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 #include "util/u_helpers.h"
  35
  36 #include "freedreno_draw.h"
  37 #include "freedreno_context.h"
  38 #include "freedreno_state.h"
  39 #include "freedreno_resource.h"
  40 #include "freedreno_query_acc.h"
  41 #include "freedreno_query_hw.h"
  42 #include "freedreno_util.h"
  43
  44 static void
  45 resource_read(struct fd_batch *batch, struct pipe_resource *prsc)
  46 {
  47         if (!prsc)
  48                 return;
  49         fd_batch_resource_used(batch, fd_resource(prsc), false);
  50 }
  51
  52 static void
  53 resource_written(struct fd_batch *batch, struct pipe_resource *prsc)
  54 {
  55         if (!prsc)
  56                 return;
  57         fd_batch_resource_used(batch, fd_resource(prsc), true);
  58 }
  59
  60 static void
  61 fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
  62 {
  63         struct fd_context *ctx = fd_context(pctx);
  64         struct fd_batch *batch = ctx->batch;
  65         struct pipe_framebuffer_state *pfb = &batch->framebuffer;
  66         struct pipe_scissor_state *scissor = fd_context_get_scissor(ctx);
  67         unsigned i, prims, buffers = 0;
  68
  69         if (!info->count_from_stream_output && !info->indirect &&
  70             !info->primitive_restart &&
  71             !u_trim_pipe_prim(info->mode, (unsigned*)&info->count))
  72                 return;
  73
  74         /* if we supported transform feedback, we'd have to disable this: */
  75         if (((scissor->maxx - scissor->minx) *
  76                         (scissor->maxy - scissor->miny)) == 0) {
  77                 return;
  78         }
  79
  80         /* TODO: push down the region versions into the tiles */
  81         if (!fd_render_condition_check(pctx))
  82                 return;
  83
  84         /* emulate unsupported primitives: */
  85         if (!fd_supported_prim(ctx, info->mode)) {
  86                 if (ctx->streamout.num_targets > 0)
  87                         debug_error("stream-out with emulated prims");
  88                 util_primconvert_save_index_buffer(ctx->primconvert, &ctx->indexbuf);
  89                 util_primconvert_save_rasterizer_state(ctx->primconvert, ctx->rasterizer);
  90                 util_primconvert_draw_vbo(ctx->primconvert, info);
  91                 return;
  92         }
  93
  94         /* Upload a user index buffer. */
  95         struct pipe_index_buffer ibuffer_saved = {};
  96         if (info->indexed && ctx->indexbuf.user_buffer &&
  97             !util_save_and_upload_index_buffer(pctx, info, &ctx->indexbuf,
  98                                                &ibuffer_saved)) {
  99                 return;
 100         }
 101
 102         if (ctx->in_blit) {
 103                 fd_batch_reset(batch);
 104                 fd_context_all_dirty(ctx);
 105         }
 106
 107         batch->blit = ctx->in_blit;
 108         batch->back_blit = ctx->in_shadow;
 109
 110         /* NOTE: needs to be before resource_written(batch->query_buf), otherwise
 111          * query_buf may not be created yet.
 112          */
 113         fd_batch_set_stage(batch, FD_STAGE_DRAW);
 114
 115         /*
 116          * Figure out the buffers/features we need:
 117          */
 118
 119         mtx_lock(&ctx->screen->lock);
 120
 121         if (fd_depth_enabled(ctx)) {
 122                 buffers |= FD_BUFFER_DEPTH;
 123                 resource_written(batch, pfb->zsbuf->texture);
 124                 batch->gmem_reason |= FD_GMEM_DEPTH_ENABLED;
 125         }
 126
 127         if (fd_stencil_enabled(ctx)) {
 128                 buffers |= FD_BUFFER_STENCIL;
 129                 resource_written(batch, pfb->zsbuf->texture);
 130                 batch->gmem_reason |= FD_GMEM_STENCIL_ENABLED;
 131         }
 132
 133         if (fd_logicop_enabled(ctx))
 134                 batch->gmem_reason |= FD_GMEM_LOGICOP_ENABLED;
 135
 136         for (i = 0; i < pfb->nr_cbufs; i++) {
 137                 struct pipe_resource *surf;
 138
 139                 if (!pfb->cbufs[i])
 140                         continue;
 141
 142                 surf = pfb->cbufs[i]->texture;
 143
 144                 resource_written(batch, surf);
 145                 buffers |= PIPE_CLEAR_COLOR0 << i;
 146
 147                 if (surf->nr_samples > 1)
 148                         batch->gmem_reason |= FD_GMEM_MSAA_ENABLED;
 149
 150                 if (fd_blend_enabled(ctx, i))
 151                         batch->gmem_reason |= FD_GMEM_BLEND_ENABLED;
 152         }
 153
 154         /* Mark SSBOs as being written.. we don't actually know which ones are
 155          * read vs written, so just assume the worst
 156          */
 157         foreach_bit(i, ctx->shaderbuf[PIPE_SHADER_FRAGMENT].enabled_mask)
 158                 resource_read(batch, ctx->shaderbuf[PIPE_SHADER_FRAGMENT].sb[i].buffer);
 159
 160         foreach_bit(i, ctx->constbuf[PIPE_SHADER_VERTEX].enabled_mask)
 161                 resource_read(batch, ctx->constbuf[PIPE_SHADER_VERTEX].cb[i].buffer);
 162         foreach_bit(i, ctx->constbuf[PIPE_SHADER_FRAGMENT].enabled_mask)
 163                 resource_read(batch, ctx->constbuf[PIPE_SHADER_FRAGMENT].cb[i].buffer);
 164
 165         /* Mark VBOs as being read */
 166         foreach_bit(i, ctx->vtx.vertexbuf.enabled_mask) {
 167                 assert(!ctx->vtx.vertexbuf.vb[i].user_buffer);
 168                 resource_read(batch, ctx->vtx.vertexbuf.vb[i].buffer);
 169         }
 170
 171         /* Mark index buffer as being read */
 172         resource_read(batch, ctx->indexbuf.buffer);
 173
 174         /* Mark textures as being read */
 175         foreach_bit(i, ctx->tex[PIPE_SHADER_VERTEX].valid_textures)
 176                 resource_read(batch, ctx->tex[PIPE_SHADER_VERTEX].textures[i]->texture);
 177         foreach_bit(i, ctx->tex[PIPE_SHADER_FRAGMENT].valid_textures)
 178                 resource_read(batch, ctx->tex[PIPE_SHADER_FRAGMENT].textures[i]->texture);
 179
 180         /* Mark streamout buffers as being written.. */
 181         for (i = 0; i < ctx->streamout.num_targets; i++)
 182                 if (ctx->streamout.targets[i])
 183                         resource_written(batch, ctx->streamout.targets[i]->buffer);
 184
 185         resource_written(batch, batch->query_buf);
 186
 187         list_for_each_entry(struct fd_acc_query, aq, &ctx->acc_active_queries, node)
 188                 resource_written(batch, aq->prsc);
 189
 190         mtx_unlock(&ctx->screen->lock);
 191
 192         batch->num_draws++;
 193
 194         prims = u_reduced_prims_for_vertices(info->mode, info->count);
 195
 196         ctx->stats.draw_calls++;
 197
 198         /* TODO prims_emitted should be clipped when the stream-out buffer is
 199          * not large enough.  See max_tf_vtx().. probably need to move that
 200          * into common code.  Although a bit more annoying since a2xx doesn't
 201          * use ir3 so no common way to get at the pipe_stream_output_info
 202          * which is needed for this calculation.
 203          */
 204         if (ctx->streamout.num_targets > 0)
 205                 ctx->stats.prims_emitted += prims;
 206         ctx->stats.prims_generated += prims;
 207
 208         /* any buffers that haven't been cleared yet, we need to restore: */
 209         batch->restore |= buffers & (FD_BUFFER_ALL & ~batch->cleared);
 210         /* and any buffers used, need to be resolved: */
 211         batch->resolve |= buffers;
 212
 213         DBG("%p: %x %ux%u num_draws=%u (%s/%s)", batch, buffers,
 214                 pfb->width, pfb->height, batch->num_draws,
 215                 util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
 216                 util_format_short_name(pipe_surface_format(pfb->zsbuf)));
 217
 218         if (ctx->draw_vbo(ctx, info))
 219                 batch->needs_flush = true;
 220
 221         for (i = 0; i < ctx->streamout.num_targets; i++)
 222                 ctx->streamout.offsets[i] += info->count;
 223
 224         if (fd_mesa_debug & FD_DBG_DDRAW)
 225                 fd_context_all_dirty(ctx);
 226
 227         fd_batch_check_size(batch);
 228
 229         if (info->indexed && ibuffer_saved.user_buffer)
 230            pctx->set_index_buffer(pctx, &ibuffer_saved);
 231 }
 232
 233 /* Generic clear implementation (partially) using u_blitter: */
 234 static void
 235 fd_blitter_clear(struct pipe_context *pctx, unsigned buffers,
 236                 const union pipe_color_union *color, double depth, unsigned stencil)
 237 {
 238         struct fd_context *ctx = fd_context(pctx);
 239         struct pipe_framebuffer_state *pfb = &ctx->batch->framebuffer;
 240         struct blitter_context *blitter = ctx->blitter;
 241
 242         fd_blitter_pipe_begin(ctx, false, true, FD_STAGE_CLEAR);
 243
 244         util_blitter_common_clear_setup(blitter, pfb->width, pfb->height,
 245                         buffers, NULL, NULL);
 246
 247         struct pipe_stencil_ref sr = {
 248                 .ref_value = { stencil & 0xff }
 249         };
 250         pctx->set_stencil_ref(pctx, &sr);
 251
 252         struct pipe_constant_buffer cb = {
 253                 .buffer_size = 16,
 254                 .user_buffer = &color->ui,
 255         };
 256         pctx->set_constant_buffer(pctx, PIPE_SHADER_FRAGMENT, 0, &cb);
 257
 258         if (!ctx->clear_rs_state) {
 259                 const struct pipe_rasterizer_state tmpl = {
 260                         .cull_face = PIPE_FACE_NONE,
 261                         .half_pixel_center = 1,
 262                         .bottom_edge_rule = 1,
 263                         .flatshade = 1,
 264                         .depth_clip = 1,
 265                 };
 266                 ctx->clear_rs_state = pctx->create_rasterizer_state(pctx, &tmpl);
 267         }
 268         pctx->bind_rasterizer_state(pctx, ctx->clear_rs_state);
 269
 270         struct pipe_viewport_state vp = {
 271                 .scale     = { 0.5f * pfb->width, -0.5f * pfb->height, depth },
 272                 .translate = { 0.5f * pfb->width,  0.5f * pfb->height, 0.0f },
 273         };
 274         pctx->set_viewport_states(pctx, 0, 1, &vp);
 275
 276         pctx->bind_vertex_elements_state(pctx, ctx->solid_vbuf_state.vtx);
 277         pctx->set_vertex_buffers(pctx, blitter->vb_slot, 1,
 278                         &ctx->solid_vbuf_state.vertexbuf.vb[0]);
 279         pctx->set_stream_output_targets(pctx, 0, NULL, NULL);
 280         pctx->bind_vs_state(pctx, ctx->solid_prog.vp);
 281         pctx->bind_fs_state(pctx, ctx->solid_prog.fp);
 282
 283         struct pipe_draw_info info = {
 284                 .mode = PIPE_PRIM_MAX,    /* maps to DI_PT_RECTLIST */
 285                 .count = 2,
 286                 .max_index = 1,
 287                 .instance_count = 1,
 288         };
 289         ctx->draw_vbo(ctx, &info);
 290
 291         util_blitter_restore_constant_buffer_state(blitter);
 292         util_blitter_restore_vertex_states(blitter);
 293         util_blitter_restore_fragment_states(blitter);
 294         util_blitter_restore_textures(blitter);
 295         util_blitter_restore_fb_state(blitter);
 296         util_blitter_restore_render_cond(blitter);
 297         util_blitter_unset_running_flag(blitter);
 298
 299         fd_blitter_pipe_end(ctx);
 300 }
 301
 302 /* TODO figure out how to make better use of existing state mechanism
 303  * for clear (and possibly gmem->mem / mem->gmem) so we can (a) keep
 304  * track of what state really actually changes, and (b) reduce the code
 305  * in the a2xx/a3xx parts.
 306  */
 307
 308 static void
 309 fd_clear(struct pipe_context *pctx, unsigned buffers,
 310                 const union pipe_color_union *color, double depth, unsigned stencil)
 311 {
 312         struct fd_context *ctx = fd_context(pctx);
 313         struct fd_batch *batch = ctx->batch;
 314         struct pipe_framebuffer_state *pfb = &batch->framebuffer;
 315         struct pipe_scissor_state *scissor = fd_context_get_scissor(ctx);
 316         unsigned cleared_buffers;
 317         int i;
 318
 319         /* TODO: push down the region versions into the tiles */
 320         if (!fd_render_condition_check(pctx))
 321                 return;
 322
 323         if (ctx->in_blit) {
 324                 fd_batch_reset(batch);
 325                 fd_context_all_dirty(ctx);
 326         }
 327
 328         /* for bookkeeping about which buffers have been cleared (and thus
 329          * can fully or partially skip mem2gmem) we need to ignore buffers
 330          * that have already had a draw, in case apps do silly things like
 331          * clear after draw (ie. if you only clear the color buffer, but
 332          * something like alpha-test causes side effects from the draw in
 333          * the depth buffer, etc)
 334          */
 335         cleared_buffers = buffers & (FD_BUFFER_ALL & ~batch->restore);
 336
 337         /* do we have full-screen scissor? */
 338         if (!memcmp(scissor, &ctx->disabled_scissor, sizeof(*scissor))) {
 339                 batch->cleared |= cleared_buffers;
 340         } else {
 341                 batch->partial_cleared |= cleared_buffers;
 342                 if (cleared_buffers & PIPE_CLEAR_COLOR)
 343                         batch->cleared_scissor.color = *scissor;
 344                 if (cleared_buffers & PIPE_CLEAR_DEPTH)
 345                         batch->cleared_scissor.depth = *scissor;
 346                 if (cleared_buffers & PIPE_CLEAR_STENCIL)
 347                         batch->cleared_scissor.stencil = *scissor;
 348         }
 349         batch->resolve |= buffers;
 350         batch->needs_flush = true;
 351
 352         mtx_lock(&ctx->screen->lock);
 353
 354         if (buffers & PIPE_CLEAR_COLOR)
 355                 for (i = 0; i < pfb->nr_cbufs; i++)
 356                         if (buffers & (PIPE_CLEAR_COLOR0 << i))
 357                                 resource_written(batch, pfb->cbufs[i]->texture);
 358
 359         if (buffers & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL)) {
 360                 resource_written(batch, pfb->zsbuf->texture);
 361                 batch->gmem_reason |= FD_GMEM_CLEARS_DEPTH_STENCIL;
 362         }
 363
 364         resource_written(batch, batch->query_buf);
 365
 366         list_for_each_entry(struct fd_acc_query, aq, &ctx->acc_active_queries, node)
 367                 resource_written(batch, aq->prsc);
 368
 369         mtx_unlock(&ctx->screen->lock);
 370
 371         DBG("%p: %x %ux%u depth=%f, stencil=%u (%s/%s)", batch, buffers,
 372                 pfb->width, pfb->height, depth, stencil,
 373                 util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
 374                 util_format_short_name(pipe_surface_format(pfb->zsbuf)));
 375
 376         /* if per-gen backend doesn't implement ctx->clear() generic
 377          * blitter clear:
 378          */
 379         if (!ctx->clear) {
 380                 fd_blitter_clear(pctx, buffers, color, depth, stencil);
 381                 return;
 382         }
 383
 384         fd_batch_set_stage(batch, FD_STAGE_CLEAR);
 385
 386         ctx->clear(ctx, buffers, color, depth, stencil);
 387
 388         if (fd_mesa_debug & FD_DBG_DCLEAR)
 389                 fd_context_all_dirty(ctx);
 390 }
 391
 392 static void
 393 fd_clear_render_target(struct pipe_context *pctx, struct pipe_surface *ps,
 394                 const union pipe_color_union *color,
 395                 unsigned x, unsigned y, unsigned w, unsigned h,
 396                 bool render_condition_enabled)
 397 {
 398         DBG("TODO: x=%u, y=%u, w=%u, h=%u", x, y, w, h);
 399 }
 400
 401 static void
 402 fd_clear_depth_stencil(struct pipe_context *pctx, struct pipe_surface *ps,
 403                 unsigned buffers, double depth, unsigned stencil,
 404                 unsigned x, unsigned y, unsigned w, unsigned h,
 405                 bool render_condition_enabled)
 406 {
 407         DBG("TODO: buffers=%u, depth=%f, stencil=%u, x=%u, y=%u, w=%u, h=%u",
 408                         buffers, depth, stencil, x, y, w, h);
 409 }
 410
 411 void
 412 fd_draw_init(struct pipe_context *pctx)
 413 {
 414         pctx->draw_vbo = fd_draw_vbo;
 415         pctx->clear = fd_clear;
 416         pctx->clear_render_target = fd_clear_render_target;
 417         pctx->clear_depth_stencil = fd_clear_depth_stencil;
 418 }