[GitHub] [tvm] Laurawly commented on a change in pull request #7147: [CUDA][PASS]Legalize tensorcore

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

Laurawly commented on a change in pull request #7147:
URL: https://github.com/apache/tvm/pull/7147#discussion_r564104044



##########
File path: tests/python/relay/test_pass_legalize_tensorcore.py
##########
@@ -0,0 +1,227 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+"""Test legalize pass"""
+import numpy as np
+import tvm
+from tvm import te
+from tvm import topi
+from tvm import relay
+from tvm.contrib import graph_runtime
+from tvm.relay import transform, analysis
+from tvm.relay.testing.temp_op_attr import TempOpAttr
+
+
+def run_opt_pass(expr, passes):
+    passes = passes if isinstance(passes, list) else [passes]
+    mod = tvm.IRModule.from_expr(expr)
+    seq = tvm.transform.Sequential(passes)
+    with tvm.transform.PassContext(opt_level=3):
+        mod = seq(mod)
+    entry = mod["main"]
+    return entry if isinstance(expr, relay.Function) else entry.body
+
+
+def test_legalize_conv2d(data_shape, kernel_shape, pad_shape, do_pad=True):

Review comment:
       Please reference [`test_legalize_pass.py`](https://github.com/apache/tvm/blob/01c469d35ec035d0ceff2d7d8c09255d65e7af38/tests/python/relay/test_pass_legalize.py) for the CI issue.

##########
File path: python/tvm/topi/cuda/conv2d_alter_op.py
##########
@@ -345,4 +347,51 @@ def _conv2d_legalize(attrs, inputs, arg_types):
             else:
                 out = relay.nn.conv2d(data, kernel, **new_attrs)
             return out
+    elif data_dtype in ["float16"]:  # todo: support int8/int4
+        if data_layout == "NHWC" and kernel_layout == "HWIO":
+            batch = data_tensor.shape[0].value
+            in_channel = data_tensor.shape[3].value
+            out_channel = kernel_tensor.shape[3].value
+
+            if (
+                (batch % 8 == 0 and in_channel % 16 == 0 and out_channel % 32 == 0)
+                or (batch % 16 == 0 and in_channel % 16 == 0 and out_channel % 16 == 0)
+                or (batch % 32 == 0 and in_channel % 16 == 0 and out_channel % 8 == 0)
+            ):
+                # no need to pad
+                return None
+
+            (db, di, do), extra_flops = pad_to_tensorcore(batch, in_channel, out_channel)
+
+            if extra_flops > 2:
+                logger.info("conv2d pad_to_tensorcore skipped, extra_flops %s", extra_flops)
+                return None
+
+            logger.info("conv2d pad_to_tensorcore, extra_flops %s", extra_flops)
+
+            # Pad batch size
+            if db != 0:
+                data = relay.nn.pad(data, pad_width=((0, db), (0, 0), (0, 0), (0, 0)))
+
+            # Pad input channel
+            if di != 0:
+                data = relay.nn.pad(data, pad_width=((0, 0), (0, 0), (0, 0), (0, di)))
+                kernel = relay.nn.pad(kernel, pad_width=((0, 0), (0, 0), (0, di), (0, 0)))
+
+            # Pad output channel
+            if do != 0:
+                kernel = relay.nn.pad(kernel, pad_width=((0, 0), (0, 0), (0, 0), (0, do)))
+
+            if do != 0:
+                new_out_channel = out_channel + do
+                new_attrs["channels"] = new_out_channel
+                out = tvm.relay.nn.conv2d(data, kernel, **new_attrs)
+            else:
+                out = relay.nn.conv2d(data, kernel, **new_attrs)

Review comment:
       ```suggestion
               if do != 0:
                   new_out_channel = out_channel + do
                   new_attrs["channels"] = new_out_channel
   
               out = relay.nn.conv2d(data, kernel, **new_attrs)
   ```




----------------------------------------------------------------
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.

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