Method getUserCode

tfjs-backend-webgpu/src/softmax_webgpu.ts:41–103

()

  }

  getUserCode(): string {
    const userCode = `
    var<workgroup> buf : array<f32, ${this.workgroupSize[0]}>;
    var<workgroup> rowMaxShared : f32;
    var<workgroup> rowSumShared : f32;
    const blockSize = ${this.workgroupSize[0]};
    ${main('index')} {
      let row = index / blockSize;
      let tid = i32(localId.x);
      let cols = uniforms.outShape[1];

      var threadMax = -3.402823e+38f;
      for (var col = tid; col < cols; col += blockSize) {
        let value = getLogits(row, col);
        threadMax = max(threadMax, value);
      }
      if (tid < cols) {
        buf[tid] = threadMax;
      }
      workgroupBarrier();

      var reduceSize = min(cols, blockSize);
      for (var currSize = reduceSize >> 1;  currSize > 0; currSize = reduceSize >> 1) {
        reduceSize = currSize + (reduceSize & 1);
        if (tid < currSize) {
          buf[tid] = max(buf[tid], buf[tid + reduceSize]);
        }
        workgroupBarrier();
      }

      if (tid == 0) {
        rowMaxShared = buf[0];
      }
      workgroupBarrier();

      var threadSum = 0.0;
      for (var col = tid; col < cols; col += blockSize) {
        let subExp = exp(getLogits(row, col) - rowMaxShared);
        threadSum += subExp;
      }
      buf[tid] = threadSum;
      workgroupBarrier();

      for (var currSize = blockSize >> 1;  currSize > 0; currSize = currSize >> 1) {
        if (tid < currSize) {
          buf[tid] = buf[tid] + buf[tid + currSize];
        }
        workgroupBarrier();
      }

      if (tid == 0) {
        rowSumShared = buf[0];
      }
      workgroupBarrier();

      for (var col = tid; col < cols; col += blockSize) {
        let value = exp(getLogits(row, col) - rowMaxShared) / rowSumShared;
        setOutputAtCoords(row, col, value);