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_VPX_DSP_X86_HIGHBD_INV_TXFM_SSE2_H_
12 #define VPX_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
22 // Note: There is no 64-bit bit-level shifting SIMD instruction. All
23 // coefficients are left shifted by 2, so that dct_const_round_shift() can be
24 // done by right shifting 2 bytes.
25
extend_64bit(const __m128i in,__m128i * const out)26 static INLINE void extend_64bit(const __m128i in,
27 __m128i *const out /*out[2]*/) {
28 out[0] = _mm_unpacklo_epi32(in, in); // 0, 0, 1, 1
29 out[1] = _mm_unpackhi_epi32(in, in); // 2, 2, 3, 3
30 }
31
wraplow_16bit_shift4(const __m128i in0,const __m128i in1,const __m128i rounding)32 static INLINE __m128i wraplow_16bit_shift4(const __m128i in0, const __m128i in1,
33 const __m128i rounding) {
34 __m128i temp[2];
35 temp[0] = _mm_add_epi32(in0, rounding);
36 temp[1] = _mm_add_epi32(in1, rounding);
37 temp[0] = _mm_srai_epi32(temp[0], 4);
38 temp[1] = _mm_srai_epi32(temp[1], 4);
39 return _mm_packs_epi32(temp[0], temp[1]);
40 }
41
wraplow_16bit_shift5(const __m128i in0,const __m128i in1,const __m128i rounding)42 static INLINE __m128i wraplow_16bit_shift5(const __m128i in0, const __m128i in1,
43 const __m128i rounding) {
44 __m128i temp[2];
45 temp[0] = _mm_add_epi32(in0, rounding);
46 temp[1] = _mm_add_epi32(in1, rounding);
47 temp[0] = _mm_srai_epi32(temp[0], 5);
48 temp[1] = _mm_srai_epi32(temp[1], 5);
49 return _mm_packs_epi32(temp[0], temp[1]);
50 }
51
dct_const_round_shift_64bit(const __m128i in)52 static INLINE __m128i dct_const_round_shift_64bit(const __m128i in) {
53 const __m128i t =
54 _mm_add_epi64(in, pair_set_epi32(DCT_CONST_ROUNDING << 2, 0));
55 return _mm_srli_si128(t, 2);
56 }
57
pack_4(const __m128i in0,const __m128i in1)58 static INLINE __m128i pack_4(const __m128i in0, const __m128i in1) {
59 const __m128i t0 = _mm_unpacklo_epi32(in0, in1); // 0, 2
60 const __m128i t1 = _mm_unpackhi_epi32(in0, in1); // 1, 3
61 return _mm_unpacklo_epi32(t0, t1); // 0, 1, 2, 3
62 }
63
abs_extend_64bit_sse2(const __m128i in,__m128i * const out,__m128i * const sign)64 static INLINE void abs_extend_64bit_sse2(const __m128i in,
65 __m128i *const out /*out[2]*/,
66 __m128i *const sign /*sign[2]*/) {
67 sign[0] = _mm_srai_epi32(in, 31);
68 out[0] = _mm_xor_si128(in, sign[0]);
69 out[0] = _mm_sub_epi32(out[0], sign[0]);
70 sign[1] = _mm_unpackhi_epi32(sign[0], sign[0]); // 64-bit sign of 2, 3
71 sign[0] = _mm_unpacklo_epi32(sign[0], sign[0]); // 64-bit sign of 0, 1
72 out[1] = _mm_unpackhi_epi32(out[0], out[0]); // 2, 3
73 out[0] = _mm_unpacklo_epi32(out[0], out[0]); // 0, 1
74 }
75
76 // Note: cospi must be non negative.
multiply_apply_sign_sse2(const __m128i in,const __m128i sign,const __m128i cospi)77 static INLINE __m128i multiply_apply_sign_sse2(const __m128i in,
78 const __m128i sign,
79 const __m128i cospi) {
80 __m128i out = _mm_mul_epu32(in, cospi);
81 out = _mm_xor_si128(out, sign);
82 return _mm_sub_epi64(out, sign);
83 }
84
85 // Note: c must be non negative.
multiplication_round_shift_sse2(const __m128i * const in,const __m128i * const sign,const int c)86 static INLINE __m128i multiplication_round_shift_sse2(
87 const __m128i *const in /*in[2]*/, const __m128i *const sign /*sign[2]*/,
88 const int c) {
89 const __m128i pair_c = pair_set_epi32(c << 2, 0);
90 __m128i t0, t1;
91
92 assert(c >= 0);
93 t0 = multiply_apply_sign_sse2(in[0], sign[0], pair_c);
94 t1 = multiply_apply_sign_sse2(in[1], sign[1], pair_c);
95 t0 = dct_const_round_shift_64bit(t0);
96 t1 = dct_const_round_shift_64bit(t1);
97
98 return pack_4(t0, t1);
99 }
100
101 // Note: c must be non negative.
multiplication_neg_round_shift_sse2(const __m128i * const in,const __m128i * const sign,const int c)102 static INLINE __m128i multiplication_neg_round_shift_sse2(
103 const __m128i *const in /*in[2]*/, const __m128i *const sign /*sign[2]*/,
104 const int c) {
105 const __m128i pair_c = pair_set_epi32(c << 2, 0);
106 __m128i t0, t1;
107
108 assert(c >= 0);
109 t0 = multiply_apply_sign_sse2(in[0], sign[0], pair_c);
110 t1 = multiply_apply_sign_sse2(in[1], sign[1], pair_c);
111 t0 = _mm_sub_epi64(_mm_setzero_si128(), t0);
112 t1 = _mm_sub_epi64(_mm_setzero_si128(), t1);
113 t0 = dct_const_round_shift_64bit(t0);
114 t1 = dct_const_round_shift_64bit(t1);
115
116 return pack_4(t0, t1);
117 }
118
119 // 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)120 static INLINE void highbd_butterfly_sse2(const __m128i in0, const __m128i in1,
121 const int c0, const int c1,
122 __m128i *const out0,
123 __m128i *const out1) {
124 const __m128i pair_c0 = pair_set_epi32(c0 << 2, 0);
125 const __m128i pair_c1 = pair_set_epi32(c1 << 2, 0);
126 __m128i temp1[4], temp2[4], sign1[2], sign2[2];
127
128 assert(c0 >= 0);
129 assert(c1 >= 0);
130 abs_extend_64bit_sse2(in0, temp1, sign1);
131 abs_extend_64bit_sse2(in1, temp2, sign2);
132 temp1[2] = multiply_apply_sign_sse2(temp1[0], sign1[0], pair_c1);
133 temp1[3] = multiply_apply_sign_sse2(temp1[1], sign1[1], pair_c1);
134 temp1[0] = multiply_apply_sign_sse2(temp1[0], sign1[0], pair_c0);
135 temp1[1] = multiply_apply_sign_sse2(temp1[1], sign1[1], pair_c0);
136 temp2[2] = multiply_apply_sign_sse2(temp2[0], sign2[0], pair_c0);
137 temp2[3] = multiply_apply_sign_sse2(temp2[1], sign2[1], pair_c0);
138 temp2[0] = multiply_apply_sign_sse2(temp2[0], sign2[0], pair_c1);
139 temp2[1] = multiply_apply_sign_sse2(temp2[1], sign2[1], pair_c1);
140 temp1[0] = _mm_sub_epi64(temp1[0], temp2[0]);
141 temp1[1] = _mm_sub_epi64(temp1[1], temp2[1]);
142 temp2[0] = _mm_add_epi64(temp1[2], temp2[2]);
143 temp2[1] = _mm_add_epi64(temp1[3], temp2[3]);
144 temp1[0] = dct_const_round_shift_64bit(temp1[0]);
145 temp1[1] = dct_const_round_shift_64bit(temp1[1]);
146 temp2[0] = dct_const_round_shift_64bit(temp2[0]);
147 temp2[1] = dct_const_round_shift_64bit(temp2[1]);
148 *out0 = pack_4(temp1[0], temp1[1]);
149 *out1 = pack_4(temp2[0], temp2[1]);
150 }
151
152 // 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)153 static INLINE void highbd_partial_butterfly_sse2(const __m128i in, const int c0,
154 const int c1,
155 __m128i *const out0,
156 __m128i *const out1) {
157 __m128i temp[2], sign[2];
158
159 assert(c0 >= 0);
160 assert(c1 >= 0);
161 abs_extend_64bit_sse2(in, temp, sign);
162 *out0 = multiplication_round_shift_sse2(temp, sign, c0);
163 *out1 = multiplication_round_shift_sse2(temp, sign, c1);
164 }
165
166 // 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)167 static INLINE void highbd_partial_butterfly_neg_sse2(const __m128i in,
168 const int c0, const int c1,
169 __m128i *const out0,
170 __m128i *const out1) {
171 __m128i temp[2], sign[2];
172
173 assert(c0 >= 0);
174 assert(c1 >= 0);
175 abs_extend_64bit_sse2(in, temp, sign);
176 *out0 = multiplication_neg_round_shift_sse2(temp, sign, c1);
177 *out1 = multiplication_round_shift_sse2(temp, sign, c0);
178 }
179
highbd_butterfly_cospi16_sse2(const __m128i in0,const __m128i in1,__m128i * const out0,__m128i * const out1)180 static INLINE void highbd_butterfly_cospi16_sse2(const __m128i in0,
181 const __m128i in1,
182 __m128i *const out0,
183 __m128i *const out1) {
184 __m128i temp1[2], temp2, sign[2];
185
186 temp2 = _mm_add_epi32(in0, in1);
187 abs_extend_64bit_sse2(temp2, temp1, sign);
188 *out0 = multiplication_round_shift_sse2(temp1, sign, cospi_16_64);
189 temp2 = _mm_sub_epi32(in0, in1);
190 abs_extend_64bit_sse2(temp2, temp1, sign);
191 *out1 = multiplication_round_shift_sse2(temp1, sign, cospi_16_64);
192 }
193
194 // Only do addition and subtraction butterfly, size = 16, 32
highbd_add_sub_butterfly(const __m128i * in,__m128i * out,int size)195 static INLINE void highbd_add_sub_butterfly(const __m128i *in, __m128i *out,
196 int size) {
197 int i = 0;
198 const int num = size >> 1;
199 const int bound = size - 1;
200 while (i < num) {
201 out[i] = _mm_add_epi32(in[i], in[bound - i]);
202 out[bound - i] = _mm_sub_epi32(in[i], in[bound - i]);
203 i++;
204 }
205 }
206
highbd_idct8_stage4(const __m128i * const in,__m128i * const out)207 static INLINE void highbd_idct8_stage4(const __m128i *const in,
208 __m128i *const out) {
209 out[0] = _mm_add_epi32(in[0], in[7]);
210 out[1] = _mm_add_epi32(in[1], in[6]);
211 out[2] = _mm_add_epi32(in[2], in[5]);
212 out[3] = _mm_add_epi32(in[3], in[4]);
213 out[4] = _mm_sub_epi32(in[3], in[4]);
214 out[5] = _mm_sub_epi32(in[2], in[5]);
215 out[6] = _mm_sub_epi32(in[1], in[6]);
216 out[7] = _mm_sub_epi32(in[0], in[7]);
217 }
218
highbd_idct8x8_final_round(__m128i * const io)219 static INLINE void highbd_idct8x8_final_round(__m128i *const io) {
220 io[0] = wraplow_16bit_shift5(io[0], io[8], _mm_set1_epi32(16));
221 io[1] = wraplow_16bit_shift5(io[1], io[9], _mm_set1_epi32(16));
222 io[2] = wraplow_16bit_shift5(io[2], io[10], _mm_set1_epi32(16));
223 io[3] = wraplow_16bit_shift5(io[3], io[11], _mm_set1_epi32(16));
224 io[4] = wraplow_16bit_shift5(io[4], io[12], _mm_set1_epi32(16));
225 io[5] = wraplow_16bit_shift5(io[5], io[13], _mm_set1_epi32(16));
226 io[6] = wraplow_16bit_shift5(io[6], io[14], _mm_set1_epi32(16));
227 io[7] = wraplow_16bit_shift5(io[7], io[15], _mm_set1_epi32(16));
228 }
229
highbd_idct16_4col_stage7(const __m128i * const in,__m128i * const out)230 static INLINE void highbd_idct16_4col_stage7(const __m128i *const in,
231 __m128i *const out) {
232 out[0] = _mm_add_epi32(in[0], in[15]);
233 out[1] = _mm_add_epi32(in[1], in[14]);
234 out[2] = _mm_add_epi32(in[2], in[13]);
235 out[3] = _mm_add_epi32(in[3], in[12]);
236 out[4] = _mm_add_epi32(in[4], in[11]);
237 out[5] = _mm_add_epi32(in[5], in[10]);
238 out[6] = _mm_add_epi32(in[6], in[9]);
239 out[7] = _mm_add_epi32(in[7], in[8]);
240 out[8] = _mm_sub_epi32(in[7], in[8]);
241 out[9] = _mm_sub_epi32(in[6], in[9]);
242 out[10] = _mm_sub_epi32(in[5], in[10]);
243 out[11] = _mm_sub_epi32(in[4], in[11]);
244 out[12] = _mm_sub_epi32(in[3], in[12]);
245 out[13] = _mm_sub_epi32(in[2], in[13]);
246 out[14] = _mm_sub_epi32(in[1], in[14]);
247 out[15] = _mm_sub_epi32(in[0], in[15]);
248 }
249
add_clamp(const __m128i in0,const __m128i in1,const int bd)250 static INLINE __m128i add_clamp(const __m128i in0, const __m128i in1,
251 const int bd) {
252 const __m128i zero = _mm_setzero_si128();
253 // Faster than _mm_set1_epi16((1 << bd) - 1).
254 const __m128i one = _mm_set1_epi16(1);
255 const __m128i max = _mm_sub_epi16(_mm_slli_epi16(one, bd), one);
256 __m128i d;
257
258 d = _mm_adds_epi16(in0, in1);
259 d = _mm_max_epi16(d, zero);
260 d = _mm_min_epi16(d, max);
261
262 return d;
263 }
264
highbd_idct_1_add_kernel(const tran_low_t * input,uint16_t * dest,int stride,int bd,const int size)265 static INLINE void highbd_idct_1_add_kernel(const tran_low_t *input,
266 uint16_t *dest, int stride, int bd,
267 const int size) {
268 int a1, i, j;
269 tran_low_t out;
270 __m128i dc, d;
271
272 out = HIGHBD_WRAPLOW(
273 dct_const_round_shift(input[0] * (tran_high_t)cospi_16_64), bd);
274 out =
275 HIGHBD_WRAPLOW(dct_const_round_shift(out * (tran_high_t)cospi_16_64), bd);
276 a1 = ROUND_POWER_OF_TWO(out, (size == 8) ? 5 : 6);
277 dc = _mm_set1_epi16(a1);
278
279 for (i = 0; i < size; ++i) {
280 for (j = 0; j < size; j += 8) {
281 d = _mm_load_si128((const __m128i *)(&dest[j]));
282 d = add_clamp(d, dc, bd);
283 _mm_store_si128((__m128i *)(&dest[j]), d);
284 }
285 dest += stride;
286 }
287 }
288
recon_and_store_4(const __m128i in,uint16_t * const dest,const int bd)289 static INLINE void recon_and_store_4(const __m128i in, uint16_t *const dest,
290 const int bd) {
291 __m128i d;
292
293 d = _mm_loadl_epi64((const __m128i *)dest);
294 d = add_clamp(d, in, bd);
295 _mm_storel_epi64((__m128i *)dest, d);
296 }
297
recon_and_store_4x2(const __m128i in,uint16_t * const dest,const int stride,const int bd)298 static INLINE void recon_and_store_4x2(const __m128i in, uint16_t *const dest,
299 const int stride, const int bd) {
300 __m128i d;
301
302 d = _mm_loadl_epi64((const __m128i *)(dest + 0 * stride));
303 d = _mm_castps_si128(
304 _mm_loadh_pi(_mm_castsi128_ps(d), (const __m64 *)(dest + 1 * stride)));
305 d = add_clamp(d, in, bd);
306 _mm_storel_epi64((__m128i *)(dest + 0 * stride), d);
307 _mm_storeh_pi((__m64 *)(dest + 1 * stride), _mm_castsi128_ps(d));
308 }
309
recon_and_store_4x4(const __m128i * const in,uint16_t * dest,const int stride,const int bd)310 static INLINE void recon_and_store_4x4(const __m128i *const in, uint16_t *dest,
311 const int stride, const int bd) {
312 recon_and_store_4x2(in[0], dest, stride, bd);
313 dest += 2 * stride;
314 recon_and_store_4x2(in[1], dest, stride, bd);
315 }
316
recon_and_store_8(const __m128i in,uint16_t ** const dest,const int stride,const int bd)317 static INLINE void recon_and_store_8(const __m128i in, uint16_t **const dest,
318 const int stride, const int bd) {
319 __m128i d;
320
321 d = _mm_load_si128((const __m128i *)(*dest));
322 d = add_clamp(d, in, bd);
323 _mm_store_si128((__m128i *)(*dest), d);
324 *dest += stride;
325 }
326
recon_and_store_8x8(const __m128i * const in,uint16_t * dest,const int stride,const int bd)327 static INLINE void recon_and_store_8x8(const __m128i *const in, uint16_t *dest,
328 const int stride, const int bd) {
329 recon_and_store_8(in[0], &dest, stride, bd);
330 recon_and_store_8(in[1], &dest, stride, bd);
331 recon_and_store_8(in[2], &dest, stride, bd);
332 recon_and_store_8(in[3], &dest, stride, bd);
333 recon_and_store_8(in[4], &dest, stride, bd);
334 recon_and_store_8(in[5], &dest, stride, bd);
335 recon_and_store_8(in[6], &dest, stride, bd);
336 recon_and_store_8(in[7], &dest, stride, bd);
337 }
338
load_pack_8_32bit(const tran_low_t * const input)339 static INLINE __m128i load_pack_8_32bit(const tran_low_t *const input) {
340 const __m128i t0 = _mm_load_si128((const __m128i *)(input + 0));
341 const __m128i t1 = _mm_load_si128((const __m128i *)(input + 4));
342 return _mm_packs_epi32(t0, t1);
343 }
344
highbd_load_pack_transpose_32bit_8x8(const tran_low_t * input,const int stride,__m128i * const in)345 static INLINE void highbd_load_pack_transpose_32bit_8x8(const tran_low_t *input,
346 const int stride,
347 __m128i *const in) {
348 in[0] = load_pack_8_32bit(input + 0 * stride);
349 in[1] = load_pack_8_32bit(input + 1 * stride);
350 in[2] = load_pack_8_32bit(input + 2 * stride);
351 in[3] = load_pack_8_32bit(input + 3 * stride);
352 in[4] = load_pack_8_32bit(input + 4 * stride);
353 in[5] = load_pack_8_32bit(input + 5 * stride);
354 in[6] = load_pack_8_32bit(input + 6 * stride);
355 in[7] = load_pack_8_32bit(input + 7 * stride);
356 transpose_16bit_8x8(in, in);
357 }
358
highbd_load_transpose_32bit_8x4(const tran_low_t * input,const int stride,__m128i * in)359 static INLINE void highbd_load_transpose_32bit_8x4(const tran_low_t *input,
360 const int stride,
361 __m128i *in) {
362 in[0] = _mm_load_si128((const __m128i *)(input + 0 * stride + 0));
363 in[1] = _mm_load_si128((const __m128i *)(input + 0 * stride + 4));
364 in[2] = _mm_load_si128((const __m128i *)(input + 1 * stride + 0));
365 in[3] = _mm_load_si128((const __m128i *)(input + 1 * stride + 4));
366 in[4] = _mm_load_si128((const __m128i *)(input + 2 * stride + 0));
367 in[5] = _mm_load_si128((const __m128i *)(input + 2 * stride + 4));
368 in[6] = _mm_load_si128((const __m128i *)(input + 3 * stride + 0));
369 in[7] = _mm_load_si128((const __m128i *)(input + 3 * stride + 4));
370 transpose_32bit_8x4(in, in);
371 }
372
highbd_load_transpose_32bit_4x4(const tran_low_t * input,const int stride,__m128i * in)373 static INLINE void highbd_load_transpose_32bit_4x4(const tran_low_t *input,
374 const int stride,
375 __m128i *in) {
376 in[0] = _mm_load_si128((const __m128i *)(input + 0 * stride));
377 in[1] = _mm_load_si128((const __m128i *)(input + 1 * stride));
378 in[2] = _mm_load_si128((const __m128i *)(input + 2 * stride));
379 in[3] = _mm_load_si128((const __m128i *)(input + 3 * stride));
380 transpose_32bit_4x4(in, in);
381 }
382
highbd_write_buffer_8(uint16_t * dest,const __m128i in,const int bd)383 static INLINE void highbd_write_buffer_8(uint16_t *dest, const __m128i in,
384 const int bd) {
385 const __m128i final_rounding = _mm_set1_epi16(1 << 5);
386 __m128i out;
387
388 out = _mm_adds_epi16(in, final_rounding);
389 out = _mm_srai_epi16(out, 6);
390 recon_and_store_8(out, &dest, 0, bd);
391 }
392
highbd_write_buffer_4(uint16_t * const dest,const __m128i in,const int bd)393 static INLINE void highbd_write_buffer_4(uint16_t *const dest, const __m128i in,
394 const int bd) {
395 const __m128i final_rounding = _mm_set1_epi32(1 << 5);
396 __m128i out;
397
398 out = _mm_add_epi32(in, final_rounding);
399 out = _mm_srai_epi32(out, 6);
400 out = _mm_packs_epi32(out, out);
401 recon_and_store_4(out, dest, bd);
402 }
403
404 #endif // VPX_VPX_DSP_X86_HIGHBD_INV_TXFM_SSE2_H_
405