fix int32-ops-nn-AtomGit

Star887

Star887

已合并

fix int32 #5525

liuyun_nj创建于 6月1日从liuyun_nj/ops-nn_1： master 合入到 cann/ops-nn： master

fix int32 #5525

已合并

liuyun_nj创建于 6月1日

从liuyun_nj/ops-nn_1 master合入到cann/ops-nn master

共 7 个文件变更+8-8

Mnorm/add_rms_norm/op_kernel/add_rms_norm.h+1-1

	@@ -90,7 +90,7 @@ public:
90	{	90	{
91	LocalTensor<float> rstdLocal = outQueueRstd.AllocTensor<float>();	91	LocalTensor<float> rstdLocal = outQueueRstd.AllocTensor<float>();
92	for (uint32_t i_i = 0; i_i < calc_row_num; i_i++) {	92	for (uint32_t i_i = 0; i_i < calc_row_num; i_i++) {
93	- uint32_t gm_bias = (i_o * this->rowFactor + i_i) * this->numCol;	93	+ uint64_t gm_bias = (static_cast<uint64_t>(i_o) * static_cast<uint64_t>(this->rowFactor) + static_cast<uint64_t>(i_i)) * static_cast<uint64_t>(this->numCol);
94	CopyIn(gm_bias);	94	CopyIn(gm_bias);
95	Compute(i_i, gammaLocal, rstdLocal);	95	Compute(i_i, gammaLocal, rstdLocal);
96	CopyOutY(gm_bias);	96	CopyOutY(gm_bias);

Mnorm/add_rms_norm/op_kernel/add_rms_norm_merge_n.h+1-1

	@@ -104,7 +104,7 @@ public:
104		104
105	__aicore__ inline void MainCompute(uint32_t i_o, uint32_t calc_row_num, LocalTensor<T>& gammaLocal)	105	__aicore__ inline void MainCompute(uint32_t i_o, uint32_t calc_row_num, LocalTensor<T>& gammaLocal)
106	{	106	{
107	- uint32_t gm_bias = i_o * rowFactor * numCol;	107	+ uint64_t gm_bias = static_cast<uint64_t>(i_o) * static_cast<uint64_t>(rowFactor) * static_cast<uint64_t>(numCol);
108	uint32_t elementNum = calc_row_num * numColAlign;	108	uint32_t elementNum = calc_row_num * numColAlign;
109	CopyInX(gm_bias, calc_row_num);	109	CopyInX(gm_bias, calc_row_num);
110	LocalTensor<T> xLocal = ComputeX(elementNum);	110	LocalTensor<T> xLocal = ComputeX(elementNum);

Mnorm/add_rms_norm/op_kernel/add_rms_norm_multi_n.h+1-1

	@@ -95,7 +95,7 @@ public:
95		95
96	__aicore__ inline void SubProcessHalf(uint32_t i_o, uint32_t calc_row_num, LocalTensor<T>& gammaLocal)	96	__aicore__ inline void SubProcessHalf(uint32_t i_o, uint32_t calc_row_num, LocalTensor<T>& gammaLocal)
97	{	97	{
98	- uint32_t gm_bias = i_o * rowFactor * numCol;	98	+ uint64_t gm_bias = static_cast<uint64_t>(i_o) * static_cast<uint64_t>(rowFactor) * static_cast<uint64_t>(numCol);
99	CopyInX(gm_bias, calc_row_num);	99	CopyInX(gm_bias, calc_row_num);
100	LocalTensor<T> xLocal = ComputeX(calc_row_num);	100	LocalTensor<T> xLocal = ComputeX(calc_row_num);
101	if constexpr (MODE == ADD_RMS_NORM_MODE \|\| MODE == PRE_RMS_NORM_MODE) {	101	if constexpr (MODE == ADD_RMS_NORM_MODE \|\| MODE == PRE_RMS_NORM_MODE) {

Mnorm/add_rms_norm_cast/op_kernel/add_rms_norm_cast.h+1-1

	@@ -83,7 +83,7 @@ public:
83	__aicore__ inline void SubProcess(uint32_t i_o, uint32_t calc_row_num)	83	__aicore__ inline void SubProcess(uint32_t i_o, uint32_t calc_row_num)
84	{	84	{
85	for (uint32_t i_i = 0; i_i < calc_row_num; i_i++) {	85	for (uint32_t i_i = 0; i_i < calc_row_num; i_i++) {
86	- uint32_t gm_bias = (i_o * rowFactor + i_i) * numCol;	86	+ uint32_t gm_bias = (static_cast<uint64_t>(i_o) * static_cast<uint64_t>(rowFactor) + static_cast<uint64_t>(i_i)) * static_cast<uint64_t>(numCol);
87	CopyIn(gm_bias);	87	CopyIn(gm_bias);
88	if constexpr (is_same<T, half>::value) {	88	if constexpr (is_same<T, half>::value) {
89	Computefp16(i_o, i_i, gm_bias);	89	Computefp16(i_o, i_i, gm_bias);

Mnorm/add_rms_norm_cast/op_kernel/add_rms_norm_cast_multi_n.h+1-1

	@@ -94,7 +94,7 @@ public:
94		94
95	__aicore__ inline void SubProcess(uint32_t i_o, uint32_t calc_row_num, LocalTensor<T>& gammaLocal)	95	__aicore__ inline void SubProcess(uint32_t i_o, uint32_t calc_row_num, LocalTensor<T>& gammaLocal)
96	{	96	{
97	- uint32_t gm_bias = i_o * rowFactor * numCol;	97	+ uint64_t gm_bias = static_cast<uint64_t>(i_o) * static_cast<uint64_t>(rowFactor) * static_cast<uint64_t>(numCol);
98	CopyInX(gm_bias, calc_row_num);	98	CopyInX(gm_bias, calc_row_num);
99	LocalTensor<T> xLocal = ComputeX(calc_row_num);	99	LocalTensor<T> xLocal = ComputeX(calc_row_num);
100	CopyOutX(gm_bias, calc_row_num);	100	CopyOutX(gm_bias, calc_row_num);

Mnorm/add_rms_norm_dynamic_quant/op_kernel/add_rms_norm_dynamic_quant_normal_kernel.h+1-1

	@@ -63,7 +63,7 @@ public:
63		63
64	LocalTensor<T> gammaLocal = weightBuf01.template Get<T>();	64	LocalTensor<T> gammaLocal = weightBuf01.template Get<T>();
65		65
66	- int32_t gmOffset = 0;	66	+ uint64_t gmOffset = 0;
67	int32_t gmOffsetScale = 0;	67	int32_t gmOffsetScale = 0;
68	int32_t elementCount = this->numLastDimAligned * this->rowStep;	68	int32_t elementCount = this->numLastDimAligned * this->rowStep;
69		69

Mnorm/rms_norm/op_kernel/rms_norm_merge_n.h+2-2

	@@ -97,8 +97,8 @@ public:
97		97
98	__aicore__ inline void DoMainCompute(uint32_t i_o, uint32_t calcRowNum, LocalTensor<float>& gammaLocal)	98	__aicore__ inline void DoMainCompute(uint32_t i_o, uint32_t calcRowNum, LocalTensor<float>& gammaLocal)
99	{	99	{
100	- uint32_t gmBias = i_o * rowFactor * numCol;	100	+ uint64_t gmBias = static_cast<uint64_t>(i_o) * static_cast<uint64_t>(rowFactor) * static_cast<uint64_t>(numCol);
101	- uint32_t elementNum = calcRowNum * numColAlign;	101	+ uint64_t elementNum = static_cast<uint64_t>(calcRowNum) * static_cast<uint64_t>(numColAlign);
102	CopyIn(gmBias, calcRowNum);	102	CopyIn(gmBias, calcRowNum);
103	if (isNorm == 1) {	103	if (isNorm == 1) {
104	ComputeSingleRow(i_o, calcRowNum, gammaLocal, gmBias);	104	ComputeSingleRow(i_o, calcRowNum, gammaLocal, gmBias);