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 "pan_allocate.h"
  26 #include "pan_bo.h"
  27 #include "pan_cmdstream.h"
  28 #include "pan_context.h"
  29 #include "pan_job.h"
  30
  31 static void
  32 panfrost_mali_viewport_init(struct panfrost_context *ctx,
  33                             struct mali_viewport *mvp)
  34 {
  35         const struct pipe_viewport_state *vp = &ctx->pipe_viewport;
  36
  37         /* Clip bounds are encoded as floats. The viewport itself is encoded as
  38          * (somewhat) asymmetric ints. */
  39
  40         const struct pipe_scissor_state *ss = &ctx->scissor;
  41
  42         memset(mvp, 0, sizeof(*mvp));
  43
  44         /* By default, do no viewport clipping, i.e. clip to (-inf, inf) in
  45          * each direction. Clipping to the viewport in theory should work, but
  46          * in practice causes issues when we're not explicitly trying to
  47          * scissor */
  48
  49         *mvp = (struct mali_viewport) {
  50                 .clip_minx = -INFINITY,
  51                 .clip_miny = -INFINITY,
  52                 .clip_maxx = INFINITY,
  53                 .clip_maxy = INFINITY,
  54         };
  55
  56         /* Always scissor to the viewport by default. */
  57         float vp_minx = (int) (vp->translate[0] - fabsf(vp->scale[0]));
  58         float vp_maxx = (int) (vp->translate[0] + fabsf(vp->scale[0]));
  59
  60         float vp_miny = (int) (vp->translate[1] - fabsf(vp->scale[1]));
  61         float vp_maxy = (int) (vp->translate[1] + fabsf(vp->scale[1]));
  62
  63         float minz = (vp->translate[2] - fabsf(vp->scale[2]));
  64         float maxz = (vp->translate[2] + fabsf(vp->scale[2]));
  65
  66         /* Apply the scissor test */
  67
  68         unsigned minx, miny, maxx, maxy;
  69
  70         if (ss && ctx->rasterizer && ctx->rasterizer->base.scissor) {
  71                 minx = MAX2(ss->minx, vp_minx);
  72                 miny = MAX2(ss->miny, vp_miny);
  73                 maxx = MIN2(ss->maxx, vp_maxx);
  74                 maxy = MIN2(ss->maxy, vp_maxy);
  75         } else {
  76                 minx = vp_minx;
  77                 miny = vp_miny;
  78                 maxx = vp_maxx;
  79                 maxy = vp_maxy;
  80         }
  81
  82         /* Hardware needs the min/max to be strictly ordered, so flip if we
  83          * need to. The viewport transformation in the vertex shader will
  84          * handle the negatives if we don't */
  85
  86         if (miny > maxy) {
  87                 unsigned temp = miny;
  88                 miny = maxy;
  89                 maxy = temp;
  90         }
  91
  92         if (minx > maxx) {
  93                 unsigned temp = minx;
  94                 minx = maxx;
  95                 maxx = temp;
  96         }
  97
  98         if (minz > maxz) {
  99                 float temp = minz;
 100                 minz = maxz;
 101                 maxz = temp;
 102         }
 103
 104         /* Clamp to the framebuffer size as a last check */
 105
 106         minx = MIN2(ctx->pipe_framebuffer.width, minx);
 107         maxx = MIN2(ctx->pipe_framebuffer.width, maxx);
 108
 109         miny = MIN2(ctx->pipe_framebuffer.height, miny);
 110         maxy = MIN2(ctx->pipe_framebuffer.height, maxy);
 111
 112         /* Upload */
 113
 114         mvp->viewport0[0] = minx;
 115         mvp->viewport1[0] = MALI_POSITIVE(maxx);
 116
 117         mvp->viewport0[1] = miny;
 118         mvp->viewport1[1] = MALI_POSITIVE(maxy);
 119
 120         mvp->clip_minz = minz;
 121         mvp->clip_maxz = maxz;
 122 }
 123
 124 void
 125 panfrost_emit_viewport(struct panfrost_batch *batch,
 126                        struct midgard_payload_vertex_tiler *tp)
 127 {
 128         struct panfrost_context *ctx = batch->ctx;
 129         struct mali_viewport mvp;
 130
 131         panfrost_mali_viewport_init(batch->ctx,  &mvp);
 132
 133         /* Update the job, unless we're doing wallpapering (whose lack of
 134          * scissor we can ignore, since if we "miss" a tile of wallpaper, it'll
 135          * just... be faster :) */
 136
 137         if (!ctx->wallpaper_batch)
 138                 panfrost_batch_union_scissor(batch, mvp.viewport0[0],
 139                                              mvp.viewport0[1],
 140                                              mvp.viewport1[0] + 1,
 141                                              mvp.viewport1[1] + 1);
 142
 143         tp->postfix.viewport = panfrost_upload_transient(batch, &mvp,
 144                                                          sizeof(mvp));
 145 }
 146
 147 static mali_ptr
 148 panfrost_map_constant_buffer_gpu(struct panfrost_batch *batch,
 149                                  enum pipe_shader_type st,
 150                                  struct panfrost_constant_buffer *buf,
 151                                  unsigned index)
 152 {
 153         struct pipe_constant_buffer *cb = &buf->cb[index];
 154         struct panfrost_resource *rsrc = pan_resource(cb->buffer);
 155
 156         if (rsrc) {
 157                 panfrost_batch_add_bo(batch, rsrc->bo,
 158                                       PAN_BO_ACCESS_SHARED |
 159                                       PAN_BO_ACCESS_READ |
 160                                       panfrost_bo_access_for_stage(st));
 161
 162                 /* Alignment gauranteed by
 163                  * PIPE_CAP_CONSTANT_BUFFER_OFFSET_ALIGNMENT */
 164                 return rsrc->bo->gpu + cb->buffer_offset;
 165         } else if (cb->user_buffer) {
 166                 return panfrost_upload_transient(batch,
 167                                                  cb->user_buffer +
 168                                                  cb->buffer_offset,
 169                                                  cb->buffer_size);
 170         } else {
 171                 unreachable("No constant buffer");
 172         }
 173 }
 174
 175 struct sysval_uniform {
 176         union {
 177                 float f[4];
 178                 int32_t i[4];
 179                 uint32_t u[4];
 180                 uint64_t du[2];
 181         };
 182 };
 183
 184 static void
 185 panfrost_upload_viewport_scale_sysval(struct panfrost_batch *batch,
 186                                       struct sysval_uniform *uniform)
 187 {
 188         struct panfrost_context *ctx = batch->ctx;
 189         const struct pipe_viewport_state *vp = &ctx->pipe_viewport;
 190
 191         uniform->f[0] = vp->scale[0];
 192         uniform->f[1] = vp->scale[1];
 193         uniform->f[2] = vp->scale[2];
 194 }
 195
 196 static void
 197 panfrost_upload_viewport_offset_sysval(struct panfrost_batch *batch,
 198                                        struct sysval_uniform *uniform)
 199 {
 200         struct panfrost_context *ctx = batch->ctx;
 201         const struct pipe_viewport_state *vp = &ctx->pipe_viewport;
 202
 203         uniform->f[0] = vp->translate[0];
 204         uniform->f[1] = vp->translate[1];
 205         uniform->f[2] = vp->translate[2];
 206 }
 207
 208 static void panfrost_upload_txs_sysval(struct panfrost_batch *batch,
 209                                        enum pipe_shader_type st,
 210                                        unsigned int sysvalid,
 211                                        struct sysval_uniform *uniform)
 212 {
 213         struct panfrost_context *ctx = batch->ctx;
 214         unsigned texidx = PAN_SYSVAL_ID_TO_TXS_TEX_IDX(sysvalid);
 215         unsigned dim = PAN_SYSVAL_ID_TO_TXS_DIM(sysvalid);
 216         bool is_array = PAN_SYSVAL_ID_TO_TXS_IS_ARRAY(sysvalid);
 217         struct pipe_sampler_view *tex = &ctx->sampler_views[st][texidx]->base;
 218
 219         assert(dim);
 220         uniform->i[0] = u_minify(tex->texture->width0, tex->u.tex.first_level);
 221
 222         if (dim > 1)
 223                 uniform->i[1] = u_minify(tex->texture->height0,
 224                                          tex->u.tex.first_level);
 225
 226         if (dim > 2)
 227                 uniform->i[2] = u_minify(tex->texture->depth0,
 228                                          tex->u.tex.first_level);
 229
 230         if (is_array)
 231                 uniform->i[dim] = tex->texture->array_size;
 232 }
 233
 234 static void
 235 panfrost_upload_ssbo_sysval(struct panfrost_batch *batch,
 236                             enum pipe_shader_type st,
 237                             unsigned ssbo_id,
 238                             struct sysval_uniform *uniform)
 239 {
 240         struct panfrost_context *ctx = batch->ctx;
 241
 242         assert(ctx->ssbo_mask[st] & (1 << ssbo_id));
 243         struct pipe_shader_buffer sb = ctx->ssbo[st][ssbo_id];
 244
 245         /* Compute address */
 246         struct panfrost_bo *bo = pan_resource(sb.buffer)->bo;
 247
 248         panfrost_batch_add_bo(batch, bo,
 249                               PAN_BO_ACCESS_SHARED | PAN_BO_ACCESS_RW |
 250                               panfrost_bo_access_for_stage(st));
 251
 252         /* Upload address and size as sysval */
 253         uniform->du[0] = bo->gpu + sb.buffer_offset;
 254         uniform->u[2] = sb.buffer_size;
 255 }
 256
 257 static void
 258 panfrost_upload_sampler_sysval(struct panfrost_batch *batch,
 259                                enum pipe_shader_type st,
 260                                unsigned samp_idx,
 261                                struct sysval_uniform *uniform)
 262 {
 263         struct panfrost_context *ctx = batch->ctx;
 264         struct pipe_sampler_state *sampl = &ctx->samplers[st][samp_idx]->base;
 265
 266         uniform->f[0] = sampl->min_lod;
 267         uniform->f[1] = sampl->max_lod;
 268         uniform->f[2] = sampl->lod_bias;
 269
 270         /* Even without any errata, Midgard represents "no mipmapping" as
 271          * fixing the LOD with the clamps; keep behaviour consistent. c.f.
 272          * panfrost_create_sampler_state which also explains our choice of
 273          * epsilon value (again to keep behaviour consistent) */
 274
 275         if (sampl->min_mip_filter == PIPE_TEX_MIPFILTER_NONE)
 276                 uniform->f[1] = uniform->f[0] + (1.0/256.0);
 277 }
 278
 279 static void
 280 panfrost_upload_num_work_groups_sysval(struct panfrost_batch *batch,
 281                                        struct sysval_uniform *uniform)
 282 {
 283         struct panfrost_context *ctx = batch->ctx;
 284
 285         uniform->u[0] = ctx->compute_grid->grid[0];
 286         uniform->u[1] = ctx->compute_grid->grid[1];
 287         uniform->u[2] = ctx->compute_grid->grid[2];
 288 }
 289
 290 static void
 291 panfrost_upload_sysvals(struct panfrost_batch *batch, void *buf,
 292                         struct panfrost_shader_state *ss,
 293                         enum pipe_shader_type st)
 294 {
 295         struct sysval_uniform *uniforms = (void *)buf;
 296
 297         for (unsigned i = 0; i < ss->sysval_count; ++i) {
 298                 int sysval = ss->sysval[i];
 299
 300                 switch (PAN_SYSVAL_TYPE(sysval)) {
 301                 case PAN_SYSVAL_VIEWPORT_SCALE:
 302                         panfrost_upload_viewport_scale_sysval(batch,
 303                                                               &uniforms[i]);
 304                         break;
 305                 case PAN_SYSVAL_VIEWPORT_OFFSET:
 306                         panfrost_upload_viewport_offset_sysval(batch,
 307                                                                &uniforms[i]);
 308                         break;
 309                 case PAN_SYSVAL_TEXTURE_SIZE:
 310                         panfrost_upload_txs_sysval(batch, st,
 311                                                    PAN_SYSVAL_ID(sysval),
 312                                                    &uniforms[i]);
 313                         break;
 314                 case PAN_SYSVAL_SSBO:
 315                         panfrost_upload_ssbo_sysval(batch, st,
 316                                                     PAN_SYSVAL_ID(sysval),
 317                                                     &uniforms[i]);
 318                         break;
 319                 case PAN_SYSVAL_NUM_WORK_GROUPS:
 320                         panfrost_upload_num_work_groups_sysval(batch,
 321                                                                &uniforms[i]);
 322                         break;
 323                 case PAN_SYSVAL_SAMPLER:
 324                         panfrost_upload_sampler_sysval(batch, st,
 325                                                        PAN_SYSVAL_ID(sysval),
 326                                                        &uniforms[i]);
 327                         break;
 328                 default:
 329                         assert(0);
 330                 }
 331         }
 332 }
 333
 334 static const void *
 335 panfrost_map_constant_buffer_cpu(struct panfrost_constant_buffer *buf,
 336                                  unsigned index)
 337 {
 338         struct pipe_constant_buffer *cb = &buf->cb[index];
 339         struct panfrost_resource *rsrc = pan_resource(cb->buffer);
 340
 341         if (rsrc)
 342                 return rsrc->bo->cpu;
 343         else if (cb->user_buffer)
 344                 return cb->user_buffer;
 345         else
 346                 unreachable("No constant buffer");
 347 }
 348
 349 void
 350 panfrost_emit_const_buf(struct panfrost_batch *batch,
 351                         enum pipe_shader_type stage,
 352                         struct midgard_payload_vertex_tiler *vtp)
 353 {
 354         struct panfrost_context *ctx = batch->ctx;
 355         struct panfrost_shader_variants *all = ctx->shader[stage];
 356
 357         if (!all)
 358                 return;
 359
 360         struct panfrost_constant_buffer *buf = &ctx->constant_buffer[stage];
 361
 362         struct panfrost_shader_state *ss = &all->variants[all->active_variant];
 363
 364         /* Uniforms are implicitly UBO #0 */
 365         bool has_uniforms = buf->enabled_mask & (1 << 0);
 366
 367         /* Allocate room for the sysval and the uniforms */
 368         size_t sys_size = sizeof(float) * 4 * ss->sysval_count;
 369         size_t uniform_size = has_uniforms ? (buf->cb[0].buffer_size) : 0;
 370         size_t size = sys_size + uniform_size;
 371         struct panfrost_transfer transfer = panfrost_allocate_transient(batch,
 372                                                                         size);
 373
 374         /* Upload sysvals requested by the shader */
 375         panfrost_upload_sysvals(batch, transfer.cpu, ss, stage);
 376
 377         /* Upload uniforms */
 378         if (has_uniforms && uniform_size) {
 379                 const void *cpu = panfrost_map_constant_buffer_cpu(buf, 0);
 380                 memcpy(transfer.cpu + sys_size, cpu, uniform_size);
 381         }
 382
 383         struct mali_vertex_tiler_postfix *postfix = &vtp->postfix;
 384
 385         /* Next up, attach UBOs. UBO #0 is the uniforms we just
 386          * uploaded */
 387
 388         unsigned ubo_count = panfrost_ubo_count(ctx, stage);
 389         assert(ubo_count >= 1);
 390
 391         size_t sz = sizeof(uint64_t) * ubo_count;
 392         uint64_t ubos[PAN_MAX_CONST_BUFFERS];
 393         int uniform_count = ss->uniform_count;
 394
 395         /* Upload uniforms as a UBO */
 396         ubos[0] = MALI_MAKE_UBO(2 + uniform_count, transfer.gpu);
 397
 398         /* The rest are honest-to-goodness UBOs */
 399
 400         for (unsigned ubo = 1; ubo < ubo_count; ++ubo) {
 401                 size_t usz = buf->cb[ubo].buffer_size;
 402                 bool enabled = buf->enabled_mask & (1 << ubo);
 403                 bool empty = usz == 0;
 404
 405                 if (!enabled || empty) {
 406                         /* Stub out disabled UBOs to catch accesses */
 407                         ubos[ubo] = MALI_MAKE_UBO(0, 0xDEAD0000);
 408                         continue;
 409                 }
 410
 411                 mali_ptr gpu = panfrost_map_constant_buffer_gpu(batch, stage,
 412                                                                 buf, ubo);
 413
 414                 unsigned bytes_per_field = 16;
 415                 unsigned aligned = ALIGN_POT(usz, bytes_per_field);
 416                 ubos[ubo] = MALI_MAKE_UBO(aligned / bytes_per_field, gpu);
 417         }
 418
 419         mali_ptr ubufs = panfrost_upload_transient(batch, ubos, sz);
 420         postfix->uniforms = transfer.gpu;
 421         postfix->uniform_buffers = ubufs;
 422
 423         buf->dirty_mask = 0;
 424 }
 425
 426 void
 427 panfrost_emit_shared_memory(struct panfrost_batch *batch,
 428                             const struct pipe_grid_info *info,
 429                             struct midgard_payload_vertex_tiler *vtp)
 430 {
 431         struct panfrost_context *ctx = batch->ctx;
 432         struct panfrost_shader_variants *all = ctx->shader[PIPE_SHADER_COMPUTE];
 433         struct panfrost_shader_state *ss = &all->variants[all->active_variant];
 434         unsigned single_size = util_next_power_of_two(MAX2(ss->shared_size,
 435                                                            128));
 436         unsigned shared_size = single_size * info->grid[0] * info->grid[1] *
 437                                info->grid[2] * 4;
 438         struct panfrost_bo *bo = panfrost_batch_get_shared_memory(batch,
 439                                                                   shared_size,
 440                                                                   1);
 441
 442         struct mali_shared_memory shared = {
 443                 .shared_memory = bo->gpu,
 444                 .shared_workgroup_count =
 445                         util_logbase2_ceil(info->grid[0]) +
 446                         util_logbase2_ceil(info->grid[1]) +
 447                         util_logbase2_ceil(info->grid[2]),
 448                 .shared_unk1 = 0x2,
 449                 .shared_shift = util_logbase2(single_size) - 1
 450         };
 451
 452         vtp->postfix.shared_memory = panfrost_upload_transient(batch, &shared,
 453                                                                sizeof(shared));
 454 }