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
40static struct v3d_compiled_shader *
41v3d_get_compiled_shader(struct v3d_context *v3d,
42                        struct v3d_key *key, size_t key_size);
43static void
44v3d_setup_shared_precompile_key(struct v3d_uncompiled_shader *uncompiled,
45                                struct v3d_key *key);
46
47static gl_varying_slot
48v3d_get_slot_for_driver_location(nir_shader *s, uint32_t driver_location)
49{
50        nir_foreach_shader_out_variable(var, s) {
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 */
67static void
68v3d_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
172static int
173type_size(const struct glsl_type *type, bool bindless)
174{
175        return glsl_count_attribute_slots(type, false);
176}
177
178static void
179precompile_all_outputs(nir_shader *s,
180                       struct v3d_varying_slot *outputs,
181                       uint8_t *num_outputs)
182{
183        nir_foreach_shader_out_variable(var, s) {
184                const int array_len = MAX2(glsl_get_length(var->type), 1);
185                for (int j = 0; j < array_len; j++) {
186                        const int slot = var->data.location + j;
187                        const int num_components =
188                                glsl_get_components(var->type);
189                        for (int i = 0; i < num_components; i++) {
190                                const int swiz = var->data.location_frac + i;
191                                outputs[(*num_outputs)++] =
192                                        v3d_slot_from_slot_and_component(slot,
193                                                                         swiz);
194                        }
195                }
196        }
197}
198
199/**
200 * Precompiles a shader variant at shader state creation time if
201 * V3D_DEBUG=precompile is set.  Used for shader-db
202 * (https://gitlab.freedesktop.org/mesa/shader-db)
203 */
204static void
205v3d_shader_precompile(struct v3d_context *v3d,
206                      struct v3d_uncompiled_shader *so)
207{
208        nir_shader *s = so->base.ir.nir;
209
210        if (s->info.stage == MESA_SHADER_FRAGMENT) {
211                struct v3d_fs_key key = {
212                        .base.shader_state = so,
213                };
214
215                nir_foreach_shader_out_variable(var, s) {
216                        if (var->data.location == FRAG_RESULT_COLOR) {
217                                key.cbufs |= 1 << 0;
218                        } else if (var->data.location >= FRAG_RESULT_DATA0) {
219                                key.cbufs |= 1 << (var->data.location -
220                                                   FRAG_RESULT_DATA0);
221                        }
222                }
223
224                key.logicop_func = PIPE_LOGICOP_COPY;
225
226                v3d_setup_shared_precompile_key(so, &key.base);
227                v3d_get_compiled_shader(v3d, &key.base, sizeof(key));
228        } else if (s->info.stage == MESA_SHADER_GEOMETRY) {
229                struct v3d_gs_key key = {
230                        .base.shader_state = so,
231                        .base.is_last_geometry_stage = true,
232                };
233
234                v3d_setup_shared_precompile_key(so, &key.base);
235
236                precompile_all_outputs(s,
237                                       key.used_outputs,
238                                       &key.num_used_outputs);
239
240                v3d_get_compiled_shader(v3d, &key.base, sizeof(key));
241
242                /* Compile GS bin shader: only position (XXX: include TF) */
243                key.is_coord = true;
244                key.num_used_outputs = 0;
245                for (int i = 0; i < 4; i++) {
246                        key.used_outputs[key.num_used_outputs++] =
247                                v3d_slot_from_slot_and_component(VARYING_SLOT_POS,
248                                                                 i);
249                }
250                v3d_get_compiled_shader(v3d, &key.base, sizeof(key));
251        } else {
252                assert(s->info.stage == MESA_SHADER_VERTEX);
253                struct v3d_vs_key key = {
254                        .base.shader_state = so,
255                        /* Emit fixed function outputs */
256                        .base.is_last_geometry_stage = true,
257                };
258
259                v3d_setup_shared_precompile_key(so, &key.base);
260
261                precompile_all_outputs(s,
262                                       key.used_outputs,
263                                       &key.num_used_outputs);
264
265                v3d_get_compiled_shader(v3d, &key.base, sizeof(key));
266
267                /* Compile VS bin shader: only position (XXX: include TF) */
268                key.is_coord = true;
269                key.num_used_outputs = 0;
270                for (int i = 0; i < 4; i++) {
271                        key.used_outputs[key.num_used_outputs++] =
272                                v3d_slot_from_slot_and_component(VARYING_SLOT_POS,
273                                                                 i);
274                }
275                v3d_get_compiled_shader(v3d, &key.base, sizeof(key));
276        }
277}
278
279static void *
280v3d_uncompiled_shader_create(struct pipe_context *pctx,
281                             enum pipe_shader_ir type, void *ir)
282{
283        struct v3d_context *v3d = v3d_context(pctx);
284        struct v3d_uncompiled_shader *so = CALLOC_STRUCT(v3d_uncompiled_shader);
285        if (!so)
286                return NULL;
287
288        so->program_id = v3d->next_uncompiled_program_id++;
289
290        nir_shader *s;
291
292        if (type == PIPE_SHADER_IR_NIR) {
293                /* The backend takes ownership of the NIR shader on state
294                 * creation.
295                 */
296                s = ir;
297        } else {
298                assert(type == PIPE_SHADER_IR_TGSI);
299
300                if (unlikely(V3D_DEBUG & V3D_DEBUG_TGSI)) {
301                        fprintf(stderr, "prog %d TGSI:\n",
302                                so->program_id);
303                        tgsi_dump(ir, 0);
304                        fprintf(stderr, "\n");
305                }
306                s = tgsi_to_nir(ir, pctx->screen, false);
307        }
308
309        if (s->info.stage != MESA_SHADER_VERTEX &&
310            s->info.stage != MESA_SHADER_GEOMETRY) {
311                NIR_PASS_V(s, nir_lower_io,
312                           nir_var_shader_in | nir_var_shader_out,
313                           type_size, (nir_lower_io_options)0);
314        }
315
316        NIR_PASS_V(s, nir_lower_regs_to_ssa);
317        NIR_PASS_V(s, nir_normalize_cubemap_coords);
318
319        NIR_PASS_V(s, nir_lower_load_const_to_scalar);
320
321        v3d_optimize_nir(NULL, s);
322
323        NIR_PASS_V(s, nir_remove_dead_variables, nir_var_function_temp, NULL);
324
325        /* Garbage collect dead instructions */
326        nir_sweep(s);
327
328        so->base.type = PIPE_SHADER_IR_NIR;
329        so->base.ir.nir = s;
330
331        if (unlikely(V3D_DEBUG & (V3D_DEBUG_NIR |
332                                  v3d_debug_flag_for_shader_stage(s->info.stage)))) {
333                fprintf(stderr, "%s prog %d NIR:\n",
334                        gl_shader_stage_name(s->info.stage),
335                        so->program_id);
336                nir_print_shader(s, stderr);
337                fprintf(stderr, "\n");
338        }
339
340        if (unlikely(V3D_DEBUG & V3D_DEBUG_PRECOMPILE))
341                v3d_shader_precompile(v3d, so);
342
343        return so;
344}
345
346static void
347v3d_shader_debug_output(const char *message, void *data)
348{
349        struct v3d_context *v3d = data;
350
351        pipe_debug_message(&v3d->debug, SHADER_INFO, "%s", message);
352}
353
354static void *
355v3d_shader_state_create(struct pipe_context *pctx,
356                        const struct pipe_shader_state *cso)
357{
358        struct v3d_uncompiled_shader *so =
359                v3d_uncompiled_shader_create(pctx,
360                                             cso->type,
361                                             (cso->type == PIPE_SHADER_IR_TGSI ?
362                                              (void *)cso->tokens :
363                                              cso->ir.nir));
364
365        v3d_set_transform_feedback_outputs(so, &cso->stream_output);
366
367        return so;
368}
369
370struct v3d_compiled_shader *
371v3d_get_compiled_shader(struct v3d_context *v3d,
372                        struct v3d_key *key,
373                        size_t key_size)
374{
375        struct v3d_uncompiled_shader *shader_state = key->shader_state;
376        nir_shader *s = shader_state->base.ir.nir;
377
378        struct hash_table *ht = v3d->prog.cache[s->info.stage];
379        struct hash_entry *entry = _mesa_hash_table_search(ht, key);
380        if (entry)
381                return entry->data;
382
383        struct v3d_compiled_shader *shader =
384                rzalloc(NULL, struct v3d_compiled_shader);
385
386        int program_id = shader_state->program_id;
387        int variant_id =
388                p_atomic_inc_return(&shader_state->compiled_variant_count);
389        uint64_t *qpu_insts;
390        uint32_t shader_size;
391
392        qpu_insts = v3d_compile(v3d->screen->compiler, key,
393                                &shader->prog_data.base, s,
394                                v3d_shader_debug_output,
395                                v3d,
396                                program_id, variant_id, &shader_size);
397        ralloc_steal(shader, shader->prog_data.base);
398
399        v3d_set_shader_uniform_dirty_flags(shader);
400
401        if (shader_size) {
402                u_upload_data(v3d->state_uploader, 0, shader_size, 8,
403                              qpu_insts, &shader->offset, &shader->resource);
404        }
405
406        free(qpu_insts);
407
408        if (ht) {
409                struct v3d_key *dup_key;
410                dup_key = ralloc_size(shader, key_size);
411                memcpy(dup_key, key, key_size);
412                _mesa_hash_table_insert(ht, dup_key, shader);
413        }
414
415        if (shader->prog_data.base->spill_size >
416            v3d->prog.spill_size_per_thread) {
417                /* The TIDX register we use for choosing the area to access
418                 * for scratch space is: (core << 6) | (qpu << 2) | thread.
419                 * Even at minimum threadcount in a particular shader, that
420                 * means we still multiply by qpus by 4.
421                 */
422                int total_spill_size = (v3d->screen->devinfo.qpu_count * 4 *
423                                        shader->prog_data.base->spill_size);
424
425                v3d_bo_unreference(&v3d->prog.spill_bo);
426                v3d->prog.spill_bo = v3d_bo_alloc(v3d->screen,
427                                                  total_spill_size, "spill");
428                v3d->prog.spill_size_per_thread =
429                        shader->prog_data.base->spill_size;
430        }
431
432        return shader;
433}
434
435static void
436v3d_free_compiled_shader(struct v3d_compiled_shader *shader)
437{
438        pipe_resource_reference(&shader->resource, NULL);
439        ralloc_free(shader);
440}
441
442static void
443v3d_setup_shared_key(struct v3d_context *v3d, struct v3d_key *key,
444                     struct v3d_texture_stateobj *texstate)
445{
446        const struct v3d_device_info *devinfo = &v3d->screen->devinfo;
447
448        key->num_tex_used = texstate->num_textures;
449        key->num_samplers_used = texstate->num_textures;
450        assert(key->num_tex_used == key->num_samplers_used);
451        for (int i = 0; i < texstate->num_textures; i++) {
452                struct pipe_sampler_view *sampler = texstate->textures[i];
453                struct v3d_sampler_view *v3d_sampler = v3d_sampler_view(sampler);
454                struct pipe_sampler_state *sampler_state =
455                        texstate->samplers[i];
456
457                if (!sampler)
458                        continue;
459
460                key->sampler[i].return_size =
461                        v3d_get_tex_return_size(devinfo,
462                                                sampler->format,
463                                                sampler_state->compare_mode);
464
465                /* For 16-bit, we set up the sampler to always return 2
466                 * channels (meaning no recompiles for most statechanges),
467                 * while for 32 we actually scale the returns with channels.
468                 */
469                if (key->sampler[i].return_size == 16) {
470                        key->sampler[i].return_channels = 2;
471                } else if (devinfo->ver > 40) {
472                        key->sampler[i].return_channels = 4;
473                } else {
474                        key->sampler[i].return_channels =
475                                v3d_get_tex_return_channels(devinfo,
476                                                            sampler->format);
477                }
478
479                if (key->sampler[i].return_size == 32 && devinfo->ver < 40) {
480                        memcpy(key->tex[i].swizzle,
481                               v3d_sampler->swizzle,
482                               sizeof(v3d_sampler->swizzle));
483                } else {
484                        /* For 16-bit returns, we let the sampler state handle
485                         * the swizzle.
486                         */
487                        key->tex[i].swizzle[0] = PIPE_SWIZZLE_X;
488                        key->tex[i].swizzle[1] = PIPE_SWIZZLE_Y;
489                        key->tex[i].swizzle[2] = PIPE_SWIZZLE_Z;
490                        key->tex[i].swizzle[3] = PIPE_SWIZZLE_W;
491                }
492        }
493}
494
495static void
496v3d_setup_shared_precompile_key(struct v3d_uncompiled_shader *uncompiled,
497                                struct v3d_key *key)
498{
499        nir_shader *s = uncompiled->base.ir.nir;
500
501        /* Note that below we access they key's texture and sampler fields
502         * using the same index. On OpenGL they are the same (they are
503         * combined)
504         */
505        key->num_tex_used = s->info.num_textures;
506        key->num_samplers_used = s->info.num_textures;
507        for (int i = 0; i < s->info.num_textures; i++) {
508                key->sampler[i].return_size = 16;
509                key->sampler[i].return_channels = 2;
510
511                key->tex[i].swizzle[0] = PIPE_SWIZZLE_X;
512                key->tex[i].swizzle[1] = PIPE_SWIZZLE_Y;
513                key->tex[i].swizzle[2] = PIPE_SWIZZLE_Z;
514                key->tex[i].swizzle[3] = PIPE_SWIZZLE_W;
515        }
516}
517
518static void
519v3d_update_compiled_fs(struct v3d_context *v3d, uint8_t prim_mode)
520{
521        struct v3d_job *job = v3d->job;
522        struct v3d_fs_key local_key;
523        struct v3d_fs_key *key = &local_key;
524        nir_shader *s = v3d->prog.bind_fs->base.ir.nir;
525
526        if (!(v3d->dirty & (V3D_DIRTY_PRIM_MODE |
527                            V3D_DIRTY_BLEND |
528                            V3D_DIRTY_FRAMEBUFFER |
529                            V3D_DIRTY_ZSA |
530                            V3D_DIRTY_RASTERIZER |
531                            V3D_DIRTY_SAMPLE_STATE |
532                            V3D_DIRTY_FRAGTEX |
533                            V3D_DIRTY_UNCOMPILED_FS))) {
534                return;
535        }
536
537        memset(key, 0, sizeof(*key));
538        v3d_setup_shared_key(v3d, &key->base, &v3d->tex[PIPE_SHADER_FRAGMENT]);
539        key->base.shader_state = v3d->prog.bind_fs;
540        key->base.ucp_enables = v3d->rasterizer->base.clip_plane_enable;
541        key->is_points = (prim_mode == PIPE_PRIM_POINTS);
542        key->is_lines = (prim_mode >= PIPE_PRIM_LINES &&
543                         prim_mode <= PIPE_PRIM_LINE_STRIP);
544        key->line_smoothing = (key->is_lines &&
545                               v3d_line_smoothing_enabled(v3d));
546        key->has_gs = v3d->prog.bind_gs != NULL;
547        if (v3d->blend->base.logicop_enable) {
548                key->logicop_func = v3d->blend->base.logicop_func;
549        } else {
550                key->logicop_func = PIPE_LOGICOP_COPY;
551        }
552        if (job->msaa) {
553                key->msaa = v3d->rasterizer->base.multisample;
554                key->sample_coverage = (v3d->rasterizer->base.multisample &&
555                                        v3d->sample_mask != (1 << V3D_MAX_SAMPLES) - 1);
556                key->sample_alpha_to_coverage = v3d->blend->base.alpha_to_coverage;
557                key->sample_alpha_to_one = v3d->blend->base.alpha_to_one;
558        }
559
560        key->swap_color_rb = v3d->swap_color_rb;
561
562        for (int i = 0; i < v3d->framebuffer.nr_cbufs; i++) {
563                struct pipe_surface *cbuf = v3d->framebuffer.cbufs[i];
564                if (!cbuf)
565                        continue;
566
567                /* gl_FragColor's propagation to however many bound color
568                 * buffers there are means that the shader compile needs to
569                 * know what buffers are present.
570                 */
571                key->cbufs |= 1 << i;
572
573                /* If logic operations are enabled then we might emit color
574                 * reads and we need to know the color buffer format and
575                 * swizzle for that.
576                 */
577                if (key->logicop_func != PIPE_LOGICOP_COPY) {
578                        key->color_fmt[i].format = cbuf->format;
579                        key->color_fmt[i].swizzle =
580                                v3d_get_format_swizzle(&v3d->screen->devinfo,
581                                                       cbuf->format);
582                }
583
584                const struct util_format_description *desc =
585                        util_format_description(cbuf->format);
586
587                if (desc->channel[0].type == UTIL_FORMAT_TYPE_FLOAT &&
588                    desc->channel[0].size == 32) {
589                        key->f32_color_rb |= 1 << i;
590                }
591
592                if (s->info.fs.untyped_color_outputs) {
593                        if (util_format_is_pure_uint(cbuf->format))
594                                key->uint_color_rb |= 1 << i;
595                        else if (util_format_is_pure_sint(cbuf->format))
596                                key->int_color_rb |= 1 << i;
597                }
598        }
599
600        if (key->is_points) {
601                key->point_sprite_mask =
602                        v3d->rasterizer->base.sprite_coord_enable;
603                /* this is handled by lower_wpos_pntc */
604                key->point_coord_upper_left = false;
605        }
606
607        struct v3d_compiled_shader *old_fs = v3d->prog.fs;
608        v3d->prog.fs = v3d_get_compiled_shader(v3d, &key->base, sizeof(*key));
609        if (v3d->prog.fs == old_fs)
610                return;
611
612        v3d->dirty |= V3D_DIRTY_COMPILED_FS;
613
614        if (old_fs) {
615                if (v3d->prog.fs->prog_data.fs->flat_shade_flags !=
616                    old_fs->prog_data.fs->flat_shade_flags) {
617                        v3d->dirty |= V3D_DIRTY_FLAT_SHADE_FLAGS;
618                }
619
620                if (v3d->prog.fs->prog_data.fs->noperspective_flags !=
621                    old_fs->prog_data.fs->noperspective_flags) {
622                        v3d->dirty |= V3D_DIRTY_NOPERSPECTIVE_FLAGS;
623                }
624
625                if (v3d->prog.fs->prog_data.fs->centroid_flags !=
626                    old_fs->prog_data.fs->centroid_flags) {
627                        v3d->dirty |= V3D_DIRTY_CENTROID_FLAGS;
628                }
629        }
630
631        if (old_fs && memcmp(v3d->prog.fs->prog_data.fs->input_slots,
632                             old_fs->prog_data.fs->input_slots,
633                             sizeof(v3d->prog.fs->prog_data.fs->input_slots))) {
634                v3d->dirty |= V3D_DIRTY_FS_INPUTS;
635        }
636}
637
638static void
639v3d_update_compiled_gs(struct v3d_context *v3d, uint8_t prim_mode)
640{
641        struct v3d_gs_key local_key;
642        struct v3d_gs_key *key = &local_key;
643
644        if (!(v3d->dirty & (V3D_DIRTY_GEOMTEX |
645                            V3D_DIRTY_RASTERIZER |
646                            V3D_DIRTY_UNCOMPILED_GS |
647                            V3D_DIRTY_PRIM_MODE |
648                            V3D_DIRTY_FS_INPUTS))) {
649                return;
650        }
651
652        if (!v3d->prog.bind_gs) {
653                v3d->prog.gs = NULL;
654                v3d->prog.gs_bin = NULL;
655                return;
656        }
657
658        memset(key, 0, sizeof(*key));
659        v3d_setup_shared_key(v3d, &key->base, &v3d->tex[PIPE_SHADER_GEOMETRY]);
660        key->base.shader_state = v3d->prog.bind_gs;
661        key->base.ucp_enables = v3d->rasterizer->base.clip_plane_enable;
662        key->base.is_last_geometry_stage = true;
663        key->num_used_outputs = v3d->prog.fs->prog_data.fs->num_inputs;
664        STATIC_ASSERT(sizeof(key->used_outputs) ==
665                      sizeof(v3d->prog.fs->prog_data.fs->input_slots));
666        memcpy(key->used_outputs, v3d->prog.fs->prog_data.fs->input_slots,
667               sizeof(key->used_outputs));
668
669        key->per_vertex_point_size =
670                (prim_mode == PIPE_PRIM_POINTS &&
671                 v3d->rasterizer->base.point_size_per_vertex);
672
673        struct v3d_compiled_shader *gs =
674                v3d_get_compiled_shader(v3d, &key->base, sizeof(*key));
675        if (gs != v3d->prog.gs) {
676                v3d->prog.gs = gs;
677                v3d->dirty |= V3D_DIRTY_COMPILED_GS;
678        }
679
680        key->is_coord = true;
681
682        /* The last bin-mode shader in the geometry pipeline only outputs
683         * varyings used by transform feedback.
684         */
685        struct v3d_uncompiled_shader *shader_state = key->base.shader_state;
686        memcpy(key->used_outputs, shader_state->tf_outputs,
687               sizeof(*key->used_outputs) * shader_state->num_tf_outputs);
688        if (shader_state->num_tf_outputs < key->num_used_outputs) {
689                uint32_t size = sizeof(*key->used_outputs) *
690                                (key->num_used_outputs -
691                                 shader_state->num_tf_outputs);
692                memset(&key->used_outputs[shader_state->num_tf_outputs],
693                       0, size);
694        }
695        key->num_used_outputs = shader_state->num_tf_outputs;
696
697        struct v3d_compiled_shader *old_gs = v3d->prog.gs;
698        struct v3d_compiled_shader *gs_bin =
699                v3d_get_compiled_shader(v3d, &key->base, sizeof(*key));
700        if (gs_bin != old_gs) {
701                v3d->prog.gs_bin = gs_bin;
702                v3d->dirty |= V3D_DIRTY_COMPILED_GS_BIN;
703        }
704
705        if (old_gs && memcmp(v3d->prog.gs->prog_data.gs->input_slots,
706                             old_gs->prog_data.gs->input_slots,
707                             sizeof(v3d->prog.gs->prog_data.gs->input_slots))) {
708                v3d->dirty |= V3D_DIRTY_GS_INPUTS;
709        }
710}
711
712static void
713v3d_update_compiled_vs(struct v3d_context *v3d, uint8_t prim_mode)
714{
715        struct v3d_vs_key local_key;
716        struct v3d_vs_key *key = &local_key;
717
718        if (!(v3d->dirty & (V3D_DIRTY_VERTTEX |
719                            V3D_DIRTY_VTXSTATE |
720                            V3D_DIRTY_UNCOMPILED_VS |
721                            (v3d->prog.bind_gs ? 0 : V3D_DIRTY_RASTERIZER) |
722                            (v3d->prog.bind_gs ? 0 : V3D_DIRTY_PRIM_MODE) |
723                            (v3d->prog.bind_gs ? V3D_DIRTY_GS_INPUTS :
724                                                 V3D_DIRTY_FS_INPUTS)))) {
725                return;
726        }
727
728        memset(key, 0, sizeof(*key));
729        v3d_setup_shared_key(v3d, &key->base, &v3d->tex[PIPE_SHADER_VERTEX]);
730        key->base.shader_state = v3d->prog.bind_vs;
731        key->base.ucp_enables = v3d->rasterizer->base.clip_plane_enable;
732        key->base.is_last_geometry_stage = !v3d->prog.bind_gs;
733
734        if (!v3d->prog.bind_gs) {
735            key->num_used_outputs = v3d->prog.fs->prog_data.fs->num_inputs;
736            STATIC_ASSERT(sizeof(key->used_outputs) ==
737                          sizeof(v3d->prog.fs->prog_data.fs->input_slots));
738            memcpy(key->used_outputs, v3d->prog.fs->prog_data.fs->input_slots,
739                   sizeof(key->used_outputs));
740        } else {
741            key->num_used_outputs = v3d->prog.gs->prog_data.gs->num_inputs;
742            STATIC_ASSERT(sizeof(key->used_outputs) ==
743                          sizeof(v3d->prog.gs->prog_data.gs->input_slots));
744            memcpy(key->used_outputs, v3d->prog.gs->prog_data.gs->input_slots,
745                   sizeof(key->used_outputs));
746        }
747
748        key->per_vertex_point_size =
749                (prim_mode == PIPE_PRIM_POINTS &&
750                 v3d->rasterizer->base.point_size_per_vertex);
751
752        nir_shader *s = v3d->prog.bind_vs->base.ir.nir;
753        uint64_t inputs_read = s->info.inputs_read;
754        assert(util_bitcount(inputs_read) <= v3d->vtx->num_elements);
755
756        while (inputs_read) {
757                int location = u_bit_scan64(&inputs_read);
758                nir_variable *var =
759                        nir_find_variable_with_location(s, nir_var_shader_in, location);
760                assert (var != NULL);
761                int driver_location = var->data.driver_location;
762                switch (v3d->vtx->pipe[driver_location].src_format) {
763                case PIPE_FORMAT_B8G8R8A8_UNORM:
764                case PIPE_FORMAT_B10G10R10A2_UNORM:
765                case PIPE_FORMAT_B10G10R10A2_SNORM:
766                case PIPE_FORMAT_B10G10R10A2_USCALED:
767                case PIPE_FORMAT_B10G10R10A2_SSCALED:
768                        key->va_swap_rb_mask |= 1 << location;
769                        break;
770                default:
771                        break;
772                }
773        }
774
775        struct v3d_compiled_shader *vs =
776                v3d_get_compiled_shader(v3d, &key->base, sizeof(*key));
777        if (vs != v3d->prog.vs) {
778                v3d->prog.vs = vs;
779                v3d->dirty |= V3D_DIRTY_COMPILED_VS;
780        }
781
782        key->is_coord = true;
783
784        /* Coord shaders only output varyings used by transform feedback,
785         * unless they are linked to other shaders in the geometry side
786         * of the pipeline, since in that case any of the output varyings
787         * could be required in later geometry stages to compute
788         * gl_Position or TF outputs.
789         */
790        if (!v3d->prog.bind_gs) {
791                struct v3d_uncompiled_shader *shader_state =
792                        key->base.shader_state;
793                memcpy(key->used_outputs, shader_state->tf_outputs,
794                       sizeof(*key->used_outputs) *
795                       shader_state->num_tf_outputs);
796                if (shader_state->num_tf_outputs < key->num_used_outputs) {
797                        uint32_t tail_bytes =
798                                sizeof(*key->used_outputs) *
799                                (key->num_used_outputs -
800                                 shader_state->num_tf_outputs);
801                        memset(&key->used_outputs[shader_state->num_tf_outputs],
802                               0, tail_bytes);
803                }
804                key->num_used_outputs = shader_state->num_tf_outputs;
805        } else {
806                key->num_used_outputs = v3d->prog.gs_bin->prog_data.gs->num_inputs;
807                STATIC_ASSERT(sizeof(key->used_outputs) ==
808                              sizeof(v3d->prog.gs_bin->prog_data.gs->input_slots));
809                memcpy(key->used_outputs, v3d->prog.gs_bin->prog_data.gs->input_slots,
810                       sizeof(key->used_outputs));
811        }
812
813        struct v3d_compiled_shader *cs =
814                v3d_get_compiled_shader(v3d, &key->base, sizeof(*key));
815        if (cs != v3d->prog.cs) {
816                v3d->prog.cs = cs;
817                v3d->dirty |= V3D_DIRTY_COMPILED_CS;
818        }
819}
820
821void
822v3d_update_compiled_shaders(struct v3d_context *v3d, uint8_t prim_mode)
823{
824        v3d_update_compiled_fs(v3d, prim_mode);
825        v3d_update_compiled_gs(v3d, prim_mode);
826        v3d_update_compiled_vs(v3d, prim_mode);
827}
828
829void
830v3d_update_compiled_cs(struct v3d_context *v3d)
831{
832        struct v3d_key local_key;
833        struct v3d_key *key = &local_key;
834
835        if (!(v3d->dirty & (V3D_DIRTY_UNCOMPILED_CS |
836                            V3D_DIRTY_COMPTEX))) {
837                return;
838        }
839
840        memset(key, 0, sizeof(*key));
841        v3d_setup_shared_key(v3d, key, &v3d->tex[PIPE_SHADER_COMPUTE]);
842        key->shader_state = v3d->prog.bind_compute;
843
844        struct v3d_compiled_shader *cs =
845                v3d_get_compiled_shader(v3d, key, sizeof(*key));
846        if (cs != v3d->prog.compute) {
847                v3d->prog.compute = cs;
848                v3d->dirty |= V3D_DIRTY_COMPILED_CS; /* XXX */
849        }
850}
851
852static uint32_t
853fs_cache_hash(const void *key)
854{
855        return _mesa_hash_data(key, sizeof(struct v3d_fs_key));
856}
857
858static uint32_t
859gs_cache_hash(const void *key)
860{
861        return _mesa_hash_data(key, sizeof(struct v3d_gs_key));
862}
863
864static uint32_t
865vs_cache_hash(const void *key)
866{
867        return _mesa_hash_data(key, sizeof(struct v3d_vs_key));
868}
869
870static uint32_t
871cs_cache_hash(const void *key)
872{
873        return _mesa_hash_data(key, sizeof(struct v3d_key));
874}
875
876static bool
877fs_cache_compare(const void *key1, const void *key2)
878{
879        return memcmp(key1, key2, sizeof(struct v3d_fs_key)) == 0;
880}
881
882static bool
883gs_cache_compare(const void *key1, const void *key2)
884{
885        return memcmp(key1, key2, sizeof(struct v3d_gs_key)) == 0;
886}
887
888static bool
889vs_cache_compare(const void *key1, const void *key2)
890{
891        return memcmp(key1, key2, sizeof(struct v3d_vs_key)) == 0;
892}
893
894static bool
895cs_cache_compare(const void *key1, const void *key2)
896{
897        return memcmp(key1, key2, sizeof(struct v3d_key)) == 0;
898}
899
900static void
901v3d_shader_state_delete(struct pipe_context *pctx, void *hwcso)
902{
903        struct v3d_context *v3d = v3d_context(pctx);
904        struct v3d_uncompiled_shader *so = hwcso;
905        nir_shader *s = so->base.ir.nir;
906
907        hash_table_foreach(v3d->prog.cache[s->info.stage], entry) {
908                const struct v3d_key *key = entry->key;
909                struct v3d_compiled_shader *shader = entry->data;
910
911                if (key->shader_state != so)
912                        continue;
913
914                if (v3d->prog.fs == shader)
915                        v3d->prog.fs = NULL;
916                if (v3d->prog.vs == shader)
917                        v3d->prog.vs = NULL;
918                if (v3d->prog.cs == shader)
919                        v3d->prog.cs = NULL;
920                if (v3d->prog.compute == shader)
921                        v3d->prog.compute = NULL;
922
923                _mesa_hash_table_remove(v3d->prog.cache[s->info.stage], entry);
924                v3d_free_compiled_shader(shader);
925        }
926
927        ralloc_free(so->base.ir.nir);
928        free(so);
929}
930
931static void
932v3d_fp_state_bind(struct pipe_context *pctx, void *hwcso)
933{
934        struct v3d_context *v3d = v3d_context(pctx);
935        v3d->prog.bind_fs = hwcso;
936        v3d->dirty |= V3D_DIRTY_UNCOMPILED_FS;
937}
938
939static void
940v3d_gp_state_bind(struct pipe_context *pctx, void *hwcso)
941{
942        struct v3d_context *v3d = v3d_context(pctx);
943        v3d->prog.bind_gs = hwcso;
944        v3d->dirty |= V3D_DIRTY_UNCOMPILED_GS;
945}
946
947static void
948v3d_vp_state_bind(struct pipe_context *pctx, void *hwcso)
949{
950        struct v3d_context *v3d = v3d_context(pctx);
951        v3d->prog.bind_vs = hwcso;
952        v3d->dirty |= V3D_DIRTY_UNCOMPILED_VS;
953}
954
955static void
956v3d_compute_state_bind(struct pipe_context *pctx, void *state)
957{
958        struct v3d_context *v3d = v3d_context(pctx);
959
960        v3d->prog.bind_compute = state;
961        v3d->dirty |= V3D_DIRTY_UNCOMPILED_CS;
962}
963
964static void *
965v3d_create_compute_state(struct pipe_context *pctx,
966                         const struct pipe_compute_state *cso)
967{
968        return v3d_uncompiled_shader_create(pctx, cso->ir_type,
969                                            (void *)cso->prog);
970}
971
972void
973v3d_program_init(struct pipe_context *pctx)
974{
975        struct v3d_context *v3d = v3d_context(pctx);
976
977        pctx->create_vs_state = v3d_shader_state_create;
978        pctx->delete_vs_state = v3d_shader_state_delete;
979
980        pctx->create_gs_state = v3d_shader_state_create;
981        pctx->delete_gs_state = v3d_shader_state_delete;
982
983        pctx->create_fs_state = v3d_shader_state_create;
984        pctx->delete_fs_state = v3d_shader_state_delete;
985
986        pctx->bind_fs_state = v3d_fp_state_bind;
987        pctx->bind_gs_state = v3d_gp_state_bind;
988        pctx->bind_vs_state = v3d_vp_state_bind;
989
990        if (v3d->screen->has_csd) {
991                pctx->create_compute_state = v3d_create_compute_state;
992                pctx->delete_compute_state = v3d_shader_state_delete;
993                pctx->bind_compute_state = v3d_compute_state_bind;
994        }
995
996        v3d->prog.cache[MESA_SHADER_VERTEX] =
997                _mesa_hash_table_create(pctx, vs_cache_hash, vs_cache_compare);
998        v3d->prog.cache[MESA_SHADER_GEOMETRY] =
999                _mesa_hash_table_create(pctx, gs_cache_hash, gs_cache_compare);
1000        v3d->prog.cache[MESA_SHADER_FRAGMENT] =
1001                _mesa_hash_table_create(pctx, fs_cache_hash, fs_cache_compare);
1002        v3d->prog.cache[MESA_SHADER_COMPUTE] =
1003                _mesa_hash_table_create(pctx, cs_cache_hash, cs_cache_compare);
1004}
1005
1006void
1007v3d_program_fini(struct pipe_context *pctx)
1008{
1009        struct v3d_context *v3d = v3d_context(pctx);
1010
1011        for (int i = 0; i < MESA_SHADER_STAGES; i++) {
1012                struct hash_table *cache = v3d->prog.cache[i];
1013                if (!cache)
1014                        continue;
1015
1016                hash_table_foreach(cache, entry) {
1017                        struct v3d_compiled_shader *shader = entry->data;
1018                        v3d_free_compiled_shader(shader);
1019                        _mesa_hash_table_remove(cache, entry);
1020                }
1021        }
1022
1023        v3d_bo_unreference(&v3d->prog.spill_bo);
1024}
1025