From e9a20cb985c5686fd2777540dbbd2a13c9525ee0 Mon Sep 17 00:00:00 2001
From: Chris Robinson <chris.kcat@gmail.com>
Date: Tue, 14 Aug 2012 03:53:07 -0700
Subject: [PATCH] Add an SSE-enhanced HRTF mixer using intrinsics

Not as good as it could be, but it's something
---
 Alc/ALc.c       |  4 +++-
 Alc/mixer.c     | 48 ++++++++++++++++++++++++++++++++++++++++++++++++
 CMakeLists.txt  |  1 +
 alsoftrc.sample |  4 ++--
 config.h.in     |  3 +++
 5 files changed, 57 insertions(+), 3 deletions(-)

diff --git a/Alc/ALc.c b/Alc/ALc.c
index e664d25a..14a74a92 100644
--- a/Alc/ALc.c
+++ b/Alc/ALc.c
@@ -812,7 +812,9 @@ static void alc_initconfig(void)
                     continue;
 
                 len = (next ? ((size_t)(next-str)) : strlen(str));
-                if(strncasecmp(str, "neon", len) == 0)
+                if(strncasecmp(str, "sse", len) == 0)
+                    capfilter &= ~CPU_CAP_SSE;
+                else if(strncasecmp(str, "neon", len) == 0)
                     capfilter &= ~CPU_CAP_NEON;
                 else
                     WARN("Invalid CPU extension \"%s\"\n", str);
diff --git a/Alc/mixer.c b/Alc/mixer.c
index e7fada8f..f5d9bb09 100644
--- a/Alc/mixer.c
+++ b/Alc/mixer.c
@@ -25,6 +25,9 @@
 #include <string.h>
 #include <ctype.h>
 #include <assert.h>
+#ifdef HAVE_XMMINTRIN_H
+#include <xmmintrin.h>
+#endif
 #ifdef HAVE_ARM_NEON_H
 #include <arm_neon.h>
 #endif
@@ -221,6 +224,39 @@ DECL_TEMPLATE(point32, C)
 DECL_TEMPLATE(lerp32, C)
 DECL_TEMPLATE(cubic32, C)
 
+#ifdef HAVE_XMMINTRIN_H
+
+static __inline void ApplyCoeffsSSE(ALuint Offset, ALfloat (*RESTRICT Values)[2],
+                                    ALfloat (*RESTRICT Coeffs)[2],
+                                    ALfloat left, ALfloat right)
+{
+    const __m128 lrlr = { left, right, left, right };
+    ALuint c;
+    for(c = 0;c < HRIR_LENGTH;c += 2)
+    {
+        const ALuint o0 = (Offset++)&HRIR_MASK;
+        const ALuint o1 = (Offset++)&HRIR_MASK;
+        __m128 vals = { 0.0f, 0.0f, 0.0f, 0.0f };
+        __m128 coeffs = { 0.0f, 0.0f, 0.0f, 0.0f };
+
+        vals = _mm_loadl_pi(vals, (__m64*)&Values[o0][0]);
+        vals = _mm_loadh_pi(vals, (__m64*)&Values[o1][0]);
+        coeffs = _mm_loadl_pi(coeffs, (__m64*)&Coeffs[c  ][0]);
+        coeffs = _mm_loadh_pi(coeffs, (__m64*)&Coeffs[c+1][0]);
+
+        vals = _mm_add_ps(vals, _mm_mul_ps(coeffs, lrlr));
+
+        _mm_storel_pi((__m64*)&Values[o0][0], vals);
+        _mm_storeh_pi((__m64*)&Values[o1][0], vals);
+    }
+}
+
+DECL_TEMPLATE(point32, SSE)
+DECL_TEMPLATE(lerp32, SSE)
+DECL_TEMPLATE(cubic32, SSE)
+
+#endif
+
 #ifdef HAVE_ARM_NEON_H
 
 static __inline void ApplyCoeffsNeon(ALuint Offset, ALfloat (*RESTRICT Values)[2],
@@ -419,18 +455,30 @@ DryMixerFunc SelectHrtfMixer(enum Resampler Resampler)
     switch(Resampler)
     {
         case PointResampler:
+#ifdef HAVE_XMMINTRIN_H
+            if((CPUCapFlags&CPU_CAP_SSE))
+                return MixDirect_Hrtf_point32_SSE;
+#endif
 #ifdef HAVE_ARM_NEON_H
             if((CPUCapFlags&CPU_CAP_NEON))
                 return MixDirect_Hrtf_point32_Neon;
 #endif
             return MixDirect_Hrtf_point32_C;
         case LinearResampler:
+#ifdef HAVE_XMMINTRIN_H
+            if((CPUCapFlags&CPU_CAP_SSE))
+                return MixDirect_Hrtf_lerp32_SSE;
+#endif
 #ifdef HAVE_ARM_NEON_H
             if((CPUCapFlags&CPU_CAP_NEON))
                 return MixDirect_Hrtf_lerp32_Neon;
 #endif
             return MixDirect_Hrtf_lerp32_C;
         case CubicResampler:
+#ifdef HAVE_XMMINTRIN_H
+            if((CPUCapFlags&CPU_CAP_SSE))
+                return MixDirect_Hrtf_cubic32_SSE;
+#endif
 #ifdef HAVE_ARM_NEON_H
             if((CPUCapFlags&CPU_CAP_NEON))
                 return MixDirect_Hrtf_cubic32_Neon;
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 96dd4347..f4a0d1f7 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -225,6 +225,7 @@ CHECK_INCLUDE_FILE(guiddef.h HAVE_GUIDDEF_H)
 IF(NOT HAVE_GUIDDEF_H)
     CHECK_INCLUDE_FILE(initguid.h HAVE_INITGUID_H)
 ENDIF()
+CHECK_INCLUDE_FILE(xmmintrin.h HAVE_XMMINTRIN_H)
 CHECK_INCLUDE_FILE(arm_neon.h HAVE_ARM_NEON_H)
 CHECK_INCLUDE_FILE(cpuid.h HAVE_CPUID_H)
 
diff --git a/alsoftrc.sample b/alsoftrc.sample
index bcfbe5e2..866154ef 100644
--- a/alsoftrc.sample
+++ b/alsoftrc.sample
@@ -14,8 +14,8 @@
 ## disable-cpu-exts:
 #  Disables use of the listed CPU extensions. Certain methods may utilize CPU
 #  extensions when detected, and this option is useful for preventing those
-#  extensions from being used. The available extensions are: neon. Specifying
-#  'all' disables use of all extensions.
+#  extensions from being used. The available extensions are: sse, neon.
+#  Specifying 'all' disables use of all extensions.
 #disable-cpu-exts =
 
 ## channels:
diff --git a/config.h.in b/config.h.in
index 82b6c334..af7420b9 100644
--- a/config.h.in
+++ b/config.h.in
@@ -104,6 +104,9 @@
 /* Define if we have pthread_np.h */
 #cmakedefine HAVE_PTHREAD_NP_H
 
+/* Define if we have xmmintrin.h */
+#cmakedefine HAVE_XMMINTRIN_H
+
 /* Define if we have arm_neon.h */
 #cmakedefine HAVE_ARM_NEON_H
 
-- 
2.11.4.GIT