TEST/CUDA: Test CUDA Memcpy with various contexts #10531
Conversation
tvegas1
force-pushed
the
cuda_ctx_memcpy
branch
2 times, most recently
from
837f312 to
700702c
Compare
test/apps/test_cuda_ctx_memcpy.cu
Outdated
| printf("Using %d CUDA device(s)\n", device_count); | ||
|
|
||
| for (i = 0; i < MAX_THREADS; i++) { | ||
| thread[i].tid = i; |
test/apps/test_cuda_ctx_memcpy.cu
Outdated
|
|
||
|
|
||
| /* Context and associated resources */ | ||
| struct context { |
There was a problem hiding this comment.
maybe use typedef for both structs
test/apps/test_cuda_ctx_memcpy.cu
Outdated
| void *mem_managed; | ||
| }; | ||
|
|
||
| struct context context[MAX_CTX * MAX_THREADS * MAX_DEV]; |
There was a problem hiding this comment.
maybe better to have an independent context per each thread?
There was a problem hiding this comment.
they share the content
test/apps/test_cuda_ctx_memcpy.cu
Outdated
|
|
||
| ptr_a = context[j].mem; | ||
| ptr_b = context[k].mem; | ||
| ptr_a_managed = context[j].mem_managed; |
test/apps/test_cuda_ctx_memcpy.cu
Outdated
| #include <unistd.h> | ||
|
|
||
|
|
||
| #define MAX_THREADS 2 |
There was a problem hiding this comment.
should we pass these from cmd line?
There was a problem hiding this comment.
as it is scale is already huge so i will add if we need other parameters?
test/apps/test_cuda_ctx_memcpy.cu
Outdated
| for (k = 0; k < MAX_CTX * MAX_THREADS * count; k++) { | ||
| /* each stream (every thread every device) */ | ||
| for (l = 0; l < MAX_CTX * MAX_THREADS * count; l++) { | ||
| CHECK_D(cuCtxSetCurrent(context[i].ctx)); |
There was a problem hiding this comment.
There is no need to set CUDA context here. Even in case of CUDA Driver API.
The context is required for stream creation using CUDA Driver API.
There was a problem hiding this comment.
intent was to confirm that we do not care about current ctx being set, even if it is different from stream's ctx
There was a problem hiding this comment.
The context is required only for stream creation. Memory copies can be executed even without setting any context.
There was a problem hiding this comment.
removed setting the context
| test_cuda_ctx_memcpy_DEPBASE = $(DEPDIR)/test_cuda_ctx_memcpy | ||
| test_cuda_ctx_memcpy_COMPILE = \ | ||
| $(NVCC) $(DEFS) $(DEFAULT_INCLUDES) $(INCLUDES) \ | ||
| $(test_cuda_ctx_memcpy_CPPFLAGS) |
There was a problem hiding this comment.
You can make your file .c and remove this.
test/apps/test_cuda_ctx_memcpy.cu
Outdated
|
|
||
| for (j = 1; j < MAX_CTX; j++) { | ||
| CHECK_D(cuCtxCreate(&context[index + j].ctx, 0, i)); | ||
| CHECK_D(cuCtxSetCurrent(context[index + j].ctx)); |
There was a problem hiding this comment.
The line can be removed since cuCtxCreate pushes the newly created context.
test/apps/test_cuda_ctx_memcpy.cu
Outdated
| CUcontext ctx; | ||
|
|
||
| for (i = 0; i < count; i++) { | ||
| CHECK(cudaSetDevice(i)); |
There was a problem hiding this comment.
This call retains the primary device context, and sets it as current context.
Maybe replace by cuDeviceGet(&device, i). And pass device to cuCtxCreate.
What?
Exercise cuda copies with various combinations of parameters (device/ctx/stream/memcpy).