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(__DJGPP__) && !defined(__MINGW32__) /* this one cries for assyntax.h */
50 /* I implemented these as macros because the appear in quite a few places,
51 * and I've tweaked them a number of times. I got tired of changing every
52 * place they appear. :)
55 #define DO_ONE_PIXEL() \
58 bswap %eax /* ARGB -> BGRA */ ; \
59 rorl $8, %eax /* BGRA -> ABGR */ ; \
60 movl %eax, (%ecx) /* ABGR -> R, G, B, A */ ; \
63 #define DO_ONE_LAST_PIXEL() \
65 bswap %eax /* ARGB -> BGRA */ ; \
66 rorl $8, %eax /* BGRA -> ABGR */ ; \
67 movl %eax, (%ecx) /* ABGR -> R, G, B, A */ ; \
71 * MMX optimized version of the BGRA8888_REV to RGBA copy routine.
74 * This function assumes that the caller will issue the EMMS instruction
75 * at the correct places.
78 .globl _generic_read_RGBA_span_BGRA8888_REV_MMX
79 .hidden _generic_read_RGBA_span_BGRA8888_REV_MMX
80 .type _generic_read_RGBA_span_BGRA8888_REV_MMX, @function
81 _generic_read_RGBA_span_BGRA8888_REV_MMX:
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
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:
345 movdqa mask+16, %xmm2
347 movl 12(%esp), %ebx /* source pointer */
348 movl 20(%esp), %edx /* number of pixels to copy */
349 movl 16(%esp), %ecx /* destination pointer */
355 jle .L46 /* Bail if there's nothing to do. */
357 /* If the source pointer isn't a multiple of 16 we have to process
358 * a few pixels the "slow" way to get the address aligned for
359 * the SSE fetch intsructions.
397 /* Would it be worth having a specialized version of this loop for
398 * the case where the destination is 16-byte aligned? That version
399 * would be identical except that it could use movedqa instead of
429 /* There may be upto 3 pixels remaining to be copied. Take care
430 * of them now. We do the 2 pixel case first because the data
463 .size _generic_read_RGBA_span_BGRA8888_REV_SSE2, .-_generic_read_RGBA_span_BGRA8888_REV_SSE2
476 /* Setting SCALE_ADJUST to 5 gives a perfect match with the classic C
477 * implementation in Mesa. Setting SCALE_ADJUST to 0 is slightly faster but
478 * at a small cost to accuracy.
481 #define SCALE_ADJUST 5
482 #if SCALE_ADJUST == 5
490 .word 0x20e8 /* (0x00ff0000 / 0x000007c0) + 1 */
491 .word 0x40c5 /* (0x00ff0000 / 0x000003f0) + 1 */
492 .word 0x839d /* (0x00ff0000 / 0x000001f0) + 1 */
494 #elif SCALE_ADJUST == 0
502 .word 0x0108 /* (0x00ff0000 / 0x0000f800) + 1 */
503 .word 0x0104 /* (0x00ff0000 / 0x0000fc00) + 1 */
504 .word 0x0108 /* (0x00ff0000 / 0x0000f800) + 1 */
507 #error SCALE_ADJUST must either be 5 or 0.
511 alpha: .long 0x00000000
515 * MMX optimized version of the RGB565 to RGBA copy routine.
519 .globl _generic_read_RGBA_span_RGB565_MMX
520 .hidden _generic_read_RGBA_span_RGB565_MMX
521 .type _generic_read_RGBA_span_RGB565_MMX, @function
523 _generic_read_RGBA_span_RGB565_MMX:
525 #ifdef USE_INNER_EMMS
529 movl 4(%esp), %eax /* source pointer */
530 movl 8(%esp), %edx /* destination pointer */
531 movl 12(%esp), %ecx /* number of pixels to copy */
538 jle .L01 /* Bail early if the count is negative. */
542 /* Fetch 4 RGB565 pixels into %mm4. Distribute the first and
543 * second pixels into the four words of %mm0 and %mm2.
549 pshufw $0x00, %mm4, %mm0
550 pshufw $0x55, %mm4, %mm2
553 /* Mask the pixels so that each word of each register contains only
554 * one color component.
561 /* Adjust the component values so that they are as small as possible,
562 * but large enough so that we can multiply them by an unsigned 16-bit
563 * number and get a value as large as 0x00ff0000.
569 psrlw $SCALE_ADJUST, %mm0
570 psrlw $SCALE_ADJUST, %mm2
573 /* Scale the input component values to be on the range
574 * [0, 0x00ff0000]. This it the real magic of the whole routine.
581 /* Always set the alpha value to 0xff.
588 /* Pack the 16-bit values to 8-bit values and store the converted
598 pshufw $0xaa, %mm4, %mm0
599 pshufw $0xff, %mm4, %mm2
606 psrlw $SCALE_ADJUST, %mm0
607 psrlw $SCALE_ADJUST, %mm2
625 /* At this point there can be at most 3 pixels left to process. If
626 * there is either 2 or 3 left, process 2.
636 pshufw $0x00, %mm4, %mm0
637 pshufw $0x55, %mm4, %mm2
644 psrlw $SCALE_ADJUST, %mm0
645 psrlw $SCALE_ADJUST, %mm2
659 /* At this point there can be at most 1 pixel left to process.
660 * Process it if needed.
669 pshufw $0x00, %mm4, %mm0
674 psrlw $SCALE_ADJUST, %mm0
685 #ifdef USE_INNER_EMMS
689 #endif /* !defined(__DJGPP__) && !defined(__MINGW32__) */
691 #if defined (__ELF__) && defined (__linux__)
692 .section .note.GNU-stack,"",%progbits