[GitHub] [tvm] jcf94 commented on a change in pull request #7146: [CUDA]batch_matmul tensorcore schedule

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

jcf94 commented on a change in pull request #7146:
URL: https://github.com/apache/tvm/pull/7146#discussion_r547081523



##########
File path: python/tvm/topi/cuda/batch_matmul_tensorcore.py
##########
@@ -0,0 +1,275 @@
+# 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.
+# pylint: disable=invalid-name,too-many-locals,unused-variable,unused-argument
+"""cuda batch_matmul operators"""
+import tvm
+from tvm import autotvm
+from tvm import te
+from ..utils import traverse_inline, get_const_tuple
+from .tensor_intrin import intrin_wmma_load_matrix_A, \
+        intrin_wmma_load_matrix_W, intrin_wmma_store_matrix, intrin_wmma_gemm
+
+@autotvm.register_topi_compute("batch_matmul_tensorcore.cuda")
+def batch_matmul_tensorcore(cfg, x, y, out_shape=None):
+    """batch matmul tensorcore operator on cuda"""
+    # todo: deal with out_shape for broadcast, liuxin.ai
+    return batch_matmul_tensorcore_cuda(x, y)
+
+
+@autotvm.register_topi_schedule("batch_matmul_tensorcore.cuda")
+def schedule_batch_matmul_tensorcore(cfg, outs):
+    """Schedule for batch_matmul operator using Tensorcore
+
+    Parameters
+    ----------
+    outs: Array of Tensor
+          The computation graph description of batch_matmul
+          in the format of an array of tensors.
+
+    Returns
+    -------
+    s: Schedule
+        The computation schedule for the op.
+    """
+    outs = [outs] if isinstance(outs, te.tensor.Tensor) else outs
+    s = te.create_schedule([x.op for x in outs])
+
+    def _schedule(cfg, s, C):
+        A, B = s[C].op.input_tensors
+        batch, m_dim, k_dim = get_const_tuple(A.shape)
+        batch, n_dim, k_dim = get_const_tuple(B.shape)
+        out_dtype = C.dtype
+        # inline astype fp16
+        s[A].compute_inline()
+        s[B].compute_inline()
+
+        # Explicit memory access
+        AS = s.cache_read(A, 'shared', [C])
+        BS = s.cache_read(B, 'shared', [C])
+        AF = s.cache_read(AS, 'wmma.matrix_a', [C])
+        BF = s.cache_read(BS, 'wmma.matrix_b', [C])
+        CF = s.cache_write(C, 'wmma.accumulator')
+        CS = s.cache_read(CF, 'shared', [C])
+
+        # fallback support
+        target = tvm.target.Target.current()
+        if cfg.is_fallback:
+            ref_log = autotvm.tophub.load_reference_log(
+                target.kind.name, target.model, 'batch_matmul_tensorcore.cuda')
+            cfg.fallback_with_reference_log(ref_log)
+
+        # ??? Deal with op fusion, such as bias and relu ??? is this needed?

Review comment:
       typo？




----------------------------------------------------------------
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