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

support softmax

Browse Source 
Signed-off-by: Jinjie Liu <jjliu@baai.ac.cn>
This commit is contained in:
Jinjie Liu
2026-02-04 16:38:33 +08:00
parent 192dc95ac0
commit b7bf598fde
4 changed files with 125 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

35
examples/softmax/softmax.cc Normal file
View File

@@ -0,0 +1,35 @@
#include <ATen/DLConvertor.h>
#include <ATen/dlpack.h>
#include <tvm/ffi/extra/cuda/cubin_launcher.h>
#include <tvm/ffi/tvm_ffi.h>
#ifndef SOFTMAX_KERNEL_STUB
#define SOFTMAX_KERNEL_STUB(grid, stream, numWarps, numStages, output, input, \
inputStride, outputStride, nRows, nCols, \
BLOCK_SIZE)
#endif
#ifndef SOFTMAX_NAME
#define SOFTMAX_NAME ""
#endif
tvm::ffi::Tensor Softmax(tvm::ffi::Tensor x) {
at::Tensor xtorch = at::fromDLPack(x.ToDLPack());
at::Tensor ytorch = at::empty_like(xtorch);
uint32_t nRows = xtorch.size(0), nCols = xtorch.size(1), numWarps = 8,
numStages = 4, xStride = xtorch.stride(0),
yStride = ytorch.stride(0),
BLOCK_SIZE = 1u << (32 - __builtin_clz(nCols - 1));
tvm::ffi::Tensor y = tvm::ffi::Tensor::FromDLPack(at::toDLPack(ytorch));
tvm::ffi::Tuple<int32_t, int32_t, int32_t> grid{nRows / 1024, 1, 1};
DLDevice device = x.device();
void *stream = TVMFFIEnvGetStream(device.device_type, device.device_id);
SOFTMAX_KERNEL_STUB(grid, stream, numWarps, numStages, y, x, xStride, yStride,
nRows, nCols, BLOCK_SIZE);
return y;
}
TVM_FFI_STATIC_INIT_BLOCK() {
namespace refl = tvm::ffi::reflection;
refl::GlobalDef().def(SOFTMAX_NAME, Softmax);
}