Parent Directory | Revision Log
Revision 1.1 - (view) (download)
1 : | ia64p | 1.1 | // *********************************** |
2 : | // interpolate8x8_ia64.s | ||
3 : | // optimized for IA-64 | ||
4 : | // Authors : Kai Kühn | ||
5 : | // Alexander Viehl | ||
6 : | // last update : 13.7.2002 | ||
7 : | // *********************************** | ||
8 : | .file "interpolate8x8_ia64.s" | ||
9 : | .pred.safe_across_calls p1-p5,p16-p63 | ||
10 : | .text | ||
11 : | .align 16 | ||
12 : | .global interpolate8x8_halfpel_h_ia64# | ||
13 : | .proc interpolate8x8_halfpel_h_ia64# | ||
14 : | interpolate8x8_halfpel_h_ia64: | ||
15 : | LL=3 | ||
16 : | SL=1 | ||
17 : | SL2=1 | ||
18 : | OL=1 | ||
19 : | OL2=1 | ||
20 : | AVL=1 | ||
21 : | AL=1 | ||
22 : | PSH=1 | ||
23 : | ML=1 | ||
24 : | STL=3 | ||
25 : | |||
26 : | alloc r9=ar.pfs,4, 60,0,64 | ||
27 : | |||
28 : | mov r20 = ar.lc | ||
29 : | mov r21 = pr | ||
30 : | |||
31 : | dep.z r22 = r33,3,3 // rshift of src | ||
32 : | |||
33 : | and r14 = -8,r33 // align src | ||
34 : | mov r15 = r32 // get dest | ||
35 : | mov r16 = r34 // stride | ||
36 : | sub r17 = 1,r35 // 1-rounding | ||
37 : | |||
38 : | ;; | ||
39 : | |||
40 : | add r18 = 8,r14 | ||
41 : | mux2 r17 = r17, 0x00 // broadcast 1-rounding | ||
42 : | |||
43 : | sub r24 = 64,r22 // lshift of src | ||
44 : | add r26 = 8,r22 // rshift of src+1 | ||
45 : | sub r27 = 56,r22 // lshift of src+1 | ||
46 : | |||
47 : | mov ar.lc = 7 // loopcounter | ||
48 : | mov ar.ec = LL + SL +OL + AVL + STL + 2*PSH + 2*AL + ML // sum of latencies | ||
49 : | mov pr.rot = 1 << 16 // init pr regs for sw-pipeling | ||
50 : | |||
51 : | ;; | ||
52 : | .rotr ald1[LL+1],ald2[LL+1],shru1[SL+1],shl1[SL+1],shru2[SL+1],shl2[SL+1],or1[OL+1+PSH],or2[OL+1+PSH],pshb1[PSH+1],pshb2[PSH+1],pshe1[PSH+1],pshe2[PSH+1+AL],pshe3[PSH+1],pshe4[PSH+1+AL],add1[AL+1],add2[AL+1],add3[AL+1],add4[AL+1],avg1[AVL+1],avg2[AVL+1],pmix1[ML+1] | ||
53 : | .rotp aldp[LL], sh1p[SL], or1p[OL], pshb[PSH], pshe[PSH],addp[AL],add2p[AL],pavg1p[AVL],mixp[ML],stp[STL] | ||
54 : | |||
55 : | .Lloop_interpolate: | ||
56 : | (aldp[0]) ld8 ald1[0] = [r14],r16 // load aligned src | ||
57 : | (aldp[0]) ld8 ald2[0] = [r18],r16 // and aligned src+8 | ||
58 : | |||
59 : | (sh1p[0]) shr.u shru1[0] = ald1[LL],r22 // get src | ||
60 : | (sh1p[0]) shl shl1[0] = ald2[LL],r27 | ||
61 : | (sh1p[0]) shr.u shru2[0] = ald1[LL],r26 // get src+1 | ||
62 : | (sh1p[0]) shl shl2[0] = ald2[LL],r24 | ||
63 : | |||
64 : | (or1p[0]) or or1[0] = shru1[SL],shl2[SL] // merge things | ||
65 : | (or1p[0]) or or2[0] = shru2[SL],shl1[SL] | ||
66 : | |||
67 : | (pshb[0]) pshl2 pshb1[0] = or1[OL],8 | ||
68 : | (pshb[0]) pshl2 pshb2[0] = or2[OL],8 | ||
69 : | |||
70 : | (pshe[0]) pshr2.u pshe1[0] = pshb1[PSH],8 | ||
71 : | (pshe[0]) pshr2.u pshe2[0] = pshb2[PSH],8 | ||
72 : | (pshe[0]) pshr2.u pshe3[0] = or1[PSH+OL],8 | ||
73 : | (pshe[0]) pshr2.u pshe4[0] = or2[PSH+OL],8 | ||
74 : | |||
75 : | (addp[0]) padd2.sss add1[0] = pshe1[PSH],r17 // add 1-rounding | ||
76 : | (addp[0]) padd2.sss add2[0] = pshe3[PSH],r17 // add 1-rounding | ||
77 : | |||
78 : | (add2p[0]) padd2.uus add3[0] = pshe2[AL+PSH],add1[AL] | ||
79 : | (add2p[0]) padd2.uus add4[0] = pshe4[AL+PSH],add2[AL] | ||
80 : | |||
81 : | (pavg1p[0]) pshr2.u avg1[0] = add3[AL],1 // parallel average | ||
82 : | (pavg1p[0]) pshr2.u avg2[0] = add4[AL],1 // parallel average | ||
83 : | |||
84 : | (mixp[0]) mix1.r pmix1[0] = avg2[AVL],avg1[AVL] | ||
85 : | |||
86 : | (stp[0]) st8 [r15] = pmix1[ML] // store results | ||
87 : | (stp[0]) add r15 = r15,r16 | ||
88 : | |||
89 : | |||
90 : | |||
91 : | |||
92 : | br.ctop.sptk.few .Lloop_interpolate | ||
93 : | ;; | ||
94 : | mov ar.lc = r20 | ||
95 : | mov pr = r21,-1 | ||
96 : | br.ret.sptk.many b0 | ||
97 : | .endp interpolate8x8_halfpel_h_ia64# | ||
98 : | |||
99 : | .align 16 | ||
100 : | .global interpolate8x8_halfpel_v_ia64# | ||
101 : | .proc interpolate8x8_halfpel_v_ia64# | ||
102 : | interpolate8x8_halfpel_v_ia64: | ||
103 : | LL=3 | ||
104 : | SL=1 | ||
105 : | SL2=1 | ||
106 : | OL=1 | ||
107 : | OL2=1 | ||
108 : | AVL=1 | ||
109 : | AL=1 | ||
110 : | PSH=1 | ||
111 : | ML=1 | ||
112 : | STL=3 | ||
113 : | |||
114 : | alloc r9=ar.pfs,4, 60,0,64 | ||
115 : | |||
116 : | mov r20 = ar.lc | ||
117 : | mov r21 = pr | ||
118 : | |||
119 : | dep.z r22 = r33,3,3 | ||
120 : | |||
121 : | and r14 = -8,r33 | ||
122 : | mov r15 = r32 | ||
123 : | mov r16 = r34 | ||
124 : | sub r17 = 1,r35 | ||
125 : | ;; | ||
126 : | |||
127 : | add r18 = 8,r14 | ||
128 : | add r19 = r14,r16 // src + stride | ||
129 : | mux2 r17 = r17, 0x00 | ||
130 : | |||
131 : | sub r24 = 64,r22 | ||
132 : | ;; | ||
133 : | add r26 = 8,r19 // src + stride + 8 | ||
134 : | |||
135 : | mov ar.lc = 7 | ||
136 : | mov ar.ec = LL + SL +OL + AVL + STL + 2*PSH + 2*AL + ML // sum of latencies | ||
137 : | mov pr.rot = 1 << 16 | ||
138 : | |||
139 : | ;; | ||
140 : | .rotr ald1[LL+1],ald2[LL+1],ald3[LL+1],ald4[LL+1],shru1[SL+1],shl1[SL+1],shru2[SL+1],shl2[SL+1],or1[OL+1+PSH],or2[OL+1+PSH],pshb1[PSH+1],pshb2[PSH+1],pshe1[PSH+1],pshe2[PSH+1+AL],pshe3[PSH+1],pshe4[PSH+1+AL],add1[AL+1],add2[AL+1],add3[AL+1],add4[AL+1],avg1[AVL+1],avg2[AVL+1],pmix1[ML+1] | ||
141 : | .rotp aldp[LL], sh1p[SL], or1p[OL], pshb[PSH], pshe[PSH],addp[AL],add2p[AL],pavg1p[AVL],mixp[ML],stp[STL] | ||
142 : | |||
143 : | |||
144 : | |||
145 : | .Lloop_interpolate2: | ||
146 : | (aldp[0]) ld8 ald1[0] = [r14],r16 | ||
147 : | (aldp[0]) ld8 ald2[0] = [r18],r16 | ||
148 : | (aldp[0]) ld8 ald3[0] = [r19],r16 | ||
149 : | (aldp[0]) ld8 ald4[0] = [r26],r16 | ||
150 : | |||
151 : | (sh1p[0]) shr.u shru1[0] = ald1[LL],r22 | ||
152 : | (sh1p[0]) shl shl1[0] = ald2[LL],r24 | ||
153 : | (sh1p[0]) shr.u shru2[0] = ald3[LL],r22 | ||
154 : | (sh1p[0]) shl shl2[0] = ald4[LL],r24 | ||
155 : | |||
156 : | (or1p[0]) or or1[0] = shru1[SL],shl1[SL] | ||
157 : | (or1p[0]) or or2[0] = shru2[SL],shl2[SL] | ||
158 : | |||
159 : | (pshb[0]) pshl2 pshb1[0] = or1[OL],8 | ||
160 : | (pshb[0]) pshl2 pshb2[0] = or2[OL],8 | ||
161 : | |||
162 : | (pshe[0]) pshr2.u pshe1[0] = pshb1[PSH],8 | ||
163 : | (pshe[0]) pshr2.u pshe2[0] = pshb2[PSH],8 | ||
164 : | (pshe[0]) pshr2.u pshe3[0] = or1[PSH+OL],8 | ||
165 : | (pshe[0]) pshr2.u pshe4[0] = or2[PSH+OL],8 | ||
166 : | |||
167 : | (addp[0]) padd2.sss add1[0] = pshe1[PSH],r17 // add 1-rounding | ||
168 : | (addp[0]) padd2.sss add2[0] = pshe3[PSH],r17 // add 1-rounding | ||
169 : | |||
170 : | (add2p[0]) padd2.uus add3[0] = pshe2[AL+PSH],add1[AL] | ||
171 : | (add2p[0]) padd2.uus add4[0] = pshe4[AL+PSH],add2[AL] | ||
172 : | |||
173 : | (pavg1p[0]) pshr2.u avg1[0] = add3[AL],1 // parallel average | ||
174 : | (pavg1p[0]) pshr2.u avg2[0] = add4[AL],1 // parallel average | ||
175 : | |||
176 : | (mixp[0]) mix1.r pmix1[0] = avg2[AVL],avg1[AVL] | ||
177 : | |||
178 : | (stp[0]) st8 [r15] = pmix1[ML] | ||
179 : | (stp[0]) add r15 = r15,r16 | ||
180 : | |||
181 : | |||
182 : | |||
183 : | |||
184 : | br.ctop.sptk.few .Lloop_interpolate2 | ||
185 : | ;; | ||
186 : | mov ar.lc = r20 | ||
187 : | mov pr = r21,-1 | ||
188 : | br.ret.sptk.many b0 | ||
189 : | .endp interpolate8x8_halfpel_v_ia64# | ||
190 : | |||
191 : | .align 16 | ||
192 : | .global interpolate8x8_halfpel_hv_ia64# | ||
193 : | .proc interpolate8x8_halfpel_hv_ia64# | ||
194 : | interpolate8x8_halfpel_hv_ia64: | ||
195 : | LL=3 | ||
196 : | SL=1 | ||
197 : | SL2=1 | ||
198 : | OL=1 | ||
199 : | OL2=1 | ||
200 : | AVL=1 | ||
201 : | AL=1 | ||
202 : | PSH=1 | ||
203 : | ML=1 | ||
204 : | STL=3 | ||
205 : | |||
206 : | alloc r9=ar.pfs,4, 92,0,96 | ||
207 : | |||
208 : | mov r20 = ar.lc | ||
209 : | mov r21 = pr | ||
210 : | |||
211 : | dep.z r22 = r33,3,3 | ||
212 : | |||
213 : | and r14 = -8,r33 | ||
214 : | mov r15 = r32 | ||
215 : | mov r16 = r34 | ||
216 : | sub r17 = 2,r35 | ||
217 : | ;; | ||
218 : | |||
219 : | add r18 = 8,r14 | ||
220 : | add r19 = r14,r16 | ||
221 : | mux2 r17 = r17, 0x00 | ||
222 : | |||
223 : | add r27 = 8,r22 | ||
224 : | sub r28 = 56,r22 | ||
225 : | sub r24 = 64,r22 | ||
226 : | ;; | ||
227 : | add r26 = 8,r19 | ||
228 : | |||
229 : | mov ar.lc = 7 | ||
230 : | mov ar.ec = LL + SL +OL + AVL + STL + 2*PSH + 2*AL + ML // sum of latencies | ||
231 : | |||
232 : | mov pr.rot = 1 << 16 | ||
233 : | |||
234 : | ;; | ||
235 : | .rotr ald1[LL+1],ald2[LL+1],ald3[LL+1],ald4[LL+1],shru1[SL+1],shl1[SL+1],shru2[SL+1],shl2[SL+1],shl3[SL+1],shru3[SL+1],shl4[SL+1],shru4[SL+1],or1[OL+1+PSH],or2[OL+1+PSH],or3[OL+1+PSH],or4[OL+1+PSH],pshb1[PSH+1],pshb2[PSH+1],pshb3[PSH+1],pshb4[PSH+1],pshe1[PSH+1],pshe2[PSH+1+AL],pshe3[PSH+1],pshe4[PSH+1+AL],pshe5[PSH+1],pshe6[PSH+1+AL],pshe7[PSH+1],pshe8[PSH+1+AL],add1[AL+1],add2[AL+1],add3[AL+1],add4[AL+1],add5[AL+1],add6[AL+1],add7[AL+1],add8[AL+1],avg1[AVL+1],avg2[AVL+1],pmix1[ML+1] | ||
236 : | .rotp aldp[LL], sh1p[SL],or1p[OL],pshb[PSH],pshe[PSH],addp[AL],add2p[AL],add3p[AL],pavg1p[AVL],mixp[ML],stp[STL] | ||
237 : | |||
238 : | |||
239 : | |||
240 : | .Lloop_interpolate3: | ||
241 : | (aldp[0]) ld8 ald1[0] = [r14],r16 | ||
242 : | (aldp[0]) ld8 ald2[0] = [r18],r16 | ||
243 : | (aldp[0]) ld8 ald3[0] = [r19],r16 | ||
244 : | (aldp[0]) ld8 ald4[0] = [r26],r16 | ||
245 : | |||
246 : | (sh1p[0]) shr.u shru1[0] = ald1[LL],r22 | ||
247 : | (sh1p[0]) shl shl1[0] = ald2[LL],r24 | ||
248 : | (sh1p[0]) shr.u shru2[0] = ald3[LL],r22 | ||
249 : | (sh1p[0]) shl shl2[0] = ald4[LL],r24 | ||
250 : | (sh1p[0]) shr.u shru3[0] = ald1[LL],r27 | ||
251 : | (sh1p[0]) shl shl3[0] = ald2[LL],r28 | ||
252 : | (sh1p[0]) shr.u shru4[0] = ald3[LL],r27 | ||
253 : | (sh1p[0]) shl shl4[0] = ald4[LL],r28 | ||
254 : | |||
255 : | |||
256 : | (or1p[0]) or or1[0] = shru1[SL],shl1[SL] | ||
257 : | (or1p[0]) or or2[0] = shru2[SL],shl2[SL] | ||
258 : | (or1p[0]) or or3[0] = shru3[SL],shl3[SL] | ||
259 : | (or1p[0]) or or4[0] = shru4[SL],shl4[SL] | ||
260 : | |||
261 : | (pshb[0]) pshl2 pshb1[0] = or1[OL],8 | ||
262 : | (pshb[0]) pshl2 pshb2[0] = or2[OL],8 | ||
263 : | (pshb[0]) pshl2 pshb3[0] = or3[OL],8 | ||
264 : | (pshb[0]) pshl2 pshb4[0] = or4[OL],8 | ||
265 : | |||
266 : | |||
267 : | (pshe[0]) pshr2.u pshe1[0] = pshb1[PSH],8 | ||
268 : | (pshe[0]) pshr2.u pshe2[0] = pshb2[PSH],8 | ||
269 : | (pshe[0]) pshr2.u pshe3[0] = or1[PSH+OL],8 | ||
270 : | (pshe[0]) pshr2.u pshe4[0] = or2[PSH+OL],8 | ||
271 : | (pshe[0]) pshr2.u pshe5[0] = pshb3[PSH],8 | ||
272 : | (pshe[0]) pshr2.u pshe6[0] = pshb4[PSH],8 | ||
273 : | (pshe[0]) pshr2.u pshe7[0] = or3[PSH+OL],8 | ||
274 : | (pshe[0]) pshr2.u pshe8[0] = or4[PSH+OL],8 | ||
275 : | |||
276 : | |||
277 : | |||
278 : | (addp[0]) padd2.sss add1[0] = pshe1[PSH],pshe2[PSH] // add 1-rounding | ||
279 : | (addp[0]) padd2.sss add2[0] = pshe3[PSH],pshe4[PSH] // add 1-rounding | ||
280 : | (addp[0]) padd2.sss add5[0] = pshe5[PSH],pshe6[PSH] // add 1-rounding | ||
281 : | (addp[0]) padd2.sss add6[0] = pshe7[PSH],pshe8[PSH] // add 1-rounding | ||
282 : | |||
283 : | |||
284 : | (add2p[0]) padd2.uus add3[0] = add1[AL],add5[AL] | ||
285 : | (add2p[0]) padd2.uus add4[0] = add2[AL],add6[AL] | ||
286 : | |||
287 : | (add3p[0]) padd2.uus add7[0] = add3[AL],r17 | ||
288 : | (add3p[0]) padd2.uus add8[0] = add4[AL],r17 | ||
289 : | |||
290 : | |||
291 : | (pavg1p[0]) pshr2.u avg1[0] = add7[AL],2 // parallel average | ||
292 : | (pavg1p[0]) pshr2.u avg2[0] = add8[AL],2 // parallel average | ||
293 : | |||
294 : | (mixp[0]) mix1.r pmix1[0] = avg2[AVL],avg1[AVL] | ||
295 : | |||
296 : | |||
297 : | (stp[0]) st8 [r15] = pmix1[ML] | ||
298 : | (stp[0]) add r15 = r15,r16 | ||
299 : | |||
300 : | |||
301 : | |||
302 : | |||
303 : | br.ctop.sptk.few .Lloop_interpolate3 | ||
304 : | ;; | ||
305 : | mov ar.lc = r20 | ||
306 : | mov pr = r21,-1 | ||
307 : | br.ret.sptk.many b0 | ||
308 : | .endp interpolate8x8_halfpel_hv_ia64# | ||
309 : | |||
310 : |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |