Commit 1632152b authored by Radu Velea's avatar Radu Velea Committed by Timothy B. Terriberry

Adding AVX config switches

Signed-off-by: Timothy B. Terriberry's avatarTimothy B. Terriberry <tterribe@xiph.org>
parent bb0e1e0d
......@@ -48,13 +48,14 @@
(defined(OPUS_X86_MAY_HAVE_SSE4_1) && !defined(OPUS_X86_PRESUME_SSE4_1))
#include "x86/x86cpu.h"
/* We currently support 4 x86 variants:
/* We currently support 5 x86 variants:
* arch[0] -> non-sse
* arch[1] -> sse
* arch[2] -> sse2
* arch[3] -> sse4.1
* arch[4] -> avx
*/
#define OPUS_ARCHMASK 3
#define OPUS_ARCHMASK 7
int opus_select_arch(void);
#else
......
......@@ -53,6 +53,7 @@ void (*const CELT_FIR_IMPL[OPUS_ARCHMASK + 1])(
celt_fir_c,
celt_fir_c,
MAY_HAVE_SSE4_1(celt_fir), /* sse4.1 */
MAY_HAVE_SSE4_1(celt_fir) /* avx */
};
void (*const XCORR_KERNEL_IMPL[OPUS_ARCHMASK + 1])(
......@@ -65,6 +66,7 @@ void (*const XCORR_KERNEL_IMPL[OPUS_ARCHMASK + 1])(
xcorr_kernel_c,
xcorr_kernel_c,
MAY_HAVE_SSE4_1(xcorr_kernel), /* sse4.1 */
MAY_HAVE_SSE4_1(xcorr_kernel) /* avx */
};
#endif
......@@ -81,6 +83,7 @@ opus_val32 (*const CELT_INNER_PROD_IMPL[OPUS_ARCHMASK + 1])(
celt_inner_prod_c,
MAY_HAVE_SSE2(celt_inner_prod),
MAY_HAVE_SSE4_1(celt_inner_prod), /* sse4.1 */
MAY_HAVE_SSE4_1(celt_inner_prod) /* avx */
};
#endif
......@@ -99,6 +102,7 @@ void (*const XCORR_KERNEL_IMPL[OPUS_ARCHMASK + 1])(
MAY_HAVE_SSE(xcorr_kernel),
MAY_HAVE_SSE(xcorr_kernel),
MAY_HAVE_SSE(xcorr_kernel),
MAY_HAVE_SSE(xcorr_kernel)
};
opus_val32 (*const CELT_INNER_PROD_IMPL[OPUS_ARCHMASK + 1])(
......@@ -110,6 +114,7 @@ opus_val32 (*const CELT_INNER_PROD_IMPL[OPUS_ARCHMASK + 1])(
MAY_HAVE_SSE(celt_inner_prod),
MAY_HAVE_SSE(celt_inner_prod),
MAY_HAVE_SSE(celt_inner_prod),
MAY_HAVE_SSE(celt_inner_prod)
};
void (*const DUAL_INNER_PROD_IMPL[OPUS_ARCHMASK + 1])(
......@@ -124,6 +129,7 @@ void (*const DUAL_INNER_PROD_IMPL[OPUS_ARCHMASK + 1])(
MAY_HAVE_SSE(dual_inner_prod),
MAY_HAVE_SSE(dual_inner_prod),
MAY_HAVE_SSE(dual_inner_prod),
MAY_HAVE_SSE(dual_inner_prod)
};
void (*const COMB_FILTER_CONST_IMPL[OPUS_ARCHMASK + 1])(
......@@ -139,6 +145,7 @@ void (*const COMB_FILTER_CONST_IMPL[OPUS_ARCHMASK + 1])(
MAY_HAVE_SSE(comb_filter_const),
MAY_HAVE_SSE(comb_filter_const),
MAY_HAVE_SSE(comb_filter_const),
MAY_HAVE_SSE(comb_filter_const)
};
......
......@@ -91,6 +91,8 @@ typedef struct CPU_Feature{
int HW_SSE;
int HW_SSE2;
int HW_SSE41;
/* SIMD: 256-bit */
int HW_AVX;
} CPU_Feature;
static void opus_cpu_feature_check(CPU_Feature *cpu_feature)
......@@ -106,11 +108,13 @@ static void opus_cpu_feature_check(CPU_Feature *cpu_feature)
cpu_feature->HW_SSE = (info[3] & (1 << 25)) != 0;
cpu_feature->HW_SSE2 = (info[3] & (1 << 26)) != 0;
cpu_feature->HW_SSE41 = (info[2] & (1 << 19)) != 0;
cpu_feature->HW_AVX = (info[2] & (1 << 28)) != 0;
}
else {
cpu_feature->HW_SSE = 0;
cpu_feature->HW_SSE2 = 0;
cpu_feature->HW_SSE41 = 0;
cpu_feature->HW_AVX = 0;
}
}
......@@ -140,6 +144,12 @@ int opus_select_arch(void)
}
arch++;
if (!cpu_feature.HW_AVX)
{
return arch;
}
arch++;
return arch;
}
......
......@@ -351,10 +351,12 @@ AM_CONDITIONAL([OPUS_ARM_EXTERNAL_ASM],
AM_CONDITIONAL([HAVE_SSE], [false])
AM_CONDITIONAL([HAVE_SSE2], [false])
AM_CONDITIONAL([HAVE_SSE4_1], [false])
AM_CONDITIONAL([HAVE_AVX], [false])
m4_define([DEFAULT_X86_SSE_CFLAGS], [-msse])
m4_define([DEFAULT_X86_SSE2_CFLAGS], [-msse2])
m4_define([DEFAULT_X86_SSE4_1_CFLAGS], [-msse4.1])
m4_define([DEFAULT_X86_AVX_CFLAGS], [-mavx])
m4_define([DEFAULT_ARM_NEON_INTR_CFLAGS], [-mfpu=neon])
# With GCC on ARM32 softfp architectures (e.g. Android, or older Ubuntu) you need to specify
# -mfloat-abi=softfp for -mfpu=neon to work. However, on ARM32 hardfp architectures (e.g. newer Ubuntu),
......@@ -371,11 +373,13 @@ AS_CASE([$host],
AC_ARG_VAR([X86_SSE_CFLAGS], [C compiler flags to compile SSE intrinsics @<:@default=]DEFAULT_X86_SSE_CFLAGS[@:>@])
AC_ARG_VAR([X86_SSE2_CFLAGS], [C compiler flags to compile SSE2 intrinsics @<:@default=]DEFAULT_X86_SSE2_CFLAGS[@:>@])
AC_ARG_VAR([X86_SSE4_1_CFLAGS], [C compiler flags to compile SSE4.1 intrinsics @<:@default=]DEFAULT_X86_SSE4_1_CFLAGS[@:>@])
AC_ARG_VAR([X86_AVX_CFLAGS], [C compiler flags to compile AVX intrinsics @<:@default=]DEFAULT_X86_AVX_CFLAGS[@:>@])
AC_ARG_VAR([ARM_NEON_INTR_CFLAGS], [C compiler flags to compile ARM NEON intrinsics @<:@default=]DEFAULT_ARM_NEON_INTR_CFLAGS / DEFAULT_ARM_NEON_SOFTFP_INTR_CFLAGS[@:>@])
AS_VAR_SET_IF([X86_SSE_CFLAGS], [], [AS_VAR_SET([X86_SSE_CFLAGS], "DEFAULT_X86_SSE_CFLAGS")])
AS_VAR_SET_IF([X86_SSE2_CFLAGS], [], [AS_VAR_SET([X86_SSE2_CFLAGS], "DEFAULT_X86_SSE2_CFLAGS")])
AS_VAR_SET_IF([X86_SSE4_1_CFLAGS], [], [AS_VAR_SET([X86_SSE4_1_CFLAGS], "DEFAULT_X86_SSE4_1_CFLAGS")])
AS_VAR_SET_IF([X86_AVX_CFLAGS], [], [AS_VAR_SET([X86_AVX_CFLAGS], "DEFAULT_X86_AVX_CFLAGS")])
AS_VAR_SET_IF([ARM_NEON_INTR_CFLAGS], [], [AS_VAR_SET([ARM_NEON_INTR_CFLAGS], ["$RESOLVED_DEFAULT_ARM_NEON_INTR_CFLAGS"])])
AC_DEFUN([OPUS_PATH_NE10],
......@@ -566,7 +570,24 @@ AS_IF([test x"$enable_intrinsics" = x"yes"],[
AC_SUBST([OPUS_X86_SSE4_1_CFLAGS])
]
)
OPUS_CHECK_INTRINSICS(
[AVX],
[$X86_AVX_CFLAGS],
[OPUS_X86_MAY_HAVE_AVX],
[OPUS_X86_PRESUME_AVX],
[[#include <immintrin.h>
]],
[[
static __m256 mtest;
mtest = _mm256_setzero_ps();
]]
)
AS_IF([test x"$OPUS_X86_MAY_HAVE_AVX" = x"1" && test x"$OPUS_X86_PRESUME_AVX" != x"1"],
[
OPUS_X86_AVX_CFLAGS="$X86_AVX_CFLAGS"
AC_SUBST([OPUS_X86_AVX_CFLAGS])
]
)
AS_IF([test x"$rtcd_support" = x"no"], [rtcd_support=""])
AS_IF([test x"$OPUS_X86_MAY_HAVE_SSE" = x"1"],
[
......@@ -606,6 +627,19 @@ AS_IF([test x"$enable_intrinsics" = x"yes"],[
[
AC_MSG_WARN([Compiler does not support SSE4.1 intrinsics])
])
AS_IF([test x"$OPUS_X86_MAY_HAVE_AVX" = x"1"],
[
AC_DEFINE([OPUS_X86_MAY_HAVE_AVX], 1, [Compiler supports X86 AVX Intrinsics])
intrinsics_support="$intrinsics_support AVX"
AS_IF([test x"$OPUS_X86_PRESUME_AVX" = x"1"],
[AC_DEFINE([OPUS_X86_PRESUME_AVX], 1, [Define if binary requires AVX intrinsics support])],
[rtcd_support="$rtcd_support AVX"])
],
[
AC_MSG_WARN([Compiler does not support AVX intrinsics])
])
AS_IF([test x"$intrinsics_support" = x""],
[intrinsics_support=no],
[intrinsics_support="x86$intrinsics_support"]
......@@ -672,6 +706,8 @@ AM_CONDITIONAL([HAVE_SSE2],
[test x"$OPUS_X86_MAY_HAVE_SSE2" = x"1"])
AM_CONDITIONAL([HAVE_SSE4_1],
[test x"$OPUS_X86_MAY_HAVE_SSE4_1" = x"1"])
AM_CONDITIONAL([HAVE_AVX],
[test x"$OPUS_X86_MAY_HAVE_AVX" = x"1"])
AS_IF([test x"$enable_rtcd" = x"yes"],[
AS_IF([test x"$rtcd_support" != x"no"],[
......
......@@ -50,6 +50,7 @@ opus_int64 (*const SILK_INNER_PROD16_ALIGNED_64_IMPL[ OPUS_ARCHMASK + 1 ] )(
silk_inner_prod16_aligned_64_c,
silk_inner_prod16_aligned_64_c,
MAY_HAVE_SSE4_1( silk_inner_prod16_aligned_64 ), /* sse4.1 */
MAY_HAVE_SSE4_1( silk_inner_prod16_aligned_64 ) /* avx */
};
#endif
......@@ -62,6 +63,7 @@ opus_int (*const SILK_VAD_GETSA_Q8_IMPL[ OPUS_ARCHMASK + 1 ] )(
silk_VAD_GetSA_Q8_c,
silk_VAD_GetSA_Q8_c,
MAY_HAVE_SSE4_1( silk_VAD_GetSA_Q8 ), /* sse4.1 */
MAY_HAVE_SSE4_1( silk_VAD_GetSA_Q8 ) /* avx */
};
void (*const SILK_NSQ_IMPL[ OPUS_ARCHMASK + 1 ] )(
......@@ -85,6 +87,7 @@ void (*const SILK_NSQ_IMPL[ OPUS_ARCHMASK + 1 ] )(
silk_NSQ_c,
silk_NSQ_c,
MAY_HAVE_SSE4_1( silk_NSQ ), /* sse4.1 */
MAY_HAVE_SSE4_1( silk_NSQ ) /* avx */
};
void (*const SILK_VQ_WMAT_EC_IMPL[ OPUS_ARCHMASK + 1 ] )(
......@@ -104,6 +107,7 @@ void (*const SILK_VQ_WMAT_EC_IMPL[ OPUS_ARCHMASK + 1 ] )(
silk_VQ_WMat_EC_c,
silk_VQ_WMat_EC_c,
MAY_HAVE_SSE4_1( silk_VQ_WMat_EC ), /* sse4.1 */
MAY_HAVE_SSE4_1( silk_VQ_WMat_EC ) /* avx */
};
void (*const SILK_NSQ_DEL_DEC_IMPL[ OPUS_ARCHMASK + 1 ] )(
......@@ -127,6 +131,7 @@ void (*const SILK_NSQ_DEL_DEC_IMPL[ OPUS_ARCHMASK + 1 ] )(
silk_NSQ_del_dec_c,
silk_NSQ_del_dec_c,
MAY_HAVE_SSE4_1( silk_NSQ_del_dec ), /* sse4.1 */
MAY_HAVE_SSE4_1( silk_NSQ_del_dec ) /* avx */
};
#if defined(FIXED_POINT)
......@@ -144,6 +149,7 @@ void (*const SILK_WARPED_LPC_ANALYSIS_FILTER_FIX_IMPL[ OPUS_ARCHMASK + 1 ] )(
silk_warped_LPC_analysis_filter_FIX_c,
silk_warped_LPC_analysis_filter_FIX_c,
MAY_HAVE_SSE4_1( silk_warped_LPC_analysis_filter_FIX ), /* sse4.1 */
MAY_HAVE_SSE4_1( silk_warped_LPC_analysis_filter_FIX ) /* avx */
};
void (*const SILK_BURG_MODIFIED_IMPL[ OPUS_ARCHMASK + 1 ] )(
......@@ -161,6 +167,7 @@ void (*const SILK_BURG_MODIFIED_IMPL[ OPUS_ARCHMASK + 1 ] )(
silk_burg_modified_c,
silk_burg_modified_c,
MAY_HAVE_SSE4_1( silk_burg_modified ), /* sse4.1 */
MAY_HAVE_SSE4_1( silk_burg_modified ) /* avx */
};
#endif
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment