src/gallium/drivers/svga/svga_shader.c

   1 /**********************************************************
   2  * Copyright 2008-2012 VMware, Inc.  All rights reserved.
   3  *
   4  * Permission is hereby granted, free of charge, to any person
   5  * obtaining a copy of this software and associated documentation
   6  * files (the "Software"), to deal in the Software without
   7  * restriction, including without limitation the rights to use, copy,
   8  * modify, merge, publish, distribute, sublicense, and/or sell copies
   9  * of the Software, and to permit persons to whom the Software is
  10  * furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice shall be
  13  * included in all copies or substantial portions of the Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  16  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  17  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  18  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  19  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  20  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  21  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  22  * SOFTWARE.
  23  *
  24  **********************************************************/
  25
  26 #include "util/u_bitmask.h"
  27 #include "util/u_memory.h"
  28 #include "util/u_format.h"
  29 #include "svga_context.h"
  30 #include "svga_cmd.h"
  31 #include "svga_format.h"
  32 #include "svga_shader.h"
  33 #include "svga_resource_texture.h"
  34
  35
  36 /**
  37  * This bit isn't really used anywhere.  It only serves to help
  38  * generate a unique "signature" for the vertex shader output bitmask.
  39  * Shader input/output signatures are used to resolve shader linking
  40  * issues.
  41  */
  42 #define FOG_GENERIC_BIT (((uint64_t) 1) << 63)
  43
  44
  45 /**
  46  * Use the shader info to generate a bitmask indicating which generic
  47  * inputs are used by the shader.  A set bit indicates that GENERIC[i]
  48  * is used.
  49  */
  50 uint64_t
  51 svga_get_generic_inputs_mask(const struct tgsi_shader_info *info)
  52 {
  53    unsigned i;
  54    uint64_t mask = 0x0;
  55
  56    for (i = 0; i < info->num_inputs; i++) {
  57       if (info->input_semantic_name[i] == TGSI_SEMANTIC_GENERIC) {
  58          unsigned j = info->input_semantic_index[i];
  59          assert(j < sizeof(mask) * 8);
  60          mask |= ((uint64_t) 1) << j;
  61       }
  62    }
  63
  64    return mask;
  65 }
  66
  67
  68 /**
  69  * Scan shader info to return a bitmask of written outputs.
  70  */
  71 uint64_t
  72 svga_get_generic_outputs_mask(const struct tgsi_shader_info *info)
  73 {
  74    unsigned i;
  75    uint64_t mask = 0x0;
  76
  77    for (i = 0; i < info->num_outputs; i++) {
  78       switch (info->output_semantic_name[i]) {
  79       case TGSI_SEMANTIC_GENERIC:
  80          {
  81             unsigned j = info->output_semantic_index[i];
  82             assert(j < sizeof(mask) * 8);
  83             mask |= ((uint64_t) 1) << j;
  84          }
  85          break;
  86       case TGSI_SEMANTIC_FOG:
  87          mask |= FOG_GENERIC_BIT;
  88          break;
  89       }
  90    }
  91
  92    return mask;
  93 }
  94
  95
  96
  97 /**
  98  * Given a mask of used generic variables (as returned by the above functions)
  99  * fill in a table which maps those indexes to small integers.
 100  * This table is used by the remap_generic_index() function in
 101  * svga_tgsi_decl_sm30.c
 102  * Example: if generics_mask = binary(1010) it means that GENERIC[1] and
 103  * GENERIC[3] are used.  The remap_table will contain:
 104  *   table[1] = 0;
 105  *   table[3] = 1;
 106  * The remaining table entries will be filled in with the next unused
 107  * generic index (in this example, 2).
 108  */
 109 void
 110 svga_remap_generics(uint64_t generics_mask,
 111                     int8_t remap_table[MAX_GENERIC_VARYING])
 112 {
 113    /* Note texcoord[0] is reserved so start at 1 */
 114    unsigned count = 1, i;
 115
 116    for (i = 0; i < MAX_GENERIC_VARYING; i++) {
 117       remap_table[i] = -1;
 118    }
 119
 120    /* for each bit set in generic_mask */
 121    while (generics_mask) {
 122       unsigned index = ffsll(generics_mask) - 1;
 123       remap_table[index] = count++;
 124       generics_mask &= ~((uint64_t) 1 << index);
 125    }
 126 }
 127
 128
 129 /**
 130  * Use the generic remap table to map a TGSI generic varying variable
 131  * index to a small integer.  If the remapping table doesn't have a
 132  * valid value for the given index (the table entry is -1) it means
 133  * the fragment shader doesn't use that VS output.  Just allocate
 134  * the next free value in that case.  Alternately, we could cull
 135  * VS instructions that write to register, or replace the register
 136  * with a dummy temp register.
 137  * XXX TODO: we should do one of the later as it would save precious
 138  * texcoord registers.
 139  */
 140 int
 141 svga_remap_generic_index(int8_t remap_table[MAX_GENERIC_VARYING],
 142                          int generic_index)
 143 {
 144    assert(generic_index < MAX_GENERIC_VARYING);
 145
 146    if (generic_index >= MAX_GENERIC_VARYING) {
 147       /* just don't return a random/garbage value */
 148       generic_index = MAX_GENERIC_VARYING - 1;
 149    }
 150
 151    if (remap_table[generic_index] == -1) {
 152       /* This is a VS output that has no matching PS input.  Find a
 153        * free index.
 154        */
 155       int i, max = 0;
 156       for (i = 0; i < MAX_GENERIC_VARYING; i++) {
 157          max = MAX2(max, remap_table[i]);
 158       }
 159       remap_table[generic_index] = max + 1;
 160    }
 161
 162    return remap_table[generic_index];
 163 }
 164
 165 static const enum pipe_swizzle copy_alpha[PIPE_SWIZZLE_MAX] = {
 166    PIPE_SWIZZLE_X,
 167    PIPE_SWIZZLE_Y,
 168    PIPE_SWIZZLE_Z,
 169    PIPE_SWIZZLE_W,
 170    PIPE_SWIZZLE_0,
 171    PIPE_SWIZZLE_1,
 172    PIPE_SWIZZLE_NONE
 173 };
 174
 175 static const enum pipe_swizzle set_alpha[PIPE_SWIZZLE_MAX] = {
 176    PIPE_SWIZZLE_X,
 177    PIPE_SWIZZLE_Y,
 178    PIPE_SWIZZLE_Z,
 179    PIPE_SWIZZLE_1,
 180    PIPE_SWIZZLE_0,
 181    PIPE_SWIZZLE_1,
 182    PIPE_SWIZZLE_NONE
 183 };
 184
 185 /**
 186  * Initialize the shader-neutral fields of svga_compile_key from context
 187  * state.  This is basically the texture-related state.
 188  */
 189 void
 190 svga_init_shader_key_common(const struct svga_context *svga,
 191                             enum pipe_shader_type shader,
 192                             struct svga_compile_key *key)
 193 {
 194    unsigned i, idx = 0;
 195
 196    assert(shader < ARRAY_SIZE(svga->curr.num_sampler_views));
 197
 198    /* In case the number of samplers and sampler_views doesn't match,
 199     * loop over the lower of the two counts.
 200     */
 201    key->num_textures = MAX2(svga->curr.num_sampler_views[shader],
 202                             svga->curr.num_samplers[shader]);
 203
 204    for (i = 0; i < key->num_textures; i++) {
 205       struct pipe_sampler_view *view = svga->curr.sampler_views[shader][i];
 206       const struct svga_sampler_state *sampler = svga->curr.sampler[shader][i];
 207       if (view) {
 208          assert(view->texture);
 209          assert(view->texture->target < (1 << 4)); /* texture_target:4 */
 210
 211          /* 1D/2D array textures with one slice are treated as non-arrays
 212           * by the SVGA3D device.  Convert the texture type here so that
 213           * we emit the right TEX/SAMPLE instruction in the shader.
 214           */
 215          if (view->texture->target == PIPE_TEXTURE_1D_ARRAY ||
 216              view->texture->target == PIPE_TEXTURE_2D_ARRAY) {
 217             if (view->texture->array_size == 1) {
 218                key->tex[i].is_array = 0;
 219             }
 220             else {
 221                assert(view->texture->array_size > 1);
 222                key->tex[i].is_array = 1;
 223             }
 224          }
 225
 226          /* If we have a non-alpha view into an svga3d surface with an
 227           * alpha channel, then explicitly set the alpha channel to 1
 228           * when sampling. Note that we need to check the
 229           * actual device format to cover also imported surface cases.
 230           */
 231          const enum pipe_swizzle *swizzle_tab =
 232             (view->texture->target != PIPE_BUFFER &&
 233              !util_format_has_alpha(view->format) &&
 234              svga_texture_device_format_has_alpha(view->texture)) ?
 235             set_alpha : copy_alpha;
 236
 237          key->tex[i].swizzle_r = swizzle_tab[view->swizzle_r];
 238          key->tex[i].swizzle_g = swizzle_tab[view->swizzle_g];
 239          key->tex[i].swizzle_b = swizzle_tab[view->swizzle_b];
 240          key->tex[i].swizzle_a = swizzle_tab[view->swizzle_a];
 241       }
 242
 243       if (sampler) {
 244          if (!sampler->normalized_coords) {
 245             assert(idx < (1 << 5));  /* width_height_idx:5 bitfield */
 246             key->tex[i].width_height_idx = idx++;
 247             key->tex[i].unnormalized = TRUE;
 248             ++key->num_unnormalized_coords;
 249
 250             if (sampler->magfilter == SVGA3D_TEX_FILTER_NEAREST ||
 251                 sampler->minfilter == SVGA3D_TEX_FILTER_NEAREST) {
 252                 key->tex[i].texel_bias = TRUE;
 253             }
 254          }
 255       }
 256    }
 257 }
 258
 259
 260 /** Search for a compiled shader variant with the same compile key */
 261 struct svga_shader_variant *
 262 svga_search_shader_key(const struct svga_shader *shader,
 263                        const struct svga_compile_key *key)
 264 {
 265    struct svga_shader_variant *variant = shader->variants;
 266
 267    assert(key);
 268
 269    for ( ; variant; variant = variant->next) {
 270       if (svga_compile_keys_equal(key, &variant->key))
 271          return variant;
 272    }
 273    return NULL;
 274 }
 275
 276 /** Search for a shader with the same token key */
 277 struct svga_shader *
 278 svga_search_shader_token_key(struct svga_shader *pshader,
 279                              const struct svga_token_key *key)
 280 {
 281    struct svga_shader *shader = pshader;
 282
 283    assert(key);
 284
 285    for ( ; shader; shader = shader->next) {
 286       if (memcmp(key, &shader->token_key, sizeof(struct svga_token_key)) == 0)
 287          return shader;
 288    }
 289    return NULL;
 290 }
 291
 292 /**
 293  * Helper function to define a gb shader for non-vgpu10 device
 294  */
 295 static enum pipe_error
 296 define_gb_shader_vgpu9(struct svga_context *svga,
 297                        SVGA3dShaderType type,
 298                        struct svga_shader_variant *variant,
 299                        unsigned codeLen)
 300 {
 301    struct svga_winsys_screen *sws = svga_screen(svga->pipe.screen)->sws;
 302    enum pipe_error ret;
 303
 304    /**
 305     * Create gb memory for the shader and upload the shader code.
 306     * Kernel module will allocate an id for the shader and issue
 307     * the DefineGBShader command.
 308     */
 309    variant->gb_shader = sws->shader_create(sws, type,
 310                                            variant->tokens, codeLen);
 311
 312    if (!variant->gb_shader)
 313       return PIPE_ERROR_OUT_OF_MEMORY;
 314
 315    ret = SVGA3D_BindGBShader(svga->swc, variant->gb_shader);
 316
 317    return ret;
 318 }
 319
 320 /**
 321  * Helper function to define a gb shader for vgpu10 device
 322  */
 323 static enum pipe_error
 324 define_gb_shader_vgpu10(struct svga_context *svga,
 325                         SVGA3dShaderType type,
 326                         struct svga_shader_variant *variant,
 327                         unsigned codeLen)
 328 {
 329    struct svga_winsys_context *swc = svga->swc;
 330    enum pipe_error ret;
 331
 332    /**
 333     * Shaders in VGPU10 enabled device reside in the device COTable.
 334     * SVGA driver will allocate an integer ID for the shader and
 335     * issue DXDefineShader and DXBindShader commands.
 336     */
 337    variant->id = util_bitmask_add(svga->shader_id_bm);
 338    if (variant->id == UTIL_BITMASK_INVALID_INDEX) {
 339       return PIPE_ERROR_OUT_OF_MEMORY;
 340    }
 341
 342    /* Create gb memory for the shader and upload the shader code */
 343    variant->gb_shader = swc->shader_create(swc,
 344                                            variant->id, type,
 345                                            variant->tokens, codeLen);
 346
 347    if (!variant->gb_shader) {
 348       /* Free the shader ID */
 349       assert(variant->id != UTIL_BITMASK_INVALID_INDEX);
 350       goto fail_no_allocation;
 351    }
 352
 353    /**
 354     * Since we don't want to do any flush within state emission to avoid
 355     * partial state in a command buffer, it's important to make sure that
 356     * there is enough room to send both the DXDefineShader & DXBindShader
 357     * commands in the same command buffer. So let's send both
 358     * commands in one command reservation. If it fails, we'll undo
 359     * the shader creation and return an error.
 360     */
 361    ret = SVGA3D_vgpu10_DefineAndBindShader(swc, variant->gb_shader,
 362                                            variant->id, type, codeLen);
 363
 364    if (ret != PIPE_OK)
 365       goto fail;
 366
 367    return PIPE_OK;
 368
 369 fail:
 370    swc->shader_destroy(swc, variant->gb_shader);
 371    variant->gb_shader = NULL;
 372
 373 fail_no_allocation:
 374    util_bitmask_clear(svga->shader_id_bm, variant->id);
 375    variant->id = UTIL_BITMASK_INVALID_INDEX;
 376
 377    return PIPE_ERROR_OUT_OF_MEMORY;
 378 }
 379
 380 /**
 381  * Issue the SVGA3D commands to define a new shader.
 382  * \param variant  contains the shader tokens, etc.  The result->id field will
 383  *                 be set here.
 384  */
 385 enum pipe_error
 386 svga_define_shader(struct svga_context *svga,
 387                    SVGA3dShaderType type,
 388                    struct svga_shader_variant *variant)
 389 {
 390    unsigned codeLen = variant->nr_tokens * sizeof(variant->tokens[0]);
 391    enum pipe_error ret;
 392
 393    SVGA_STATS_TIME_PUSH(svga_sws(svga), SVGA_STATS_TIME_DEFINESHADER);
 394
 395    variant->id = UTIL_BITMASK_INVALID_INDEX;
 396
 397    if (svga_have_gb_objects(svga)) {
 398       if (svga_have_vgpu10(svga))
 399          ret = define_gb_shader_vgpu10(svga, type, variant, codeLen);
 400       else
 401          ret = define_gb_shader_vgpu9(svga, type, variant, codeLen);
 402    }
 403    else {
 404       /* Allocate an integer ID for the shader */
 405       variant->id = util_bitmask_add(svga->shader_id_bm);
 406       if (variant->id == UTIL_BITMASK_INVALID_INDEX) {
 407          ret = PIPE_ERROR_OUT_OF_MEMORY;
 408          goto done;
 409       }
 410
 411       /* Issue SVGA3D device command to define the shader */
 412       ret = SVGA3D_DefineShader(svga->swc,
 413                                 variant->id,
 414                                 type,
 415                                 variant->tokens,
 416                                 codeLen);
 417       if (ret != PIPE_OK) {
 418          /* free the ID */
 419          assert(variant->id != UTIL_BITMASK_INVALID_INDEX);
 420          util_bitmask_clear(svga->shader_id_bm, variant->id);
 421          variant->id = UTIL_BITMASK_INVALID_INDEX;
 422       }
 423    }
 424
 425 done:
 426    SVGA_STATS_TIME_POP(svga_sws(svga));
 427    return ret;
 428 }
 429
 430
 431 /**
 432  * Issue the SVGA3D commands to set/bind a shader.
 433  * \param result  the shader to bind.
 434  */
 435 enum pipe_error
 436 svga_set_shader(struct svga_context *svga,
 437                 SVGA3dShaderType type,
 438                 struct svga_shader_variant *variant)
 439 {
 440    enum pipe_error ret;
 441    unsigned id = variant ? variant->id : SVGA3D_INVALID_ID;
 442
 443    assert(type == SVGA3D_SHADERTYPE_VS ||
 444           type == SVGA3D_SHADERTYPE_GS ||
 445           type == SVGA3D_SHADERTYPE_PS);
 446
 447    if (svga_have_gb_objects(svga)) {
 448       struct svga_winsys_gb_shader *gbshader =
 449          variant ? variant->gb_shader : NULL;
 450
 451       if (svga_have_vgpu10(svga))
 452          ret = SVGA3D_vgpu10_SetShader(svga->swc, type, gbshader, id);
 453       else
 454          ret = SVGA3D_SetGBShader(svga->swc, type, gbshader);
 455    }
 456    else {
 457       ret = SVGA3D_SetShader(svga->swc, type, id);
 458    }
 459
 460    return ret;
 461 }
 462
 463
 464 struct svga_shader_variant *
 465 svga_new_shader_variant(struct svga_context *svga)
 466 {
 467    svga->hud.num_shaders++;
 468    return CALLOC_STRUCT(svga_shader_variant);
 469 }
 470
 471
 472 enum pipe_error
 473 svga_destroy_shader_variant(struct svga_context *svga,
 474                             SVGA3dShaderType type,
 475                             struct svga_shader_variant *variant)
 476 {
 477    enum pipe_error ret = PIPE_OK;
 478
 479    if (svga_have_gb_objects(svga) && variant->gb_shader) {
 480       if (svga_have_vgpu10(svga)) {
 481          struct svga_winsys_context *swc = svga->swc;
 482          swc->shader_destroy(swc, variant->gb_shader);
 483          ret = SVGA3D_vgpu10_DestroyShader(svga->swc, variant->id);
 484          if (ret != PIPE_OK) {
 485             /* flush and try again */
 486             svga_context_flush(svga, NULL);
 487             ret = SVGA3D_vgpu10_DestroyShader(svga->swc, variant->id);
 488          }
 489          util_bitmask_clear(svga->shader_id_bm, variant->id);
 490       }
 491       else {
 492          struct svga_winsys_screen *sws = svga_screen(svga->pipe.screen)->sws;
 493          sws->shader_destroy(sws, variant->gb_shader);
 494       }
 495       variant->gb_shader = NULL;
 496    }
 497    else {
 498       if (variant->id != UTIL_BITMASK_INVALID_INDEX) {
 499          ret = SVGA3D_DestroyShader(svga->swc, variant->id, type);
 500          if (ret != PIPE_OK) {
 501             /* flush and try again */
 502             svga_context_flush(svga, NULL);
 503             ret = SVGA3D_DestroyShader(svga->swc, variant->id, type);
 504             assert(ret == PIPE_OK);
 505          }
 506          util_bitmask_clear(svga->shader_id_bm, variant->id);
 507       }
 508    }
 509
 510    FREE((unsigned *)variant->tokens);
 511    FREE(variant);
 512
 513    svga->hud.num_shaders--;
 514
 515    return ret;
 516 }
 517
 518 /*
 519  * Rebind shaders.
 520  * Called at the beginning of every new command buffer to ensure that
 521  * shaders are properly paged-in. Instead of sending the SetShader
 522  * command, this function sends a private allocation command to
 523  * page in a shader. This avoids emitting redundant state to the device
 524  * just to page in a resource.
 525  */
 526 enum pipe_error
 527 svga_rebind_shaders(struct svga_context *svga)
 528 {
 529    struct svga_winsys_context *swc = svga->swc;
 530    struct svga_hw_draw_state *hw = &svga->state.hw_draw;
 531    enum pipe_error ret;
 532
 533    assert(svga_have_vgpu10(svga));
 534
 535    /**
 536     * If the underlying winsys layer does not need resource rebinding,
 537     * just clear the rebind flags and return.
 538     */
 539    if (swc->resource_rebind == NULL) {
 540       svga->rebind.flags.vs = 0;
 541       svga->rebind.flags.gs = 0;
 542       svga->rebind.flags.fs = 0;
 543
 544       return PIPE_OK;
 545    }
 546
 547    if (svga->rebind.flags.vs && hw->vs && hw->vs->gb_shader) {
 548       ret = swc->resource_rebind(swc, NULL, hw->vs->gb_shader, SVGA_RELOC_READ);
 549       if (ret != PIPE_OK)
 550          return ret;
 551    }
 552    svga->rebind.flags.vs = 0;
 553
 554    if (svga->rebind.flags.gs && hw->gs && hw->gs->gb_shader) {
 555       ret = swc->resource_rebind(swc, NULL, hw->gs->gb_shader, SVGA_RELOC_READ);
 556       if (ret != PIPE_OK)
 557          return ret;
 558    }
 559    svga->rebind.flags.gs = 0;
 560
 561    if (svga->rebind.flags.fs && hw->fs && hw->fs->gb_shader) {
 562       ret = swc->resource_rebind(swc, NULL, hw->fs->gb_shader, SVGA_RELOC_READ);
 563       if (ret != PIPE_OK)
 564          return ret;
 565    }
 566    svga->rebind.flags.fs = 0;
 567
 568    return PIPE_OK;
 569 }