JinjieLiu/triton-tvm-ffi
1
0
Fork 0
mirror of https://github.com/sgjzfzzf/triton-tvm-ffi.git synced 2026-05-02 03:52:11 +08:00
Code Issues Packages Projects Releases Wiki Activity

verify tvm-ffi cpp wrapper on vector-add.py

Browse Source 
Signed-off-by: jinjieliu <jinjie.liu@usc.edu>
This commit is contained in:
jinjieliu
2026-02-04 02:30:26 +08:00
commit dc8c2c17e0
10 changed files with 423 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
examples/add/add.cc Normal file
View File

@@ -0,0 +1,31 @@
#include <ATen/DLConvertor.h>
#include <ATen/dlpack.h>
#include <tvm/ffi/extra/cuda/cubin_launcher.h>
#include <tvm/ffi/tvm_ffi.h>
#ifndef ADD_KERNEL_STUB
#define ADD_KERNEL_STUB(grid, stream, numWarps, numStages, x, y, output, \
numel, BLOCK_SIZE)
#endif
#ifndef ADD_NAME
#define ADD_NAME ""
#endif
tvm::ffi::Tensor Add(tvm::ffi::Tensor x, tvm::ffi::Tensor y) {
at::Tensor xtorch = at::fromDLPack(x.ToDLPack());
at::Tensor otorch = at::empty_like(xtorch);
int64_t numel = otorch.numel();
tvm::ffi::Tensor output = tvm::ffi::Tensor::FromDLPack(at::toDLPack(otorch));
tvm::ffi::Tuple<int32_t, int32_t, int32_t> grid{(numel + 1023) / 1024, 1, 1};
size_t numWarps = 4, numStages = 3;
DLDevice device = x.device();
void *stream = TVMFFIEnvGetStream(device.device_type, device.device_id);
ADD_KERNEL_STUB(grid, stream, numWarps, numStages, x, y, output, numel, 1024);
return output;
}
TVM_FFI_STATIC_INIT_BLOCK() {
namespace refl = tvm::ffi::reflection;
refl::GlobalDef().def(ADD_NAME, Add);
}