#include "op_plugin/AclOpsInterface.h"
#include "op_plugin/OpApiInterface.h"
#include "op_plugin/utils/op_api_common.h"
#include "op_plugin/utils/OpUtils.h"
namespace op_api {
using npu_preparation = at_npu::native::OpPreparation;
static bool is_normal_broadcast_expanded(const at::Tensor &t)
{
return t.stride(0) == 0 && t.size(0) != 0;
}
static at::Tensor restore_broadcast_tensor(const at::Tensor &t)
{
auto sizes = std::array<int64_t, 3>{1, t.size(1), t.size(2)};
auto strides = std::array<int64_t, 3>{t.size(1) * t.size(2), t.stride(1), t.stride(2)};
return t.as_strided(sizes, strides);
}
static bool is_compatible_impl_enabled()
{
static auto compatible_env = std::getenv("TORCH_NPU_USE_COMPATIBLE_IMPL");
return compatible_env != nullptr && std::string(compatible_env) == "1";
}
static bool is_dim1_slice_non_contiguous(const at::Tensor &t)
{
if (t.dim() != 3) return false;
if (t.stride(0) == 0) return false;
if (t.stride(2) != 1) return false;
if (t.stride(1) != t.size(2)) return false;
if (t.stride(0) == t.stride(1) * t.size(1)) return false;
if (t.size(1) <= 1 || 16 % t.size(1) != 0) return false;
return true;
}
static std::pair<at::Tensor, at::Tensor> maybe_restore_broadcast(
const at::Tensor &self, const at::Tensor &mat2)
{
if (!is_compatible_impl_enabled()) {
return {self, mat2};
}
bool both_broadcast = is_normal_broadcast_expanded(self) && is_normal_broadcast_expanded(mat2);
if (both_broadcast) {
return {self, mat2};
}
at::Tensor self_in = is_normal_broadcast_expanded(self) ? restore_broadcast_tensor(self) : self;
at::Tensor mat2_in = is_normal_broadcast_expanded(mat2) ? restore_broadcast_tensor(mat2) : mat2;
return {self_in, mat2_in};
}
at::Tensor &bmm_out(
const at::Tensor &self,
const at::Tensor &mat2,
const at::ScalarType output_dtype,
at::Tensor &result)
{
TORCH_CHECK(self.dim() == 3, "self must be a 3D tensor");
TORCH_CHECK(mat2.dim() == 3, "mat2 must be a 3D tensor");
DO_MATMUL_COMPATIBILITY(aclnnBatchMatMulWeightNz, aclnnBatchMatMul, self, mat2, acl_op::bmm_out(self, mat2, result));
auto output_size = {self.size(0), self.size(1), mat2.size(2)};
npu_preparation::check_tensor({self, mat2}, result, output_dtype, output_size);
int8_t cube_math_type = op_plugin::utils::get_cube_math_type_with_passthrough();
if (op_plugin::utils::is_nz_format(mat2) && !op_plugin::utils::is_nz_format(self)) {
EXEC_NPU_CMD(aclnnBatchMatMulWeightNz, self, mat2, result, cube_math_type);
} else if (is_compatible_impl_enabled() && is_dim1_slice_non_contiguous(self) &&
is_normal_broadcast_expanded(mat2)) {
auto mat2_input = mat2[0];
EXEC_NPU_CMD(aclnnMatmul, self, mat2_input, result, cube_math_type);
} else {
auto [self_input, mat2_input] = maybe_restore_broadcast(self, mat2);
EXEC_NPU_CMD(aclnnBatchMatMul, self_input, mat2_input, result, cube_math_type);
}
auto outnames = at::namedinference::compute_bmm_outnames(result, self, mat2);
at::namedinference::propagate_names_if_nonempty(result, outnames);
return result;
}
at::Tensor &bmm_out(
const at::Tensor &self,
const at::Tensor &mat2,
at::Tensor &result)
{
TORCH_CHECK(self.dim() == 3, "self must be a 3D tensor");
TORCH_CHECK(mat2.dim() == 3, "mat2 must be a 3D tensor");
DO_MATMUL_COMPATIBILITY(aclnnBatchMatMulWeightNz, aclnnBatchMatMul, self, mat2, acl_op::bmm_out(self, mat2, result));
auto output_size = {self.size(0), self.size(1), mat2.size(2)};
npu_preparation::check_tensor({self, mat2}, result, self.scalar_type(), output_size);
int8_t cube_math_type = op_plugin::utils::get_cube_math_type_with_passthrough();
if (op_plugin::utils::is_nz_format(mat2) && !op_plugin::utils::is_nz_format(self)) {
EXEC_NPU_CMD(aclnnBatchMatMulWeightNz, self, mat2, result, cube_math_type);
} else if (is_compatible_impl_enabled() && is_dim1_slice_non_contiguous(self) &&
is_normal_broadcast_expanded(mat2)) {
auto mat2_input = mat2[0];
EXEC_NPU_CMD(aclnnMatmul, self, mat2_input, result, cube_math_type);
} else {
auto [self_input, mat2_input] = maybe_restore_broadcast(self, mat2);
EXEC_NPU_CMD(aclnnBatchMatMul, self_input, mat2_input, result, cube_math_type);
}
auto outnames = at::namedinference::compute_bmm_outnames(result, self, mat2);
at::namedinference::propagate_names_if_nonempty(result, outnames);
return result;
}
at::Tensor bmm(const at::Tensor &self, const at::Tensor &mat2, const at::ScalarType output_dtype)
{
TORCH_CHECK(self.dim() == 3, "self must be a 3D tensor");
TORCH_CHECK(mat2.dim() == 3, "mat2 must be a 3D tensor");
DO_MATMUL_COMPATIBILITY(aclnnBatchMatMulWeightNz, aclnnBatchMatMul, self, mat2, acl_op::bmm(self, mat2));
auto output_size = {self.size(0), self.size(1), mat2.size(2)};
at::Tensor result = npu_preparation::apply_tensor_without_format(output_size, self.options().dtype(output_dtype));
int8_t cube_math_type = op_plugin::utils::get_cube_math_type_with_passthrough();
if (op_plugin::utils::is_nz_format(mat2) && !op_plugin::utils::is_nz_format(self)) {
EXEC_NPU_CMD(aclnnBatchMatMulWeightNz, self, mat2, result, cube_math_type);
} else if (is_compatible_impl_enabled() && is_dim1_slice_non_contiguous(self) &&
is_normal_broadcast_expanded(mat2)) {
auto mat2_input = mat2[0];
EXEC_NPU_CMD(aclnnMatmul, self, mat2_input, result, cube_math_type);
} else {
auto [self_input, mat2_input] = maybe_restore_broadcast(self, mat2);
EXEC_NPU_CMD(aclnnBatchMatMul, self_input, mat2_input, result, cube_math_type);
}
auto outnames = at::namedinference::compute_bmm_outnames(result, self, mat2);
at::namedinference::propagate_names_if_nonempty(result, outnames);
FLOP_COUNT(FlopCounter::bmm_flop, self, mat2);
return result;
}
at::Tensor bmm(const at::Tensor &self, const at::Tensor &mat2)
{
TORCH_CHECK(self.dim() == 3, "self must be a 3D tensor");
TORCH_CHECK(mat2.dim() == 3, "mat2 must be a 3D tensor");
DO_MATMUL_COMPATIBILITY(aclnnBatchMatMulWeightNz, aclnnBatchMatMul, self, mat2, acl_op::bmm(self, mat2));
auto output_size = {self.size(0), self.size(1), mat2.size(2)};
at::Tensor result = npu_preparation::apply_tensor_without_format(output_size, self.options());
int8_t cube_math_type = op_plugin::utils::get_cube_math_type_with_passthrough();
if (op_plugin::utils::is_nz_format(mat2) && !op_plugin::utils::is_nz_format(self)) {
EXEC_NPU_CMD(aclnnBatchMatMulWeightNz, self, mat2, result, cube_math_type);
} else if (is_compatible_impl_enabled() && is_dim1_slice_non_contiguous(self) &&
is_normal_broadcast_expanded(mat2)) {
auto mat2_input = mat2[0];
EXEC_NPU_CMD(aclnnMatmul, self, mat2_input, result, cube_math_type);
} else {
auto [self_input, mat2_input] = maybe_restore_broadcast(self, mat2);
EXEC_NPU_CMD(aclnnBatchMatMul, self_input, mat2_input, result, cube_math_type);
}
auto outnames = at::namedinference::compute_bmm_outnames(result, self, mat2);
at::namedinference::propagate_names_if_nonempty(result, outnames);
FLOP_COUNT(FlopCounter::bmm_flop, self, mat2);
return result;
}
}