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