Add network communication through another GPU connected with NVLink
(PXN).
Add aggregation of messages coming from different local GPUs through
PXN and going to the same destination.
Add new v5 plugin API with grouped receives and tags.
Add compat for v4 plugins.
Add naming of NCCL threads to help debugging.
Fix NVLink detection and avoid data corruption when some NVLinks are
down.
Add support for Relaxed Ordering for IB.
Add profiling and timing infrastructure.
This commit is contained in:
Sylvain Jeaugey
2022-01-07 06:39:55 -08:00
parent 0144073673
commit 3c223c105a
70 changed files with 6453 additions and 3430 deletions
+28
View File
@@ -0,0 +1,28 @@
/*************************************************************************
* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
*
* See LICENSE.txt for license information
************************************************************************/
/* Define weak symbols used to allow libnccl_static.a to work with older libcudart_static.a */
enum cudaError_t { cudaErrorStubLibrary = 34 };
extern "C" {
cudaError_t cudaStreamGetCaptureInfo_v2(...) __attribute__((visibility("hidden"))) __attribute((weak));
cudaError_t cudaStreamGetCaptureInfo_v2(...) { return cudaErrorStubLibrary; }
cudaError_t cudaUserObjectCreate(...) __attribute__((visibility("hidden"))) __attribute((weak));
cudaError_t cudaUserObjectCreate(...) { return cudaErrorStubLibrary; }
cudaError_t cudaGraphRetainUserObject(...) __attribute__((visibility("hidden"))) __attribute((weak));
cudaError_t cudaGraphRetainUserObject(...) { return cudaErrorStubLibrary; }
cudaError_t cudaStreamUpdateCaptureDependencies(...) __attribute__((visibility("hidden"))) __attribute((weak));
cudaError_t cudaStreamUpdateCaptureDependencies(...) { return cudaErrorStubLibrary; }
cudaError_t cudaGetDriverEntryPoint(...) __attribute__((visibility("hidden"))) __attribute((weak));
cudaError_t cudaGetDriverEntryPoint(...) { return cudaErrorStubLibrary; }
}