src/gallium/drivers/llvmpipe/lp_tile_soa.py

   1 #!/usr/bin/env python
   2
   3 '''
   4 /**************************************************************************
   5  *
   6  * Copyright 2009 VMware, Inc.
   7  * All Rights Reserved.
   8  *
   9  * Permission is hereby granted, free of charge, to any person obtaining a
  10  * copy of this software and associated documentation files (the
  11  * "Software"), to deal in the Software without restriction, including
  12  * without limitation the rights to use, copy, modify, merge, publish,
  13  * distribute, sub license, and/or sell copies of the Software, and to
  14  * permit persons to whom the Software is furnished to do so, subject to
  15  * the following conditions:
  16  *
  17  * The above copyright notice and this permission notice (including the
  18  * next paragraph) shall be included in all copies or substantial portions
  19  * of the Software.
  20  *
  21  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  22  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  23  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  24  * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
  25  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  26  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  27  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  28  *
  29  **************************************************************************/
  30
  31 /**
  32  * @file
  33  * Pixel format accessor functions.
  34  *
  35  * @author Jose Fonseca <jfonseca@vmware.com>
  36  */
  37 '''
  38
  39
  40 import sys
  41 import os.path
  42
  43 sys.path.insert(0, os.path.join(os.path.dirname(sys.argv[0]), '../../auxiliary/util'))
  44
  45 from u_format_pack import *
  46
  47
  48 def is_format_supported(format):
  49     '''Determines whether we actually have the plumbing necessary to generate the
  50     to read/write to/from this format.'''
  51
  52     # FIXME: Ideally we would support any format combination here.
  53
  54     if format.layout != PLAIN:
  55         return False
  56
  57     for i in range(4):
  58         channel = format.channels[i]
  59         if channel.type not in (VOID, UNSIGNED, SIGNED, FLOAT):
  60             return False
  61         if channel.type == FLOAT and channel.size not in (16, 32 ,64):
  62             return False
  63
  64     if format.colorspace not in ('rgb', 'srgb'):
  65         return False
  66
  67     return True
  68
  69
  70 def generate_format_read(format, dst_channel, dst_native_type, dst_suffix):
  71     '''Generate the function to read pixels from a particular format'''
  72
  73     name = format.short_name()
  74
  75     src_native_type = native_type(format)
  76
  77     print 'static void'
  78     print 'lp_tile_%s_swizzle_%s(%s *dst, const uint8_t *src, unsigned src_stride, unsigned x0, unsigned y0)' % (name, dst_suffix, dst_native_type)
  79     print '{'
  80     print '   unsigned x, y;'
  81     print '   const uint8_t *src_row = src + y0*src_stride;'
  82     print '   for (y = 0; y < TILE_SIZE; ++y) {'
  83     print '      const %s *src_pixel = (const %s *)(src_row + x0*%u);' % (src_native_type, src_native_type, format.stride())
  84     print '      for (x = 0; x < TILE_SIZE; ++x) {'
  85
  86     names = ['']*4
  87     if format.colorspace in ('rgb', 'srgb'):
  88         for i in range(4):
  89             swizzle = format.swizzles[i]
  90             if swizzle < 4:
  91                 names[swizzle] += 'rgba'[i]
  92     elif format.colorspace == 'zs':
  93         swizzle = format.swizzles[0]
  94         if swizzle < 4:
  95             names[swizzle] = 'z'
  96         else:
  97             assert False
  98     else:
  99         assert False
 100
 101     if format.layout == PLAIN:
 102         if not format.is_array():
 103             print '         %s pixel = *src_pixel++;' % src_native_type
 104             shift = 0;
 105             for i in range(4):
 106                 src_channel = format.channels[i]
 107                 width = src_channel.size
 108                 if names[i]:
 109                     value = 'pixel'
 110                     mask = (1 << width) - 1
 111                     if shift:
 112                         value = '(%s >> %u)' % (value, shift)
 113                     if shift + width < format.block_size():
 114                         value = '(%s & 0x%x)' % (value, mask)
 115                     value = conversion_expr(src_channel, dst_channel, dst_native_type, value, clamp=False)
 116                     print '         %s %s = %s;' % (dst_native_type, names[i], value)
 117                 shift += width
 118         else:
 119             for i in range(4):
 120                 if names[i]:
 121                     print '         %s %s;' % (dst_native_type, names[i])
 122             for i in range(4):
 123                 src_channel = format.channels[i]
 124                 if names[i]:
 125                     value = '(*src_pixel++)'
 126                     value = conversion_expr(src_channel, dst_channel, dst_native_type, value, clamp=False)
 127                     print '         %s = %s;' % (names[i], value)
 128                 elif src_channel.size:
 129                     print '         ++src_pixel;'
 130     else:
 131         assert False
 132
 133     for i in range(4):
 134         if format.colorspace in ('rgb', 'srgb'):
 135             swizzle = format.swizzles[i]
 136             if swizzle < 4:
 137                 value = names[swizzle]
 138             elif swizzle == SWIZZLE_0:
 139                 value = '0'
 140             elif swizzle == SWIZZLE_1:
 141                 value = get_one(dst_channel)
 142             else:
 143                 assert False
 144         elif format.colorspace == 'zs':
 145             if i < 3:
 146                 value = 'z'
 147             else:
 148                 value = get_one(dst_channel)
 149         else:
 150             assert False
 151         print '         TILE_PIXEL(dst, x, y, %u) = %s; /* %s */' % (i, value, 'rgba'[i])
 152
 153     print '      }'
 154     print '      src_row += src_stride;'
 155     print '   }'
 156     print '}'
 157     print
 158
 159
 160 def pack_rgba(format, src_channel, r, g, b, a):
 161     """Return an expression for packing r, g, b, a into a pixel of the
 162     given format.  Ex: '(b << 24) | (g << 16) | (r << 8) | (a << 0)'
 163     """
 164     assert format.colorspace in ('rgb', 'srgb')
 165     inv_swizzle = format.inv_swizzles()
 166     shift = 0
 167     expr = None
 168     for i in range(4):
 169         # choose r, g, b, or a depending on the inverse swizzle term
 170         if inv_swizzle[i] == 0:
 171             value = r
 172         elif inv_swizzle[i] == 1:
 173             value = g
 174         elif inv_swizzle[i] == 2:
 175             value = b
 176         elif inv_swizzle[i] == 3:
 177             value = a
 178         else:
 179             value = None
 180
 181         if value:
 182             dst_channel = format.channels[i]
 183             dst_native_type = native_type(format)
 184             value = conversion_expr(src_channel, dst_channel, dst_native_type, value, clamp=False)
 185             term = "((%s) << %d)" % (value, shift)
 186             if expr:
 187                 expr = expr + " | " + term
 188             else:
 189                 expr = term
 190
 191         width = format.channels[i].size
 192         shift = shift + width
 193     return expr
 194
 195
 196 def emit_unrolled_unswizzle_code(format, src_channel):
 197     '''Emit code for writing a block based on unrolled loops.
 198     This is considerably faster than the TILE_PIXEL-based code below.
 199     '''
 200     dst_native_type = 'uint%u_t' % format.block_size()
 201     print '   const unsigned dstpix_stride = dst_stride / %d;' % format.stride()
 202     print '   %s *dstpix = (%s *) dst;' % (dst_native_type, dst_native_type)
 203     print '   unsigned int qx, qy, i;'
 204     print
 205     print '   for (qy = 0; qy < TILE_SIZE; qy += TILE_VECTOR_HEIGHT) {'
 206     print '      const unsigned py = y0 + qy;'
 207     print '      for (qx = 0; qx < TILE_SIZE; qx += TILE_VECTOR_WIDTH) {'
 208     print '         const unsigned px = x0 + qx;'
 209     print '         const uint8_t *r = src + 0 * TILE_C_STRIDE;'
 210     print '         const uint8_t *g = src + 1 * TILE_C_STRIDE;'
 211     print '         const uint8_t *b = src + 2 * TILE_C_STRIDE;'
 212     print '         const uint8_t *a = src + 3 * TILE_C_STRIDE;'
 213     print '         (void) r; (void) g; (void) b; (void) a; /* silence warnings */'
 214     print '         for (i = 0; i < TILE_C_STRIDE; i += 2) {'
 215     print '            const uint32_t pixel0 = %s;' % pack_rgba(format, src_channel, "r[i+0]", "g[i+0]", "b[i+0]", "a[i+0]")
 216     print '            const uint32_t pixel1 = %s;' % pack_rgba(format, src_channel, "r[i+1]", "g[i+1]", "b[i+1]", "a[i+1]")
 217     print '            const unsigned offset = (py + tile_y_offset[i]) * dstpix_stride + (px + tile_x_offset[i]);'
 218     print '            dstpix[offset + 0] = pixel0;'
 219     print '            dstpix[offset + 1] = pixel1;'
 220     print '         }'
 221     print '         src += TILE_X_STRIDE;'
 222     print '      }'
 223     print '   }'
 224
 225
 226 def emit_tile_pixel_unswizzle_code(format, src_channel):
 227     '''Emit code for writing a block based on the TILE_PIXEL macro.'''
 228     dst_native_type = native_type(format)
 229
 230     inv_swizzle = format.inv_swizzles()
 231
 232     print '   unsigned x, y;'
 233     print '   uint8_t *dst_row = dst + y0*dst_stride;'
 234     print '   for (y = 0; y < TILE_SIZE; ++y) {'
 235     print '      %s *dst_pixel = (%s *)(dst_row + x0*%u);' % (dst_native_type, dst_native_type, format.stride())
 236     print '      for (x = 0; x < TILE_SIZE; ++x) {'
 237
 238     if format.layout == PLAIN:
 239         if not format.is_array():
 240             print '         %s pixel = 0;' % dst_native_type
 241             shift = 0;
 242             for i in range(4):
 243                 dst_channel = format.channels[i]
 244                 width = dst_channel.size
 245                 if inv_swizzle[i] is not None:
 246                     value = 'TILE_PIXEL(src, x, y, %u)' % inv_swizzle[i]
 247                     value = conversion_expr(src_channel, dst_channel, dst_native_type, value, clamp=False)
 248                     if shift:
 249                         value = '(%s << %u)' % (value, shift)
 250                     print '         pixel |= %s;' % value
 251                 shift += width
 252             print '         *dst_pixel++ = pixel;'
 253         else:
 254             for i in range(4):
 255                 dst_channel = format.channels[i]
 256                 if inv_swizzle[i] is not None:
 257                     value = 'TILE_PIXEL(src, x, y, %u)' % inv_swizzle[i]
 258                     value = conversion_expr(src_channel, dst_channel, dst_native_type, value, clamp=False)
 259                     print '         *dst_pixel++ = %s;' % value
 260                 elif dst_channel.size:
 261                     print '         ++dst_pixel;'
 262     else:
 263         assert False
 264
 265     print '      }'
 266     print '      dst_row += dst_stride;'
 267     print '   }'
 268
 269
 270 def generate_format_write(format, src_channel, src_native_type, src_suffix):
 271     '''Generate the function to write pixels to a particular format'''
 272
 273     name = format.short_name()
 274
 275     print 'static void'
 276     print 'lp_tile_%s_unswizzle_%s(const %s *src, uint8_t *dst, unsigned dst_stride, unsigned x0, unsigned y0)' % (name, src_suffix, src_native_type)
 277     print '{'
 278     if format.layout == PLAIN \
 279         and format.colorspace == 'rgb' \
 280         and format.block_size() <= 32 \
 281         and format.is_pot() \
 282         and not format.is_mixed() \
 283         and (format.channels[0].type == UNSIGNED \
 284              or format.channels[1].type == UNSIGNED):
 285         emit_unrolled_unswizzle_code(format, src_channel)
 286     else:
 287         emit_tile_pixel_unswizzle_code(format, src_channel)
 288     print '}'
 289     print
 290
 291
 292 def generate_ssse3():
 293     print '''
 294 #if defined(PIPE_ARCH_SSE)
 295
 296 #include "util/u_sse.h"
 297
 298 static INLINE void swz4( __m128i x,
 299                          __m128i y,
 300                          __m128i z,
 301                          __m128i w,
 302                          __m128i *a,
 303                          __m128i *b,
 304                          __m128i *c,
 305                          __m128i *d)
 306 {
 307    __m128i i, j, k, l;
 308    __m128i m, n, o, p;
 309    __m128i e, f, g, h;
 310
 311    m = _mm_unpacklo_epi8(x,y);
 312    n = _mm_unpackhi_epi8(x,y);
 313    o = _mm_unpacklo_epi8(z,w);
 314    p = _mm_unpackhi_epi8(z,w);
 315
 316    i = _mm_unpacklo_epi16(m,n);
 317    j = _mm_unpackhi_epi16(m,n);
 318    k = _mm_unpacklo_epi16(o,p);
 319    l = _mm_unpackhi_epi16(o,p);
 320
 321    e = _mm_unpacklo_epi8(i,j);
 322    f = _mm_unpackhi_epi8(i,j);
 323    g = _mm_unpacklo_epi8(k,l);
 324    h = _mm_unpackhi_epi8(k,l);
 325
 326    *a = _mm_unpacklo_epi64(e,g);
 327    *b = _mm_unpackhi_epi64(e,g);
 328    *c = _mm_unpacklo_epi64(f,h);
 329    *d = _mm_unpackhi_epi64(f,h);
 330 }
 331
 332 static INLINE void unswz4( __m128i a,
 333                            __m128i b,
 334                            __m128i c,
 335                            __m128i d,
 336                            __m128i *x,
 337                            __m128i *y,
 338                            __m128i *z,
 339                            __m128i *w)
 340 {
 341    __m128i i, j, k, l;
 342    __m128i m, n, o, p;
 343
 344    i = _mm_unpacklo_epi8(a,b);
 345    j = _mm_unpackhi_epi8(a,b);
 346    k = _mm_unpacklo_epi8(c,d);
 347    l = _mm_unpackhi_epi8(c,d);
 348
 349    m = _mm_unpacklo_epi16(i,k);
 350    n = _mm_unpackhi_epi16(i,k);
 351    o = _mm_unpacklo_epi16(j,l);
 352    p = _mm_unpackhi_epi16(j,l);
 353
 354    *x = _mm_unpacklo_epi64(m,n);
 355    *y = _mm_unpackhi_epi64(m,n);
 356    *z = _mm_unpacklo_epi64(o,p);
 357    *w = _mm_unpackhi_epi64(o,p);
 358 }
 359
 360 static void
 361 lp_tile_b8g8r8a8_unorm_swizzle_4ub_ssse3(uint8_t *dst,
 362                                          const uint8_t *src, unsigned src_stride,
 363                                          unsigned x0, unsigned y0)
 364 {
 365    __m128i *dst128 = (__m128i *) dst;
 366    unsigned x, y;
 367
 368    src += y0 * src_stride;
 369    src += x0 * sizeof(uint32_t);
 370
 371    for (y = 0; y < TILE_SIZE; y += 4) {
 372       const uint8_t *src_row = src;
 373
 374       for (x = 0; x < TILE_SIZE; x += 4) {
 375          swz4(*(__m128i *) (src_row + 0 * src_stride),
 376               *(__m128i *) (src_row + 1 * src_stride),
 377               *(__m128i *) (src_row + 2 * src_stride),
 378               *(__m128i *) (src_row + 3 * src_stride),
 379               dst128 + 2,     /* b */
 380               dst128 + 1,     /* g */
 381               dst128 + 0,     /* r */
 382               dst128 + 3);    /* a */
 383
 384          dst128 += 4;
 385          src_row += sizeof(__m128i);
 386       }
 387
 388       src += 4 * src_stride;
 389    }
 390 }
 391
 392 static void
 393 lp_tile_b8g8r8a8_unorm_unswizzle_4ub_ssse3(const uint8_t *src,
 394                                           uint8_t *dst, unsigned dst_stride,
 395                                           unsigned x0, unsigned y0)
 396 {
 397    unsigned int x, y;
 398    const __m128i *src128 = (const __m128i *) src;
 399
 400    dst += y0 * dst_stride;
 401    dst += x0 * sizeof(uint32_t);
 402
 403    for (y = 0; y < TILE_SIZE; y += 4) {
 404       const uint8_t *dst_row = dst;
 405
 406       for (x = 0; x < TILE_SIZE; x += 4) {
 407          unswz4( src128[2],     /* b */
 408                  src128[1],     /* g */
 409                  src128[0],     /* r */
 410                  src128[3],     /* a */
 411                  (__m128i *) (dst_row + 0 * dst_stride),
 412                  (__m128i *) (dst_row + 1 * dst_stride),
 413                  (__m128i *) (dst_row + 2 * dst_stride),
 414                  (__m128i *) (dst_row + 3 * dst_stride));
 415
 416          src128 += 4;
 417          dst_row += sizeof(__m128i);;
 418       }
 419
 420       dst += 4 * dst_stride;
 421    }
 422 }
 423
 424 #endif /* PIPE_ARCH_SSSE3 */
 425 '''
 426
 427
 428 def generate_swizzle(formats, dst_channel, dst_native_type, dst_suffix):
 429     '''Generate the dispatch function to read pixels from any format'''
 430
 431     for format in formats:
 432         if is_format_supported(format):
 433             generate_format_read(format, dst_channel, dst_native_type, dst_suffix)
 434
 435     print 'void'
 436     print 'lp_tile_swizzle_%s(enum pipe_format format, %s *dst, const void *src, unsigned src_stride, unsigned x, unsigned y)' % (dst_suffix, dst_native_type)
 437     print '{'
 438     print '   void (*func)(%s *dst, const uint8_t *src, unsigned src_stride, unsigned x0, unsigned y0);' % dst_native_type
 439     print '#ifdef DEBUG'
 440     print '   lp_tile_swizzle_count += 1;'
 441     print '#endif'
 442     print '   switch(format) {'
 443     for format in formats:
 444         if is_format_supported(format):
 445             print '   case %s:' % format.name
 446             func_name = 'lp_tile_%s_swizzle_%s' % (format.short_name(), dst_suffix)
 447             if format.name == 'PIPE_FORMAT_B8G8R8A8_UNORM':
 448                 print '#ifdef PIPE_ARCH_SSE'
 449                 print '      func = util_cpu_caps.has_ssse3 ? %s_ssse3 : %s;' % (func_name, func_name)
 450                 print '#else'
 451                 print '      func = %s;' % (func_name,)
 452                 print '#endif'
 453             else:
 454                 print '      func = %s;' % (func_name,)
 455             print '      break;'
 456     print '   default:'
 457     print '      debug_printf("%s: unsupported format %s\\n", __FUNCTION__, util_format_name(format));'
 458     print '      return;'
 459     print '   }'
 460     print '   func(dst, (const uint8_t *)src, src_stride, x, y);'
 461     print '}'
 462     print
 463
 464
 465 def generate_unswizzle(formats, src_channel, src_native_type, src_suffix):
 466     '''Generate the dispatch function to write pixels to any format'''
 467
 468     for format in formats:
 469         if is_format_supported(format):
 470             generate_format_write(format, src_channel, src_native_type, src_suffix)
 471
 472     print 'void'
 473     print 'lp_tile_unswizzle_%s(enum pipe_format format, const %s *src, void *dst, unsigned dst_stride, unsigned x, unsigned y)' % (src_suffix, src_native_type)
 474
 475     print '{'
 476     print '   void (*func)(const %s *src, uint8_t *dst, unsigned dst_stride, unsigned x0, unsigned y0);' % src_native_type
 477     print '#ifdef DEBUG'
 478     print '   lp_tile_unswizzle_count += 1;'
 479     print '#endif'
 480     print '   switch(format) {'
 481     for format in formats:
 482         if is_format_supported(format):
 483             print '   case %s:' % format.name
 484             func_name = 'lp_tile_%s_unswizzle_%s' % (format.short_name(), src_suffix)
 485             if format.name == 'PIPE_FORMAT_B8G8R8A8_UNORM':
 486                 print '#ifdef PIPE_ARCH_SSE'
 487                 print '      func = util_cpu_caps.has_ssse3 ? %s_ssse3 : %s;' % (func_name, func_name)
 488                 print '#else'
 489                 print '      func = %s;' % (func_name,)
 490                 print '#endif'
 491             else:
 492                 print '      func = %s;' % (func_name,)
 493             print '      break;'
 494     print '   default:'
 495     print '      debug_printf("%s: unsupported format %s\\n", __FUNCTION__, util_format_name(format));'
 496     print '      return;'
 497     print '   }'
 498     print '   func(src, (uint8_t *)dst, dst_stride, x, y);'
 499     print '}'
 500     print
 501
 502
 503 def main():
 504     formats = []
 505     for arg in sys.argv[1:]:
 506         formats.extend(parse(arg))
 507
 508     print '/* This file is autogenerated by lp_tile_soa.py from u_format.csv. Do not edit directly. */'
 509     print
 510     # This will print the copyright message on the top of this file
 511     print __doc__.strip()
 512     print
 513     print '#include "pipe/p_compiler.h"'
 514     print '#include "util/u_format.h"'
 515     print '#include "util/u_math.h"'
 516     print '#include "util/u_half.h"'
 517     print '#include "util/u_cpu_detect.h"'
 518     print '#include "lp_tile_soa.h"'
 519     print
 520     print '#ifdef DEBUG'
 521     print 'unsigned lp_tile_unswizzle_count = 0;'
 522     print 'unsigned lp_tile_swizzle_count = 0;'
 523     print '#endif'
 524     print
 525     print 'const unsigned char'
 526     print 'tile_offset[TILE_VECTOR_HEIGHT][TILE_VECTOR_WIDTH] = {'
 527     print '   {  0,  1,  4,  5},'
 528     print '   {  2,  3,  6,  7},'
 529     print '   {  8,  9, 12, 13},'
 530     print '   { 10, 11, 14, 15}'
 531     print '};'
 532     print
 533     print '/* Note: these lookup tables could be replaced with some'
 534     print ' * bit-twiddling code, but this is a little faster.'
 535     print ' */'
 536     print 'static unsigned tile_x_offset[TILE_VECTOR_WIDTH * TILE_VECTOR_HEIGHT] = {'
 537     print '   0, 1, 0, 1, 2, 3, 2, 3,'
 538     print '   0, 1, 0, 1, 2, 3, 2, 3'
 539     print '};'
 540     print
 541     print 'static unsigned tile_y_offset[TILE_VECTOR_WIDTH * TILE_VECTOR_HEIGHT] = {'
 542     print '   0, 0, 1, 1, 0, 0, 1, 1,'
 543     print '   2, 2, 3, 3, 2, 2, 3, 3'
 544     print '};'
 545     print
 546
 547     generate_ssse3()
 548
 549     channel = Channel(UNSIGNED, True, 8)
 550     native_type = 'uint8_t'
 551     suffix = '4ub'
 552
 553     generate_swizzle(formats, channel, native_type, suffix)
 554     generate_unswizzle(formats, channel, native_type, suffix)
 555
 556
 557 if __name__ == '__main__':
 558     main()