src/mesa/drivers/dri/i965/brw_vs.c

   1 /*
   2  Copyright (C) Intel Corp.  2006.  All Rights Reserved.
   3  Intel funded Tungsten Graphics to
   4  develop this 3D driver.
   5
   6  Permission is hereby granted, free of charge, to any person obtaining
   7  a copy of this software and associated documentation files (the
   8  "Software"), to deal in the Software without restriction, including
   9  without limitation the rights to use, copy, modify, merge, publish,
  10  distribute, sublicense, and/or sell copies of the Software, and to
  11  permit persons to whom the Software is furnished to do so, subject to
  12  the following conditions:
  13
  14  The above copyright notice and this permission notice (including the
  15  next paragraph) shall be included in all copies or substantial
  16  portions of the Software.
  17
  18  THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  19  EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
  21  IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
  22  LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
  23  OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
  24  WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25
  26  **********************************************************************/
  27  /*
  28   * Authors:
  29   *   Keith Whitwell <keithw@vmware.com>
  30   */
  31
  32
  33 #include "main/compiler.h"
  34 #include "main/context.h"
  35 #include "brw_context.h"
  36 #include "brw_vs.h"
  37 #include "brw_util.h"
  38 #include "brw_state.h"
  39 #include "program/prog_print.h"
  40 #include "program/prog_parameter.h"
  41 #include "compiler/brw_nir.h"
  42 #include "brw_program.h"
  43
  44 #include "util/ralloc.h"
  45
  46 /**
  47  * Decide which set of clip planes should be used when clipping via
  48  * gl_Position or gl_ClipVertex.
  49  */
  50 gl_clip_plane *
  51 brw_select_clip_planes(struct gl_context *ctx)
  52 {
  53    if (ctx->_Shader->CurrentProgram[MESA_SHADER_VERTEX]) {
  54       /* There is currently a GLSL vertex shader, so clip according to GLSL
  55        * rules, which means compare gl_ClipVertex (or gl_Position, if
  56        * gl_ClipVertex wasn't assigned) against the eye-coordinate clip planes
  57        * that were stored in EyeUserPlane at the time the clip planes were
  58        * specified.
  59        */
  60       return ctx->Transform.EyeUserPlane;
  61    } else {
  62       /* Either we are using fixed function or an ARB vertex program.  In
  63        * either case the clip planes are going to be compared against
  64        * gl_Position (which is in clip coordinates) so we have to clip using
  65        * _ClipUserPlane, which was transformed into clip coordinates by Mesa
  66        * core.
  67        */
  68       return ctx->Transform._ClipUserPlane;
  69    }
  70 }
  71
  72 GLbitfield64
  73 brw_vs_outputs_written(struct brw_context *brw, struct brw_vs_prog_key *key,
  74                        GLbitfield64 user_varyings)
  75 {
  76    GLbitfield64 outputs_written = user_varyings;
  77
  78    if (key->copy_edgeflag) {
  79       outputs_written |= BITFIELD64_BIT(VARYING_SLOT_EDGE);
  80    }
  81
  82    if (brw->gen < 6) {
  83       /* Put dummy slots into the VUE for the SF to put the replaced
  84        * point sprite coords in.  We shouldn't need these dummy slots,
  85        * which take up precious URB space, but it would mean that the SF
  86        * doesn't get nice aligned pairs of input coords into output
  87        * coords, which would be a pain to handle.
  88        */
  89       for (unsigned i = 0; i < 8; i++) {
  90          if (key->point_coord_replace & (1 << i))
  91             outputs_written |= BITFIELD64_BIT(VARYING_SLOT_TEX0 + i);
  92       }
  93
  94       /* if back colors are written, allocate slots for front colors too */
  95       if (outputs_written & BITFIELD64_BIT(VARYING_SLOT_BFC0))
  96          outputs_written |= BITFIELD64_BIT(VARYING_SLOT_COL0);
  97       if (outputs_written & BITFIELD64_BIT(VARYING_SLOT_BFC1))
  98          outputs_written |= BITFIELD64_BIT(VARYING_SLOT_COL1);
  99    }
 100
 101    /* In order for legacy clipping to work, we need to populate the clip
 102     * distance varying slots whenever clipping is enabled, even if the vertex
 103     * shader doesn't write to gl_ClipDistance.
 104     */
 105    if (key->nr_userclip_plane_consts > 0) {
 106       outputs_written |= BITFIELD64_BIT(VARYING_SLOT_CLIP_DIST0);
 107       outputs_written |= BITFIELD64_BIT(VARYING_SLOT_CLIP_DIST1);
 108    }
 109
 110    return outputs_written;
 111 }
 112
 113 static void
 114 brw_vs_debug_recompile(struct brw_context *brw, struct gl_program *prog,
 115                        const struct brw_vs_prog_key *key)
 116 {
 117    perf_debug("Recompiling vertex shader for program %d\n", prog->Id);
 118
 119    bool found = false;
 120    const struct brw_vs_prog_key *old_key =
 121       brw_find_previous_compile(&brw->cache, BRW_CACHE_VS_PROG,
 122                                 key->program_string_id);
 123
 124    if (!old_key) {
 125       perf_debug("  Didn't find previous compile in the shader cache for "
 126                  "debug\n");
 127       return;
 128    }
 129
 130    for (unsigned int i = 0; i < VERT_ATTRIB_MAX; i++) {
 131       found |= key_debug(brw, "Vertex attrib w/a flags",
 132                          old_key->gl_attrib_wa_flags[i],
 133                          key->gl_attrib_wa_flags[i]);
 134    }
 135
 136    found |= key_debug(brw, "legacy user clipping",
 137                       old_key->nr_userclip_plane_consts,
 138                       key->nr_userclip_plane_consts);
 139
 140    found |= key_debug(brw, "copy edgeflag",
 141                       old_key->copy_edgeflag, key->copy_edgeflag);
 142    found |= key_debug(brw, "PointCoord replace",
 143                       old_key->point_coord_replace, key->point_coord_replace);
 144    found |= key_debug(brw, "vertex color clamping",
 145                       old_key->clamp_vertex_color, key->clamp_vertex_color);
 146
 147    found |= brw_debug_recompile_sampler_key(brw, &old_key->tex, &key->tex);
 148
 149    if (!found) {
 150       perf_debug("  Something else\n");
 151    }
 152 }
 153
 154 static bool
 155 brw_codegen_vs_prog(struct brw_context *brw,
 156                     struct brw_program *vp,
 157                     struct brw_vs_prog_key *key)
 158 {
 159    const struct brw_compiler *compiler = brw->screen->compiler;
 160    const struct gen_device_info *devinfo = &brw->screen->devinfo;
 161    GLuint program_size;
 162    const GLuint *program;
 163    struct brw_vs_prog_data prog_data;
 164    struct brw_stage_prog_data *stage_prog_data = &prog_data.base.base;
 165    void *mem_ctx;
 166    bool start_busy = false;
 167    double start_time = 0;
 168
 169    memset(&prog_data, 0, sizeof(prog_data));
 170
 171    /* Use ALT floating point mode for ARB programs so that 0^0 == 1. */
 172    if (vp->program.is_arb_asm)
 173       stage_prog_data->use_alt_mode = true;
 174
 175    mem_ctx = ralloc_context(NULL);
 176
 177    brw_assign_common_binding_table_offsets(devinfo, &vp->program,
 178                                            &prog_data.base.base, 0);
 179
 180    /* Allocate the references to the uniforms that will end up in the
 181     * prog_data associated with the compiled program, and which will be freed
 182     * by the state cache.
 183     */
 184    int param_count = vp->program.nir->num_uniforms / 4;
 185
 186    prog_data.base.base.nr_image_params = vp->program.info.num_images;
 187
 188    /* vec4_visitor::setup_uniform_clipplane_values() also uploads user clip
 189     * planes as uniforms.
 190     */
 191    param_count += key->nr_userclip_plane_consts * 4;
 192
 193    stage_prog_data->param =
 194       rzalloc_array(NULL, const gl_constant_value *, param_count);
 195    stage_prog_data->pull_param =
 196       rzalloc_array(NULL, const gl_constant_value *, param_count);
 197    stage_prog_data->image_param =
 198       rzalloc_array(NULL, struct brw_image_param,
 199                     stage_prog_data->nr_image_params);
 200    stage_prog_data->nr_params = param_count;
 201
 202    if (!vp->program.is_arb_asm) {
 203       brw_nir_setup_glsl_uniforms(vp->program.nir, &vp->program,
 204                                   &prog_data.base.base,
 205                                   compiler->scalar_stage[MESA_SHADER_VERTEX]);
 206    } else {
 207       brw_nir_setup_arb_uniforms(vp->program.nir, &vp->program,
 208                                  &prog_data.base.base);
 209    }
 210
 211    uint64_t outputs_written =
 212       brw_vs_outputs_written(brw, key, vp->program.info.outputs_written);
 213
 214    brw_compute_vue_map(devinfo,
 215                        &prog_data.base.vue_map, outputs_written,
 216                        vp->program.nir->info.separate_shader);
 217
 218    if (0) {
 219       _mesa_fprint_program_opt(stderr, &vp->program, PROG_PRINT_DEBUG, true);
 220    }
 221
 222    if (unlikely(brw->perf_debug)) {
 223       start_busy = (brw->batch.last_bo &&
 224                     brw_bo_busy(brw->batch.last_bo));
 225       start_time = get_time();
 226    }
 227
 228    if (unlikely(INTEL_DEBUG & DEBUG_VS)) {
 229       if (vp->program.is_arb_asm)
 230          brw_dump_arb_asm("vertex", &vp->program);
 231    }
 232
 233    int st_index = -1;
 234    if (INTEL_DEBUG & DEBUG_SHADER_TIME) {
 235       st_index = brw_get_shader_time_index(brw, &vp->program, ST_VS,
 236                                            !vp->program.is_arb_asm);
 237    }
 238
 239    /* Emit GEN4 code.
 240     */
 241    char *error_str;
 242    program = brw_compile_vs(compiler, brw, mem_ctx, key, &prog_data,
 243                             vp->program.nir,
 244                             brw_select_clip_planes(&brw->ctx),
 245                             !_mesa_is_gles3(&brw->ctx),
 246                             st_index, &program_size, &error_str);
 247    if (program == NULL) {
 248       if (!vp->program.is_arb_asm) {
 249          vp->program.sh.data->LinkStatus = linking_failure;
 250          ralloc_strcat(&vp->program.sh.data->InfoLog, error_str);
 251       }
 252
 253       _mesa_problem(NULL, "Failed to compile vertex shader: %s\n", error_str);
 254
 255       ralloc_free(mem_ctx);
 256       return false;
 257    }
 258
 259    if (unlikely(brw->perf_debug)) {
 260       if (vp->compiled_once) {
 261          brw_vs_debug_recompile(brw, &vp->program, key);
 262       }
 263       if (start_busy && !brw_bo_busy(brw->batch.last_bo)) {
 264          perf_debug("VS compile took %.03f ms and stalled the GPU\n",
 265                     (get_time() - start_time) * 1000);
 266       }
 267       vp->compiled_once = true;
 268    }
 269
 270    /* Scratch space is used for register spilling */
 271    brw_alloc_stage_scratch(brw, &brw->vs.base,
 272                            prog_data.base.base.total_scratch,
 273                            devinfo->max_vs_threads);
 274
 275    brw_upload_cache(&brw->cache, BRW_CACHE_VS_PROG,
 276                     key, sizeof(struct brw_vs_prog_key),
 277                     program, program_size,
 278                     &prog_data, sizeof(prog_data),
 279                     &brw->vs.base.prog_offset, &brw->vs.base.prog_data);
 280    ralloc_free(mem_ctx);
 281
 282    return true;
 283 }
 284
 285 static bool
 286 brw_vs_state_dirty(const struct brw_context *brw)
 287 {
 288    return brw_state_dirty(brw,
 289                           _NEW_BUFFERS |
 290                           _NEW_LIGHT |
 291                           _NEW_POINT |
 292                           _NEW_POLYGON |
 293                           _NEW_TEXTURE |
 294                           _NEW_TRANSFORM,
 295                           BRW_NEW_VERTEX_PROGRAM |
 296                           BRW_NEW_VS_ATTRIB_WORKAROUNDS);
 297 }
 298
 299 void
 300 brw_vs_populate_key(struct brw_context *brw,
 301                     struct brw_vs_prog_key *key)
 302 {
 303    struct gl_context *ctx = &brw->ctx;
 304    /* BRW_NEW_VERTEX_PROGRAM */
 305    struct brw_program *vp = (struct brw_program *)brw->vertex_program;
 306    struct gl_program *prog = (struct gl_program *) brw->vertex_program;
 307
 308    memset(key, 0, sizeof(*key));
 309
 310    /* Just upload the program verbatim for now.  Always send it all
 311     * the inputs it asks for, whether they are varying or not.
 312     */
 313    key->program_string_id = vp->id;
 314
 315    if (ctx->Transform.ClipPlanesEnabled != 0 &&
 316        (ctx->API == API_OPENGL_COMPAT || ctx->API == API_OPENGLES) &&
 317        vp->program.info.clip_distance_array_size == 0) {
 318       key->nr_userclip_plane_consts =
 319          _mesa_logbase2(ctx->Transform.ClipPlanesEnabled) + 1;
 320    }
 321
 322    if (brw->gen < 6) {
 323       /* _NEW_POLYGON */
 324       key->copy_edgeflag = (ctx->Polygon.FrontMode != GL_FILL ||
 325                             ctx->Polygon.BackMode != GL_FILL);
 326
 327       /* _NEW_POINT */
 328       if (ctx->Point.PointSprite) {
 329          key->point_coord_replace = ctx->Point.CoordReplace & 0xff;
 330       }
 331    }
 332
 333    if (prog->info.outputs_written &
 334        (VARYING_BIT_COL0 | VARYING_BIT_COL1 | VARYING_BIT_BFC0 |
 335         VARYING_BIT_BFC1)) {
 336       /* _NEW_LIGHT | _NEW_BUFFERS */
 337       key->clamp_vertex_color = ctx->Light._ClampVertexColor;
 338    }
 339
 340    /* _NEW_TEXTURE */
 341    brw_populate_sampler_prog_key_data(ctx, prog, &key->tex);
 342
 343    /* BRW_NEW_VS_ATTRIB_WORKAROUNDS */
 344    if (brw->gen < 8 && !brw->is_haswell) {
 345       memcpy(key->gl_attrib_wa_flags, brw->vb.attrib_wa_flags,
 346              sizeof(brw->vb.attrib_wa_flags));
 347    }
 348 }
 349
 350 void
 351 brw_upload_vs_prog(struct brw_context *brw)
 352 {
 353    struct brw_vs_prog_key key;
 354    /* BRW_NEW_VERTEX_PROGRAM */
 355    struct brw_program *vp = (struct brw_program *)brw->vertex_program;
 356
 357    if (!brw_vs_state_dirty(brw))
 358       return;
 359
 360    brw_vs_populate_key(brw, &key);
 361
 362    if (!brw_search_cache(&brw->cache, BRW_CACHE_VS_PROG,
 363                          &key, sizeof(key),
 364                          &brw->vs.base.prog_offset, &brw->vs.base.prog_data)) {
 365       bool success = brw_codegen_vs_prog(brw, vp, &key);
 366       (void) success;
 367       assert(success);
 368    }
 369 }
 370
 371 bool
 372 brw_vs_precompile(struct gl_context *ctx, struct gl_program *prog)
 373 {
 374    struct brw_context *brw = brw_context(ctx);
 375    struct brw_vs_prog_key key;
 376    uint32_t old_prog_offset = brw->vs.base.prog_offset;
 377    struct brw_stage_prog_data *old_prog_data = brw->vs.base.prog_data;
 378    bool success;
 379
 380    struct brw_program *bvp = brw_program(prog);
 381
 382    memset(&key, 0, sizeof(key));
 383
 384    brw_setup_tex_for_precompile(brw, &key.tex, prog);
 385    key.program_string_id = bvp->id;
 386    key.clamp_vertex_color =
 387       (prog->info.outputs_written &
 388        (VARYING_BIT_COL0 | VARYING_BIT_COL1 | VARYING_BIT_BFC0 |
 389         VARYING_BIT_BFC1));
 390
 391    success = brw_codegen_vs_prog(brw, bvp, &key);
 392
 393    brw->vs.base.prog_offset = old_prog_offset;
 394    brw->vs.base.prog_data = old_prog_data;
 395
 396    return success;
 397 }