* 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_fixpipe_impl.h
* \brief
*/
#if !defined(__ASCENDC_INCLUDE_INTERNAL_HEADERS__)
#pragma message("impl/basic_api/dav_m310/kernel_operator_fixpipe_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_FIXPIPE_IMPL_H__
#endif
#ifndef ASCENDC_MODULE_OPERATOR_FIXPIPE_IMPL_H
#define ASCENDC_MODULE_OPERATOR_FIXPIPE_IMPL_H
#include "kernel_struct_fixpipe.h"
#include "kernel_operator_block_sync_intf.h"
namespace AscendC {
* SPR *
* ************************************************************************************************* */
template <typename T>
__aicore__ inline void SetFixPipeConfigImpl(const LocalTensor<T> &reluPre, const LocalTensor<T> &quantPre,
bool isUnitFlag = false)
{
uint64_t config = 0;
config = config | ((uint64_t)reluPre.GetPhyAddr() >> 6);
config = config | (((uint64_t)quantPre.GetPhyAddr() >> 7) << 8);
config = config | (static_cast<uint64_t>(isUnitFlag) << 63);
set_fpc(config);
}
template <typename T, bool setRelu = false>
__aicore__ inline void SetFixPipeConfigImpl(const LocalTensor<T> &pre, bool isUnitFlag = false)
{
uint64_t config = 0;
if constexpr (setRelu) {
config = config | ((uint64_t)pre.GetPhyAddr() >> 6);
} else {
config =
config | (((uint64_t)pre.GetPhyAddr() >> 7) << 8);
}
config = config | (static_cast<uint64_t>(isUnitFlag) << 63);
set_fpc(config);
}
__aicore__ inline void SetFixpipeNz2ndFlagImpl(uint16_t ndNum, uint16_t srcNdStride, uint16_t dstNdStride)
{
uint64_t config = (static_cast<uint64_t>(dstNdStride) << 32) | (static_cast<uint64_t>(srcNdStride) << 16) |
(static_cast<uint64_t>(ndNum));
set_loop3_para(config);
}
__aicore__ inline void SetFixpipePreQuantFlagImpl(uint64_t config)
{
set_quant_pre(config);
}
__aicore__ inline void SetFixPipeClipReluImpl(uint64_t config)
{
(void)(config);
ASCENDC_REPORT_NOT_SUPPORT(false, "SetFixPipeClipRelu");
}
template <typename T>
__aicore__ inline void SetFixPipeAddrImpl(const LocalTensor<T> &eleWise, uint16_t c0ChStride)
{
ASCENDC_REPORT_NOT_SUPPORT(false, "SetFixPipeAddr");
}
* Fixpipe *
* ************************************************************************************************* */
const uint32_t L0C_SRC_ALIGN = 16 * sizeof(float);
__aicore__ inline bool IsVectorQuantMode(QuantMode_t quantPre)
{
return (quantPre == QuantMode_t::VDEQF16 || quantPre == QuantMode_t::VQF322B8_PRE ||
quantPre == QuantMode_t::VREQ8);
}
__aicore__ inline bool IsScalarQuantMode(QuantMode_t quantPre)
{
return (quantPre == QuantMode_t::DEQF16 || quantPre == QuantMode_t::QF322B8_PRE || quantPre == QuantMode_t::REQ8);
}
__aicore__ inline void SetLoop3Para(const FixpipeParamsM310& intriParams)
{
ASCENDC_ASSERT((intriParams.ndNum > 0), { KERNEL_LOG(KERNEL_ERROR, "ndNum must be larger than 0"); });
uint64_t loop3Para = static_cast<uint64_t>(intriParams.dstNdStride) << 32;
loop3Para |= static_cast<uint64_t>(intriParams.srcNdStride) << 16;
loop3Para |= static_cast<uint64_t>(intriParams.ndNum);
set_loop3_para(loop3Para);
}
template <typename T, typename U, const FixpipeConfig& config>
__aicore__ inline void CheckCommonFixpipeParam(__cc__ U *src, const FixpipeParamsM310 ¶ms)
{
#if ASCENDC_CPU_DEBUG
ASCENDC_CHECK_TENSOR_PTR_ALIGN(src, TPosition::CO1, L0C_SRC_ALIGN, "srcLocal", "Fixpipe");
if (params.isChannelSplit) {
ASCENDC_ASSERT((params.nSize <= UINT12_MAX && params.nSize >=1 && params.nSize % 8 == 0),
{KERNEL_LOG(KERNEL_ERROR,"Failed to check nSize value in Fixpipe, when isChannelSplit is true, its valid "
"range is 1 ~ 4095 and must be divisible by 8, current value is %u", params.nSize); });
} else if (config.format == CO2Layout::ROW_MAJOR) {
ASCENDC_CHECK_VALUE_RANGE(params.nSize, 1, UINT12_MAX, "nSize",
"Fixpipe when isChannelSplit is false and format is NZ2ND");
} else {
ASCENDC_ASSERT((params.nSize <= UINT12_MAX && params.nSize >=1 && params.nSize % 16 == 0),
{KERNEL_LOG(KERNEL_ERROR, "Failed to check nSize value in Fixpipe, when isChannelSplit is false and format "
"is NZ, its valid range is 1 ~ 4095 and must be divisible by 16, current value is %u", params.nSize); });
}
if constexpr(config.format == CO2Layout::ROW_MAJOR) {
ASCENDC_CHECK_VALUE_RANGE(params.mSize, 1, 8192, "mSize", "Fixpipe when format is ROW_MAJOR");
} else {
ASCENDC_CHECK_VALUE_RANGE(params.mSize, 1, UINT16_MAX, "mSize", "Fixpipe when format is NZ");
}
ASCENDC_ASSERT((params.dstStride != 0), {KERNEL_LOG(KERNEL_ERROR, "Failed to check dstStride value in Fixpipe, "\
"its valid range is 1 ~ 4294967295, current value is %u", params.dstStride);});
if (params.ndNum > 1) {
ASCENDC_CHECK_VALUE_RANGE(params.srcNdStride, 1, VALUE_512, "srcNdStride", "Fixpipe when ndNum is > 1");
ASCENDC_CHECK_VALUE_RANGE(params.dstNdStride, 1, UINT16_MAX, "dstNdStride", "Fixpipe when ndNum is > 1");
}
if constexpr(IsSameType<U, float>::value && SupportType<T, int8_t, uint8_t>()) {
ASCENDC_ASSERT((params.quantPre == QuantMode_t::QF322B8_PRE || params.quantPre == QuantMode_t::VQF322B8_PRE),
{KERNEL_LOG(KERNEL_ERROR, "Failed to check quantPre value in Fixpipe, when src is float, dst is "
"int8_t / uint8_t, supported values are QF322B8_PRE and VQF322B8_PRE");});
} else if constexpr(IsSameType<U, float>::value && IsSameType<T, half>::value) {
ASCENDC_ASSERT((params.quantPre == QuantMode_t::F322F16), {KERNEL_LOG(KERNEL_ERROR, "Failed to check quantPre "\
"value in Fixpipe, when src is float, dst is half, supported value is F322F16");});
} else if constexpr(IsSameType<U, int32_t>::value && SupportType<T, int8_t, uint8_t>()) {
ASCENDC_ASSERT((params.quantPre == QuantMode_t::REQ8 || params.quantPre == QuantMode_t::VREQ8),
{KERNEL_LOG(KERNEL_ERROR, "Failed to check quantPre value in Fixpipe, when src is int32_t, dst is "\
"int8_t / uint8_t, supported values are REQ8 and VREQ8");});
} else if constexpr(IsSameType<U, int32_t>::value && IsSameType<T, half>::value) {
ASCENDC_ASSERT((params.quantPre == QuantMode_t::DEQF16 || params.quantPre == QuantMode_t::VDEQF16),
{KERNEL_LOG(KERNEL_ERROR, "Failed to check quantPre value in Fixpipe, when src is int32_t, dst is half, "\
"supported values are DEQF16 and VDEQF16");});
}
#endif
}
template <typename T, typename U, const FixpipeConfig& config>
__aicore__ inline void CheckFixpipeL0C2UBParam(__ubuf__ T *dst, __cc__ U *src, const FixpipeParamsM310 ¶ms)
{
CheckCommonFixpipeParam<T, U, config>(src, params);
ASCENDC_CHECK_TENSOR_PTR_ALIGN(
dst, TPosition::VECCALC, ONE_BLK_SIZE, "dstLocal", "Fixpipe when dst position is VECIN/VECCALC/VECOUT");
static_assert((SupportType<Tuple<U, T>, Tuple<float, int8_t>, Tuple<float, uint8_t>, Tuple<float, half>,
Tuple<float, float>, Tuple<int32_t, int8_t>, Tuple<int32_t, uint8_t>,
Tuple<int32_t, half>, Tuple<int32_t, int32_t>>()), "Failed to check dtype in Fixpipe, when src position is "\
"CO1 and dst position is GM, support dtype combinations are src: float, dst: int8_t / uint8_t / half "\
"/ float; src: int32_t, dst: int8_t / uint8_t / half / int32_t");
if constexpr(IsSameType<U, float>::value && IsSameType<T, float>::value) {
ASCENDC_ASSERT((params.quantPre == QuantMode_t::NoQuant), {KERNEL_LOG(KERNEL_ERROR, "Failed to check quantPre "\
"value in Fixpipe, when src is float, dst is float, supported value is NoQuant");});
} else if constexpr(IsSameType<U, int32_t>::value && IsSameType<T, int32_t>::value) {
ASCENDC_ASSERT((params.quantPre == QuantMode_t::NoQuant), {KERNEL_LOG(KERNEL_ERROR, "Failed to check quantPre "\
"value in Fixpipe, when src is int32_t, dst is int32_t, supported value is NoQuant");});
}
if (params.isChannelSplit) {
ASCENDC_DEBUG_ASSERT((IsSameType<U, float>::value && IsSameType<T, float>::value), KERNEL_LOG_INTERNAL(KERNEL_ERROR, "Failed to check "\
"isChannelSplit value in Fixpipe, isChannelSplit can be set true only when src and dst are both float \n"));
ASCENDC_DEBUG_ASSERT((config.format != CO2Layout::ROW_MAJOR), KERNEL_LOG_INTERNAL(KERNEL_ERROR, "Failed to check format value in Fixpipe, "\
"when isChannelSplit is set true, format must be set as NZ \n"));
}
}
template <typename T, typename U, const FixpipeConfig& config>
__aicore__ inline void CheckFixpipeL0C2L1Param(__cbuf__ T *dst, __cc__ U *src, const FixpipeParamsM310 ¶ms)
{
CheckCommonFixpipeParam<T, U, config>(src, params);
#if ASCENDC_CPU_DEBUG
uint64_t dstAbsAddr = dst - (T*)(GetTPipePtr()->GetBaseAddr(int8_t(TPosition::A1)));
ASCENDC_ASSERT((dstAbsAddr % ONE_BLK_SIZE == 0), {KERNEL_LOG(KERNEL_ERROR, "Failed to check dstLocal start "\
"address alignment in Fixpipe, when dst position is A1, its start address must align with 32B");});
#endif
ASCENDC_DEBUG_ASSERT((!(params.isChannelSplit)), KERNEL_LOG_INTERNAL(KERNEL_ERROR, "Failed to check isChannelSplit in Fixpipe, when src position is "\
"CO1 and dst position is A1, isChannelSplit must be set as false \n"));
static_assert((SupportType<Tuple<U, T>, Tuple<float, int8_t>, Tuple<float, uint8_t>, Tuple<float, half>,
Tuple<float, float>, Tuple<int32_t, int32_t>, Tuple<int32_t, int8_t>, Tuple<int32_t, uint8_t>, Tuple<int32_t, half>>()),
"Failed to check dtype in Fixpipe, when src position is CO1 and dst position is A1, "\
"support dtype combinations are src: float, dst: int8_t / uint8_t / half; src: int32_t, dst: "\
"int8_t / uint8_t / half");
}
template <typename T, typename U, const FixpipeConfig& config>
__aicore__ inline void CheckFixpipeL0C2GMParam(__gm__ T *dst, __cc__ U *src, const FixpipeParamsM310 ¶ms)
{
CheckCommonFixpipeParam<T, U, config>(src, params);
static_assert((SupportType<Tuple<U, T>, Tuple<float, int8_t>, Tuple<float, uint8_t>, Tuple<float, half>,
Tuple<float, float>, Tuple<int32_t, int8_t>, Tuple<int32_t, uint8_t>, Tuple<int32_t, half>, Tuple<int32_t, int32_t>>()),
"Failed to check dtype in Fixpipe, when src position is CO1 "\
"and dst position is GM, support dtype combinations are src: float, dst: int8_t / uint8_t / half "\
"/ float; src: int32_t, dst: int8_t / uint8_t / half / int32_t");
if constexpr(IsSameType<U, float>::value && IsSameType<T, float>::value) {
ASCENDC_ASSERT((params.quantPre == QuantMode_t::NoQuant), {KERNEL_LOG(KERNEL_ERROR, "Failed to check quantPre "\
"value in Fixpipe, when src is float, dst is float, supported value is NoQuant");});
} else if constexpr(IsSameType<U, int32_t>::value && IsSameType<T, int32_t>::value) {
ASCENDC_ASSERT((params.quantPre == QuantMode_t::NoQuant), {KERNEL_LOG(KERNEL_ERROR, "Failed to check quantPre "\
"value in Fixpipe, when src is int32_t, dst is int32_t, supported value is NoQuant");});
}
if (params.isChannelSplit) {
ASCENDC_DEBUG_ASSERT((IsSameType<U, float>::value && IsSameType<T, float>::value), KERNEL_LOG_INTERNAL(KERNEL_ERROR, "Failed to check "\
"isChannelSplit value in Fixpipe, isChannelSplit can be set true only when src and dst are both float \n"));
ASCENDC_DEBUG_ASSERT((config.format != CO2Layout::ROW_MAJOR), KERNEL_LOG_INTERNAL(KERNEL_ERROR, "Failed to check format value in Fixpipe, "\
"when isChannelSplit is set true, format must be set as NZ \n"));
}
}
template <typename T, typename U, const FixpipeConfig& config>
__aicore__ inline FixpipeParamsM310 AdaptFixpipeParams(const FixpipeParamsM310 &intriParams)
{
FixpipeParamsM310 params(intriParams);
if constexpr (config.format == CO2Layout::NZ) {
params.dstStride = intriParams.dstStride * ONE_BLK_SIZE / sizeof(T);
}
if constexpr(config.format == CO2Layout::ROW_MAJOR) {
constexpr uint16_t oriUnit = 1024;
constexpr uint16_t multiples = oriUnit / (BLOCK_CUBE * sizeof(U));
params.srcNdStride = params.srcNdStride * multiples;
}
return params;
}
struct FixpipeTilingM310 {
uint16_t nIterNum = 0;
uint16_t nSize = 0;
bool isDb = false;
uint16_t tailNSize = 0;
};
__aicore__ inline FixpipeTilingM310 GenFixpipeTilingM310(uint16_t n)
{
FixpipeTilingM310 tiling;
uint16_t maxDeqNums = 256;
if (n <= maxDeqNums) {
tiling.nIterNum = 1;
tiling.nSize = n;
tiling.isDb = false;
tiling.tailNSize = 0;
} else {
tiling.isDb = true;
uint16_t dbMaxDeqNums = maxDeqNums / 2;
tiling.nIterNum = n / dbMaxDeqNums;
tiling.nSize = dbMaxDeqNums;
tiling.tailNSize = n % dbMaxDeqNums;
}
return tiling;
}
__aicore__ inline void CopyDeqTensorToFbuf(
__cbuf__ uint64_t *cbufWorkspace, const FixpipeTilingM310 &fixpipeTiling, uint16_t calNSize, uint16_t nIterIndex)
{
uint16_t deqDataSize = DivCeil(calNSize * sizeof(uint64_t), 128) * 128;
__fbuf__ uint64_t *deqTensorTempBuf =
AscendCUtils::GetTemporaryFbBufferAddr<uint64_t>(0, deqDataSize / sizeof(uint64_t));
uint32_t deqValueOffset = nIterIndex * fixpipeTiling.nSize;
constexpr uint16_t fbufBurstLenUnit = 64;
uint16_t fbufBurstLen = deqDataSize / fbufBurstLenUnit;
copy_cbuf_to_fbuf(deqTensorTempBuf, cbufWorkspace + deqValueOffset, 1, fbufBurstLen, 0, 0);
uint64_t deqTensorAddr = ((uint64_t)deqTensorTempBuf >> static_cast<uint64_t>(7)) << 8;
set_fpc(deqTensorAddr);
AscendCUtils::FreeTemporaryFbBuffer<uint64_t>(deqTensorTempBuf);
}
template <typename T, typename U, const FixpipeConfig &config>
__aicore__ inline void FixpipeL0cToUB(__ubuf__ T *dst, __cc__ U *src, const FixpipeParamsM310 &intriParams,
const FixpipeTilingM310 &fixpipeTiling, uint16_t calNSize, uint16_t nIterIndex = 0)
{
uint16_t cburstNum = fixpipeTiling.nSize / 16;
uint32_t srcOffset = cburstNum * nIterIndex * intriParams.srcStride * BLOCK_CUBE;
uint32_t dstOffset = 0;
bool nz2ndEn = false;
if constexpr (config.format == CO2Layout::ROW_MAJOR) {
dstOffset = nIterIndex * fixpipeTiling.nSize;
nz2ndEn = true;
} else {
dstOffset = cburstNum * nIterIndex * intriParams.dstStride * 32 / sizeof(T);
}
return copy_matrix_cc_to_ub((__ubuf__ T*)(dst + dstOffset), (__cc__ U*)(src + srcOffset), 0,
calNSize, intriParams.mSize, intriParams.dstStride, intriParams.srcStride, 0, intriParams.unitFlag,
intriParams.quantPre, static_cast<uint8_t>(intriParams.reluEn),
intriParams.isChannelSplit, nz2ndEn, false);
}
template <typename T, typename U, const FixpipeConfig& config>
__aicore__ inline void FixpipeL0C2UBImpl(__ubuf__ T *dst, __cc__ U *src, const FixpipeParamsM310 &intriParams)
{
const FixpipeParamsM310& params = AdaptFixpipeParams<T, U, config>(intriParams);
CheckFixpipeL0C2UBParam<T, U, config>(dst, src, params);
if constexpr (config.format == CO2Layout::ROW_MAJOR) {
SetLoop3Para(params);
}
make code for scalar quant mode:
1. copy deq scalar u64 immediate
2. code gen: move data from l0c to ub
*/
if (IsScalarQuantMode(params.quantPre)) {
set_quant_pre(params.deqScalar);
}
PipeBarrier<PIPE_FIX>();
FixpipeTilingM310 fixpipeTiling;
FixpipeL0cToUB<T, U, config>(dst, src, params, fixpipeTiling, params.nSize, 0);
}
template <typename T, typename U, const FixpipeConfig &config>
__aicore__ inline void FixpipeL0C2UBImplN(__ubuf__ T *dst, __cc__ U *src, __cbuf__ uint64_t *cbufWorkspace,
const FixpipeParamsM310 &intriParams, const FixpipeTilingM310 &fixpipeTiling, uint16_t calNSize,
uint16_t nIterIndex)
{
CopyDeqTensorToFbuf(cbufWorkspace, fixpipeTiling, calNSize, nIterIndex);
PipeBarrier<PIPE_FIX>();
FixpipeL0cToUB<T, U, config>(dst, src, intriParams, fixpipeTiling, calNSize, nIterIndex);
}
template <typename T, typename U, const FixpipeConfig &config>
__aicore__ inline void FixpipeL0C2UBImpl(
__ubuf__ T *dst, __cc__ U *src, __cbuf__ uint64_t *cbufWorkspace, const FixpipeParamsM310 &intriParams)
{
const FixpipeParamsM310& params = AdaptFixpipeParams<T, U, config>(intriParams);
CheckFixpipeL0C2UBParam<T, U, config>(dst, src, params);
if constexpr (config.format == CO2Layout::ROW_MAJOR) {
SetLoop3Para(params);
}
make code for vector quant mode:
1. generate tiling
2. copy deq tensor from l1 to fb0
3. code gen: move data from l0c to ub
*/
FixpipeTilingM310 fixpipeTiling = GenFixpipeTilingM310(params.nSize);
if (IsVectorQuantMode(params.quantPre)) {
for (uint16_t i = 0; i < fixpipeTiling.nIterNum; ++i) {
FixpipeL0C2UBImplN<T, U, config>(
dst, src, cbufWorkspace, params, fixpipeTiling, fixpipeTiling.nSize, i);
}
if (fixpipeTiling.tailNSize > 0) {
FixpipeL0C2UBImplN<T, U, config>(
dst, src, cbufWorkspace, params, fixpipeTiling, fixpipeTiling.tailNSize, fixpipeTiling.nIterNum);
}
return;
}
}
template <typename T, typename U, const FixpipeConfig &config>
__aicore__ inline void FixpipeL0cToL1(__cbuf__ T *dst, __cc__ U *src, const FixpipeParamsM310 &intriParams,
const FixpipeTilingM310 &fixpipeTiling, uint16_t calNSize, uint16_t nIterIndex = 0)
{
uint16_t cburstNum = fixpipeTiling.nSize / 16;
uint32_t srcOffset = cburstNum * nIterIndex * intriParams.srcStride * BLOCK_CUBE;
uint32_t dstOffset = 0;
bool nz2ndEn = false;
if constexpr (config.format == CO2Layout::ROW_MAJOR) {
dstOffset = nIterIndex * fixpipeTiling.nSize;
nz2ndEn = true;
} else {
dstOffset = cburstNum * nIterIndex * intriParams.dstStride * 32 / sizeof(T);
}
return copy_matrix_cc_to_cbuf((__cbuf__ T*)(dst + dstOffset), (__cc__ U*)(src + srcOffset), 0,
calNSize, intriParams.mSize, intriParams.dstStride, intriParams.srcStride, 0, intriParams.unitFlag,
intriParams.quantPre, static_cast<uint8_t>(intriParams.reluEn),
false, nz2ndEn, false);
}
template <typename T, typename U, const FixpipeConfig& config>
__aicore__ inline void FixpipeL0C2L1Impl(__cbuf__ T *dst, __cc__ U *src, const FixpipeParamsM310 &intriParams)
{
const FixpipeParamsM310& params = AdaptFixpipeParams<T, U, config>(intriParams);
CheckFixpipeL0C2L1Param<T, U, config>(dst, src, intriParams);
if constexpr (config.format == CO2Layout::ROW_MAJOR) {
SetLoop3Para(params);
}
make code for scalar quant mode:
1. copy deq scalar u64 immediate
2. code gen: move data from l0c to l1
*/
if (IsScalarQuantMode(params.quantPre)) {
set_quant_pre(params.deqScalar);
}
PipeBarrier<PIPE_FIX>();
FixpipeTilingM310 fixpipeTiling;
FixpipeL0cToL1<T, U, config>(dst, src, params, fixpipeTiling, params.nSize);
}
template <typename T, typename U, const FixpipeConfig &config>
__aicore__ inline void FixpipeL0C2L1ImplN(__cbuf__ T *dst, __cc__ U *src, __cbuf__ uint64_t *cbufWorkspace,
const FixpipeParamsM310 &intriParams, const FixpipeTilingM310 &fixpipeTiling, uint16_t calNSize,
uint16_t nIterIndex)
{
CopyDeqTensorToFbuf(cbufWorkspace, fixpipeTiling, calNSize, nIterIndex);
PipeBarrier<PIPE_FIX>();
FixpipeL0cToL1<T, U, config>(dst, src, intriParams, fixpipeTiling, calNSize, nIterIndex);
}
template <typename T, typename U, const FixpipeConfig &config>
__aicore__ inline void FixpipeL0C2L1Impl(
__cbuf__ T *dst, __cc__ U *src, __cbuf__ uint64_t *cbufWorkspace, const FixpipeParamsM310 &intriParams)
{
const FixpipeParamsM310& params = AdaptFixpipeParams<T, U, config>(intriParams);
CheckFixpipeL0C2L1Param<T, U, config>(dst, src, params);
if constexpr (config.format == CO2Layout::ROW_MAJOR) {
SetLoop3Para(params);
}
make code for vector quant mode:
1. generate tiling
2. copy deq tensor from l1 to fb0
3. code gen: move data from l0c to l1
*/
FixpipeTilingM310 fixpipeTiling = GenFixpipeTilingM310(params.nSize);
if (IsVectorQuantMode(params.quantPre)) {
for (uint16_t i = 0; i < fixpipeTiling.nIterNum; ++i) {
FixpipeL0C2L1ImplN<T, U, config>(
dst, src, cbufWorkspace, params, fixpipeTiling, fixpipeTiling.nSize, i);
}
if (fixpipeTiling.tailNSize > 0) {
FixpipeL0C2L1ImplN<T, U, config>(
dst, src, cbufWorkspace, params, fixpipeTiling, fixpipeTiling.tailNSize, fixpipeTiling.nIterNum);
}
return;
}
}
template <typename T>
__aicore__ inline uint64_t GetGMLength(
const FixpipeParamsM310 &intriParams, const uint16_t &calNSize, const bool &nz2ndEn)
{
const uint16_t dstEleSize = sizeof(T);
constexpr uint16_t dstStrideUnit = 32;
constexpr uint16_t fractalNsize = 16;
uint64_t cburstNum = calNSize / fractalNsize;
uint64_t gmLen;
gmLen =
(cburstNum - 1) * intriParams.dstStride * dstStrideUnit + intriParams.mSize * fractalNsize * dstEleSize;
if (nz2ndEn) {
gmLen = (static_cast<uint64_t>(intriParams.ndNum) - 1) * dstEleSize * intriParams.dstNdStride +
(intriParams.mSize - 1) * intriParams.dstStride * dstEleSize + cburstNum * fractalNsize * dstEleSize;
}
return gmLen;
}
template <typename T, typename U, const FixpipeConfig &config>
__aicore__ inline void FixpipeL0cToOut(__gm__ T *dst, __cc__ U *src, const FixpipeParamsM310 &intriParams,
const FixpipeTilingM310 &fixpipeTiling, uint16_t calNSize, uint16_t nIterIndex = 0)
{
uint16_t cburstNum = fixpipeTiling.nSize / 16;
uint32_t srcOffset = cburstNum * nIterIndex * intriParams.srcStride * BLOCK_CUBE;
uint32_t dstOffset = 0;
bool nz2ndEn = false;
if constexpr (config.format == CO2Layout::ROW_MAJOR) {
dstOffset = nIterIndex * fixpipeTiling.nSize;
nz2ndEn = true;
} else {
dstOffset = cburstNum * nIterIndex * intriParams.dstStride * 32 / sizeof(T);
}
if constexpr (g_gm_overflow_check) {
uint64_t gmLen = GetGMLength<T>(intriParams, calNSize, nz2ndEn);
AscendCUtils::CheckGmMemOverflow((__gm__ T *)(dst + dstOffset), false, gmLen);
}
return copy_matrix_cc_to_gm((__gm__ T*)(dst + dstOffset), (__cc__ U*)(src + srcOffset), 0,
calNSize, intriParams.mSize, intriParams.dstStride, intriParams.srcStride, 0, intriParams.unitFlag,
intriParams.quantPre, static_cast<uint8_t>(intriParams.reluEn),
intriParams.isChannelSplit, nz2ndEn, false);
}
template <typename T, typename U, const FixpipeConfig& config>
__aicore__ inline void FixpipeL0C2GMImpl(__gm__ T *dst, __cc__ U *src, const FixpipeParamsM310 &intriParams)
{
const FixpipeParamsM310& params = AdaptFixpipeParams<T, U, config>(intriParams);
CheckFixpipeL0C2GMParam<T, U, config>(dst, src, params);
if constexpr (config.format == CO2Layout::ROW_MAJOR) {
SetLoop3Para(params);
}
make code for scalar quant mode:
1. copy deq scalar u64 immediate
2. code gen: move data from l0c to gm
*/
if (IsScalarQuantMode(params.quantPre)) {
set_quant_pre(params.deqScalar);
}
PipeBarrier<PIPE_FIX>();
FixpipeTilingM310 fixpipeTiling;
FixpipeL0cToOut<T, U, config>(dst, src, params, fixpipeTiling, params.nSize);
}
template <typename T, typename U, const FixpipeConfig &config>
__aicore__ inline void FixpipeL0C2GMImplN(__gm__ T *dst, __cc__ U *src, __cbuf__ uint64_t *cbufWorkspace,
const FixpipeParamsM310 &intriParams, const FixpipeTilingM310 &fixpipeTiling, uint16_t calNSize,
uint16_t nIterIndex)
{
CopyDeqTensorToFbuf(cbufWorkspace, fixpipeTiling, calNSize, nIterIndex);
PipeBarrier<PIPE_FIX>();
FixpipeL0cToOut<T, U, config>(dst, src, intriParams, fixpipeTiling, calNSize, nIterIndex);
}
template <typename T, typename U, const FixpipeConfig &config>
__aicore__ inline void FixpipeL0C2GMImpl(
__gm__ T *dst, __cc__ U *src, __cbuf__ uint64_t *cbufWorkspace, const FixpipeParamsM310 &intriParams)
{
const FixpipeParamsM310& params = AdaptFixpipeParams<T, U, config>(intriParams);
CheckFixpipeL0C2GMParam<T, U, config>(dst, src, params);
if constexpr (config.format == CO2Layout::ROW_MAJOR) {
SetLoop3Para(params);
}
make code for vector quant mode:
1. generate tiling
2. copy deq tensor from gm to fb0 (l1 -> fb0)
3. code gen: move data from l0c to gm
*/
FixpipeTilingM310 fixpipeTiling = GenFixpipeTilingM310(params.nSize);
if (IsVectorQuantMode(params.quantPre)) {
for (uint16_t i = 0; i < fixpipeTiling.nIterNum; ++i) {
FixpipeL0C2GMImplN<T, U, config>(
dst, src, cbufWorkspace, params, fixpipeTiling, fixpipeTiling.nSize, i);
}
if (fixpipeTiling.tailNSize > 0) {
FixpipeL0C2GMImplN<T, U, config>(dst, src, cbufWorkspace, params,
fixpipeTiling, fixpipeTiling.tailNSize, fixpipeTiling.nIterNum);
}
return;
}
}
}
#endif
#if defined(__UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_KERNEL_OPERATOR_FIXPIPE_IMPL_H__)
#undef __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#undef __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_KERNEL_OPERATOR_FIXPIPE_IMPL_H__
#endif