[ExecuTorch] Optimized op_mm using CPUBlas gemm

kernels/optimized/cpu/op_mm.cpp

@@ -0,0 +1,66 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#include <executorch/kernels/optimized/blas/CPUBlas.h>
+#include <executorch/kernels/portable/cpu/util/matmul_ops_util.h>
+#include <executorch/runtime/kernel/kernel_includes.h>
+
+namespace torch {
+namespace executor {
+namespace native {
+
+using Tensor = exec_aten::Tensor;
+
+Tensor& opt_mm_out(
+    RuntimeContext& ctx,
+    const Tensor& in,
+    const Tensor& mat2,
+    Tensor& out) {
+  ET_KERNEL_CHECK(ctx, check_mm_args(in, mat2, out), InvalidArgument, out);
+
+  size_t output_ndim = 0;
+  exec_aten::SizesType output_sizes[kTensorDimensionLimit];
+  get_mm_out_target_size(in, mat2, output_sizes, &output_ndim);
+  ET_KERNEL_CHECK(
+      ctx,
+      resize_tensor(out, {output_sizes, output_ndim}) == Error::Ok,
+      InvalidArgument,
+      out);
+
+  ET_SWITCH_REAL_TYPES_AND2(
+      Half, BFloat16, in.scalar_type(), ctx, "mm.out", CTYPE, [&]() {
+        size_t n = in.size(0);
+        size_t k = in.size(1);
+        size_t m = mat2.size(1);
+
+        // gemm expects column-major inputs and produces column-major
+        // output. So, we take advantage of the identity (A @ B).t()
+        // = B.t() @ A.t() here; row-major B is B.t() from gemm's
+        // column-major perspective, etc.
+        executorch::cpublas::gemm(
+            executorch::cpublas::TransposeType::NoTranspose,
+            executorch::cpublas::TransposeType::NoTranspose,
+            m,
+            n,
+            k,
+            static_cast<CTYPE>(1),
+            mat2.const_data_ptr<CTYPE>(),
+            m,
+            in.const_data_ptr<CTYPE>(),
+            k,
+            static_cast<CTYPE>(0),
+            out.mutable_data_ptr<CTYPE>(),
+            m);
+      });
+
+  return out;
+}
+
+} // namespace native
+} // namespace executor
+} // namespace torch

kernels/optimized/cpu/targets.bzl

@@ -52,6 +52,13 @@ _OPTIMIZED_ATEN_OPS = (
             ],
         }),
     ),
+    op_target(
+        name = "op_mm",
+        deps = [
+            "//executorch/kernels/optimized:libblas",
+            "//executorch/kernels/portable/cpu/util:matmul_ops_util",
+        ],
+    ),
     op_target(
         name = "op_mul",
         deps = [

kernels/optimized/optimized.yaml

@@ -52,6 +52,11 @@
     - arg_meta: null
       kernel_name: torch::executor::opt_le_tensor_out
 
+- op: mm.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::opt_mm_out
+
 - op: mul.out
   kernels:
     - arg_meta: null

kernels/test/targets.bzl

@@ -244,7 +244,7 @@ def define_common_targets():
     _common_op_test("op_mean_test", ["aten", "portable"])
     _common_op_test("op_min_test", ["aten", "portable"])
     _common_op_test("op_minimum_test", ["aten", "portable"])
-    _common_op_test("op_mm_test", ["aten", "portable"])
+    _common_op_test("op_mm_test", ["aten", "portable", "optimized"])
     _common_op_test("op_mul_test", ["aten", "portable", "optimized"])
     _common_op_test("op_narrow_copy_test", ["aten", "portable"])
     _common_op_test("op_native_batch_norm_test", ["aten", "portable"])