2 * (C) Copyright IBM Corporation 2004
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * on the rights to use, copy, modify, merge, publish, distribute, sub
9 * license, and/or sell copies of the Software, and to permit persons to whom
10 * the Software is furnished to do so, subject to the following conditions:
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19 * IBM AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22 * USE OR OTHER DEALINGS IN THE SOFTWARE.
26 * \file read_rgba_span_x86.S
27 * Optimized routines to transfer pixel data from the framebuffer to a
28 * buffer in main memory.
30 * \author Ian Romanick <idr@us.ibm.com>
33 .file "read_rgba_span_x86.S"
34 #if !defined(__MINGW32__) && !defined(__APPLE__) /* this one cries for assyntax.h */
35 /* Kevin F. Quinn 2nd July 2006
36 * Replaced data segment constants with text-segment instructions.
38 #define LOAD_MASK(mvins,m1,m2) \
51 /* I implemented these as macros because they appear in several places,
52 * and I've tweaked them a number of times. I got tired of changing every
53 * place they appear. :)
56 #define DO_ONE_PIXEL() \
59 bswap %eax /* ARGB -> BGRA */ ; \
60 rorl $8, %eax /* BGRA -> ABGR */ ; \
61 movl %eax, (%ecx) /* ABGR -> R, G, B, A */ ; \
64 #define DO_ONE_LAST_PIXEL() \
66 bswap %eax /* ARGB -> BGRA */ ; \
67 rorl $8, %eax /* BGRA -> ABGR */ ; \
68 movl %eax, (%ecx) /* ABGR -> R, G, B, A */ ;
72 * MMX optimized version of the BGRA8888_REV to RGBA copy routine.
75 * This function assumes that the caller will issue the EMMS instruction
76 * at the correct places.
79 .globl _generic_read_RGBA_span_BGRA8888_REV_MMX
80 .hidden _generic_read_RGBA_span_BGRA8888_REV_MMX
81 .type _generic_read_RGBA_span_BGRA8888_REV_MMX, @function
82 _generic_read_RGBA_span_BGRA8888_REV_MMX:
88 LOAD_MASK(movq,%mm1,%mm2)
90 movl 8(%esp), %ebx /* source pointer */
91 movl 16(%esp), %edx /* number of pixels to copy */
92 movl 12(%esp), %ecx /* destination pointer */
95 jle .L20 /* Bail if there's nothing to do. */
108 /* Would it be faster to unroll this loop once and process 4 pixels
109 * per pass, instead of just two?
119 /* These 9 instructions do what PSHUFB (if there were such an
120 * instruction) could do in 1. :(
141 #ifdef USE_INNER_EMMS
145 /* At this point there are either 1 or 0 pixels remaining to be
146 * converted. Convert the last pixel, if needed.
157 .size _generic_read_RGBA_span_BGRA8888_REV_MMX, .-_generic_read_RGBA_span_BGRA8888_REV_MMX
161 * SSE optimized version of the BGRA8888_REV to RGBA copy routine. SSE
162 * instructions are only actually used to read data from the framebuffer.
163 * In practice, the speed-up is pretty small.
166 * Do some more testing and determine if there's any reason to have this
167 * function in addition to the MMX version.
170 * This function assumes that the caller will issue the EMMS instruction
171 * at the correct places.
174 .globl _generic_read_RGBA_span_BGRA8888_REV_SSE
175 .hidden _generic_read_RGBA_span_BGRA8888_REV_SSE
176 .type _generic_read_RGBA_span_BGRA8888_REV_SSE, @function
177 _generic_read_RGBA_span_BGRA8888_REV_SSE:
182 #ifdef USE_INNER_EMMS
186 LOAD_MASK(movq,%mm1,%mm2)
188 movl 16(%esp), %ebx /* source pointer */
189 movl 24(%esp), %edx /* number of pixels to copy */
190 movl 20(%esp), %ecx /* destination pointer */
193 jle .L35 /* Bail if there's nothing to do. */
197 andl $0xfffffff0, %esp
245 /* This would be so much better if we could just move directly from
246 * an SSE register to an MMX register. Unfortunately, that
247 * functionality wasn't introduced until SSE2 with the MOVDQ2Q
289 #ifdef USE_INNER_EMMS
294 /* At this point there are either [0, 3] pixels remaining to be
329 .size _generic_read_RGBA_span_BGRA8888_REV_SSE, .-_generic_read_RGBA_span_BGRA8888_REV_SSE
333 * SSE2 optimized version of the BGRA8888_REV to RGBA copy routine.
337 .globl _generic_read_RGBA_span_BGRA8888_REV_SSE2
338 .hidden _generic_read_RGBA_span_BGRA8888_REV_SSE2
339 .type _generic_read_RGBA_span_BGRA8888_REV_SSE2, @function
340 _generic_read_RGBA_span_BGRA8888_REV_SSE2:
344 LOAD_MASK(movdqu,%xmm1,%xmm2)
346 movl 12(%esp), %ebx /* source pointer */
347 movl 20(%esp), %edx /* number of pixels to copy */
348 movl 16(%esp), %ecx /* destination pointer */
354 jle .L46 /* Bail if there's nothing to do. */
356 /* If the source pointer isn't a multiple of 16 we have to process
357 * a few pixels the "slow" way to get the address aligned for
358 * the SSE fetch intsructions.
396 /* Would it be worth having a specialized version of this loop for
397 * the case where the destination is 16-byte aligned? That version
398 * would be identical except that it could use movedqa instead of
428 /* There may be upto 3 pixels remaining to be copied. Take care
429 * of them now. We do the 2 pixel case first because the data
464 .size _generic_read_RGBA_span_BGRA8888_REV_SSE2, .-_generic_read_RGBA_span_BGRA8888_REV_SSE2
468 #define MASK_565_L 0x07e0f800
469 #define MASK_565_H 0x0000001f
470 /* Setting SCALE_ADJUST to 5 gives a perfect match with the
471 * classic C implementation in Mesa. Setting SCALE_ADJUST
472 * to 0 is slightly faster but at a small cost to accuracy.
474 #define SCALE_ADJUST 5
475 #if SCALE_ADJUST == 5
476 #define PRESCALE_L 0x00100001
477 #define PRESCALE_H 0x00000200
478 #define SCALE_L 0x40C620E8
479 #define SCALE_H 0x0000839d
480 #elif SCALE_ADJUST == 0
481 #define PRESCALE_L 0x00200001
482 #define PRESCALE_H 0x00000800
483 #define SCALE_L 0x01040108
484 #define SCALE_H 0x00000108
486 #error SCALE_ADJUST must either be 5 or 0.
488 #define ALPHA_L 0x00000000
489 #define ALPHA_H 0x00ff0000
492 * MMX optimized version of the RGB565 to RGBA copy routine.
496 .globl _generic_read_RGBA_span_RGB565_MMX
497 .hidden _generic_read_RGBA_span_RGB565_MMX
498 .type _generic_read_RGBA_span_RGB565_MMX, @function
500 _generic_read_RGBA_span_RGB565_MMX:
502 #ifdef USE_INNER_EMMS
506 movl 4(%esp), %eax /* source pointer */
507 movl 8(%esp), %edx /* destination pointer */
508 movl 12(%esp), %ecx /* number of pixels to copy */
525 jl .L01 /* Bail early if the count is negative. */
529 /* Fetch 4 RGB565 pixels into %mm4. Distribute the first and
530 * second pixels into the four words of %mm0 and %mm2.
536 pshufw $0x00, %mm4, %mm0
537 pshufw $0x55, %mm4, %mm2
540 /* Mask the pixels so that each word of each register contains only
541 * one color component.
548 /* Adjust the component values so that they are as small as possible,
549 * but large enough so that we can multiply them by an unsigned 16-bit
550 * number and get a value as large as 0x00ff0000.
556 psrlw $SCALE_ADJUST, %mm0
557 psrlw $SCALE_ADJUST, %mm2
560 /* Scale the input component values to be on the range
561 * [0, 0x00ff0000]. This it the real magic of the whole routine.
568 /* Always set the alpha value to 0xff.
575 /* Pack the 16-bit values to 8-bit values and store the converted
583 pshufw $0xaa, %mm4, %mm0
584 pshufw $0xff, %mm4, %mm2
591 psrlw $SCALE_ADJUST, %mm0
592 psrlw $SCALE_ADJUST, %mm2
610 /* At this point there can be at most 3 pixels left to process. If
611 * there is either 2 or 3 left, process 2.
621 pshufw $0x00, %mm4, %mm0
622 pshufw $0x55, %mm4, %mm2
629 psrlw $SCALE_ADJUST, %mm0
630 psrlw $SCALE_ADJUST, %mm2
644 /* At this point there can be at most 1 pixel left to process.
645 * Process it if needed.
654 pshufw $0x00, %mm4, %mm0
659 psrlw $SCALE_ADJUST, %mm0
670 #ifdef USE_INNER_EMMS
674 #endif /* !defined(__MINGW32__) && !defined(__APPLE__) */
676 #if defined (__ELF__) && defined (__linux__)
677 .section .note.GNU-stack,"",%progbits