diff options
author | Michael Sevakis <jethead71@rockbox.org> | 2007-02-19 02:49:26 +0000 |
---|---|---|
committer | Michael Sevakis <jethead71@rockbox.org> | 2007-02-19 02:49:26 +0000 |
commit | 36175ac9453999d2d079c521126ecc5ac7a8d984 (patch) | |
tree | a37e87b5fd7283d1456b7a346e16c1a5ed590a2c /apps/dsp_cf.S | |
parent | 2801a87d543f38cadd076330f329c84e23852997 (diff) | |
download | rockbox-36175ac9453999d2d079c521126ecc5ac7a8d984.tar.gz rockbox-36175ac9453999d2d079c521126ecc5ac7a8d984.zip |
SWCODEC: DSP optimizations for conversion to internal format and resampling. Assembly resampling for Coldfire. Word has it ARM will get that soon.
git-svn-id: svn://svn.rockbox.org/rockbox/trunk@12399 a1c6a512-1295-4272-9138-f99709370657
Diffstat (limited to 'apps/dsp_cf.S')
-rw-r--r-- | apps/dsp_cf.S | 145 |
1 files changed, 141 insertions, 4 deletions
diff --git a/apps/dsp_cf.S b/apps/dsp_cf.S index 719d1db1d5..233be82860 100644 --- a/apps/dsp_cf.S +++ b/apps/dsp_cf.S | |||
@@ -17,8 +17,11 @@ | |||
17 | * | 17 | * |
18 | ****************************************************************************/ | 18 | ****************************************************************************/ |
19 | 19 | ||
20 | .section .text | 20 | /**************************************************************************** |
21 | .global apply_crossfeed | 21 | * apply_crossfeed(int32_t* src[], int count) |
22 | */ | ||
23 | .section .text | ||
24 | .global apply_crossfeed | ||
22 | apply_crossfeed: | 25 | apply_crossfeed: |
23 | lea.l (-44, %sp), %sp | 26 | lea.l (-44, %sp), %sp |
24 | movem.l %d2-%d7/%a2-%a6, (%sp) | save all regs | 27 | movem.l %d2-%d7/%a2-%a6, (%sp) | save all regs |
@@ -67,11 +70,11 @@ apply_crossfeed: | |||
67 | addq.l #1, %d4 | index++ | 70 | addq.l #1, %d4 | index++ |
68 | moveq.l #13, %d6 | 71 | moveq.l #13, %d6 |
69 | cmp.l %d6, %d4 | wrap index to 0 if it overflows | 72 | cmp.l %d6, %d4 | wrap index to 0 if it overflows |
70 | jlt .nowrap | 73 | jlt .cfnowrap |
71 | moveq.l #13*8, %d4 | 74 | moveq.l #13*8, %d4 |
72 | sub.l %d4, %a0 | wrap back delay line ptr as well | 75 | sub.l %d4, %a0 | wrap back delay line ptr as well |
73 | clr.l %d4 | 76 | clr.l %d4 |
74 | .nowrap: | 77 | .cfnowrap: |
75 | subq.l #1, %d7 | 78 | subq.l #1, %d7 |
76 | jne .cfloop | 79 | jne .cfloop |
77 | | save data back to struct | 80 | | save data back to struct |
@@ -81,4 +84,138 @@ apply_crossfeed: | |||
81 | movem.l (%sp), %d2-%d7/%a2-%a6 | 84 | movem.l (%sp), %d2-%d7/%a2-%a6 |
82 | lea.l (44, %sp), %sp | 85 | lea.l (44, %sp), %sp |
83 | rts | 86 | rts |
87 | .cfend: | ||
88 | .size apply_crossfeed,.cfend-apply_crossfeed | ||
84 | 89 | ||
90 | /**************************************************************************** | ||
91 | * dsp_downsample(int channels, int count, struct resample_data *r, | ||
92 | * in32_t **src, int32_t **dst) | ||
93 | */ | ||
94 | .section .text | ||
95 | .global dsp_downsample | ||
96 | dsp_downsample: | ||
97 | lea.l -40(%sp), %sp | save non-clobberables | ||
98 | movem.l %d2-%d7/%a2-%a5, (%sp) | | ||
99 | movem.l 44(%sp), %d2-%d3/%a0-%a2| %d2 = ch = channels | ||
100 | | %d3 = count | ||
101 | | %a0 = r | ||
102 | | %a1 = src | ||
103 | | %a2 = dst | ||
104 | move.l 4(%a0), %d4 | %d4 = delta = r->delta | ||
105 | move.l #16, %d7 | %d7 = shift | ||
106 | .dschannel_loop: | ||
107 | move.l (%a0), %d5 | %d5 = phase = r->phase | ||
108 | move.l -4(%a1, %d2.l*4), %a3 | %a3 = s = src[ch-1] | ||
109 | move.l -4(%a2, %d2.l*4), %a4 | %a4 = d = dst[ch-1] | ||
110 | lea.l 4(%a0, %d2.l*4), %a5 | %a5 = &r->last_sample[ch-1] | ||
111 | move.l (%a5), %d0 | %d0 = last = r->last_sample[ch-1] | ||
112 | move.l -4(%a3, %d3.l*4), %d1 | r->last_sample[ch-1] = s[count-1] | ||
113 | move.l %d1, (%a5) | | ||
114 | move.l %d5, %d6 | %d6 = pos = phase >> 16 | ||
115 | lsr.l %d7, %d6 | | ||
116 | cmp.l %d3, %d6 | past end of samples? | ||
117 | bge.b .dsloop_skip | yes? skip loop | ||
118 | tst.l %d6 | need last sample of prev. frame? | ||
119 | bne.b .dsloop | no? start main loop | ||
120 | move.l (%a3, %d6.l*4), %d1 | %d1 = s[pos] | ||
121 | bra.b .dsuse_last_start | start with last (last in %d0) | ||
122 | .dsloop: | ||
123 | lea.l -4(%a3, %d6.l*4), %a5 | load s[pos-1] and s[pos] | ||
124 | movem.l (%a5), %d0-%d1 | | ||
125 | .dsuse_last_start: | ||
126 | sub.l %d0, %d1 | %d1 = diff = s[pos] - s[pos-1] | ||
127 | move.l %d0, %acc0 | %acc0 = previous sample | ||
128 | move.l %d5, %d0 | frac = (phase << 16) >> 1 | ||
129 | lsl.l %d7, %d0 | | ||
130 | lsr.l #1, %d0 | | ||
131 | mac.l %d0, %d1, %acc0 | %acc0 += frac * diff | ||
132 | move.l %acc0, %d0 | | ||
133 | add.l %d4, %d5 | phase += delta | ||
134 | move.l %d5, %d6 | pos = phase >> 16 | ||
135 | lsr.l %d7, %d6 | | ||
136 | move.l %d0, (%a4)+ | *d++ = %d0 | ||
137 | cmp.l %d3, %d6 | pos < count? | ||
138 | blt.b .dsloop | yes? continue resampling | ||
139 | .dsloop_skip: | ||
140 | subq.l #1, %d2 | ch > 0? | ||
141 | bgt.b .dschannel_loop | yes? process next channel | ||
142 | asl.l %d7, %d3 | wrap phase to start of next frame | ||
143 | sub.l %d3, %d5 | r->phase = phase - (count << 16) | ||
144 | move.l %d5, (%a0) | | ||
145 | move.l %a4, %d0 | return d - d[0] | ||
146 | sub.l (%a2), %d0 | | ||
147 | asr.l #2, %d0 | convert bytes->samples | ||
148 | movem.l (%sp), %d2-%d7/%a2-%a5 | restore non-clobberables | ||
149 | move.l %acc1, %acc0 | clear %acc0 | ||
150 | lea.l 40(%sp), %sp | cleanup stack | ||
151 | rts | buh-bye | ||
152 | .dsend: | ||
153 | .size dsp_downsample,.dsend-dsp_downsample | ||
154 | |||
155 | /**************************************************************************** | ||
156 | * dsp_upsample(int channels, int count, struct resample_data *r, | ||
157 | * in32_t **src, int32_t **dst) | ||
158 | */ | ||
159 | .section .text | ||
160 | .global dsp_upsample | ||
161 | dsp_upsample: | ||
162 | lea.l -40(%sp), %sp | save non-clobberables | ||
163 | movem.l %d2-%d7/%a2-%a5, (%sp) | | ||
164 | movem.l 44(%sp), %d2-%d3/%a0-%a2| %d2 = ch = channels | ||
165 | | %d3 = count | ||
166 | | %a0 = r | ||
167 | | %a1 = src | ||
168 | | %a2 = dst | ||
169 | move.l 4(%a0), %d4 | %d4 = delta = r->delta | ||
170 | swap %d4 | swap delta to high word to use | ||
171 | | carries to increment position | ||
172 | .uschannel_loop: | ||
173 | move.l (%a0), %d5 | %d5 = phase = r->phase | ||
174 | move.l -4(%a1, %d2.l*4), %a3 | %a3 = s = src[ch-1] | ||
175 | move.l -4(%a2, %d2.l*4), %a4 | %a4 = d = dst[ch-1] | ||
176 | lea.l 4(%a0, %d2.l*4), %a5 | %a5 = &r->last_sample[ch-1] | ||
177 | move.l (%a5), %d0 | %d0 = last = r->last_sample[ch-1] | ||
178 | move.l -4(%a3, %d3.l*4), %d1 | r->last_sample[ch-1] = s[count-1] | ||
179 | move.l %d1, (%a5) | | ||
180 | moveq.l #16, %d1 | %d0 = shift | ||
181 | move.l %d5, %d6 | %d6 = pos = phase >> 16 | ||
182 | lsl.l %d1, %d5 | swap phase to high word to use | ||
183 | | carries to increment position | ||
184 | lsr.l %d1, %d6 | pos == 0? | ||
185 | bne.b .usstart_1 | no? transistion from down | ||
186 | move.l (%a3), %d1 | %d1 = s[0] | ||
187 | sub.l %d0, %d1 | diff = s[pos] - last | ||
188 | bra.b .usloop_0 | jump to typical start point | ||
189 | .usstart_1: | ||
190 | cmp.l %d3, %d6 | past end of samples? | ||
191 | bge.b .usloop_skip | yes? skip loop | ||
192 | .usloop_1: | ||
193 | lea.l -4(%a3, %d6.l*4), %a5 | load s[pos-1] and s[pos] | ||
194 | movem.l (%a5), %d0-%d1 | | ||
195 | sub.l %d0, %d1 | %d1 = diff = s[pos] - s[pos-1] | ||
196 | .usloop_0: | ||
197 | move.l %d0, %acc0 | %acc0 = previous sample | ||
198 | lsr.l #1, %d5 | make phase into frac | ||
199 | mac.l %d1, %d5, %acc0 | %acc0 += diff * frac | ||
200 | move.l %acc0, %d7 | | ||
201 | lsl.l #1, %d5 | restore frac to phase | ||
202 | move.l %d7, (%a4)+ | *d++ = %d0 | ||
203 | add.l %d4, %d5 | phase += delta | ||
204 | bcc.b .usloop_0 | load next values? | ||
205 | addq.l #1, %d6 | increment position | ||
206 | cmp.l %d3, %d6 | pos < count? | ||
207 | blt.b .usloop_1 | yes? continue resampling | ||
208 | .usloop_skip: | ||
209 | subq.l #1, %d2 | ch > 0? | ||
210 | bgt.b .uschannel_loop | yes? process next channel | ||
211 | swap %d5 | wrap phase to start of next frame | ||
212 | move.l %d5, (%a0) | ...and save in r->phase | ||
213 | move.l %a4, %d0 | return d - d[0] | ||
214 | sub.l (%a2), %d0 | | ||
215 | asr.l #2, %d0 | convert bytes->samples | ||
216 | movem.l (%sp), %d2-%d7/%a2-%a5 | restore non-clobberables | ||
217 | move.l %acc1, %acc0 | clear %acc0 | ||
218 | lea.l 40(%sp), %sp | cleanup stack | ||
219 | rts | buh-bye | ||
220 | .usend: | ||
221 | .size dsp_upsample,.usend-dsp_upsample | ||