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 < ARRAY_SIZE(svga->curr.num_sampler_views));
 175
 176    /* In case the number of samplers and sampler_views doesn't match,
 177     * loop over the lower of the two counts.
 178     */
 179    key->num_textures = MIN2(svga->curr.num_sampler_views[shader],
 180                             svga->curr.num_samplers[shader]);
 181
 182    for (i = 0; i < key->num_textures; i++) {
 183       struct pipe_sampler_view *view = svga->curr.sampler_views[shader][i];
 184       const struct svga_sampler_state *sampler = svga->curr.sampler[shader][i];
 185       if (view && sampler) {
 186          assert(view->texture);
 187          assert(view->texture->target < (1 << 4)); /* texture_target:4 */
 188
 189          /* 1D/2D array textures with one slice are treated as non-arrays
 190           * by the SVGA3D device.  Convert the texture type here so that
 191           * we emit the right TEX/SAMPLE instruction in the shader.
 192           */
 193          if (view->texture->target == PIPE_TEXTURE_1D_ARRAY ||
 194              view->texture->target == PIPE_TEXTURE_2D_ARRAY) {
 195             if (view->texture->array_size == 1) {
 196                key->tex[i].is_array = 0;
 197             }
 198             else {
 199                assert(view->texture->array_size > 1);
 200                key->tex[i].is_array = 1;
 201             }
 202          }
 203
 204          if (!sampler->normalized_coords) {
 205             assert(idx < (1 << 5));  /* width_height_idx:5 bitfield */
 206             key->tex[i].width_height_idx = idx++;
 207             key->tex[i].unnormalized = TRUE;
 208             ++key->num_unnormalized_coords;
 209          }
 210
 211          key->tex[i].swizzle_r = view->swizzle_r;
 212          key->tex[i].swizzle_g = view->swizzle_g;
 213          key->tex[i].swizzle_b = view->swizzle_b;
 214          key->tex[i].swizzle_a = view->swizzle_a;
 215       }
 216    }
 217 }
 218
 219
 220 /** Search for a compiled shader variant with the same compile key */
 221 struct svga_shader_variant *
 222 svga_search_shader_key(const struct svga_shader *shader,
 223                        const struct svga_compile_key *key)
 224 {
 225    struct svga_shader_variant *variant = shader->variants;
 226
 227    assert(key);
 228
 229    for ( ; variant; variant = variant->next) {
 230       if (svga_compile_keys_equal(key, &variant->key))
 231          return variant;
 232    }
 233    return NULL;
 234 }
 235
 236 /** Search for a shader with the same token key */
 237 struct svga_shader *
 238 svga_search_shader_token_key(struct svga_shader *pshader,
 239                              const struct svga_token_key *key)
 240 {
 241    struct svga_shader *shader = pshader;
 242
 243    assert(key);
 244
 245    for ( ; shader; shader = shader->next) {
 246       if (memcmp(key, &shader->token_key, sizeof(struct svga_token_key)) == 0)
 247          return shader;
 248    }
 249    return NULL;
 250 }
 251
 252 /**
 253  * Helper function to define a gb shader for non-vgpu10 device
 254  */
 255 static enum pipe_error
 256 define_gb_shader_vgpu9(struct svga_context *svga,
 257                        SVGA3dShaderType type,
 258                        struct svga_shader_variant *variant,
 259                        unsigned codeLen)
 260 {
 261    struct svga_winsys_screen *sws = svga_screen(svga->pipe.screen)->sws;
 262    enum pipe_error ret;
 263
 264    /**
 265     * Create gb memory for the shader and upload the shader code.
 266     * Kernel module will allocate an id for the shader and issue
 267     * the DefineGBShader command.
 268     */
 269    variant->gb_shader = sws->shader_create(sws, type,
 270                                            variant->tokens, codeLen);
 271
 272    if (!variant->gb_shader)
 273       return PIPE_ERROR_OUT_OF_MEMORY;
 274
 275    ret = SVGA3D_BindGBShader(svga->swc, variant->gb_shader);
 276
 277    return ret;
 278 }
 279
 280 /**
 281  * Helper function to define a gb shader for vgpu10 device
 282  */
 283 static enum pipe_error
 284 define_gb_shader_vgpu10(struct svga_context *svga,
 285                         SVGA3dShaderType type,
 286                         struct svga_shader_variant *variant,
 287                         unsigned codeLen)
 288 {
 289    struct svga_winsys_context *swc = svga->swc;
 290    enum pipe_error ret;
 291
 292    /**
 293     * Shaders in VGPU10 enabled device reside in the device COTable.
 294     * SVGA driver will allocate an integer ID for the shader and
 295     * issue DXDefineShader and DXBindShader commands.
 296     */
 297    variant->id = util_bitmask_add(svga->shader_id_bm);
 298    if (variant->id == UTIL_BITMASK_INVALID_INDEX) {
 299       return PIPE_ERROR_OUT_OF_MEMORY;
 300    }
 301
 302    /* Create gb memory for the shader and upload the shader code */
 303    variant->gb_shader = swc->shader_create(swc,
 304                                            variant->id, type,
 305                                            variant->tokens, codeLen);
 306
 307    if (!variant->gb_shader) {
 308       /* Free the shader ID */
 309       assert(variant->id != UTIL_BITMASK_INVALID_INDEX);
 310       goto fail_no_allocation;
 311    }
 312
 313    /**
 314     * Since we don't want to do any flush within state emission to avoid
 315     * partial state in a command buffer, it's important to make sure that
 316     * there is enough room to send both the DXDefineShader & DXBindShader
 317     * commands in the same command buffer. So let's send both
 318     * commands in one command reservation. If it fails, we'll undo
 319     * the shader creation and return an error.
 320     */
 321    ret = SVGA3D_vgpu10_DefineAndBindShader(swc, variant->gb_shader,
 322                                            variant->id, type, codeLen);
 323
 324    if (ret != PIPE_OK)
 325       goto fail;
 326
 327    return PIPE_OK;
 328
 329 fail:
 330    swc->shader_destroy(swc, variant->gb_shader);
 331    variant->gb_shader = NULL;
 332
 333 fail_no_allocation:
 334    util_bitmask_clear(svga->shader_id_bm, variant->id);
 335    variant->id = UTIL_BITMASK_INVALID_INDEX;
 336
 337    return PIPE_ERROR_OUT_OF_MEMORY;
 338 }
 339
 340 /**
 341  * Issue the SVGA3D commands to define a new shader.
 342  * \param variant  contains the shader tokens, etc.  The result->id field will
 343  *                 be set here.
 344  */
 345 enum pipe_error
 346 svga_define_shader(struct svga_context *svga,
 347                    SVGA3dShaderType type,
 348                    struct svga_shader_variant *variant)
 349 {
 350    unsigned codeLen = variant->nr_tokens * sizeof(variant->tokens[0]);
 351    enum pipe_error ret;
 352
 353    variant->id = UTIL_BITMASK_INVALID_INDEX;
 354
 355    if (svga_have_gb_objects(svga)) {
 356       if (svga_have_vgpu10(svga))
 357          return define_gb_shader_vgpu10(svga, type, variant, codeLen);
 358       else
 359          return define_gb_shader_vgpu9(svga, type, variant, codeLen);
 360    }
 361    else {
 362       /* Allocate an integer ID for the shader */
 363       variant->id = util_bitmask_add(svga->shader_id_bm);
 364       if (variant->id == UTIL_BITMASK_INVALID_INDEX) {
 365          return PIPE_ERROR_OUT_OF_MEMORY;
 366       }
 367
 368       /* Issue SVGA3D device command to define the shader */
 369       ret = SVGA3D_DefineShader(svga->swc,
 370                                 variant->id,
 371                                 type,
 372                                 variant->tokens,
 373                                 codeLen);
 374       if (ret != PIPE_OK) {
 375          /* free the ID */
 376          assert(variant->id != UTIL_BITMASK_INVALID_INDEX);
 377          util_bitmask_clear(svga->shader_id_bm, variant->id);
 378          variant->id = UTIL_BITMASK_INVALID_INDEX;
 379       }
 380    }
 381
 382    return ret;
 383 }
 384
 385
 386 /**
 387  * Issue the SVGA3D commands to set/bind a shader.
 388  * \param result  the shader to bind.
 389  */
 390 enum pipe_error
 391 svga_set_shader(struct svga_context *svga,
 392                 SVGA3dShaderType type,
 393                 struct svga_shader_variant *variant)
 394 {
 395    enum pipe_error ret;
 396    unsigned id = variant ? variant->id : SVGA3D_INVALID_ID;
 397
 398    assert(type == SVGA3D_SHADERTYPE_VS ||
 399           type == SVGA3D_SHADERTYPE_GS ||
 400           type == SVGA3D_SHADERTYPE_PS);
 401
 402    if (svga_have_gb_objects(svga)) {
 403       struct svga_winsys_gb_shader *gbshader =
 404          variant ? variant->gb_shader : NULL;
 405
 406       if (svga_have_vgpu10(svga))
 407          ret = SVGA3D_vgpu10_SetShader(svga->swc, type, gbshader, id);
 408       else
 409          ret = SVGA3D_SetGBShader(svga->swc, type, gbshader);
 410    }
 411    else {
 412       ret = SVGA3D_SetShader(svga->swc, type, id);
 413    }
 414
 415    return ret;
 416 }
 417
 418
 419 struct svga_shader_variant *
 420 svga_new_shader_variant(struct svga_context *svga)
 421 {
 422    svga->hud.num_shaders++;
 423    return CALLOC_STRUCT(svga_shader_variant);
 424 }
 425
 426
 427 enum pipe_error
 428 svga_destroy_shader_variant(struct svga_context *svga,
 429                             SVGA3dShaderType type,
 430                             struct svga_shader_variant *variant)
 431 {
 432    enum pipe_error ret = PIPE_OK;
 433
 434    if (svga_have_gb_objects(svga) && variant->gb_shader) {
 435       if (svga_have_vgpu10(svga)) {
 436          struct svga_winsys_context *swc = svga->swc;
 437          swc->shader_destroy(swc, variant->gb_shader);
 438          ret = SVGA3D_vgpu10_DestroyShader(svga->swc, variant->id);
 439          if (ret != PIPE_OK) {
 440             /* flush and try again */
 441             svga_context_flush(svga, NULL);
 442             ret = SVGA3D_vgpu10_DestroyShader(svga->swc, variant->id);
 443          }
 444          util_bitmask_clear(svga->shader_id_bm, variant->id);
 445       }
 446       else {
 447          struct svga_winsys_screen *sws = svga_screen(svga->pipe.screen)->sws;
 448          sws->shader_destroy(sws, variant->gb_shader);
 449       }
 450       variant->gb_shader = NULL;
 451    }
 452    else {
 453       if (variant->id != UTIL_BITMASK_INVALID_INDEX) {
 454          ret = SVGA3D_DestroyShader(svga->swc, variant->id, type);
 455          if (ret != PIPE_OK) {
 456             /* flush and try again */
 457             svga_context_flush(svga, NULL);
 458             ret = SVGA3D_DestroyShader(svga->swc, variant->id, type);
 459             assert(ret == PIPE_OK);
 460          }
 461          util_bitmask_clear(svga->shader_id_bm, variant->id);
 462       }
 463    }
 464
 465    FREE((unsigned *)variant->tokens);
 466    FREE(variant);
 467
 468    svga->hud.num_shaders--;
 469
 470    return ret;
 471 }
 472
 473 /*
 474  * Rebind shaders.
 475  * Called at the beginning of every new command buffer to ensure that
 476  * shaders are properly paged-in. Instead of sending the SetShader
 477  * command, this function sends a private allocation command to
 478  * page in a shader. This avoids emitting redundant state to the device
 479  * just to page in a resource.
 480  */
 481 enum pipe_error
 482 svga_rebind_shaders(struct svga_context *svga)
 483 {
 484    struct svga_winsys_context *swc = svga->swc;
 485    struct svga_hw_draw_state *hw = &svga->state.hw_draw;
 486    enum pipe_error ret;
 487
 488    assert(svga_have_vgpu10(svga));
 489
 490    /**
 491     * If the underlying winsys layer does not need resource rebinding,
 492     * just clear the rebind flags and return.
 493     */
 494    if (swc->resource_rebind == NULL) {
 495       svga->rebind.flags.vs = 0;
 496       svga->rebind.flags.gs = 0;
 497       svga->rebind.flags.fs = 0;
 498
 499       return PIPE_OK;
 500    }
 501
 502    if (svga->rebind.flags.vs && hw->vs && hw->vs->gb_shader) {
 503       ret = swc->resource_rebind(swc, NULL, hw->vs->gb_shader, SVGA_RELOC_READ);
 504       if (ret != PIPE_OK)
 505          return ret;
 506    }
 507    svga->rebind.flags.vs = 0;
 508
 509    if (svga->rebind.flags.gs && hw->gs && hw->gs->gb_shader) {
 510       ret = swc->resource_rebind(swc, NULL, hw->gs->gb_shader, SVGA_RELOC_READ);
 511       if (ret != PIPE_OK)
 512          return ret;
 513    }
 514    svga->rebind.flags.gs = 0;
 515
 516    if (svga->rebind.flags.fs && hw->fs && hw->fs->gb_shader) {
 517       ret = swc->resource_rebind(swc, NULL, hw->fs->gb_shader, SVGA_RELOC_READ);
 518       if (ret != PIPE_OK)
 519          return ret;
 520    }
 521    svga->rebind.flags.fs = 0;
 522
 523    return PIPE_OK;
 524 }