Where can I find the implementation for OpenCL kernel for MaxPool gradient calculation?

skoppula · August 5, 2019, 8:14pm

I’m looking to understand how backpropagation through a MaxPool layer is implemented in GPU operations, and would be interested to look at TVM’s implementation of this.

Thanks.

vinx13 · August 5, 2019, 8:44pm

github.com

dmlc/tvm/blob/master/topi/include/topi/nn/pooling.h#L171


                                         make_const(Int(32), 1));
      return pool_sum(indices) / divide_factor;
    }
  }, "tensor", kElementWise);
} else {
  LOG(ERROR) << "Unrecognized pool_type: " << pool_type;
  return x;
}
}


inline Tensor pool_grad_impl(const Tensor& out_grad, const Tensor& x,
                           const Array<Expr>& kernel_size, const Array<Expr>& stride_size,
                           const Array<Expr>& padding_size, PoolType pool_type, bool ceil_mode,
                           const size_t height_axis, const size_t width_axis,
                           bool count_include_pad) {
CHECK(out_grad->shape.size() >= 2) << "Pooling grad output must >= 2-D (H, W)";
CHECK(x->shape.size() >= 2) << "Pooling input must >= 2-D (H, W)";
CHECK_EQ(kernel_size.size(), 2) << "Pooling kernel_size must have 2 elements";
CHECK_EQ(stride_size.size(), 2) << "Pooling stride_size must have 2 elements";
CHECK_EQ(padding_size.size(), 4) << "Pooling padding_size must have 4 elements";