Mercurial > hg > dmlib
annotate dmblitfunc.h @ 12:35b0c83af3ca
MMX-version of scaled 32-bit RGBA alpha blending blit works now.
| author | Matti Hamalainen <ccr@tnsp.org> |
|---|---|
| date | Fri, 28 Sep 2012 09:35:25 +0300 |
| parents | 62b2c3de93a8 |
| children | 07bace8425d9 |
| rev | line source |
|---|---|
| 0 | 1 /* |
| 2 * DMLib | |
| 3 * -- Sprite / surface blitting functions | |
| 4 * Programmed and designed by Matti 'ccr' Hamalainen | |
| 5 * (C) Copyright 2011-2012 Tecnic Software productions (TNSP) | |
| 6 */ | |
| 7 | |
| 8 // ======================================================================= | |
| 9 // DMD_NONE | |
| 10 // ======================================================================= | |
| 11 | |
| 12 #define DM_SCALED_NAME dmScaledBlitSurface8to8 | |
| 13 #define DM_SCALED_SRC_BYTES 1 | |
| 14 #define DM_SCALED_DST_BYTES 1 | |
| 15 #define DM_SCALED_SRC_TYPE Uint8 | |
| 16 #define DM_SCALED_DST_TYPE Uint8 | |
| 17 | |
| 18 #define DM_SCALED_INNER *dp++ = sp[FP_GETH(xv)]; | |
| 19 | |
| 20 #include "dmscaledblit.h" | |
| 21 | |
| 22 // ----------------------------------------------------------------------- | |
| 23 | |
| 24 #define DM_SCALED_NAME dmScaledBlitSurface8to32 | |
| 25 #define DM_SCALED_SRC_BYTES 1 | |
| 26 #define DM_SCALED_DST_BYTES 4 | |
| 27 #define DM_SCALED_SRC_TYPE Uint8 | |
| 28 #define DM_SCALED_DST_TYPE Uint32 | |
| 29 | |
| 30 #define DM_SCALED_VARS const Uint32 *pal; | |
| 31 #define DM_SCALED_INIT \ | |
| 32 if (src->format->palette == NULL || src->format->palette->ncolors < 256) return -2; \ | |
| 33 pal = (Uint32 *) src->format->palette->colors; | |
| 34 | |
| 35 #define DM_SCALED_INNER *dp++ = pal[sp[FP_GETH(xv)]]; | |
| 36 | |
| 37 #include "dmscaledblit.h" | |
| 38 | |
| 39 // ----------------------------------------------------------------------- | |
| 40 | |
| 41 #define DM_SCALED_NAME dmScaledBlitSurface32to32 | |
| 42 #define DM_SCALED_SRC_BYTES 4 | |
| 43 #define DM_SCALED_DST_BYTES 4 | |
| 44 #define DM_SCALED_SRC_TYPE Uint32 | |
| 45 #define DM_SCALED_DST_TYPE Uint32 | |
| 46 #define DM_SCALED_INIT | |
| 47 | |
| 48 #define DM_SCALED_INNER *dp++ = sp[FP_GETH(xv)]; | |
| 49 | |
| 50 #include "dmscaledblit.h" | |
| 51 | |
| 52 | |
| 53 // ======================================================================= | |
| 54 // DMD_TRANSPARENT | |
| 55 // ======================================================================= | |
| 56 | |
| 57 #define DM_SCALED_NAME dmScaledBlitSurface8to8Transparent | |
| 58 #define DM_SCALED_SRC_BYTES 1 | |
| 59 #define DM_SCALED_DST_BYTES 1 | |
| 60 #define DM_SCALED_SRC_TYPE Uint8 | |
| 61 #define DM_SCALED_DST_TYPE Uint8 | |
| 62 | |
| 63 #define DM_SCALED_INNER \ | |
| 64 *dp = sp[FP_GETH(xv)] ? sp[FP_GETH(xv)] : *dp; dp++; | |
| 65 | |
| 66 #include "dmscaledblit.h" | |
| 67 | |
| 68 // ----------------------------------------------------------------------- | |
| 69 | |
| 70 #define DM_SCALED_NAME dmScaledBlitSurface8to32Transparent | |
| 71 #define DM_SCALED_SRC_BYTES 1 | |
| 72 #define DM_SCALED_DST_BYTES 4 | |
| 73 #define DM_SCALED_SRC_TYPE Uint8 | |
| 74 #define DM_SCALED_DST_TYPE DMRGBA32 | |
| 75 #define DM_SCALED_VARS const DMRGBA32 *pal; | |
| 76 #define DM_SCALED_INIT \ | |
| 77 if (src->format->palette == NULL || src->format->palette->ncolors < 256) return -2; \ | |
| 78 pal = (DMRGBA32 *) src->format->palette->colors; | |
| 79 | |
| 80 #define DM_SCALED_INNER \ | |
| 81 const DMRGBA32 q = pal[sp[FP_GETH(xv)]]; \ | |
| 82 dp->r += ((q.r - dp->r) * q.a) >> 8; \ | |
| 83 dp->g += ((q.g - dp->g) * q.a) >> 8; \ | |
| 84 dp->b += ((q.b - dp->b) * q.a) >> 8; \ | |
| 85 dp->a = q.a; \ | |
| 86 dp++; | |
| 87 | |
| 88 #include "dmscaledblit.h" | |
| 89 | |
| 90 | |
| 91 // ----------------------------------------------------------------------- | |
| 92 | |
| 93 #define DM_SCALED_NAME dmScaledBlitSurface32to32Transparent | |
| 94 #define DM_SCALED_SRC_BYTES 4 | |
| 95 #define DM_SCALED_DST_BYTES 4 | |
| 96 #define DM_SCALED_INIT | |
| 97 | |
| 98 #ifdef DM_USE_SIMD | |
| 99 #define DM_SCALED_INNER_INIT \ | |
|
12
35b0c83af3ca
MMX-version of scaled 32-bit RGBA alpha blending blit works now.
Matti Hamalainen <ccr@tnsp.org>
parents:
10
diff
changeset
|
100 const Uint32 qpdmask = 0xff000000; \ |
|
35b0c83af3ca
MMX-version of scaled 32-bit RGBA alpha blending blit works now.
Matti Hamalainen <ccr@tnsp.org>
parents:
10
diff
changeset
|
101 const Uint64 qpdrm = 0xff00ff00ff00ffULL; |
| 0 | 102 |
| 103 #define DM_SCALED_SRC_TYPE Uint32 | |
| 104 #define DM_SCALED_DST_TYPE Uint32 | |
| 105 #define DM_SCALED_INNER \ | |
| 106 asm( \ | |
| 107 "movd %2, %%mm1\n" \ | |
|
12
35b0c83af3ca
MMX-version of scaled 32-bit RGBA alpha blending blit works now.
Matti Hamalainen <ccr@tnsp.org>
parents:
10
diff
changeset
|
108 \ |
|
35b0c83af3ca
MMX-version of scaled 32-bit RGBA alpha blending blit works now.
Matti Hamalainen <ccr@tnsp.org>
parents:
10
diff
changeset
|
109 "movd %3, %%mm2\n" \ |
| 0 | 110 "movq %%mm1, %%mm3\n" \ |
|
12
35b0c83af3ca
MMX-version of scaled 32-bit RGBA alpha blending blit works now.
Matti Hamalainen <ccr@tnsp.org>
parents:
10
diff
changeset
|
111 "pand %%mm2, %%mm3\n" \ |
| 0 | 112 "movq %%mm3, %%mm4\n" \ |
|
12
35b0c83af3ca
MMX-version of scaled 32-bit RGBA alpha blending blit works now.
Matti Hamalainen <ccr@tnsp.org>
parents:
10
diff
changeset
|
113 "psrlq $8, %%mm3\n" \ |
|
35b0c83af3ca
MMX-version of scaled 32-bit RGBA alpha blending blit works now.
Matti Hamalainen <ccr@tnsp.org>
parents:
10
diff
changeset
|
114 "por %%mm3, %%mm4\n" \ |
|
35b0c83af3ca
MMX-version of scaled 32-bit RGBA alpha blending blit works now.
Matti Hamalainen <ccr@tnsp.org>
parents:
10
diff
changeset
|
115 "movq %%mm4, %%mm3\n" \ |
|
35b0c83af3ca
MMX-version of scaled 32-bit RGBA alpha blending blit works now.
Matti Hamalainen <ccr@tnsp.org>
parents:
10
diff
changeset
|
116 "psrlq $16, %%mm4\n" \ |
| 0 | 117 "por %%mm3, %%mm4\n" \ |
|
12
35b0c83af3ca
MMX-version of scaled 32-bit RGBA alpha blending blit works now.
Matti Hamalainen <ccr@tnsp.org>
parents:
10
diff
changeset
|
118 "pxor %%mm5, %%mm5\n" \ |
|
35b0c83af3ca
MMX-version of scaled 32-bit RGBA alpha blending blit works now.
Matti Hamalainen <ccr@tnsp.org>
parents:
10
diff
changeset
|
119 "movq %%mm5, %%mm2\n" \ |
|
35b0c83af3ca
MMX-version of scaled 32-bit RGBA alpha blending blit works now.
Matti Hamalainen <ccr@tnsp.org>
parents:
10
diff
changeset
|
120 "punpcklbw %%mm5, %%mm4\n" \ |
|
35b0c83af3ca
MMX-version of scaled 32-bit RGBA alpha blending blit works now.
Matti Hamalainen <ccr@tnsp.org>
parents:
10
diff
changeset
|
121 "movq %%mm4, %%mm5\n" \ |
| 0 | 122 \ |
|
12
35b0c83af3ca
MMX-version of scaled 32-bit RGBA alpha blending blit works now.
Matti Hamalainen <ccr@tnsp.org>
parents:
10
diff
changeset
|
123 "movq %%mm2, %%mm4\n" \ |
| 0 | 124 \ |
| 125 "movd %1, %%mm3\n" \ | |
| 126 "punpcklbw %%mm2, %%mm1\n" \ | |
|
12
35b0c83af3ca
MMX-version of scaled 32-bit RGBA alpha blending blit works now.
Matti Hamalainen <ccr@tnsp.org>
parents:
10
diff
changeset
|
127 "punpcklbw %%mm4, %%mm3\n" \ |
|
35b0c83af3ca
MMX-version of scaled 32-bit RGBA alpha blending blit works now.
Matti Hamalainen <ccr@tnsp.org>
parents:
10
diff
changeset
|
128 \ |
| 0 | 129 "psubw %%mm3, %%mm1\n" \ |
|
12
35b0c83af3ca
MMX-version of scaled 32-bit RGBA alpha blending blit works now.
Matti Hamalainen <ccr@tnsp.org>
parents:
10
diff
changeset
|
130 "pmullw %%mm5, %%mm1\n" \ |
|
35b0c83af3ca
MMX-version of scaled 32-bit RGBA alpha blending blit works now.
Matti Hamalainen <ccr@tnsp.org>
parents:
10
diff
changeset
|
131 "psraw $8, %%mm1\n" \ |
|
35b0c83af3ca
MMX-version of scaled 32-bit RGBA alpha blending blit works now.
Matti Hamalainen <ccr@tnsp.org>
parents:
10
diff
changeset
|
132 "paddw %%mm3, %%mm1\n" \ |
|
35b0c83af3ca
MMX-version of scaled 32-bit RGBA alpha blending blit works now.
Matti Hamalainen <ccr@tnsp.org>
parents:
10
diff
changeset
|
133 "pand %4, %%mm1\n" \ |
| 0 | 134 "packuswb %%mm2, %%mm1\n" \ |
| 135 "movd %%mm1, %0\n" \ | |
| 136 : "=m" (*dp) \ | |
|
12
35b0c83af3ca
MMX-version of scaled 32-bit RGBA alpha blending blit works now.
Matti Hamalainen <ccr@tnsp.org>
parents:
10
diff
changeset
|
137 : "m" (*dp), "m" (sp[FP_GETH(xv)]), "m" (qpdmask), "m" (qpdrm) \ |
| 0 | 138 : "memory" ); dp++; |
| 139 | |
| 140 #define DM_SCALED_FINISH asm("emms\n"); | |
| 141 #else | |
| 142 #define DM_SCALED_SRC_TYPE DMRGBA32 | |
| 143 #define DM_SCALED_DST_TYPE DMRGBA32 | |
| 144 #define DM_SCALED_INNER \ | |
| 145 const DMRGBA32 q = sp[FP_GETH(xv)]; \ | |
| 146 dp->r += ((q.r - dp->r) * q.a) >> 8; \ | |
| 147 dp->g += ((q.g - dp->g) * q.a) >> 8; \ | |
| 148 dp->b += ((q.b - dp->b) * q.a) >> 8; \ | |
| 149 dp->a = q.a; \ | |
| 150 dp++; | |
| 151 #endif | |
| 152 | |
| 153 #include "dmscaledblit.h" | |
| 154 | |
| 155 // ----------------------------------------------------------------------- | |
| 156 | |
| 157 #define DM_SCALED_NAME dmScaledBlitSurface32to32TransparentX | |
| 158 #define DM_SCALED_SRC_BYTES 4 | |
| 159 #define DM_SCALED_DST_BYTES 4 | |
| 160 #define DM_SCALED_SRC_TYPE DMRGBA32 | |
| 161 #define DM_SCALED_DST_TYPE DMRGBA32 | |
| 162 #define DM_SCALED_INIT | |
| 163 | |
| 164 #define DM_SCALED_INNER \ | |
| 165 const DMRGBA32 q = sp[FP_GETH(xv)]; \ | |
| 166 dp->r = (q.r * q.a + dp->r * dp->a) >> 9; \ | |
| 167 dp->g = (q.g * q.a + dp->g * dp->a) >> 9; \ | |
| 168 dp->b = (q.b * q.a + dp->b * dp->a) >> 9; \ | |
| 169 dp->a = q.a ? q.a : dp->a; \ | |
| 170 dp++; | |
| 171 | |
| 172 #include "dmscaledblit.h" | |
| 173 | |
| 174 | |
| 175 // ======================================================================= | |
| 176 // DMD_SATURATE | |
| 177 // ======================================================================= | |
| 178 | |
| 179 #define DM_SCALED_NAME dmScaledBlitSurface8to8Saturate | |
| 180 #define DM_SCALED_SRC_BYTES 1 | |
| 181 #define DM_SCALED_DST_BYTES 1 | |
| 182 #define DM_SCALED_SRC_TYPE Uint8 | |
| 183 #define DM_SCALED_DST_TYPE Uint8 | |
| 184 #define DM_SCALED_INNER *(dp++) += sp[FP_GETH(xv)]; | |
| 185 #include "dmscaledblit.h" | |
| 186 | |
| 187 // ----------------------------------------------------------------------- | |
| 188 | |
| 189 #define DM_SCALED_NAME dmScaledBlitSurface8to32Saturate | |
| 190 #define DM_SCALED_SRC_BYTES 1 | |
| 191 #define DM_SCALED_DST_BYTES 4 | |
| 192 #define DM_SCALED_SRC_TYPE Uint8 | |
| 193 #define DM_SCALED_DST_TYPE DMRGBA32 | |
| 194 #define DM_SCALED_VARS const DMRGBA32 *pal; | |
| 195 #define DM_SCALED_INIT \ | |
| 196 if (src->format->palette == NULL || src->format->palette->ncolors < 256) return -2; \ | |
| 197 pal = (DMRGBA32 *) src->format->palette->colors; | |
| 198 | |
| 199 #ifdef DM_USE_SIMD | |
| 200 #define DM_SCALED_INNER \ | |
| 201 asm("movd %2, %%mm1\n" \ | |
| 202 "movd %1, %%mm2\n" \ | |
| 203 "paddusb %%mm2, %%mm1\n" \ | |
| 204 "movd %%mm1, %0\n" \ | |
| 205 : "=m" (*dp) \ | |
| 206 : "m" (*dp), "m" (pal[sp[FP_GETH(xv)]]) \ | |
| 207 : "memory" ); dp++; | |
| 208 | |
| 209 #define DM_SCALED_FINISH asm("emms\n"); | |
| 210 #else | |
| 211 #define DM_SCALED_INNER \ | |
| 212 const DMRGBA32 q = pal[sp[FP_GETH(xv)]]; \ | |
| 213 const int qr = dp->r + q.r, qg = dp->g + q.g, qb = dp->b + q.b; \ | |
| 214 dp->r = qr < 256 ? qr : 255; \ | |
| 215 dp->g = qg < 256 ? qg : 255; \ | |
| 216 dp->b = qb < 256 ? qb : 255; \ | |
| 217 dp->a = q.a; \ | |
| 218 dp++; | |
| 219 #endif | |
| 220 | |
| 221 #include "dmscaledblit.h" | |
| 222 | |
| 223 // ----------------------------------------------------------------------- | |
| 224 | |
| 225 #define DM_SCALED_NAME dmScaledBlitSurface32to32Saturate | |
| 226 #define DM_SCALED_SRC_BYTES 4 | |
| 227 #define DM_SCALED_DST_BYTES 4 | |
| 228 #define DM_SCALED_INIT | |
| 229 | |
| 230 #ifdef DM_USE_SIMD | |
| 231 #define DM_SCALED_SRC_TYPE Uint32 | |
| 232 #define DM_SCALED_DST_TYPE Uint32 | |
| 233 #define DM_SCALED_INNER \ | |
| 234 asm("movd %2, %%mm1\n" \ | |
| 235 "movd %1, %%mm2\n" \ | |
| 236 "paddusb %%mm2, %%mm1\n" \ | |
| 237 "movd %%mm1, %0\n" \ | |
| 238 : "=m" (*dp) \ | |
| 239 : "m" (*dp), "m" (sp[FP_GETH(xv)]) \ | |
| 240 : "memory" ); dp++; | |
| 241 | |
| 242 #define DM_SCALED_FINISH asm("emms\n"); | |
| 243 #else | |
| 244 #define DM_SCALED_SRC_TYPE DMRGBA32 | |
| 245 #define DM_SCALED_DST_TYPE DMRGBA32 | |
| 246 #define DM_SCALED_INNER \ | |
| 247 const DMRGBA32 q = sp[FP_GETH(xv)]; \ | |
| 248 const int qr = dp->r + q.r, qg = dp->g + q.g, qb = dp->b + q.b; \ | |
| 249 dp->r = qr < 256 ? qr : 255; \ | |
| 250 dp->g = qg < 256 ? qg : 255; \ | |
| 251 dp->b = qb < 256 ? qb : 255; \ | |
| 252 dp->a = q.a; \ | |
| 253 dp++; | |
| 254 #endif | |
| 255 | |
| 256 #include "dmscaledblit.h" | |
| 257 | |
| 258 | |
| 259 // ======================================================================= | |
| 260 // ======================================================================= |