fix bugs on failure of mm example

Signed-off-by: Jinjie Liu <jjliu@baai.ac.cn>

python/triton_tvm_ffi/driver.py

@@ -34,7 +34,35 @@ class TVMLauncher(object):
                 ],
             )
             launch: tvm_ffi.Function = mod.get_function("launch")
-            self.launch = launch
+            self.launch = (
+                lambda grid_x,
+                grid_y,
+                grid_z,
+                stream,
+                function,
+                num_warps,
+                num_ctas,
+                shared_memory,
+                global_scratch,
+                profile_scratch,
+                *args: launch(
+                    grid_x,
+                    grid_y,
+                    grid_z,
+                    stream,
+                    function,
+                    num_warps,
+                    num_ctas,
+                    shared_memory,
+                    global_scratch,
+                    profile_scratch,
+                    *(
+                        arg
+                        for arg, type in zip(args, self.signature)
+                        if type != "constexpr"
+                    ),
+                )
+            )
         else:
             self.impl: TVMFFILauncherImpl = TVMFFILauncherImpl(
                 [string_to_type(t) for t in self.signature],
@@ -129,19 +157,19 @@ class TVMLauncher(object):
 
     @cached_property
     def codegen(self) -> str:
-        env: Final[jinja2.Environment] = jinja2.Environment(
+        env: jinja2.Environment = jinja2.Environment(
             loader=jinja2.PackageLoader("triton_tvm_ffi", "templates"),
             trim_blocks=True,
             lstrip_blocks=True,
         )
-        template = env.get_template("launch.c.j2")
+        template: jinja2.Template = env.get_template("launch.c.j2")
-        signature = list(
+        signature: List[int] = list(
             filter(
                 lambda t: t != "void",
                 map(lambda t: type_to_ctype(string_to_type(t)), self.signature),
             )
         )
-        html = template.render(signature=signature)
+        html: str = template.render(signature=signature)
         return html
 
 

python/triton_tvm_ffi/templates/launch.c.j2

@@ -1,11 +1,13 @@
-#include <cassert>
+#include <assert.h>
 #include <cuda.h>
 #include <tvm/ffi/tvm_ffi.h>
 
 #ifdef __cplusplus
 extern "C"
 #endif
-TVM_FFI_DLL_EXPORT void __tvm_ffi_launch(void *handle, const TVMFFIAny *args, int32_t num_args, TVMFFIAny *result) {
+    TVM_FFI_DLL_EXPORT void
+    __tvm_ffi_launch(void *handle, const TVMFFIAny *args, int32_t num_args,
+                     TVMFFIAny *result) {
   int32_t gridX = args[0].v_int64;
   int32_t gridY = args[1].v_int64;
   int32_t gridZ = args[2].v_int64;
@@ -22,8 +24,7 @@ TVM_FFI_DLL_EXPORT void __tvm_ffi_launch(void *handle, const TVMFFIAny *args, in
     config.gridDimX = gridX * numCtas;
     config.gridDimY = gridY;
     config.gridDimZ = gridZ;
-    static constexpr int32_t kThreadsPerWarp = 32;
+    config.blockDimX = 32 * numWarps;
-    config.blockDimX = kThreadsPerWarp * numWarps;
     config.blockDimY = 1;
     config.blockDimZ = 1;
     config.sharedMemBytes = sharedMemory;
@@ -56,8 +57,7 @@ TVM_FFI_DLL_EXPORT void __tvm_ffi_launch(void *handle, const TVMFFIAny *args, in
       assert(false, "unsupported type yet {{ type }}");
   {% endif %}
 {% endfor %}
-    void *foo = NULL, *bar = NULL;
+    void *params[] = { {% for type in signature %} &arg{{ loop.index0 }}, {% endfor %}&globalScratch, &profileScratch };
-    void *params[] = { {% for type in signature %} &arg{{ loop.index0 }}, {% endfor %}&foo, &bar };
     cuLaunchKernelEx(&config, function, params, NULL);
   }
 }