projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
v3d: support rendering to multi-layered framebuffers
[mesa.git] / src / gallium / drivers / v3d / v3d_program.c
1 /*
2 * Copyright © 2014-2017 Broadcom
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24 #include <inttypes.h>
25 #include "util/format/u_format.h"
26 #include "util/u_math.h"
27 #include "util/u_memory.h"
28 #include "util/ralloc.h"
29 #include "util/hash_table.h"
30 #include "util/u_upload_mgr.h"
31 #include "tgsi/tgsi_dump.h"
32 #include "tgsi/tgsi_parse.h"
33 #include "compiler/nir/nir.h"
34 #include "compiler/nir/nir_builder.h"
35 #include "nir/tgsi_to_nir.h"
36 #include "compiler/v3d_compiler.h"
37 #include "v3d_context.h"
38 #include "broadcom/cle/v3d_packet_v33_pack.h"
39
40 static struct v3d_compiled_shader *
41 v3d_get_compiled_shader(struct v3d_context *v3d,
42 struct v3d_key *key, size_t key_size);
43 static void
44 v3d_setup_shared_precompile_key(struct v3d_uncompiled_shader *uncompiled,
45 struct v3d_key *key);
46
47 static gl_varying_slot
48 v3d_get_slot_for_driver_location(nir_shader *s, uint32_t driver_location)
49 {
50 nir_foreach_variable(var, &s->outputs) {
51 if (var->data.driver_location == driver_location) {
52 return var->data.location;
53 }
54 }
55
56 return -1;
57 }
58
59 /**
60 * Precomputes the TRANSFORM_FEEDBACK_OUTPUT_DATA_SPEC array for the shader.
61 *
62 * A shader can have 16 of these specs, and each one of them can write up to
63 * 16 dwords. Since we allow a total of 64 transform feedback output
64 * components (not 16 vectors), we have to group the writes of multiple
65 * varyings together in a single data spec.
66 */
67 static void
68 v3d_set_transform_feedback_outputs(struct v3d_uncompiled_shader *so,
69 const struct pipe_stream_output_info *stream_output)
70 {
71 if (!stream_output->num_outputs)
72 return;
73
74 struct v3d_varying_slot slots[PIPE_MAX_SO_OUTPUTS * 4];
75 int slot_count = 0;
76
77 for (int buffer = 0; buffer < PIPE_MAX_SO_BUFFERS; buffer++) {
78 uint32_t buffer_offset = 0;
79 uint32_t vpm_start = slot_count;
80
81 for (int i = 0; i < stream_output->num_outputs; i++) {
82 const struct pipe_stream_output *output =
83 &stream_output->output[i];
84
85 if (output->output_buffer != buffer)
86 continue;
87
88 /* We assume that the SO outputs appear in increasing
89 * order in the buffer.
90 */
91 assert(output->dst_offset >= buffer_offset);
92
93 /* Pad any undefined slots in the output */
94 for (int j = buffer_offset; j < output->dst_offset; j++) {
95 slots[slot_count] =
96 v3d_slot_from_slot_and_component(VARYING_SLOT_POS, 0);
97 slot_count++;
98 buffer_offset++;
99 }
100
101 /* Set the coordinate shader up to output the
102 * components of this varying.
103 */
104 for (int j = 0; j < output->num_components; j++) {
105 gl_varying_slot slot =
106 v3d_get_slot_for_driver_location(so->base.ir.nir, output->register_index);
107
108 slots[slot_count] =
109 v3d_slot_from_slot_and_component(slot,
110 output->start_component + j);
111 slot_count++;
112 buffer_offset++;
113 }
114 }
115
116 uint32_t vpm_size = slot_count - vpm_start;
117 if (!vpm_size)
118 continue;
119
120 uint32_t vpm_start_offset = vpm_start + 6;
121
122 while (vpm_size) {
123 uint32_t write_size = MIN2(vpm_size, 1 << 4);
124
125 struct V3D33_TRANSFORM_FEEDBACK_OUTPUT_DATA_SPEC unpacked = {
126 /* We need the offset from the coordinate shader's VPM
127 * output block, which has the [X, Y, Z, W, Xs, Ys]
128 * values at the start.
129 */
130 .first_shaded_vertex_value_to_output = vpm_start_offset,
131 .number_of_consecutive_vertex_values_to_output_as_32_bit_values = write_size,
132 .output_buffer_to_write_to = buffer,
133 };
134
135 /* GFXH-1559 */
136 assert(unpacked.first_shaded_vertex_value_to_output != 8 ||
137 so->num_tf_specs != 0);
138
139 assert(so->num_tf_specs != ARRAY_SIZE(so->tf_specs));
140 V3D33_TRANSFORM_FEEDBACK_OUTPUT_DATA_SPEC_pack(NULL,
141 (void *)&so->tf_specs[so->num_tf_specs],
142 &unpacked);
143
144 /* If point size is being written by the shader, then
145 * all the VPM start offsets are shifted up by one.
146 * We won't know that until the variant is compiled,
147 * though.
148 */
149 unpacked.first_shaded_vertex_value_to_output++;
150
151 /* GFXH-1559 */
152 assert(unpacked.first_shaded_vertex_value_to_output != 8 ||
153 so->num_tf_specs != 0);
154
155 V3D33_TRANSFORM_FEEDBACK_OUTPUT_DATA_SPEC_pack(NULL,
156 (void *)&so->tf_specs_psiz[so->num_tf_specs],
157 &unpacked);
158 so->num_tf_specs++;
159 vpm_start_offset += write_size;
160 vpm_size -= write_size;
161 }
162 so->base.stream_output.stride[buffer] =
163 stream_output->stride[buffer];
164 }
165
166 so->num_tf_outputs = slot_count;
167 so->tf_outputs = ralloc_array(so->base.ir.nir, struct v3d_varying_slot,
168 slot_count);
169 memcpy(so->tf_outputs, slots, sizeof(*slots) * slot_count);
170 }
171
172 static int
173 type_size(const struct glsl_type *type, bool bindless)
174 {
175 return glsl_count_attribute_slots(type, false);
176 }
177
178 /**
179 * Precompiles a shader variant at shader state creation time if
180 * V3D_DEBUG=precompile is set. Used for shader-db
181 * (https://gitlab.freedesktop.org/mesa/shader-db)
182 */
183 static void
184 v3d_shader_precompile(struct v3d_context *v3d,
185 struct v3d_uncompiled_shader *so)
186 {
187 nir_shader *s = so->base.ir.nir;
188
189 if (s->info.stage == MESA_SHADER_FRAGMENT) {
190 struct v3d_fs_key key = {
191 .base.shader_state = so,
192 };
193
194 nir_foreach_variable(var, &s->outputs) {
195 if (var->data.location == FRAG_RESULT_COLOR) {
196 key.cbufs |= 1 << 0;
197 } else if (var->data.location >= FRAG_RESULT_DATA0) {
198 key.cbufs |= 1 << (var->data.location -
199 FRAG_RESULT_DATA0);
200 }
201 }
202
203 key.logicop_func = PIPE_LOGICOP_COPY;
204
205 v3d_setup_shared_precompile_key(so, &key.base);
206 v3d_get_compiled_shader(v3d, &key.base, sizeof(key));
207 } else {
208 /* FIXME: add geometry shaders */
209
210 struct v3d_vs_key key = {
211 .base.shader_state = so,
212 /* Emit fixed function outputs */
213 .base.is_last_geometry_stage = true,
214 };
215
216 v3d_setup_shared_precompile_key(so, &key.base);
217
218 /* Compile VS: All outputs */
219 nir_foreach_variable(var, &s->outputs) {
220 unsigned array_len = MAX2(glsl_get_length(var->type), 1);
221 assert(array_len == 1);
222 (void)array_len;
223
224 int slot = var->data.location;
225 for (int i = 0; i < glsl_get_components(var->type); i++) {
226 int swiz = var->data.location_frac + i;
227 key.used_outputs[key.num_used_outputs++] =
228 v3d_slot_from_slot_and_component(slot,
229 swiz);
230 }
231 }
232
233 v3d_get_compiled_shader(v3d, &key.base, sizeof(key));
234
235 /* Compile VS bin shader: only position (XXX: include TF) */
236 key.is_coord = true;
237 key.num_used_outputs = 0;
238 for (int i = 0; i < 4; i++) {
239 key.used_outputs[key.num_used_outputs++] =
240 v3d_slot_from_slot_and_component(VARYING_SLOT_POS,
241 i);
242 }
243 v3d_get_compiled_shader(v3d, &key.base, sizeof(key));
244 }
245 }
246
247 static void *
248 v3d_uncompiled_shader_create(struct pipe_context *pctx,
249 enum pipe_shader_ir type, void *ir)
250 {
251 struct v3d_context *v3d = v3d_context(pctx);
252 struct v3d_uncompiled_shader *so = CALLOC_STRUCT(v3d_uncompiled_shader);
253 if (!so)
254 return NULL;
255
256 so->program_id = v3d->next_uncompiled_program_id++;
257
258 nir_shader *s;
259
260 if (type == PIPE_SHADER_IR_NIR) {
261 /* The backend takes ownership of the NIR shader on state
262 * creation.
263 */
264 s = ir;
265 } else {
266 assert(type == PIPE_SHADER_IR_TGSI);
267
268 if (V3D_DEBUG & V3D_DEBUG_TGSI) {
269 fprintf(stderr, "prog %d TGSI:\n",
270 so->program_id);
271 tgsi_dump(ir, 0);
272 fprintf(stderr, "\n");
273 }
274 s = tgsi_to_nir(ir, pctx->screen);
275 }
276
277 nir_variable_mode lower_mode = nir_var_all & ~nir_var_uniform;
278 if (s->info.stage == MESA_SHADER_VERTEX ||
279 s->info.stage == MESA_SHADER_GEOMETRY) {
280 lower_mode &= ~(nir_var_shader_in | nir_var_shader_out);
281 }
282 NIR_PASS_V(s, nir_lower_io, lower_mode,
283 type_size,
284 (nir_lower_io_options)0);
285
286 NIR_PASS_V(s, nir_lower_regs_to_ssa);
287 NIR_PASS_V(s, nir_normalize_cubemap_coords);
288
289 NIR_PASS_V(s, nir_lower_load_const_to_scalar);
290
291 v3d_optimize_nir(s);
292
293 NIR_PASS_V(s, nir_remove_dead_variables, nir_var_function_temp);
294
295 /* Garbage collect dead instructions */
296 nir_sweep(s);
297
298 so->base.type = PIPE_SHADER_IR_NIR;
299 so->base.ir.nir = s;
300
301 if (V3D_DEBUG & (V3D_DEBUG_NIR |
302 v3d_debug_flag_for_shader_stage(s->info.stage))) {
303 fprintf(stderr, "%s prog %d NIR:\n",
304 gl_shader_stage_name(s->info.stage),
305 so->program_id);
306 nir_print_shader(s, stderr);
307 fprintf(stderr, "\n");
308 }
309
310 if (V3D_DEBUG & V3D_DEBUG_PRECOMPILE)
311 v3d_shader_precompile(v3d, so);
312
313 return so;
314 }
315
316 static void
317 v3d_shader_debug_output(const char *message, void *data)
318 {
319 struct v3d_context *v3d = data;
320
321 pipe_debug_message(&v3d->debug, SHADER_INFO, "%s", message);
322 }
323
324 static void *
325 v3d_shader_state_create(struct pipe_context *pctx,
326 const struct pipe_shader_state *cso)
327 {
328 struct v3d_uncompiled_shader *so =
329 v3d_uncompiled_shader_create(pctx,
330 cso->type,
331 (cso->type == PIPE_SHADER_IR_TGSI ?
332 (void *)cso->tokens :
333 cso->ir.nir));
334
335 v3d_set_transform_feedback_outputs(so, &cso->stream_output);
336
337 return so;
338 }
339
340 struct v3d_compiled_shader *
341 v3d_get_compiled_shader(struct v3d_context *v3d,
342 struct v3d_key *key,
343 size_t key_size)
344 {
345 struct v3d_uncompiled_shader *shader_state = key->shader_state;
346 nir_shader *s = shader_state->base.ir.nir;
347
348 struct hash_table *ht = v3d->prog.cache[s->info.stage];
349 struct hash_entry *entry = _mesa_hash_table_search(ht, key);
350 if (entry)
351 return entry->data;
352
353 struct v3d_compiled_shader *shader =
354 rzalloc(NULL, struct v3d_compiled_shader);
355
356 int program_id = shader_state->program_id;
357 int variant_id =
358 p_atomic_inc_return(&shader_state->compiled_variant_count);
359 uint64_t *qpu_insts;
360 uint32_t shader_size;
361
362 qpu_insts = v3d_compile(v3d->screen->compiler, key,
363 &shader->prog_data.base, s,
364 v3d_shader_debug_output,
365 v3d,
366 program_id, variant_id, &shader_size);
367 ralloc_steal(shader, shader->prog_data.base);
368
369 v3d_set_shader_uniform_dirty_flags(shader);
370
371 if (shader_size) {
372 u_upload_data(v3d->state_uploader, 0, shader_size, 8,
373 qpu_insts, &shader->offset, &shader->resource);
374 }
375
376 free(qpu_insts);
377
378 if (ht) {
379 struct v3d_key *dup_key;
380 dup_key = ralloc_size(shader, key_size);
381 memcpy(dup_key, key, key_size);
382 _mesa_hash_table_insert(ht, dup_key, shader);
383 }
384
385 if (shader->prog_data.base->spill_size >
386 v3d->prog.spill_size_per_thread) {
387 /* The TIDX register we use for choosing the area to access
388 * for scratch space is: (core << 6) | (qpu << 2) | thread.
389 * Even at minimum threadcount in a particular shader, that
390 * means we still multiply by qpus by 4.
391 */
392 int total_spill_size = (v3d->screen->devinfo.qpu_count * 4 *
393 shader->prog_data.base->spill_size);
394
395 v3d_bo_unreference(&v3d->prog.spill_bo);
396 v3d->prog.spill_bo = v3d_bo_alloc(v3d->screen,
397 total_spill_size, "spill");
398 v3d->prog.spill_size_per_thread =
399 shader->prog_data.base->spill_size;
400 }
401
402 return shader;
403 }
404
405 static void
406 v3d_free_compiled_shader(struct v3d_compiled_shader *shader)
407 {
408 pipe_resource_reference(&shader->resource, NULL);
409 ralloc_free(shader);
410 }
411
412 static void
413 v3d_setup_shared_key(struct v3d_context *v3d, struct v3d_key *key,
414 struct v3d_texture_stateobj *texstate)
415 {
416 const struct v3d_device_info *devinfo = &v3d->screen->devinfo;
417
418 for (int i = 0; i < texstate->num_textures; i++) {
419 struct pipe_sampler_view *sampler = texstate->textures[i];
420 struct v3d_sampler_view *v3d_sampler = v3d_sampler_view(sampler);
421 struct pipe_sampler_state *sampler_state =
422 texstate->samplers[i];
423
424 if (!sampler)
425 continue;
426
427 key->tex[i].return_size =
428 v3d_get_tex_return_size(devinfo,
429 sampler->format,
430 sampler_state->compare_mode);
431
432 /* For 16-bit, we set up the sampler to always return 2
433 * channels (meaning no recompiles for most statechanges),
434 * while for 32 we actually scale the returns with channels.
435 */
436 if (key->tex[i].return_size == 16) {
437 key->tex[i].return_channels = 2;
438 } else if (devinfo->ver > 40) {
439 key->tex[i].return_channels = 4;
440 } else {
441 key->tex[i].return_channels =
442 v3d_get_tex_return_channels(devinfo,
443 sampler->format);
444 }
445
446 if (key->tex[i].return_size == 32 && devinfo->ver < 40) {
447 memcpy(key->tex[i].swizzle,
448 v3d_sampler->swizzle,
449 sizeof(v3d_sampler->swizzle));
450 } else {
451 /* For 16-bit returns, we let the sampler state handle
452 * the swizzle.
453 */
454 key->tex[i].swizzle[0] = PIPE_SWIZZLE_X;
455 key->tex[i].swizzle[1] = PIPE_SWIZZLE_Y;
456 key->tex[i].swizzle[2] = PIPE_SWIZZLE_Z;
457 key->tex[i].swizzle[3] = PIPE_SWIZZLE_W;
458 }
459
460 if (sampler) {
461 key->tex[i].clamp_s =
462 sampler_state->wrap_s == PIPE_TEX_WRAP_CLAMP;
463 key->tex[i].clamp_t =
464 sampler_state->wrap_t == PIPE_TEX_WRAP_CLAMP;
465 key->tex[i].clamp_r =
466 sampler_state->wrap_r == PIPE_TEX_WRAP_CLAMP;
467 }
468 }
469 }
470
471 static void
472 v3d_setup_shared_precompile_key(struct v3d_uncompiled_shader *uncompiled,
473 struct v3d_key *key)
474 {
475 nir_shader *s = uncompiled->base.ir.nir;
476
477 for (int i = 0; i < s->info.num_textures; i++) {
478 key->tex[i].return_size = 16;
479 key->tex[i].return_channels = 2;
480
481 key->tex[i].swizzle[0] = PIPE_SWIZZLE_X;
482 key->tex[i].swizzle[1] = PIPE_SWIZZLE_Y;
483 key->tex[i].swizzle[2] = PIPE_SWIZZLE_Z;
484 key->tex[i].swizzle[3] = PIPE_SWIZZLE_W;
485 }
486 }
487
488 static void
489 v3d_update_compiled_fs(struct v3d_context *v3d, uint8_t prim_mode)
490 {
491 struct v3d_job *job = v3d->job;
492 struct v3d_fs_key local_key;
493 struct v3d_fs_key *key = &local_key;
494 nir_shader *s = v3d->prog.bind_fs->base.ir.nir;
495
496 if (!(v3d->dirty & (VC5_DIRTY_PRIM_MODE |
497 VC5_DIRTY_BLEND |
498 VC5_DIRTY_FRAMEBUFFER |
499 VC5_DIRTY_ZSA |
500 VC5_DIRTY_RASTERIZER |
501 VC5_DIRTY_SAMPLE_STATE |
502 VC5_DIRTY_FRAGTEX |
503 VC5_DIRTY_UNCOMPILED_FS))) {
504 return;
505 }
506
507 memset(key, 0, sizeof(*key));
508 v3d_setup_shared_key(v3d, &key->base, &v3d->tex[PIPE_SHADER_FRAGMENT]);
509 key->base.shader_state = v3d->prog.bind_fs;
510 key->base.ucp_enables = v3d->rasterizer->base.clip_plane_enable;
511 key->is_points = (prim_mode == PIPE_PRIM_POINTS);
512 key->is_lines = (prim_mode >= PIPE_PRIM_LINES &&
513 prim_mode <= PIPE_PRIM_LINE_STRIP);
514 key->clamp_color = v3d->rasterizer->base.clamp_fragment_color;
515 if (v3d->blend->base.logicop_enable) {
516 key->logicop_func = v3d->blend->base.logicop_func;
517 } else {
518 key->logicop_func = PIPE_LOGICOP_COPY;
519 }
520 if (job->msaa) {
521 key->msaa = v3d->rasterizer->base.multisample;
522 key->sample_coverage = (v3d->rasterizer->base.multisample &&
523 v3d->sample_mask != (1 << V3D_MAX_SAMPLES) - 1);
524 key->sample_alpha_to_coverage = v3d->blend->base.alpha_to_coverage;
525 key->sample_alpha_to_one = v3d->blend->base.alpha_to_one;
526 }
527
528 key->depth_enabled = (v3d->zsa->base.depth.enabled ||
529 v3d->zsa->base.stencil[0].enabled);
530 if (v3d->zsa->base.alpha.enabled) {
531 key->alpha_test = true;
532 key->alpha_test_func = v3d->zsa->base.alpha.func;
533 }
534
535 key->swap_color_rb = v3d->swap_color_rb;
536
537 for (int i = 0; i < v3d->framebuffer.nr_cbufs; i++) {
538 struct pipe_surface *cbuf = v3d->framebuffer.cbufs[i];
539 if (!cbuf)
540 continue;
541
542 /* gl_FragColor's propagation to however many bound color
543 * buffers there are means that the shader compile needs to
544 * know what buffers are present.
545 */
546 key->cbufs |= 1 << i;
547
548 /* If logic operations are enabled then we might emit color
549 * reads and we need to know the color buffer format and
550 * swizzle for that.
551 */
552 if (key->logicop_func != PIPE_LOGICOP_COPY) {
553 key->color_fmt[i].format = cbuf->format;
554 key->color_fmt[i].swizzle =
555 v3d_get_format_swizzle(&v3d->screen->devinfo,
556 cbuf->format);
557 }
558
559 const struct util_format_description *desc =
560 util_format_description(cbuf->format);
561
562 if (desc->channel[0].type == UTIL_FORMAT_TYPE_FLOAT &&
563 desc->channel[0].size == 32) {
564 key->f32_color_rb |= 1 << i;
565 }
566
567 if (s->info.fs.untyped_color_outputs) {
568 if (util_format_is_pure_uint(cbuf->format))
569 key->uint_color_rb |= 1 << i;
570 else if (util_format_is_pure_sint(cbuf->format))
571 key->int_color_rb |= 1 << i;
572 }
573 }
574
575 if (key->is_points) {
576 key->point_sprite_mask =
577 v3d->rasterizer->base.sprite_coord_enable;
578 key->point_coord_upper_left =
579 (v3d->rasterizer->base.sprite_coord_mode ==
580 PIPE_SPRITE_COORD_UPPER_LEFT);
581 }
582
583 key->light_twoside = v3d->rasterizer->base.light_twoside;
584 key->shade_model_flat = v3d->rasterizer->base.flatshade;
585
586 struct v3d_compiled_shader *old_fs = v3d->prog.fs;
587 v3d->prog.fs = v3d_get_compiled_shader(v3d, &key->base, sizeof(*key));
588 if (v3d->prog.fs == old_fs)
589 return;
590
591 v3d->dirty |= VC5_DIRTY_COMPILED_FS;
592
593 if (old_fs) {
594 if (v3d->prog.fs->prog_data.fs->flat_shade_flags !=
595 old_fs->prog_data.fs->flat_shade_flags) {
596 v3d->dirty |= VC5_DIRTY_FLAT_SHADE_FLAGS;
597 }
598
599 if (v3d->prog.fs->prog_data.fs->noperspective_flags !=
600 old_fs->prog_data.fs->noperspective_flags) {
601 v3d->dirty |= VC5_DIRTY_NOPERSPECTIVE_FLAGS;
602 }
603
604 if (v3d->prog.fs->prog_data.fs->centroid_flags !=
605 old_fs->prog_data.fs->centroid_flags) {
606 v3d->dirty |= VC5_DIRTY_CENTROID_FLAGS;
607 }
608 }
609
610 if (old_fs && memcmp(v3d->prog.fs->prog_data.fs->input_slots,
611 old_fs->prog_data.fs->input_slots,
612 sizeof(v3d->prog.fs->prog_data.fs->input_slots))) {
613 v3d->dirty |= VC5_DIRTY_FS_INPUTS;
614 }
615 }
616
617 static void
618 v3d_update_compiled_gs(struct v3d_context *v3d, uint8_t prim_mode)
619 {
620 struct v3d_gs_key local_key;
621 struct v3d_gs_key *key = &local_key;
622
623 if (!(v3d->dirty & (VC5_DIRTY_GEOMTEX |
624 VC5_DIRTY_RASTERIZER |
625 VC5_DIRTY_UNCOMPILED_GS |
626 VC5_DIRTY_PRIM_MODE |
627 VC5_DIRTY_FS_INPUTS))) {
628 return;
629 }
630
631 if (!v3d->prog.bind_gs) {
632 v3d->prog.gs = NULL;
633 v3d->prog.gs_bin = NULL;
634 return;
635 }
636
637 memset(key, 0, sizeof(*key));
638 v3d_setup_shared_key(v3d, &key->base, &v3d->tex[PIPE_SHADER_GEOMETRY]);
639 key->base.shader_state = v3d->prog.bind_gs;
640 key->base.ucp_enables = v3d->rasterizer->base.clip_plane_enable;
641 key->base.is_last_geometry_stage = true;
642 key->num_used_outputs = v3d->prog.fs->prog_data.fs->num_inputs;
643 STATIC_ASSERT(sizeof(key->used_outputs) ==
644 sizeof(v3d->prog.fs->prog_data.fs->input_slots));
645 memcpy(key->used_outputs, v3d->prog.fs->prog_data.fs->input_slots,
646 sizeof(key->used_outputs));
647
648 key->per_vertex_point_size =
649 (prim_mode == PIPE_PRIM_POINTS &&
650 v3d->rasterizer->base.point_size_per_vertex);
651
652 struct v3d_compiled_shader *gs =
653 v3d_get_compiled_shader(v3d, &key->base, sizeof(*key));
654 if (gs != v3d->prog.gs) {
655 v3d->prog.gs = gs;
656 v3d->dirty |= VC5_DIRTY_COMPILED_GS;
657 }
658
659 key->is_coord = true;
660
661 /* The last bin-mode shader in the geometry pipeline only outputs
662 * varyings used by transform feedback.
663 */
664 struct v3d_uncompiled_shader *shader_state = key->base.shader_state;
665 memcpy(key->used_outputs, shader_state->tf_outputs,
666 sizeof(*key->used_outputs) * shader_state->num_tf_outputs);
667 if (shader_state->num_tf_outputs < key->num_used_outputs) {
668 uint32_t size = sizeof(*key->used_outputs) *
669 (key->num_used_outputs -
670 shader_state->num_tf_outputs);
671 memset(&key->used_outputs[shader_state->num_tf_outputs],
672 0, size);
673 }
674 key->num_used_outputs = shader_state->num_tf_outputs;
675
676 struct v3d_compiled_shader *old_gs = v3d->prog.gs;
677 struct v3d_compiled_shader *gs_bin =
678 v3d_get_compiled_shader(v3d, &key->base, sizeof(*key));
679 if (gs_bin != old_gs) {
680 v3d->prog.gs_bin = gs_bin;
681 v3d->dirty |= VC5_DIRTY_COMPILED_GS_BIN;
682 }
683
684 if (old_gs && memcmp(v3d->prog.gs->prog_data.gs->input_slots,
685 old_gs->prog_data.gs->input_slots,
686 sizeof(v3d->prog.gs->prog_data.gs->input_slots))) {
687 v3d->dirty |= VC5_DIRTY_GS_INPUTS;
688 }
689 }
690
691 static void
692 v3d_update_compiled_vs(struct v3d_context *v3d, uint8_t prim_mode)
693 {
694 struct v3d_vs_key local_key;
695 struct v3d_vs_key *key = &local_key;
696
697 if (!(v3d->dirty & (VC5_DIRTY_VERTTEX |
698 VC5_DIRTY_VTXSTATE |
699 VC5_DIRTY_UNCOMPILED_VS |
700 (v3d->prog.bind_gs ? 0 : VC5_DIRTY_RASTERIZER) |
701 (v3d->prog.bind_gs ? 0 : VC5_DIRTY_PRIM_MODE) |
702 (v3d->prog.bind_gs ? VC5_DIRTY_GS_INPUTS :
703 VC5_DIRTY_FS_INPUTS)))) {
704 return;
705 }
706
707 memset(key, 0, sizeof(*key));
708 v3d_setup_shared_key(v3d, &key->base, &v3d->tex[PIPE_SHADER_VERTEX]);
709 key->base.shader_state = v3d->prog.bind_vs;
710 key->base.ucp_enables = v3d->rasterizer->base.clip_plane_enable;
711 key->base.is_last_geometry_stage = !v3d->prog.bind_gs;
712
713 if (!v3d->prog.bind_gs) {
714 key->num_used_outputs = v3d->prog.fs->prog_data.fs->num_inputs;
715 STATIC_ASSERT(sizeof(key->used_outputs) ==
716 sizeof(v3d->prog.fs->prog_data.fs->input_slots));
717 memcpy(key->used_outputs, v3d->prog.fs->prog_data.fs->input_slots,
718 sizeof(key->used_outputs));
719 } else {
720 key->num_used_outputs = v3d->prog.gs->prog_data.gs->num_inputs;
721 STATIC_ASSERT(sizeof(key->used_outputs) ==
722 sizeof(v3d->prog.gs->prog_data.gs->input_slots));
723 memcpy(key->used_outputs, v3d->prog.gs->prog_data.gs->input_slots,
724 sizeof(key->used_outputs));
725 }
726
727 key->clamp_color = v3d->rasterizer->base.clamp_vertex_color;
728
729 key->per_vertex_point_size =
730 (prim_mode == PIPE_PRIM_POINTS &&
731 v3d->rasterizer->base.point_size_per_vertex);
732
733 struct v3d_compiled_shader *vs =
734 v3d_get_compiled_shader(v3d, &key->base, sizeof(*key));
735 if (vs != v3d->prog.vs) {
736 v3d->prog.vs = vs;
737 v3d->dirty |= VC5_DIRTY_COMPILED_VS;
738 }
739
740 key->is_coord = true;
741
742 /* Coord shaders only output varyings used by transform feedback,
743 * unless they are linked to other shaders in the geometry side
744 * of the pipeline, since in that case any of the output varyings
745 * could be required in later geometry stages to compute
746 * gl_Position or TF outputs.
747 */
748 if (!v3d->prog.bind_gs) {
749 struct v3d_uncompiled_shader *shader_state =
750 key->base.shader_state;
751 memcpy(key->used_outputs, shader_state->tf_outputs,
752 sizeof(*key->used_outputs) *
753 shader_state->num_tf_outputs);
754 if (shader_state->num_tf_outputs < key->num_used_outputs) {
755 uint32_t tail_bytes =
756 sizeof(*key->used_outputs) *
757 (key->num_used_outputs -
758 shader_state->num_tf_outputs);
759 memset(&key->used_outputs[shader_state->num_tf_outputs],
760 0, tail_bytes);
761 }
762 key->num_used_outputs = shader_state->num_tf_outputs;
763 }
764
765 struct v3d_compiled_shader *cs =
766 v3d_get_compiled_shader(v3d, &key->base, sizeof(*key));
767 if (cs != v3d->prog.cs) {
768 v3d->prog.cs = cs;
769 v3d->dirty |= VC5_DIRTY_COMPILED_CS;
770 }
771 }
772
773 void
774 v3d_update_compiled_shaders(struct v3d_context *v3d, uint8_t prim_mode)
775 {
776 v3d_update_compiled_fs(v3d, prim_mode);
777 v3d_update_compiled_gs(v3d, prim_mode);
778 v3d_update_compiled_vs(v3d, prim_mode);
779 }
780
781 void
782 v3d_update_compiled_cs(struct v3d_context *v3d)
783 {
784 struct v3d_key local_key;
785 struct v3d_key *key = &local_key;
786
787 if (!(v3d->dirty & (VC5_DIRTY_UNCOMPILED_CS |
788 VC5_DIRTY_COMPTEX))) {
789 return;
790 }
791
792 memset(key, 0, sizeof(*key));
793 v3d_setup_shared_key(v3d, key, &v3d->tex[PIPE_SHADER_COMPUTE]);
794 key->shader_state = v3d->prog.bind_compute;
795
796 struct v3d_compiled_shader *cs =
797 v3d_get_compiled_shader(v3d, key, sizeof(*key));
798 if (cs != v3d->prog.compute) {
799 v3d->prog.compute = cs;
800 v3d->dirty |= VC5_DIRTY_COMPILED_CS; /* XXX */
801 }
802 }
803
804 static uint32_t
805 fs_cache_hash(const void *key)
806 {
807 return _mesa_hash_data(key, sizeof(struct v3d_fs_key));
808 }
809
810 static uint32_t
811 gs_cache_hash(const void *key)
812 {
813 return _mesa_hash_data(key, sizeof(struct v3d_gs_key));
814 }
815
816 static uint32_t
817 vs_cache_hash(const void *key)
818 {
819 return _mesa_hash_data(key, sizeof(struct v3d_vs_key));
820 }
821
822 static uint32_t
823 cs_cache_hash(const void *key)
824 {
825 return _mesa_hash_data(key, sizeof(struct v3d_key));
826 }
827
828 static bool
829 fs_cache_compare(const void *key1, const void *key2)
830 {
831 return memcmp(key1, key2, sizeof(struct v3d_fs_key)) == 0;
832 }
833
834 static bool
835 gs_cache_compare(const void *key1, const void *key2)
836 {
837 return memcmp(key1, key2, sizeof(struct v3d_gs_key)) == 0;
838 }
839
840 static bool
841 vs_cache_compare(const void *key1, const void *key2)
842 {
843 return memcmp(key1, key2, sizeof(struct v3d_vs_key)) == 0;
844 }
845
846 static bool
847 cs_cache_compare(const void *key1, const void *key2)
848 {
849 return memcmp(key1, key2, sizeof(struct v3d_key)) == 0;
850 }
851
852 static void
853 v3d_shader_state_delete(struct pipe_context *pctx, void *hwcso)
854 {
855 struct v3d_context *v3d = v3d_context(pctx);
856 struct v3d_uncompiled_shader *so = hwcso;
857 nir_shader *s = so->base.ir.nir;
858
859 hash_table_foreach(v3d->prog.cache[s->info.stage], entry) {
860 const struct v3d_key *key = entry->key;
861 struct v3d_compiled_shader *shader = entry->data;
862
863 if (key->shader_state != so)
864 continue;
865
866 if (v3d->prog.fs == shader)
867 v3d->prog.fs = NULL;
868 if (v3d->prog.vs == shader)
869 v3d->prog.vs = NULL;
870 if (v3d->prog.cs == shader)
871 v3d->prog.cs = NULL;
872 if (v3d->prog.compute == shader)
873 v3d->prog.compute = NULL;
874
875 _mesa_hash_table_remove(v3d->prog.cache[s->info.stage], entry);
876 v3d_free_compiled_shader(shader);
877 }
878
879 ralloc_free(so->base.ir.nir);
880 free(so);
881 }
882
883 static void
884 v3d_fp_state_bind(struct pipe_context *pctx, void *hwcso)
885 {
886 struct v3d_context *v3d = v3d_context(pctx);
887 v3d->prog.bind_fs = hwcso;
888 v3d->dirty |= VC5_DIRTY_UNCOMPILED_FS;
889 }
890
891 static void
892 v3d_gp_state_bind(struct pipe_context *pctx, void *hwcso)
893 {
894 struct v3d_context *v3d = v3d_context(pctx);
895 v3d->prog.bind_gs = hwcso;
896 v3d->dirty |= VC5_DIRTY_UNCOMPILED_GS;
897 }
898
899 static void
900 v3d_vp_state_bind(struct pipe_context *pctx, void *hwcso)
901 {
902 struct v3d_context *v3d = v3d_context(pctx);
903 v3d->prog.bind_vs = hwcso;
904 v3d->dirty |= VC5_DIRTY_UNCOMPILED_VS;
905 }
906
907 static void
908 v3d_compute_state_bind(struct pipe_context *pctx, void *state)
909 {
910 struct v3d_context *v3d = v3d_context(pctx);
911
912 v3d->prog.bind_compute = state;
913 v3d->dirty |= VC5_DIRTY_UNCOMPILED_CS;
914 }
915
916 static void *
917 v3d_create_compute_state(struct pipe_context *pctx,
918 const struct pipe_compute_state *cso)
919 {
920 return v3d_uncompiled_shader_create(pctx, cso->ir_type,
921 (void *)cso->prog);
922 }
923
924 void
925 v3d_program_init(struct pipe_context *pctx)
926 {
927 struct v3d_context *v3d = v3d_context(pctx);
928
929 pctx->create_vs_state = v3d_shader_state_create;
930 pctx->delete_vs_state = v3d_shader_state_delete;
931
932 pctx->create_gs_state = v3d_shader_state_create;
933 pctx->delete_gs_state = v3d_shader_state_delete;
934
935 pctx->create_fs_state = v3d_shader_state_create;
936 pctx->delete_fs_state = v3d_shader_state_delete;
937
938 pctx->bind_fs_state = v3d_fp_state_bind;
939 pctx->bind_gs_state = v3d_gp_state_bind;
940 pctx->bind_vs_state = v3d_vp_state_bind;
941
942 if (v3d->screen->has_csd) {
943 pctx->create_compute_state = v3d_create_compute_state;
944 pctx->delete_compute_state = v3d_shader_state_delete;
945 pctx->bind_compute_state = v3d_compute_state_bind;
946 }
947
948 v3d->prog.cache[MESA_SHADER_VERTEX] =
949 _mesa_hash_table_create(pctx, vs_cache_hash, vs_cache_compare);
950 v3d->prog.cache[MESA_SHADER_GEOMETRY] =
951 _mesa_hash_table_create(pctx, gs_cache_hash, gs_cache_compare);
952 v3d->prog.cache[MESA_SHADER_FRAGMENT] =
953 _mesa_hash_table_create(pctx, fs_cache_hash, fs_cache_compare);
954 v3d->prog.cache[MESA_SHADER_COMPUTE] =
955 _mesa_hash_table_create(pctx, cs_cache_hash, cs_cache_compare);
956 }
957
958 void
959 v3d_program_fini(struct pipe_context *pctx)
960 {
961 struct v3d_context *v3d = v3d_context(pctx);
962
963 for (int i = 0; i < MESA_SHADER_STAGES; i++) {
964 struct hash_table *cache = v3d->prog.cache[i];
965 if (!cache)
966 continue;
967
968 hash_table_foreach(cache, entry) {
969 struct v3d_compiled_shader *shader = entry->data;
970 v3d_free_compiled_shader(shader);
971 _mesa_hash_table_remove(cache, entry);
972 }
973 }
974
975 v3d_bo_unreference(&v3d->prog.spill_bo);
976 }