src/gallium/drivers/panfrost/pan_cmdstream.c

   1 /*
   2  * Copyright (C) 2018 Alyssa Rosenzweig
   3  * Copyright (C) 2020 Collabora Ltd.
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9  * and/or sell copies of the Software, and to permit persons to whom the
  10  * Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice (including the next
  13  * paragraph) shall be included in all copies or substantial portions of the
  14  * Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  22  * SOFTWARE.
  23  */
  24
  25 #include "util/macros.h"
  26
  27 #include "panfrost-quirks.h"
  28
  29 #include "pan_allocate.h"
  30 #include "pan_bo.h"
  31 #include "pan_cmdstream.h"
  32 #include "pan_context.h"
  33 #include "pan_job.h"
  34
  35 /* TODO: Bifrost requires just a mali_shared_memory, without the rest of the
  36  * framebuffer */
  37
  38 void
  39 panfrost_vt_attach_framebuffer(struct panfrost_context *ctx,
  40                                struct midgard_payload_vertex_tiler *vt)
  41 {
  42         struct panfrost_screen *screen = pan_screen(ctx->base.screen);
  43         struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx);
  44
  45         /* If we haven't, reserve space for the framebuffer */
  46
  47         if (!batch->framebuffer.gpu) {
  48                 unsigned size = (screen->quirks & MIDGARD_SFBD) ?
  49                         sizeof(struct mali_single_framebuffer) :
  50                         sizeof(struct mali_framebuffer);
  51
  52                 batch->framebuffer = panfrost_allocate_transient(batch, size);
  53
  54                 /* Tag the pointer */
  55                 if (!(screen->quirks & MIDGARD_SFBD))
  56                         batch->framebuffer.gpu |= MALI_MFBD;
  57         }
  58
  59         vt->postfix.shared_memory = batch->framebuffer.gpu;
  60 }
  61
  62 void
  63 panfrost_emit_shader_meta(struct panfrost_batch *batch,
  64                           enum pipe_shader_type st,
  65                           struct midgard_payload_vertex_tiler *vtp)
  66 {
  67         struct panfrost_context *ctx = batch->ctx;
  68         struct panfrost_shader_state *ss = panfrost_get_shader_state(ctx, st);
  69
  70         if (!ss) {
  71                 vtp->postfix.shader = 0;
  72                 return;
  73         }
  74
  75         /* Add the shader BO to the batch. */
  76         panfrost_batch_add_bo(batch, ss->bo,
  77                               PAN_BO_ACCESS_PRIVATE |
  78                               PAN_BO_ACCESS_READ |
  79                               panfrost_bo_access_for_stage(st));
  80
  81         vtp->postfix.shader = panfrost_upload_transient(batch, ss->tripipe,
  82                                                         sizeof(*ss->tripipe));
  83 }
  84
  85 static void
  86 panfrost_mali_viewport_init(struct panfrost_context *ctx,
  87                             struct mali_viewport *mvp)
  88 {
  89         const struct pipe_viewport_state *vp = &ctx->pipe_viewport;
  90
  91         /* Clip bounds are encoded as floats. The viewport itself is encoded as
  92          * (somewhat) asymmetric ints. */
  93
  94         const struct pipe_scissor_state *ss = &ctx->scissor;
  95
  96         memset(mvp, 0, sizeof(*mvp));
  97
  98         /* By default, do no viewport clipping, i.e. clip to (-inf, inf) in
  99          * each direction. Clipping to the viewport in theory should work, but
 100          * in practice causes issues when we're not explicitly trying to
 101          * scissor */
 102
 103         *mvp = (struct mali_viewport) {
 104                 .clip_minx = -INFINITY,
 105                 .clip_miny = -INFINITY,
 106                 .clip_maxx = INFINITY,
 107                 .clip_maxy = INFINITY,
 108         };
 109
 110         /* Always scissor to the viewport by default. */
 111         float vp_minx = (int) (vp->translate[0] - fabsf(vp->scale[0]));
 112         float vp_maxx = (int) (vp->translate[0] + fabsf(vp->scale[0]));
 113
 114         float vp_miny = (int) (vp->translate[1] - fabsf(vp->scale[1]));
 115         float vp_maxy = (int) (vp->translate[1] + fabsf(vp->scale[1]));
 116
 117         float minz = (vp->translate[2] - fabsf(vp->scale[2]));
 118         float maxz = (vp->translate[2] + fabsf(vp->scale[2]));
 119
 120         /* Apply the scissor test */
 121
 122         unsigned minx, miny, maxx, maxy;
 123
 124         if (ss && ctx->rasterizer && ctx->rasterizer->base.scissor) {
 125                 minx = MAX2(ss->minx, vp_minx);
 126                 miny = MAX2(ss->miny, vp_miny);
 127                 maxx = MIN2(ss->maxx, vp_maxx);
 128                 maxy = MIN2(ss->maxy, vp_maxy);
 129         } else {
 130                 minx = vp_minx;
 131                 miny = vp_miny;
 132                 maxx = vp_maxx;
 133                 maxy = vp_maxy;
 134         }
 135
 136         /* Hardware needs the min/max to be strictly ordered, so flip if we
 137          * need to. The viewport transformation in the vertex shader will
 138          * handle the negatives if we don't */
 139
 140         if (miny > maxy) {
 141                 unsigned temp = miny;
 142                 miny = maxy;
 143                 maxy = temp;
 144         }
 145
 146         if (minx > maxx) {
 147                 unsigned temp = minx;
 148                 minx = maxx;
 149                 maxx = temp;
 150         }
 151
 152         if (minz > maxz) {
 153                 float temp = minz;
 154                 minz = maxz;
 155                 maxz = temp;
 156         }
 157
 158         /* Clamp to the framebuffer size as a last check */
 159
 160         minx = MIN2(ctx->pipe_framebuffer.width, minx);
 161         maxx = MIN2(ctx->pipe_framebuffer.width, maxx);
 162
 163         miny = MIN2(ctx->pipe_framebuffer.height, miny);
 164         maxy = MIN2(ctx->pipe_framebuffer.height, maxy);
 165
 166         /* Upload */
 167
 168         mvp->viewport0[0] = minx;
 169         mvp->viewport1[0] = MALI_POSITIVE(maxx);
 170
 171         mvp->viewport0[1] = miny;
 172         mvp->viewport1[1] = MALI_POSITIVE(maxy);
 173
 174         mvp->clip_minz = minz;
 175         mvp->clip_maxz = maxz;
 176 }
 177
 178 void
 179 panfrost_emit_viewport(struct panfrost_batch *batch,
 180                        struct midgard_payload_vertex_tiler *tp)
 181 {
 182         struct panfrost_context *ctx = batch->ctx;
 183         struct mali_viewport mvp;
 184
 185         panfrost_mali_viewport_init(batch->ctx,  &mvp);
 186
 187         /* Update the job, unless we're doing wallpapering (whose lack of
 188          * scissor we can ignore, since if we "miss" a tile of wallpaper, it'll
 189          * just... be faster :) */
 190
 191         if (!ctx->wallpaper_batch)
 192                 panfrost_batch_union_scissor(batch, mvp.viewport0[0],
 193                                              mvp.viewport0[1],
 194                                              mvp.viewport1[0] + 1,
 195                                              mvp.viewport1[1] + 1);
 196
 197         tp->postfix.viewport = panfrost_upload_transient(batch, &mvp,
 198                                                          sizeof(mvp));
 199 }
 200
 201 static mali_ptr
 202 panfrost_map_constant_buffer_gpu(struct panfrost_batch *batch,
 203                                  enum pipe_shader_type st,
 204                                  struct panfrost_constant_buffer *buf,
 205                                  unsigned index)
 206 {
 207         struct pipe_constant_buffer *cb = &buf->cb[index];
 208         struct panfrost_resource *rsrc = pan_resource(cb->buffer);
 209
 210         if (rsrc) {
 211                 panfrost_batch_add_bo(batch, rsrc->bo,
 212                                       PAN_BO_ACCESS_SHARED |
 213                                       PAN_BO_ACCESS_READ |
 214                                       panfrost_bo_access_for_stage(st));
 215
 216                 /* Alignment gauranteed by
 217                  * PIPE_CAP_CONSTANT_BUFFER_OFFSET_ALIGNMENT */
 218                 return rsrc->bo->gpu + cb->buffer_offset;
 219         } else if (cb->user_buffer) {
 220                 return panfrost_upload_transient(batch,
 221                                                  cb->user_buffer +
 222                                                  cb->buffer_offset,
 223                                                  cb->buffer_size);
 224         } else {
 225                 unreachable("No constant buffer");
 226         }
 227 }
 228
 229 struct sysval_uniform {
 230         union {
 231                 float f[4];
 232                 int32_t i[4];
 233                 uint32_t u[4];
 234                 uint64_t du[2];
 235         };
 236 };
 237
 238 static void
 239 panfrost_upload_viewport_scale_sysval(struct panfrost_batch *batch,
 240                                       struct sysval_uniform *uniform)
 241 {
 242         struct panfrost_context *ctx = batch->ctx;
 243         const struct pipe_viewport_state *vp = &ctx->pipe_viewport;
 244
 245         uniform->f[0] = vp->scale[0];
 246         uniform->f[1] = vp->scale[1];
 247         uniform->f[2] = vp->scale[2];
 248 }
 249
 250 static void
 251 panfrost_upload_viewport_offset_sysval(struct panfrost_batch *batch,
 252                                        struct sysval_uniform *uniform)
 253 {
 254         struct panfrost_context *ctx = batch->ctx;
 255         const struct pipe_viewport_state *vp = &ctx->pipe_viewport;
 256
 257         uniform->f[0] = vp->translate[0];
 258         uniform->f[1] = vp->translate[1];
 259         uniform->f[2] = vp->translate[2];
 260 }
 261
 262 static void panfrost_upload_txs_sysval(struct panfrost_batch *batch,
 263                                        enum pipe_shader_type st,
 264                                        unsigned int sysvalid,
 265                                        struct sysval_uniform *uniform)
 266 {
 267         struct panfrost_context *ctx = batch->ctx;
 268         unsigned texidx = PAN_SYSVAL_ID_TO_TXS_TEX_IDX(sysvalid);
 269         unsigned dim = PAN_SYSVAL_ID_TO_TXS_DIM(sysvalid);
 270         bool is_array = PAN_SYSVAL_ID_TO_TXS_IS_ARRAY(sysvalid);
 271         struct pipe_sampler_view *tex = &ctx->sampler_views[st][texidx]->base;
 272
 273         assert(dim);
 274         uniform->i[0] = u_minify(tex->texture->width0, tex->u.tex.first_level);
 275
 276         if (dim > 1)
 277                 uniform->i[1] = u_minify(tex->texture->height0,
 278                                          tex->u.tex.first_level);
 279
 280         if (dim > 2)
 281                 uniform->i[2] = u_minify(tex->texture->depth0,
 282                                          tex->u.tex.first_level);
 283
 284         if (is_array)
 285                 uniform->i[dim] = tex->texture->array_size;
 286 }
 287
 288 static void
 289 panfrost_upload_ssbo_sysval(struct panfrost_batch *batch,
 290                             enum pipe_shader_type st,
 291                             unsigned ssbo_id,
 292                             struct sysval_uniform *uniform)
 293 {
 294         struct panfrost_context *ctx = batch->ctx;
 295
 296         assert(ctx->ssbo_mask[st] & (1 << ssbo_id));
 297         struct pipe_shader_buffer sb = ctx->ssbo[st][ssbo_id];
 298
 299         /* Compute address */
 300         struct panfrost_bo *bo = pan_resource(sb.buffer)->bo;
 301
 302         panfrost_batch_add_bo(batch, bo,
 303                               PAN_BO_ACCESS_SHARED | PAN_BO_ACCESS_RW |
 304                               panfrost_bo_access_for_stage(st));
 305
 306         /* Upload address and size as sysval */
 307         uniform->du[0] = bo->gpu + sb.buffer_offset;
 308         uniform->u[2] = sb.buffer_size;
 309 }
 310
 311 static void
 312 panfrost_upload_sampler_sysval(struct panfrost_batch *batch,
 313                                enum pipe_shader_type st,
 314                                unsigned samp_idx,
 315                                struct sysval_uniform *uniform)
 316 {
 317         struct panfrost_context *ctx = batch->ctx;
 318         struct pipe_sampler_state *sampl = &ctx->samplers[st][samp_idx]->base;
 319
 320         uniform->f[0] = sampl->min_lod;
 321         uniform->f[1] = sampl->max_lod;
 322         uniform->f[2] = sampl->lod_bias;
 323
 324         /* Even without any errata, Midgard represents "no mipmapping" as
 325          * fixing the LOD with the clamps; keep behaviour consistent. c.f.
 326          * panfrost_create_sampler_state which also explains our choice of
 327          * epsilon value (again to keep behaviour consistent) */
 328
 329         if (sampl->min_mip_filter == PIPE_TEX_MIPFILTER_NONE)
 330                 uniform->f[1] = uniform->f[0] + (1.0/256.0);
 331 }
 332
 333 static void
 334 panfrost_upload_num_work_groups_sysval(struct panfrost_batch *batch,
 335                                        struct sysval_uniform *uniform)
 336 {
 337         struct panfrost_context *ctx = batch->ctx;
 338
 339         uniform->u[0] = ctx->compute_grid->grid[0];
 340         uniform->u[1] = ctx->compute_grid->grid[1];
 341         uniform->u[2] = ctx->compute_grid->grid[2];
 342 }
 343
 344 static void
 345 panfrost_upload_sysvals(struct panfrost_batch *batch, void *buf,
 346                         struct panfrost_shader_state *ss,
 347                         enum pipe_shader_type st)
 348 {
 349         struct sysval_uniform *uniforms = (void *)buf;
 350
 351         for (unsigned i = 0; i < ss->sysval_count; ++i) {
 352                 int sysval = ss->sysval[i];
 353
 354                 switch (PAN_SYSVAL_TYPE(sysval)) {
 355                 case PAN_SYSVAL_VIEWPORT_SCALE:
 356                         panfrost_upload_viewport_scale_sysval(batch,
 357                                                               &uniforms[i]);
 358                         break;
 359                 case PAN_SYSVAL_VIEWPORT_OFFSET:
 360                         panfrost_upload_viewport_offset_sysval(batch,
 361                                                                &uniforms[i]);
 362                         break;
 363                 case PAN_SYSVAL_TEXTURE_SIZE:
 364                         panfrost_upload_txs_sysval(batch, st,
 365                                                    PAN_SYSVAL_ID(sysval),
 366                                                    &uniforms[i]);
 367                         break;
 368                 case PAN_SYSVAL_SSBO:
 369                         panfrost_upload_ssbo_sysval(batch, st,
 370                                                     PAN_SYSVAL_ID(sysval),
 371                                                     &uniforms[i]);
 372                         break;
 373                 case PAN_SYSVAL_NUM_WORK_GROUPS:
 374                         panfrost_upload_num_work_groups_sysval(batch,
 375                                                                &uniforms[i]);
 376                         break;
 377                 case PAN_SYSVAL_SAMPLER:
 378                         panfrost_upload_sampler_sysval(batch, st,
 379                                                        PAN_SYSVAL_ID(sysval),
 380                                                        &uniforms[i]);
 381                         break;
 382                 default:
 383                         assert(0);
 384                 }
 385         }
 386 }
 387
 388 static const void *
 389 panfrost_map_constant_buffer_cpu(struct panfrost_constant_buffer *buf,
 390                                  unsigned index)
 391 {
 392         struct pipe_constant_buffer *cb = &buf->cb[index];
 393         struct panfrost_resource *rsrc = pan_resource(cb->buffer);
 394
 395         if (rsrc)
 396                 return rsrc->bo->cpu;
 397         else if (cb->user_buffer)
 398                 return cb->user_buffer;
 399         else
 400                 unreachable("No constant buffer");
 401 }
 402
 403 void
 404 panfrost_emit_const_buf(struct panfrost_batch *batch,
 405                         enum pipe_shader_type stage,
 406                         struct midgard_payload_vertex_tiler *vtp)
 407 {
 408         struct panfrost_context *ctx = batch->ctx;
 409         struct panfrost_shader_variants *all = ctx->shader[stage];
 410
 411         if (!all)
 412                 return;
 413
 414         struct panfrost_constant_buffer *buf = &ctx->constant_buffer[stage];
 415
 416         struct panfrost_shader_state *ss = &all->variants[all->active_variant];
 417
 418         /* Uniforms are implicitly UBO #0 */
 419         bool has_uniforms = buf->enabled_mask & (1 << 0);
 420
 421         /* Allocate room for the sysval and the uniforms */
 422         size_t sys_size = sizeof(float) * 4 * ss->sysval_count;
 423         size_t uniform_size = has_uniforms ? (buf->cb[0].buffer_size) : 0;
 424         size_t size = sys_size + uniform_size;
 425         struct panfrost_transfer transfer = panfrost_allocate_transient(batch,
 426                                                                         size);
 427
 428         /* Upload sysvals requested by the shader */
 429         panfrost_upload_sysvals(batch, transfer.cpu, ss, stage);
 430
 431         /* Upload uniforms */
 432         if (has_uniforms && uniform_size) {
 433                 const void *cpu = panfrost_map_constant_buffer_cpu(buf, 0);
 434                 memcpy(transfer.cpu + sys_size, cpu, uniform_size);
 435         }
 436
 437         struct mali_vertex_tiler_postfix *postfix = &vtp->postfix;
 438
 439         /* Next up, attach UBOs. UBO #0 is the uniforms we just
 440          * uploaded */
 441
 442         unsigned ubo_count = panfrost_ubo_count(ctx, stage);
 443         assert(ubo_count >= 1);
 444
 445         size_t sz = sizeof(uint64_t) * ubo_count;
 446         uint64_t ubos[PAN_MAX_CONST_BUFFERS];
 447         int uniform_count = ss->uniform_count;
 448
 449         /* Upload uniforms as a UBO */
 450         ubos[0] = MALI_MAKE_UBO(2 + uniform_count, transfer.gpu);
 451
 452         /* The rest are honest-to-goodness UBOs */
 453
 454         for (unsigned ubo = 1; ubo < ubo_count; ++ubo) {
 455                 size_t usz = buf->cb[ubo].buffer_size;
 456                 bool enabled = buf->enabled_mask & (1 << ubo);
 457                 bool empty = usz == 0;
 458
 459                 if (!enabled || empty) {
 460                         /* Stub out disabled UBOs to catch accesses */
 461                         ubos[ubo] = MALI_MAKE_UBO(0, 0xDEAD0000);
 462                         continue;
 463                 }
 464
 465                 mali_ptr gpu = panfrost_map_constant_buffer_gpu(batch, stage,
 466                                                                 buf, ubo);
 467
 468                 unsigned bytes_per_field = 16;
 469                 unsigned aligned = ALIGN_POT(usz, bytes_per_field);
 470                 ubos[ubo] = MALI_MAKE_UBO(aligned / bytes_per_field, gpu);
 471         }
 472
 473         mali_ptr ubufs = panfrost_upload_transient(batch, ubos, sz);
 474         postfix->uniforms = transfer.gpu;
 475         postfix->uniform_buffers = ubufs;
 476
 477         buf->dirty_mask = 0;
 478 }
 479
 480 void
 481 panfrost_emit_shared_memory(struct panfrost_batch *batch,
 482                             const struct pipe_grid_info *info,
 483                             struct midgard_payload_vertex_tiler *vtp)
 484 {
 485         struct panfrost_context *ctx = batch->ctx;
 486         struct panfrost_shader_variants *all = ctx->shader[PIPE_SHADER_COMPUTE];
 487         struct panfrost_shader_state *ss = &all->variants[all->active_variant];
 488         unsigned single_size = util_next_power_of_two(MAX2(ss->shared_size,
 489                                                            128));
 490         unsigned shared_size = single_size * info->grid[0] * info->grid[1] *
 491                                info->grid[2] * 4;
 492         struct panfrost_bo *bo = panfrost_batch_get_shared_memory(batch,
 493                                                                   shared_size,
 494                                                                   1);
 495
 496         struct mali_shared_memory shared = {
 497                 .shared_memory = bo->gpu,
 498                 .shared_workgroup_count =
 499                         util_logbase2_ceil(info->grid[0]) +
 500                         util_logbase2_ceil(info->grid[1]) +
 501                         util_logbase2_ceil(info->grid[2]),
 502                 .shared_unk1 = 0x2,
 503                 .shared_shift = util_logbase2(single_size) - 1
 504         };
 505
 506         vtp->postfix.shared_memory = panfrost_upload_transient(batch, &shared,
 507                                                                sizeof(shared));
 508 }