* 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.
*/
* \file asc_aicore_dump_impl.h
* \brief
*/
#ifndef IMPL_UTILS_DEBUG_ASC_AICORE_DUMP_IMPL_H
#define IMPL_UTILS_DEBUG_ASC_AICORE_DUMP_IMPL_H
#ifndef ASCENDC_CPU_DEBUG
#include "impl/utils/sys_macros.h"
#include "impl/utils/common_types.h"
#include "impl/utils/debug/asc_debug_types.h"
#include "impl/utils/debug/asc_debug_utils.h"
#if __NPU_ARCH__ == 2201
#include "impl/utils/debug/npu_arch_2201/asc_aicore_dump_utils.h"
#elif __NPU_ARCH__ == 3510
#include "impl/utils/debug/npu_arch_3510/asc_aicore_dump_utils.h"
#elif __NPU_ARCH__ == 5102
#include "impl/utils/debug/npu_arch_5102/asc_aicore_dump_utils.h"
#endif
namespace __asc_aicore {
template <AscendC::Hardware hardware, typename T, typename U>
__aicore__ inline void set_dump_tlv_info(U src, __gm__ DumpTensorTlv* dumpTlv, uint32_t alignDumpLen, uint32_t desc, uint32_t dump_size)
{
dumpTlv->type = static_cast<uint32_t>(DumpType::DUMP_TENSOR);
dumpTlv->length = sizeof(DumpTensorTlv) - sizeof(uint32_t[2]) + alignDumpLen;
dumpTlv->tensorAddr = static_cast<uint32_t>(reinterpret_cast<uintptr_t>(src));
dumpTlv->dataType = static_cast<uint32_t>(get_dump_datatype<T>());
dumpTlv->desc = desc;
dumpTlv->blockIdx = static_cast<uint32_t>(asc_debug_get_block_idx());
dumpTlv->bufferId = static_cast<uint32_t>(0U);
dumpTlv->position = static_cast<uint16_t>(hardware);
dumpTlv->dim = static_cast<uint32_t>(0U);
for (uint32_t i = 0; i < K_MAX_SHAPE_DIM; ++i) {
dumpTlv->shape[i] = static_cast<uint32_t>(0U);
}
dumpTlv->resv1 = static_cast<uint32_t>(0U);
dumpTlv->dumpSize = dump_size * sizeof(T);
asc_entire_dcci(reinterpret_cast<__gm__ uint64_t*>(dumpTlv));
}
template <AscendC::Hardware hardware, typename T, typename U>
__aicore__ inline uint32_t set_dump_tlv_data(U src, __gm__ DumpTensorTlv* dumpTlv, uint32_t alignDumpLen, uint32_t dump_size)
{
__gm__ T* dumpDstAddr = reinterpret_cast<__gm__ T*>(dumpTlv + 1);
if (dumpDstAddr == nullptr) {
return 1;
}
if (hardware == AscendC::Hardware::GM && src == nullptr) {
return 1;
}
sync_all();
uint32_t ret = 0;
uint32_t dumpLen = 0;
if constexpr (hardware == AscendC::Hardware::GM) {
dumpLen = dump_size * sizeof(T);
ret = mem_copy_gm_to_gm(reinterpret_cast<__gm__ uint8_t*>(dumpDstAddr), reinterpret_cast<__gm__ const uint8_t*>(src), dumpLen);
} else if constexpr (hardware == AscendC::Hardware::UB) {
dumpLen = alignDumpLen / ASC_ONE_DATABLOCK_SIZE;
ret = mem_copy_ub_to_gm_impl(dumpDstAddr, src, static_cast<uint16_t>(dumpLen));
} else if constexpr (hardware == AscendC::Hardware::L1) {
ret = mem_copy_l1buf_to_gm_impl(dumpDstAddr, src, alignDumpLen);
} else if constexpr (hardware == AscendC::Hardware::L0A) {
ret = mem_copy_abuf_to_gm_impl(dumpDstAddr, src, alignDumpLen);
} else if constexpr (hardware == AscendC::Hardware::L0B) {
ret = mem_copy_bbuf_to_gm_impl(dumpDstAddr, src, alignDumpLen);
} else if constexpr (hardware == AscendC::Hardware::L0C) {
ret = mem_copy_cbuf_to_gm_impl(dumpDstAddr, src, alignDumpLen);
}
sync_all();
return ret;
}
template <AscendC::Hardware hardware, typename T, typename U>
__aicore__ inline void asc_dump_impl(U src, uint32_t desc, uint32_t dump_size)
{
__gm__ DebugBlockHeadInfo* blockInfo = get_block_info();
if (dump_size <= 0 || blockInfo == nullptr) {
return;
}
constexpr uint16_t dataBlockSize = 32;
uint32_t alignDumpLen = align_up(dump_size * sizeof(T), dataBlockSize);
uint32_t tlvLen = sizeof(DumpTensorTlv) + alignDumpLen;
if (!check_ringbuf_space(blockInfo, tlvLen)) {
return;
}
__gm__ DumpTensorTlv* dumpTlv = reinterpret_cast<__gm__ DumpTensorTlv*>(get_ringbuf_tlv_addr(blockInfo));
set_dump_tlv_info<hardware, T>(src, dumpTlv, alignDumpLen, desc, dump_size);
if (set_dump_tlv_data<hardware, T>(src, dumpTlv, alignDumpLen, dump_size) != 0) {
return;
}
__gm__ DebugBlockWriteInfo* writeInfo = get_block_write_info(blockInfo);
update_write_info(writeInfo, tlvLen);
}
template <AscendC::Hardware hardware, typename T, typename U>
__aicore__ inline void asc_dump_local_impl(U input, uint32_t desc, uint32_t dump_size)
{
uint64_t ctrlValue = get_ctrl();
set_atomic_none();
enable_asc_diagnostics();
if (g_sysPrintFifoSpace != nullptr) {
asc_dump_impl<hardware, T>(input, desc, dump_size);
}
set_ctrl(ctrlValue);
}
template<typename T>
__aicore__ inline void asc_dump_gm(__gm__ T* input, uint32_t desc, uint32_t dump_size)
{
asc_dump_local_impl<AscendC::Hardware::GM, T>(input, desc, dump_size);
}
template<typename T>
__aicore__ inline void asc_dump_ubuf(__ubuf__ T* input, uint32_t desc, uint32_t dump_size)
{
asc_dump_local_impl<AscendC::Hardware::UB, T>(input, desc, dump_size);
}
template<typename T>
__aicore__ inline void asc_dump_abuf(__ca__ T* input, uint32_t desc, uint32_t dump_size)
{
asc_dump_local_impl<AscendC::Hardware::L0A, T>(input, desc, dump_size);
}
template<typename T>
__aicore__ inline void asc_dump_bbuf(__cb__ T* input, uint32_t desc, uint32_t dump_size)
{
asc_dump_local_impl<AscendC::Hardware::L0B, T>(input, desc, dump_size);
}
template<typename T>
__aicore__ inline void asc_dump_cbuf(__cc__ T* input, uint32_t desc, uint32_t dump_size)
{
asc_dump_local_impl<AscendC::Hardware::L0C, T>(input, desc, dump_size);
}
template<typename T>
__aicore__ inline void asc_dump_l1buf(__cbuf__ T* input, uint32_t desc, uint32_t dump_size)
{
asc_dump_local_impl<AscendC::Hardware::L1, T>(input, desc, dump_size);
}
template<typename T>
__aicore__ inline void asc_dump(__gm__ T* input, uint32_t desc, uint32_t dump_size)
{
asc_dump_gm(input, desc, dump_size);
}
template<typename T>
__aicore__ inline void asc_dump(__ubuf__ T* input, uint32_t desc, uint32_t dump_size)
{
asc_dump_ubuf(input, desc, dump_size);
}
template<typename T>
__aicore__ inline void asc_dump(__cc__ T* input, uint32_t desc, uint32_t dump_size)
{
asc_dump_cbuf(input, desc, dump_size);
}
template<typename T>
__aicore__ inline void asc_dump(__cbuf__ T* input, uint32_t desc, uint32_t dump_size)
{
asc_dump_l1buf(input, desc, dump_size);
}
}
#else
#include "kernel_log.h"
namespace __asc_aicore {
template<typename T>
__aicore__ inline void asc_dump_gm(__gm__ T* input, uint32_t desc, uint32_t dump_size) {
ASCENDC_ASSERT((false), "asc_dump_gm is not supported in cpu mode.");
}
template<typename T>
__aicore__ inline void asc_dump_ubuf(__ubuf__ T* input, uint32_t desc, uint32_t dump_size) {
ASCENDC_ASSERT((false), "asc_dump_ubuf is not supported in cpu mode.");
}
template<typename T>
__aicore__ inline void asc_dump_cbuf(__cc__ T* input, uint32_t desc, uint32_t dump_size) {
ASCENDC_ASSERT((false), "asc_dump_cbuf is not supported in cpu mode.");
}
template<typename T>
__aicore__ inline void asc_dump_l1buf(__cbuf__ T* input, uint32_t desc, uint32_t dump_size) {
ASCENDC_ASSERT((false), "asc_dump_l1buf is not supported in cpu mode.");
}
template<typename T>
__aicore__ inline void asc_dump(T* input, uint32_t desc, uint32_t dump_size) {
ASCENDC_ASSERT((false), "asc_dump is not supported in cpu mode.");
}
}
using namespace __asc_aicore;
#endif
#if defined(__UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_ASC_AICORE_DUMP_IMPL__)
#undef __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#undef __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_ASC_AICORE_DUMP_IMPL__
#endif
#endif
