Star1296
1410
代码
代码
Issues108
Pull Requests213
Actions
项目讨论
Wiki
分析
Star1296
1410
已合并
修复STFT算子int32/uint32溢出风险 #3133
east_yang创建于 6月4日
修复STFT算子int32/uint32溢出风险 #3133
已合并
east_yang创建于 6月4日
从east_yang/ops-mathmaster合入到cann/ops-mathmaster
共 6 个文件变更+88-85
Mmath/stft/op_host/stft_tiling_align.cpp+3-3
@@ -284,13 +284,13 @@ ge::graphStatus STFTTiling::GetWorkspaceSize()
284 // 每块workspace地址需要512B对齐284 // 每块workspace地址需要512B对齐
285 // 第0块workspace用于存储按照窗口拆分之后的input data285 // 第0块workspace用于存储按照窗口拆分之后的input data
286 size_t windowSplitWorkspaceSize =286 size_t windowSplitWorkspaceSize =
287- batchLoop *287+ (uint64_t)batchLoop *
288- ((aicCoreNum * frameCount * nfft * sizeof(float) + WORKSPACE_ALIGN_SIZE - 1) / WORKSPACE_ALIGN_SIZE) *288+ (((uint64_t)aicCoreNum * frameCount * nfft * sizeof(float) + WORKSPACE_ALIGN_SIZE - 1) / WORKSPACE_ALIGN_SIZE) *
289 WORKSPACE_ALIGN_SIZE;289 WORKSPACE_ALIGN_SIZE;
290 290 
291 // 第一块workspace用于存储input data和plan mm运算之后的结果291 // 第一块workspace用于存储input data和plan mm运算之后的结果
292 size_t matmulWorkspaceSize =292 size_t matmulWorkspaceSize =
293- ((batch * frameCount * matmulM * 2 * sizeof(float) + WORKSPACE_ALIGN_SIZE - 1) / WORKSPACE_ALIGN_SIZE) *293+ (((uint64_t)batch * frameCount * matmulM * 2 * sizeof(float) + WORKSPACE_ALIGN_SIZE - 1) / WORKSPACE_ALIGN_SIZE) *
294 WORKSPACE_ALIGN_SIZE;294 WORKSPACE_ALIGN_SIZE;
295 workspaceSize_ = windowSplitWorkspaceSize + matmulWorkspaceSize + EXTRA_WORKSPACE_SIZE;295 workspaceSize_ = windowSplitWorkspaceSize + matmulWorkspaceSize + EXTRA_WORKSPACE_SIZE;
296 return ge::GRAPH_SUCCESS;296 return ge::GRAPH_SUCCESS;
Mmath/stft/op_host/stft_tiling_generalized.cpp+13-13
@@ -134,22 +134,22 @@ private:
134 134 
135void STFTGeneralizedTiling::GetPlanSplitStrategy()135void STFTGeneralizedTiling::GetPlanSplitStrategy()
136{136{
137- int32_t numBlocks = aivCoreNum;137+ uint32_t numBlocks = aivCoreNum;
138- int32_t oneRowSize = nfftAlign * 4;138+ uint32_t oneRowSize = nfftAlign * 4;
139- int32_t halfUbSize = (ubSize - oneRowSize) / 2;139+ uint32_t halfUbSize = (ubSize - oneRowSize) / 2;
140- int32_t mFactor = CeilDiv(2 * matmulM, numBlocks);140+ uint32_t mFactor = CeilDiv(2 * matmulM, numBlocks);
141- int32_t prevCnt = numBlocks * (mFactor - 1);141+ uint32_t prevCnt = numBlocks * (mFactor - 1);
142- int32_t remainCnt = 2 * matmulM - prevCnt;142+ uint32_t remainCnt = 2 * matmulM - prevCnt;
143- int32_t totalLine = mFactor;143+ uint32_t totalLine = mFactor;
144- int32_t tailBlockIdx = remainCnt;144+ uint32_t tailBlockIdx = remainCnt;
145- int32_t tailLine = mFactor;145+ uint32_t tailLine = mFactor;
146 if (remainCnt < numBlocks) {146 if (remainCnt < numBlocks) {
147 tailLine = mFactor - 1;147 tailLine = mFactor - 1;
148 }148 }
149- int32_t ubMaxLine = halfUbSize / oneRowSize;149+ uint32_t ubMaxLine = halfUbSize / oneRowSize;
150- int32_t numsInOneRepeat = 64;150+ uint32_t numsInOneRepeat = 64;
151- int32_t totalInCol = nfftAlign / numsInOneRepeat;151+ uint32_t totalInCol = nfftAlign / numsInOneRepeat;
152- int32_t tailInCol = nfftAlign % numsInOneRepeat;152+ uint32_t tailInCol = nfftAlign % numsInOneRepeat;
153 153 
154 tilingData.planTilingData.set_totalInCol(totalInCol);154 tilingData.planTilingData.set_totalInCol(totalInCol);
155 tilingData.planTilingData.set_tailInCol(tailInCol);155 tilingData.planTilingData.set_tailInCol(tailInCol);
Mmath/stft/op_kernel/stft.h+27-27
@@ -60,21 +60,21 @@ public:
60 {60 {
61 inTilingData = tilingData;61 inTilingData = tilingData;
62 size_t windowSplitWorkspaceSize =62 size_t windowSplitWorkspaceSize =
63- (((inTilingData->blockNum * inTilingData->frameCount * inTilingData->nfft * sizeof(T) + 511) / 512) * 512) *63+ ((((uint64_t)inTilingData->blockNum * inTilingData->frameCount * inTilingData->nfft * sizeof(T) + 511) / 512) * 512) *
64 inTilingData->aivBatchLoop / sizeof(T);64 inTilingData->aivBatchLoop / sizeof(T);
65 if (g_coreType == AIV) {65 if (g_coreType == AIV) {
66- inputGm.SetGlobalBuffer((__gm__ T*)x, inTilingData->batch * inTilingData->inputSize);66+ inputGm.SetGlobalBuffer((__gm__ T*)x, (uint64_t)inTilingData->batch * inTilingData->inputSize);
67 windowSplitWorkspaceGm.SetGlobalBuffer(67 windowSplitWorkspaceGm.SetGlobalBuffer(
68 (__gm__ T*)workspace,68 (__gm__ T*)workspace,
69- inTilingData->blockNum * inTilingData->aivBatchLoop * inTilingData->frameCount * inTilingData->nfft);69+ (uint64_t)inTilingData->blockNum * inTilingData->aivBatchLoop * inTilingData->frameCount * inTilingData->nfft);
70 outputGm.SetGlobalBuffer(70 outputGm.SetGlobalBuffer(
71- (__gm__ T*)y, inTilingData->matmulM * inTilingData->frameCount * inTilingData->batch * DOUBLE_BUFFER);71+ (__gm__ T*)y, (uint64_t)inTilingData->matmulM * inTilingData->frameCount * inTilingData->batch * DOUBLE_BUFFER);
72 gmReal.SetGlobalBuffer(72 gmReal.SetGlobalBuffer(
73 reinterpret_cast<__gm__ T*>(workspace) + windowSplitWorkspaceSize,73 reinterpret_cast<__gm__ T*>(workspace) + windowSplitWorkspaceSize,
74- inTilingData->batch * inTilingData->matmulM * inTilingData->frameCount * DOUBLE_BUFFER);74+ (uint64_t)inTilingData->batch * inTilingData->matmulM * inTilingData->frameCount * DOUBLE_BUFFER);
75 gmImag.SetGlobalBuffer(75 gmImag.SetGlobalBuffer(
76 reinterpret_cast<__gm__ T*>(workspace) + windowSplitWorkspaceSize,76 reinterpret_cast<__gm__ T*>(workspace) + windowSplitWorkspaceSize,
77- inTilingData->batch * inTilingData->matmulM * inTilingData->frameCount * DOUBLE_BUFFER);77+ (uint64_t)inTilingData->batch * inTilingData->matmulM * inTilingData->frameCount * DOUBLE_BUFFER);
78 pipe.InitBuffer(78 pipe.InitBuffer(
79 inQueueInput, bufferNum,79 inQueueInput, bufferNum,
80 (inTilingData->blkFrame * inTilingData->hop + (inTilingData->nfft - inTilingData->hop)) * sizeof(T));80 (inTilingData->blkFrame * inTilingData->hop + (inTilingData->nfft - inTilingData->hop)) * sizeof(T));
@@ -84,13 +84,13 @@ public:
84 if (g_coreType == AIC) {84 if (g_coreType == AIC) {
85 auto blockIdx = GetBlockIdx();85 auto blockIdx = GetBlockIdx();
86 a1Global.SetGlobalBuffer(86 a1Global.SetGlobalBuffer(
87- reinterpret_cast<__gm__ T*>(window), inTilingData->matmulM * inTilingData->nfft * DOUBLE_BUFFER);87+ reinterpret_cast<__gm__ T*>(window), (uint64_t)inTilingData->matmulM * inTilingData->nfft * DOUBLE_BUFFER);
88 bGlobal.SetGlobalBuffer(88 bGlobal.SetGlobalBuffer(
89 reinterpret_cast<__gm__ T*>(workspace),89 reinterpret_cast<__gm__ T*>(workspace),
90- inTilingData->blockNum * inTilingData->nfft * inTilingData->frameCount * inTilingData->aivBatchLoop);90+ (uint64_t)inTilingData->blockNum * inTilingData->nfft * inTilingData->frameCount * inTilingData->aivBatchLoop);
91 matMulWorkspaceGm.SetGlobalBuffer(91 matMulWorkspaceGm.SetGlobalBuffer(
92 reinterpret_cast<__gm__ T*>(workspace) + windowSplitWorkspaceSize,92 reinterpret_cast<__gm__ T*>(workspace) + windowSplitWorkspaceSize,
93- inTilingData->matmulM * inTilingData->frameCount * inTilingData->batch * DOUBLE_BUFFER);93+ (uint64_t)inTilingData->matmulM * inTilingData->frameCount * inTilingData->batch * DOUBLE_BUFFER);
94 94 
95 curCoreM_ = inTilingData->aicTotalLen;95 curCoreM_ = inTilingData->aicTotalLen;
96 curCoreN_ = inTilingData->mmTilingData.N;96 curCoreN_ = inTilingData->mmTilingData.N;
@@ -180,15 +180,15 @@ public:
180 CreateGatherMask(maskTemp, maskCol, REAL_IMAG_COLS, 0, REAL_IMAG * gatherSizePerRepeat);180 CreateGatherMask(maskTemp, maskCol, REAL_IMAG_COLS, 0, REAL_IMAG * gatherSizePerRepeat);
181 181 
182 // 生成mask等差数列182 // 生成mask等差数列
183- int32_t offsetBase = 0;183+ int64_t offsetBase = 0;
184 if (unlikely(blockIdx % windowLoop < aivMTailIdx)) {184 if (unlikely(blockIdx % windowLoop < aivMTailIdx)) {
185- offsetBase = ((blockIdx % windowLoop) / C_V_DOUBLE) * totalMLen * frameCount +185+ offsetBase = (int64_t)((blockIdx % windowLoop) / C_V_DOUBLE) * totalMLen * frameCount +
186- (blockIdx % C_V_DOUBLE) * (aivTotalEvenMLen * REAL_IMAG) * frameCount;186+ (int64_t)(blockIdx % C_V_DOUBLE) * (aivTotalEvenMLen * REAL_IMAG) * frameCount;
187 } else {187 } else {
188- offsetBase = (aivMTailIdx / C_V_DOUBLE) * totalMLen * frameCount +188+ offsetBase = (int64_t)(aivMTailIdx / C_V_DOUBLE) * totalMLen * frameCount +
189- (((blockIdx % windowLoop) - aivMTailIdx) % C_V_DOUBLE) * (aivTailEvenMLen * REAL_IMAG) *189+ (int64_t)(((blockIdx % windowLoop) - aivMTailIdx) % C_V_DOUBLE) * (aivTailEvenMLen * REAL_IMAG) *
190 frameCount +190 frameCount +
191- ((blockIdx % windowLoop) - aivMTailIdx) / C_V_DOUBLE * aicTailLen * frameCount;191+ (int64_t)(((blockIdx % windowLoop) - aivMTailIdx) / C_V_DOUBLE) * aicTailLen * frameCount;
192 }192 }
193 193 
194 int repeats = (N * sizeof(T) + gatherSizePerRepeat - 1) / gatherSizePerRepeat;194 int repeats = (N * sizeof(T) + gatherSizePerRepeat - 1) / gatherSizePerRepeat;
@@ -203,9 +203,9 @@ public:
203 AscendC::WaitEvent(flag_id_fix);203 AscendC::WaitEvent(flag_id_fix);
204 204 
205 int32_t curBatch = (blockIdx / windowLoop) * aicBatchLoop + i;205 int32_t curBatch = (blockIdx / windowLoop) * aicBatchLoop + i;
206- int32_t gmRealOffset = curBatch * matmulM * frameCount * DOUBLE_BUFFER;206+ int64_t gmRealOffset = (int64_t)curBatch * matmulM * frameCount * DOUBLE_BUFFER;
207- int32_t gmImagOffset = gmRealOffset + frameCount;207+ int64_t gmImagOffset = gmRealOffset + frameCount;
208- int32_t outputOffset = curBatch * matmulM * frameCount * DOUBLE_BUFFER;208+ int64_t outputOffset = (int64_t)curBatch * matmulM * frameCount * DOUBLE_BUFFER;
209 209 
210 gmRealOffset = gmRealOffset + offsetBase;210 gmRealOffset = gmRealOffset + offsetBase;
211 gmImagOffset = gmImagOffset + offsetBase;211 gmImagOffset = gmImagOffset + offsetBase;
@@ -266,15 +266,15 @@ public:
266 globalN_ = inTilingData->mmTilingData.N;266 globalN_ = inTilingData->mmTilingData.N;
267 globalK_ = inTilingData->mmTilingData.Ka;267 globalK_ = inTilingData->mmTilingData.Ka;
268 268 
269- int32_t a1GlobalOffset = 0;269+ int64_t a1GlobalOffset = 0;
270- int32_t outputOffsetBase = 0;270+ int64_t outputOffsetBase = 0;
271 if (unlikely(blockIdx % aicMatmulMCore < aicMTailIdx)) {271 if (unlikely(blockIdx % aicMatmulMCore < aicMTailIdx)) {
272- a1GlobalOffset = innerReminder * totalMLen * nfft;272+ a1GlobalOffset = (int64_t)innerReminder * totalMLen * nfft;
273- outputOffsetBase = innerReminder * totalMLen * frameCount;273+ outputOffsetBase = (int64_t)innerReminder * totalMLen * frameCount;
274 } else {274 } else {
275- a1GlobalOffset = aicMTailIdx * totalMLen * nfft + (innerReminder - aicMTailIdx) * aicTailLen * nfft;275+ a1GlobalOffset = (int64_t)aicMTailIdx * totalMLen * nfft + (int64_t)(innerReminder - aicMTailIdx) * aicTailLen * nfft;
276 outputOffsetBase =276 outputOffsetBase =
277- aicMTailIdx * totalMLen * frameCount + (innerReminder - aicMTailIdx) * aicTailLen * frameCount;277+ (int64_t)aicMTailIdx * totalMLen * frameCount + (int64_t)(innerReminder - aicMTailIdx) * aicTailLen * frameCount;
278 }278 }
279 279 
280 uint64_t flag_id_mte3 = 3;280 uint64_t flag_id_mte3 = 3;
@@ -286,9 +286,9 @@ public:
286 }286 }
287 287 
288 int32_t curBatch = curBatchBase + i;288 int32_t curBatch = curBatchBase + i;
289- int32_t bGlobalOffset =289+ int64_t bGlobalOffset =
290- blockIdx * aicBatchLoop * frameCount * inTilingData->nfft + i * frameCount * inTilingData->nfft;290+ (int64_t)blockIdx * aicBatchLoop * frameCount * inTilingData->nfft + (int64_t)i * frameCount * inTilingData->nfft;
291- int32_t outputOffset = curBatch * matmulM * frameCount * REAL_IMAG + outputOffsetBase;291+ int64_t outputOffset = (int64_t)curBatch * matmulM * frameCount * REAL_IMAG + outputOffsetBase;
292 bGM = bGlobal[bGlobalOffset];292 bGM = bGlobal[bGlobalOffset];
293 aGM = a1Global[a1GlobalOffset];293 aGM = a1Global[a1GlobalOffset];
294 cGM = matMulWorkspaceGm[outputOffset];294 cGM = matMulWorkspaceGm[outputOffset];
Mmath/stft/op_kernel/stft_generalized.h+12-10
@@ -48,8 +48,8 @@ public:
48 return;48 return;
49 }49 }
50 50 
51- uint64_t inputGmSize = tiling->batch * ((tiling->inputSize + tiling->nfft) * sizeof(T) + BLOCK_SIZE - 1) /51+ uint64_t inputGmSize = (uint64_t)tiling->batch * (((uint64_t)(tiling->inputSize + tiling->nfft) * sizeof(T) + BLOCK_SIZE - 1) /
52- BLOCK_SIZE * BLOCK_SIZE / sizeof(T);52+ BLOCK_SIZE * BLOCK_SIZE / sizeof(T));
53 inputGm.SetGlobalBuffer((__gm__ T*)x, inputGmSize);53 inputGm.SetGlobalBuffer((__gm__ T*)x, inputGmSize);
54 54 
55 uint64_t splitWindowGmSize = (uint64_t)tiling->batch * tiling->matmulN * tiling->nfftAlign;55 uint64_t splitWindowGmSize = (uint64_t)tiling->batch * tiling->matmulN * tiling->nfftAlign;
@@ -159,14 +159,16 @@ public:
159 }159 }
160 160 
161 for (int i = 0; i < bFactor; i++) {161 for (int i = 0; i < bFactor; i++) {
162- int64_t inputOffset = (bOffset + i) * (tiling->inputSize + tiling->nfft) + nOffset * tiling->hopLength;162+ int64_t inputOffset = (int64_t)(bOffset + i) * (tiling->inputSize + tiling->nfft) +
163- int64_t splitWindowOffset = ((bOffset + i) * tiling->matmulN + nOffset) * tiling->nfftAlign;163+ (int64_t)nOffset * tiling->hopLength;
164+ int64_t splitWindowOffset = ((int64_t)(bOffset + i) * tiling->matmulN + nOffset) * tiling->nfftAlign;
164 int64_t outputOffset =165 int64_t outputOffset =
165- (((bOffset + i) * tiling->matmulM + mOffset) * tiling->matmulN + nOffset) * IMAG_AND_REAL;166+ (((int64_t)(bOffset + i) * tiling->matmulM + mOffset) * tiling->matmulN + nOffset) * IMAG_AND_REAL;
166- int64_t realOffset = (bOffset + i) * tiling->matmulM * tiling->matmulN + mOffset * tiling->matmulN +167+ int64_t realOffset = (int64_t)(bOffset + i) * tiling->matmulM * tiling->matmulN +
167- nIdx * mFactor * tiling->matmulNCoreFactor;168+ (int64_t)mOffset * tiling->matmulN +
169+ (int64_t)nIdx * mFactor * tiling->matmulNCoreFactor;
168 int64_t imagOffset = realOffset;170 int64_t imagOffset = realOffset;
169- int64_t a1Offset = mOffset * tiling->nfftAlign;171+ int64_t a1Offset = (int64_t)mOffset * tiling->nfftAlign;
170 int64_t a2Offset = a1Offset;172 int64_t a2Offset = a1Offset;
171 173 
172 int64_t planOffset = IMAG_AND_REAL * mOffset * tiling->nfftAlign;174 int64_t planOffset = IMAG_AND_REAL * mOffset * tiling->nfftAlign;
@@ -715,8 +717,8 @@ private:
715 int32_t total = 0;717 int32_t total = 0;
716 while (total < nFactor) {718 while (total < nFactor) {
717 LocalTensor<T> inputLocal = inCopy.template AllocTensor<T>();719 LocalTensor<T> inputLocal = inCopy.template AllocTensor<T>();
718- int32_t inputLeft =720+ int64_t inputLeft =
719- (tiling->batch * (tiling->inputSize + tiling->nfft) - total * tiling->hopLength - inputOffset) *721+ ((int64_t)tiling->batch * (tiling->inputSize + tiling->nfft) - (int64_t)total * tiling->hopLength - inputOffset) *
720 sizeof(T);722 sizeof(T);
721 int32_t copyLength = inputLeft > bufferSize ? bufferSize / sizeof(T) : inputLeft / sizeof(T);723 int32_t copyLength = inputLeft > bufferSize ? bufferSize / sizeof(T) : inputLeft / sizeof(T);
722 DataCopyPadExtParams<T> dataCopyPadExtParams;724 DataCopyPadExtParams<T> dataCopyPadExtParams;
Mmath/stft/op_kernel/stft_generalized_complex.h+28-27
@@ -53,11 +53,11 @@ public:
53 tiling = tilingData;53 tiling = tilingData;
54 54 
55 inputGm.SetGlobalBuffer(55 inputGm.SetGlobalBuffer(
56- (__gm__ T*)x, tiling->batch *56+ (__gm__ T*)x, (uint64_t)tiling->batch *
57- ((tiling->inputSize + tiling->nfft) * COMPLEX_COEFFICIENT * sizeof(T) + BLOCK_SIZE - 1) /57+ (((uint64_t)(tiling->inputSize + tiling->nfft) * COMPLEX_COEFFICIENT * sizeof(T) + BLOCK_SIZE - 1) /
58- BLOCK_SIZE * BLOCK_SIZE / sizeof(T));58+ BLOCK_SIZE * BLOCK_SIZE / sizeof(T)));
59- size_t splitWindowWorkspaceSize = tiling->batch * tiling->matmulN * tiling->nfftAlign;59+ uint64_t splitWindowWorkspaceSize = (uint64_t)tiling->batch * tiling->matmulN * tiling->nfftAlign;
60- size_t splitWindowWorkspaceSizeAlign =60+ uint64_t splitWindowWorkspaceSizeAlign =
61 (((splitWindowWorkspaceSize * sizeof(T) * COMPLEX_COEFFICIENT + WORKSPACE_ALIGN_SIZE - 1) /61 (((splitWindowWorkspaceSize * sizeof(T) * COMPLEX_COEFFICIENT + WORKSPACE_ALIGN_SIZE - 1) /
62 WORKSPACE_ALIGN_SIZE) *62 WORKSPACE_ALIGN_SIZE) *
63 WORKSPACE_ALIGN_SIZE) /63 WORKSPACE_ALIGN_SIZE) /
@@ -66,7 +66,7 @@ public:
66 splitRealWindowGm.SetGlobalBuffer((__gm__ T*)workspace, splitWindowWorkspaceSize);66 splitRealWindowGm.SetGlobalBuffer((__gm__ T*)workspace, splitWindowWorkspaceSize);
67 splitImagWindowGm.SetGlobalBuffer((__gm__ T*)workspace + splitWindowWorkspaceSize, splitWindowWorkspaceSize);67 splitImagWindowGm.SetGlobalBuffer((__gm__ T*)workspace + splitWindowWorkspaceSize, splitWindowWorkspaceSize);
68 68 
69- size_t matmulWorkspaceSize = tiling->batch * tiling->matmulM * tiling->matmulN;69+ uint64_t matmulWorkspaceSize = (uint64_t)tiling->batch * tiling->matmulM * tiling->matmulN;
70 70 
71 aRealGm.SetGlobalBuffer((__gm__ T*)workspace + splitWindowWorkspaceSizeAlign, matmulWorkspaceSize);71 aRealGm.SetGlobalBuffer((__gm__ T*)workspace + splitWindowWorkspaceSizeAlign, matmulWorkspaceSize);
72 aImagGm.SetGlobalBuffer(72 aImagGm.SetGlobalBuffer(
@@ -76,11 +76,11 @@ public:
76 matmulWorkspaceSize);76 matmulWorkspaceSize);
77 bImagGm.SetGlobalBuffer(77 bImagGm.SetGlobalBuffer(
78 (__gm__ T*)workspace + splitWindowWorkspaceSizeAlign + matmulWorkspaceSize * 3, matmulWorkspaceSize);78 (__gm__ T*)workspace + splitWindowWorkspaceSizeAlign + matmulWorkspaceSize * 3, matmulWorkspaceSize);
79- outputGm.SetGlobalBuffer((__gm__ T*)y, tiling->batch * tiling->matmulM * tiling->matmulN * DOUBLE_BUFFER);79+ outputGm.SetGlobalBuffer((__gm__ T*)y, (uint64_t)tiling->batch * tiling->matmulM * tiling->matmulN * DOUBLE_BUFFER);
80- a1Global.SetGlobalBuffer(reinterpret_cast<__gm__ T*>(window), tiling->matmulM * tiling->nfftAlign);80+ a1Global.SetGlobalBuffer(reinterpret_cast<__gm__ T*>(window), (uint64_t)tiling->matmulM * tiling->nfftAlign);
81 a2Global.SetGlobalBuffer(81 a2Global.SetGlobalBuffer(
82- reinterpret_cast<__gm__ T*>(window) + tiling->matmulM * tiling->nfftAlign,82+ reinterpret_cast<__gm__ T*>(window) + (uint64_t)tiling->matmulM * tiling->nfftAlign,
83- tiling->matmulM * tiling->nfftAlign);83+ (uint64_t)tiling->matmulM * tiling->nfftAlign);
84 84 
85 size_t ubAlignBufferSize = (tiling->nFactorUbFormer * COMPLEX_COEFFICIENT + REPEAT_NUM_FOR_FP32 - 1) /85 size_t ubAlignBufferSize = (tiling->nFactorUbFormer * COMPLEX_COEFFICIENT + REPEAT_NUM_FOR_FP32 - 1) /
86 REPEAT_NUM_FOR_FP32 * REPEAT_NUM_FOR_FP32 * sizeof(T);86 REPEAT_NUM_FOR_FP32 * REPEAT_NUM_FOR_FP32 * sizeof(T);
@@ -153,16 +153,17 @@ public:
153 }153 }
154 154 
155 for (uint32_t i = 0; i < bFactor; i++) {155 for (uint32_t i = 0; i < bFactor; i++) {
156- int64_t inputOffset = (bOffset + i) * (tiling->inputSize + tiling->nfft) * COMPLEX_COEFFICIENT +156+ int64_t inputOffset = (int64_t)(bOffset + i) * (tiling->inputSize + tiling->nfft) * COMPLEX_COEFFICIENT +
157- nOffset * tiling->hopLength * COMPLEX_COEFFICIENT;157+ (int64_t)nOffset * tiling->hopLength * COMPLEX_COEFFICIENT;
158- int64_t realSplitWindowOffset = ((bOffset + i) * tiling->matmulN + nOffset) * tiling->nfftAlign;158+ int64_t realSplitWindowOffset = ((int64_t)(bOffset + i) * tiling->matmulN + nOffset) * tiling->nfftAlign;
159 int64_t imagSplitWindowOffset = realSplitWindowOffset;159 int64_t imagSplitWindowOffset = realSplitWindowOffset;
160 int64_t outputOffset =160 int64_t outputOffset =
161- (((bOffset + i) * tiling->matmulM + mOffset) * tiling->matmulN + nOffset) * DOUBLE_BUFFER;161+ (((int64_t)(bOffset + i) * tiling->matmulM + mOffset) * tiling->matmulN + nOffset) * DOUBLE_BUFFER;
162- int64_t realOffset = (bOffset + i) * tiling->matmulM * tiling->matmulN + mOffset * tiling->matmulN +162+ int64_t realOffset = (int64_t)(bOffset + i) * tiling->matmulM * tiling->matmulN +
163- nIdx * mFactor * tiling->matmulNCoreFactor;163+ (int64_t)mOffset * tiling->matmulN +
164+ (int64_t)nIdx * mFactor * tiling->matmulNCoreFactor;
164 int64_t imagOffset = realOffset;165 int64_t imagOffset = realOffset;
165- int64_t a1Offset = mOffset * tiling->nfftAlign;166+ int64_t a1Offset = (int64_t)mOffset * tiling->nfftAlign;
166 int64_t a2Offset = a1Offset;167 int64_t a2Offset = a1Offset;
167 168 
168 SplitWindows(inputOffset, realSplitWindowOffset, imagSplitWindowOffset, nFactor);169 SplitWindows(inputOffset, realSplitWindowOffset, imagSplitWindowOffset, nFactor);
@@ -228,9 +229,9 @@ private:
228 __aicore__ inline void GatherForSmallNFactorAlign(229 __aicore__ inline void GatherForSmallNFactorAlign(
229 int64_t realOffset, int64_t imagOffset, int64_t outputOffset, uint32_t mFactor, uint32_t nFactor)230 int64_t realOffset, int64_t imagOffset, int64_t outputOffset, uint32_t mFactor, uint32_t nFactor)
230 {231 {
231- int32_t complexCount = mFactor * nFactor * DOUBLE_BUFFER;232+ int64_t complexCount = (int64_t)mFactor * nFactor * DOUBLE_BUFFER;
232 int32_t ubCount = tiling->maskUBSize / sizeof(int32_t) / DOUBLE_BUFFER;233 int32_t ubCount = tiling->maskUBSize / sizeof(int32_t) / DOUBLE_BUFFER;
233- int32_t gatherCountPerLoop = complexCount > maskCount ? maskCount : complexCount;234+ int32_t gatherCountPerLoop = (int32_t)(complexCount > maskCount ? maskCount : complexCount);
234 235 
235 gatherCountPerLoop = gatherCountPerLoop - gatherCountPerLoop % (nFactor * DOUBLE_BUFFER);236 gatherCountPerLoop = gatherCountPerLoop - gatherCountPerLoop % (nFactor * DOUBLE_BUFFER);
236 int32_t realCountPerLoop = gatherCountPerLoop / DOUBLE_BUFFER;237 int32_t realCountPerLoop = gatherCountPerLoop / DOUBLE_BUFFER;
@@ -239,7 +240,7 @@ private:
239 AscendC::SetFlag<AscendC::HardEvent::MTE3_MTE2>(EVENT_ID0);240 AscendC::SetFlag<AscendC::HardEvent::MTE3_MTE2>(EVENT_ID0);
240 AscendC::SetFlag<AscendC::HardEvent::MTE3_MTE2>(EVENT_ID1);241 AscendC::SetFlag<AscendC::HardEvent::MTE3_MTE2>(EVENT_ID1);
241 int ping = 1;242 int ping = 1;
242- int repeats = (complexCount + gatherCountPerLoop - 1) / gatherCountPerLoop;243+ int repeats = (int)((complexCount + gatherCountPerLoop - 1) / gatherCountPerLoop);
243 244 
244 for (int i = 0; i < repeats; i++) {245 for (int i = 0; i < repeats; i++) {
245 event_t event_id = ping ? EVENT_ID0 : EVENT_ID1;246 event_t event_id = ping ? EVENT_ID0 : EVENT_ID1;
@@ -252,7 +253,7 @@ private:
252 int32_t copyLen = realCountPerLoop * sizeof(T);253 int32_t copyLen = realCountPerLoop * sizeof(T);
253 254 
254 if (i == repeats - 1) {255 if (i == repeats - 1) {
255- copyLen = (mFactor * nFactor - realCountPerLoop * i) * sizeof(T);256+ copyLen = ((int64_t)mFactor * nFactor - (int64_t)realCountPerLoop * i) * sizeof(T);
256 }257 }
257 258 
258 int32_t nBlocks = (copyLen + BLOCK_SIZE - 1) / BLOCK_SIZE;259 int32_t nBlocks = (copyLen + BLOCK_SIZE - 1) / BLOCK_SIZE;
@@ -304,8 +305,8 @@ private:
304 int64_t realOffset, int64_t imagOffset, int64_t outputOffset, uint32_t mFactor, uint32_t nFactor)305 int64_t realOffset, int64_t imagOffset, int64_t outputOffset, uint32_t mFactor, uint32_t nFactor)
305 {306 {
306 int32_t nFactorAlign = (nFactor * sizeof(T) + BLOCK_SIZE - 1) / BLOCK_SIZE * BLOCK_SIZE / sizeof(T);307 int32_t nFactorAlign = (nFactor * sizeof(T) + BLOCK_SIZE - 1) / BLOCK_SIZE * BLOCK_SIZE / sizeof(T);
307- int32_t complexCount = mFactor * nFactorAlign * DOUBLE_BUFFER;308+ int64_t complexCount = (int64_t)mFactor * nFactorAlign * DOUBLE_BUFFER;
308- int32_t gatherCountPerLoop = complexCount > maskCount ? maskCount : complexCount;309+ int32_t gatherCountPerLoop = (int32_t)(complexCount > maskCount ? maskCount : complexCount);
309 gatherCountPerLoop = gatherCountPerLoop - gatherCountPerLoop % (nFactorAlign * DOUBLE_BUFFER);310 gatherCountPerLoop = gatherCountPerLoop - gatherCountPerLoop % (nFactorAlign * DOUBLE_BUFFER);
310 int32_t realCountPerLoop = gatherCountPerLoop / DOUBLE_BUFFER;311 int32_t realCountPerLoop = gatherCountPerLoop / DOUBLE_BUFFER;
311 int32_t imagCountPerLoop = gatherCountPerLoop / DOUBLE_BUFFER;312 int32_t imagCountPerLoop = gatherCountPerLoop / DOUBLE_BUFFER;
@@ -314,7 +315,7 @@ private:
314 AscendC::SetFlag<AscendC::HardEvent::MTE3_MTE2>(EVENT_ID1);315 AscendC::SetFlag<AscendC::HardEvent::MTE3_MTE2>(EVENT_ID1);
315 316 
316 int ping = 1;317 int ping = 1;
317- int repeats = (complexCount + gatherCountPerLoop - 1) / gatherCountPerLoop;318+ int repeats = (int)((complexCount + gatherCountPerLoop - 1) / gatherCountPerLoop);
318 319 
319 for (int i = 0; i < repeats; i++) {320 for (int i = 0; i < repeats; i++) {
320 event_t event_id = ping ? EVENT_ID0 : EVENT_ID1;321 event_t event_id = ping ? EVENT_ID0 : EVENT_ID1;
@@ -326,7 +327,7 @@ private:
326 327 
327 int32_t copyLen = realCountPerLoop * sizeof(T);328 int32_t copyLen = realCountPerLoop * sizeof(T);
328 if (i == repeats - 1) {329 if (i == repeats - 1) {
329- copyLen = (mFactor * nFactorAlign - realCountPerLoop * i) * sizeof(T);330+ copyLen = ((int64_t)mFactor * nFactorAlign - (int64_t)realCountPerLoop * i) * sizeof(T);
330 }331 }
331 332 
332 AscendC::WaitFlag<AscendC::HardEvent::MTE3_MTE2>(event_id);333 AscendC::WaitFlag<AscendC::HardEvent::MTE3_MTE2>(event_id);
@@ -402,7 +403,7 @@ private:
402 403 
403 int32_t copyLen = realCountPerLoop * sizeof(T);404 int32_t copyLen = realCountPerLoop * sizeof(T);
404 if (i == repeats - 1) {405 if (i == repeats - 1) {
405- copyLen = (mFactor * nFactor - realCountPerLoop * i) * sizeof(T);406+ copyLen = ((int64_t)mFactor * nFactor - (int64_t)realCountPerLoop * i) * sizeof(T);
406 }407 }
407 408 
408 int32_t nBlocks = (copyLen + BLOCK_SIZE - 1) / BLOCK_SIZE;409 int32_t nBlocks = (copyLen + BLOCK_SIZE - 1) / BLOCK_SIZE;
@@ -476,7 +477,7 @@ private:
476 477 
477 int32_t copyLen = realCountPerLoop * sizeof(T);478 int32_t copyLen = realCountPerLoop * sizeof(T);
478 if (i == repeats - 1) {479 if (i == repeats - 1) {
479- copyLen = (mFactor * nFactor - realCountPerLoop * i) * sizeof(T);480+ copyLen = ((int64_t)mFactor * nFactor - (int64_t)realCountPerLoop * i) * sizeof(T);
480 }481 }
481 482 
482 int32_t nBlocks = (copyLen + BLOCK_SIZE - 1) / BLOCK_SIZE;483 int32_t nBlocks = (copyLen + BLOCK_SIZE - 1) / BLOCK_SIZE;
Mmath/stft/op_kernel/stft_plan_mul.h+5-5
@@ -40,10 +40,10 @@ public:
40 col = inTilingData->oneRowLen;40 col = inTilingData->oneRowLen;
41 ubMaxLine = inTilingData->ubMaxLine;41 ubMaxLine = inTilingData->ubMaxLine;
42 row = totalRow;42 row = totalRow;
43- int32_t planGlobalOffset = blockIdx * totalRow * col;43+ int64_t planGlobalOffset = (int64_t)blockIdx * totalRow * col;
44 if (blockIdx >= inTilingData->tailBlockIdx) {44 if (blockIdx >= inTilingData->tailBlockIdx) {
45- planGlobalOffset = inTilingData->tailBlockIdx * totalRow * col +45+ planGlobalOffset = (int64_t)inTilingData->tailBlockIdx * totalRow * col +
46- (blockIdx - inTilingData->tailBlockIdx) * tailRow * col;46+ (int64_t)(blockIdx - inTilingData->tailBlockIdx) * tailRow * col;
47 row = tailRow;47 row = tailRow;
48 }48 }
49 uint64_t splitWindowWorkspaceSize =49 uint64_t splitWindowWorkspaceSize =
@@ -56,9 +56,9 @@ public:
56 WORKSPACE_ALIGN_SIZE) *56 WORKSPACE_ALIGN_SIZE) *
57 WORKSPACE_ALIGN_SIZE / sizeof(T);57 WORKSPACE_ALIGN_SIZE / sizeof(T);
58 uint64_t planOffset = splitWindowWorkspaceSize + matmulWorkspaceSize;58 uint64_t planOffset = splitWindowWorkspaceSize + matmulWorkspaceSize;
59- planGm.SetGlobalBuffer((__gm__ T*)plan + planGlobalOffset, row * col);59+ planGm.SetGlobalBuffer((__gm__ T*)plan + planGlobalOffset, (int64_t)row * col);
60 windowGm.SetGlobalBuffer((__gm__ T*)window, col);60 windowGm.SetGlobalBuffer((__gm__ T*)window, col);
61- outputGm.SetGlobalBuffer((__gm__ T*)workspace + planOffset + planGlobalOffset, row * col);61+ outputGm.SetGlobalBuffer((__gm__ T*)workspace + planOffset + planGlobalOffset, (int64_t)row * col);
62 pipe->InitBuffer(planInQue, bufferNum, ubMaxLine * col * sizeof(T));62 pipe->InitBuffer(planInQue, bufferNum, ubMaxLine * col * sizeof(T));
63 pipe->InitBuffer(planOutQue, bufferNum, ubMaxLine * col * sizeof(T));63 pipe->InitBuffer(planOutQue, bufferNum, ubMaxLine * col * sizeof(T));
64 pipe->InitBuffer(windowInQue, 1, col * sizeof(T));64 pipe->InitBuffer(windowInQue, 1, col * sizeof(T));