* 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 logical_template_impl.h
* \brief
*/
#if !defined(__ASCENDC_INCLUDE_INTERNAL_HEADERS__)
#pragma message( \
"impl/adv_api/detail/math/logical_template/logical_template.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/math/logical_and.h\"\" and use public functions or variables defined in interface headers files.")
#define __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#define __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_MATH_LOGICAL_TEMPLATE_LOGICAL_TEMPLATE_H__
#endif
#ifndef LIB_MATH_LOGICAL_TEMPLATE_IMPL_H
#define LIB_MATH_LOGICAL_TEMPLATE_IMPL_H
#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 3510 || __NPU_ARCH__ == 5102)
#include "kernel_tensor.h"
#include "kernel_basic_intf.h"
namespace AscendC {
constexpr uint32_t LOGICAL_TEMPLATE_B64_REPEAT_STRIDE = 2;
template <
auto func, typename T, typename U, typename RegT, typename RegU, const Reg::RegTrait& Trait = Reg::RegTraitNumOne>
__simd_vf__ inline void LogicalTemplateVF(
__ubuf__ T* dst, __ubuf__ U* src0, __ubuf__ U* src1, uint16_t repeatTime, uint32_t count, uint32_t oneRepElm)
{
RegT dstVreg;
RegT brcOneReg;
RegT brcZeroReg;
RegU src0Vreg;
RegU src1Vreg;
Reg::MaskReg mask;
Reg::MaskReg cmpMask0;
Reg::MaskReg cmpMask1;
Reg::MaskReg cmpMask2;
Reg::Duplicate(brcOneReg, 1u);
Reg::Duplicate(brcZeroReg, 0u);
for (uint16_t i = 0; i < repeatTime; ++i) {
mask = Reg::UpdateMask<U, Trait>(count);
Reg::LoadAlign(src0Vreg, src0 + i * oneRepElm);
Reg::LoadAlign(src1Vreg, src1 + i * oneRepElm);
Reg::CompareScalar<U, CMPMODE::NE>(cmpMask0, src0Vreg, static_cast<U>(0), mask);
Reg::CompareScalar<U, CMPMODE::NE>(cmpMask1, src1Vreg, static_cast<U>(0), mask);
func(cmpMask2, cmpMask0, cmpMask1, mask);
if constexpr (sizeof(U) == 2) {
Reg::MaskPack(cmpMask2, cmpMask2);
Reg::MaskPack(mask, mask);
} else if constexpr (sizeof(U) == 4 || sizeof(U) == 8) {
Reg::MaskPack(cmpMask2, cmpMask2);
Reg::MaskPack(cmpMask2, cmpMask2);
Reg::MaskPack(mask, mask);
Reg::MaskPack(mask, mask);
}
Reg::Select(dstVreg, brcOneReg, brcZeroReg, cmpMask2);
Reg::StoreAlign(dst + i * oneRepElm, dstVreg, mask);
}
}
template <auto func, typename T, typename U>
__aicore__ inline void LogicalTemplateImpl(
const LocalTensor<T>& dst, const LocalTensor<U>& src0, const LocalTensor<U>& src1, const uint32_t count)
{
static_assert(SupportType<T, bool>(), "only support bool data type on current device!");
static_assert(
SupportType<
U, bool, uint8_t, int8_t, half, bfloat16_t, uint16_t, int16_t, float, uint32_t, int32_t, uint64_t,
int64_t>(),
"only support bool/uint8_t/int8_t/half/bfloat16_t/"
"uint16_t/int16_t/float/uint32_t/int32_t/uint64_t/int64_t data type on current device!");
using RegT = Reg::RegTensor<T>;
if constexpr (sizeof(U) == 8) {
using RegU = Reg::RegTensor<U, Reg::RegTraitNumTwo>;
constexpr uint32_t oneRepElm =
static_cast<uint32_t>(GetVecLen() / sizeof(U) * LOGICAL_TEMPLATE_B64_REPEAT_STRIDE);
uint16_t repeatTime = static_cast<uint16_t>(CeilDivision(count, oneRepElm));
LogicalTemplateVF<func, T, U, RegT, RegU, Reg::RegTraitNumTwo>(
(__ubuf__ T*)dst.GetPhyAddr(), (__ubuf__ U*)src0.GetPhyAddr(), (__ubuf__ U*)src1.GetPhyAddr(), repeatTime,
count, oneRepElm);
} else {
constexpr uint32_t oneRepElm = static_cast<uint32_t>(GetVecLen() / sizeof(U));
uint16_t repeatTime = static_cast<uint16_t>(CeilDivision(count, oneRepElm));
if constexpr (Std::is_same_v<U, bool>) {
using RegU = Reg::RegTensor<uint8_t>;
LogicalTemplateVF<func, T, uint8_t, RegT, RegU>(
(__ubuf__ T*)dst.GetPhyAddr(), (__ubuf__ uint8_t*)src0.GetPhyAddr(),
(__ubuf__ uint8_t*)src1.GetPhyAddr(), repeatTime, count, oneRepElm);
} else {
using RegU = Reg::RegTensor<U>;
LogicalTemplateVF<func, T, U, RegT, RegU>(
(__ubuf__ T*)dst.GetPhyAddr(), (__ubuf__ U*)src0.GetPhyAddr(), (__ubuf__ U*)src1.GetPhyAddr(),
repeatTime, count, oneRepElm);
}
}
}
template <
auto func, typename T, typename U, typename RegT, typename RegU, const Reg::RegTrait& Trait = Reg::RegTraitNumOne>
__simd_vf__ inline void LogicalTemplateBothTensorVF(
__ubuf__ T* dst, __ubuf__ U* src, __ubuf__ U* scalar, uint16_t repeatTime, uint32_t count, uint32_t oneRepElm)
{
RegT dstVreg;
RegT brcOneReg;
RegT brcZeroReg;
RegU srcVreg;
RegU dupVreg;
Reg::MaskReg mask;
Reg::MaskReg cmpMask0;
Reg::MaskReg cmpMask1;
Reg::MaskReg cmpMask2;
Reg::MaskReg fullMask;
Reg::Duplicate(brcOneReg, 1u);
Reg::Duplicate(brcZeroReg, 0u);
fullMask = Reg::CreateMask<U, Reg::MaskPattern::ALL, Trait>();
if constexpr (sizeof(U) == 1) {
Reg::LoadAlign<U, Reg::LoadDist::DIST_BRC_B8>(dupVreg, scalar);
} else if constexpr (sizeof(U) == 2) {
Reg::LoadAlign<U, Reg::LoadDist::DIST_BRC_B16>(dupVreg, scalar);
} else if constexpr (sizeof(U) == 4) {
Reg::LoadAlign<U, Reg::LoadDist::DIST_BRC_B32>(dupVreg, scalar);
} else if constexpr (sizeof(U) == 8) {
Reg::LoadAlign(dupVreg, scalar);
Reg::Duplicate(dupVreg, dupVreg, fullMask);
}
Reg::CompareScalar<U, CMPMODE::NE>(cmpMask1, dupVreg, static_cast<U>(0), fullMask);
for (uint16_t i = 0; i < repeatTime; ++i) {
mask = Reg::UpdateMask<U, Trait>(count);
Reg::LoadAlign(srcVreg, src + i * oneRepElm);
Reg::CompareScalar<U, CMPMODE::NE>(cmpMask0, srcVreg, static_cast<U>(0), mask);
func(cmpMask2, cmpMask0, cmpMask1, mask);
if constexpr (sizeof(U) == 2) {
Reg::MaskPack(cmpMask2, cmpMask2);
Reg::MaskPack(mask, mask);
} else if constexpr (sizeof(U) == 4 || sizeof(U) == 8) {
Reg::MaskPack(cmpMask2, cmpMask2);
Reg::MaskPack(cmpMask2, cmpMask2);
Reg::MaskPack(mask, mask);
Reg::MaskPack(mask, mask);
}
Reg::Select(dstVreg, brcOneReg, brcZeroReg, cmpMask2);
Reg::StoreAlign(dst + i * oneRepElm, dstVreg, mask);
}
}
template <
auto func, typename T, typename U, typename RegT, typename RegU, const Reg::RegTrait& Trait = Reg::RegTraitNumOne>
__simd_vf__ inline void LogicalTemplateSingleScalarVF(
__ubuf__ T* dst, __ubuf__ U* src, U scalar, uint16_t repeatTime, uint32_t count, uint32_t oneRepElm)
{
RegT dstVreg;
RegT brcZeroReg;
RegT brcOneReg;
RegU srcVreg;
RegU dupVreg;
Reg::MaskReg mask;
Reg::MaskReg fullMask;
Reg::MaskReg cmpMask0;
Reg::MaskReg cmpMask1;
Reg::MaskReg cmpMask2;
Reg::Duplicate(brcOneReg, 1u);
Reg::Duplicate(brcZeroReg, 0u);
fullMask = Reg::CreateMask<U, Reg::MaskPattern::ALL, Trait>();
Reg::Duplicate(dupVreg, scalar);
Reg::CompareScalar<U, CMPMODE::NE>(cmpMask0, dupVreg, static_cast<U>(0), fullMask);
for (uint16_t i = 0; i < repeatTime; ++i) {
mask = Reg::UpdateMask<U, Trait>(count);
Reg::LoadAlign(srcVreg, src + i * oneRepElm);
Reg::CompareScalar<U, CMPMODE::NE>(cmpMask1, srcVreg, static_cast<U>(0), mask);
func(cmpMask2, cmpMask0, cmpMask1, mask);
if constexpr (sizeof(U) == 2) {
Reg::MaskPack(cmpMask2, cmpMask2);
Reg::MaskPack(mask, mask);
} else if constexpr (sizeof(U) == 4 || sizeof(U) == 8) {
Reg::MaskPack(cmpMask2, cmpMask2);
Reg::MaskPack(cmpMask2, cmpMask2);
Reg::MaskPack(mask, mask);
Reg::MaskPack(mask, mask);
}
Reg::Select(dstVreg, brcOneReg, brcZeroReg, cmpMask2);
Reg::StoreAlign(dst + i * oneRepElm, dstVreg, mask);
}
}
template <auto func, typename T, typename U, typename S, int8_t scalarTensorIndex>
__aicore__ inline void LogicalTemplateBothTensorCompute(
const LocalTensor<T>& dst, const U& src0, const S& src1, const uint32_t count)
{
using RegT = Reg::RegTensor<T>;
using ActualU = typename U::PrimType;
using ActualS = typename S::PrimType;
static_assert(
SupportType<
ActualU, bool, uint8_t, int8_t, half, bfloat16_t, uint16_t, int16_t, float, uint32_t, int32_t, uint64_t,
int64_t>(),
"only support bool/uint8_t/int8_t/half/bfloat16_t/"
"uint16_t/int16_t/float/uint32_t/int32_t/uint64_t/int64_t data type on current device!");
static_assert(Std::is_same_v<ActualU, ActualS>, "The dataType ActualU and ActualS should be the same");
static_assert((scalarTensorIndex == 0 || scalarTensorIndex == 1), "scalarTensorIndex out of range");
if constexpr (sizeof(ActualU) == 8) {
using RegU = Reg::RegTensor<ActualU, Reg::RegTraitNumTwo>;
constexpr uint32_t oneRepElm =
static_cast<uint32_t>(GetVecLen() / sizeof(ActualU) * LOGICAL_TEMPLATE_B64_REPEAT_STRIDE);
uint16_t repeatTime = static_cast<uint16_t>(CeilDivision(count, oneRepElm));
if constexpr (scalarTensorIndex == 0) {
LogicalTemplateBothTensorVF<func, T, ActualU, RegT, RegU, Reg::RegTraitNumTwo>(
(__ubuf__ T*)dst.GetPhyAddr(), (__ubuf__ ActualU*)src1.GetPhyAddr(),
(__ubuf__ ActualU*)src0.GetPhyAddr(), repeatTime, count, oneRepElm);
} else {
LogicalTemplateBothTensorVF<func, T, ActualU, RegT, RegU, Reg::RegTraitNumTwo>(
(__ubuf__ T*)dst.GetPhyAddr(), (__ubuf__ ActualU*)src0.GetPhyAddr(),
(__ubuf__ ActualU*)src1.GetPhyAddr(), repeatTime, count, oneRepElm);
}
} else {
constexpr uint32_t oneRepElm = static_cast<uint32_t>(GetVecLen() / sizeof(ActualU));
uint16_t repeatTime = static_cast<uint16_t>(CeilDivision(count, oneRepElm));
if constexpr (Std::is_same_v<ActualU, bool>) {
using RegU = Reg::RegTensor<uint8_t>;
if constexpr (scalarTensorIndex == 0) {
LogicalTemplateBothTensorVF<func, T, uint8_t, RegT, RegU>(
(__ubuf__ T*)dst.GetPhyAddr(), (__ubuf__ uint8_t*)src1.GetPhyAddr(),
(__ubuf__ uint8_t*)src0.GetPhyAddr(), repeatTime, count, oneRepElm);
} else {
LogicalTemplateBothTensorVF<func, T, uint8_t, RegT, RegU>(
(__ubuf__ T*)dst.GetPhyAddr(), (__ubuf__ uint8_t*)src0.GetPhyAddr(),
(__ubuf__ uint8_t*)src1.GetPhyAddr(), repeatTime, count, oneRepElm);
}
} else {
using RegU = Reg::RegTensor<ActualU>;
if constexpr (scalarTensorIndex == 0) {
LogicalTemplateBothTensorVF<func, T, ActualU, RegT, RegU>(
(__ubuf__ T*)dst.GetPhyAddr(), (__ubuf__ ActualU*)src1.GetPhyAddr(),
(__ubuf__ ActualU*)src0.GetPhyAddr(), repeatTime, count, oneRepElm);
} else {
LogicalTemplateBothTensorVF<func, T, ActualU, RegT, RegU>(
(__ubuf__ T*)dst.GetPhyAddr(), (__ubuf__ ActualU*)src0.GetPhyAddr(),
(__ubuf__ ActualU*)src1.GetPhyAddr(), repeatTime, count, oneRepElm);
}
}
}
}
template <auto func, typename T, typename U, typename S>
__aicore__ inline void LogicalTemplateTensorScalarCompute(
const LocalTensor<T>& dst, const U& src0, const S& src1, const uint32_t count)
{
using RegT = Reg::RegTensor<T>;
using ActualU = typename U::PrimType;
static_assert(
SupportType<
ActualU, bool, uint8_t, int8_t, half, bfloat16_t, uint16_t, int16_t, float, uint32_t, int32_t, uint64_t,
int64_t>(),
"only support bool/uint8_t/int8_t/half/bfloat16_t/"
"uint16_t/int16_t/float/uint32_t/int32_t/uint64_t/int64_t data type on current device!");
static_assert(Std::is_same_v<ActualU, S>, "The dataType ActualU and S should be the same");
if constexpr (sizeof(ActualU) == 8) {
using RegU = Reg::RegTensor<ActualU, Reg::RegTraitNumTwo>;
constexpr uint32_t oneRepElm =
static_cast<uint32_t>(GetVecLen() / sizeof(ActualU) * LOGICAL_TEMPLATE_B64_REPEAT_STRIDE);
uint16_t repeatTime = static_cast<uint16_t>(CeilDivision(count, oneRepElm));
LogicalTemplateSingleScalarVF<func, T, ActualU, RegT, RegU, Reg::RegTraitNumTwo>(
(__ubuf__ T*)dst.GetPhyAddr(), (__ubuf__ ActualU*)src0.GetPhyAddr(), src1, repeatTime, count, oneRepElm);
} else {
constexpr uint32_t oneRepElm = static_cast<uint32_t>(GetVecLen() / sizeof(ActualU));
uint16_t repeatTime = static_cast<uint16_t>(CeilDivision(count, oneRepElm));
if constexpr (Std::is_same_v<ActualU, bool>) {
using RegU = Reg::RegTensor<uint8_t>;
LogicalTemplateSingleScalarVF<func, T, uint8_t, RegT, RegU>(
(__ubuf__ T*)dst.GetPhyAddr(), (__ubuf__ uint8_t*)src0.GetPhyAddr(), static_cast<uint8_t>(src1),
repeatTime, count, oneRepElm);
} else {
using RegU = Reg::RegTensor<ActualU>;
LogicalTemplateSingleScalarVF<func, T, ActualU, RegT, RegU>(
(__ubuf__ T*)dst.GetPhyAddr(), (__ubuf__ ActualU*)src0.GetPhyAddr(), src1, repeatTime, count,
oneRepElm);
}
}
}
template <auto func, typename T, typename U, typename S>
__aicore__ inline void LogicalTemplateScalarTensorCompute(
const LocalTensor<T>& dst, const U& src0, const S& src1, const uint32_t count)
{
using RegT = Reg::RegTensor<T>;
using ActualS = typename S::PrimType;
static_assert(
SupportType<
U, bool, uint8_t, int8_t, half, bfloat16_t, uint16_t, int16_t, float, uint32_t, int32_t, uint64_t,
int64_t>(),
"only support bool/uint8_t/int8_t/half/bfloat16_t/"
"uint16_t/int16_t/float/uint32_t/int32_t/uint64_t/int64_t data type on current device!");
static_assert(Std::is_same_v<ActualS, U>, "The dataType S and ActualU should be the same");
if constexpr (sizeof(U) == 8) {
using RegU = Reg::RegTensor<U, Reg::RegTraitNumTwo>;
constexpr uint32_t oneRepElm =
static_cast<uint32_t>(GetVecLen() / sizeof(U) * LOGICAL_TEMPLATE_B64_REPEAT_STRIDE);
uint16_t repeatTime = static_cast<uint16_t>(CeilDivision(count, oneRepElm));
LogicalTemplateSingleScalarVF<func, T, U, RegT, RegU, Reg::RegTraitNumTwo>(
(__ubuf__ T*)dst.GetPhyAddr(), (__ubuf__ U*)src1.GetPhyAddr(), src0, repeatTime, count, oneRepElm);
} else {
constexpr uint32_t oneRepElm = static_cast<uint32_t>(GetVecLen() / sizeof(U));
uint16_t repeatTime = static_cast<uint16_t>(CeilDivision(count, oneRepElm));
if constexpr (Std::is_same_v<U, bool>) {
using RegU = Reg::RegTensor<uint8_t>;
LogicalTemplateSingleScalarVF<func, T, uint8_t, RegT, RegU>(
(__ubuf__ T*)dst.GetPhyAddr(), (__ubuf__ uint8_t*)src1.GetPhyAddr(), static_cast<uint8_t>(src0),
repeatTime, count, oneRepElm);
} else {
using RegU = Reg::RegTensor<U>;
LogicalTemplateSingleScalarVF<func, T, U, RegT, RegU>(
(__ubuf__ T*)dst.GetPhyAddr(), (__ubuf__ U*)src1.GetPhyAddr(), src0, repeatTime, count, oneRepElm);
}
}
}
template <auto func, typename T, typename U, typename S, int8_t scalarTensorIndex>
__aicore__ inline void LogicalTemplateScalarImpl(
const LocalTensor<T>& dst, const U& src0, const S& src1, const uint32_t count)
{
static_assert(SupportType<T, bool>(), "only support bool data type on current device!");
static_assert(!TypeUtils::IsInnerDefaultType<U, S>(), "One of src0 and src1 should be Tensor");
using RegT = Reg::RegTensor<T>;
if constexpr (Std::is_same_v<S, U>) {
LogicalTemplateBothTensorCompute<func, T, U, S, scalarTensorIndex>(dst, src0, src1, count);
} else if constexpr (TypeUtils::IsLocalTensorType<U>() && TypeUtils::IsInnerDefaultType<S>()) {
LogicalTemplateTensorScalarCompute<func, T, U, S>(dst, src0, src1, count);
} else if constexpr (TypeUtils::IsLocalTensorType<S>() && TypeUtils::IsInnerDefaultType<U>()) {
LogicalTemplateScalarTensorCompute<func, T, U, S>(dst, src0, src1, count);
}
}
}
#endif
#endif
#if defined(__UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_MATH_LOGICAL_TEMPLATE_LOGICAL_TEMPLATE_H__)
#undef __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#undef __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_MATH_LOGICAL_TEMPLATE_LOGICAL_TEMPLATE_H__
#endif