706411c13c6f79247f4cfe61a9050943597f7a04
4 /**************************************************************************
6 * Copyright 2009 VMware, Inc.
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:
17 * The above copyright notice and this permission notice (including the
18 * next paragraph) shall be included in all copies or substantial portions
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.
29 **************************************************************************/
33 * Pixel format accessor functions.
35 * @author Jose Fonseca <jfonseca@vmware.com>
43 sys
.path
.insert(0, os
.path
.join(os
.path
.dirname(sys
.argv
[0]), '../../auxiliary/util'))
45 from u_format_pack
import *
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.'''
52 # FIXME: Ideally we would support any format combination here.
54 if format
.layout
!= PLAIN
:
58 channel
= format
.channels
[i
]
59 if channel
.type not in (VOID
, UNSIGNED
, SIGNED
, FLOAT
):
61 if channel
.type == FLOAT
and channel
.size
not in (16, 32 ,64):
64 if format
.colorspace
not in ('rgb', 'srgb'):
70 def generate_format_read(format
, dst_channel
, dst_native_type
, dst_suffix
):
71 '''Generate the function to read pixels from a particular format'''
73 name
= format
.short_name()
75 src_native_type
= native_type(format
)
78 print 'lp_tile_%s_swizzle_%s(%s * restrict dst, const uint8_t * restrict src, unsigned src_stride, unsigned x0, unsigned y0)' % (name
, dst_suffix
, dst_native_type
)
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) {'
87 if format
.colorspace
in ('rgb', 'srgb'):
89 swizzle
= format
.swizzles
[i
]
91 names
[swizzle
] += 'rgba'[i
]
92 elif format
.colorspace
== 'zs':
93 swizzle
= format
.swizzles
[0]
101 if format
.layout
== PLAIN
:
102 if not format
.is_array():
103 print ' %s pixel = *src_pixel++;' % src_native_type
106 src_channel
= format
.channels
[i
]
107 width
= src_channel
.size
110 mask
= (1 << width
) - 1
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
)
121 print ' %s %s;' % (dst_native_type
, names
[i
])
123 src_channel
= format
.channels
[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;'
134 if format
.colorspace
in ('rgb', 'srgb'):
135 swizzle
= format
.swizzles
[i
]
137 value
= names
[swizzle
]
138 elif swizzle
== SWIZZLE_0
:
140 elif swizzle
== SWIZZLE_1
:
141 value
= get_one(dst_channel
)
144 elif format
.colorspace
== 'zs':
148 value
= get_one(dst_channel
)
151 print ' TILE_PIXEL(dst, x, y, %u) = %s; /* %s */' % (i
, value
, 'rgba'[i
])
154 print ' src_row += src_stride;'
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)'
164 assert format
.colorspace
in ('rgb', 'srgb')
165 inv_swizzle
= format
.inv_swizzles()
169 # choose r, g, b, or a depending on the inverse swizzle term
170 if inv_swizzle
[i
] == 0:
172 elif inv_swizzle
[i
] == 1:
174 elif inv_swizzle
[i
] == 2:
176 elif inv_swizzle
[i
] == 3:
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
)
187 expr
= expr
+ " | " + term
191 width
= format
.channels
[i
].size
192 shift
= shift
+ width
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.
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;'
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;'
221 print ' src += TILE_X_STRIDE;'
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
)
230 inv_swizzle
= format
.inv_swizzles()
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) {'
238 if format
.layout
== PLAIN
:
239 if not format
.is_array():
240 print ' %s pixel = 0;' % dst_native_type
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)
249 value
= '(%s << %u)' % (value
, shift
)
250 print ' pixel |= %s;' % value
252 print ' *dst_pixel++ = pixel;'
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;'
266 print ' dst_row += dst_stride;'
270 def generate_format_write(format
, src_channel
, src_native_type
, src_suffix
):
271 '''Generate the function to write pixels to a particular format'''
273 name
= format
.short_name()
276 print 'lp_tile_%s_unswizzle_%s(const %s * restrict src, uint8_t * restrict dst, unsigned dst_stride, unsigned x0, unsigned y0)' % (name
, src_suffix
, src_native_type
)
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
)
287 emit_tile_pixel_unswizzle_code(format
, src_channel
)
294 #if defined(PIPE_ARCH_SSE)
296 #include "util/u_sse.h"
298 static ALWAYS_INLINE void
299 swz4( const __m128i * restrict x,
300 const __m128i * restrict y,
301 const __m128i * restrict z,
302 const __m128i * restrict w,
303 __m128i * restrict a,
304 __m128i * restrict b,
305 __m128i * restrict c,
306 __m128i * restrict d)
312 m = _mm_unpacklo_epi8(*x,*y);
313 n = _mm_unpackhi_epi8(*x,*y);
314 o = _mm_unpacklo_epi8(*z,*w);
315 p = _mm_unpackhi_epi8(*z,*w);
317 i = _mm_unpacklo_epi16(m,n);
318 j = _mm_unpackhi_epi16(m,n);
319 k = _mm_unpacklo_epi16(o,p);
320 l = _mm_unpackhi_epi16(o,p);
322 e = _mm_unpacklo_epi8(i,j);
323 f = _mm_unpackhi_epi8(i,j);
324 g = _mm_unpacklo_epi8(k,l);
325 h = _mm_unpackhi_epi8(k,l);
327 *a = _mm_unpacklo_epi64(e,g);
328 *b = _mm_unpackhi_epi64(e,g);
329 *c = _mm_unpacklo_epi64(f,h);
330 *d = _mm_unpackhi_epi64(f,h);
333 static ALWAYS_INLINE void
334 unswz4( const __m128i * restrict a,
335 const __m128i * restrict b,
336 const __m128i * restrict c,
337 const __m128i * restrict d,
338 __m128i * restrict x,
339 __m128i * restrict y,
340 __m128i * restrict z,
341 __m128i * restrict w)
346 i = _mm_unpacklo_epi8(*a,*b);
347 j = _mm_unpackhi_epi8(*a,*b);
348 k = _mm_unpacklo_epi8(*c,*d);
349 l = _mm_unpackhi_epi8(*c,*d);
351 m = _mm_unpacklo_epi16(i,k);
352 n = _mm_unpackhi_epi16(i,k);
353 o = _mm_unpacklo_epi16(j,l);
354 p = _mm_unpackhi_epi16(j,l);
356 *x = _mm_unpacklo_epi64(m,n);
357 *y = _mm_unpackhi_epi64(m,n);
358 *z = _mm_unpacklo_epi64(o,p);
359 *w = _mm_unpackhi_epi64(o,p);
363 lp_tile_b8g8r8a8_unorm_swizzle_4ub_sse2(uint8_t * restrict dst,
364 const uint8_t * restrict src, unsigned src_stride,
365 unsigned x0, unsigned y0)
367 __m128i *dst128 = (__m128i *) dst;
370 src += y0 * src_stride;
371 src += x0 * sizeof(uint32_t);
373 for (y = 0; y < TILE_SIZE; y += 4) {
374 const uint8_t *src_row = src;
376 for (x = 0; x < TILE_SIZE; x += 4) {
377 swz4((const __m128i *) (src_row + 0 * src_stride),
378 (const __m128i *) (src_row + 1 * src_stride),
379 (const __m128i *) (src_row + 2 * src_stride),
380 (const __m128i *) (src_row + 3 * src_stride),
387 src_row += sizeof(__m128i);
390 src += 4 * src_stride;
395 lp_tile_b8g8r8a8_unorm_unswizzle_4ub_sse2(const uint8_t * restrict src,
396 uint8_t * restrict dst, unsigned dst_stride,
397 unsigned x0, unsigned y0)
400 const __m128i *src128 = (const __m128i *) src;
402 dst += y0 * dst_stride;
403 dst += x0 * sizeof(uint32_t);
405 for (y = 0; y < TILE_SIZE; y += 4) {
406 const uint8_t *dst_row = dst;
408 for (x = 0; x < TILE_SIZE; x += 4) {
409 unswz4( &src128[2], /* b */
413 (__m128i *) (dst_row + 0 * dst_stride),
414 (__m128i *) (dst_row + 1 * dst_stride),
415 (__m128i *) (dst_row + 2 * dst_stride),
416 (__m128i *) (dst_row + 3 * dst_stride));
419 dst_row += sizeof(__m128i);;
422 dst += 4 * dst_stride;
427 lp_tile_b8g8r8x8_unorm_swizzle_4ub_sse2(uint8_t * restrict dst,
428 const uint8_t * restrict src, unsigned src_stride,
429 unsigned x0, unsigned y0)
431 __m128i *dst128 = (__m128i *) dst;
434 src += y0 * src_stride;
435 src += x0 * sizeof(uint32_t);
437 for (y = 0; y < TILE_SIZE; y += 4) {
438 const uint8_t *src_row = src;
440 for (x = 0; x < TILE_SIZE; x += 4) {
441 swz4((const __m128i *) (src_row + 0 * src_stride),
442 (const __m128i *) (src_row + 1 * src_stride),
443 (const __m128i *) (src_row + 2 * src_stride),
444 (const __m128i *) (src_row + 3 * src_stride),
451 src_row += sizeof(__m128i);
454 src += 4 * src_stride;
459 lp_tile_b8g8r8x8_unorm_unswizzle_4ub_sse2(const uint8_t * restrict src,
460 uint8_t * restrict dst, unsigned dst_stride,
461 unsigned x0, unsigned y0)
464 const __m128i *src128 = (const __m128i *) src;
466 dst += y0 * dst_stride;
467 dst += x0 * sizeof(uint32_t);
469 for (y = 0; y < TILE_SIZE; y += 4) {
470 const uint8_t *dst_row = dst;
472 for (x = 0; x < TILE_SIZE; x += 4) {
473 unswz4( &src128[2], /* b */
477 (__m128i *) (dst_row + 0 * dst_stride),
478 (__m128i *) (dst_row + 1 * dst_stride),
479 (__m128i *) (dst_row + 2 * dst_stride),
480 (__m128i *) (dst_row + 3 * dst_stride));
483 dst_row += sizeof(__m128i);;
486 dst += 4 * dst_stride;
490 #endif /* PIPE_ARCH_SSE */
494 def generate_swizzle(formats
, dst_channel
, dst_native_type
, dst_suffix
):
495 '''Generate the dispatch function to read pixels from any format'''
497 for format
in formats
:
498 if is_format_supported(format
):
499 generate_format_read(format
, dst_channel
, dst_native_type
, dst_suffix
)
502 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
)
504 print ' void (*func)(%s * restrict dst, const uint8_t * restrict src, unsigned src_stride, unsigned x0, unsigned y0);' % dst_native_type
506 print ' lp_tile_swizzle_count += 1;'
508 print ' switch(format) {'
509 for format
in formats
:
510 if is_format_supported(format
):
511 print ' case %s:' % format
.name
512 func_name
= 'lp_tile_%s_swizzle_%s' % (format
.short_name(), dst_suffix
)
513 if format
.name
== 'PIPE_FORMAT_B8G8R8A8_UNORM' or format
.name
== 'PIPE_FORMAT_B8G8R8X8_UNORM':
514 print '#ifdef PIPE_ARCH_SSE'
515 print ' func = util_cpu_caps.has_sse2 ? %s_sse2 : %s;' % (func_name
, func_name
)
517 print ' func = %s;' % (func_name
,)
520 print ' func = %s;' % (func_name
,)
523 print ' debug_printf("%s: unsupported format %s\\n", __FUNCTION__, util_format_name(format));'
526 print ' func(dst, (const uint8_t *)src, src_stride, x, y);'
531 def generate_unswizzle(formats
, src_channel
, src_native_type
, src_suffix
):
532 '''Generate the dispatch function to write pixels to any format'''
534 for format
in formats
:
535 if is_format_supported(format
):
536 generate_format_write(format
, src_channel
, src_native_type
, src_suffix
)
539 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
)
542 print ' void (*func)(const %s * restrict src, uint8_t * restrict dst, unsigned dst_stride, unsigned x0, unsigned y0);' % src_native_type
544 print ' lp_tile_unswizzle_count += 1;'
546 print ' switch(format) {'
547 for format
in formats
:
548 if is_format_supported(format
):
549 print ' case %s:' % format
.name
550 func_name
= 'lp_tile_%s_unswizzle_%s' % (format
.short_name(), src_suffix
)
551 if format
.name
== 'PIPE_FORMAT_B8G8R8A8_UNORM' or format
.name
== 'PIPE_FORMAT_B8G8R8X8_UNORM':
552 print '#ifdef PIPE_ARCH_SSE'
553 print ' func = util_cpu_caps.has_sse2 ? %s_sse2 : %s;' % (func_name
, func_name
)
555 print ' func = %s;' % (func_name
,)
558 print ' func = %s;' % (func_name
,)
561 print ' debug_printf("%s: unsupported format %s\\n", __FUNCTION__, util_format_name(format));'
564 print ' func(src, (uint8_t *)dst, dst_stride, x, y);'
571 for arg
in sys
.argv
[1:]:
572 formats
.extend(parse(arg
))
574 print '/* This file is autogenerated by lp_tile_soa.py from u_format.csv. Do not edit directly. */'
576 # This will print the copyright message on the top of this file
577 print CopyRight
.strip()
579 print '#include "pipe/p_compiler.h"'
580 print '#include "util/u_format.h"'
581 print '#include "util/u_math.h"'
582 print '#include "util/u_half.h"'
583 print '#include "util/u_cpu_detect.h"'
584 print '#include "lp_tile_soa.h"'
587 print 'unsigned lp_tile_unswizzle_count = 0;'
588 print 'unsigned lp_tile_swizzle_count = 0;'
591 print 'const unsigned char'
592 print 'tile_offset[TILE_VECTOR_HEIGHT][TILE_VECTOR_WIDTH] = {'
593 print ' { 0, 1, 4, 5},'
594 print ' { 2, 3, 6, 7},'
595 print ' { 8, 9, 12, 13},'
596 print ' { 10, 11, 14, 15}'
599 print '/* Note: these lookup tables could be replaced with some'
600 print ' * bit-twiddling code, but this is a little faster.'
602 print 'static unsigned tile_x_offset[TILE_VECTOR_WIDTH * TILE_VECTOR_HEIGHT] = {'
603 print ' 0, 1, 0, 1, 2, 3, 2, 3,'
604 print ' 0, 1, 0, 1, 2, 3, 2, 3'
607 print 'static unsigned tile_y_offset[TILE_VECTOR_WIDTH * TILE_VECTOR_HEIGHT] = {'
608 print ' 0, 0, 1, 1, 0, 0, 1, 1,'
609 print ' 2, 2, 3, 3, 2, 2, 3, 3'
615 channel
= Channel(UNSIGNED
, True, False, 8)
616 native_type
= 'uint8_t'
619 generate_swizzle(formats
, channel
, native_type
, suffix
)
620 generate_unswizzle(formats
, channel
, native_type
, suffix
)
623 if __name__
== '__main__':