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 "svga_context.h"
  29 #include "svga_cmd.h"
  30 #include "svga_format.h"
  31 #include "svga_shader.h"
  32
  33
  34 /**
  35  * This bit isn't really used anywhere.  It only serves to help
  36  * generate a unique "signature" for the vertex shader output bitmask.
  37  * Shader input/output signatures are used to resolve shader linking
  38  * issues.
  39  */
  40 #define FOG_GENERIC_BIT (((uint64_t) 1) << 63)
  41
  42
  43 /**
  44  * Use the shader info to generate a bitmask indicating which generic
  45  * inputs are used by the shader.  A set bit indicates that GENERIC[i]
  46  * is used.
  47  */
  48 uint64_t
  49 svga_get_generic_inputs_mask(const struct tgsi_shader_info *info)
  50 {
  51    unsigned i;
  52    uint64_t mask = 0x0;
  53
  54    for (i = 0; i < info->num_inputs; i++) {
  55       if (info->input_semantic_name[i] == TGSI_SEMANTIC_GENERIC) {
  56          unsigned j = info->input_semantic_index[i];
  57          assert(j < sizeof(mask) * 8);
  58          mask |= ((uint64_t) 1) << j;
  59       }
  60    }
  61
  62    return mask;
  63 }
  64
  65
  66 /**
  67  * Scan shader info to return a bitmask of written outputs.
  68  */
  69 uint64_t
  70 svga_get_generic_outputs_mask(const struct tgsi_shader_info *info)
  71 {
  72    unsigned i;
  73    uint64_t mask = 0x0;
  74
  75    for (i = 0; i < info->num_outputs; i++) {
  76       switch (info->output_semantic_name[i]) {
  77       case TGSI_SEMANTIC_GENERIC:
  78          {
  79             unsigned j = info->output_semantic_index[i];
  80             assert(j < sizeof(mask) * 8);
  81             mask |= ((uint64_t) 1) << j;
  82          }
  83          break;
  84       case TGSI_SEMANTIC_FOG:
  85          mask |= FOG_GENERIC_BIT;
  86          break;
  87       }
  88    }
  89
  90    return mask;
  91 }
  92
  93
  94
  95 /**
  96  * Given a mask of used generic variables (as returned by the above functions)
  97  * fill in a table which maps those indexes to small integers.
  98  * This table is used by the remap_generic_index() function in
  99  * svga_tgsi_decl_sm30.c
 100  * Example: if generics_mask = binary(1010) it means that GENERIC[1] and
 101  * GENERIC[3] are used.  The remap_table will contain:
 102  *   table[1] = 0;
 103  *   table[3] = 1;
 104  * The remaining table entries will be filled in with the next unused
 105  * generic index (in this example, 2).
 106  */
 107 void
 108 svga_remap_generics(uint64_t generics_mask,
 109                     int8_t remap_table[MAX_GENERIC_VARYING])
 110 {
 111    /* Note texcoord[0] is reserved so start at 1 */
 112    unsigned count = 1, i;
 113
 114    for (i = 0; i < MAX_GENERIC_VARYING; i++) {
 115       remap_table[i] = -1;
 116    }
 117
 118    /* for each bit set in generic_mask */
 119    while (generics_mask) {
 120       unsigned index = ffsll(generics_mask) - 1;
 121       remap_table[index] = count++;
 122       generics_mask &= ~((uint64_t) 1 << index);
 123    }
 124 }
 125
 126
 127 /**
 128  * Use the generic remap table to map a TGSI generic varying variable
 129  * index to a small integer.  If the remapping table doesn't have a
 130  * valid value for the given index (the table entry is -1) it means
 131  * the fragment shader doesn't use that VS output.  Just allocate
 132  * the next free value in that case.  Alternately, we could cull
 133  * VS instructions that write to register, or replace the register
 134  * with a dummy temp register.
 135  * XXX TODO: we should do one of the later as it would save precious
 136  * texcoord registers.
 137  */
 138 int
 139 svga_remap_generic_index(int8_t remap_table[MAX_GENERIC_VARYING],
 140                          int generic_index)
 141 {
 142    assert(generic_index < MAX_GENERIC_VARYING);
 143
 144    if (generic_index >= MAX_GENERIC_VARYING) {
 145       /* just don't return a random/garbage value */
 146       generic_index = MAX_GENERIC_VARYING - 1;
 147    }
 148
 149    if (remap_table[generic_index] == -1) {
 150       /* This is a VS output that has no matching PS input.  Find a
 151        * free index.
 152        */
 153       int i, max = 0;
 154       for (i = 0; i < MAX_GENERIC_VARYING; i++) {
 155          max = MAX2(max, remap_table[i]);
 156       }
 157       remap_table[generic_index] = max + 1;
 158    }
 159
 160    return remap_table[generic_index];
 161 }
 162
 163
 164 /**
 165  * Initialize the shader-neutral fields of svga_compile_key from context
 166  * state.  This is basically the texture-related state.
 167  */
 168 void
 169 svga_init_shader_key_common(const struct svga_context *svga, unsigned shader,
 170                             struct svga_compile_key *key)
 171 {
 172    unsigned i, idx = 0;
 173
 174    assert(shader < Elements(svga->curr.num_sampler_views));
 175
 176    for (i = 0; i < svga->curr.num_sampler_views[shader]; i++) {
 177       struct pipe_sampler_view *view = svga->curr.sampler_views[shader][i];
 178       if (view) {
 179          assert(svga->curr.sampler[shader][i]);
 180          assert(view->texture);
 181          assert(view->texture->target < (1 << 4)); /* texture_target:4 */
 182
 183          key->tex[i].texture_target = view->texture->target;
 184
 185          /* 1D/2D array textures with one slice are treated as non-arrays
 186           * by the SVGA3D device.  Convert the texture type here so that
 187           * we emit the right TEX/SAMPLE instruction in the shader.
 188           */
 189          if (view->texture->array_size == 1) {
 190             if (view->texture->target == PIPE_TEXTURE_1D_ARRAY) {
 191                key->tex[i].texture_target = PIPE_TEXTURE_1D;
 192             }
 193             else if (view->texture->target == PIPE_TEXTURE_2D_ARRAY) {
 194                key->tex[i].texture_target = PIPE_TEXTURE_2D;
 195             }
 196          }
 197
 198          key->tex[i].texture_msaa = view->texture->nr_samples > 1;
 199          if (!svga->curr.sampler[shader][i]->normalized_coords) {
 200             assert(idx < (1 << 5));  /* width_height_idx:5 bitfield */
 201             key->tex[i].width_height_idx = idx++;
 202             key->tex[i].unnormalized = TRUE;
 203             ++key->num_unnormalized_coords;
 204          }
 205
 206          key->tex[i].swizzle_r = view->swizzle_r;
 207          key->tex[i].swizzle_g = view->swizzle_g;
 208          key->tex[i].swizzle_b = view->swizzle_b;
 209          key->tex[i].swizzle_a = view->swizzle_a;
 210
 211          key->tex[i].return_type = svga_get_texture_datatype(view->format);
 212       }
 213    }
 214    key->num_textures = svga->curr.num_sampler_views[shader];
 215 }
 216
 217
 218 /** Search for a compiled shader variant with the same compile key */
 219 struct svga_shader_variant *
 220 svga_search_shader_key(const struct svga_shader *shader,
 221                        const struct svga_compile_key *key)
 222 {
 223    struct svga_shader_variant *variant = shader->variants;
 224
 225    assert(key);
 226
 227    for ( ; variant; variant = variant->next) {
 228       if (svga_compile_keys_equal(key, &variant->key))
 229          return variant;
 230    }
 231    return NULL;
 232 }
 233
 234 /** Search for a shader with the same token key */
 235 struct svga_shader *
 236 svga_search_shader_token_key(struct svga_shader *pshader,
 237                              const struct svga_token_key *key)
 238 {
 239    struct svga_shader *shader = pshader;
 240
 241    assert(key);
 242
 243    for ( ; shader; shader = shader->next) {
 244       if (memcmp(key, &shader->token_key, sizeof(struct svga_token_key)) == 0)
 245          return shader;
 246    }
 247    return NULL;
 248 }
 249
 250 /**
 251  * Helper function to define a gb shader for non-vgpu10 device
 252  */
 253 static enum pipe_error
 254 define_gb_shader_vgpu9(struct svga_context *svga,
 255                        SVGA3dShaderType type,
 256                        struct svga_shader_variant *variant,
 257                        unsigned codeLen)
 258 {
 259    struct svga_winsys_screen *sws = svga_screen(svga->pipe.screen)->sws;
 260    enum pipe_error ret;
 261
 262    /**
 263     * Create gb memory for the shader and upload the shader code.
 264     * Kernel module will allocate an id for the shader and issue
 265     * the DefineGBShader command.
 266     */
 267    variant->gb_shader = sws->shader_create(sws, type,
 268                                            variant->tokens, codeLen);
 269
 270    if (!variant->gb_shader)
 271       return PIPE_ERROR_OUT_OF_MEMORY;
 272
 273    ret = SVGA3D_BindGBShader(svga->swc, variant->gb_shader);
 274
 275    return ret;
 276 }
 277
 278 /**
 279  * Helper function to define a gb shader for vgpu10 device
 280  */
 281 static enum pipe_error
 282 define_gb_shader_vgpu10(struct svga_context *svga,
 283                         SVGA3dShaderType type,
 284                         struct svga_shader_variant *variant,
 285                         unsigned codeLen)
 286 {
 287    struct svga_winsys_context *swc = svga->swc;
 288    enum pipe_error ret;
 289
 290    /**
 291     * Shaders in VGPU10 enabled device reside in the device COTable.
 292     * SVGA driver will allocate an integer ID for the shader and
 293     * issue DXDefineShader and DXBindShader commands.
 294     */
 295    variant->id = util_bitmask_add(svga->shader_id_bm);
 296    if (variant->id == UTIL_BITMASK_INVALID_INDEX) {
 297       return PIPE_ERROR_OUT_OF_MEMORY;
 298    }
 299
 300    /* Create gb memory for the shader and upload the shader code */
 301    variant->gb_shader = swc->shader_create(swc,
 302                                            variant->id, type,
 303                                            variant->tokens, codeLen);
 304
 305    if (!variant->gb_shader) {
 306       /* Free the shader ID */
 307       assert(variant->id != UTIL_BITMASK_INVALID_INDEX);
 308       goto fail_no_allocation;
 309    }
 310
 311    /**
 312     * Since we don't want to do any flush within state emission to avoid
 313     * partial state in a command buffer, it's important to make sure that
 314     * there is enough room to send both the DXDefineShader & DXBindShader
 315     * commands in the same command buffer. So let's send both
 316     * commands in one command reservation. If it fails, we'll undo
 317     * the shader creation and return an error.
 318     */
 319    ret = SVGA3D_vgpu10_DefineAndBindShader(swc, variant->gb_shader,
 320                                            variant->id, type, codeLen);
 321
 322    if (ret != PIPE_OK)
 323       goto fail;
 324
 325    return PIPE_OK;
 326
 327 fail:
 328    swc->shader_destroy(swc, variant->gb_shader);
 329    variant->gb_shader = NULL;
 330
 331 fail_no_allocation:
 332    util_bitmask_clear(svga->shader_id_bm, variant->id);
 333    variant->id = UTIL_BITMASK_INVALID_INDEX;
 334
 335    return PIPE_ERROR_OUT_OF_MEMORY;
 336 }
 337
 338 /**
 339  * Issue the SVGA3D commands to define a new shader.
 340  * \param variant  contains the shader tokens, etc.  The result->id field will
 341  *                 be set here.
 342  */
 343 enum pipe_error
 344 svga_define_shader(struct svga_context *svga,
 345                    SVGA3dShaderType type,
 346                    struct svga_shader_variant *variant)
 347 {
 348    unsigned codeLen = variant->nr_tokens * sizeof(variant->tokens[0]);
 349    enum pipe_error ret;
 350
 351    variant->id = UTIL_BITMASK_INVALID_INDEX;
 352
 353    if (svga_have_gb_objects(svga)) {
 354       if (svga_have_vgpu10(svga))
 355          return define_gb_shader_vgpu10(svga, type, variant, codeLen);
 356       else
 357          return define_gb_shader_vgpu9(svga, type, variant, codeLen);
 358    }
 359    else {
 360       /* Allocate an integer ID for the shader */
 361       variant->id = util_bitmask_add(svga->shader_id_bm);
 362       if (variant->id == UTIL_BITMASK_INVALID_INDEX) {
 363          return PIPE_ERROR_OUT_OF_MEMORY;
 364       }
 365
 366       /* Issue SVGA3D device command to define the shader */
 367       ret = SVGA3D_DefineShader(svga->swc,
 368                                 variant->id,
 369                                 type,
 370                                 variant->tokens,
 371                                 codeLen);
 372       if (ret != PIPE_OK) {
 373          /* free the ID */
 374          assert(variant->id != UTIL_BITMASK_INVALID_INDEX);
 375          util_bitmask_clear(svga->shader_id_bm, variant->id);
 376          variant->id = UTIL_BITMASK_INVALID_INDEX;
 377       }
 378    }
 379
 380    return ret;
 381 }
 382
 383
 384 /**
 385  * Issue the SVGA3D commands to set/bind a shader.
 386  * \param result  the shader to bind.
 387  */
 388 enum pipe_error
 389 svga_set_shader(struct svga_context *svga,
 390                 SVGA3dShaderType type,
 391                 struct svga_shader_variant *variant)
 392 {
 393    enum pipe_error ret;
 394    unsigned id = variant ? variant->id : SVGA3D_INVALID_ID;
 395
 396    assert(type == SVGA3D_SHADERTYPE_VS ||
 397           type == SVGA3D_SHADERTYPE_GS ||
 398           type == SVGA3D_SHADERTYPE_PS);
 399
 400    if (svga_have_gb_objects(svga)) {
 401       struct svga_winsys_gb_shader *gbshader =
 402          variant ? variant->gb_shader : NULL;
 403
 404       if (svga_have_vgpu10(svga))
 405          ret = SVGA3D_vgpu10_SetShader(svga->swc, type, gbshader, id);
 406       else
 407          ret = SVGA3D_SetGBShader(svga->swc, type, gbshader);
 408    }
 409    else {
 410       ret = SVGA3D_SetShader(svga->swc, type, id);
 411    }
 412
 413    return ret;
 414 }
 415
 416
 417 struct svga_shader_variant *
 418 svga_new_shader_variant(struct svga_context *svga)
 419 {
 420    svga->hud.num_shaders++;
 421    return CALLOC_STRUCT(svga_shader_variant);
 422 }
 423
 424
 425 enum pipe_error
 426 svga_destroy_shader_variant(struct svga_context *svga,
 427                             SVGA3dShaderType type,
 428                             struct svga_shader_variant *variant)
 429 {
 430    enum pipe_error ret = PIPE_OK;
 431
 432    if (svga_have_gb_objects(svga) && variant->gb_shader) {
 433       if (svga_have_vgpu10(svga)) {
 434          struct svga_winsys_context *swc = svga->swc;
 435          swc->shader_destroy(swc, variant->gb_shader);
 436          ret = SVGA3D_vgpu10_DestroyShader(svga->swc, variant->id);
 437          if (ret != PIPE_OK) {
 438             /* flush and try again */
 439             svga_context_flush(svga, NULL);
 440             ret = SVGA3D_vgpu10_DestroyShader(svga->swc, variant->id);
 441          }
 442          util_bitmask_clear(svga->shader_id_bm, variant->id);
 443       }
 444       else {
 445          struct svga_winsys_screen *sws = svga_screen(svga->pipe.screen)->sws;
 446          sws->shader_destroy(sws, variant->gb_shader);
 447       }
 448       variant->gb_shader = NULL;
 449    }
 450    else {
 451       if (variant->id != UTIL_BITMASK_INVALID_INDEX) {
 452          ret = SVGA3D_DestroyShader(svga->swc, variant->id, type);
 453          if (ret != PIPE_OK) {
 454             /* flush and try again */
 455             svga_context_flush(svga, NULL);
 456             ret = SVGA3D_DestroyShader(svga->swc, variant->id, type);
 457             assert(ret == PIPE_OK);
 458          }
 459          util_bitmask_clear(svga->shader_id_bm, variant->id);
 460       }
 461    }
 462
 463    FREE((unsigned *)variant->tokens);
 464    FREE(variant);
 465
 466    svga->hud.num_shaders--;
 467
 468    return ret;
 469 }
 470
 471 /*
 472  * Rebind shaders.
 473  * Called at the beginning of every new command buffer to ensure that
 474  * shaders are properly paged-in. Instead of sending the SetShader
 475  * command, this function sends a private allocation command to
 476  * page in a shader. This avoids emitting redundant state to the device
 477  * just to page in a resource.
 478  */
 479 enum pipe_error
 480 svga_rebind_shaders(struct svga_context *svga)
 481 {
 482    struct svga_winsys_context *swc = svga->swc;
 483    struct svga_hw_draw_state *hw = &svga->state.hw_draw;
 484    enum pipe_error ret;
 485
 486    assert(svga_have_vgpu10(svga));
 487
 488    /**
 489     * If the underlying winsys layer does not need resource rebinding,
 490     * just clear the rebind flags and return.
 491     */
 492    if (swc->resource_rebind == NULL) {
 493       svga->rebind.flags.vs = 0;
 494       svga->rebind.flags.gs = 0;
 495       svga->rebind.flags.fs = 0;
 496
 497       return PIPE_OK;
 498    }
 499
 500    if (svga->rebind.flags.vs && hw->vs && hw->vs->gb_shader) {
 501       ret = swc->resource_rebind(swc, NULL, hw->vs->gb_shader, SVGA_RELOC_READ);
 502       if (ret != PIPE_OK)
 503          return ret;
 504    }
 505    svga->rebind.flags.vs = 0;
 506
 507    if (svga->rebind.flags.gs && hw->gs && hw->gs->gb_shader) {
 508       ret = swc->resource_rebind(swc, NULL, hw->gs->gb_shader, SVGA_RELOC_READ);
 509       if (ret != PIPE_OK)
 510          return ret;
 511    }
 512    svga->rebind.flags.gs = 0;
 513
 514    if (svga->rebind.flags.fs && hw->fs && hw->fs->gb_shader) {
 515       ret = swc->resource_rebind(swc, NULL, hw->fs->gb_shader, SVGA_RELOC_READ);
 516       if (ret != PIPE_OK)
 517          return ret;
 518    }
 519    svga->rebind.flags.fs = 0;
 520
 521    return PIPE_OK;
 522 }