[openal] [PATCH] Add AVX linear resampler
Timothy Arceri
t_arceri at yahoo.com.au
Sat Jun 7 05:59:45 EDT 2014
---
Alc/ALc.c | 6 ++-
Alc/ALu.c | 4 ++
Alc/helpers.c | 8 +++-
Alc/mixer_avx.c | 115 ++++++++++++++++++++++++++++++++++++++++++++++
Alc/mixer_defs.h | 2 +
CMakeLists.txt | 26 +++++++++++
OpenAL32/Include/alMain.h | 3 +-
config.h.in | 1 +
8 files changed, 161 insertions(+), 4 deletions(-)
create mode 100644 Alc/mixer_avx.c
diff --git a/Alc/ALc.c b/Alc/ALc.c
index 7c8c232..7b60df9 100644
--- a/Alc/ALc.c
+++ b/Alc/ALc.c
@@ -909,7 +909,9 @@ static void alc_initconfig(void)
ReadALConfig();
capfilter = 0;
-#if defined(HAVE_SSE4_1)
+#if defined(HAVE_AVX)
+ capfilter |= CPU_CAP_SSE | CPU_CAP_SSE2 | CPU_CAP_SSE4_1 | CPU_CAP_AVX;
+#elif defined(HAVE_SSE4_1)
capfilter |= CPU_CAP_SSE | CPU_CAP_SSE2 | CPU_CAP_SSE4_1;
#elif defined(HAVE_SSE2)
capfilter |= CPU_CAP_SSE | CPU_CAP_SSE2;
@@ -946,6 +948,8 @@ static void alc_initconfig(void)
capfilter &= ~CPU_CAP_SSE2;
else if(len == 6 && strncasecmp(str, "sse4.1", len) == 0)
capfilter &= ~CPU_CAP_SSE4_1;
+ else if(len == 3 && strncasecmp(str, "avx", len) == 0)
+ capfilter &= ~CPU_CAP_AVX;
else if(len == 4 && strncasecmp(str, "neon", len) == 0)
capfilter &= ~CPU_CAP_NEON;
else
diff --git a/Alc/ALu.c b/Alc/ALu.c
index 488a727..0212474 100644
--- a/Alc/ALu.c
+++ b/Alc/ALu.c
@@ -91,6 +91,10 @@ static ResamplerFunc SelectResampler(enum Resampler Resampler, ALuint increment)
case PointResampler:
return Resample_point32_C;
case LinearResampler:
+#ifdef HAVE_AVX
+ if((CPUCapFlags&CPU_CAP_AVX))
+ return Resample_lerp32_AVX;
+#endif
#ifdef HAVE_SSE4_1
if((CPUCapFlags&CPU_CAP_SSE4_1))
return Resample_lerp32_SSE41;
diff --git a/Alc/helpers.c b/Alc/helpers.c
index ff28bf0..f3a741c 100644
--- a/Alc/helpers.c
+++ b/Alc/helpers.c
@@ -141,8 +141,11 @@ void FillCPUCaps(ALuint capfilter)
if((cpuinf[0].regs[3]&(1<<26)))
{
caps |= CPU_CAP_SSE2;
- if((cpuinf[0].regs[2]&(1<<19)))
+ if((cpuinf[0].regs[2]&(1<<19))) {
caps |= CPU_CAP_SSE4_1;
+ if((cpuinf[0].regs[2]&(1<<28)))
+ caps |= CPU_CAP_AVX;
+ }
}
}
}
@@ -168,10 +171,11 @@ void FillCPUCaps(ALuint capfilter)
caps |= CPU_CAP_NEON;
#endif
- TRACE("Got caps:%s%s%s%s%s\n",
+ TRACE("Got caps:%s%s%s%s%s%s\n",
((caps&CPU_CAP_SSE) ? ((capfilter&CPU_CAP_SSE) ? " SSE" : " (SSE)") : ""),
((caps&CPU_CAP_SSE2) ? ((capfilter&CPU_CAP_SSE2) ? " SSE2" : " (SSE2)") : ""),
((caps&CPU_CAP_SSE4_1) ? ((capfilter&CPU_CAP_SSE4_1) ? " SSE4.1" : " (SSE4.1)") : ""),
+ ((caps&CPU_CAP_AVX) ? ((capfilter&CPU_CAP_AVX) ? " AVX" : " (AVX)") : ""),
((caps&CPU_CAP_NEON) ? ((capfilter&CPU_CAP_NEON) ? " Neon" : " (Neon)") : ""),
((!caps) ? " -none-" : "")
);
diff --git a/Alc/mixer_avx.c b/Alc/mixer_avx.c
new file mode 100644
index 0000000..c224931
--- /dev/null
+++ b/Alc/mixer_avx.c
@@ -0,0 +1,115 @@
+/**
+ * OpenAL cross platform audio library
+ * Copyright (C) 1999-2014 by authors.
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Library General Public
+ * License as published by the Free Software Foundation; either
+ * version 2 of the License, or (at your option) any later version.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * Library General Public License for more details.
+ *
+ * You should have received a copy of the GNU Library General Public
+ * License along with this library; if not, write to the
+ * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+ * Boston, MA 02111-1307, USA.
+ * Or go to http://www.gnu.org/copyleft/lgpl.html
+ */
+
+#include "config.h"
+
+#include <xmmintrin.h>
+#include <emmintrin.h>
+#include <smmintrin.h>
+#include <immintrin.h>
+
+#include "alu.h"
+#include "mixer_defs.h"
+
+/*****************************************************************************
+*
+* Join two 128-bit vectors
+*
+*****************************************************************************/
+#define _mm256_set_m128i(hi,lo) _mm256_insertf128_si256(_mm256_castsi128_si256(lo),(hi),1)
+ // _mm256_set_m128i(hi,lo); // not defined in all versions of immintrin.h
+
+
+const ALfloat *Resample_lerp32_AVX(const ALfloat *src, ALuint frac, ALuint increment,
+ ALfloat *restrict dst, ALuint numsamples)
+{
+ float const fracOne = 1.0f/FRACTIONONE;
+ const __m128i increment4_8 = _mm_set1_epi32(increment*8);
+ const __m256 fracOne8 = _mm256_broadcast_ss(&fracOne);
+ const __m128i fracMask4 = _mm_set1_epi32(FRACTIONMASK);
+ alignas(16) union { ALuint i[8]; float f[8]; } pos_;
+ alignas(16) union { ALuint i[8]; float f[8]; } frac_;
+ __m128i frac4_2, pos4_2, frac4_1, pos4_1;
+ __m256i frac8;
+ ALuint pos;
+ ALuint i;
+
+ InitiatePositionArrays(frac, increment, frac_.i, pos_.i, 8);
+
+ frac4_1 = _mm_castps_si128(_mm_load_ps(frac_.f));
+ frac4_2 = _mm_castps_si128(_mm_load_ps(&frac_.f[4]));
+ pos4_1 = _mm_castps_si128(_mm_load_ps(pos_.f));
+ pos4_2 = _mm_castps_si128(_mm_load_ps(&pos_.f[4]));
+
+ frac8 = _mm256_set_m128i(frac4_2, frac4_1);
+
+ /*frac4_1 = _mm_set_epi32(frac_arr[3], frac_arr[2], frac_arr[1], frac_arr[0]);
+ frac4_2 = _mm_set_epi32(frac_arr[7], frac_arr[6], frac_arr[5], frac_arr[4]);
+
+ pos4_1 = _mm_set_epi32(pos_arr[3], pos_arr[2], pos_arr[1], pos_arr[0]);
+ pos4_2 = _mm_set_epi32(pos_arr[7], pos_arr[6], pos_arr[5], pos_arr[4]);*/
+
+ for(i = 0;i < numsamples-7;i += 8)
+ {
+ __m256 val1 = _mm256_set_ps(src[pos_.i[7]], src[pos_.i[6]], src[pos_.i[5]], src[pos_.i[4]],
+ src[pos_.i[3]], src[pos_.i[2]], src[pos_.i[1]], src[pos_.i[0]]);
+ __m256 val2 = _mm256_set_ps(src[pos_.i[7]+1], src[pos_.i[6]+1], src[pos_.i[5]+1], src[pos_.i[4]+1],
+ src[pos_.i[3]+1], src[pos_.i[2]+1], src[pos_.i[1]+1], src[pos_.i[0]+1]);
+
+ /* val1 + (val2-val1)*mu */
+ const __m256 r0 = _mm256_sub_ps(val2, val1);
+ const __m256 mu = _mm256_mul_ps(_mm256_cvtepi32_ps(frac8), fracOne8);
+ const __m256 out = _mm256_add_ps(val1, _mm256_mul_ps(mu, r0));
+
+ _mm256_store_ps(&dst[i], out);
+
+ frac4_1 = _mm_add_epi32(frac4_1, increment4_8);
+ pos4_1 = _mm_add_epi32(pos4_1, _mm_srli_epi32(frac4_1, FRACTIONBITS));
+ frac4_1 = _mm_and_si128(frac4_1, fracMask4);
+
+ frac4_2 = _mm_add_epi32(frac4_2, increment4_8);
+ pos4_2 = _mm_add_epi32(pos4_2, _mm_srli_epi32(frac4_2, FRACTIONBITS));
+ frac4_2 = _mm_and_si128(frac4_2, fracMask4);
+
+ frac8 = _mm256_set_m128i(frac4_2, frac4_1);
+
+ pos_.i[0] = _mm_extract_epi32(pos4_1, 0);
+ pos_.i[1] = _mm_extract_epi32(pos4_1, 1);
+ pos_.i[2] = _mm_extract_epi32(pos4_1, 2);
+ pos_.i[3] = _mm_extract_epi32(pos4_1, 3);
+ pos_.i[4] = _mm_extract_epi32(pos4_2, 0);
+ pos_.i[5] = _mm_extract_epi32(pos4_2, 1);
+ pos_.i[6] = _mm_extract_epi32(pos4_2, 2);
+ pos_.i[7] = _mm_extract_epi32(pos4_2, 3);
+ }
+
+ pos = pos_.i[0];
+ frac = _mm_cvtsi128_si32(_mm256_extractf128_si256(frac8, 0));
+
+ for(;i < numsamples;i++)
+ {
+ dst[i] = lerp(src[pos], src[pos+1], frac * (1.0f/FRACTIONONE));
+
+ frac += increment;
+ pos += frac>>FRACTIONBITS;
+ frac &= FRACTIONMASK;
+ }
+ return dst;
+}
diff --git a/Alc/mixer_defs.h b/Alc/mixer_defs.h
index caa06c2..4f2565b 100644
--- a/Alc/mixer_defs.h
+++ b/Alc/mixer_defs.h
@@ -62,6 +62,8 @@ const ALfloat *Resample_lerp32_SSE2(const ALfloat *src, ALuint frac, ALuint incr
ALfloat *restrict dst, ALuint numsamples);
const ALfloat *Resample_lerp32_SSE41(const ALfloat *src, ALuint frac, ALuint increment,
ALfloat *restrict dst, ALuint numsamples);
+const ALfloat *Resample_lerp32_AVX(const ALfloat *src, ALuint frac, ALuint increment,
+ ALfloat *restrict dst, ALuint numsamples);
/* Neon mixers */
void MixDirect_Hrtf_Neon(ALfloat (*restrict OutBuffer)[BUFFERSIZE], const ALfloat *data,
diff --git a/CMakeLists.txt b/CMakeLists.txt
index af41c00..6fd182a 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -302,6 +302,7 @@ ENDIF()
SET(SSE_SWITCH "")
SET(SSE2_SWITCH "")
SET(SSE4_1_SWITCH "")
+SET(AVX_SWITCH "")
IF(NOT MSVC)
CHECK_C_COMPILER_FLAG(-msse HAVE_MSSE_SWITCH)
IF(HAVE_MSSE_SWITCH)
@@ -315,6 +316,10 @@ IF(NOT MSVC)
IF(HAVE_MSSE4_1_SWITCH)
SET(SSE4_1_SWITCH "-msse4.1")
ENDIF()
+ CHECK_C_COMPILER_FLAG(-mavx HAVE_MAVX_SWITCH)
+ IF(HAVE_MAVX_SWITCH)
+ SET(AVX_SWITCH "-mavx")
+ ENDIF()
ENDIF()
CHECK_C_SOURCE_COMPILES("int foo(const char *str, ...) __attribute__((format(printf, 1, 2)));
@@ -553,6 +558,7 @@ SET(CPU_EXTS "Default")
SET(HAVE_SSE 0)
SET(HAVE_SSE2 0)
SET(HAVE_SSE4_1 0)
+SET(HAVE_AVX 0)
SET(HAVE_NEON 0)
# Check for SSE support
@@ -616,6 +622,26 @@ IF(ALSOFT_REQUIRE_SSE4_1 AND NOT HAVE_SSE4_1)
MESSAGE(FATAL_ERROR "Failed to enable required SSE4.1 CPU extensions")
ENDIF()
+OPTION(ALSOFT_REQUIRE_AVX "Require AVX support" OFF)
+CHECK_INCLUDE_FILE(immintrin.h HAVE_IMMINTRIN_H "${AVX_SWITCH}")
+IF(HAVE_IMMINTRIN_H)
+ OPTION(ALSOFT_CPUEXT_AVX "Enable AVX support" ON)
+ IF(ALSOFT_CPUEXT_AVX)
+ IF(ALIGN_DECL OR HAVE_C11_ALIGNAS)
+ SET(HAVE_AVX 1)
+ SET(ALC_OBJS ${ALC_OBJS} Alc/mixer_avx.c)
+ IF(AVX_SWITCH)
+ SET_SOURCE_FILES_PROPERTIES(Alc/mixer_avx.c PROPERTIES
+ COMPILE_FLAGS "${AVX_SWITCH}")
+ ENDIF()
+ SET(CPU_EXTS "${CPU_EXTS}, AVX")
+ ENDIF()
+ ENDIF()
+ENDIF()
+IF(ALSOFT_REQUIRE_AVX AND NOT HAVE_AVX)
+ MESSAGE(FATAL_ERROR "Failed to enabled required AVX CPU extensions")
+ENDIF()
+
# Check for ARM Neon support
OPTION(ALSOFT_REQUIRE_NEON "Require ARM Neon support" OFF)
CHECK_INCLUDE_FILE(arm_neon.h HAVE_ARM_NEON_H)
diff --git a/OpenAL32/Include/alMain.h b/OpenAL32/Include/alMain.h
index 232c438..b4e9fe6 100644
--- a/OpenAL32/Include/alMain.h
+++ b/OpenAL32/Include/alMain.h
@@ -888,7 +888,8 @@ enum {
CPU_CAP_SSE = 1<<0,
CPU_CAP_SSE2 = 1<<1,
CPU_CAP_SSE4_1 = 1<<2,
- CPU_CAP_NEON = 1<<3,
+ CPU_CAP_AVX = 1<<3,
+ CPU_CAP_NEON = 1<<4,
};
void FillCPUCaps(ALuint capfilter);
diff --git a/config.h.in b/config.h.in
index 3fdc0c7..2eb1c0a 100644
--- a/config.h.in
+++ b/config.h.in
@@ -27,6 +27,7 @@
#cmakedefine HAVE_SSE
#cmakedefine HAVE_SSE2
#cmakedefine HAVE_SSE4_1
+#cmakedefine HAVE_AVX
/* Define if we have ARM Neon CPU extensions */
#cmakedefine HAVE_NEON
--
1.9.0
More information about the openal
mailing list