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