Balar mmio with Vanadis #2428
Balar mmio with Vanadis #2428
Conversation
* Add a new CUDA API id "GPU_PARAM_CONFIG" to support querying kernel function argument size and alignment information from GPGPU-Sim. * Add param "cuda_executable" to BalarMMIO so that it can know the CUDA binary path when running LLVM CUDA code (Vanadis cannot know the host file structure). * Add all the CUDA API implementations needed to link the test program inside tests/vanadisLLVMRISCV. * Minor formatting changes.
…hake to riscv gcc
|
Status Flag 'Pre-Test Inspection' - - This Pull Request Requires Inspection... The code must be inspected by a member of the Team before Testing/Merging |
There was a problem hiding this comment.
@gvoskuilen @feldergast Do we want to keep the prerequisites in this readme or remove them in favor of the list that we test against? Already discussed what testing we want in the nightlies versus weeklies.
|
|
||
| - Tested on commit `0f358dda178f96db3b0da88b2b965492c4be187d` | ||
| - Use `./configure --prefix=$SST_CORE_HOME --disable-mpi --disable-mem-pools` for sst-core config |
There was a problem hiding this comment.
@William-An Did you test at all with mem pools enabled?
| balar->cuda_ret.is_cuda_call_done = false; | ||
|
|
||
| // Create a DMA request to read the cuda call packet from cache to balar | ||
| DMAEngine::DMAEngineControlRegisters dma_registers; |
There was a problem hiding this comment.
@William-An Did we discuss putting this in memH or vanadis?
@gvoskuilen
| gridDim, | ||
| blockDim, | ||
| packet->configure_call.sharedMem, | ||
| packet->configure_call.stream | ||
| (cudaStream_t) packet->configure_call.stream |
There was a problem hiding this comment.
Do CUDA streams work in this framework?
| GPU_MALLOC_HOST_RET, | ||
| }; | ||
|
|
||
| // Future: Make this into a class with additional serialization methods? |
There was a problem hiding this comment.
@gvoskuilen @feldergast Is this going to be necessary for checkpointing/debug?
| # Constans shared across components | ||
| network_bw = "25GB/s" | ||
| clock = "2GHz" | ||
| balar_mmio_testcpu_addr = 4096 |
There was a problem hiding this comment.
@William-An How configurable are the mmio addresses?
| clock = "2GHz" | ||
| balar_mmio_testcpu_addr = 4096 | ||
| balar_mmio_vanadis_addr = 0x80100000 | ||
| balar_mmio_size = 1024 |
| uint64_t size; | ||
| uint64_t offset; | ||
| uint8_t value[200]; |
There was a problem hiding this comment.
@William-An If this is related to the array from above, we should find a way to ensure that this is propagated everywhere that relies on it.
| @@ -43,7 +48,8 @@ int main( int argc, char* argv[] ) { | |||
|
|
|||
| // Preparing the data | |||
There was a problem hiding this comment.
Why only five updates? And why is n = 10k?
|
|
||
| /** | ||
| * @file cuda_runtime_api.h | ||
| * @author Weili An (an107@purdue.edu) |
There was a problem hiding this comment.
@William-An You should probably remove your email address from these unless you want users bugging you directly. ^-^
Balar mmio with Vanadis