产品 |
是否支持 |
|---|---|
Ascend 950PR/Ascend 950DT |
√ |
√ |
|
√ |
|
x |
|
√ |
|
x |
|
x |
为Select不传入mask参数的接口设置比较寄存器。配合不同的selMode传入不同的数据。
1 2 | template <typename T> __aicore__ inline void SetCmpMask(const LocalTensor<T>& src) |
参数名 |
描述 |
|---|---|
T |
操作数的数据类型。 |
参数名 |
输入/输出 |
描述 |
|---|---|---|
src |
输入 |
类型为LocalTensor，支持的TPosition为VECIN/VECCALC/VECOUT。 LocalTensor的起始地址需要16字节对齐。 |
无
无
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 | uint32_t dataSize = 256; uint32_t selDataSize = 8; TPipe pipe; TQue<TPosition::VECIN, 1> inQueueX; TQue<TPosition::VECIN, 1> inQueueY; TQue<TPosition::VECIN, 1> inQueueSel; TQue<TPosition::VECOUT, 1> outQueue; pipe.InitBuffer(inQueueX, 1, dataSize * sizeof(float)); pipe.InitBuffer(inQueueY, 1, dataSize * sizeof(float)); pipe.InitBuffer(inQueueSel, 1, selDataSize * sizeof(uint8_t)); pipe.InitBuffer(outQueue, 1, dataSize * sizeof(float)); AscendC::LocalTensor<float> dst = outQueue.AllocTensor<float>(); AscendC::LocalTensor<uint8_t> sel = inQueueSel.AllocTensor<uint8_t>(); AscendC::LocalTensor<float> src0 = inQueueX.AllocTensor<float>(); AscendC::LocalTensor<float> src1 = inQueueY.AllocTensor<float>(); uint8_t repeat = 4; uint32_t mask = 64; AscendC::BinaryRepeatParams repeatParams = { 1, 1, 1, 8, 8, 8 }; // selMode为模式0（SELMODE::VSEL_CMPMASK_SPR） AscendC::SetCmpMask(sel); AscendC::PipeBarrier<PIPE_V>(); AscendC::SetVectorMask<float>(mask); AscendC::Select<float, AscendC::SELMODE::VSEL_CMPMASK_SPR>(dst, src0, src1, repeat, repeatParams); // selMode为模式2（SELMODE::VSEL_TENSOR_TENSOR_MODE） AscendC::LocalTensor<int32_t> tempBuf; #if defined(ASCENDC_CPU_DEBUG) && (ASCENDC_CPU_DEBUG == 1) // cpu调试 tempBuf.ReinterpretCast<int64_t>().SetValue(0, reinterpret_cast<int64_t>(reinterpret_cast<__ubuf__ int64_t*>(sel.GetPhyAddr()))); event_t eventIdSToV = static_cast<event_t>(AscendC::GetTPipePtr()->FetchEventID(AscendC::HardEvent::S_V)); AscendC::SetFlag<AscendC::HardEvent::S_V>(eventIdSToV); AscendC::WaitFlag<AscendC::HardEvent::S_V>(eventIdSToV); #else // npu调试 uint32_t selAddr = static_cast<uint32_t>(reinterpret_cast<int64_t>(reinterpret_cast<__ubuf__ int64_t*>(sel.GetPhyAddr()))); AscendC::SetVectorMask<uint32_t>(32); AscendC::Duplicate<uint32_t, false>(tempBuf.ReinterpretCast<uint32_t>(), selAddr, AscendC::MASK_PLACEHOLDER, 1, 1, 8); AscendC::PipeBarrier<PIPE_V>(); #endif AscendC::SetCmpMask<int64_t>(tempBuf.ReinterpretCast<int64_t>()); AscendC::PipeBarrier<PIPE_V>(); AscendC::SetVectorMask<float>(mask); AscendC::Select<float, AscendC::SELMODE::VSEL_TENSOR_TENSOR_MODE>(dst, src0, src1, repeat, repeatParams); |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 | uint32_t dataSize = 256; uint32_t selDataSize = 8; TPipe pipe; TQue<TPosition::VECIN, 1> inQueueX; TQue<TPosition::VECIN, 1> inQueueY; TQue<TPosition::VECIN, 1> inQueueSel; TQue<TPosition::VECOUT, 1> outQueue; pipe.InitBuffer(inQueueX, 1, dataSize * sizeof(float)); pipe.InitBuffer(inQueueY, 1, dataSize * sizeof(float)); pipe.InitBuffer(inQueueSel, 1, selDataSize * sizeof(uint8_t)); pipe.InitBuffer(outQueue, 1, dataSize * sizeof(float)); AscendC::LocalTensor<float> dst = outQueue.AllocTensor<float>(); AscendC::LocalTensor<uint8_t> sel = inQueueSel.AllocTensor<uint8_t>(); AscendC::LocalTensor<float> src0 = inQueueX.AllocTensor<float>(); AscendC::LocalTensor<float> tmpScalar = inQueueY.AllocTensor<float>(); uint8_t repeat = 4; uint32_t mask = 64; AscendC::BinaryRepeatParams repeatParams = { 1, 1, 1, 8, 8, 8 }; // selMode为模式1（SELMODE::VSEL_TENSOR_SCALAR_MODE） AscendC::SetVectorMask<uint32_t>(32); AscendC::Duplicate<float, false>(tmpScalar, static_cast<float>(1.0), MASK_PLACEHOLDER, 1, 1, 8); AscendC::PipeBarrier<PIPE_V>(); AscendC::SetCmpMask(tmpScalar); AscendC::PipeBarrier<PIPE_V>(); AscendC::SetVectorMask<float>(mask); AscendC::Select(dst, sel, src0, repeat, repeatParams); |