[GitHub] [tvm] Hzfengsy commented on a diff in pull request #15106: [TensorIR][ROCm] AMD Matrix Core Support

["Hzfengsy (via GitHub)" <gi...@apache.org>]

Hzfengsy commented on code in PR #15106:
URL: https://github.com/apache/tvm/pull/15106#discussion_r1232987090


##########
tests/python/unittest/test_tir_schedule_tensorize_mfma.py:
##########
@@ -0,0 +1,322 @@
+# 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=missing-docstring
+import tvm
+from tvm import te
+from tvm.tir.tensor_intrin.rocm import (
+    shared_16x4_to_local_64x1_layout_A,

Review Comment:
   Could we consider change it to 
   ```python
   from tvm.tir.tensor_intrin import rocm
   # or
   from tvm.tir.tensor_intrin.rocm import *
   ```



##########
python/tvm/testing/tir.py:
##########
@@ -128,3 +128,114 @@ def tile_wmma_fragment(block_read, height, width):
     sch.tensorize(sch.get_loops(C_warp)[-2], mma_store_intrin)
 
     return sch
+
+
+def mfma_schedule(
+    workload,
+    k_inner,
+    in_dtype,
+    b_transposed,
+    i_factors,
+    j_factors,
+    k_factors,
+    index_map_A,
+    index_map_B,
+    index_map_C,
+    ldmatrix_a_intrin,
+    ldmatrix_b_intrin,
+    mfma_intrin,
+    mfma_fill_intrin,
+    mfma_store_intrin,
+    shared_scope="shared",
+):
+    import tvm
+    ir_module = tvm.IRModule({"main": workload})
+    sch = tvm.tir.Schedule(ir_module)
+
+    wmma_m = 16
+    wmma_n = 16
+    wmma_k = k_inner
+    warp_size = 64
+    block = sch.get_block("C")
+    i, j, k = sch.get_loops(block)
+    i, i_tc = sch.split(i, factors=[None, wmma_m])
+    j, j_tc = sch.split(j, factors=[None, wmma_n])
+    k, k_tc = sch.split(k, factors=[None, k_inner])
+
+    sch.reorder(i, j, k, i_tc, j_tc, k_tc)
+
+    block_inner = sch.blockize(i_tc)
+    block_outer, block_inner = block_inner, block
+
+    num_ty = i_factors[2] * j_factors[2]
+
+    i0, i1, i2, i3, i4 = sch.split(i, factors=i_factors)
+    j0, j1, j2, j3, j4 = sch.split(j, factors=j_factors)
+    k0, k1, k2 = sch.split(k, k_factors)
+
+    sch.reorder(i0, j0, i1, j1, j2, i2, k0, k1, i3, j3, k2, i4, j4)
+
+    block_idx = sch.fuse(i0, j0)
+    block_idy = sch.fuse(i1, j1)
+    thread_idy = sch.fuse(j2, i2)
+    sch.bind(block_idx, "blockIdx.x")
+    sch.bind(block_idy, "blockIdx.y")
+    sch.bind(thread_idy, "threadIdx.y")
+
+    def fetch_to_shared(block, idx, ndim):
+        block_read = sch.cache_read(block, idx, shared_scope)
+        sch.compute_at(block_read, k0)
+        vector_size = 16 if in_dtype == "int8" else 8
+        fused = sch.fuse(*sch.get_loops(block_read)[-ndim:])
+        _, f_1, f_2, f_3 = sch.split(
+            fused, factors=[None, num_ty, warp_size, vector_size])
+        sch.bind(f_2, "threadIdx.x")
+        sch.bind(f_1, "threadIdx.y")
+        sch.vectorize(f_3)
+        return block_read
+
+    fetch_to_shared(block_outer, 0, 2)
+    fetch_to_shared(block_outer, 1, 2)
+
+    A_warp = sch.cache_read(block_outer, 0, "local")
+    B_warp = sch.cache_read(block_outer, 1, "local")
+
+    sch.compute_at(A_warp, k1)
+    sch.compute_at(B_warp, k1)
+    C_warp = sch.cache_write(block_outer, 0, "local")
+    sch.reverse_compute_at(C_warp, thread_idy)
+
+    ii, jj = sch.get_loops(C_warp)[-2:]
+    io, ii = sch.split(ii, factors=[None, 16])
+    jo, ji = sch.split(jj, factors=[None, 16])
+    sch.reorder(io, jo, ii, ji)
+
+    sch.decompose_reduction(block_outer, sch.get_loops(block_outer)[3])
+    block_init_c = sch.get_block("C_init")
+
+    def tile_wmma_fragment(block_read, height, width):
+        i, j = sch.get_loops(block_read)[-2:]
+        i0, i1 = sch.split(i, factors=[None, height])
+        j0, j1 = sch.split(j, factors=[None, width])
+        sch.reorder(i0, j0, i1, j1)
+        return i1
+
+    loop_a = tile_wmma_fragment(A_warp, 16, k_inner)
+
+    if b_transposed:
+        loop_b = tile_wmma_fragment(B_warp, 16, k_inner)
+    else:
+        loop_b = tile_wmma_fragment(B_warp, k_inner, 16)
+
+    sch.transform_layout(A_warp, ("write", 0), index_map_A)
+    sch.transform_layout(B_warp, ("write", 0), index_map_B)
+    sch.transform_layout(C_warp, ("read", 0), index_map_C)
+    print(sch.mod["main"].script())

Review Comment:
   is it necessary? 



-- 
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