#include "op_plugin/AclOpsInterface.h"
#include "op_plugin/OpApiInterface.h"
#include "op_plugin/utils/op_api_common.h"
namespace op_api {
using npu_preparation = at_npu::native::OpPreparation;
#if VERSION_BETWEEN(V2R1, V2R6)
const at::Tensor& _conv_depthwise2d_out(
const at::Tensor& self,
const at::Tensor& weight,
at::IntArrayRef kernel_size,
const c10::optional<at::Tensor>& bias_opt,
at::IntArrayRef stride,
at::IntArrayRef padding,
at::IntArrayRef dilation,
const at::Tensor& out) {
DO_COMPATIBILITY(aclnnConvDepthwise2d, acl_op::_conv_depthwise2d_out(self, weight, kernel_size, bias_opt,
stride, padding, dilation, out));
bool is_jit_enable = !at_npu::native::env::CheckJitDisable();
bool is_allow_internel_format = !at_npu::native::env::CheckForbidInternalFormat();
ASCEND_LOGI("_conv_depthwise2d_out exec with jit compile: %d, allow internal format: %d",
is_jit_enable, is_allow_internel_format);
if (is_allow_internel_format || is_jit_enable) {
return acl_op::_conv_depthwise2d_out(self, weight, kernel_size, bias_opt, stride, padding, dilation, out);
}
const at::Tensor& bias = c10::value_or_else(bias_opt, [] {return at::Tensor();});
int8_t cube_math_type = npu_preparation::get_cube_math_type(at_npu::native::env::IsAllowConvHF32());
EXEC_NPU_CMD(aclnnConvDepthwise2d, self, weight, kernel_size, bias, stride, padding, dilation, out, cube_math_type);
return out;
}
#endif
#if VERSION_BETWEEN(V2R7, VERSION_NEWEST)
at::Tensor& _conv_depthwise2d_out(
const at::Tensor& self,
const at::Tensor& weight,
at::IntArrayRef kernel_size,
const c10::optional<at::Tensor>& bias_opt,
at::IntArrayRef stride,
at::IntArrayRef padding,
at::IntArrayRef dilation,
at::Tensor& out)
{
DO_COMPATIBILITY(aclnnConvDepthwise2d, acl_op::_conv_depthwise2d_out(self, weight, kernel_size, bias_opt,
stride, padding, dilation, out));
bool is_jit_enable = !at_npu::native::env::CheckJitDisable();
bool is_allow_internel_format = !at_npu::native::env::CheckForbidInternalFormat();
ASCEND_LOGI("_conv_depthwise2d_out exec with jit compile: %d, allow internal format: %d",
is_jit_enable, is_allow_internel_format);
if (is_allow_internel_format || is_jit_enable) {
return acl_op::_conv_depthwise2d_out(self, weight, kernel_size, bias_opt, stride, padding, dilation, out);
}
const at::Tensor& bias = c10::value_or_else(bias_opt, [] {return at::Tensor();});
int8_t cube_math_type = npu_preparation::get_cube_math_type(at_npu::native::env::IsAllowConvHF32());
EXEC_NPU_CMD(aclnnConvDepthwise2d, self, weight, kernel_size, bias, stride, padding, dilation, out, cube_math_type);
return out;
}
#endif
at::Tensor _conv_depthwise2d(
const at::Tensor& self,
const at::Tensor& weight,
at::IntArrayRef kernel_size,
const c10::optional<at::Tensor>& bias_opt,
at::IntArrayRef stride,
at::IntArrayRef padding,
at::IntArrayRef dilation) {
DO_COMPATIBILITY(aclnnConvDepthwise2d, acl_op::_conv_depthwise2d(self, weight, kernel_size, bias_opt,
stride, padding, dilation));
bool is_jit_enable = !at_npu::native::env::CheckJitDisable();
bool is_allow_internel_format = !at_npu::native::env::CheckForbidInternalFormat();
ASCEND_LOGI("_conv_depthwise2d exec with jit compile: %d, allow internal format: %d",
is_jit_enable, is_allow_internel_format);
if (is_allow_internel_format || is_jit_enable) {
return acl_op::_conv_depthwise2d(self, weight, kernel_size, bias_opt, stride, padding, dilation);
}
const at::Tensor& bias = c10::value_or_else(bias_opt, [] {return at::Tensor();});
auto output_size = op_infer::conv_depthwise2d_npu_output_size(self, weight, kernel_size, stride, padding, dilation);
at::Tensor out = npu_preparation::apply_tensor_without_format(self, output_size);
int8_t cube_math_type = npu_preparation::get_cube_math_type(at_npu::native::env::IsAllowConvHF32());
EXEC_NPU_CMD(aclnnConvDepthwise2d, self, weight, kernel_size, bias, stride, padding, dilation, out, cube_math_type);
return out;
}
}