Compute nodes

Supported bindings: ossia

Compute is simpler than draw, as the pipeline only has one shader (the compute shader). Instead of draw, the method in which to run compute dispatch calls is called dispatch.

Here is an example:

#pragma once
#include <avnd/common/member_reflection.hpp>
#include <fmt/format.h>
#include <fmt/printf.h>
#include <gpp/commands.hpp>
#include <gpp/meta.hpp>
#include <gpp/ports.hpp>
#include <halp/controls.hpp>
#include <halp/static_string.hpp>

#include <vector>
namespace examples
{

struct GpuComputeExample
{
  // halp_meta is a short hand for defining a static function:
  // #define halp_meta(name, val) static constexpr auto name() return { val; }
  halp_meta(name, "Average color");
  halp_meta(uuid, "03bce361-a2ca-4959-95b4-6aac3b6c07b5");
  halp_meta(category, "Visuals/Analysis")
  halp_meta(c_name, "average_color")
  halp_meta(author, "Jean-Michaël Celerier")
  halp_meta(
      manual_url,
      "https://ossia.io/score-docs/processes/"
      "computer-vision-utilities.html#average-color")
  halp_meta(description, "Extract the average color of an input video feed")

  static constexpr int downscale = 16;

  // Define the layout of our pipeline in C++ simply through the structure of a struct
  struct layout
  {
    halp_meta(local_size_x, 16)
    halp_meta(local_size_y, 16)
    halp_meta(local_size_z, 1)
    halp_flags(compute);

    struct bindings
    {
      // Each binding is a struct member
      struct
      {
        halp_meta(name, "my_buf");
        halp_meta(binding, 0);
        halp_flags(std140, buffer, load, store);

        using color = float[4];
        gpp::uniform<"result", color*> values;
      } my_buf;

      // Define the members of our ubos
      struct custom_ubo
      {
        halp_meta(name, "custom");
        halp_meta(binding, 1);
        halp_flags(std140, ubo);

        gpp::uniform<"width", int> width;
        gpp::uniform<"height", int> height;
      } ubo;

      struct
      {
        halp_meta(name, "img")
        halp_meta(format, "rgba32f")
        halp_meta(binding, 2);
        halp_flags(image2D, readonly);
      } image;
    } bindings;
  };

  using bindings = decltype(layout::bindings);
  using uniforms = decltype(bindings::ubo);

  // Definition of our ports which will get parsed by the
  // software that instantiate this class
  struct
  {
    // Here we use some helper types in the usual fashion
    gpp::image_input_port<"Image", &bindings::image> tex;

    gpp::uniform_control_port<
        halp::hslider_i32<"Width", halp::range{1, 1000, 100}>, &uniforms::width>
        width;

    gpp::uniform_control_port<
        halp::hslider_i32<"Height", halp::range{1, 1000, 100}>, &uniforms::height>
        height;
  } inputs;

  // The output port on which we write the average color
  struct
  {
    struct
    {
      halp_meta(name, "color")
      float value[4];
    } color_out;
  } outputs;

  std::string_view compute()
  {
    return R"_(
void main()
{
  // Note: the algorithm is most likely wrong as I know FUCK ALL
  // about compute shaders ; fixes welcome ;p

  ivec2 call = ivec2(gl_GlobalInvocationID.xy);
  vec4 color = vec4(0.0, 0.0,0,0);

  for(int i = 0; i < gl_WorkGroupSize.x; i++)
  {
    for(int j = 0; j < gl_WorkGroupSize.y; j++)
    {
      uint x = call.x * gl_WorkGroupSize.x + i;
      uint y = call.y * gl_WorkGroupSize.y + j;

      if (x < width && y < height)
      {
        color += imageLoad(img, ivec2(x,y));
      }
    }
  }

  if(gl_LocalInvocationIndex < ((width * height) / gl_WorkGroupSize.x * gl_WorkGroupSize.y))
  {
    result[gl_GlobalInvocationID.y * gl_WorkGroupSize.x + gl_GlobalInvocationID.x] = color;
  }
}
)_";
  }

  // Allocate and update buffers
  gpp::co_update update()
  {
    // Deallocate if the size changed
    const int w = this->inputs.width / downscale;
    const int h = this->inputs.height / downscale;

    if(last_w != w || last_h != h)
    {
      if(this->buf)
      {
        co_yield gpp::buffer_release{.handle = buf};
        buf = nullptr;
      }
      last_w = w;
      last_h = h;
    }

    if(w > 0 && h > 0)
    {
      // No buffer: reallocate
      const int bytes = w * h * sizeof(float) * 4;
      if(!this->buf)
      {
        this->buf = co_yield gpp::static_allocation{
            .binding = lay.bindings.my_buf.binding(), .size = bytes};
      }
    }
  }

  // Relaease allocated data
  gpp::co_release release()
  {
    if(buf)
    {
      co_yield gpp::buffer_release{.handle = buf};
      buf = nullptr;
    }
  }

  // Do the GPU dispatch call
  gpp::co_dispatch dispatch()
  {
    if(!buf)
      co_return;

    const int w = this->inputs.width / downscale;
    const int h = this->inputs.height / downscale;
    const int downscaled_pixels_count = w * h;
    const int bytes = downscaled_pixels_count * sizeof(float) * 4;

    // Run a pass
    co_yield gpp::begin_compute_pass{};

    co_yield gpp::compute_dispatch{.x = 1, .y = 1, .z = 1};

    // Request an asynchronous readback
    gpp::buffer_awaiter readback
        = co_yield gpp::readback_buffer{.handle = buf, .offset = 0, .size = bytes};

    co_yield gpp::end_compute_pass{};

    // The readback can be fetched once the compute pass is done
    // (this needs to be improved in terms of asyncness)
    auto [data, size] = co_yield readback;

    using color = float[4];
    auto flt = reinterpret_cast<const color*>(data);

    // finish summing on the cpu
    auto& final = outputs.color_out.value;

    final[0] = 0.f;
    final[1] = 0.f;
    final[2] = 0.f;
    final[3] = 0.f;

    for(int i = 0; i < downscaled_pixels_count; i++)
    {
      for(int j = 0; j < 4; j++)
      {
        final[j] += flt[i][j];
      }
    }

    final[0] /= downscaled_pixels_count;
    final[1] /= downscaled_pixels_count;
    final[2] /= downscaled_pixels_count;
    final[3] /= downscaled_pixels_count;
  }

private:
  static constexpr auto lay = layout{};
  int last_w{}, last_h{};
  gpp::buffer_handle buf{};
  std::vector<float> zeros{};
};

}