#include "op_plugin/OpApiInterface.h"
#include "op_plugin/utils/op_api_common.h"
namespace op_api {
using npu_preparation = at_npu::native::OpPreparation;
namespace {
constexpr int64_t DIMENSION_2D = 2;
constexpr int64_t DIMENSION_3D = 3;
constexpr int64_t DIM_0 = 0;
constexpr int64_t DIM_1 = 1;
constexpr int64_t DIM_2 = 2;
constexpr int64_t ROW_BLOCK_SIZE_INVALID = 0;
};
std::tuple<at::Tensor, at::Tensor> npu_grouped_dynamic_block_quant(const at::Tensor &x, const at::Tensor &group_list,
double min_scale, c10::string_view round_mode, int64_t dst_type, int64_t row_block_size, int64_t col_block_size,
int64_t group_list_type, double dst_type_max) {
at::Tensor y;
at::Tensor scale;
auto y_shape = op_infer::array_to_small_vector(x.sizes());
auto scale_shape = op_infer::array_to_small_vector(x.sizes());
TORCH_CHECK(row_block_size != ROW_BLOCK_SIZE_INVALID,
"[npu_grouped_dynamic_block_quant]: row_block_size cannot be zero." + OPS_ERROR(ErrCode::PARAM));
int64_t group_list_dim = group_list.dim();
TORCH_CHECK(group_list_dim == 1, "group_list must be 1D tensor, got dim = ", group_list.dim());
int64_t group_list_shape = group_list.sizes()[0];
ASCEND_LOGI("[npu_grouped_dynamic_block_quant]: group_list shape is %ld.", group_list_shape);
if (group_list_type == 0) {
if (scale_shape.size() == DIMENSION_2D) {
scale_shape[DIM_0] = scale_shape[DIM_0] / row_block_size + group_list_shape;
scale_shape[DIM_1] = op_infer::CeilDiv(scale_shape[DIM_1], col_block_size);
} else if (scale_shape.size() == DIMENSION_3D) {
scale_shape[DIM_1] = scale_shape[DIM_1] / row_block_size + group_list_shape;
scale_shape[DIM_2] = op_infer::CeilDiv(scale_shape[DIM_2], col_block_size);
} else {
TORCH_CHECK(false, "x must be 2 or 3 dimensional.", OPS_ERROR(ErrCode::NOT_SUPPORT));
}
} else {
ASCEND_LOGI("[npu_grouped_dynamic_block_quant]: group_list_type only supports value 0.");
}
ASCEND_LOGI("[npu_grouped_dynamic_block_quant]: Getting aclTensor y dtype by Parameter(dst_type): %ld", dst_type);
aclDataType y_acltype = c10_npu::GetAclDataType(dst_type);
at::ScalarType dtype = npu_preparation::convert_to_scalar_type(y_acltype);
y = npu_preparation::apply_tensor_without_format(y_shape, c10::dtype(dtype));
scale = npu_preparation::apply_tensor_without_format(scale_shape, c10::dtype(c10::ScalarType::Float));
char *round_mode_ptr = const_cast<char *>(round_mode.data());
ASCEND_LOGI("[npu_grouped_dynamic_block_quant]: Setting aclTensor y dtype to: %s",
at_npu::native::AclDataTypeToString(y_acltype).c_str());
TensorWrapper y_wrapper = {y, y_acltype};
static bool npu_support_v2 = check_aclnn_kernel_available("aclnnGroupedDynamicBlockQuantV2");
if (npu_support_v2) {
EXEC_NPU_CMD(aclnnGroupedDynamicBlockQuantV2, x, group_list, min_scale, round_mode_ptr, y_acltype,
row_block_size, col_block_size, group_list_type, dst_type_max, y_wrapper, scale);
} else {
EXEC_NPU_CMD(aclnnGroupedDynamicBlockQuant, x, group_list, min_scale, round_mode_ptr, y_acltype, row_block_size,
col_block_size, group_list_type, y_wrapper, scale);
}
return std::tie(y, scale);
}
}