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_vt_update_occlusion_query(struct panfrost_context *ctx,
  64                                    struct midgard_payload_vertex_tiler *tp)
  65 {
  66         SET_BIT(tp->gl_enables, MALI_OCCLUSION_QUERY, ctx->occlusion_query);
  67         if (ctx->occlusion_query)
  68                 tp->postfix.occlusion_counter = ctx->occlusion_query->bo->gpu;
  69         else
  70                 tp->postfix.occlusion_counter = 0;
  71 }
  72
  73 void
  74 panfrost_emit_shader_meta(struct panfrost_batch *batch,
  75                           enum pipe_shader_type st,
  76                           struct midgard_payload_vertex_tiler *vtp)
  77 {
  78         struct panfrost_context *ctx = batch->ctx;
  79         struct panfrost_shader_state *ss = panfrost_get_shader_state(ctx, st);
  80
  81         if (!ss) {
  82                 vtp->postfix.shader = 0;
  83                 return;
  84         }
  85
  86         /* Add the shader BO to the batch. */
  87         panfrost_batch_add_bo(batch, ss->bo,
  88                               PAN_BO_ACCESS_PRIVATE |
  89                               PAN_BO_ACCESS_READ |
  90                               panfrost_bo_access_for_stage(st));
  91
  92         vtp->postfix.shader = panfrost_upload_transient(batch, ss->tripipe,
  93                                                         sizeof(*ss->tripipe));
  94 }
  95
  96 static void
  97 panfrost_mali_viewport_init(struct panfrost_context *ctx,
  98                             struct mali_viewport *mvp)
  99 {
 100         const struct pipe_viewport_state *vp = &ctx->pipe_viewport;
 101
 102         /* Clip bounds are encoded as floats. The viewport itself is encoded as
 103          * (somewhat) asymmetric ints. */
 104
 105         const struct pipe_scissor_state *ss = &ctx->scissor;
 106
 107         memset(mvp, 0, sizeof(*mvp));
 108
 109         /* By default, do no viewport clipping, i.e. clip to (-inf, inf) in
 110          * each direction. Clipping to the viewport in theory should work, but
 111          * in practice causes issues when we're not explicitly trying to
 112          * scissor */
 113
 114         *mvp = (struct mali_viewport) {
 115                 .clip_minx = -INFINITY,
 116                 .clip_miny = -INFINITY,
 117                 .clip_maxx = INFINITY,
 118                 .clip_maxy = INFINITY,
 119         };
 120
 121         /* Always scissor to the viewport by default. */
 122         float vp_minx = (int) (vp->translate[0] - fabsf(vp->scale[0]));
 123         float vp_maxx = (int) (vp->translate[0] + fabsf(vp->scale[0]));
 124
 125         float vp_miny = (int) (vp->translate[1] - fabsf(vp->scale[1]));
 126         float vp_maxy = (int) (vp->translate[1] + fabsf(vp->scale[1]));
 127
 128         float minz = (vp->translate[2] - fabsf(vp->scale[2]));
 129         float maxz = (vp->translate[2] + fabsf(vp->scale[2]));
 130
 131         /* Apply the scissor test */
 132
 133         unsigned minx, miny, maxx, maxy;
 134
 135         if (ss && ctx->rasterizer && ctx->rasterizer->base.scissor) {
 136                 minx = MAX2(ss->minx, vp_minx);
 137                 miny = MAX2(ss->miny, vp_miny);
 138                 maxx = MIN2(ss->maxx, vp_maxx);
 139                 maxy = MIN2(ss->maxy, vp_maxy);
 140         } else {
 141                 minx = vp_minx;
 142                 miny = vp_miny;
 143                 maxx = vp_maxx;
 144                 maxy = vp_maxy;
 145         }
 146
 147         /* Hardware needs the min/max to be strictly ordered, so flip if we
 148          * need to. The viewport transformation in the vertex shader will
 149          * handle the negatives if we don't */
 150
 151         if (miny > maxy) {
 152                 unsigned temp = miny;
 153                 miny = maxy;
 154                 maxy = temp;
 155         }
 156
 157         if (minx > maxx) {
 158                 unsigned temp = minx;
 159                 minx = maxx;
 160                 maxx = temp;
 161         }
 162
 163         if (minz > maxz) {
 164                 float temp = minz;
 165                 minz = maxz;
 166                 maxz = temp;
 167         }
 168
 169         /* Clamp to the framebuffer size as a last check */
 170
 171         minx = MIN2(ctx->pipe_framebuffer.width, minx);
 172         maxx = MIN2(ctx->pipe_framebuffer.width, maxx);
 173
 174         miny = MIN2(ctx->pipe_framebuffer.height, miny);
 175         maxy = MIN2(ctx->pipe_framebuffer.height, maxy);
 176
 177         /* Upload */
 178
 179         mvp->viewport0[0] = minx;
 180         mvp->viewport1[0] = MALI_POSITIVE(maxx);
 181
 182         mvp->viewport0[1] = miny;
 183         mvp->viewport1[1] = MALI_POSITIVE(maxy);
 184
 185         mvp->clip_minz = minz;
 186         mvp->clip_maxz = maxz;
 187 }
 188
 189 void
 190 panfrost_emit_viewport(struct panfrost_batch *batch,
 191                        struct midgard_payload_vertex_tiler *tp)
 192 {
 193         struct panfrost_context *ctx = batch->ctx;
 194         struct mali_viewport mvp;
 195
 196         panfrost_mali_viewport_init(batch->ctx,  &mvp);
 197
 198         /* Update the job, unless we're doing wallpapering (whose lack of
 199          * scissor we can ignore, since if we "miss" a tile of wallpaper, it'll
 200          * just... be faster :) */
 201
 202         if (!ctx->wallpaper_batch)
 203                 panfrost_batch_union_scissor(batch, mvp.viewport0[0],
 204                                              mvp.viewport0[1],
 205                                              mvp.viewport1[0] + 1,
 206                                              mvp.viewport1[1] + 1);
 207
 208         tp->postfix.viewport = panfrost_upload_transient(batch, &mvp,
 209                                                          sizeof(mvp));
 210 }
 211
 212 static mali_ptr
 213 panfrost_map_constant_buffer_gpu(struct panfrost_batch *batch,
 214                                  enum pipe_shader_type st,
 215                                  struct panfrost_constant_buffer *buf,
 216                                  unsigned index)
 217 {
 218         struct pipe_constant_buffer *cb = &buf->cb[index];
 219         struct panfrost_resource *rsrc = pan_resource(cb->buffer);
 220
 221         if (rsrc) {
 222                 panfrost_batch_add_bo(batch, rsrc->bo,
 223                                       PAN_BO_ACCESS_SHARED |
 224                                       PAN_BO_ACCESS_READ |
 225                                       panfrost_bo_access_for_stage(st));
 226
 227                 /* Alignment gauranteed by
 228                  * PIPE_CAP_CONSTANT_BUFFER_OFFSET_ALIGNMENT */
 229                 return rsrc->bo->gpu + cb->buffer_offset;
 230         } else if (cb->user_buffer) {
 231                 return panfrost_upload_transient(batch,
 232                                                  cb->user_buffer +
 233                                                  cb->buffer_offset,
 234                                                  cb->buffer_size);
 235         } else {
 236                 unreachable("No constant buffer");
 237         }
 238 }
 239
 240 struct sysval_uniform {
 241         union {
 242                 float f[4];
 243                 int32_t i[4];
 244                 uint32_t u[4];
 245                 uint64_t du[2];
 246         };
 247 };
 248
 249 static void
 250 panfrost_upload_viewport_scale_sysval(struct panfrost_batch *batch,
 251                                       struct sysval_uniform *uniform)
 252 {
 253         struct panfrost_context *ctx = batch->ctx;
 254         const struct pipe_viewport_state *vp = &ctx->pipe_viewport;
 255
 256         uniform->f[0] = vp->scale[0];
 257         uniform->f[1] = vp->scale[1];
 258         uniform->f[2] = vp->scale[2];
 259 }
 260
 261 static void
 262 panfrost_upload_viewport_offset_sysval(struct panfrost_batch *batch,
 263                                        struct sysval_uniform *uniform)
 264 {
 265         struct panfrost_context *ctx = batch->ctx;
 266         const struct pipe_viewport_state *vp = &ctx->pipe_viewport;
 267
 268         uniform->f[0] = vp->translate[0];
 269         uniform->f[1] = vp->translate[1];
 270         uniform->f[2] = vp->translate[2];
 271 }
 272
 273 static void panfrost_upload_txs_sysval(struct panfrost_batch *batch,
 274                                        enum pipe_shader_type st,
 275                                        unsigned int sysvalid,
 276                                        struct sysval_uniform *uniform)
 277 {
 278         struct panfrost_context *ctx = batch->ctx;
 279         unsigned texidx = PAN_SYSVAL_ID_TO_TXS_TEX_IDX(sysvalid);
 280         unsigned dim = PAN_SYSVAL_ID_TO_TXS_DIM(sysvalid);
 281         bool is_array = PAN_SYSVAL_ID_TO_TXS_IS_ARRAY(sysvalid);
 282         struct pipe_sampler_view *tex = &ctx->sampler_views[st][texidx]->base;
 283
 284         assert(dim);
 285         uniform->i[0] = u_minify(tex->texture->width0, tex->u.tex.first_level);
 286
 287         if (dim > 1)
 288                 uniform->i[1] = u_minify(tex->texture->height0,
 289                                          tex->u.tex.first_level);
 290
 291         if (dim > 2)
 292                 uniform->i[2] = u_minify(tex->texture->depth0,
 293                                          tex->u.tex.first_level);
 294
 295         if (is_array)
 296                 uniform->i[dim] = tex->texture->array_size;
 297 }
 298
 299 static void
 300 panfrost_upload_ssbo_sysval(struct panfrost_batch *batch,
 301                             enum pipe_shader_type st,
 302                             unsigned ssbo_id,
 303                             struct sysval_uniform *uniform)
 304 {
 305         struct panfrost_context *ctx = batch->ctx;
 306
 307         assert(ctx->ssbo_mask[st] & (1 << ssbo_id));
 308         struct pipe_shader_buffer sb = ctx->ssbo[st][ssbo_id];
 309
 310         /* Compute address */
 311         struct panfrost_bo *bo = pan_resource(sb.buffer)->bo;
 312
 313         panfrost_batch_add_bo(batch, bo,
 314                               PAN_BO_ACCESS_SHARED | PAN_BO_ACCESS_RW |
 315                               panfrost_bo_access_for_stage(st));
 316
 317         /* Upload address and size as sysval */
 318         uniform->du[0] = bo->gpu + sb.buffer_offset;
 319         uniform->u[2] = sb.buffer_size;
 320 }
 321
 322 static void
 323 panfrost_upload_sampler_sysval(struct panfrost_batch *batch,
 324                                enum pipe_shader_type st,
 325                                unsigned samp_idx,
 326                                struct sysval_uniform *uniform)
 327 {
 328         struct panfrost_context *ctx = batch->ctx;
 329         struct pipe_sampler_state *sampl = &ctx->samplers[st][samp_idx]->base;
 330
 331         uniform->f[0] = sampl->min_lod;
 332         uniform->f[1] = sampl->max_lod;
 333         uniform->f[2] = sampl->lod_bias;
 334
 335         /* Even without any errata, Midgard represents "no mipmapping" as
 336          * fixing the LOD with the clamps; keep behaviour consistent. c.f.
 337          * panfrost_create_sampler_state which also explains our choice of
 338          * epsilon value (again to keep behaviour consistent) */
 339
 340         if (sampl->min_mip_filter == PIPE_TEX_MIPFILTER_NONE)
 341                 uniform->f[1] = uniform->f[0] + (1.0/256.0);
 342 }
 343
 344 static void
 345 panfrost_upload_num_work_groups_sysval(struct panfrost_batch *batch,
 346                                        struct sysval_uniform *uniform)
 347 {
 348         struct panfrost_context *ctx = batch->ctx;
 349
 350         uniform->u[0] = ctx->compute_grid->grid[0];
 351         uniform->u[1] = ctx->compute_grid->grid[1];
 352         uniform->u[2] = ctx->compute_grid->grid[2];
 353 }
 354
 355 static void
 356 panfrost_upload_sysvals(struct panfrost_batch *batch, void *buf,
 357                         struct panfrost_shader_state *ss,
 358                         enum pipe_shader_type st)
 359 {
 360         struct sysval_uniform *uniforms = (void *)buf;
 361
 362         for (unsigned i = 0; i < ss->sysval_count; ++i) {
 363                 int sysval = ss->sysval[i];
 364
 365                 switch (PAN_SYSVAL_TYPE(sysval)) {
 366                 case PAN_SYSVAL_VIEWPORT_SCALE:
 367                         panfrost_upload_viewport_scale_sysval(batch,
 368                                                               &uniforms[i]);
 369                         break;
 370                 case PAN_SYSVAL_VIEWPORT_OFFSET:
 371                         panfrost_upload_viewport_offset_sysval(batch,
 372                                                                &uniforms[i]);
 373                         break;
 374                 case PAN_SYSVAL_TEXTURE_SIZE:
 375                         panfrost_upload_txs_sysval(batch, st,
 376                                                    PAN_SYSVAL_ID(sysval),
 377                                                    &uniforms[i]);
 378                         break;
 379                 case PAN_SYSVAL_SSBO:
 380                         panfrost_upload_ssbo_sysval(batch, st,
 381                                                     PAN_SYSVAL_ID(sysval),
 382                                                     &uniforms[i]);
 383                         break;
 384                 case PAN_SYSVAL_NUM_WORK_GROUPS:
 385                         panfrost_upload_num_work_groups_sysval(batch,
 386                                                                &uniforms[i]);
 387                         break;
 388                 case PAN_SYSVAL_SAMPLER:
 389                         panfrost_upload_sampler_sysval(batch, st,
 390                                                        PAN_SYSVAL_ID(sysval),
 391                                                        &uniforms[i]);
 392                         break;
 393                 default:
 394                         assert(0);
 395                 }
 396         }
 397 }
 398
 399 static const void *
 400 panfrost_map_constant_buffer_cpu(struct panfrost_constant_buffer *buf,
 401                                  unsigned index)
 402 {
 403         struct pipe_constant_buffer *cb = &buf->cb[index];
 404         struct panfrost_resource *rsrc = pan_resource(cb->buffer);
 405
 406         if (rsrc)
 407                 return rsrc->bo->cpu;
 408         else if (cb->user_buffer)
 409                 return cb->user_buffer;
 410         else
 411                 unreachable("No constant buffer");
 412 }
 413
 414 void
 415 panfrost_emit_const_buf(struct panfrost_batch *batch,
 416                         enum pipe_shader_type stage,
 417                         struct midgard_payload_vertex_tiler *vtp)
 418 {
 419         struct panfrost_context *ctx = batch->ctx;
 420         struct panfrost_shader_variants *all = ctx->shader[stage];
 421
 422         if (!all)
 423                 return;
 424
 425         struct panfrost_constant_buffer *buf = &ctx->constant_buffer[stage];
 426
 427         struct panfrost_shader_state *ss = &all->variants[all->active_variant];
 428
 429         /* Uniforms are implicitly UBO #0 */
 430         bool has_uniforms = buf->enabled_mask & (1 << 0);
 431
 432         /* Allocate room for the sysval and the uniforms */
 433         size_t sys_size = sizeof(float) * 4 * ss->sysval_count;
 434         size_t uniform_size = has_uniforms ? (buf->cb[0].buffer_size) : 0;
 435         size_t size = sys_size + uniform_size;
 436         struct panfrost_transfer transfer = panfrost_allocate_transient(batch,
 437                                                                         size);
 438
 439         /* Upload sysvals requested by the shader */
 440         panfrost_upload_sysvals(batch, transfer.cpu, ss, stage);
 441
 442         /* Upload uniforms */
 443         if (has_uniforms && uniform_size) {
 444                 const void *cpu = panfrost_map_constant_buffer_cpu(buf, 0);
 445                 memcpy(transfer.cpu + sys_size, cpu, uniform_size);
 446         }
 447
 448         struct mali_vertex_tiler_postfix *postfix = &vtp->postfix;
 449
 450         /* Next up, attach UBOs. UBO #0 is the uniforms we just
 451          * uploaded */
 452
 453         unsigned ubo_count = panfrost_ubo_count(ctx, stage);
 454         assert(ubo_count >= 1);
 455
 456         size_t sz = sizeof(uint64_t) * ubo_count;
 457         uint64_t ubos[PAN_MAX_CONST_BUFFERS];
 458         int uniform_count = ss->uniform_count;
 459
 460         /* Upload uniforms as a UBO */
 461         ubos[0] = MALI_MAKE_UBO(2 + uniform_count, transfer.gpu);
 462
 463         /* The rest are honest-to-goodness UBOs */
 464
 465         for (unsigned ubo = 1; ubo < ubo_count; ++ubo) {
 466                 size_t usz = buf->cb[ubo].buffer_size;
 467                 bool enabled = buf->enabled_mask & (1 << ubo);
 468                 bool empty = usz == 0;
 469
 470                 if (!enabled || empty) {
 471                         /* Stub out disabled UBOs to catch accesses */
 472                         ubos[ubo] = MALI_MAKE_UBO(0, 0xDEAD0000);
 473                         continue;
 474                 }
 475
 476                 mali_ptr gpu = panfrost_map_constant_buffer_gpu(batch, stage,
 477                                                                 buf, ubo);
 478
 479                 unsigned bytes_per_field = 16;
 480                 unsigned aligned = ALIGN_POT(usz, bytes_per_field);
 481                 ubos[ubo] = MALI_MAKE_UBO(aligned / bytes_per_field, gpu);
 482         }
 483
 484         mali_ptr ubufs = panfrost_upload_transient(batch, ubos, sz);
 485         postfix->uniforms = transfer.gpu;
 486         postfix->uniform_buffers = ubufs;
 487
 488         buf->dirty_mask = 0;
 489 }
 490
 491 void
 492 panfrost_emit_shared_memory(struct panfrost_batch *batch,
 493                             const struct pipe_grid_info *info,
 494                             struct midgard_payload_vertex_tiler *vtp)
 495 {
 496         struct panfrost_context *ctx = batch->ctx;
 497         struct panfrost_shader_variants *all = ctx->shader[PIPE_SHADER_COMPUTE];
 498         struct panfrost_shader_state *ss = &all->variants[all->active_variant];
 499         unsigned single_size = util_next_power_of_two(MAX2(ss->shared_size,
 500                                                            128));
 501         unsigned shared_size = single_size * info->grid[0] * info->grid[1] *
 502                                info->grid[2] * 4;
 503         struct panfrost_bo *bo = panfrost_batch_get_shared_memory(batch,
 504                                                                   shared_size,
 505                                                                   1);
 506
 507         struct mali_shared_memory shared = {
 508                 .shared_memory = bo->gpu,
 509                 .shared_workgroup_count =
 510                         util_logbase2_ceil(info->grid[0]) +
 511                         util_logbase2_ceil(info->grid[1]) +
 512                         util_logbase2_ceil(info->grid[2]),
 513                 .shared_unk1 = 0x2,
 514                 .shared_shift = util_logbase2(single_size) - 1
 515         };
 516
 517         vtp->postfix.shared_memory = panfrost_upload_transient(batch, &shared,
 518                                                                sizeof(shared));
 519 }