i965: Shorten sampler loops in key setup.
[mesa.git] / src / mesa / drivers / dri / i965 / brw_vs.c
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
43 static inline void assign_vue_slot(struct brw_vue_map *vue_map,
44 int varying)
45 {
46 /* Make sure this varying hasn't been assigned a slot already */
47 assert (vue_map->varying_to_slot[varying] == -1);
48
49 vue_map->varying_to_slot[varying] = vue_map->num_slots;
50 vue_map->slot_to_varying[vue_map->num_slots++] = varying;
51 }
52
53 /**
54 * Compute the VUE map for vertex shader program.
55 *
56 * Note that consumers of this map using cache keys must include
57 * prog_data->userclip and prog_data->outputs_written in their key
58 * (generated by CACHE_NEW_VS_PROG).
59 */
60 void
61 brw_compute_vue_map(struct brw_context *brw, struct brw_vue_map *vue_map,
62 GLbitfield64 slots_valid, bool userclip_active)
63 {
64 vue_map->slots_valid = slots_valid;
65 int i;
66
67 /* Make sure that the values we store in vue_map->varying_to_slot and
68 * vue_map->slot_to_varying won't overflow the signed chars that are used
69 * to store them. Note that since vue_map->slot_to_varying sometimes holds
70 * values equal to BRW_VARYING_SLOT_COUNT, we need to ensure that
71 * BRW_VARYING_SLOT_COUNT is <= 127, not 128.
72 */
73 STATIC_ASSERT(BRW_VARYING_SLOT_COUNT <= 127);
74
75 vue_map->num_slots = 0;
76 for (i = 0; i < BRW_VARYING_SLOT_COUNT; ++i) {
77 vue_map->varying_to_slot[i] = -1;
78 vue_map->slot_to_varying[i] = BRW_VARYING_SLOT_COUNT;
79 }
80
81 /* VUE header: format depends on chip generation and whether clipping is
82 * enabled.
83 */
84 switch (brw->gen) {
85 case 4:
86 case 5:
87 /* There are 8 dwords in VUE header pre-Ironlake:
88 * dword 0-3 is indices, point width, clip flags.
89 * dword 4-7 is ndc position
90 * dword 8-11 is the first vertex data.
91 *
92 * On Ironlake the VUE header is nominally 20 dwords, but the hardware
93 * will accept the same header layout as Gen4 [and should be a bit faster]
94 */
95 assign_vue_slot(vue_map, VARYING_SLOT_PSIZ);
96 assign_vue_slot(vue_map, BRW_VARYING_SLOT_NDC);
97 assign_vue_slot(vue_map, VARYING_SLOT_POS);
98 break;
99 case 6:
100 case 7:
101 /* There are 8 or 16 DWs (D0-D15) in VUE header on Sandybridge:
102 * dword 0-3 of the header is indices, point width, clip flags.
103 * dword 4-7 is the 4D space position
104 * dword 8-15 of the vertex header is the user clip distance if
105 * enabled.
106 * dword 8-11 or 16-19 is the first vertex element data we fill.
107 */
108 assign_vue_slot(vue_map, VARYING_SLOT_PSIZ);
109 assign_vue_slot(vue_map, VARYING_SLOT_POS);
110 if (userclip_active) {
111 assign_vue_slot(vue_map, VARYING_SLOT_CLIP_DIST0);
112 assign_vue_slot(vue_map, VARYING_SLOT_CLIP_DIST1);
113 }
114 /* front and back colors need to be consecutive so that we can use
115 * ATTRIBUTE_SWIZZLE_INPUTATTR_FACING to swizzle them when doing
116 * two-sided color.
117 */
118 if (slots_valid & BITFIELD64_BIT(VARYING_SLOT_COL0))
119 assign_vue_slot(vue_map, VARYING_SLOT_COL0);
120 if (slots_valid & BITFIELD64_BIT(VARYING_SLOT_BFC0))
121 assign_vue_slot(vue_map, VARYING_SLOT_BFC0);
122 if (slots_valid & BITFIELD64_BIT(VARYING_SLOT_COL1))
123 assign_vue_slot(vue_map, VARYING_SLOT_COL1);
124 if (slots_valid & BITFIELD64_BIT(VARYING_SLOT_BFC1))
125 assign_vue_slot(vue_map, VARYING_SLOT_BFC1);
126 break;
127 default:
128 assert (!"VUE map not known for this chip generation");
129 break;
130 }
131
132 /* The hardware doesn't care about the rest of the vertex outputs, so just
133 * assign them contiguously. Don't reassign outputs that already have a
134 * slot.
135 *
136 * We generally don't need to assign a slot for VARYING_SLOT_CLIP_VERTEX,
137 * since it's encoded as the clip distances by emit_clip_distances().
138 * However, it may be output by transform feedback, and we'd rather not
139 * recompute state when TF changes, so we just always include it.
140 */
141 for (int i = 0; i < VARYING_SLOT_MAX; ++i) {
142 if ((slots_valid & BITFIELD64_BIT(i)) &&
143 vue_map->varying_to_slot[i] == -1) {
144 assign_vue_slot(vue_map, i);
145 }
146 }
147 }
148
149
150 /**
151 * Decide which set of clip planes should be used when clipping via
152 * gl_Position or gl_ClipVertex.
153 */
154 gl_clip_plane *brw_select_clip_planes(struct gl_context *ctx)
155 {
156 if (ctx->Shader.CurrentVertexProgram) {
157 /* There is currently a GLSL vertex shader, so clip according to GLSL
158 * rules, which means compare gl_ClipVertex (or gl_Position, if
159 * gl_ClipVertex wasn't assigned) against the eye-coordinate clip planes
160 * that were stored in EyeUserPlane at the time the clip planes were
161 * specified.
162 */
163 return ctx->Transform.EyeUserPlane;
164 } else {
165 /* Either we are using fixed function or an ARB vertex program. In
166 * either case the clip planes are going to be compared against
167 * gl_Position (which is in clip coordinates) so we have to clip using
168 * _ClipUserPlane, which was transformed into clip coordinates by Mesa
169 * core.
170 */
171 return ctx->Transform._ClipUserPlane;
172 }
173 }
174
175
176 bool
177 brw_vec4_prog_data_compare(const struct brw_vec4_prog_data *a,
178 const struct brw_vec4_prog_data *b)
179 {
180 /* Compare all the struct up to the pointers. */
181 if (memcmp(a, b, offsetof(struct brw_vec4_prog_data, param)))
182 return false;
183
184 if (memcmp(a->param, b->param, a->nr_params * sizeof(void *)))
185 return false;
186
187 if (memcmp(a->pull_param, b->pull_param, a->nr_pull_params * sizeof(void *)))
188 return false;
189
190 return true;
191 }
192
193
194 bool
195 brw_vs_prog_data_compare(const void *in_a, const void *in_b,
196 int aux_size, const void *in_key)
197 {
198 const struct brw_vs_prog_data *a = in_a;
199 const struct brw_vs_prog_data *b = in_b;
200
201 /* Compare the base vec4 structure. */
202 if (!brw_vec4_prog_data_compare(&a->base, &b->base))
203 return false;
204
205 /* Compare the rest of the struct. */
206 const unsigned offset = sizeof(struct brw_vec4_prog_data);
207 if (memcmp(((char *) &a) + offset, ((char *) &b) + offset,
208 sizeof(struct brw_vs_prog_data) - offset)) {
209 return false;
210 }
211
212 return true;
213 }
214
215 static bool
216 do_vs_prog(struct brw_context *brw,
217 struct gl_shader_program *prog,
218 struct brw_vertex_program *vp,
219 struct brw_vs_prog_key *key)
220 {
221 GLuint program_size;
222 const GLuint *program;
223 struct brw_vs_compile c;
224 struct brw_vs_prog_data prog_data;
225 void *mem_ctx;
226 int i;
227 struct gl_shader *vs = NULL;
228
229 if (prog)
230 vs = prog->_LinkedShaders[MESA_SHADER_VERTEX];
231
232 memset(&c, 0, sizeof(c));
233 memcpy(&c.key, key, sizeof(*key));
234 memset(&prog_data, 0, sizeof(prog_data));
235
236 mem_ctx = ralloc_context(NULL);
237
238 c.vp = vp;
239
240 /* Allocate the references to the uniforms that will end up in the
241 * prog_data associated with the compiled program, and which will be freed
242 * by the state cache.
243 */
244 int param_count;
245 if (vs) {
246 /* We add padding around uniform values below vec4 size, with the worst
247 * case being a float value that gets blown up to a vec4, so be
248 * conservative here.
249 */
250 param_count = vs->num_uniform_components * 4;
251
252 } else {
253 param_count = vp->program.Base.Parameters->NumParameters * 4;
254 }
255 /* We also upload clip plane data as uniforms */
256 param_count += MAX_CLIP_PLANES * 4;
257
258 prog_data.base.param = rzalloc_array(NULL, const float *, param_count);
259 prog_data.base.pull_param = rzalloc_array(NULL, const float *, param_count);
260
261 GLbitfield64 outputs_written = vp->program.Base.OutputsWritten;
262 prog_data.inputs_read = vp->program.Base.InputsRead;
263
264 if (c.key.copy_edgeflag) {
265 outputs_written |= BITFIELD64_BIT(VARYING_SLOT_EDGE);
266 prog_data.inputs_read |= VERT_BIT_EDGEFLAG;
267 }
268
269 if (brw->gen < 6) {
270 /* Put dummy slots into the VUE for the SF to put the replaced
271 * point sprite coords in. We shouldn't need these dummy slots,
272 * which take up precious URB space, but it would mean that the SF
273 * doesn't get nice aligned pairs of input coords into output
274 * coords, which would be a pain to handle.
275 */
276 for (i = 0; i < 8; i++) {
277 if (c.key.point_coord_replace & (1 << i))
278 outputs_written |= BITFIELD64_BIT(VARYING_SLOT_TEX0 + i);
279 }
280
281 /* if back colors are written, allocate slots for front colors too */
282 if (outputs_written & BITFIELD64_BIT(VARYING_SLOT_BFC0))
283 outputs_written |= BITFIELD64_BIT(VARYING_SLOT_COL0);
284 if (outputs_written & BITFIELD64_BIT(VARYING_SLOT_BFC1))
285 outputs_written |= BITFIELD64_BIT(VARYING_SLOT_COL1);
286
287 if (c.key.base.userclip_active) {
288 outputs_written |= BITFIELD64_BIT(VARYING_SLOT_CLIP_DIST0);
289 outputs_written |= BITFIELD64_BIT(VARYING_SLOT_CLIP_DIST1);
290 }
291 }
292
293 brw_compute_vue_map(brw, &prog_data.base.vue_map, outputs_written,
294 c.key.base.userclip_active);
295
296 if (0) {
297 _mesa_fprint_program_opt(stdout, &c.vp->program.Base, PROG_PRINT_DEBUG,
298 true);
299 }
300
301 /* Emit GEN4 code.
302 */
303 program = brw_vs_emit(brw, prog, &c, &prog_data, mem_ctx, &program_size);
304 if (program == NULL) {
305 ralloc_free(mem_ctx);
306 return false;
307 }
308
309 if (prog_data.base.nr_pull_params)
310 prog_data.base.num_surfaces = 1;
311 if (c.vp->program.Base.SamplersUsed)
312 prog_data.base.num_surfaces = SURF_INDEX_VS_TEXTURE(BRW_MAX_TEX_UNIT);
313 if (prog &&
314 prog->_LinkedShaders[MESA_SHADER_VERTEX]->NumUniformBlocks) {
315 prog_data.base.num_surfaces =
316 SURF_INDEX_VS_UBO(prog->_LinkedShaders[MESA_SHADER_VERTEX]->NumUniformBlocks);
317 }
318
319 /* Scratch space is used for register spilling */
320 if (c.base.last_scratch) {
321 perf_debug("Vertex shader triggered register spilling. "
322 "Try reducing the number of live vec4 values to "
323 "improve performance.\n");
324
325 prog_data.base.total_scratch
326 = brw_get_scratch_size(c.base.last_scratch*REG_SIZE);
327
328 brw_get_scratch_bo(brw, &brw->vs.scratch_bo,
329 prog_data.base.total_scratch * brw->max_vs_threads);
330 }
331
332 brw_upload_cache(&brw->cache, BRW_VS_PROG,
333 &c.key, sizeof(c.key),
334 program, program_size,
335 &prog_data, sizeof(prog_data),
336 &brw->vs.prog_offset, &brw->vs.prog_data);
337 ralloc_free(mem_ctx);
338
339 return true;
340 }
341
342 static bool
343 key_debug(struct brw_context *brw, const char *name, int a, int b)
344 {
345 if (a != b) {
346 perf_debug(" %s %d->%d\n", name, a, b);
347 return true;
348 }
349 return false;
350 }
351
352 void
353 brw_vs_debug_recompile(struct brw_context *brw,
354 struct gl_shader_program *prog,
355 const struct brw_vs_prog_key *key)
356 {
357 struct brw_cache_item *c = NULL;
358 const struct brw_vs_prog_key *old_key = NULL;
359 bool found = false;
360
361 perf_debug("Recompiling vertex shader for program %d\n", prog->Name);
362
363 for (unsigned int i = 0; i < brw->cache.size; i++) {
364 for (c = brw->cache.items[i]; c; c = c->next) {
365 if (c->cache_id == BRW_VS_PROG) {
366 old_key = c->key;
367
368 if (old_key->base.program_string_id == key->base.program_string_id)
369 break;
370 }
371 }
372 if (c)
373 break;
374 }
375
376 if (!c) {
377 perf_debug(" Didn't find previous compile in the shader cache for "
378 "debug\n");
379 return;
380 }
381
382 for (unsigned int i = 0; i < VERT_ATTRIB_MAX; i++) {
383 found |= key_debug(brw, "Vertex attrib w/a flags",
384 old_key->gl_attrib_wa_flags[i],
385 key->gl_attrib_wa_flags[i]);
386 }
387
388 found |= key_debug(brw, "user clip flags",
389 old_key->base.userclip_active, key->base.userclip_active);
390
391 found |= key_debug(brw, "user clipping planes as push constants",
392 old_key->base.nr_userclip_plane_consts,
393 key->base.nr_userclip_plane_consts);
394
395 found |= key_debug(brw, "clip distance enable",
396 old_key->base.uses_clip_distance, key->base.uses_clip_distance);
397 found |= key_debug(brw, "copy edgeflag",
398 old_key->copy_edgeflag, key->copy_edgeflag);
399 found |= key_debug(brw, "PointCoord replace",
400 old_key->point_coord_replace, key->point_coord_replace);
401 found |= key_debug(brw, "vertex color clamping",
402 old_key->base.clamp_vertex_color, key->base.clamp_vertex_color);
403
404 found |= brw_debug_recompile_sampler_key(brw, &old_key->base.tex,
405 &key->base.tex);
406
407 if (!found) {
408 perf_debug(" Something else\n");
409 }
410 }
411
412 static void brw_upload_vs_prog(struct brw_context *brw)
413 {
414 struct gl_context *ctx = &brw->ctx;
415 struct brw_vs_prog_key key;
416 /* BRW_NEW_VERTEX_PROGRAM */
417 struct brw_vertex_program *vp =
418 (struct brw_vertex_program *)brw->vertex_program;
419 struct gl_program *prog = (struct gl_program *) brw->vertex_program;
420 int i;
421
422 memset(&key, 0, sizeof(key));
423
424 /* Just upload the program verbatim for now. Always send it all
425 * the inputs it asks for, whether they are varying or not.
426 */
427 key.base.program_string_id = vp->id;
428 key.base.userclip_active = (ctx->Transform.ClipPlanesEnabled != 0);
429 key.base.uses_clip_distance = vp->program.UsesClipDistance;
430 if (key.base.userclip_active && !key.base.uses_clip_distance) {
431 key.base.nr_userclip_plane_consts
432 = _mesa_logbase2(ctx->Transform.ClipPlanesEnabled) + 1;
433 }
434
435 /* _NEW_POLYGON */
436 if (brw->gen < 6) {
437 key.copy_edgeflag = (ctx->Polygon.FrontMode != GL_FILL ||
438 ctx->Polygon.BackMode != GL_FILL);
439 }
440
441 /* _NEW_LIGHT | _NEW_BUFFERS */
442 key.base.clamp_vertex_color = ctx->Light._ClampVertexColor;
443
444 /* _NEW_POINT */
445 if (brw->gen < 6 && ctx->Point.PointSprite) {
446 for (i = 0; i < 8; i++) {
447 if (ctx->Point.CoordReplace[i])
448 key.point_coord_replace |= (1 << i);
449 }
450 }
451
452 /* _NEW_TEXTURE */
453 brw_populate_sampler_prog_key_data(ctx, prog, brw->vs.sampler_count,
454 &key.base.tex);
455
456 /* BRW_NEW_VERTICES */
457 if (brw->gen < 8 && !brw->is_haswell) {
458 /* Prior to Haswell, the hardware can't natively support GL_FIXED or
459 * 2_10_10_10_REV vertex formats. Set appropriate workaround flags.
460 */
461 for (i = 0; i < VERT_ATTRIB_MAX; i++) {
462 if (!(vp->program.Base.InputsRead & BITFIELD64_BIT(i)))
463 continue;
464
465 uint8_t wa_flags = 0;
466
467 switch (brw->vb.inputs[i].glarray->Type) {
468
469 case GL_FIXED:
470 wa_flags = brw->vb.inputs[i].glarray->Size;
471 break;
472
473 case GL_INT_2_10_10_10_REV:
474 wa_flags |= BRW_ATTRIB_WA_SIGN;
475 /* fallthough */
476
477 case GL_UNSIGNED_INT_2_10_10_10_REV:
478 if (brw->vb.inputs[i].glarray->Format == GL_BGRA)
479 wa_flags |= BRW_ATTRIB_WA_BGRA;
480
481 if (brw->vb.inputs[i].glarray->Normalized)
482 wa_flags |= BRW_ATTRIB_WA_NORMALIZE;
483 else if (!brw->vb.inputs[i].glarray->Integer)
484 wa_flags |= BRW_ATTRIB_WA_SCALE;
485
486 break;
487 }
488
489 key.gl_attrib_wa_flags[i] = wa_flags;
490 }
491 }
492
493 if (!brw_search_cache(&brw->cache, BRW_VS_PROG,
494 &key, sizeof(key),
495 &brw->vs.prog_offset, &brw->vs.prog_data)) {
496 bool success = do_vs_prog(brw, ctx->Shader.CurrentVertexProgram,
497 vp, &key);
498
499 assert(success);
500 }
501 if (memcmp(&brw->vs.prog_data->base.vue_map, &brw->vue_map_geom_out,
502 sizeof(brw->vue_map_geom_out)) != 0) {
503 brw->vue_map_geom_out = brw->vs.prog_data->base.vue_map;
504 brw->state.dirty.brw |= BRW_NEW_VUE_MAP_GEOM_OUT;
505 }
506 }
507
508 /* See brw_vs.c:
509 */
510 const struct brw_tracked_state brw_vs_prog = {
511 .dirty = {
512 .mesa = (_NEW_TRANSFORM | _NEW_POLYGON | _NEW_POINT | _NEW_LIGHT |
513 _NEW_TEXTURE |
514 _NEW_BUFFERS),
515 .brw = (BRW_NEW_VERTEX_PROGRAM |
516 BRW_NEW_VERTICES),
517 .cache = 0
518 },
519 .emit = brw_upload_vs_prog
520 };
521
522 bool
523 brw_vs_precompile(struct gl_context *ctx, struct gl_shader_program *prog)
524 {
525 struct brw_context *brw = brw_context(ctx);
526 struct brw_vs_prog_key key;
527 uint32_t old_prog_offset = brw->vs.prog_offset;
528 struct brw_vs_prog_data *old_prog_data = brw->vs.prog_data;
529 bool success;
530
531 if (!prog->_LinkedShaders[MESA_SHADER_VERTEX])
532 return true;
533
534 struct gl_vertex_program *vp = (struct gl_vertex_program *)
535 prog->_LinkedShaders[MESA_SHADER_VERTEX]->Program;
536 struct brw_vertex_program *bvp = brw_vertex_program(vp);
537
538 memset(&key, 0, sizeof(key));
539
540 key.base.program_string_id = bvp->id;
541 key.base.clamp_vertex_color = ctx->API == API_OPENGL_COMPAT;
542
543 for (int i = 0; i < MAX_SAMPLERS; i++) {
544 if (vp->Base.ShadowSamplers & (1 << i)) {
545 /* Assume DEPTH_TEXTURE_MODE is the default: X, X, X, 1 */
546 key.base.tex.swizzles[i] =
547 MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_X, SWIZZLE_X, SWIZZLE_ONE);
548 } else {
549 /* Color sampler: assume no swizzling. */
550 key.base.tex.swizzles[i] = SWIZZLE_XYZW;
551 }
552 }
553
554 success = do_vs_prog(brw, prog, bvp, &key);
555
556 brw->vs.prog_offset = old_prog_offset;
557 brw->vs.prog_data = old_prog_data;
558
559 return success;
560 }
561
562
563 void
564 brw_vec4_prog_data_free(const struct brw_vec4_prog_data *prog_data)
565 {
566 ralloc_free((void *)prog_data->param);
567 ralloc_free((void *)prog_data->pull_param);
568 }
569
570
571 void
572 brw_vs_prog_data_free(const void *in_prog_data)
573 {
574 const struct brw_vs_prog_data *prog_data = in_prog_data;
575
576 brw_vec4_prog_data_free(&prog_data->base);
577 }