1;
2;  Copyright (c) 2010 The Webm project authors. All Rights Reserved.
3;
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.
9;
10
11
12    EXPORT  |idct_dequant_full_2x_neon|
13    ARM
14    REQUIRE8
15    PRESERVE8
16
17    AREA ||.text||, CODE, READONLY, ALIGN=2
18;void idct_dequant_full_2x_neon(short *q, short *dq,
19;                               unsigned char *dst, int stride);
20; r0    *q,
21; r1    *dq,
22; r2    *dst
23; r3    stride
24|idct_dequant_full_2x_neon| PROC
25    vld1.16         {q0, q1}, [r1]          ; dq (same l/r)
26    vld1.16         {q2, q3}, [r0]          ; l q
27    add             r0, r0, #32
28    vld1.16         {q4, q5}, [r0]          ; r q
29    add             r12, r2, #4
30
31    ; interleave the predictors
32    vld1.32         {d28[0]}, [r2],  r3     ; l pre
33    vld1.32         {d28[1]}, [r12], r3     ; r pre
34    vld1.32         {d29[0]}, [r2],  r3
35    vld1.32         {d29[1]}, [r12], r3
36    vld1.32         {d30[0]}, [r2],  r3
37    vld1.32         {d30[1]}, [r12], r3
38    vld1.32         {d31[0]}, [r2],  r3
39    vld1.32         {d31[1]}, [r12]
40
41    adr             r1, cospi8sqrt2minus1   ; pointer to the first constant
42
43    ; dequant: q[i] = q[i] * dq[i]
44    vmul.i16        q2, q2, q0
45    vmul.i16        q3, q3, q1
46    vmul.i16        q4, q4, q0
47    vmul.i16        q5, q5, q1
48
49    vld1.16         {d0}, [r1]
50
51    ; q2: l0r0  q3: l8r8
52    ; q4: l4r4  q5: l12r12
53    vswp            d5, d8
54    vswp            d7, d10
55
56    ; _CONSTANTS_ * 4,12 >> 16
57    ; q6:  4 * sinpi : c1/temp1
58    ; q7: 12 * sinpi : d1/temp2
59    ; q8:  4 * cospi
60    ; q9: 12 * cospi
61    vqdmulh.s16     q6, q4, d0[2]           ; sinpi8sqrt2
62    vqdmulh.s16     q7, q5, d0[2]
63    vqdmulh.s16     q8, q4, d0[0]           ; cospi8sqrt2minus1
64    vqdmulh.s16     q9, q5, d0[0]
65
66    vqadd.s16       q10, q2, q3             ; a1 = 0 + 8
67    vqsub.s16       q11, q2, q3             ; b1 = 0 - 8
68
69    ; vqdmulh only accepts signed values. this was a problem because
70    ; our constant had the high bit set, and was treated as a negative value.
71    ; vqdmulh also doubles the value before it shifts by 16. we need to
72    ; compensate for this. in the case of sinpi8sqrt2, the lowest bit is 0,
73    ; so we can shift the constant without losing precision. this avoids
74    ; shift again afterward, but also avoids the sign issue. win win!
75    ; for cospi8sqrt2minus1 the lowest bit is 1, so we lose precision if we
76    ; pre-shift it
77    vshr.s16        q8, q8, #1
78    vshr.s16        q9, q9, #1
79
80    ; q4:  4 +  4 * cospi : d1/temp1
81    ; q5: 12 + 12 * cospi : c1/temp2
82    vqadd.s16       q4, q4, q8
83    vqadd.s16       q5, q5, q9
84
85    ; c1 = temp1 - temp2
86    ; d1 = temp1 + temp2
87    vqsub.s16       q2, q6, q5
88    vqadd.s16       q3, q4, q7
89
90    ; [0]: a1+d1
91    ; [1]: b1+c1
92    ; [2]: b1-c1
93    ; [3]: a1-d1
94    vqadd.s16       q4, q10, q3
95    vqadd.s16       q5, q11, q2
96    vqsub.s16       q6, q11, q2
97    vqsub.s16       q7, q10, q3
98
99    ; rotate
100    vtrn.32         q4, q6
101    vtrn.32         q5, q7
102    vtrn.16         q4, q5
103    vtrn.16         q6, q7
104    ; idct loop 2
105    ; q4: l 0, 4, 8,12 r 0, 4, 8,12
106    ; q5: l 1, 5, 9,13 r 1, 5, 9,13
107    ; q6: l 2, 6,10,14 r 2, 6,10,14
108    ; q7: l 3, 7,11,15 r 3, 7,11,15
109
110    ; q8:  1 * sinpi : c1/temp1
111    ; q9:  3 * sinpi : d1/temp2
112    ; q10: 1 * cospi
113    ; q11: 3 * cospi
114    vqdmulh.s16     q8, q5, d0[2]           ; sinpi8sqrt2
115    vqdmulh.s16     q9, q7, d0[2]
116    vqdmulh.s16     q10, q5, d0[0]          ; cospi8sqrt2minus1
117    vqdmulh.s16     q11, q7, d0[0]
118
119    vqadd.s16       q2, q4, q6             ; a1 = 0 + 2
120    vqsub.s16       q3, q4, q6             ; b1 = 0 - 2
121
122    ; see note on shifting above
123    vshr.s16        q10, q10, #1
124    vshr.s16        q11, q11, #1
125
126    ; q10: 1 + 1 * cospi : d1/temp1
127    ; q11: 3 + 3 * cospi : c1/temp2
128    vqadd.s16       q10, q5, q10
129    vqadd.s16       q11, q7, q11
130
131    ; q8: c1 = temp1 - temp2
132    ; q9: d1 = temp1 + temp2
133    vqsub.s16       q8, q8, q11
134    vqadd.s16       q9, q10, q9
135
136    ; a1+d1
137    ; b1+c1
138    ; b1-c1
139    ; a1-d1
140    vqadd.s16       q4, q2, q9
141    vqadd.s16       q5, q3, q8
142    vqsub.s16       q6, q3, q8
143    vqsub.s16       q7, q2, q9
144
145    ; +4 >> 3 (rounding)
146    vrshr.s16       q4, q4, #3              ; lo
147    vrshr.s16       q5, q5, #3
148    vrshr.s16       q6, q6, #3              ; hi
149    vrshr.s16       q7, q7, #3
150
151    vtrn.32         q4, q6
152    vtrn.32         q5, q7
153    vtrn.16         q4, q5
154    vtrn.16         q6, q7
155
156    ; adding pre
157    ; input is still packed. pre was read interleaved
158    vaddw.u8        q4, q4, d28
159    vaddw.u8        q5, q5, d29
160    vaddw.u8        q6, q6, d30
161    vaddw.u8        q7, q7, d31
162
163    vmov.i16        q14, #0
164    vmov            q15, q14
165    vst1.16         {q14, q15}, [r0]        ; write over high input
166    sub             r0, r0, #32
167    vst1.16         {q14, q15}, [r0]        ; write over low input
168
169    sub             r2, r2, r3, lsl #2      ; dst - 4*stride
170    add             r1, r2, #4              ; hi
171
172    ;saturate and narrow
173    vqmovun.s16     d0, q4                  ; lo
174    vqmovun.s16     d1, q5
175    vqmovun.s16     d2, q6                  ; hi
176    vqmovun.s16     d3, q7
177
178    vst1.32         {d0[0]}, [r2], r3       ; lo
179    vst1.32         {d0[1]}, [r1], r3       ; hi
180    vst1.32         {d1[0]}, [r2], r3
181    vst1.32         {d1[1]}, [r1], r3
182    vst1.32         {d2[0]}, [r2], r3
183    vst1.32         {d2[1]}, [r1], r3
184    vst1.32         {d3[0]}, [r2]
185    vst1.32         {d3[1]}, [r1]
186
187    bx             lr
188
189    ENDP           ; |idct_dequant_full_2x_neon|
190
191; Constant Pool
192cospi8sqrt2minus1 DCD 0x4e7b
193; because the lowest bit in 0x8a8c is 0, we can pre-shift this
194sinpi8sqrt2       DCD 0x4546
195
196    END
197