From baed0193d34f4b54814c3cedc3f63ca96b516e58 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E9=99=88=E4=B8=80=E6=BA=90?= Date: Mon, 15 Sep 2025 15:01:30 +0800 Subject: [PATCH] =?UTF-8?q?clamp=20=E4=BF=AE=E6=94=B9=E5=AE=9E=E7=8E=B0?= =?UTF-8?q?=E6=96=B9=E5=BC=8F?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- impl/math/clamp/clamp_common_impl.h | 155 ++-------------------------- 1 file changed, 11 insertions(+), 144 deletions(-) diff --git a/impl/math/clamp/clamp_common_impl.h b/impl/math/clamp/clamp_common_impl.h index f48692fa..2c5984e7 100644 --- a/impl/math/clamp/clamp_common_impl.h +++ b/impl/math/clamp/clamp_common_impl.h @@ -16,150 +16,19 @@ #define IMPL_MATH_CLAMP_CLAMP_COMMON_IMPL_H #include "kernel_tensor.h" #include "../../common/check.h" +#include "../../api_check/kernel_api_check.h" namespace AscendC { -constexpr uint32_t CLAMP_FLOAT_MASK = 64; -constexpr uint32_t CLAMP_HALF_MASK = 128; -constexpr uint32_t CLAMP_BYTE_PER_REPEAT = 512; - -struct ClampParams { - __aicore__ ClampParams(){}; - uint32_t vcmpvsRepeat = 0; - uint64_t ClampMask = 0; - uint64_t selectTailElement = 0; - uint32_t selectTailOffset = 0; - uint32_t clampSplitCount = 0; - uint32_t selectTailRepeatLoop = 0; - uint32_t selectTailRepeatTail = 0; - uint32_t selectTailPreRepeatOffset = 0; - uint32_t selectTailMainRepeatOffset = 0; - uint32_t selectTailTailRepeatOffset = 0; - uint32_t loopCount = 0; - uint32_t calcTail = 0; - uint32_t vcmpvsRepeatLoop = 0; - uint32_t vcmpvsRepeatTail = 0; - uint32_t vcmpvsPreRepeatOffset = 0; - uint32_t vcmpvsMainRepeatOffset = 0; -}; - -template -__aicore__ inline void ClampComputeCount(const LocalTensor& dstTensor, const LocalTensor& srcTensor, - const LocalTensor& sharedTmpBuffer, const T scalar, const uint32_t repeat, const uint64_t mask, - CLAMPMODE selMode, const BinaryRepeatParams& repeatParams) -{ - if (selMode == CLAMPMODE::CLAMP_MAX) { - CompareScalar(sharedTmpBuffer, srcTensor, static_cast(scalar), CMPMODE::LT, mask, (uint8_t)repeat, - { DEFAULT_BLK_STRIDE, DEFAULT_BLK_STRIDE, DEFAULT_REPEAT_STRIDE, DEFAULT_REPEAT_STRIDE }); - } else if (selMode == CLAMPMODE::CLAMP_MIN) { - CompareScalar(sharedTmpBuffer, srcTensor, static_cast(scalar), CMPMODE::GT, mask, (uint8_t)repeat, - { DEFAULT_BLK_STRIDE, DEFAULT_BLK_STRIDE, DEFAULT_REPEAT_STRIDE, DEFAULT_REPEAT_STRIDE }); - } else { - ASCENDC_ASSERT(false, { KERNEL_LOG(KERNEL_ERROR, "selMode is not supported"); }); - } - PipeBarrier(); - // 2、selsct 1 - Select(dstTensor, sharedTmpBuffer, srcTensor, static_cast(scalar), SELMODE::VSEL_TENSOR_SCALAR_MODE, mask, - (uint8_t)repeat, repeatParams); - PipeBarrier(); -} - -template __aicore__ inline void GetMainParams(const uint32_t calCount, ClampParams& params, - const uint32_t sharedTmpBufferSize) -{ - if constexpr (sizeof(T) == sizeof(uint16_t)) { - params.clampSplitCount = params.clampSplitCount / CLAMP_HALF_MASK * CLAMP_HALF_MASK; - params.vcmpvsRepeat = params.clampSplitCount / CLAMP_HALF_MASK; - params.ClampMask = CLAMP_HALF_MASK; - } else { - params.clampSplitCount = params.clampSplitCount / CLAMP_FLOAT_MASK * CLAMP_FLOAT_MASK; - params.vcmpvsRepeat = params.clampSplitCount / CLAMP_FLOAT_MASK; - params.ClampMask = CLAMP_FLOAT_MASK; - } - CheckTmpBufferSize(params.clampSplitCount, 0, sharedTmpBufferSize); - params.loopCount = calCount / params.clampSplitCount; - params.calcTail = calCount % params.clampSplitCount; - params.vcmpvsRepeatLoop = params.vcmpvsRepeat / MAX_REPEAT_TIMES; - params.vcmpvsRepeatTail = params.vcmpvsRepeat % MAX_REPEAT_TIMES; - params.vcmpvsPreRepeatOffset = MAX_REPEAT_TIMES * params.ClampMask; - params.vcmpvsMainRepeatOffset = params.vcmpvsRepeatLoop * MAX_REPEAT_TIMES * params.ClampMask; -} - -template __aicore__ inline void GetTailParams(const uint32_t calcTail, ClampParams& params) -{ - if constexpr (sizeof(T) == sizeof(uint16_t)) { - params.vcmpvsRepeat = calcTail / CLAMP_HALF_MASK; - params.ClampMask = (calcTail < CLAMP_HALF_MASK) ? calcTail : CLAMP_HALF_MASK; - params.selectTailElement = calcTail % CLAMP_HALF_MASK; - params.selectTailOffset = params.vcmpvsRepeat * CLAMP_HALF_MASK; - } else { - params.vcmpvsRepeat = calcTail / CLAMP_FLOAT_MASK; - params.ClampMask = (calcTail < CLAMP_FLOAT_MASK) ? calcTail : CLAMP_FLOAT_MASK; - params.selectTailElement = calcTail % CLAMP_FLOAT_MASK; - params.selectTailOffset = params.vcmpvsRepeat * CLAMP_FLOAT_MASK; - } - params.selectTailRepeatLoop = params.vcmpvsRepeat / MAX_REPEAT_TIMES; - params.selectTailRepeatTail = params.vcmpvsRepeat % MAX_REPEAT_TIMES; - params.selectTailPreRepeatOffset = MAX_REPEAT_TIMES * params.ClampMask; - params.selectTailMainRepeatOffset = params.selectTailRepeatLoop * MAX_REPEAT_TIMES * params.ClampMask; - params.selectTailTailRepeatOffset = params.selectTailRepeatLoop * MAX_REPEAT_TIMES * params.ClampMask + - params.selectTailRepeatTail * params.ClampMask; -} - -template +template __aicore__ inline void ClampCompute(const LocalTensor& dstTensor, const LocalTensor& srcTensor, - const LocalTensor& sharedTmpBuffer, const T scalar, const uint32_t calCount, CLAMPMODE selMode, - ClampParams& params) + const LocalTensor& sharedTmpBuffer, const T scalar, const uint32_t calCount, CLAMPMODE selMode) { - CheckTensorPosition(dstTensor, "dstTensor", "VECIN, VECOUT, VECCALC"); - CheckTensorPosition(srcTensor, "srcTensor", "VECIN, VECOUT, VECCALC"); - CheckTensorPosition(sharedTmpBuffer, "sharedTmpBuffer", "VECIN, VECOUT, VECCALC"); - - CheckCalCount(calCount, "calCount", srcTensor, "srcTensor", "Clamp"); - CheckCalCount(calCount, "calCount", dstTensor, "dstTensor", "Clamp"); - - ASCENDC_ASSERT((std::is_same::value || std::is_same::value), { - KERNEL_LOG( KERNEL_ERROR, "Failed to check the data types, current api support data types are half/float."); }); - uint32_t sharedTmpBufferSize = sharedTmpBuffer.GetSize(); - params.clampSplitCount = sharedTmpBufferSize * sizeof(uint8_t) / sizeof(uint8_t); - // split the input based on the stack buffer. - GetMainParams(calCount, params, sharedTmpBufferSize); - BinaryRepeatParams vselRepeatParams; - // clampSplitCount is full mask aligned. The loop in loopCount does not have a tail block. - for (uint32_t i = 0; i < params.loopCount; i++) { - for (uint32_t j = 0; j < params.vcmpvsRepeatLoop; j++) { - ClampComputeCount(dstTensor[i * params.clampSplitCount + j * params.vcmpvsPreRepeatOffset], - srcTensor[i * params.clampSplitCount + j * params.vcmpvsPreRepeatOffset], sharedTmpBuffer, scalar, - MAX_REPEAT_TIMES, params.ClampMask, selMode, vselRepeatParams); - } - if (params.vcmpvsRepeatTail) { - ClampComputeCount(dstTensor[i * params.clampSplitCount + params.vcmpvsMainRepeatOffset], - srcTensor[i * params.clampSplitCount + params.vcmpvsMainRepeatOffset], sharedTmpBuffer, scalar, - params.vcmpvsRepeatTail, params.ClampMask, selMode, vselRepeatParams); - } - } - // calcTail is a tail block smaller than clampSplitCount： - // 1. calcTail < 128 , vcmpvs_lt calculate 128 element, select mask=calcTail； - // 2. 128 <= calcTail < clampSplitCount： - // 1) main, vcmpvs_lt calculate 128 element, select mask=128； - // 2) tail, vcmpvs_lt calculate 128 element, select mask=tailElement； - uint32_t mainCount = params.loopCount * params.clampSplitCount; - if (params.calcTail > 0) { - GetTailParams(params.calcTail, params); - for (uint32_t j = 0; j < params.selectTailRepeatLoop; j++) { - ClampComputeCount(dstTensor[mainCount + j * params.selectTailPreRepeatOffset], - srcTensor[mainCount + j * params.selectTailPreRepeatOffset], sharedTmpBuffer, scalar, MAX_REPEAT_TIMES, - params.ClampMask, selMode, vselRepeatParams); - } - if (params.selectTailRepeatTail) { - ClampComputeCount(dstTensor[mainCount + params.selectTailMainRepeatOffset], - srcTensor[mainCount + params.selectTailMainRepeatOffset], sharedTmpBuffer, scalar, - params.selectTailRepeatTail, params.ClampMask, selMode, vselRepeatParams); - } - if (params.selectTailElement) { - ClampComputeCount(dstTensor[mainCount + params.selectTailTailRepeatOffset], - srcTensor[mainCount + params.selectTailTailRepeatOffset], sharedTmpBuffer, scalar, 1, - params.selectTailElement, selMode, vselRepeatParams); - } + if (selMode == CLAMPMODE::CLAMP_MIN) { + CHECK_FUNC_HIGHLEVEL_API(ClampMin, (T, isReuseSource), (dstTensor, srcTensor, sharedTmpBuffer, scalar, calCount)); + Maxs(dstTensor, srcTensor, scalar, calCount); + } else if (selMode == CLAMPMODE::CLAMP_MAX) { + CHECK_FUNC_HIGHLEVEL_API(ClampMax, (T, isReuseSource), (dstTensor, srcTensor, sharedTmpBuffer, scalar, calCount)); + Mins(dstTensor, srcTensor, scalar, calCount); } } /* ************************************************************************************************** @@ -170,8 +39,7 @@ template __aicore__ inline void ClampMaxImpl(const LocalTensor& dstTensor, const LocalTensor& srcTensor, const LocalTensor& sharedTmpBuffer, const T scalar, const uint32_t calCount) { - ClampParams params; - ClampCompute(dstTensor, srcTensor, sharedTmpBuffer, scalar, calCount, CLAMPMODE::CLAMP_MAX, params); + ClampCompute(dstTensor, srcTensor, sharedTmpBuffer, scalar, calCount, CLAMPMODE::CLAMP_MAX); } /* ************************************************************************************************** @@ -182,8 +50,7 @@ template __aicore__ inline void ClampMinImpl(const LocalTensor& dstTensor, const LocalTensor& srcTensor, const LocalTensor& sharedTmpBuffer, const T scalar, const uint32_t calCount) { - ClampParams params; - ClampCompute(dstTensor, srcTensor, sharedTmpBuffer, scalar, calCount, CLAMPMODE::CLAMP_MIN, params); + ClampCompute(dstTensor, srcTensor, sharedTmpBuffer, scalar, calCount, CLAMPMODE::CLAMP_MIN); } } // namespace AscendC #endif // IMPL_MATH_CLAMP_CLAMP_COMMON_IMPL_H \ No newline at end of file -- Gitee