| 1 |
|
/* |
| 2 |
|
* video_blit.h - Video/graphics emulation, blitters |
| 3 |
|
* |
| 4 |
< |
* Basilisk II (C) 1997-2000 Christian Bauer |
| 4 |
> |
* Basilisk II (C) 1997-2001 Christian Bauer |
| 5 |
|
* |
| 6 |
|
* This program is free software; you can redistribute it and/or modify |
| 7 |
|
* it under the terms of the GNU General Public License as published by |
| 18 |
|
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
| 19 |
|
*/ |
| 20 |
|
|
| 21 |
< |
#ifndef FB_BLIT_1 |
| 21 |
> |
#ifndef FB_DEPTH |
| 22 |
> |
# error "Undefined screen depth" |
| 23 |
> |
#endif |
| 24 |
> |
|
| 25 |
> |
#if !defined(FB_BLIT_1) && (FB_DEPTH <= 16) |
| 26 |
|
# error "Undefined 16-bit word blit function" |
| 27 |
|
#endif |
| 28 |
|
|
| 29 |
< |
#ifndef FB_BLIT_2 |
| 29 |
> |
#if !defined(FB_BLIT_2) |
| 30 |
|
# error "Undefined 32-bit word blit function" |
| 31 |
|
#endif |
| 32 |
|
|
| 33 |
< |
#ifndef FB_DEPTH |
| 34 |
< |
# error "Undefined screen depth" |
| 33 |
> |
#if !defined(FB_BLIT_4) |
| 34 |
> |
# error "Undefined 64-bit word blit function" |
| 35 |
|
#endif |
| 36 |
|
|
| 37 |
|
static void FB_FUNC_NAME(uint8 * dest, const uint8 * source, uint32 length) |
| 38 |
|
{ |
| 39 |
+ |
#define DEREF_WORD_PTR(ptr, ofs) (((uint16 *)(ptr))[(ofs)]) |
| 40 |
+ |
#define DEREF_LONG_PTR(ptr, ofs) (((uint32 *)(ptr))[(ofs)]) |
| 41 |
+ |
#define DEREF_QUAD_PTR(ptr, ofs) (((uint64 *)(ptr))[(ofs)]) |
| 42 |
+ |
|
| 43 |
+ |
#ifndef UNALIGNED_PROFITABLE |
| 44 |
|
#if FB_DEPTH <= 8 |
| 45 |
|
// Align source and dest to 16-bit word boundaries |
| 46 |
< |
if (FB_DEPTH <= 8 && ((unsigned long) source) & 1) { |
| 46 |
> |
if (((unsigned long) source) & 1) { |
| 47 |
|
*dest++ = *source++; |
| 48 |
|
length -= 1; |
| 49 |
|
} |
| 50 |
|
#endif |
| 51 |
|
|
| 43 |
– |
// source and dest are mutually aligned |
| 44 |
– |
uint16 * swp = ((uint16 *)source); |
| 45 |
– |
uint16 * dwp = ((uint16 *) dest ); |
| 46 |
– |
|
| 47 |
– |
#if FB_DEPTH <= 8 |
| 48 |
– |
if (length >= 2) { |
| 49 |
– |
#endif |
| 50 |
– |
|
| 52 |
|
#if FB_DEPTH <= 16 |
| 53 |
|
// Align source and dest to 32-bit word boundaries |
| 54 |
|
if (((unsigned long) source) & 2) { |
| 55 |
< |
const uint16 val = *swp++; |
| 56 |
< |
FB_BLIT_1(*dwp++, val); |
| 55 |
> |
FB_BLIT_1(DEREF_WORD_PTR(dest, 0), DEREF_WORD_PTR(source, 0)); |
| 56 |
> |
dest += 2; source += 2; |
| 57 |
|
length -= 2; |
| 58 |
|
} |
| 59 |
|
#endif |
| 60 |
+ |
#endif |
| 61 |
|
|
| 62 |
< |
// Blit 4-byte words |
| 63 |
< |
if (length >= 4) { |
| 64 |
< |
const int remainder = (length / 4) % 8; |
| 65 |
< |
uint32 * slp = (uint32 *)swp + remainder; |
| 66 |
< |
uint32 * dlp = (uint32 *)dwp + remainder; |
| 62 |
> |
// Blit 8-byte words |
| 63 |
> |
if (length >= 8) { |
| 64 |
> |
const int remainder = (length / 8) % 8; |
| 65 |
> |
source += remainder * 8; |
| 66 |
> |
dest += remainder * 8; |
| 67 |
|
|
| 68 |
< |
int n = ((length / 4) + 7) / 8; |
| 68 |
> |
int n = ((length / 8) + 7) / 8; |
| 69 |
|
switch (remainder) { |
| 70 |
|
case 0: do { |
| 71 |
< |
slp += 8; dlp += 8; |
| 72 |
< |
FB_BLIT_2(dlp[-8], slp[-8]); |
| 73 |
< |
case 7: FB_BLIT_2(dlp[-7], slp[-7]); |
| 74 |
< |
case 6: FB_BLIT_2(dlp[-6], slp[-6]); |
| 75 |
< |
case 5: FB_BLIT_2(dlp[-5], slp[-5]); |
| 76 |
< |
case 4: FB_BLIT_2(dlp[-4], slp[-4]); |
| 77 |
< |
case 3: FB_BLIT_2(dlp[-3], slp[-3]); |
| 78 |
< |
case 2: FB_BLIT_2(dlp[-2], slp[-2]); |
| 79 |
< |
case 1: FB_BLIT_2(dlp[-1], slp[-1]); |
| 71 |
> |
dest += 64; source += 64; |
| 72 |
> |
FB_BLIT_4(DEREF_QUAD_PTR(dest, -8), DEREF_QUAD_PTR(source, -8)); |
| 73 |
> |
case 7: FB_BLIT_4(DEREF_QUAD_PTR(dest, -7), DEREF_QUAD_PTR(source, -7)); |
| 74 |
> |
case 6: FB_BLIT_4(DEREF_QUAD_PTR(dest, -6), DEREF_QUAD_PTR(source, -6)); |
| 75 |
> |
case 5: FB_BLIT_4(DEREF_QUAD_PTR(dest, -5), DEREF_QUAD_PTR(source, -5)); |
| 76 |
> |
case 4: FB_BLIT_4(DEREF_QUAD_PTR(dest, -4), DEREF_QUAD_PTR(source, -4)); |
| 77 |
> |
case 3: FB_BLIT_4(DEREF_QUAD_PTR(dest, -3), DEREF_QUAD_PTR(source, -3)); |
| 78 |
> |
case 2: FB_BLIT_4(DEREF_QUAD_PTR(dest, -2), DEREF_QUAD_PTR(source, -2)); |
| 79 |
> |
case 1: FB_BLIT_4(DEREF_QUAD_PTR(dest, -1), DEREF_QUAD_PTR(source, -1)); |
| 80 |
|
} while (--n > 0); |
| 81 |
|
} |
| 82 |
|
} |
| 83 |
|
|
| 84 |
+ |
// There could be one long left to blit |
| 85 |
+ |
if (length & 4) { |
| 86 |
+ |
FB_BLIT_2(DEREF_LONG_PTR(dest, 0), DEREF_LONG_PTR(source, 0)); |
| 87 |
+ |
#if FB_DEPTH <= 16 |
| 88 |
+ |
dest += 4; |
| 89 |
+ |
source += 4; |
| 90 |
+ |
#endif |
| 91 |
+ |
} |
| 92 |
+ |
|
| 93 |
|
#if FB_DEPTH <= 16 |
| 94 |
< |
// There might remain one word to blit |
| 94 |
> |
// There could be one word left to blit |
| 95 |
|
if (length & 2) { |
| 96 |
< |
uint16 * const s = (uint16 *)(((uint8 *)swp) + length - 2); |
| 97 |
< |
uint16 * const d = (uint16 *)(((uint8 *)dwp) + length - 2); |
| 98 |
< |
FB_BLIT_1(*d, *s); |
| 96 |
> |
FB_BLIT_1(DEREF_WORD_PTR(dest, 0), DEREF_WORD_PTR(source, 0)); |
| 97 |
> |
#if FB_DEPTH <= 8 |
| 98 |
> |
dest += 2; |
| 99 |
> |
source += 2; |
| 100 |
> |
#endif |
| 101 |
|
} |
| 102 |
|
#endif |
| 103 |
|
|
| 104 |
|
#if FB_DEPTH <= 8 |
| 105 |
< |
} |
| 105 |
> |
// There could be one byte left to blit |
| 106 |
> |
if (length & 1) |
| 107 |
> |
*dest = *source; |
| 108 |
|
#endif |
| 109 |
+ |
|
| 110 |
+ |
#undef DEREF_LONG_PTR |
| 111 |
+ |
#undef DEREF_WORD_PTR |
| 112 |
|
} |
| 113 |
|
|
| 114 |
|
#undef FB_FUNC_NAME |
| 121 |
|
#undef FB_BLIT_2 |
| 122 |
|
#endif |
| 123 |
|
|
| 124 |
+ |
#ifdef FB_BLIT_4 |
| 125 |
+ |
#undef FB_BLIT_4 |
| 126 |
+ |
#endif |
| 127 |
+ |
|
| 128 |
|
#ifdef FB_DEPTH |
| 129 |
|
#undef FB_DEPTH |
| 130 |
|
#endif |
