src/gallium/drivers/freedreno/freedreno_context.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 "freedreno_context.h"
  30 #include "freedreno_blitter.h"
  31 #include "freedreno_draw.h"
  32 #include "freedreno_fence.h"
  33 #include "freedreno_program.h"
  34 #include "freedreno_resource.h"
  35 #include "freedreno_texture.h"
  36 #include "freedreno_state.h"
  37 #include "freedreno_gmem.h"
  38 #include "freedreno_query.h"
  39 #include "freedreno_query_hw.h"
  40 #include "freedreno_util.h"
  41 #include "util/u_upload_mgr.h"
  42
  43 static void
  44 fd_context_flush(struct pipe_context *pctx, struct pipe_fence_handle **fencep,
  45                 unsigned flags)
  46 {
  47         struct fd_context *ctx = fd_context(pctx);
  48         struct pipe_fence_handle *fence = NULL;
  49
  50         DBG("%p: flush: flags=%x\n", ctx->batch, flags);
  51
  52         /* Take a ref to the batch's fence (batch can be unref'd when flushed: */
  53         fd_fence_ref(pctx->screen, &fence, ctx->batch->fence);
  54
  55         if (flags & PIPE_FLUSH_FENCE_FD)
  56                 ctx->batch->needs_out_fence_fd = true;
  57
  58         if (!ctx->screen->reorder) {
  59                 fd_batch_flush(ctx->batch, true, false);
  60         } else if (flags & PIPE_FLUSH_DEFERRED) {
  61                 fd_bc_flush_deferred(&ctx->screen->batch_cache, ctx);
  62         } else {
  63                 fd_bc_flush(&ctx->screen->batch_cache, ctx);
  64         }
  65
  66         if (fencep)
  67                 fd_fence_ref(pctx->screen, fencep, fence);
  68
  69         fd_fence_ref(pctx->screen, &fence, NULL);
  70 }
  71
  72 static void
  73 fd_texture_barrier(struct pipe_context *pctx, unsigned flags)
  74 {
  75         /* On devices that could sample from GMEM we could possibly do better.
  76          * Or if we knew that we were doing GMEM bypass we could just emit a
  77          * cache flush, perhaps?  But we don't know if future draws would cause
  78          * us to use GMEM, and a flush in bypass isn't the end of the world.
  79          */
  80         fd_context_flush(pctx, NULL, 0);
  81 }
  82
  83 static void
  84 fd_memory_barrier(struct pipe_context *pctx, unsigned flags)
  85 {
  86         fd_context_flush(pctx, NULL, 0);
  87         /* TODO do we need to check for persistently mapped buffers and fd_bo_cpu_prep()?? */
  88 }
  89
  90 /**
  91  * emit marker string as payload of a no-op packet, which can be
  92  * decoded by cffdump.
  93  */
  94 static void
  95 fd_emit_string_marker(struct pipe_context *pctx, const char *string, int len)
  96 {
  97         struct fd_context *ctx = fd_context(pctx);
  98         struct fd_ringbuffer *ring;
  99         const uint32_t *buf = (const void *)string;
 100
 101         if (!ctx->batch)
 102                 return;
 103
 104         ctx->batch->needs_flush = true;
 105
 106         ring = ctx->batch->draw;
 107
 108         /* max packet size is 0x3fff dwords: */
 109         len = MIN2(len, 0x3fff * 4);
 110
 111         if (ctx->screen->gpu_id >= 500)
 112                 OUT_PKT7(ring, CP_NOP, align(len, 4) / 4);
 113         else
 114                 OUT_PKT3(ring, CP_NOP, align(len, 4) / 4);
 115         while (len >= 4) {
 116                 OUT_RING(ring, *buf);
 117                 buf++;
 118                 len -= 4;
 119         }
 120
 121         /* copy remainder bytes without reading past end of input string: */
 122         if (len > 0) {
 123                 uint32_t w = 0;
 124                 memcpy(&w, buf, len);
 125                 OUT_RING(ring, w);
 126         }
 127 }
 128
 129 void
 130 fd_context_destroy(struct pipe_context *pctx)
 131 {
 132         struct fd_context *ctx = fd_context(pctx);
 133         unsigned i;
 134
 135         DBG("");
 136
 137         if (ctx->screen->reorder && util_queue_is_initialized(&ctx->flush_queue))
 138                 util_queue_destroy(&ctx->flush_queue);
 139
 140         fd_batch_reference(&ctx->batch, NULL);  /* unref current batch */
 141         fd_bc_invalidate_context(ctx);
 142
 143         fd_prog_fini(pctx);
 144
 145         if (ctx->blitter)
 146                 util_blitter_destroy(ctx->blitter);
 147
 148         if (pctx->stream_uploader)
 149                 u_upload_destroy(pctx->stream_uploader);
 150
 151         if (ctx->clear_rs_state)
 152                 pctx->delete_rasterizer_state(pctx, ctx->clear_rs_state);
 153
 154         if (ctx->primconvert)
 155                 util_primconvert_destroy(ctx->primconvert);
 156
 157         slab_destroy_child(&ctx->transfer_pool);
 158
 159         for (i = 0; i < ARRAY_SIZE(ctx->vsc_pipe); i++) {
 160                 struct fd_vsc_pipe *pipe = &ctx->vsc_pipe[i];
 161                 if (!pipe->bo)
 162                         break;
 163                 fd_bo_del(pipe->bo);
 164         }
 165
 166         fd_device_del(ctx->dev);
 167         fd_pipe_del(ctx->pipe);
 168
 169         if (fd_mesa_debug & (FD_DBG_BSTAT | FD_DBG_MSGS)) {
 170                 printf("batch_total=%u, batch_sysmem=%u, batch_gmem=%u, batch_nondraw=%u, batch_restore=%u\n",
 171                         (uint32_t)ctx->stats.batch_total, (uint32_t)ctx->stats.batch_sysmem,
 172                         (uint32_t)ctx->stats.batch_gmem, (uint32_t)ctx->stats.batch_nondraw,
 173                         (uint32_t)ctx->stats.batch_restore);
 174         }
 175
 176         FREE(ctx);
 177 }
 178
 179 static void
 180 fd_set_debug_callback(struct pipe_context *pctx,
 181                 const struct pipe_debug_callback *cb)
 182 {
 183         struct fd_context *ctx = fd_context(pctx);
 184
 185         if (cb)
 186                 ctx->debug = *cb;
 187         else
 188                 memset(&ctx->debug, 0, sizeof(ctx->debug));
 189 }
 190
 191 /* TODO we could combine a few of these small buffers (solid_vbuf,
 192  * blit_texcoord_vbuf, and vsc_size_mem, into a single buffer and
 193  * save a tiny bit of memory
 194  */
 195
 196 static struct pipe_resource *
 197 create_solid_vertexbuf(struct pipe_context *pctx)
 198 {
 199         static const float init_shader_const[] = {
 200                         -1.000000, +1.000000, +1.000000,
 201                         +1.000000, -1.000000, +1.000000,
 202         };
 203         struct pipe_resource *prsc = pipe_buffer_create(pctx->screen,
 204                         PIPE_BIND_CUSTOM, PIPE_USAGE_IMMUTABLE, sizeof(init_shader_const));
 205         pipe_buffer_write(pctx, prsc, 0,
 206                         sizeof(init_shader_const), init_shader_const);
 207         return prsc;
 208 }
 209
 210 static struct pipe_resource *
 211 create_blit_texcoord_vertexbuf(struct pipe_context *pctx)
 212 {
 213         struct pipe_resource *prsc = pipe_buffer_create(pctx->screen,
 214                         PIPE_BIND_CUSTOM, PIPE_USAGE_DYNAMIC, 16);
 215         return prsc;
 216 }
 217
 218 void
 219 fd_context_setup_common_vbos(struct fd_context *ctx)
 220 {
 221         struct pipe_context *pctx = &ctx->base;
 222
 223         ctx->solid_vbuf = create_solid_vertexbuf(pctx);
 224         ctx->blit_texcoord_vbuf = create_blit_texcoord_vertexbuf(pctx);
 225
 226         /* setup solid_vbuf_state: */
 227         ctx->solid_vbuf_state.vtx = pctx->create_vertex_elements_state(
 228                         pctx, 1, (struct pipe_vertex_element[]){{
 229                                 .vertex_buffer_index = 0,
 230                                 .src_offset = 0,
 231                                 .src_format = PIPE_FORMAT_R32G32B32_FLOAT,
 232                         }});
 233         ctx->solid_vbuf_state.vertexbuf.count = 1;
 234         ctx->solid_vbuf_state.vertexbuf.vb[0].stride = 12;
 235         ctx->solid_vbuf_state.vertexbuf.vb[0].buffer.resource = ctx->solid_vbuf;
 236
 237         /* setup blit_vbuf_state: */
 238         ctx->blit_vbuf_state.vtx = pctx->create_vertex_elements_state(
 239                         pctx, 2, (struct pipe_vertex_element[]){{
 240                                 .vertex_buffer_index = 0,
 241                                 .src_offset = 0,
 242                                 .src_format = PIPE_FORMAT_R32G32_FLOAT,
 243                         }, {
 244                                 .vertex_buffer_index = 1,
 245                                 .src_offset = 0,
 246                                 .src_format = PIPE_FORMAT_R32G32B32_FLOAT,
 247                         }});
 248         ctx->blit_vbuf_state.vertexbuf.count = 2;
 249         ctx->blit_vbuf_state.vertexbuf.vb[0].stride = 8;
 250         ctx->blit_vbuf_state.vertexbuf.vb[0].buffer.resource = ctx->blit_texcoord_vbuf;
 251         ctx->blit_vbuf_state.vertexbuf.vb[1].stride = 12;
 252         ctx->blit_vbuf_state.vertexbuf.vb[1].buffer.resource = ctx->solid_vbuf;
 253 }
 254
 255 void
 256 fd_context_cleanup_common_vbos(struct fd_context *ctx)
 257 {
 258         struct pipe_context *pctx = &ctx->base;
 259
 260         pctx->delete_vertex_elements_state(pctx, ctx->solid_vbuf_state.vtx);
 261         pctx->delete_vertex_elements_state(pctx, ctx->blit_vbuf_state.vtx);
 262
 263         pipe_resource_reference(&ctx->solid_vbuf, NULL);
 264         pipe_resource_reference(&ctx->blit_texcoord_vbuf, NULL);
 265 }
 266
 267 struct pipe_context *
 268 fd_context_init(struct fd_context *ctx, struct pipe_screen *pscreen,
 269                 const uint8_t *primtypes, void *priv, unsigned flags)
 270 {
 271         struct fd_screen *screen = fd_screen(pscreen);
 272         struct pipe_context *pctx;
 273         unsigned prio = 1;
 274         int i;
 275
 276         /* lower numerical value == higher priority: */
 277         if (fd_mesa_debug & FD_DBG_HIPRIO)
 278                 prio = 0;
 279         else if (flags & PIPE_CONTEXT_HIGH_PRIORITY)
 280                 prio = 0;
 281         else if (flags & PIPE_CONTEXT_LOW_PRIORITY)
 282                 prio = 2;
 283
 284         ctx->screen = screen;
 285         ctx->pipe = fd_pipe_new2(screen->dev, FD_PIPE_3D, prio);
 286
 287         ctx->primtypes = primtypes;
 288         ctx->primtype_mask = 0;
 289         for (i = 0; i < PIPE_PRIM_MAX; i++)
 290                 if (primtypes[i])
 291                         ctx->primtype_mask |= (1 << i);
 292
 293         /* need some sane default in case state tracker doesn't
 294          * set some state:
 295          */
 296         ctx->sample_mask = 0xffff;
 297
 298         pctx = &ctx->base;
 299         pctx->screen = pscreen;
 300         pctx->priv = priv;
 301         pctx->flush = fd_context_flush;
 302         pctx->emit_string_marker = fd_emit_string_marker;
 303         pctx->set_debug_callback = fd_set_debug_callback;
 304         pctx->create_fence_fd = fd_create_fence_fd;
 305         pctx->fence_server_sync = fd_fence_server_sync;
 306         pctx->texture_barrier = fd_texture_barrier;
 307         pctx->memory_barrier = fd_memory_barrier;
 308
 309         pctx->stream_uploader = u_upload_create_default(pctx);
 310         if (!pctx->stream_uploader)
 311                 goto fail;
 312         pctx->const_uploader = pctx->stream_uploader;
 313
 314         ctx->batch = fd_bc_alloc_batch(&screen->batch_cache, ctx);
 315
 316         slab_create_child(&ctx->transfer_pool, &screen->transfer_pool);
 317
 318         if (!ctx->blit)
 319                 ctx->blit = fd_blitter_blit;
 320
 321         fd_draw_init(pctx);
 322         fd_resource_context_init(pctx);
 323         fd_query_context_init(pctx);
 324         fd_texture_init(pctx);
 325         fd_state_init(pctx);
 326
 327         ctx->blitter = util_blitter_create(pctx);
 328         if (!ctx->blitter)
 329                 goto fail;
 330
 331         ctx->primconvert = util_primconvert_create(pctx, ctx->primtype_mask);
 332         if (!ctx->primconvert)
 333                 goto fail;
 334
 335         list_inithead(&ctx->hw_active_queries);
 336         list_inithead(&ctx->acc_active_queries);
 337
 338         return pctx;
 339
 340 fail:
 341         pctx->destroy(pctx);
 342         return NULL;
 343 }