Fix typos in END_FUNCTION arguments in ARM64 assembly kernels
Fix compilation errors with GNU ARM64 toolchain
PiperOrigin-RevId: 281967270
diff --git a/src/f32-gemm/1x8s4-neon.c b/src/f32-gemm/1x8s4-neon.c
index 6986389..884a6f4 100644
--- a/src/f32-gemm/1x8s4-neon.c
+++ b/src/f32-gemm/1x8s4-neon.c
@@ -7,6 +7,7 @@
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
+
#include <assert.h>
#include <arm_neon.h>
diff --git a/src/f32-gemm/1x8s4-neonfma.c b/src/f32-gemm/1x8s4-neonfma.c
index e993dd6..fb8b0af 100644
--- a/src/f32-gemm/1x8s4-neonfma.c
+++ b/src/f32-gemm/1x8s4-neonfma.c
@@ -7,6 +7,7 @@
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
+
#include <assert.h>
#include <arm_neon.h>
diff --git a/src/f32-gemm/4x8s4-neon.c b/src/f32-gemm/4x8s4-neon.c
index 7783666..e462dc0 100644
--- a/src/f32-gemm/4x8s4-neon.c
+++ b/src/f32-gemm/4x8s4-neon.c
@@ -7,6 +7,7 @@
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
+
#include <assert.h>
#include <arm_neon.h>
diff --git a/src/f32-gemm/4x8s4-neonfma.c b/src/f32-gemm/4x8s4-neonfma.c
index a5e37c1..88d9db4 100644
--- a/src/f32-gemm/4x8s4-neonfma.c
+++ b/src/f32-gemm/4x8s4-neonfma.c
@@ -7,6 +7,7 @@
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
+
#include <assert.h>
#include <arm_neon.h>
diff --git a/src/f32-gemm/5x8-aarch64-neonfma-cortex-a75.S b/src/f32-gemm/5x8-aarch64-neonfma-cortex-a75.S
index abd5f8f..e71982c 100644
--- a/src/f32-gemm/5x8-aarch64-neonfma-cortex-a75.S
+++ b/src/f32-gemm/5x8-aarch64-neonfma-cortex-a75.S
@@ -577,7 +577,7 @@
LDP d8, d9, [sp], 48
RET
-END_FUNCTION f32_gemm_ukernel_5x8__aarch64_neonfma_cortex_a75
+END_FUNCTION xnn_f32_gemm_ukernel_5x8__aarch64_neonfma_cortex_a75
#ifdef __ELF__
.section ".note.GNU-stack","",%progbits
diff --git a/src/f32-gemm/5x8-aarch64-neonfma-cortex-a75.S.in b/src/f32-gemm/5x8-aarch64-neonfma-cortex-a75.S.in
index f16a2b3..8edfd93 100644
--- a/src/f32-gemm/5x8-aarch64-neonfma-cortex-a75.S.in
+++ b/src/f32-gemm/5x8-aarch64-neonfma-cortex-a75.S.in
@@ -646,7 +646,7 @@
LDP d8, d9, [sp], 48
RET
-END_FUNCTION f32_gemm${"inc" if INC else ""}_ukernel_5x8__aarch64_neonfma_cortex_a75
+END_FUNCTION xnn_f32_gemm${"inc" if INC else ""}_ukernel_5x8__aarch64_neonfma_cortex_a75
#ifdef __ELF__
.section ".note.GNU-stack","",%progbits
diff --git a/src/f32-gemm/6x8-aarch64-neonfma-cortex-a73.S b/src/f32-gemm/6x8-aarch64-neonfma-cortex-a73.S
index 67b43a0..96e0e3a 100644
--- a/src/f32-gemm/6x8-aarch64-neonfma-cortex-a73.S
+++ b/src/f32-gemm/6x8-aarch64-neonfma-cortex-a73.S
@@ -655,7 +655,7 @@
LDP d8, d9, [sp], 64
RET
-END_FUNCTION f32_gemm_ukernel_6x8__aarch64_neonfma_cortex_a73
+END_FUNCTION xnn_f32_gemm_ukernel_6x8__aarch64_neonfma_cortex_a73
#ifdef __ELF__
.section ".note.GNU-stack","",%progbits
diff --git a/src/f32-gemm/6x8-aarch64-neonfma-cortex-a73.S.in b/src/f32-gemm/6x8-aarch64-neonfma-cortex-a73.S.in
index 23daaf6..3479796 100644
--- a/src/f32-gemm/6x8-aarch64-neonfma-cortex-a73.S.in
+++ b/src/f32-gemm/6x8-aarch64-neonfma-cortex-a73.S.in
@@ -734,7 +734,7 @@
LDP d8, d9, [sp], 64
RET
-END_FUNCTION f32_gemm${"inc" if INC else ""}_ukernel_6x8__aarch64_neonfma_cortex_a73
+END_FUNCTION xnn_f32_gemm${"inc" if INC else ""}_ukernel_6x8__aarch64_neonfma_cortex_a73
#ifdef __ELF__
.section ".note.GNU-stack","",%progbits
diff --git a/src/f32-gemm/6x8-aarch64-neonfma-cortex-a75.S b/src/f32-gemm/6x8-aarch64-neonfma-cortex-a75.S
index bd81473..ab56616 100644
--- a/src/f32-gemm/6x8-aarch64-neonfma-cortex-a75.S
+++ b/src/f32-gemm/6x8-aarch64-neonfma-cortex-a75.S
@@ -656,7 +656,7 @@
LDP d8, d9, [sp], 64
RET
-END_FUNCTION f32_gemm_ukernel_6x8__aarch64_neonfma_cortex_a75
+END_FUNCTION xnn_f32_gemm_ukernel_6x8__aarch64_neonfma_cortex_a75
#ifdef __ELF__
.section ".note.GNU-stack","",%progbits
diff --git a/src/f32-gemm/6x8-aarch64-neonfma-cortex-a75.S.in b/src/f32-gemm/6x8-aarch64-neonfma-cortex-a75.S.in
index 33dae71..4d5552c 100644
--- a/src/f32-gemm/6x8-aarch64-neonfma-cortex-a75.S.in
+++ b/src/f32-gemm/6x8-aarch64-neonfma-cortex-a75.S.in
@@ -735,7 +735,7 @@
LDP d8, d9, [sp], 64
RET
-END_FUNCTION f32_gemm${"inc" if INC else ""}_ukernel_6x8__aarch64_neonfma_cortex_a75
+END_FUNCTION xnn_f32_gemm${"inc" if INC else ""}_ukernel_6x8__aarch64_neonfma_cortex_a75
#ifdef __ELF__
.section ".note.GNU-stack","",%progbits
diff --git a/src/f32-gemm/6x8-aarch64-neonfma-ld128.S b/src/f32-gemm/6x8-aarch64-neonfma-ld128.S
index 6a1b035..433eee6 100644
--- a/src/f32-gemm/6x8-aarch64-neonfma-ld128.S
+++ b/src/f32-gemm/6x8-aarch64-neonfma-ld128.S
@@ -351,7 +351,7 @@
9:
RET
-END_FUNCTION xnn_f32_gemm_ukernel_6x8__aarch64_neonfma__ld128
+END_FUNCTION xnn_f32_gemm_ukernel_6x8__aarch64_neonfma_ld128
#ifdef __ELF__
.section ".note.GNU-stack","",%progbits
diff --git a/src/f32-gemm/6x8-aarch64-neonfma-ld128.S.in b/src/f32-gemm/6x8-aarch64-neonfma-ld128.S.in
index 8d18b74..1d9013c 100644
--- a/src/f32-gemm/6x8-aarch64-neonfma-ld128.S.in
+++ b/src/f32-gemm/6x8-aarch64-neonfma-ld128.S.in
@@ -424,7 +424,7 @@
9:
RET
-END_FUNCTION xnn_f32_gemm${"inc" if INC else ""}_ukernel_6x8__aarch64_neonfma__ld128
+END_FUNCTION xnn_f32_gemm${"inc" if INC else ""}_ukernel_6x8__aarch64_neonfma_ld128
#ifdef __ELF__
.section ".note.GNU-stack","",%progbits
diff --git a/src/f32-gemm/6x8-aarch64-neonfma-ld64.S b/src/f32-gemm/6x8-aarch64-neonfma-ld64.S
index 0c84b01..32c6420 100644
--- a/src/f32-gemm/6x8-aarch64-neonfma-ld64.S
+++ b/src/f32-gemm/6x8-aarch64-neonfma-ld64.S
@@ -284,7 +284,7 @@
8:
RET
-END_FUNCTION xnn_f32_gemm_ukernel_6x8__aarch64_neonfma__ld64
+END_FUNCTION xnn_f32_gemm_ukernel_6x8__aarch64_neonfma_ld64
#ifdef __ELF__
.section ".note.GNU-stack","",%progbits
diff --git a/src/f32-gemm/6x8-aarch64-neonfma-ld64.S.in b/src/f32-gemm/6x8-aarch64-neonfma-ld64.S.in
index eb0ecc2..c1b8950 100644
--- a/src/f32-gemm/6x8-aarch64-neonfma-ld64.S.in
+++ b/src/f32-gemm/6x8-aarch64-neonfma-ld64.S.in
@@ -357,7 +357,7 @@
8:
RET
-END_FUNCTION xnn_f32_gemm${"inc" if INC else ""}_ukernel_6x8__aarch64_neonfma__ld64
+END_FUNCTION xnn_f32_gemm${"inc" if INC else ""}_ukernel_6x8__aarch64_neonfma_ld64
#ifdef __ELF__
.section ".note.GNU-stack","",%progbits
diff --git a/src/f32-gemm/6x8s4-neon.c b/src/f32-gemm/6x8s4-neon.c
index 706d1f6..be25d07 100644
--- a/src/f32-gemm/6x8s4-neon.c
+++ b/src/f32-gemm/6x8s4-neon.c
@@ -7,6 +7,7 @@
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
+
#include <assert.h>
#include <arm_neon.h>
diff --git a/src/f32-gemm/6x8s4-neonfma.c b/src/f32-gemm/6x8s4-neonfma.c
index 9a1e7c5..310afec 100644
--- a/src/f32-gemm/6x8s4-neonfma.c
+++ b/src/f32-gemm/6x8s4-neonfma.c
@@ -7,6 +7,7 @@
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
+
#include <assert.h>
#include <arm_neon.h>
diff --git a/src/f32-gemm/8x8s4-neon.c b/src/f32-gemm/8x8s4-neon.c
index fe3b8a7..75e8c1f 100644
--- a/src/f32-gemm/8x8s4-neon.c
+++ b/src/f32-gemm/8x8s4-neon.c
@@ -7,6 +7,7 @@
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
+
#include <assert.h>
#include <arm_neon.h>
diff --git a/src/f32-gemm/8x8s4-neonfma.c b/src/f32-gemm/8x8s4-neonfma.c
index c2aeda4..e208a2e 100644
--- a/src/f32-gemm/8x8s4-neonfma.c
+++ b/src/f32-gemm/8x8s4-neonfma.c
@@ -7,6 +7,7 @@
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
+
#include <assert.h>
#include <arm_neon.h>
diff --git a/src/f32-gemminc/1x8s4-neon.c b/src/f32-gemminc/1x8s4-neon.c
index 4a6f670..c85e4e8 100644
--- a/src/f32-gemminc/1x8s4-neon.c
+++ b/src/f32-gemminc/1x8s4-neon.c
@@ -7,6 +7,7 @@
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
+
#include <assert.h>
#include <arm_neon.h>
diff --git a/src/f32-gemminc/1x8s4-neonfma.c b/src/f32-gemminc/1x8s4-neonfma.c
index c7dbc9f..ab988bd 100644
--- a/src/f32-gemminc/1x8s4-neonfma.c
+++ b/src/f32-gemminc/1x8s4-neonfma.c
@@ -7,6 +7,7 @@
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
+
#include <assert.h>
#include <arm_neon.h>
diff --git a/src/f32-gemminc/4x8s4-neon.c b/src/f32-gemminc/4x8s4-neon.c
index 80dfb29..e77cb53 100644
--- a/src/f32-gemminc/4x8s4-neon.c
+++ b/src/f32-gemminc/4x8s4-neon.c
@@ -7,6 +7,7 @@
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
+
#include <assert.h>
#include <arm_neon.h>
diff --git a/src/f32-gemminc/4x8s4-neonfma.c b/src/f32-gemminc/4x8s4-neonfma.c
index 2356442..ceab0ba 100644
--- a/src/f32-gemminc/4x8s4-neonfma.c
+++ b/src/f32-gemminc/4x8s4-neonfma.c
@@ -7,6 +7,7 @@
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
+
#include <assert.h>
#include <arm_neon.h>
diff --git a/src/f32-gemminc/5x8-aarch64-neonfma-cortex-a75.S b/src/f32-gemminc/5x8-aarch64-neonfma-cortex-a75.S
index 4eb7aaf..fe2afb1 100644
--- a/src/f32-gemminc/5x8-aarch64-neonfma-cortex-a75.S
+++ b/src/f32-gemminc/5x8-aarch64-neonfma-cortex-a75.S
@@ -574,7 +574,7 @@
LDP d8, d9, [sp], 48
RET
-END_FUNCTION f32_gemminc_ukernel_5x8__aarch64_neonfma_cortex_a75
+END_FUNCTION xnn_f32_gemminc_ukernel_5x8__aarch64_neonfma_cortex_a75
#ifdef __ELF__
.section ".note.GNU-stack","",%progbits
diff --git a/src/f32-gemminc/6x8-aarch64-neonfma-cortex-a73.S b/src/f32-gemminc/6x8-aarch64-neonfma-cortex-a73.S
index 9341b52..e0a842b 100644
--- a/src/f32-gemminc/6x8-aarch64-neonfma-cortex-a73.S
+++ b/src/f32-gemminc/6x8-aarch64-neonfma-cortex-a73.S
@@ -651,7 +651,7 @@
LDP d8, d9, [sp], 64
RET
-END_FUNCTION f32_gemminc_ukernel_6x8__aarch64_neonfma_cortex_a73
+END_FUNCTION xnn_f32_gemminc_ukernel_6x8__aarch64_neonfma_cortex_a73
#ifdef __ELF__
.section ".note.GNU-stack","",%progbits
diff --git a/src/f32-gemminc/6x8-aarch64-neonfma-cortex-a75.S b/src/f32-gemminc/6x8-aarch64-neonfma-cortex-a75.S
index b181ecf..004fde2 100644
--- a/src/f32-gemminc/6x8-aarch64-neonfma-cortex-a75.S
+++ b/src/f32-gemminc/6x8-aarch64-neonfma-cortex-a75.S
@@ -652,7 +652,7 @@
LDP d8, d9, [sp], 64
RET
-END_FUNCTION f32_gemminc_ukernel_6x8__aarch64_neonfma_cortex_a75
+END_FUNCTION xnn_f32_gemminc_ukernel_6x8__aarch64_neonfma_cortex_a75
#ifdef __ELF__
.section ".note.GNU-stack","",%progbits
diff --git a/src/f32-gemminc/6x8-aarch64-neonfma-ld128.S b/src/f32-gemminc/6x8-aarch64-neonfma-ld128.S
index 838043a..a4fa41e 100644
--- a/src/f32-gemminc/6x8-aarch64-neonfma-ld128.S
+++ b/src/f32-gemminc/6x8-aarch64-neonfma-ld128.S
@@ -347,7 +347,7 @@
9:
RET
-END_FUNCTION xnn_f32_gemminc_ukernel_6x8__aarch64_neonfma__ld128
+END_FUNCTION xnn_f32_gemminc_ukernel_6x8__aarch64_neonfma_ld128
#ifdef __ELF__
.section ".note.GNU-stack","",%progbits
diff --git a/src/f32-gemminc/6x8-aarch64-neonfma-ld64.S b/src/f32-gemminc/6x8-aarch64-neonfma-ld64.S
index bfe514a..bb1e374 100644
--- a/src/f32-gemminc/6x8-aarch64-neonfma-ld64.S
+++ b/src/f32-gemminc/6x8-aarch64-neonfma-ld64.S
@@ -280,7 +280,7 @@
8:
RET
-END_FUNCTION xnn_f32_gemminc_ukernel_6x8__aarch64_neonfma__ld64
+END_FUNCTION xnn_f32_gemminc_ukernel_6x8__aarch64_neonfma_ld64
#ifdef __ELF__
.section ".note.GNU-stack","",%progbits
diff --git a/src/f32-gemminc/6x8s4-neon.c b/src/f32-gemminc/6x8s4-neon.c
index 6340245..e6b9ce3 100644
--- a/src/f32-gemminc/6x8s4-neon.c
+++ b/src/f32-gemminc/6x8s4-neon.c
@@ -7,6 +7,7 @@
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
+
#include <assert.h>
#include <arm_neon.h>
diff --git a/src/f32-gemminc/6x8s4-neonfma.c b/src/f32-gemminc/6x8s4-neonfma.c
index ccc6a38..dc75c2f 100644
--- a/src/f32-gemminc/6x8s4-neonfma.c
+++ b/src/f32-gemminc/6x8s4-neonfma.c
@@ -7,6 +7,7 @@
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
+
#include <assert.h>
#include <arm_neon.h>
diff --git a/src/f32-gemminc/8x8s4-neon.c b/src/f32-gemminc/8x8s4-neon.c
index 77c6228..a072d6d 100644
--- a/src/f32-gemminc/8x8s4-neon.c
+++ b/src/f32-gemminc/8x8s4-neon.c
@@ -7,6 +7,7 @@
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
+
#include <assert.h>
#include <arm_neon.h>
diff --git a/src/f32-gemminc/8x8s4-neonfma.c b/src/f32-gemminc/8x8s4-neonfma.c
index 0b169f0..b611084 100644
--- a/src/f32-gemminc/8x8s4-neonfma.c
+++ b/src/f32-gemminc/8x8s4-neonfma.c
@@ -7,6 +7,7 @@
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
+
#include <assert.h>
#include <arm_neon.h>