1/*
2 *  Copyright (c) 2017 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 <arm_neon.h>
12#include <assert.h>
13#include <string.h>
14
15#include "./vpx_config.h"
16#include "./vpx_dsp_rtcd.h"
17#include "vpx/vpx_integer.h"
18#include "vpx_dsp/arm/transpose_neon.h"
19#include "vpx_dsp/arm/vpx_convolve8_neon.h"
20#include "vpx_ports/mem.h"
21
22static INLINE void scaledconvolve_horiz_w4(
23    const uint8_t *src, const ptrdiff_t src_stride, uint8_t *dst,
24    const ptrdiff_t dst_stride, const InterpKernel *const x_filters,
25    const int x0_q4, const int x_step_q4, const int w, const int h) {
26  DECLARE_ALIGNED(16, uint8_t, temp[4 * 4]);
27  int x, y, z;
28
29  src -= SUBPEL_TAPS / 2 - 1;
30
31  y = h;
32  do {
33    int x_q4 = x0_q4;
34    x = 0;
35    do {
36      // process 4 src_x steps
37      for (z = 0; z < 4; ++z) {
38        const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
39        if (x_q4 & SUBPEL_MASK) {
40          const int16x8_t filters = vld1q_s16(x_filters[x_q4 & SUBPEL_MASK]);
41          const int16x4_t filter3 = vdup_lane_s16(vget_low_s16(filters), 3);
42          const int16x4_t filter4 = vdup_lane_s16(vget_high_s16(filters), 0);
43          uint8x8_t s[8], d;
44          int16x8_t ss[4];
45          int16x4_t t[8], tt;
46
47          load_u8_8x4(src_x, src_stride, &s[0], &s[1], &s[2], &s[3]);
48          transpose_u8_8x4(&s[0], &s[1], &s[2], &s[3]);
49
50          ss[0] = vreinterpretq_s16_u16(vmovl_u8(s[0]));
51          ss[1] = vreinterpretq_s16_u16(vmovl_u8(s[1]));
52          ss[2] = vreinterpretq_s16_u16(vmovl_u8(s[2]));
53          ss[3] = vreinterpretq_s16_u16(vmovl_u8(s[3]));
54          t[0] = vget_low_s16(ss[0]);
55          t[1] = vget_low_s16(ss[1]);
56          t[2] = vget_low_s16(ss[2]);
57          t[3] = vget_low_s16(ss[3]);
58          t[4] = vget_high_s16(ss[0]);
59          t[5] = vget_high_s16(ss[1]);
60          t[6] = vget_high_s16(ss[2]);
61          t[7] = vget_high_s16(ss[3]);
62
63          tt = convolve8_4(t[0], t[1], t[2], t[3], t[4], t[5], t[6], t[7],
64                           filters, filter3, filter4);
65          d = vqrshrun_n_s16(vcombine_s16(tt, tt), 7);
66          vst1_lane_u32((uint32_t *)&temp[4 * z], vreinterpret_u32_u8(d), 0);
67        } else {
68          int i;
69          for (i = 0; i < 4; ++i) {
70            temp[z * 4 + i] = src_x[i * src_stride + 3];
71          }
72        }
73        x_q4 += x_step_q4;
74      }
75
76      // transpose the 4x4 filters values back to dst
77      {
78        const uint8x8x4_t d4 = vld4_u8(temp);
79        vst1_lane_u32((uint32_t *)&dst[x + 0 * dst_stride],
80                      vreinterpret_u32_u8(d4.val[0]), 0);
81        vst1_lane_u32((uint32_t *)&dst[x + 1 * dst_stride],
82                      vreinterpret_u32_u8(d4.val[1]), 0);
83        vst1_lane_u32((uint32_t *)&dst[x + 2 * dst_stride],
84                      vreinterpret_u32_u8(d4.val[2]), 0);
85        vst1_lane_u32((uint32_t *)&dst[x + 3 * dst_stride],
86                      vreinterpret_u32_u8(d4.val[3]), 0);
87      }
88      x += 4;
89    } while (x < w);
90
91    src += src_stride * 4;
92    dst += dst_stride * 4;
93    y -= 4;
94  } while (y > 0);
95}
96
97static INLINE void scaledconvolve_horiz_w8(
98    const uint8_t *src, const ptrdiff_t src_stride, uint8_t *dst,
99    const ptrdiff_t dst_stride, const InterpKernel *const x_filters,
100    const int x0_q4, const int x_step_q4, const int w, const int h) {
101  DECLARE_ALIGNED(16, uint8_t, temp[8 * 8]);
102  int x, y, z;
103  src -= SUBPEL_TAPS / 2 - 1;
104
105  // This function processes 8x8 areas. The intermediate height is not always
106  // a multiple of 8, so force it to be a multiple of 8 here.
107  y = (h + 7) & ~7;
108
109  do {
110    int x_q4 = x0_q4;
111    x = 0;
112    do {
113      uint8x8_t d[8];
114      // process 8 src_x steps
115      for (z = 0; z < 8; ++z) {
116        const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
117
118        if (x_q4 & SUBPEL_MASK) {
119          const int16x8_t filters = vld1q_s16(x_filters[x_q4 & SUBPEL_MASK]);
120          uint8x8_t s[8];
121          load_u8_8x8(src_x, src_stride, &s[0], &s[1], &s[2], &s[3], &s[4],
122                      &s[5], &s[6], &s[7]);
123          transpose_u8_8x8(&s[0], &s[1], &s[2], &s[3], &s[4], &s[5], &s[6],
124                           &s[7]);
125          d[0] = scale_filter_8(s, filters);
126          vst1_u8(&temp[8 * z], d[0]);
127        } else {
128          int i;
129          for (i = 0; i < 8; ++i) {
130            temp[z * 8 + i] = src_x[i * src_stride + 3];
131          }
132        }
133        x_q4 += x_step_q4;
134      }
135
136      // transpose the 8x8 filters values back to dst
137      load_u8_8x8(temp, 8, &d[0], &d[1], &d[2], &d[3], &d[4], &d[5], &d[6],
138                  &d[7]);
139      transpose_u8_8x8(&d[0], &d[1], &d[2], &d[3], &d[4], &d[5], &d[6], &d[7]);
140      vst1_u8(&dst[x + 0 * dst_stride], d[0]);
141      vst1_u8(&dst[x + 1 * dst_stride], d[1]);
142      vst1_u8(&dst[x + 2 * dst_stride], d[2]);
143      vst1_u8(&dst[x + 3 * dst_stride], d[3]);
144      vst1_u8(&dst[x + 4 * dst_stride], d[4]);
145      vst1_u8(&dst[x + 5 * dst_stride], d[5]);
146      vst1_u8(&dst[x + 6 * dst_stride], d[6]);
147      vst1_u8(&dst[x + 7 * dst_stride], d[7]);
148      x += 8;
149    } while (x < w);
150
151    src += src_stride * 8;
152    dst += dst_stride * 8;
153  } while (y -= 8);
154}
155
156static INLINE void scaledconvolve_vert_w4(
157    const uint8_t *src, const ptrdiff_t src_stride, uint8_t *dst,
158    const ptrdiff_t dst_stride, const InterpKernel *const y_filters,
159    const int y0_q4, const int y_step_q4, const int w, const int h) {
160  int y;
161  int y_q4 = y0_q4;
162
163  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
164  y = h;
165  do {
166    const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
167
168    if (y_q4 & SUBPEL_MASK) {
169      const int16x8_t filters = vld1q_s16(y_filters[y_q4 & SUBPEL_MASK]);
170      const int16x4_t filter3 = vdup_lane_s16(vget_low_s16(filters), 3);
171      const int16x4_t filter4 = vdup_lane_s16(vget_high_s16(filters), 0);
172      uint8x8_t s[8], d;
173      int16x4_t t[8], tt;
174
175      load_u8_8x8(src_y, src_stride, &s[0], &s[1], &s[2], &s[3], &s[4], &s[5],
176                  &s[6], &s[7]);
177      t[0] = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(s[0])));
178      t[1] = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(s[1])));
179      t[2] = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(s[2])));
180      t[3] = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(s[3])));
181      t[4] = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(s[4])));
182      t[5] = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(s[5])));
183      t[6] = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(s[6])));
184      t[7] = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(s[7])));
185
186      tt = convolve8_4(t[0], t[1], t[2], t[3], t[4], t[5], t[6], t[7], filters,
187                       filter3, filter4);
188      d = vqrshrun_n_s16(vcombine_s16(tt, tt), 7);
189      vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d), 0);
190    } else {
191      memcpy(dst, &src_y[3 * src_stride], w);
192    }
193
194    dst += dst_stride;
195    y_q4 += y_step_q4;
196  } while (--y);
197}
198
199static INLINE void scaledconvolve_vert_w8(
200    const uint8_t *src, const ptrdiff_t src_stride, uint8_t *dst,
201    const ptrdiff_t dst_stride, const InterpKernel *const y_filters,
202    const int y0_q4, const int y_step_q4, const int w, const int h) {
203  int y;
204  int y_q4 = y0_q4;
205
206  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
207  y = h;
208  do {
209    const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
210    if (y_q4 & SUBPEL_MASK) {
211      const int16x8_t filters = vld1q_s16(y_filters[y_q4 & SUBPEL_MASK]);
212      uint8x8_t s[8], d;
213      load_u8_8x8(src_y, src_stride, &s[0], &s[1], &s[2], &s[3], &s[4], &s[5],
214                  &s[6], &s[7]);
215      d = scale_filter_8(s, filters);
216      vst1_u8(dst, d);
217    } else {
218      memcpy(dst, &src_y[3 * src_stride], w);
219    }
220    dst += dst_stride;
221    y_q4 += y_step_q4;
222  } while (--y);
223}
224
225static INLINE void scaledconvolve_vert_w16(
226    const uint8_t *src, const ptrdiff_t src_stride, uint8_t *dst,
227    const ptrdiff_t dst_stride, const InterpKernel *const y_filters,
228    const int y0_q4, const int y_step_q4, const int w, const int h) {
229  int x, y;
230  int y_q4 = y0_q4;
231
232  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
233  y = h;
234  do {
235    const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
236    if (y_q4 & SUBPEL_MASK) {
237      x = 0;
238      do {
239        const int16x8_t filters = vld1q_s16(y_filters[y_q4 & SUBPEL_MASK]);
240        uint8x16_t ss[8];
241        uint8x8_t s[8], d[2];
242        load_u8_16x8(src_y, src_stride, &ss[0], &ss[1], &ss[2], &ss[3], &ss[4],
243                     &ss[5], &ss[6], &ss[7]);
244        s[0] = vget_low_u8(ss[0]);
245        s[1] = vget_low_u8(ss[1]);
246        s[2] = vget_low_u8(ss[2]);
247        s[3] = vget_low_u8(ss[3]);
248        s[4] = vget_low_u8(ss[4]);
249        s[5] = vget_low_u8(ss[5]);
250        s[6] = vget_low_u8(ss[6]);
251        s[7] = vget_low_u8(ss[7]);
252        d[0] = scale_filter_8(s, filters);
253
254        s[0] = vget_high_u8(ss[0]);
255        s[1] = vget_high_u8(ss[1]);
256        s[2] = vget_high_u8(ss[2]);
257        s[3] = vget_high_u8(ss[3]);
258        s[4] = vget_high_u8(ss[4]);
259        s[5] = vget_high_u8(ss[5]);
260        s[6] = vget_high_u8(ss[6]);
261        s[7] = vget_high_u8(ss[7]);
262        d[1] = scale_filter_8(s, filters);
263        vst1q_u8(&dst[x], vcombine_u8(d[0], d[1]));
264        src_y += 16;
265        x += 16;
266      } while (x < w);
267    } else {
268      memcpy(dst, &src_y[3 * src_stride], w);
269    }
270    dst += dst_stride;
271    y_q4 += y_step_q4;
272  } while (--y);
273}
274
275void vpx_scaled_2d_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
276                        ptrdiff_t dst_stride, const InterpKernel *filter,
277                        int x0_q4, int x_step_q4, int y0_q4, int y_step_q4,
278                        int w, int h) {
279  // Note: Fixed size intermediate buffer, temp, places limits on parameters.
280  // 2d filtering proceeds in 2 steps:
281  //   (1) Interpolate horizontally into an intermediate buffer, temp.
282  //   (2) Interpolate temp vertically to derive the sub-pixel result.
283  // Deriving the maximum number of rows in the temp buffer (135):
284  // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
285  // --Largest block size is 64x64 pixels.
286  // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
287  //   original frame (in 1/16th pixel units).
288  // --Must round-up because block may be located at sub-pixel position.
289  // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
290  // --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
291  // --Require an additional 8 rows for the horiz_w8 transpose tail.
292  // When calling in frame scaling function, the smallest scaling factor is x1/4
293  // ==> y_step_q4 = 64. Since w and h are at most 16, the temp buffer is still
294  // big enough.
295  DECLARE_ALIGNED(16, uint8_t, temp[(135 + 8) * 64]);
296  const int intermediate_height =
297      (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
298
299  assert(w <= 64);
300  assert(h <= 64);
301  assert(y_step_q4 <= 32 || (y_step_q4 <= 64 && h <= 32));
302  assert(x_step_q4 <= 64);
303
304  if (w >= 8) {
305    scaledconvolve_horiz_w8(src - src_stride * (SUBPEL_TAPS / 2 - 1),
306                            src_stride, temp, 64, filter, x0_q4, x_step_q4, w,
307                            intermediate_height);
308  } else {
309    scaledconvolve_horiz_w4(src - src_stride * (SUBPEL_TAPS / 2 - 1),
310                            src_stride, temp, 64, filter, x0_q4, x_step_q4, w,
311                            intermediate_height);
312  }
313
314  if (w >= 16) {
315    scaledconvolve_vert_w16(temp + 64 * (SUBPEL_TAPS / 2 - 1), 64, dst,
316                            dst_stride, filter, y0_q4, y_step_q4, w, h);
317  } else if (w == 8) {
318    scaledconvolve_vert_w8(temp + 64 * (SUBPEL_TAPS / 2 - 1), 64, dst,
319                           dst_stride, filter, y0_q4, y_step_q4, w, h);
320  } else {
321    scaledconvolve_vert_w4(temp + 64 * (SUBPEL_TAPS / 2 - 1), 64, dst,
322                           dst_stride, filter, y0_q4, y_step_q4, w, h);
323  }
324}
325