Mercurial > libavcodec.hg

comparison armv4l/dsputil_arm.c @ 8072:4b32e8762864 libavcodec

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
ARM: move add_pixels_clamped_ARM() to dsputil_arm_s.S
author mru
date Mon, 27 Oct 2008 00:25:19 +0000
parents 59be7e4941e8
children 0d108ec85620
comparison
equal deleted inserted replaced
8071:2487a9db02a0 8072:4b32e8762864
64 CALL_2X_PIXELS(put_pixels16_xy2_arm, put_pixels8_xy2_arm, 8) 64 CALL_2X_PIXELS(put_pixels16_xy2_arm, put_pixels8_xy2_arm, 8)
65 CALL_2X_PIXELS(put_no_rnd_pixels16_x2_arm , put_no_rnd_pixels8_x2_arm , 8) 65 CALL_2X_PIXELS(put_no_rnd_pixels16_x2_arm , put_no_rnd_pixels8_x2_arm , 8)
66 CALL_2X_PIXELS(put_no_rnd_pixels16_y2_arm , put_no_rnd_pixels8_y2_arm , 8) 66 CALL_2X_PIXELS(put_no_rnd_pixels16_y2_arm , put_no_rnd_pixels8_y2_arm , 8)
67 CALL_2X_PIXELS(put_no_rnd_pixels16_xy2_arm, put_no_rnd_pixels8_xy2_arm, 8) 67 CALL_2X_PIXELS(put_no_rnd_pixels16_xy2_arm, put_no_rnd_pixels8_xy2_arm, 8)
68 68
69 static void add_pixels_clamped_ARM(short *block, unsigned char *dest, int line_size) 69 extern void ff_add_pixels_clamped_ARM(short *block, unsigned char *dest,
70 { 70 int line_size);
71 __asm__ volatile (
72 "mov r10, #8 \n\t"
73
74 "1: \n\t"
75
76 /* load dest */
77 "ldr r4, [%1] \n\t"
78 /* block[0] and block[1]*/
79 "ldrsh r5, [%0] \n\t"
80 "ldrsh r7, [%0, #2] \n\t"
81 "and r6, r4, #0xFF \n\t"
82 "and r8, r4, #0xFF00 \n\t"
83 "add r6, r5, r6 \n\t"
84 "add r8, r7, r8, lsr #8 \n\t"
85 "mvn r5, r5 \n\t"
86 "mvn r7, r7 \n\t"
87 "tst r6, #0x100 \n\t"
88 "movne r6, r5, lsr #24 \n\t"
89 "tst r8, #0x100 \n\t"
90 "movne r8, r7, lsr #24 \n\t"
91 "mov r9, r6 \n\t"
92 "ldrsh r5, [%0, #4] \n\t" /* moved form [A] */
93 "orr r9, r9, r8, lsl #8 \n\t"
94 /* block[2] and block[3] */
95 /* [A] */
96 "ldrsh r7, [%0, #6] \n\t"
97 "and r6, r4, #0xFF0000 \n\t"
98 "and r8, r4, #0xFF000000 \n\t"
99 "add r6, r5, r6, lsr #16 \n\t"
100 "add r8, r7, r8, lsr #24 \n\t"
101 "mvn r5, r5 \n\t"
102 "mvn r7, r7 \n\t"
103 "tst r6, #0x100 \n\t"
104 "movne r6, r5, lsr #24 \n\t"
105 "tst r8, #0x100 \n\t"
106 "movne r8, r7, lsr #24 \n\t"
107 "orr r9, r9, r6, lsl #16 \n\t"
108 "ldr r4, [%1, #4] \n\t" /* moved form [B] */
109 "orr r9, r9, r8, lsl #24 \n\t"
110 /* store dest */
111 "ldrsh r5, [%0, #8] \n\t" /* moved form [C] */
112 "str r9, [%1] \n\t"
113
114 /* load dest */
115 /* [B] */
116 /* block[4] and block[5] */
117 /* [C] */
118 "ldrsh r7, [%0, #10] \n\t"
119 "and r6, r4, #0xFF \n\t"
120 "and r8, r4, #0xFF00 \n\t"
121 "add r6, r5, r6 \n\t"
122 "add r8, r7, r8, lsr #8 \n\t"
123 "mvn r5, r5 \n\t"
124 "mvn r7, r7 \n\t"
125 "tst r6, #0x100 \n\t"
126 "movne r6, r5, lsr #24 \n\t"
127 "tst r8, #0x100 \n\t"
128 "movne r8, r7, lsr #24 \n\t"
129 "mov r9, r6 \n\t"
130 "ldrsh r5, [%0, #12] \n\t" /* moved from [D] */
131 "orr r9, r9, r8, lsl #8 \n\t"
132 /* block[6] and block[7] */
133 /* [D] */
134 "ldrsh r7, [%0, #14] \n\t"
135 "and r6, r4, #0xFF0000 \n\t"
136 "and r8, r4, #0xFF000000 \n\t"
137 "add r6, r5, r6, lsr #16 \n\t"
138 "add r8, r7, r8, lsr #24 \n\t"
139 "mvn r5, r5 \n\t"
140 "mvn r7, r7 \n\t"
141 "tst r6, #0x100 \n\t"
142 "movne r6, r5, lsr #24 \n\t"
143 "tst r8, #0x100 \n\t"
144 "movne r8, r7, lsr #24 \n\t"
145 "orr r9, r9, r6, lsl #16 \n\t"
146 "add %0, %0, #16 \n\t" /* moved from [E] */
147 "orr r9, r9, r8, lsl #24 \n\t"
148 "subs r10, r10, #1 \n\t" /* moved from [F] */
149 /* store dest */
150 "str r9, [%1, #4] \n\t"
151
152 /* [E] */
153 /* [F] */
154 "add %1, %1, %2 \n\t"
155 "bne 1b \n\t"
156 : "+r"(block),
157 "+r"(dest)
158 : "r"(line_size)
159 : "r4", "r5", "r6", "r7", "r8", "r9", "r10", "cc", "memory" );
160 }
161 71
162 /* XXX: those functions should be suppressed ASAP when all IDCTs are 72 /* XXX: those functions should be suppressed ASAP when all IDCTs are
163 converted */ 73 converted */
164 static void j_rev_dct_ARM_put(uint8_t *dest, int line_size, DCTELEM *block) 74 static void j_rev_dct_ARM_put(uint8_t *dest, int line_size, DCTELEM *block)
165 { 75 {
198 { 108 {
199 ippiDCT8x8Inv_Video_16s_C1I(block); 109 ippiDCT8x8Inv_Video_16s_C1I(block);
200 #ifdef HAVE_IWMMXT 110 #ifdef HAVE_IWMMXT
201 add_pixels_clamped_iwmmxt(block, dest, line_size); 111 add_pixels_clamped_iwmmxt(block, dest, line_size);
202 #else 112 #else
203 add_pixels_clamped_ARM(block, dest, line_size); 113 ff_add_pixels_clamped_ARM(block, dest, line_size);
204 #endif 114 #endif
205 } 115 }
206 #endif 116 #endif
207 117
208 int mm_support(void) 118 int mm_support(void)