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>
32 /* Control flow enforcement support */
39 .file "read_rgba_span_x86.S"
40 #if !defined(__MINGW32__) && !defined(__APPLE__) /* this one cries for assyntax.h */
41 /* Kevin F. Quinn 2nd July 2006
42 * Replaced data segment constants with text-segment instructions.
44 #define LOAD_MASK(mvins,m1,m2) \
57 /* I implemented these as macros because they appear in several places,
58 * and I've tweaked them a number of times. I got tired of changing every
59 * place they appear. :)
62 #define DO_ONE_PIXEL() \
65 bswap %eax /* ARGB -> BGRA */ ; \
66 rorl $8, %eax /* BGRA -> ABGR */ ; \
67 movl %eax, (%ecx) /* ABGR -> R, G, B, A */ ; \
70 #define DO_ONE_LAST_PIXEL() \
72 bswap %eax /* ARGB -> BGRA */ ; \
73 rorl $8, %eax /* BGRA -> ABGR */ ; \
74 movl %eax, (%ecx) /* ABGR -> R, G, B, A */ ;
78 * MMX optimized version of the BGRA8888_REV to RGBA copy routine.
81 * This function assumes that the caller will issue the EMMS instruction
82 * at the correct places.
85 .globl _generic_read_RGBA_span_BGRA8888_REV_MMX
86 .hidden _generic_read_RGBA_span_BGRA8888_REV_MMX
87 .type _generic_read_RGBA_span_BGRA8888_REV_MMX, @function
88 _generic_read_RGBA_span_BGRA8888_REV_MMX:
95 LOAD_MASK(movq,%mm1,%mm2)
97 movl 8(%esp), %ebx /* source pointer */
98 movl 16(%esp), %edx /* number of pixels to copy */
99 movl 12(%esp), %ecx /* destination pointer */
102 jle .L20 /* Bail if there's nothing to do. */
115 /* Would it be faster to unroll this loop once and process 4 pixels
116 * per pass, instead of just two?
126 /* These 9 instructions do what PSHUFB (if there were such an
127 * instruction) could do in 1. :(
148 #ifdef USE_INNER_EMMS
152 /* At this point there are either 1 or 0 pixels remaining to be
153 * converted. Convert the last pixel, if needed.
164 .size _generic_read_RGBA_span_BGRA8888_REV_MMX, .-_generic_read_RGBA_span_BGRA8888_REV_MMX
168 * SSE optimized version of the BGRA8888_REV to RGBA copy routine. SSE
169 * instructions are only actually used to read data from the framebuffer.
170 * In practice, the speed-up is pretty small.
173 * Do some more testing and determine if there's any reason to have this
174 * function in addition to the MMX version.
177 * This function assumes that the caller will issue the EMMS instruction
178 * at the correct places.
181 .globl _generic_read_RGBA_span_BGRA8888_REV_SSE
182 .hidden _generic_read_RGBA_span_BGRA8888_REV_SSE
183 .type _generic_read_RGBA_span_BGRA8888_REV_SSE, @function
184 _generic_read_RGBA_span_BGRA8888_REV_SSE:
190 #ifdef USE_INNER_EMMS
194 LOAD_MASK(movq,%mm1,%mm2)
196 movl 16(%esp), %ebx /* source pointer */
197 movl 24(%esp), %edx /* number of pixels to copy */
198 movl 20(%esp), %ecx /* destination pointer */
201 jle .L35 /* Bail if there's nothing to do. */
205 andl $0xfffffff0, %esp
253 /* This would be so much better if we could just move directly from
254 * an SSE register to an MMX register. Unfortunately, that
255 * functionality wasn't introduced until SSE2 with the MOVDQ2Q
297 #ifdef USE_INNER_EMMS
302 /* At this point there are either [0, 3] pixels remaining to be
337 .size _generic_read_RGBA_span_BGRA8888_REV_SSE, .-_generic_read_RGBA_span_BGRA8888_REV_SSE
341 * SSE2 optimized version of the BGRA8888_REV to RGBA copy routine.
345 .globl _generic_read_RGBA_span_BGRA8888_REV_SSE2
346 .hidden _generic_read_RGBA_span_BGRA8888_REV_SSE2
347 .type _generic_read_RGBA_span_BGRA8888_REV_SSE2, @function
348 _generic_read_RGBA_span_BGRA8888_REV_SSE2:
353 LOAD_MASK(movdqu,%xmm1,%xmm2)
355 movl 12(%esp), %ebx /* source pointer */
356 movl 20(%esp), %edx /* number of pixels to copy */
357 movl 16(%esp), %ecx /* destination pointer */
363 jle .L46 /* Bail if there's nothing to do. */
365 /* If the source pointer isn't a multiple of 16 we have to process
366 * a few pixels the "slow" way to get the address aligned for
367 * the SSE fetch intsructions.
405 /* Would it be worth having a specialized version of this loop for
406 * the case where the destination is 16-byte aligned? That version
407 * would be identical except that it could use movedqa instead of
437 /* There may be upto 3 pixels remaining to be copied. Take care
438 * of them now. We do the 2 pixel case first because the data
473 .size _generic_read_RGBA_span_BGRA8888_REV_SSE2, .-_generic_read_RGBA_span_BGRA8888_REV_SSE2
477 #define MASK_565_L 0x07e0f800
478 #define MASK_565_H 0x0000001f
479 /* Setting SCALE_ADJUST to 5 gives a perfect match with the
480 * classic C implementation in Mesa. Setting SCALE_ADJUST
481 * to 0 is slightly faster but at a small cost to accuracy.
483 #define SCALE_ADJUST 5
484 #if SCALE_ADJUST == 5
485 #define PRESCALE_L 0x00100001
486 #define PRESCALE_H 0x00000200
487 #define SCALE_L 0x40C620E8
488 #define SCALE_H 0x0000839d
489 #elif SCALE_ADJUST == 0
490 #define PRESCALE_L 0x00200001
491 #define PRESCALE_H 0x00000800
492 #define SCALE_L 0x01040108
493 #define SCALE_H 0x00000108
495 #error SCALE_ADJUST must either be 5 or 0.
497 #define ALPHA_L 0x00000000
498 #define ALPHA_H 0x00ff0000
501 * MMX optimized version of the RGB565 to RGBA copy routine.
505 .globl _generic_read_RGBA_span_RGB565_MMX
506 .hidden _generic_read_RGBA_span_RGB565_MMX
507 .type _generic_read_RGBA_span_RGB565_MMX, @function
509 _generic_read_RGBA_span_RGB565_MMX:
511 #ifdef USE_INNER_EMMS
515 movl 4(%esp), %eax /* source pointer */
516 movl 8(%esp), %edx /* destination pointer */
517 movl 12(%esp), %ecx /* number of pixels to copy */
534 jl .L01 /* Bail early if the count is negative. */
538 /* Fetch 4 RGB565 pixels into %mm4. Distribute the first and
539 * second pixels into the four words of %mm0 and %mm2.
545 pshufw $0x00, %mm4, %mm0
546 pshufw $0x55, %mm4, %mm2
549 /* Mask the pixels so that each word of each register contains only
550 * one color component.
557 /* Adjust the component values so that they are as small as possible,
558 * but large enough so that we can multiply them by an unsigned 16-bit
559 * number and get a value as large as 0x00ff0000.
565 psrlw $SCALE_ADJUST, %mm0
566 psrlw $SCALE_ADJUST, %mm2
569 /* Scale the input component values to be on the range
570 * [0, 0x00ff0000]. This it the real magic of the whole routine.
577 /* Always set the alpha value to 0xff.
584 /* Pack the 16-bit values to 8-bit values and store the converted
592 pshufw $0xaa, %mm4, %mm0
593 pshufw $0xff, %mm4, %mm2
600 psrlw $SCALE_ADJUST, %mm0
601 psrlw $SCALE_ADJUST, %mm2
619 /* At this point there can be at most 3 pixels left to process. If
620 * there is either 2 or 3 left, process 2.
630 pshufw $0x00, %mm4, %mm0
631 pshufw $0x55, %mm4, %mm2
638 psrlw $SCALE_ADJUST, %mm0
639 psrlw $SCALE_ADJUST, %mm2
653 /* At this point there can be at most 1 pixel left to process.
654 * Process it if needed.
663 pshufw $0x00, %mm4, %mm0
668 psrlw $SCALE_ADJUST, %mm0
679 #ifdef USE_INNER_EMMS
683 #endif /* !defined(__MINGW32__) && !defined(__APPLE__) */
685 #if defined (__ELF__) && defined (__linux__)
686 .section .note.GNU-stack,"",%progbits