| #include <cstdio> |
| #include <cuda.h> |
| #include <cuda_fp16.h> |
| #include <cstring> |
| #define CHECK_CU(x) do { CUresult r = x; if (r != CUDA_SUCCESS) { printf("CUDA err %d at %d\n", r, __LINE__); return 1; } } while(0) |
| int main() { |
| CHECK_CU(cuInit(0)); CUdevice dev; CHECK_CU(cuDeviceGet(&dev, 0)); |
| CUcontext ctx; CHECK_CU(cuDevicePrimaryCtxRetain(&ctx, dev)); CHECK_CU(cuCtxSetCurrent(ctx)); |
| CUmodule mod; CHECK_CU(cuModuleLoad(&mod, "conv2d_implicit_gemm.sm_86.cubin")); |
| CUfunction fn; CHECK_CU(cuModuleGetFunction(&fn, mod, "implicit_gemm_conv")); |
|
|
| int N=1, H=64, W=64, Cin=64, kH=3, kW=3, pad=1; |
| int out_H=64, out_W=64, M=4096, Kdim=576, Cout=64; |
| size_t Xe = (size_t)N*H*W*Cin; |
| size_t Ye = (size_t)M*Cout; |
| size_t Be = (size_t)Kdim*Cout; |
| CUdeviceptr dX, dB, dY; |
| CHECK_CU(cuMemAlloc(&dX, Xe*sizeof(float))); |
| CHECK_CU(cuMemAlloc(&dB, Be*sizeof(__half))); |
| CHECK_CU(cuMemAlloc(&dY, Ye*sizeof(float))); |
| CHECK_CU(cuMemsetD32(dX, 0, Xe)); |
| CHECK_CU(cuMemsetD32(dB, 0, Be)); |
|
|
| void* args[] = {&dX, &dB, &dY, &N, &H, &W, &Cin, &kH, &kW, &pad, &out_H, &out_W, &M, &Kdim, &Cout}; |
| int grid_m = (M + 64 - 1) / 64; |
| int grid_n = (Cout + 64 - 1) / 64; |
| printf("Grid: %d x %d\n", grid_m, grid_n); |
| printf("Smem: %d bytes\n", 8192); |
|
|
| CHECK_CU(cuFuncSetAttribute(fn, CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES, 8192)); |
| printf("Launching implicit...\n"); |
| CHECK_CU(cuLaunchKernel(fn, grid_m, grid_n, 1, 128, 1, 1, 8192, 0, args, 0)); |
| printf("Syncing...\n"); |
| CHECK_CU(cuCtxSynchronize()); |
| printf("OK!\n"); |
| return 0; |
| } |
|
|
