* 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 softmax_common.h
* \brief
*/
#if !defined(__ASCENDC_INCLUDE_INTERNAL_HEADERS__)
#pragma message( \
"impl/adv_api/detail/activation/softmax/softmax_common.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 \"adv_api/activation/softmax.h\"\" and use public functions or variables defined in interface headers files.")
#define __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#define __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_SOFTMAX_COMMON_H__
#endif
#ifndef IMPL_ACTIVATION_SOFTMAX_SOFTMAX_COMMON_IMPL_H
#define IMPL_ACTIVATION_SOFTMAX_SOFTMAX_COMMON_IMPL_H
#include "kernel_pop_stack_buffer.h"
#include "softmax_common/softmax_common_utils.h"
#include "softmax_common/softmax_common_shape_process.h"
#include "softmax_common/softmax_tiling_func.h"
#include "softmax_common/softmax_common_broadcast.h"
#include "softmax_common/softmax_common_reduce.h"
#include "softmax_common/softmax_common_arithmetic.h"
namespace AscendC {
#if defined(__NPU_ARCH__) && \
(__NPU_ARCH__ == 3510 || __NPU_ARCH__ == 5102 || __NPU_ARCH__ == 3003 || __NPU_ARCH__ == 3113)
template <typename T>
constexpr __aicore__ inline void SoftmaxApiSupportedTypeCheck()
{
static_assert(
std::is_same<T, half>::value || std::is_same<T, float>::value,
"This Related Api of Softmax only support half/float input dtype");
}
#endif
__aicore__ inline void CreateSpecialFormatMask(
uint64_t& lowMask, const uint32_t& maskLen, const uint32_t& nzBlockCount,
const uint32_t& totalLen = SOFTMAX_SHAPE_NZ_BASIC_COUNT)
{
ASCENDC_ASSERT((maskLen <= totalLen), { KERNEL_LOG(KERNEL_ERROR, "maskLen must be less than totalLen"); });
if (totalLen == SOFTMAX_SHAPE_NZ_BASIC_COUNT) {
ASCENDC_ASSERT((nzBlockCount <= B32_BYTE_SIZE), {
KERNEL_LOG(KERNEL_ERROR, "nzBlockCount must be less than 4 when totalLen is 16");
});
}
if (totalLen >= B32_DATA_NUM_PER_BLOCK) {
ASCENDC_ASSERT((nzBlockCount <= B64_BYTE_SIZE), {
KERNEL_LOG(KERNEL_ERROR, "nzBlockCount must be less than 8 when totalLen is no greater than 8");
});
}
ASCENDC_ASSERT((nzBlockCount >= 1), { KERNEL_LOG(KERNEL_ERROR, "nzBlockCount must be large than 1"); });
uint16_t originalMask = totalLen == SOFTMAX_SHAPE_NZ_BASIC_COUNT ? 0xFFFF : 0xFF;
uint64_t defaultMask = originalMask >> (totalLen - maskLen);
lowMask = defaultMask;
for (uint32_t i = 0; i < nzBlockCount - 1; i++) {
lowMask = lowMask << totalLen;
lowMask = lowMask | defaultMask;
}
}
__aicore__ inline void BinaryComputeWithSpecialMask(
const LocalTensor<float>& dst, const LocalTensor<float>& src0, const LocalTensor<float>& src1, uint64_t mask[2],
const uint32_t& lastBlockMaskLen, const uint32_t& splitCount,
void (*func)(
const LocalTensor<float>&, const LocalTensor<float>&, const LocalTensor<float>&, uint64_t*, const uint8_t,
const BinaryRepeatParams&))
{
uint32_t repeat = splitCount / FLOAT_REPEAT_SIZE;
uint32_t tail = splitCount % FLOAT_REPEAT_SIZE;
uint32_t repeatRange = repeat / MAX_REPEAT_TIMES;
uint32_t repeatTail = repeat % MAX_REPEAT_TIMES;
const auto offsetCount = MAX_REPEAT_TIMES * FLOAT_REPEAT_SIZE;
uint32_t dstOffset = 0;
uint32_t src0Offset = 0;
uint32_t src1Offset = 0;
for (uint32_t i = 0; i < repeatRange; i++) {
func(
dst[i * offsetCount], src0[i * offsetCount], src1[i * offsetCount], mask, MAX_REPEAT_TIMES,
{1, 1, 1, DEFAULT_REPEAT_STRIDE, DEFAULT_REPEAT_STRIDE, DEFAULT_REPEAT_STRIDE});
}
if (repeatTail != 0) {
func(
dst[repeatRange * offsetCount], src0[repeatRange * offsetCount], src1[repeatRange * offsetCount], mask,
repeatTail, {1, 1, 1, DEFAULT_REPEAT_STRIDE, DEFAULT_REPEAT_STRIDE, DEFAULT_REPEAT_STRIDE});
}
if (tail != 0) {
uint64_t tailMask[2] = {0, 0};
CreateSpecialFormatMask(tailMask[0], lastBlockMaskLen, tail / SOFTMAX_SHAPE_NZ_BASIC_COUNT);
func(
dst[repeat * FLOAT_REPEAT_SIZE], src0[repeat * FLOAT_REPEAT_SIZE], src1[repeat * FLOAT_REPEAT_SIZE],
tailMask, 1, {1, 1, 1, DEFAULT_REPEAT_STRIDE, DEFAULT_REPEAT_STRIDE, DEFAULT_REPEAT_STRIDE});
}
}
__aicore__ inline void UnaryComputeWithSpecialMask(
const LocalTensor<float>& dst, const LocalTensor<float>& src, uint64_t mask[2], const uint32_t& lastBlockMaskLen,
const uint32_t& splitCount,
void (*func)(
const LocalTensor<float>&, const LocalTensor<float>&, uint64_t*, const uint8_t, const UnaryRepeatParams&))
{
uint32_t repeat = splitCount / FLOAT_REPEAT_SIZE;
uint32_t tail = splitCount % FLOAT_REPEAT_SIZE;
uint32_t repeatRange = repeat / MAX_REPEAT_TIMES;
uint32_t repeatTail = repeat % MAX_REPEAT_TIMES;
const auto offsetCount = MAX_REPEAT_TIMES * FLOAT_REPEAT_SIZE;
uint32_t dstOffset = 0;
uint32_t src0Offset = 0;
uint32_t src1Offset = 0;
for (uint32_t i = 0; i < repeatRange; i++) {
func(
dst[i * offsetCount], src[i * offsetCount], mask, MAX_REPEAT_TIMES,
{1, 1, DEFAULT_REPEAT_STRIDE, DEFAULT_REPEAT_STRIDE});
}
if (repeatTail != 0) {
func(
dst[repeatRange * offsetCount], src[repeatRange * offsetCount], mask, repeatTail,
{1, 1, DEFAULT_REPEAT_STRIDE, DEFAULT_REPEAT_STRIDE});
}
if (tail != 0) {
uint64_t tailMask[2] = {0, 0};
CreateSpecialFormatMask(tailMask[0], lastBlockMaskLen, tail / SOFTMAX_SHAPE_NZ_BASIC_COUNT);
func(
dst[repeat * FLOAT_REPEAT_SIZE], src[repeat * FLOAT_REPEAT_SIZE], tailMask, 1,
{1, 1, DEFAULT_REPEAT_STRIDE, DEFAULT_REPEAT_STRIDE});
}
}
};
#endif
#if defined(__UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_SOFTMAX_COMMON_H__)
#undef __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#undef __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_SOFTMAX_COMMON_H__
#endif
