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