Star1329
1447
代码
代码
Issues117
Pull Requests256
Actions
项目讨论
Wiki
分析
Star1329
1447
已合并
hans_encode/hans_decode算子bug修复 #449
Nythas N创建于 2025年12月20日
hans_encode/hans_decode算子bug修复 #449
已合并
Nythas N创建于 2025年12月20日
从NingGuangyou/ops-math_3565master合入到cann/ops-mathmaster
共 5 个文件变更+49-38
Amath/hans_decode/docs/aclnnHansDecode.md+3-0
@@ -0,0 +1,3 @@
1+# aclnnHansDecode
2+ 
3+本文档内容正按全新接口模板整改中，将陆续上线，如需使用该接口请访问昇腾社区[《算子库接口》](https://hiascend.com/document/redirect/CannCommunityOplist)对应的aclnnHansDecode章节。
Mmath/hans_decode/op_kernel/hans_decode_base.h+13-14
@@ -362,10 +362,10 @@ __aicore__ inline void HansDecode<IF_BF16>::GetNegativeBitNum(DecodeLoopConfig*
362 stateBufferBitNumberUb.template ReinterpretCast<float>(), const16Ub.template ReinterpretCast<float>(),362 stateBufferBitNumberUb.template ReinterpretCast<float>(), const16Ub.template ReinterpretCast<float>(),
363 AscendC::CMPMODE::LE, EACH_LOOOP_STATE_READ_NUM);363 AscendC::CMPMODE::LE, EACH_LOOOP_STATE_READ_NUM);
364 PipeBarrier<PIPE_V>();364 PipeBarrier<PIPE_V>();
365- GatherMask(negativeBitTempSpace.template ReinterpretCast<uint32_t>(),365+ GatherMask(
366- constOneUb.template ReinterpretCast<uint32_t>(),366+ negativeBitTempSpace.template ReinterpretCast<uint32_t>(), constOneUb.template ReinterpretCast<uint32_t>(),
367- negativeZeroCmpBitMask.template ReinterpretCast<uint32_t>(),367+ negativeZeroCmpBitMask.template ReinterpretCast<uint32_t>(), true, EACH_REPEAT_BYTES / sizeof(uint32_t),
368- false, 0, {1, 1, static_cast<uint8_t>(REPEAT_STRIDE), 1}, this->rsvdCnt);368+ {1, 1, REPEAT_STRIDE, 0}, this->rsvdCnt);
369 PipeBarrier<PIPE_V>();369 PipeBarrier<PIPE_V>();
370 config->negativeZeroSum = AscendC::AscendCUtils::GetRsvdCnt();370 config->negativeZeroSum = AscendC::AscendCUtils::GetRsvdCnt();
371 config->loopDataIndex = config->loopDataIndex - EACH_LOOOP_STATE_READ_NUM;371 config->loopDataIndex = config->loopDataIndex - EACH_LOOOP_STATE_READ_NUM;
@@ -561,11 +561,11 @@ __aicore__ inline void HansDecode<IF_BF16>::MoveOutFp32(
561 PipeBarrier<PIPE_V>();561 PipeBarrier<PIPE_V>();
562 Cast(recoverMantissaFp162s32Ub, recoverMantissaFp16Ub, AscendC::RoundMode::CAST_CEIL, PROCESS_3072);562 Cast(recoverMantissaFp162s32Ub, recoverMantissaFp16Ub, AscendC::RoundMode::CAST_CEIL, PROCESS_3072);
563 PipeBarrier<PIPE_V>();563 PipeBarrier<PIPE_V>();
564- GatherMask(recoverMantissaFp16GatherUb.template ReinterpretCast<uint16_t>(),564+ GatherMask(
565- recoverMantissaFp162s32Ub.template ReinterpretCast<uint16_t>(),565+ recoverMantissaFp16GatherUb.template ReinterpretCast<uint16_t>(),
566- mantissaBitMask.template ReinterpretCast<uint16_t>(), false, 0,566+ recoverMantissaFp162s32Ub.template ReinterpretCast<uint16_t>(),
567- {static_cast<uint8_t>(3 * mantissaRepeatTimes), 1, static_cast<uint8_t>(REPEAT_STRIDE), 1},567+ mantissaBitMask.template ReinterpretCast<uint16_t>(), true, 3 * mantissaRepeatTimes * EACH_REPEAT_BYTES / sizeof(uint16_t),
568- this->rsvdCnt);568+ {1, 1, REPEAT_STRIDE, 0}, this->rsvdCnt);
569 PipeBarrier<PIPE_V>();569 PipeBarrier<PIPE_V>();
570 Cast(570 Cast(
571 recoverMantissaFp16GatherUb.template ReinterpretCast<half>(),571 recoverMantissaFp16GatherUb.template ReinterpretCast<half>(),
@@ -724,11 +724,10 @@ __aicore__ inline void HansDecode<IF_BF16>::ProcessFp32BeforeH2D(
724 AscendC::RoundMode::CAST_TRUNC, (size * (sizeof(float) - 1)));724 AscendC::RoundMode::CAST_TRUNC, (size * (sizeof(float) - 1)));
725 PipeBarrier<PIPE_V>();725 PipeBarrier<PIPE_V>();
726 SetVectorMask<uint64_t, MaskMode::NORMAL>((uint64_t)-1, (uint64_t)-1);726 SetVectorMask<uint64_t, MaskMode::NORMAL>((uint64_t)-1, (uint64_t)-1);
727- GatherMask(deviceExpUb[index].template ReinterpretCast<uint16_t>(),727+ GatherMask(
728- recoverUb[index].template ReinterpretCast<uint16_t>(),728+ deviceExpUb[index].template ReinterpretCast<uint16_t>(), recoverUb[index].template ReinterpretCast<uint16_t>(),
729- mantissaBitMask.template ReinterpretCast<uint16_t>(), false, 0,729+ mantissaBitMask.template ReinterpretCast<uint16_t>(), true, (size * (sizeof(float) - 1) * 2) / CONST_128 * EACH_REPEAT_BYTES / sizeof(uint16_t),
730- {static_cast<uint8_t>((size * (sizeof(float) - 1) * 2) / CONST_128), 1, static_cast<uint8_t>(REPEAT_STRIDE), 1},730+ {1, 1, REPEAT_STRIDE, 0}, this->rsvdCnt);
731- this->rsvdCnt);
732 PipeBarrier<PIPE_V>();731 PipeBarrier<PIPE_V>();
733 Cast(732 Cast(
734 deviceExpUb[index].template ReinterpretCast<half>(), deviceExpUb[index].template ReinterpretCast<int16_t>(),733 deviceExpUb[index].template ReinterpretCast<half>(), deviceExpUb[index].template ReinterpretCast<int16_t>(),
Amath/hans_encode/docs/aclnnHansEncode.md+3-0
@@ -0,0 +1,3 @@
1+# aclnnHansEncode
2+ 
3+本文档内容正按全新接口模板整改中，将陆续上线，如需使用该接口请访问昇腾社区[《算子库接口》](https://hiascend.com/document/redirect/CannCommunityOplist)对应的aclnnHansEncode章节。
Mmath/hans_encode/op_kernel/hans_encode_base.h+25-20
@@ -287,10 +287,11 @@ class HansEncode {
287 GatherMask(this->inputExpHalfLocal, this->inputHalfLocal, expPattern, false, 0, expParams, this->rsvdCnt);287 GatherMask(this->inputExpHalfLocal, this->inputHalfLocal, expPattern, false, 0, expParams, this->rsvdCnt);
288 PipeBarrier<PIPE_V>();288 PipeBarrier<PIPE_V>();
289 if (this->dtypeSize == sizeof(float)) {289 if (this->dtypeSize == sizeof(float)) {
290- GatherMask(this->inputHalfLocal.template ReinterpretCast<uint16_t>(),290+ GatherMask(
291- this->inputHalfLocal.template ReinterpretCast<uint16_t>(),291+ this->inputHalfLocal.template ReinterpretCast<uint16_t>(),
292- this->fp32MantissaMaskLocal.template ReinterpretCast<uint16_t>(),292+ this->inputHalfLocal.template ReinterpretCast<uint16_t>(),
293- false, 0, {static_cast<uint8_t>(expParams.repeatTimes), 1, 8, 1}, this->rsvdCnt);293+ this->fp32MantissaMaskLocal.template ReinterpretCast<uint16_t>(), true,
294+ expParams.repeatTimes * EACH_REPEAT_BYTES / sizeof(uint16_t), {1, 1, 8, 0}, this->rsvdCnt);
294 } else {295 } else {
295 GatherMask(this->inputHalfLocal, this->inputHalfLocal, 1, false, 0, expParams, this->rsvdCnt);296 GatherMask(this->inputHalfLocal, this->inputHalfLocal, 1, false, 0, expParams, this->rsvdCnt);
296 }297 }
@@ -346,10 +347,11 @@ class HansEncode {
346 CompareScalar(stateBufferBitCmpLocal.ReinterpretCast<uint8_t>(), overflowCheckLocal.ReinterpretCast<float>(),347 CompareScalar(stateBufferBitCmpLocal.ReinterpretCast<uint8_t>(), overflowCheckLocal.ReinterpretCast<float>(),
347 this->const16Float, AscendC::CMPMODE::GT, EACH_LOOOP_REPEAT_TIMES);348 this->const16Float, AscendC::CMPMODE::GT, EACH_LOOOP_REPEAT_TIMES);
348 PipeBarrier<PIPE_V>();349 PipeBarrier<PIPE_V>();
349- GatherMask(this->inputHalfLocal.template ReinterpretCast<uint32_t>(),350+ GatherMask(
350- overflowCheckLocal.template ReinterpretCast<uint32_t>(),351+ this->inputHalfLocal.template ReinterpretCast<uint32_t>(),
351- stateBufferBitCmpLocal.template ReinterpretCast<uint32_t>(),352+ overflowCheckLocal.template ReinterpretCast<uint32_t>(),
352- false, 0, {1, 1, 8, 1}, this->rsvdCnt);353+ stateBufferBitCmpLocal.template ReinterpretCast<uint32_t>(), true, EACH_REPEAT_BYTES / sizeof(uint32_t),
354+ {1, 1, 8, 0}, this->rsvdCnt);
353 uint64_t compressBufferOffset = AscendC::AscendCUtils::GetRsvdCnt();355 uint64_t compressBufferOffset = AscendC::AscendCUtils::GetRsvdCnt();
354 this->bufOffset -= (EACH_LOOOP_REPEAT_TIMES * sizeof(int32_t) + EACH_LOOOP_REPEAT_TIMES * sizeof(int32_t));356 this->bufOffset -= (EACH_LOOOP_REPEAT_TIMES * sizeof(int32_t) + EACH_LOOOP_REPEAT_TIMES * sizeof(int32_t));
355 if ((currentCoreOutputAcculmulateSize + EACH_LOOOP_REPEAT_TIMES * sizeof(int32_t) +357 if ((currentCoreOutputAcculmulateSize + EACH_LOOOP_REPEAT_TIMES * sizeof(int32_t) +
@@ -410,18 +412,20 @@ class HansEncode {
410 PipeBarrier<PIPE_V>();412 PipeBarrier<PIPE_V>();
411 Brcb(overflowFlagBrcbLocal, overflowFlagLocal, BYTE_BIT_NUM, {1, 8});413 Brcb(overflowFlagBrcbLocal, overflowFlagLocal, BYTE_BIT_NUM, {1, 8});
412 PipeBarrier<PIPE_V>();414 PipeBarrier<PIPE_V>();
413- GatherMask(overflowFlagBrcbReduceLocal.template ReinterpretCast<uint32_t>(),415+ GatherMask(
414- overflowFlagBrcbLocal.template ReinterpretCast<uint32_t>(),416+ overflowFlagBrcbReduceLocal.template ReinterpretCast<uint32_t>(),
415- this->reduceLowBlockMaskLocal.template ReinterpretCast<uint32_t>(),417+ overflowFlagBrcbLocal.template ReinterpretCast<uint32_t>(),
416- false, 0, {8, 1, 8, 1}, this->rsvdCnt);418+ this->reduceLowBlockMaskLocal.template ReinterpretCast<uint32_t>(), true, CONST_8 * EACH_REPEAT_BYTES / sizeof(uint32_t),
419+ {1, 1, 8, 0}, this->rsvdCnt);
417 PipeBarrier<PIPE_V>();420 PipeBarrier<PIPE_V>();
418 Muls(overflowFlagBrcbReduceLocal, overflowFlagBrcbReduceLocal, (int32_t)INT32_LOW_16_BIT_MASK,421 Muls(overflowFlagBrcbReduceLocal, overflowFlagBrcbReduceLocal, (int32_t)INT32_LOW_16_BIT_MASK,
419 EACH_LOOOP_PROCESS_NUM / (sizeof(uint16_t) * BYTE_BIT_NUM));422 EACH_LOOOP_PROCESS_NUM / (sizeof(uint16_t) * BYTE_BIT_NUM));
420 PipeBarrier<PIPE_V>();423 PipeBarrier<PIPE_V>();
421- GatherMask(this->outputDeviceLocal.template ReinterpretCast<uint16_t>(),424+ GatherMask(
422- this->stateBufferLocal.template ReinterpretCast<uint16_t>(),425+ this->outputDeviceLocal.template ReinterpretCast<uint16_t>(),
423- overflowFlagBrcbReduceLocal.template ReinterpretCast<uint16_t>(),426+ this->stateBufferLocal.template ReinterpretCast<uint16_t>(),
424- false, 0, {static_cast<uint8_t>(processDataLength / BLOCK_SIZE), 1, 8, 1}, this->rsvdCnt);427+ overflowFlagBrcbReduceLocal.template ReinterpretCast<uint16_t>(), true,
428+ processDataLength / BLOCK_SIZE * EACH_REPEAT_BYTES / sizeof(uint16_t), {1, 1, 8, 0}, this->rsvdCnt);
425 this->compressBufferOffset = AscendC::AscendCUtils::GetRsvdCnt();429 this->compressBufferOffset = AscendC::AscendCUtils::GetRsvdCnt();
426 GatherOutput(overflowFlagBrcbLocal, overflowFlagLocal, processDataLength);430 GatherOutput(overflowFlagBrcbLocal, overflowFlagLocal, processDataLength);
427 this->bufOffset -= (BLOCK_SIZE * sizeof(int32_t) + EACH_LOOOP_REPEAT_TIMES * sizeof(int32_t) +431 this->bufOffset -= (BLOCK_SIZE * sizeof(int32_t) + EACH_LOOOP_REPEAT_TIMES * sizeof(int32_t) +
@@ -626,10 +630,11 @@ class HansEncode {
626 this->rsvdCnt);630 this->rsvdCnt);
627 PipeBarrier<PIPE_V>();631 PipeBarrier<PIPE_V>();
628 if (this->dtypeSize == sizeof(int32_t)) {632 if (this->dtypeSize == sizeof(int32_t)) {
629- GatherMask(inputReaminMantissaFp16LocalList[index].template ReinterpretCast<uint16_t>(),633+ GatherMask(
630- inputReaminFp16LocalList[index].template ReinterpretCast<uint16_t>(),634+ inputReaminMantissaFp16LocalList[index].template ReinterpretCast<uint16_t>(),
631- this->fp32MantissaMaskLocal.template ReinterpretCast<uint16_t>(),635+ inputReaminFp16LocalList[index].template ReinterpretCast<uint16_t>(),
632- false, 0, {static_cast<uint8_t>(expParams.repeatTimes), 1, 8, 1}, this->rsvdCnt);636+ this->fp32MantissaMaskLocal.template ReinterpretCast<uint16_t>(), true, expParams.repeatTimes * EACH_REPEAT_BYTES / sizeof(uint16_t),
637+ {1, 1, 8, 0}, this->rsvdCnt);
633 } else {638 } else {
634 GatherMask(inputReaminMantissaFp16LocalList[index], inputReaminFp16LocalList[index], 1, false, 0, expParams,639 GatherMask(inputReaminMantissaFp16LocalList[index], inputReaminFp16LocalList[index], 1, false, 0, expParams,
635 this->rsvdCnt);640 this->rsvdCnt);
Mmath/hans_encode/op_kernel/hans_pdf_statistic_base.h+5-4
@@ -141,10 +141,11 @@ protected:
141 compareLocalInt.template ReinterpretCast<uint8_t>(), this->inputLocal, scalarValue, CMPMODE::EQ,141 compareLocalInt.template ReinterpretCast<uint8_t>(), this->inputLocal, scalarValue, CMPMODE::EQ,
142 computeLength);142 computeLength);
143 PipeBarrier<PIPE_V>();143 PipeBarrier<PIPE_V>();
144- GatherMask(this->expLocal.template ReinterpretCast<intType>(),144+ GatherMask(
145- this->inputLocal.template ReinterpretCast<intType>(),145+ this->expLocal.template ReinterpretCast<intType>(),
146- compareLocalInt.template ReinterpretCast<intType>(),146+ this->inputLocal.template ReinterpretCast<intType>(),
147- false, 0, {static_cast<uint8_t>(repeatTimes), 1, 8, 1}, this->rsvdCnt);147+ compareLocalInt.template ReinterpretCast<intType>(), true, repeatTimes * EACH_REPEAT_BYTES / sizeof(intType),
148+ {1, 1, 8, 0}, this->rsvdCnt);
148 rsvdCnt = AscendC::AscendCUtils::GetRsvdCnt();149 rsvdCnt = AscendC::AscendCUtils::GetRsvdCnt();
149 PipeBarrier<PIPE_V>();150 PipeBarrier<PIPE_V>();
150 SetFlag<HardEvent::V_S>(eventVS);151 SetFlag<HardEvent::V_S>(eventVS);