* Copyright (c) 2025 Huawei Technologies Co., Ltd.
* This program is free software, you can redistribute it and/or modify it under the terms and conditions of
* CANN Open Software License Agreement Version 2.0 (the "License").
* Please refer to the License for details. You may not use this file except in compliance with the License.
* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
* See LICENSE in the root of the software repository for the full text of the License.
*/
* \file kernel_operator_mm_impl.h
* \brief
*/
#if !defined(__ASCENDC_INCLUDE_INTERNAL_HEADERS__)
#pragma message( \
"impl/basic_api/dav_c220/kernel_operator_mm_impl.h is an internal header file and must not be used directly. Functions or variables defined in this file may be removed in the future. Please use \"#include \"basic_api/kernel_tensor.h\"\" and use public functions or variables defined in interface headers files.")
#define __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#define __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_KERNEL_OPERATOR_MM_IMPL_H__
#endif
#ifndef ASCENDC_MODULE_OPERATOR_MM_IMPL_H
#define ASCENDC_MODULE_OPERATOR_MM_IMPL_H
#include "../../../include/basic_api/kernel_struct_mm.h"
#include "../kernel_npu_debug.h"
namespace AscendC {
* LoadData 2dv2 *
* ************************************************************************************************* */
template <typename T>
__aicore__ inline void LoadData2DL12L0ACal(__ca__ T* dst, __cbuf__ T* src, const LoadData2DParams& loadDataParam)
{
if ASCEND_IS_AIC {
if constexpr (IsSameType<T, int4b_t>::value) {
load_cbuf_to_ca_s4(
(__ca__ void*)dst, (__cbuf__ void*)src, loadDataParam.startIndex, loadDataParam.repeatTimes,
loadDataParam.srcStride, loadDataParam.dstGap, loadDataParam.sid, 0, inc);
} else {
using U = typename Conditional<SupportType<T, uint16_t, int16_t>(), half, T>::type;
if (loadDataParam.ifTranspose) {
load_cbuf_to_ca(
(__ca__ U*)dst, (__cbuf__ U*)src, loadDataParam.startIndex, loadDataParam.repeatTimes,
loadDataParam.srcStride, loadDataParam.dstGap, loadDataParam.sid, 1, inc);
} else {
load_cbuf_to_ca(
(__ca__ U*)dst, (__cbuf__ U*)src, loadDataParam.startIndex, loadDataParam.repeatTimes,
loadDataParam.srcStride, loadDataParam.dstGap, loadDataParam.sid, 0, inc);
}
}
}
}
template <typename T>
__aicore__ inline void LoadData2DL12L0BCal(__cb__ T* dst, __cbuf__ T* src, const LoadData2DParams& loadDataParam)
{
if ASCEND_IS_AIC {
if constexpr (IsSameType<T, int4b_t>::value) {
load_cbuf_to_cb_s4(
(__cb__ void*)dst, (__cbuf__ void*)src, loadDataParam.startIndex, loadDataParam.repeatTimes,
loadDataParam.srcStride, loadDataParam.dstGap, loadDataParam.sid, 0, inc);
} else {
using U = typename Conditional<SupportType<T, uint16_t, int16_t>(), half, T>::type;
if (loadDataParam.ifTranspose) {
load_cbuf_to_cb(
(__cb__ U*)dst, (__cbuf__ U*)src, loadDataParam.startIndex, loadDataParam.repeatTimes,
loadDataParam.srcStride, loadDataParam.dstGap, loadDataParam.sid, 1, inc);
} else {
load_cbuf_to_cb(
(__cb__ U*)dst, (__cbuf__ U*)src, loadDataParam.startIndex, loadDataParam.repeatTimes,
loadDataParam.srcStride, loadDataParam.dstGap, loadDataParam.sid, 0, inc);
}
}
}
}
template <typename T>
__aicore__ inline void LoadData2DGM2L0ACal(__ca__ T* dst, __gm__ T* src, const LoadData2DParams& loadDataParam)
{
if ASCEND_IS_AIC {
ASCENDC_DEBUG_ASSERT(
(!SupportType<T, int4b_t>()),
KERNEL_LOG_INTERNAL(
KERNEL_ERROR, "Failed to check dtype in "
"LoadData with LoadData2DParams, GM -> L0A Buffer(A2) does not support int4b_t.\n"));
using U = typename Conditional<SupportType<T, uint16_t, int16_t>(), half, T>::type;
load_gm_to_ca(
(__ca__ U*)dst, (__gm__ U*)src, loadDataParam.startIndex, loadDataParam.repeatTimes,
loadDataParam.srcStride, loadDataParam.dstGap, loadDataParam.sid, (addr_cal_mode_t)0);
}
}
template <typename T>
__aicore__ inline void LoadData2DGM2L0BCal(__cb__ T* dst, __gm__ T* src, const LoadData2DParams& loadDataParam)
{
if ASCEND_IS_AIC {
ASCENDC_DEBUG_ASSERT(
(!SupportType<T, int4b_t>()),
KERNEL_LOG_INTERNAL(
KERNEL_ERROR, "Failed to check dtype in "
"LoadData with LoadData2DParams, GM -> L0B Buffer(B2) does not support int4b_t.\n"));
using U = typename Conditional<SupportType<T, uint16_t, int16_t>(), half, T>::type;
load_gm_to_cb(
(__cb__ U*)dst, (__gm__ U*)src, loadDataParam.startIndex, loadDataParam.repeatTimes,
loadDataParam.srcStride, loadDataParam.dstGap, loadDataParam.sid, (addr_cal_mode_t)0);
}
}
template <typename T>
__aicore__ inline void LoadData2DGM2L1Cal(__cbuf__ T* dst, __gm__ T* src, const LoadData2DParams& loadDataParam)
{
if ASCEND_IS_AIC {
ASCENDC_DEBUG_ASSERT(
(!SupportType<T, int4b_t>()),
KERNEL_LOG_INTERNAL(
KERNEL_ERROR, "Failed to check dtype in "
"LoadData with LoadData2DParams, GM -> L1 Buffer(A1/B1) does not support int4b_t.\n"));
using U = typename Conditional<SupportType<T, uint16_t, int16_t>(), half, T>::type;
if (loadDataParam.addrMode == 0) {
load_gm_to_cbuf(
(__cbuf__ U*)dst, (__gm__ U*)src, loadDataParam.startIndex, loadDataParam.repeatTimes,
loadDataParam.srcStride, loadDataParam.dstGap, loadDataParam.sid, inc);
} else {
load_gm_to_cbuf(
(__cbuf__ U*)dst, (__gm__ U*)src, loadDataParam.startIndex, loadDataParam.repeatTimes,
loadDataParam.srcStride, loadDataParam.dstGap, loadDataParam.sid, dec);
}
}
}
template <typename T>
__aicore__ inline void LoadData2DL12L0ACal(__ca__ T* dst, __cbuf__ T* src, const LoadData2DParamsV2& loadDataParam)
{
#if defined(ASCENDC_DEBUG) || defined(ASCENDC_CPU_DEBUG)
ReportNotSupport(false, "LoadData with LoadData2DParamsV2 from L1 Buffer(A1) to L0A Buffer(A2)");
#endif
}
template <typename T>
__aicore__ inline void LoadData2DL12L0BCal(__cb__ T* dst, __cbuf__ T* src, const LoadData2DParamsV2& loadDataParam)
{
#if defined(ASCENDC_DEBUG) || defined(ASCENDC_CPU_DEBUG)
ReportNotSupport(false, "LoadData with LoadData2DParamsV2 from L1 Buffer(B1) to L0B Buffer(B2)");
#endif
}
template <typename T>
__aicore__ inline void LoadData2DGM2L0ACal(__ca__ T* dst, __gm__ T* src, const LoadData2DParamsV2& loadDataParam)
{
#if defined(ASCENDC_DEBUG) || defined(ASCENDC_CPU_DEBUG)
ReportNotSupport(false, "LoadData with LoadData2DParamsV2 from GM to L0A Buffer(A2)");
#endif
}
template <typename T>
__aicore__ inline void LoadData2DGM2L0BCal(__cb__ T* dst, __gm__ T* src, const LoadData2DParamsV2& loadDataParam)
{
#if defined(ASCENDC_DEBUG) || defined(ASCENDC_CPU_DEBUG)
ReportNotSupport(false, "LoadData with LoadData2DParamsV2 from GM to L0B Buffer(B2)");
#endif
}
template <typename T>
__aicore__ inline void LoadData2DGM2L1Cal(__cbuf__ T* dst, __gm__ T* src, const LoadData2DParamsV2& loadDataParam)
{
#if defined(ASCENDC_DEBUG) || defined(ASCENDC_CPU_DEBUG)
ReportNotSupport(false, "LoadData with LoadData2DParamsV2 from GM to L1 Buffer(A1/B1)");
#endif
}
template <typename T>
__aicore__ inline void LoadData2DL12L0ATransposeCal(
__ca__ T* dst, __cbuf__ T* src, const LoadData2dTransposeParams& loadDataParam)
{
if ASCEND_IS_AIC {
if constexpr (!IsSameType<T, int4b_t>::value) {
load_cbuf_to_ca_transpose(
dst, src, loadDataParam.startIndex, loadDataParam.repeatTimes, loadDataParam.srcStride,
loadDataParam.dstGap, inc, loadDataParam.dstFracGap);
}
}
}
template <typename T>
__aicore__ inline void LoadData2DL12L0BTransposeCal(
__cb__ T* dst, __cbuf__ T* src, const LoadData2dTransposeParams& loadDataParam)
{
if ASCEND_IS_AIC {
if constexpr (IsSameType<T, int4b_t>::value) {
load_cbuf_to_cb_transpose_s4(
(__cb__ void*)dst, (__cbuf__ void*)src, loadDataParam.startIndex, loadDataParam.repeatTimes,
loadDataParam.srcStride, loadDataParam.dstGap, inc, loadDataParam.dstFracGap);
} else {
load_cbuf_to_cb_transpose(
dst, src, loadDataParam.startIndex, loadDataParam.repeatTimes, loadDataParam.srcStride,
loadDataParam.dstGap, inc, loadDataParam.dstFracGap);
}
}
}
template <typename T>
__aicore__ inline void LoadData2DL12L0BTransposeCal(
__cb__ T* dst, __cbuf__ T* src, const LoadData2dTransposeParamsV2& loadDataParam)
{
#if defined(ASCENDC_DEBUG) || defined(ASCENDC_CPU_DEBUG)
ReportNotSupport(
false, "LoadDataWithTranspose with LoadData2dTransposeParamsV2 from L1 Buffer(B1) to L0B Buffer(B2)");
#endif
}
* LoadData 3dv2 *
* ************************************************************************************************* */
template <typename T>
__aicore__ inline void LoadData3DV2L12L0ACal(
__ca__ T* dst, __cbuf__ T* src, const LoadData3DParamsV2<T>& loadDataParams)
{
if ASCEND_IS_AIC {
if constexpr (IsSameType<T, int4b_t>::value) {
img2colv2_cbuf_to_ca_s4(
(__ca__ void*)dst, (__cbuf__ void*)src, loadDataParams.kExtension, loadDataParams.mExtension,
loadDataParams.kStartPt, loadDataParams.mStartPt, loadDataParams.strideW, loadDataParams.strideH,
loadDataParams.filterW, loadDataParams.filterH, loadDataParams.dilationFilterW,
loadDataParams.dilationFilterH, loadDataParams.filterSizeW, loadDataParams.filterSizeH,
loadDataParams.enTranspose, loadDataParams.fMatrixCtrl, loadDataParams.channelSize);
} else {
img2colv2_cbuf_to_ca(
dst, src, loadDataParams.kExtension, loadDataParams.mExtension, loadDataParams.kStartPt,
loadDataParams.mStartPt, loadDataParams.strideW, loadDataParams.strideH, loadDataParams.filterW,
loadDataParams.filterH, loadDataParams.dilationFilterW, loadDataParams.dilationFilterH,
loadDataParams.filterSizeW, loadDataParams.filterSizeH, loadDataParams.enTranspose,
loadDataParams.fMatrixCtrl, loadDataParams.channelSize);
}
}
}
template <typename T>
__aicore__ inline void LoadData3DV2L12L0BCal(
__cb__ T* dst, __cbuf__ T* src, const LoadData3DParamsV2<T>& loadDataParams)
{
if ASCEND_IS_AIC {
if constexpr (!IsSameType<T, int4b_t>::value) {
img2colv2_cbuf_to_cb(
dst, src, loadDataParams.kExtension, loadDataParams.mExtension, loadDataParams.kStartPt,
loadDataParams.mStartPt, loadDataParams.strideW, loadDataParams.strideH, loadDataParams.filterW,
loadDataParams.filterH, loadDataParams.dilationFilterW, loadDataParams.dilationFilterH,
loadDataParams.filterSizeW, loadDataParams.filterSizeH, loadDataParams.enTranspose,
loadDataParams.fMatrixCtrl, loadDataParams.channelSize);
}
}
}
template <>
__aicore__ inline void LoadData3DV2L12L0ACal(
__ca__ bfloat16_t* dst, __cbuf__ bfloat16_t* src, const LoadData3DParamsV2<bfloat16_t>& loadDataParams)
{
if ASCEND_IS_AIC {
img2colv2_cbuf_to_ca(
reinterpret_cast<__ca__ half*>(dst), reinterpret_cast<__cbuf__ half*>(src), loadDataParams.kExtension,
loadDataParams.mExtension, loadDataParams.kStartPt, loadDataParams.mStartPt, loadDataParams.strideW,
loadDataParams.strideH, loadDataParams.filterW, loadDataParams.filterH, loadDataParams.dilationFilterW,
loadDataParams.dilationFilterH, loadDataParams.filterSizeW, loadDataParams.filterSizeH,
loadDataParams.enTranspose, loadDataParams.fMatrixCtrl, loadDataParams.channelSize);
}
}
template <>
__aicore__ inline void LoadData3DV2L12L0BCal(
__cb__ bfloat16_t* dst, __cbuf__ bfloat16_t* src, const LoadData3DParamsV2<bfloat16_t>& loadDataParams)
{
if ASCEND_IS_AIC {
img2colv2_cbuf_to_cb(
reinterpret_cast<__cb__ half*>(dst), reinterpret_cast<__cbuf__ half*>(src), loadDataParams.kExtension,
loadDataParams.mExtension, loadDataParams.kStartPt, loadDataParams.mStartPt, loadDataParams.strideW,
loadDataParams.strideH, loadDataParams.filterW, loadDataParams.filterH, loadDataParams.dilationFilterW,
loadDataParams.dilationFilterH, loadDataParams.filterSizeW, loadDataParams.filterSizeH,
loadDataParams.enTranspose, loadDataParams.fMatrixCtrl, loadDataParams.channelSize);
}
}
* LoadData 3dv2Pro *
* ************************************************************************************************* */
template <typename T>
__aicore__ inline void LoadData3DV2L12L0ACal(
__ca__ T* dst, __cbuf__ T* src, const LoadData3DParamsV2Pro& loadDataParams)
{
if ASCEND_IS_AIC {
if constexpr (IsSameType<T, int4b_t>::value) {
img2colv2_cbuf_to_ca_s4(
(__ca__ void*)dst, (__cbuf__ void*)src, loadDataParams.extConfig,
loadDataParams.extConfig >> LOAD_M_EXTENSION, loadDataParams.extConfig >> LOAD_K_START_POSITION,
loadDataParams.extConfig >> LOAD_M_START_POSITION, loadDataParams.filterConfig,
loadDataParams.filterConfig >> LOAD_STRIDE_H, loadDataParams.filterConfig >> LOAD_FILTER_W,
loadDataParams.filterConfig >> LOAD_FILTER_H, loadDataParams.filterConfig >> LOAD_DILATION_FILTER_W,
loadDataParams.filterConfig >> LOAD_DILATION_FILTER_H, loadDataParams.filterSizeW,
loadDataParams.filterSizeH, loadDataParams.enTranspose, loadDataParams.fMatrixCtrl,
loadDataParams.channelSize);
} else {
img2colv2_cbuf_to_ca(
dst, src, loadDataParams.extConfig, loadDataParams.extConfig >> LOAD_M_EXTENSION,
loadDataParams.extConfig >> LOAD_K_START_POSITION, loadDataParams.extConfig >> LOAD_M_START_POSITION,
loadDataParams.filterConfig, loadDataParams.filterConfig >> LOAD_STRIDE_H,
loadDataParams.filterConfig >> LOAD_FILTER_W, loadDataParams.filterConfig >> LOAD_FILTER_H,
loadDataParams.filterConfig >> LOAD_DILATION_FILTER_W,
loadDataParams.filterConfig >> LOAD_DILATION_FILTER_H, loadDataParams.filterSizeW,
loadDataParams.filterSizeH, loadDataParams.enTranspose, loadDataParams.fMatrixCtrl,
loadDataParams.channelSize);
}
}
}
template <>
__aicore__ inline void LoadData3DV2L12L0ACal(
__ca__ bfloat16_t* dst, __cbuf__ bfloat16_t* src, const LoadData3DParamsV2Pro& loadDataParams)
{
if ASCEND_IS_AIC {
img2colv2_cbuf_to_ca(
(__ca__ half*)dst, (__cbuf__ half*)src, loadDataParams.extConfig,
loadDataParams.extConfig >> LOAD_M_EXTENSION, loadDataParams.extConfig >> LOAD_K_START_POSITION,
loadDataParams.extConfig >> LOAD_M_START_POSITION, loadDataParams.filterConfig,
loadDataParams.filterConfig >> LOAD_STRIDE_H, loadDataParams.filterConfig >> LOAD_FILTER_W,
loadDataParams.filterConfig >> LOAD_FILTER_H, loadDataParams.filterConfig >> LOAD_DILATION_FILTER_W,
loadDataParams.filterConfig >> LOAD_DILATION_FILTER_H, loadDataParams.filterSizeW,
loadDataParams.filterSizeH, loadDataParams.enTranspose, loadDataParams.fMatrixCtrl,
loadDataParams.channelSize);
}
}
template <typename T>
__aicore__ inline void LoadData3DV2L12L0BCal(
__cb__ T* dst, __cbuf__ T* src, const LoadData3DParamsV2Pro& loadDataParams)
{
if ASCEND_IS_AIC {
if constexpr (!IsSameType<T, int4b_t>::value) {
img2colv2_cbuf_to_cb(
dst, src, loadDataParams.extConfig, loadDataParams.extConfig >> LOAD_M_EXTENSION,
loadDataParams.extConfig >> LOAD_K_START_POSITION, loadDataParams.extConfig >> LOAD_M_START_POSITION,
loadDataParams.filterConfig, loadDataParams.filterConfig >> LOAD_STRIDE_H,
loadDataParams.filterConfig >> LOAD_FILTER_W, loadDataParams.filterConfig >> LOAD_FILTER_H,
loadDataParams.filterConfig >> LOAD_DILATION_FILTER_W,
loadDataParams.filterConfig >> LOAD_DILATION_FILTER_H, loadDataParams.filterSizeW,
loadDataParams.filterSizeH, loadDataParams.enTranspose, loadDataParams.fMatrixCtrl,
loadDataParams.channelSize);
}
}
}
template <>
__aicore__ inline void LoadData3DV2L12L0BCal(
__cb__ bfloat16_t* dst, __cbuf__ bfloat16_t* src, const LoadData3DParamsV2Pro& loadDataParams)
{
if ASCEND_IS_AIC {
img2colv2_cbuf_to_cb(
(__cb__ half*)dst, (__cbuf__ half*)src, loadDataParams.extConfig,
loadDataParams.extConfig >> LOAD_M_EXTENSION, loadDataParams.extConfig >> LOAD_K_START_POSITION,
loadDataParams.extConfig >> LOAD_M_START_POSITION, loadDataParams.filterConfig,
loadDataParams.filterConfig >> LOAD_STRIDE_H, loadDataParams.filterConfig >> LOAD_FILTER_W,
loadDataParams.filterConfig >> LOAD_FILTER_H, loadDataParams.filterConfig >> LOAD_DILATION_FILTER_W,
loadDataParams.filterConfig >> LOAD_DILATION_FILTER_H, loadDataParams.filterSizeW,
loadDataParams.filterSizeH, loadDataParams.enTranspose, loadDataParams.fMatrixCtrl,
loadDataParams.channelSize);
}
}
* Mmad *
* ************************************************************************************************* */
template <typename T, typename U, typename S>
__aicore__ inline void MmadCal(__cc__ T* c, __ca__ U* a, __cb__ S* b, const MmadParams& mmadParams)
{
if ASCEND_IS_AIC {
ASCENDC_DEBUG_ASSERT(
(SupportType<
Tuple<T, U, S>, Tuple<int32_t, int8_t, int8_t>, Tuple<float, half, half>, Tuple<float, float, float>,
Tuple<float, bfloat16_t, bfloat16_t>, Tuple<int32_t, int4b_t, int4b_t>>()),
KERNEL_LOG_INTERNAL(
KERNEL_ERROR,
"Failed to check dtype in "
"Mmad, current api support dtype combination is dst: int32_t, fm: int8_t, filter: int8_t; dst: "
"float, fm: half, filter: half; dst: float, fm: float, filter: float; dst: float, fm: "
"bfloat16_t, filter: bfloat16_t; dst: int32_t, fm: int4b_t, filter: int4b_t"));
bool cmatrixInitVal = mmadParams.cmatrixInitVal && (!mmadParams.isBias);
if constexpr ((IsSameType<U, int4b_t>::value) && (IsSameType<S, int4b_t>::value)) {
ASCENDC_DEBUG_ASSERT(
(mmadParams.k % 2 == 0), KERNEL_LOG_INTERNAL(
KERNEL_ERROR, "MmadParams.k must be an "
"even number when U and S are int4b_t.\n"));
mad_s4(
c, (__ca__ void*)a, (__cb__ void*)b, mmadParams.m, mmadParams.k, mmadParams.n, mmadParams.unitFlag,
mmadParams.kDirectionAlign, mmadParams.cmatrixSource, cmatrixInitVal);
} else {
mad(c, a, b, mmadParams.m, mmadParams.k, mmadParams.n, mmadParams.unitFlag, mmadParams.kDirectionAlign,
mmadParams.cmatrixSource, cmatrixInitVal);
}
}
}
template <typename T, typename U, typename S>
__aicore__ inline void MmadCal(
__cc__ T* c, __ca__ U* a, __cb__ S* b, uint64_t bias, const MmadParams& mmadParams, bool cmatrixSource)
{
if ASCEND_IS_AIC {
ASCENDC_DEBUG_ASSERT(
(SupportType<
Tuple<T, U, S>, Tuple<int32_t, int8_t, int8_t>, Tuple<float, half, half>, Tuple<float, float, float>,
Tuple<float, bfloat16_t, bfloat16_t>>()),
KERNEL_LOG_INTERNAL(
KERNEL_ERROR,
"Failed to check dtype in Mmad, current api support dtype "
"combination is dst: int32_t, fm: int8_t, filter: int8_t; dst: float, fm: half, filter: half; dst: "
"float, fm: float, filter: float; dst: float, fm: bfloat16_t, filter: bfloat16_t;"));
if constexpr ((IsSameType<U, int4b_t>::value) && (IsSameType<S, int4b_t>::value)) {
mad_s4(
c, (__ca__ void*)a, (__cb__ void*)b, mmadParams.m, mmadParams.k, mmadParams.n, mmadParams.unitFlag,
mmadParams.kDirectionAlign, cmatrixSource, mmadParams.cmatrixInitVal);
} else {
mad(c, a, b, bias, mmadParams.m, mmadParams.k, mmadParams.n, mmadParams.unitFlag,
mmadParams.kDirectionAlign, cmatrixSource, mmadParams.cmatrixInitVal);
}
}
}
__aicore__ inline void MmadSpCal(__cc__ int32_t* c, __ca__ int8_t* a, __cb__ int8_t* b, const MmadParams& mmadParams)
{
if ASCEND_IS_AIC {
mad_sp(
c, a, b, mmadParams.m, mmadParams.k, mmadParams.n, mmadParams.unitFlag, mmadParams.cmatrixSource,
mmadParams.cmatrixInitVal);
}
}
template <
typename T = int8_t, typename U = uint8_t,
typename std::enable_if<IsSameType<PrimT<T>, int8_t>::value, bool>::type = true,
typename std::enable_if<IsSameType<PrimT<U>, uint8_t>::value, bool>::type = true>
__aicore__ inline void LoadDataWithSparseCal(
const LocalTensor<T>& dst, const LocalTensor<T>& src, const LocalTensor<U>& idx,
const LoadData2dParams& loadDataParam)
{
if ASCEND_IS_AIC {
#ifdef ASCENDC_CPU_DEBUG
SetIndexMatrix((void*)idx.GetPhyAddr());
__cbuf__ int8_t* srcAddr = (__cbuf__ int8_t*)src.GetPhyAddr();
#else
uint64_t src0tmp = reinterpret_cast<uint64_t>(src.GetPhyAddr());
uint64_t src1tmp = reinterpret_cast<uint64_t>(idx.GetPhyAddr());
uint64_t srctmp = (src0tmp & 0xffffffff) | ((src1tmp & 0xffffffff) << 32);
__cbuf__ int8_t* srcAddr = reinterpret_cast<__cbuf__ int8_t*>(srctmp);
#endif
load_cbuf_to_cb_sp(
(__cb__ int8_t*)dst.GetPhyAddr(), srcAddr, loadDataParam.startIndex, loadDataParam.repeatTimes);
}
}
template <typename T = int8_t, typename std::enable_if<IsSameType<PrimT<T>, int8_t>::value, bool>::type = true>
__aicore__ inline void LoadUnzipIndexCal(const GlobalTensor<T>& src, uint32_t numOfIndexTabEntry)
{
ASCENDC_REPORT_NOT_SUPPORT(false, "LoadUnzipIndex");
}
* LoadData 3dv1 *
* ************************************************************************************************* */
__aicore__ inline void Load3DSetFMatrixCal(uint16_t l1H, uint16_t l1W, const uint8_t padList[4])
{
if ASCEND_IS_AIC {
uint64_t regFMatrix = 0;
regFMatrix |= uint64_t(l1W & 0xFFFF);
uint32_t l1HShiftBit = 16;
regFMatrix |= uint64_t(l1H & 0xFFFF) << l1HShiftBit;
uint32_t padNumber = 4;
uint32_t padListShiftBit = 8;
uint32_t padListShiftBase = 32;
for (uint32_t i = 0; i < padNumber; i++) {
regFMatrix |= uint64_t(padList[i] & 0xFF) << (padListShiftBase + i * padListShiftBit);
}
set_fmatrix(regFMatrix);
}
}
__aicore__ inline void Load3DSetFMatrixBCal(uint16_t l1H, uint16_t l1W, const uint8_t padList[4])
{
if ASCEND_IS_AIC {
uint64_t regFMatrix = 0;
regFMatrix |= static_cast<uint64_t>(l1W);
uint32_t l1HShiftBit = 16;
regFMatrix |= static_cast<uint64_t>(l1H) << l1HShiftBit;
uint32_t padNumber = 4;
uint32_t padListShiftBit = 8;
uint32_t padListShiftBase = 32;
for (uint32_t i = 0; i < padNumber; i++) {
regFMatrix |= uint64_t(padList[i] & 0xFF) << (padListShiftBase + i * padListShiftBit);
}
set_fmatrix_b(regFMatrix);
}
}
template <typename T>
__aicore__ inline void Load3DSetPaddingCal(const T padValue)
{
if ASCEND_IS_AIC {
uint64_t paddingValue = 0;
uint64_t padValueShiftBit = 8;
if constexpr (sizeof(T) == B16_BYTE_SIZE || sizeof(T) == B32_BYTE_SIZE) {
paddingValue = static_cast<uint64_t>(GetScalarBitcodeValue((T)padValue));
} else {
paddingValue =
((static_cast<uint64_t>(padValue)) << padValueShiftBit) | (static_cast<uint64_t>(padValue) & 0xFF);
}
set_padding(paddingValue);
}
}
* LoadData 3dv1 *
* ************************************************************************************************* */
template <typename T>
__aicore__ inline void LoadData3DV1L12L0ACal(
__ca__ T* dst, __cbuf__ T* src, const LoadData3DParamsV1<T>& loadDataParams)
{
#if defined(ASCENDC_DEBUG) || defined(ASCENDC_CPU_DEBUG)
ReportNotSupport(false, "LoadData with LoadData3DParamsV1 from L1 Buffer(A1) to L0A Buffer(A2)");
#endif
}
template <typename T>
__aicore__ inline void LoadData3DV1L12L0BCal(
__cb__ T* dst, __cbuf__ T* src, const LoadData3DParamsV1<T>& loadDataParams)
{
#if defined(ASCENDC_DEBUG) || defined(ASCENDC_CPU_DEBUG)
ReportNotSupport(false, "LoadData with LoadData3DParamsV1 from L1 Buffer(B1) to L0B Buffer(B2)");
#endif
}
template <typename T>
__aicore__ inline void LoadData3DV1L12UBCal(
__ubuf__ T* dst, __cbuf__ T* src, const LoadData3DParamsV1<T>& loadDataParams)
{
#if defined(ASCENDC_DEBUG) || defined(ASCENDC_CPU_DEBUG)
ReportNotSupport(false, "LoadData with LoadData3DParamsV1 from L1 to UB");
#endif
}
* LoadData 3dv2 *
* ************************************************************************************************* */
template <typename T>
__aicore__ inline void LoadData3DV2L12UBCal(
__ubuf__ T* dst, __cbuf__ T* src, const LoadData3DParamsV2<T>& loadDataParams)
{
#if defined(ASCENDC_DEBUG) || defined(ASCENDC_CPU_DEBUG)
ReportNotSupport(false, "LoadData with LoadData3DParamsV2 from L1 to UB");
#endif
}
template <>
__aicore__ inline void LoadData3DV2L12L0BCal(
__cb__ int8_t* dst, __cbuf__ int8_t* src, const LoadData3DParamsV2<int8_t>& loadDataParams)
{
#if defined(ASCENDC_DEBUG) || defined(ASCENDC_CPU_DEBUG)
ReportNotSupport(false, "LoadData with LoadData3DParamsV2 from L1 Buffer(B1) to L0B Buffer(B2) with type int8_t");
#endif
}
template <>
__aicore__ inline void LoadData3DV2L12L0BCal(
__cb__ uint8_t* dst, __cbuf__ uint8_t* src, const LoadData3DParamsV2<uint8_t>& loadDataParams)
{
#if defined(ASCENDC_DEBUG) || defined(ASCENDC_CPU_DEBUG)
ReportNotSupport(false, "LoadData with LoadData3DParamsV2 from L1 Buffer(B1) to L0B Buffer(B2) with type uint8_t");
#endif
}
* LoadData 3dv2Pro *
* ************************************************************************************************* */
template <typename T>
__aicore__ inline void LoadData3DV2L12UBCal(
__ubuf__ T* dst, __cbuf__ T* src, const LoadData3DParamsV2Pro& loadDataParams)
{
ASCENDC_REPORT_NOT_SUPPORT(false, "LoadData with LoadData3DParamsV2Pro from L1 to UB");
}
template <>
__aicore__ inline void LoadData3DV2L12L0BCal(
__cb__ int8_t* dst, __cbuf__ int8_t* src, const LoadData3DParamsV2Pro& loadDataParams)
{
ASCENDC_REPORT_NOT_SUPPORT(
false, "LoadData with LoadData3DParamsV2Pro from L1 Buffer(B1) to L0B Buffer(B2) with type int8_t");
}
template <>
__aicore__ inline void LoadData3DV2L12L0BCal(
__cb__ uint8_t* dst, __cbuf__ uint8_t* src, const LoadData3DParamsV2Pro& loadDataParams)
{
ASCENDC_REPORT_NOT_SUPPORT(
false, "LoadData with LoadData3DParamsV2Pro from L1 Buffer(B1) to L0B Buffer(B2) with type uint8_t");
}
* BroadCastVecToMM *
* ************************************************************************************************* */
template <typename T>
__aicore__ inline void BroadCastVecToMMCal(
__cc__ T* dstLocal, __ubuf__ T* srcLocal, const int32_t blockCount, const uint8_t blockLen, const uint8_t srcGap,
const uint8_t dstGap)
{
ASCENDC_REPORT_NOT_SUPPORT(false, "BroadCastVecToMM");
}
* InitL1Buffer *
* ************************************************************************************************* */
__aicore__ inline void CheckInitConstValueParams(uint16_t repeatTime, uint16_t blockNum, uint16_t dstGap)
{
ASCENDC_CHECK_VALUE_RANGE(repeatTime, 0, UINT15_MAX, "repeatTime", "InitConstValue");
ASCENDC_CHECK_VALUE_RANGE(blockNum, 0, UINT15_MAX, "blockNum", "InitConstValue");
ASCENDC_CHECK_VALUE_RANGE(dstGap, 0, UINT15_MAX, "dstGap", "InitConstValue");
}
template <typename T>
__aicore__ inline void InitL1BufferCal(__cbuf__ T* dst, const InitConstValueParams<T>& initConstValueParams)
{
if ASCEND_IS_AIC {
CheckInitConstValueParams(
initConstValueParams.repeatTimes, initConstValueParams.blockNum, initConstValueParams.dstGap);
int64_t repeatBit = (static_cast<uint64_t>(initConstValueParams.blockNum) << 16) |
(static_cast<uint64_t>(initConstValueParams.dstGap) << 32) |
initConstValueParams.repeatTimes;
if constexpr (IsSameType<T, bfloat16_t>::value) {
create_cbuf_matrix_bf16(dst, repeatBit, initConstValueParams.initValue);
} else if constexpr (IsSameType<T, uint32_t>::value || IsSameType<T, half>::value) {
create_cbuf_matrix(dst, repeatBit, (T)initConstValueParams.initValue);
} else if constexpr (IsSameType<T, int16_t>::value || IsSameType<T, uint16_t>::value) {
create_cbuf_matrix(dst, repeatBit, GetScalarBitcodeToHalf(initConstValueParams.initValue));
} else if constexpr (IsSameType<T, float>::value || IsSameType<T, int32_t>::value) {
create_cbuf_matrix(
dst, repeatBit, static_cast<uint32_t>(GetScalarBitcodeValue(initConstValueParams.initValue)));
} else {
ASCENDC_ASSERT(false, {
KERNEL_LOG(
KERNEL_ERROR,
"Failed to check dtype in InitConstValue, current api "
"support dtype combination is dst: half, int16_t, uint16_t, bfloat16_t, int32_t, uint32_t, float");
});
}
}
}
* InitL0ANzMatrix *
* ************************************************************************************************* */
template <typename T>
__aicore__ inline void InitL0ANzMatrixCal(__ca__ T* dst, const InitConstValueParams<T>& initConstValueParams)
{
if ASCEND_IS_AIC {
CheckInitConstValueParams(
initConstValueParams.repeatTimes, initConstValueParams.blockNum, initConstValueParams.dstGap);
int64_t repeatBit = (static_cast<uint64_t>(initConstValueParams.blockNum) << 16) |
(static_cast<uint64_t>(initConstValueParams.dstGap) << 32) |
initConstValueParams.repeatTimes;
if constexpr (IsSameType<T, bfloat16_t>::value) {
create_ca_matrix_bf16(dst, repeatBit, initConstValueParams.initValue);
} else if constexpr (IsSameType<T, uint32_t>::value || IsSameType<T, half>::value) {
create_ca_matrix(dst, repeatBit, (T)initConstValueParams.initValue);
} else if constexpr (IsSameType<T, int16_t>::value || IsSameType<T, uint16_t>::value) {
create_ca_matrix(dst, repeatBit, GetScalarBitcodeToHalf(initConstValueParams.initValue));
} else if constexpr (IsSameType<T, float>::value || IsSameType<T, int32_t>::value) {
create_ca_matrix(
dst, repeatBit, static_cast<uint32_t>(GetScalarBitcodeValue(initConstValueParams.initValue)));
} else {
ASCENDC_ASSERT(false, {
KERNEL_LOG(
KERNEL_ERROR,
"Failed to check dtype in InitConstValue, current api "
"support dtype combination is dst: half, int16_t, uint16_t, bfloat16_t, int32_t, uint32_t, float");
});
}
}
}
* InitL0BNzMatrix *
* ************************************************************************************************* */
template <typename T>
__aicore__ inline void InitL0BNzMatrixCal(__cb__ T* dst, const InitConstValueParams<T>& initConstValueParams)
{
if ASCEND_IS_AIC {
CheckInitConstValueParams(
initConstValueParams.repeatTimes, initConstValueParams.blockNum, initConstValueParams.dstGap);
int64_t repeatBit = (static_cast<uint64_t>(initConstValueParams.blockNum) << 16) |
(static_cast<uint64_t>(initConstValueParams.dstGap) << 32) |
initConstValueParams.repeatTimes;
if constexpr (IsSameType<T, bfloat16_t>::value) {
create_cb_matrix_bf16(dst, repeatBit, initConstValueParams.initValue);
} else if constexpr (IsSameType<T, uint32_t>::value || IsSameType<T, half>::value) {
create_cb_matrix(dst, repeatBit, (T)initConstValueParams.initValue);
} else if constexpr (IsSameType<T, int16_t>::value || IsSameType<T, uint16_t>::value) {
create_cb_matrix(dst, repeatBit, GetScalarBitcodeToHalf(initConstValueParams.initValue));
} else if constexpr (IsSameType<T, float>::value || IsSameType<T, int32_t>::value) {
create_cb_matrix(
dst, repeatBit, static_cast<uint32_t>(GetScalarBitcodeValue(initConstValueParams.initValue)));
} else {
ASCENDC_ASSERT(false, {
KERNEL_LOG(
KERNEL_ERROR,
"Failed to check dtype in InitConstValue, current api "
"support dtype combination is dst: half, int16_t, uint16_t, bfloat16_t, int32_t, uint32_t, float");
});
}
}
}
* SetLoadDataRepeat *
* ************************************************************************************************* */
__aicore__ inline void SetLoadDataRepeatCal(const LoadDataRepeatParam& repeatParams)
{
if ASCEND_IS_AIC {
uint64_t rptConfig = static_cast<uint64_t>(repeatParams.repeatStride) |
(static_cast<uint64_t>(repeatParams.repeatTime) << 16) |
(static_cast<uint64_t>(repeatParams.repeatMode) << 24);
set_l3d_rpt(rptConfig);
}
}
* SetLoadDataBoundary *
* ************************************************************************************************* */
__aicore__ inline void SetLoadDataBoundaryCal(uint32_t boundaryValue)
{
if ASCEND_IS_AIC {
set_l1_3d_size(static_cast<uint64_t>(boundaryValue));
}
}
* LoadImageToLocalCal *
* ************************************************************************************************* */
template <typename T>
__aicore__ inline void LoadImageToLocalCal(__cbuf__ T* dst, const LoadImageToLocalParams& loadDataParams)
{
if ASCEND_IS_AIC {
ASCENDC_ASSERT((SupportType<T, int8_t, half>()), {
KERNEL_LOG(
KERNEL_ERROR, "Failed to check dtype in "
"LoadImageToLocal, current api support dtype combination is dst: int8_t / half.");
});
load_image_to_cbuf(
dst, static_cast<uint16_t>(loadDataParams.horizSize - 1),
static_cast<uint16_t>(loadDataParams.vertSize - 1), loadDataParams.horizStartPos,
loadDataParams.vertStartPos, static_cast<uint16_t>(loadDataParams.srcHorizSize - 1),
loadDataParams.topPadSize, loadDataParams.botPadSize, loadDataParams.leftPadSize,
loadDataParams.rightPadSize, loadDataParams.sid);
}
}
* LoadDataUnzip *
* ************************************************************************************************* */
template <typename T>
__aicore__ inline void LoadDataUnzipToL1Cal(__cbuf__ T* dst, __gm__ T* src)
{
ASCENDC_REPORT_NOT_SUPPORT(false, "LoadDataUnzip");
}
template <typename T>
__aicore__ inline void LoadDataUnzipToL0BCal(__cb__ T* dst, __gm__ T* src)
{
ASCENDC_REPORT_NOT_SUPPORT(false, "LoadDataUnzip");
}
template <typename T>
__aicore__ inline void LoadDataUnzipToL0ACal(__ca__ T* dst, __gm__ T* src)
{
ASCENDC_REPORT_NOT_SUPPORT(false, "LoadDataUnzip");
}
}
#endif
#if defined(__UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_KERNEL_OPERATOR_MM_IMPL_H__)
#undef __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#undef __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_KERNEL_OPERATOR_MM_IMPL_H__
#endif
