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_pack_color.h"
  34
  35 #include "freedreno_gmem.h"
  36 #include "freedreno_context.h"
  37 #include "freedreno_state.h"
  38 #include "freedreno_program.h"
  39 #include "freedreno_resource.h"
  40 #include "freedreno_zsa.h"
  41 #include "freedreno_util.h"
  42
  43 /*
  44  * GMEM is the small (ie. 256KiB for a200, 512KiB for a220, etc) tile buffer
  45  * inside the GPU.  All rendering happens to GMEM.  Larger render targets
  46  * are split into tiles that are small enough for the color (and depth and/or
  47  * stencil, if enabled) buffers to fit within GMEM.  Before rendering a tile,
  48  * if there was not a clear invalidating the previous tile contents, we need
  49  * to restore the previous tiles contents (system mem -> GMEM), and after all
  50  * the draw calls, before moving to the next tile, we need to save the tile
  51  * contents (GMEM -> system mem).
  52  *
  53  * The code in this file handles dealing with GMEM and tiling.
  54  *
  55  * The structure of the ringbuffer ends up being:
  56  *
  57  *     +--<---<-- IB ---<---+---<---+---<---<---<--+
  58  *     |                    |       |              |
  59  *     v                    ^       ^              ^
  60  *   ------------------------------------------------------
  61  *     | clear/draw cmds | Tile0 | Tile1 | .... | TileN |
  62  *   ------------------------------------------------------
  63  *                       ^
  64  *                       |
  65  *                       address submitted in issueibcmds
  66  *
  67  * Where the per-tile section handles scissor setup, mem2gmem restore (if
  68  * needed), IB to draw cmds earlier in the ringbuffer, and then gmem2mem
  69  * resolve.
  70  */
  71
  72 /* transfer from gmem to system memory (ie. normal RAM) */
  73
  74 static void
  75 emit_gmem2mem_surf(struct fd_ringbuffer *ring, uint32_t swap, uint32_t base,
  76                 struct pipe_surface *psurf)
  77 {
  78         struct fd_resource *rsc = fd_resource(psurf->texture);
  79
  80         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
  81         OUT_RING(ring, CP_REG(REG_A2XX_RB_COLOR_INFO));
  82         OUT_RING(ring, A2XX_RB_COLOR_INFO_SWAP(swap) |
  83                         A2XX_RB_COLOR_INFO_BASE(base / 1024) |
  84                         A2XX_RB_COLOR_INFO_FORMAT(fd_pipe2color(psurf->format)));
  85
  86         OUT_PKT3(ring, CP_SET_CONSTANT, 5);
  87         OUT_RING(ring, CP_REG(REG_A2XX_RB_COPY_CONTROL));
  88         OUT_RING(ring, 0x00000000);             /* RB_COPY_CONTROL */
  89         OUT_RELOC(ring, rsc->bo, 0, 0);         /* RB_COPY_DEST_BASE */
  90         OUT_RING(ring, rsc->pitch >> 5);        /* RB_COPY_DEST_PITCH */
  91         OUT_RING(ring,                          /* RB_COPY_DEST_INFO */
  92                         A2XX_RB_COPY_DEST_INFO_FORMAT(fd_pipe2color(psurf->format)) |
  93                         A2XX_RB_COPY_DEST_INFO_LINEAR |
  94                         A2XX_RB_COPY_DEST_INFO_SWAP(swap) |
  95                         A2XX_RB_COPY_DEST_INFO_WRITE_RED |
  96                         A2XX_RB_COPY_DEST_INFO_WRITE_GREEN |
  97                         A2XX_RB_COPY_DEST_INFO_WRITE_BLUE |
  98                         A2XX_RB_COPY_DEST_INFO_WRITE_ALPHA);
  99
 100         OUT_PKT3(ring, CP_WAIT_FOR_IDLE, 1);
 101         OUT_RING(ring, 0x0000000);
 102
 103         OUT_PKT3(ring, CP_DRAW_INDX, 3);
 104         OUT_RING(ring, 0x00000000);
 105         OUT_RING(ring, DRAW(DI_PT_RECTLIST, DI_SRC_SEL_AUTO_INDEX,
 106                         INDEX_SIZE_IGN, IGNORE_VISIBILITY));
 107         OUT_RING(ring, 3);                                      /* NumIndices */
 108 }
 109
 110 static void
 111 emit_gmem2mem(struct fd_context *ctx, struct fd_ringbuffer *ring,
 112                 uint32_t xoff, uint32_t yoff, uint32_t bin_w, uint32_t bin_h)
 113 {
 114         struct pipe_framebuffer_state *pfb = &ctx->framebuffer;
 115
 116         fd_emit_vertex_bufs(ring, 0x9c, (struct fd_vertex_buf[]) {
 117                         { .prsc = ctx->solid_vertexbuf, .size = 48 },
 118                 }, 1);
 119
 120         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 121         OUT_RING(ring, CP_REG(REG_A2XX_VGT_INDX_OFFSET));
 122         OUT_RING(ring, 0);
 123
 124         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 125         OUT_RING(ring, CP_REG(REG_A2XX_VGT_VERTEX_REUSE_BLOCK_CNTL));
 126         OUT_RING(ring, 0x0000028f);
 127
 128         fd_program_emit(ring, &ctx->solid_prog);
 129
 130         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 131         OUT_RING(ring, CP_REG(REG_A2XX_PA_SC_AA_MASK));
 132         OUT_RING(ring, 0x0000ffff);
 133
 134         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 135         OUT_RING(ring, CP_REG(REG_A2XX_RB_DEPTHCONTROL));
 136         OUT_RING(ring, A2XX_RB_DEPTHCONTROL_EARLY_Z_ENABLE);
 137
 138         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 139         OUT_RING(ring, CP_REG(REG_A2XX_PA_SU_SC_MODE_CNTL));
 140         OUT_RING(ring, A2XX_PA_SU_SC_MODE_CNTL_PROVOKING_VTX_LAST |  /* PA_SU_SC_MODE_CNTL */
 141                         A2XX_PA_SU_SC_MODE_CNTL_FRONT_PTYPE(PC_DRAW_TRIANGLES) |
 142                         A2XX_PA_SU_SC_MODE_CNTL_BACK_PTYPE(PC_DRAW_TRIANGLES));
 143
 144         OUT_PKT3(ring, CP_SET_CONSTANT, 3);
 145         OUT_RING(ring, CP_REG(REG_A2XX_PA_SC_WINDOW_SCISSOR_TL));
 146         OUT_RING(ring, xy2d(0, 0));                       /* PA_SC_WINDOW_SCISSOR_TL */
 147         OUT_RING(ring, xy2d(pfb->width, pfb->height));    /* PA_SC_WINDOW_SCISSOR_BR */
 148
 149         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 150         OUT_RING(ring, CP_REG(REG_A2XX_PA_CL_VTE_CNTL));
 151         OUT_RING(ring, A2XX_PA_CL_VTE_CNTL_VTX_W0_FMT |
 152                         A2XX_PA_CL_VTE_CNTL_VPORT_X_SCALE_ENA |
 153                         A2XX_PA_CL_VTE_CNTL_VPORT_X_OFFSET_ENA |
 154                         A2XX_PA_CL_VTE_CNTL_VPORT_Y_SCALE_ENA |
 155                         A2XX_PA_CL_VTE_CNTL_VPORT_Y_OFFSET_ENA);
 156
 157         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 158         OUT_RING(ring, CP_REG(REG_A2XX_PA_CL_CLIP_CNTL));
 159         OUT_RING(ring, 0x00000000);
 160
 161         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 162         OUT_RING(ring, CP_REG(REG_A2XX_RB_MODECONTROL));
 163         OUT_RING(ring, A2XX_RB_MODECONTROL_EDRAM_MODE(EDRAM_COPY));
 164
 165         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 166         OUT_RING(ring, CP_REG(REG_A2XX_RB_COPY_DEST_OFFSET));
 167         OUT_RING(ring, A2XX_RB_COPY_DEST_OFFSET_X(xoff) |
 168                         A2XX_RB_COPY_DEST_OFFSET_Y(yoff));
 169
 170         if (ctx->resolve & (FD_BUFFER_DEPTH | FD_BUFFER_STENCIL))
 171                 emit_gmem2mem_surf(ring, 0, bin_w * bin_h, pfb->zsbuf);
 172
 173         if (ctx->resolve & FD_BUFFER_COLOR)
 174                 emit_gmem2mem_surf(ring, 1, 0, pfb->cbufs[0]);
 175
 176         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 177         OUT_RING(ring, CP_REG(REG_A2XX_RB_MODECONTROL));
 178         OUT_RING(ring, A2XX_RB_MODECONTROL_EDRAM_MODE(COLOR_DEPTH));
 179 }
 180
 181 /* transfer from system memory to gmem */
 182
 183 static void
 184 emit_mem2gmem_surf(struct fd_ringbuffer *ring, uint32_t swap, uint32_t base,
 185                 struct pipe_surface *psurf)
 186 {
 187         struct fd_resource *rsc = fd_resource(psurf->texture);
 188         uint32_t swiz;
 189
 190         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 191         OUT_RING(ring, CP_REG(REG_A2XX_RB_COLOR_INFO));
 192         OUT_RING(ring, A2XX_RB_COLOR_INFO_SWAP(swap) |
 193                         A2XX_RB_COLOR_INFO_BASE(base) |
 194                         A2XX_RB_COLOR_INFO_FORMAT(fd_pipe2color(psurf->format)));
 195
 196         swiz = fd_tex_swiz(psurf->format, PIPE_SWIZZLE_RED, PIPE_SWIZZLE_GREEN,
 197                         PIPE_SWIZZLE_BLUE, PIPE_SWIZZLE_ALPHA);
 198
 199         /* emit fb as a texture: */
 200         OUT_PKT3(ring, CP_SET_CONSTANT, 7);
 201         OUT_RING(ring, 0x00010000);
 202         OUT_RING(ring, A2XX_SQ_TEX_0_CLAMP_X(SQ_TEX_WRAP) |
 203                         A2XX_SQ_TEX_0_CLAMP_Y(SQ_TEX_WRAP) |
 204                         A2XX_SQ_TEX_0_CLAMP_Z(SQ_TEX_WRAP) |
 205                         A2XX_SQ_TEX_0_PITCH(rsc->pitch));
 206         OUT_RELOC(ring, rsc->bo, 0,
 207                         fd_pipe2surface(psurf->format) | 0x800);
 208         OUT_RING(ring, A2XX_SQ_TEX_2_WIDTH(psurf->width - 1) |
 209                         A2XX_SQ_TEX_2_HEIGHT(psurf->height - 1));
 210         OUT_RING(ring, 0x01000000 | // XXX
 211                         swiz |
 212                         A2XX_SQ_TEX_3_XY_MAG_FILTER(SQ_TEX_FILTER_POINT) |
 213                         A2XX_SQ_TEX_3_XY_MIN_FILTER(SQ_TEX_FILTER_POINT));
 214         OUT_RING(ring, 0x00000000);
 215         OUT_RING(ring, 0x00000200);
 216
 217         OUT_PKT3(ring, CP_DRAW_INDX, 3);
 218         OUT_RING(ring, 0x00000000);
 219         OUT_RING(ring, DRAW(DI_PT_RECTLIST, DI_SRC_SEL_AUTO_INDEX,
 220                         INDEX_SIZE_IGN, IGNORE_VISIBILITY));
 221         OUT_RING(ring, 3);                                      /* NumIndices */
 222 }
 223
 224 static void
 225 emit_mem2gmem(struct fd_context *ctx, struct fd_ringbuffer *ring,
 226                 uint32_t xoff, uint32_t yoff, uint32_t bin_w, uint32_t bin_h)
 227 {
 228         struct pipe_framebuffer_state *pfb = &ctx->framebuffer;
 229         float x0, y0, x1, y1;
 230
 231         fd_emit_vertex_bufs(ring, 0x9c, (struct fd_vertex_buf[]) {
 232                         { .prsc = ctx->solid_vertexbuf, .size = 48, .offset = 0x30 },
 233                         { .prsc = ctx->solid_vertexbuf, .size = 32, .offset = 0x60 },
 234                 }, 2);
 235
 236         /* write texture coordinates to vertexbuf: */
 237         x0 = ((float)xoff) / ((float)pfb->width);
 238         x1 = ((float)xoff + bin_w) / ((float)pfb->width);
 239         y0 = ((float)yoff) / ((float)pfb->height);
 240         y1 = ((float)yoff + bin_h) / ((float)pfb->height);
 241         OUT_PKT3(ring, CP_MEM_WRITE, 9);
 242         OUT_RELOC(ring, fd_resource(ctx->solid_vertexbuf)->bo, 0x60, 0);
 243         OUT_RING(ring, fui(x0));
 244         OUT_RING(ring, fui(y0));
 245         OUT_RING(ring, fui(x1));
 246         OUT_RING(ring, fui(y0));
 247         OUT_RING(ring, fui(x0));
 248         OUT_RING(ring, fui(y1));
 249         OUT_RING(ring, fui(x1));
 250         OUT_RING(ring, fui(y1));
 251
 252         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 253         OUT_RING(ring, CP_REG(REG_A2XX_VGT_INDX_OFFSET));
 254         OUT_RING(ring, 0);
 255
 256         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 257         OUT_RING(ring, CP_REG(REG_A2XX_VGT_VERTEX_REUSE_BLOCK_CNTL));
 258         OUT_RING(ring, 0x0000003b);
 259
 260         fd_program_emit(ring, &ctx->blit_prog);
 261
 262         OUT_PKT0(ring, REG_A2XX_TC_CNTL_STATUS, 1);
 263         OUT_RING(ring, A2XX_TC_CNTL_STATUS_L2_INVALIDATE);
 264
 265         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 266         OUT_RING(ring, CP_REG(REG_A2XX_RB_DEPTHCONTROL));
 267         OUT_RING(ring, A2XX_RB_DEPTHCONTROL_EARLY_Z_ENABLE);
 268
 269         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 270         OUT_RING(ring, CP_REG(REG_A2XX_PA_SU_SC_MODE_CNTL));
 271         OUT_RING(ring, A2XX_PA_SU_SC_MODE_CNTL_PROVOKING_VTX_LAST |
 272                         A2XX_PA_SU_SC_MODE_CNTL_FRONT_PTYPE(PC_DRAW_TRIANGLES) |
 273                         A2XX_PA_SU_SC_MODE_CNTL_BACK_PTYPE(PC_DRAW_TRIANGLES));
 274
 275         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 276         OUT_RING(ring, CP_REG(REG_A2XX_PA_SC_AA_MASK));
 277         OUT_RING(ring, 0x0000ffff);
 278
 279         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 280         OUT_RING(ring, CP_REG(REG_A2XX_RB_COLORCONTROL));
 281         OUT_RING(ring, A2XX_RB_COLORCONTROL_ALPHA_FUNC(PIPE_FUNC_ALWAYS) |
 282                         A2XX_RB_COLORCONTROL_BLEND_DISABLE |
 283                         A2XX_RB_COLORCONTROL_ROP_CODE(12) |
 284                         A2XX_RB_COLORCONTROL_DITHER_MODE(DITHER_DISABLE) |
 285                         A2XX_RB_COLORCONTROL_DITHER_TYPE(DITHER_PIXEL));
 286
 287         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 288         OUT_RING(ring, CP_REG(REG_A2XX_RB_BLEND_CONTROL));
 289         OUT_RING(ring, A2XX_RB_BLEND_CONTROL_COLOR_SRCBLEND(FACTOR_ONE) |
 290                         A2XX_RB_BLEND_CONTROL_COLOR_COMB_FCN(BLEND_DST_PLUS_SRC) |
 291                         A2XX_RB_BLEND_CONTROL_COLOR_DESTBLEND(FACTOR_ZERO) |
 292                         A2XX_RB_BLEND_CONTROL_ALPHA_SRCBLEND(FACTOR_ONE) |
 293                         A2XX_RB_BLEND_CONTROL_ALPHA_COMB_FCN(BLEND_DST_PLUS_SRC) |
 294                         A2XX_RB_BLEND_CONTROL_ALPHA_DESTBLEND(FACTOR_ZERO));
 295
 296         OUT_PKT3(ring, CP_SET_CONSTANT, 3);
 297         OUT_RING(ring, CP_REG(REG_A2XX_PA_SC_WINDOW_SCISSOR_TL));
 298         OUT_RING(ring, A2XX_PA_SC_WINDOW_OFFSET_DISABLE |
 299                         xy2d(0,0));                     /* PA_SC_WINDOW_SCISSOR_TL */
 300         OUT_RING(ring, xy2d(bin_w, bin_h));     /* PA_SC_WINDOW_SCISSOR_BR */
 301
 302         OUT_PKT3(ring, CP_SET_CONSTANT, 5);
 303         OUT_RING(ring, CP_REG(REG_A2XX_PA_CL_VPORT_XSCALE));
 304         OUT_RING(ring, fui((float)bin_w/2.0));  /* PA_CL_VPORT_XSCALE */
 305         OUT_RING(ring, fui((float)bin_w/2.0));  /* PA_CL_VPORT_XOFFSET */
 306         OUT_RING(ring, fui(-(float)bin_h/2.0)); /* PA_CL_VPORT_YSCALE */
 307         OUT_RING(ring, fui((float)bin_h/2.0));  /* PA_CL_VPORT_YOFFSET */
 308
 309         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 310         OUT_RING(ring, CP_REG(REG_A2XX_PA_CL_VTE_CNTL));
 311         OUT_RING(ring, A2XX_PA_CL_VTE_CNTL_VTX_XY_FMT |
 312                         A2XX_PA_CL_VTE_CNTL_VTX_Z_FMT |       // XXX check this???
 313                         A2XX_PA_CL_VTE_CNTL_VPORT_X_SCALE_ENA |
 314                         A2XX_PA_CL_VTE_CNTL_VPORT_X_OFFSET_ENA |
 315                         A2XX_PA_CL_VTE_CNTL_VPORT_Y_SCALE_ENA |
 316                         A2XX_PA_CL_VTE_CNTL_VPORT_Y_OFFSET_ENA);
 317
 318         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 319         OUT_RING(ring, CP_REG(REG_A2XX_PA_CL_CLIP_CNTL));
 320         OUT_RING(ring, 0x00000000);
 321
 322         if (ctx->restore & (FD_BUFFER_DEPTH | FD_BUFFER_STENCIL))
 323                 emit_mem2gmem_surf(ring, 0, bin_w * bin_h, pfb->zsbuf);
 324
 325         if (ctx->restore & FD_BUFFER_COLOR)
 326                 emit_mem2gmem_surf(ring, 1, 0, pfb->cbufs[0]);
 327
 328         /* TODO blob driver seems to toss in a CACHE_FLUSH after each DRAW_INDX.. */
 329 }
 330
 331 static void
 332 calculate_tiles(struct fd_context *ctx)
 333 {
 334         struct fd_gmem_stateobj *gmem = &ctx->gmem;
 335         struct pipe_scissor_state *scissor = &ctx->max_scissor;
 336         uint32_t cpp = util_format_get_blocksize(ctx->framebuffer.cbufs[0]->format);
 337         uint32_t gmem_size = ctx->screen->gmemsize_bytes;
 338         uint32_t minx, miny, width, height;
 339         uint32_t nbins_x = 1, nbins_y = 1;
 340         uint32_t bin_w, bin_h;
 341         uint32_t max_width = 992;
 342
 343         if ((gmem->cpp == cpp) &&
 344                         !memcmp(&gmem->scissor, scissor, sizeof(gmem->scissor))) {
 345                 /* everything is up-to-date */
 346                 return;
 347         }
 348
 349         minx = scissor->minx & ~31; /* round down to multiple of 32 */
 350         miny = scissor->miny & ~31;
 351         width = scissor->maxx - minx;
 352         height = scissor->maxy - miny;
 353
 354 // TODO we probably could optimize this a bit if we know that
 355 // Z or stencil is not enabled for any of the draw calls..
 356 //      if (fd_stencil_enabled(ctx->zsa) || fd_depth_enabled(ctx->zsa)) {
 357                 gmem_size /= 2;
 358                 max_width = 256;
 359 //      }
 360
 361         bin_w = ALIGN(width, 32);
 362         bin_h = ALIGN(height, 32);
 363
 364         /* first, find a bin width that satisfies the maximum width
 365          * restrictions:
 366          */
 367         while (bin_w > max_width) {
 368                 nbins_x++;
 369                 bin_w = ALIGN(width / nbins_x, 32);
 370         }
 371
 372         /* then find a bin height that satisfies the memory constraints:
 373          */
 374         while ((bin_w * bin_h * cpp) > gmem_size) {
 375                 nbins_y++;
 376                 bin_h = ALIGN(height / nbins_y, 32);
 377         }
 378
 379         DBG("using %d bins of size %dx%d", nbins_x*nbins_y, bin_w, bin_h);
 380
 381         gmem->scissor = *scissor;
 382         gmem->cpp = cpp;
 383         gmem->minx = minx;
 384         gmem->miny = miny;
 385         gmem->bin_h = bin_h;
 386         gmem->bin_w = bin_w;
 387         gmem->nbins_x = nbins_x;
 388         gmem->nbins_y = nbins_y;
 389         gmem->width = width;
 390         gmem->height = height;
 391 }
 392
 393 void
 394 fd_gmem_render_tiles(struct pipe_context *pctx)
 395 {
 396         struct fd_context *ctx = fd_context(pctx);
 397         struct pipe_framebuffer_state *pfb = &ctx->framebuffer;
 398         struct fd_gmem_stateobj *gmem = &ctx->gmem;
 399         struct fd_ringbuffer *ring = ctx->ring;
 400         enum a2xx_colorformatx colorformatx = fd_pipe2color(pfb->cbufs[0]->format);
 401         uint32_t i, timestamp, yoff = 0;
 402         uint32_t reg;
 403
 404         calculate_tiles(ctx);
 405
 406         DBG("rendering %dx%d tiles (%s/%s)", gmem->nbins_x, gmem->nbins_y,
 407                         util_format_name(pfb->cbufs[0]->format),
 408                         pfb->zsbuf ? util_format_name(pfb->zsbuf->format) : "none");
 409
 410         /* mark the end of the clear/draw cmds before emitting per-tile cmds: */
 411         fd_ringmarker_mark(ctx->draw_end);
 412
 413         /* RB_SURFACE_INFO / RB_DEPTH_INFO can be emitted once per tile pass,
 414          * but RB_COLOR_INFO gets overwritten by gmem2mem and mem2gmem and so
 415          * needs to be emitted for each tile:
 416          */
 417         OUT_PKT3(ring, CP_SET_CONSTANT, 4);
 418         OUT_RING(ring, CP_REG(REG_A2XX_RB_SURFACE_INFO));
 419         OUT_RING(ring, gmem->bin_w);                 /* RB_SURFACE_INFO */
 420         OUT_RING(ring, A2XX_RB_COLOR_INFO_SWAP(1) | /* RB_COLOR_INFO */
 421                         A2XX_RB_COLOR_INFO_FORMAT(colorformatx));
 422         reg = A2XX_RB_DEPTH_INFO_DEPTH_BASE(ALIGN(gmem->bin_w * gmem->bin_h, 4));
 423         if (pfb->zsbuf)
 424                 reg |= A2XX_RB_DEPTH_INFO_DEPTH_FORMAT(fd_pipe2depth(pfb->zsbuf->format));
 425         OUT_RING(ring, reg);                         /* RB_DEPTH_INFO */
 426
 427         yoff= gmem->miny;
 428         for (i = 0; i < gmem->nbins_y; i++) {
 429                 uint32_t j, xoff = gmem->minx;
 430                 uint32_t bh = gmem->bin_h;
 431
 432                 /* clip bin height: */
 433                 bh = min(bh, gmem->height - yoff);
 434
 435                 for (j = 0; j < gmem->nbins_x; j++) {
 436                         uint32_t bw = gmem->bin_w;
 437
 438                         /* clip bin width: */
 439                         bw = min(bw, gmem->width - xoff);
 440
 441                         DBG("bin_h=%d, yoff=%d, bin_w=%d, xoff=%d",
 442                                         bh, yoff, bw, xoff);
 443
 444                         /* setup screen scissor for current tile (same for mem2gmem): */
 445                         OUT_PKT3(ring, CP_SET_CONSTANT, 3);
 446                         OUT_RING(ring, CP_REG(REG_A2XX_PA_SC_SCREEN_SCISSOR_TL));
 447                         OUT_RING(ring, xy2d(0,0));           /* PA_SC_SCREEN_SCISSOR_TL */
 448                         OUT_RING(ring, xy2d(bw, bh));        /* PA_SC_SCREEN_SCISSOR_BR */
 449
 450                         if (ctx->restore)
 451                                 emit_mem2gmem(ctx, ring, xoff, yoff, bw, bh);
 452
 453                         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 454                         OUT_RING(ring, CP_REG(REG_A2XX_RB_COLOR_INFO));
 455                         OUT_RING(ring, A2XX_RB_COLOR_INFO_SWAP(1) | /* RB_COLOR_INFO */
 456                                         A2XX_RB_COLOR_INFO_FORMAT(colorformatx));
 457
 458                         /* setup window scissor and offset for current tile (different
 459                          * from mem2gmem):
 460                          */
 461                         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 462                         OUT_RING(ring, CP_REG(REG_A2XX_PA_SC_WINDOW_OFFSET));
 463                         OUT_RING(ring, A2XX_PA_SC_WINDOW_OFFSET_X(-xoff) |
 464                                         A2XX_PA_SC_WINDOW_OFFSET_Y(-yoff));/* PA_SC_WINDOW_OFFSET */
 465
 466                         /* emit IB to drawcmds: */
 467                         OUT_IB  (ring, ctx->draw_start, ctx->draw_end);
 468
 469                         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 470                         OUT_RING(ring, CP_REG(REG_A2XX_PA_SC_WINDOW_OFFSET));
 471                         OUT_RING(ring, 0x00000000);          /* PA_SC_WINDOW_OFFSET */
 472
 473                         /* emit gmem2mem to transfer tile back to system memory: */
 474                         emit_gmem2mem(ctx, ring, xoff, yoff, bw, bh);
 475
 476                         xoff += bw;
 477                 }
 478
 479                 yoff += bh;
 480         }
 481
 482         /* GPU executes starting from tile cmds, which IB back to draw cmds: */
 483         fd_ringmarker_flush(ctx->draw_end);
 484
 485         /* mark start for next draw cmds: */
 486         fd_ringmarker_mark(ctx->draw_start);
 487
 488         /* update timestamps on render targets: */
 489         fd_pipe_timestamp(ctx->screen->pipe, &timestamp);
 490         fd_resource(pfb->cbufs[0]->texture)->timestamp = timestamp;
 491         if (pfb->zsbuf)
 492                 fd_resource(pfb->zsbuf->texture)->timestamp = timestamp;
 493
 494         /* reset maximal bounds: */
 495         ctx->max_scissor.minx = ctx->max_scissor.miny = ~0;
 496         ctx->max_scissor.maxx = ctx->max_scissor.maxy = 0;
 497
 498         /* Note that because the per-tile setup and mem2gmem/gmem2mem are emitted
 499          * after the draw/clear calls, but executed before, we need to preemptively
 500          * flag some state as dirty before the first draw/clear call.
 501          *
 502          * TODO maybe we need to mark all state as dirty to not worry about state
 503          * being clobbered by other contexts?
 504          */
 505         ctx->dirty |= FD_DIRTY_ZSA |
 506                         FD_DIRTY_RASTERIZER |
 507                         FD_DIRTY_FRAMEBUFFER |
 508                         FD_DIRTY_SAMPLE_MASK |
 509                         FD_DIRTY_VIEWPORT |
 510                         FD_DIRTY_CONSTBUF |
 511                         FD_DIRTY_PROG |
 512                         FD_DIRTY_SCISSOR |
 513                         /* probably only needed if we need to mem2gmem on the next
 514                          * draw..  but not sure if there is a good way to know?
 515                          */
 516                         FD_DIRTY_VERTTEX |
 517                         FD_DIRTY_FRAGTEX |
 518                         FD_DIRTY_BLEND;
 519 }