brw_vs.c revision bdedd03b701781c8b71e162f7eb834e6a11105de
1/* 2 Copyright (C) Intel Corp. 2006. All Rights Reserved. 3 Intel funded Tungsten Graphics (http://www.tungstengraphics.com) 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 <keith@tungstengraphics.com> 30 */ 31 32 33#include "main/compiler.h" 34#include "brw_context.h" 35#include "brw_vs.h" 36#include "brw_util.h" 37#include "brw_state.h" 38#include "program/prog_print.h" 39#include "program/prog_parameter.h" 40 41#include "glsl/ralloc.h" 42 43static inline void assign_vue_slot(struct brw_vue_map *vue_map, 44 int vert_result) 45{ 46 /* Make sure this vert_result hasn't been assigned a slot already */ 47 assert (vue_map->vert_result_to_slot[vert_result] == -1); 48 49 vue_map->vert_result_to_slot[vert_result] = vue_map->num_slots; 50 vue_map->slot_to_vert_result[vue_map->num_slots++] = vert_result; 51} 52 53/** 54 * Compute the VUE map for vertex shader program. 55 */ 56void 57brw_compute_vue_map(struct brw_vue_map *vue_map, 58 const struct intel_context *intel, 59 bool userclip_active, 60 GLbitfield64 outputs_written) 61{ 62 int i; 63 64 vue_map->num_slots = 0; 65 for (i = 0; i < BRW_VERT_RESULT_MAX; ++i) { 66 vue_map->vert_result_to_slot[i] = -1; 67 vue_map->slot_to_vert_result[i] = BRW_VERT_RESULT_MAX; 68 } 69 70 /* VUE header: format depends on chip generation and whether clipping is 71 * enabled. 72 */ 73 switch (intel->gen) { 74 case 4: 75 /* There are 8 dwords in VUE header pre-Ironlake: 76 * dword 0-3 is indices, point width, clip flags. 77 * dword 4-7 is ndc position 78 * dword 8-11 is the first vertex data. 79 */ 80 assign_vue_slot(vue_map, VERT_RESULT_PSIZ); 81 assign_vue_slot(vue_map, BRW_VERT_RESULT_NDC); 82 assign_vue_slot(vue_map, VERT_RESULT_HPOS); 83 break; 84 case 5: 85 /* There are 20 DWs (D0-D19) in VUE header on Ironlake: 86 * dword 0-3 of the header is indices, point width, clip flags. 87 * dword 4-7 is the ndc position 88 * dword 8-11 of the vertex header is the 4D space position 89 * dword 12-19 of the vertex header is the user clip distance. 90 * dword 20-23 is a pad so that the vertex element data is aligned 91 * dword 24-27 is the first vertex data we fill. 92 * 93 * Note: future pipeline stages expect 4D space position to be 94 * contiguous with the other vert_results, so we make dword 24-27 a 95 * duplicate copy of the 4D space position. 96 */ 97 assign_vue_slot(vue_map, VERT_RESULT_PSIZ); 98 assign_vue_slot(vue_map, BRW_VERT_RESULT_NDC); 99 assign_vue_slot(vue_map, BRW_VERT_RESULT_HPOS_DUPLICATE); 100 assign_vue_slot(vue_map, VERT_RESULT_CLIP_DIST0); 101 assign_vue_slot(vue_map, VERT_RESULT_CLIP_DIST1); 102 assign_vue_slot(vue_map, BRW_VERT_RESULT_PAD); 103 assign_vue_slot(vue_map, VERT_RESULT_HPOS); 104 break; 105 case 6: 106 case 7: 107 /* There are 8 or 16 DWs (D0-D15) in VUE header on Sandybridge: 108 * dword 0-3 of the header is indices, point width, clip flags. 109 * dword 4-7 is the 4D space position 110 * dword 8-15 of the vertex header is the user clip distance if 111 * enabled. 112 * dword 8-11 or 16-19 is the first vertex element data we fill. 113 */ 114 assign_vue_slot(vue_map, VERT_RESULT_PSIZ); 115 assign_vue_slot(vue_map, VERT_RESULT_HPOS); 116 if (userclip_active) { 117 assign_vue_slot(vue_map, VERT_RESULT_CLIP_DIST0); 118 assign_vue_slot(vue_map, VERT_RESULT_CLIP_DIST1); 119 } 120 /* front and back colors need to be consecutive so that we can use 121 * ATTRIBUTE_SWIZZLE_INPUTATTR_FACING to swizzle them when doing 122 * two-sided color. 123 */ 124 if (outputs_written & BITFIELD64_BIT(VERT_RESULT_COL0)) 125 assign_vue_slot(vue_map, VERT_RESULT_COL0); 126 if (outputs_written & BITFIELD64_BIT(VERT_RESULT_BFC0)) 127 assign_vue_slot(vue_map, VERT_RESULT_BFC0); 128 if (outputs_written & BITFIELD64_BIT(VERT_RESULT_COL1)) 129 assign_vue_slot(vue_map, VERT_RESULT_COL1); 130 if (outputs_written & BITFIELD64_BIT(VERT_RESULT_BFC1)) 131 assign_vue_slot(vue_map, VERT_RESULT_BFC1); 132 break; 133 default: 134 assert (!"VUE map not known for this chip generation"); 135 break; 136 } 137 138 /* The hardware doesn't care about the rest of the vertex outputs, so just 139 * assign them contiguously. Don't reassign outputs that already have a 140 * slot. 141 * 142 * Also, prior to Gen6, don't assign a slot for VERT_RESULT_CLIP_VERTEX, 143 * since it is unsupported. In Gen6 and above, VERT_RESULT_CLIP_VERTEX may 144 * be needed for transform feedback; since we don't want to have to 145 * recompute the VUE map (and everything that depends on it) when transform 146 * feedback is enabled or disabled, just go ahead and assign a slot for it. 147 */ 148 for (int i = 0; i < VERT_RESULT_MAX; ++i) { 149 if (intel->gen < 6 && i == VERT_RESULT_CLIP_VERTEX) 150 continue; 151 if ((outputs_written & BITFIELD64_BIT(i)) && 152 vue_map->vert_result_to_slot[i] == -1) { 153 assign_vue_slot(vue_map, i); 154 } 155 } 156} 157 158 159/** 160 * Decide which set of clip planes should be used when clipping via 161 * gl_Position or gl_ClipVertex. 162 */ 163gl_clip_plane *brw_select_clip_planes(struct gl_context *ctx) 164{ 165 if (ctx->Shader.CurrentVertexProgram) { 166 /* There is currently a GLSL vertex shader, so clip according to GLSL 167 * rules, which means compare gl_ClipVertex (or gl_Position, if 168 * gl_ClipVertex wasn't assigned) against the eye-coordinate clip planes 169 * that were stored in EyeUserPlane at the time the clip planes were 170 * specified. 171 */ 172 return ctx->Transform.EyeUserPlane; 173 } else { 174 /* Either we are using fixed function or an ARB vertex program. In 175 * either case the clip planes are going to be compared against 176 * gl_Position (which is in clip coordinates) so we have to clip using 177 * _ClipUserPlane, which was transformed into clip coordinates by Mesa 178 * core. 179 */ 180 return ctx->Transform._ClipUserPlane; 181 } 182} 183 184 185static bool 186do_vs_prog(struct brw_context *brw, 187 struct gl_shader_program *prog, 188 struct brw_vertex_program *vp, 189 struct brw_vs_prog_key *key) 190{ 191 struct gl_context *ctx = &brw->intel.ctx; 192 struct intel_context *intel = &brw->intel; 193 GLuint program_size; 194 const GLuint *program; 195 struct brw_vs_compile c; 196 void *mem_ctx; 197 int aux_size; 198 int i; 199 200 memset(&c, 0, sizeof(c)); 201 memcpy(&c.key, key, sizeof(*key)); 202 203 mem_ctx = ralloc_context(NULL); 204 205 brw_init_compile(brw, &c.func, mem_ctx); 206 c.vp = vp; 207 208 c.prog_data.outputs_written = vp->program.Base.OutputsWritten; 209 c.prog_data.inputs_read = vp->program.Base.InputsRead; 210 211 if (c.key.copy_edgeflag) { 212 c.prog_data.outputs_written |= BITFIELD64_BIT(VERT_RESULT_EDGE); 213 c.prog_data.inputs_read |= VERT_BIT_EDGEFLAG; 214 } 215 216 /* Put dummy slots into the VUE for the SF to put the replaced 217 * point sprite coords in. We shouldn't need these dummy slots, 218 * which take up precious URB space, but it would mean that the SF 219 * doesn't get nice aligned pairs of input coords into output 220 * coords, which would be a pain to handle. 221 */ 222 for (i = 0; i < 8; i++) { 223 if (c.key.point_coord_replace & (1 << i)) 224 c.prog_data.outputs_written |= BITFIELD64_BIT(VERT_RESULT_TEX0 + i); 225 } 226 227 if (0) { 228 _mesa_fprint_program_opt(stdout, &c.vp->program.Base, PROG_PRINT_DEBUG, 229 true); 230 } 231 232 /* Emit GEN4 code. 233 */ 234 if (prog) { 235 if (!brw_vs_emit(prog, &c)) { 236 ralloc_free(mem_ctx); 237 return false; 238 } 239 } else { 240 brw_old_vs_emit(&c); 241 } 242 243 /* Scratch space is used for register spilling */ 244 if (c.last_scratch) { 245 c.prog_data.total_scratch = brw_get_scratch_size(c.last_scratch); 246 247 brw_get_scratch_bo(intel, &brw->vs.scratch_bo, 248 c.prog_data.total_scratch * brw->max_vs_threads); 249 } 250 251 /* get the program 252 */ 253 program = brw_get_program(&c.func, &program_size); 254 255 /* We upload from &c.prog_data including the constant_map assuming 256 * they're packed together. It would be nice to have a 257 * compile-time assert macro here. 258 */ 259 assert(c.constant_map == (int8_t *)&c.prog_data + 260 sizeof(c.prog_data)); 261 assert(ctx->Const.VertexProgram.MaxNativeParameters == 262 ARRAY_SIZE(c.constant_map)); 263 (void) ctx; 264 265 aux_size = sizeof(c.prog_data); 266 /* constant_map */ 267 aux_size += c.vp->program.Base.Parameters->NumParameters; 268 269 brw_upload_cache(&brw->cache, BRW_VS_PROG, 270 &c.key, sizeof(c.key), 271 program, program_size, 272 &c.prog_data, aux_size, 273 &brw->vs.prog_offset, &brw->vs.prog_data); 274 ralloc_free(mem_ctx); 275 276 return true; 277} 278 279 280static void brw_upload_vs_prog(struct brw_context *brw) 281{ 282 struct intel_context *intel = &brw->intel; 283 struct gl_context *ctx = &intel->ctx; 284 struct brw_vs_prog_key key; 285 /* BRW_NEW_VERTEX_PROGRAM */ 286 struct brw_vertex_program *vp = 287 (struct brw_vertex_program *)brw->vertex_program; 288 struct gl_program *prog = (struct gl_program *) brw->vertex_program; 289 int i; 290 291 memset(&key, 0, sizeof(key)); 292 293 /* Just upload the program verbatim for now. Always send it all 294 * the inputs it asks for, whether they are varying or not. 295 */ 296 key.program_string_id = vp->id; 297 key.userclip_active = (ctx->Transform.ClipPlanesEnabled != 0); 298 key.uses_clip_distance = vp->program.UsesClipDistance; 299 if (key.userclip_active && !key.uses_clip_distance) { 300 if (intel->gen < 6) { 301 key.nr_userclip_plane_consts 302 = _mesa_bitcount_64(ctx->Transform.ClipPlanesEnabled); 303 key.userclip_planes_enabled_gen_4_5 304 = ctx->Transform.ClipPlanesEnabled; 305 } else { 306 key.nr_userclip_plane_consts 307 = _mesa_logbase2(ctx->Transform.ClipPlanesEnabled) + 1; 308 } 309 } 310 key.copy_edgeflag = (ctx->Polygon.FrontMode != GL_FILL || 311 ctx->Polygon.BackMode != GL_FILL); 312 313 /* _NEW_LIGHT | _NEW_BUFFERS */ 314 key.clamp_vertex_color = ctx->Light._ClampVertexColor; 315 316 /* _NEW_POINT */ 317 if (ctx->Point.PointSprite) { 318 for (i = 0; i < 8; i++) { 319 if (ctx->Point.CoordReplace[i]) 320 key.point_coord_replace |= (1 << i); 321 } 322 } 323 324 /* _NEW_TEXTURE */ 325 for (i = 0; i < BRW_MAX_TEX_UNIT; i++) { 326 if (prog->TexturesUsed[i]) 327 brw_populate_sampler_prog_key_data(ctx, &key.tex, i); 328 } 329 330 /* BRW_NEW_VERTICES */ 331 for (i = 0; i < VERT_ATTRIB_MAX; i++) { 332 if (vp->program.Base.InputsRead & BITFIELD64_BIT(i) && 333 brw->vb.inputs[i].glarray->Type == GL_FIXED) { 334 key.gl_fixed_input_size[i] = brw->vb.inputs[i].glarray->Size; 335 } 336 } 337 338 if (!brw_search_cache(&brw->cache, BRW_VS_PROG, 339 &key, sizeof(key), 340 &brw->vs.prog_offset, &brw->vs.prog_data)) { 341 bool success = do_vs_prog(brw, ctx->Shader.CurrentVertexProgram, 342 vp, &key); 343 344 assert(success); 345 } 346 brw->vs.constant_map = ((int8_t *)brw->vs.prog_data + 347 sizeof(*brw->vs.prog_data)); 348} 349 350/* See brw_vs.c: 351 */ 352const struct brw_tracked_state brw_vs_prog = { 353 .dirty = { 354 .mesa = (_NEW_TRANSFORM | _NEW_POLYGON | _NEW_POINT | _NEW_LIGHT | 355 _NEW_TEXTURE | 356 _NEW_BUFFERS), 357 .brw = (BRW_NEW_VERTEX_PROGRAM | 358 BRW_NEW_VERTICES), 359 .cache = 0 360 }, 361 .emit = brw_upload_vs_prog 362}; 363 364bool 365brw_vs_precompile(struct gl_context *ctx, struct gl_shader_program *prog) 366{ 367 struct brw_context *brw = brw_context(ctx); 368 struct brw_vs_prog_key key; 369 uint32_t old_prog_offset = brw->vs.prog_offset; 370 struct brw_vs_prog_data *old_prog_data = brw->vs.prog_data; 371 bool success; 372 373 if (!prog->_LinkedShaders[MESA_SHADER_VERTEX]) 374 return true; 375 376 struct gl_vertex_program *vp = (struct gl_vertex_program *) 377 prog->_LinkedShaders[MESA_SHADER_VERTEX]->Program; 378 struct brw_vertex_program *bvp = brw_vertex_program(vp); 379 380 memset(&key, 0, sizeof(key)); 381 382 key.program_string_id = bvp->id; 383 key.clamp_vertex_color = true; 384 385 success = do_vs_prog(brw, prog, bvp, &key); 386 387 brw->vs.prog_offset = old_prog_offset; 388 brw->vs.prog_data = old_prog_data; 389 390 return success; 391} 392