Automated rollback of commit 3708f53219aa2b201e82e7172c5064c1eb9d483b

PiperOrigin-RevId: 260136255

diff --git a/mlir/test/mlir-cpu-runner/cblas.cpp b/mlir/test/mlir-cpu-runner/cblas.cpp
index afc8a6d..d219b7b 100644 (file)
--- a/mlir/test/mlir-cpu-runner/cblas.cpp
+++ b/mlir/test/mlir-cpu-runner/cblas.cpp
@@ -21,7 +21,6 @@
 
 #include "include/cblas.h"
 #include <assert.h>
-#include <cstdint>
 
 extern "C" float cblas_sdot(const int N, const float *X, const int incX,
                             const float *Y, const int incY) {
@@ -31,79 +30,16 @@ extern "C" float cblas_sdot(const int N, const float *X, const int incX,
   return res;
 }
 
-#if defined(__GNUC__) || defined(__clang__)
-#define __RESTRICT__ __restrict__
-#define __ALIGN__(X) __attribute((aligned(X)))
-#define __NOINLINE__ __attribute((noinline))
-#elif defined(_MSC_VER)
-#define __RESTRICT__ __restrict
-#define __ALIGN__(X) __declspec(align(X))
-#define __NOINLINE__ __declspec(noinline)
-#endif
-
-// Implements a fastpath, specialized matrix multiplication where A, B and C are
-// assumed to be small enough to fit within L1 cache. A, B and C are assumed not
-// to alias and be aligned modulo 64.
-// This is an implementation meant to be compiled for vectorization with e.g.:
-//   `clang -mavx2 -mfma  -ffp-contract=fast`
-template <int VAL_M, int VAL_N, int VAL_K>
-__NOINLINE__ static void
-cblas_sgemm_impl(const float alpha, const float *__RESTRICT__ A __ALIGN__(64),
-                 const int lda, const float *__RESTRICT__ B __ALIGN__(64),
-                 const int ldb, const float beta,
-                 float *__RESTRICT__ C __ALIGN__(64), const int ldc) {
-  for (int m = 0; m < VAL_M; ++m) {
-    auto *pA = A + m * lda;
-    auto *pC = C + m * ldc;
-    float res[VAL_N];
-    for (int n = 0; n < VAL_N; ++n) {
-      res[n] = 0.0f;
-    }
-    for (int k = 0; k < VAL_K; ++k) {
-      auto *pB = B + k * ldb;
-      float a = pA[k];
-      for (int n = 0; n < VAL_N; ++n) {
-        res[n] += a * pB[n];
-      }
-    }
-    for (int n = 0; n < VAL_N; ++n) {
-      pC[n] = alpha * pC[n] + beta * res[n];
-    }
-  }
-}
-
-extern "C" void
-cblas_sgemm(const enum CBLAS_ORDER Order, const enum CBLAS_TRANSPOSE TransA,
-            const enum CBLAS_TRANSPOSE TransB, const int M, const int N,
-            const int K, const float alpha, const float *__RESTRICT__ A,
-            const int lda, const float *__RESTRICT__ B, const int ldb,
-            const float beta, float *__RESTRICT__ C, const int ldc) {
+extern "C" void cblas_sgemm(const enum CBLAS_ORDER Order,
+                            const enum CBLAS_TRANSPOSE TransA,
+                            const enum CBLAS_TRANSPOSE TransB, const int M,
+                            const int N, const int K, const float alpha,
+                            const float *A, const int lda, const float *B,
+                            const int ldb, const float beta, float *C,
+                            const int ldc) {
   assert(Order == CBLAS_ORDER::CblasRowMajor);
   assert(TransA == CBLAS_TRANSPOSE::CblasNoTrans);
   assert(TransB == CBLAS_TRANSPOSE::CblasNoTrans);
-
-  bool aligned64 = !(reinterpret_cast<std::uintptr_t>(A) & 63) &&
-                   !(reinterpret_cast<std::uintptr_t>(B) & 63) &&
-                   !(reinterpret_cast<std::uintptr_t>(C) & 63);
-  if (aligned64) {
-    if (M == 16 && N == 16 && K == 32)
-      return cblas_sgemm_impl<16, 16, 32>(alpha, A, lda, B, ldb, beta, C, ldc);
-    else if (M == 16 && N == 16 && K == 64)
-      return cblas_sgemm_impl<16, 16, 64>(alpha, A, lda, B, ldb, beta, C, ldc);
-    else if (M == 16 && N == 32 && K == 16)
-      return cblas_sgemm_impl<16, 32, 16>(alpha, A, lda, B, ldb, beta, C, ldc);
-    else if (M == 16 && N == 64 && K == 16)
-      return cblas_sgemm_impl<16, 64, 16>(alpha, A, lda, B, ldb, beta, C, ldc);
-
-    else if (M == 32 && N == 32 && K == 32)
-      return cblas_sgemm_impl<32, 32, 32>(alpha, A, lda, B, ldb, beta, C, ldc);
-    else if (M == 32 && N == 32 && K == 64)
-      return cblas_sgemm_impl<32, 32, 64>(alpha, A, lda, B, ldb, beta, C, ldc);
-    else if (M == 32 && N == 64 && K == 32)
-      return cblas_sgemm_impl<32, 64, 32>(alpha, A, lda, B, ldb, beta, C, ldc);
-  }
-
-  // Slow path.
   for (int m = 0; m < M; ++m) {
     auto *pA = A + m * lda;
     auto *pC = C + m * ldc;

diff --git a/mlir/test/mlir-cpu-runner/cblas_interface.cpp b/mlir/test/mlir-cpu-runner/cblas_interface.cpp
index dfb6cb5..1a63237 100644 (file)
--- a/mlir/test/mlir-cpu-runner/cblas_interface.cpp
+++ b/mlir/test/mlir-cpu-runner/cblas_interface.cpp
@@ -46,8 +46,8 @@ extern "C" void linalg_dot_impl(ViewType<float, 1> *X, ViewType<float, 1> *Y,
 
 extern "C" void linalg_matmul_impl(ViewType<float, 2> *A, ViewType<float, 2> *B,
                                    ViewType<float, 2> *C) {
-  assert(A->strides[1] == 1);
-  assert(A->strides[1] == 1);
+  assert(A->strides[1] == B->strides[1]);
+  assert(A->strides[1] == C->strides[1]);
   assert(A->strides[1] == 1);
   assert(A->sizes[0] >= A->strides[1]);
   assert(B->sizes[0] >= B->strides[1]);