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 mm and autotune

Browse Source 
Signed-off-by: jinjieliu <jinjie.liu@usc.edu>
This commit is contained in:
jinjieliu
2026-02-07 00:41:23 +08:00
parent f6c7a48c1b
commit 2298b6f8c8
7 changed files with 486 additions and 56 deletions
Download Patch File Download Diff File Expand all files Collapse all files

80
python/triton_tvm_ffi/templates/gendef.cc.j2
View File

@@ -1,4 +1,6 @@
#include <cassert>
#include <cuda.h>
#include <optional>
#include <tvm/ffi/extra/cuda/cubin_launcher.h>
#include <tvm/ffi/function.h>
@@ -9,32 +11,60 @@
{% if fn.ctypes is none %}
#define {{ fn.fnname | upper }}_STUB tvm::ffi::Function::GetGlobalRequired("{{ fn.fullname }}")
{% else %}
TVM_FFI_EMBED_CUBIN(triton_{{ fn.fnname }});
#define {{ fn.fnname | upper}}_STUB(__grid, __stream, __numWarps, __numStages{% for ctype in fn.ctypes %}, {{ "__arg" ~ loop.index0 }}{% endfor %}) do { \
const tvm::ffi::Map<tvm::ffi::String, tvm::ffi::Any> __meta = { \
{% for name in fn.signature %}
{ "{{ name }}", __arg{{ loop.index0 }} }, \
static const char __cubin[] = "{{ fn.kernel_cstr }}";
#define __CUDA_CHECK(code) assert((code) == CUDA_SUCCESS)
static CUfunction __Get{{ fn.fnname }}Kernel() {
static std::optional<CUfunction> function = std::nullopt;
if (!function) {
CUmodule module;
CUfunction func;
__CUDA_CHECK(cuModuleLoadData(&module, __cubin));
__CUDA_CHECK(cuModuleGetFunction(&func, module, "{{ fn.fnname }}"));
{% if fn.shmem > 49152 %}
int shared_optin, shared_static;
__CUDA_CHECK(cuDeviceGetAttribute(&shared_optin, CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK_OPTIN, /* TODO: we assume the device id is 0 here, but this may not work on devices with more than one gpu */0));
if (shared_optin >= 49152) {
__CUDA_CHECK(cuFuncGetAttribute(&shared_static, CU_FUNC_ATTRIBUTE_SHARED_SIZE_BYTES, func));
__CUDA_CHECK(cuFuncSetAttribute(func, CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES, shared_optin - shared_static));
}
{% endif %}
function = func;
}
return *function;
}
#define {{ fn.fnname | upper }}_STUB(__grid, __stream, __numWarps, __numStages, __args, __kwargs) do { \
const char *__signature[] = { "{{ fn.signature | join("\", \"") }}" }; \
tvm::ffi::Map<tvm::ffi::String, tvm::ffi::Any> __meta = { \
{% if fn.best_config != none %}
{% for k, v in fn.best_config.items() %}
{ "{{ k }}", {{ v }} }, \
{% endfor %}
}; \
static auto __kernel = TVM_FFI_EMBED_CUBIN_GET_KERNEL(triton_{{ fn.fnname }}, "{{ fn.fnname }}"); \
tvm::ffi::dim3 __gridDim = MakeGridDim(__grid, __meta); \
tvm::ffi::dim3 __block({% if fn.num_warps != none %}{{ fn.num_warps }}{% else %}__numWarps{% endif %} * 32, 1, 1); \
void *dummy = nullptr
{%- for ctype in fn.ctypes -%}
{%- if ctype == "CUdeviceptr" -%}
, *__arg{{ loop.index0 }}_ptr=__arg{{ loop.index0 }}.data_ptr()
{%- endif -%}
{%- endfor -%}; \
void *__params[] = {
{%- for ctype in fn.ctypes -%}
{%- if ctype != none -%}
&__arg{{ loop.index0 }}
{%- if ctype == "CUdeviceptr" -%}
_ptr
{%- endif -%},
{%- endif -%}
{%- endfor -%}&dummy, &dummy }; \
TVM_FFI_CHECK_CUBIN_LAUNCHER_CUDA_ERROR(__kernel.Launch(__params, __gridDim, __block, static_cast<tvm::ffi::cuda_api::StreamHandle>(__stream))); \
{% endif %}
}; \
for (size_t __i = 0, __size_args = __args.size(); __i < sizeof(__signature) / sizeof(const char *); ++__i) { \
if (__i < __size_args) { \
__meta.Set(__signature[__i], __args[__i]); \
} else if (auto __val = __kwargs.Get(__signature[__i])) { \
__meta.Set(__signature[__i], *__val); \
} \
} \
CUfunction __function = __Get{{ fn.fnname }}Kernel(); \
tvm::ffi::Tuple<int32_t, int32_t, int32_t> __gridDim = MakeGridDim(__grid, __meta); \
void *dummy = nullptr; \
{% for ctype in fn.ctypes %}
{% if ctype != none %}
{% if ctype == "CUdeviceptr" %}
void *__arg{{ loop.index0 }} = __args[{{ loop.index0 }}].cast<tvm::ffi::TensorView>().data_ptr(); \
{% else %}
{{ ctype }} __arg{{ loop.index0 }} = __args[{{ loop.index0 }}].cast<{{ ctype }}>(); \
{% endif %}
{% endif %}
{% endfor %}
void *__params[] = { {% for ctype in fn.ctypes %}{% if ctype != none %}&__arg{{ loop.index0 }}, {% endif %}{% endfor %}&dummy, &dummy }; \
__CUDA_CHECK(cuLaunchKernel(__function, __gridDim.get<0>(), __gridDim.get<1>(), __gridDim.get<2>(), 32 * {{ fn.num_warps }}, 1, 1, {{ fn.shmem }}, reinterpret_cast<CUstream>(__stream), __params, nullptr)); \
} while (false)
{% endif %}
{% endfor %}

10
python/triton_tvm_ffi/templates/grid.h
View File

@@ -6,19 +6,21 @@
#include <tvm/ffi/tvm_ffi.h>
template <typename T>
inline tvm::ffi::dim3
inline tvm::ffi::Tuple<int32_t, int32_t, int32_t>
MakeGridDim(const T &grid,
const tvm::ffi::Map<tvm::ffi::String, tvm::ffi::Any> &meta);
template <>
inline tvm::ffi::dim3 MakeGridDim<tvm::ffi::Tuple<int32_t, int32_t, int32_t>>(
inline tvm::ffi::Tuple<int32_t, int32_t, int32_t>
MakeGridDim<tvm::ffi::Tuple<int32_t, int32_t, int32_t>>(
const tvm::ffi::Tuple<int32_t, int32_t, int32_t> &grid,
const tvm::ffi::Map<tvm::ffi::String, tvm::ffi::Any> &) {
return tvm::ffi::dim3(grid.get<0>(), grid.get<1>(), grid.get<2>());
return grid;
}
template <>
inline tvm::ffi::dim3 MakeGridDim<tvm::ffi::Function>(
inline tvm::ffi::Tuple<int32_t, int32_t, int32_t>
MakeGridDim<tvm::ffi::Function>(
const tvm::ffi::Function &grid,
const tvm::ffi::Map<tvm::ffi::String, tvm::ffi::Any> &meta) {
tvm::ffi::Tuple<int32_t, int32_t, int32_t> tuple =