Fix crashes on 2k3000 (#3897)

* Fix tests crash on 2k3000

Forgot to add runtime check here, resulting in crash happend on 2k3000.

* Not apply -mflags globaly on loongarch

This can fix crashes caused by automatic vectorization.

---------

Co-authored-by: jinboson <[email protected]>

Author: jinboson

Makefile

@@ -115,6 +115,30 @@ ifneq ($(V),Yes)
     QUIET      = @
 endif
 
+### Not apply -mflags globaly
+ifeq (loongarch64, $(ARCH))
+LSX_OBJS := \
+    $(SRC_PATH)codec/common/loongarch/copy_mb_lsx.o \
+    $(SRC_PATH)codec/common/loongarch/deblock_lsx.o \
+    $(SRC_PATH)codec/common/loongarch/intra_pred_com_lsx.o \
+    $(SRC_PATH)codec/common/loongarch/mc_chroma_lsx.o \
+    $(SRC_PATH)codec/common/loongarch/mc_horver_lsx.o \
+    $(SRC_PATH)codec/decoder/core/loongarch/mb_aux_lsx.o \
+    $(SRC_PATH)codec/encoder/core/loongarch/get_intra_predictor_lsx.o \
+    $(SRC_PATH)codec/encoder/core/loongarch/quant_lsx.o \
+    $(SRC_PATH)codec/encoder/core/loongarch/svc_motion_estimate_lsx.o \
+    $(SRC_PATH)codec/processing/src/loongarch/vaa_lsx.o
+
+LASX_OBJS := \
+    $(SRC_PATH)codec/common/loongarch/intra_pred_com_lasx.o \
+    $(SRC_PATH)codec/common/loongarch/satd_sad_lasx.o \
+    $(SRC_PATH)codec/encoder/core/loongarch/dct_lasx.o \
+    $(SRC_PATH)codec/encoder/core/loongarch/sample_lasx.o \
+    $(SRC_PATH)codec/processing/src/loongarch/vaa_lasx.o
+
+$(LSX_OBJS): CFLAGS += -mlsx
+$(LASX_OBJS): CFLAGS += -mlasx
+endif
 
 INCLUDES += -I$(SRC_PATH)codec/api/wels -I$(SRC_PATH)codec/common/inc -Icodec/common/inc
 

build/arch.mk

@@ -65,14 +65,14 @@ ASMFLAGS += -I$(SRC_PATH)codec/common/loongarch/
 ifeq ($(ENABLE_LSX), Yes)
 ENABLE_LSX = $(shell $(SRC_PATH)build/loongarch-simd-check.sh $(CC) lsx)
 ifeq ($(ENABLE_LSX), Yes)
-CFLAGS += -DHAVE_LSX -mlsx
+CFLAGS += -DHAVE_LSX
 endif
 endif
 #lasx
 ifeq ($(ENABLE_LASX), Yes)
 ENABLE_LASX = $(shell $(SRC_PATH)build/loongarch-simd-check.sh $(CC) lasx)
 ifeq ($(ENABLE_LASX), Yes)
-CFLAGS += -DHAVE_LASX -mlasx
+CFLAGS += -DHAVE_LASX
 endif
 endif
 endif

test/encoder/EncUT_DecodeMbAux.cpp

@@ -253,7 +253,8 @@ TEST (DecodeMbAuxTest, WelsIDctT4Rec_mmi) {
 #endif
 #if defined(HAVE_LASX)
 TEST (DecodeMbAuxTest, WelsIDctT4Rec_lasx) {
-  TestIDctT4Rec<int16_t> (WelsIDctT4Rec_lasx);
+  if (WelsCPUFeatureDetect (0) & WELS_CPU_LASX)
+    TestIDctT4Rec<int16_t> (WelsIDctT4Rec_lasx);
 }
 #endif
 template<typename clip_t>
@@ -428,6 +429,7 @@ TEST (DecodeMbAuxTest, WelsIDctRecI16x16Dc_mmi) {
 #endif
 #if defined(HAVE_LASX)
 TEST (DecodeMbAuxTest, WelsIDctFourT4Rec_lasx) {
-  TestIDctFourT4Rec<int16_t> (WelsIDctFourT4Rec_lasx);
+  if (WelsCPUFeatureDetect (0) & WELS_CPU_LASX)
+    TestIDctFourT4Rec<int16_t> (WelsIDctFourT4Rec_lasx);
 }
 #endif

test/encoder/EncUT_EncoderMbAux.cpp

@@ -294,11 +294,13 @@ TEST (EncodeMbAuxTest, WelsCalculateSingleCtr4x4_mmi) {
 #endif
 #ifdef HAVE_LASX
 TEST (EncodeMbAuxTest, WelsDctT4_lasx) {
-  TestDctT4 (WelsDctT4_lasx);
+  if (WelsCPUFeatureDetect (0) & WELS_CPU_LASX)
+    TestDctT4 (WelsDctT4_lasx);
 }
 
 TEST (EncodeMbAuxTest, WelsDctFourT4_lasx) {
-  TestDctFourT4 (WelsDctFourT4_lasx);
+  if (WelsCPUFeatureDetect (0) & WELS_CPU_LASX)
+    TestDctFourT4 (WelsDctFourT4_lasx);
 }
 #endif