Move the FIR4 from SSE2 to SSE3

SSE3 can avoid the slow _MM_TRANSPOSE_PS4 call thanks to the inclusion of
horizontal adds.

6 files changed, 120 insertions, 73 deletions

diff --git a/Alc/ALc.c b/Alc/ALc.c
index 2610bf45..6a69fdb6 100644
--- a/Alc/ALc.c
+++ b/Alc/ALc.c
@@ -955,7 +955,9 @@ static void alc_initconfig(void)
 
     capfilter = 0;
 #if defined(HAVE_SSE4_1)
-    capfilter |= CPU_CAP_SSE | CPU_CAP_SSE2 | CPU_CAP_SSE4_1;
+    capfilter |= CPU_CAP_SSE | CPU_CAP_SSE2 | CPU_CAP_SSE3 | CPU_CAP_SSE4_1;
+#elif defined(HAVE_SSE3)
+    capfilter |= CPU_CAP_SSE | CPU_CAP_SSE2 | CPU_CAP_SSE3;
 #elif defined(HAVE_SSE2)
     capfilter |= CPU_CAP_SSE | CPU_CAP_SSE2;
 #elif defined(HAVE_SSE)
@@ -989,6 +991,8 @@ static void alc_initconfig(void)
                     capfilter &= ~CPU_CAP_SSE;
                 else if(len == 4 && strncasecmp(str, "sse2", len) == 0)
                     capfilter &= ~CPU_CAP_SSE2;
+                else if(len == 4 && strncasecmp(str, "sse3", len) == 0)
+                    capfilter &= ~CPU_CAP_SSE3;
                 else if(len == 6 && strncasecmp(str, "sse4.1", len) == 0)
                     capfilter &= ~CPU_CAP_SSE4_1;
                 else if(len == 4 && strncasecmp(str, "neon", len) == 0)
diff --git a/Alc/helpers.c b/Alc/helpers.c
index dac29296..c789a426 100644
--- a/Alc/helpers.c
+++ b/Alc/helpers.c
@@ -152,8 +152,12 @@ void FillCPUCaps(ALuint capfilter)
                 if((cpuinf[0].regs[3]&(1<<26)))
                 {
                     caps |= CPU_CAP_SSE2;
-                    if((cpuinf[0].regs[2]&(1<<19)))
-                        caps |= CPU_CAP_SSE4_1;
+                    if((cpuinf[0].regs[2]&(1<<0)))
+                    {
+                        caps |= CPU_CAP_SSE3;
+                        if((cpuinf[0].regs[2]&(1<<19)))
+                            caps |= CPU_CAP_SSE4_1;
+                    }
                 }
             }
         }
@@ -196,8 +200,12 @@ void FillCPUCaps(ALuint capfilter)
                 if((cpuinf[0].regs[3]&(1<<26)))
                 {
                     caps |= CPU_CAP_SSE2;
-                    if((cpuinf[0].regs[2]&(1<<19)))
-                        caps |= CPU_CAP_SSE4_1;
+                    if((cpuinf[0].regs[2]&(1<<0)))
+                    {
+                        caps |= CPU_CAP_SSE3;
+                        if((cpuinf[0].regs[2]&(1<<19)))
+                            caps |= CPU_CAP_SSE4_1;
+                    }
                 }
             }
         }
@@ -206,7 +214,10 @@ void FillCPUCaps(ALuint capfilter)
     /* Assume support for whatever's supported if we can't check for it */
 #if defined(HAVE_SSE4_1)
 #warning "Assuming SSE 4.1 run-time support!"
-    caps |= CPU_CAP_SSE | CPU_CAP_SSE2 | CPU_CAP_SSE4_1;
+    caps |= CPU_CAP_SSE | CPU_CAP_SSE2 | CPU_CAP_SSE3 | CPU_CAP_SSE4_1;
+#elif defined(HAVE_SSE3)
+#warning "Assuming SSE 3 run-time support!"
+    caps |= CPU_CAP_SSE | CPU_CAP_SSE2 | CPU_CAP_SSE3;
 #elif defined(HAVE_SSE2)
 #warning "Assuming SSE 2 run-time support!"
     caps |= CPU_CAP_SSE | CPU_CAP_SSE2;
@@ -220,9 +231,10 @@ void FillCPUCaps(ALuint capfilter)
     caps |= CPU_CAP_NEON;
 #endif
 
-    TRACE("Extensions:%s%s%s%s%s\n",
+    TRACE("Extensions:%s%s%s%s%s%s\n",
         ((capfilter&CPU_CAP_SSE)    ? ((caps&CPU_CAP_SSE)    ? " +SSE"    : " -SSE")    : ""),
         ((capfilter&CPU_CAP_SSE2)   ? ((caps&CPU_CAP_SSE2)   ? " +SSE2"   : " -SSE2")   : ""),
+        ((capfilter&CPU_CAP_SSE3)   ? ((caps&CPU_CAP_SSE3)   ? " +SSE3"   : " -SSE3")   : ""),
         ((capfilter&CPU_CAP_SSE4_1) ? ((caps&CPU_CAP_SSE4_1) ? " +SSE4.1" : " -SSE4.1") : ""),
         ((capfilter&CPU_CAP_NEON)   ? ((caps&CPU_CAP_NEON)   ? " +Neon"   : " -Neon")   : ""),
         ((!capfilter) ? " -none-" : "")
diff --git a/Alc/mixer.c b/Alc/mixer.c
index fa060958..8061ab73 100644
--- a/Alc/mixer.c
+++ b/Alc/mixer.c
@@ -122,9 +122,9 @@ static inline ResamplerFunc SelectResampler(enum Resampler resampler)
             if((CPUCapFlags&CPU_CAP_SSE4_1))
                 return Resample_fir4_32_SSE41;
 #endif
-#ifdef HAVE_SSE2
-            if((CPUCapFlags&CPU_CAP_SSE2))
-                return Resample_fir4_32_SSE2;
+#ifdef HAVE_SSE3
+            if((CPUCapFlags&CPU_CAP_SSE3))
+                return Resample_fir4_32_SSE3;
 #endif
             return Resample_fir4_32_C;
         case FIR6Resampler:
diff --git a/Alc/mixer_defs.h b/Alc/mixer_defs.h
index 4c94954f..bba8e9a1 100644
--- a/Alc/mixer_defs.h
+++ b/Alc/mixer_defs.h
@@ -55,7 +55,7 @@ const ALfloat *Resample_lerp32_SSE2(const ALfloat *src, ALuint frac, ALuint incr
 const ALfloat *Resample_lerp32_SSE41(const ALfloat *src, ALuint frac, ALuint increment,
                                      ALfloat *restrict dst, ALuint numsamples);
 
-const ALfloat *Resample_fir4_32_SSE2(const ALfloat *src, ALuint frac, ALuint increment,
+const ALfloat *Resample_fir4_32_SSE3(const ALfloat *src, ALuint frac, ALuint increment,
                                      ALfloat *restrict dst, ALuint numsamples);
 const ALfloat *Resample_fir4_32_SSE41(const ALfloat *src, ALuint frac, ALuint increment,
                                       ALfloat *restrict dst, ALuint numsamples);
diff --git a/Alc/mixer_sse2.c b/Alc/mixer_sse2.c
index 1321f5f2..6759f795 100644
--- a/Alc/mixer_sse2.c
+++ b/Alc/mixer_sse2.c
@@ -76,65 +76,3 @@ const ALfloat *Resample_lerp32_SSE2(const ALfloat *src, ALuint frac, ALuint incr
     }
     return dst;
 }
-
-const ALfloat *Resample_fir4_32_SSE2(const ALfloat *src, ALuint frac, ALuint increment,
-                                     ALfloat *restrict dst, ALuint numsamples)
-{
-    const __m128i increment4 = _mm_set1_epi32(increment*4);
-    const __m128i fracMask4 = _mm_set1_epi32(FRACTIONMASK);
-    alignas(16) union { ALuint i[4]; float f[4]; } pos_;
-    alignas(16) union { ALuint i[4]; float f[4]; } frac_;
-    __m128i frac4, pos4;
-    ALuint pos;
-    ALuint i;
-
-    InitiatePositionArrays(frac, increment, frac_.i, pos_.i, 4);
-
-    frac4 = _mm_castps_si128(_mm_load_ps(frac_.f));
-    pos4 = _mm_castps_si128(_mm_load_ps(pos_.f));
-
-    --src;
-    for(i = 0;numsamples-i > 3;i += 4)
-    {
-        const __m128 val0 = _mm_loadu_ps(&src[pos_.i[0]]);
-        const __m128 val1 = _mm_loadu_ps(&src[pos_.i[1]]);
-        const __m128 val2 = _mm_loadu_ps(&src[pos_.i[2]]);
-        const __m128 val3 = _mm_loadu_ps(&src[pos_.i[3]]);
-        __m128 k0 = _mm_load_ps(ResampleCoeffs.FIR4[frac_.i[0]]);
-        __m128 k1 = _mm_load_ps(ResampleCoeffs.FIR4[frac_.i[1]]);
-        __m128 k2 = _mm_load_ps(ResampleCoeffs.FIR4[frac_.i[2]]);
-        __m128 k3 = _mm_load_ps(ResampleCoeffs.FIR4[frac_.i[3]]);
-        __m128 out;
-
-        k0 = _mm_mul_ps(k0, val0);
-        k1 = _mm_mul_ps(k1, val1);
-        k2 = _mm_mul_ps(k2, val2);
-        k3 = _mm_mul_ps(k3, val3);
-        _MM_TRANSPOSE4_PS(k0, k1, k2, k3);
-        k0 = _mm_add_ps(k0, k1);
-        k2 = _mm_add_ps(k2, k3);
-        out = _mm_add_ps(k0, k2);
-
-        _mm_store_ps(&dst[i], out);
-
-        frac4 = _mm_add_epi32(frac4, increment4);
-        pos4 = _mm_add_epi32(pos4, _mm_srli_epi32(frac4, FRACTIONBITS));
-        frac4 = _mm_and_si128(frac4, fracMask4);
-
-        _mm_store_ps(pos_.f, _mm_castsi128_ps(pos4));
-        _mm_store_ps(frac_.f, _mm_castsi128_ps(frac4));
-    }
-
-    pos = pos_.i[0];
-    frac = frac_.i[0];
-
-    for(;i < numsamples;i++)
-    {
-        dst[i] = resample_fir4(src[pos], src[pos+1], src[pos+2], src[pos+3], frac);
-
-        frac += increment;
-        pos  += frac>>FRACTIONBITS;
-        frac &= FRACTIONMASK;
-    }
-    return dst;
-}
diff --git a/Alc/mixer_sse3.c b/Alc/mixer_sse3.c
new file mode 100644
index 00000000..ced90593
--- /dev/null
+++ b/Alc/mixer_sse3.c
@@ -0,0 +1,93 @@
+/**
+ * OpenAL cross platform audio library, SSE3 mixer functions
+ *
+ * Copyright (C) 2014 by Timothy Arceri <[email protected]>.
+ * Copyright (C) 2015 by Chris Robinson <[email protected]>.
+ *
+ * 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.,
+ *  51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+ * Or go to http://www.gnu.org/copyleft/lgpl.html
+ */
+
+#include "config.h"
+
+#include <xmmintrin.h>
+#include <emmintrin.h>
+#include <pmmintrin.h>
+
+#include "alu.h"
+#include "mixer_defs.h"
+
+
+const ALfloat *Resample_fir4_32_SSE3(const ALfloat *src, ALuint frac, ALuint increment,
+                                     ALfloat *restrict dst, ALuint numsamples)
+{
+    const __m128i increment4 = _mm_set1_epi32(increment*4);
+    const __m128i fracMask4 = _mm_set1_epi32(FRACTIONMASK);
+    alignas(16) union { ALuint i[4]; float f[4]; } pos_;
+    alignas(16) union { ALuint i[4]; float f[4]; } frac_;
+    __m128i frac4, pos4;
+    ALuint pos;
+    ALuint i;
+
+    InitiatePositionArrays(frac, increment, frac_.i, pos_.i, 4);
+
+    frac4 = _mm_castps_si128(_mm_load_ps(frac_.f));
+    pos4 = _mm_castps_si128(_mm_load_ps(pos_.f));
+
+    --src;
+    for(i = 0;numsamples-i > 3;i += 4)
+    {
+        const __m128 val0 = _mm_loadu_ps(&src[pos_.i[0]]);
+        const __m128 val1 = _mm_loadu_ps(&src[pos_.i[1]]);
+        const __m128 val2 = _mm_loadu_ps(&src[pos_.i[2]]);
+        const __m128 val3 = _mm_loadu_ps(&src[pos_.i[3]]);
+        __m128 k0 = _mm_load_ps(ResampleCoeffs.FIR4[frac_.i[0]]);
+        __m128 k1 = _mm_load_ps(ResampleCoeffs.FIR4[frac_.i[1]]);
+        __m128 k2 = _mm_load_ps(ResampleCoeffs.FIR4[frac_.i[2]]);
+        __m128 k3 = _mm_load_ps(ResampleCoeffs.FIR4[frac_.i[3]]);
+        __m128 out;
+
+        k0 = _mm_mul_ps(k0, val0);
+        k1 = _mm_mul_ps(k1, val1);
+        k2 = _mm_mul_ps(k2, val2);
+        k3 = _mm_mul_ps(k3, val3);
+        k0 = _mm_hadd_ps(k0, k1);
+        k2 = _mm_hadd_ps(k2, k3);
+        out = _mm_hadd_ps(k0, k2);
+
+        _mm_store_ps(&dst[i], out);
+
+        frac4 = _mm_add_epi32(frac4, increment4);
+        pos4 = _mm_add_epi32(pos4, _mm_srli_epi32(frac4, FRACTIONBITS));
+        frac4 = _mm_and_si128(frac4, fracMask4);
+
+        _mm_store_ps(pos_.f, _mm_castsi128_ps(pos4));
+        _mm_store_ps(frac_.f, _mm_castsi128_ps(frac4));
+    }
+
+    pos = pos_.i[0];
+    frac = frac_.i[0];
+
+    for(;i < numsamples;i++)
+    {
+        dst[i] = resample_fir4(src[pos], src[pos+1], src[pos+2], src[pos+3], frac);
+
+        frac += increment;
+        pos  += frac>>FRACTIONBITS;
+        frac &= FRACTIONMASK;
+    }
+    return dst;
+}