x86: add detection for FMA3 instruction set

Based on x264 code

Signed-off-by: James Almer <jamrial@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>

configure

@@ -323,6 +323,7 @@ Optimization options (experts only):
   --disable-sse42          disable SSE4.2 optimizations
   --disable-avx            disable AVX optimizations
   --disable-xop            disable XOP optimizations
+  --disable-fma3           disable FMA3 optimizations
   --disable-fma4           disable FMA4 optimizations
   --disable-avx2           disable AVX2 optimizations
   --disable-armv5te        disable armv5te optimizations
@@ -1455,6 +1456,7 @@ ARCH_EXT_LIST_X86='
     avx
     avx2
     cpunop
+    fma3
     fma4
     i686
     mmx
@@ -1828,6 +1830,7 @@ sse4_deps="ssse3"
 sse42_deps="sse4"
 avx_deps="sse42"
 xop_deps="avx"
+fma3_deps="avx"
 fma4_deps="avx"
 avx2_deps="avx"
 
@@ -4252,6 +4255,7 @@ EOF
             die "yasm/nasm not found or too old. Use --disable-yasm for a crippled build."
         check_yasm "vextractf128 xmm0, ymm0, 0"      || disable avx_external avresample
         check_yasm "vpmacsdd xmm0, xmm1, xmm2, xmm3" || disable xop_external
+        check_yasm "vfmadd132ps ymm0, ymm1, ymm2"    || disable fma3_external
         check_yasm "vfmaddps ymm0, ymm1, ymm2, ymm3" || disable fma4_external
         check_yasm "CPU amdnop" && enable cpunop
     fi
@@ -4937,6 +4941,7 @@ if enabled x86; then
     echo "SSSE3 enabled             ${ssse3-no}"
     echo "AVX enabled               ${avx-no}"
     echo "XOP enabled               ${xop-no}"
+    echo "FMA3 enabled              ${fma3-no}"
     echo "FMA4 enabled              ${fma4-no}"
     echo "i686 features enabled     ${i686-no}"
     echo "CMOV is fast              ${fast_cmov-no}"

libavutil/cpu.c

@@ -91,6 +91,7 @@ int av_parse_cpu_flags(const char *s)
 #define CPUFLAG_SSE42    (AV_CPU_FLAG_SSE42    | CPUFLAG_SSE4)
 #define CPUFLAG_AVX      (AV_CPU_FLAG_AVX      | CPUFLAG_SSE42)
 #define CPUFLAG_XOP      (AV_CPU_FLAG_XOP      | CPUFLAG_AVX)
+#define CPUFLAG_FMA3     (AV_CPU_FLAG_FMA3     | CPUFLAG_AVX)
 #define CPUFLAG_FMA4     (AV_CPU_FLAG_FMA4     | CPUFLAG_AVX)
 #define CPUFLAG_AVX2     (AV_CPU_FLAG_AVX2     | CPUFLAG_AVX)
     static const AVOption cpuflags_opts[] = {
@@ -111,6 +112,7 @@ int av_parse_cpu_flags(const char *s)
         { "sse4.2"  , NULL, 0, AV_OPT_TYPE_CONST, { .i64 = CPUFLAG_SSE42        },    .unit = "flags" },
         { "avx"     , NULL, 0, AV_OPT_TYPE_CONST, { .i64 = CPUFLAG_AVX          },    .unit = "flags" },
         { "xop"     , NULL, 0, AV_OPT_TYPE_CONST, { .i64 = CPUFLAG_XOP          },    .unit = "flags" },
+        { "fma3"    , NULL, 0, AV_OPT_TYPE_CONST, { .i64 = CPUFLAG_FMA3         },    .unit = "flags" },
         { "fma4"    , NULL, 0, AV_OPT_TYPE_CONST, { .i64 = CPUFLAG_FMA4         },    .unit = "flags" },
         { "avx2"    , NULL, 0, AV_OPT_TYPE_CONST, { .i64 = CPUFLAG_AVX2         },    .unit = "flags" },
         { "3dnow"   , NULL, 0, AV_OPT_TYPE_CONST, { .i64 = CPUFLAG_3DNOW        },    .unit = "flags" },
@@ -166,6 +168,7 @@ int av_parse_cpu_caps(unsigned *flags, const char *s)
         { "sse4.2"  , NULL, 0, AV_OPT_TYPE_CONST, { .i64 = AV_CPU_FLAG_SSE42    },    .unit = "flags" },
         { "avx"     , NULL, 0, AV_OPT_TYPE_CONST, { .i64 = AV_CPU_FLAG_AVX      },    .unit = "flags" },
         { "xop"     , NULL, 0, AV_OPT_TYPE_CONST, { .i64 = AV_CPU_FLAG_XOP      },    .unit = "flags" },
+        { "fma3"    , NULL, 0, AV_OPT_TYPE_CONST, { .i64 = AV_CPU_FLAG_FMA3     },    .unit = "flags" },
         { "fma4"    , NULL, 0, AV_OPT_TYPE_CONST, { .i64 = AV_CPU_FLAG_FMA4     },    .unit = "flags" },
         { "avx2"    , NULL, 0, AV_OPT_TYPE_CONST, { .i64 = AV_CPU_FLAG_AVX2     },    .unit = "flags" },
         { "3dnow"   , NULL, 0, AV_OPT_TYPE_CONST, { .i64 = AV_CPU_FLAG_3DNOW    },    .unit = "flags" },
@@ -279,6 +282,7 @@ static const struct {
     { AV_CPU_FLAG_SSE42,     "sse4.2"     },
     { AV_CPU_FLAG_AVX,       "avx"        },
     { AV_CPU_FLAG_XOP,       "xop"        },
+    { AV_CPU_FLAG_FMA3,      "fma3"       },
     { AV_CPU_FLAG_FMA4,      "fma4"       },
     { AV_CPU_FLAG_3DNOW,     "3dnow"      },
     { AV_CPU_FLAG_3DNOWEXT,  "3dnowext"   },

libavutil/cpu.h

@@ -51,6 +51,7 @@
 // #define AV_CPU_FLAG_CMOV         0x1000 ///< supports cmov instruction
 // #endif
 #define AV_CPU_FLAG_AVX2         0x8000 ///< AVX2 functions: requires OS support even if YMM registers aren't used
+#define AV_CPU_FLAG_FMA3        0x10000 ///< Haswell FMA3 functions
 
 #define AV_CPU_FLAG_ALTIVEC      0x0001 ///< standard
 

libavutil/x86/cpu.c

@@ -131,8 +131,11 @@ int ff_get_cpu_flags_x86(void)
         if ((ecx & 0x18000000) == 0x18000000) {
             /* Check for OS support */
             xgetbv(0, eax, edx);
-            if ((eax & 0x6) == 0x6)
+            if ((eax & 0x6) == 0x6) {
                 rval |= AV_CPU_FLAG_AVX;
+                if (ecx&0x00001000)
+                    rval |= AV_CPU_FLAG_FMA3;
+            }
         }
 #if HAVE_AVX2
     if (max_std_level >= 7) {

libavutil/x86/cpu.h

@@ -38,6 +38,7 @@
 #define X86_SSE42(flags)            CPUEXT(flags, SSE42)
 #define X86_AVX(flags)              CPUEXT(flags, AVX)
 #define X86_XOP(flags)              CPUEXT(flags, XOP)
+#define X86_FMA3(flags)             CPUEXT(flags, FMA3)
 #define X86_FMA4(flags)             CPUEXT(flags, FMA4)
 #define X86_AVX2(flags)             CPUEXT(flags, AVX2)
 
@@ -53,6 +54,7 @@
 #define EXTERNAL_SSE42(flags)       CPUEXT_SUFFIX(flags, _EXTERNAL, SSE42)
 #define EXTERNAL_AVX(flags)         CPUEXT_SUFFIX(flags, _EXTERNAL, AVX)
 #define EXTERNAL_XOP(flags)         CPUEXT_SUFFIX(flags, _EXTERNAL, XOP)
+#define EXTERNAL_FMA3(flags)        CPUEXT_SUFFIX(flags, _EXTERNAL, FMA3)
 #define EXTERNAL_FMA4(flags)        CPUEXT_SUFFIX(flags, _EXTERNAL, FMA4)
 #define EXTERNAL_AVX2(flags)        CPUEXT_SUFFIX(flags, _EXTERNAL, AVX2)
 
@@ -68,6 +70,7 @@
 #define INLINE_SSE42(flags)         CPUEXT_SUFFIX(flags, _INLINE, SSE42)
 #define INLINE_AVX(flags)           CPUEXT_SUFFIX(flags, _INLINE, AVX)
 #define INLINE_XOP(flags)           CPUEXT_SUFFIX(flags, _INLINE, XOP)
+#define INLINE_FMA3(flags)          CPUEXT_SUFFIX(flags, _INLINE, FMA3)
 #define INLINE_FMA4(flags)          CPUEXT_SUFFIX(flags, _INLINE, FMA4)
 #define INLINE_AVX2(flags)          CPUEXT_SUFFIX(flags, _INLINE, AVX2)