DeforaNetworks/hashcat
3
0
Fork 0
Code Releases Activity

Add async CUDA memcpy functions: hc_cuMemcpyDtoDAsync(), hc_cuMemcpyDtoHAsync() and hc_cuMemcpyHtoDAsync(). Implement partially async CUDA memset and bzero kernels.

Browse Source
This commit is contained in:
Jukka Ojanen
2021-07-20 12:23:39 +03:00
parent f07ff6f03d
commit 4263cafdcf
3 changed files with 103 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files
include/backend.h
+3
View File
@@ -69,8 +69,11 @@ int hc_cuInit (hashcat_ctx_t *hashcat_ctx, unsigned int Flags
int hc_cuLaunchKernel (hashcat_ctx_t *hashcat_ctx, CUfunction f, unsigned int gridDimX, unsigned int gridDimY, unsigned int gridDimZ, unsigned int blockDimX, unsigned int blockDimY, unsigned int blockDimZ, unsigned int sharedMemBytes, CUstream hStream, void **kernelParams, void **extra);
int hc_cuMemAlloc (hashcat_ctx_t *hashcat_ctx, CUdeviceptr *dptr, size_t bytesize);
int hc_cuMemcpyDtoD (hashcat_ctx_t *hashcat_ctx, CUdeviceptr dstDevice, CUdeviceptr srcDevice, size_t ByteCount);
int hc_cuMemcpyDtoDAsync (hashcat_ctx_t *hashcat_ctx, CUdeviceptr dstDevice, CUdeviceptr srcDevice, size_t ByteCount, CUstream hStream);
int hc_cuMemcpyDtoH (hashcat_ctx_t *hashcat_ctx, void *dstHost, CUdeviceptr srcDevice, size_t ByteCount);
int hc_cuMemcpyDtoHAsync (hashcat_ctx_t *hashcat_ctx, void *dstHost, CUdeviceptr srcDevice, size_t ByteCount, CUstream hStream);
int hc_cuMemcpyHtoD (hashcat_ctx_t *hashcat_ctx, CUdeviceptr dstDevice, const void *srcHost, size_t ByteCount);
int hc_cuMemcpyHtoDAsync (hashcat_ctx_t *hashcat_ctx, CUdeviceptr dstDevice, const void *srcHost, size_t ByteCount, CUstream hStream);
int hc_cuMemFree (hashcat_ctx_t *hashcat_ctx, CUdeviceptr dptr);
int hc_cuModuleGetFunction (hashcat_ctx_t *hashcat_ctx, CUfunction *hfunc, CUmodule hmod, const char *name);
int hc_cuModuleLoadDataEx (hashcat_ctx_t *hashcat_ctx, CUmodule *module, const void *image, unsigned int numOptions, CUjit_option *options, void **optionValues);
include/ext_cuda.h
+6
View File
@@ -1029,8 +1029,11 @@ typedef CUresult (CUDA_API_CALL *CUDA_CULAUNCHKERNEL) (CUfunction, uns
typedef CUresult (CUDA_API_CALL *CUDA_CUMEMALLOC) (CUdeviceptr *, size_t);
typedef CUresult (CUDA_API_CALL *CUDA_CUMEMALLOCHOST) (void **, size_t);
typedef CUresult (CUDA_API_CALL *CUDA_CUMEMCPYDTOD) (CUdeviceptr, CUdeviceptr, size_t);
typedef CUresult (CUDA_API_CALL *CUDA_CUMEMCPYDTODASYNC) (CUdeviceptr, CUdeviceptr, size_t, CUstream);
typedef CUresult (CUDA_API_CALL *CUDA_CUMEMCPYDTOH) (void *, CUdeviceptr, size_t);
typedef CUresult (CUDA_API_CALL *CUDA_CUMEMCPYDTOHASYNC) (void *, CUdeviceptr, size_t, CUstream);
typedef CUresult (CUDA_API_CALL *CUDA_CUMEMCPYHTOD) (CUdeviceptr, const void *, size_t);
typedef CUresult (CUDA_API_CALL *CUDA_CUMEMCPYHTODASYNC) (CUdeviceptr, const void *, size_t, CUstream);
typedef CUresult (CUDA_API_CALL *CUDA_CUMEMFREE) (CUdeviceptr);
typedef CUresult (CUDA_API_CALL *CUDA_CUMEMFREEHOST) (void *);
typedef CUresult (CUDA_API_CALL *CUDA_CUMEMGETINFO) (size_t *, size_t *);
@@ -1091,8 +1094,11 @@ typedef struct hc_cuda_lib
CUDA_CUMEMALLOC cuMemAlloc;
CUDA_CUMEMALLOCHOST cuMemAllocHost;
CUDA_CUMEMCPYDTOD cuMemcpyDtoD;
CUDA_CUMEMCPYDTODASYNC cuMemcpyDtoDAsync;
CUDA_CUMEMCPYDTOH cuMemcpyDtoH;
CUDA_CUMEMCPYDTOHASYNC cuMemcpyDtoHAsync;
CUDA_CUMEMCPYHTOD cuMemcpyHtoD;
CUDA_CUMEMCPYHTODASYNC cuMemcpyHtoDAsync;
CUDA_CUMEMFREE cuMemFree;
CUDA_CUMEMFREEHOST cuMemFreeHost;
CUDA_CUMEMGETINFO cuMemGetInfo;