src/gallium/drivers/freedreno/freedreno_gmem.c

   1 /* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
   2
   3 /*
   4  * Copyright (C) 2012 Rob Clark <robclark@freedesktop.org>
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the "Software"),
   8  * to deal in the Software without restriction, including without limitation
   9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10  * and/or sell copies of the Software, and to permit persons to whom the
  11  * Software is furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice (including the next
  14  * paragraph) shall be included in all copies or substantial portions of the
  15  * Software.
  16  *
  17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  18  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  20  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  21  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  22  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  23  * SOFTWARE.
  24  *
  25  * Authors:
  26  *    Rob Clark <robclark@freedesktop.org>
  27  */
  28
  29 #include "pipe/p_state.h"
  30 #include "util/u_string.h"
  31 #include "util/u_memory.h"
  32 #include "util/u_inlines.h"
  33 #include "util/u_format.h"
  34
  35 #include "freedreno_gmem.h"
  36 #include "freedreno_context.h"
  37 #include "freedreno_fence.h"
  38 #include "freedreno_resource.h"
  39 #include "freedreno_query_hw.h"
  40 #include "freedreno_util.h"
  41
  42 /*
  43  * GMEM is the small (ie. 256KiB for a200, 512KiB for a220, etc) tile buffer
  44  * inside the GPU.  All rendering happens to GMEM.  Larger render targets
  45  * are split into tiles that are small enough for the color (and depth and/or
  46  * stencil, if enabled) buffers to fit within GMEM.  Before rendering a tile,
  47  * if there was not a clear invalidating the previous tile contents, we need
  48  * to restore the previous tiles contents (system mem -> GMEM), and after all
  49  * the draw calls, before moving to the next tile, we need to save the tile
  50  * contents (GMEM -> system mem).
  51  *
  52  * The code in this file handles dealing with GMEM and tiling.
  53  *
  54  * The structure of the ringbuffer ends up being:
  55  *
  56  *     +--<---<-- IB ---<---+---<---+---<---<---<--+
  57  *     |                    |       |              |
  58  *     v                    ^       ^              ^
  59  *   ------------------------------------------------------
  60  *     | clear/draw cmds | Tile0 | Tile1 | .... | TileN |
  61  *   ------------------------------------------------------
  62  *                       ^
  63  *                       |
  64  *                       address submitted in issueibcmds
  65  *
  66  * Where the per-tile section handles scissor setup, mem2gmem restore (if
  67  * needed), IB to draw cmds earlier in the ringbuffer, and then gmem2mem
  68  * resolve.
  69  */
  70
  71 static uint32_t bin_width(struct fd_screen *screen)
  72 {
  73         if (is_a4xx(screen) || is_a5xx(screen) || is_a6xx(screen))
  74                 return 1024;
  75         if (is_a3xx(screen))
  76                 return 992;
  77         return 512;
  78 }
  79
  80 static uint32_t
  81 total_size(uint8_t cbuf_cpp[], uint8_t zsbuf_cpp[2],
  82                    uint32_t bin_w, uint32_t bin_h, struct fd_gmem_stateobj *gmem)
  83 {
  84         uint32_t total = 0, i;
  85
  86         for (i = 0; i < MAX_RENDER_TARGETS; i++) {
  87                 if (cbuf_cpp[i]) {
  88                         gmem->cbuf_base[i] = align(total, 0x4000);
  89                         total = gmem->cbuf_base[i] + cbuf_cpp[i] * bin_w * bin_h;
  90                 }
  91         }
  92
  93         if (zsbuf_cpp[0]) {
  94                 gmem->zsbuf_base[0] = align(total, 0x4000);
  95                 total = gmem->zsbuf_base[0] + zsbuf_cpp[0] * bin_w * bin_h;
  96         }
  97
  98         if (zsbuf_cpp[1]) {
  99                 gmem->zsbuf_base[1] = align(total, 0x4000);
 100                 total = gmem->zsbuf_base[1] + zsbuf_cpp[1] * bin_w * bin_h;
 101         }
 102
 103         return total;
 104 }
 105
 106 static void
 107 calculate_tiles(struct fd_batch *batch)
 108 {
 109         struct fd_context *ctx = batch->ctx;
 110         struct fd_gmem_stateobj *gmem = &ctx->gmem;
 111         struct pipe_scissor_state *scissor = &batch->max_scissor;
 112         struct pipe_framebuffer_state *pfb = &batch->framebuffer;
 113         const uint32_t gmem_alignw = ctx->screen->gmem_alignw;
 114         const uint32_t gmem_alignh = ctx->screen->gmem_alignh;
 115         const unsigned npipes = ctx->screen->num_vsc_pipes;
 116         const uint32_t gmem_size = ctx->screen->gmemsize_bytes;
 117         uint32_t minx, miny, width, height;
 118         uint32_t nbins_x = 1, nbins_y = 1;
 119         uint32_t bin_w, bin_h;
 120         uint32_t max_width = bin_width(ctx->screen);
 121         uint8_t cbuf_cpp[MAX_RENDER_TARGETS] = {0}, zsbuf_cpp[2] = {0};
 122         uint32_t i, j, t, xoff, yoff;
 123         uint32_t tpp_x, tpp_y;
 124         bool has_zs = !!(batch->resolve & (FD_BUFFER_DEPTH | FD_BUFFER_STENCIL));
 125         int tile_n[npipes];
 126
 127         if (has_zs) {
 128                 struct fd_resource *rsc = fd_resource(pfb->zsbuf->texture);
 129                 zsbuf_cpp[0] = rsc->cpp;
 130                 if (rsc->stencil)
 131                         zsbuf_cpp[1] = rsc->stencil->cpp;
 132         }
 133         for (i = 0; i < pfb->nr_cbufs; i++) {
 134                 if (pfb->cbufs[i])
 135                         cbuf_cpp[i] = util_format_get_blocksize(pfb->cbufs[i]->format);
 136                 else
 137                         cbuf_cpp[i] = 4;
 138                 /* if MSAA, color buffers are super-sampled in GMEM: */
 139                 cbuf_cpp[i] *= pfb->samples;
 140         }
 141
 142         if (!memcmp(gmem->zsbuf_cpp, zsbuf_cpp, sizeof(zsbuf_cpp)) &&
 143                 !memcmp(gmem->cbuf_cpp, cbuf_cpp, sizeof(cbuf_cpp)) &&
 144                 !memcmp(&gmem->scissor, scissor, sizeof(gmem->scissor))) {
 145                 /* everything is up-to-date */
 146                 return;
 147         }
 148
 149         if (fd_mesa_debug & FD_DBG_NOSCIS) {
 150                 minx = 0;
 151                 miny = 0;
 152                 width = pfb->width;
 153                 height = pfb->height;
 154         } else {
 155                 /* round down to multiple of alignment: */
 156                 minx = scissor->minx & ~(gmem_alignw - 1);
 157                 miny = scissor->miny & ~(gmem_alignh - 1);
 158                 width = scissor->maxx - minx;
 159                 height = scissor->maxy - miny;
 160         }
 161
 162         bin_w = align(width, gmem_alignw);
 163         bin_h = align(height, gmem_alignh);
 164
 165         /* first, find a bin width that satisfies the maximum width
 166          * restrictions:
 167          */
 168         while (bin_w > max_width) {
 169                 nbins_x++;
 170                 bin_w = align(width / nbins_x, gmem_alignw);
 171         }
 172
 173         if (fd_mesa_debug & FD_DBG_MSGS) {
 174                 debug_printf("binning input: cbuf cpp:");
 175                 for (i = 0; i < pfb->nr_cbufs; i++)
 176                         debug_printf(" %d", cbuf_cpp[i]);
 177                 debug_printf(", zsbuf cpp: %d; %dx%d\n",
 178                                 zsbuf_cpp[0], width, height);
 179         }
 180
 181         /* then find a bin width/height that satisfies the memory
 182          * constraints:
 183          */
 184         while (total_size(cbuf_cpp, zsbuf_cpp, bin_w, bin_h, gmem) > gmem_size) {
 185                 if (bin_w > bin_h) {
 186                         nbins_x++;
 187                         bin_w = align(width / nbins_x, gmem_alignw);
 188                 } else {
 189                         nbins_y++;
 190                         bin_h = align(height / nbins_y, gmem_alignh);
 191                 }
 192         }
 193
 194         DBG("using %d bins of size %dx%d", nbins_x*nbins_y, bin_w, bin_h);
 195
 196         gmem->scissor = *scissor;
 197         memcpy(gmem->cbuf_cpp, cbuf_cpp, sizeof(cbuf_cpp));
 198         memcpy(gmem->zsbuf_cpp, zsbuf_cpp, sizeof(zsbuf_cpp));
 199         gmem->bin_h = bin_h;
 200         gmem->bin_w = bin_w;
 201         gmem->nbins_x = nbins_x;
 202         gmem->nbins_y = nbins_y;
 203         gmem->minx = minx;
 204         gmem->miny = miny;
 205         gmem->width = width;
 206         gmem->height = height;
 207
 208         /*
 209          * Assign tiles and pipes:
 210          *
 211          * At some point it might be worth playing with different
 212          * strategies and seeing if that makes much impact on
 213          * performance.
 214          */
 215
 216 #define div_round_up(v, a)  (((v) + (a) - 1) / (a))
 217         /* figure out number of tiles per pipe: */
 218         tpp_x = tpp_y = 1;
 219         while (div_round_up(nbins_y, tpp_y) > 8)
 220                 tpp_y += 2;
 221         while ((div_round_up(nbins_y, tpp_y) *
 222                         div_round_up(nbins_x, tpp_x)) > 8)
 223                 tpp_x += 1;
 224
 225         gmem->maxpw = tpp_x;
 226         gmem->maxph = tpp_y;
 227
 228         /* configure pipes: */
 229         xoff = yoff = 0;
 230         for (i = 0; i < npipes; i++) {
 231                 struct fd_vsc_pipe *pipe = &ctx->vsc_pipe[i];
 232
 233                 if (xoff >= nbins_x) {
 234                         xoff = 0;
 235                         yoff += tpp_y;
 236                 }
 237
 238                 if (yoff >= nbins_y) {
 239                         break;
 240                 }
 241
 242                 pipe->x = xoff;
 243                 pipe->y = yoff;
 244                 pipe->w = MIN2(tpp_x, nbins_x - xoff);
 245                 pipe->h = MIN2(tpp_y, nbins_y - yoff);
 246
 247                 xoff += tpp_x;
 248         }
 249
 250         for (; i < npipes; i++) {
 251                 struct fd_vsc_pipe *pipe = &ctx->vsc_pipe[i];
 252                 pipe->x = pipe->y = pipe->w = pipe->h = 0;
 253         }
 254
 255 #if 0 /* debug */
 256         printf("%dx%d ... tpp=%dx%d\n", nbins_x, nbins_y, tpp_x, tpp_y);
 257         for (i = 0; i < 8; i++) {
 258                 struct fd_vsc_pipe *pipe = &ctx->pipe[i];
 259                 printf("pipe[%d]: %ux%u @ %u,%u\n", i,
 260                                 pipe->w, pipe->h, pipe->x, pipe->y);
 261         }
 262 #endif
 263
 264         /* configure tiles: */
 265         t = 0;
 266         yoff = miny;
 267         memset(tile_n, 0, sizeof(tile_n));
 268         for (i = 0; i < nbins_y; i++) {
 269                 uint32_t bw, bh;
 270
 271                 xoff = minx;
 272
 273                 /* clip bin height: */
 274                 bh = MIN2(bin_h, miny + height - yoff);
 275
 276                 for (j = 0; j < nbins_x; j++) {
 277                         struct fd_tile *tile = &ctx->tile[t];
 278                         uint32_t p;
 279
 280                         assert(t < ARRAY_SIZE(ctx->tile));
 281
 282                         /* pipe number: */
 283                         p = ((i / tpp_y) * div_round_up(nbins_x, tpp_x)) + (j / tpp_x);
 284
 285                         /* clip bin width: */
 286                         bw = MIN2(bin_w, minx + width - xoff);
 287
 288                         tile->n = tile_n[p]++;
 289                         tile->p = p;
 290                         tile->bin_w = bw;
 291                         tile->bin_h = bh;
 292                         tile->xoff = xoff;
 293                         tile->yoff = yoff;
 294
 295                         t++;
 296
 297                         xoff += bw;
 298                 }
 299
 300                 yoff += bh;
 301         }
 302
 303 #if 0 /* debug */
 304         t = 0;
 305         for (i = 0; i < nbins_y; i++) {
 306                 for (j = 0; j < nbins_x; j++) {
 307                         struct fd_tile *tile = &ctx->tile[t++];
 308                         printf("|p:%u n:%u|", tile->p, tile->n);
 309                 }
 310                 printf("\n");
 311         }
 312 #endif
 313 }
 314
 315 static void
 316 render_tiles(struct fd_batch *batch)
 317 {
 318         struct fd_context *ctx = batch->ctx;
 319         struct fd_gmem_stateobj *gmem = &ctx->gmem;
 320         int i;
 321
 322         ctx->emit_tile_init(batch);
 323
 324         if (batch->restore)
 325                 ctx->stats.batch_restore++;
 326
 327         for (i = 0; i < (gmem->nbins_x * gmem->nbins_y); i++) {
 328                 struct fd_tile *tile = &ctx->tile[i];
 329
 330                 DBG("bin_h=%d, yoff=%d, bin_w=%d, xoff=%d",
 331                         tile->bin_h, tile->yoff, tile->bin_w, tile->xoff);
 332
 333                 ctx->emit_tile_prep(batch, tile);
 334
 335                 if (batch->restore) {
 336                         ctx->emit_tile_mem2gmem(batch, tile);
 337                 }
 338
 339                 ctx->emit_tile_renderprep(batch, tile);
 340
 341                 if (ctx->query_prepare_tile)
 342                         ctx->query_prepare_tile(batch, i, batch->gmem);
 343
 344                 /* emit IB to drawcmds: */
 345                 ctx->emit_ib(batch->gmem, batch->draw);
 346                 fd_reset_wfi(batch);
 347
 348                 /* emit gmem2mem to transfer tile back to system memory: */
 349                 ctx->emit_tile_gmem2mem(batch, tile);
 350         }
 351
 352         if (ctx->emit_tile_fini)
 353                 ctx->emit_tile_fini(batch);
 354 }
 355
 356 static void
 357 render_sysmem(struct fd_batch *batch)
 358 {
 359         struct fd_context *ctx = batch->ctx;
 360
 361         ctx->emit_sysmem_prep(batch);
 362
 363         if (ctx->query_prepare_tile)
 364                 ctx->query_prepare_tile(batch, 0, batch->gmem);
 365
 366         /* emit IB to drawcmds: */
 367         ctx->emit_ib(batch->gmem, batch->draw);
 368         fd_reset_wfi(batch);
 369
 370         if (ctx->emit_sysmem_fini)
 371                 ctx->emit_sysmem_fini(batch);
 372 }
 373
 374 static void
 375 flush_ring(struct fd_batch *batch)
 376 {
 377         /* for compute/blit batch, there is no batch->gmem, only batch->draw: */
 378         struct fd_ringbuffer *ring = batch->nondraw ? batch->draw : batch->gmem;
 379         uint32_t timestamp;
 380         int out_fence_fd = -1;
 381
 382         fd_ringbuffer_flush2(ring, batch->in_fence_fd,
 383                         batch->needs_out_fence_fd ? &out_fence_fd : NULL);
 384
 385         timestamp = fd_ringbuffer_timestamp(ring);
 386         fd_fence_populate(batch->fence, timestamp, out_fence_fd);
 387 }
 388
 389 void
 390 fd_gmem_render_tiles(struct fd_batch *batch)
 391 {
 392         struct fd_context *ctx = batch->ctx;
 393         struct pipe_framebuffer_state *pfb = &batch->framebuffer;
 394         bool sysmem = false;
 395
 396         if (ctx->emit_sysmem_prep && !batch->nondraw) {
 397                 if (batch->cleared || batch->gmem_reason ||
 398                                 ((batch->num_draws > 5) && !batch->blit) ||
 399                                 (pfb->samples > 1)) {
 400                         DBG("GMEM: cleared=%x, gmem_reason=%x, num_draws=%u, samples=%u",
 401                                 batch->cleared, batch->gmem_reason, batch->num_draws,
 402                                 pfb->samples);
 403                 } else if (!(fd_mesa_debug & FD_DBG_NOBYPASS)) {
 404                         sysmem = true;
 405                 }
 406
 407                 /* For ARB_framebuffer_no_attachments: */
 408                 if ((pfb->nr_cbufs == 0) && !pfb->zsbuf) {
 409                         sysmem = true;
 410                 }
 411         }
 412
 413         fd_reset_wfi(batch);
 414
 415         ctx->stats.batch_total++;
 416
 417         if (batch->nondraw) {
 418                 DBG("%p: rendering non-draw", batch);
 419                 ctx->stats.batch_nondraw++;
 420         } else if (sysmem) {
 421                 DBG("%p: rendering sysmem %ux%u (%s/%s)",
 422                         batch, pfb->width, pfb->height,
 423                         util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
 424                         util_format_short_name(pipe_surface_format(pfb->zsbuf)));
 425                 if (ctx->query_prepare)
 426                         ctx->query_prepare(batch, 1);
 427                 render_sysmem(batch);
 428                 ctx->stats.batch_sysmem++;
 429         } else {
 430                 struct fd_gmem_stateobj *gmem = &ctx->gmem;
 431                 calculate_tiles(batch);
 432                 DBG("%p: rendering %dx%d tiles %ux%u (%s/%s)",
 433                         batch, pfb->width, pfb->height, gmem->nbins_x, gmem->nbins_y,
 434                         util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
 435                         util_format_short_name(pipe_surface_format(pfb->zsbuf)));
 436                 if (ctx->query_prepare)
 437                         ctx->query_prepare(batch, gmem->nbins_x * gmem->nbins_y);
 438                 render_tiles(batch);
 439                 ctx->stats.batch_gmem++;
 440         }
 441
 442         flush_ring(batch);
 443 }
 444
 445 /* When deciding whether a tile needs mem2gmem, we need to take into
 446  * account the scissor rect(s) that were cleared.  To simplify we only
 447  * consider the last scissor rect for each buffer, since the common
 448  * case would be a single clear.
 449  */
 450 bool
 451 fd_gmem_needs_restore(struct fd_batch *batch, struct fd_tile *tile,
 452                 uint32_t buffers)
 453 {
 454         if (!(batch->restore & buffers))
 455                 return false;
 456
 457         return true;
 458 }