[GitHub] [tvm] comaniac commented on a change in pull request #9419: [CUTLASS] Initial support for dynamic shape dense

[GitBox <gi...@apache.org>]

comaniac commented on a change in pull request #9419:
URL: https://github.com/apache/tvm/pull/9419#discussion_r740532471



##########
File path: python/tvm/contrib/cutlass/build.py
##########
@@ -105,9 +131,19 @@ def tune_cutlass_kernels(mod, sm, profile_all=True, use_multiprocessing=False, t
             MM = arg0_shape[0]
             KK = arg0_shape[1]
             NN = arg1_shape[0]
-            out = cutlass_profiler.profile(
-                MM, NN, KK, annotator.signature["ret_dtype"], profile_all, use_multiprocessing
-            )
+            out_dtype = annotator.signature["ret_dtype"]
+            if any(isinstance(s, tvm.tir.Any) for s in [MM, KK, NN]):
+                out = cutlass_profiler.get_default(out_dtype)
+                print("Picked the default kernel " + out["name"])

Review comment:
       Proper logging.

##########
File path: python/tvm/contrib/cutlass/gen_gemm.py
##########
@@ -335,6 +347,16 @@ def check_align(self, op_name, M):
             return False
         return True
 
+    def get_default(self, out_dtype):
+        """Return the default kernel for the requested architecture.
+        For now, the default kernel was picked arbitrary.
+        """
+        ops = GENERATOR_FUNC_TABLE[self.sm](out_dtype)
+        default_kernel_name = DEFAULT_KERNELS[self.sm][out_dtype]

Review comment:
       Check whether the SM is listed.




-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org