src/gallium/drivers/freedreno/freedreno_context.c

   1 /*
   2  * Copyright (C) 2012 Rob Clark <robclark@freedesktop.org>
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  20  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  21  * SOFTWARE.
  22  *
  23  * Authors:
  24  *    Rob Clark <robclark@freedesktop.org>
  25  */
  26
  27 #include "freedreno_context.h"
  28 #include "freedreno_blitter.h"
  29 #include "freedreno_draw.h"
  30 #include "freedreno_fence.h"
  31 #include "freedreno_program.h"
  32 #include "freedreno_resource.h"
  33 #include "freedreno_texture.h"
  34 #include "freedreno_state.h"
  35 #include "freedreno_gmem.h"
  36 #include "freedreno_query.h"
  37 #include "freedreno_query_hw.h"
  38 #include "freedreno_util.h"
  39 #include "util/u_upload_mgr.h"
  40
  41 static void
  42 fd_context_flush(struct pipe_context *pctx, struct pipe_fence_handle **fencep,
  43                 unsigned flags)
  44 {
  45         struct fd_context *ctx = fd_context(pctx);
  46         struct pipe_fence_handle *fence = NULL;
  47         // TODO we want to lookup batch if it exists, but not create one if not.
  48         struct fd_batch *batch = fd_context_batch(ctx);
  49
  50         DBG("%p: flush: flags=%x\n", ctx->batch, flags);
  51
  52         /* if no rendering since last flush, ie. app just decided it needed
  53          * a fence, re-use the last one:
  54          */
  55         if (ctx->last_fence) {
  56                 fd_fence_ref(&fence, ctx->last_fence);
  57                 goto out;
  58         }
  59
  60         if (!batch)
  61                 return;
  62
  63         /* Take a ref to the batch's fence (batch can be unref'd when flushed: */
  64         fd_fence_ref(&fence, batch->fence);
  65
  66         /* TODO is it worth trying to figure out if app is using fence-fd's, to
  67          * avoid requesting one every batch?
  68          */
  69         batch->needs_out_fence_fd = true;
  70
  71         if (!ctx->screen->reorder) {
  72                 fd_batch_flush(batch, true);
  73         } else if (flags & PIPE_FLUSH_DEFERRED) {
  74                 fd_bc_flush_deferred(&ctx->screen->batch_cache, ctx);
  75         } else {
  76                 fd_bc_flush(&ctx->screen->batch_cache, ctx);
  77         }
  78
  79 out:
  80         if (fencep)
  81                 fd_fence_ref(fencep, fence);
  82
  83         fd_fence_ref(&ctx->last_fence, fence);
  84
  85         fd_fence_ref(&fence, NULL);
  86 }
  87
  88 static void
  89 fd_texture_barrier(struct pipe_context *pctx, unsigned flags)
  90 {
  91         if (flags == PIPE_TEXTURE_BARRIER_FRAMEBUFFER) {
  92                 struct fd_context *ctx = fd_context(pctx);
  93
  94                 if (ctx->framebuffer_barrier) {
  95                         ctx->framebuffer_barrier(ctx);
  96                         return;
  97                 }
  98         }
  99
 100         /* On devices that could sample from GMEM we could possibly do better.
 101          * Or if we knew that we were doing GMEM bypass we could just emit a
 102          * cache flush, perhaps?  But we don't know if future draws would cause
 103          * us to use GMEM, and a flush in bypass isn't the end of the world.
 104          */
 105         fd_context_flush(pctx, NULL, 0);
 106 }
 107
 108 static void
 109 fd_memory_barrier(struct pipe_context *pctx, unsigned flags)
 110 {
 111         if (!(flags & ~PIPE_BARRIER_UPDATE))
 112                 return;
 113
 114         fd_context_flush(pctx, NULL, 0);
 115         /* TODO do we need to check for persistently mapped buffers and fd_bo_cpu_prep()?? */
 116 }
 117
 118 /**
 119  * emit marker string as payload of a no-op packet, which can be
 120  * decoded by cffdump.
 121  */
 122 static void
 123 fd_emit_string_marker(struct pipe_context *pctx, const char *string, int len)
 124 {
 125         struct fd_context *ctx = fd_context(pctx);
 126         struct fd_ringbuffer *ring;
 127         const uint32_t *buf = (const void *)string;
 128
 129         if (!ctx->batch)
 130                 return;
 131
 132         ctx->batch->needs_flush = true;
 133
 134         ring = ctx->batch->draw;
 135
 136         /* max packet size is 0x3fff dwords: */
 137         len = MIN2(len, 0x3fff * 4);
 138
 139         if (ctx->screen->gpu_id >= 500)
 140                 OUT_PKT7(ring, CP_NOP, align(len, 4) / 4);
 141         else
 142                 OUT_PKT3(ring, CP_NOP, align(len, 4) / 4);
 143         while (len >= 4) {
 144                 OUT_RING(ring, *buf);
 145                 buf++;
 146                 len -= 4;
 147         }
 148
 149         /* copy remainder bytes without reading past end of input string: */
 150         if (len > 0) {
 151                 uint32_t w = 0;
 152                 memcpy(&w, buf, len);
 153                 OUT_RING(ring, w);
 154         }
 155 }
 156
 157 void
 158 fd_context_destroy(struct pipe_context *pctx)
 159 {
 160         struct fd_context *ctx = fd_context(pctx);
 161         unsigned i;
 162
 163         DBG("");
 164
 165         fd_fence_ref(&ctx->last_fence, NULL);
 166
 167         if (ctx->screen->reorder && util_queue_is_initialized(&ctx->flush_queue))
 168                 util_queue_destroy(&ctx->flush_queue);
 169
 170         util_copy_framebuffer_state(&ctx->framebuffer, NULL);
 171         fd_batch_reference(&ctx->batch, NULL);  /* unref current batch */
 172         fd_bc_invalidate_context(ctx);
 173
 174         fd_prog_fini(pctx);
 175
 176         if (ctx->blitter)
 177                 util_blitter_destroy(ctx->blitter);
 178
 179         if (pctx->stream_uploader)
 180                 u_upload_destroy(pctx->stream_uploader);
 181
 182         if (ctx->clear_rs_state)
 183                 pctx->delete_rasterizer_state(pctx, ctx->clear_rs_state);
 184
 185         if (ctx->primconvert)
 186                 util_primconvert_destroy(ctx->primconvert);
 187
 188         slab_destroy_child(&ctx->transfer_pool);
 189
 190         for (i = 0; i < ARRAY_SIZE(ctx->vsc_pipe); i++) {
 191                 struct fd_vsc_pipe *pipe = &ctx->vsc_pipe[i];
 192                 if (!pipe->bo)
 193                         break;
 194                 fd_bo_del(pipe->bo);
 195         }
 196
 197         fd_device_del(ctx->dev);
 198         fd_pipe_del(ctx->pipe);
 199
 200         if (fd_mesa_debug & (FD_DBG_BSTAT | FD_DBG_MSGS)) {
 201                 printf("batch_total=%u, batch_sysmem=%u, batch_gmem=%u, batch_nondraw=%u, batch_restore=%u\n",
 202                         (uint32_t)ctx->stats.batch_total, (uint32_t)ctx->stats.batch_sysmem,
 203                         (uint32_t)ctx->stats.batch_gmem, (uint32_t)ctx->stats.batch_nondraw,
 204                         (uint32_t)ctx->stats.batch_restore);
 205         }
 206 }
 207
 208 static void
 209 fd_set_debug_callback(struct pipe_context *pctx,
 210                 const struct pipe_debug_callback *cb)
 211 {
 212         struct fd_context *ctx = fd_context(pctx);
 213
 214         if (cb)
 215                 ctx->debug = *cb;
 216         else
 217                 memset(&ctx->debug, 0, sizeof(ctx->debug));
 218 }
 219
 220 static uint32_t
 221 fd_get_reset_count(struct fd_context *ctx, bool per_context)
 222 {
 223         uint64_t val;
 224         enum fd_param_id param =
 225                 per_context ? FD_CTX_FAULTS : FD_GLOBAL_FAULTS;
 226         int ret = fd_pipe_get_param(ctx->pipe, param, &val);
 227         debug_assert(!ret);
 228         return val;
 229 }
 230
 231 static enum pipe_reset_status
 232 fd_get_device_reset_status(struct pipe_context *pctx)
 233 {
 234         struct fd_context *ctx = fd_context(pctx);
 235         int context_faults = fd_get_reset_count(ctx, true);
 236         int global_faults  = fd_get_reset_count(ctx, false);
 237         enum pipe_reset_status status;
 238
 239         if (context_faults != ctx->context_reset_count) {
 240                 status = PIPE_GUILTY_CONTEXT_RESET;
 241         } else if (global_faults != ctx->global_reset_count) {
 242                 status = PIPE_INNOCENT_CONTEXT_RESET;
 243         } else {
 244                 status = PIPE_NO_RESET;
 245         }
 246
 247         ctx->context_reset_count = context_faults;
 248         ctx->global_reset_count = global_faults;
 249
 250         return status;
 251 }
 252
 253 /* TODO we could combine a few of these small buffers (solid_vbuf,
 254  * blit_texcoord_vbuf, and vsc_size_mem, into a single buffer and
 255  * save a tiny bit of memory
 256  */
 257
 258 static struct pipe_resource *
 259 create_solid_vertexbuf(struct pipe_context *pctx)
 260 {
 261         static const float init_shader_const[] = {
 262                         -1.000000, +1.000000, +1.000000,
 263                         +1.000000, -1.000000, +1.000000,
 264         };
 265         struct pipe_resource *prsc = pipe_buffer_create(pctx->screen,
 266                         PIPE_BIND_CUSTOM, PIPE_USAGE_IMMUTABLE, sizeof(init_shader_const));
 267         pipe_buffer_write(pctx, prsc, 0,
 268                         sizeof(init_shader_const), init_shader_const);
 269         return prsc;
 270 }
 271
 272 static struct pipe_resource *
 273 create_blit_texcoord_vertexbuf(struct pipe_context *pctx)
 274 {
 275         struct pipe_resource *prsc = pipe_buffer_create(pctx->screen,
 276                         PIPE_BIND_CUSTOM, PIPE_USAGE_DYNAMIC, 16);
 277         return prsc;
 278 }
 279
 280 void
 281 fd_context_setup_common_vbos(struct fd_context *ctx)
 282 {
 283         struct pipe_context *pctx = &ctx->base;
 284
 285         ctx->solid_vbuf = create_solid_vertexbuf(pctx);
 286         ctx->blit_texcoord_vbuf = create_blit_texcoord_vertexbuf(pctx);
 287
 288         /* setup solid_vbuf_state: */
 289         ctx->solid_vbuf_state.vtx = pctx->create_vertex_elements_state(
 290                         pctx, 1, (struct pipe_vertex_element[]){{
 291                                 .vertex_buffer_index = 0,
 292                                 .src_offset = 0,
 293                                 .src_format = PIPE_FORMAT_R32G32B32_FLOAT,
 294                         }});
 295         ctx->solid_vbuf_state.vertexbuf.count = 1;
 296         ctx->solid_vbuf_state.vertexbuf.vb[0].stride = 12;
 297         ctx->solid_vbuf_state.vertexbuf.vb[0].buffer.resource = ctx->solid_vbuf;
 298
 299         /* setup blit_vbuf_state: */
 300         ctx->blit_vbuf_state.vtx = pctx->create_vertex_elements_state(
 301                         pctx, 2, (struct pipe_vertex_element[]){{
 302                                 .vertex_buffer_index = 0,
 303                                 .src_offset = 0,
 304                                 .src_format = PIPE_FORMAT_R32G32_FLOAT,
 305                         }, {
 306                                 .vertex_buffer_index = 1,
 307                                 .src_offset = 0,
 308                                 .src_format = PIPE_FORMAT_R32G32B32_FLOAT,
 309                         }});
 310         ctx->blit_vbuf_state.vertexbuf.count = 2;
 311         ctx->blit_vbuf_state.vertexbuf.vb[0].stride = 8;
 312         ctx->blit_vbuf_state.vertexbuf.vb[0].buffer.resource = ctx->blit_texcoord_vbuf;
 313         ctx->blit_vbuf_state.vertexbuf.vb[1].stride = 12;
 314         ctx->blit_vbuf_state.vertexbuf.vb[1].buffer.resource = ctx->solid_vbuf;
 315 }
 316
 317 void
 318 fd_context_cleanup_common_vbos(struct fd_context *ctx)
 319 {
 320         struct pipe_context *pctx = &ctx->base;
 321
 322         pctx->delete_vertex_elements_state(pctx, ctx->solid_vbuf_state.vtx);
 323         pctx->delete_vertex_elements_state(pctx, ctx->blit_vbuf_state.vtx);
 324
 325         pipe_resource_reference(&ctx->solid_vbuf, NULL);
 326         pipe_resource_reference(&ctx->blit_texcoord_vbuf, NULL);
 327 }
 328
 329 struct pipe_context *
 330 fd_context_init(struct fd_context *ctx, struct pipe_screen *pscreen,
 331                 const uint8_t *primtypes, void *priv, unsigned flags)
 332 {
 333         struct fd_screen *screen = fd_screen(pscreen);
 334         struct pipe_context *pctx;
 335         unsigned prio = 1;
 336         int i;
 337
 338         /* lower numerical value == higher priority: */
 339         if (fd_mesa_debug & FD_DBG_HIPRIO)
 340                 prio = 0;
 341         else if (flags & PIPE_CONTEXT_HIGH_PRIORITY)
 342                 prio = 0;
 343         else if (flags & PIPE_CONTEXT_LOW_PRIORITY)
 344                 prio = 2;
 345
 346         ctx->screen = screen;
 347         ctx->pipe = fd_pipe_new2(screen->dev, FD_PIPE_3D, prio);
 348
 349         if (fd_device_version(screen->dev) >= FD_VERSION_ROBUSTNESS) {
 350                 ctx->context_reset_count = fd_get_reset_count(ctx, true);
 351                 ctx->global_reset_count = fd_get_reset_count(ctx, false);
 352         }
 353
 354         ctx->primtypes = primtypes;
 355         ctx->primtype_mask = 0;
 356         for (i = 0; i < PIPE_PRIM_MAX; i++)
 357                 if (primtypes[i])
 358                         ctx->primtype_mask |= (1 << i);
 359
 360         /* need some sane default in case state tracker doesn't
 361          * set some state:
 362          */
 363         ctx->sample_mask = 0xffff;
 364
 365         pctx = &ctx->base;
 366         pctx->screen = pscreen;
 367         pctx->priv = priv;
 368         pctx->flush = fd_context_flush;
 369         pctx->emit_string_marker = fd_emit_string_marker;
 370         pctx->set_debug_callback = fd_set_debug_callback;
 371         pctx->get_device_reset_status = fd_get_device_reset_status;
 372         pctx->create_fence_fd = fd_create_fence_fd;
 373         pctx->fence_server_sync = fd_fence_server_sync;
 374         pctx->texture_barrier = fd_texture_barrier;
 375         pctx->memory_barrier = fd_memory_barrier;
 376
 377         pctx->stream_uploader = u_upload_create_default(pctx);
 378         if (!pctx->stream_uploader)
 379                 goto fail;
 380         pctx->const_uploader = pctx->stream_uploader;
 381
 382         if (!ctx->screen->reorder)
 383                 ctx->batch = fd_bc_alloc_batch(&screen->batch_cache, ctx, false);
 384
 385         slab_create_child(&ctx->transfer_pool, &screen->transfer_pool);
 386
 387         fd_draw_init(pctx);
 388         fd_resource_context_init(pctx);
 389         fd_query_context_init(pctx);
 390         fd_texture_init(pctx);
 391         fd_state_init(pctx);
 392
 393         ctx->blitter = util_blitter_create(pctx);
 394         if (!ctx->blitter)
 395                 goto fail;
 396
 397         ctx->primconvert = util_primconvert_create(pctx, ctx->primtype_mask);
 398         if (!ctx->primconvert)
 399                 goto fail;
 400
 401         list_inithead(&ctx->hw_active_queries);
 402         list_inithead(&ctx->acc_active_queries);
 403
 404         return pctx;
 405
 406 fail:
 407         pctx->destroy(pctx);
 408         return NULL;
 409 }