src/gallium/drivers/r600/evergreen_compute_internal.c

   1 /*
   2  * Permission is hereby granted, free of charge, to any person obtaining a
   3  * copy of this software and associated documentation files (the "Software"),
   4  * to deal in the Software without restriction, including without limitation
   5  * on the rights to use, copy, modify, merge, publish, distribute, sub
   6  * license, and/or sell copies of the Software, and to permit persons to whom
   7  * the Software is furnished to do so, subject to the following conditions:
   8  *
   9  * The above copyright notice and this permission notice (including the next
  10  * paragraph) shall be included in all copies or substantial portions of the
  11  * Software.
  12  *
  13  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  14  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  15  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  16  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
  17  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  18  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  19  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  20  *
  21  * Authors:
  22  *      Adam Rak <adam.rak@streamnovation.com>
  23  */
  24
  25 #include <stdlib.h>
  26 #include <stdio.h>
  27
  28 #include "pipe/p_defines.h"
  29 #include "pipe/p_state.h"
  30 #include "pipe/p_context.h"
  31 #include "util/u_blitter.h"
  32 #include "util/u_double_list.h"
  33 #include "util/u_transfer.h"
  34 #include "util/u_surface.h"
  35 #include "util/u_pack_color.h"
  36 #include "util/u_memory.h"
  37 #include "util/u_inlines.h"
  38 #include "util/u_framebuffer.h"
  39 #include "r600.h"
  40 #include "r600_resource.h"
  41 #include "r600_shader.h"
  42 #include "r600_pipe.h"
  43 #include "r600_formats.h"
  44 #include "evergreend.h"
  45 #include "evergreen_compute_internal.h"
  46 #include "r600_hw_context_priv.h"
  47
  48 int get_compute_resource_num(void)
  49 {
  50         int num = 0;
  51 #define DECL_COMPUTE_RESOURCE(name, n) num += n;
  52 #include "compute_resource.def"
  53 #undef DECL_COMPUTE_RESOURCE
  54         return num;
  55 }
  56
  57 void evergreen_emit_raw_value(
  58         struct evergreen_compute_resource* res,
  59         unsigned value)
  60 {
  61         res->cs[res->cs_end++] = value;
  62 }
  63
  64 void evergreen_emit_ctx_value(struct r600_context *ctx, unsigned value)
  65 {
  66         ctx->cs->buf[ctx->cs->cdw++] = value;
  67 }
  68
  69 void evergreen_mult_reg_set_(
  70         struct evergreen_compute_resource* res,
  71         int index,
  72         u32* array,
  73         int size)
  74 {
  75         int i = 0;
  76
  77         evergreen_emit_raw_reg_set(res, index, size / 4);
  78
  79         for (i = 0; i < size; i+=4) {
  80                 res->cs[res->cs_end++] = array[i / 4];
  81         }
  82 }
  83
  84 void evergreen_reg_set(
  85         struct evergreen_compute_resource* res,
  86         unsigned index,
  87         unsigned value)
  88 {
  89         evergreen_emit_raw_reg_set(res, index, 1);
  90         res->cs[res->cs_end++] = value;
  91 }
  92
  93 struct evergreen_compute_resource* get_empty_res(
  94         struct r600_pipe_compute* pipe,
  95         enum evergreen_compute_resources res_code,
  96         int offset_index)
  97 {
  98         int code_index = -1;
  99         int code_size = -1;
 100
 101         {
 102                 int i = 0;
 103                 #define DECL_COMPUTE_RESOURCE(name, n) if (COMPUTE_RESOURCE_ ## name    == res_code) {code_index = i; code_size = n;} i += n;
 104                 #include "compute_resource.def"
 105                 #undef DECL_COMPUTE_RESOURCE
 106         }
 107
 108         assert(code_index != -1 && "internal error: resouce index not found");
 109         assert(offset_index < code_size && "internal error: overindexing resource");
 110
 111         int index = code_index + offset_index;
 112
 113         struct evergreen_compute_resource* res = &pipe->resources[index];
 114
 115         res->enabled = true;
 116         res->bo = NULL;
 117         res->cs_end = 0;
 118         bzero(&res->do_reloc, sizeof(res->do_reloc));
 119
 120         return res;
 121 }
 122
 123 void evergreen_emit_raw_reg_set(
 124         struct evergreen_compute_resource* res,
 125         unsigned index,
 126         int num)
 127 {
 128         res->enabled = 1;
 129         int cs_end = res->cs_end;
 130
 131         if (index >= EVERGREEN_CONFIG_REG_OFFSET
 132                         && index < EVERGREEN_CONFIG_REG_END) {
 133                 res->cs[cs_end] = PKT3C(PKT3_SET_CONFIG_REG, num, 0);
 134                 res->cs[cs_end+1] = (index - EVERGREEN_CONFIG_REG_OFFSET) >> 2;
 135         } else if (index >= EVERGREEN_CONTEXT_REG_OFFSET
 136                         && index < EVERGREEN_CONTEXT_REG_END) {
 137                 res->cs[cs_end] = PKT3C(PKT3_SET_CONTEXT_REG, num, 0);
 138                 res->cs[cs_end+1] = (index - EVERGREEN_CONTEXT_REG_OFFSET) >> 2;
 139         } else if (index >= EVERGREEN_RESOURCE_OFFSET
 140                         && index < EVERGREEN_RESOURCE_END) {
 141                 res->cs[cs_end] = PKT3C(PKT3_SET_RESOURCE, num, 0);
 142                 res->cs[cs_end+1] = (index - EVERGREEN_RESOURCE_OFFSET) >> 2;
 143         } else if (index >= EVERGREEN_SAMPLER_OFFSET
 144                         && index < EVERGREEN_SAMPLER_END) {
 145                 res->cs[cs_end] = PKT3C(PKT3_SET_SAMPLER, num, 0);
 146                 res->cs[cs_end+1] = (index - EVERGREEN_SAMPLER_OFFSET) >> 2;
 147         } else if (index >= EVERGREEN_CTL_CONST_OFFSET
 148                         && index < EVERGREEN_CTL_CONST_END) {
 149                 res->cs[cs_end] = PKT3C(PKT3_SET_CTL_CONST, num, 0);
 150                 res->cs[cs_end+1] = (index - EVERGREEN_CTL_CONST_OFFSET) >> 2;
 151         } else if (index >= EVERGREEN_LOOP_CONST_OFFSET
 152                         && index < EVERGREEN_LOOP_CONST_END) {
 153                 res->cs[cs_end] = PKT3C(PKT3_SET_LOOP_CONST, num, 0);
 154                 res->cs[cs_end+1] = (index - EVERGREEN_LOOP_CONST_OFFSET) >> 2;
 155         } else if (index >= EVERGREEN_BOOL_CONST_OFFSET
 156                         && index < EVERGREEN_BOOL_CONST_END) {
 157                 res->cs[cs_end] = PKT3C(PKT3_SET_BOOL_CONST, num, 0);
 158                 res->cs[cs_end+1] = (index - EVERGREEN_BOOL_CONST_OFFSET) >> 2;
 159         } else {
 160                 res->cs[cs_end] = PKT0(index, num-1);
 161                 res->cs_end--;
 162         }
 163
 164         res->cs_end += 2;
 165 }
 166
 167 void evergreen_emit_force_reloc(struct evergreen_compute_resource* res)
 168 {
 169         res->do_reloc[res->cs_end] += 1;
 170 }
 171
 172 void evergreen_emit_ctx_reg_set(
 173         struct r600_context *ctx,
 174         unsigned index,
 175         int num)
 176 {
 177
 178         if (index >= EVERGREEN_CONFIG_REG_OFFSET
 179                         && index < EVERGREEN_CONFIG_REG_END) {
 180                 ctx->cs->buf[ctx->cs->cdw++] = PKT3C(PKT3_SET_CONFIG_REG, num, 0);
 181                 ctx->cs->buf[ctx->cs->cdw++] = (index - EVERGREEN_CONFIG_REG_OFFSET) >> 2;
 182         } else if (index >= EVERGREEN_CONTEXT_REG_OFFSET
 183                         && index < EVERGREEN_CONTEXT_REG_END) {
 184                 ctx->cs->buf[ctx->cs->cdw++] = PKT3C(PKT3_SET_CONTEXT_REG, num, 0);
 185                 ctx->cs->buf[ctx->cs->cdw++] = (index - EVERGREEN_CONTEXT_REG_OFFSET) >> 2;
 186         } else if (index >= EVERGREEN_RESOURCE_OFFSET
 187                         && index < EVERGREEN_RESOURCE_END) {
 188                 ctx->cs->buf[ctx->cs->cdw++] = PKT3C(PKT3_SET_RESOURCE, num, 0);
 189                 ctx->cs->buf[ctx->cs->cdw++] = (index - EVERGREEN_RESOURCE_OFFSET) >> 2;
 190         } else if (index >= EVERGREEN_SAMPLER_OFFSET
 191                         && index < EVERGREEN_SAMPLER_END) {
 192                 ctx->cs->buf[ctx->cs->cdw++] = PKT3C(PKT3_SET_SAMPLER, num, 0);
 193                 ctx->cs->buf[ctx->cs->cdw++] = (index - EVERGREEN_SAMPLER_OFFSET) >> 2;
 194         } else if (index >= EVERGREEN_CTL_CONST_OFFSET
 195                         && index < EVERGREEN_CTL_CONST_END) {
 196                 ctx->cs->buf[ctx->cs->cdw++] = PKT3C(PKT3_SET_CTL_CONST, num, 0);
 197                 ctx->cs->buf[ctx->cs->cdw++] = (index - EVERGREEN_CTL_CONST_OFFSET) >> 2;
 198         } else if (index >= EVERGREEN_LOOP_CONST_OFFSET
 199                         && index < EVERGREEN_LOOP_CONST_END) {
 200                 ctx->cs->buf[ctx->cs->cdw++] = PKT3C(PKT3_SET_LOOP_CONST, num, 0);
 201                 ctx->cs->buf[ctx->cs->cdw++] = (index - EVERGREEN_LOOP_CONST_OFFSET) >> 2;
 202         } else if (index >= EVERGREEN_BOOL_CONST_OFFSET
 203                         && index < EVERGREEN_BOOL_CONST_END) {
 204                 ctx->cs->buf[ctx->cs->cdw++] = PKT3C(PKT3_SET_BOOL_CONST, num, 0);
 205                 ctx->cs->buf[ctx->cs->cdw++] = (index - EVERGREEN_BOOL_CONST_OFFSET) >> 2;
 206         } else {
 207                 ctx->cs->buf[ctx->cs->cdw++] = PKT0(index, num-1);
 208         }
 209 }
 210
 211 void evergreen_emit_ctx_reloc(
 212         struct r600_context *ctx,
 213         struct r600_resource *bo,
 214         enum radeon_bo_usage usage)
 215 {
 216         assert(bo);
 217
 218         ctx->cs->buf[ctx->cs->cdw++] = PKT3(PKT3_NOP, 0, 0);
 219         u32 rr = r600_context_bo_reloc(ctx, bo, usage);
 220         ctx->cs->buf[ctx->cs->cdw++] = rr;
 221 }
 222
 223 int evergreen_compute_get_gpu_format(
 224         struct number_type_and_format* fmt,
 225         struct r600_resource *bo)
 226 {
 227         switch (bo->b.b.format)
 228         {
 229                 case PIPE_FORMAT_R8_UNORM:
 230                 case PIPE_FORMAT_R32_UNORM:
 231                 case PIPE_FORMAT_R32_UINT:
 232                         fmt->format = V_028C70_COLOR_32;
 233                         fmt->number_type = V_028C70_NUMBER_UNORM;
 234                         fmt->num_format_all = 0;
 235                 break;
 236                 case PIPE_FORMAT_R32_FLOAT:
 237                         fmt->format = V_028C70_COLOR_32_FLOAT;
 238                         fmt->number_type = V_028C70_NUMBER_FLOAT;
 239                         fmt->num_format_all = 0;
 240                 break;
 241                 case PIPE_FORMAT_R32G32B32A32_FLOAT:
 242                         fmt->format = V_028C70_COLOR_32_32_32_32_FLOAT;
 243                         fmt->number_type = V_028C70_NUMBER_FLOAT;
 244                         fmt->num_format_all = 0;
 245                 break;
 246
 247                 ///TODO: other formats...
 248
 249                 default:
 250                         return 0;
 251         }
 252
 253         return 1;
 254 }
 255
 256 void evergreen_set_rat(
 257         struct r600_pipe_compute *pipe,
 258         int id,
 259         struct r600_resource* bo,
 260         int start,
 261         int size)
 262 {
 263         assert(id < 12);
 264         assert((size & 3) == 0);
 265         assert((start & 0xFF) == 0);
 266
 267         struct pipe_surface rat_templ;
 268         struct r600_surface *surf;
 269         struct r600_context *rctx = pipe->ctx;
 270
 271         COMPUTE_DBG("bind rat: %i \n", id);
 272
 273         /* Create the RAT surface */
 274         memset(&rat_templ, 0, sizeof(rat_templ));
 275         rat_templ.usage = RADEON_USAGE_READWRITE;
 276         rat_templ.format = PIPE_FORMAT_R32_UINT;
 277         rat_templ.u.tex.level = 0;
 278         rat_templ.u.tex.first_layer = 0;
 279         rat_templ.u.tex.last_layer = 0;
 280
 281         /* Add the RAT the list of color buffers */
 282         pipe->ctx->framebuffer.state.cbufs[id] = pipe->ctx->context.create_surface(
 283                 (struct pipe_context *)pipe->ctx,
 284                 (struct pipe_resource *)bo, &rat_templ);
 285
 286         /* Update the number of color buffers */
 287         pipe->ctx->framebuffer.state.nr_cbufs =
 288                 MAX2(id + 1, pipe->ctx->framebuffer.state.nr_cbufs);
 289
 290         /* Update the cb_target_mask
 291          * XXX: I think this is a potential spot for bugs once we start doing
 292          * GL interop.  cb_target_mask may be modified in the 3D sections
 293          * of this driver. */
 294         pipe->ctx->compute_cb_target_mask |= (0xf << (id * 4));
 295
 296         surf = (struct r600_surface*)pipe->ctx->framebuffer.state.cbufs[id];
 297         evergreen_init_color_surface_rat(rctx, surf);
 298 }
 299
 300 void evergreen_set_gds(
 301         struct r600_pipe_compute *pipe,
 302         uint32_t addr,
 303         uint32_t size)
 304 {
 305         struct evergreen_compute_resource* res =
 306                 get_empty_res(pipe, COMPUTE_RESOURCE_GDS, 0);
 307
 308         evergreen_reg_set(res, R_028728_GDS_ORDERED_WAVE_PER_SE, 1);
 309         evergreen_reg_set(res, R_028720_GDS_ADDR_BASE, addr);
 310         evergreen_reg_set(res, R_028724_GDS_ADDR_SIZE, size);
 311 }
 312
 313 void evergreen_set_export(
 314         struct r600_pipe_compute *pipe,
 315         struct r600_resource* bo,
 316         int offset, int size)
 317 {
 318         #define SX_MEMORY_EXPORT_BASE 0x9010
 319         #define SX_MEMORY_EXPORT_SIZE 0x9014
 320
 321         struct evergreen_compute_resource* res =
 322                 get_empty_res(pipe, COMPUTE_RESOURCE_EXPORT, 0);
 323
 324         evergreen_reg_set(res, SX_MEMORY_EXPORT_SIZE, size);
 325
 326         if (size) {
 327                 evergreen_reg_set(res, SX_MEMORY_EXPORT_BASE, offset);
 328                 res->bo = bo;
 329                 res->usage = RADEON_USAGE_WRITE;
 330                 res->coher_bo_size = size;
 331                 res->flags = 0;
 332         }
 333 }
 334
 335 void evergreen_set_loop_const(
 336         struct r600_pipe_compute *pipe,
 337         int id, int count, int init, int inc) {
 338
 339         struct evergreen_compute_resource* res =
 340                 get_empty_res(pipe, COMPUTE_RESOURCE_LOOP, id);
 341
 342         assert(id < 32);
 343         assert(count <= 0xFFF);
 344         assert(init <= 0xFF);
 345         assert(inc <= 0xFF);
 346
 347         /* Compute shaders use LOOP_CONST registers SQ_LOOP_CONST_160 to
 348          * SQ_LOOP_CONST_191 */
 349         evergreen_reg_set(res, R_03A200_SQ_LOOP_CONST_0 + (160 * 4) + (id * 4),
 350                 count | init << 12 | inc << 24);
 351 }
 352
 353 void evergreen_set_tmp_ring(
 354         struct r600_pipe_compute *pipe,
 355         struct r600_resource* bo,
 356         int offset, int size, int se)
 357 {
 358         #define SQ_LSTMP_RING_BASE 0x00008e10
 359         #define SQ_LSTMP_RING_SIZE 0x00008e14
 360         #define GRBM_GFX_INDEX                                  0x802C
 361         #define         INSTANCE_INDEX(x)                       ((x) << 0)
 362         #define         SE_INDEX(x)                             ((x) << 16)
 363         #define         INSTANCE_BROADCAST_WRITES               (1 << 30)
 364         #define         SE_BROADCAST_WRITES                     (1 << 31)
 365
 366         struct evergreen_compute_resource* res =
 367                 get_empty_res(pipe, COMPUTE_RESOURCE_TMPRING, se);
 368
 369         evergreen_reg_set(res,
 370                 GRBM_GFX_INDEX,INSTANCE_INDEX(0)
 371                 | SE_INDEX(se)
 372                 | INSTANCE_BROADCAST_WRITES);
 373         evergreen_reg_set(res, SQ_LSTMP_RING_SIZE, size);
 374
 375         if (size) {
 376                 assert(bo);
 377
 378                 evergreen_reg_set(res, SQ_LSTMP_RING_BASE, offset);
 379                 res->bo = bo;
 380                 res->usage = RADEON_USAGE_WRITE;
 381                 res->coher_bo_size = 0;
 382                 res->flags = 0;
 383         }
 384
 385         if (size) {
 386                 evergreen_emit_force_reloc(res);
 387         }
 388
 389         evergreen_reg_set(res,
 390                 GRBM_GFX_INDEX,INSTANCE_INDEX(0)
 391                 | SE_INDEX(0)
 392                 | INSTANCE_BROADCAST_WRITES
 393                 | SE_BROADCAST_WRITES);
 394 }
 395
 396 static uint32_t r600_colorformat_endian_swap(uint32_t colorformat)
 397 {
 398         if (R600_BIG_ENDIAN) {
 399                 switch(colorformat) {
 400                 case V_028C70_COLOR_4_4:
 401                         return ENDIAN_NONE;
 402
 403                 /* 8-bit buffers. */
 404                 case V_028C70_COLOR_8:
 405                         return ENDIAN_NONE;
 406
 407                 /* 16-bit buffers. */
 408                 case V_028C70_COLOR_5_6_5:
 409                 case V_028C70_COLOR_1_5_5_5:
 410                 case V_028C70_COLOR_4_4_4_4:
 411                 case V_028C70_COLOR_16:
 412                 case V_028C70_COLOR_8_8:
 413                         return ENDIAN_8IN16;
 414
 415                 /* 32-bit buffers. */
 416                 case V_028C70_COLOR_8_8_8_8:
 417                 case V_028C70_COLOR_2_10_10_10:
 418                 case V_028C70_COLOR_8_24:
 419                 case V_028C70_COLOR_24_8:
 420                 case V_028C70_COLOR_32_FLOAT:
 421                 case V_028C70_COLOR_16_16_FLOAT:
 422                 case V_028C70_COLOR_16_16:
 423                         return ENDIAN_8IN32;
 424
 425                 /* 64-bit buffers. */
 426                 case V_028C70_COLOR_16_16_16_16:
 427                 case V_028C70_COLOR_16_16_16_16_FLOAT:
 428                         return ENDIAN_8IN16;
 429
 430                 case V_028C70_COLOR_32_32_FLOAT:
 431                 case V_028C70_COLOR_32_32:
 432                 case V_028C70_COLOR_X24_8_32_FLOAT:
 433                         return ENDIAN_8IN32;
 434
 435                 /* 96-bit buffers. */
 436                 case V_028C70_COLOR_32_32_32_FLOAT:
 437                 /* 128-bit buffers. */
 438                 case V_028C70_COLOR_32_32_32_32_FLOAT:
 439                 case V_028C70_COLOR_32_32_32_32:
 440                         return ENDIAN_8IN32;
 441                 default:
 442                         return ENDIAN_NONE; /* Unsupported. */
 443                 }
 444         } else {
 445                 return ENDIAN_NONE;
 446         }
 447 }
 448
 449 static unsigned r600_tex_dim(unsigned dim)
 450 {
 451         switch (dim) {
 452         default:
 453         case PIPE_TEXTURE_1D:
 454                 return V_030000_SQ_TEX_DIM_1D;
 455         case PIPE_TEXTURE_1D_ARRAY:
 456                 return V_030000_SQ_TEX_DIM_1D_ARRAY;
 457         case PIPE_TEXTURE_2D:
 458         case PIPE_TEXTURE_RECT:
 459                 return V_030000_SQ_TEX_DIM_2D;
 460         case PIPE_TEXTURE_2D_ARRAY:
 461                 return V_030000_SQ_TEX_DIM_2D_ARRAY;
 462         case PIPE_TEXTURE_3D:
 463                 return V_030000_SQ_TEX_DIM_3D;
 464         case PIPE_TEXTURE_CUBE:
 465                 return V_030000_SQ_TEX_DIM_CUBEMAP;
 466         }
 467 }
 468
 469 void evergreen_set_tex_resource(
 470         struct r600_pipe_compute *pipe,
 471         struct r600_pipe_sampler_view* view,
 472         int id)
 473 {
 474         struct evergreen_compute_resource* res =
 475                 get_empty_res(pipe, COMPUTE_RESOURCE_TEX, id);
 476         struct r600_texture *tmp =
 477                 (struct r600_texture*)view->base.texture;
 478
 479         unsigned format, endian;
 480         uint32_t word4 = 0, yuv_format = 0, pitch = 0;
 481         unsigned char swizzle[4], array_mode = 0, non_disp_tiling = 0;
 482         unsigned height, depth;
 483
 484         swizzle[0] = 0;
 485         swizzle[1] = 1;
 486         swizzle[2] = 2;
 487         swizzle[3] = 3;
 488
 489         format = r600_translate_texformat((struct pipe_screen *)pipe->ctx->screen,
 490                 view->base.format, swizzle, &word4, &yuv_format);
 491
 492         if (format == ~0) {
 493                 format = 0;
 494         }
 495
 496         endian = r600_colorformat_endian_swap(format);
 497
 498         height = view->base.texture->height0;
 499         depth = view->base.texture->depth0;
 500
 501         pitch = align(tmp->surface.level[0].nblk_x *
 502                 util_format_get_blockwidth(tmp->resource.b.b.format), 8);
 503         array_mode = tmp->array_mode[0];
 504         non_disp_tiling = tmp->non_disp_tiling;
 505
 506         assert(view->base.texture->target != PIPE_TEXTURE_1D_ARRAY);
 507         assert(view->base.texture->target != PIPE_TEXTURE_2D_ARRAY);
 508
 509         evergreen_emit_raw_value(res, PKT3C(PKT3_SET_RESOURCE, 8, 0));
 510         evergreen_emit_raw_value(res, (id+816)*32 >> 2); ///TODO: check this line
 511         evergreen_emit_raw_value(res,
 512                                 (S_030000_DIM(r600_tex_dim(view->base.texture->target)) |
 513                                 S_030000_PITCH((pitch / 8) - 1) |
 514                                 S_030000_NON_DISP_TILING_ORDER(non_disp_tiling) |
 515                                 S_030000_TEX_WIDTH(view->base.texture->width0 - 1)));
 516         evergreen_emit_raw_value(res, (S_030004_TEX_HEIGHT(height - 1) |
 517                                 S_030004_TEX_DEPTH(depth - 1) |
 518                                 S_030004_ARRAY_MODE(array_mode)));
 519         evergreen_emit_raw_value(res, tmp->surface.level[0].offset >> 8);
 520         evergreen_emit_raw_value(res, tmp->surface.level[0].offset >> 8);
 521         evergreen_emit_raw_value(res, (word4 |
 522                                 S_030010_SRF_MODE_ALL(V_030010_SRF_MODE_ZERO_CLAMP_MINUS_ONE) |
 523                                 S_030010_ENDIAN_SWAP(endian) |
 524                                 S_030010_BASE_LEVEL(0)));
 525         evergreen_emit_raw_value(res, (S_030014_LAST_LEVEL(0) |
 526                                 S_030014_BASE_ARRAY(0) |
 527                                 S_030014_LAST_ARRAY(0)));
 528         evergreen_emit_raw_value(res, (S_030018_MAX_ANISO(4 /* max 16 samples */)));
 529         evergreen_emit_raw_value(res,
 530                 S_03001C_TYPE(V_03001C_SQ_TEX_VTX_VALID_TEXTURE)
 531                 | S_03001C_DATA_FORMAT(format));
 532
 533         res->bo = (struct r600_resource*)view->base.texture;
 534
 535         res->usage = RADEON_USAGE_READ;
 536
 537         res->coher_bo_size = tmp->surface.level[0].offset +
 538                              util_format_get_blockwidth(tmp->resource.b.b.format) *
 539                              view->base.texture->width0*height*depth;
 540
 541         pipe->ctx->flags |= R600_CONTEXT_TEX_FLUSH;
 542
 543         evergreen_emit_force_reloc(res);
 544         evergreen_emit_force_reloc(res);
 545 }
 546
 547 void evergreen_set_sampler_resource(
 548         struct r600_pipe_compute *pipe,
 549         struct compute_sampler_state *sampler,
 550         int id)
 551 {
 552         struct evergreen_compute_resource* res =
 553                 get_empty_res(pipe, COMPUTE_RESOURCE_SAMPLER, id);
 554
 555         unsigned aniso_flag_offset = sampler->state.max_anisotropy > 1 ? 2 : 0;
 556
 557         evergreen_emit_raw_value(res, PKT3C(PKT3_SET_SAMPLER, 3, 0));
 558         evergreen_emit_raw_value(res, (id + 90)*3);
 559         evergreen_emit_raw_value(res,
 560                 S_03C000_CLAMP_X(r600_tex_wrap(sampler->state.wrap_s)) |
 561                 S_03C000_CLAMP_Y(r600_tex_wrap(sampler->state.wrap_t)) |
 562                 S_03C000_CLAMP_Z(r600_tex_wrap(sampler->state.wrap_r)) |
 563                 S_03C000_XY_MAG_FILTER(r600_tex_filter(sampler->state.mag_img_filter) | aniso_flag_offset) |
 564                 S_03C000_XY_MIN_FILTER(r600_tex_filter(sampler->state.min_img_filter) | aniso_flag_offset) |
 565                 S_03C000_BORDER_COLOR_TYPE(V_03C000_SQ_TEX_BORDER_COLOR_OPAQUE_BLACK)
 566         );
 567         evergreen_emit_raw_value(res,
 568                 S_03C004_MIN_LOD(S_FIXED(CLAMP(sampler->state.min_lod, 0, 15), 8)) |
 569                 S_03C004_MAX_LOD(S_FIXED(CLAMP(sampler->state.max_lod, 0, 15), 8))
 570         );
 571         evergreen_emit_raw_value(res,
 572                 S_03C008_LOD_BIAS(S_FIXED(CLAMP(sampler->state.lod_bias, -16, 16), 8)) |
 573                 (sampler->state.seamless_cube_map ? 0 : S_03C008_DISABLE_CUBE_WRAP(1)) |
 574                 S_03C008_TYPE(1)
 575         );
 576 }
 577
 578 void evergreen_set_const_cache(
 579         struct r600_pipe_compute *pipe,
 580         int cache_id,
 581         struct r600_resource* cbo,
 582         int size, int offset)
 583 {
 584         #define SQ_ALU_CONST_BUFFER_SIZE_LS_0 0x00028fc0
 585         #define SQ_ALU_CONST_CACHE_LS_0 0x00028f40
 586
 587         struct evergreen_compute_resource* res =
 588                 get_empty_res(pipe, COMPUTE_RESOURCE_CONST_MEM, cache_id);
 589
 590         assert(size < 0x200);
 591         assert((offset & 0xFF) == 0);
 592         assert(cache_id < 16);
 593
 594         evergreen_reg_set(res, SQ_ALU_CONST_BUFFER_SIZE_LS_0 + cache_id*4, size);
 595         evergreen_reg_set(res, SQ_ALU_CONST_CACHE_LS_0 + cache_id*4, offset >> 8);
 596         res->bo = cbo;
 597         res->usage = RADEON_USAGE_READ;
 598         res->coher_bo_size = size;
 599
 600         pipe->ctx->flags |= R600_CONTEXT_SHADERCONST_FLUSH;
 601 }
 602
 603 struct r600_resource* r600_compute_buffer_alloc_vram(
 604         struct r600_screen *screen,
 605         unsigned size)
 606 {
 607         assert(size);
 608
 609         struct pipe_resource * buffer = pipe_buffer_create(
 610                         (struct pipe_screen*) screen,
 611                         PIPE_BIND_CUSTOM,
 612                         PIPE_USAGE_IMMUTABLE,
 613                         size);
 614
 615         return (struct r600_resource *)buffer;
 616 }