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_resource.h"
  38 #include "freedreno_query_hw.h"
  39 #include "freedreno_util.h"
  40
  41 /*
  42  * GMEM is the small (ie. 256KiB for a200, 512KiB for a220, etc) tile buffer
  43  * inside the GPU.  All rendering happens to GMEM.  Larger render targets
  44  * are split into tiles that are small enough for the color (and depth and/or
  45  * stencil, if enabled) buffers to fit within GMEM.  Before rendering a tile,
  46  * if there was not a clear invalidating the previous tile contents, we need
  47  * to restore the previous tiles contents (system mem -> GMEM), and after all
  48  * the draw calls, before moving to the next tile, we need to save the tile
  49  * contents (GMEM -> system mem).
  50  *
  51  * The code in this file handles dealing with GMEM and tiling.
  52  *
  53  * The structure of the ringbuffer ends up being:
  54  *
  55  *     +--<---<-- IB ---<---+---<---+---<---<---<--+
  56  *     |                    |       |              |
  57  *     v                    ^       ^              ^
  58  *   ------------------------------------------------------
  59  *     | clear/draw cmds | Tile0 | Tile1 | .... | TileN |
  60  *   ------------------------------------------------------
  61  *                       ^
  62  *                       |
  63  *                       address submitted in issueibcmds
  64  *
  65  * Where the per-tile section handles scissor setup, mem2gmem restore (if
  66  * needed), IB to draw cmds earlier in the ringbuffer, and then gmem2mem
  67  * resolve.
  68  */
  69
  70 static uint32_t bin_width(struct fd_context *ctx)
  71 {
  72         if (is_a4xx(ctx->screen))
  73                 return 1024;
  74         if (is_a3xx(ctx->screen))
  75                 return 992;
  76         return 512;
  77 }
  78
  79 static uint32_t
  80 total_size(uint8_t cbuf_cpp[], uint8_t zsbuf_cpp[2],
  81                    uint32_t bin_w, uint32_t bin_h, struct fd_gmem_stateobj *gmem)
  82 {
  83         uint32_t total = 0, i;
  84
  85         for (i = 0; i < 4; i++) {
  86                 if (cbuf_cpp[i]) {
  87                         gmem->cbuf_base[i] = align(total, 0x4000);
  88                         total = gmem->cbuf_base[i] + cbuf_cpp[i] * bin_w * bin_h;
  89                 }
  90         }
  91
  92         if (zsbuf_cpp[0]) {
  93                 gmem->zsbuf_base[0] = align(total, 0x4000);
  94                 total = gmem->zsbuf_base[0] + zsbuf_cpp[0] * bin_w * bin_h;
  95         }
  96
  97         if (zsbuf_cpp[1]) {
  98                 gmem->zsbuf_base[1] = align(total, 0x4000);
  99                 total = gmem->zsbuf_base[1] + zsbuf_cpp[1] * bin_w * bin_h;
 100         }
 101
 102         return total;
 103 }
 104
 105 static void
 106 calculate_tiles(struct fd_context *ctx)
 107 {
 108         struct fd_gmem_stateobj *gmem = &ctx->gmem;
 109         struct pipe_scissor_state *scissor = &ctx->max_scissor;
 110         struct pipe_framebuffer_state *pfb = &ctx->framebuffer;
 111         uint32_t gmem_size = ctx->screen->gmemsize_bytes;
 112         uint32_t minx, miny, width, height;
 113         uint32_t nbins_x = 1, nbins_y = 1;
 114         uint32_t bin_w, bin_h;
 115         uint32_t max_width = bin_width(ctx);
 116         uint8_t cbuf_cpp[4] = {0}, zsbuf_cpp[2] = {0};
 117         uint32_t i, j, t, xoff, yoff;
 118         uint32_t tpp_x, tpp_y;
 119         bool has_zs = !!(ctx->resolve & (FD_BUFFER_DEPTH | FD_BUFFER_STENCIL));
 120         int tile_n[ARRAY_SIZE(ctx->pipe)];
 121
 122         if (has_zs) {
 123                 struct fd_resource *rsc = fd_resource(pfb->zsbuf->texture);
 124                 zsbuf_cpp[0] = rsc->cpp;
 125                 if (rsc->stencil)
 126                         zsbuf_cpp[1] = rsc->stencil->cpp;
 127         }
 128         for (i = 0; i < pfb->nr_cbufs; i++) {
 129                 if (pfb->cbufs[i])
 130                         cbuf_cpp[i] = util_format_get_blocksize(pfb->cbufs[i]->format);
 131                 else
 132                         cbuf_cpp[i] = 4;
 133         }
 134
 135         if (!memcmp(gmem->zsbuf_cpp, zsbuf_cpp, sizeof(zsbuf_cpp)) &&
 136                 !memcmp(gmem->cbuf_cpp, cbuf_cpp, sizeof(cbuf_cpp)) &&
 137                 !memcmp(&gmem->scissor, scissor, sizeof(gmem->scissor))) {
 138                 /* everything is up-to-date */
 139                 return;
 140         }
 141
 142         if (fd_mesa_debug & FD_DBG_NOSCIS) {
 143                 minx = 0;
 144                 miny = 0;
 145                 width = pfb->width;
 146                 height = pfb->height;
 147         } else {
 148                 minx = scissor->minx & ~31; /* round down to multiple of 32 */
 149                 miny = scissor->miny & ~31;
 150                 width = scissor->maxx - minx;
 151                 height = scissor->maxy - miny;
 152         }
 153
 154         bin_w = align(width, 32);
 155         bin_h = align(height, 32);
 156
 157         /* first, find a bin width that satisfies the maximum width
 158          * restrictions:
 159          */
 160         while (bin_w > max_width) {
 161                 nbins_x++;
 162                 bin_w = align(width / nbins_x, 32);
 163         }
 164
 165         /* then find a bin width/height that satisfies the memory
 166          * constraints:
 167          */
 168         DBG("binning input: cbuf cpp: %d %d %d %d, zsbuf cpp: %d; %dx%d",
 169                 cbuf_cpp[0], cbuf_cpp[1], cbuf_cpp[2], cbuf_cpp[3], zsbuf_cpp[0],
 170                 width, height);
 171         while (total_size(cbuf_cpp, zsbuf_cpp, bin_w, bin_h, gmem) > gmem_size) {
 172                 if (bin_w > bin_h) {
 173                         nbins_x++;
 174                         bin_w = align(width / nbins_x, 32);
 175                 } else {
 176                         nbins_y++;
 177                         bin_h = align(height / nbins_y, 32);
 178                 }
 179         }
 180
 181         DBG("using %d bins of size %dx%d", nbins_x*nbins_y, bin_w, bin_h);
 182
 183         gmem->scissor = *scissor;
 184         memcpy(gmem->cbuf_cpp, cbuf_cpp, sizeof(cbuf_cpp));
 185         memcpy(gmem->zsbuf_cpp, zsbuf_cpp, sizeof(zsbuf_cpp));
 186         gmem->bin_h = bin_h;
 187         gmem->bin_w = bin_w;
 188         gmem->nbins_x = nbins_x;
 189         gmem->nbins_y = nbins_y;
 190         gmem->minx = minx;
 191         gmem->miny = miny;
 192         gmem->width = width;
 193         gmem->height = height;
 194
 195         /*
 196          * Assign tiles and pipes:
 197          *
 198          * At some point it might be worth playing with different
 199          * strategies and seeing if that makes much impact on
 200          * performance.
 201          */
 202
 203 #define div_round_up(v, a)  (((v) + (a) - 1) / (a))
 204         /* figure out number of tiles per pipe: */
 205         tpp_x = tpp_y = 1;
 206         while (div_round_up(nbins_y, tpp_y) > 8)
 207                 tpp_y += 2;
 208         while ((div_round_up(nbins_y, tpp_y) *
 209                         div_round_up(nbins_x, tpp_x)) > 8)
 210                 tpp_x += 1;
 211
 212         /* configure pipes: */
 213         xoff = yoff = 0;
 214         for (i = 0; i < ARRAY_SIZE(ctx->pipe); i++) {
 215                 struct fd_vsc_pipe *pipe = &ctx->pipe[i];
 216
 217                 if (xoff >= nbins_x) {
 218                         xoff = 0;
 219                         yoff += tpp_y;
 220                 }
 221
 222                 if (yoff >= nbins_y) {
 223                         break;
 224                 }
 225
 226                 pipe->x = xoff;
 227                 pipe->y = yoff;
 228                 pipe->w = MIN2(tpp_x, nbins_x - xoff);
 229                 pipe->h = MIN2(tpp_y, nbins_y - yoff);
 230
 231                 xoff += tpp_x;
 232         }
 233
 234         for (; i < ARRAY_SIZE(ctx->pipe); i++) {
 235                 struct fd_vsc_pipe *pipe = &ctx->pipe[i];
 236                 pipe->x = pipe->y = pipe->w = pipe->h = 0;
 237         }
 238
 239 #if 0 /* debug */
 240         printf("%dx%d ... tpp=%dx%d\n", nbins_x, nbins_y, tpp_x, tpp_y);
 241         for (i = 0; i < 8; i++) {
 242                 struct fd_vsc_pipe *pipe = &ctx->pipe[i];
 243                 printf("pipe[%d]: %ux%u @ %u,%u\n", i,
 244                                 pipe->w, pipe->h, pipe->x, pipe->y);
 245         }
 246 #endif
 247
 248         /* configure tiles: */
 249         t = 0;
 250         yoff = miny;
 251         memset(tile_n, 0, sizeof(tile_n));
 252         for (i = 0; i < nbins_y; i++) {
 253                 uint32_t bw, bh;
 254
 255                 xoff = minx;
 256
 257                 /* clip bin height: */
 258                 bh = MIN2(bin_h, miny + height - yoff);
 259
 260                 for (j = 0; j < nbins_x; j++) {
 261                         struct fd_tile *tile = &ctx->tile[t];
 262                         uint32_t p;
 263
 264                         assert(t < ARRAY_SIZE(ctx->tile));
 265
 266                         /* pipe number: */
 267                         p = ((i / tpp_y) * div_round_up(nbins_x, tpp_x)) + (j / tpp_x);
 268
 269                         /* clip bin width: */
 270                         bw = MIN2(bin_w, minx + width - xoff);
 271
 272                         tile->n = tile_n[p]++;
 273                         tile->p = p;
 274                         tile->bin_w = bw;
 275                         tile->bin_h = bh;
 276                         tile->xoff = xoff;
 277                         tile->yoff = yoff;
 278
 279                         t++;
 280
 281                         xoff += bw;
 282                 }
 283
 284                 yoff += bh;
 285         }
 286
 287 #if 0 /* debug */
 288         t = 0;
 289         for (i = 0; i < nbins_y; i++) {
 290                 for (j = 0; j < nbins_x; j++) {
 291                         struct fd_tile *tile = &ctx->tile[t++];
 292                         printf("|p:%u n:%u|", tile->p, tile->n);
 293                 }
 294                 printf("\n");
 295         }
 296 #endif
 297 }
 298
 299 static void
 300 render_tiles(struct fd_context *ctx)
 301 {
 302         struct fd_gmem_stateobj *gmem = &ctx->gmem;
 303         int i;
 304
 305         ctx->emit_tile_init(ctx);
 306
 307         if (ctx->restore)
 308                 ctx->stats.batch_restore++;
 309
 310         for (i = 0; i < (gmem->nbins_x * gmem->nbins_y); i++) {
 311                 struct fd_tile *tile = &ctx->tile[i];
 312
 313                 DBG("bin_h=%d, yoff=%d, bin_w=%d, xoff=%d",
 314                         tile->bin_h, tile->yoff, tile->bin_w, tile->xoff);
 315
 316                 ctx->emit_tile_prep(ctx, tile);
 317
 318                 if (ctx->restore) {
 319                         fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_MEM2GMEM);
 320                         ctx->emit_tile_mem2gmem(ctx, tile);
 321                         fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_NULL);
 322                 }
 323
 324                 ctx->emit_tile_renderprep(ctx, tile);
 325
 326                 fd_hw_query_prepare_tile(ctx, i, ctx->ring);
 327
 328                 /* emit IB to drawcmds: */
 329                 OUT_IB(ctx->ring, ctx->draw_start, ctx->draw_end);
 330                 fd_reset_wfi(ctx);
 331
 332                 /* emit gmem2mem to transfer tile back to system memory: */
 333                 fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_GMEM2MEM);
 334                 ctx->emit_tile_gmem2mem(ctx, tile);
 335                 fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_NULL);
 336         }
 337 }
 338
 339 static void
 340 render_sysmem(struct fd_context *ctx)
 341 {
 342         ctx->emit_sysmem_prep(ctx);
 343
 344         fd_hw_query_prepare_tile(ctx, 0, ctx->ring);
 345
 346         /* emit IB to drawcmds: */
 347         OUT_IB(ctx->ring, ctx->draw_start, ctx->draw_end);
 348         fd_reset_wfi(ctx);
 349 }
 350
 351 void
 352 fd_gmem_render_tiles(struct fd_context *ctx)
 353 {
 354         struct pipe_framebuffer_state *pfb = &ctx->framebuffer;
 355         bool sysmem = false;
 356
 357         if (ctx->emit_sysmem_prep) {
 358                 if (ctx->cleared || ctx->gmem_reason || (ctx->num_draws > 5)) {
 359                         DBG("GMEM: cleared=%x, gmem_reason=%x, num_draws=%u",
 360                                 ctx->cleared, ctx->gmem_reason, ctx->num_draws);
 361                 } else if (!(fd_mesa_debug & FD_DBG_NOBYPASS)) {
 362                         sysmem = true;
 363                 }
 364         }
 365
 366         /* close out the draw cmds by making sure any active queries are
 367          * paused:
 368          */
 369         fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_NULL);
 370
 371         /* mark the end of the clear/draw cmds before emitting per-tile cmds: */
 372         fd_ringmarker_mark(ctx->draw_end);
 373         fd_ringmarker_mark(ctx->binning_end);
 374
 375         fd_reset_wfi(ctx);
 376
 377         ctx->stats.batch_total++;
 378
 379         if (sysmem) {
 380                 DBG("rendering sysmem (%s/%s)",
 381                         util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
 382                         util_format_short_name(pipe_surface_format(pfb->zsbuf)));
 383                 fd_hw_query_prepare(ctx, 1);
 384                 render_sysmem(ctx);
 385                 ctx->stats.batch_sysmem++;
 386         } else {
 387                 struct fd_gmem_stateobj *gmem = &ctx->gmem;
 388                 calculate_tiles(ctx);
 389                 DBG("rendering %dx%d tiles (%s/%s)", gmem->nbins_x, gmem->nbins_y,
 390                         util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
 391                         util_format_short_name(pipe_surface_format(pfb->zsbuf)));
 392                 fd_hw_query_prepare(ctx, gmem->nbins_x * gmem->nbins_y);
 393                 render_tiles(ctx);
 394                 ctx->stats.batch_gmem++;
 395         }
 396
 397         /* GPU executes starting from tile cmds, which IB back to draw cmds: */
 398         fd_ringmarker_flush(ctx->draw_end);
 399
 400         /* mark start for next draw/binning cmds: */
 401         fd_ringmarker_mark(ctx->draw_start);
 402         fd_ringmarker_mark(ctx->binning_start);
 403
 404         fd_reset_wfi(ctx);
 405
 406         /* reset maximal bounds: */
 407         ctx->max_scissor.minx = ctx->max_scissor.miny = ~0;
 408         ctx->max_scissor.maxx = ctx->max_scissor.maxy = 0;
 409
 410         ctx->dirty = ~0;
 411 }
 412
 413 /* tile needs restore if it isn't completely contained within the
 414  * cleared scissor:
 415  */
 416 static bool
 417 skip_restore(struct pipe_scissor_state *scissor, struct fd_tile *tile)
 418 {
 419         unsigned minx = tile->xoff;
 420         unsigned maxx = tile->xoff + tile->bin_w;
 421         unsigned miny = tile->yoff;
 422         unsigned maxy = tile->yoff + tile->bin_h;
 423         return (minx >= scissor->minx) && (maxx <= scissor->maxx) &&
 424                         (miny >= scissor->miny) && (maxy <= scissor->maxy);
 425 }
 426
 427 /* When deciding whether a tile needs mem2gmem, we need to take into
 428  * account the scissor rect(s) that were cleared.  To simplify we only
 429  * consider the last scissor rect for each buffer, since the common
 430  * case would be a single clear.
 431  */
 432 bool
 433 fd_gmem_needs_restore(struct fd_context *ctx, struct fd_tile *tile,
 434                 uint32_t buffers)
 435 {
 436         if (!(ctx->restore & buffers))
 437                 return false;
 438
 439         /* if buffers partially cleared, then slow-path to figure out
 440          * if this particular tile needs restoring:
 441          */
 442         if ((buffers & FD_BUFFER_COLOR) &&
 443                         (ctx->partial_cleared & FD_BUFFER_COLOR) &&
 444                         skip_restore(&ctx->cleared_scissor.color, tile))
 445                 return false;
 446         if ((buffers & FD_BUFFER_DEPTH) &&
 447                         (ctx->partial_cleared & FD_BUFFER_DEPTH) &&
 448                         skip_restore(&ctx->cleared_scissor.depth, tile))
 449                 return false;
 450         if ((buffers & FD_BUFFER_STENCIL) &&
 451                         (ctx->partial_cleared & FD_BUFFER_STENCIL) &&
 452                         skip_restore(&ctx->cleared_scissor.stencil, tile))
 453                 return false;
 454
 455         return true;
 456 }