1. ffmpeg
  2. Commit f9bb4bdffcbde7362db2a0e041a2893dde0ace6f
Browse code

Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win on ev6.

Originally committed as revision 979 to svn://svn.ffmpeg.org/ffmpeg/trunk

Falk Hüffner authored on 2002/09/30 00:14:28
Showing 4 changed files
libavcodec/Makefile
History View file @ f9bb4bd
... ... 
@@ -68,7 +68,7 @@ endif
68 68 
 # alpha specific stuff
69 69 
 ifeq ($(TARGET_ARCH_ALPHA),yes)
70 70 
 OBJS += alpha/dsputil_alpha.o alpha/mpegvideo_alpha.o alpha/motion_est_alpha.o
71 
-ASM_OBJS += alpha/dsputil_alpha_asm.o
71 
+ASM_OBJS += alpha/dsputil_alpha_asm.o alpha/motion_est_mvi_asm.o
72 72 
 CFLAGS += -fforce-addr -freduce-all-givs
73 73 
 endif
74 74
libavcodec/alpha/dsputil_alpha.c
History View file @ f9bb4bd
... ... 
@@ -34,7 +34,7 @@ void get_pixels_mvi(DCTELEM *restrict block,
34 34 
 void diff_pixels_mvi(DCTELEM *block, const uint8_t *s1, const uint8_t *s2,
35 35 
                      int stride);
36 36 
 int pix_abs8x8_mvi(uint8_t *pix1, uint8_t *pix2, int line_size);
37 
-int pix_abs16x16_mvi(uint8_t *pix1, uint8_t *pix2, int line_size);
37 
+int pix_abs16x16_mvi_asm(uint8_t *pix1, uint8_t *pix2, int line_size);
38 38 
 int pix_abs16x16_x2_mvi(uint8_t *pix1, uint8_t *pix2, int line_size);
39 39 
 int pix_abs16x16_y2_mvi(uint8_t *pix1, uint8_t *pix2, int line_size);
40 40 
 int pix_abs16x16_xy2_mvi(uint8_t *pix1, uint8_t *pix2, int line_size);
... ... 
@@ -335,7 +335,7 @@ void dsputil_init_alpha(void)
335 335 
         get_pixels       = get_pixels_mvi;
336 336 
         diff_pixels      = diff_pixels_mvi;
337 337 
         pix_abs8x8       = pix_abs8x8_mvi;
338 
-        pix_abs16x16     = pix_abs16x16_mvi;
338 
+        pix_abs16x16     = pix_abs16x16_mvi_asm;
339 339 
         pix_abs16x16_x2  = pix_abs16x16_x2_mvi;
340 340 
         pix_abs16x16_y2  = pix_abs16x16_y2_mvi;
341 341 
         pix_abs16x16_xy2 = pix_abs16x16_xy2_mvi;
libavcodec/alpha/motion_est_alpha.c
History View file @ f9bb4bd
... ... 
@@ -117,6 +117,7 @@ int pix_abs8x8_mvi(uint8_t *pix1, uint8_t *pix2, int line_size)
117 117 
     return result;
118 118 
 }
119 119
120 
+#if 0				/* now done in assembly */
120 121 
 int pix_abs16x16_mvi(uint8_t *pix1, uint8_t *pix2, int line_size)
121 122 
 {
122 123 
     int result = 0;
... ... 
@@ -157,6 +158,7 @@ int pix_abs16x16_mvi(uint8_t *pix1, uint8_t *pix2, int line_size)
157 157
158 158 
     return result;
159 159 
 }
160 
+#endif
160 161
161 162 
 int pix_abs16x16_x2_mvi(uint8_t *pix1, uint8_t *pix2, int line_size)
162 163 
 {
libavcodec/alpha/motion_est_mvi_asm.S
History View file @ f9bb4bd
163 164 
new file mode 100644
... ... 
@@ -0,0 +1,186 @@
0 
+/*
1 
+ * Alpha optimized DSP utils
2 
+ * Copyright (c) 2002 Falk Hueffner <falk@debian.org>
3 
+ *
4 
+ * This program is free software; you can redistribute it and/or modify
5 
+ * it under the terms of the GNU General Public License as published by
6 
+ * the Free Software Foundation; either version 2 of the License, or
7 
+ * (at your option) any later version.
8 
+ *
9 
+ * This program is distributed in the hope that it will be useful,
10 
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12 
+ * GNU General Public License for more details.
13 
+ *
14 
+ * You should have received a copy of the GNU General Public License
15 
+ * along with this program; if not, write to the Free Software
16 
+ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
17 
+ */
18 
+
19 
+#include "regdef.h"
20 
+#ifdef HAVE_AV_CONFIG_H
21 
+#include "config.h"
22 
+#endif
23 
+
24 
+/* Some nicer register names.  */
25 
+#define ta t10
26 
+#define tb t11
27 
+#define tc t12
28 
+#define td AT
29 
+/* Danger: these overlap with the argument list and the return value */
30 
+#define te a5
31 
+#define tf a4
32 
+#define tg a3
33 
+#define th v0
34 
+
35 
+        .set noat
36 
+        .set noreorder
37 
+        .arch pca56
38 
+        .text
39 
+
40 
+/*****************************************************************************
41 
+ * int pix_abs16x16_mvi_asm(uint8_t *pix1, uint8_t *pix2, int line_size)
42 
+ *
43 
+ * This code is written with a pca56 in mind. For ev6, one should
44 
+ * really take the increased latency of 3 cycles for MVI instructions
45 
+ * into account.
46 
+ *
47 
+ * It is important to keep the loading and first use of a register as
48 
+ * far apart as possible, because if a register is accessed before it
49 
+ * has been fetched from memory, the CPU will stall.
50 
+ */
51 
+        .align 4
52 
+        .globl pix_abs16x16_mvi_asm
53 
+        .ent pix_abs16x16_mvi_asm
54 
+pix_abs16x16_mvi_asm:
55 
+        .frame sp, 0, ra, 0
56 
+        .prologue 0
57 
+
58 
+#ifdef HAVE_GPROF
59 
+        lda     AT, _mcount
60 
+        jsr     AT, (AT), _mcount
61 
+#endif
62 
+
63 
+        and     a1, 7, t0
64 
+        clr     v0
65 
+        lda     a3, 16
66 
+        beq     t0, $aligned
67 
+        .align 4
68 
+$unaligned:
69 
+        /* Registers:
70 
+           line 0:
71 
+           t0:  left_u -> left lo -> left
72 
+           t1:  mid
73 
+           t2:  right_u -> right hi -> right
74 
+           t3:  ref left
75 
+           t4:  ref right
76 
+           line 1:
77 
+           t5:  left_u -> left lo -> left
78 
+           t6:  mid
79 
+           t7:  right_u -> right hi -> right
80 
+           t8:  ref left
81 
+           t9:  ref right
82 
+           temp:
83 
+           ta:  left hi
84 
+           tb:  right lo
85 
+           tc:  error left
86 
+           td:  error right  */
87 
+
88 
+        /* load line 0 */
89 
+        ldq_u   t0, 0(a1)       # left_u
90 
+        ldq_u   t1, 8(a1)       # mid
91 
+        ldq_u   t2, 16(a1)      # right_u
92 
+        ldq     t3, 0(a0)       # ref left
93 
+        ldq     t4, 8(a0)       # ref right
94 
+        addq    a0, a2, a0      # pix1
95 
+        addq    a1, a2, a1      # pix2
96 
+        /* load line 1 */
97 
+        ldq_u   t5, 0(a1)       # left_u
98 
+        ldq_u   t6, 8(a1)       # mid
99 
+        ldq_u   t7, 16(a1)      # right_u
100 
+        ldq     t8, 0(a0)       # ref left
101 
+        ldq     t9, 8(a0)       # ref right
102 
+        addq    a0, a2, a0      # pix1
103 
+        addq    a1, a2, a1      # pix2
104 
+        /* calc line 0 */
105 
+        extql   t0, a1, t0      # left lo
106 
+        extqh   t1, a1, ta      # left hi
107 
+        extql   t1, a1, tb      # right lo
108 
+        or      t0, ta, t0      # left
109 
+        extqh   t2, a1, t2      # right hi
110 
+        perr    t3, t0, tc      # error left
111 
+        or      t2, tb, t2      # right
112 
+        perr    t4, t2, td      # error right
113 
+        addq    v0, tc, v0      # add error left
114 
+        addq    v0, td, v0      # add error left
115 
+        /* calc line 1 */
116 
+        extql   t5, a1, t5      # left lo
117 
+        extqh   t6, a1, ta      # left hi
118 
+        extql   t6, a1, tb      # right lo
119 
+        or      t5, ta, t5      # left
120 
+        extqh   t7, a1, t7      # right hi
121 
+        perr    t8, t5, tc      # error left
122 
+        or      t7, tb, t7      # right
123 
+        perr    t9, t7, td      # error right
124 
+        addq    v0, tc, v0      # add error left
125 
+        addq    v0, td, v0      # add error left
126 
+        /* loop */
127 
+        subq    a3,  2, a3      # h -= 2
128 
+        bne     a3, $unaligned
129 
+        ret
130 
+
131 
+        .align 4
132 
+$aligned:
133 
+        /* load line 0 */
134 
+        ldq     t0, 0(a1)       # left
135 
+        ldq     t1, 8(a1)       # right
136 
+        addq    a1, a2, a1      # pix2
137 
+        ldq     t2, 0(a0)       # ref left
138 
+        ldq     t3, 8(a0)       # ref right
139 
+        addq    a0, a2, a0      # pix1
140 
+        /* load line 1 */
141 
+        ldq     t4, 0(a1)       # left
142 
+        ldq     t5, 8(a1)       # right
143 
+        addq    a1, a2, a1      # pix2
144 
+        ldq     t6, 0(a0)       # ref left
145 
+        ldq     t7, 8(a0)       # ref right
146 
+        addq    a0, a2, a0      # pix1
147 
+        /* load line 2 */
148 
+        ldq     t8, 0(a1)       # left
149 
+        ldq     t9, 8(a1)       # right
150 
+        addq    a1, a2, a1      # pix2
151 
+        ldq     ta, 0(a0)       # ref left
152 
+        ldq     tb, 8(a0)       # ref right
153 
+        addq    a0, a2, a0      # pix1
154 
+        /* load line 3 */
155 
+        ldq     tc, 0(a1)       # left
156 
+        ldq     td, 8(a1)       # right
157 
+        addq    a1, a2, a1      # pix2
158 
+        ldq     te, 0(a0)       # ref left
159 
+        ldq     tf, 8(a0)       # ref right
160 
+        /* calc line 0 */
161 
+        perr    t0, t2, t0      # error left
162 
+        addq    a0, a2, a0      # pix1
163 
+        perr    t1, t3, t1      # error right
164 
+        addq    v0, t0, v0      # add error left
165 
+        /* calc line 1 */
166 
+        perr    t4, t6, t0      # error left
167 
+        addq    v0, t1, v0      # add error right
168 
+        perr    t5, t7, t1      # error right
169 
+        addq    v0, t0, v0      # add error left
170 
+        /* calc line 2 */
171 
+        perr    t8, ta, t0      # error left
172 
+        addq    v0, t1, v0      # add error right
173 
+        perr    t9, tb, t1      # error right
174 
+        addq    v0, t0, v0      # add error left
175 
+        /* calc line 3 */
176 
+        perr    tc, te, t0      # error left
177 
+        addq    v0, t1, v0      # add error right
178 
+        perr    td, tf, t1      # error right
179 
+        addq    v0, t0, v0      # add error left
180 
+        addq    v0, t1, v0      # add error right
181 
+        /* loop */
182 
+        subq    a3,  4, a3      # h -= 4
183 
+        bne     a3, $aligned
184 
+        ret
185 
+        .end pix_abs16x16_mvi_asm