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_util.h"
  39
  40 /*
  41  * GMEM is the small (ie. 256KiB for a200, 512KiB for a220, etc) tile buffer
  42  * inside the GPU.  All rendering happens to GMEM.  Larger render targets
  43  * are split into tiles that are small enough for the color (and depth and/or
  44  * stencil, if enabled) buffers to fit within GMEM.  Before rendering a tile,
  45  * if there was not a clear invalidating the previous tile contents, we need
  46  * to restore the previous tiles contents (system mem -> GMEM), and after all
  47  * the draw calls, before moving to the next tile, we need to save the tile
  48  * contents (GMEM -> system mem).
  49  *
  50  * The code in this file handles dealing with GMEM and tiling.
  51  *
  52  * The structure of the ringbuffer ends up being:
  53  *
  54  *     +--<---<-- IB ---<---+---<---+---<---<---<--+
  55  *     |                    |       |              |
  56  *     v                    ^       ^              ^
  57  *   ------------------------------------------------------
  58  *     | clear/draw cmds | Tile0 | Tile1 | .... | TileN |
  59  *   ------------------------------------------------------
  60  *                       ^
  61  *                       |
  62  *                       address submitted in issueibcmds
  63  *
  64  * Where the per-tile section handles scissor setup, mem2gmem restore (if
  65  * needed), IB to draw cmds earlier in the ringbuffer, and then gmem2mem
  66  * resolve.
  67  */
  68
  69 static void
  70 calculate_tiles(struct fd_context *ctx)
  71 {
  72         struct fd_gmem_stateobj *gmem = &ctx->gmem;
  73         struct pipe_scissor_state *scissor = &ctx->max_scissor;
  74         uint32_t cpp = util_format_get_blocksize(ctx->framebuffer.cbufs[0]->format);
  75         uint32_t gmem_size = ctx->screen->gmemsize_bytes;
  76         uint32_t minx, miny, width, height;
  77         uint32_t nbins_x = 1, nbins_y = 1;
  78         uint32_t bin_w, bin_h;
  79         uint32_t max_width = 992;
  80
  81         if ((gmem->cpp == cpp) &&
  82                         !memcmp(&gmem->scissor, scissor, sizeof(gmem->scissor))) {
  83                 /* everything is up-to-date */
  84                 return;
  85         }
  86
  87         minx = scissor->minx & ~31; /* round down to multiple of 32 */
  88         miny = scissor->miny & ~31;
  89         width = scissor->maxx - minx;
  90         height = scissor->maxy - miny;
  91
  92 // TODO we probably could optimize this a bit if we know that
  93 // Z or stencil is not enabled for any of the draw calls..
  94 //      if (fd_stencil_enabled(ctx->zsa) || fd_depth_enabled(ctx->zsa)) {
  95                 gmem_size /= 2;
  96                 max_width = 256;
  97 //      }
  98
  99         bin_w = align(width, 32);
 100         bin_h = align(height, 32);
 101
 102         /* first, find a bin width that satisfies the maximum width
 103          * restrictions:
 104          */
 105         while (bin_w > max_width) {
 106                 nbins_x++;
 107                 bin_w = align(width / nbins_x, 32);
 108         }
 109
 110         /* then find a bin height that satisfies the memory constraints:
 111          */
 112         while ((bin_w * bin_h * cpp) > gmem_size) {
 113                 nbins_y++;
 114                 bin_h = align(height / nbins_y, 32);
 115         }
 116
 117         DBG("using %d bins of size %dx%d", nbins_x*nbins_y, bin_w, bin_h);
 118
 119         gmem->scissor = *scissor;
 120         gmem->cpp = cpp;
 121         gmem->minx = minx;
 122         gmem->miny = miny;
 123         gmem->bin_h = bin_h;
 124         gmem->bin_w = bin_w;
 125         gmem->nbins_x = nbins_x;
 126         gmem->nbins_y = nbins_y;
 127         gmem->width = width;
 128         gmem->height = height;
 129 }
 130
 131 static void
 132 render_tiles(struct fd_context *ctx)
 133 {
 134         struct fd_gmem_stateobj *gmem = &ctx->gmem;
 135         uint32_t i, yoff = 0;
 136
 137         yoff= gmem->miny;
 138
 139         ctx->emit_tile_init(ctx);
 140
 141         for (i = 0; i < gmem->nbins_y; i++) {
 142                 uint32_t j, xoff = gmem->minx;
 143                 uint32_t bh = gmem->bin_h;
 144
 145                 /* clip bin height: */
 146                 bh = MIN2(bh, gmem->height - yoff);
 147
 148                 for (j = 0; j < gmem->nbins_x; j++) {
 149                         uint32_t bw = gmem->bin_w;
 150
 151                         /* clip bin width: */
 152                         bw = MIN2(bw, gmem->width - xoff);
 153
 154                         DBG("bin_h=%d, yoff=%d, bin_w=%d, xoff=%d",
 155                                         bh, yoff, bw, xoff);
 156
 157                         ctx->emit_tile_prep(ctx, xoff, yoff, bw, bh);
 158
 159                         if (ctx->restore)
 160                                 ctx->emit_tile_mem2gmem(ctx, xoff, yoff, bw, bh);
 161
 162                         ctx->emit_tile_renderprep(ctx, xoff, yoff, bw, bh);
 163
 164                         /* emit IB to drawcmds: */
 165                         OUT_IB(ctx->ring, ctx->draw_start, ctx->draw_end);
 166
 167                         /* emit gmem2mem to transfer tile back to system memory: */
 168                         ctx->emit_tile_gmem2mem(ctx, xoff, yoff, bw, bh);
 169
 170                         xoff += bw;
 171                 }
 172
 173                 yoff += bh;
 174         }
 175 }
 176
 177 static void
 178 render_sysmem(struct fd_context *ctx)
 179 {
 180         ctx->emit_sysmem_prep(ctx);
 181
 182         /* emit IB to drawcmds: */
 183         OUT_IB(ctx->ring, ctx->draw_start, ctx->draw_end);
 184 }
 185
 186 void
 187 fd_gmem_render_tiles(struct pipe_context *pctx)
 188 {
 189         struct fd_context *ctx = fd_context(pctx);
 190         struct pipe_framebuffer_state *pfb = &ctx->framebuffer;
 191         uint32_t timestamp = 0;
 192         bool sysmem = false;
 193
 194         if (ctx->emit_sysmem_prep) {
 195                 if (ctx->cleared || ctx->gmem_reason || (ctx->num_draws > 5)) {
 196                         DBG("GMEM: cleared=%x, gmem_reason=%x, num_draws=%u",
 197                                 ctx->cleared, ctx->gmem_reason, ctx->num_draws);
 198                 } else {
 199                         sysmem = true;
 200                 }
 201         }
 202
 203         /* mark the end of the clear/draw cmds before emitting per-tile cmds: */
 204         fd_ringmarker_mark(ctx->draw_end);
 205
 206         if (sysmem) {
 207                 DBG("rendering sysmem (%s/%s)",
 208                         util_format_name(pfb->cbufs[0]->format),
 209                         pfb->zsbuf ? util_format_name(pfb->zsbuf->format) : "none");
 210                 render_sysmem(ctx);
 211         } else {
 212                 struct fd_gmem_stateobj *gmem = &ctx->gmem;
 213                 DBG("rendering %dx%d tiles (%s/%s)", gmem->nbins_x, gmem->nbins_y,
 214                         util_format_name(pfb->cbufs[0]->format),
 215                         pfb->zsbuf ? util_format_name(pfb->zsbuf->format) : "none");
 216                 calculate_tiles(ctx);
 217                 render_tiles(ctx);
 218         }
 219
 220         /* GPU executes starting from tile cmds, which IB back to draw cmds: */
 221         fd_ringmarker_flush(ctx->draw_end);
 222
 223         /* mark start for next draw cmds: */
 224         fd_ringmarker_mark(ctx->draw_start);
 225
 226         /* update timestamps on render targets: */
 227         timestamp = fd_ringbuffer_timestamp(ctx->ring);
 228         fd_resource(pfb->cbufs[0]->texture)->timestamp = timestamp;
 229         if (pfb->zsbuf)
 230                 fd_resource(pfb->zsbuf->texture)->timestamp = timestamp;
 231
 232         /* reset maximal bounds: */
 233         ctx->max_scissor.minx = ctx->max_scissor.miny = ~0;
 234         ctx->max_scissor.maxx = ctx->max_scissor.maxy = 0;
 235
 236         /* Note that because the per-tile setup and mem2gmem/gmem2mem are emitted
 237          * after the draw/clear calls, but executed before, we need to preemptively
 238          * flag some state as dirty before the first draw/clear call.
 239          *
 240          * TODO maybe we need to mark all state as dirty to not worry about state
 241          * being clobbered by other contexts?
 242          */
 243         ctx->dirty |= FD_DIRTY_ZSA |
 244                         FD_DIRTY_RASTERIZER |
 245                         FD_DIRTY_FRAMEBUFFER |
 246                         FD_DIRTY_SAMPLE_MASK |
 247                         FD_DIRTY_VIEWPORT |
 248                         FD_DIRTY_CONSTBUF |
 249                         FD_DIRTY_PROG |
 250                         FD_DIRTY_SCISSOR |
 251                         /* probably only needed if we need to mem2gmem on the next
 252                          * draw..  but not sure if there is a good way to know?
 253                          */
 254                         FD_DIRTY_VERTTEX |
 255                         FD_DIRTY_FRAGTEX |
 256                         FD_DIRTY_BLEND;
 257
 258         if (fd_mesa_debug & FD_DBG_DGMEM)
 259                 ctx->dirty = 0xffffffff;
 260 }