merge main

DeepLink-org · Mar 24, 2024 · d18234e · d18234e
1 parent 102b1c4
commit d18234e
Showing 1 changed file with 5 additions and 5 deletions.
diff --git a/dipu/torch_dipu/csrc_dipu/vendor/ascend/deviceimpl.cpp b/dipu/torch_dipu/csrc_dipu/vendor/ascend/deviceimpl.cpp
@@ -135,7 +135,10 @@ OpStatus mallocDevice(void** p, size_t nbytes, bool throwExcepion) {
     *p = nullptr;
     return OpStatus::SUCCESS;
   }
-  DIPU_CALLACLRT(::aclrtMalloc(p, nbytes, ACL_MEM_MALLOC_HUGE_FIRST));
+  // 若用户需申请大块内存并自行划分、管理内存时，建议使用aclrtMallocAlign32接	  DIPU_CALLACLRT(::aclrtMalloc(p, nbytes, ACL_MEM_MALLOC_HUGE_FIRST));
+  // 口，该接口相比aclrtMalloc接口，只会对用户申请的size向上对齐成32字节整数
+  // 倍，不会再多加32字节。 大块内存用作缓存时，无需多加32字节
+  DIPU_CALLACLRT(::aclrtMallocAlign32(p, nbytes, ACL_MEM_MALLOC_HUGE_FIRST));
   return OpStatus::SUCCESS;
 }
 
@@ -149,21 +152,18 @@ void freeDevice(void* p) {
 // (synchronous) copy from device to a device
 void memCopyD2D(size_t nbytes, deviceId_t dstDevId, void* dst,
                 deviceId_t srcDevId, const void* src) {
-  syncDevice();
   DIPU_CALLACLRT(
       ::aclrtMemcpy(dst, nbytes, src, nbytes, ACL_MEMCPY_DEVICE_TO_DEVICE));
 }
 
 // (synchronous) copy from host to a device
 void memCopyH2D(size_t nbytes, void* dst, const void* src) {
-  syncDevice();
   DIPU_CALLACLRT(
       ::aclrtMemcpy(dst, nbytes, src, nbytes, ACL_MEMCPY_HOST_TO_DEVICE));
 }
 
 // (synchronous) copy from a device to host
 void memCopyD2H(size_t nbytes, void* dst, const void* src) {
-  syncDevice();
   DIPU_CALLACLRT(
       ::aclrtMemcpy(dst, nbytes, src, nbytes, ACL_MEMCPY_DEVICE_TO_HOST));
 }
@@ -251,7 +251,7 @@ bool streamNotNull(deviceStream_t stream) {
 }
 
 void streamWaitEvent(deviceStream_t stream, deviceEvent_t event) {
-  DIPU_CALLACLRT(::aclrtSynchronizeEvent(event))
+  DIPU_CALLACLRT(::aclrtStreamWaitEvent(stream, event))
 }
 
 // =====================