freedreno_context.c

/*
 * Copyright (C) 2012 Rob Clark <[email protected]>
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice (including the next
 * paragraph) shall be included in all copies or substantial portions of the
 * Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 *
 * Authors:
 *    Rob Clark <[email protected]>
 */

#include "freedreno_context.h"
#include "ir3/ir3_cache.h"
#include "util/u_upload_mgr.h"
#include "freedreno_blitter.h"
#include "freedreno_draw.h"
#include "freedreno_fence.h"
#include "freedreno_gmem.h"
#include "freedreno_program.h"
#include "freedreno_query.h"
#include "freedreno_query_hw.h"
#include "freedreno_resource.h"
#include "freedreno_state.h"
#include "freedreno_texture.h"
#include "freedreno_util.h"

static void
fd_context_flush(struct pipe_context *pctx, struct pipe_fence_handle **fencep,
                 unsigned flags) in_dt
{
   struct fd_context *ctx = fd_context(pctx);
   struct pipe_fence_handle *fence = NULL;
   struct fd_batch *batch = NULL;

   /* We want to lookup current batch if it exists, but not create a new
    * one if not (unless we need a fence)
    */
   fd_batch_reference(&batch, ctx->batch);

   DBG("%p: flush: flags=%x, fencep=%p", batch, flags, fencep);

   if (fencep && !batch) {
      batch = fd_context_batch(ctx);
   } else if (!batch) {
      if (ctx->screen->reorder)
         fd_bc_flush(ctx, flags & PIPE_FLUSH_DEFERRED);
      fd_bc_dump(ctx, "%p: NULL batch, remaining:\n", ctx);
      return;
   }

   /* With TC_FLUSH_ASYNC, the fence will have been pre-created from
    * the front-end thread.  But not yet associated with a batch,
    * because we cannot safely access ctx->batch outside of the driver
    * thread.  So instead, replace the existing batch->fence with the
    * one created earlier
    */
   if ((flags & TC_FLUSH_ASYNC) && fencep) {
      /* We don't currently expect async+flush in the fence-fd
       * case.. for that to work properly we'd need TC to tell
       * us in the create_fence callback that it needs an fd.
       */
      assert(!(flags & PIPE_FLUSH_FENCE_FD));

      fd_fence_set_batch(*fencep, batch);
      fd_fence_ref(&batch->fence, *fencep);

      /* If we have nothing to flush, update the pre-created unflushed
       * fence with the current state of the last-fence:
       */
      if (ctx->last_fence) {
         fd_fence_repopulate(*fencep, ctx->last_fence);
         fd_fence_ref(&fence, *fencep);
         fd_bc_dump(ctx, "%p: (deferred) reuse last_fence, remaining:\n", ctx);
         goto out;
      }

      /* async flush is not compatible with deferred flush, since
       * nothing triggers the batch flush which fence_flush() would
       * be waiting for
       */
      flags &= ~PIPE_FLUSH_DEFERRED;
   } else if (!batch->fence) {
      batch->fence = fd_fence_create(batch);
   }

   /* In some sequence of events, we can end up with a last_fence that is
    * not an "fd" fence, which results in eglDupNativeFenceFDANDROID()
    * errors.
    */
   if ((flags & PIPE_FLUSH_FENCE_FD) && ctx->last_fence &&
       !fd_fence_is_fd(ctx->last_fence))
      fd_fence_ref(&ctx->last_fence, NULL);

   /* if no rendering since last flush, ie. app just decided it needed
    * a fence, re-use the last one:
    */
   if (ctx->last_fence) {
      fd_fence_ref(&fence, ctx->last_fence);
      fd_bc_dump(ctx, "%p: reuse last_fence, remaining:\n", ctx);
      goto out;
   }

   /* Take a ref to the batch's fence (batch can be unref'd when flushed: */
   fd_fence_ref(&fence, batch->fence);

   if (flags & PIPE_FLUSH_FENCE_FD)
      fence->submit_fence.use_fence_fd = true;

   fd_bc_dump(ctx, "%p: flushing %p<%u>, flags=0x%x, pending:\n", ctx,
              batch, batch->seqno, flags);

   /* If we get here, we need to flush for a fence, even if there is
    * no rendering yet:
    */
   batch->needs_flush = true;

   if (!ctx->screen->reorder) {
      fd_batch_flush(batch);
   } else {
      fd_bc_flush(ctx, flags & PIPE_FLUSH_DEFERRED);
   }

   fd_bc_dump(ctx, "%p: remaining:\n", ctx);

out:
   if (fencep)
      fd_fence_ref(fencep, fence);

   fd_fence_ref(&ctx->last_fence, fence);

   fd_fence_ref(&fence, NULL);

   fd_batch_reference(&batch, NULL);

   u_trace_context_process(&ctx->trace_context,
                           !!(flags & PIPE_FLUSH_END_OF_FRAME));
}

static void
fd_texture_barrier(struct pipe_context *pctx, unsigned flags) in_dt
{
   if (flags == PIPE_TEXTURE_BARRIER_FRAMEBUFFER) {
      struct fd_context *ctx = fd_context(pctx);

      if (ctx->framebuffer_barrier) {
         ctx->framebuffer_barrier(ctx);
         return;
      }
   }

   /* On devices that could sample from GMEM we could possibly do better.
    * Or if we knew that we were doing GMEM bypass we could just emit a
    * cache flush, perhaps?  But we don't know if future draws would cause
    * us to use GMEM, and a flush in bypass isn't the end of the world.
    */
   fd_context_flush(pctx, NULL, 0);
}

static void
fd_memory_barrier(struct pipe_context *pctx, unsigned flags)
{
   if (!(flags & ~PIPE_BARRIER_UPDATE))
      return;

   fd_context_flush(pctx, NULL, 0);

   /* TODO do we need to check for persistently mapped buffers and
    * fd_bo_cpu_prep()??
    */
}

static void
emit_string_tail(struct fd_ringbuffer *ring, const char *string, int len)
{
   const uint32_t *buf = (const void *)string;

   while (len >= 4) {
      OUT_RING(ring, *buf);
      buf++;
      len -= 4;
   }

   /* copy remainder bytes without reading past end of input string: */
   if (len > 0) {
      uint32_t w = 0;
      memcpy(&w, buf, len);
      OUT_RING(ring, w);
   }
}

/* for prior to a5xx: */
void
fd_emit_string(struct fd_ringbuffer *ring, const char *string, int len)
{
   /* max packet size is 0x3fff+1 dwords: */
   len = MIN2(len, 0x4000 * 4);

   OUT_PKT3(ring, CP_NOP, align(len, 4) / 4);
   emit_string_tail(ring, string, len);
}

/* for a5xx+ */
void
fd_emit_string5(struct fd_ringbuffer *ring, const char *string, int len)
{
   /* max packet size is 0x3fff dwords: */
   len = MIN2(len, 0x3fff * 4);

   OUT_PKT7(ring, CP_NOP, align(len, 4) / 4);
   emit_string_tail(ring, string, len);
}

/**
 * emit marker string as payload of a no-op packet, which can be
 * decoded by cffdump.
 */
static void
fd_emit_string_marker(struct pipe_context *pctx, const char *string,
                      int len) in_dt
{
   struct fd_context *ctx = fd_context(pctx);

   DBG("%.*s", len, string);

   if (!ctx->batch)
      return;

   struct fd_batch *batch = fd_context_batch_locked(ctx);

   fd_batch_needs_flush(batch);

   if (ctx->screen->gpu_id >= 500) {
      fd_emit_string5(batch->draw, string, len);
   } else {
      fd_emit_string(batch->draw, string, len);
   }

   fd_batch_unlock_submit(batch);
   fd_batch_reference(&batch, NULL);
}

/**
 * If we have a pending fence_server_sync() (GPU side sync), flush now.
 * The alternative to try to track this with batch dependencies gets
 * hairy quickly.
 *
 * Call this before switching to a different batch, to handle this case.
 */
void
fd_context_switch_from(struct fd_context *ctx)
{
   if (ctx->batch && (ctx->batch->in_fence_fd != -1))
      fd_batch_flush(ctx->batch);
}

/**
 * If there is a pending fence-fd that we need to sync on, this will
 * transfer the reference to the next batch we are going to render
 * to.
 */
void
fd_context_switch_to(struct fd_context *ctx, struct fd_batch *batch)
{
   if (ctx->in_fence_fd != -1) {
      sync_accumulate("freedreno", &batch->in_fence_fd, ctx->in_fence_fd);
      close(ctx->in_fence_fd);
      ctx->in_fence_fd = -1;
   }
}

/**
 * Return a reference to the current batch, caller must unref.
 */
struct fd_batch *
fd_context_batch(struct fd_context *ctx)
{
   struct fd_batch *batch = NULL;

   tc_assert_driver_thread(ctx->tc);

   fd_batch_reference(&batch, ctx->batch);

   if (unlikely(!batch)) {
      batch =
         fd_batch_from_fb(ctx, &ctx->framebuffer);
      util_copy_framebuffer_state(&batch->framebuffer, &ctx->framebuffer);
      fd_batch_reference(&ctx->batch, batch);
      fd_context_all_dirty(ctx);
   }
   fd_context_switch_to(ctx, batch);

   return batch;
}

/**
 * Return a locked reference to the current batch.  A batch with emit
 * lock held is protected against flushing while the lock is held.
 * The emit-lock should be acquired before screen-lock.  The emit-lock
 * should be held while emitting cmdstream.
 */
struct fd_batch *
fd_context_batch_locked(struct fd_context *ctx)
{
   struct fd_batch *batch = NULL;

   while (!batch) {
      batch = fd_context_batch(ctx);
      if (!fd_batch_lock_submit(batch)) {
         fd_batch_reference(&batch, NULL);
      }
   }

   return batch;
}

void
fd_context_destroy(struct pipe_context *pctx)
{
   struct fd_context *ctx = fd_context(pctx);
   unsigned i;

   DBG("");

   fd_screen_lock(ctx->screen);
   list_del(&ctx->node);
   fd_screen_unlock(ctx->screen);

   fd_fence_ref(&ctx->last_fence, NULL);

   if (ctx->in_fence_fd != -1)
      close(ctx->in_fence_fd);

   for (i = 0; i < ARRAY_SIZE(ctx->pvtmem); i++) {
      if (ctx->pvtmem[i].bo)
         fd_bo_del(ctx->pvtmem[i].bo);
   }

   util_copy_framebuffer_state(&ctx->framebuffer, NULL);
   fd_batch_reference(&ctx->batch, NULL); /* unref current batch */

   /* Make sure nothing in the batch cache references our context any more. */
   fd_bc_flush(ctx, false);

   fd_prog_fini(pctx);

   if (ctx->blitter)
      util_blitter_destroy(ctx->blitter);

   if (pctx->stream_uploader)
      u_upload_destroy(pctx->stream_uploader);

   for (i = 0; i < ARRAY_SIZE(ctx->clear_rs_state); i++)
      if (ctx->clear_rs_state[i])
         pctx->delete_rasterizer_state(pctx, ctx->clear_rs_state[i]);

   if (ctx->primconvert)
      util_primconvert_destroy(ctx->primconvert);

   slab_destroy_child(&ctx->transfer_pool);
   slab_destroy_child(&ctx->transfer_pool_unsync);

   for (i = 0; i < ARRAY_SIZE(ctx->vsc_pipe_bo); i++) {
      if (!ctx->vsc_pipe_bo[i])
         break;
      fd_bo_del(ctx->vsc_pipe_bo[i]);
   }

   fd_device_del(ctx->dev);
   fd_pipe_purge(ctx->pipe);
   fd_pipe_del(ctx->pipe);

   simple_mtx_destroy(&ctx->gmem_lock);

   u_trace_context_fini(&ctx->trace_context);

   fd_autotune_fini(&ctx->autotune);

   ir3_cache_destroy(ctx->shader_cache);

   if (FD_DBG(BSTAT) || FD_DBG(MSGS)) {
      mesa_logi(
         "batch_total=%u, batch_sysmem=%u, batch_gmem=%u, batch_nondraw=%u, "
         "batch_restore=%u\n",
         (uint32_t)ctx->stats.batch_total, (uint32_t)ctx->stats.batch_sysmem,
         (uint32_t)ctx->stats.batch_gmem, (uint32_t)ctx->stats.batch_nondraw,
         (uint32_t)ctx->stats.batch_restore);
   }
}

static void
fd_set_debug_callback(struct pipe_context *pctx,
                      const struct pipe_debug_callback *cb)
{
   struct fd_context *ctx = fd_context(pctx);

   if (cb)
      ctx->debug = *cb;
   else
      memset(&ctx->debug, 0, sizeof(ctx->debug));
}

static uint32_t
fd_get_reset_count(struct fd_context *ctx, bool per_context)
{
   uint64_t val;
   enum fd_param_id param = per_context ? FD_CTX_FAULTS : FD_GLOBAL_FAULTS;
   int ret = fd_pipe_get_param(ctx->pipe, param, &val);
   debug_assert(!ret);
   return val;
}

static enum pipe_reset_status
fd_get_device_reset_status(struct pipe_context *pctx)
{
   struct fd_context *ctx = fd_context(pctx);
   int context_faults = fd_get_reset_count(ctx, true);
   int global_faults = fd_get_reset_count(ctx, false);
   enum pipe_reset_status status;

   /* Not called in driver thread, but threaded_context syncs
    * before calling this:
    */
   fd_context_access_begin(ctx);

   if (context_faults != ctx->context_reset_count) {
      status = PIPE_GUILTY_CONTEXT_RESET;
   } else if (global_faults != ctx->global_reset_count) {
      status = PIPE_INNOCENT_CONTEXT_RESET;
   } else {
      status = PIPE_NO_RESET;
   }

   ctx->context_reset_count = context_faults;
   ctx->global_reset_count = global_faults;

   fd_context_access_end(ctx);

   return status;
}

static void
fd_trace_record_ts(struct u_trace *ut, struct pipe_resource *timestamps,
                   unsigned idx)
{
   struct fd_batch *batch = container_of(ut, struct fd_batch, trace);
   struct fd_ringbuffer *ring = batch->nondraw ? batch->draw : batch->gmem;

   if (ring->cur == batch->last_timestamp_cmd) {
      uint64_t *ts = fd_bo_map(fd_resource(timestamps)->bo);
      ts[idx] = U_TRACE_NO_TIMESTAMP;
      return;
   }

   unsigned ts_offset = idx * sizeof(uint64_t);
   batch->ctx->record_timestamp(ring, fd_resource(timestamps)->bo, ts_offset);
   batch->last_timestamp_cmd = ring->cur;
}

static uint64_t
fd_trace_read_ts(struct u_trace_context *utctx,
                 struct pipe_resource *timestamps, unsigned idx)
{
   struct fd_context *ctx =
      container_of(utctx, struct fd_context, trace_context);
   struct fd_bo *ts_bo = fd_resource(timestamps)->bo;

   /* Only need to stall on results for the first entry: */
   if (idx == 0) {
      /* Avoid triggering deferred submits from flushing, since that
       * changes the behavior of what we are trying to measure:
       */
      while (fd_bo_cpu_prep(ts_bo, ctx->pipe, FD_BO_PREP_NOSYNC))
         usleep(10000);
      int ret = fd_bo_cpu_prep(ts_bo, ctx->pipe, FD_BO_PREP_READ);
      if (ret)
         return U_TRACE_NO_TIMESTAMP;
   }

   uint64_t *ts = fd_bo_map(ts_bo);

   /* Don't translate the no-timestamp marker: */
   if (ts[idx] == U_TRACE_NO_TIMESTAMP)
      return U_TRACE_NO_TIMESTAMP;

   return ctx->ts_to_ns(ts[idx]);
}

/* TODO we could combine a few of these small buffers (solid_vbuf,
 * blit_texcoord_vbuf, and vsc_size_mem, into a single buffer and
 * save a tiny bit of memory
 */

static struct pipe_resource *
create_solid_vertexbuf(struct pipe_context *pctx)
{
   static const float init_shader_const[] = {
      -1.000000, +1.000000, +1.000000, +1.000000, -1.000000, +1.000000,
   };
   struct pipe_resource *prsc =
      pipe_buffer_create(pctx->screen, PIPE_BIND_CUSTOM, PIPE_USAGE_IMMUTABLE,
                         sizeof(init_shader_const));
   pipe_buffer_write(pctx, prsc, 0, sizeof(init_shader_const),
                     init_shader_const);
   return prsc;
}

static struct pipe_resource *
create_blit_texcoord_vertexbuf(struct pipe_context *pctx)
{
   struct pipe_resource *prsc = pipe_buffer_create(
      pctx->screen, PIPE_BIND_CUSTOM, PIPE_USAGE_DYNAMIC, 16);
   return prsc;
}

void
fd_context_setup_common_vbos(struct fd_context *ctx)
{
   struct pipe_context *pctx = &ctx->base;

   ctx->solid_vbuf = create_solid_vertexbuf(pctx);
   ctx->blit_texcoord_vbuf = create_blit_texcoord_vertexbuf(pctx);

   /* setup solid_vbuf_state: */
   ctx->solid_vbuf_state.vtx = pctx->create_vertex_elements_state(
      pctx, 1,
      (struct pipe_vertex_element[]){{
         .vertex_buffer_index = 0,
         .src_offset = 0,
         .src_format = PIPE_FORMAT_R32G32B32_FLOAT,
      }});
   ctx->solid_vbuf_state.vertexbuf.count = 1;
   ctx->solid_vbuf_state.vertexbuf.vb[0].stride = 12;
   ctx->solid_vbuf_state.vertexbuf.vb[0].buffer.resource = ctx->solid_vbuf;

   /* setup blit_vbuf_state: */
   ctx->blit_vbuf_state.vtx = pctx->create_vertex_elements_state(
      pctx, 2,
      (struct pipe_vertex_element[]){
         {
            .vertex_buffer_index = 0,
            .src_offset = 0,
            .src_format = PIPE_FORMAT_R32G32_FLOAT,
         },
         {
            .vertex_buffer_index = 1,
            .src_offset = 0,
            .src_format = PIPE_FORMAT_R32G32B32_FLOAT,
         }});
   ctx->blit_vbuf_state.vertexbuf.count = 2;
   ctx->blit_vbuf_state.vertexbuf.vb[0].stride = 8;
   ctx->blit_vbuf_state.vertexbuf.vb[0].buffer.resource =
      ctx->blit_texcoord_vbuf;
   ctx->blit_vbuf_state.vertexbuf.vb[1].stride = 12;
   ctx->blit_vbuf_state.vertexbuf.vb[1].buffer.resource = ctx->solid_vbuf;
}

void
fd_context_cleanup_common_vbos(struct fd_context *ctx)
{
   struct pipe_context *pctx = &ctx->base;

   pctx->delete_vertex_elements_state(pctx, ctx->solid_vbuf_state.vtx);
   pctx->delete_vertex_elements_state(pctx, ctx->blit_vbuf_state.vtx);

   pipe_resource_reference(&ctx->solid_vbuf, NULL);
   pipe_resource_reference(&ctx->blit_texcoord_vbuf, NULL);
}

struct pipe_context *
fd_context_init(struct fd_context *ctx, struct pipe_screen *pscreen,
                const uint8_t *primtypes, void *priv,
                unsigned flags) disable_thread_safety_analysis
{
   struct fd_screen *screen = fd_screen(pscreen);
   struct pipe_context *pctx;
   unsigned prio = 1;
   int i;

   /* lower numerical value == higher priority: */
   if (FD_DBG(HIPRIO))
      prio = 0;
   else if (flags & PIPE_CONTEXT_HIGH_PRIORITY)
      prio = 0;
   else if (flags & PIPE_CONTEXT_LOW_PRIORITY)
      prio = 2;

   /* Some of the stats will get printed out at context destroy, so
    * make sure they are collected:
    */
   if (FD_DBG(BSTAT) || FD_DBG(MSGS))
      ctx->stats_users++;

   ctx->screen = screen;
   ctx->pipe = fd_pipe_new2(screen->dev, FD_PIPE_3D, prio);

   ctx->in_fence_fd = -1;

   if (fd_device_version(screen->dev) >= FD_VERSION_ROBUSTNESS) {
      ctx->context_reset_count = fd_get_reset_count(ctx, true);
      ctx->global_reset_count = fd_get_reset_count(ctx, false);
   }

   ctx->primtypes = primtypes;
   ctx->primtype_mask = 0;
   for (i = 0; i <= PIPE_PRIM_MAX; i++)
      if (primtypes[i])
         ctx->primtype_mask |= (1 << i);

   simple_mtx_init(&ctx->gmem_lock, mtx_plain);

   /* need some sane default in case gallium frontends don't
    * set some state:
    */
   ctx->sample_mask = 0xffff;
   ctx->active_queries = true;

   pctx = &ctx->base;
   pctx->screen = pscreen;
   pctx->priv = priv;
   pctx->flush = fd_context_flush;
   pctx->emit_string_marker = fd_emit_string_marker;
   pctx->set_debug_callback = fd_set_debug_callback;
   pctx->get_device_reset_status = fd_get_device_reset_status;
   pctx->create_fence_fd = fd_create_fence_fd;
   pctx->fence_server_sync = fd_fence_server_sync;
   pctx->fence_server_signal = fd_fence_server_signal;
   pctx->texture_barrier = fd_texture_barrier;
   pctx->memory_barrier = fd_memory_barrier;

   pctx->stream_uploader = u_upload_create_default(pctx);
   if (!pctx->stream_uploader)
      goto fail;
   pctx->const_uploader = pctx->stream_uploader;

   slab_create_child(&ctx->transfer_pool, &screen->transfer_pool);
   slab_create_child(&ctx->transfer_pool_unsync, &screen->transfer_pool);

   fd_draw_init(pctx);
   fd_resource_context_init(pctx);
   fd_query_context_init(pctx);
   fd_texture_init(pctx);
   fd_state_init(pctx);

   ctx->blitter = util_blitter_create(pctx);
   if (!ctx->blitter)
      goto fail;

   ctx->primconvert = util_primconvert_create(pctx, ctx->primtype_mask);
   if (!ctx->primconvert)
      goto fail;

   list_inithead(&ctx->hw_active_queries);
   list_inithead(&ctx->acc_active_queries);

   fd_screen_lock(ctx->screen);
   ctx->seqno = ++screen->ctx_seqno;
   list_add(&ctx->node, &ctx->screen->context_list);
   fd_screen_unlock(ctx->screen);

   ctx->current_scissor = &ctx->disabled_scissor;

   u_trace_context_init(&ctx->trace_context, pctx, fd_trace_record_ts,
                        fd_trace_read_ts);

   fd_autotune_init(&ctx->autotune, screen->dev);

   return pctx;

fail:
   pctx->destroy(pctx);
   return NULL;
}

struct pipe_context *
fd_context_init_tc(struct pipe_context *pctx, unsigned flags)
{
   struct fd_context *ctx = fd_context(pctx);

   if (!(flags & PIPE_CONTEXT_PREFER_THREADED))
      return pctx;

   /* Clover (compute-only) is unsupported. */
   if (flags & PIPE_CONTEXT_COMPUTE_ONLY)
      return pctx;

   struct pipe_context *tc = threaded_context_create(
      pctx, &ctx->screen->transfer_pool,
      fd_replace_buffer_storage,
      fd_fence_create_unflushed,
      fd_resource_busy,
      false,
      &ctx->tc);

   uint64_t total_ram;
   if (tc && tc != pctx && os_get_total_physical_memory(&total_ram)) {
      ((struct threaded_context *)tc)->bytes_mapped_limit = total_ram / 16;
   }

   return tc;
}