1/*
2 * Mesa 3-D graphics library
3 *
4 * Copyright 2012 Intel Corporation
5 * Copyright 2013 Google
6 *
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sublicense, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
14 *
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
17 * of the Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
20 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
21 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
22 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
23 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
24 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
25 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 *
27 * Authors:
28 *    Chad Versace <chad.versace@linux.intel.com>
29 *    Frank Henigman <fjhenigman@google.com>
30 */
31
32#include <string.h>
33
34#include "util/macros.h"
35#include "util/u_math.h"
36#include "util/rounding.h"
37#include "isl_priv.h"
38
39#if defined(__SSSE3__)
40#include <tmmintrin.h>
41#elif defined(__SSE2__)
42#include <emmintrin.h>
43#endif
44
45#define FILE_DEBUG_FLAG DEBUG_TEXTURE
46
47#define ALIGN_DOWN(a, b) ROUND_DOWN_TO(a, b)
48#define ALIGN_UP(a, b) ALIGN(a, b)
49
50/* Tile dimensions.  Width and span are in bytes, height is in pixels (i.e.
51 * unitless).  A "span" is the most number of bytes we can copy from linear
52 * to tiled without needing to calculate a new destination address.
53 */
54static const uint32_t xtile_width = 512;
55static const uint32_t xtile_height = 8;
56static const uint32_t xtile_span = 64;
57static const uint32_t ytile_width = 128;
58static const uint32_t ytile_height = 32;
59static const uint32_t ytile_span = 16;
60
61static inline uint32_t
62ror(uint32_t n, uint32_t d)
63{
64   return (n >> d) | (n << (32 - d));
65}
66
67// bswap32 already exists as a macro on some platforms (FreeBSD)
68#ifndef bswap32
69static inline uint32_t
70bswap32(uint32_t n)
71{
72#if defined(HAVE___BUILTIN_BSWAP32)
73   return __builtin_bswap32(n);
74#else
75   return (n >> 24) |
76          ((n >> 8) & 0x0000ff00) |
77          ((n << 8) & 0x00ff0000) |
78          (n << 24);
79#endif
80}
81#endif
82
83/**
84 * Copy RGBA to BGRA - swap R and B.
85 */
86static inline void *
87rgba8_copy(void *dst, const void *src, size_t bytes)
88{
89   uint32_t *d = dst;
90   uint32_t const *s = src;
91
92   assert(bytes % 4 == 0);
93
94   while (bytes >= 4) {
95      *d = ror(bswap32(*s), 8);
96      d += 1;
97      s += 1;
98      bytes -= 4;
99   }
100   return dst;
101}
102
103#ifdef __SSSE3__
104static const uint8_t rgba8_permutation[16] =
105   { 2,1,0,3, 6,5,4,7, 10,9,8,11, 14,13,12,15 };
106
107static inline void
108rgba8_copy_16_aligned_dst(void *dst, const void *src)
109{
110   _mm_store_si128(dst,
111                   _mm_shuffle_epi8(_mm_loadu_si128(src),
112                                    *(__m128i *)rgba8_permutation));
113}
114
115static inline void
116rgba8_copy_16_aligned_src(void *dst, const void *src)
117{
118   _mm_storeu_si128(dst,
119                    _mm_shuffle_epi8(_mm_load_si128(src),
120                                     *(__m128i *)rgba8_permutation));
121}
122
123#elif defined(__SSE2__)
124static inline void
125rgba8_copy_16_aligned_dst(void *dst, const void *src)
126{
127   __m128i srcreg, dstreg, agmask, ag, rb, br;
128
129   agmask = _mm_set1_epi32(0xFF00FF00);
130   srcreg = _mm_loadu_si128((__m128i *)src);
131
132   rb = _mm_andnot_si128(agmask, srcreg);
133   ag = _mm_and_si128(agmask, srcreg);
134   br = _mm_shufflehi_epi16(_mm_shufflelo_epi16(rb, _MM_SHUFFLE(2, 3, 0, 1)),
135                            _MM_SHUFFLE(2, 3, 0, 1));
136   dstreg = _mm_or_si128(ag, br);
137
138   _mm_store_si128((__m128i *)dst, dstreg);
139}
140
141static inline void
142rgba8_copy_16_aligned_src(void *dst, const void *src)
143{
144   __m128i srcreg, dstreg, agmask, ag, rb, br;
145
146   agmask = _mm_set1_epi32(0xFF00FF00);
147   srcreg = _mm_load_si128((__m128i *)src);
148
149   rb = _mm_andnot_si128(agmask, srcreg);
150   ag = _mm_and_si128(agmask, srcreg);
151   br = _mm_shufflehi_epi16(_mm_shufflelo_epi16(rb, _MM_SHUFFLE(2, 3, 0, 1)),
152                            _MM_SHUFFLE(2, 3, 0, 1));
153   dstreg = _mm_or_si128(ag, br);
154
155   _mm_storeu_si128((__m128i *)dst, dstreg);
156}
157#endif
158
159/**
160 * Copy RGBA to BGRA - swap R and B, with the destination 16-byte aligned.
161 */
162static inline void *
163rgba8_copy_aligned_dst(void *dst, const void *src, size_t bytes)
164{
165   assert(bytes == 0 || !(((uintptr_t)dst) & 0xf));
166
167#if defined(__SSSE3__) || defined(__SSE2__)
168   if (bytes == 64) {
169      rgba8_copy_16_aligned_dst(dst +  0, src +  0);
170      rgba8_copy_16_aligned_dst(dst + 16, src + 16);
171      rgba8_copy_16_aligned_dst(dst + 32, src + 32);
172      rgba8_copy_16_aligned_dst(dst + 48, src + 48);
173      return dst;
174   }
175
176   while (bytes >= 16) {
177      rgba8_copy_16_aligned_dst(dst, src);
178      src += 16;
179      dst += 16;
180      bytes -= 16;
181   }
182#endif
183
184   rgba8_copy(dst, src, bytes);
185
186   return dst;
187}
188
189/**
190 * Copy RGBA to BGRA - swap R and B, with the source 16-byte aligned.
191 */
192static inline void *
193rgba8_copy_aligned_src(void *dst, const void *src, size_t bytes)
194{
195   assert(bytes == 0 || !(((uintptr_t)src) & 0xf));
196
197#if defined(__SSSE3__) || defined(__SSE2__)
198   if (bytes == 64) {
199      rgba8_copy_16_aligned_src(dst +  0, src +  0);
200      rgba8_copy_16_aligned_src(dst + 16, src + 16);
201      rgba8_copy_16_aligned_src(dst + 32, src + 32);
202      rgba8_copy_16_aligned_src(dst + 48, src + 48);
203      return dst;
204   }
205
206   while (bytes >= 16) {
207      rgba8_copy_16_aligned_src(dst, src);
208      src += 16;
209      dst += 16;
210      bytes -= 16;
211   }
212#endif
213
214   rgba8_copy(dst, src, bytes);
215
216   return dst;
217}
218
219/**
220 * Each row from y0 to y1 is copied in three parts: [x0,x1), [x1,x2), [x2,x3).
221 * These ranges are in bytes, i.e. pixels * bytes-per-pixel.
222 * The first and last ranges must be shorter than a "span" (the longest linear
223 * stretch within a tile) and the middle must equal a whole number of spans.
224 * Ranges may be empty.  The region copied must land entirely within one tile.
225 * 'dst' is the start of the tile and 'src' is the corresponding
226 * address to copy from, though copying begins at (x0, y0).
227 * To enable swizzling 'swizzle_bit' must be 1<<6, otherwise zero.
228 * Swizzling flips bit 6 in the copy destination offset, when certain other
229 * bits are set in it.
230 */
231typedef void (*tile_copy_fn)(uint32_t x0, uint32_t x1, uint32_t x2, uint32_t x3,
232                             uint32_t y0, uint32_t y1,
233                             char *dst, const char *src,
234                             int32_t linear_pitch,
235                             uint32_t swizzle_bit,
236                             isl_memcpy_type copy_type);
237
238/**
239 * Copy texture data from linear to X tile layout.
240 *
241 * \copydoc tile_copy_fn
242 *
243 * The mem_copy parameters allow the user to specify an alternative mem_copy
244 * function that, for instance, may do RGBA -> BGRA swizzling.  The first
245 * function must handle any memory alignment while the second function must
246 * only handle 16-byte alignment in whichever side (source or destination) is
247 * tiled.
248 */
249static inline void
250linear_to_xtiled(uint32_t x0, uint32_t x1, uint32_t x2, uint32_t x3,
251                 uint32_t y0, uint32_t y1,
252                 char *dst, const char *src,
253                 int32_t src_pitch,
254                 uint32_t swizzle_bit,
255                 isl_mem_copy_fn mem_copy,
256                 isl_mem_copy_fn mem_copy_align16)
257{
258   /* The copy destination offset for each range copied is the sum of
259    * an X offset 'x0' or 'xo' and a Y offset 'yo.'
260    */
261   uint32_t xo, yo;
262
263   src += (ptrdiff_t)y0 * src_pitch;
264
265   for (yo = y0 * xtile_width; yo < y1 * xtile_width; yo += xtile_width) {
266      /* Bits 9 and 10 of the copy destination offset control swizzling.
267       * Only 'yo' contributes to those bits in the total offset,
268       * so calculate 'swizzle' just once per row.
269       * Move bits 9 and 10 three and four places respectively down
270       * to bit 6 and xor them.
271       */
272      uint32_t swizzle = ((yo >> 3) ^ (yo >> 4)) & swizzle_bit;
273
274      mem_copy(dst + ((x0 + yo) ^ swizzle), src + x0, x1 - x0);
275
276      for (xo = x1; xo < x2; xo += xtile_span) {
277         mem_copy_align16(dst + ((xo + yo) ^ swizzle), src + xo, xtile_span);
278      }
279
280      mem_copy_align16(dst + ((xo + yo) ^ swizzle), src + x2, x3 - x2);
281
282      src += src_pitch;
283   }
284}
285
286/**
287 * Copy texture data from linear to Y tile layout.
288 *
289 * \copydoc tile_copy_fn
290 */
291static inline void
292linear_to_ytiled(uint32_t x0, uint32_t x1, uint32_t x2, uint32_t x3,
293                 uint32_t y0, uint32_t y3,
294                 char *dst, const char *src,
295                 int32_t src_pitch,
296                 uint32_t swizzle_bit,
297                 isl_mem_copy_fn mem_copy,
298                 isl_mem_copy_fn mem_copy_align16)
299{
300   /* Y tiles consist of columns that are 'ytile_span' wide (and the same height
301    * as the tile).  Thus the destination offset for (x,y) is the sum of:
302    *   (x % column_width)                    // position within column
303    *   (x / column_width) * bytes_per_column // column number * bytes per column
304    *   y * column_width
305    *
306    * The copy destination offset for each range copied is the sum of
307    * an X offset 'xo0' or 'xo' and a Y offset 'yo.'
308    */
309   const uint32_t column_width = ytile_span;
310   const uint32_t bytes_per_column = column_width * ytile_height;
311
312   uint32_t y1 = MIN2(y3, ALIGN_UP(y0, 4));
313   uint32_t y2 = MAX2(y1, ALIGN_DOWN(y3, 4));
314
315   uint32_t xo0 = (x0 % ytile_span) + (x0 / ytile_span) * bytes_per_column;
316   uint32_t xo1 = (x1 % ytile_span) + (x1 / ytile_span) * bytes_per_column;
317
318   /* Bit 9 of the destination offset control swizzling.
319    * Only the X offset contributes to bit 9 of the total offset,
320    * so swizzle can be calculated in advance for these X positions.
321    * Move bit 9 three places down to bit 6.
322    */
323   uint32_t swizzle0 = (xo0 >> 3) & swizzle_bit;
324   uint32_t swizzle1 = (xo1 >> 3) & swizzle_bit;
325
326   uint32_t x, yo;
327
328   src += (ptrdiff_t)y0 * src_pitch;
329
330   if (y0 != y1) {
331      for (yo = y0 * column_width; yo < y1 * column_width; yo += column_width) {
332         uint32_t xo = xo1;
333         uint32_t swizzle = swizzle1;
334
335         mem_copy(dst + ((xo0 + yo) ^ swizzle0), src + x0, x1 - x0);
336
337         /* Step by spans/columns.  As it happens, the swizzle bit flips
338          * at each step so we don't need to calculate it explicitly.
339          */
340         for (x = x1; x < x2; x += ytile_span) {
341            mem_copy_align16(dst + ((xo + yo) ^ swizzle), src + x, ytile_span);
342            xo += bytes_per_column;
343            swizzle ^= swizzle_bit;
344         }
345
346         mem_copy_align16(dst + ((xo + yo) ^ swizzle), src + x2, x3 - x2);
347
348         src += src_pitch;
349      }
350   }
351
352   for (yo = y1 * column_width; yo < y2 * column_width; yo += 4 * column_width) {
353      uint32_t xo = xo1;
354      uint32_t swizzle = swizzle1;
355
356      if (x0 != x1) {
357         mem_copy(dst + ((xo0 + yo + 0 * column_width) ^ swizzle0), src + x0 + 0 * src_pitch, x1 - x0);
358         mem_copy(dst + ((xo0 + yo + 1 * column_width) ^ swizzle0), src + x0 + 1 * src_pitch, x1 - x0);
359         mem_copy(dst + ((xo0 + yo + 2 * column_width) ^ swizzle0), src + x0 + 2 * src_pitch, x1 - x0);
360         mem_copy(dst + ((xo0 + yo + 3 * column_width) ^ swizzle0), src + x0 + 3 * src_pitch, x1 - x0);
361      }
362
363      /* Step by spans/columns.  As it happens, the swizzle bit flips
364       * at each step so we don't need to calculate it explicitly.
365       */
366      for (x = x1; x < x2; x += ytile_span) {
367         mem_copy_align16(dst + ((xo + yo + 0 * column_width) ^ swizzle), src + x + 0 * src_pitch, ytile_span);
368         mem_copy_align16(dst + ((xo + yo + 1 * column_width) ^ swizzle), src + x + 1 * src_pitch, ytile_span);
369         mem_copy_align16(dst + ((xo + yo + 2 * column_width) ^ swizzle), src + x + 2 * src_pitch, ytile_span);
370         mem_copy_align16(dst + ((xo + yo + 3 * column_width) ^ swizzle), src + x + 3 * src_pitch, ytile_span);
371         xo += bytes_per_column;
372         swizzle ^= swizzle_bit;
373      }
374
375      if (x2 != x3) {
376         mem_copy_align16(dst + ((xo + yo + 0 * column_width) ^ swizzle), src + x2 + 0 * src_pitch, x3 - x2);
377         mem_copy_align16(dst + ((xo + yo + 1 * column_width) ^ swizzle), src + x2 + 1 * src_pitch, x3 - x2);
378         mem_copy_align16(dst + ((xo + yo + 2 * column_width) ^ swizzle), src + x2 + 2 * src_pitch, x3 - x2);
379         mem_copy_align16(dst + ((xo + yo + 3 * column_width) ^ swizzle), src + x2 + 3 * src_pitch, x3 - x2);
380      }
381
382      src += 4 * src_pitch;
383   }
384
385   if (y2 != y3) {
386      for (yo = y2 * column_width; yo < y3 * column_width; yo += column_width) {
387         uint32_t xo = xo1;
388         uint32_t swizzle = swizzle1;
389
390         mem_copy(dst + ((xo0 + yo) ^ swizzle0), src + x0, x1 - x0);
391
392         /* Step by spans/columns.  As it happens, the swizzle bit flips
393          * at each step so we don't need to calculate it explicitly.
394          */
395         for (x = x1; x < x2; x += ytile_span) {
396            mem_copy_align16(dst + ((xo + yo) ^ swizzle), src + x, ytile_span);
397            xo += bytes_per_column;
398            swizzle ^= swizzle_bit;
399         }
400
401         mem_copy_align16(dst + ((xo + yo) ^ swizzle), src + x2, x3 - x2);
402
403         src += src_pitch;
404      }
405   }
406}
407
408/**
409 * Copy texture data from X tile layout to linear.
410 *
411 * \copydoc tile_copy_fn
412 */
413static inline void
414xtiled_to_linear(uint32_t x0, uint32_t x1, uint32_t x2, uint32_t x3,
415                 uint32_t y0, uint32_t y1,
416                 char *dst, const char *src,
417                 int32_t dst_pitch,
418                 uint32_t swizzle_bit,
419                 isl_mem_copy_fn mem_copy,
420                 isl_mem_copy_fn mem_copy_align16)
421{
422   /* The copy destination offset for each range copied is the sum of
423    * an X offset 'x0' or 'xo' and a Y offset 'yo.'
424    */
425   uint32_t xo, yo;
426
427   dst += (ptrdiff_t)y0 * dst_pitch;
428
429   for (yo = y0 * xtile_width; yo < y1 * xtile_width; yo += xtile_width) {
430      /* Bits 9 and 10 of the copy destination offset control swizzling.
431       * Only 'yo' contributes to those bits in the total offset,
432       * so calculate 'swizzle' just once per row.
433       * Move bits 9 and 10 three and four places respectively down
434       * to bit 6 and xor them.
435       */
436      uint32_t swizzle = ((yo >> 3) ^ (yo >> 4)) & swizzle_bit;
437
438      mem_copy(dst + x0, src + ((x0 + yo) ^ swizzle), x1 - x0);
439
440      for (xo = x1; xo < x2; xo += xtile_span) {
441         mem_copy_align16(dst + xo, src + ((xo + yo) ^ swizzle), xtile_span);
442      }
443
444      mem_copy_align16(dst + x2, src + ((xo + yo) ^ swizzle), x3 - x2);
445
446      dst += dst_pitch;
447   }
448}
449
450 /**
451 * Copy texture data from Y tile layout to linear.
452 *
453 * \copydoc tile_copy_fn
454 */
455static inline void
456ytiled_to_linear(uint32_t x0, uint32_t x1, uint32_t x2, uint32_t x3,
457                 uint32_t y0, uint32_t y3,
458                 char *dst, const char *src,
459                 int32_t dst_pitch,
460                 uint32_t swizzle_bit,
461                 isl_mem_copy_fn mem_copy,
462                 isl_mem_copy_fn mem_copy_align16)
463{
464   /* Y tiles consist of columns that are 'ytile_span' wide (and the same height
465    * as the tile).  Thus the destination offset for (x,y) is the sum of:
466    *   (x % column_width)                    // position within column
467    *   (x / column_width) * bytes_per_column // column number * bytes per column
468    *   y * column_width
469    *
470    * The copy destination offset for each range copied is the sum of
471    * an X offset 'xo0' or 'xo' and a Y offset 'yo.'
472    */
473   const uint32_t column_width = ytile_span;
474   const uint32_t bytes_per_column = column_width * ytile_height;
475
476   uint32_t y1 = MIN2(y3, ALIGN_UP(y0, 4));
477   uint32_t y2 = MAX2(y1, ALIGN_DOWN(y3, 4));
478
479   uint32_t xo0 = (x0 % ytile_span) + (x0 / ytile_span) * bytes_per_column;
480   uint32_t xo1 = (x1 % ytile_span) + (x1 / ytile_span) * bytes_per_column;
481
482   /* Bit 9 of the destination offset control swizzling.
483    * Only the X offset contributes to bit 9 of the total offset,
484    * so swizzle can be calculated in advance for these X positions.
485    * Move bit 9 three places down to bit 6.
486    */
487   uint32_t swizzle0 = (xo0 >> 3) & swizzle_bit;
488   uint32_t swizzle1 = (xo1 >> 3) & swizzle_bit;
489
490   uint32_t x, yo;
491
492   dst += (ptrdiff_t)y0 * dst_pitch;
493
494   if (y0 != y1) {
495      for (yo = y0 * column_width; yo < y1 * column_width; yo += column_width) {
496         uint32_t xo = xo1;
497         uint32_t swizzle = swizzle1;
498
499         mem_copy(dst + x0, src + ((xo0 + yo) ^ swizzle0), x1 - x0);
500
501         /* Step by spans/columns.  As it happens, the swizzle bit flips
502          * at each step so we don't need to calculate it explicitly.
503          */
504         for (x = x1; x < x2; x += ytile_span) {
505            mem_copy_align16(dst + x, src + ((xo + yo) ^ swizzle), ytile_span);
506            xo += bytes_per_column;
507            swizzle ^= swizzle_bit;
508         }
509
510         mem_copy_align16(dst + x2, src + ((xo + yo) ^ swizzle), x3 - x2);
511
512         dst += dst_pitch;
513      }
514   }
515
516   for (yo = y1 * column_width; yo < y2 * column_width; yo += 4 * column_width) {
517      uint32_t xo = xo1;
518      uint32_t swizzle = swizzle1;
519
520      if (x0 != x1) {
521         mem_copy(dst + x0 + 0 * dst_pitch, src + ((xo0 + yo + 0 * column_width) ^ swizzle0), x1 - x0);
522         mem_copy(dst + x0 + 1 * dst_pitch, src + ((xo0 + yo + 1 * column_width) ^ swizzle0), x1 - x0);
523         mem_copy(dst + x0 + 2 * dst_pitch, src + ((xo0 + yo + 2 * column_width) ^ swizzle0), x1 - x0);
524         mem_copy(dst + x0 + 3 * dst_pitch, src + ((xo0 + yo + 3 * column_width) ^ swizzle0), x1 - x0);
525      }
526
527      /* Step by spans/columns.  As it happens, the swizzle bit flips
528       * at each step so we don't need to calculate it explicitly.
529       */
530      for (x = x1; x < x2; x += ytile_span) {
531         mem_copy_align16(dst + x + 0 * dst_pitch, src + ((xo + yo + 0 * column_width) ^ swizzle), ytile_span);
532         mem_copy_align16(dst + x + 1 * dst_pitch, src + ((xo + yo + 1 * column_width) ^ swizzle), ytile_span);
533         mem_copy_align16(dst + x + 2 * dst_pitch, src + ((xo + yo + 2 * column_width) ^ swizzle), ytile_span);
534         mem_copy_align16(dst + x + 3 * dst_pitch, src + ((xo + yo + 3 * column_width) ^ swizzle), ytile_span);
535         xo += bytes_per_column;
536         swizzle ^= swizzle_bit;
537      }
538
539      if (x2 != x3) {
540         mem_copy_align16(dst + x2 + 0 * dst_pitch, src + ((xo + yo + 0 * column_width) ^ swizzle), x3 - x2);
541         mem_copy_align16(dst + x2 + 1 * dst_pitch, src + ((xo + yo + 1 * column_width) ^ swizzle), x3 - x2);
542         mem_copy_align16(dst + x2 + 2 * dst_pitch, src + ((xo + yo + 2 * column_width) ^ swizzle), x3 - x2);
543         mem_copy_align16(dst + x2 + 3 * dst_pitch, src + ((xo + yo + 3 * column_width) ^ swizzle), x3 - x2);
544      }
545
546      dst += 4 * dst_pitch;
547   }
548
549   if (y2 != y3) {
550      for (yo = y2 * column_width; yo < y3 * column_width; yo += column_width) {
551         uint32_t xo = xo1;
552         uint32_t swizzle = swizzle1;
553
554         mem_copy(dst + x0, src + ((xo0 + yo) ^ swizzle0), x1 - x0);
555
556         /* Step by spans/columns.  As it happens, the swizzle bit flips
557          * at each step so we don't need to calculate it explicitly.
558          */
559         for (x = x1; x < x2; x += ytile_span) {
560            mem_copy_align16(dst + x, src + ((xo + yo) ^ swizzle), ytile_span);
561            xo += bytes_per_column;
562            swizzle ^= swizzle_bit;
563         }
564
565         mem_copy_align16(dst + x2, src + ((xo + yo) ^ swizzle), x3 - x2);
566
567         dst += dst_pitch;
568      }
569   }
570}
571
572#if defined(INLINE_SSE41)
573static ALWAYS_INLINE void *
574_memcpy_streaming_load(void *dest, const void *src, size_t count)
575{
576   if (count == 16) {
577      __m128i val = _mm_stream_load_si128((__m128i *)src);
578      _mm_storeu_si128((__m128i *)dest, val);
579      return dest;
580   } else if (count == 64) {
581      __m128i val0 = _mm_stream_load_si128(((__m128i *)src) + 0);
582      __m128i val1 = _mm_stream_load_si128(((__m128i *)src) + 1);
583      __m128i val2 = _mm_stream_load_si128(((__m128i *)src) + 2);
584      __m128i val3 = _mm_stream_load_si128(((__m128i *)src) + 3);
585      _mm_storeu_si128(((__m128i *)dest) + 0, val0);
586      _mm_storeu_si128(((__m128i *)dest) + 1, val1);
587      _mm_storeu_si128(((__m128i *)dest) + 2, val2);
588      _mm_storeu_si128(((__m128i *)dest) + 3, val3);
589      return dest;
590   } else {
591      assert(count < 64); /* and (count < 16) for ytiled */
592      return memcpy(dest, src, count);
593   }
594}
595#endif
596
597static isl_mem_copy_fn
598choose_copy_function(isl_memcpy_type copy_type)
599{
600   switch(copy_type) {
601   case ISL_MEMCPY:
602      return memcpy;
603   case ISL_MEMCPY_BGRA8:
604      return rgba8_copy;
605   case ISL_MEMCPY_STREAMING_LOAD:
606#if defined(INLINE_SSE41)
607      return _memcpy_streaming_load;
608#else
609      unreachable("ISL_MEMCOPY_STREAMING_LOAD requires sse4.1");
610#endif
611   case ISL_MEMCPY_INVALID:
612      unreachable("invalid copy_type");
613   }
614   unreachable("unhandled copy_type");
615   return NULL;
616}
617
618/**
619 * Copy texture data from linear to X tile layout, faster.
620 *
621 * Same as \ref linear_to_xtiled but faster, because it passes constant
622 * parameters for common cases, allowing the compiler to inline code
623 * optimized for those cases.
624 *
625 * \copydoc tile_copy_fn
626 */
627static FLATTEN void
628linear_to_xtiled_faster(uint32_t x0, uint32_t x1, uint32_t x2, uint32_t x3,
629                        uint32_t y0, uint32_t y1,
630                        char *dst, const char *src,
631                        int32_t src_pitch,
632                        uint32_t swizzle_bit,
633                        isl_memcpy_type copy_type)
634{
635   isl_mem_copy_fn mem_copy = choose_copy_function(copy_type);
636
637   if (x0 == 0 && x3 == xtile_width && y0 == 0 && y1 == xtile_height) {
638      if (mem_copy == memcpy)
639         return linear_to_xtiled(0, 0, xtile_width, xtile_width, 0, xtile_height,
640                                 dst, src, src_pitch, swizzle_bit, memcpy, memcpy);
641      else if (mem_copy == rgba8_copy)
642         return linear_to_xtiled(0, 0, xtile_width, xtile_width, 0, xtile_height,
643                                 dst, src, src_pitch, swizzle_bit,
644                                 rgba8_copy, rgba8_copy_aligned_dst);
645      else
646         unreachable("not reached");
647   } else {
648      if (mem_copy == memcpy)
649         return linear_to_xtiled(x0, x1, x2, x3, y0, y1,
650                                 dst, src, src_pitch, swizzle_bit,
651                                 memcpy, memcpy);
652      else if (mem_copy == rgba8_copy)
653         return linear_to_xtiled(x0, x1, x2, x3, y0, y1,
654                                 dst, src, src_pitch, swizzle_bit,
655                                 rgba8_copy, rgba8_copy_aligned_dst);
656      else
657         unreachable("not reached");
658   }
659   linear_to_xtiled(x0, x1, x2, x3, y0, y1,
660                    dst, src, src_pitch, swizzle_bit, mem_copy, mem_copy);
661}
662
663/**
664 * Copy texture data from linear to Y tile layout, faster.
665 *
666 * Same as \ref linear_to_ytiled but faster, because it passes constant
667 * parameters for common cases, allowing the compiler to inline code
668 * optimized for those cases.
669 *
670 * \copydoc tile_copy_fn
671 */
672static FLATTEN void
673linear_to_ytiled_faster(uint32_t x0, uint32_t x1, uint32_t x2, uint32_t x3,
674                        uint32_t y0, uint32_t y1,
675                        char *dst, const char *src,
676                        int32_t src_pitch,
677                        uint32_t swizzle_bit,
678                        isl_memcpy_type copy_type)
679{
680   isl_mem_copy_fn mem_copy = choose_copy_function(copy_type);
681
682   if (x0 == 0 && x3 == ytile_width && y0 == 0 && y1 == ytile_height) {
683      if (mem_copy == memcpy)
684         return linear_to_ytiled(0, 0, ytile_width, ytile_width, 0, ytile_height,
685                                 dst, src, src_pitch, swizzle_bit, memcpy, memcpy);
686      else if (mem_copy == rgba8_copy)
687         return linear_to_ytiled(0, 0, ytile_width, ytile_width, 0, ytile_height,
688                                 dst, src, src_pitch, swizzle_bit,
689                                 rgba8_copy, rgba8_copy_aligned_dst);
690      else
691         unreachable("not reached");
692   } else {
693      if (mem_copy == memcpy)
694         return linear_to_ytiled(x0, x1, x2, x3, y0, y1,
695                                 dst, src, src_pitch, swizzle_bit, memcpy, memcpy);
696      else if (mem_copy == rgba8_copy)
697         return linear_to_ytiled(x0, x1, x2, x3, y0, y1,
698                                 dst, src, src_pitch, swizzle_bit,
699                                 rgba8_copy, rgba8_copy_aligned_dst);
700      else
701         unreachable("not reached");
702   }
703   linear_to_ytiled(x0, x1, x2, x3, y0, y1,
704                    dst, src, src_pitch, swizzle_bit, mem_copy, mem_copy);
705}
706
707/**
708 * Copy texture data from X tile layout to linear, faster.
709 *
710 * Same as \ref xtile_to_linear but faster, because it passes constant
711 * parameters for common cases, allowing the compiler to inline code
712 * optimized for those cases.
713 *
714 * \copydoc tile_copy_fn
715 */
716static FLATTEN void
717xtiled_to_linear_faster(uint32_t x0, uint32_t x1, uint32_t x2, uint32_t x3,
718                        uint32_t y0, uint32_t y1,
719                        char *dst, const char *src,
720                        int32_t dst_pitch,
721                        uint32_t swizzle_bit,
722                        isl_memcpy_type copy_type)
723{
724   isl_mem_copy_fn mem_copy = choose_copy_function(copy_type);
725
726   if (x0 == 0 && x3 == xtile_width && y0 == 0 && y1 == xtile_height) {
727      if (mem_copy == memcpy)
728         return xtiled_to_linear(0, 0, xtile_width, xtile_width, 0, xtile_height,
729                                 dst, src, dst_pitch, swizzle_bit, memcpy, memcpy);
730      else if (mem_copy == rgba8_copy)
731         return xtiled_to_linear(0, 0, xtile_width, xtile_width, 0, xtile_height,
732                                 dst, src, dst_pitch, swizzle_bit,
733                                 rgba8_copy, rgba8_copy_aligned_src);
734#if defined(INLINE_SSE41)
735      else if (mem_copy == _memcpy_streaming_load)
736         return xtiled_to_linear(0, 0, xtile_width, xtile_width, 0, xtile_height,
737                                 dst, src, dst_pitch, swizzle_bit,
738                                 memcpy, _memcpy_streaming_load);
739#endif
740      else
741         unreachable("not reached");
742   } else {
743      if (mem_copy == memcpy)
744         return xtiled_to_linear(x0, x1, x2, x3, y0, y1,
745                                 dst, src, dst_pitch, swizzle_bit, memcpy, memcpy);
746      else if (mem_copy == rgba8_copy)
747         return xtiled_to_linear(x0, x1, x2, x3, y0, y1,
748                                 dst, src, dst_pitch, swizzle_bit,
749                                 rgba8_copy, rgba8_copy_aligned_src);
750#if defined(INLINE_SSE41)
751      else if (mem_copy == _memcpy_streaming_load)
752         return xtiled_to_linear(x0, x1, x2, x3, y0, y1,
753                                 dst, src, dst_pitch, swizzle_bit,
754                                 memcpy, _memcpy_streaming_load);
755#endif
756      else
757         unreachable("not reached");
758   }
759   xtiled_to_linear(x0, x1, x2, x3, y0, y1,
760                    dst, src, dst_pitch, swizzle_bit, mem_copy, mem_copy);
761}
762
763/**
764 * Copy texture data from Y tile layout to linear, faster.
765 *
766 * Same as \ref ytile_to_linear but faster, because it passes constant
767 * parameters for common cases, allowing the compiler to inline code
768 * optimized for those cases.
769 *
770 * \copydoc tile_copy_fn
771 */
772static FLATTEN void
773ytiled_to_linear_faster(uint32_t x0, uint32_t x1, uint32_t x2, uint32_t x3,
774                        uint32_t y0, uint32_t y1,
775                        char *dst, const char *src,
776                        int32_t dst_pitch,
777                        uint32_t swizzle_bit,
778                        isl_memcpy_type copy_type)
779{
780   isl_mem_copy_fn mem_copy = choose_copy_function(copy_type);
781
782   if (x0 == 0 && x3 == ytile_width && y0 == 0 && y1 == ytile_height) {
783      if (mem_copy == memcpy)
784         return ytiled_to_linear(0, 0, ytile_width, ytile_width, 0, ytile_height,
785                                 dst, src, dst_pitch, swizzle_bit, memcpy, memcpy);
786      else if (mem_copy == rgba8_copy)
787         return ytiled_to_linear(0, 0, ytile_width, ytile_width, 0, ytile_height,
788                                 dst, src, dst_pitch, swizzle_bit,
789                                 rgba8_copy, rgba8_copy_aligned_src);
790#if defined(INLINE_SSE41)
791      else if (copy_type == ISL_MEMCPY_STREAMING_LOAD)
792         return ytiled_to_linear(0, 0, ytile_width, ytile_width, 0, ytile_height,
793                                 dst, src, dst_pitch, swizzle_bit,
794                                 memcpy, _memcpy_streaming_load);
795#endif
796      else
797         unreachable("not reached");
798   } else {
799      if (mem_copy == memcpy)
800         return ytiled_to_linear(x0, x1, x2, x3, y0, y1,
801                                 dst, src, dst_pitch, swizzle_bit, memcpy, memcpy);
802      else if (mem_copy == rgba8_copy)
803         return ytiled_to_linear(x0, x1, x2, x3, y0, y1,
804                                 dst, src, dst_pitch, swizzle_bit,
805                                 rgba8_copy, rgba8_copy_aligned_src);
806#if defined(INLINE_SSE41)
807      else if (copy_type == ISL_MEMCPY_STREAMING_LOAD)
808         return ytiled_to_linear(x0, x1, x2, x3, y0, y1,
809                                 dst, src, dst_pitch, swizzle_bit,
810                                 memcpy, _memcpy_streaming_load);
811#endif
812      else
813         unreachable("not reached");
814   }
815   ytiled_to_linear(x0, x1, x2, x3, y0, y1,
816                    dst, src, dst_pitch, swizzle_bit, mem_copy, mem_copy);
817}
818
819/**
820 * Copy from linear to tiled texture.
821 *
822 * Divide the region given by X range [xt1, xt2) and Y range [yt1, yt2) into
823 * pieces that do not cross tile boundaries and copy each piece with a tile
824 * copy function (\ref tile_copy_fn).
825 * The X range is in bytes, i.e. pixels * bytes-per-pixel.
826 * The Y range is in pixels (i.e. unitless).
827 * 'dst' is the address of (0, 0) in the destination tiled texture.
828 * 'src' is the address of (xt1, yt1) in the source linear texture.
829 */
830static void
831linear_to_tiled(uint32_t xt1, uint32_t xt2,
832                      uint32_t yt1, uint32_t yt2,
833                      char *dst, const char *src,
834                      uint32_t dst_pitch, int32_t src_pitch,
835                      bool has_swizzling,
836                      enum isl_tiling tiling,
837                      isl_memcpy_type copy_type)
838{
839   tile_copy_fn tile_copy;
840   uint32_t xt0, xt3;
841   uint32_t yt0, yt3;
842   uint32_t xt, yt;
843   uint32_t tw, th, span;
844   uint32_t swizzle_bit = has_swizzling ? 1<<6 : 0;
845
846   if (tiling == ISL_TILING_X) {
847      tw = xtile_width;
848      th = xtile_height;
849      span = xtile_span;
850      tile_copy = linear_to_xtiled_faster;
851   } else if (tiling == ISL_TILING_Y0) {
852      tw = ytile_width;
853      th = ytile_height;
854      span = ytile_span;
855      tile_copy = linear_to_ytiled_faster;
856   } else {
857      unreachable("unsupported tiling");
858   }
859
860   /* Round out to tile boundaries. */
861   xt0 = ALIGN_DOWN(xt1, tw);
862   xt3 = ALIGN_UP  (xt2, tw);
863   yt0 = ALIGN_DOWN(yt1, th);
864   yt3 = ALIGN_UP  (yt2, th);
865
866   /* Loop over all tiles to which we have something to copy.
867    * 'xt' and 'yt' are the origin of the destination tile, whether copying
868    * copying a full or partial tile.
869    * tile_copy() copies one tile or partial tile.
870    * Looping x inside y is the faster memory access pattern.
871    */
872   for (yt = yt0; yt < yt3; yt += th) {
873      for (xt = xt0; xt < xt3; xt += tw) {
874         /* The area to update is [x0,x3) x [y0,y1).
875          * May not want the whole tile, hence the min and max.
876          */
877         uint32_t x0 = MAX2(xt1, xt);
878         uint32_t y0 = MAX2(yt1, yt);
879         uint32_t x3 = MIN2(xt2, xt + tw);
880         uint32_t y1 = MIN2(yt2, yt + th);
881
882         /* [x0,x3) is split into [x0,x1), [x1,x2), [x2,x3) such that
883          * the middle interval is the longest span-aligned part.
884          * The sub-ranges could be empty.
885          */
886         uint32_t x1, x2;
887         x1 = ALIGN_UP(x0, span);
888         if (x1 > x3)
889            x1 = x2 = x3;
890         else
891            x2 = ALIGN_DOWN(x3, span);
892
893         assert(x0 <= x1 && x1 <= x2 && x2 <= x3);
894         assert(x1 - x0 < span && x3 - x2 < span);
895         assert(x3 - x0 <= tw);
896         assert((x2 - x1) % span == 0);
897
898         /* Translate by (xt,yt) for single-tile copier. */
899         tile_copy(x0-xt, x1-xt, x2-xt, x3-xt,
900                   y0-yt, y1-yt,
901                   dst + (ptrdiff_t)xt * th  +  (ptrdiff_t)yt        * dst_pitch,
902                   src + (ptrdiff_t)xt - xt1 + ((ptrdiff_t)yt - yt1) * src_pitch,
903                   src_pitch,
904                   swizzle_bit,
905                   copy_type);
906      }
907   }
908}
909
910/**
911 * Copy from tiled to linear texture.
912 *
913 * Divide the region given by X range [xt1, xt2) and Y range [yt1, yt2) into
914 * pieces that do not cross tile boundaries and copy each piece with a tile
915 * copy function (\ref tile_copy_fn).
916 * The X range is in bytes, i.e. pixels * bytes-per-pixel.
917 * The Y range is in pixels (i.e. unitless).
918 * 'dst' is the address of (xt1, yt1) in the destination linear texture.
919 * 'src' is the address of (0, 0) in the source tiled texture.
920 */
921static void
922tiled_to_linear(uint32_t xt1, uint32_t xt2,
923                      uint32_t yt1, uint32_t yt2,
924                      char *dst, const char *src,
925                      int32_t dst_pitch, uint32_t src_pitch,
926                      bool has_swizzling,
927                      enum isl_tiling tiling,
928                      isl_memcpy_type copy_type)
929{
930   tile_copy_fn tile_copy;
931   uint32_t xt0, xt3;
932   uint32_t yt0, yt3;
933   uint32_t xt, yt;
934   uint32_t tw, th, span;
935   uint32_t swizzle_bit = has_swizzling ? 1<<6 : 0;
936
937   if (tiling == ISL_TILING_X) {
938      tw = xtile_width;
939      th = xtile_height;
940      span = xtile_span;
941      tile_copy = xtiled_to_linear_faster;
942   } else if (tiling == ISL_TILING_Y0) {
943      tw = ytile_width;
944      th = ytile_height;
945      span = ytile_span;
946      tile_copy = ytiled_to_linear_faster;
947   } else {
948      unreachable("unsupported tiling");
949   }
950
951#if defined(INLINE_SSE41)
952   if (copy_type == ISL_MEMCPY_STREAMING_LOAD) {
953      /* The hidden cacheline sized register used by movntdqa can apparently
954       * give you stale data, so do an mfence to invalidate it.
955       */
956      _mm_mfence();
957   }
958#endif
959
960   /* Round out to tile boundaries. */
961   xt0 = ALIGN_DOWN(xt1, tw);
962   xt3 = ALIGN_UP  (xt2, tw);
963   yt0 = ALIGN_DOWN(yt1, th);
964   yt3 = ALIGN_UP  (yt2, th);
965
966   /* Loop over all tiles to which we have something to copy.
967    * 'xt' and 'yt' are the origin of the destination tile, whether copying
968    * copying a full or partial tile.
969    * tile_copy() copies one tile or partial tile.
970    * Looping x inside y is the faster memory access pattern.
971    */
972   for (yt = yt0; yt < yt3; yt += th) {
973      for (xt = xt0; xt < xt3; xt += tw) {
974         /* The area to update is [x0,x3) x [y0,y1).
975          * May not want the whole tile, hence the min and max.
976          */
977         uint32_t x0 = MAX2(xt1, xt);
978         uint32_t y0 = MAX2(yt1, yt);
979         uint32_t x3 = MIN2(xt2, xt + tw);
980         uint32_t y1 = MIN2(yt2, yt + th);
981
982         /* [x0,x3) is split into [x0,x1), [x1,x2), [x2,x3) such that
983          * the middle interval is the longest span-aligned part.
984          * The sub-ranges could be empty.
985          */
986         uint32_t x1, x2;
987         x1 = ALIGN_UP(x0, span);
988         if (x1 > x3)
989            x1 = x2 = x3;
990         else
991            x2 = ALIGN_DOWN(x3, span);
992
993         assert(x0 <= x1 && x1 <= x2 && x2 <= x3);
994         assert(x1 - x0 < span && x3 - x2 < span);
995         assert(x3 - x0 <= tw);
996         assert((x2 - x1) % span == 0);
997
998         /* Translate by (xt,yt) for single-tile copier. */
999         tile_copy(x0-xt, x1-xt, x2-xt, x3-xt,
1000                   y0-yt, y1-yt,
1001                   dst + (ptrdiff_t)xt - xt1 + ((ptrdiff_t)yt - yt1) * dst_pitch,
1002                   src + (ptrdiff_t)xt * th  +  (ptrdiff_t)yt        * src_pitch,
1003                   dst_pitch,
1004                   swizzle_bit,
1005                   copy_type);
1006      }
1007   }
1008}
1009