Parent Directory | Revision Log
Revision 1.2 - (view) (download) (as text)
1 : | ia64p | 1.1 | #! /usr/bin/python |
2 : | |||
3 : | ia64p | 1.2 | # generate the linzer-feig multiply-add idct for ia64 |
4 : | # (c) 2002 Christian Schwarz <schwarz@ira.uka.de>, | ||
5 : | # Haiko Gaisser <haiko@gaisser.de>, | ||
6 : | # Sebastian Hack <mail@s-hack.de> | ||
7 : | |||
8 : | |||
9 : | ia64p | 1.1 | import math |
10 : | |||
11 : | pre_shuffle = [ 0, 4, 2, 6, 1, 7, 3, 5 ] | ||
12 : | post_shuffle = [ 0, 1, 6, 3, 7, 2, 5, 4 ] | ||
13 : | |||
14 : | constants = 16 | ||
15 : | float_scratch = range(32, 32+constants) | ||
16 : | regbase = max(float_scratch)+1 | ||
17 : | intregbase = 33 | ||
18 : | |||
19 : | def print_matrix(matrix,s=''): | ||
20 : | if s != '': | ||
21 : | print "\n\t// %s" % s | ||
22 : | for i in range(0, 8): | ||
23 : | print "\t// ", | ||
24 : | for j in range(0, 4): | ||
25 : | print "%2d" % matrix[i*4+j], | ||
26 : | print "" | ||
27 : | |||
28 : | def exchange_elements(list, a, b): | ||
29 : | """ Exchange two list elements | ||
30 : | """ | ||
31 : | (list[a], list[b]) = (list[b], list[a]) | ||
32 : | |||
33 : | def alloc_regs(matrix, n): | ||
34 : | """ get the smallest register not used by the matrix | ||
35 : | """ | ||
36 : | |||
37 : | regs = [ ] | ||
38 : | for i in range(0, n): | ||
39 : | m = regbase | ||
40 : | while m in matrix or m in regs: | ||
41 : | m = m + 1 | ||
42 : | regs.append(m) | ||
43 : | return regs | ||
44 : | |||
45 : | def transpose_2x2_submatrix(matrix, i, j): | ||
46 : | """ transpose a 2x2 submatrix in the 8x8 matrix | ||
47 : | """ | ||
48 : | a = j | ||
49 : | b = i | ||
50 : | |||
51 : | tmp = matrix[i*8+j] | ||
52 : | matrix[i*8+j] = matrix[a*8+b] | ||
53 : | matrix[a*8+b] = tmp | ||
54 : | |||
55 : | tmp = matrix[i*8+j+4] | ||
56 : | matrix[i*8+j+4] = matrix[a*8+b+4] | ||
57 : | matrix[a*8+b+4] = tmp | ||
58 : | |||
59 : | |||
60 : | def transpose(matrix): | ||
61 : | """ register renaming for transpose | ||
62 : | """ | ||
63 : | regs = alloc_regs(matrix, 16) | ||
64 : | save_regs = regs[:] | ||
65 : | |||
66 : | # emit code ... | ||
67 : | for i in range(1,8,2): | ||
68 : | for j in range(0,4): | ||
69 : | r1 = matrix[(i-1)*4+j] | ||
70 : | r2 = matrix[i*4+j] | ||
71 : | print '\tfmix.r f%d = f%d, f%d' % (save_regs.pop(0), r1, r2) | ||
72 : | |||
73 : | print '\t;;' | ||
74 : | |||
75 : | for i in range(0,8,2): | ||
76 : | for j in range(0,4): | ||
77 : | r1 = matrix[i*4+j] | ||
78 : | r2 = matrix[(i+1)*4+j] | ||
79 : | print '\tfmix.l f%d = f%d, f%d' % (r1, r1, r2) | ||
80 : | |||
81 : | print '\t;;' | ||
82 : | |||
83 : | # first stage, transpose the 2x2 matrices | ||
84 : | for i in range(1,8,2): | ||
85 : | for j in range(0,4): | ||
86 : | r = matrix[i*4+j] | ||
87 : | matrix[i*4+j] = regs.pop(0) | ||
88 : | |||
89 : | # print_matrix(matrix) | ||
90 : | |||
91 : | # exchange the 2x2 matrices by renaming the registers | ||
92 : | for i in range(0, 4): | ||
93 : | for j in range(i+1, 4): | ||
94 : | transpose_2x2_submatrix(matrix, i, j) | ||
95 : | |||
96 : | # print '' | ||
97 : | # print_matrix(matrix) | ||
98 : | # print "transpose" | ||
99 : | # print_matrix(matrix) | ||
100 : | |||
101 : | # register renaming for 8 regs containing a column | ||
102 : | def shuffle_column(matrix, col, permutation): | ||
103 : | l = [ ] | ||
104 : | for i in range(0,8): | ||
105 : | l.append(matrix[i*4+col]) | ||
106 : | for i in range(0,8): | ||
107 : | matrix[i*4+col] = l[permutation[i]] | ||
108 : | |||
109 : | def butterfly(matrix, col, i, j, c1, c2): | ||
110 : | """ register renaming for a butterfly operation in a column | ||
111 : | """ | ||
112 : | ri = matrix[i*4+col] | ||
113 : | rj = matrix[j*4+col] | ||
114 : | regs = alloc_regs(matrix, 1) | ||
115 : | |||
116 : | print '\t// (f%d, f%d) = (f%d, f%d) $ (%s, %s), (line %d, %d)' % \ | ||
117 : | (regs[0], rj, ri, rj, c1, c2, i, j) | ||
118 : | print '\tfpma f%d = f%d, %s, f%d' % (regs[0], rj, c1, ri) | ||
119 : | print '\tfpnma f%d = f%d, %s, f%d' % (rj, rj, c2, ri) | ||
120 : | print '\t;;' | ||
121 : | |||
122 : | matrix[i*4+col] = regs[0] | ||
123 : | |||
124 : | |||
125 : | def column_idct(matrix, col): | ||
126 : | |||
127 : | print_matrix(matrix, "before pre shuffle") | ||
128 : | shuffle_column(matrix, col, pre_shuffle) | ||
129 : | print_matrix(matrix, "after pre shuffle") | ||
130 : | |||
131 : | butterfly(matrix, col, 0, 1, 'c0', 'c0') | ||
132 : | butterfly(matrix, col, 2, 3, 'c1', 'c2') | ||
133 : | butterfly(matrix, col, 4, 5, 'c3', 'c4') | ||
134 : | butterfly(matrix, col, 6, 7, 'c5', 'c6') | ||
135 : | print '\t;;' | ||
136 : | butterfly(matrix, col, 0, 3, 'c7', 'c7') | ||
137 : | butterfly(matrix, col, 1, 2, 'c8', 'c8') | ||
138 : | butterfly(matrix, col, 4, 6, 'c9', 'c9') | ||
139 : | butterfly(matrix, col, 5, 7, 'c10', 'c10') | ||
140 : | print '\t;;' | ||
141 : | butterfly(matrix, col, 5, 6, 'c11', 'c11') | ||
142 : | butterfly(matrix, col, 0, 4, 'c12', 'c12') | ||
143 : | butterfly(matrix, col, 3, 7, 'c14', 'c14') | ||
144 : | print '\t;;' | ||
145 : | butterfly(matrix, col, 1, 5, 'c13', 'c13') | ||
146 : | butterfly(matrix, col, 2, 6, 'c13', 'c13') | ||
147 : | |||
148 : | print_matrix(matrix, "before post shuffle") | ||
149 : | shuffle_column(matrix, col, post_shuffle) | ||
150 : | print_matrix(matrix, "after post shuffle") | ||
151 : | |||
152 : | def gen_idct(matrix): | ||
153 : | |||
154 : | for j in range(0, 2): | ||
155 : | for i in range(0, 4): | ||
156 : | print '\tfpma f%d = f%d, c0, f0' \ | ||
157 : | % (2 * (matrix[i],)) | ||
158 : | print '\t;;' | ||
159 : | for i in range(0,4): | ||
160 : | column_idct(matrix, i) | ||
161 : | print '\t;;' | ||
162 : | transpose(matrix) | ||
163 : | |||
164 : | def gen_consts(): | ||
165 : | print 'addreg1 = r14' | ||
166 : | print 'addreg2 = r15' | ||
167 : | |||
168 : | for i in range(0, constants): | ||
169 : | print 'c%d = f%d' % (i, float_scratch.pop(0)) | ||
170 : | |||
171 : | sqrt2 = math.sqrt(2.0) | ||
172 : | t = [ ] | ||
173 : | s = [ ] | ||
174 : | c = [ ] | ||
175 : | for i in range(0,5): | ||
176 : | t.append(math.tan(i * math.pi / 16)) | ||
177 : | s.append(math.sin(i * math.pi / 16)) | ||
178 : | c.append(math.cos(i * math.pi / 16)) | ||
179 : | |||
180 : | consts = [ ] | ||
181 : | consts.append(1.0 / (2.0 * sqrt2)) | ||
182 : | consts.append(-1 / t[2]) | ||
183 : | consts.append(-t[2]) | ||
184 : | consts.append(t[1]) | ||
185 : | consts.append(1 / t[1]) | ||
186 : | consts.append(t[3]) | ||
187 : | consts.append(1 / t[3]) | ||
188 : | consts.append(0.5 * c[2]) | ||
189 : | consts.append(0.5 * s[2]) | ||
190 : | consts.append(c[3] / c[1]) | ||
191 : | consts.append(s[3] / s[1]) | ||
192 : | consts.append(c[1] / s[1]) | ||
193 : | consts.append(0.5 * c[1]) | ||
194 : | consts.append(0.5 * s[1] * c[4]) | ||
195 : | consts.append(0.5 * s[1]) | ||
196 : | consts.append(1.0) | ||
197 : | |||
198 : | print '.sdata' | ||
199 : | for i in range(0, constants): | ||
200 : | if i % 2 == 0: | ||
201 : | print '.align 16' | ||
202 : | print '.data_c%d:' % i | ||
203 : | print '.single %.30f, %.30f' % (consts[i], consts[i]) | ||
204 : | print '' | ||
205 : | |||
206 : | def gen_load(matrix): | ||
207 : | |||
208 : | for i in range(0, 64, 2): | ||
209 : | print '\tld2 r%d = [addreg1], 4' % (intregbase+i) | ||
210 : | print '\tld2 r%d = [addreg2], 4' % (intregbase+i+1) | ||
211 : | print '\t;;' | ||
212 : | |||
213 : | for i in range(0, 64, 2): | ||
214 : | print '\tsxt2 r%d = r%d' % (2*(intregbase+i,)) | ||
215 : | print '\tsxt2 r%d = r%d' % (2*(intregbase+i+1,)) | ||
216 : | print '\t;;' | ||
217 : | |||
218 : | for i in range(0, 64, 2): | ||
219 : | print '\tsetf.sig f%d = r%d' % (regbase+i, intregbase+i) | ||
220 : | print '\tsetf.sig f%d = r%d' % (regbase+i+1, intregbase+i+1) | ||
221 : | print '\t;;' | ||
222 : | |||
223 : | for i in range(0, 64, 2): | ||
224 : | print '\tfcvt.xf f%d = f%d' % (2*(regbase+i,)) | ||
225 : | print '\tfcvt.xf f%d = f%d' % (2*(regbase+i+1,)) | ||
226 : | print '\t;;' | ||
227 : | |||
228 : | for i in range(0, 32): | ||
229 : | print '\tfpack f%d = f%d, f%d' \ | ||
230 : | % (regbase+i, regbase+2*i, regbase+2*i+1) | ||
231 : | print '\t;;' | ||
232 : | |||
233 : | """ | ||
234 : | for i in range(0, len(matrix)): | ||
235 : | print '\tld2 r18 = [addreg1], 4' | ||
236 : | print '\tld2 r19 = [addreg2], 4' | ||
237 : | print '\t;;' | ||
238 : | print '\tsxt2 r18 = r18' | ||
239 : | print '\tsxt2 r19 = r19' | ||
240 : | print '\t;;' | ||
241 : | print '\tsetf.sig f18 = r18' | ||
242 : | print '\tsetf.sig f19 = r19' | ||
243 : | print '\t;;' | ||
244 : | print '\tfcvt.xf f18 = f18' | ||
245 : | print '\tfcvt.xf f19 = f19' | ||
246 : | print '\t;;' | ||
247 : | print '\tfpack f%d = f18, f19' % (matrix[i]) | ||
248 : | print '\t;;' | ||
249 : | """ | ||
250 : | |||
251 : | def gen_store(matrix): | ||
252 : | print '\tmov addreg1 = in0' | ||
253 : | print '\tadd addreg2 = 4, in0' | ||
254 : | print '\t;;' | ||
255 : | |||
256 : | for i in range(0, len(matrix)): | ||
257 : | print '\tfpcvt.fx f%d = f%d' % (2*(matrix[i],)) | ||
258 : | print '\t;;' | ||
259 : | |||
260 : | for i in range(0, len(matrix)): | ||
261 : | print '\tgetf.sig r%d = f%d' % (intregbase+i, matrix[i]) | ||
262 : | print '\t;;' | ||
263 : | |||
264 : | for i in range(0, len(matrix)): | ||
265 : | print '\tshl r%d = r%d, 7' % (2*(intregbase+i,)) | ||
266 : | print '\t;;' | ||
267 : | |||
268 : | for i in range(0, len(matrix)): | ||
269 : | print '\tpack4.sss r%d = r%d, r0' % (2*(intregbase+i,)) | ||
270 : | print '\t;;' | ||
271 : | |||
272 : | for i in range(0, len(matrix)): | ||
273 : | print '\tpshr2 r%d = r%d, 7' % (2*(intregbase+i,)) | ||
274 : | print '\t;;' | ||
275 : | |||
276 : | for i in range(0, len(matrix)): | ||
277 : | print '\tmux2 r%d = r%d, 0xe1' % (2*(intregbase+i,)) | ||
278 : | print '\t;;' | ||
279 : | |||
280 : | for i in range(0, len(matrix), 2): | ||
281 : | print '\tst4 [addreg1] = r%d, 8' % (intregbase+i) | ||
282 : | print '\tst4 [addreg2] = r%d, 8' % (intregbase+i+1) | ||
283 : | ia64p | 1.2 | print '\t;;' |
284 : | ia64p | 1.1 | |
285 : | def main(): | ||
286 : | gen_consts() | ||
287 : | |||
288 : | print '.text' | ||
289 : | print '.global idct_ia64' | ||
290 : | ia64p | 1.2 | print '.global idct_ia64_init' |
291 : | print '.align 16' | ||
292 : | print '.proc idct_ia64_init' | ||
293 : | print 'idct_ia64_init:' | ||
294 : | print 'br.ret.sptk.few b0' | ||
295 : | print '.endp' | ||
296 : | ia64p | 1.1 | print '.align 16' |
297 : | print '.proc idct_ia64' | ||
298 : | print 'idct_ia64:' | ||
299 : | |||
300 : | f = open('idct_init.s') | ||
301 : | print f.read() | ||
302 : | f.close() | ||
303 : | |||
304 : | matrix = [ ] | ||
305 : | for i in range(0,32): | ||
306 : | matrix.append(regbase + i) | ||
307 : | |||
308 : | gen_load(matrix) | ||
309 : | # print_matrix(matrix) | ||
310 : | gen_idct(matrix) | ||
311 : | # transpose(matrix) | ||
312 : | print_matrix(matrix) | ||
313 : | gen_store(matrix) | ||
314 : | |||
315 : | f = open('idct_fini.s') | ||
316 : | print f.read() | ||
317 : | f.close() | ||
318 : | |||
319 : | print '.endp' | ||
320 : | |||
321 : | |||
322 : | if __name__ == "__main__": | ||
323 : | main() |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |