随机数生成（Random）
除了可以下发Kernel执行任务外，Runtime还提供下发Reduce和随机数生成的内置系统任务的功能。（系统任务区别于Kernel任务在于无需用户提供执行代码）。 系统任务可以下发到某条Stream异步执行，同样遵循同一流上任务保序执行的规则。
通过aclrtReduceAsync接口可以下发执行Reduce操作任务，调用代码示例如下：
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 43 44 | aclInit(NULL); aclrtSetDevice(0); aclrtStream stream; aclrtCreateStream(&stream); // 准备 Host 数据 const int count = 4; float hostInput[4] = {1.0, 2.0, 3.0, 4.0}; float hostOutput[4] = {0, 0, 0, 0}; size_t size = count * sizeof(float); // 申请 Device 内存 void *devInput = NULL; void *devOutput = NULL; aclrtMalloc(&devInput, size, ACL_MEM_MALLOC_HUGE_FIRST); aclrtMalloc(&devOutput, size, ACL_MEM_MALLOC_HUGE_FIRST); // 拷贝数据到 Device aclrtMemcpy(devInput, size, hostInput, size, ACL_MEMCPY_HOST_TO_DEVICE); aclrtMemcpy(devOutput, size, hostInput, size, ACL_MEMCPY_HOST_TO_DEVICE); // 调用 aclrtReduceAsync aclrtReduceAsync(devOutput, devInput, size, ACL_RT_MEMCPY_SDMA_AUTOMATIC_SUM, // 归约类型 ACL_FLOAT, // 数据类型 stream, NULL); // 同步 stream aclrtSynchronizeStream(stream); // 拷回结果 aclrtMemcpy(hostOutput, size, devOutput, size, ACL_MEMCPY_DEVICE_TO_HOST); for (int i = 0; i < count; i++) { printf("Reduce SUM result[%d] = %f\n", i, hostOutput[i]); } /* 预期如下结果 Reduce SUM result[0] = 2.000000 Reduce SUM result[1] = 4.000000 Reduce SUM result[2] = 6.000000 Reduce SUM result[3] = 8.000000 */ // 释放资源 aclrtFree(devInput); aclrtFree(devOutput); aclrtDestroyStream(stream); aclrtResetDeviceForce(0); aclFinalize(); |
通过aclrtRandomNumAsync执行随机数生成任务，调用代码示例如下：
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 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 | aclError NormalFloatAsync( float mean, float stddev, uint64_t seed, uint64_t num, void *counterDevAddr, void *devOutput, aclrtStream stream) { aclrtRandomNumTaskInfo taskInfo = {}; taskInfo.dataType = ACL_FLOAT; taskInfo.randomNumFuncParaInfo.funcType = ACL_RT_RANDOM_NUM_FUNC_TYPE_NORMAL_DIS; taskInfo.randomParaAddr = NULL; taskInfo.randomCounterAddr = counterDevAddr; taskInfo.randomResultAddr = devOutput; memcpy(taskInfo.randomNumFuncParaInfo.paramInfo.normalDisInfo.mean.valueOrAddr, &mean, sizeof(float)); taskInfo.randomNumFuncParaInfo.paramInfo.normalDisInfo.mean.size = sizeof(float); taskInfo.randomNumFuncParaInfo.paramInfo.normalDisInfo.mean.isAddr = 0; memcpy(taskInfo.randomNumFuncParaInfo.paramInfo.normalDisInfo.stddev.valueOrAddr, &stddev, sizeof(float)); taskInfo.randomNumFuncParaInfo.paramInfo.normalDisInfo.stddev.size = sizeof(float); taskInfo.randomNumFuncParaInfo.paramInfo.normalDisInfo.stddev.isAddr = 0; memcpy(taskInfo.randomSeed.valueOrAddr, &seed, sizeof(uint64_t)); taskInfo.randomSeed.size = sizeof(uint64_t); taskInfo.randomSeed.isAddr = 0; memcpy(taskInfo.randomNum.valueOrAddr, &num, sizeof(uint64_t)); taskInfo.randomNum.size = sizeof(uint64_t); taskInfo.randomNum.isAddr = 0; return aclrtRandomNumAsync(&taskInfo, stream, NULL); } int main() { aclError ret; // 初始化 ACL ret = aclInit(NULL); ret = aclrtSetDevice(0); aclrtStream stream; ret = aclrtCreateStream(&stream); uint64_t num = 128; size_t size = num * sizeof(uint64_t); // 申请足够大内存 // 申请 Device 内存 void *devOutput = NULL; ret = aclrtMalloc(&devOutput, size, ACL_MEM_MALLOC_NORMAL_ONLY); // 准备 Host 数据 void *hostOutput = malloc(size); // 申请存放随机数状态 counter 的device内存 （要求 16Byte) void *counterAddr = NULL; ret = aclrtMalloc((void **)&counterAddr, 16, ACL_MEM_MALLOC_NORMAL_ONLY); float mean = 3.0; float stddev = 2.0; ret =NormalFloatAsync(mean, stddev, 0, num, counterAddr, devOutput, stream); // 同步 stream aclrtSynchronizeStream(stream); // 拷回结果 aclrtMemcpy(hostOutput, size, devOutput, size, ACL_MEMCPY_DEVICE_TO_HOST); // 释放资源 free(hostOutput); aclrtFree(devOutput); aclrtFree(counterAddr); aclrtDestroyStream(stream); aclrtResetDeviceForce(0); aclFinalize(); return 0; } |
父主题： 专用加速器