sh4 optimized idct & bswap patch by (BERO <bero at geocities dot co dot jp>)

git-svn-id: file:///var/local/repositories/ffmpeg/trunk@1877 9553f0bf-9b14-0410-a0b8-cfaf0461ba5b

libavcodec/Makefile

@@ -124,6 +124,11 @@ OBJS += ppc/dsputil_altivec.o ppc/mpegvideo_altivec.o ppc/idct_altivec.o \
         ppc/fft_altivec.o ppc/gmc_altivec.o
 endif
 
+ifeq ($(TARGET_ARCH_SH4),yes)
+OBJS+= sh4/idct_sh4.o sh4/dsputil_sh4.o sh4/dsputil_align.o
+endif
+
+
 SRCS := $(OBJS:.o=.c) $(ASM_OBJS:.o=.S)
 OBJS := $(OBJS) $(ASM_OBJS)
 
@@ -179,6 +184,7 @@ clean:
 	   alpha/*.o alpha/*~ \
 	   ppc/*.o ppc/*~ \
 	   ps2/*.o ps2/*~ \
+	   sh4/*.o sh4/*~ \
 	   liba52/*.o liba52/*~ \
 	   apiexample $(TESTS)
 	$(MAKE) -C libpostproc clean

libavcodec/bswap.h

@@ -47,6 +47,39 @@ inline static unsigned long long int ByteSwap64(unsigned long long int x)
 }
 #define bswap_64(x) ByteSwap64(x)
 
+#elif defined(ARCH_SH4)
+
+static inline uint16_t ByteSwap16(uint16_t x) {
+	__asm__("swap.b %0,%0":"=r"(x):"0"(x));
+	return x;
+}
+
+static inline uint32_t ByteSwap32(uint32_t x) {
+	__asm__(
+	"swap.b %0,%0\n"
+	"swap.w %0,%0\n"
+	"swap.b %0,%0\n"
+	:"=r"(x):"0"(x));
+	return x;
+}
+
+#define bswap_16(x) ByteSwap16(x)
+#define bswap_32(x) ByteSwap32(x)
+
+inline static uint64_t ByteSwap64(uint64_t x)
+{
+    union { 
+        uint64_t ll;
+        struct {
+           uint32_t l,h;
+        } l;
+    } r;
+    r.l.l = bswap_32 (x);
+    r.l.h = bswap_32 (x>>32);
+    return r.ll;
+}
+#define bswap_64(x) ByteSwap64(x)
+
 #else
 
 #define bswap_16(x) (((x) & 0x00ff) << 8 | ((x) & 0xff00) >> 8)

libavcodec/dsputil.c

@@ -2662,6 +2662,9 @@ void dsputil_init(DSPContext* c, AVCodecContext *avctx)
 #ifdef HAVE_MMI
     dsputil_init_mmi(c, avctx);
 #endif
+#ifdef ARCH_SH4
+    dsputil_init_sh4(c,avctx);
+#endif
 
     switch(c->idct_permutation_type){
     case FF_NO_IDCT_PERM:

libavcodec/dsputil.h

@@ -352,6 +352,12 @@ void dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx);
 
 void dsputil_init_mmi(DSPContext* c, AVCodecContext *avctx);
 
+#elif defined(ARCH_SH4)
+
+#define __align8 __attribute__ ((aligned (8)))
+
+void dsputil_init_sh4(DSPContext* c, AVCodecContext *avctx);
+
 #else
 
 #define __align8

libavcodec/sh4/dsputil_sh4.c

+/*
+ *  sh4 dsputil
+ *
+ * Copyright (c) 2003 BERO <bero@geocities.co.jp>
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser 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
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser 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
+ */
+
+#include "../avcodec.h"
+#include "../dsputil.h"
+
+static void memzero_align8(void *dst,size_t size)
+{
+#if defined(__SH4__) || defined(__SH4_SINGLE__) || defined(__SH4_SINGLE_ONLY__)
+	(char*)dst+=size;
+	size/=8*4;
+	asm(
+#if defined(__SH4__)
+	" fschg\n"  //single float mode
+#endif
+	" fldi0 fr0\n"
+	" fldi0 fr1\n"
+	" fschg\n"  // double
+	"1: \n" \
+	" dt %1\n"
+	" fmov	dr0,@-%0\n"
+	" fmov	dr0,@-%0\n"
+	" fmov	dr0,@-%0\n"
+	" bf.s 1b\n"
+	" fmov	dr0,@-%0\n"
+#if defined(__SH4_SINGLE__) || defined(__SH4_SINGLE_ONLY__)
+	" fschg" //back to single
+#endif
+	: : "r"(dst),"r"(size): "memory" );
+#else
+	double *d = dst;
+	size/=8*4;
+	do {
+		d[0] = 0.0;
+		d[1] = 0.0;
+		d[2] = 0.0;
+		d[3] = 0.0;
+		d+=4;
+	} while(--size);
+#endif
+}
+
+static void clear_blocks_sh4(DCTELEM *blocks)
+{
+//	if (((int)blocks&7)==0) 
+	memzero_align8(blocks,sizeof(DCTELEM)*6*64);
+}
+
+extern void idct_sh4(DCTELEM *block);
+static void idct_put(uint8_t *dest, int line_size, DCTELEM *block)
+{
+	idct_sh4(block);
+	int i;
+	uint8_t *cm = cropTbl + MAX_NEG_CROP;
+	for(i=0;i<8;i++) {
+		dest[0] = cm[block[0]];
+		dest[1] = cm[block[1]];
+		dest[2] = cm[block[2]];
+		dest[3] = cm[block[3]];
+		dest[4] = cm[block[4]];
+		dest[5] = cm[block[5]];
+		dest[6] = cm[block[6]];
+		dest[7] = cm[block[7]];
+		dest+=line_size;
+		block+=8;
+	}
+}
+static void idct_add(uint8_t *dest, int line_size, DCTELEM *block)
+{
+	idct_sh4(block);
+	int i;
+	uint8_t *cm = cropTbl + MAX_NEG_CROP;
+	for(i=0;i<8;i++) {
+		dest[0] = cm[dest[0]+block[0]];
+		dest[1] = cm[dest[1]+block[1]];
+		dest[2] = cm[dest[2]+block[2]];
+		dest[3] = cm[dest[3]+block[3]];
+		dest[4] = cm[dest[4]+block[4]];
+		dest[5] = cm[dest[5]+block[5]];
+		dest[6] = cm[dest[6]+block[6]];
+		dest[7] = cm[dest[7]+block[7]];
+		dest+=line_size;
+		block+=8;
+	}
+}
+
+extern void dsputil_init_align(DSPContext* c, AVCodecContext *avctx);
+
+void dsputil_init_sh4(DSPContext* c, AVCodecContext *avctx)
+{
+	dsputil_init_align(c,avctx);
+
+	c->clear_blocks = clear_blocks_sh4;
+	c->idct_put = idct_put;
+	c->idct_add = idct_add;
+	c->idct_permutation_type= FF_NO_IDCT_PERM; //FF_SIMPLE_IDCT_PERM; //FF_LIBMPEG2_IDCT_PERM;
+}

libavcodec/sh4/idct_sh4.c

+/*
+ * idct for sh4
+ *
+ * Copyright (c) 2001-2003 BERO <bero@geocities.co.jp>
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser 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
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser 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
+ */
+
+#include "../dsputil.h"
+#define c1	1.38703984532214752434	/* sqrt(2)*cos(1*pi/16) */
+#define c2	1.30656296487637657577	/* sqrt(2)*cos(2*pi/16) */
+#define c3	1.17587560241935884520	/* sqrt(2)*cos(3*pi/16) */
+#define c4	1.00000000000000000000	/* sqrt(2)*cos(4*pi/16) */
+#define c5	0.78569495838710234903	/* sqrt(2)*cos(5*pi/16) */
+#define c6	0.54119610014619712324	/* sqrt(2)*cos(6*pi/16) */
+#define c7	0.27589937928294311353	/* sqrt(2)*cos(7*pi/16) */
+
+const	static float even_table[] __attribute__ ((aligned(8))) = {
+	 c4, c4, c4, c4,
+	 c2, c6,-c6,-c2,
+	 c4,-c4,-c4, c4,
+	 c6,-c2, c2,-c6
+};
+
+const	static float odd_table[] __attribute__ ((aligned(8))) = {
+	 c1, c3, c5, c7,
+	 c3,-c7,-c1,-c5,
+	 c5,-c1, c7, c3,
+	 c7,-c5, c3,-c1
+};
+
+#undef	c1
+#undef	c2
+#undef	c3
+#undef	c4
+#undef	c5
+#undef	c6
+#undef	c7
+
+#if defined(__SH4_SINGLE__) || defined(__SH4_SINGLE_ONLY__)
+
+#define	load_matrix(table) \
+	__asm__ volatile( \
+	"	fschg\n" \
+	"	fmov	@%0+,xd0\n" \
+	"	fmov	@%0+,xd2\n" \
+	"	fmov	@%0+,xd4\n" \
+	"	fmov	@%0+,xd6\n" \
+	"	fmov	@%0+,xd8\n" \
+	"	fmov	@%0+,xd10\n" \
+	"	fmov	@%0+,xd12\n" \
+	"	fmov	@%0+,xd14\n" \
+	"	fschg\n" \
+	:\
+	: "r"(table)\
+	: "0" \
+	)
+
+#define	ftrv() \
+		__asm__ volatile("ftrv xmtrx,fv0" \
+		: "=f"(fr0),"=f"(fr1),"=f"(fr2),"=f"(fr3) \
+		:  "0"(fr0), "1"(fr1), "2"(fr2), "3"(fr3) );
+
+#define	DEFREG	\
+	register float fr0 __asm__("fr0"); \
+	register float fr1 __asm__("fr1"); \
+	register float fr2 __asm__("fr2"); \
+	register float fr3 __asm__("fr3")
+
+#else
+
+/* generic C code for check */
+
+static void ftrv_(const float xf[],float fv[])
+{
+	float f0,f1,f2,f3;
+	f0 = fv[0];
+	f1 = fv[1];
+	f2 = fv[2];
+	f3 = fv[3];
+	fv[0] = xf[0]*f0 + xf[4]*f1 + xf[ 8]*f2 + xf[12]*f3;
+	fv[1] = xf[1]*f0 + xf[5]*f1 + xf[ 9]*f2 + xf[13]*f3;
+	fv[2] = xf[2]*f0 + xf[6]*f1 + xf[10]*f2 + xf[14]*f3;
+	fv[3] = xf[3]*f0 + xf[7]*f1 + xf[11]*f2 + xf[15]*f3;
+}
+
+static void load_matrix_(float xf[],const float table[])
+{
+	int i;
+	for(i=0;i<16;i++) xf[i]=table[i];
+}
+
+#define	ftrv()	ftrv_(xf,fv)
+#define	load_matrix(table)	load_matrix_(xf,table)
+
+#define	DEFREG \
+	float fv[4],xf[16]
+
+#define	fr0	fv[0]
+#define	fr1	fv[1]
+#define	fr2	fv[2]
+#define	fr3	fv[3]
+
+#endif
+
+#if 1
+#define	DESCALE(x,n)	(x)*(1.0f/(1<<(n)))
+#else
+#define	DESCALE(x,n)	(((int)(x)+(1<<(n-1)))>>(n))
+#endif
+
+/* this code work worse on gcc cvs. 3.2.3 work fine */
+
+
+#if 1
+//optimized 
+
+void idct_sh4(DCTELEM *block)
+{
+	DEFREG;
+
+	int i;
+	float	tblock[8*8],*fblock;
+	int ofs1,ofs2,ofs3;
+
+#if defined(__SH4__)
+#error	"FIXME!! change to single float"
+#endif
+
+	/* row */
+
+	/* even part */
+	load_matrix(even_table);
+
+	fblock = tblock+4;
+	i = 8;
+	do {
+		fr0 = block[0];
+		fr1 = block[2];
+		fr2 = block[4];
+		fr3 = block[6];
+		block+=8;
+		ftrv();
+		*--fblock = fr3;
+		*--fblock = fr2;
+		*--fblock = fr1;
+		*--fblock = fr0;
+		fblock+=8+4;
+	} while(--i);
+	block-=8*8;
+	fblock-=8*8+4;
+
+	load_matrix(odd_table);
+
+	i = 8;
+
+//	ofs1 = sizeof(float)*1;
+//	ofs2 = sizeof(float)*2;
+//	ofs3 = sizeof(float)*3;
+
+	do {
+		float t0,t1,t2,t3;
+		fr0 = block[1];
+		fr1 = block[3];
+		fr2 = block[5];
+		fr3 = block[7];
+		block+=8;
+		ftrv();
+		t0 = *fblock++;
+		t1 = *fblock++;
+		t2 = *fblock++;
+		t3 = *fblock++;
+		fblock+=4;
+		*--fblock = t0 - fr0;
+		*--fblock = t1 - fr1;
+		*--fblock = t2 - fr2;
+		*--fblock = t3 - fr3;
+		*--fblock = t3 + fr3;
+		*--fblock = t2 + fr2;
+		*--fblock = t1 + fr1;
+		*--fblock = t0 + fr0;
+		fblock+=8;
+	} while(--i);
+	block-=8*8;
+	fblock-=8*8;
+
+	/* col */
+
+	/* even part */
+	load_matrix(even_table);
+
+	ofs1 = sizeof(float)*2*8;
+	ofs2 = sizeof(float)*4*8;
+	ofs3 = sizeof(float)*6*8;
+
+	i = 8;
+
+#define	OA(fblock,ofs)	*(float*)((char*)fblock + ofs)
+
+	do {
+		fr0 = OA(fblock,   0);
+		fr1 = OA(fblock,ofs1);
+		fr2 = OA(fblock,ofs2);
+		fr3 = OA(fblock,ofs3);
+		ftrv();
+		OA(fblock,0   ) = fr0;
+		OA(fblock,ofs1) = fr1;
+		OA(fblock,ofs2) = fr2;
+		OA(fblock,ofs3) = fr3;
+		fblock++;
+	} while(--i);
+	fblock-=8;
+
+	load_matrix(odd_table);
+
+	i=8;
+	do {
+		float t0,t1,t2,t3;
+		t0 = OA(fblock,   0); /* [8*0] */
+		t1 = OA(fblock,ofs1); /* [8*2] */
+		t2 = OA(fblock,ofs2); /* [8*4] */
+		t3 = OA(fblock,ofs3); /* [8*6] */
+		fblock+=8;
+		fr0 = OA(fblock,   0); /* [8*1] */
+		fr1 = OA(fblock,ofs1); /* [8*3] */
+		fr2 = OA(fblock,ofs2); /* [8*5] */
+		fr3 = OA(fblock,ofs3); /* [8*7] */
+		fblock+=-8+1;
+		ftrv();
+		block[8*0] = DESCALE(t0 + fr0,3);
+		block[8*7] = DESCALE(t0 - fr0,3);
+		block[8*1] = DESCALE(t1 + fr1,3);
+		block[8*6] = DESCALE(t1 - fr1,3);
+		block[8*2] = DESCALE(t2 + fr2,3);
+		block[8*5] = DESCALE(t2 - fr2,3);
+		block[8*3] = DESCALE(t3 + fr3,3);
+		block[8*4] = DESCALE(t3 - fr3,3);
+		block++;
+	} while(--i);
+
+#if defined(__SH4__)
+#error	"FIXME!! change to double"
+#endif
+}
+#else
+void idct_sh4(DCTELEM *block)
+{
+	DEFREG;
+
+	int i;
+	float	tblock[8*8],*fblock;
+
+	/* row */
+
+	/* even part */
+	load_matrix(even_table);
+
+	fblock = tblock;
+	i = 8;
+	do {
+		fr0 = block[0];
+		fr1 = block[2];
+		fr2 = block[4];
+		fr3 = block[6];
+		block+=8;
+		ftrv();
+		fblock[0] = fr0;
+		fblock[2] = fr1;
+		fblock[4] = fr2;
+		fblock[6] = fr3;
+		fblock+=8;
+	} while(--i);
+	block-=8*8;
+	fblock-=8*8;
+
+	load_matrix(odd_table);
+
+	i = 8;
+
+	do {
+		float t0,t1,t2,t3;
+		fr0 = block[1];
+		fr1 = block[3];
+		fr2 = block[5];
+		fr3 = block[7];
+		block+=8;
+		ftrv();
+		t0 = fblock[0];
+		t1 = fblock[2];
+		t2 = fblock[4];
+		t3 = fblock[6];
+		fblock[0] = t0 + fr0;
+		fblock[7] = t0 - fr0;
+		fblock[1] = t1 + fr1;
+		fblock[6] = t1 - fr1;
+		fblock[2] = t2 + fr2;
+		fblock[5] = t2 - fr2;
+		fblock[3] = t3 + fr3;
+		fblock[4] = t3 - fr3;
+		fblock+=8;
+	} while(--i);
+	block-=8*8;
+	fblock-=8*8;
+
+	/* col */
+
+	/* even part */
+	load_matrix(even_table);
+
+	i = 8;
+
+	do {
+		fr0 = fblock[8*0];
+		fr1 = fblock[8*2];
+		fr2 = fblock[8*4];
+		fr3 = fblock[8*6];
+		ftrv();
+		fblock[8*0] = fr0;
+		fblock[8*2] = fr1;
+		fblock[8*4] = fr2;
+		fblock[8*6] = fr3;
+		fblock++;
+	} while(--i);
+	fblock-=8;
+
+	load_matrix(odd_table);
+
+	i=8;
+	do {
+		float t0,t1,t2,t3;
+		fr0 = fblock[8*1];
+		fr1 = fblock[8*3];
+		fr2 = fblock[8*5];
+		fr3 = fblock[8*7];
+		ftrv();
+		t0 = fblock[8*0];
+		t1 = fblock[8*2];
+		t2 = fblock[8*4];
+		t3 = fblock[8*6];
+		fblock++;
+		block[8*0] = DESCALE(t0 + fr0,3);
+		block[8*7] = DESCALE(t0 - fr0,3);
+		block[8*1] = DESCALE(t1 + fr1,3);
+		block[8*6] = DESCALE(t1 - fr1,3);
+		block[8*2] = DESCALE(t2 + fr2,3);
+		block[8*5] = DESCALE(t2 - fr2,3);
+		block[8*3] = DESCALE(t3 + fr3,3);
+		block[8*4] = DESCALE(t3 - fr3,3);
+		block++;
+	} while(--i);
+}
+#endif