大家好,本文学习Chrome->webgpu-samplers->twoCubes和instancedCube示例。

这两个示例都与“rotatingCube”示例差不多。建议大家先学习该示例,再学习本文的两个示例

上一篇博文:
WebGPU学习(六):学习“rotatingCube”示例

学习twoCubes.ts

该示例绘制了两个立方体。

与“rotatingCube”示例相比,该示例增加了以下的内容:

  • 一个ubo保存两个立方体的mvp矩阵
  • 每帧更新两个mvp矩阵数据
  • draw两次,分别设置对应的uniformBindGroup

下面,我们打开twoCubes.ts文件,依次来看下新增内容:

一个ubo保存两个立方体的mvp矩阵

  • vertex shader定义uniform block

因为只有一个ubo,所以只有一个uniform block,代码与rotatingCube示例相同:

  const vertexShaderGLSL = `#version 450
  layout(set = 0, binding = 0) uniform Uniforms {
    mat4 modelViewProjectionMatrix;
  } uniforms;
  ...
  void main() {
    gl_Position = uniforms.modelViewProjectionMatrix * position;
    ...
  }
  `;
  • 创建uniform buffer

代码如下:

  const matrixSize = 4 * 16; // BYTES_PER_ELEMENT(4) * matrix length(4 * 4 = 16)
  const offset = 256; // uniformBindGroup offset must be 256-byte aligned
  const uniformBufferSize = offset + matrixSize;

  const uniformBuffer = device.createBuffer({
    size: uniformBufferSize,
    usage: GPUBufferUsage.UNIFORM | GPUBufferUsage.COPY_DST,
  });

uniform buffer要保存两个mvp矩阵的数据,但是它们不能连续存放,它们的起始位置必须为256的倍数,所以uniform buffer实际的内存布局为:
0-63:第一个mvp矩阵
64-255:0(占位)
256-319:第二个mvp矩阵

uniform buffer的size为256+64=320

  • 创建uniform bind group

创建两个uniform bind group,通过指定offset和size,对应到同一个uniform buffer:

  const uniformBindGroup1 = device.createBindGroup({
    layout: uniformsBindGroupLayout,
    bindings: [{
      binding: 0,
      resource: {
        buffer: uniformBuffer,
        offset: 0,
        size: matrixSize
      }
    }],
  });

  const uniformBindGroup2 = device.createBindGroup({
    layout: uniformsBindGroupLayout,
    bindings: [{
      binding: 0,
      resource: {
        buffer: uniformBuffer,
        offset: offset,
        size: matrixSize
      }
    }]
  });
  • 创建2个mvp矩阵

代码如下:

  //因为是固定相机,所以只需要计算一次projection矩阵
  const aspect = Math.abs(canvas.width / canvas.height);
  let projectionMatrix = mat4.create();
  mat4.perspective(projectionMatrix, (2 * Math.PI) / 5, aspect, 1, 100.0);
  ...
  
  let modelMatrix1 = mat4.create();
  mat4.translate(modelMatrix1, modelMatrix1, vec3.fromValues(-2, 0, 0));
  let modelMatrix2 = mat4.create();
  mat4.translate(modelMatrix2, modelMatrix2, vec3.fromValues(2, 0, 0));
  //创建两个mvp矩阵
  let modelViewProjectionMatrix1 = mat4.create();
  let modelViewProjectionMatrix2 = mat4.create();
  //因为是固定相机,所以只需要计算一次view矩阵
  let viewMatrix = mat4.create();
  mat4.translate(viewMatrix, viewMatrix, vec3.fromValues(0, 0, -7));

  let tmpMat41 = mat4.create();
  let tmpMat42 = mat4.create();

每帧更新两个mvp矩阵数据

相关代码如下所示:

  function updateTransformationMatrix() {
    let now = Date.now() / 1000;

    mat4.rotate(tmpMat41, modelMatrix1, 1, vec3.fromValues(Math.sin(now), Math.cos(now), 0));
    mat4.rotate(tmpMat42, modelMatrix2, 1, vec3.fromValues(Math.cos(now), Math.sin(now), 0));

    mat4.multiply(modelViewProjectionMatrix1, viewMatrix, tmpMat41);
    mat4.multiply(modelViewProjectionMatrix1, projectionMatrix, modelViewProjectionMatrix1);
    mat4.multiply(modelViewProjectionMatrix2, viewMatrix, tmpMat42);
    mat4.multiply(modelViewProjectionMatrix2, projectionMatrix, modelViewProjectionMatrix2);
  }

  return function frame() {
    updateTransformationMatrix();

    ...

    uniformBuffer.setSubData(0, modelViewProjectionMatrix1);
    uniformBuffer.setSubData(offset, modelViewProjectionMatrix2);
    ...
  }

updateTransformationMatrix函数更新两个mvp矩阵;
调用两次setSubData,分别将更新后的mvp矩阵数据更新到同一个uniform buffer中。

draw两次,分别设置对应的uniformBindGroup

代码如下:

  return function frame() {
    ...
    const passEncoder = commandEncoder.beginRenderPass(renderPassDescriptor);
    ...

    passEncoder.setBindGroup(0, uniformBindGroup1);
    passEncoder.draw(36, 1, 0, 0);

    passEncoder.setBindGroup(0, uniformBindGroup2);
    passEncoder.draw(36, 1, 0, 0);

    passEncoder.endPass();

    ...
  }

第一次draw,绘制第一个cube,设置对应的uniformBindGroup1;
第二次draw,绘制第二个cube,设置对应的uniformBindGroup2。

最终渲染结果

学习instancedCube.ts

该示例使用instance技术,通过一次draw,绘制了多个立方体实例。

与“rotatingCube”示例相比,该示例增加了以下的内容:

  • 一个ubo保存所有立方体实例的mvp矩阵
  • 每帧更新所有立方体实例的mvp矩阵数据
  • 指定实例个数,draw一次

下面,我们打开instancedCube.ts文件,依次来看下新增内容:

一个ubo保存所有立方体实例的mvp矩阵

  • vertex shader定义uniform block

代码如下:

  const vertexShaderGLSL = `#version 450
  //总共16个实例
  #define MAX_NUM_INSTANCES 16
  layout(set = 0, binding = 0) uniform Uniforms {
    //ubo包含mvp矩阵数组,数组长度为16
    mat4 modelViewProjectionMatrix[MAX_NUM_INSTANCES];
  } uniforms;
  layout(location = 0) in vec4 position;
  layout(location = 1) in vec4 color;
  ...
  void main() {
    //使用gl_InstanceIndex取到当前实例的序号(0-15),通过它获取对应的mvp矩阵
    gl_Position = uniforms.modelViewProjectionMatrix[gl_InstanceIndex] * position;
    ...
  }`;
  • 创建uniform buffer

代码如下:

  //16个立方体的排列顺序是x方向4个、y方向4个
  const xCount = 4;
  const yCount = 4;
  const numInstances = xCount * yCount;
  const matrixFloatCount = 16;
  // BYTES_PER_ELEMENT(4) * matrix length(4 * 4 = 16)
  const matrixSize = 4 * matrixFloatCount;
  const uniformBufferSize = numInstances * matrixSize;

  const uniformBuffer = device.createBuffer({
    size: uniformBufferSize,
    usage: GPUBufferUsage.UNIFORM | GPUBufferUsage.COPY_DST,
  });

这里与twoCubes不同的是,不同实例的mvp矩阵的数据是连续存放的,所以uniform buffer的size为numInstances(16个)* matrixSize。

  • 创建uniform bind group

只创建一个:

  const uniformBindGroup = device.createBindGroup({
    layout: uniformsBindGroupLayout,
    bindings: [{
      binding: 0,
      resource: {
        buffer: uniformBuffer,
      }
    }],
  });
  • 准备mvp矩阵数据

代码如下:

  //因为是固定相机,所以只需要计算一次projection矩阵
  const aspect = Math.abs(canvas.width / canvas.height);
  let projectionMatrix = mat4.create();
  mat4.perspective(projectionMatrix, (2 * Math.PI) / 5, aspect, 1, 100.0);
  ...
  
  
  
  let modelMatrices = new Array(numInstances);
  //mvpMatricesData用来依次存放所有立方体实例的mvp矩阵数据
  let mvpMatricesData = new Float32Array(matrixFloatCount * numInstances);

  let step = 4.0;

  let m = 0;
  //准备modelMatrices数据
  for (let x = 0; x < xCount; x++) {
    for (let y = 0; y < yCount; y++) {
      modelMatrices[m] = mat4.create();
      mat4.translate(modelMatrices[m], modelMatrices[m], vec3.fromValues(
        step * (x - xCount / 2 + 0.5),
        step * (y - yCount / 2 + 0.5),
        0
      ));
      m++;
    }
  }

  //因为是固定相机,所以只需要计算一次view矩阵
  let viewMatrix = mat4.create();
  mat4.translate(viewMatrix, viewMatrix, vec3.fromValues(0, 0, -12));

  let tmpMat4 = mat4.create();

每帧更新所有立方体实例的mvp矩阵数据

相关代码如下所示:

  function updateTransformationMatrix() {

    let now = Date.now() / 1000;

    let m = 0, i = 0;
    for (let x = 0; x < xCount; x++) {
      for (let y = 0; y < yCount; y++) {
        mat4.rotate(tmpMat4, modelMatrices[i], 1, vec3.fromValues(Math.sin((x + 0.5) * now), Math.cos((y + 0.5) * now), 0));

        mat4.multiply(tmpMat4, viewMatrix, tmpMat4);
        mat4.multiply(tmpMat4, projectionMatrix, tmpMat4);

        mvpMatricesData.set(tmpMat4, m);

        i++;
        m += matrixFloatCount;
      }
    }
  }

  return function frame() {
    updateTransformationMatrix();
    ...
    uniformBuffer.setSubData(0, mvpMatricesData);
    ...
  }

updateTransformationMatrix函数更新mvpMatricesData;
调用一次setSubData,将更新后的mvpMatricesData设置到uniform buffer中。

指定实例个数,draw一次

代码如下:

  return function frame() {
    ...
    const passEncoder = commandEncoder.beginRenderPass(renderPassDescriptor);
    ...


    //设置对应的uniformBindGroup
    passEncoder.setBindGroup(0, uniformBindGroup);
    //指定实例个数为numInstances
    passEncoder.draw(36, numInstances, 0, 0);
    ...
  }

最终渲染结果

参考资料

WebGPU规范
webgpu-samplers Github Repo