[ExecuTorch] Support BFloat16 in CPUBlas gemm

kernels/optimized/blas/CPUBlas.cpp

@@ -173,5 +173,28 @@ void gemm(
 }
 // clang-format on
 
+// clang-format off
+void gemm(
+    TransposeType transa, TransposeType transb,
+    int64_t m, int64_t n, int64_t k,
+    const BFloat16 alpha,
+    const BFloat16 *a, int64_t lda,
+    const BFloat16 *b, int64_t ldb,
+    const BFloat16 beta,
+    BFloat16 *c, int64_t ldc) {
+  normalize_last_dims(transa, transb, m, n, k, &lda, &ldb, &ldc);
+
+  using acc_type = utils::compute_dtype<BFloat16>;
+  gemm_impl(
+      transa, transb,
+      m, n, k,
+      static_cast<const acc_type>(alpha),
+      a, lda,
+      b, ldb,
+      static_cast<const acc_type>(beta),
+      c, ldc);
+}
+// clang-format on
+
 } // namespace cpublas
 } // namespace executorch

kernels/optimized/blas/CPUBlas.h

@@ -17,6 +17,7 @@
 namespace executorch {
 namespace cpublas {
 
+using BFloat16 = torch::executor::BFloat16;
 using Half = torch::executor::Half;
 
 enum class TransposeType {
@@ -104,6 +105,15 @@ void gemm(
     const Half *b, int64_t ldb,
     const Half beta,
     Half *c, int64_t ldc);
+
+void gemm(
+    TransposeType transa, TransposeType transb,
+    int64_t m, int64_t n, int64_t k,
+    const BFloat16 alpha,
+    const BFloat16 *a, int64_t lda,
+    const BFloat16 *b, int64_t ldb,
+    const BFloat16 beta,
+    BFloat16 *c, int64_t ldc);
 // clang-format on
 
 // clang-format off

kernels/optimized/test/libblas_test.cpp

@@ -9,6 +9,7 @@
 #include <gtest/gtest.h>
 
 #include <executorch/kernels/optimized/blas/CPUBlas.h>
+#include <executorch/runtime/core/exec_aten/exec_aten.h>
 
 #include <vector>
 
@@ -17,7 +18,8 @@
   _<float, N>();                           \
   _<int64_t, N>();                         \
   _<uint8_t, N>();                         \
-  _<int32_t, N>();
+  _<int32_t, N>();                         \
+  _<exec_aten::BFloat16, N>();
 
 namespace {