Move Alpha optimized IDCT to own file. Based on a patch by MÃ¥ns

Rullgård <mru@users.sourceforge.net>.

I've left out the idctCol2 part, because W4 has recently been decreed
to be 16383, and also I doubt it will give a noticeable speedup.

Originally committed as revision 1029 to svn://svn.ffmpeg.org/ffmpeg/trunk

5 files changed, 313 insertions, 267 deletions

diff --git a/libavcodec/Makefile b/libavcodec/Makefile
index c956f052d7..286e5285b7 100644
--- a/libavcodec/Makefile
+++ b/libavcodec/Makefile
@@ -67,7 +67,8 @@ endif
 
 # alpha specific stuff
 ifeq ($(TARGET_ARCH_ALPHA),yes)
-OBJS += alpha/dsputil_alpha.o alpha/mpegvideo_alpha.o alpha/motion_est_alpha.o
+OBJS += alpha/dsputil_alpha.o alpha/mpegvideo_alpha.o \
+	alpha/simple_idct_alpha.o alpha/motion_est_alpha.o
 ASM_OBJS += alpha/dsputil_alpha_asm.o alpha/motion_est_mvi_asm.o
 CFLAGS += -fforce-addr -freduce-all-givs
 endif
diff --git a/libavcodec/alpha/dsputil_alpha.c b/libavcodec/alpha/dsputil_alpha.c
index 3bf2290a24..371e20269b 100644
--- a/libavcodec/alpha/dsputil_alpha.c
+++ b/libavcodec/alpha/dsputil_alpha.c
@@ -20,8 +20,6 @@
 #include "asm.h"
 #include "../dsputil.h"
 
-void simple_idct_axp(DCTELEM *block);
-
 void put_pixels_axp_asm(uint8_t *block, const uint8_t *pixels,
 			int line_size, int h);
 void put_pixels_clamped_mvi_asm(const DCTELEM *block, uint8_t *pixels,
diff --git a/libavcodec/alpha/mpegvideo_alpha.c b/libavcodec/alpha/mpegvideo_alpha.c
index 1a58e5dfd0..350b53f62f 100644
--- a/libavcodec/alpha/mpegvideo_alpha.c
+++ b/libavcodec/alpha/mpegvideo_alpha.c
@@ -21,6 +21,9 @@
 #include "../dsputil.h"
 #include "../mpegvideo.h"
 
+extern void simple_idct_put_axp(uint8_t *dest, int line_size, DCTELEM *block);
+extern void simple_idct_add_axp(uint8_t *dest, int line_size, DCTELEM *block);
+  
 static void dct_unquantize_h263_axp(MpegEncContext *s, DCTELEM *block,
                                     int n, int qscale)
 {
@@ -94,4 +97,6 @@ static void dct_unquantize_h263_axp(MpegEncContext *s, DCTELEM *block,
 void MPV_common_init_axp(MpegEncContext *s)
 {
     s->dct_unquantize_h263 = dct_unquantize_h263_axp;
+    s->idct_put = simple_idct_put_axp;
+    s->idct_add = simple_idct_add_axp;
 }
diff --git a/libavcodec/alpha/simple_idct_alpha.c b/libavcodec/alpha/simple_idct_alpha.c
new file mode 100644
index 0000000000..31677184e0
--- /dev/null
+++ b/libavcodec/alpha/simple_idct_alpha.c
@@ -0,0 +1,306 @@
+/*
+ * Simple IDCT (Alpha optimized)
+ *
+ * Copyright (c) 2001 Michael Niedermayer <michaelni@gmx.at>
+ *
+ * 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
+ *
+ * based upon some outcommented c code from mpeg2dec (idct_mmx.c
+ * written by Aaron Holtzman <aholtzma@ess.engr.uvic.ca>)
+ *
+ * Alpha optimiziations by Måns Rullgård <mru@users.sourceforge.net>
+ *                     and Falk Hueffner <falk@debian.org>
+ */
+
+#include "asm.h"
+#include "../dsputil.h"
+
+// cos(i * M_PI / 16) * sqrt(2) * (1 << 14)
+// W4 is actually exactly 16384, but using 16383 works around
+// accumulating rounding errors for some encoders
+#define W1 ((int_fast32_t) 22725)
+#define W2 ((int_fast32_t) 21407)
+#define W3 ((int_fast32_t) 19266)
+#define W4 ((int_fast32_t) 16383)
+#define W5 ((int_fast32_t) 12873)
+#define W6 ((int_fast32_t)  8867)
+#define W7 ((int_fast32_t)  4520)
+#define ROW_SHIFT 11
+#define COL_SHIFT 20
+
+/* 0: all entries 0, 1: only first entry nonzero, 2: otherwise  */
+static inline int idct_row(DCTELEM *row)
+{
+    int_fast32_t a0, a1, a2, a3, b0, b1, b2, b3, t;
+    uint64_t l, r;
+    l = ldq(row);
+    r = ldq(row + 4);
+
+    if (l == 0 && r == 0)
+        return 0;
+    
+    a0 = W4 * sextw(l) + (1 << (ROW_SHIFT - 1));
+
+    if (((l & ~0xffffUL) | r) == 0) {
+        a0 >>= ROW_SHIFT;
+        a0 = (uint16_t) a0;
+        a0 |= a0 << 16;
+        a0 |= a0 << 32;
+        
+        stq(a0, row);
+        stq(a0, row + 4);
+        return 1;
+    }
+
+    a1 = a0;
+    a2 = a0;
+    a3 = a0;
+
+    t = extwl(l, 4);            /* row[2] */
+    if (t != 0) {
+        t = sextw(t);
+        a0 += W2 * t;
+        a1 += W6 * t;
+        a2 -= W6 * t;
+        a3 -= W2 * t;
+    }
+
+    t = extwl(r, 0);            /* row[4] */
+    if (t != 0) {
+        t = sextw(t);
+        a0 += W4 * t;
+        a1 -= W4 * t;
+        a2 -= W4 * t;
+        a3 += W4 * t;
+    }
+
+    t = extwl(r, 4);            /* row[6] */
+    if (t != 0) {
+        t = sextw(t);
+        a0 += W6 * t;
+        a1 -= W2 * t;
+        a2 += W2 * t;
+        a3 -= W6 * t;
+    }
+
+    t = extwl(l, 2);            /* row[1] */
+    if (t != 0) {
+        t = sextw(t);
+        b0 = W1 * t;
+        b1 = W3 * t;
+        b2 = W5 * t;
+        b3 = W7 * t;
+    } else {
+        b0 = 0;
+        b1 = 0;
+        b2 = 0;
+        b3 = 0;
+    }
+
+    t = extwl(l, 6);            /* row[3] */
+    if (t) {
+        t = sextw(t);
+        b0 += W3 * t;
+        b1 -= W7 * t;
+        b2 -= W1 * t;
+        b3 -= W5 * t;
+    }
+
+    
+    t = extwl(r, 2);            /* row[5] */
+    if (t) {
+        t = sextw(t);
+        b0 += W5 * t;
+        b1 -= W1 * t;
+        b2 += W7 * t;
+        b3 += W3 * t;
+    }
+
+    t = extwl(r, 6);            /* row[7] */
+    if (t) {
+        t = sextw(t);
+        b0 += W7 * t;
+        b1 -= W5 * t;
+        b2 += W3 * t;
+        b3 -= W1 * t;
+    }
+
+    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 2;
+}
+
+static inline void idct_col(DCTELEM *col)
+{
+    int_fast32_t a0, a1, a2, a3, b0, b1, b2, b3;
+
+    col[0] += (1 << (COL_SHIFT - 1)) / W4;
+
+    a0 = W4 * col[8 * 0];
+    a1 = W4 * col[8 * 0];
+    a2 = W4 * col[8 * 0];
+    a3 = W4 * col[8 * 0];
+
+    if (col[8 * 2]) {
+        a0 += W2 * col[8 * 2];
+        a1 += W6 * col[8 * 2];
+        a2 -= W6 * col[8 * 2];
+        a3 -= W2 * col[8 * 2];
+    }
+
+    if (col[8 * 4]) {
+        a0 += W4 * col[8 * 4];
+        a1 -= W4 * col[8 * 4];
+        a2 -= W4 * col[8 * 4];
+        a3 += W4 * col[8 * 4];
+    }
+
+    if (col[8 * 6]) {
+        a0 += W6 * col[8 * 6];
+        a1 -= W2 * col[8 * 6];
+        a2 += W2 * col[8 * 6];
+        a3 -= W6 * col[8 * 6];
+    }
+
+    if (col[8 * 1]) {
+        b0 = W1 * col[8 * 1];
+        b1 = W3 * col[8 * 1];
+        b2 = W5 * col[8 * 1];
+        b3 = W7 * col[8 * 1];
+    } else {
+        b0 = 0;
+        b1 = 0;
+        b2 = 0;
+        b3 = 0;
+    }
+
+    if (col[8 * 3]) {
+        b0 += W3 * col[8 * 3];
+        b1 -= W7 * col[8 * 3];
+        b2 -= W1 * col[8 * 3];
+        b3 -= W5 * col[8 * 3];
+    }
+
+    if (col[8 * 5]) {
+        b0 += W5 * col[8 * 5];
+        b1 -= W1 * col[8 * 5];
+        b2 += W7 * col[8 * 5];
+        b3 += W3 * col[8 * 5];
+    }
+
+    if (col[8 * 7]) {
+        b0 += W7 * col[8 * 7];
+        b1 -= W5 * col[8 * 7];
+        b2 += W3 * col[8 * 7];
+        b3 -= W1 * col[8 * 7];
+    }
+
+    col[8 * 0] = (a0 + b0) >> COL_SHIFT;
+    col[8 * 7] = (a0 - b0) >> COL_SHIFT;
+    col[8 * 1] = (a1 + b1) >> COL_SHIFT;
+    col[8 * 6] = (a1 - b1) >> COL_SHIFT;
+    col[8 * 2] = (a2 + b2) >> COL_SHIFT;
+    col[8 * 5] = (a2 - b2) >> COL_SHIFT;
+    col[8 * 3] = (a3 + b3) >> COL_SHIFT;
+    col[8 * 4] = (a3 - b3) >> COL_SHIFT;
+}
+
+/* If all rows but the first one are zero after row transformation,
+   all rows will be identical after column transformation.  */
+static inline void idct_col2(DCTELEM *col)
+{
+    int i;
+    uint64_t l, r;
+    uint64_t *lcol = (uint64_t *) col;
+
+    for (i = 0; i < 8; ++i) {
+        int_fast32_t 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;
+}
+
+void simple_idct_axp(DCTELEM *block)
+{
+
+    int i;
+    int rowsZero = 1;           /* all rows except row 0 zero */
+    int rowsConstant = 1;       /* all rows consist of a constant value */
+
+    for (i = 0; i < 8; i++) {
+        int sparseness = idct_row(block + 8 * i);
+
+        if (i > 0 && sparseness > 0)
+            rowsZero = 0;
+        if (sparseness == 2)
+            rowsConstant = 0;
+    }
+
+    if (rowsZero) {
+        idct_col2(block);
+    } else if (rowsConstant) {
+        uint64_t *lblock = (uint64_t *) block;
+
+        idct_col(block);
+        for (i = 0; i < 8; i += 2) {
+            uint64_t v = (uint16_t) block[i * 8];
+            uint64_t w = (uint16_t) block[i * 8 + 8];
+
+            v |= v << 16;
+            w |= w << 16;
+            v |= v << 32;
+            w |= w << 32;
+            lblock[0] = v;
+            lblock[1] = v;
+            lblock[2] = w;
+            lblock[3] = w;
+            lblock += 4;
+        }
+    } else {
+        for (i = 0; i < 8; i++)
+            idct_col(block + i);
+    }
+}
+
+void simple_idct_put_axp(uint8_t *dest, int line_size, DCTELEM *block)
+{
+    simple_idct_axp(block);
+    put_pixels_clamped(block, dest, line_size);
+}
+
+void simple_idct_add_axp(uint8_t *dest, int line_size, DCTELEM *block)
+{
+    simple_idct_axp(block);
+    add_pixels_clamped(block, dest, line_size);
+}
diff --git a/libavcodec/simple_idct.c b/libavcodec/simple_idct.c
index ffb79f6db3..ad27ac5942 100644
--- a/libavcodec/simple_idct.c
+++ b/libavcodec/simple_idct.c
@@ -25,8 +25,6 @@
 #include "dsputil.h"
 #include "simple_idct.h"
 
-//#define ARCH_ALPHA
-
 #if 0
 #define W1 2841 /* 2048*sqrt (2)*cos (1*pi/16) */
 #define W2 2676 /* 2048*sqrt (2)*cos (2*pi/16) */
@@ -49,10 +47,6 @@
 #define COL_SHIFT 20 // 6
 #endif
 
-#ifdef ARCH_ALPHA
-#define FAST_64BIT
-#endif
-
 #if defined(ARCH_POWERPC_405)
 
 /* signed 16x16 -> 32 multiply add accumulate */
@@ -73,180 +67,6 @@
 
 #endif
 
-#ifdef ARCH_ALPHA
-/* 0: all entries 0, 1: only first entry nonzero, 2: otherwise  */
-static inline int 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;
-#if 1			//is ok if |a0| < 1024 than theres an +-1 error (for the *W4 case for W4=16383 !!!)
-			a0 = row[0]<<3;
-#else
-			a0 = W4 * row[0];
-			a0 += 1 << (ROW_SHIFT - 1);
-			a0 >>= ROW_SHIFT;
-#endif
-			v = (uint16_t) a0;
-			v += v << 16;
-			v += v << 32;
-			lrow[0] = v;
-			lrow[1] = v;
-
-			return 1;
-		}
-	}
-
-        a0 = (W4 * row[0]) + (1 << (ROW_SHIFT - 1));
-	a1 = a0;
-	a2 = a0;
-	a3 = a0;
-
-	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];
-	}
-
-	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 2;
-}
-
-inline static void idctSparseCol2(int16_t *col)
-{
-        int a0, a1, a2, a3, b0, b1, b2, b3;
-
-        col[0] += (1 << (COL_SHIFT - 1)) / W4;
-
-        a0 = W4 * col[8 * 0];
-        a1 = W4 * col[8 * 0];
-        a2 = W4 * col[8 * 0];
-        a3 = W4 * col[8 * 0];
-
-        if (col[8 * 2]) {
-                a0 += W2 * col[8 * 2];
-                a1 += W6 * col[8 * 2];
-                a2 -= W6 * col[8 * 2];
-                a3 -= W2 * col[8 * 2];
-        }
-
-        if (col[8 * 4]) {
-                a0 += W4 * col[8 * 4];
-                a1 -= W4 * col[8 * 4];
-                a2 -= W4 * col[8 * 4];
-                a3 += W4 * col[8 * 4];
-        }
-
-        if (col[8 * 6]) {
-                a0 += W6 * col[8 * 6];
-                a1 -= W2 * col[8 * 6];
-                a2 += W2 * col[8 * 6];
-                a3 -= W6 * col[8 * 6];
-        }
-
-        if (col[8 * 1]) {
-                b0 = W1 * col[8 * 1];
-                b1 = W3 * col[8 * 1];
-                b2 = W5 * col[8 * 1];
-                b3 = W7 * col[8 * 1];
-        } else {
-                b0 = b1 = b2 = b3 = 0;
-        }
-
-        if (col[8 * 3]) {
-                b0 += W3 * col[8 * 3];
-                b1 -= W7 * col[8 * 3];
-                b2 -= W1 * col[8 * 3];
-                b3 -= W5 * col[8 * 3];
-        }
-
-        if (col[8 * 5]) {
-                b0 += W5 * col[8 * 5];
-                b1 -= W1 * col[8 * 5];
-                b2 += W7 * col[8 * 5];
-                b3 += W3 * col[8 * 5];
-        }
-
-        if (col[8 * 7]) {
-                b0 += W7 * col[8 * 7];
-                b1 -= W5 * col[8 * 7];
-                b2 += W3 * col[8 * 7];
-                b3 -= W1 * col[8 * 7];
-        }
-
-        col[8 * 0] = (a0 + b0) >> COL_SHIFT;
-        col[8 * 7] = (a0 - b0) >> COL_SHIFT;
-        col[8 * 1] = (a1 + b1) >> COL_SHIFT;
-        col[8 * 6] = (a1 - b1) >> COL_SHIFT;
-        col[8 * 2] = (a2 + b2) >> COL_SHIFT;
-        col[8 * 5] = (a2 - b2) >> COL_SHIFT;
-        col[8 * 3] = (a3 + b3) >> COL_SHIFT;
-        col[8 * 4] = (a3 - b3) >> COL_SHIFT;
-}
-
-#else  /* not ARCH_ALPHA */
-
 static inline void idctRowCondDC (int16_t * row)
 {
 	int a0, a1, a2, a3, b0, b1, b2, b3;
@@ -337,7 +157,6 @@ static inline void idctRowCondDC (int16_t * row)
 	row[3] = (a3 + b3) >> ROW_SHIFT;
 	row[4] = (a3 - b3) >> ROW_SHIFT;
 }
-#endif /* not ARCH_ALPHA */
 
 static inline void idctSparseColPut (UINT8 *dest, int line_size, 
                                      int16_t * col)
@@ -546,87 +365,6 @@ static inline void idctSparseCol (int16_t * col)
         col[56] = ((a0 - b0) >> 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;
-}
-
-void simple_idct (short *block)
-{
-
-	int i;
-        int rowsZero = 1;       /* all rows except row 0 zero */
-        int rowsConstant = 1;	/* all rows consist of a constant value */
-
-	for (i = 0; i < 8; i++) {
-		int sparseness = idctRowCondDC(block + 8 * i);
-
-		if (i > 0 && sparseness > 0)
-                        rowsZero = 0;
-                if (sparseness == 2)
-                        rowsConstant = 0;
-	}
-
-        if (rowsZero) {
-                idctCol2(block);
-        } else if (rowsConstant) {
-		uint64_t *lblock = (uint64_t *) block;
-
-		idctSparseCol2(block);
-		for (i = 0; i < 8; i++) {
-			uint64_t v = (uint16_t) block[i * 8];
-
-			v += v << 16;
-			v += v << 32;
-			lblock[0] = v;
-			lblock[1] = v;
-			lblock += 2;
-		}
-	} else {
-		for (i = 0; i < 8; i++)
-			idctSparseCol2(block + i);
-	}
-}
-
-/* XXX: suppress this mess */
-void simple_idct_put(UINT8 *dest, int line_size, DCTELEM *block)
-{
-    simple_idct(block);
-    put_pixels_clamped(block, dest, line_size);
-}
-
-void simple_idct_add(UINT8 *dest, int line_size, DCTELEM *block)
-{
-    simple_idct(block);
-    add_pixels_clamped(block, dest, line_size);
-}
-
-#else
-
 void simple_idct_put(UINT8 *dest, int line_size, INT16 *block)
 {
     int i;
@@ -657,8 +395,6 @@ void simple_idct(INT16 *block)
         idctSparseCol(block + i);
 }
 
-#endif
-
 /* 2x4x8 idct */
 
 #define CN_SHIFT 12