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
.name
== 'PIPE_FORMAT_R11G11B10_FLOAT':
57 if format
.name
== 'PIPE_FORMAT_R9G9B9E5_FLOAT':
60 if format
.layout
!= PLAIN
:
64 channel
= format
.channels
[i
]
65 if channel
.type not in (VOID
, UNSIGNED
, SIGNED
, FLOAT
):
67 if channel
.type == FLOAT
and channel
.size
not in (16, 32 ,64):
70 if format
.colorspace
not in ('rgb', 'srgb'):
76 def generate_format_read(format
, dst_channel
, dst_native_type
, dst_suffix
):
77 '''Generate the function to read pixels from a particular format'''
79 name
= format
.short_name()
81 src_native_type
= native_type(format
)
84 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
)
86 print ' unsigned x, y;'
87 print ' const uint8_t *src_row = src + y0*src_stride;'
88 print ' for (y = 0; y < TILE_SIZE; ++y) {'
89 print ' const %s *src_pixel = (const %s *)(src_row + x0*%u);' % (src_native_type
, src_native_type
, format
.stride())
90 print ' for (x = 0; x < TILE_SIZE; ++x) {'
93 if format
.colorspace
in ('rgb', 'srgb'):
95 swizzle
= format
.swizzles
[i
]
97 names
[swizzle
] += 'rgba'[i
]
98 elif format
.colorspace
== 'zs':
99 swizzle
= format
.swizzles
[0]
107 if format
.name
== 'PIPE_FORMAT_R11G11B10_FLOAT':
108 print ' float tmp[3];'
109 print ' uint8_t r, g, b;'
110 print ' r11g11b10f_to_float3(*src_pixel++, tmp);'
112 print ' %s = tmp[%d] * 0xff;' % (names
[i
], i
)
113 elif format
.name
== 'PIPE_FORMAT_R9G9B9E5_FLOAT':
114 print ' float tmp[3];'
115 print ' uint8_t r, g, b;'
116 print ' rgb9e5_to_float3(*src_pixel++, tmp);'
118 print ' %s = tmp[%d] * 0xff;' % (names
[i
], i
)
119 elif format
.layout
== PLAIN
:
120 if not format
.is_array():
121 print ' %s pixel = *src_pixel++;' % src_native_type
124 src_channel
= format
.channels
[i
]
125 width
= src_channel
.size
128 mask
= (1 << width
) - 1
130 value
= '(%s >> %u)' % (value
, shift
)
131 if shift
+ width
< format
.block_size():
132 value
= '(%s & 0x%x)' % (value
, mask
)
133 value
= conversion_expr(src_channel
, dst_channel
, dst_native_type
, value
, clamp
=False)
134 print ' %s %s = %s;' % (dst_native_type
, names
[i
], value
)
139 print ' %s %s;' % (dst_native_type
, names
[i
])
141 src_channel
= format
.channels
[i
]
143 value
= '(*src_pixel++)'
144 value
= conversion_expr(src_channel
, dst_channel
, dst_native_type
, value
, clamp
=False)
145 print ' %s = %s;' % (names
[i
], value
)
146 elif src_channel
.size
:
147 print ' ++src_pixel;'
152 if format
.colorspace
in ('rgb', 'srgb'):
153 swizzle
= format
.swizzles
[i
]
155 value
= names
[swizzle
]
156 elif swizzle
== SWIZZLE_0
:
158 elif swizzle
== SWIZZLE_1
:
159 value
= get_one(dst_channel
)
162 elif format
.colorspace
== 'zs':
166 value
= get_one(dst_channel
)
169 print ' TILE_PIXEL(dst, x, y, %u) = %s; /* %s */' % (i
, value
, 'rgba'[i
])
172 print ' src_row += src_stride;'
178 def pack_rgba(format
, src_channel
, r
, g
, b
, a
):
179 """Return an expression for packing r, g, b, a into a pixel of the
180 given format. Ex: '(b << 24) | (g << 16) | (r << 8) | (a << 0)'
182 assert format
.colorspace
in ('rgb', 'srgb')
183 inv_swizzle
= format
.inv_swizzles()
187 # choose r, g, b, or a depending on the inverse swizzle term
188 if inv_swizzle
[i
] == 0:
190 elif inv_swizzle
[i
] == 1:
192 elif inv_swizzle
[i
] == 2:
194 elif inv_swizzle
[i
] == 3:
200 dst_channel
= format
.channels
[i
]
201 dst_native_type
= native_type(format
)
202 value
= conversion_expr(src_channel
, dst_channel
, dst_native_type
, value
, clamp
=False)
203 term
= "((%s) << %d)" % (value
, shift
)
205 expr
= expr
+ " | " + term
209 width
= format
.channels
[i
].size
210 shift
= shift
+ width
214 def emit_unrolled_unswizzle_code(format
, src_channel
):
215 '''Emit code for writing a block based on unrolled loops.
216 This is considerably faster than the TILE_PIXEL-based code below.
218 dst_native_type
= 'uint%u_t' % format
.block_size()
219 print ' const unsigned dstpix_stride = dst_stride / %d;' % format
.stride()
220 print ' %s *dstpix = (%s *) dst;' % (dst_native_type
, dst_native_type
)
221 print ' unsigned int qx, qy, i;'
223 print ' for (qy = 0; qy < TILE_SIZE; qy += TILE_VECTOR_HEIGHT) {'
224 print ' const unsigned py = y0 + qy;'
225 print ' for (qx = 0; qx < TILE_SIZE; qx += TILE_VECTOR_WIDTH) {'
226 print ' const unsigned px = x0 + qx;'
227 print ' const uint8_t *r = src + 0 * TILE_C_STRIDE;'
228 print ' const uint8_t *g = src + 1 * TILE_C_STRIDE;'
229 print ' const uint8_t *b = src + 2 * TILE_C_STRIDE;'
230 print ' const uint8_t *a = src + 3 * TILE_C_STRIDE;'
231 print ' (void) r; (void) g; (void) b; (void) a; /* silence warnings */'
232 print ' for (i = 0; i < TILE_C_STRIDE; i += 2) {'
233 print ' const uint32_t pixel0 = %s;' % pack_rgba(format
, src_channel
, "r[i+0]", "g[i+0]", "b[i+0]", "a[i+0]")
234 print ' const uint32_t pixel1 = %s;' % pack_rgba(format
, src_channel
, "r[i+1]", "g[i+1]", "b[i+1]", "a[i+1]")
235 print ' const unsigned offset = (py + tile_y_offset[i]) * dstpix_stride + (px + tile_x_offset[i]);'
236 print ' dstpix[offset + 0] = pixel0;'
237 print ' dstpix[offset + 1] = pixel1;'
239 print ' src += TILE_X_STRIDE;'
244 def emit_tile_pixel_unswizzle_code(format
, src_channel
):
245 '''Emit code for writing a block based on the TILE_PIXEL macro.'''
246 dst_native_type
= native_type(format
)
248 inv_swizzle
= format
.inv_swizzles()
250 print ' unsigned x, y;'
251 print ' uint8_t *dst_row = dst + y0*dst_stride;'
252 print ' for (y = 0; y < TILE_SIZE; ++y) {'
253 print ' %s *dst_pixel = (%s *)(dst_row + x0*%u);' % (dst_native_type
, dst_native_type
, format
.stride())
254 print ' for (x = 0; x < TILE_SIZE; ++x) {'
256 if format
.name
== 'PIPE_FORMAT_R11G11B10_FLOAT':
257 print ' float tmp[3];'
259 print ' tmp[%d] = ubyte_to_float(TILE_PIXEL(src, x, y, %u));' % (i
, inv_swizzle
[i
])
260 print ' *dst_pixel++ = float3_to_r11g11b10f(tmp);'
261 elif format
.name
== 'PIPE_FORMAT_R9G9B9E5_FLOAT':
262 print ' float tmp[3];'
264 print ' tmp[%d] = ubyte_to_float(TILE_PIXEL(src, x, y, %u));' % (i
, inv_swizzle
[i
])
265 print ' *dst_pixel++ = float3_to_rgb9e5(tmp);'
266 elif format
.layout
== PLAIN
:
267 if not format
.is_array():
268 print ' %s pixel = 0;' % dst_native_type
271 dst_channel
= format
.channels
[i
]
272 width
= dst_channel
.size
273 if inv_swizzle
[i
] is not None:
274 value
= 'TILE_PIXEL(src, x, y, %u)' % inv_swizzle
[i
]
275 value
= conversion_expr(src_channel
, dst_channel
, dst_native_type
, value
, clamp
=False)
277 value
= '(%s << %u)' % (value
, shift
)
278 print ' pixel |= %s;' % value
280 print ' *dst_pixel++ = pixel;'
283 dst_channel
= format
.channels
[i
]
284 if inv_swizzle
[i
] is not None:
285 value
= 'TILE_PIXEL(src, x, y, %u)' % inv_swizzle
[i
]
286 value
= conversion_expr(src_channel
, dst_channel
, dst_native_type
, value
, clamp
=False)
287 print ' *dst_pixel++ = %s;' % value
288 elif dst_channel
.size
:
289 print ' ++dst_pixel;'
294 print ' dst_row += dst_stride;'
298 def generate_format_write(format
, src_channel
, src_native_type
, src_suffix
):
299 '''Generate the function to write pixels to a particular format'''
301 name
= format
.short_name()
304 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
)
306 if format
.layout
== PLAIN \
307 and format
.colorspace
== 'rgb' \
308 and format
.block_size() <= 32 \
309 and format
.is_pot() \
310 and not format
.is_mixed() \
311 and (format
.channels
[0].type == UNSIGNED \
312 or format
.channels
[1].type == UNSIGNED
):
313 emit_unrolled_unswizzle_code(format
, src_channel
)
315 emit_tile_pixel_unswizzle_code(format
, src_channel
)
322 #if defined(PIPE_ARCH_SSE)
324 #include "util/u_sse.h"
326 static ALWAYS_INLINE void
327 swz4( const __m128i * restrict x,
328 const __m128i * restrict y,
329 const __m128i * restrict z,
330 const __m128i * restrict w,
331 __m128i * restrict a,
332 __m128i * restrict b,
333 __m128i * restrict c,
334 __m128i * restrict d)
340 m = _mm_unpacklo_epi8(*x,*y);
341 n = _mm_unpackhi_epi8(*x,*y);
342 o = _mm_unpacklo_epi8(*z,*w);
343 p = _mm_unpackhi_epi8(*z,*w);
345 i = _mm_unpacklo_epi16(m,n);
346 j = _mm_unpackhi_epi16(m,n);
347 k = _mm_unpacklo_epi16(o,p);
348 l = _mm_unpackhi_epi16(o,p);
350 e = _mm_unpacklo_epi8(i,j);
351 f = _mm_unpackhi_epi8(i,j);
352 g = _mm_unpacklo_epi8(k,l);
353 h = _mm_unpackhi_epi8(k,l);
355 *a = _mm_unpacklo_epi64(e,g);
356 *b = _mm_unpackhi_epi64(e,g);
357 *c = _mm_unpacklo_epi64(f,h);
358 *d = _mm_unpackhi_epi64(f,h);
361 static ALWAYS_INLINE void
362 unswz4( const __m128i * restrict a,
363 const __m128i * restrict b,
364 const __m128i * restrict c,
365 const __m128i * restrict d,
366 __m128i * restrict x,
367 __m128i * restrict y,
368 __m128i * restrict z,
369 __m128i * restrict w)
374 i = _mm_unpacklo_epi8(*a,*b);
375 j = _mm_unpackhi_epi8(*a,*b);
376 k = _mm_unpacklo_epi8(*c,*d);
377 l = _mm_unpackhi_epi8(*c,*d);
379 m = _mm_unpacklo_epi16(i,k);
380 n = _mm_unpackhi_epi16(i,k);
381 o = _mm_unpacklo_epi16(j,l);
382 p = _mm_unpackhi_epi16(j,l);
384 *x = _mm_unpacklo_epi64(m,n);
385 *y = _mm_unpackhi_epi64(m,n);
386 *z = _mm_unpacklo_epi64(o,p);
387 *w = _mm_unpackhi_epi64(o,p);
391 lp_tile_b8g8r8a8_unorm_swizzle_4ub_sse2(uint8_t * restrict dst,
392 const uint8_t * restrict src, unsigned src_stride,
393 unsigned x0, unsigned y0)
395 __m128i *dst128 = (__m128i *) dst;
398 src += y0 * src_stride;
399 src += x0 * sizeof(uint32_t);
401 for (y = 0; y < TILE_SIZE; y += 4) {
402 const uint8_t *src_row = src;
404 for (x = 0; x < TILE_SIZE; x += 4) {
405 swz4((const __m128i *) (src_row + 0 * src_stride),
406 (const __m128i *) (src_row + 1 * src_stride),
407 (const __m128i *) (src_row + 2 * src_stride),
408 (const __m128i *) (src_row + 3 * src_stride),
415 src_row += sizeof(__m128i);
418 src += 4 * src_stride;
423 lp_tile_b8g8r8a8_unorm_unswizzle_4ub_sse2(const uint8_t * restrict src,
424 uint8_t * restrict dst, unsigned dst_stride,
425 unsigned x0, unsigned y0)
428 const __m128i *src128 = (const __m128i *) src;
430 dst += y0 * dst_stride;
431 dst += x0 * sizeof(uint32_t);
433 for (y = 0; y < TILE_SIZE; y += 4) {
434 const uint8_t *dst_row = dst;
436 for (x = 0; x < TILE_SIZE; x += 4) {
437 unswz4( &src128[2], /* b */
441 (__m128i *) (dst_row + 0 * dst_stride),
442 (__m128i *) (dst_row + 1 * dst_stride),
443 (__m128i *) (dst_row + 2 * dst_stride),
444 (__m128i *) (dst_row + 3 * dst_stride));
447 dst_row += sizeof(__m128i);;
450 dst += 4 * dst_stride;
455 lp_tile_b8g8r8x8_unorm_swizzle_4ub_sse2(uint8_t * restrict dst,
456 const uint8_t * restrict src, unsigned src_stride,
457 unsigned x0, unsigned y0)
459 __m128i *dst128 = (__m128i *) dst;
462 src += y0 * src_stride;
463 src += x0 * sizeof(uint32_t);
465 for (y = 0; y < TILE_SIZE; y += 4) {
466 const uint8_t *src_row = src;
468 for (x = 0; x < TILE_SIZE; x += 4) {
469 swz4((const __m128i *) (src_row + 0 * src_stride),
470 (const __m128i *) (src_row + 1 * src_stride),
471 (const __m128i *) (src_row + 2 * src_stride),
472 (const __m128i *) (src_row + 3 * src_stride),
479 src_row += sizeof(__m128i);
482 src += 4 * src_stride;
487 lp_tile_b8g8r8x8_unorm_unswizzle_4ub_sse2(const uint8_t * restrict src,
488 uint8_t * restrict dst, unsigned dst_stride,
489 unsigned x0, unsigned y0)
492 const __m128i *src128 = (const __m128i *) src;
494 dst += y0 * dst_stride;
495 dst += x0 * sizeof(uint32_t);
497 for (y = 0; y < TILE_SIZE; y += 4) {
498 const uint8_t *dst_row = dst;
500 for (x = 0; x < TILE_SIZE; x += 4) {
501 unswz4( &src128[2], /* b */
505 (__m128i *) (dst_row + 0 * dst_stride),
506 (__m128i *) (dst_row + 1 * dst_stride),
507 (__m128i *) (dst_row + 2 * dst_stride),
508 (__m128i *) (dst_row + 3 * dst_stride));
511 dst_row += sizeof(__m128i);;
514 dst += 4 * dst_stride;
518 #endif /* PIPE_ARCH_SSE */
522 def generate_swizzle(formats
, dst_channel
, dst_native_type
, dst_suffix
):
523 '''Generate the dispatch function to read pixels from any format'''
525 for format
in formats
:
526 if is_format_supported(format
):
527 generate_format_read(format
, dst_channel
, dst_native_type
, dst_suffix
)
530 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
)
532 print ' void (*func)(%s * restrict dst, const uint8_t * restrict src, unsigned src_stride, unsigned x0, unsigned y0);' % dst_native_type
534 print ' lp_tile_swizzle_count += 1;'
536 print ' switch(format) {'
537 for format
in formats
:
538 if is_format_supported(format
):
539 print ' case %s:' % format
.name
540 func_name
= 'lp_tile_%s_swizzle_%s' % (format
.short_name(), dst_suffix
)
541 if format
.name
== 'PIPE_FORMAT_B8G8R8A8_UNORM' or format
.name
== 'PIPE_FORMAT_B8G8R8X8_UNORM':
542 print '#ifdef PIPE_ARCH_SSE'
543 print ' func = util_cpu_caps.has_sse2 ? %s_sse2 : %s;' % (func_name
, func_name
)
545 print ' func = %s;' % (func_name
,)
548 print ' func = %s;' % (func_name
,)
551 print ' debug_printf("%s: unsupported format %s\\n", __FUNCTION__, util_format_name(format));'
554 print ' func(dst, (const uint8_t *)src, src_stride, x, y);'
559 def generate_unswizzle(formats
, src_channel
, src_native_type
, src_suffix
):
560 '''Generate the dispatch function to write pixels to any format'''
562 for format
in formats
:
563 if is_format_supported(format
):
564 generate_format_write(format
, src_channel
, src_native_type
, src_suffix
)
567 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
)
570 print ' void (*func)(const %s * restrict src, uint8_t * restrict dst, unsigned dst_stride, unsigned x0, unsigned y0);' % src_native_type
572 print ' lp_tile_unswizzle_count += 1;'
574 print ' switch(format) {'
575 for format
in formats
:
576 if is_format_supported(format
):
577 print ' case %s:' % format
.name
578 func_name
= 'lp_tile_%s_unswizzle_%s' % (format
.short_name(), src_suffix
)
579 if format
.name
== 'PIPE_FORMAT_B8G8R8A8_UNORM' or format
.name
== 'PIPE_FORMAT_B8G8R8X8_UNORM':
580 print '#ifdef PIPE_ARCH_SSE'
581 print ' func = util_cpu_caps.has_sse2 ? %s_sse2 : %s;' % (func_name
, func_name
)
583 print ' func = %s;' % (func_name
,)
586 print ' func = %s;' % (func_name
,)
589 print ' debug_printf("%s: unsupported format %s\\n", __FUNCTION__, util_format_name(format));'
592 print ' func(src, (uint8_t *)dst, dst_stride, x, y);'
599 for arg
in sys
.argv
[1:]:
600 formats
.extend(parse(arg
))
602 print '/* This file is autogenerated by lp_tile_soa.py from u_format.csv. Do not edit directly. */'
604 # This will print the copyright message on the top of this file
605 print CopyRight
.strip()
607 print '#include "pipe/p_compiler.h"'
608 print '#include "util/u_math.h"'
609 print '#include "util/u_format.h"'
610 print '#include "util/u_format_r11g11b10f.h"'
611 print '#include "util/u_format_rgb9e5.h"'
612 print '#include "util/u_half.h"'
613 print '#include "util/u_cpu_detect.h"'
614 print '#include "lp_tile_soa.h"'
617 print 'unsigned lp_tile_unswizzle_count = 0;'
618 print 'unsigned lp_tile_swizzle_count = 0;'
621 print 'const unsigned char'
622 print 'tile_offset[TILE_VECTOR_HEIGHT][TILE_VECTOR_WIDTH] = {'
623 print ' { 0, 1, 4, 5},'
624 print ' { 2, 3, 6, 7},'
625 print ' { 8, 9, 12, 13},'
626 print ' { 10, 11, 14, 15}'
629 print '/* Note: these lookup tables could be replaced with some'
630 print ' * bit-twiddling code, but this is a little faster.'
632 print 'static unsigned tile_x_offset[TILE_VECTOR_WIDTH * TILE_VECTOR_HEIGHT] = {'
633 print ' 0, 1, 0, 1, 2, 3, 2, 3,'
634 print ' 0, 1, 0, 1, 2, 3, 2, 3'
637 print 'static unsigned tile_y_offset[TILE_VECTOR_WIDTH * TILE_VECTOR_HEIGHT] = {'
638 print ' 0, 0, 1, 1, 0, 0, 1, 1,'
639 print ' 2, 2, 3, 3, 2, 2, 3, 3'
645 channel
= Channel(UNSIGNED
, True, False, 8)
646 native_type
= 'uint8_t'
649 generate_swizzle(formats
, channel
, native_type
, suffix
)
650 generate_unswizzle(formats
, channel
, native_type
, suffix
)
653 if __name__
== '__main__':