1 /*
2 * Copyright (c) 2015 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 #ifndef VPX_DSP_X86_HIGHBD_INV_TXFM_SSE2_H_
12 #define VPX_DSP_X86_HIGHBD_INV_TXFM_SSE2_H_
13
14 #include <emmintrin.h> // SSE2
15
16 #include "./vpx_config.h"
17 #include "vpx/vpx_integer.h"
18 #include "vpx_dsp/inv_txfm.h"
19 #include "vpx_dsp/x86/transpose_sse2.h"
20 #include "vpx_dsp/x86/txfm_common_sse2.h"
21
extend_64bit(const __m128i in,__m128i * const out)22 static INLINE void extend_64bit(const __m128i in,
23 __m128i *const out /*out[2]*/) {
24 out[0] = _mm_unpacklo_epi32(in, in); // 0, 0, 1, 1
25 out[1] = _mm_unpackhi_epi32(in, in); // 2, 2, 3, 3
26 }
27
wraplow_16bit_shift4(const __m128i in0,const __m128i in1,const __m128i rounding)28 static INLINE __m128i wraplow_16bit_shift4(const __m128i in0, const __m128i in1,
29 const __m128i rounding) {
30 __m128i temp[2];
31 temp[0] = _mm_add_epi32(in0, rounding);
32 temp[1] = _mm_add_epi32(in1, rounding);
33 temp[0] = _mm_srai_epi32(temp[0], 4);
34 temp[1] = _mm_srai_epi32(temp[1], 4);
35 return _mm_packs_epi32(temp[0], temp[1]);
36 }
37
wraplow_16bit_shift5(const __m128i in0,const __m128i in1,const __m128i rounding)38 static INLINE __m128i wraplow_16bit_shift5(const __m128i in0, const __m128i in1,
39 const __m128i rounding) {
40 __m128i temp[2];
41 temp[0] = _mm_add_epi32(in0, rounding);
42 temp[1] = _mm_add_epi32(in1, rounding);
43 temp[0] = _mm_srai_epi32(temp[0], 5);
44 temp[1] = _mm_srai_epi32(temp[1], 5);
45 return _mm_packs_epi32(temp[0], temp[1]);
46 }
47
dct_const_round_shift_64bit(const __m128i in)48 static INLINE __m128i dct_const_round_shift_64bit(const __m128i in) {
49 const __m128i t =
50 _mm_add_epi64(in, pair_set_epi32(DCT_CONST_ROUNDING << 2, 0));
51 return _mm_srli_si128(t, 2);
52 }
53
pack_4(const __m128i in0,const __m128i in1)54 static INLINE __m128i pack_4(const __m128i in0, const __m128i in1) {
55 const __m128i t0 = _mm_unpacklo_epi32(in0, in1); // 0, 2
56 const __m128i t1 = _mm_unpackhi_epi32(in0, in1); // 1, 3
57 return _mm_unpacklo_epi32(t0, t1); // 0, 1, 2, 3
58 }
59
abs_extend_64bit_sse2(const __m128i in,__m128i * const out,__m128i * const sign)60 static INLINE void abs_extend_64bit_sse2(const __m128i in,
61 __m128i *const out /*out[2]*/,
62 __m128i *const sign /*sign[2]*/) {
63 sign[0] = _mm_srai_epi32(in, 31);
64 out[0] = _mm_xor_si128(in, sign[0]);
65 out[0] = _mm_sub_epi32(out[0], sign[0]);
66 sign[1] = _mm_unpackhi_epi32(sign[0], sign[0]); // 64-bit sign of 2, 3
67 sign[0] = _mm_unpacklo_epi32(sign[0], sign[0]); // 64-bit sign of 0, 1
68 out[1] = _mm_unpackhi_epi32(out[0], out[0]); // 2, 3
69 out[0] = _mm_unpacklo_epi32(out[0], out[0]); // 0, 1
70 }
71
72 // Note: cospi must be non negative.
multiply_apply_sign_sse2(const __m128i in,const __m128i sign,const __m128i cospi)73 static INLINE __m128i multiply_apply_sign_sse2(const __m128i in,
74 const __m128i sign,
75 const __m128i cospi) {
76 __m128i out = _mm_mul_epu32(in, cospi);
77 out = _mm_xor_si128(out, sign);
78 return _mm_sub_epi64(out, sign);
79 }
80
81 // Note: c must be non negative.
multiplication_round_shift_sse2(const __m128i * const in,const __m128i * const sign,const int c)82 static INLINE __m128i multiplication_round_shift_sse2(
83 const __m128i *const in /*in[2]*/, const __m128i *const sign /*sign[2]*/,
84 const int c) {
85 const __m128i pair_c = pair_set_epi32(c << 2, 0);
86 __m128i t0, t1;
87
88 assert(c >= 0);
89 t0 = multiply_apply_sign_sse2(in[0], sign[0], pair_c);
90 t1 = multiply_apply_sign_sse2(in[1], sign[1], pair_c);
91 t0 = dct_const_round_shift_64bit(t0);
92 t1 = dct_const_round_shift_64bit(t1);
93
94 return pack_4(t0, t1);
95 }
96
97 // Note: c must be non negative.
multiplication_neg_round_shift_sse2(const __m128i * const in,const __m128i * const sign,const int c)98 static INLINE __m128i multiplication_neg_round_shift_sse2(
99 const __m128i *const in /*in[2]*/, const __m128i *const sign /*sign[2]*/,
100 const int c) {
101 const __m128i pair_c = pair_set_epi32(c << 2, 0);
102 __m128i t0, t1;
103
104 assert(c >= 0);
105 t0 = multiply_apply_sign_sse2(in[0], sign[0], pair_c);
106 t1 = multiply_apply_sign_sse2(in[1], sign[1], pair_c);
107 t0 = _mm_sub_epi64(_mm_setzero_si128(), t0);
108 t1 = _mm_sub_epi64(_mm_setzero_si128(), t1);
109 t0 = dct_const_round_shift_64bit(t0);
110 t1 = dct_const_round_shift_64bit(t1);
111
112 return pack_4(t0, t1);
113 }
114
115 // Note: c0 and c1 must be non negative.
highbd_butterfly_sse2(const __m128i in0,const __m128i in1,const int c0,const int c1,__m128i * const out0,__m128i * const out1)116 static INLINE void highbd_butterfly_sse2(const __m128i in0, const __m128i in1,
117 const int c0, const int c1,
118 __m128i *const out0,
119 __m128i *const out1) {
120 const __m128i pair_c0 = pair_set_epi32(c0 << 2, 0);
121 const __m128i pair_c1 = pair_set_epi32(c1 << 2, 0);
122 __m128i temp1[4], temp2[4], sign1[2], sign2[2];
123
124 assert(c0 >= 0);
125 assert(c1 >= 0);
126 abs_extend_64bit_sse2(in0, temp1, sign1);
127 abs_extend_64bit_sse2(in1, temp2, sign2);
128 temp1[2] = multiply_apply_sign_sse2(temp1[0], sign1[0], pair_c1);
129 temp1[3] = multiply_apply_sign_sse2(temp1[1], sign1[1], pair_c1);
130 temp1[0] = multiply_apply_sign_sse2(temp1[0], sign1[0], pair_c0);
131 temp1[1] = multiply_apply_sign_sse2(temp1[1], sign1[1], pair_c0);
132 temp2[2] = multiply_apply_sign_sse2(temp2[0], sign2[0], pair_c0);
133 temp2[3] = multiply_apply_sign_sse2(temp2[1], sign2[1], pair_c0);
134 temp2[0] = multiply_apply_sign_sse2(temp2[0], sign2[0], pair_c1);
135 temp2[1] = multiply_apply_sign_sse2(temp2[1], sign2[1], pair_c1);
136 temp1[0] = _mm_sub_epi64(temp1[0], temp2[0]);
137 temp1[1] = _mm_sub_epi64(temp1[1], temp2[1]);
138 temp2[0] = _mm_add_epi64(temp1[2], temp2[2]);
139 temp2[1] = _mm_add_epi64(temp1[3], temp2[3]);
140 temp1[0] = dct_const_round_shift_64bit(temp1[0]);
141 temp1[1] = dct_const_round_shift_64bit(temp1[1]);
142 temp2[0] = dct_const_round_shift_64bit(temp2[0]);
143 temp2[1] = dct_const_round_shift_64bit(temp2[1]);
144 *out0 = pack_4(temp1[0], temp1[1]);
145 *out1 = pack_4(temp2[0], temp2[1]);
146 }
147
148 // Note: c0 and c1 must be non negative.
highbd_partial_butterfly_sse2(const __m128i in,const int c0,const int c1,__m128i * const out0,__m128i * const out1)149 static INLINE void highbd_partial_butterfly_sse2(const __m128i in, const int c0,
150 const int c1,
151 __m128i *const out0,
152 __m128i *const out1) {
153 __m128i temp[2], sign[2];
154
155 assert(c0 >= 0);
156 assert(c1 >= 0);
157 abs_extend_64bit_sse2(in, temp, sign);
158 *out0 = multiplication_round_shift_sse2(temp, sign, c0);
159 *out1 = multiplication_round_shift_sse2(temp, sign, c1);
160 }
161
162 // Note: c0 and c1 must be non negative.
highbd_partial_butterfly_neg_sse2(const __m128i in,const int c0,const int c1,__m128i * const out0,__m128i * const out1)163 static INLINE void highbd_partial_butterfly_neg_sse2(const __m128i in,
164 const int c0, const int c1,
165 __m128i *const out0,
166 __m128i *const out1) {
167 __m128i temp[2], sign[2];
168
169 assert(c0 >= 0);
170 assert(c1 >= 0);
171 abs_extend_64bit_sse2(in, temp, sign);
172 *out0 = multiplication_neg_round_shift_sse2(temp, sign, c1);
173 *out1 = multiplication_round_shift_sse2(temp, sign, c0);
174 }
175
highbd_butterfly_cospi16_sse2(const __m128i in0,const __m128i in1,__m128i * const out0,__m128i * const out1)176 static INLINE void highbd_butterfly_cospi16_sse2(const __m128i in0,
177 const __m128i in1,
178 __m128i *const out0,
179 __m128i *const out1) {
180 __m128i temp1[2], temp2, sign[2];
181
182 temp2 = _mm_add_epi32(in0, in1);
183 abs_extend_64bit_sse2(temp2, temp1, sign);
184 *out0 = multiplication_round_shift_sse2(temp1, sign, cospi_16_64);
185 temp2 = _mm_sub_epi32(in0, in1);
186 abs_extend_64bit_sse2(temp2, temp1, sign);
187 *out1 = multiplication_round_shift_sse2(temp1, sign, cospi_16_64);
188 }
189
190 // Only do addition and subtraction butterfly, size = 16, 32
highbd_add_sub_butterfly(const __m128i * in,__m128i * out,int size)191 static INLINE void highbd_add_sub_butterfly(const __m128i *in, __m128i *out,
192 int size) {
193 int i = 0;
194 const int num = size >> 1;
195 const int bound = size - 1;
196 while (i < num) {
197 out[i] = _mm_add_epi32(in[i], in[bound - i]);
198 out[bound - i] = _mm_sub_epi32(in[i], in[bound - i]);
199 i++;
200 }
201 }
202
highbd_idct8_stage4(const __m128i * const in,__m128i * const out)203 static INLINE void highbd_idct8_stage4(const __m128i *const in,
204 __m128i *const out) {
205 out[0] = _mm_add_epi32(in[0], in[7]);
206 out[1] = _mm_add_epi32(in[1], in[6]);
207 out[2] = _mm_add_epi32(in[2], in[5]);
208 out[3] = _mm_add_epi32(in[3], in[4]);
209 out[4] = _mm_sub_epi32(in[3], in[4]);
210 out[5] = _mm_sub_epi32(in[2], in[5]);
211 out[6] = _mm_sub_epi32(in[1], in[6]);
212 out[7] = _mm_sub_epi32(in[0], in[7]);
213 }
214
highbd_idct8x8_final_round(__m128i * const io)215 static INLINE void highbd_idct8x8_final_round(__m128i *const io) {
216 io[0] = wraplow_16bit_shift5(io[0], io[8], _mm_set1_epi32(16));
217 io[1] = wraplow_16bit_shift5(io[1], io[9], _mm_set1_epi32(16));
218 io[2] = wraplow_16bit_shift5(io[2], io[10], _mm_set1_epi32(16));
219 io[3] = wraplow_16bit_shift5(io[3], io[11], _mm_set1_epi32(16));
220 io[4] = wraplow_16bit_shift5(io[4], io[12], _mm_set1_epi32(16));
221 io[5] = wraplow_16bit_shift5(io[5], io[13], _mm_set1_epi32(16));
222 io[6] = wraplow_16bit_shift5(io[6], io[14], _mm_set1_epi32(16));
223 io[7] = wraplow_16bit_shift5(io[7], io[15], _mm_set1_epi32(16));
224 }
225
highbd_idct16_4col_stage7(const __m128i * const in,__m128i * const out)226 static INLINE void highbd_idct16_4col_stage7(const __m128i *const in,
227 __m128i *const out) {
228 out[0] = _mm_add_epi32(in[0], in[15]);
229 out[1] = _mm_add_epi32(in[1], in[14]);
230 out[2] = _mm_add_epi32(in[2], in[13]);
231 out[3] = _mm_add_epi32(in[3], in[12]);
232 out[4] = _mm_add_epi32(in[4], in[11]);
233 out[5] = _mm_add_epi32(in[5], in[10]);
234 out[6] = _mm_add_epi32(in[6], in[9]);
235 out[7] = _mm_add_epi32(in[7], in[8]);
236 out[8] = _mm_sub_epi32(in[7], in[8]);
237 out[9] = _mm_sub_epi32(in[6], in[9]);
238 out[10] = _mm_sub_epi32(in[5], in[10]);
239 out[11] = _mm_sub_epi32(in[4], in[11]);
240 out[12] = _mm_sub_epi32(in[3], in[12]);
241 out[13] = _mm_sub_epi32(in[2], in[13]);
242 out[14] = _mm_sub_epi32(in[1], in[14]);
243 out[15] = _mm_sub_epi32(in[0], in[15]);
244 }
245
add_clamp(const __m128i in0,const __m128i in1,const int bd)246 static INLINE __m128i add_clamp(const __m128i in0, const __m128i in1,
247 const int bd) {
248 const __m128i zero = _mm_set1_epi16(0);
249 // Faster than _mm_set1_epi16((1 << bd) - 1).
250 const __m128i one = _mm_set1_epi16(1);
251 const __m128i max = _mm_sub_epi16(_mm_slli_epi16(one, bd), one);
252 __m128i d;
253
254 d = _mm_adds_epi16(in0, in1);
255 d = _mm_max_epi16(d, zero);
256 d = _mm_min_epi16(d, max);
257
258 return d;
259 }
260
highbd_idct_1_add_kernel(const tran_low_t * input,uint16_t * dest,int stride,int bd,const int size)261 static INLINE void highbd_idct_1_add_kernel(const tran_low_t *input,
262 uint16_t *dest, int stride, int bd,
263 const int size) {
264 int a1, i, j;
265 tran_low_t out;
266 __m128i dc, d;
267
268 out = HIGHBD_WRAPLOW(
269 dct_const_round_shift(input[0] * (tran_high_t)cospi_16_64), bd);
270 out =
271 HIGHBD_WRAPLOW(dct_const_round_shift(out * (tran_high_t)cospi_16_64), bd);
272 a1 = ROUND_POWER_OF_TWO(out, (size == 8) ? 5 : 6);
273 dc = _mm_set1_epi16(a1);
274
275 for (i = 0; i < size; ++i) {
276 for (j = 0; j < size; j += 8) {
277 d = _mm_load_si128((const __m128i *)(&dest[j]));
278 d = add_clamp(d, dc, bd);
279 _mm_store_si128((__m128i *)(&dest[j]), d);
280 }
281 dest += stride;
282 }
283 }
284
recon_and_store_4(const __m128i in,uint16_t * const dest,const int bd)285 static INLINE void recon_and_store_4(const __m128i in, uint16_t *const dest,
286 const int bd) {
287 __m128i d;
288
289 d = _mm_loadl_epi64((const __m128i *)dest);
290 d = add_clamp(d, in, bd);
291 _mm_storel_epi64((__m128i *)dest, d);
292 }
293
recon_and_store_4x2(const __m128i in,uint16_t * const dest,const int stride,const int bd)294 static INLINE void recon_and_store_4x2(const __m128i in, uint16_t *const dest,
295 const int stride, const int bd) {
296 __m128i d;
297
298 d = _mm_loadl_epi64((const __m128i *)(dest + 0 * stride));
299 d = _mm_castps_si128(
300 _mm_loadh_pi(_mm_castsi128_ps(d), (const __m64 *)(dest + 1 * stride)));
301 d = add_clamp(d, in, bd);
302 _mm_storel_epi64((__m128i *)(dest + 0 * stride), d);
303 _mm_storeh_pi((__m64 *)(dest + 1 * stride), _mm_castsi128_ps(d));
304 }
305
recon_and_store_4x4(const __m128i * const in,uint16_t * dest,const int stride,const int bd)306 static INLINE void recon_and_store_4x4(const __m128i *const in, uint16_t *dest,
307 const int stride, const int bd) {
308 recon_and_store_4x2(in[0], dest, stride, bd);
309 dest += 2 * stride;
310 recon_and_store_4x2(in[1], dest, stride, bd);
311 }
312
recon_and_store_8(const __m128i in,uint16_t ** const dest,const int stride,const int bd)313 static INLINE void recon_and_store_8(const __m128i in, uint16_t **const dest,
314 const int stride, const int bd) {
315 __m128i d;
316
317 d = _mm_load_si128((const __m128i *)(*dest));
318 d = add_clamp(d, in, bd);
319 _mm_store_si128((__m128i *)(*dest), d);
320 *dest += stride;
321 }
322
recon_and_store_8x8(const __m128i * const in,uint16_t * dest,const int stride,const int bd)323 static INLINE void recon_and_store_8x8(const __m128i *const in, uint16_t *dest,
324 const int stride, const int bd) {
325 recon_and_store_8(in[0], &dest, stride, bd);
326 recon_and_store_8(in[1], &dest, stride, bd);
327 recon_and_store_8(in[2], &dest, stride, bd);
328 recon_and_store_8(in[3], &dest, stride, bd);
329 recon_and_store_8(in[4], &dest, stride, bd);
330 recon_and_store_8(in[5], &dest, stride, bd);
331 recon_and_store_8(in[6], &dest, stride, bd);
332 recon_and_store_8(in[7], &dest, stride, bd);
333 }
334
load_pack_8_32bit(const tran_low_t * const input)335 static INLINE __m128i load_pack_8_32bit(const tran_low_t *const input) {
336 const __m128i t0 = _mm_load_si128((const __m128i *)(input + 0));
337 const __m128i t1 = _mm_load_si128((const __m128i *)(input + 4));
338 return _mm_packs_epi32(t0, t1);
339 }
340
highbd_load_pack_transpose_32bit_8x8(const tran_low_t * input,const int stride,__m128i * const in)341 static INLINE void highbd_load_pack_transpose_32bit_8x8(const tran_low_t *input,
342 const int stride,
343 __m128i *const in) {
344 in[0] = load_pack_8_32bit(input + 0 * stride);
345 in[1] = load_pack_8_32bit(input + 1 * stride);
346 in[2] = load_pack_8_32bit(input + 2 * stride);
347 in[3] = load_pack_8_32bit(input + 3 * stride);
348 in[4] = load_pack_8_32bit(input + 4 * stride);
349 in[5] = load_pack_8_32bit(input + 5 * stride);
350 in[6] = load_pack_8_32bit(input + 6 * stride);
351 in[7] = load_pack_8_32bit(input + 7 * stride);
352 transpose_16bit_8x8(in, in);
353 }
354
highbd_load_transpose_32bit_8x4(const tran_low_t * input,const int stride,__m128i * in)355 static INLINE void highbd_load_transpose_32bit_8x4(const tran_low_t *input,
356 const int stride,
357 __m128i *in) {
358 in[0] = _mm_load_si128((const __m128i *)(input + 0 * stride + 0));
359 in[1] = _mm_load_si128((const __m128i *)(input + 0 * stride + 4));
360 in[2] = _mm_load_si128((const __m128i *)(input + 1 * stride + 0));
361 in[3] = _mm_load_si128((const __m128i *)(input + 1 * stride + 4));
362 in[4] = _mm_load_si128((const __m128i *)(input + 2 * stride + 0));
363 in[5] = _mm_load_si128((const __m128i *)(input + 2 * stride + 4));
364 in[6] = _mm_load_si128((const __m128i *)(input + 3 * stride + 0));
365 in[7] = _mm_load_si128((const __m128i *)(input + 3 * stride + 4));
366 transpose_32bit_8x4(in, in);
367 }
368
highbd_load_transpose_32bit_4x4(const tran_low_t * input,const int stride,__m128i * in)369 static INLINE void highbd_load_transpose_32bit_4x4(const tran_low_t *input,
370 const int stride,
371 __m128i *in) {
372 in[0] = _mm_load_si128((const __m128i *)(input + 0 * stride));
373 in[1] = _mm_load_si128((const __m128i *)(input + 1 * stride));
374 in[2] = _mm_load_si128((const __m128i *)(input + 2 * stride));
375 in[3] = _mm_load_si128((const __m128i *)(input + 3 * stride));
376 transpose_32bit_4x4(in, in);
377 }
378
highbd_write_buffer_8(uint16_t * dest,const __m128i in,const int bd)379 static INLINE void highbd_write_buffer_8(uint16_t *dest, const __m128i in,
380 const int bd) {
381 const __m128i final_rounding = _mm_set1_epi16(1 << 5);
382 __m128i out;
383
384 out = _mm_adds_epi16(in, final_rounding);
385 out = _mm_srai_epi16(out, 6);
386 recon_and_store_8(out, &dest, 0, bd);
387 }
388
highbd_write_buffer_4(uint16_t * const dest,const __m128i in,const int bd)389 static INLINE void highbd_write_buffer_4(uint16_t *const dest, const __m128i in,
390 const int bd) {
391 const __m128i final_rounding = _mm_set1_epi32(1 << 5);
392 __m128i out;
393
394 out = _mm_add_epi32(in, final_rounding);
395 out = _mm_srai_epi32(out, 6);
396 out = _mm_packs_epi32(out, out);
397 recon_and_store_4(out, dest, bd);
398 }
399
400 #endif // VPX_DSP_X86_HIGHBD_INV_TXFM_SSE2_H_
401