drivers/gpu/drm/radeon/r600_blit.c

   1 /*
   2  * Copyright 2009 Advanced Micro Devices, Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE COPYRIGHT HOLDER(S) AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
  19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
  21  * DEALINGS IN THE SOFTWARE.
  22  *
  23  * Authors:
  24  *     Alex Deucher <alexander.deucher@amd.com>
  25  */
  26 #include "drmP.h"
  27 #include "drm.h"
  28 #include "radeon_drm.h"
  29 #include "radeon_drv.h"
  30
  31 #include "r600_blit_shaders.h"
  32
  33 #define DI_PT_RECTLIST        0x11
  34 #define DI_INDEX_SIZE_16_BIT  0x0
  35 #define DI_SRC_SEL_AUTO_INDEX 0x2
  36
  37 #define FMT_8                 0x1
  38 #define FMT_5_6_5             0x8
  39 #define FMT_8_8_8_8           0x1a
  40 #define COLOR_8               0x1
  41 #define COLOR_5_6_5           0x8
  42 #define COLOR_8_8_8_8         0x1a
  43
  44 static inline void
  45 set_render_target(drm_radeon_private_t *dev_priv, int format, int w, int h, u64 gpu_addr)
  46 {
  47         u32 cb_color_info;
  48         int pitch, slice;
  49         RING_LOCALS;
  50         DRM_DEBUG("\n");
  51
  52         h = ALIGN(h, 8);
  53         if (h < 8)
  54                 h = 8;
  55
  56         cb_color_info = ((format << 2) | (1 << 27));
  57         pitch = (w / 8) - 1;
  58         slice = ((w * h) / 64) - 1;
  59
  60         if (((dev_priv->flags & RADEON_FAMILY_MASK) > CHIP_R600) &&
  61             ((dev_priv->flags & RADEON_FAMILY_MASK) < CHIP_RV770)) {
  62                 BEGIN_RING(21 + 2);
  63                 OUT_RING(CP_PACKET3(R600_IT_SET_CONTEXT_REG, 1));
  64                 OUT_RING((R600_CB_COLOR0_BASE - R600_SET_CONTEXT_REG_OFFSET) >> 2);
  65                 OUT_RING(gpu_addr >> 8);
  66                 OUT_RING(CP_PACKET3(R600_IT_SURFACE_BASE_UPDATE, 0));
  67                 OUT_RING(2 << 0);
  68         } else {
  69                 BEGIN_RING(21);
  70                 OUT_RING(CP_PACKET3(R600_IT_SET_CONTEXT_REG, 1));
  71                 OUT_RING((R600_CB_COLOR0_BASE - R600_SET_CONTEXT_REG_OFFSET) >> 2);
  72                 OUT_RING(gpu_addr >> 8);
  73         }
  74
  75         OUT_RING(CP_PACKET3(R600_IT_SET_CONTEXT_REG, 1));
  76         OUT_RING((R600_CB_COLOR0_SIZE - R600_SET_CONTEXT_REG_OFFSET) >> 2);
  77         OUT_RING((pitch << 0) | (slice << 10));
  78
  79         OUT_RING(CP_PACKET3(R600_IT_SET_CONTEXT_REG, 1));
  80         OUT_RING((R600_CB_COLOR0_VIEW - R600_SET_CONTEXT_REG_OFFSET) >> 2);
  81         OUT_RING(0);
  82
  83         OUT_RING(CP_PACKET3(R600_IT_SET_CONTEXT_REG, 1));
  84         OUT_RING((R600_CB_COLOR0_INFO - R600_SET_CONTEXT_REG_OFFSET) >> 2);
  85         OUT_RING(cb_color_info);
  86
  87         OUT_RING(CP_PACKET3(R600_IT_SET_CONTEXT_REG, 1));
  88         OUT_RING((R600_CB_COLOR0_TILE - R600_SET_CONTEXT_REG_OFFSET) >> 2);
  89         OUT_RING(0);
  90
  91         OUT_RING(CP_PACKET3(R600_IT_SET_CONTEXT_REG, 1));
  92         OUT_RING((R600_CB_COLOR0_FRAG - R600_SET_CONTEXT_REG_OFFSET) >> 2);
  93         OUT_RING(0);
  94
  95         OUT_RING(CP_PACKET3(R600_IT_SET_CONTEXT_REG, 1));
  96         OUT_RING((R600_CB_COLOR0_MASK - R600_SET_CONTEXT_REG_OFFSET) >> 2);
  97         OUT_RING(0);
  98
  99         ADVANCE_RING();
 100 }
 101
 102 static inline void
 103 cp_set_surface_sync(drm_radeon_private_t *dev_priv,
 104                     u32 sync_type, u32 size, u64 mc_addr)
 105 {
 106         u32 cp_coher_size;
 107         RING_LOCALS;
 108         DRM_DEBUG("\n");
 109
 110         if (size == 0xffffffff)
 111                 cp_coher_size = 0xffffffff;
 112         else
 113                 cp_coher_size = ((size + 255) >> 8);
 114
 115         BEGIN_RING(5);
 116         OUT_RING(CP_PACKET3(R600_IT_SURFACE_SYNC, 3));
 117         OUT_RING(sync_type);
 118         OUT_RING(cp_coher_size);
 119         OUT_RING((mc_addr >> 8));
 120         OUT_RING(10); /* poll interval */
 121         ADVANCE_RING();
 122 }
 123
 124 static inline void
 125 set_shaders(struct drm_device *dev)
 126 {
 127         drm_radeon_private_t *dev_priv = dev->dev_private;
 128         u64 gpu_addr;
 129         int i;
 130         u32 *vs, *ps;
 131         uint32_t sq_pgm_resources;
 132         RING_LOCALS;
 133         DRM_DEBUG("\n");
 134
 135         /* load shaders */
 136         vs = (u32 *) ((char *)dev->agp_buffer_map->handle + dev_priv->blit_vb->offset);
 137         ps = (u32 *) ((char *)dev->agp_buffer_map->handle + dev_priv->blit_vb->offset + 256);
 138
 139         for (i = 0; i < r6xx_vs_size; i++)
 140                 vs[i] = cpu_to_le32(r6xx_vs[i]);
 141         for (i = 0; i < r6xx_ps_size; i++)
 142                 ps[i] = cpu_to_le32(r6xx_ps[i]);
 143
 144         dev_priv->blit_vb->used = 512;
 145
 146         gpu_addr = dev_priv->gart_buffers_offset + dev_priv->blit_vb->offset;
 147
 148         /* setup shader regs */
 149         sq_pgm_resources = (1 << 0);
 150
 151         BEGIN_RING(9 + 12);
 152         /* VS */
 153         OUT_RING(CP_PACKET3(R600_IT_SET_CONTEXT_REG, 1));
 154         OUT_RING((R600_SQ_PGM_START_VS - R600_SET_CONTEXT_REG_OFFSET) >> 2);
 155         OUT_RING(gpu_addr >> 8);
 156
 157         OUT_RING(CP_PACKET3(R600_IT_SET_CONTEXT_REG, 1));
 158         OUT_RING((R600_SQ_PGM_RESOURCES_VS - R600_SET_CONTEXT_REG_OFFSET) >> 2);
 159         OUT_RING(sq_pgm_resources);
 160
 161         OUT_RING(CP_PACKET3(R600_IT_SET_CONTEXT_REG, 1));
 162         OUT_RING((R600_SQ_PGM_CF_OFFSET_VS - R600_SET_CONTEXT_REG_OFFSET) >> 2);
 163         OUT_RING(0);
 164
 165         /* PS */
 166         OUT_RING(CP_PACKET3(R600_IT_SET_CONTEXT_REG, 1));
 167         OUT_RING((R600_SQ_PGM_START_PS - R600_SET_CONTEXT_REG_OFFSET) >> 2);
 168         OUT_RING((gpu_addr + 256) >> 8);
 169
 170         OUT_RING(CP_PACKET3(R600_IT_SET_CONTEXT_REG, 1));
 171         OUT_RING((R600_SQ_PGM_RESOURCES_PS - R600_SET_CONTEXT_REG_OFFSET) >> 2);
 172         OUT_RING(sq_pgm_resources | (1 << 28));
 173
 174         OUT_RING(CP_PACKET3(R600_IT_SET_CONTEXT_REG, 1));
 175         OUT_RING((R600_SQ_PGM_EXPORTS_PS - R600_SET_CONTEXT_REG_OFFSET) >> 2);
 176         OUT_RING(2);
 177
 178         OUT_RING(CP_PACKET3(R600_IT_SET_CONTEXT_REG, 1));
 179         OUT_RING((R600_SQ_PGM_CF_OFFSET_PS - R600_SET_CONTEXT_REG_OFFSET) >> 2);
 180         OUT_RING(0);
 181         ADVANCE_RING();
 182
 183         cp_set_surface_sync(dev_priv,
 184                             R600_SH_ACTION_ENA, 512, gpu_addr);
 185 }
 186
 187 static inline void
 188 set_vtx_resource(drm_radeon_private_t *dev_priv, u64 gpu_addr)
 189 {
 190         uint32_t sq_vtx_constant_word2;
 191         RING_LOCALS;
 192         DRM_DEBUG("\n");
 193
 194         sq_vtx_constant_word2 = (((gpu_addr >> 32) & 0xff) | (16 << 8));
 195 #ifdef __BIG_ENDIAN
 196         sq_vtx_constant_word2 |= (2 << 30);
 197 #endif
 198
 199         BEGIN_RING(9);
 200         OUT_RING(CP_PACKET3(R600_IT_SET_RESOURCE, 7));
 201         OUT_RING(0x460);
 202         OUT_RING(gpu_addr & 0xffffffff);
 203         OUT_RING(48 - 1);
 204         OUT_RING(sq_vtx_constant_word2);
 205         OUT_RING(1 << 0);
 206         OUT_RING(0);
 207         OUT_RING(0);
 208         OUT_RING(R600_SQ_TEX_VTX_VALID_BUFFER << 30);
 209         ADVANCE_RING();
 210
 211         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV610) ||
 212             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV620) ||
 213             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS780) ||
 214             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS880) ||
 215             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV710))
 216                 cp_set_surface_sync(dev_priv,
 217                                     R600_TC_ACTION_ENA, 48, gpu_addr);
 218         else
 219                 cp_set_surface_sync(dev_priv,
 220                                     R600_VC_ACTION_ENA, 48, gpu_addr);
 221 }
 222
 223 static inline void
 224 set_tex_resource(drm_radeon_private_t *dev_priv,
 225                  int format, int w, int h, int pitch, u64 gpu_addr)
 226 {
 227         uint32_t sq_tex_resource_word0, sq_tex_resource_word1, sq_tex_resource_word4;
 228         RING_LOCALS;
 229         DRM_DEBUG("\n");
 230
 231         if (h < 1)
 232                 h = 1;
 233
 234         sq_tex_resource_word0 = (1 << 0);
 235         sq_tex_resource_word0 |= ((((pitch >> 3) - 1) << 8) |
 236                                   ((w - 1) << 19));
 237
 238         sq_tex_resource_word1 = (format << 26);
 239         sq_tex_resource_word1 |= ((h - 1) << 0);
 240
 241         sq_tex_resource_word4 = ((1 << 14) |
 242                                  (0 << 16) |
 243                                  (1 << 19) |
 244                                  (2 << 22) |
 245                                  (3 << 25));
 246
 247         BEGIN_RING(9);
 248         OUT_RING(CP_PACKET3(R600_IT_SET_RESOURCE, 7));
 249         OUT_RING(0);
 250         OUT_RING(sq_tex_resource_word0);
 251         OUT_RING(sq_tex_resource_word1);
 252         OUT_RING(gpu_addr >> 8);
 253         OUT_RING(gpu_addr >> 8);
 254         OUT_RING(sq_tex_resource_word4);
 255         OUT_RING(0);
 256         OUT_RING(R600_SQ_TEX_VTX_VALID_TEXTURE << 30);
 257         ADVANCE_RING();
 258
 259 }
 260
 261 static inline void
 262 set_scissors(drm_radeon_private_t *dev_priv, int x1, int y1, int x2, int y2)
 263 {
 264         RING_LOCALS;
 265         DRM_DEBUG("\n");
 266
 267         BEGIN_RING(12);
 268         OUT_RING(CP_PACKET3(R600_IT_SET_CONTEXT_REG, 2));
 269         OUT_RING((R600_PA_SC_SCREEN_SCISSOR_TL - R600_SET_CONTEXT_REG_OFFSET) >> 2);
 270         OUT_RING((x1 << 0) | (y1 << 16));
 271         OUT_RING((x2 << 0) | (y2 << 16));
 272
 273         OUT_RING(CP_PACKET3(R600_IT_SET_CONTEXT_REG, 2));
 274         OUT_RING((R600_PA_SC_GENERIC_SCISSOR_TL - R600_SET_CONTEXT_REG_OFFSET) >> 2);
 275         OUT_RING((x1 << 0) | (y1 << 16) | (1 << 31));
 276         OUT_RING((x2 << 0) | (y2 << 16));
 277
 278         OUT_RING(CP_PACKET3(R600_IT_SET_CONTEXT_REG, 2));
 279         OUT_RING((R600_PA_SC_WINDOW_SCISSOR_TL - R600_SET_CONTEXT_REG_OFFSET) >> 2);
 280         OUT_RING((x1 << 0) | (y1 << 16) | (1 << 31));
 281         OUT_RING((x2 << 0) | (y2 << 16));
 282         ADVANCE_RING();
 283 }
 284
 285 static inline void
 286 draw_auto(drm_radeon_private_t *dev_priv)
 287 {
 288         RING_LOCALS;
 289         DRM_DEBUG("\n");
 290
 291         BEGIN_RING(10);
 292         OUT_RING(CP_PACKET3(R600_IT_SET_CONFIG_REG, 1));
 293         OUT_RING((R600_VGT_PRIMITIVE_TYPE - R600_SET_CONFIG_REG_OFFSET) >> 2);
 294         OUT_RING(DI_PT_RECTLIST);
 295
 296         OUT_RING(CP_PACKET3(R600_IT_INDEX_TYPE, 0));
 297 #ifdef __BIG_ENDIAN
 298         OUT_RING((2 << 2) | DI_INDEX_SIZE_16_BIT);
 299 #else
 300         OUT_RING(DI_INDEX_SIZE_16_BIT);
 301 #endif
 302
 303         OUT_RING(CP_PACKET3(R600_IT_NUM_INSTANCES, 0));
 304         OUT_RING(1);
 305
 306         OUT_RING(CP_PACKET3(R600_IT_DRAW_INDEX_AUTO, 1));
 307         OUT_RING(3);
 308         OUT_RING(DI_SRC_SEL_AUTO_INDEX);
 309
 310         ADVANCE_RING();
 311         COMMIT_RING();
 312 }
 313
 314 static inline void
 315 set_default_state(drm_radeon_private_t *dev_priv)
 316 {
 317         int i;
 318         u32 sq_config, sq_gpr_resource_mgmt_1, sq_gpr_resource_mgmt_2;
 319         u32 sq_thread_resource_mgmt, sq_stack_resource_mgmt_1, sq_stack_resource_mgmt_2;
 320         int num_ps_gprs, num_vs_gprs, num_temp_gprs, num_gs_gprs, num_es_gprs;
 321         int num_ps_threads, num_vs_threads, num_gs_threads, num_es_threads;
 322         int num_ps_stack_entries, num_vs_stack_entries, num_gs_stack_entries, num_es_stack_entries;
 323         RING_LOCALS;
 324
 325         switch ((dev_priv->flags & RADEON_FAMILY_MASK)) {
 326         case CHIP_R600:
 327                 num_ps_gprs = 192;
 328                 num_vs_gprs = 56;
 329                 num_temp_gprs = 4;
 330                 num_gs_gprs = 0;
 331                 num_es_gprs = 0;
 332                 num_ps_threads = 136;
 333                 num_vs_threads = 48;
 334                 num_gs_threads = 4;
 335                 num_es_threads = 4;
 336                 num_ps_stack_entries = 128;
 337                 num_vs_stack_entries = 128;
 338                 num_gs_stack_entries = 0;
 339                 num_es_stack_entries = 0;
 340                 break;
 341         case CHIP_RV630:
 342         case CHIP_RV635:
 343                 num_ps_gprs = 84;
 344                 num_vs_gprs = 36;
 345                 num_temp_gprs = 4;
 346                 num_gs_gprs = 0;
 347                 num_es_gprs = 0;
 348                 num_ps_threads = 144;
 349                 num_vs_threads = 40;
 350                 num_gs_threads = 4;
 351                 num_es_threads = 4;
 352                 num_ps_stack_entries = 40;
 353                 num_vs_stack_entries = 40;
 354                 num_gs_stack_entries = 32;
 355                 num_es_stack_entries = 16;
 356                 break;
 357         case CHIP_RV610:
 358         case CHIP_RV620:
 359         case CHIP_RS780:
 360         case CHIP_RS880:
 361         default:
 362                 num_ps_gprs = 84;
 363                 num_vs_gprs = 36;
 364                 num_temp_gprs = 4;
 365                 num_gs_gprs = 0;
 366                 num_es_gprs = 0;
 367                 num_ps_threads = 136;
 368                 num_vs_threads = 48;
 369                 num_gs_threads = 4;
 370                 num_es_threads = 4;
 371                 num_ps_stack_entries = 40;
 372                 num_vs_stack_entries = 40;
 373                 num_gs_stack_entries = 32;
 374                 num_es_stack_entries = 16;
 375                 break;
 376         case CHIP_RV670:
 377                 num_ps_gprs = 144;
 378                 num_vs_gprs = 40;
 379                 num_temp_gprs = 4;
 380                 num_gs_gprs = 0;
 381                 num_es_gprs = 0;
 382                 num_ps_threads = 136;
 383                 num_vs_threads = 48;
 384                 num_gs_threads = 4;
 385                 num_es_threads = 4;
 386                 num_ps_stack_entries = 40;
 387                 num_vs_stack_entries = 40;
 388                 num_gs_stack_entries = 32;
 389                 num_es_stack_entries = 16;
 390                 break;
 391         case CHIP_RV770:
 392                 num_ps_gprs = 192;
 393                 num_vs_gprs = 56;
 394                 num_temp_gprs = 4;
 395                 num_gs_gprs = 0;
 396                 num_es_gprs = 0;
 397                 num_ps_threads = 188;
 398                 num_vs_threads = 60;
 399                 num_gs_threads = 0;
 400                 num_es_threads = 0;
 401                 num_ps_stack_entries = 256;
 402                 num_vs_stack_entries = 256;
 403                 num_gs_stack_entries = 0;
 404                 num_es_stack_entries = 0;
 405                 break;
 406         case CHIP_RV730:
 407         case CHIP_RV740:
 408                 num_ps_gprs = 84;
 409                 num_vs_gprs = 36;
 410                 num_temp_gprs = 4;
 411                 num_gs_gprs = 0;
 412                 num_es_gprs = 0;
 413                 num_ps_threads = 188;
 414                 num_vs_threads = 60;
 415                 num_gs_threads = 0;
 416                 num_es_threads = 0;
 417                 num_ps_stack_entries = 128;
 418                 num_vs_stack_entries = 128;
 419                 num_gs_stack_entries = 0;
 420                 num_es_stack_entries = 0;
 421                 break;
 422         case CHIP_RV710:
 423                 num_ps_gprs = 192;
 424                 num_vs_gprs = 56;
 425                 num_temp_gprs = 4;
 426                 num_gs_gprs = 0;
 427                 num_es_gprs = 0;
 428                 num_ps_threads = 144;
 429                 num_vs_threads = 48;
 430                 num_gs_threads = 0;
 431                 num_es_threads = 0;
 432                 num_ps_stack_entries = 128;
 433                 num_vs_stack_entries = 128;
 434                 num_gs_stack_entries = 0;
 435                 num_es_stack_entries = 0;
 436                 break;
 437         }
 438
 439         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV610) ||
 440             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV620) ||
 441             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS780) ||
 442             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS880) ||
 443             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV710))
 444                 sq_config = 0;
 445         else
 446                 sq_config = R600_VC_ENABLE;
 447
 448         sq_config |= (R600_DX9_CONSTS |
 449                       R600_ALU_INST_PREFER_VECTOR |
 450                       R600_PS_PRIO(0) |
 451                       R600_VS_PRIO(1) |
 452                       R600_GS_PRIO(2) |
 453                       R600_ES_PRIO(3));
 454
 455         sq_gpr_resource_mgmt_1 = (R600_NUM_PS_GPRS(num_ps_gprs) |
 456                                   R600_NUM_VS_GPRS(num_vs_gprs) |
 457                                   R600_NUM_CLAUSE_TEMP_GPRS(num_temp_gprs));
 458         sq_gpr_resource_mgmt_2 = (R600_NUM_GS_GPRS(num_gs_gprs) |
 459                                   R600_NUM_ES_GPRS(num_es_gprs));
 460         sq_thread_resource_mgmt = (R600_NUM_PS_THREADS(num_ps_threads) |
 461                                    R600_NUM_VS_THREADS(num_vs_threads) |
 462                                    R600_NUM_GS_THREADS(num_gs_threads) |
 463                                    R600_NUM_ES_THREADS(num_es_threads));
 464         sq_stack_resource_mgmt_1 = (R600_NUM_PS_STACK_ENTRIES(num_ps_stack_entries) |
 465                                     R600_NUM_VS_STACK_ENTRIES(num_vs_stack_entries));
 466         sq_stack_resource_mgmt_2 = (R600_NUM_GS_STACK_ENTRIES(num_gs_stack_entries) |
 467                                     R600_NUM_ES_STACK_ENTRIES(num_es_stack_entries));
 468
 469         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV770) {
 470                 BEGIN_RING(r7xx_default_size + 10);
 471                 for (i = 0; i < r7xx_default_size; i++)
 472                         OUT_RING(r7xx_default_state[i]);
 473         } else {
 474                 BEGIN_RING(r6xx_default_size + 10);
 475                 for (i = 0; i < r6xx_default_size; i++)
 476                         OUT_RING(r6xx_default_state[i]);
 477         }
 478         OUT_RING(CP_PACKET3(R600_IT_EVENT_WRITE, 0));
 479         OUT_RING(R600_CACHE_FLUSH_AND_INV_EVENT);
 480         /* SQ config */
 481         OUT_RING(CP_PACKET3(R600_IT_SET_CONFIG_REG, 6));
 482         OUT_RING((R600_SQ_CONFIG - R600_SET_CONFIG_REG_OFFSET) >> 2);
 483         OUT_RING(sq_config);
 484         OUT_RING(sq_gpr_resource_mgmt_1);
 485         OUT_RING(sq_gpr_resource_mgmt_2);
 486         OUT_RING(sq_thread_resource_mgmt);
 487         OUT_RING(sq_stack_resource_mgmt_1);
 488         OUT_RING(sq_stack_resource_mgmt_2);
 489         ADVANCE_RING();
 490 }
 491
 492 static inline uint32_t i2f(uint32_t input)
 493 {
 494         u32 result, i, exponent, fraction;
 495
 496         if ((input & 0x3fff) == 0)
 497                 result = 0; /* 0 is a special case */
 498         else {
 499                 exponent = 140; /* exponent biased by 127; */
 500                 fraction = (input & 0x3fff) << 10; /* cheat and only
 501                                                       handle numbers below 2^^15 */
 502                 for (i = 0; i < 14; i++) {
 503                         if (fraction & 0x800000)
 504                                 break;
 505                         else {
 506                                 fraction = fraction << 1; /* keep
 507                                                              shifting left until top bit = 1 */
 508                                 exponent = exponent - 1;
 509                         }
 510                 }
 511                 result = exponent << 23 | (fraction & 0x7fffff); /* mask
 512                                                                     off top bit; assumed 1 */
 513         }
 514         return result;
 515 }
 516
 517
 518 static inline int r600_nomm_get_vb(struct drm_device *dev)
 519 {
 520         drm_radeon_private_t *dev_priv = dev->dev_private;
 521         dev_priv->blit_vb = radeon_freelist_get(dev);
 522         if (!dev_priv->blit_vb) {
 523                 DRM_ERROR("Unable to allocate vertex buffer for blit\n");
 524                 return -EAGAIN;
 525         }
 526         return 0;
 527 }
 528
 529 static inline void r600_nomm_put_vb(struct drm_device *dev)
 530 {
 531         drm_radeon_private_t *dev_priv = dev->dev_private;
 532
 533         dev_priv->blit_vb->used = 0;
 534         radeon_cp_discard_buffer(dev, dev_priv->blit_vb->file_priv->master, dev_priv->blit_vb);
 535 }
 536
 537 static inline void *r600_nomm_get_vb_ptr(struct drm_device *dev)
 538 {
 539         drm_radeon_private_t *dev_priv = dev->dev_private;
 540         return (((char *)dev->agp_buffer_map->handle +
 541                  dev_priv->blit_vb->offset + dev_priv->blit_vb->used));
 542 }
 543
 544 int
 545 r600_prepare_blit_copy(struct drm_device *dev, struct drm_file *file_priv)
 546 {
 547         drm_radeon_private_t *dev_priv = dev->dev_private;
 548         int ret;
 549         DRM_DEBUG("\n");
 550
 551         ret = r600_nomm_get_vb(dev);
 552         if (ret)
 553                 return ret;
 554
 555         dev_priv->blit_vb->file_priv = file_priv;
 556
 557         set_default_state(dev_priv);
 558         set_shaders(dev);
 559
 560         return 0;
 561 }
 562
 563
 564 void
 565 r600_done_blit_copy(struct drm_device *dev)
 566 {
 567         drm_radeon_private_t *dev_priv = dev->dev_private;
 568         RING_LOCALS;
 569         DRM_DEBUG("\n");
 570
 571         BEGIN_RING(5);
 572         OUT_RING(CP_PACKET3(R600_IT_EVENT_WRITE, 0));
 573         OUT_RING(R600_CACHE_FLUSH_AND_INV_EVENT);
 574         /* wait for 3D idle clean */
 575         OUT_RING(CP_PACKET3(R600_IT_SET_CONFIG_REG, 1));
 576         OUT_RING((R600_WAIT_UNTIL - R600_SET_CONFIG_REG_OFFSET) >> 2);
 577         OUT_RING(RADEON_WAIT_3D_IDLE | RADEON_WAIT_3D_IDLECLEAN);
 578
 579         ADVANCE_RING();
 580         COMMIT_RING();
 581
 582         r600_nomm_put_vb(dev);
 583 }
 584
 585 void
 586 r600_blit_copy(struct drm_device *dev,
 587                uint64_t src_gpu_addr, uint64_t dst_gpu_addr,
 588                int size_bytes)
 589 {
 590         drm_radeon_private_t *dev_priv = dev->dev_private;
 591         int max_bytes;
 592         u64 vb_addr;
 593         u32 *vb;
 594
 595         vb = r600_nomm_get_vb_ptr(dev);
 596
 597         if ((size_bytes & 3) || (src_gpu_addr & 3) || (dst_gpu_addr & 3)) {
 598                 max_bytes = 8192;
 599
 600                 while (size_bytes) {
 601                         int cur_size = size_bytes;
 602                         int src_x = src_gpu_addr & 255;
 603                         int dst_x = dst_gpu_addr & 255;
 604                         int h = 1;
 605                         src_gpu_addr = src_gpu_addr & ~255;
 606                         dst_gpu_addr = dst_gpu_addr & ~255;
 607
 608                         if (!src_x && !dst_x) {
 609                                 h = (cur_size / max_bytes);
 610                                 if (h > 8192)
 611                                         h = 8192;
 612                                 if (h == 0)
 613                                         h = 1;
 614                                 else
 615                                         cur_size = max_bytes;
 616                         } else {
 617                                 if (cur_size > max_bytes)
 618                                         cur_size = max_bytes;
 619                                 if (cur_size > (max_bytes - dst_x))
 620                                         cur_size = (max_bytes - dst_x);
 621                                 if (cur_size > (max_bytes - src_x))
 622                                         cur_size = (max_bytes - src_x);
 623                         }
 624
 625                         if ((dev_priv->blit_vb->used + 48) > dev_priv->blit_vb->total) {
 626
 627                                 r600_nomm_put_vb(dev);
 628                                 r600_nomm_get_vb(dev);
 629                                 if (!dev_priv->blit_vb)
 630                                         return;
 631                                 set_shaders(dev);
 632                                 vb = r600_nomm_get_vb_ptr(dev);
 633                         }
 634
 635                         vb[0] = i2f(dst_x);
 636                         vb[1] = 0;
 637                         vb[2] = i2f(src_x);
 638                         vb[3] = 0;
 639
 640                         vb[4] = i2f(dst_x);
 641                         vb[5] = i2f(h);
 642                         vb[6] = i2f(src_x);
 643                         vb[7] = i2f(h);
 644
 645                         vb[8] = i2f(dst_x + cur_size);
 646                         vb[9] = i2f(h);
 647                         vb[10] = i2f(src_x + cur_size);
 648                         vb[11] = i2f(h);
 649
 650                         /* src */
 651                         set_tex_resource(dev_priv, FMT_8,
 652                                          src_x + cur_size, h, src_x + cur_size,
 653                                          src_gpu_addr);
 654
 655                         cp_set_surface_sync(dev_priv,
 656                                             R600_TC_ACTION_ENA, (src_x + cur_size * h), src_gpu_addr);
 657
 658                         /* dst */
 659                         set_render_target(dev_priv, COLOR_8,
 660                                           dst_x + cur_size, h,
 661                                           dst_gpu_addr);
 662
 663                         /* scissors */
 664                         set_scissors(dev_priv, dst_x, 0, dst_x + cur_size, h);
 665
 666                         /* Vertex buffer setup */
 667                         vb_addr = dev_priv->gart_buffers_offset +
 668                                 dev_priv->blit_vb->offset +
 669                                 dev_priv->blit_vb->used;
 670                         set_vtx_resource(dev_priv, vb_addr);
 671
 672                         /* draw */
 673                         draw_auto(dev_priv);
 674
 675                         cp_set_surface_sync(dev_priv,
 676                                             R600_CB_ACTION_ENA | R600_CB0_DEST_BASE_ENA,
 677                                             cur_size * h, dst_gpu_addr);
 678
 679                         vb += 12;
 680                         dev_priv->blit_vb->used += 12 * 4;
 681
 682                         src_gpu_addr += cur_size * h;
 683                         dst_gpu_addr += cur_size * h;
 684                         size_bytes -= cur_size * h;
 685                 }
 686         } else {
 687                 max_bytes = 8192 * 4;
 688
 689                 while (size_bytes) {
 690                         int cur_size = size_bytes;
 691                         int src_x = (src_gpu_addr & 255);
 692                         int dst_x = (dst_gpu_addr & 255);
 693                         int h = 1;
 694                         src_gpu_addr = src_gpu_addr & ~255;
 695                         dst_gpu_addr = dst_gpu_addr & ~255;
 696
 697                         if (!src_x && !dst_x) {
 698                                 h = (cur_size / max_bytes);
 699                                 if (h > 8192)
 700                                         h = 8192;
 701                                 if (h == 0)
 702                                         h = 1;
 703                                 else
 704                                         cur_size = max_bytes;
 705                         } else {
 706                                 if (cur_size > max_bytes)
 707                                         cur_size = max_bytes;
 708                                 if (cur_size > (max_bytes - dst_x))
 709                                         cur_size = (max_bytes - dst_x);
 710                                 if (cur_size > (max_bytes - src_x))
 711                                         cur_size = (max_bytes - src_x);
 712                         }
 713
 714                         if ((dev_priv->blit_vb->used + 48) > dev_priv->blit_vb->total) {
 715                                 r600_nomm_put_vb(dev);
 716                                 r600_nomm_get_vb(dev);
 717                                 if (!dev_priv->blit_vb)
 718                                         return;
 719
 720                                 set_shaders(dev);
 721                                 vb = r600_nomm_get_vb_ptr(dev);
 722                         }
 723
 724                         vb[0] = i2f(dst_x / 4);
 725                         vb[1] = 0;
 726                         vb[2] = i2f(src_x / 4);
 727                         vb[3] = 0;
 728
 729                         vb[4] = i2f(dst_x / 4);
 730                         vb[5] = i2f(h);
 731                         vb[6] = i2f(src_x / 4);
 732                         vb[7] = i2f(h);
 733
 734                         vb[8] = i2f((dst_x + cur_size) / 4);
 735                         vb[9] = i2f(h);
 736                         vb[10] = i2f((src_x + cur_size) / 4);
 737                         vb[11] = i2f(h);
 738
 739                         /* src */
 740                         set_tex_resource(dev_priv, FMT_8_8_8_8,
 741                                          (src_x + cur_size) / 4,
 742                                          h, (src_x + cur_size) / 4,
 743                                          src_gpu_addr);
 744
 745                         cp_set_surface_sync(dev_priv,
 746                                             R600_TC_ACTION_ENA, (src_x + cur_size * h), src_gpu_addr);
 747
 748                         /* dst */
 749                         set_render_target(dev_priv, COLOR_8_8_8_8,
 750                                           (dst_x + cur_size) / 4, h,
 751                                           dst_gpu_addr);
 752
 753                         /* scissors */
 754                         set_scissors(dev_priv, (dst_x / 4), 0, (dst_x + cur_size / 4), h);
 755
 756                         /* Vertex buffer setup */
 757                         vb_addr = dev_priv->gart_buffers_offset +
 758                                 dev_priv->blit_vb->offset +
 759                                 dev_priv->blit_vb->used;
 760                         set_vtx_resource(dev_priv, vb_addr);
 761
 762                         /* draw */
 763                         draw_auto(dev_priv);
 764
 765                         cp_set_surface_sync(dev_priv,
 766                                             R600_CB_ACTION_ENA | R600_CB0_DEST_BASE_ENA,
 767                                             cur_size * h, dst_gpu_addr);
 768
 769                         vb += 12;
 770                         dev_priv->blit_vb->used += 12 * 4;
 771
 772                         src_gpu_addr += cur_size * h;
 773                         dst_gpu_addr += cur_size * h;
 774                         size_bytes -= cur_size * h;
 775                 }
 776         }
 777 }
 778
 779 void
 780 r600_blit_swap(struct drm_device *dev,
 781                uint64_t src_gpu_addr, uint64_t dst_gpu_addr,
 782                int sx, int sy, int dx, int dy,
 783                int w, int h, int src_pitch, int dst_pitch, int cpp)
 784 {
 785         drm_radeon_private_t *dev_priv = dev->dev_private;
 786         int cb_format, tex_format;
 787         int sx2, sy2, dx2, dy2;
 788         u64 vb_addr;
 789         u32 *vb;
 790
 791         if ((dev_priv->blit_vb->used + 48) > dev_priv->blit_vb->total) {
 792
 793                 r600_nomm_put_vb(dev);
 794                 r600_nomm_get_vb(dev);
 795                 if (!dev_priv->blit_vb)
 796                         return;
 797
 798                 set_shaders(dev);
 799         }
 800         vb = r600_nomm_get_vb_ptr(dev);
 801
 802         sx2 = sx + w;
 803         sy2 = sy + h;
 804         dx2 = dx + w;
 805         dy2 = dy + h;
 806
 807         vb[0] = i2f(dx);
 808         vb[1] = i2f(dy);
 809         vb[2] = i2f(sx);
 810         vb[3] = i2f(sy);
 811
 812         vb[4] = i2f(dx);
 813         vb[5] = i2f(dy2);
 814         vb[6] = i2f(sx);
 815         vb[7] = i2f(sy2);
 816
 817         vb[8] = i2f(dx2);
 818         vb[9] = i2f(dy2);
 819         vb[10] = i2f(sx2);
 820         vb[11] = i2f(sy2);
 821
 822         switch(cpp) {
 823         case 4:
 824                 cb_format = COLOR_8_8_8_8;
 825                 tex_format = FMT_8_8_8_8;
 826                 break;
 827         case 2:
 828                 cb_format = COLOR_5_6_5;
 829                 tex_format = FMT_5_6_5;
 830                 break;
 831         default:
 832                 cb_format = COLOR_8;
 833                 tex_format = FMT_8;
 834                 break;
 835         }
 836
 837         /* src */
 838         set_tex_resource(dev_priv, tex_format,
 839                          src_pitch / cpp,
 840                          sy2, src_pitch / cpp,
 841                          src_gpu_addr);
 842
 843         cp_set_surface_sync(dev_priv,
 844                             R600_TC_ACTION_ENA, src_pitch * sy2, src_gpu_addr);
 845
 846         /* dst */
 847         set_render_target(dev_priv, cb_format,
 848                           dst_pitch / cpp, dy2,
 849                           dst_gpu_addr);
 850
 851         /* scissors */
 852         set_scissors(dev_priv, dx, dy, dx2, dy2);
 853
 854         /* Vertex buffer setup */
 855         vb_addr = dev_priv->gart_buffers_offset +
 856                 dev_priv->blit_vb->offset +
 857                 dev_priv->blit_vb->used;
 858         set_vtx_resource(dev_priv, vb_addr);
 859
 860         /* draw */
 861         draw_auto(dev_priv);
 862
 863         cp_set_surface_sync(dev_priv,
 864                             R600_CB_ACTION_ENA | R600_CB0_DEST_BASE_ENA,
 865                             dst_pitch * dy2, dst_gpu_addr);
 866
 867         dev_priv->blit_vb->used += 12 * 4;
 868 }