#include "csrc/OpApiCommon.h"
#include "csrc/functions.h"
namespace {
constexpr int64_t MINI_CHANNELS = 16;
}
std::tuple<at::Tensor, at::Tensor> npu_sparse_conv3d(const at::Tensor& indices, at::IntArrayRef kernel_size,
at::IntArrayRef stride, at::IntArrayRef padding, int out_channel, at::IntArrayRef outSpatialShape, int batch_size)
{
TORCH_CHECK_NPU(indices);
TORCH_CHECK(
out_channel <= 1024, "out_channel must less or equal than 1024 expected but got out_channel: ", out_channel);
TORCH_CHECK(out_channel % 8 == 0, "out_channel must be divisible by 8 but got out_channel: ", out_channel);
auto indices_size = indices.sizes();
int64_t kernelsum = 1;
for (int32_t i = 0; i < static_cast<int32_t>(kernel_size.size()); i++) {
kernelsum *= kernel_size[i];
}
int64_t outputsum = indices_size[0] * kernelsum;
c10::SmallVector<int64_t, 8> indices_out_size = {outputsum};
c10::SmallVector<int64_t, 8> indices_pairs_size = {outputsum, indices_size[1]};
c10::SmallVector<int64_t, 8> spatial_size = {
batch_size, outSpatialShape[0], outSpatialShape[1], outSpatialShape[2], out_channel};
at::IntArrayRef outputShape = at::IntArrayRef(spatial_size);
at::Tensor indices_out = at::empty(indices_out_size, indices.options()).fill_(-1);
at::Tensor indices_pairs = at::empty(indices_pairs_size, indices.options()).fill_(-1);
EXEC_NPU_CMD(aclnnSparseConv3d, indices, kernel_size, outputShape, stride, padding, indices_out, indices_pairs);
return std::tie(indices_pairs, indices_out);
}
std::tuple<at::Tensor, at::Tensor> npu_sparse_conv3d_grad(const at::Tensor& former_sorted_indices,
const at::Tensor& indices_offset, const at::Tensor& feature, const at::Tensor& weight, const at::Tensor& grad)
{
TORCH_CHECK_NPU(former_sorted_indices);
TORCH_CHECK_NPU(indices_offset);
TORCH_CHECK_NPU(feature);
TORCH_CHECK_NPU(weight);
TORCH_CHECK_NPU(grad);
TORCH_CHECK(indices_offset.size(0) > 1,
"The length of 'indices_offset' must be greater than 1. Got: indices_offset length = ", indices_offset.size(0));
TORCH_CHECK(former_sorted_indices.scalar_type() == at::kInt && indices_offset.scalar_type() == at::kInt,
"'former_sorted_indices' and 'indices_offset' must be 'int32', but got: ", former_sorted_indices.scalar_type(),
", ", indices_offset.scalar_type());
TORCH_CHECK(feature.scalar_type() == grad.scalar_type() && weight.scalar_type() == grad.scalar_type(),
"Input data types must match, but got: feature type = ", feature.scalar_type(),
", weight type = ", weight.scalar_type(), ", grad type = ", grad.scalar_type());
TORCH_CHECK(grad.size(0) + 1 == indices_offset.size(0),
"The length of 'indices_offset' must match the number of output points. Got: indices_offset length = ",
indices_offset.size(0), ", grad length = ", grad.size(0));
TORCH_CHECK(feature.size(1) % MINI_CHANNELS == 0 && grad.size(1) % MINI_CHANNELS == 0,
"Channels must be aligned to 16. Got: feature channels = ", feature.size(1),
", grad channels = ", grad.size(1));
auto feature_size = feature.sizes();
auto weight_size = weight.sizes();
int64_t kernelIC = weight_size[3];
int64_t kernelOC = weight_size[4];
int64_t start_offset = indices_offset[0].item<int64_t>();
int64_t end_offset = indices_offset[-1].item<int64_t>();
c10::SmallVector<int64_t, SIZE> feature_grad_size = {feature_size[0], kernelIC};
auto dtype = feature.dtype();
if (dtype == at::kHalf) {
auto feature_fp32 = feature.to(at::kFloat);
auto weight_fp32 = weight.to(at::kFloat);
auto grad_fp32 = grad.to(at::kFloat);
at::Tensor feature_grad_fp32 = at::zeros(feature_grad_size, feature.options().dtype(at::kFloat));
at::Tensor weight_grad_fp32 = at::zeros(weight_size, feature.options().dtype(at::kFloat));
EXEC_NPU_CMD(aclnnSparseConv3dGrad, feature_fp32, weight_fp32, grad_fp32, former_sorted_indices, indices_offset,
start_offset, end_offset, feature_grad_fp32, weight_grad_fp32);
at::Tensor feature_grad = feature_grad_fp32.to(at::kHalf);
at::Tensor weight_grad = weight_grad_fp32.to(at::kHalf);
return std::tie(feature_grad, weight_grad);
} else {
at::Tensor feature_grad = at::zeros(feature_grad_size, feature.options());
at::Tensor weight_grad = at::zeros(weight_size, feature.options());
EXEC_NPU_CMD(aclnnSparseConv3dGrad, feature, weight, grad, former_sorted_indices, indices_offset, start_offset,
end_offset, feature_grad, weight_grad);
return std::tie(feature_grad, weight_grad);
}
}