* 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 dynamic_quant_v2_regbase_apt.cpp
* \brief dynamic_quant_v2 kernel enter
*/
#include "kernel_operator.h"
#include "../dynamic_quant/arch35/dynamic_quant_struct.h"
#include "../dynamic_quant/arch35/dynamic_quant_regbase_full_load.h"
#include "../dynamic_quant/arch35/dynamic_quant_regbase_moe_full_load.h"
#include "../dynamic_quant/arch35/dynamic_quant_regbase_large_shape_db.h"
#include "../dynamic_quant/arch35/dynamic_quant_regbase_moe_large_shape.h"
#include "../dynamic_quant/arch35/dynamic_quant_regbase_full_load_pertensor.h"
#include "../dynamic_quant/arch35/dynamic_quant_regbase_large_shape_db_pertensor.h"
#include "../dynamic_quant/arch35/dynamic_quant_regbase_perchannel_full_load.h"
#include "../dynamic_quant/arch35/dynamic_quant_regbase_perchannel_recompute.h"
#include "../dynamic_quant/arch35/dynamic_quant_regbase_perchannel_split_m.h"
#define FLOAT_OVERFLOW_MODE_CTRL 60
namespace
{
using namespace AscendC;
using namespace DynamicQuantNDOpt;
using namespace DynamicQuantNDOpt2;
using namespace DynamicQuantNDPerTensorOpt;
using namespace DynamicQuantNDPerTensorOpt2;
using namespace DynamicQuantPerChannel;
template<uint64_t V>
using UIntAsBool = std::integral_constant<bool, V != 0>;
template <uint64_t useDb, uint64_t quantMode, uint64_t hasSmooth, uint64_t isSymmetrical>
__global__ __aicore__ void dynamic_quant_v2(GM_ADDR x, GM_ADDR smooth_scales, GM_ADDR group_index, GM_ADDR y,
GM_ADDR scale, GM_ADDR offset, GM_ADDR workSpace, GM_ADDR tiling)
{
#if (__NPU_ARCH__ == 3510)
int64_t oriOverflowMode = AscendC::GetCtrlSpr<FLOAT_OVERFLOW_MODE_CTRL, FLOAT_OVERFLOW_MODE_CTRL>();
#endif
KERNEL_TASK_TYPE_DEFAULT(KERNEL_TYPE_MIX_AIV_1_0);
TPipe pipe;
GET_TILING_DATA(tilingData, tiling);
GM_ADDR usrWorkspace = GetUserWorkspace(workSpace);
if constexpr (quantMode == TPL_COMMON_FULL_LOAD) {
DynamicQuantRegbaseFullLoad<
DTYPE_X, DTYPE_Y,
UIntAsBool<hasSmooth>::value,
static_cast<uint32_t>(useDb + 1),
UIntAsBool<isSymmetrical>::value
> op(&pipe);
op.Init(x, smooth_scales, y, scale, offset, workSpace, &tilingData);
op.Process();
} else if constexpr (quantMode == TPL_COMMON_LARGE_SHAPE) {
DynamicQuantLargeShapeDb<DTYPE_X, DTYPE_Y,
static_cast<int64_t>(hasSmooth),
UIntAsBool<isSymmetrical>::value
> op(&pipe);
op.Init(x, smooth_scales, y, scale, offset, workSpace, tilingData);
op.Process();
} else if constexpr (quantMode == TPL_MOE_FULL_LOAD) {
DynamicQuantV2Op::DynamicQuantRegbaseFullLoadMOE<
DTYPE_X, DTYPE_Y,
UIntAsBool<hasSmooth>::value,
static_cast<uint32_t>(useDb + 1),
UIntAsBool<isSymmetrical>::value
> op(&pipe);
op.Init(x, smooth_scales, group_index, y, scale, offset, workSpace, &tilingData);
op.Process();
} else if constexpr (quantMode == TPL_MOE_LARGE_SHAPE) {
DynamicQuantRegBase::DynamicQuantLargeShapeMOE<
DTYPE_X, DTYPE_Y,
static_cast<int64_t>(hasSmooth),
UIntAsBool<isSymmetrical>::value
> op(&pipe);
op.Init(x, smooth_scales, group_index, y, scale, offset, workSpace, tilingData);
op.Process();
} else if constexpr (quantMode == TPL_PER_TENSOR_FULL_LOAD) {
DynamicQuantRegbaseFullLoadPertensor<DTYPE_X, DTYPE_Y, static_cast<int64_t>(hasSmooth),static_cast<uint32_t>(useDb + 1),UIntAsBool<isSymmetrical>::value> op(&pipe);
op.Init(x, smooth_scales, y, scale, offset, usrWorkspace, &tilingData);
op.Process();
} else if constexpr (quantMode == TPL_PER_TENSOR_LARGE_SHAPE) {
DynamicQuantLargeShapeDbPertensor<DTYPE_X, DTYPE_Y, static_cast<int64_t>(hasSmooth),UIntAsBool<isSymmetrical>::value> op(&pipe);
op.Init(x, smooth_scales, y, scale, offset, usrWorkspace, tilingData);
op.Process();
} else if constexpr (quantMode == TPL_EMPTY_TENSOR) {
return ;
} else if constexpr (quantMode == TPL_PER_CHANNEL_FULL_LOAD) {
DynamicQuantRegbasePerChannnelFullLoad<DTYPE_X, DTYPE_Y, static_cast<int64_t>(hasSmooth), UIntAsBool<isSymmetrical>::value> op(&pipe);
op.Init(x, smooth_scales, y, scale, offset, usrWorkspace, &tilingData);
op.Process();
} else if constexpr (quantMode == TPL_PER_CHANNEL_RECOMPUTE) {
DynamicQuantRegbasePerChannnelRecompute<DTYPE_X, DTYPE_Y, static_cast<int64_t>(hasSmooth), UIntAsBool<isSymmetrical>::value> op(&pipe);
op.Init(x, smooth_scales, y, scale, offset, usrWorkspace, &tilingData);
op.Process();
} else if constexpr (quantMode == TPL_PER_CHANNEL_SPLIT_M) {
DynamicQuantRegbasePerChannnelSplitM<DTYPE_X, DTYPE_Y, static_cast<int64_t>(hasSmooth), UIntAsBool<isSymmetrical>::value> op(&pipe);
op.Init(x, smooth_scales, y, scale, offset, usrWorkspace, &tilingData);
op.Process();
}
#if (__NPU_ARCH__ == 3510)
AscendC::SetCtrlSpr<FLOAT_OVERFLOW_MODE_CTRL, FLOAT_OVERFLOW_MODE_CTRL>(oriOverflowMode);
#endif
}
}