src/gallium/drivers/freedreno/freedreno_draw.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_prim.h"
  33 #include "util/u_format.h"
  34
  35 #include "freedreno_draw.h"
  36 #include "freedreno_context.h"
  37 #include "freedreno_state.h"
  38 #include "freedreno_resource.h"
  39 #include "freedreno_query_hw.h"
  40 #include "freedreno_util.h"
  41
  42 static void
  43 resource_used(struct fd_context *ctx, struct pipe_resource *prsc,
  44                 enum fd_resource_status status)
  45 {
  46         struct fd_resource *rsc;
  47
  48         if (!prsc)
  49                 return;
  50
  51         rsc = fd_resource(prsc);
  52         rsc->status |= status;
  53         if (rsc->stencil)
  54                 rsc->stencil->status |= status;
  55
  56         /* TODO resources can actually be shared across contexts,
  57          * so I'm not sure a single list-head will do the trick?
  58          */
  59         debug_assert((rsc->pending_ctx == ctx) || !rsc->pending_ctx);
  60         list_delinit(&rsc->list);
  61         list_addtail(&rsc->list, &ctx->used_resources);
  62         rsc->pending_ctx = ctx;
  63 }
  64
  65 static void
  66 resource_read(struct fd_context *ctx, struct pipe_resource *prsc)
  67 {
  68         resource_used(ctx, prsc, FD_PENDING_READ);
  69 }
  70
  71 static void
  72 resource_written(struct fd_context *ctx, struct pipe_resource *prsc)
  73 {
  74         resource_used(ctx, prsc, FD_PENDING_WRITE);
  75 }
  76
  77 static void
  78 fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
  79 {
  80         struct fd_context *ctx = fd_context(pctx);
  81         struct pipe_framebuffer_state *pfb = &ctx->framebuffer;
  82         struct pipe_scissor_state *scissor = fd_context_get_scissor(ctx);
  83         unsigned i, prims, buffers = 0;
  84
  85         /* if we supported transform feedback, we'd have to disable this: */
  86         if (((scissor->maxx - scissor->minx) *
  87                         (scissor->maxy - scissor->miny)) == 0) {
  88                 return;
  89         }
  90
  91         /* TODO: push down the region versions into the tiles */
  92         if (!fd_render_condition_check(pctx))
  93                 return;
  94
  95         /* emulate unsupported primitives: */
  96         if (!fd_supported_prim(ctx, info->mode)) {
  97                 if (ctx->streamout.num_targets > 0)
  98                         debug_error("stream-out with emulated prims");
  99                 util_primconvert_save_index_buffer(ctx->primconvert, &ctx->indexbuf);
 100                 util_primconvert_save_rasterizer_state(ctx->primconvert, ctx->rasterizer);
 101                 util_primconvert_draw_vbo(ctx->primconvert, info);
 102                 return;
 103         }
 104
 105         ctx->needs_flush = true;
 106
 107         /*
 108          * Figure out the buffers/features we need:
 109          */
 110
 111         if (fd_depth_enabled(ctx)) {
 112                 buffers |= FD_BUFFER_DEPTH;
 113                 resource_written(ctx, pfb->zsbuf->texture);
 114                 ctx->gmem_reason |= FD_GMEM_DEPTH_ENABLED;
 115         }
 116
 117         if (fd_stencil_enabled(ctx)) {
 118                 buffers |= FD_BUFFER_STENCIL;
 119                 resource_written(ctx, pfb->zsbuf->texture);
 120                 ctx->gmem_reason |= FD_GMEM_STENCIL_ENABLED;
 121         }
 122
 123         if (fd_logicop_enabled(ctx))
 124                 ctx->gmem_reason |= FD_GMEM_LOGICOP_ENABLED;
 125
 126         for (i = 0; i < pfb->nr_cbufs; i++) {
 127                 struct pipe_resource *surf;
 128
 129                 if (!pfb->cbufs[i])
 130                         continue;
 131
 132                 surf = pfb->cbufs[i]->texture;
 133
 134                 resource_written(ctx, surf);
 135                 buffers |= PIPE_CLEAR_COLOR0 << i;
 136
 137                 if (surf->nr_samples > 1)
 138                         ctx->gmem_reason |= FD_GMEM_MSAA_ENABLED;
 139
 140                 if (fd_blend_enabled(ctx, i))
 141                         ctx->gmem_reason |= FD_GMEM_BLEND_ENABLED;
 142         }
 143
 144         /* Skip over buffer 0, that is sent along with the command stream */
 145         for (i = 1; i < PIPE_MAX_CONSTANT_BUFFERS; i++) {
 146                 resource_read(ctx, ctx->constbuf[PIPE_SHADER_VERTEX].cb[i].buffer);
 147                 resource_read(ctx, ctx->constbuf[PIPE_SHADER_FRAGMENT].cb[i].buffer);
 148         }
 149
 150         /* Mark VBOs as being read */
 151         for (i = 0; i < ctx->vtx.vertexbuf.count; i++) {
 152                 assert(!ctx->vtx.vertexbuf.vb[i].user_buffer);
 153                 resource_read(ctx, ctx->vtx.vertexbuf.vb[i].buffer);
 154         }
 155
 156         /* Mark index buffer as being read */
 157         resource_read(ctx, ctx->indexbuf.buffer);
 158
 159         /* Mark textures as being read */
 160         for (i = 0; i < ctx->verttex.num_textures; i++)
 161                 if (ctx->verttex.textures[i])
 162                         resource_read(ctx, ctx->verttex.textures[i]->texture);
 163         for (i = 0; i < ctx->fragtex.num_textures; i++)
 164                 if (ctx->fragtex.textures[i])
 165                         resource_read(ctx, ctx->fragtex.textures[i]->texture);
 166
 167         /* Mark streamout buffers as being written.. */
 168         for (i = 0; i < ctx->streamout.num_targets; i++)
 169                 if (ctx->streamout.targets[i])
 170                         resource_written(ctx, ctx->streamout.targets[i]->buffer);
 171
 172         ctx->num_draws++;
 173
 174         prims = u_reduced_prims_for_vertices(info->mode, info->count);
 175
 176         ctx->stats.draw_calls++;
 177
 178         /* TODO prims_emitted should be clipped when the stream-out buffer is
 179          * not large enough.  See max_tf_vtx().. probably need to move that
 180          * into common code.  Although a bit more annoying since a2xx doesn't
 181          * use ir3 so no common way to get at the pipe_stream_output_info
 182          * which is needed for this calculation.
 183          */
 184         if (ctx->streamout.num_targets > 0)
 185                 ctx->stats.prims_emitted += prims;
 186         ctx->stats.prims_generated += prims;
 187
 188         /* any buffers that haven't been cleared yet, we need to restore: */
 189         ctx->restore |= buffers & (FD_BUFFER_ALL & ~ctx->cleared);
 190         /* and any buffers used, need to be resolved: */
 191         ctx->resolve |= buffers;
 192
 193         DBG("%x num_draws=%u (%s/%s)", buffers, ctx->num_draws,
 194                 util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
 195                 util_format_short_name(pipe_surface_format(pfb->zsbuf)));
 196
 197         fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_DRAW);
 198         ctx->draw_vbo(ctx, info);
 199
 200         for (i = 0; i < ctx->streamout.num_targets; i++)
 201                 ctx->streamout.offsets[i] += info->count;
 202
 203         if (fd_mesa_debug & FD_DBG_DDRAW)
 204                 ctx->dirty = 0xffffffff;
 205
 206         /* if an app (or, well, piglit test) does many thousands of draws
 207          * without flush (or anything which implicitly flushes, like
 208          * changing render targets), we can exceed the ringbuffer size.
 209          * Since we don't currently have a sane way to wrapparound, and
 210          * we use the same buffer for both draw and tiling commands, for
 211          * now we need to do this hack and trigger flush if we are running
 212          * low on remaining space for cmds:
 213          */
 214         if (((ctx->ring->cur - ctx->ring->start) >
 215                                 (ctx->ring->size/4 - FD_TILING_COMMANDS_DWORDS)) ||
 216                         (fd_mesa_debug & FD_DBG_FLUSH))
 217                 fd_context_render(pctx);
 218 }
 219
 220 /* TODO figure out how to make better use of existing state mechanism
 221  * for clear (and possibly gmem->mem / mem->gmem) so we can (a) keep
 222  * track of what state really actually changes, and (b) reduce the code
 223  * in the a2xx/a3xx parts.
 224  */
 225
 226 static void
 227 fd_clear(struct pipe_context *pctx, unsigned buffers,
 228                 const union pipe_color_union *color, double depth, unsigned stencil)
 229 {
 230         struct fd_context *ctx = fd_context(pctx);
 231         struct pipe_framebuffer_state *pfb = &ctx->framebuffer;
 232         struct pipe_scissor_state *scissor = fd_context_get_scissor(ctx);
 233         unsigned cleared_buffers;
 234         int i;
 235
 236         /* TODO: push down the region versions into the tiles */
 237         if (!fd_render_condition_check(pctx))
 238                 return;
 239
 240         /* for bookkeeping about which buffers have been cleared (and thus
 241          * can fully or partially skip mem2gmem) we need to ignore buffers
 242          * that have already had a draw, in case apps do silly things like
 243          * clear after draw (ie. if you only clear the color buffer, but
 244          * something like alpha-test causes side effects from the draw in
 245          * the depth buffer, etc)
 246          */
 247         cleared_buffers = buffers & (FD_BUFFER_ALL & ~ctx->restore);
 248
 249         /* do we have full-screen scissor? */
 250         if (!memcmp(scissor, &ctx->disabled_scissor, sizeof(*scissor))) {
 251                 ctx->cleared |= cleared_buffers;
 252         } else {
 253                 ctx->partial_cleared |= cleared_buffers;
 254                 if (cleared_buffers & PIPE_CLEAR_COLOR)
 255                         ctx->cleared_scissor.color = *scissor;
 256                 if (cleared_buffers & PIPE_CLEAR_DEPTH)
 257                         ctx->cleared_scissor.depth = *scissor;
 258                 if (cleared_buffers & PIPE_CLEAR_STENCIL)
 259                         ctx->cleared_scissor.stencil = *scissor;
 260         }
 261         ctx->resolve |= buffers;
 262         ctx->needs_flush = true;
 263
 264         if (buffers & PIPE_CLEAR_COLOR)
 265                 for (i = 0; i < pfb->nr_cbufs; i++)
 266                         if (buffers & (PIPE_CLEAR_COLOR0 << i))
 267                                 resource_written(ctx, pfb->cbufs[i]->texture);
 268
 269         if (buffers & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL)) {
 270                 resource_written(ctx, pfb->zsbuf->texture);
 271                 ctx->gmem_reason |= FD_GMEM_CLEARS_DEPTH_STENCIL;
 272         }
 273
 274         DBG("%x depth=%f, stencil=%u (%s/%s)", buffers, depth, stencil,
 275                 util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
 276                 util_format_short_name(pipe_surface_format(pfb->zsbuf)));
 277
 278         fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_CLEAR);
 279
 280         ctx->clear(ctx, buffers, color, depth, stencil);
 281
 282         ctx->dirty |= FD_DIRTY_ZSA |
 283                         FD_DIRTY_VIEWPORT |
 284                         FD_DIRTY_RASTERIZER |
 285                         FD_DIRTY_SAMPLE_MASK |
 286                         FD_DIRTY_PROG |
 287                         FD_DIRTY_CONSTBUF |
 288                         FD_DIRTY_BLEND |
 289                         FD_DIRTY_FRAMEBUFFER;
 290
 291         if (fd_mesa_debug & FD_DBG_DCLEAR)
 292                 ctx->dirty = 0xffffffff;
 293 }
 294
 295 static void
 296 fd_clear_render_target(struct pipe_context *pctx, struct pipe_surface *ps,
 297                 const union pipe_color_union *color,
 298                 unsigned x, unsigned y, unsigned w, unsigned h)
 299 {
 300         DBG("TODO: x=%u, y=%u, w=%u, h=%u", x, y, w, h);
 301 }
 302
 303 static void
 304 fd_clear_depth_stencil(struct pipe_context *pctx, struct pipe_surface *ps,
 305                 unsigned buffers, double depth, unsigned stencil,
 306                 unsigned x, unsigned y, unsigned w, unsigned h)
 307 {
 308         DBG("TODO: buffers=%u, depth=%f, stencil=%u, x=%u, y=%u, w=%u, h=%u",
 309                         buffers, depth, stencil, x, y, w, h);
 310 }
 311
 312 void
 313 fd_draw_init(struct pipe_context *pctx)
 314 {
 315         list_inithead(&fd_context(pctx)->used_resources);
 316
 317         pctx->draw_vbo = fd_draw_vbo;
 318         pctx->clear = fd_clear;
 319         pctx->clear_render_target = fd_clear_render_target;
 320         pctx->clear_depth_stencil = fd_clear_depth_stencil;
 321 }