Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>

Originally committed as revision 274 to svn://svn.ffmpeg.org/ffmpeg/trunk

libavcodec/Makefile

@@ -37,6 +37,12 @@ OBJS += mlib/dsputil_mlib.o
 CFLAGS += $(MLIB_INC)
 endif
 
+# alpha specific stuff
+ifeq ($(TARGET_ARCH_ALPHA),yes)
+OBJS += alpha/dsputil_alpha.o alpha/mpegvideo_alpha.o
+CFLAGS += -Wa,-mpca56
+endif
+
 SRCS = $(OBJS:.o=.c) $(ASM_OBJS:.o=.s)
 
 LIB= libavcodec.a
@@ -74,6 +80,7 @@ clean:
 	rm -f *.o *~ $(LIB) $(SLIB) *.so i386/*.o i386/*~ \
 	   armv4l/*.o armv4l/*~ \
 	   mlib/*.o mlib/*~ \
+	   alpha/*.o alpha/*~ \
            libac3/*.o libac3/*~ \
            apiexample $(TESTS)
 

libavcodec/alpha/asm.h

+/*
+ * Alpha optimized DSP utils
+ * Copyright (c) 2002 Falk Hueffner <falk@debian.org>
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program 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 General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ */
+
+#ifndef LIBAVCODEC_ALPHA_ASM_H
+#define LIBAVCODEC_ALPHA_ASM_H
+
+#include <stdint.h>
+
+#define AMASK_BWX (1 << 0)
+#define AMASK_FIX (1 << 1)
+#define AMASK_MVI (1 << 8)
+
+static inline uint64_t BYTE_VEC(uint64_t x)
+{
+    x |= x <<  8;
+    x |= x << 16;
+    x |= x << 32;
+    return x;
+}
+static inline uint64_t WORD_VEC(uint64_t x)
+{
+    x |= x << 16;
+    x |= x << 32;
+    return x;
+}
+
+static inline int32_t ldl(const void* p)
+{
+    return *(const int32_t*) p;
+}
+static inline uint64_t ldq(const void* p)
+{
+    return *(const uint64_t*) p;
+}
+/* FIXME ccc doesn't seem to get it? Use inline asm?  */
+static inline uint64_t ldq_u(const void* p)
+{
+    return *(const uint64_t*) ((uintptr_t) p & ~7ul);
+}
+static inline void stl(uint32_t l, void* p)
+{
+    *(uint32_t*) p = l;
+}
+static inline void stq(uint64_t l, void* p)
+{
+    *(uint64_t*) p = l;
+}
+
+#ifdef __GNUC__
+#define OPCODE1(name)						\
+static inline uint64_t name(uint64_t l)				\
+{								\
+    uint64_t r;							\
+    asm (#name " %1, %0" : "=r" (r) : "r" (l));			\
+    return r;							\
+}
+
+#define OPCODE2(name)						\
+static inline uint64_t name(uint64_t l1, uint64_t l2)		\
+{								\
+    uint64_t r;							\
+    asm (#name " %1, %2, %0" : "=r" (r) : "r" (l1), "rI" (l2));	\
+    return r;							\
+}
+
+/* We don't want gcc to move this around or combine it with another
+   rpcc, so mark it volatile.  */
+static inline uint64_t rpcc(void)
+{
+    uint64_t r;
+    asm volatile ("rpcc %0" : "=r" (r));
+    return r;
+}
+
+static inline uint64_t uldq(const void* v)
+{
+    struct foo {
+	unsigned long l;
+    } __attribute__((packed));
+
+    return ((const struct foo*) v)->l;
+}
+
+#elif defined(__DECC)		/* Compaq "ccc" compiler */
+
+#include <c_asm.h>
+#define OPCODE1(name)							\
+static inline uint64_t name(uint64_t l)					\
+{									\
+    return asm (#name " %a0, %v0", l);					\
+}
+
+#define OPCODE2(name)							\
+static inline uint64_t name(uint64_t l1, uint64_t l2)			\
+{									\
+    return asm (#name " %a0, %a1, %v0", l1, l2);			\
+}
+
+static inline uint64_t rpcc(void)
+{
+    return asm  ("rpcc %v0");
+}
+
+static inline uint64_t uldq(const void* v)
+{
+    return *(const __unaligned uint64_t *) v;
+}
+
+#endif
+
+OPCODE1(amask);
+OPCODE1(unpkbw);
+OPCODE1(pkwb);
+OPCODE2(extql);
+OPCODE2(extqh);
+OPCODE2(zap);
+OPCODE2(cmpbge);
+OPCODE2(minsw4);
+OPCODE2(minuw4);
+OPCODE2(minub8);
+OPCODE2(maxsw4);
+OPCODE2(maxuw4);
+OPCODE2(perr);
+
+#endif /* LIBAVCODEC_ALPHA_ASM_H */

libavcodec/alpha/dsputil_alpha.c

+/*
+ * Alpha optimized DSP utils
+ * Copyright (c) 2002 Falk Hueffner <falk@debian.org>
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program 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 General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ */
+
+#include "asm.h"
+#include "../dsputil.h"
+
+void simple_idct_axp(DCTELEM *block);
+
+static void put_pixels_clamped_axp(const DCTELEM *block, UINT8 *pixels, 
+				   int line_size)
+{
+    int i = 8;
+    do {
+	UINT64 shorts;
+
+	shorts = ldq(block);
+	shorts = maxsw4(shorts, 0);
+	shorts = minsw4(shorts, WORD_VEC(0x00ff));
+	stl(pkwb(shorts), pixels);
+
+	shorts = ldq(block + 4);
+	shorts = maxsw4(shorts, 0);
+	shorts = minsw4(shorts, WORD_VEC(0x00ff));
+	stl(pkwb(shorts), pixels + 4);
+
+	pixels += line_size;
+	block += 8;
+    } while (--i);
+}
+
+static void add_pixels_clamped_axp(const DCTELEM *block, UINT8 *pixels, 
+				   int line_size)
+{
+    int i = 8;
+    do {
+	UINT64 shorts; 
+
+	shorts = ldq(block);
+	shorts &= ~WORD_VEC(0x8000); /* clear highest bit to avoid overflow */
+	shorts += unpkbw(ldl(pixels));
+	shorts &= ~WORD_VEC(0x8000); /* hibit would be set for e. g. -2 + 3 */
+	shorts = minuw4(shorts, WORD_VEC(0x4000)); /* set neg. to 0x4000 */
+	shorts &= ~WORD_VEC(0x4000); /* ...and zap them */
+	shorts = minsw4(shorts, WORD_VEC(0x00ff)); /* clamp to 255 */
+	stl(pkwb(shorts), pixels);
+
+	/* next 4 */
+	shorts = ldq(block + 4);
+	shorts &= ~WORD_VEC(0x8000);
+	shorts += unpkbw(ldl(pixels + 4));
+	shorts &= ~WORD_VEC(0x8000);
+	shorts = minuw4(shorts, WORD_VEC(0x4000));
+	shorts &= ~WORD_VEC(0x4000);
+	shorts = minsw4(shorts, WORD_VEC(0x00ff));
+	stl(pkwb(shorts), pixels + 4);
+
+	pixels += line_size;
+	block += 8;
+    } while (--i);
+}
+
+/* Average 8 unsigned bytes in parallel: (b1 + b2) >> 1
+   Since the immediate result could be greater than 255, we do the
+   shift first. The result is too low by one if the bytes were both
+   odd, so we need to add (l1 & l2) & BYTE_VEC(0x01).  */
+static inline UINT64 avg2_no_rnd(UINT64 l1, UINT64 l2)
+{
+    UINT64 correction = (l1 & l2) & BYTE_VEC(0x01);
+    l1 = (l1 & ~BYTE_VEC(0x01)) >> 1;
+    l2 = (l2 & ~BYTE_VEC(0x01)) >> 1;
+    return l1 + l2 + correction;
+}
+
+/* Average 8 bytes with rounding: (b1 + b2 + 1) >> 1
+   The '1' only has an effect when one byte is even and the other odd,
+   i. e. we also need to add (l1 ^ l2) & BYTE_VEC(0x01).
+   Incidentally, that is equivalent to (l1 | l2) & BYTE_VEC(0x01).  */
+static inline UINT64 avg2(UINT64 l1, UINT64 l2)
+{
+    UINT64 correction = (l1 | l2) & BYTE_VEC(0x01);
+    l1 = (l1 & ~BYTE_VEC(0x01)) >> 1;
+    l2 = (l2 & ~BYTE_VEC(0x01)) >> 1;
+    return l1 + l2 + correction;
+}
+
+static inline UINT64 avg4(UINT64 l1, UINT64 l2, UINT64 l3, UINT64 l4)
+{
+    UINT64 r1 = ((l1 & ~BYTE_VEC(0x03)) >> 2)
+	      + ((l2 & ~BYTE_VEC(0x03)) >> 2)
+	      + ((l3 & ~BYTE_VEC(0x03)) >> 2)
+	      + ((l4 & ~BYTE_VEC(0x03)) >> 2);
+    UINT64 r2 = ((  (l1 & BYTE_VEC(0x03))
+		  + (l2 & BYTE_VEC(0x03))
+		  + (l3 & BYTE_VEC(0x03))
+		  + (l4 & BYTE_VEC(0x03))
+		  + BYTE_VEC(0x02)) >> 2) & BYTE_VEC(0x03);
+    return r1 + r2;
+}
+
+static inline UINT64 avg4_no_rnd(UINT64 l1, UINT64 l2, UINT64 l3, UINT64 l4)
+{
+    UINT64 r1 = ((l1 & ~BYTE_VEC(0x03)) >> 2)
+	      + ((l2 & ~BYTE_VEC(0x03)) >> 2)
+	      + ((l3 & ~BYTE_VEC(0x03)) >> 2)
+	      + ((l4 & ~BYTE_VEC(0x03)) >> 2);
+    UINT64 r2 = (( (l1 & BYTE_VEC(0x03))
+		 + (l2 & BYTE_VEC(0x03))
+		 + (l3 & BYTE_VEC(0x03))
+		 + (l4 & BYTE_VEC(0x03))
+		 + BYTE_VEC(0x01)) >> 2) & BYTE_VEC(0x03);
+    return r1 + r2;
+}
+
+#define PIXOPNAME(suffix) put ## suffix
+#define BTYPE UINT8
+#define AVG2 avg2
+#define AVG4 avg4
+#define STORE(l, b) stq(l, b)
+#include "pixops.h"
+#undef PIXOPNAME
+#undef BTYPE
+#undef AVG2
+#undef AVG4
+#undef STORE
+
+#define PIXOPNAME(suffix) put_no_rnd ## suffix
+#define BTYPE UINT8
+#define AVG2 avg2_no_rnd
+#define AVG4 avg4_no_rnd
+#define STORE(l, b) stq(l, b)
+#include "pixops.h"
+#undef PIXOPNAME
+#undef BTYPE
+#undef AVG2
+#undef AVG4
+#undef STORE
+
+/* The following functions are untested.  */
+#if 0
+
+#define PIXOPNAME(suffix) avg ## suffix
+#define BTYPE UINT8
+#define AVG2 avg2
+#define AVG4 avg4
+#define STORE(l, b) stq(AVG2(l, ldq(b)), b);
+#include "pixops.h"
+#undef PIXOPNAME
+#undef BTYPE
+#undef AVG2
+#undef AVG4
+#undef STORE
+
+#define PIXOPNAME(suffix) avg_no_rnd ## suffix
+#define BTYPE UINT8
+#define AVG2 avg2_no_rnd
+#define AVG4 avg4_no_rnd
+#define STORE(l, b) stq(AVG2(l, ldq(b)), b);
+#include "pixops.h"
+#undef PIXOPNAME
+#undef BTYPE
+#undef AVG2
+#undef AVG4
+#undef STORE
+
+#define PIXOPNAME(suffix) sub ## suffix
+#define BTYPE DCTELEM
+#define AVG2 avg2
+#define AVG4 avg4
+#define STORE(l, block) do {		\
+    UINT64 xxx = l;			\
+    (block)[0] -= (xxx >>  0) & 0xff;	\
+    (block)[1] -= (xxx >>  8) & 0xff;	\
+    (block)[2] -= (xxx >> 16) & 0xff;	\
+    (block)[3] -= (xxx >> 24) & 0xff;	\
+    (block)[4] -= (xxx >> 32) & 0xff;	\
+    (block)[5] -= (xxx >> 40) & 0xff;	\
+    (block)[6] -= (xxx >> 48) & 0xff;	\
+    (block)[7] -= (xxx >> 56) & 0xff;	\
+} while (0)
+#include "pixops.h"
+#undef PIXOPNAME
+#undef BTYPE
+#undef AVG2
+#undef AVG4
+#undef STORE
+
+#endif
+
+void dsputil_init_alpha(void)
+{
+    put_pixels_tab[0] = put_pixels_axp;
+    put_pixels_tab[1] = put_pixels_x2_axp;
+    put_pixels_tab[2] = put_pixels_y2_axp;
+    put_pixels_tab[3] = put_pixels_xy2_axp;
+
+    put_no_rnd_pixels_tab[0] = put_pixels_axp;
+    put_no_rnd_pixels_tab[1] = put_no_rnd_pixels_x2_axp;
+    put_no_rnd_pixels_tab[2] = put_no_rnd_pixels_y2_axp;
+    put_no_rnd_pixels_tab[3] = put_no_rnd_pixels_xy2_axp;
+
+    /* amask clears all bits that correspond to present features.  */
+    if (amask(AMASK_MVI) == 0) {
+	fprintf(stderr, "MVI extension detected\n");
+	put_pixels_clamped = put_pixels_clamped_axp;
+	add_pixels_clamped = add_pixels_clamped_axp;
+    }
+}

libavcodec/alpha/mpegvideo_alpha.c

+/*
+ * Alpha optimized DSP utils
+ * Copyright (c) 2002 Falk Hueffner <falk@debian.org>
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program 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 General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ */
+
+#include "asm.h"
+#include "../dsputil.h"
+#include "../mpegvideo.h"
+
+extern UINT8 zigzag_end[64];
+
+static void dct_unquantize_h263_axp(MpegEncContext *s, 
+				    DCTELEM *block, int n, int qscale)
+{
+    int i, level;
+    UINT64 qmul, qadd;
+    if (s->mb_intra) {
+        if (n < 4) 
+            block[0] = block[0] * s->y_dc_scale;
+        else
+            block[0] = block[0] * s->c_dc_scale;
+	/* Catch up to aligned point.  */
+	qmul = s->qscale << 1;
+	qadd = (s->qscale - 1) | 1;
+	for (i = 1; i < 4; ++i) {
+	    level = block[i];
+	    if (level) {
+		if (level < 0) {
+		    level = level * qmul - qadd;
+		} else {
+		    level = level * qmul + qadd;
+		}
+		block[i] = level;
+	    }
+	}
+	block += 4;
+	i = 60 / 4;
+    } else {
+        i = zigzag_end[s->block_last_index[n]] / 4;
+    }
+    qmul = s->qscale << 1;
+    qadd = WORD_VEC((qscale - 1) | 1);
+    do {
+	UINT64 levels, negmask, zeromask, corr;
+	levels = ldq(block);
+	if (levels == 0)
+	    continue;
+	zeromask = cmpbge(0, levels);
+	zeromask &= zeromask >> 1;
+	/* Negate all negative words.  */
+	negmask = maxsw4(levels, WORD_VEC(0xffff)); /* negative -> ffff (-1) */
+	negmask = minsw4(negmask, 0);		    /* positive -> 0000 (0) */
+	corr    = negmask & WORD_VEC(0x0001); /* twos-complement correction */
+	levels ^= negmask;
+	levels += corr;
+
+	levels = levels * qmul;
+	levels += zap(qadd, zeromask);
+
+	/* Re-negate negative words.  */
+	levels -= corr;
+	levels ^= negmask;
+
+	stq(levels, block);
+    } while (block += 4, --i);
+}
+
+void MPV_common_init_axp(MpegEncContext *s)
+{
+    if (amask(AMASK_MVI) == 0) {
+        if (s->out_format == FMT_H263)
+	    s->dct_unquantize = dct_unquantize_h263_axp;
+    }
+}

libavcodec/alpha/pixops.h

+/*
+ * Alpha optimized DSP utils
+ * Copyright (c) 2002 Falk Hueffner <falk@debian.org>
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program 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 General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ */
+
+/* This file is intended to be #included with proper definitions of
+ * PIXOPNAME, BTYPE, AVG2, AVG4 and STORE.  */
+
+static void PIXOPNAME(_pixels_axp)(BTYPE *block, const UINT8 *pixels,
+				   int line_size, int h)
+{
+    if ((size_t) pixels & 0x7) {
+	do {
+	    STORE(uldq(pixels), block);
+	    pixels += line_size;
+	    block  += line_size;
+	} while (--h);
+    } else {
+	do {
+	    STORE(ldq(pixels), block);
+	    pixels += line_size;
+	    block  += line_size;
+	} while (--h);
+    }
+}
+
+static void PIXOPNAME(_pixels_x2_axp)(BTYPE *block, const UINT8 *pixels,
+				      int line_size, int h)
+{
+    if ((size_t) pixels & 0x7) {
+	do {
+	    UINT64 pix1, pix2;
+
+	    pix1 = uldq(pixels);
+	    pix2 = pix1 >> 8 | ((UINT64) pixels[8] << 56);
+	    STORE(AVG2(pix1, pix2), block);
+	    pixels += line_size;
+	    block += line_size;
+	} while (--h);
+    } else {
+	do {
+	    UINT64 pix1, pix2;
+
+	    pix1 = ldq(pixels);
+	    pix2 = pix1 >> 8 | ((UINT64) pixels[8] << 56);
+	    STORE(AVG2(pix1, pix2), block);
+	    pixels += line_size;
+	    block += line_size;
+	} while (--h);
+    }
+}
+
+static void PIXOPNAME(_pixels_y2_axp)(BTYPE *block, const UINT8 *pixels,
+				      int line_size, int h)
+{
+    if ((size_t) pixels & 0x7) {
+	UINT64 pix = uldq(pixels);
+	do {
+	    UINT64 next_pix;
+
+	    pixels += line_size;
+	    next_pix = uldq(pixels);
+	    STORE(AVG2(pix, next_pix), block);
+	    block += line_size;
+	    pix = next_pix;
+	} while (--h);
+    } else {
+	UINT64 pix = ldq(pixels);
+	do {
+	    UINT64 next_pix;
+
+	    pixels += line_size;
+	    next_pix = ldq(pixels);
+	    STORE(AVG2(pix, next_pix), block);
+	    block += line_size;
+	    pix = next_pix;
+	} while (--h);
+    }
+}
+
+/* This could be further sped up by recycling AVG4 intermediate
+  results from the previous loop pass.  */
+static void PIXOPNAME(_pixels_xy2_axp)(BTYPE *block, const UINT8 *pixels,
+				       int line_size, int h)
+{
+    if ((size_t) pixels & 0x7) {
+	UINT64 pix1 = uldq(pixels);
+	UINT64 pix2 = pix1 >> 8 | ((UINT64) pixels[8] << 56);
+
+	do {
+	    UINT64 next_pix1, next_pix2;
+
+	    pixels += line_size;
+	    next_pix1 = uldq(pixels);
+	    next_pix2 = next_pix1 >> 8 | ((UINT64) pixels[8] << 56);
+
+	    STORE(AVG4(pix1, pix2, next_pix1, next_pix2), block);
+
+	    block += line_size;
+	    pix1 = next_pix1;
+	    pix2 = next_pix2;
+	} while (--h);
+    } else {
+	UINT64 pix1 = ldq(pixels);
+	UINT64 pix2 = pix1 >> 8 | ((UINT64) pixels[8] << 56);
+
+	do {
+	    UINT64 next_pix1, next_pix2;
+
+	    pixels += line_size;
+	    next_pix1 = ldq(pixels);
+	    next_pix2 = next_pix1 >> 8 | ((UINT64) pixels[8] << 56);
+
+	    STORE(AVG4(pix1, pix2, next_pix1, next_pix2), block);
+
+	    block += line_size;
+	    pix1 = next_pix1;
+	    pix2 = next_pix2;
+	} while (--h);
+    }
+}

libavcodec/dsputil.c

@@ -497,6 +497,10 @@ void dsputil_init(void)
     dsputil_init_mlib();
     use_permuted_idct = 0;
 #endif
+#ifdef ARCH_ALPHA
+    dsputil_init_alpha();
+    use_permuted_idct = 0;
+#endif
 
 #ifdef SIMPLE_IDCT
     if(ff_idct == simple_idct) use_permuted_idct=0;

libavcodec/dsputil.h

@@ -123,6 +123,13 @@ void dsputil_init_armv4l(void);
 
 void dsputil_init_mlib(void);   
 
+#elif defined(ARCH_ALPHA)
+
+#define emms_c()
+#define __align8 __attribute__ ((aligned (8)))
+
+void dsputil_init_alpha(void);
+
 #else
 
 #define emms_c()

libavcodec/msmpeg4.c

@@ -460,6 +460,18 @@ static int msmpeg4_pred_dc(MpegEncContext * s, int n,
 	: "r" (scale)
 	: "%eax", "%edx"
     );
+#elif defined (ARCH_ALPHA)
+    /* Divisions are extremely costly on Alpha; optimize the most
+       common case.  */
+    if (scale == 8) {
+	a = (a + (8 >> 1)) / 8;
+	b = (b + (8 >> 1)) / 8;
+	c = (c + (8 >> 1)) / 8;
+    } else {
+	a = (a + (scale >> 1)) / scale;
+	b = (b + (scale >> 1)) / scale;
+	c = (c + (scale >> 1)) / scale;
+    }
 #else
     a = (a + (scale >> 1)) / scale;
     b = (b + (scale >> 1)) / scale;

libavcodec/simple_idct.c

@@ -23,6 +23,7 @@
 #include <inttypes.h>
 
 #include "simple_idct.h"
+#include "../config.h"
 
 #if 0
 #define W1 2841 /* 2048*sqrt (2)*cos (1*pi/16) */
@@ -102,6 +103,107 @@ static int inline idctRowCondZ (int16_t * row)
 	return 1;
 }
 
+#ifdef ARCH_ALPHA
+static int inline idctRowCondDC(int16_t *row)
+{
+	int_fast32_t a0, a1, a2, a3, b0, b1, b2, b3;
+	uint64_t *lrow = (uint64_t *) row;
+
+	if (lrow[1] == 0) {
+		if (lrow[0] == 0)
+			return 0;
+		if ((lrow[0] & ~0xffffULL) == 0) {
+			uint64_t v;
+
+			a0 = W4 * row[0];
+			a0 += 1 << (ROW_SHIFT - 1);
+			a0 >>= ROW_SHIFT;
+			v = (uint16_t) a0;
+			v += v << 16;
+			v += v << 32;
+			lrow[0] = v;
+			lrow[1] = v;
+
+			return 1;
+		}
+	}
+
+	a0 = W4 * row[0];
+	a1 = W4 * row[0];
+	a2 = W4 * row[0];
+	a3 = W4 * row[0];
+
+	if (row[2]) {
+		a0 += W2 * row[2];
+		a1 += W6 * row[2];
+		a2 -= W6 * row[2];
+		a3 -= W2 * row[2];
+	}
+
+	if (row[4]) {
+		a0 += W4 * row[4];
+		a1 -= W4 * row[4];
+		a2 -= W4 * row[4];
+		a3 += W4 * row[4];
+	}
+
+	if (row[6]) {
+		a0 += W6 * row[6];
+		a1 -= W2 * row[6];
+		a2 += W2 * row[6];
+		a3 -= W6 * row[6];
+	}
+
+	a0 += 1 << (ROW_SHIFT - 1);
+	a1 += 1 << (ROW_SHIFT - 1);
+	a2 += 1 << (ROW_SHIFT - 1);
+	a3 += 1 << (ROW_SHIFT - 1);
+
+	if (row[1]) {
+		b0 = W1 * row[1];
+		b1 = W3 * row[1];
+		b2 = W5 * row[1];
+		b3 = W7 * row[1];
+	} else {
+		b0 = 0;
+		b1 = 0;
+		b2 = 0;
+		b3 = 0;
+	}
+
+	if (row[3]) {
+		b0 += W3 * row[3];
+		b1 -= W7 * row[3];
+		b2 -= W1 * row[3];
+		b3 -= W5 * row[3];
+	}
+
+	if (row[5]) {
+		b0 += W5 * row[5];
+		b1 -= W1 * row[5];
+		b2 += W7 * row[5];
+		b3 += W3 * row[5];
+	}
+
+	if (row[7]) {
+		b0 += W7 * row[7];
+		b1 -= W5 * row[7];
+		b2 += W3 * row[7];
+		b3 -= W1 * row[7];
+	}
+
+	row[0] = (a0 + b0) >> ROW_SHIFT;
+	row[1] = (a1 + b1) >> ROW_SHIFT;
+	row[2] = (a2 + b2) >> ROW_SHIFT;
+	row[3] = (a3 + b3) >> ROW_SHIFT;
+	row[4] = (a3 - b3) >> ROW_SHIFT;
+	row[5] = (a2 - b2) >> ROW_SHIFT;
+	row[6] = (a1 - b1) >> ROW_SHIFT;
+	row[7] = (a0 - b0) >> ROW_SHIFT;
+
+	return 1;
+}
+#else  /* not ARCH_ALPHA */
 static int inline idctRowCondDC (int16_t * row)
 {
 	int a0, a1, a2, a3, b0, b1, b2, b3;
@@ -147,6 +249,7 @@ static int inline idctRowCondDC (int16_t * row)
 	
 	return 1;
 }
+#endif /* not ARCH_ALPHA */
 
 static void inline idctCol (int16_t * col)
 {
@@ -243,6 +346,7 @@ static void inline idctSparseCol (int16_t * col)
 		b3 += - W1*col[8*7];
 	}
 
+#ifndef ARCH_ALPHA
 	if(!(b0|b1|b2|b3)){
 		col[8*0] = (a0) >> COL_SHIFT;
 		col[8*7] = (a0) >> COL_SHIFT;
@@ -253,6 +357,7 @@ static void inline idctSparseCol (int16_t * col)
 		col[8*3] = (a3) >> COL_SHIFT;
 		col[8*4] = (a3) >> COL_SHIFT;
 	}else{
+#endif
 		col[8*0] = (a0 + b0) >> COL_SHIFT;
 		col[8*7] = (a0 - b0) >> COL_SHIFT;
 		col[8*1] = (a1 + b1) >> COL_SHIFT;
@@ -261,7 +366,9 @@ static void inline idctSparseCol (int16_t * col)
 		col[8*5] = (a2 - b2) >> COL_SHIFT;
 		col[8*3] = (a3 + b3) >> COL_SHIFT;
 		col[8*4] = (a3 - b3) >> COL_SHIFT;
+#ifndef ARCH_ALPHA
 	}
+#endif
 }
 
 static void inline idctSparse2Col (int16_t * col)
@@ -337,6 +444,34 @@ static void inline idctSparse2Col (int16_t * col)
 	col[8*4] = (a3 - b3) >> COL_SHIFT;
 }
 
+#ifdef ARCH_ALPHA
+/* If all rows but the first one are zero after row transformation,
+   all rows will be identical after column transformation.  */
+static inline void idctCol2(int16_t *col)
+{
+	int i;
+	uint64_t l, r;
+	uint64_t *lcol = (uint64_t *) col;
+
+	for (i = 0; i < 8; ++i) {
+		int a0 = col[0] + (1 << (COL_SHIFT - 1)) / W4;
+
+		a0 *= W4;
+		col[0] = a0 >> COL_SHIFT;
+		++col;
+	}
+
+	l = lcol[0];
+	r = lcol[1];
+	lcol[ 2] = l; lcol[ 3] = r;
+	lcol[ 4] = l; lcol[ 5] = r;
+	lcol[ 6] = l; lcol[ 7] = r;
+	lcol[ 8] = l; lcol[ 9] = r;
+	lcol[10] = l; lcol[11] = r;
+	lcol[12] = l; lcol[13] = r;
+	lcol[14] = l; lcol[15] = r;
+}
+#endif
 
 void simple_idct (short *block)
 {
@@ -411,6 +546,21 @@ void simple_idct (short *block)
 		for(i=0; i<8; i++)
 			idctSparse2Col(block + i);
 	}
+#elif defined(ARCH_ALPHA)
+	int shortcut = 1;
+
+	for (i = 0; i < 8; i++) {
+		int anynonzero = idctRowCondDC(block + 8 * i);
+		if (i > 0 && anynonzero)
+			shortcut = 0;
+	}
+
+	if (shortcut) {
+		idctCol2(block);
+	} else {
+		for (i = 0; i < 8; i++)
+			idctSparseCol(block + i);
+	}
 #else
 	for(i=0; i<8; i++)
 		idctRowCondDC(block + i*8);