MulAddDst【免费下载链接】asc-devkit本项目是CANN 推出的昇腾AI处理器专用的算子程序开发语言原生支持C和C标准规范主要由类库和语言扩展层构成提供多层级API满足多维场景算子开发诉求。项目地址: https://gitcode.com/cann/asc-devkit产品支持情况Ascend 950PR/Ascend 950DT支持Atlas A3 训练系列产品/Atlas A3 推理系列产品支持Atlas A2 训练系列产品/Atlas A2 推理系列产品支持Atlas 200I/500 A2 推理产品支持Atlas 推理系列产品 AI Core支持Atlas 推理系列产品 Vector Core不支持Atlas 训练系列产品不支持Kirin X90支持Kirin 9030支持功能说明头文件路径为basic_api/kernel_operator_vec_binary_intf.h。按元素将src0和src1相乘并和dst相加将最终结果存放进dst中。计算公式如下$$ dst_j (src0_i * src1_i) dst_i $$函数原型tensor前n个数据计算template typename T, typename U __aicore__ inline void MulAddDst(const LocalTensorT dst, const LocalTensorU src0, const LocalTensorU src1, const int32_t count)tensor高维切分计算mask逐bit模式template typename T, typename U, bool isSetMask true __aicore__ inline void MulAddDst(const LocalTensorT dst, const LocalTensorU src0,const LocalTensorU src1, const uint64_t mask[], const uint8_t repeatTime, const BinaryRepeatParams repeatParams)mask连续模式template typename T, typename U, bool isSetMask true __aicore__ inline void MulAddDst(const LocalTensorT dst, const LocalTensorU src0,const LocalTensorU src1, uint64_t mask, const uint8_t repeatTime, const BinaryRepeatParams repeatParams)参数说明表模板参数说明参数名描述T目的操作数数据类型。目的操作数和源操作数的数据类型约束请参考数据类型约束。U源操作数数据类型。isSetMask是否在接口内部设置mask。• true表示在接口内部设置mask。• false表示在接口外部设置mask开发者需要使用SetVectorMask接口设置mask值。这种模式下接口入参中的mask值设置为占位符MASK_PLACEHOLDER用于占位无实际含义。表参数说明参数名输入/输出描述dst输出目的操作数。类型为LocalTensor支持的TPosition为VECIN/VECCALC/VECOUT。src0、src1输入源操作数。源操作数数据类型和目的操作数数据类型可以不一致。类型为LocalTensor支持的TPosition为VECIN/VECCALC/VECOUT。count输入参与计算的元素个数。关于该参数的具体说明请参考连续计算。mask[]/mask输入mask用于控制每次迭代内参与计算的元素。设置详见掩码操作。repeatTime输入重复迭代次数。矢量计算单元每次读取连续的256Bytes数据进行计算为完成对输入数据的处理必须通过多次迭代repeat才能完成所有数据的读取与计算。repeatTime表示迭代的次数。关于该参数的具体描述请参考高维切分。repeatParams输入控制操作数地址步长的参数。BinaryRepeatParams类型包含操作数相邻迭代间相同datablock的地址步长操作数同一迭代内不同datablock的地址步长等参数。相邻迭代间的地址步长参数说明请参考repeatStride同一迭代内DataBlock的地址步长参数说明请参考dataBlockStride。数据类型PAR列表示矢量计算单元一个迭代能够处理的元素个数。表Ascend 950PR/Ascend 950DT支持的数据类型约束src0数据类型src1数据类型dst数据类型PARhalfhalfhalf128floatfloatfloat64halfhalffloat64int64_tint64_tint64_t32uint64_tuint64_tuint64_t32针对Ascend 950PR/Ascend 950DTint64_t、uint64_t数据类型仅支持tensor前n个数据计算接口。表Atlas A3 训练系列产品/Atlas A3 推理系列产品支持的数据类型约束src0数据类型src1数据类型dst数据类型PARhalfhalfhalf128floatfloatfloat64halfhalffloat64表Atlas A2 训练系列产品/Atlas A2 推理系列产品支持的数据类型约束src0数据类型src1数据类型dst数据类型PARhalfhalfhalf128floatfloatfloat64halfhalffloat64表Atlas 200I/500 A2 推理产品支持的数据类型约束src0数据类型src1数据类型dst数据类型PARhalfhalfhalf128floatfloatfloat64int16_tint16_tint16_t128uint16_tuint16_tuint16_t128int32_tint32_tint32_t64uint32_tuint32_tuint32_t64表Atlas 推理系列产品AI Core支持的数据类型约束src0数据类型src1数据类型dst数据类型PARhalfhalfhalf128floatfloatfloat64halfhalffloat64表Kirin X90支持的数据类型约束src0数据类型src1数据类型dst数据类型PARhalfhalfhalf128floatfloatfloat64halfhalffloat64表Kirin 9030支持的数据类型约束src0数据类型src1数据类型dst数据类型PARhalfhalfhalf128floatfloatfloat64halfhalffloat64返回值说明无约束说明操作数地址对齐要求请参见通用地址对齐约束。操作数地址重叠约束请参考通用地址重叠约束。源操作数的数据类型为half、目的操作数的数据类型为float的情况下不支持地址重叠。当参数count或repeatTime取值为0时该接口的行为如下针对如下型号该接口不会执行计算操作不会对目的操作数进行写入该接口将被视为NOP空操作。Atlas A3 训练系列产品/Atlas A3 推理系列产品Atlas A2 训练系列产品/Atlas A2 推理系列产品针对Ascend 950PR/Ascend 950DT该接口通过VF调用Reg矢量计算API实现兼容当参数count或repeatTime取值为0时不保证该接口将被视为NOP空操作。MulAddDst指令实际使用会受到bank冲突影响。地址不重叠场景下无法在一拍读取dst、src0、src1三块不同地址下的数据因此只能达到一半的理论并行度理论并行度将在原有基础上减半在地址重叠场景下则保持原有理论并行度。使用tensor高维切分计算接口时src和scalar的数据类型为half、dst的数据类型为float的情况下一个迭代处理内最多处理64个输入数据。对UB空间的占用说明。针对Ascend 950PR/Ascend 950DTtensor高维切分计算接口占用8KB Unified Buffer临时空间。tensor前n个数据连续计算接口不涉及8KB Unified Buffer临时空间的占用。调用示例高维切分计算接口样例-mask连续模式源操作数的数据类型为half、目的操作数的数据类型为floatuint64_t mask 64; // repeatTime 4一次迭代计算64个数共计算256个数 // dstBlkStride, src0BlkStride, src1BlkStride 1单次迭代内数据连续读取和写入 // dstRepStride 8src0RepStride, src1RepStride 4相邻迭代间数据连续读取和写入 AscendC::MulAddDst(dstLocal, src0Local, src1Local, mask, 4, { 1, 1, 1, 8, 4, 4 });高维切分计算接口样例-mask逐bit模式源操作数的数据类型为half、目的操作数的数据类型为floatuint64_t mask[2] { UINT64_MAX, 0 }; // repeatTime 4一次迭代计算64个数共计算256个数 // dstBlkStride, src0BlkStride, src1BlkStride 1单次迭代内数据连续读取和写入 // dstRepStride 8src0RepStride, src1RepStride 4相邻迭代间数据连续读取和写入 AscendC::MulAddDst(dstLocal, src0Local, src1Local, mask, 4, { 1, 1, 1, 8, 4, 4 });tensor前n个数据计算样例源操作数的数据类型为half、目的操作数的数据类型为floatAscendC::MulAddDst(dstLocal, src0Local, src1Local, 256);结果示例如下输入数据src0Local [-83. 58.2 -14.28 -43.12 20.72 -79.9 54.16 31.56 -1.464 68.25 -28.31 -93.5 -4.2 -46.56 -22.23 78.5 -69.56 -37.03 -53.12 58.28 -71.56 -34.44 85.94 96.3 66.06 99.94 -45.94 8.75 -93.9 35.56 82.56 -70.8 -68.75 -35.4 95.3 -49.1 -56.34 86.75 90.25 24.17 79.06 -49.66 -95.3 -6.965 -63.72 -33.16 -15.56 -43.28 51.28 40.1 83.25 49.72 55.47 -53.7 17.55 -36.06 63. 59.16 -66.8 -9.01 25.56 44.28 22.12 -33.84 -31.9 -74.2 79.94 -34.94 1.119 18.45 -92.75 -83.25 42.66 -77.6 33.28 0.709 -19.3 44.44 45.28 -33.4 -55.94 -42.22 -37.72 39.4 87.25 23.19 34.16 51.3 -22.16 15.234 59. -20.45 -63.9 41.84 -14.63 -80.94 47.8 -36.84 8.47 -60.66 -26.06 -42.78 30.5 -91.3 55.84 -85.44 -99.44 68.2 -71.7 27.45 -11.48 -48.03 71. 71.5 -59.2 14.67 79.25 32.7 -54.22 6.17 -69.94 -49.22 87.7 -61.53 36.25 -57.84 -81.75 -24.84 -35. -62.44 -47.22 19.95 21.16 -31.56 13.38 72.4 -64.06 -89.75 -28.17 34.4 -68.06 -46.94 16.06 65.56 3.16 -59.88 -32.97 30.69 89.5 16.66 25.05 -1.988 5.27 -23.14 -26.89 -24.72 1.427 -14.46 81.9 -59.94 68.7 -83.2 -75.44 88.6 27.62 -58.06 -36.1 -49.53 27.73 89.5 -51.5 90. 67.94 -70.8 24.2 -75.8 -96.75 -22.66 33.03 6.293 -87.5 36.56 36.06 -76.8 1.786 82.9 87.6 -63.94 -4.51 -89.06 -56.06 75.2 -31.89 27.44 35.22 -27.19 37.53 96.94 -83.25 -49.6 31.78 -50.25 65.2 69.9 63.03 53. -70.1 -57.22 -11.99 -23.14 44.28 -77.3 77.25 10.805 16.3 -96.6 -94.9 34.1 -40.25 -99.7 -6.156 44.97 82.7 51.1 -53.28 85.44 -80.94 -47. -53.47 -35.22 76.75 -28.38 26.48 -67.06 34.28 -54.6 21.52 -38.9 79.75 51.7 -39.44 48.56 -91.7 -44.06 92.9 11.79 8.98 -5.074 12.375 -24.77 -27.31 76.2 39.8 -5.46 25.17 47. ] 输入数据src1Local [-57.97 43.5 8.08 72.4 -81.44 -52. 69.1 -84.25 31.12 34.34 74.75 83.56 -83. 80.1 42.84 -31.6 88.56 47.34 18.89 -95.25 16.88 -85.75 76.75 -17.19 23.39 92.56 22.81 77.94 38.62 -55.8 38.22 -88.6 -99.4 -66.75 90.44 80.56 12.78 -12.6 -68.4 2.816 27.45 -60.88 70. 61.78 -90.56 -99.25 38.25 -14.49 -35.88 38.1 13. 29.22 -57.06 -44.7 6.535 -44.6 -76.3 91.7 36.66 83.9 66. -81.25 -50.06 68. 2.705 -51.72 66.9 49.03 15.76 9.37 33.2 99.56 -20.55 83.3 -57.1 37.06 68.94 -91.9 -46.06 -92.7 64.4 8.164 8.98 10.76 -75.6 26.94 46.8 62. 8.734 -69.25 -70.2 -59. 67.25 87.6 48.72 60.16 19.39 48.62 21.64 25.06 1.013 -36.6 -46.28 -29.14 67.44 56.7 32.03 -28.81 -94.44 49.6 0.583 -84.4 -51.53 -43. 66. -68. 77.44 -50.16 -90.4 -46.22 90.25 88. 79.25 -40.84 -71.7 -27.03 19.53 85.44 45.06 60.72 19.22 -28.95 -47.72 97.8 -51.6 31.42 31.75 -21.84 -71.4 77.9 43.12 35.66 -50.84 -52. -48.84 -53.97 -59.56 31.2 -64.3 -10.47 86.25 -84.44 -56.4 -63.03 -99.9 54.44 40.72 74.94 8.305 18.52 -47.34 -74.06 79.1 92.44 84.94 -98.7 -41.06 -80.2 -71.06 89.06 96.2 -19.83 -51.03 -92. 82.25 -75.75 58.66 22.72 -89.06 -83.06 -73.5 18.75 -0.939 -96.4 50.12 -73.9 -56.97 52.34 -95.56 11.02 -46.3 -52.2 -8.46 80.56 77. -51.72 38.8 -66.44 -69. -30.33 -53.3 5.406 74.8 52.25 -35.88 92.5 51.38 40.47 43.94 -29.05 89.7 -74.5 -83.5 81.75 -56.6 -13.625 86.9 -4.58 -67.5 -6.67 -59.53 -30.4 -91.75 -84.3 -66.6 -28.61 -13.79 -70.75 -90.2 -47.94 59.56 84.2 0.7085 -57.44 -24.94 -11.875 -90.4 54.22 -44.16 -36.34 -31.64 72.1 -81.25 75.8 93.9 -28.28 -20.53 90.2 -58.97 -95.7 59.22 -37.8 94.9 -86.7 36.16 26.47 ] 输入数据dstLocal [-97.94773 -61.303955 32.56878 -87.50743 -78.92147 59.20739 50.336506 49.039738 -76.2525 0.25441223 -71.73807 6.481831 -55.5052 -51.057415 31.403702 63.285076 98.1897 86.71727 -50.16466 88.94256 72.111435 8.4164915 34.524082 73.14016 4.838548 69.67902 -97.855736 90.358696 9.051491 37.595695 -66.01661 -97.110634 82.84477 69.46122 25.561102 47.926853 -10.202202 78.2545 31.339691 12.940468 -31.499294 -3.351652 62.46355 45.0427 -86.02812 -43.48385 -62.274956 -36.077827 51.81446 32.47797 59.10228 68.18655 9.3604145 -76.47674 -50.29268 94.496346 30.837933 -48.315712 -44.92399 -62.369625 47.578724 84.84092 -66.64584 88.376434 95.05615 -92.37309 3.0038757 85.21814 -6.688882 97.74142 20.733965 -5.62451 69.6166 -64.435455 94.09325 -63.13334 89.150345 -17.61865 32.776333 27.28345 31.288876 -9.983517 -46.39662 -37.025536 47.853374 -30.384796 -79.801544 -11.131944 -36.417023 84.25002 -74.19904 -86.72338 -6.5878353 26.253004 -28.112898 -64.88305 -40.56897 -65.849686 22.276798 -3.356709 -78.41364 -67.26924 -10.346288 -43.172684 10.149812 -22.575602 -28.780804 -64.24396 -14.579756 -30.369322 -59.28742 -37.098255 31.078829 29.901808 50.531147 -88.35735 -45.65366 -6.7495203 6.8026304 56.172153 -0.8727364 9.618746 89.294815 75.4403 81.63827 -61.722088 -72.85743 9.296161 -69.17855 2.3497865 20.234892 -13.279363 -44.531677 55.188084 -45.736256 -30.018398 27.09971 28.841034 35.764072 21.457811 -15.206495 94.05271 79.9942 -36.39198 38.40136 5.2365685 -11.435508 67.15551 87.03286 7.9285994 78.32062 97.863335 -28.68556 -72.658554 -79.39075 -82.65206 39.52689 -22.053177 30.602457 -26.158005 49.83525 -72.24563 -97.10148 54.803936 65.070786 -57.019573 35.972733 6.694148 -74.88097 -71.13884 -84.549545 -26.875593 -3.2775877 -8.592472 -5.248627 -22.2127 98.26377 -51.741936 -69.48398 -47.230175 92.72371 18.192408 -39.66745 44.556633 -21.733562 15.191482 5.9535656 41.23602 89.30139 -32.57541 -47.595608 -50.371124 -87.899666 57.644466 38.85747 47.65093 49.42874 -32.424126 -22.5012 78.78245 -70.6598 -87.218544 50.347565 55.945244 -3.4658287 17.902784 -30.977674 53.424767 -82.00753 2.9060571 -1.010124 -94.316765 13.186674 -52.089214 58.975357 48.281635 26.436571 -27.11565 89.21593 -10.962796 49.347828 21.556795 78.163956 35.06028 10.803711 53.231297 -44.78757 -0.6473386 26.717777 63.757347 -4.90904 21.724916 37.443634 -89.250656 62.98874 72.13095 -12.19138 84.16487 71.54008 -73.41178 -97.612564 39.947853 -1.3887504 -5.6196795 -54.509125 -28.877354 26.259935 42.28702 -38.848114 -76.46558 -91.69401 71.27111 89.36143 -65.70425 -31.810083 82.811226 ] 输出数据dstLocal [ 4.71345850e03 2.46985229e03 -8.27969437e01 -3.20867920e03 -1.76620471e03 4.21270752e03 3.79388721e03 -2.61010083e03 -1.21815369e02 2.34421533e03 -2.18809741e03 -7.80661182e03 2.93029968e02 -3.78187769e03 -9.21200317e02 -2.41682422e03 -6.06243945e03 -1.66648096e03 -1.05372913e03 -5.46234668e03 -1.13550574e03 2.96143213e03 6.63022705e03 -1.58223096e03 1.55008167e03 9.32014355e03 -1.14580493e03 7.72311829e02 -3.61687036e03 -1.94723633e03 3.08941895e03 6.17864697e03 6.91487598e03 2.43282837e03 8.64538574e03 -3.90718848e03 -7.30345764e02 -1.01493103e03 -6.13950391e03 8.10182877e01 2.13901343e03 3.01947266e03 -6.60941162e03 -3.85254059e02 5.68450098e03 3.24727393e03 -6.57540588e02 5.91162170e02 -1.78790039e03 1.55979932e03 1.14135229e03 1.52090625e03 -3.15582520e03 2.32268335e03 6.43788910e01 1.70265845e03 -4.77684961e03 5.37557275e03 -2.49401978e03 -8.17899902e02 1.73470374e03 -3.51301074e03 -1.17427869e03 -2.21299854e03 8.74725342e00 3.74451172e03 5.34882422e03 -1.62781116e03 1.09463263e01 2.70595306e02 -3.05740674e03 -8.29420215e03 -8.06836060e02 -6.53156836e03 -1.80605811e03 -3.68566055e01 -1.24112793e03 -4.10031396e03 -2.05299121e03 3.12362524e03 -3.56968774e03 -3.54660034e02 -3.84981323e02 3.86899506e02 -6.55042773e03 5.94228516e02 1.51913794e03 3.17024316e03 -2.29938019e02 -9.70730469e02 -4.21526172e03 1.12001099e03 -4.30428320e03 3.69281152e03 -7.41005249e02 -4.93377930e03 8.86545288e02 -1.85737708e03 2.05545837e02 -1.52355396e03 -1.04807014e02 1.49825708e03 -1.42192444e03 2.61773071e03 3.77611279e03 -4.86581396e03 -3.21388818e03 -2.02889624e03 6.75540869e03 1.33113416e03 -6.59783478e01 4.01553857e03 -3.62763989e03 -3.04459814e03 -3.85584375e03 -1.08604480e03 6.09126807e03 -1.64623193e03 4.90682227e03 -2.29084198e02 -6.31273193e03 -4.32163135e03 7.03852930e03 2.58860718e03 -2.51703369e03 1.50186682e03 -1.66953711e03 -2.11329175e03 -1.64636609e03 -3.78877710e03 -8.87250488e02 -5.90984680e02 -1.05408154e03 -3.03201904e03 -7.36205750e02 2.24413989e03 -2.00688464e03 1.98931763e03 2.04653162e03 2.70084448e03 -2.95040186e03 -1.57956250e03 -7.36683533e02 -3.44564209e03 -1.15952522e02 3.23661548e03 1.95226562e03 1.02470142e03 -5.66893604e03 -1.66441513e02 2.23861353e03 2.65748840e02 -3.25920044e02 1.38592395e03 2.60631055e03 -1.42827905e03 9.76226807e01 -1.10571973e03 7.10548767e02 -1.13593823e03 -3.20208862e03 6.08882861e03 -6.06609375e03 8.24707715e03 2.41146924e03 5.67302344e03 1.51807239e03 3.97848120e03 -2.04575500e03 7.89995508e03 -5.03820557e03 -1.81140686e03 -3.47021313e03 6.50615771e03 1.98545837e03 5.72058398e03 -5.57672852e03 -5.66463623e02 -3.01132959e03 -5.69939880e02 6.52397363e03 7.03739258e02 -7.35288696e01 7.44736133e03 6.77963409e01 -6.10719922e03 -4.98593311e03 -3.30549243e03 5.20452515e02 -1.01435034e03 2.54879883e03 -3.97421875e03 1.81924927e02 2.26807812e03 2.75070117e03 1.45375439e03 1.50611035e03 -6.47270947e03 5.72174902e03 1.58286792e03 -1.76499768e03 -3.58882599e02 4.92718750e03 3.70691406e03 -2.26471191e03 4.92040283e03 -3.63364966e03 -2.26214648e03 -6.08914246e02 6.75068909e02 3.97046460e03 5.66546436e03 -6.43718848e03 8.31193970e02 -8.63325928e02 1.36479724e03 -8.21582910e03 -1.83201096e02 2.80609082e03 6.54139771e02 4.15837097e02 -1.34577429e03 -7.50841406e03 -4.27278174e03 3.56066699e03 -2.39108228e03 1.07126465e03 3.32460278e03 4.84893066e03 1.75205615e03 4.56651270e03 -2.36709546e03 5.62077103e01 3.76265161e03 -7.91899902e02 7.20431763e02 -1.95666602e03 -2.02528333e03 -3.44992090e03 -1.95192932e03 1.15021460e03 3.54251807e03 7.44822070e03 -3.34610791e03 8.66413184e03 -3.62286774e02 -1.58040115e02 -4.15344086e02 -7.68586426e02 2.29329517e03 -1.70910608e03 -2.80956885e03 3.86657227e03 4.07690765e02 8.78310547e02 1.32684253e03]【免费下载链接】asc-devkit本项目是CANN 推出的昇腾AI处理器专用的算子程序开发语言原生支持C和C标准规范主要由类库和语言扩展层构成提供多层级API满足多维场景算子开发诉求。项目地址: https://gitcode.com/cann/asc-devkit创作声明：本文部分内容由AI辅助生成（AIGC），仅供参考