From 4cbe470278ddcfb4f0d0267e7a1295d7b72d661a Mon Sep 17 00:00:00 2001
From: Sam Lantinga <slouken@libsdl.org>
Date: Sat, 9 Nov 2002 06:13:28 +0000
Subject: [PATCH] Added MMX audio mixing code for gcc (thanks Stephane!)

---
 docs.html                 |   1 +
 src/audio/Makefile.am     |   4 +-
 src/audio/SDL_mixer.c     |  36 ++++++++
 src/audio/SDL_mixer_MMX.c | 185 ++++++++++++++++++++++++++++++++++++++
 src/audio/SDL_mixer_MMX.h |  13 +++
 5 files changed, 238 insertions(+), 1 deletion(-)
 create mode 100644 src/audio/SDL_mixer_MMX.c
 create mode 100644 src/audio/SDL_mixer_MMX.h

diff --git a/docs.html b/docs.html
index cdced29e8..2f330d578 100644
--- a/docs.html
+++ b/docs.html
@@ -20,6 +20,7 @@ <H2>
 Major changes since SDL 1.0.0:
 </H2>
 <UL>
+	<LI> 1.2.6: Added MMX audio mixing code for gcc (thanks Stephane!)
 	<LI> 1.2.6: Fixed potential dropped events under DirectInput
 	<LI> 1.2.6: Added Visual C++ 7 (.NET) projects (thanks James!)
 	<LI> 1.2.5: Added an environment variable SDL_HAS3BUTTONMOUSE for Quartz
diff --git a/src/audio/Makefile.am b/src/audio/Makefile.am
index 9b3a2afa0..91fe93be4 100644
--- a/src/audio/Makefile.am
+++ b/src/audio/Makefile.am
@@ -22,7 +22,9 @@ COMMON_SRCS =			\
 	SDL_mixer.c		\
 	SDL_sysaudio.h		\
 	SDL_wave.c		\
-	SDL_wave.h
+	SDL_wave.h              \
+	SDL_mixer_MMX.c         \
+	SDL_mixer_MMX.h
 
 libaudio_la_SOURCES = $(COMMON_SRCS)
 libaudio_la_LIBADD = $(DRIVERS)
diff --git a/src/audio/SDL_mixer.c b/src/audio/SDL_mixer.c
index 50839338f..c77391a04 100644
--- a/src/audio/SDL_mixer.c
+++ b/src/audio/SDL_mixer.c
@@ -35,7 +35,24 @@ static char rcsid =
 #include "SDL_mutex.h"
 #include "SDL_timer.h"
 #include "SDL_sysaudio.h"
+#include "SDL_mixer_MMX.h"
+
+/* Function to check the CPU flags */
+#define MMX_CPU		0x800000
+#ifdef USE_ASMBLIT
+#define CPU_Flags()	Hermes_X86_CPU()
+#else
+#define CPU_Flags()	0L
+#endif
+
+#ifdef USE_ASMBLIT
+#define X86_ASSEMBLER
+#define HermesConverterInterface	void
+#define HermesClearInterface		void
+#define STACKCALL
 
+#include "HeadX86.h"
+#endif
 
 /* This table is used to add two sound values together and pin
  * the value to avoid overflow.  (used with permission from ARDI)
@@ -130,6 +147,15 @@ void SDL_MixAudio (Uint8 *dst, const Uint8 *src, Uint32 len, int volume)
 		break;
 
 		case AUDIO_S8: {
+
+#if defined(i386) && defined(__GNUC__) && defined(USE_ASMBLIT)
+			if (CPU_Flags() & MMX_CPU)
+			{
+				SDL_MixAudio_MMX_S8((char*)dst,(char*)src,(unsigned int)len,(int)volume);
+			}
+			else
+#endif
+			{
 			Sint8 *dst8, *src8;
 			Sint8 src_sample;
 			int dst_sample;
@@ -153,10 +179,19 @@ void SDL_MixAudio (Uint8 *dst, const Uint8 *src, Uint32 len, int volume)
 				++dst8;
 				++src8;
 			}
+			}
 		}
 		break;
 
 		case AUDIO_S16LSB: {
+#if defined(i386) && defined(__GNUC__) && defined(USE_ASMBLIT)
+			if (CPU_Flags() & MMX_CPU)
+			{
+				SDL_MixAudio_MMX_S16((char*)dst,(char*)src,(unsigned int)len,(int)volume);
+			}
+			else
+#endif
+			{
 			Sint16 src1, src2;
 			int dst_sample;
 			const int max_audioval = ((1<<(16-1))-1);
@@ -180,6 +215,7 @@ void SDL_MixAudio (Uint8 *dst, const Uint8 *src, Uint32 len, int volume)
 				dst[1] = dst_sample&0xFF;
 				dst += 2;
 			}
+			}
 		}
 		break;
 
diff --git a/src/audio/SDL_mixer_MMX.c b/src/audio/SDL_mixer_MMX.c
new file mode 100644
index 000000000..179aa92be
--- /dev/null
+++ b/src/audio/SDL_mixer_MMX.c
@@ -0,0 +1,185 @@
+// MMX assembler version of SDL_MixAudio for signed little endian 16 bit samples and signed 8 bit samples
+// Copyright 2002 Stephane Marchesin (stephane.marchesin@wanadoo.fr)
+// This code is licensed under the LGPL (see COPYING for details)
+// 
+// Assumes buffer size in bytes is a multiple of 16
+// Assumes SDL_MIX_MAXVOLUME = 128
+
+
+////////////////////////////////////////////////
+// Mixing for 16 bit signed buffers
+////////////////////////////////////////////////
+
+#if defined(i386) && defined(__GNUC__) && defined(USE_ASMBLIT)
+void SDL_MixAudio_MMX_S16(char* dst,char* src,unsigned int size,int volume)
+{
+    __asm__ __volatile__ (
+
+"	movl %0,%%edi\n"	// edi = dst
+"	movl %1,%%esi\n"	// esi = src
+"	movl %3,%%eax\n"	// eax = volume
+
+"	movl %2,%%ebx\n"	// ebx = size
+
+"	shrl $4,%%ebx\n"	// process 16 bytes per iteration = 8 samples
+
+"	jz .endS16\n"
+
+"	pxor %%mm0,%%mm0\n"
+
+"	movd %%eax,%%mm0\n"
+"	movq %%mm0,%%mm1\n"
+"	psllq $16,%%mm0\n"
+"	por %%mm1,%%mm0\n"
+"	psllq $16,%%mm0\n"
+"	por %%mm1,%%mm0\n"
+"	psllq $16,%%mm0\n"
+"	por %%mm1,%%mm0\n"		// mm0 = vol|vol|vol|vol
+
+".align 16\n"
+"	.mixloopS16:\n"
+
+"	movq (%%esi),%%mm1\n" // mm1 = a|b|c|d
+
+"	movq %%mm1,%%mm2\n" // mm2 = a|b|c|d
+
+"	movq 8(%%esi),%%mm4\n" // mm4 = e|f|g|h
+
+	// pré charger le buffer dst dans mm7
+"	movq (%%edi),%%mm7\n" // mm7 = dst[0]"
+
+	// multiplier par le volume
+"	pmullw %%mm0,%%mm1\n" // mm1 = l(a*v)|l(b*v)|l(c*v)|l(d*v)
+
+"	pmulhw %%mm0,%%mm2\n" // mm2 = h(a*v)|h(b*v)|h(c*v)|h(d*v)
+"	movq %%mm4,%%mm5\n" // mm5 = e|f|g|h
+
+"	pmullw %%mm0,%%mm4\n" // mm4 = l(e*v)|l(f*v)|l(g*v)|l(h*v)
+
+"	pmulhw %%mm0,%%mm5\n" // mm5 = h(e*v)|h(f*v)|h(g*v)|h(h*v)
+"	movq %%mm1,%%mm3\n" // mm3 = l(a*v)|l(b*v)|l(c*v)|l(d*v)
+
+"	punpckhwd %%mm2,%%mm1\n" // mm1 = a*v|b*v
+
+"	movq %%mm4,%%mm6\n" // mm6 = l(e*v)|l(f*v)|l(g*v)|l(h*v)
+"	punpcklwd %%mm2,%%mm3\n" // mm3 = c*v|d*v
+
+"	punpckhwd %%mm5,%%mm4\n" // mm4 = e*f|f*v
+
+"	punpcklwd %%mm5,%%mm6\n" // mm6 = g*v|h*v
+
+	// pré charger le buffer dst dans mm5
+"	movq 8(%%edi),%%mm5\n" // mm5 = dst[1]
+
+	// diviser par 128
+"	psrad $7,%%mm1\n" // mm1 = a*v/128|b*v/128 , 128 = SDL_MIX_MAXVOLUME
+"	addl $16,%%esi\n"
+
+"	psrad $7,%%mm3\n" // mm3 = c*v/128|d*v/128
+
+"	psrad $7,%%mm4\n" // mm4 = e*v/128|f*v/128
+
+	// mm1 = le sample avec le volume modifié
+"	packssdw %%mm1,%%mm3\n" // mm3 = s(a*v|b*v|c*v|d*v)
+
+"	psrad $7,%%mm6\n" // mm6= g*v/128|h*v/128
+"	paddsw %%mm7,%%mm3\n" // mm3 = adjust_volume(src)+dst
+
+	// mm4 = le sample avec le volume modifié
+"	packssdw %%mm4,%%mm6\n" // mm6 = s(e*v|f*v|g*v|h*v)
+"	movq %%mm3,(%%edi)\n"
+
+"	paddsw %%mm5,%%mm6\n" // mm6 = adjust_volume(src)+dst
+
+"	movq %%mm6,8(%%edi)\n"
+
+"	addl $16,%%edi\n"
+
+"	dec %%ebx\n"
+
+"	jnz .mixloopS16\n"
+
+"	emms\n"
+
+".endS16:\n"
+	 :
+	 : "m" (dst), "m"(src),"m"(size),
+	 "m"(volume)
+	 : "eax","ebx", "esi", "edi","memory"
+	 );
+}
+
+
+
+////////////////////////////////////////////////
+// Mixing for 8 bit signed buffers
+////////////////////////////////////////////////
+
+void SDL_MixAudio_MMX_S8(char* dst,char* src,unsigned int size,int volume)
+{
+    __asm__ __volatile__ (
+
+"	movl %0,%%edi\n"	// edi = dst
+"	movl %1,%%esi\n"	// esi = src
+"	movl %3,%%eax\n"	// eax = volume
+
+"	movd %%ebx,%%mm0\n"
+"	movq %%mm0,%%mm1\n"
+"	psllq $16,%%mm0\n"
+"	por %%mm1,%%mm0\n"
+"	psllq $16,%%mm0\n"
+"	por %%mm1,%%mm0\n"
+"	psllq $16,%%mm0\n"
+"	por %%mm1,%%mm0\n"
+
+"	movl %2,%%ebx\n"	// ebx = size
+"	shr $3,%%ebx\n"	// process 8 bytes per iteration = 8 samples
+
+"	cmp $0,%%ebx\n"
+"	je .endS8\n"
+
+".align 16\n"
+"	.mixloopS8:\n"
+
+"	pxor %%mm2,%%mm2\n"		// mm2 = 0
+"	movq (%%esi),%%mm1\n"	// mm1 = a|b|c|d|e|f|g|h
+
+"	movq %%mm1,%%mm3\n" 	// mm3 = a|b|c|d|e|f|g|h
+
+	// on va faire le "sign extension" en faisant un cmp avec 0 qui retourne 1 si <0, 0 si >0
+"	pcmpgtb %%mm1,%%mm2\n"	// mm2 = 11111111|00000000|00000000....
+
+"	punpckhbw %%mm2,%%mm1\n"	// mm1 = 0|a|0|b|0|c|0|d
+
+"	punpcklbw %%mm2,%%mm3\n"	// mm3 = 0|e|0|f|0|g|0|h
+"	movq (%%edi),%%mm2\n"	// mm2 = destination
+
+"	pmullw %%mm0,%%mm1\n"	// mm1 = v*a|v*b|v*c|v*d
+"	addl $8,%%esi\n"
+
+"	pmullw %%mm0,%%mm3\n"	// mm3 = v*e|v*f|v*g|v*h
+"	psraw $7,%%mm1\n"		// mm1 = v*a/128|v*b/128|v*c/128|v*d/128 
+
+"	psraw $7,%%mm3\n"		// mm3 = v*e/128|v*f/128|v*g/128|v*h/128
+
+"	packsswb %%mm1,%%mm3\n"	// mm1 = v*a/128|v*b/128|v*c/128|v*d/128|v*e/128|v*f/128|v*g/128|v*h/128
+
+"	paddsb %%mm2,%%mm3\n"	// add to destination buffer
+
+"	movq %%mm3,(%%edi)\n"	// store back to ram
+"	addl $8,%%edi\n"
+
+"	dec %%ebx\n"
+
+"	jnz .mixloopS8\n"
+
+".endS8:\n"
+"	emms\n"
+	 :
+	 : "m" (dst), "m"(src),"m"(size),
+	 "m"(volume)
+	 : "eax","ebx", "esi", "edi","memory"
+	 );
+}
+#endif
+
diff --git a/src/audio/SDL_mixer_MMX.h b/src/audio/SDL_mixer_MMX.h
new file mode 100644
index 000000000..8cd29235b
--- /dev/null
+++ b/src/audio/SDL_mixer_MMX.h
@@ -0,0 +1,13 @@
+// headers for MMX assembler version of SDL_MixAudio
+// Copyright 2002 Stephane Marchesin (stephane.marchesin@wanadoo.fr)
+// This code is licensed under the LGPL (see COPYING for details)
+// 
+// Assumes buffer size in bytes is a multiple of 16
+// Assumes SDL_MIX_MAXVOLUME = 128
+
+
+#if defined(i386) && defined(__GNUC__) && defined(USE_ASMBLIT)
+void SDL_MixAudio_MMX_S16(char* ,char* ,unsigned int ,int );
+void SDL_MixAudio_MMX_S8(char* ,char* ,unsigned int ,int );
+#endif
+