Allow opt into GEMM -> GEMV forwarding

Makefile.system

@@ -274,9 +274,18 @@ endif
 ifeq ($(ARCH), loongarch64)
 SMALL_MATRIX_OPT = 1
 endif
+ifeq ($(ARCH), arm64)
+GEMM_GEMV_FORWARD = 1
+endif
+
 ifeq ($(SMALL_MATRIX_OPT), 1)
 CCOMMON_OPT += -DSMALL_MATRIX_OPT
 endif
+ifeq ($(GEMM_GEMV_FORWARD), 1)
+ifneq ($(ONLY_CBLAS), 1)
+CCOMMON_OPT += -DGEMM_GEMV_FORWARD
+endif
+endif
 
 # This operation is expensive, so execution should be once.
 ifndef GOTOBLAS_MAKEFILE

cmake/system.cmake

@@ -391,6 +391,13 @@ endif ()
 if (X86_64 OR ${CORE} STREQUAL POWER10)
   set(SMALL_MATRIX_OPT TRUE)
 endif ()
+if (ARM64)
+  set(GEMM_GEMV_FORWARD TRUE)
+endif ()
+
+if (GEMM_GEMV_FORWARD AND NOT ONLY_CBLAS)
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DGEMM_GEMV_FORWARD")
+endif ()
 if (SMALL_MATRIX_OPT)
   set(CCOMMON_OPT "${CCOMMON_OPT} -DSMALL_MATRIX_OPT")
 endif ()

interface/gemm.c

@@ -498,7 +498,7 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANS
 	 args.m, args.n, args.k, args.lda, args.ldb, args.ldc);
 #endif
 
-#if !defined(GEMM3M) && !defined(COMPLEX)
+#if defined(GEMM_GEMV_FORWARD) && !defined(GEMM3M) && !defined(COMPLEX)
   // Check if we can convert GEMM -> GEMV
   if (args.k != 0) {
     if (args.n == 1) {