2 ; Copyright (c) 2010 The WebM project authors. All Rights Reserved.
4 ; Use of this source code is governed by a BSD-style license
5 ; that can be found in the LICENSE file in the root of the source
6 ; tree. An additional intellectual property rights grant can be found
7 ; in the file PATENTS. All contributing project authors may
8 ; be found in the AUTHORS file in the root of the source tree.
11 %include "third_party/x86inc/x86inc.asm"
19 cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
20 shift, qcoeff, dqcoeff, dequant, zbin_oq, \
25 ; actual quantize loop - setup pointers, rounders, etc.
26 movifnidn coeffq, coeffmp
27 movifnidn ncoeffq, ncoeffmp
29 movifnidn zbinq, zbinmp
30 movifnidn roundq, roundmp
31 movifnidn quantq, quantmp
32 movd m4, dword zbin_oqm ; m4 = zbin_oq
33 mova m0, [zbinq] ; m0 = zbin
35 mova m1, [roundq] ; m1 = round
37 mova m2, [quantq] ; m2 = quant
38 paddw m0, m4 ; m0 = zbin + zbin_oq
44 psrlw m0, 1 ; m0 = (m0 + 1) / 2
45 psrlw m1, 1 ; m1 = (m1 + 1) / 2
47 mova m3, [r2q] ; m3 = dequant
51 mova m4, [r2] ; m4 = shift
57 pxor m5, m5 ; m5 = dedicated zero
58 DEFINE_ARGS coeff, ncoeff, d1, qcoeff, dqcoeff, iscan, d2, d3, d4, d5, d6, eob
59 lea coeffq, [ coeffq+ncoeffq*2]
60 lea iscanq, [ iscanq+ncoeffq*2]
61 lea qcoeffq, [ qcoeffq+ncoeffq*2]
62 lea dqcoeffq, [dqcoeffq+ncoeffq*2]
65 ; get DC and first 15 AC coeffs
66 mova m9, [ coeffq+ncoeffq*2+ 0] ; m9 = c[i]
67 mova m10, [ coeffq+ncoeffq*2+16] ; m10 = c[i]
68 pabsw m6, m9 ; m6 = abs(m9)
69 pabsw m11, m10 ; m11 = abs(m10)
70 pcmpgtw m7, m6, m0 ; m7 = c[i] >= zbin
72 pcmpgtw m12, m11, m0 ; m12 = c[i] >= zbin
73 paddsw m6, m1 ; m6 += round
75 paddsw m11, m1 ; m11 += round
76 pmulhw m8, m6, m2 ; m8 = m6*q>>16
78 pmulhw m13, m11, m2 ; m13 = m11*q>>16
79 paddw m8, m6 ; m8 += m6
80 paddw m13, m11 ; m13 += m11
81 pmulhw m8, m4 ; m8 = m8*qsh>>16
83 pmulhw m13, m4 ; m13 = m13*qsh>>16
84 psignw m8, m9 ; m8 = reinsert sign
85 psignw m13, m10 ; m13 = reinsert sign
88 mova [qcoeffq+ncoeffq*2+ 0], m8
89 mova [qcoeffq+ncoeffq*2+16], m13
94 pmullw m8, m3 ; dqc[i] = qc[i] * q
96 pmullw m13, m3 ; dqc[i] = qc[i] * q
103 mova [dqcoeffq+ncoeffq*2+ 0], m8
104 mova [dqcoeffq+ncoeffq*2+16], m13
105 pcmpeqw m8, m5 ; m8 = c[i] == 0
106 pcmpeqw m13, m5 ; m13 = c[i] == 0
107 mova m6, [ iscanq+ncoeffq*2+ 0] ; m6 = scan[i]
108 mova m11, [ iscanq+ncoeffq*2+16] ; m11 = scan[i]
109 psubw m6, m7 ; m6 = scan[i] + 1
110 psubw m11, m12 ; m11 = scan[i] + 1
111 pandn m8, m6 ; m8 = max(eob)
112 pandn m13, m11 ; m13 = max(eob)
118 mova m9, [ coeffq+ncoeffq*2+ 0] ; m9 = c[i]
119 mova m10, [ coeffq+ncoeffq*2+16] ; m10 = c[i]
120 pabsw m6, m9 ; m6 = abs(m9)
121 pabsw m11, m10 ; m11 = abs(m10)
122 pcmpgtw m7, m6, m0 ; m7 = c[i] >= zbin
123 pcmpgtw m12, m11, m0 ; m12 = c[i] >= zbin
130 paddsw m6, m1 ; m6 += round
131 paddsw m11, m1 ; m11 += round
132 pmulhw m14, m6, m2 ; m14 = m6*q>>16
133 pmulhw m13, m11, m2 ; m13 = m11*q>>16
134 paddw m14, m6 ; m14 += m6
135 paddw m13, m11 ; m13 += m11
136 pmulhw m14, m4 ; m14 = m14*qsh>>16
137 pmulhw m13, m4 ; m13 = m13*qsh>>16
138 psignw m14, m9 ; m14 = reinsert sign
139 psignw m13, m10 ; m13 = reinsert sign
142 mova [qcoeffq+ncoeffq*2+ 0], m14
143 mova [qcoeffq+ncoeffq*2+16], m13
148 pmullw m14, m3 ; dqc[i] = qc[i] * q
149 pmullw m13, m3 ; dqc[i] = qc[i] * q
156 mova [dqcoeffq+ncoeffq*2+ 0], m14
157 mova [dqcoeffq+ncoeffq*2+16], m13
158 pcmpeqw m14, m5 ; m14 = c[i] == 0
159 pcmpeqw m13, m5 ; m13 = c[i] == 0
160 mova m6, [ iscanq+ncoeffq*2+ 0] ; m6 = scan[i]
161 mova m11, [ iscanq+ncoeffq*2+16] ; m11 = scan[i]
162 psubw m6, m7 ; m6 = scan[i] + 1
163 psubw m11, m12 ; m11 = scan[i] + 1
164 pandn m14, m6 ; m14 = max(eob)
165 pandn m13, m11 ; m13 = max(eob)
174 mova [qcoeffq+ncoeffq*2+ 0], m5
175 mova [qcoeffq+ncoeffq*2+16], m5
176 mova [dqcoeffq+ncoeffq*2+ 0], m5
177 mova [dqcoeffq+ncoeffq*2+16], m5
183 ; horizontally accumulate/max eobs and write into [eob] memory pointer
195 ; skip-block, i.e. just write all zeroes
198 movifnidn ncoeffq, ncoeffmp
201 DEFINE_ARGS dqcoeff, ncoeff, qcoeff, eob
202 lea dqcoeffq, [dqcoeffq+ncoeffq*2]
203 lea qcoeffq, [ qcoeffq+ncoeffq*2]
207 mova [dqcoeffq+ncoeffq*2+ 0], m7
208 mova [dqcoeffq+ncoeffq*2+16], m7
209 mova [qcoeffq+ncoeffq*2+ 0], m7
210 mova [qcoeffq+ncoeffq*2+16], m7
219 QUANTIZE_FN b_32x32, 7
222 cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
223 shift, qcoeff, dqcoeff, dequant, zbin_oq, \
228 ; actual quantize loop - setup pointers, rounders, etc.
229 movifnidn coeffq, coeffmp
230 movifnidn ncoeffq, ncoeffmp
232 movifnidn zbinq, zbinmp
233 movifnidn roundq, roundmp
234 movifnidn quantq, quantmp
235 mova m1, [roundq] ; m1 = round
236 mova m2, [quantq] ; m2 = quant
241 psrlw m1, 1 ; m1 = (m1 + 1) / 2
243 mova m3, [r2q] ; m3 = dequant
250 pxor m5, m5 ; m5 = dedicated zero
251 DEFINE_ARGS coeff, ncoeff, d1, qcoeff, dqcoeff, iscan, d2, d3, d4, d5, d6, eob
252 lea coeffq, [ coeffq+ncoeffq*2]
253 lea iscanq, [ iscanq+ncoeffq*2]
254 lea qcoeffq, [ qcoeffq+ncoeffq*2]
255 lea dqcoeffq, [dqcoeffq+ncoeffq*2]
258 ; get DC and first 15 AC coeffs
259 mova m9, [ coeffq+ncoeffq*2+ 0] ; m9 = c[i]
260 mova m10, [ coeffq+ncoeffq*2+16] ; m10 = c[i]
261 pabsw m6, m9 ; m6 = abs(m9)
262 pabsw m11, m10 ; m11 = abs(m10)
265 paddsw m6, m1 ; m6 += round
267 paddsw m11, m1 ; m11 += round
268 pmulhw m8, m6, m2 ; m8 = m6*q>>16
270 pmulhw m13, m11, m2 ; m13 = m11*q>>16
271 psignw m8, m9 ; m8 = reinsert sign
272 psignw m13, m10 ; m13 = reinsert sign
273 mova [qcoeffq+ncoeffq*2+ 0], m8
274 mova [qcoeffq+ncoeffq*2+16], m13
279 pmullw m8, m3 ; dqc[i] = qc[i] * q
281 pmullw m13, m3 ; dqc[i] = qc[i] * q
289 mova [dqcoeffq+ncoeffq*2+ 0], m8
290 mova [dqcoeffq+ncoeffq*2+16], m13
291 pcmpeqw m8, m5 ; m8 = c[i] == 0
292 pcmpeqw m13, m5 ; m13 = c[i] == 0
293 mova m6, [ iscanq+ncoeffq*2+ 0] ; m6 = scan[i]
294 mova m11, [ iscanq+ncoeffq*2+16] ; m11 = scan[i]
295 psubw m6, m7 ; m6 = scan[i] + 1
296 psubw m11, m7 ; m11 = scan[i] + 1
297 pandn m8, m6 ; m8 = max(eob)
298 pandn m13, m11 ; m13 = max(eob)
304 mova m9, [ coeffq+ncoeffq*2+ 0] ; m9 = c[i]
305 mova m10, [ coeffq+ncoeffq*2+16] ; m10 = c[i]
306 pabsw m6, m9 ; m6 = abs(m9)
307 pabsw m11, m10 ; m11 = abs(m10)
319 paddsw m6, m1 ; m6 += round
320 paddsw m11, m1 ; m11 += round
321 pmulhw m14, m6, m2 ; m14 = m6*q>>16
322 pmulhw m13, m11, m2 ; m13 = m11*q>>16
323 psignw m14, m9 ; m14 = reinsert sign
324 psignw m13, m10 ; m13 = reinsert sign
325 mova [qcoeffq+ncoeffq*2+ 0], m14
326 mova [qcoeffq+ncoeffq*2+16], m13
331 pmullw m14, m3 ; dqc[i] = qc[i] * q
332 pmullw m13, m3 ; dqc[i] = qc[i] * q
339 mova [dqcoeffq+ncoeffq*2+ 0], m14
340 mova [dqcoeffq+ncoeffq*2+16], m13
341 pcmpeqw m14, m5 ; m14 = c[i] == 0
342 pcmpeqw m13, m5 ; m13 = c[i] == 0
343 mova m6, [ iscanq+ncoeffq*2+ 0] ; m6 = scan[i]
344 mova m11, [ iscanq+ncoeffq*2+16] ; m11 = scan[i]
345 psubw m6, m7 ; m6 = scan[i] + 1
346 psubw m11, m7 ; m11 = scan[i] + 1
347 pandn m14, m6 ; m14 = max(eob)
348 pandn m13, m11 ; m13 = max(eob)
357 mova [qcoeffq+ncoeffq*2+ 0], m5
358 mova [qcoeffq+ncoeffq*2+16], m5
359 mova [dqcoeffq+ncoeffq*2+ 0], m5
360 mova [dqcoeffq+ncoeffq*2+16], m5
366 ; horizontally accumulate/max eobs and write into [eob] memory pointer
378 ; skip-block, i.e. just write all zeroes
381 movifnidn ncoeffq, ncoeffmp
384 DEFINE_ARGS dqcoeff, ncoeff, qcoeff, eob
385 lea dqcoeffq, [dqcoeffq+ncoeffq*2]
386 lea qcoeffq, [ qcoeffq+ncoeffq*2]
390 mova [dqcoeffq+ncoeffq*2+ 0], m7
391 mova [dqcoeffq+ncoeffq*2+16], m7
392 mova [qcoeffq+ncoeffq*2+ 0], m7
393 mova [qcoeffq+ncoeffq*2+16], m7
402 QUANTIZE_FP fp_32x32, 7