Use cuda virtual memory management and merge blocks #36189
Conversation
… auto_growth_v2
| auto result = | ||
| paddle::platform::dynload::cuDeviceGet(&device, p.GetDeviceId()); | ||
| PADDLE_ENFORCE_EQ( | ||
| result, CUDA_SUCCESS, | ||
| platform::errors::Fatal("Call CUDA API cuDeviceGet faild, return %d.", | ||
| result)); |
There was a problem hiding this comment.
PADDLE_ENFORCE_CUDA_SUCCESS ？
| result, CUDA_SUCCESS, | ||
| platform::errors::Fatal( | ||
| "Call CUDA API cuDeviceGetAttribute faild, return %d.", result)); | ||
| } catch (...) { |
There was a problem hiding this comment.
in which case it may raise exception?
There was a problem hiding this comment.
There was a problem hiding this comment.
I see, plz add comments on that.
| @@ -131,6 +131,20 @@ gpuError_t RecordedCudaMalloc(void **ptr, size_t size, int dev_id); | |||
| //! CudaFree with recorded info | |||
| void RecordedCudaFree(void *p, size_t size, int dev_id); | |||
|
|
|||
| #ifdef PADDLE_WITH_CUDA | |||
There was a problem hiding this comment.
我不确定在不是CUDA的情况下CUDA_VERSION是否会被初始化为乱码，因此严谨了一些。
There was a problem hiding this comment.
There is #ifdef PADDLE_WITH_CUDA at the beginning, so I think maybe it is duplicated.
| __macro(cuInit); \ | ||
| __macro(cuDriverGetVersion); \ | ||
| __macro(cuGetErrorString); \ | ||
| __macro(cuModuleLoadData); \ | ||
| __macro(cuModuleGetFunction); \ | ||
| __macro(cuModuleUnload); \ | ||
| __macro(cuOccupancyMaxActiveBlocksPerMultiprocessor); \ | ||
| __macro(cuLaunchKernel); \ | ||
| __macro(cuCtxCreate); \ | ||
| __macro(cuCtxGetCurrent); \ | ||
| __macro(cuDeviceGetCount); \ | ||
| __macro(cuDevicePrimaryCtxGetState); \ |
There was a problem hiding this comment.
These are duplicated with APIs in #else, maybe #else is not needed
There was a problem hiding this comment.
这里面有些API，只有cuda10.2以上才有，所以用宏控制了一下。
There was a problem hiding this comment.
I think you can always define the APIs that exists in version < 10.2 and > 10.2 without macro.
paddle/fluid/platform/gpu_info.cc
Outdated
| @@ -641,6 +646,30 @@ class RecordedCudaMallocHelper { | |||
|
|
|||
| uint64_t LimitSize() const { return limit_size_; } | |||
|
|
|||
| #ifdef PADDLE_WITH_CUDA | |||
| #if CUDA_VERSION >= 10020 | |||
| CUresult cuMemCreate(CUmemGenericAllocationHandle *handle, size_t size, | |||
There was a problem hiding this comment.
Better name the member function like other functions, for example, "CreateMem". (Start with upper case)
| paddle::platform::CUDADeviceGuard guard(place.device); | ||
| PADDLE_ENFORCE_CUDA_SUCCESS(cudaMemGetInfo(&actual_avail, &actual_total)); | ||
|
|
||
| virtual_mem_size_ = (actual_total + granularity_ - 1) & ~(granularity_ - 1); |
There was a problem hiding this comment.
这是为了申请一个虚拟地址空间，大小恰好等于GPU的显存大小。确保一次性开辟的虚拟地址空间够用。
There was a problem hiding this comment.
virtual_mem_size_ = (actual_total + granularity_ - 1) & ~(granularity_ - 1);
已经改成使用函数调用方式
| block--; | ||
| auto pre = block; | ||
| block++; | ||
| block++; | ||
| auto next = block; | ||
| block--; |
There was a problem hiding this comment.
cannot easily understand it...
There was a problem hiding this comment.
有pre、next还不好理解吗？
There was a problem hiding this comment.
已改成使用std::next和std::prev
PR types
New features
PR changes
Others
Describe
使用NV Virtual Memory Management（VMM）机制新写Allocator（AutoGrowthV2）。
使用VMM后，可以对向CUDA申请的显存块进行合并（而老的AutoGrowth是不能够合并显存块的）。