diff options
Diffstat (limited to 'apps')
-rw-r--r-- | apps/recorder/jpeg_idct_arm.S | 230 |
1 files changed, 215 insertions, 15 deletions
diff --git a/apps/recorder/jpeg_idct_arm.S b/apps/recorder/jpeg_idct_arm.S index 46ac479caa..75e3e4d2b3 100644 --- a/apps/recorder/jpeg_idct_arm.S +++ b/apps/recorder/jpeg_idct_arm.S | |||
@@ -429,6 +429,7 @@ jpeg_idct4h: | |||
429 | #endif | 429 | #endif |
430 | .size jpeg_idct4h, .-jpeg_idct4h | 430 | .size jpeg_idct4h, .-jpeg_idct4h |
431 | 431 | ||
432 | #if ARM_ARCH < 6 | ||
432 | jpeg_idct8v: | 433 | jpeg_idct8v: |
433 | stmdb sp!, { r4-r11, lr } | 434 | stmdb sp!, { r4-r11, lr } |
434 | add r2, r0, #128 | 435 | add r2, r0, #128 |
@@ -641,13 +642,9 @@ jpeg_idct8h: | |||
641 | orrs r9, r6, r7 | 642 | orrs r9, r6, r7 |
642 | orreqs r9, r5, r4, lsr #16 | 643 | orreqs r9, r5, r4, lsr #16 |
643 | bne 2f | 644 | bne 2f |
644 | #if ARM_ARCH < 6 | ||
645 | mov r8, r8, asr #21 | 645 | mov r8, r8, asr #21 |
646 | cmp r8, #255 | 646 | cmp r8, #255 |
647 | mvnhi r8, r8, asr #31 | 647 | mvnhi r8, r8, asr #31 |
648 | #else | ||
649 | usat r8, #8, r8, asr #21 | ||
650 | #endif | ||
651 | #ifdef HAVE_LCD_COLOR | 648 | #ifdef HAVE_LCD_COLOR |
652 | strb r8, [r1] | 649 | strb r8, [r1] |
653 | strb r8, [r1, #4] | 650 | strb r8, [r1, #4] |
@@ -848,7 +845,6 @@ jpeg_idct8h: | |||
848 | add r10, r11, r4 /* o0 */ | 845 | add r10, r11, r4 /* o0 */ |
849 | sub r11, r11, r4 /* o7 */ | 846 | sub r11, r11, r4 /* o7 */ |
850 | /* output in order: r10 r12 r8 r6 r7 r9 r14 r11 */ | 847 | /* output in order: r10 r12 r8 r6 r7 r9 r14 r11 */ |
851 | #if ARM_ARCH < 6 | ||
852 | mov r10, r10, asr #18 | 848 | mov r10, r10, asr #18 |
853 | cmp r10, #255 | 849 | cmp r10, #255 |
854 | mvnhi r10, r10, asr #31 | 850 | mvnhi r10, r10, asr #31 |
@@ -873,16 +869,6 @@ jpeg_idct8h: | |||
873 | mov r11, r11, asr #18 | 869 | mov r11, r11, asr #18 |
874 | cmp r11, #255 | 870 | cmp r11, #255 |
875 | mvnhi r11, r11, asr #31 | 871 | mvnhi r11, r11, asr #31 |
876 | #else | ||
877 | usat r10, #8, r10, asr #18 | ||
878 | usat r12, #8, r12, asr #18 | ||
879 | usat r8, #8, r8, asr #18 | ||
880 | usat r6, #8, r6, asr #18 | ||
881 | usat r7, #8, r7, asr #18 | ||
882 | usat r9, #8, r9, asr #18 | ||
883 | usat r14, #8, r14, asr #18 | ||
884 | usat r11, #8, r11, asr #18 | ||
885 | #endif | ||
886 | #ifdef HAVE_LCD_COLOR | 872 | #ifdef HAVE_LCD_COLOR |
887 | strb r10, [r1] | 873 | strb r10, [r1] |
888 | strb r12, [r1, #4] | 874 | strb r12, [r1, #4] |
@@ -908,3 +894,217 @@ jpeg_idct8h: | |||
908 | bcc 1b | 894 | bcc 1b |
909 | ldmia sp!, { r4-r11, pc } | 895 | ldmia sp!, { r4-r11, pc } |
910 | .size jpeg_idct8h, .-jpeg_idct8h | 896 | .size jpeg_idct8h, .-jpeg_idct8h |
897 | #else | ||
898 | jpeg_idct8v: | ||
899 | stmdb sp!, { r4-r11, lr } | ||
900 | add r2, r0, #128 | ||
901 | 1: | ||
902 | ldmia r0!, { r4-r7 } | ||
903 | orrs r9, r6, r7 | ||
904 | orreqs r9, r5, r4, lsr #16 | ||
905 | bne 2f | ||
906 | mov r4, r4, lsl #2 | ||
907 | strh r4, [r2] | ||
908 | strh r4, [r2, #16] | ||
909 | strh r4, [r2, #32] | ||
910 | strh r4, [r2, #48] | ||
911 | strh r4, [r2, #64] | ||
912 | strh r4, [r2, #80] | ||
913 | strh r4, [r2, #96] | ||
914 | strh r4, [r2, #112] | ||
915 | cmp r0, r1 | ||
916 | add r2, r2, #2 | ||
917 | bcc 1b | ||
918 | ldmia sp!, { r4-r11, pc } | ||
919 | 2: | ||
920 | ldrd r8, .Lpool8 | ||
921 | mov r12, r4, lsl #16 | ||
922 | add r10, r5, r7 /* r10 = d2 + d6 */ | ||
923 | add r12, r12, #8192 | ||
924 | add r3, r12, r6, lsl #16 /* tmp0 */ | ||
925 | sub r12, r12, r6, lsl #16 /* tmp1 */ | ||
926 | pkhtb r4, r5, r4, asr #16 /* r4 = (tmp3[o], tmp2[0]) = (d1, d3) */ | ||
927 | smulbb r14, r8, r10 /* r14 = z1[e] = (d2 + d6) * 4433 */ | ||
928 | pkhtb r6, r6, r7, asr #16 /* r6 = (tmp0[o], tmp1[o]) = (d7, d5) */ | ||
929 | smlatb r7, r8, r7, r14 /* r7 = tmp2[e] = z1 - d6 * 15137 */ | ||
930 | smlabb r5, r9, r5, r14 /* r5 = tmp3[e] = z1 + d2 * 6270 */ | ||
931 | pkhtb r9, r9, r9, asr #16 /* r9 = (9633, 9633) */ | ||
932 | add r10, r5, r3, asr #3 /* r10 = tmp10 */ | ||
933 | rsb r11, r5, r3, asr #3 /* r11 = tmp13 */ | ||
934 | mov r3, r4, ror #16 | ||
935 | rsb r14, r7, r12, asr #3 /* r14 = tmp12 */ | ||
936 | add r12, r7, r12, asr #3 /* r12 = tmp11 */ | ||
937 | sadd16 r8, r3, r6 /* z3, z4 */ | ||
938 | stmdb sp, { r10-r12, r14 } /* tmp10 tmp13 tmp11 tmp12 */ | ||
939 | smuad r5, r9, r8 /* r5 = z5 = (z3[o] + z4[o]) * 9633 */ | ||
940 | ldrd r10, .Lpool8+8 | ||
941 | sadd16 r7, r4, r6 /* r7 = (z1, z2) */ | ||
942 | smlatt r9, r10, r8, r5 /* r9 = z4 = z5 - z4 * 16069 */ | ||
943 | smlabb r8, r10, r8, r5 /* r8 = z3 = z5 - z3 * 3196 */ | ||
944 | smlabb r14, r11, r7, r9 /* r14 = z1 + z4 */ | ||
945 | smlabb r12, r11, r7, r8 /* r12 = z1 + z3 */ | ||
946 | smlatt r5, r11, r7, r9 /* r5 = z2 + z4 */ | ||
947 | smlatt r7, r11, r7, r8 /* r7 = z2 + z3 */ | ||
948 | ldrd r8, .Lpool8+16 | ||
949 | smlabt r7, r9, r4, r7 /* r7 = tmp2 */ | ||
950 | smlatb r14, r9, r4, r14 /* r14 = tmp3 */ | ||
951 | ldmdb sp, { r4, r9-r11 } /* tmp10 tmp13 tmp11 tmp12 */ | ||
952 | smlabb r12, r8, r6, r12 /* r12 = tmp0 */ | ||
953 | smlatt r5, r8, r6, r5 /* r5 = tmp1 */ | ||
954 | /* used: r4, r5, r7, r9-r12, r14 */ | ||
955 | add r6, r4, r14 /* o0 */ | ||
956 | sub r8, r4, r14 /* o7 */ | ||
957 | add r14, r9, r12 /* o3 */ | ||
958 | sub r12, r9, r12 /* o4 */ | ||
959 | add r4, r10, r7 /* o1 */ | ||
960 | sub r7, r10, r7 /* o6 */ | ||
961 | add r9, r11, r5 /* o2 */ | ||
962 | sub r10, r11, r5 /* o5 */ | ||
963 | mov r6, r6, asr #11 | ||
964 | mov r4, r4, asr #11 | ||
965 | mov r9, r9, asr #11 | ||
966 | mov r14, r14, asr #11 | ||
967 | mov r12, r12, asr #11 | ||
968 | mov r10, r10, asr #11 | ||
969 | mov r7, r7, asr #11 | ||
970 | mov r8, r8, asr #11 | ||
971 | strh r6, [r2] | ||
972 | strh r4, [r2, #16] | ||
973 | strh r9, [r2, #32] | ||
974 | strh r14, [r2, #48] | ||
975 | strh r12, [r2, #64] | ||
976 | strh r10, [r2, #80] | ||
977 | strh r7, [r2, #96] | ||
978 | strh r8, [r2, #112] | ||
979 | cmp r0, r1 | ||
980 | add r2, r2, #2 | ||
981 | bcc 1b | ||
982 | ldmia sp!, { r4-r11, pc } | ||
983 | .size jpeg_idct8v, .-jpeg_idct8v | ||
984 | |||
985 | .align 4 | ||
986 | .Lpool8: | ||
987 | .short 4433 | ||
988 | .short -15137 | ||
989 | .short 6270 | ||
990 | .short 9633 | ||
991 | .short -16069 | ||
992 | .short -3196 | ||
993 | .short -7373 | ||
994 | .short -20995 | ||
995 | .short 2446 | ||
996 | .short 16819 | ||
997 | .short 25172 | ||
998 | .short 12299 | ||
999 | |||
1000 | .align 2 | ||
1001 | jpeg_idct8h: | ||
1002 | stmdb sp!, { r4-r11, lr } | ||
1003 | 1: | ||
1004 | ldr r14, =4112 | ||
1005 | ldmia r0!, { r4-r7 } | ||
1006 | sadd16 r4, r4, r14 | ||
1007 | orrs r9, r6, r7 | ||
1008 | orreqs r9, r5, r4, lsr #16 | ||
1009 | bne 2f | ||
1010 | sxth r4, r4 | ||
1011 | usat r4, #8, r4, asr #5 | ||
1012 | #ifdef HAVE_LCD_COLOR | ||
1013 | strb r4, [r1] | ||
1014 | strb r4, [r1, #4] | ||
1015 | strb r4, [r1, #8] | ||
1016 | strb r4, [r1, #12] | ||
1017 | strb r4, [r1, #16] | ||
1018 | strb r4, [r1, #20] | ||
1019 | strb r4, [r1, #24] | ||
1020 | strb r4, [r1, #28] | ||
1021 | #else | ||
1022 | strb r4, [r1] | ||
1023 | strb r4, [r1, #1] | ||
1024 | strb r4, [r1, #2] | ||
1025 | strb r4, [r1, #3] | ||
1026 | strb r4, [r1, #4] | ||
1027 | strb r4, [r1, #5] | ||
1028 | strb r4, [r1, #6] | ||
1029 | strb r4, [r1, #7] | ||
1030 | #endif | ||
1031 | add r1, r1, r3 | ||
1032 | cmp r0, r2 | ||
1033 | bcc 1b | ||
1034 | ldmia sp!, { r4-r11, pc } | ||
1035 | 2: | ||
1036 | ldrd r8, .Lpool8 | ||
1037 | sadd16 r10, r5, r7 /* r10 = (d2 + d6, d3 + d7) */ | ||
1038 | ssub16 r12, r4, r6 /* r12 = (d0 - d4, d1 - d5) */ | ||
1039 | sadd16 r11, r4, r6 /* r11 = (d0 + d4, d1 + d5) */ | ||
1040 | pkhtb r4, r5, r4, asr #16 /* r4 = (tmp3[o], tmp2[0]) = (d1, d3) */ | ||
1041 | smulbb r14, r8, r10 /* r14 = z1[e] = (d2 + d6) * 4433 */ | ||
1042 | pkhtb r6, r6, r7, asr #16 /* r6 = (tmp0[o], tmp1[o]) = (d7, d5) */ | ||
1043 | smlatb r7, r8, r7, r14 /* r7 = tmp2[e] = z1 - d6 * 15137 */ | ||
1044 | smlabb r5, r9, r5, r14 /* r5 = tmp3[e] = z1 + d2 * 6270 */ | ||
1045 | sxth r12, r12 /* r12 = tmp1[e] = d0 - d4 */ | ||
1046 | pkhtb r8, r11, r10, asr #16 /* r8 = (z3[o], z4[o]) */ | ||
1047 | sxth r14, r11 /* r14 = tmp0[e] */ | ||
1048 | pkhtb r9, r9, r9, asr #16 /* r9 = (9633, 9633) */ | ||
1049 | add r10, r5, r14, lsl #13 /* r10 = tmp10 */ | ||
1050 | rsb r11, r5, r14, lsl #13 /* r11 = tmp13 */ | ||
1051 | rsb r14, r7, r12, lsl #13 /* r14 = tmp12 */ | ||
1052 | add r12, r7, r12, lsl #13 /* r12 = tmp11 */ | ||
1053 | stmdb sp, { r10-r12, r14 } /* tmp10 tmp13 tmp11 tmp12 */ | ||
1054 | smuad r5, r9, r8 /* r5 = z5 = (z3[o] + z4[o]) * 9633 */ | ||
1055 | ldrd r10, .Lpool8+8 | ||
1056 | sadd16 r7, r4, r6 /* r7 = (z1, z2) */ | ||
1057 | smlatt r9, r10, r8, r5 /* r9 = z4 = z5 - z4 * 16069 */ | ||
1058 | smlabb r8, r10, r8, r5 /* r8 = z3 = z5 - z3 * 3196 */ | ||
1059 | smlabb r14, r11, r7, r9 /* r14 = z1 + z4 */ | ||
1060 | smlabb r12, r11, r7, r8 /* r12 = z1 + z3 */ | ||
1061 | smlatt r5, r11, r7, r9 /* r5 = z2 + z4 */ | ||
1062 | smlatt r7, r11, r7, r8 /* r7 = z2 + z3 */ | ||
1063 | ldrd r8, .Lpool8+16 | ||
1064 | smlabt r7, r9, r4, r7 /* r7 = tmp2 */ | ||
1065 | smlatb r14, r9, r4, r14 /* r14 = tmp3 */ | ||
1066 | ldmdb sp, { r4, r9-r11 } /* tmp10 tmp13 tmp11 tmp12 */ | ||
1067 | smlabb r12, r8, r6, r12 /* r12 = tmp0 */ | ||
1068 | smlatt r5, r8, r6, r5 /* r5 = tmp1 */ | ||
1069 | /* used: r4, r5, r7, r9-r12, r14 */ | ||
1070 | add r6, r4, r14 /* o0 */ | ||
1071 | sub r8, r4, r14 /* o7 */ | ||
1072 | add r14, r9, r12 /* o3 */ | ||
1073 | sub r12, r9, r12 /* o4 */ | ||
1074 | add r4, r10, r7 /* o1 */ | ||
1075 | sub r7, r10, r7 /* o6 */ | ||
1076 | add r9, r11, r5 /* o2 */ | ||
1077 | sub r10, r11, r5 /* o5 */ | ||
1078 | usat r6, #8, r6, asr #18 | ||
1079 | usat r4, #8, r4, asr #18 | ||
1080 | usat r9, #8, r9, asr #18 | ||
1081 | usat r14, #8, r14, asr #18 | ||
1082 | usat r12, #8, r12, asr #18 | ||
1083 | usat r10, #8, r10, asr #18 | ||
1084 | usat r7, #8, r7, asr #18 | ||
1085 | usat r8, #8, r8, asr #18 | ||
1086 | #ifdef HAVE_LCD_COLOR | ||
1087 | strb r6, [r1] | ||
1088 | strb r4, [r1, #4] | ||
1089 | strb r9, [r1, #8] | ||
1090 | strb r14, [r1, #12] | ||
1091 | strb r12, [r1, #16] | ||
1092 | strb r10, [r1, #20] | ||
1093 | strb r7, [r1, #24] | ||
1094 | strb r8, [r1, #28] | ||
1095 | #else | ||
1096 | strb r6, [r1] | ||
1097 | strb r4, [r1, #1] | ||
1098 | strb r9, [r1, #2] | ||
1099 | strb r14, [r1, #3] | ||
1100 | strb r12, [r1, #4] | ||
1101 | strb r10, [r1, #5] | ||
1102 | strb r7, [r1, #6] | ||
1103 | strb r8, [r1, #7] | ||
1104 | #endif | ||
1105 | cmp r0, r2 | ||
1106 | add r1, r1, r3 | ||
1107 | bcc 1b | ||
1108 | ldmia sp!, { r4-r11, pc } | ||
1109 | .size jpeg_idct8h, .-jpeg_idct8h | ||
1110 | #endif | ||