-#include "../../../../../common/data_utils.h"
-
-aclrtStream CreateStream(int device)
-{
- if (aclInit(NULL) != ACL_SUCCESS) {
- printf("acl init failed\n");
- return NULL;
- }
- if (aclrtSetDevice(device) != ACL_SUCCESS) {
- printf("Set device failed\n");
- (void)aclFinalize();
- return NULL;
- }
- aclrtStream stream = nullptr;
- if (aclrtCreateStream(&stream) != ACL_SUCCESS) {
- printf("Create stream failed\n");
- return NULL;
- }
- return stream;
-}
-
-void DestroyStream(aclrtStream stream, int device)
-{
- (void)aclrtDestroyStream(stream);
- if (aclrtResetDevice(device) != ACL_SUCCESS) {
- printf("Reset device failed\n");
- }
- if (aclFinalize() != ACL_SUCCESS) {
- printf("Finalize acl failed\n");
- }
-}
-
-struct tensorInfo {
- int64_t *dims;
- int64_t dimCnt;
- aclDataType dtype;
- aclFormat fmt;
-};
-
-int64_t GetDataSize(struct tensorInfo *desc)
-{
- if (!desc->dims)
- return 0;
- int64_t size = 1;
- for (auto i = 0; i < desc->dimCnt; i++) {
- size *= desc->dims[i];
- }
- return size * sizeof(float);
-}
-
-int64_t CompareResult(void *outputData, int64_t outSize)
-{
- void *goldenData;
- CHECK_ACL(aclrtMallocHost((void **)(&goldenData), outSize));
- size_t goldenSize = outSize;
- bool ret = ReadFile("../output/golden.bin", goldenSize, goldenData, goldenSize);
- if (ret) {
- printf("ReadFile golden success!\n");
- } else {
- return -1;
- }
- constexpr float EPS = 1e-6;
- int64_t wrongNum = 0;
-
- for (int i = 0; i < outSize / sizeof(float); i++) {
- float a = ((float *)outputData)[i];
- float b = ((float *)goldenData)[i];
- float ae = std::abs(a - b);
- float re = ae / abs(b);
- if (ae > EPS && re > EPS) {
- printf("CompareResult failed output is %lf, golden is %lf\n", a, b);
- wrongNum++;
- }
- }
- CHECK_ACL(aclrtFreeHost(goldenData));
- return wrongNum;
-}
-
-int main(int32_t argc, char *argv[])
-{
- aclrtStream stream;
- int64_t inputX[] = {1, 48};
- int64_t outputY[] = {96, 48};
- if (argc == 2) {
- if (!strcmp(argv[1], "1")) {
- inputX[0] = 96;
- inputX[1] = 1;
- outputY[0] = 96;
- outputY[1] = 96;
- }
- }
- struct tensorInfo tensorDesc[] = {
- {inputX, 2, ACL_FLOAT, ACL_FORMAT_ND},
- {outputY, 2, ACL_FLOAT, ACL_FORMAT_ND},
- };
- stream = CreateStream(0);
- aclTensor *tensors[sizeof(tensorDesc) / sizeof(struct tensorInfo)];
- void *devMem[sizeof(tensorDesc) / sizeof(struct tensorInfo)];
- for (auto i = 0; i < sizeof(tensorDesc) / sizeof(struct tensorInfo); i++) {
- void *data;
- struct tensorInfo *info = &(tensorDesc[i]);
- int64_t size = GetDataSize(info);
- if (size == 0) {
- tensors[i] = NULL;
- devMem[i] = NULL;
- continue;
- }
- CHECK_ACL(aclrtMalloc(&data, size, ACL_MEM_MALLOC_HUGE_FIRST));
- // read input
- if (i == 0) {
- size_t inputSize = size;
- void *dataHost;
- CHECK_ACL(aclrtMallocHost((void **)(&dataHost), inputSize));
- ReadFile("../input/input.bin", inputSize, dataHost, inputSize);
- CHECK_ACL(aclrtMemcpy(data, size, dataHost, size, ACL_MEMCPY_HOST_TO_DEVICE));
- CHECK_ACL(aclrtFreeHost(dataHost));
- }
- devMem[i] = data;
- tensors[i] =
- aclCreateTensor(info->dims, info->dimCnt, info->dtype, NULL, 0, info->fmt, info->dims, info->dimCnt, data);
- }
- size_t workspaceSize = 0;
- aclOpExecutor *handle;
- int32_t ret;
- ret = aclnnBroadcastCustomGetWorkspaceSize(tensors[0], tensors[1], &workspaceSize, &handle);
- printf("aclnnBroadcastCustomGetWorkspaceSize ret %u workspace size %lu\n", ret, workspaceSize);
- void *workspace = NULL;
- if (workspaceSize != 0) {
- CHECK_ACL(aclrtMalloc(&workspace, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST));
- }
- ret = aclnnBroadcastCustom(workspace, workspaceSize, handle, stream);
- printf("aclnnBroadcastCustom ret %u\n", ret);
- if (aclrtSynchronizeStreamWithTimeout(stream, 5000) != ACL_SUCCESS) {
- printf("Synchronize stream failed\n");
- }
-
- uint8_t *yHost;
- int64_t yHostSize = GetDataSize(&(tensorDesc[1]));
- CHECK_ACL(aclrtMallocHost((void **)(&yHost), yHostSize));
- CHECK_ACL(aclrtMemcpy(yHost, yHostSize, devMem[1], yHostSize, ACL_MEMCPY_DEVICE_TO_HOST));
- WriteFile("../output/output.bin", yHost, yHostSize);
-
- int64_t wrongNum = CompareResult(yHost, yHostSize);
- if (wrongNum != 0) {
- printf("test failed!\n");
- } else {
- printf("test pass!\n");
- }
- printf("\n");
-
- CHECK_ACL(aclrtFreeHost(yHost));
-
- for (auto i = 0; i < sizeof(tensorDesc) / sizeof(struct tensorInfo); i++) {
- if (!tensors[i])
- continue;
- if (devMem[i]) {
- CHECK_ACL(aclrtFree(devMem[i]));
- }
- aclDestroyTensor(tensors[i]);
- }
- DestroyStream(stream, 0);
- return 0;
-}
diff --git a/examples/readme.md b/examples/readme.md
index 071b5db8136121422a09cc6f90b429fbb5d7c21e..2b64a7703b3ffc85d66ffa500cc94ab8d847752b 100644
--- a/examples/readme.md
+++ b/examples/readme.md
@@ -41,11 +41,6 @@
| layernorm |
将输入数据收敛到[0, 1]之间，每个tensor的特征值减去该特征的均值然后除以该特征的标准差，实现数据归一化。 |
-
- | pad |
- broadcast |
- 对输入tensor的shape进行广播。 |
-