drivers/video/aty/mach64_accel.c

   1
   2 /*
   3  *  ATI Mach64 Hardware Acceleration
   4  */
   5
   6 #include <linux/delay.h>
   7 #include <linux/fb.h>
   8 #include <video/mach64.h>
   9 #include "atyfb.h"
  10
  11     /*
  12      *  Generic Mach64 routines
  13      */
  14
  15 /* this is for DMA GUI engine! work in progress */
  16 typedef struct {
  17         u32 frame_buf_offset;
  18         u32 system_mem_addr;
  19         u32 command;
  20         u32 reserved;
  21 } BM_DESCRIPTOR_ENTRY;
  22
  23 #define LAST_DESCRIPTOR (1 << 31)
  24 #define SYSTEM_TO_FRAME_BUFFER 0
  25
  26 static u32 rotation24bpp(u32 dx, u32 direction)
  27 {
  28         u32 rotation;
  29         if (direction & DST_X_LEFT_TO_RIGHT) {
  30                 rotation = (dx / 4) % 6;
  31         } else {
  32                 rotation = ((dx + 2) / 4) % 6;
  33         }
  34
  35         return ((rotation << 8) | DST_24_ROTATION_ENABLE);
  36 }
  37
  38 void aty_reset_engine(const struct atyfb_par *par)
  39 {
  40         /* reset engine */
  41         aty_st_le32(GEN_TEST_CNTL,
  42                 aty_ld_le32(GEN_TEST_CNTL, par) &
  43                 ~(GUI_ENGINE_ENABLE | HWCURSOR_ENABLE), par);
  44         /* enable engine */
  45         aty_st_le32(GEN_TEST_CNTL,
  46                 aty_ld_le32(GEN_TEST_CNTL, par) | GUI_ENGINE_ENABLE, par);
  47         /* ensure engine is not locked up by clearing any FIFO or */
  48         /* HOST errors */
  49         aty_st_le32(BUS_CNTL,
  50                 aty_ld_le32(BUS_CNTL, par) | BUS_HOST_ERR_ACK | BUS_FIFO_ERR_ACK, par);
  51 }
  52
  53 static void reset_GTC_3D_engine(const struct atyfb_par *par)
  54 {
  55         aty_st_le32(SCALE_3D_CNTL, 0xc0, par);
  56         mdelay(GTC_3D_RESET_DELAY);
  57         aty_st_le32(SETUP_CNTL, 0x00, par);
  58         mdelay(GTC_3D_RESET_DELAY);
  59         aty_st_le32(SCALE_3D_CNTL, 0x00, par);
  60         mdelay(GTC_3D_RESET_DELAY);
  61 }
  62
  63 void aty_init_engine(struct atyfb_par *par, struct fb_info *info)
  64 {
  65         u32 pitch_value;
  66
  67         /* determine modal information from global mode structure */
  68         pitch_value = info->var.xres_virtual;
  69
  70         if (info->var.bits_per_pixel == 24) {
  71                 /* In 24 bpp, the engine is in 8 bpp - this requires that all */
  72                 /* horizontal coordinates and widths must be adjusted */
  73                 pitch_value *= 3;
  74         }
  75
  76         /* On GTC (RagePro), we need to reset the 3D engine before */
  77         if (M64_HAS(RESET_3D))
  78                 reset_GTC_3D_engine(par);
  79
  80         /* Reset engine, enable, and clear any engine errors */
  81         aty_reset_engine(par);
  82         /* Ensure that vga page pointers are set to zero - the upper */
  83         /* page pointers are set to 1 to handle overflows in the */
  84         /* lower page */
  85         aty_st_le32(MEM_VGA_WP_SEL, 0x00010000, par);
  86         aty_st_le32(MEM_VGA_RP_SEL, 0x00010000, par);
  87
  88         /* ---- Setup standard engine context ---- */
  89
  90         /* All GUI registers here are FIFOed - therefore, wait for */
  91         /* the appropriate number of empty FIFO entries */
  92         wait_for_fifo(14, par);
  93
  94         /* enable all registers to be loaded for context loads */
  95         aty_st_le32(CONTEXT_MASK, 0xFFFFFFFF, par);
  96
  97         /* set destination pitch to modal pitch, set offset to zero */
  98         aty_st_le32(DST_OFF_PITCH, (pitch_value / 8) << 22, par);
  99
 100         /* zero these registers (set them to a known state) */
 101         aty_st_le32(DST_Y_X, 0, par);
 102         aty_st_le32(DST_HEIGHT, 0, par);
 103         aty_st_le32(DST_BRES_ERR, 0, par);
 104         aty_st_le32(DST_BRES_INC, 0, par);
 105         aty_st_le32(DST_BRES_DEC, 0, par);
 106
 107         /* set destination drawing attributes */
 108         aty_st_le32(DST_CNTL, DST_LAST_PEL | DST_Y_TOP_TO_BOTTOM |
 109                     DST_X_LEFT_TO_RIGHT, par);
 110
 111         /* set source pitch to modal pitch, set offset to zero */
 112         aty_st_le32(SRC_OFF_PITCH, (pitch_value / 8) << 22, par);
 113
 114         /* set these registers to a known state */
 115         aty_st_le32(SRC_Y_X, 0, par);
 116         aty_st_le32(SRC_HEIGHT1_WIDTH1, 1, par);
 117         aty_st_le32(SRC_Y_X_START, 0, par);
 118         aty_st_le32(SRC_HEIGHT2_WIDTH2, 1, par);
 119
 120         /* set source pixel retrieving attributes */
 121         aty_st_le32(SRC_CNTL, SRC_LINE_X_LEFT_TO_RIGHT, par);
 122
 123         /* set host attributes */
 124         wait_for_fifo(13, par);
 125         aty_st_le32(HOST_CNTL, 0, par);
 126
 127         /* set pattern attributes */
 128         aty_st_le32(PAT_REG0, 0, par);
 129         aty_st_le32(PAT_REG1, 0, par);
 130         aty_st_le32(PAT_CNTL, 0, par);
 131
 132         /* set scissors to modal size */
 133         aty_st_le32(SC_LEFT, 0, par);
 134         aty_st_le32(SC_TOP, 0, par);
 135         aty_st_le32(SC_BOTTOM, par->crtc.vyres - 1, par);
 136         aty_st_le32(SC_RIGHT, pitch_value - 1, par);
 137
 138         /* set background color to minimum value (usually BLACK) */
 139         aty_st_le32(DP_BKGD_CLR, 0, par);
 140
 141         /* set foreground color to maximum value (usually WHITE) */
 142         aty_st_le32(DP_FRGD_CLR, 0xFFFFFFFF, par);
 143
 144         /* set write mask to effect all pixel bits */
 145         aty_st_le32(DP_WRITE_MASK, 0xFFFFFFFF, par);
 146
 147         /* set foreground mix to overpaint and background mix to */
 148         /* no-effect */
 149         aty_st_le32(DP_MIX, FRGD_MIX_S | BKGD_MIX_D, par);
 150
 151         /* set primary source pixel channel to foreground color */
 152         /* register */
 153         aty_st_le32(DP_SRC, FRGD_SRC_FRGD_CLR, par);
 154
 155         /* set compare functionality to false (no-effect on */
 156         /* destination) */
 157         wait_for_fifo(3, par);
 158         aty_st_le32(CLR_CMP_CLR, 0, par);
 159         aty_st_le32(CLR_CMP_MASK, 0xFFFFFFFF, par);
 160         aty_st_le32(CLR_CMP_CNTL, 0, par);
 161
 162         /* set pixel depth */
 163         wait_for_fifo(2, par);
 164         aty_st_le32(DP_PIX_WIDTH, par->crtc.dp_pix_width, par);
 165         aty_st_le32(DP_CHAIN_MASK, par->crtc.dp_chain_mask, par);
 166
 167         wait_for_fifo(5, par);
 168         aty_st_le32(SCALE_3D_CNTL, 0, par);
 169         aty_st_le32(Z_CNTL, 0, par);
 170         aty_st_le32(CRTC_INT_CNTL, aty_ld_le32(CRTC_INT_CNTL, par) & ~0x20,
 171                     par);
 172         aty_st_le32(GUI_TRAJ_CNTL, 0x100023, par);
 173
 174         /* insure engine is idle before leaving */
 175         wait_for_idle(par);
 176 }
 177
 178     /*
 179      *  Accelerated functions
 180      */
 181
 182 static inline void draw_rect(s16 x, s16 y, u16 width, u16 height,
 183                              struct atyfb_par *par)
 184 {
 185         /* perform rectangle fill */
 186         wait_for_fifo(2, par);
 187         aty_st_le32(DST_Y_X, (x << 16) | y, par);
 188         aty_st_le32(DST_HEIGHT_WIDTH, (width << 16) | height, par);
 189         par->blitter_may_be_busy = 1;
 190 }
 191
 192 void atyfb_copyarea(struct fb_info *info, const struct fb_copyarea *area)
 193 {
 194         struct atyfb_par *par = (struct atyfb_par *) info->par;
 195         u32 dy = area->dy, sy = area->sy, direction = DST_LAST_PEL;
 196         u32 sx = area->sx, dx = area->dx, width = area->width, rotation = 0;
 197
 198         if (par->asleep)
 199                 return;
 200         if (!area->width || !area->height)
 201                 return;
 202         if (!par->accel_flags) {
 203                 cfb_copyarea(info, area);
 204                 return;
 205         }
 206
 207         if (info->var.bits_per_pixel == 24) {
 208                 /* In 24 bpp, the engine is in 8 bpp - this requires that all */
 209                 /* horizontal coordinates and widths must be adjusted */
 210                 sx *= 3;
 211                 dx *= 3;
 212                 width *= 3;
 213         }
 214
 215         if (area->sy < area->dy) {
 216                 dy += area->height - 1;
 217                 sy += area->height - 1;
 218         } else
 219                 direction |= DST_Y_TOP_TO_BOTTOM;
 220
 221         if (sx < dx) {
 222                 dx += width - 1;
 223                 sx += width - 1;
 224         } else
 225                 direction |= DST_X_LEFT_TO_RIGHT;
 226
 227         if (info->var.bits_per_pixel == 24) {
 228                 rotation = rotation24bpp(dx, direction);
 229         }
 230
 231         wait_for_fifo(4, par);
 232         aty_st_le32(DP_SRC, FRGD_SRC_BLIT, par);
 233         aty_st_le32(SRC_Y_X, (sx << 16) | sy, par);
 234         aty_st_le32(SRC_HEIGHT1_WIDTH1, (width << 16) | area->height, par);
 235         aty_st_le32(DST_CNTL, direction | rotation, par);
 236         draw_rect(dx, dy, width, area->height, par);
 237 }
 238
 239 void atyfb_fillrect(struct fb_info *info, const struct fb_fillrect *rect)
 240 {
 241         struct atyfb_par *par = (struct atyfb_par *) info->par;
 242         u32 color = rect->color, dx = rect->dx, width = rect->width, rotation = 0;
 243
 244         if (par->asleep)
 245                 return;
 246         if (!rect->width || !rect->height)
 247                 return;
 248         if (!par->accel_flags) {
 249                 cfb_fillrect(info, rect);
 250                 return;
 251         }
 252
 253         color |= (rect->color << 8);
 254         color |= (rect->color << 16);
 255
 256         if (info->var.bits_per_pixel == 24) {
 257                 /* In 24 bpp, the engine is in 8 bpp - this requires that all */
 258                 /* horizontal coordinates and widths must be adjusted */
 259                 dx *= 3;
 260                 width *= 3;
 261                 rotation = rotation24bpp(dx, DST_X_LEFT_TO_RIGHT);
 262         }
 263
 264         wait_for_fifo(3, par);
 265         aty_st_le32(DP_FRGD_CLR, color, par);
 266         aty_st_le32(DP_SRC,
 267                     BKGD_SRC_BKGD_CLR | FRGD_SRC_FRGD_CLR | MONO_SRC_ONE,
 268                     par);
 269         aty_st_le32(DST_CNTL,
 270                     DST_LAST_PEL | DST_Y_TOP_TO_BOTTOM |
 271                     DST_X_LEFT_TO_RIGHT | rotation, par);
 272         draw_rect(dx, rect->dy, width, rect->height, par);
 273 }
 274
 275 void atyfb_imageblit(struct fb_info *info, const struct fb_image *image)
 276 {
 277         struct atyfb_par *par = (struct atyfb_par *) info->par;
 278         u32 src_bytes, dx = image->dx, dy = image->dy, width = image->width;
 279         u32 pix_width_save, pix_width, host_cntl, rotation = 0, src, mix;
 280
 281         if (par->asleep)
 282                 return;
 283         if (!image->width || !image->height)
 284                 return;
 285         if (!par->accel_flags ||
 286             (image->depth != 1 && info->var.bits_per_pixel != image->depth)) {
 287                 cfb_imageblit(info, image);
 288                 return;
 289         }
 290
 291         pix_width = pix_width_save = aty_ld_le32(DP_PIX_WIDTH, par);
 292         host_cntl = aty_ld_le32(HOST_CNTL, par) | HOST_BYTE_ALIGN;
 293
 294         switch (image->depth) {
 295         case 1:
 296             pix_width &= ~(BYTE_ORDER_MASK | HOST_MASK);
 297             pix_width |= (BYTE_ORDER_MSB_TO_LSB | HOST_1BPP);
 298             break;
 299         case 4:
 300             pix_width &= ~(BYTE_ORDER_MASK | HOST_MASK);
 301             pix_width |= (BYTE_ORDER_MSB_TO_LSB | HOST_4BPP);
 302             break;
 303         case 8:
 304             pix_width &= ~HOST_MASK;
 305             pix_width |= HOST_8BPP;
 306             break;
 307         case 15:
 308             pix_width &= ~HOST_MASK;
 309             pix_width |= HOST_15BPP;
 310             break;
 311         case 16:
 312             pix_width &= ~HOST_MASK;
 313             pix_width |= HOST_16BPP;
 314             break;
 315         case 24:
 316             pix_width &= ~HOST_MASK;
 317             pix_width |= HOST_24BPP;
 318             break;
 319         case 32:
 320             pix_width &= ~HOST_MASK;
 321             pix_width |= HOST_32BPP;
 322             break;
 323         }
 324
 325         if (info->var.bits_per_pixel == 24) {
 326                 /* In 24 bpp, the engine is in 8 bpp - this requires that all */
 327                 /* horizontal coordinates and widths must be adjusted */
 328                 dx *= 3;
 329                 width *= 3;
 330
 331                 rotation = rotation24bpp(dx, DST_X_LEFT_TO_RIGHT);
 332
 333                 pix_width &= ~DST_MASK;
 334                 pix_width |= DST_8BPP;
 335
 336                 /*
 337                  * since Rage 3D IIc we have DP_HOST_TRIPLE_EN bit
 338                  * this hwaccelerated triple has an issue with not aligned data
 339                  */
 340                 if (M64_HAS(HW_TRIPLE) && image->width % 8 == 0)
 341                         pix_width |= DP_HOST_TRIPLE_EN;
 342         }
 343
 344         if (image->depth == 1) {
 345                 u32 fg, bg;
 346                 if (info->fix.visual == FB_VISUAL_TRUECOLOR ||
 347                     info->fix.visual == FB_VISUAL_DIRECTCOLOR) {
 348                         fg = ((u32*)(info->pseudo_palette))[image->fg_color];
 349                         bg = ((u32*)(info->pseudo_palette))[image->bg_color];
 350                 } else {
 351                         fg = image->fg_color;
 352                         bg = image->bg_color;
 353                 }
 354
 355                 wait_for_fifo(2, par);
 356                 aty_st_le32(DP_BKGD_CLR, bg, par);
 357                 aty_st_le32(DP_FRGD_CLR, fg, par);
 358                 src = MONO_SRC_HOST | FRGD_SRC_FRGD_CLR | BKGD_SRC_BKGD_CLR;
 359                 mix = FRGD_MIX_S | BKGD_MIX_S;
 360         } else {
 361                 src = MONO_SRC_ONE | FRGD_SRC_HOST;
 362                 mix = FRGD_MIX_D_XOR_S | BKGD_MIX_D;
 363         }
 364
 365         wait_for_fifo(6, par);
 366         aty_st_le32(DP_WRITE_MASK, 0xFFFFFFFF, par);
 367         aty_st_le32(DP_PIX_WIDTH, pix_width, par);
 368         aty_st_le32(DP_MIX, mix, par);
 369         aty_st_le32(DP_SRC, src, par);
 370         aty_st_le32(HOST_CNTL, host_cntl, par);
 371         aty_st_le32(DST_CNTL, DST_Y_TOP_TO_BOTTOM | DST_X_LEFT_TO_RIGHT | rotation, par);
 372
 373         draw_rect(dx, dy, width, image->height, par);
 374         src_bytes = (((image->width * image->depth) + 7) / 8) * image->height;
 375
 376         /* manual triple each pixel */
 377         if (info->var.bits_per_pixel == 24 && !(pix_width & DP_HOST_TRIPLE_EN)) {
 378                 int inbit, outbit, mult24, byte_id_in_dword, width;
 379                 u8 *pbitmapin = (u8*)image->data, *pbitmapout;
 380                 u32 hostdword;
 381
 382                 for (width = image->width, inbit = 7, mult24 = 0; src_bytes; ) {
 383                         for (hostdword = 0, pbitmapout = (u8*)&hostdword, byte_id_in_dword = 0;
 384                                 byte_id_in_dword < 4 && src_bytes;
 385                                 byte_id_in_dword++, pbitmapout++) {
 386                                 for (outbit = 7; outbit >= 0; outbit--) {
 387                                         *pbitmapout |= (((*pbitmapin >> inbit) & 1) << outbit);
 388                                         mult24++;
 389                                         /* next bit */
 390                                         if (mult24 == 3) {
 391                                                 mult24 = 0;
 392                                                 inbit--;
 393                                                 width--;
 394                                         }
 395
 396                                         /* next byte */
 397                                         if (inbit < 0 || width == 0) {
 398                                                 src_bytes--;
 399                                                 pbitmapin++;
 400                                                 inbit = 7;
 401
 402                                                 if (width == 0) {
 403                                                     width = image->width;
 404                                                     outbit = 0;
 405                                                 }
 406                                         }
 407                                 }
 408                         }
 409                         wait_for_fifo(1, par);
 410                         aty_st_le32(HOST_DATA0, hostdword, par);
 411                 }
 412         } else {
 413                 u32 *pbitmap, dwords = (src_bytes + 3) / 4;
 414                 for (pbitmap = (u32*)(image->data); dwords; dwords--, pbitmap++) {
 415                         wait_for_fifo(1, par);
 416                         aty_st_le32(HOST_DATA0, le32_to_cpup(pbitmap), par);
 417                 }
 418         }
 419
 420         /* restore pix_width */
 421         wait_for_fifo(1, par);
 422         aty_st_le32(DP_PIX_WIDTH, pix_width_save, par);
 423 }