_raytrace_denoise_img.set(&tile_horizon_tracing_tx_); PassSimple &pass = tile_classify_ps_; pass.init(); /* bind_descriptor calls desc.bind(pass). * Note that it can be called before the shader binding. */ pass.bind_descriptor(desc); ``` This have the benefit of creating a compilation error if any renaming in the create info is not reflected on the C++ side. More over, it adds missing resources check. Another potential benefit of this is the possibility of completely phasing out the `gpu::ShaderInterface`. But that would require changing the C style API in the whole codebase.

👍 1

Clément Foucault added the

Type

Design

label 2023-12-28 02:07:13 +01:00

Clément Foucault added the

Module

EEVEE & Viewport

label 2023-12-28 02:17:50 +01:00

Sign in to join this conversation.