* Copyright (c) 2026 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.
*/
* NOTE: Portions of this code were AI-generated and have been
* technically reviewed for functional accuracy and security
*/
* \file inv_grad_tiling.cpp
* \brief InvGrad tiling implementation (arch35)
*
* Tiling strategy:
* 1. Multi-core: divide total elements evenly across AI Cores
* 2. UB: divide per-core elements into UB-sized chunks
* 3. Buffer layout (unified for all dtypes):
* yQueue(1) = ubFactor * sizeof(T)
* dyQueue(1) = ubFactor * sizeof(T)
* outQueue(1)= ubFactor * sizeof(T)
* tmpBuf1 = ubFactor * sizeof(float) (fp32 intermediate yy / yFloat)
* tmpBuf2 = ubFactor * sizeof(float) (fp32 intermediate dyFloat)
*
* Unified formula for all dtypes:
* bytesPerElem = 3 * sizeof(T) + 2 * sizeof(float)
* ubFactor = FloorAlign(ubSize / bytesPerElem, ubBlockSize)
*
* Shape/dtype equality between y and dy is enforced here. InferShape only
* copies y.shape to dx.shape.
*/
#include "register/op_def_registry.h"
#include "op_common/log/log.h"
#include "op_common/op_host/util/math_util.h"
#include "op_common/op_host/util/platform_util.h"
#include "../op_kernel/inv_grad_tiling_data.h"
#include "../op_kernel/inv_grad_tiling_key.h"
namespace optiling {
using Ops::Base::CeilDiv;
using Ops::Base::FloorAlign;
constexpr uint32_t WS_SYS_SIZE = 0U;
static const gert::Shape g_vec_1_shape = {1};
static inline const gert::Shape EnsureNotScalar(const gert::Shape& in_shape)
{
if (in_shape.GetDimNum() == 0) {
return g_vec_1_shape;
}
return in_shape;
}
static ge::graphStatus GetPlatformInfo(gert::TilingContext* context, uint64_t& ubSize, int64_t& coreNum)
{
fe::PlatFormInfos* platformInfoPtr = context->GetPlatformInfo();
OP_CHECK_NULL_WITH_CONTEXT(context, platformInfoPtr);
auto ascendcPlatform = platform_ascendc::PlatformAscendC(platformInfoPtr);
coreNum = ascendcPlatform.GetCoreNumAiv();
OP_CHECK_IF(coreNum == 0, OP_LOGE(context, "coreNum is 0"), return ge::GRAPH_FAILED);
ascendcPlatform.GetCoreMemSize(platform_ascendc::CoreMemType::UB, ubSize);
OP_CHECK_IF(ubSize == 0, OP_LOGE(context, "ubSize is 0"), return ge::GRAPH_FAILED);
return ge::GRAPH_SUCCESS;
}
static ge::graphStatus GetShapeInfo(gert::TilingContext* context, int64_t& totalElements,
ge::DataType& dataType)
{
auto yStorage = context->GetInputShape(0);
OP_CHECK_NULL_WITH_CONTEXT(context, yStorage);
auto yShape = EnsureNotScalar(yStorage->GetStorageShape());
totalElements = yShape.GetShapeSize();
auto yDesc = context->GetInputDesc(0);
OP_CHECK_NULL_WITH_CONTEXT(context, yDesc);
dataType = yDesc->GetDataType();
const std::set<ge::DataType> supportedDtype = {
ge::DT_FLOAT, ge::DT_FLOAT16, ge::DT_BF16
};
if (supportedDtype.count(dataType) == 0) {
OP_LOGE(context, "InvGrad: unsupported y dtype %d", static_cast<int>(dataType));
return ge::GRAPH_FAILED;
}
auto dyStorage = context->GetInputShape(1);
OP_CHECK_NULL_WITH_CONTEXT(context, dyStorage);
auto dyShape = EnsureNotScalar(dyStorage->GetStorageShape());
auto dyDesc = context->GetInputDesc(1);
OP_CHECK_NULL_WITH_CONTEXT(context, dyDesc);
ge::DataType dyDataType = dyDesc->GetDataType();
if (dyDataType != dataType) {
OP_LOGE(context, "InvGrad: y dtype %d != dy dtype %d",
static_cast<int>(dataType), static_cast<int>(dyDataType));
return ge::GRAPH_FAILED;
}
if (dyShape.GetDimNum() != yShape.GetDimNum()) {
OP_LOGE(context, "InvGrad: y.dimNum %zu != dy.dimNum %zu",
yShape.GetDimNum(), dyShape.GetDimNum());
return ge::GRAPH_FAILED;
}
for (size_t i = 0; i < yShape.GetDimNum(); ++i) {
if (yShape.GetDim(i) != dyShape.GetDim(i)) {
OP_LOGE(context, "InvGrad: y.shape[%zu] %ld != dy.shape[%zu] %ld",
i, yShape.GetDim(i), i, dyShape.GetDim(i));
return ge::GRAPH_FAILED;
}
}
return ge::GRAPH_SUCCESS;
}
static ge::graphStatus GetWorkspaceSize(gert::TilingContext* context)
{
size_t* currentWorkspace = context->GetWorkspaceSizes(1);
OP_CHECK_NULL_WITH_CONTEXT(context, currentWorkspace);
currentWorkspace[0] = WS_SYS_SIZE;
return ge::GRAPH_SUCCESS;
}
static ge::graphStatus InvGradTilingFunc(gert::TilingContext* context)
{
uint64_t ubSize = 0;
int64_t coreNum = 0;
OP_CHECK_IF(
GetPlatformInfo(context, ubSize, coreNum) != ge::GRAPH_SUCCESS,
OP_LOGE(context, "GetPlatformInfo error"), return ge::GRAPH_FAILED);
int64_t totalElements = 0;
ge::DataType dataType = ge::DT_FLOAT;
OP_CHECK_IF(
GetShapeInfo(context, totalElements, dataType) != ge::GRAPH_SUCCESS,
OP_LOGE(context, "GetShapeInfo error"), return ge::GRAPH_FAILED);
OP_CHECK_IF(
GetWorkspaceSize(context) != ge::GRAPH_SUCCESS,
OP_LOGE(context, "GetWorkspaceSize error"), return ge::GRAPH_FAILED);
InvGradTilingData* tiling = context->GetTilingData<InvGradTilingData>();
OP_CHECK_NULL_WITH_CONTEXT(context, tiling);
OP_CHECK_IF(
memset_s(tiling, sizeof(InvGradTilingData), 0, sizeof(InvGradTilingData)) != EOK,
OP_LOGE(context, "set tiling data error"), return ge::GRAPH_FAILED);
int64_t typeSize = 4;
switch (dataType) {
case ge::DT_FLOAT:
typeSize = 4;
break;
case ge::DT_FLOAT16:
typeSize = 2;
break;
case ge::DT_BF16:
typeSize = 2;
break;
default:
OP_LOGE(context, "InvGrad: unexpected dtype %d", static_cast<int>(dataType));
return ge::GRAPH_FAILED;
}
int64_t ubBlockSize = 32 / typeSize;
if (totalElements == 0) {
tiling->totalElements = 0;
tiling->blockFactor = 0;
tiling->ubFactor = 0;
context->SetBlockDim(1);
uint32_t dTypeY = static_cast<uint32_t>(dataType);
ASCENDC_TPL_SEL_PARAM(context, dTypeY);
return ge::GRAPH_SUCCESS;
}
int64_t blockFactor = CeilDiv(totalElements, coreNum);
blockFactor = ((blockFactor + ubBlockSize - 1) / ubBlockSize) * ubBlockSize;
int64_t usedCoreNum = CeilDiv(totalElements, blockFactor);
int64_t bytesPerElem = 3 * typeSize + 2 * static_cast<int64_t>(sizeof(float));
int64_t maxElementsPerCopy = static_cast<int64_t>(UINT16_MAX) / typeSize;
int64_t ubFactor = FloorAlign(
static_cast<int64_t>(ubSize) / bytesPerElem,
ubBlockSize);
if (ubFactor > maxElementsPerCopy) {
ubFactor = FloorAlign(maxElementsPerCopy, ubBlockSize);
}
OP_CHECK_IF(ubFactor <= 0,
OP_LOGE(context, "InvGrad: ubFactor is %ld, UB too small", ubFactor),
return ge::GRAPH_FAILED);
tiling->totalElements = totalElements;
tiling->blockFactor = blockFactor;
tiling->ubFactor = ubFactor;
context->SetBlockDim(usedCoreNum);
uint32_t dTypeY = static_cast<uint32_t>(dataType);
ASCENDC_TPL_SEL_PARAM(context, dTypeY);
return ge::GRAPH_SUCCESS;
}
static ge::graphStatus TilingParseForInvGrad([[maybe_unused]] gert::TilingParseContext* context)
{
return ge::GRAPH_SUCCESS;
}
struct InvGradCompileInfo {};
IMPL_OP_OPTILING(InvGrad).Tiling(InvGradTilingFunc).TilingParse<InvGradCompileInfo>(TilingParseForInvGrad);
}