ref: e43d14d55409892015ba7ec66b7a03c549700e36
parent: 6fe2299370b3ad727f11b0cd185b952ca34ab3e5
parent: 97f795c4b596f863f255bde4f64c9d7c1c711d58
author: guangwei <GuangweiWang@users.noreply.github.com>
date: Mon Dec 23 10:48:36 EST 2019
Merge pull request #3225 from gxw-loongson/master Adjust the mmi/msa detection mode for mips platform.
--- a/build/arch.mk
+++ b/build/arch.mk
@@ -30,14 +30,26 @@
endif
endif
-#for loongson
+#for mips
ifneq ($(filter mips mips64, $(ARCH)),)
ifeq ($(USE_ASM), Yes)
+ENABLE_MMI=Yes
+ENABLE_MSA=Yes
ASM_ARCH = mips
ASMFLAGS += -I$(SRC_PATH)codec/common/mips/
-LOONGSON3A = $(shell g++ -dM -E - < /dev/null | grep '_MIPS_TUNE ' | cut -f 3 -d " ")
-ifeq ($(LOONGSON3A), "loongson3a")
-CFLAGS += -DHAVE_MMI
+#mmi
+ifeq ($(ENABLE_MMI), Yes)
+ENABLE_MMI = $(shell $(SRC_PATH)build/mips-simd-check.sh $(CC) mmi)
+ifeq ($(ENABLE_MMI), Yes)
+CFLAGS += -DHAVE_MMI -Wa,-mloongson-mmi,-mloongson-ext
+endif
+endif
+#msa
+ifeq ($(ENABLE_MSA), Yes)
+ENABLE_MSA = $(shell $(SRC_PATH)build/mips-simd-check.sh $(CC) msa)
+ifeq ($(ENABLE_MSA), Yes)
+CFLAGS += -DHAVE_MSA -mmsa
+endif
endif
endif
endif
--- /dev/null
+++ b/build/mips-simd-check.sh
@@ -1,0 +1,32 @@
+#!/bin/bash
+#**********************************************************************************
+# This script is using in build/arch.mk for mips to detect the simd instructions:
+# mmi, msa (maybe more in the future).
+#
+# --usage:
+# ./mips-simd-check.sh $(CC) mmi
+# or ./mips-simd-check.sh $(CC) msa
+#
+# date: 10/17/2019 Created
+#**********************************************************************************
+
+TMPC=$(mktemp tmp.XXXXXX.c)
+TMPO=$(mktemp tmp.XXXXXX.o)
+if [ $2 == "mmi" ]
+then
+ echo "void main(void){ __asm__ volatile(\"punpcklhw \$f0, \$f0, \$f0\"); }" > $TMPC+ $1 -Wa,-mloongson-mmi $TMPC -o $TMPO &> /dev/null
+ if test -s $TMPO
+ then
+ echo "Yes"
+ fi
+elif [ $2 == "msa" ]
+then
+ echo "void main(void){ __asm__ volatile(\"addvi.b \$w0, \$w1, 1\"); }" > $TMPC+ $1 -mmsa $TMPC -o $TMPO &> /dev/null
+ if test -s $TMPO
+ then
+ echo "Yes"
+ fi
+fi
+rm -f $TMPC $TMPO
--- a/build/mktargets.py
+++ b/build/mktargets.py
@@ -119,9 +119,9 @@
armfiles.append(file)
mipsfiles = []
for file in cfiles:
- c = file.split('/')- if 'mips' in c:
- mipsfiles.append(file)
+ c = file.split('/')+ if 'mips' in c:
+ mipsfiles.append(file)
cfiles = [x for x in cfiles if x not in mipsfiles]
@@ -181,15 +181,34 @@
f.write("OBJS += $(%s_OBJSARM64)\n\n"%(PREFIX))if len(mipsfiles) > 0:
- f.write("%s_ASM_MIPS_SRCS=\\\n"%(PREFIX))- for c in mipsfiles:
- f.write("\t$(%s_SRCDIR)/%s\\\n"%(PREFIX, c))- f.write("\n")- f.write("%s_OBJSMIPS += $(%s_ASM_MIPS_SRCS:.c=.$(OBJ))\n"%(PREFIX, PREFIX))- f.write("ifeq ($(ASM_ARCH), mips)\n")- f.write("%s_OBJS += $(%s_OBJSMIPS)\n"%(PREFIX,PREFIX))- f.write("endif\n")- f.write("OBJS += $(%s_OBJSMIPS)\n\n"%(PREFIX))+ mmifiles = []
+ for file in mipsfiles:
+ if '_mmi' in file:
+ mmifiles.append(file)
+ f.write("%s_ASM_MIPS_MMI_SRCS=\\\n"%(PREFIX))+ for c in mmifiles:
+ f.write("\t$(%s_SRCDIR)/%s\\\n"%(PREFIX, c))+ f.write("\n")+ f.write("%s_OBJSMIPS_MMI += $(%s_ASM_MIPS_MMI_SRCS:.c=.$(OBJ))\n\n"%(PREFIX, PREFIX))+ msafiles = []
+ for file in mipsfiles:
+ if '_msa' in file:
+ msafiles.append(file)
+ f.write("%s_ASM_MIPS_MSA_SRCS=\\\n"%(PREFIX))+ for c in msafiles:
+ f.write("\t$(%s_SRCDIR)/%s\\\n"%(PREFIX, c))+ f.write("\n")+ f.write("%s_OBJSMIPS_MSA += $(%s_ASM_MIPS_MSA_SRCS:.c=.$(OBJ))\n"%(PREFIX, PREFIX))+ f.write("ifeq ($(ASM_ARCH), mips)\n")+ f.write("ifeq ($(ENABLE_MMI), Yes)\n")+ f.write("%s_OBJS += $(%s_OBJSMIPS_MMI)\n"%(PREFIX,PREFIX))+ f.write("endif\n")+ f.write("ifeq ($(ENABLE_MSA), Yes)\n")+ f.write("%s_OBJS += $(%s_OBJSMIPS_MSA)\n"%(PREFIX,PREFIX))+ f.write("endif\n")+ f.write("endif\n")+ f.write("OBJS += $(%s_OBJSMIPS_MMI)\n"%(PREFIX))+ f.write("OBJS += $(%s_OBJSMIPS_MSA)\n\n"%(PREFIX)) f.write("OBJS += $(%s_OBJS)\n\n"%(PREFIX))write_cpp_rule_pattern(f)
--- a/codec/common/inc/cpu_core.h
+++ b/codec/common/inc/cpu_core.h
@@ -86,6 +86,7 @@
/* For loongson */
#define WELS_CPU_MMI 0x00000001 /* mmi */
+#define WELS_CPU_MSA 0x00000002 /* msa */
/*
* Interfaces for CPU core feature detection as below
--- a/codec/common/src/cpu.cpp
+++ b/codec/common/src/cpu.cpp
@@ -309,12 +309,45 @@
#elif defined(mips)
/* for loongson */
+static uint32_t get_cpu_flags_from_cpuinfo(void)
+{+ uint32_t flags = 0;
+
+# ifdef __linux__
+ FILE* fp = fopen("/proc/cpuinfo", "r");+ if (!fp)
+ return flags;
+
+ char buf[200];
+ memset(buf, 0, sizeof(buf));
+ while (fgets(buf, sizeof(buf), fp)) {+ if (!strncmp(buf, "model name", strlen("model name"))) {+ if (strstr(buf, "Loongson-3A") || strstr(buf, "Loongson-3B") ||
+ strstr(buf, "Loongson-2K")) {+ flags |= WELS_CPU_MMI;
+ }
+ break;
+ }
+ }
+ while (fgets(buf, sizeof(buf), fp)) {+ if(!strncmp(buf, "ASEs implemented", strlen("ASEs implemented"))) {+ if (strstr(buf, "loongson-mmi") || strstr(buf, "loongson-ext")) {+ flags |= WELS_CPU_MMI;
+ }
+ if (strstr(buf, "msa")) {+ flags |= WELS_CPU_MSA;
+ }
+ break;
+ }
+ }
+ fclose(fp);
+# endif
+
+ return flags;
+}
+
uint32_t WelsCPUFeatureDetect (int32_t* pNumberOfLogicProcessors) {-#if defined(HAVE_MMI)
- return WELS_CPU_MMI;
-#else
- return 0;
-#endif
+ return get_cpu_flags_from_cpuinfo();
}
#else /* Neither X86_ASM, HAVE_NEON, HAVE_NEON_AARCH64 nor mips */
@@ -324,5 +357,3 @@
}
#endif
-
-
--- a/codec/common/targets.mk
+++ b/codec/common/targets.mk
@@ -66,7 +66,7 @@
endif
OBJS += $(COMMON_OBJSARM64)
-COMMON_ASM_MIPS_SRCS=\
+COMMON_ASM_MIPS_MMI_SRCS=\
$(COMMON_SRCDIR)/mips/copy_mb_mmi.c\
$(COMMON_SRCDIR)/mips/deblock_mmi.c\
$(COMMON_SRCDIR)/mips/expand_picture_mmi.c\
@@ -73,11 +73,21 @@
$(COMMON_SRCDIR)/mips/intra_pred_com_mmi.c\
$(COMMON_SRCDIR)/mips/satd_sad_mmi.c\
-COMMON_OBJSMIPS += $(COMMON_ASM_MIPS_SRCS:.c=.$(OBJ))
+COMMON_OBJSMIPS_MMI += $(COMMON_ASM_MIPS_MMI_SRCS:.c=.$(OBJ))
+
+COMMON_ASM_MIPS_MSA_SRCS=\
+
+COMMON_OBJSMIPS_MSA += $(COMMON_ASM_MIPS_MSA_SRCS:.c=.$(OBJ))
ifeq ($(ASM_ARCH), mips)
-COMMON_OBJS += $(COMMON_OBJSMIPS)
+ifeq ($(ENABLE_MMI), Yes)
+COMMON_OBJS += $(COMMON_OBJSMIPS_MMI)
endif
-OBJS += $(COMMON_OBJSMIPS)
+ifeq ($(ENABLE_MSA), Yes)
+COMMON_OBJS += $(COMMON_OBJSMIPS_MSA)
+endif
+endif
+OBJS += $(COMMON_OBJSMIPS_MMI)
+OBJS += $(COMMON_OBJSMIPS_MSA)
OBJS += $(COMMON_OBJS)
--- a/codec/decoder/targets.mk
+++ b/codec/decoder/targets.mk
@@ -57,14 +57,24 @@
endif
OBJS += $(DECODER_OBJSARM64)
-DECODER_ASM_MIPS_SRCS=\
+DECODER_ASM_MIPS_MMI_SRCS=\
$(DECODER_SRCDIR)/core/mips/dct_mmi.c\
-DECODER_OBJSMIPS += $(DECODER_ASM_MIPS_SRCS:.c=.$(OBJ))
+DECODER_OBJSMIPS_MMI += $(DECODER_ASM_MIPS_MMI_SRCS:.c=.$(OBJ))
+
+DECODER_ASM_MIPS_MSA_SRCS=\
+
+DECODER_OBJSMIPS_MSA += $(DECODER_ASM_MIPS_MSA_SRCS:.c=.$(OBJ))
ifeq ($(ASM_ARCH), mips)
-DECODER_OBJS += $(DECODER_OBJSMIPS)
+ifeq ($(ENABLE_MMI), Yes)
+DECODER_OBJS += $(DECODER_OBJSMIPS_MMI)
endif
-OBJS += $(DECODER_OBJSMIPS)
+ifeq ($(ENABLE_MSA), Yes)
+DECODER_OBJS += $(DECODER_OBJSMIPS_MSA)
+endif
+endif
+OBJS += $(DECODER_OBJSMIPS_MMI)
+OBJS += $(DECODER_OBJSMIPS_MSA)
OBJS += $(DECODER_OBJS)
--- a/codec/encoder/targets.mk
+++ b/codec/encoder/targets.mk
@@ -82,16 +82,26 @@
endif
OBJS += $(ENCODER_OBJSARM64)
-ENCODER_ASM_MIPS_SRCS=\
+ENCODER_ASM_MIPS_MMI_SRCS=\
$(ENCODER_SRCDIR)/core/mips/dct_mmi.c\
$(ENCODER_SRCDIR)/core/mips/quant_mmi.c\
$(ENCODER_SRCDIR)/core/mips/score_mmi.c\
-ENCODER_OBJSMIPS += $(ENCODER_ASM_MIPS_SRCS:.c=.$(OBJ))
+ENCODER_OBJSMIPS_MMI += $(ENCODER_ASM_MIPS_MMI_SRCS:.c=.$(OBJ))
+
+ENCODER_ASM_MIPS_MSA_SRCS=\
+
+ENCODER_OBJSMIPS_MSA += $(ENCODER_ASM_MIPS_MSA_SRCS:.c=.$(OBJ))
ifeq ($(ASM_ARCH), mips)
-ENCODER_OBJS += $(ENCODER_OBJSMIPS)
+ifeq ($(ENABLE_MMI), Yes)
+ENCODER_OBJS += $(ENCODER_OBJSMIPS_MMI)
endif
-OBJS += $(ENCODER_OBJSMIPS)
+ifeq ($(ENABLE_MSA), Yes)
+ENCODER_OBJS += $(ENCODER_OBJSMIPS_MSA)
+endif
+endif
+OBJS += $(ENCODER_OBJSMIPS_MMI)
+OBJS += $(ENCODER_OBJSMIPS_MSA)
OBJS += $(ENCODER_OBJS)
--- a/codec/processing/targets.mk
+++ b/codec/processing/targets.mk
@@ -58,14 +58,24 @@
endif
OBJS += $(PROCESSING_OBJSARM64)
-PROCESSING_ASM_MIPS_SRCS=\
+PROCESSING_ASM_MIPS_MMI_SRCS=\
$(PROCESSING_SRCDIR)/src/mips/vaa_mmi.c\
-PROCESSING_OBJSMIPS += $(PROCESSING_ASM_MIPS_SRCS:.c=.$(OBJ))
+PROCESSING_OBJSMIPS_MMI += $(PROCESSING_ASM_MIPS_MMI_SRCS:.c=.$(OBJ))
+
+PROCESSING_ASM_MIPS_MSA_SRCS=\
+
+PROCESSING_OBJSMIPS_MSA += $(PROCESSING_ASM_MIPS_MSA_SRCS:.c=.$(OBJ))
ifeq ($(ASM_ARCH), mips)
-PROCESSING_OBJS += $(PROCESSING_OBJSMIPS)
+ifeq ($(ENABLE_MMI), Yes)
+PROCESSING_OBJS += $(PROCESSING_OBJSMIPS_MMI)
endif
-OBJS += $(PROCESSING_OBJSMIPS)
+ifeq ($(ENABLE_MSA), Yes)
+PROCESSING_OBJS += $(PROCESSING_OBJSMIPS_MSA)
+endif
+endif
+OBJS += $(PROCESSING_OBJSMIPS_MMI)
+OBJS += $(PROCESSING_OBJSMIPS_MSA)
OBJS += $(PROCESSING_OBJS)