xref: /aosp_15_r20/external/libvpx/vp9/encoder/ppc/vp9_quantize_vsx.c (revision fb1b10ab9aebc7c7068eedab379b749d7e3900be)
1 /*
2  *  Copyright (c) 2018 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 #include "./vpx_config.h"
12 
13 #include "./vp9_rtcd.h"
14 #include "vpx_dsp/ppc/types_vsx.h"
15 
16 // Multiply the packed 16-bit integers in a and b, producing intermediate 32-bit
17 // integers, and return the high 16 bits of the intermediate integers.
18 // (a * b) >> 16
19 // Note: Because this is done in 2 operations, a and b cannot both be UINT16_MIN
vec_mulhi(int16x8_t a,int16x8_t b)20 static INLINE int16x8_t vec_mulhi(int16x8_t a, int16x8_t b) {
21   // madds does ((A * B) >> 15) + C, we need >> 16, so we perform an extra right
22   // shift.
23   return vec_sra(vec_madds(a, b, vec_zeros_s16), vec_ones_u16);
24 }
25 
26 // Negate 16-bit integers in a when the corresponding signed 16-bit
27 // integer in b is negative.
vec_sign(int16x8_t a,int16x8_t b)28 static INLINE int16x8_t vec_sign(int16x8_t a, int16x8_t b) {
29   const int16x8_t mask = vec_sra(b, vec_shift_sign_s16);
30   return vec_xor(vec_add(a, mask), mask);
31 }
32 
33 // Compare packed 16-bit integers across a, and return the maximum value in
34 // every element. Returns a vector containing the biggest value across vector a.
vec_max_across(int16x8_t a)35 static INLINE int16x8_t vec_max_across(int16x8_t a) {
36   a = vec_max(a, vec_perm(a, a, vec_perm64));
37   a = vec_max(a, vec_perm(a, a, vec_perm32));
38   return vec_max(a, vec_perm(a, a, vec_perm16));
39 }
40 
vp9_quantize_fp_vsx(const tran_low_t * coeff_ptr,intptr_t n_coeffs,const int16_t * round_ptr,const int16_t * quant_ptr,tran_low_t * qcoeff_ptr,tran_low_t * dqcoeff_ptr,const int16_t * dequant_ptr,uint16_t * eob_ptr,const int16_t * scan,const int16_t * iscan)41 void vp9_quantize_fp_vsx(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
42                          const int16_t *round_ptr, const int16_t *quant_ptr,
43                          tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
44                          const int16_t *dequant_ptr, uint16_t *eob_ptr,
45                          const int16_t *scan, const int16_t *iscan) {
46   int16x8_t qcoeff0, qcoeff1, dqcoeff0, dqcoeff1, eob;
47   bool16x8_t zero_coeff0, zero_coeff1;
48 
49   int16x8_t round = vec_vsx_ld(0, round_ptr);
50   int16x8_t quant = vec_vsx_ld(0, quant_ptr);
51   int16x8_t dequant = vec_vsx_ld(0, dequant_ptr);
52   int16x8_t coeff0 = vec_vsx_ld(0, coeff_ptr);
53   int16x8_t coeff1 = vec_vsx_ld(16, coeff_ptr);
54   int16x8_t scan0 = vec_vsx_ld(0, iscan);
55   int16x8_t scan1 = vec_vsx_ld(16, iscan);
56 
57   (void)scan;
58 
59   // First set of 8 coeff starts with DC + 7 AC
60   qcoeff0 = vec_mulhi(vec_vaddshs(vec_abs(coeff0), round), quant);
61   zero_coeff0 = vec_cmpeq(qcoeff0, vec_zeros_s16);
62   qcoeff0 = vec_sign(qcoeff0, coeff0);
63   vec_vsx_st(qcoeff0, 0, qcoeff_ptr);
64 
65   dqcoeff0 = vec_mladd(qcoeff0, dequant, vec_zeros_s16);
66   vec_vsx_st(dqcoeff0, 0, dqcoeff_ptr);
67 
68   // Remove DC value from round and quant
69   round = vec_splat(round, 1);
70   quant = vec_splat(quant, 1);
71 
72   // Remove DC value from dequant
73   dequant = vec_splat(dequant, 1);
74 
75   // Second set of 8 coeff starts with (all AC)
76   qcoeff1 = vec_mulhi(vec_vaddshs(vec_abs(coeff1), round), quant);
77   zero_coeff1 = vec_cmpeq(qcoeff1, vec_zeros_s16);
78   qcoeff1 = vec_sign(qcoeff1, coeff1);
79   vec_vsx_st(qcoeff1, 16, qcoeff_ptr);
80 
81   dqcoeff1 = vec_mladd(qcoeff1, dequant, vec_zeros_s16);
82   vec_vsx_st(dqcoeff1, 16, dqcoeff_ptr);
83 
84   eob = vec_max(vec_or(scan0, zero_coeff0), vec_or(scan1, zero_coeff1));
85 
86   // We quantize 16 coeff up front (enough for a 4x4) and process 24 coeff per
87   // loop iteration.
88   // for 8x8: 16 + 2 x 24 = 64
89   // for 16x16: 16 + 10 x 24 = 256
90   if (n_coeffs > 16) {
91     int16x8_t coeff2, qcoeff2, dqcoeff2, eob2, scan2;
92     bool16x8_t zero_coeff2;
93 
94     int index = 16;
95     int off0 = 32;
96     int off1 = 48;
97     int off2 = 64;
98 
99     do {
100       coeff0 = vec_vsx_ld(off0, coeff_ptr);
101       coeff1 = vec_vsx_ld(off1, coeff_ptr);
102       coeff2 = vec_vsx_ld(off2, coeff_ptr);
103       scan0 = vec_vsx_ld(off0, iscan);
104       scan1 = vec_vsx_ld(off1, iscan);
105       scan2 = vec_vsx_ld(off2, iscan);
106 
107       qcoeff0 = vec_mulhi(vec_vaddshs(vec_abs(coeff0), round), quant);
108       zero_coeff0 = vec_cmpeq(qcoeff0, vec_zeros_s16);
109       qcoeff0 = vec_sign(qcoeff0, coeff0);
110       vec_vsx_st(qcoeff0, off0, qcoeff_ptr);
111       dqcoeff0 = vec_mladd(qcoeff0, dequant, vec_zeros_s16);
112       vec_vsx_st(dqcoeff0, off0, dqcoeff_ptr);
113 
114       qcoeff1 = vec_mulhi(vec_vaddshs(vec_abs(coeff1), round), quant);
115       zero_coeff1 = vec_cmpeq(qcoeff1, vec_zeros_s16);
116       qcoeff1 = vec_sign(qcoeff1, coeff1);
117       vec_vsx_st(qcoeff1, off1, qcoeff_ptr);
118       dqcoeff1 = vec_mladd(qcoeff1, dequant, vec_zeros_s16);
119       vec_vsx_st(dqcoeff1, off1, dqcoeff_ptr);
120 
121       qcoeff2 = vec_mulhi(vec_vaddshs(vec_abs(coeff2), round), quant);
122       zero_coeff2 = vec_cmpeq(qcoeff2, vec_zeros_s16);
123       qcoeff2 = vec_sign(qcoeff2, coeff2);
124       vec_vsx_st(qcoeff2, off2, qcoeff_ptr);
125       dqcoeff2 = vec_mladd(qcoeff2, dequant, vec_zeros_s16);
126       vec_vsx_st(dqcoeff2, off2, dqcoeff_ptr);
127 
128       eob = vec_max(eob, vec_or(scan0, zero_coeff0));
129       eob2 = vec_max(vec_or(scan1, zero_coeff1), vec_or(scan2, zero_coeff2));
130       eob = vec_max(eob, eob2);
131 
132       index += 24;
133       off0 += 48;
134       off1 += 48;
135       off2 += 48;
136     } while (index < n_coeffs);
137   }
138 
139   eob = vec_max_across(eob);
140   *eob_ptr = eob[0] + 1;
141 }
142 
143 // Sets the value of a 32-bit integers to 1 when the corresponding value in a is
144 // negative.
vec_is_neg(int32x4_t a)145 static INLINE int32x4_t vec_is_neg(int32x4_t a) {
146   return vec_sr(a, vec_shift_sign_s32);
147 }
148 
149 // DeQuantization function used for 32x32 blocks. Quantized coeff of 32x32
150 // blocks are twice as big as for other block sizes. As such, using
151 // vec_mladd results in overflow.
dequantize_coeff_32(int16x8_t qcoeff,int16x8_t dequant)152 static INLINE int16x8_t dequantize_coeff_32(int16x8_t qcoeff,
153                                             int16x8_t dequant) {
154   int32x4_t dqcoeffe = vec_mule(qcoeff, dequant);
155   int32x4_t dqcoeffo = vec_mulo(qcoeff, dequant);
156   // Add 1 if negative to round towards zero because the C uses division.
157   dqcoeffe = vec_add(dqcoeffe, vec_is_neg(dqcoeffe));
158   dqcoeffo = vec_add(dqcoeffo, vec_is_neg(dqcoeffo));
159   dqcoeffe = vec_sra(dqcoeffe, vec_ones_u32);
160   dqcoeffo = vec_sra(dqcoeffo, vec_ones_u32);
161   return (int16x8_t)vec_perm(dqcoeffe, dqcoeffo, vec_perm_odd_even_pack);
162 }
163 
vp9_quantize_fp_32x32_vsx(const tran_low_t * coeff_ptr,intptr_t n_coeffs,const int16_t * round_ptr,const int16_t * quant_ptr,tran_low_t * qcoeff_ptr,tran_low_t * dqcoeff_ptr,const int16_t * dequant_ptr,uint16_t * eob_ptr,const int16_t * scan,const int16_t * iscan)164 void vp9_quantize_fp_32x32_vsx(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
165                                const int16_t *round_ptr,
166                                const int16_t *quant_ptr, tran_low_t *qcoeff_ptr,
167                                tran_low_t *dqcoeff_ptr,
168                                const int16_t *dequant_ptr, uint16_t *eob_ptr,
169                                const int16_t *scan, const int16_t *iscan) {
170   // In stage 1, we quantize 16 coeffs (DC + 15 AC)
171   // In stage 2, we loop 42 times and quantize 24 coeffs per iteration
172   // (32 * 32 - 16) / 24 = 42
173   int num_itr = 42;
174   // Offsets are in bytes, 16 coeffs = 32 bytes
175   int off0 = 32;
176   int off1 = 48;
177   int off2 = 64;
178 
179   int16x8_t qcoeff0, qcoeff1, dqcoeff0, dqcoeff1, eob;
180   bool16x8_t mask0, mask1, zero_coeff0, zero_coeff1;
181 
182   int16x8_t round = vec_vsx_ld(0, round_ptr);
183   int16x8_t quant = vec_vsx_ld(0, quant_ptr);
184   int16x8_t dequant = vec_vsx_ld(0, dequant_ptr);
185   int16x8_t coeff0 = vec_vsx_ld(0, coeff_ptr);
186   int16x8_t coeff1 = vec_vsx_ld(16, coeff_ptr);
187   int16x8_t scan0 = vec_vsx_ld(0, iscan);
188   int16x8_t scan1 = vec_vsx_ld(16, iscan);
189   int16x8_t thres = vec_sra(dequant, vec_splats((uint16_t)2));
190   int16x8_t abs_coeff0 = vec_abs(coeff0);
191   int16x8_t abs_coeff1 = vec_abs(coeff1);
192 
193   (void)scan;
194   (void)n_coeffs;
195 
196   mask0 = vec_cmpge(abs_coeff0, thres);
197   round = vec_sra(vec_add(round, vec_ones_s16), vec_ones_u16);
198   // First set of 8 coeff starts with DC + 7 AC
199   qcoeff0 = vec_madds(vec_vaddshs(abs_coeff0, round), quant, vec_zeros_s16);
200   qcoeff0 = vec_and(qcoeff0, mask0);
201   zero_coeff0 = vec_cmpeq(qcoeff0, vec_zeros_s16);
202   qcoeff0 = vec_sign(qcoeff0, coeff0);
203   vec_vsx_st(qcoeff0, 0, qcoeff_ptr);
204 
205   dqcoeff0 = dequantize_coeff_32(qcoeff0, dequant);
206   vec_vsx_st(dqcoeff0, 0, dqcoeff_ptr);
207 
208   // Remove DC value from thres, round, quant and dequant
209   thres = vec_splat(thres, 1);
210   round = vec_splat(round, 1);
211   quant = vec_splat(quant, 1);
212   dequant = vec_splat(dequant, 1);
213 
214   mask1 = vec_cmpge(abs_coeff1, thres);
215 
216   // Second set of 8 coeff starts with (all AC)
217   qcoeff1 =
218       vec_madds(vec_vaddshs(vec_abs(coeff1), round), quant, vec_zeros_s16);
219   qcoeff1 = vec_and(qcoeff1, mask1);
220   zero_coeff1 = vec_cmpeq(qcoeff1, vec_zeros_s16);
221   qcoeff1 = vec_sign(qcoeff1, coeff1);
222   vec_vsx_st(qcoeff1, 16, qcoeff_ptr);
223 
224   dqcoeff1 = dequantize_coeff_32(qcoeff1, dequant);
225   vec_vsx_st(dqcoeff1, 16, dqcoeff_ptr);
226 
227   eob = vec_max(vec_or(scan0, zero_coeff0), vec_or(scan1, zero_coeff1));
228 
229   do {
230     int16x8_t coeff2, abs_coeff2, qcoeff2, dqcoeff2, eob2, scan2;
231     bool16x8_t zero_coeff2, mask2;
232     coeff0 = vec_vsx_ld(off0, coeff_ptr);
233     coeff1 = vec_vsx_ld(off1, coeff_ptr);
234     coeff2 = vec_vsx_ld(off2, coeff_ptr);
235     scan0 = vec_vsx_ld(off0, iscan);
236     scan1 = vec_vsx_ld(off1, iscan);
237     scan2 = vec_vsx_ld(off2, iscan);
238 
239     abs_coeff0 = vec_abs(coeff0);
240     abs_coeff1 = vec_abs(coeff1);
241     abs_coeff2 = vec_abs(coeff2);
242 
243     qcoeff0 = vec_madds(vec_vaddshs(abs_coeff0, round), quant, vec_zeros_s16);
244     qcoeff1 = vec_madds(vec_vaddshs(abs_coeff1, round), quant, vec_zeros_s16);
245     qcoeff2 = vec_madds(vec_vaddshs(abs_coeff2, round), quant, vec_zeros_s16);
246 
247     mask0 = vec_cmpge(abs_coeff0, thres);
248     mask1 = vec_cmpge(abs_coeff1, thres);
249     mask2 = vec_cmpge(abs_coeff2, thres);
250 
251     qcoeff0 = vec_and(qcoeff0, mask0);
252     qcoeff1 = vec_and(qcoeff1, mask1);
253     qcoeff2 = vec_and(qcoeff2, mask2);
254 
255     zero_coeff0 = vec_cmpeq(qcoeff0, vec_zeros_s16);
256     zero_coeff1 = vec_cmpeq(qcoeff1, vec_zeros_s16);
257     zero_coeff2 = vec_cmpeq(qcoeff2, vec_zeros_s16);
258 
259     qcoeff0 = vec_sign(qcoeff0, coeff0);
260     qcoeff1 = vec_sign(qcoeff1, coeff1);
261     qcoeff2 = vec_sign(qcoeff2, coeff2);
262 
263     vec_vsx_st(qcoeff0, off0, qcoeff_ptr);
264     vec_vsx_st(qcoeff1, off1, qcoeff_ptr);
265     vec_vsx_st(qcoeff2, off2, qcoeff_ptr);
266 
267     dqcoeff0 = dequantize_coeff_32(qcoeff0, dequant);
268     dqcoeff1 = dequantize_coeff_32(qcoeff1, dequant);
269     dqcoeff2 = dequantize_coeff_32(qcoeff2, dequant);
270 
271     vec_vsx_st(dqcoeff0, off0, dqcoeff_ptr);
272     vec_vsx_st(dqcoeff1, off1, dqcoeff_ptr);
273     vec_vsx_st(dqcoeff2, off2, dqcoeff_ptr);
274 
275     eob = vec_max(eob, vec_or(scan0, zero_coeff0));
276     eob2 = vec_max(vec_or(scan1, zero_coeff1), vec_or(scan2, zero_coeff2));
277     eob = vec_max(eob, eob2);
278 
279     off0 += 48;
280     off1 += 48;
281     off2 += 48;
282     num_itr--;
283   } while (num_itr != 0);
284 
285   eob = vec_max_across(eob);
286   *eob_ptr = eob[0] + 1;
287 }
288