From: Erik Niemeyer <erik.a.niemeyer@gmail.com>
Date: Tue, 29 Oct 2013 15:48:12 +0000 (-0600)
Subject: CL for adding AVX-AVX2 support in libvpx.
X-Git-Tag: v1.3.0~117^2
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=e6863ef3185faf90e7be2c7f2c90bce0c08da1eb;p=libvpx

CL for adding AVX-AVX2 support in libvpx.

Change-Id: Idc03f3fca4bf2d0afd33631ea1d3caf8fc34ec29
---

diff --git a/build/make/Makefile b/build/make/Makefile
index 7a2523960..030c1b57f 100644
--- a/build/make/Makefile
+++ b/build/make/Makefile
@@ -114,6 +114,10 @@ $(BUILD_PFX)%_ssse3.c.d: CFLAGS += -mssse3
 $(BUILD_PFX)%_ssse3.c.o: CFLAGS += -mssse3
 $(BUILD_PFX)%_sse4.c.d: CFLAGS += -msse4.1
 $(BUILD_PFX)%_sse4.c.o: CFLAGS += -msse4.1
+$(BUILD_PFX)%_avx.c.d: CFLAGS += -mavx
+$(BUILD_PFX)%_avx.c.o: CFLAGS += -mavx
+$(BUILD_PFX)%_avx2.c.d: CFLAGS += -mavx2
+$(BUILD_PFX)%_avx2.c.o: CFLAGS += -mavx2
 
 $(BUILD_PFX)%.c.d: %.c
 	$(if $(quiet),@echo "    [DEP] $@")
diff --git a/build/make/configure.sh b/build/make/configure.sh
index c6c866079..83f480a42 100755
--- a/build/make/configure.sh
+++ b/build/make/configure.sh
@@ -1108,6 +1108,18 @@ EOF
             soft_enable sse4_1
         fi
 
+        if enabled gcc && ! disabled avx && ! check_cflags -mavx; then
+            RTCD_OPTIONS="${RTCD_OPTIONS}--disable-avx "
+        else
+            soft_enable avx
+        fi
+
+        if enabled gcc && ! disabled avx2 && ! check_cflags -mavx2; then
+            RTCD_OPTIONS="${RTCD_OPTIONS}--disable-avx2 "
+        else
+            soft_enable avx2
+        fi
+
         case "${AS}" in
             auto|"")
                 which nasm >/dev/null 2>&1 && AS=nasm
diff --git a/build/make/rtcd.sh b/build/make/rtcd.sh
index c531e95a1..2967b5aed 100755
--- a/build/make/rtcd.sh
+++ b/build/make/rtcd.sh
@@ -327,11 +327,11 @@ EOF
 require c
 case $arch in
   x86)
-    ALL_ARCHS=$(filter mmx sse sse2 sse3 ssse3 sse4_1)
+    ALL_ARCHS=$(filter mmx sse sse2 sse3 ssse3 sse4_1 avx avx2)
     x86
     ;;
   x86_64)
-    ALL_ARCHS=$(filter mmx sse sse2 sse3 ssse3 sse4_1)
+    ALL_ARCHS=$(filter mmx sse sse2 sse3 ssse3 sse4_1 avx avx2)
     REQUIRES=${REQUIRES:-mmx sse sse2}
     require $(filter $REQUIRES)
     x86
diff --git a/configure b/configure
index 297cec463..45d5c1204 100755
--- a/configure
+++ b/configure
@@ -234,6 +234,8 @@ ARCH_EXT_LIST="
     sse3
     ssse3
     sse4_1
+    avx
+    avx2
 
     altivec
 "
@@ -422,7 +424,7 @@ process_targets() {
     fi
 
     # The write_common_config (config.mk) logic is deferred until after the
-    # recursive calls to configure complete, becuase we want our universal
+    # recursive calls to configure complete, because we want our universal
     # targets to be executed last.
     write_common_config_targets
     enabled universal && echo "FAT_ARCHS=${fat_bin_archs}" >> config.mk
diff --git a/test/test_libvpx.cc b/test/test_libvpx.cc
index a4dbca4cd..80aca98b3 100644
--- a/test/test_libvpx.cc
+++ b/test/test_libvpx.cc
@@ -45,6 +45,10 @@ int main(int argc, char **argv) {
     append_gtest_filter(":-SSSE3/*");
   if (!(simd_caps & HAS_SSE4_1))
     append_gtest_filter(":-SSE4_1/*");
+  if (!(simd_caps & HAS_AVX))
+    append_gtest_filter(":-AVX/*");
+  if (!(simd_caps & HAS_AVX2))
+    append_gtest_filter(":-AVX2/*");
 #endif
 
 #if !CONFIG_SHARED
diff --git a/vp9/common/vp9_rtcd_defs.sh b/vp9/common/vp9_rtcd_defs.sh
index 2d9fbff97..862933826 100644
--- a/vp9/common/vp9_rtcd_defs.sh
+++ b/vp9/common/vp9_rtcd_defs.sh
@@ -22,10 +22,11 @@ forward_decls vp9_common_forward_decls
 
 # x86inc.asm doesn't work if pic is enabled on 32 bit platforms so no assembly.
 [ "$CONFIG_USE_X86INC" = "yes" ] && mmx_x86inc=mmx && sse_x86inc=sse &&
-  sse2_x86inc=sse2 && ssse3_x86inc=ssse3
+  sse2_x86inc=sse2 && ssse3_x86inc=ssse3 && avx_x86inc=avx && avx2_x86inc=avx2
 
 # this variable is for functions that are 64 bit only.
-[ $arch = "x86_64" ] && mmx_x86_64=mmx && sse2_x86_64=sse2 && ssse3_x86_64=ssse3
+[ $arch = "x86_64" ] && mmx_x86_64=mmx && sse2_x86_64=sse2 && 
+  ssse3_x86_64=ssse3 && avx_x86_64=avx && avx2_x86_64=avx2
 
 #
 # RECON
diff --git a/vpx_ports/x86.h b/vpx_ports/x86.h
index b009c3560..299058347 100644
--- a/vpx_ports/x86.h
+++ b/vpx_ports/x86.h
@@ -88,12 +88,14 @@ void __cpuid(int CPUInfo[4], int info_type);
 #endif
 #endif /* end others */
 
-#define HAS_MMX   0x01
-#define HAS_SSE   0x02
-#define HAS_SSE2  0x04
-#define HAS_SSE3  0x08
-#define HAS_SSSE3 0x10
-#define HAS_SSE4_1 0x20
+#define HAS_MMX     0x01
+#define HAS_SSE     0x02
+#define HAS_SSE2    0x04
+#define HAS_SSE3    0x08
+#define HAS_SSSE3   0x10
+#define HAS_SSE4_1  0x20
+#define HAS_AVX     0x40
+#define HAS_AVX2    0x80
 #ifndef BIT
 #define BIT(n) (1<<n)
 #endif
@@ -132,12 +134,16 @@ x86_simd_caps(void) {
 
   if (reg_edx & BIT(26)) flags |= HAS_SSE2; /* aka wmt */
 
-  if (reg_ecx & BIT(0))  flags |= HAS_SSE3;
+  if (reg_ecx & BIT(0)) flags |= HAS_SSE3;
 
-  if (reg_ecx & BIT(9))  flags |= HAS_SSSE3;
+  if (reg_ecx & BIT(9)) flags |= HAS_SSSE3;
 
   if (reg_ecx & BIT(19)) flags |= HAS_SSE4_1;
 
+  if (reg_ecx & BIT(28)) flags |= HAS_AVX;
+
+  if (reg_ebx & BIT(5)) flags |= HAS_AVX2;
+
   return flags & mask;
 }