src/video/SDL_stretch.c
author Sam Lantinga <slouken@libsdl.org>
Mon, 28 Aug 2006 03:17:39 +0000
changeset 1985 8055185ae4ed
parent 1895 c121d94672cb
child 2828 7e5ff6cd05bf
permissions -rw-r--r--
Added source color and alpha modulation support.
Added perl script to generate optimized render copy functions.
slouken@0
     1
/*
slouken@0
     2
    SDL - Simple DirectMedia Layer
slouken@1312
     3
    Copyright (C) 1997-2006 Sam Lantinga
slouken@0
     4
slouken@0
     5
    This library is free software; you can redistribute it and/or
slouken@1312
     6
    modify it under the terms of the GNU Lesser General Public
slouken@0
     7
    License as published by the Free Software Foundation; either
slouken@1312
     8
    version 2.1 of the License, or (at your option) any later version.
slouken@0
     9
slouken@0
    10
    This library is distributed in the hope that it will be useful,
slouken@0
    11
    but WITHOUT ANY WARRANTY; without even the implied warranty of
slouken@0
    12
    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
slouken@1312
    13
    Lesser General Public License for more details.
slouken@0
    14
slouken@1312
    15
    You should have received a copy of the GNU Lesser General Public
slouken@1312
    16
    License along with this library; if not, write to the Free Software
slouken@1312
    17
    Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
slouken@0
    18
slouken@0
    19
    Sam Lantinga
slouken@252
    20
    slouken@libsdl.org
slouken@0
    21
*/
slouken@1402
    22
#include "SDL_config.h"
slouken@0
    23
slouken@0
    24
/* This a stretch blit implementation based on ideas given to me by
slouken@0
    25
   Tomasz Cejner - thanks! :)
slouken@0
    26
slouken@0
    27
   April 27, 2000 - Sam Lantinga
slouken@0
    28
*/
slouken@0
    29
slouken@0
    30
#include "SDL_video.h"
slouken@0
    31
#include "SDL_blit.h"
slouken@0
    32
slouken@0
    33
/* This isn't ready for general consumption yet - it should be folded
slouken@0
    34
   into the general blitting mechanism.
slouken@0
    35
*/
slouken@0
    36
slouken@1361
    37
#if ((defined(_MFC_VER) && defined(_M_IX86)/* && !defined(_WIN32_WCE) still needed? */) || \
slouken@1442
    38
     defined(__WATCOMC__) || \
slouken@1402
    39
     (defined(__GNUC__) && defined(__i386__))) && SDL_ASSEMBLY_ROUTINES
slouken@0
    40
#define USE_ASM_STRETCH
slouken@0
    41
#endif
slouken@0
    42
slouken@0
    43
#ifdef USE_ASM_STRETCH
slouken@0
    44
slouken@1361
    45
#if defined(_M_IX86) || defined(i386)
slouken@0
    46
#define PREFIX16	0x66
slouken@0
    47
#define STORE_BYTE	0xAA
slouken@0
    48
#define STORE_WORD	0xAB
slouken@0
    49
#define LOAD_BYTE	0xAC
slouken@0
    50
#define LOAD_WORD	0xAD
slouken@0
    51
#define RETURN		0xC3
slouken@0
    52
#else
slouken@0
    53
#error Need assembly opcodes for this architecture
slouken@0
    54
#endif
slouken@0
    55
slouken@0
    56
static unsigned char copy_row[4096];
slouken@0
    57
slouken@1895
    58
static int
slouken@1895
    59
generate_rowbytes(int src_w, int dst_w, int bpp)
slouken@0
    60
{
slouken@1895
    61
    static struct
slouken@1895
    62
    {
slouken@1895
    63
        int bpp;
slouken@1895
    64
        int src_w;
slouken@1895
    65
        int dst_w;
slouken@1895
    66
    } last;
slouken@0
    67
slouken@1895
    68
    int i;
slouken@1895
    69
    int pos, inc;
slouken@1895
    70
    unsigned char *eip;
slouken@1895
    71
    unsigned char load, store;
slouken@0
    72
slouken@1895
    73
    /* See if we need to regenerate the copy buffer */
slouken@1895
    74
    if ((src_w == last.src_w) && (dst_w == last.dst_w) && (bpp == last.bpp)) {
slouken@1895
    75
        return (0);
slouken@1895
    76
    }
slouken@1895
    77
    last.bpp = bpp;
slouken@1895
    78
    last.src_w = src_w;
slouken@1895
    79
    last.dst_w = dst_w;
slouken@0
    80
slouken@1895
    81
    switch (bpp) {
slouken@1895
    82
    case 1:
slouken@1895
    83
        load = LOAD_BYTE;
slouken@1895
    84
        store = STORE_BYTE;
slouken@1895
    85
        break;
slouken@1895
    86
    case 2:
slouken@1895
    87
    case 4:
slouken@1895
    88
        load = LOAD_WORD;
slouken@1895
    89
        store = STORE_WORD;
slouken@1895
    90
        break;
slouken@1895
    91
    default:
slouken@1895
    92
        SDL_SetError("ASM stretch of %d bytes isn't supported\n", bpp);
slouken@1895
    93
        return (-1);
slouken@1895
    94
    }
slouken@1895
    95
    pos = 0x10000;
slouken@1895
    96
    inc = (src_w << 16) / dst_w;
slouken@1895
    97
    eip = copy_row;
slouken@1895
    98
    for (i = 0; i < dst_w; ++i) {
slouken@1895
    99
        while (pos >= 0x10000L) {
slouken@1895
   100
            if (bpp == 2) {
slouken@1895
   101
                *eip++ = PREFIX16;
slouken@1895
   102
            }
slouken@1895
   103
            *eip++ = load;
slouken@1895
   104
            pos -= 0x10000L;
slouken@1895
   105
        }
slouken@1895
   106
        if (bpp == 2) {
slouken@1895
   107
            *eip++ = PREFIX16;
slouken@1895
   108
        }
slouken@1895
   109
        *eip++ = store;
slouken@1895
   110
        pos += inc;
slouken@1895
   111
    }
slouken@1895
   112
    *eip++ = RETURN;
slouken@0
   113
slouken@1895
   114
    /* Verify that we didn't overflow (too late) */
slouken@1895
   115
    if (eip > (copy_row + sizeof(copy_row))) {
slouken@1895
   116
        SDL_SetError("Copy buffer overflow");
slouken@1895
   117
        return (-1);
slouken@1895
   118
    }
slouken@1895
   119
    return (0);
slouken@0
   120
}
slouken@0
   121
slouken@0
   122
#else
slouken@0
   123
slouken@1985
   124
#define DEFINE_COPY_ROW(name, type)                     \
slouken@1985
   125
void name(type *src, int src_w, type *dst, int dst_w)   \
slouken@1985
   126
{                                                       \
slouken@1985
   127
    int i;                                              \
slouken@1985
   128
    int pos, inc;                                       \
slouken@1985
   129
    type pixel = 0;                                     \
slouken@1985
   130
                                                        \
slouken@1985
   131
    pos = 0x10000;                                      \
slouken@1985
   132
    inc = (src_w << 16) / dst_w;                        \
slouken@1985
   133
    for ( i=dst_w; i>0; --i ) {                         \
slouken@1985
   134
        while ( pos >= 0x10000L ) {                     \
slouken@1985
   135
            pixel = *src++;                             \
slouken@1985
   136
            pos -= 0x10000L;                            \
slouken@1985
   137
        }                                               \
slouken@1985
   138
        *dst++ = pixel;                                 \
slouken@1985
   139
        pos += inc;                                     \
slouken@1985
   140
    }                                                   \
slouken@0
   141
}
slouken@1985
   142
/* *INDENT-OFF* */
slouken@0
   143
DEFINE_COPY_ROW(copy_row1, Uint8)
slouken@1985
   144
DEFINE_COPY_ROW(copy_row2, Uint16)
slouken@1985
   145
DEFINE_COPY_ROW(copy_row4, Uint32)
slouken@1985
   146
/* *INDENT-ON* */
slouken@1895
   147
#endif /* USE_ASM_STRETCH */
slouken@1895
   148
/* The ASM code doesn't handle 24-bpp stretch blits */
slouken@1985
   149
void
slouken@1985
   150
copy_row3(Uint8 * src, int src_w, Uint8 * dst, int dst_w)
slouken@1895
   151
{
slouken@1895
   152
    int i;
slouken@1895
   153
    int pos, inc;
slouken@1895
   154
    Uint8 pixel[3];
slouken@0
   155
slouken@1895
   156
    pos = 0x10000;
slouken@1895
   157
    inc = (src_w << 16) / dst_w;
slouken@1895
   158
    for (i = dst_w; i > 0; --i) {
slouken@1895
   159
        while (pos >= 0x10000L) {
slouken@1895
   160
            pixel[0] = *src++;
slouken@1895
   161
            pixel[1] = *src++;
slouken@1895
   162
            pixel[2] = *src++;
slouken@1895
   163
            pos -= 0x10000L;
slouken@1895
   164
        }
slouken@1895
   165
        *dst++ = pixel[0];
slouken@1895
   166
        *dst++ = pixel[1];
slouken@1895
   167
        *dst++ = pixel[2];
slouken@1895
   168
        pos += inc;
slouken@1895
   169
    }
slouken@0
   170
}
slouken@0
   171
slouken@0
   172
/* Perform a stretch blit between two surfaces of the same format.
slouken@0
   173
   NOTE:  This function is not safe to call from multiple threads!
slouken@0
   174
*/
slouken@1895
   175
int
slouken@1895
   176
SDL_SoftStretch(SDL_Surface * src, SDL_Rect * srcrect,
slouken@1895
   177
                SDL_Surface * dst, SDL_Rect * dstrect)
slouken@0
   178
{
slouken@1895
   179
    int src_locked;
slouken@1895
   180
    int dst_locked;
slouken@1895
   181
    int pos, inc;
slouken@1895
   182
    int dst_width;
slouken@1895
   183
    int dst_maxrow;
slouken@1895
   184
    int src_row, dst_row;
slouken@1895
   185
    Uint8 *srcp = NULL;
slouken@1895
   186
    Uint8 *dstp;
slouken@1895
   187
    SDL_Rect full_src;
slouken@1895
   188
    SDL_Rect full_dst;
slouken@0
   189
#if defined(USE_ASM_STRETCH) && defined(__GNUC__)
slouken@1895
   190
    int u1, u2;
slouken@0
   191
#endif
slouken@1895
   192
    const int bpp = dst->format->BytesPerPixel;
slouken@0
   193
slouken@1895
   194
    if (src->format->BitsPerPixel != dst->format->BitsPerPixel) {
slouken@1895
   195
        SDL_SetError("Only works with same format surfaces");
slouken@1895
   196
        return (-1);
slouken@1895
   197
    }
slouken@0
   198
slouken@1895
   199
    /* Verify the blit rectangles */
slouken@1895
   200
    if (srcrect) {
slouken@1895
   201
        if ((srcrect->x < 0) || (srcrect->y < 0) ||
slouken@1895
   202
            ((srcrect->x + srcrect->w) > src->w) ||
slouken@1895
   203
            ((srcrect->y + srcrect->h) > src->h)) {
slouken@1895
   204
            SDL_SetError("Invalid source blit rectangle");
slouken@1895
   205
            return (-1);
slouken@1895
   206
        }
slouken@1895
   207
    } else {
slouken@1895
   208
        full_src.x = 0;
slouken@1895
   209
        full_src.y = 0;
slouken@1895
   210
        full_src.w = src->w;
slouken@1895
   211
        full_src.h = src->h;
slouken@1895
   212
        srcrect = &full_src;
slouken@1895
   213
    }
slouken@1895
   214
    if (dstrect) {
slouken@1895
   215
        if ((dstrect->x < 0) || (dstrect->y < 0) ||
slouken@1895
   216
            ((dstrect->x + dstrect->w) > dst->w) ||
slouken@1895
   217
            ((dstrect->y + dstrect->h) > dst->h)) {
slouken@1895
   218
            SDL_SetError("Invalid destination blit rectangle");
slouken@1895
   219
            return (-1);
slouken@1895
   220
        }
slouken@1895
   221
    } else {
slouken@1895
   222
        full_dst.x = 0;
slouken@1895
   223
        full_dst.y = 0;
slouken@1895
   224
        full_dst.w = dst->w;
slouken@1895
   225
        full_dst.h = dst->h;
slouken@1895
   226
        dstrect = &full_dst;
slouken@1895
   227
    }
slouken@0
   228
slouken@1895
   229
    /* Lock the destination if it's in hardware */
slouken@1895
   230
    dst_locked = 0;
slouken@1895
   231
    if (SDL_MUSTLOCK(dst)) {
slouken@1895
   232
        if (SDL_LockSurface(dst) < 0) {
slouken@1895
   233
            SDL_SetError("Unable to lock destination surface");
slouken@1895
   234
            return (-1);
slouken@1895
   235
        }
slouken@1895
   236
        dst_locked = 1;
slouken@1895
   237
    }
slouken@1895
   238
    /* Lock the source if it's in hardware */
slouken@1895
   239
    src_locked = 0;
slouken@1895
   240
    if (SDL_MUSTLOCK(src)) {
slouken@1895
   241
        if (SDL_LockSurface(src) < 0) {
slouken@1895
   242
            if (dst_locked) {
slouken@1895
   243
                SDL_UnlockSurface(dst);
slouken@1895
   244
            }
slouken@1895
   245
            SDL_SetError("Unable to lock source surface");
slouken@1895
   246
            return (-1);
slouken@1895
   247
        }
slouken@1895
   248
        src_locked = 1;
slouken@1895
   249
    }
slouken@894
   250
slouken@1895
   251
    /* Set up the data... */
slouken@1895
   252
    pos = 0x10000;
slouken@1895
   253
    inc = (srcrect->h << 16) / dstrect->h;
slouken@1895
   254
    src_row = srcrect->y;
slouken@1895
   255
    dst_row = dstrect->y;
slouken@1895
   256
    dst_width = dstrect->w * bpp;
slouken@0
   257
slouken@0
   258
#ifdef USE_ASM_STRETCH
slouken@1895
   259
    /* Write the opcodes for this stretch */
slouken@1895
   260
    if ((bpp != 3) && (generate_rowbytes(srcrect->w, dstrect->w, bpp) < 0)) {
slouken@1895
   261
        return (-1);
slouken@1895
   262
    }
slouken@0
   263
#endif
slouken@0
   264
slouken@1895
   265
    /* Perform the stretch blit */
slouken@1895
   266
    for (dst_maxrow = dst_row + dstrect->h; dst_row < dst_maxrow; ++dst_row) {
slouken@1895
   267
        dstp = (Uint8 *) dst->pixels + (dst_row * dst->pitch)
slouken@1895
   268
            + (dstrect->x * bpp);
slouken@1895
   269
        while (pos >= 0x10000L) {
slouken@1895
   270
            srcp = (Uint8 *) src->pixels + (src_row * src->pitch)
slouken@1895
   271
                + (srcrect->x * bpp);
slouken@1895
   272
            ++src_row;
slouken@1895
   273
            pos -= 0x10000L;
slouken@1895
   274
        }
slouken@0
   275
#ifdef USE_ASM_STRETCH
slouken@1895
   276
        switch (bpp) {
slouken@1895
   277
        case 3:
slouken@1895
   278
            copy_row3(srcp, srcrect->w, dstp, dstrect->w);
slouken@1895
   279
            break;
slouken@1895
   280
        default:
slouken@0
   281
#ifdef __GNUC__
slouken@1895
   282
          __asm__ __volatile__("call *%4": "=&D"(u1), "=&S"(u2): "0"(dstp), "1"(srcp), "r"(copy_row):"memory");
slouken@1442
   283
#elif defined(_MSC_VER) || defined(__WATCOMC__)
slouken@1985
   284
            /* *INDENT-OFF* */
slouken@1895
   285
            {
slouken@1895
   286
                void *code = copy_row;
slouken@1895
   287
                __asm {
slouken@1985
   288
                    push edi
slouken@1985
   289
                    push esi
slouken@1985
   290
                    mov edi, dstp
slouken@1985
   291
                    mov esi, srcp
slouken@1985
   292
                    call dword ptr code
slouken@1985
   293
                    pop esi
slouken@1985
   294
                    pop edi
slouken@1985
   295
                }
slouken@1895
   296
            }
slouken@1985
   297
            /* *INDENT-ON* */
slouken@0
   298
#else
slouken@0
   299
#error Need inline assembly for this compiler
slouken@0
   300
#endif
slouken@1895
   301
            break;
slouken@1895
   302
        }
slouken@0
   303
#else
slouken@1895
   304
        switch (bpp) {
slouken@1895
   305
        case 1:
slouken@1895
   306
            copy_row1(srcp, srcrect->w, dstp, dstrect->w);
slouken@1895
   307
            break;
slouken@1895
   308
        case 2:
slouken@1895
   309
            copy_row2((Uint16 *) srcp, srcrect->w,
slouken@1895
   310
                      (Uint16 *) dstp, dstrect->w);
slouken@1895
   311
            break;
slouken@1895
   312
        case 3:
slouken@1895
   313
            copy_row3(srcp, srcrect->w, dstp, dstrect->w);
slouken@1895
   314
            break;
slouken@1895
   315
        case 4:
slouken@1895
   316
            copy_row4((Uint32 *) srcp, srcrect->w,
slouken@1895
   317
                      (Uint32 *) dstp, dstrect->w);
slouken@1895
   318
            break;
slouken@1895
   319
        }
slouken@0
   320
#endif
slouken@1895
   321
        pos += inc;
slouken@1895
   322
    }
slouken@894
   323
slouken@1895
   324
    /* We need to unlock the surfaces if they're locked */
slouken@1895
   325
    if (dst_locked) {
slouken@1895
   326
        SDL_UnlockSurface(dst);
slouken@1895
   327
    }
slouken@1895
   328
    if (src_locked) {
slouken@1895
   329
        SDL_UnlockSurface(src);
slouken@1895
   330
    }
slouken@1895
   331
    return (0);
slouken@0
   332
}
slouken@0
   333
slouken@1895
   334
/* vi: set ts=4 sw=4 expandtab: */