v3d_blit.c revision 7ec681f3
1/*
2 * Copyright © 2015-2017 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 "util/format/u_format.h"
25#include "util/u_surface.h"
26#include "util/u_blitter.h"
27#include "compiler/nir/nir_builder.h"
28#include "v3d_context.h"
29#include "broadcom/common/v3d_tiling.h"
30
31void
32v3d_blitter_save(struct v3d_context *v3d)
33{
34        util_blitter_save_fragment_constant_buffer_slot(v3d->blitter,
35                                                        v3d->constbuf[PIPE_SHADER_FRAGMENT].cb);
36        util_blitter_save_vertex_buffer_slot(v3d->blitter, v3d->vertexbuf.vb);
37        util_blitter_save_vertex_elements(v3d->blitter, v3d->vtx);
38        util_blitter_save_vertex_shader(v3d->blitter, v3d->prog.bind_vs);
39        util_blitter_save_geometry_shader(v3d->blitter, v3d->prog.bind_gs);
40        util_blitter_save_so_targets(v3d->blitter, v3d->streamout.num_targets,
41                                     v3d->streamout.targets);
42        util_blitter_save_rasterizer(v3d->blitter, v3d->rasterizer);
43        util_blitter_save_viewport(v3d->blitter, &v3d->viewport);
44        util_blitter_save_scissor(v3d->blitter, &v3d->scissor);
45        util_blitter_save_fragment_shader(v3d->blitter, v3d->prog.bind_fs);
46        util_blitter_save_blend(v3d->blitter, v3d->blend);
47        util_blitter_save_depth_stencil_alpha(v3d->blitter, v3d->zsa);
48        util_blitter_save_stencil_ref(v3d->blitter, &v3d->stencil_ref);
49        util_blitter_save_sample_mask(v3d->blitter, v3d->sample_mask);
50        util_blitter_save_framebuffer(v3d->blitter, &v3d->framebuffer);
51        util_blitter_save_fragment_sampler_states(v3d->blitter,
52                        v3d->tex[PIPE_SHADER_FRAGMENT].num_samplers,
53                        (void **)v3d->tex[PIPE_SHADER_FRAGMENT].samplers);
54        util_blitter_save_fragment_sampler_views(v3d->blitter,
55                        v3d->tex[PIPE_SHADER_FRAGMENT].num_textures,
56                        v3d->tex[PIPE_SHADER_FRAGMENT].textures);
57        util_blitter_save_so_targets(v3d->blitter, v3d->streamout.num_targets,
58                                     v3d->streamout.targets);
59}
60
61static void
62v3d_render_blit(struct pipe_context *ctx, struct pipe_blit_info *info)
63{
64        struct v3d_context *v3d = v3d_context(ctx);
65        struct v3d_resource *src = v3d_resource(info->src.resource);
66        struct pipe_resource *tiled = NULL;
67
68        if (!info->mask)
69                return;
70
71        if (!src->tiled) {
72                struct pipe_box box = {
73                        .x = 0,
74                        .y = 0,
75                        .width = u_minify(info->src.resource->width0,
76                                           info->src.level),
77                        .height = u_minify(info->src.resource->height0,
78                                           info->src.level),
79                        .depth = 1,
80                };
81                struct pipe_resource tmpl = {
82                        .target = info->src.resource->target,
83                        .format = info->src.resource->format,
84                        .width0 = box.width,
85                        .height0 = box.height,
86                        .depth0 = 1,
87                        .array_size = 1,
88                };
89                tiled = ctx->screen->resource_create(ctx->screen, &tmpl);
90                if (!tiled) {
91                        fprintf(stderr, "Failed to create tiled blit temp\n");
92                        return;
93                }
94                ctx->resource_copy_region(ctx,
95                                          tiled, 0,
96                                          0, 0, 0,
97                                          info->src.resource, info->src.level,
98                                          &box);
99                info->src.level = 0;
100                info->src.resource = tiled;
101        }
102
103        if (!util_blitter_is_blit_supported(v3d->blitter, info)) {
104                fprintf(stderr, "blit unsupported %s -> %s\n",
105                    util_format_short_name(info->src.resource->format),
106                    util_format_short_name(info->dst.resource->format));
107                return;
108        }
109
110        v3d_blitter_save(v3d);
111        util_blitter_blit(v3d->blitter, info);
112
113        pipe_resource_reference(&tiled, NULL);
114        info->mask = 0;
115}
116
117/* Implement stencil blits by reinterpreting the stencil data as an RGBA8888
118 * or R8 texture.
119 */
120static void
121v3d_stencil_blit(struct pipe_context *ctx, struct pipe_blit_info *info)
122{
123        struct v3d_context *v3d = v3d_context(ctx);
124        struct v3d_resource *src = v3d_resource(info->src.resource);
125        struct v3d_resource *dst = v3d_resource(info->dst.resource);
126        enum pipe_format src_format, dst_format;
127
128        if ((info->mask & PIPE_MASK_S) == 0)
129                return;
130
131        if (src->separate_stencil) {
132                src = src->separate_stencil;
133                src_format = PIPE_FORMAT_R8_UINT;
134        } else {
135                src_format = PIPE_FORMAT_RGBA8888_UINT;
136        }
137
138        if (dst->separate_stencil) {
139                dst = dst->separate_stencil;
140                dst_format = PIPE_FORMAT_R8_UINT;
141        } else {
142                dst_format = PIPE_FORMAT_RGBA8888_UINT;
143        }
144
145        /* Initialize the surface. */
146        struct pipe_surface dst_tmpl = {
147                .u.tex = {
148                        .level = info->dst.level,
149                        .first_layer = info->dst.box.z,
150                        .last_layer = info->dst.box.z,
151                },
152                .format = dst_format,
153        };
154        struct pipe_surface *dst_surf =
155                ctx->create_surface(ctx, &dst->base, &dst_tmpl);
156
157        /* Initialize the sampler view. */
158        struct pipe_sampler_view src_tmpl = {
159                .target = src->base.target,
160                .format = src_format,
161                .u.tex = {
162                        .first_level = info->src.level,
163                        .last_level = info->src.level,
164                        .first_layer = 0,
165                        .last_layer = (PIPE_TEXTURE_3D ?
166                                       u_minify(src->base.depth0,
167                                                info->src.level) - 1 :
168                                       src->base.array_size - 1),
169                },
170                .swizzle_r = PIPE_SWIZZLE_X,
171                .swizzle_g = PIPE_SWIZZLE_Y,
172                .swizzle_b = PIPE_SWIZZLE_Z,
173                .swizzle_a = PIPE_SWIZZLE_W,
174        };
175        struct pipe_sampler_view *src_view =
176                ctx->create_sampler_view(ctx, &src->base, &src_tmpl);
177
178        v3d_blitter_save(v3d);
179        util_blitter_blit_generic(v3d->blitter, dst_surf, &info->dst.box,
180                                  src_view, &info->src.box,
181                                  src->base.width0, src->base.height0,
182                                  PIPE_MASK_R,
183                                  PIPE_TEX_FILTER_NEAREST,
184                                  info->scissor_enable ? &info->scissor : NULL,
185                                  info->alpha_blend, false);
186
187        pipe_surface_reference(&dst_surf, NULL);
188        pipe_sampler_view_reference(&src_view, NULL);
189
190        info->mask &= ~PIPE_MASK_S;
191}
192
193/* Disable level 0 write, just write following mipmaps */
194#define V3D_TFU_IOA_DIMTW (1 << 0)
195#define V3D_TFU_IOA_FORMAT_SHIFT 3
196#define V3D_TFU_IOA_FORMAT_LINEARTILE 3
197#define V3D_TFU_IOA_FORMAT_UBLINEAR_1_COLUMN 4
198#define V3D_TFU_IOA_FORMAT_UBLINEAR_2_COLUMN 5
199#define V3D_TFU_IOA_FORMAT_UIF_NO_XOR 6
200#define V3D_TFU_IOA_FORMAT_UIF_XOR 7
201
202#define V3D_TFU_ICFG_NUMMM_SHIFT 5
203#define V3D_TFU_ICFG_TTYPE_SHIFT 9
204
205#define V3D_TFU_ICFG_OPAD_SHIFT 22
206
207#define V3D_TFU_ICFG_FORMAT_SHIFT 18
208#define V3D_TFU_ICFG_FORMAT_RASTER 0
209#define V3D_TFU_ICFG_FORMAT_SAND_128 1
210#define V3D_TFU_ICFG_FORMAT_SAND_256 2
211#define V3D_TFU_ICFG_FORMAT_LINEARTILE 11
212#define V3D_TFU_ICFG_FORMAT_UBLINEAR_1_COLUMN 12
213#define V3D_TFU_ICFG_FORMAT_UBLINEAR_2_COLUMN 13
214#define V3D_TFU_ICFG_FORMAT_UIF_NO_XOR 14
215#define V3D_TFU_ICFG_FORMAT_UIF_XOR 15
216
217static bool
218v3d_tfu(struct pipe_context *pctx,
219        struct pipe_resource *pdst,
220        struct pipe_resource *psrc,
221        unsigned int src_level,
222        unsigned int base_level,
223        unsigned int last_level,
224        unsigned int src_layer,
225        unsigned int dst_layer,
226        bool for_mipmap)
227{
228        struct v3d_context *v3d = v3d_context(pctx);
229        struct v3d_screen *screen = v3d->screen;
230        struct v3d_resource *src = v3d_resource(psrc);
231        struct v3d_resource *dst = v3d_resource(pdst);
232        struct v3d_resource_slice *src_base_slice = &src->slices[src_level];
233        struct v3d_resource_slice *dst_base_slice = &dst->slices[base_level];
234        int msaa_scale = pdst->nr_samples > 1 ? 2 : 1;
235        int width = u_minify(pdst->width0, base_level) * msaa_scale;
236        int height = u_minify(pdst->height0, base_level) * msaa_scale;
237        enum pipe_format pformat;
238
239        if (psrc->format != pdst->format)
240                return false;
241        if (psrc->nr_samples != pdst->nr_samples)
242                return false;
243
244        if (pdst->target != PIPE_TEXTURE_2D || psrc->target != PIPE_TEXTURE_2D)
245                return false;
246
247        /* Can't write to raster. */
248        if (dst_base_slice->tiling == V3D_TILING_RASTER)
249                return false;
250
251        /* When using TFU for blit, we are doing exact copies (both input and
252         * output format must be the same, no scaling, etc), so there is no
253         * pixel format conversions. Thus we can rewrite the format to use one
254         * that is TFU compatible based on its texel size.
255         */
256        if (for_mipmap) {
257                pformat = pdst->format;
258        } else {
259                switch (dst->cpp) {
260                case 16: pformat = PIPE_FORMAT_R32G32B32A32_FLOAT;   break;
261                case 8:  pformat = PIPE_FORMAT_R16G16B16A16_FLOAT;   break;
262                case 4:  pformat = PIPE_FORMAT_R32_FLOAT;            break;
263                case 2:  pformat = PIPE_FORMAT_R16_FLOAT;            break;
264                case 1:  pformat = PIPE_FORMAT_R8_UNORM;             break;
265                default: unreachable("unsupported format bit-size"); break;
266                };
267        }
268
269        uint32_t tex_format = v3d_get_tex_format(&screen->devinfo, pformat);
270
271        if (!v3d_tfu_supports_tex_format(&screen->devinfo, tex_format, for_mipmap)) {
272                assert(for_mipmap);
273                return false;
274        }
275
276        v3d_flush_jobs_writing_resource(v3d, psrc, V3D_FLUSH_DEFAULT, false);
277        v3d_flush_jobs_reading_resource(v3d, pdst, V3D_FLUSH_DEFAULT, false);
278
279        struct drm_v3d_submit_tfu tfu = {
280                .ios = (height << 16) | width,
281                .bo_handles = {
282                        dst->bo->handle,
283                        src != dst ? src->bo->handle : 0
284                },
285                .in_sync = v3d->out_sync,
286                .out_sync = v3d->out_sync,
287        };
288        uint32_t src_offset = (src->bo->offset +
289                               v3d_layer_offset(psrc, src_level, src_layer));
290        tfu.iia |= src_offset;
291        if (src_base_slice->tiling == V3D_TILING_RASTER) {
292                tfu.icfg |= (V3D_TFU_ICFG_FORMAT_RASTER <<
293                             V3D_TFU_ICFG_FORMAT_SHIFT);
294        } else {
295                tfu.icfg |= ((V3D_TFU_ICFG_FORMAT_LINEARTILE +
296                              (src_base_slice->tiling - V3D_TILING_LINEARTILE)) <<
297                             V3D_TFU_ICFG_FORMAT_SHIFT);
298        }
299
300        uint32_t dst_offset = (dst->bo->offset +
301                               v3d_layer_offset(pdst, base_level, dst_layer));
302        tfu.ioa |= dst_offset;
303        if (last_level != base_level)
304                tfu.ioa |= V3D_TFU_IOA_DIMTW;
305        tfu.ioa |= ((V3D_TFU_IOA_FORMAT_LINEARTILE +
306                     (dst_base_slice->tiling - V3D_TILING_LINEARTILE)) <<
307                    V3D_TFU_IOA_FORMAT_SHIFT);
308
309        tfu.icfg |= tex_format << V3D_TFU_ICFG_TTYPE_SHIFT;
310        tfu.icfg |= (last_level - base_level) << V3D_TFU_ICFG_NUMMM_SHIFT;
311
312        switch (src_base_slice->tiling) {
313        case V3D_TILING_UIF_NO_XOR:
314        case V3D_TILING_UIF_XOR:
315                tfu.iis |= (src_base_slice->padded_height /
316                            (2 * v3d_utile_height(src->cpp)));
317                break;
318        case V3D_TILING_RASTER:
319                tfu.iis |= src_base_slice->stride / src->cpp;
320                break;
321        case V3D_TILING_LINEARTILE:
322        case V3D_TILING_UBLINEAR_1_COLUMN:
323        case V3D_TILING_UBLINEAR_2_COLUMN:
324                break;
325       }
326
327        /* If we're writing level 0 (!IOA_DIMTW), then we need to supply the
328         * OPAD field for the destination (how many extra UIF blocks beyond
329         * those necessary to cover the height).  When filling mipmaps, the
330         * miplevel 1+ tiling state is inferred.
331         */
332        if (dst_base_slice->tiling == V3D_TILING_UIF_NO_XOR ||
333            dst_base_slice->tiling == V3D_TILING_UIF_XOR) {
334                int uif_block_h = 2 * v3d_utile_height(dst->cpp);
335                int implicit_padded_height = align(height, uif_block_h);
336
337                tfu.icfg |= (((dst_base_slice->padded_height -
338                               implicit_padded_height) / uif_block_h) <<
339                             V3D_TFU_ICFG_OPAD_SHIFT);
340        }
341
342        int ret = v3d_ioctl(screen->fd, DRM_IOCTL_V3D_SUBMIT_TFU, &tfu);
343        if (ret != 0) {
344                fprintf(stderr, "Failed to submit TFU job: %d\n", ret);
345                return false;
346        }
347
348        dst->writes++;
349
350        return true;
351}
352
353bool
354v3d_generate_mipmap(struct pipe_context *pctx,
355                    struct pipe_resource *prsc,
356                    enum pipe_format format,
357                    unsigned int base_level,
358                    unsigned int last_level,
359                    unsigned int first_layer,
360                    unsigned int last_layer)
361{
362        if (format != prsc->format)
363                return false;
364
365        /* We could maybe support looping over layers for array textures, but
366         * we definitely don't support 3D.
367         */
368        if (first_layer != last_layer)
369                return false;
370
371        return v3d_tfu(pctx,
372                       prsc, prsc,
373                       base_level,
374                       base_level, last_level,
375                       first_layer, first_layer,
376                       true);
377}
378
379static void
380v3d_tfu_blit(struct pipe_context *pctx, struct pipe_blit_info *info)
381{
382        int dst_width = u_minify(info->dst.resource->width0, info->dst.level);
383        int dst_height = u_minify(info->dst.resource->height0, info->dst.level);
384
385        if ((info->mask & PIPE_MASK_RGBA) == 0)
386                return;
387
388        if (info->scissor_enable ||
389            info->dst.box.x != 0 ||
390            info->dst.box.y != 0 ||
391            info->dst.box.width != dst_width ||
392            info->dst.box.height != dst_height ||
393            info->src.box.x != 0 ||
394            info->src.box.y != 0 ||
395            info->src.box.width != info->dst.box.width ||
396            info->src.box.height != info->dst.box.height) {
397                return;
398        }
399
400        if (info->dst.format != info->src.format)
401                return;
402
403        if (v3d_tfu(pctx, info->dst.resource, info->src.resource,
404                    info->src.level,
405                    info->dst.level, info->dst.level,
406                    info->src.box.z, info->dst.box.z,
407                    false)) {
408                info->mask &= ~PIPE_MASK_RGBA;
409        }
410}
411
412static struct pipe_surface *
413v3d_get_blit_surface(struct pipe_context *pctx,
414                     struct pipe_resource *prsc,
415                     unsigned level,
416                     int16_t layer)
417{
418        struct pipe_surface tmpl;
419
420        tmpl.format = prsc->format;
421        tmpl.u.tex.level = level;
422        tmpl.u.tex.first_layer = layer;
423        tmpl.u.tex.last_layer = layer;
424
425        return pctx->create_surface(pctx, prsc, &tmpl);
426}
427
428static bool
429is_tile_unaligned(unsigned size, unsigned tile_size)
430{
431        return size & (tile_size - 1);
432}
433
434static void
435v3d_tlb_blit(struct pipe_context *pctx, struct pipe_blit_info *info)
436{
437        struct v3d_context *v3d = v3d_context(pctx);
438        struct v3d_screen *screen = v3d->screen;
439
440        if (screen->devinfo.ver < 40 || !info->mask)
441                return;
442
443        bool is_color_blit = info->mask & PIPE_MASK_RGBA;
444        bool is_depth_blit = info->mask & PIPE_MASK_Z;
445        bool is_stencil_blit = info->mask & PIPE_MASK_S;
446
447        /* We should receive either a depth/stencil blit, or color blit, but
448         * not both.
449         */
450        assert ((is_color_blit && !is_depth_blit && !is_stencil_blit) ||
451                (!is_color_blit && (is_depth_blit || is_stencil_blit)));
452
453        if (info->scissor_enable)
454                return;
455
456        if (info->src.box.x != info->dst.box.x ||
457            info->src.box.y != info->dst.box.y ||
458            info->src.box.width != info->dst.box.width ||
459            info->src.box.height != info->dst.box.height)
460                return;
461
462        if (is_color_blit &&
463            util_format_is_depth_or_stencil(info->dst.resource->format))
464                return;
465
466        if (!v3d_rt_format_supported(&screen->devinfo, info->src.resource->format))
467                return;
468
469        if (v3d_get_rt_format(&screen->devinfo, info->src.resource->format) !=
470            v3d_get_rt_format(&screen->devinfo, info->dst.resource->format))
471                return;
472
473        bool msaa = (info->src.resource->nr_samples > 1 ||
474                     info->dst.resource->nr_samples > 1);
475        bool is_msaa_resolve = (info->src.resource->nr_samples > 1 &&
476                                info->dst.resource->nr_samples < 2);
477
478        if (is_msaa_resolve &&
479            !v3d_format_supports_tlb_msaa_resolve(&screen->devinfo, info->src.resource->format))
480                return;
481
482        v3d_flush_jobs_writing_resource(v3d, info->src.resource, V3D_FLUSH_DEFAULT, false);
483
484        struct pipe_surface *dst_surf =
485           v3d_get_blit_surface(pctx, info->dst.resource, info->dst.level, info->dst.box.z);
486        struct pipe_surface *src_surf =
487           v3d_get_blit_surface(pctx, info->src.resource, info->src.level, info->src.box.z);
488
489        struct pipe_surface *surfaces[V3D_MAX_DRAW_BUFFERS] = { 0 };
490        if (is_color_blit)
491                surfaces[0] = dst_surf;
492
493        uint32_t tile_width, tile_height, max_bpp;
494        v3d_get_tile_buffer_size(msaa, is_color_blit ? 1 : 0, surfaces, src_surf, &tile_width, &tile_height, &max_bpp);
495
496        int dst_surface_width = u_minify(info->dst.resource->width0,
497                                         info->dst.level);
498        int dst_surface_height = u_minify(info->dst.resource->height0,
499                                         info->dst.level);
500        if (is_tile_unaligned(info->dst.box.x, tile_width) ||
501            is_tile_unaligned(info->dst.box.y, tile_height) ||
502            (is_tile_unaligned(info->dst.box.width, tile_width) &&
503             info->dst.box.x + info->dst.box.width != dst_surface_width) ||
504            (is_tile_unaligned(info->dst.box.height, tile_height) &&
505             info->dst.box.y + info->dst.box.height != dst_surface_height)) {
506                pipe_surface_reference(&dst_surf, NULL);
507                pipe_surface_reference(&src_surf, NULL);
508                return;
509        }
510
511        struct v3d_job *job = v3d_get_job(v3d,
512                                          is_color_blit ? 1u : 0u,
513                                          surfaces,
514                                          is_color_blit ? NULL : dst_surf,
515                                          src_surf);
516        job->msaa = msaa;
517        job->tile_width = tile_width;
518        job->tile_height = tile_height;
519        job->internal_bpp = max_bpp;
520        job->draw_min_x = info->dst.box.x;
521        job->draw_min_y = info->dst.box.y;
522        job->draw_max_x = info->dst.box.x + info->dst.box.width;
523        job->draw_max_y = info->dst.box.y + info->dst.box.height;
524        job->scissor.disabled = false;
525
526        /* The simulator complains if we do a TLB load from a source with a
527         * stride that is smaller than the destination's, so we program the
528         * 'frame region' to match the smallest dimensions of the two surfaces.
529         * This should be fine because we only get here if the src and dst boxes
530         * match, so we know the blit involves the same tiles on both surfaces.
531         */
532        job->draw_width = MIN2(dst_surf->width, src_surf->width);
533        job->draw_height = MIN2(dst_surf->height, src_surf->height);
534        job->draw_tiles_x = DIV_ROUND_UP(job->draw_width,
535                                         job->tile_width);
536        job->draw_tiles_y = DIV_ROUND_UP(job->draw_height,
537                                         job->tile_height);
538
539        job->needs_flush = true;
540        job->num_layers = info->dst.box.depth;
541
542        job->store = 0;
543        if (is_color_blit) {
544                job->store |= PIPE_CLEAR_COLOR0;
545                info->mask &= ~PIPE_MASK_RGBA;
546        }
547        if (is_depth_blit) {
548                job->store |= PIPE_CLEAR_DEPTH;
549                info->mask &= ~PIPE_MASK_Z;
550        }
551        if (is_stencil_blit){
552                job->store |= PIPE_CLEAR_STENCIL;
553                info->mask &= ~PIPE_MASK_S;
554        }
555
556        v3d41_start_binning(v3d, job);
557
558        v3d_job_submit(v3d, job);
559
560        pipe_surface_reference(&dst_surf, NULL);
561        pipe_surface_reference(&src_surf, NULL);
562}
563
564/**
565 * Creates the VS of the custom blit shader to convert YUV plane from
566 * the NV12 format with BROADCOM_SAND_COL128 modifier to UIF tiled format.
567 * This vertex shader is mostly a pass-through VS.
568 */
569static void *
570v3d_get_sand8_vs(struct pipe_context *pctx)
571{
572        struct v3d_context *v3d = v3d_context(pctx);
573        struct pipe_screen *pscreen = pctx->screen;
574
575        if (v3d->sand8_blit_vs)
576                return v3d->sand8_blit_vs;
577
578        const struct nir_shader_compiler_options *options =
579                pscreen->get_compiler_options(pscreen,
580                                              PIPE_SHADER_IR_NIR,
581                                              PIPE_SHADER_VERTEX);
582
583        nir_builder b = nir_builder_init_simple_shader(MESA_SHADER_VERTEX,
584                                                       options,
585                                                       "sand8_blit_vs");
586
587        const struct glsl_type *vec4 = glsl_vec4_type();
588        nir_variable *pos_in = nir_variable_create(b.shader,
589                                                   nir_var_shader_in,
590                                                   vec4, "pos");
591
592        nir_variable *pos_out = nir_variable_create(b.shader,
593                                                    nir_var_shader_out,
594                                                    vec4, "gl_Position");
595        pos_out->data.location = VARYING_SLOT_POS;
596        nir_store_var(&b, pos_out, nir_load_var(&b, pos_in), 0xf);
597
598        struct pipe_shader_state shader_tmpl = {
599                .type = PIPE_SHADER_IR_NIR,
600                .ir.nir = b.shader,
601        };
602
603        v3d->sand8_blit_vs = pctx->create_vs_state(pctx, &shader_tmpl);
604
605        return v3d->sand8_blit_vs;
606}
607/**
608 * Creates the FS of the custom blit shader to convert YUV plane from
609 * the NV12 format with BROADCOM_SAND_COL128 modifier to UIF tiled format.
610 * The result texture is equivalent to a chroma (cpp=2) or luma (cpp=1)
611 * plane for a NV12 format without the SAND modifier.
612 */
613static void *
614v3d_get_sand8_fs(struct pipe_context *pctx, int cpp)
615{
616        struct v3d_context *v3d = v3d_context(pctx);
617        struct pipe_screen *pscreen = pctx->screen;
618        struct pipe_shader_state **cached_shader;
619        const char *name;
620
621        if (cpp == 1) {
622                cached_shader = &v3d->sand8_blit_fs_luma;
623                name = "sand8_blit_fs_luma";
624        } else {
625                cached_shader = &v3d->sand8_blit_fs_chroma;
626                name = "sand8_blit_fs_chroma";
627        }
628
629        if (*cached_shader)
630                return *cached_shader;
631
632        const struct nir_shader_compiler_options *options =
633                pscreen->get_compiler_options(pscreen,
634                                              PIPE_SHADER_IR_NIR,
635                                              PIPE_SHADER_FRAGMENT);
636
637        nir_builder b = nir_builder_init_simple_shader(MESA_SHADER_FRAGMENT,
638                                                       options, "%s", name);
639        const struct glsl_type *vec4 = glsl_vec4_type();
640
641        const struct glsl_type *glsl_int = glsl_int_type();
642
643        nir_variable *color_out =
644                nir_variable_create(b.shader, nir_var_shader_out,
645                                    vec4, "f_color");
646        color_out->data.location = FRAG_RESULT_COLOR;
647
648        nir_variable *pos_in =
649                nir_variable_create(b.shader, nir_var_shader_in, vec4, "pos");
650        pos_in->data.location = VARYING_SLOT_POS;
651        nir_ssa_def *pos = nir_load_var(&b, pos_in);
652
653        nir_ssa_def *zero = nir_imm_int(&b, 0);
654        nir_ssa_def *one = nir_imm_int(&b, 1);
655        nir_ssa_def *two = nir_imm_int(&b, 2);
656        nir_ssa_def *six = nir_imm_int(&b, 6);
657        nir_ssa_def *seven = nir_imm_int(&b, 7);
658        nir_ssa_def *eight = nir_imm_int(&b, 8);
659
660        nir_ssa_def *x = nir_f2i32(&b, nir_channel(&b, pos, 0));
661        nir_ssa_def *y = nir_f2i32(&b, nir_channel(&b, pos, 1));
662
663        nir_variable *stride_in =
664                nir_variable_create(b.shader, nir_var_uniform, glsl_int,
665                                    "sand8_stride");
666        nir_ssa_def *stride =
667                nir_load_uniform(&b, 1, 32, zero,
668                                 .base = stride_in->data.driver_location,
669                                 .range = 1,
670                                 .dest_type = nir_type_int32);
671
672        nir_ssa_def *x_offset;
673        nir_ssa_def *y_offset;
674
675        /* UIF tiled format is composed by UIF blocks, Each block has
676         * four 64 byte microtiles. Inside each microtile pixels are stored
677         * in raster format. But microtiles have different dimensions
678         * based in the bits per pixel of the image.
679         *
680         *   8bpp microtile dimensions are 8x8
681         *  16bpp microtile dimensions are 8x4
682         *  32bpp microtile dimensions are 4x4
683         *
684         * As we are reading and writing with 32bpp to optimize
685         * the number of texture operations during the blit. We need
686         * to adjust the offsets were we read and write as data will
687         * be later read using 8bpp (luma) and 16bpp (chroma).
688         *
689         * For chroma 8x4 16bpp raster order is compatible with 4x4
690         * 32bpp. In both layouts each line has 8*2 == 4*4 == 16 bytes.
691         * But luma 8x8 8bpp raster order is not compatible
692         * with 4x4 32bpp. 8bpp has 8 bytes per line, and 32bpp has
693         * 16 bytes per line. So if we read a 8bpp texture that was
694         * written as 32bpp texture. Bytes would be misplaced.
695         *
696         * inter/intra_utile_x_offests takes care of mapping the offsets
697         * between microtiles to deal with this issue for luma planes.
698         */
699        if (cpp == 1) {
700                nir_ssa_def *intra_utile_x_offset =
701                        nir_ishl(&b, nir_iand_imm(&b, x, 1), two);
702                nir_ssa_def *inter_utile_x_offset =
703                        nir_ishl(&b, nir_iand_imm(&b, x, 60), one);
704                nir_ssa_def *stripe_offset=
705                        nir_ishl(&b,nir_imul(&b,nir_ishr_imm(&b, x, 6),
706                                             stride),
707                                 seven);
708
709                x_offset = nir_iadd(&b, stripe_offset,
710                                        nir_iadd(&b, intra_utile_x_offset,
711                                                     inter_utile_x_offset));
712                y_offset = nir_iadd(&b,
713                                    nir_ishl(&b, nir_iand_imm(&b, x, 2), six),
714                                    nir_ishl(&b, y, eight));
715        } else  {
716                nir_ssa_def *stripe_offset=
717                        nir_ishl(&b,nir_imul(&b,nir_ishr_imm(&b, x, 5),
718                                                stride),
719                                 seven);
720                x_offset = nir_iadd(&b, stripe_offset,
721                               nir_ishl(&b, nir_iand_imm(&b, x, 31), two));
722                y_offset = nir_ishl(&b, y, seven);
723        }
724        nir_ssa_def *ubo_offset = nir_iadd(&b, x_offset, y_offset);
725        nir_ssa_def *load =
726        nir_load_ubo(&b, 1, 32, one, ubo_offset,
727                    .align_mul = 4,
728                    .align_offset = 0,
729                    .range_base = 0,
730                    .range = ~0);
731
732        nir_ssa_def *output = nir_unpack_unorm_4x8(&b, load);
733
734        nir_store_var(&b, color_out,
735                      output,
736                      0xF);
737
738        struct pipe_shader_state shader_tmpl = {
739                .type = PIPE_SHADER_IR_NIR,
740                .ir.nir = b.shader,
741        };
742
743        *cached_shader = pctx->create_fs_state(pctx, &shader_tmpl);
744
745        return *cached_shader;
746}
747
748/**
749 * Turns NV12 with SAND8 format modifier from raster-order with interleaved
750 * luma and chroma 128-byte-wide-columns to tiled format for luma and chroma.
751 *
752 * This implementation is based on vc4_yuv_blit.
753 */
754static void
755v3d_sand8_blit(struct pipe_context *pctx, struct pipe_blit_info *info)
756{
757        struct v3d_context *v3d = v3d_context(pctx);
758        struct v3d_resource *src = v3d_resource(info->src.resource);
759        ASSERTED struct v3d_resource *dst = v3d_resource(info->dst.resource);
760
761        if (!src->sand_col128_stride)
762                return;
763        if (src->tiled)
764                return;
765        if (src->base.format != PIPE_FORMAT_R8_UNORM &&
766            src->base.format != PIPE_FORMAT_R8G8_UNORM)
767                return;
768        if (!(info->mask & PIPE_MASK_RGBA))
769                return;
770
771        assert(dst->base.format == src->base.format);
772        assert(dst->tiled);
773
774        assert(info->src.box.x == 0 && info->dst.box.x == 0);
775        assert(info->src.box.y == 0 && info->dst.box.y == 0);
776        assert(info->src.box.width == info->dst.box.width);
777        assert(info->src.box.height == info->dst.box.height);
778
779        v3d_blitter_save(v3d);
780
781        struct pipe_surface dst_tmpl;
782        util_blitter_default_dst_texture(&dst_tmpl, info->dst.resource,
783                                         info->dst.level, info->dst.box.z);
784        /* Although the src textures are cpp=1 or cpp=2, the dst texture
785         * uses a cpp=4 dst texture. So, all read/write texture ops will
786         * be done using 32-bit read and writes.
787         */
788        dst_tmpl.format = PIPE_FORMAT_R8G8B8A8_UNORM;
789        struct pipe_surface *dst_surf =
790                pctx->create_surface(pctx, info->dst.resource, &dst_tmpl);
791        if (!dst_surf) {
792                fprintf(stderr, "Failed to create YUV dst surface\n");
793                util_blitter_unset_running_flag(v3d->blitter);
794                return;
795        }
796
797        uint32_t sand8_stride = src->sand_col128_stride;
798
799        /* Adjust the dimensions of dst luma/chroma to match src
800         * size now we are using a cpp=4 format. Next dimension take into
801         * account the UIF microtile layouts.
802         */
803        dst_surf->width = align(dst_surf->width, 8) / 2;
804        if (src->cpp == 1)
805                dst_surf->height /= 2;
806
807        /* Set the constant buffer. */
808        struct pipe_constant_buffer cb_uniforms = {
809                .user_buffer = &sand8_stride,
810                .buffer_size = sizeof(sand8_stride),
811        };
812
813        pctx->set_constant_buffer(pctx, PIPE_SHADER_FRAGMENT, 0, false,
814                                  &cb_uniforms);
815        struct pipe_constant_buffer cb_src = {
816                .buffer = info->src.resource,
817                .buffer_offset = src->slices[info->src.level].offset,
818                .buffer_size = (src->bo->size -
819                                src->slices[info->src.level].offset),
820        };
821        pctx->set_constant_buffer(pctx, PIPE_SHADER_FRAGMENT, 2, false,
822                                  &cb_src);
823        /* Unbind the textures, to make sure we don't try to recurse into the
824         * shadow blit.
825         */
826        pctx->set_sampler_views(pctx, PIPE_SHADER_FRAGMENT, 0, 0, 0, false, NULL);
827        pctx->bind_sampler_states(pctx, PIPE_SHADER_FRAGMENT, 0, 0, NULL);
828
829        util_blitter_custom_shader(v3d->blitter, dst_surf,
830                                   v3d_get_sand8_vs(pctx),
831                                   v3d_get_sand8_fs(pctx, src->cpp));
832
833        util_blitter_restore_textures(v3d->blitter);
834        util_blitter_restore_constant_buffer_state(v3d->blitter);
835
836        /* Restore cb1 (util_blitter doesn't handle this one). */
837        struct pipe_constant_buffer cb_disabled = { 0 };
838        pctx->set_constant_buffer(pctx, PIPE_SHADER_FRAGMENT, 1, false,
839                                  &cb_disabled);
840
841        pipe_surface_reference(&dst_surf, NULL);
842
843        info->mask &= ~PIPE_MASK_RGBA;
844        return;
845}
846
847
848/* Optimal hardware path for blitting pixels.
849 * Scaling, format conversion, up- and downsampling (resolve) are allowed.
850 */
851void
852v3d_blit(struct pipe_context *pctx, const struct pipe_blit_info *blit_info)
853{
854        struct v3d_context *v3d = v3d_context(pctx);
855        struct pipe_blit_info info = *blit_info;
856
857        v3d_sand8_blit(pctx, &info);
858
859        v3d_tfu_blit(pctx, &info);
860
861        v3d_tlb_blit(pctx, &info);
862
863        v3d_stencil_blit(pctx, &info);
864
865        v3d_render_blit(pctx, &info);
866
867        /* Flush our blit jobs immediately.  They're unlikely to get reused by
868         * normal drawing or other blits, and without flushing we can easily
869         * run into unexpected OOMs when blits are used for a large series of
870         * texture uploads before using the textures.
871         */
872        v3d_flush_jobs_writing_resource(v3d, info.dst.resource,
873                                        V3D_FLUSH_DEFAULT, false);
874}
875