gpu: Reuse transfer buffers more aggresively

gpu/command_buffer/client/gles2_implementation.cc

 // Copyright (c) 2012 The Chromium Authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
 // A class to emulate GLES2 over command buffers.
 
 #include "gpu/command_buffer/client/gles2_implementation.h"
 
 #include <algorithm>
 #include <map>
@@ skipping 1360 matching lines @@
   }
 
   GLuint buffer_id;
   if (GetBoundPixelTransferBuffer(target, "glBufferData", &buffer_id)) {
     if (!buffer_id) {
       return;
     }
 
     BufferTracker::Buffer* buffer = buffer_tracker_->GetBuffer(buffer_id);
     if (buffer) {
-      // Free buffer memory, pending the passage of a token.
-      buffer_tracker_->FreePendingToken(buffer, helper_->InsertToken());
-
-      // Remove old buffer.
-      buffer_tracker_->RemoveBuffer(buffer_id);
+      FreeTransferBuffer(buffer);

[piman, 2014/01/16 21:22:50]

How does synchronization work wrt asynchronous ReadPixels?

[jadahl, 2014/01/17 08:50:25]

Hmm. Good point. Reading how readpixels is implemented, it looks like it should
be enough to mark last usage after helper_->ReadPixels() when the
pack_transfer_buffer is bound, assuming service side doesn't reads async. Do you
think that'd be enough?

+      buffer_tracker_->RemoveBuffer(buffer->id());
     }
 
     // Create new buffer.
     buffer = buffer_tracker_->CreateBuffer(buffer_id, size);
     DCHECK(buffer);
     if (buffer->address() && data)
       memcpy(buffer->address(), data, size);
     return;
   }
 
@@ skipping 108 matching lines @@
     GLenum target, GLintptr offset, GLsizeiptr size, const void* data) {
   GPU_CLIENT_SINGLE_THREAD_CHECK();
   GPU_CLIENT_LOG("[" << GetLogPrefix() << "] glBufferSubData("
       << GLES2Util::GetStringBufferTarget(target) << ", "
       << offset << ", " << size << ", "
       << static_cast<const void*>(data) << ")");
   BufferSubDataHelper(target, offset, size, data);
   CheckGLError();
 }
 
+void GLES2Implementation::FreeTransferBuffer(BufferTracker::Buffer* buffer) {
+  int32 token = buffer->last_usage_token();
+
+  if (buffer->async_query_id()) {
+    QueryTracker::Query* query =
+        query_tracker_->GetQuery(buffer->async_query_id());
+    if (!query->CheckResultsAvailable(helper_))
+      buffer_tracker_->FreePendingSerial(buffer, query->serial());
+    else
+      buffer_tracker_->Free(buffer);
+  } else if (token) {
+    if (helper_->HasTokenPassed(token))
+      buffer_tracker_->Free(buffer);
+    else
+      buffer_tracker_->FreePendingToken(buffer, token);
+  }
+}
+
 bool GLES2Implementation::GetBoundPixelTransferBuffer(
     GLenum target,
     const char* function_name,
     GLuint* buffer_id) {
   *buffer_id = 0;
 
   switch (target) {
     case GL_PIXEL_PACK_TRANSFER_BUFFER_CHROMIUM:
       *buffer_id = bound_pixel_pack_transfer_buffer_id_;
       break;
@@ skipping 55 matching lines @@
   // CompressedTexImage2D.
   if (bound_pixel_unpack_transfer_buffer_id_) {
     GLuint offset = ToGLuint(data);
     BufferTracker::Buffer* buffer = GetBoundPixelUnpackTransferBufferIfValid(
         bound_pixel_unpack_transfer_buffer_id_,
         "glCompressedTexImage2D", offset, image_size);
     if (buffer && buffer->shm_id() != -1) {
       helper_->CompressedTexImage2D(
           target, level, internalformat, width, height, border, image_size,
           buffer->shm_id(), buffer->shm_offset() + offset);
-      buffer->set_transfer_ready_token(helper_->InsertToken());
+      MarkPixelTransferBufferLastUsage(buffer);
     }
     return;
   }
   SetBucketContents(kResultBucketId, data, image_size);
   helper_->CompressedTexImage2DBucket(
       target, level, internalformat, width, height, border, kResultBucketId);
   // Free the bucket. This is not required but it does free up the memory.
   // and we don't have to wait for the result so from the client's perspective
   // it's cheap.
   helper_->SetBucketSize(kResultBucketId, 0);
@@ skipping 20 matching lines @@
   // CompressedTexSubImage2D.
   if (bound_pixel_unpack_transfer_buffer_id_) {
     GLuint offset = ToGLuint(data);
     BufferTracker::Buffer* buffer = GetBoundPixelUnpackTransferBufferIfValid(
         bound_pixel_unpack_transfer_buffer_id_,
         "glCompressedTexSubImage2D", offset, image_size);
     if (buffer && buffer->shm_id() != -1) {
       helper_->CompressedTexSubImage2D(
           target, level, xoffset, yoffset, width, height, format, image_size,
           buffer->shm_id(), buffer->shm_offset() + offset);
-      buffer->set_transfer_ready_token(helper_->InsertToken());
+      MarkPixelTransferBufferLastUsage(buffer);
       CheckGLError();
     }
     return;
   }
   SetBucketContents(kResultBucketId, data, image_size);
   helper_->CompressedTexSubImage2DBucket(
       target, level, xoffset, yoffset, width, height, format, kResultBucketId);
   // Free the bucket. This is not required but it does free up the memory.
   // and we don't have to wait for the result so from the client's perspective
   // it's cheap.
@@ skipping 66 matching lines @@
   // If there's a pixel unpack buffer bound use it when issuing TexImage2D.
   if (bound_pixel_unpack_transfer_buffer_id_) {
     GLuint offset = ToGLuint(pixels);
     BufferTracker::Buffer* buffer = GetBoundPixelUnpackTransferBufferIfValid(
         bound_pixel_unpack_transfer_buffer_id_,
         "glTexImage2D", offset, size);
     if (buffer && buffer->shm_id() != -1) {
       helper_->TexImage2D(
           target, level, internalformat, width, height, border, format, type,
           buffer->shm_id(), buffer->shm_offset() + offset);
-      buffer->set_transfer_ready_token(helper_->InsertToken());
+      MarkPixelTransferBufferLastUsage(buffer);
       CheckGLError();
     }
     return;
   }
 
   // If there's no data just issue TexImage2D
   if (!pixels) {
     helper_->TexImage2D(
        target, level, internalformat, width, height, border, format, type,
        0, 0);
@@ skipping 85 matching lines @@
   // If there's a pixel unpack buffer bound use it when issuing TexSubImage2D.
   if (bound_pixel_unpack_transfer_buffer_id_) {
     GLuint offset = ToGLuint(pixels);
     BufferTracker::Buffer* buffer = GetBoundPixelUnpackTransferBufferIfValid(
         bound_pixel_unpack_transfer_buffer_id_,
         "glTexSubImage2D", offset, temp_size);
     if (buffer && buffer->shm_id() != -1) {
       helper_->TexSubImage2D(
           target, level, xoffset, yoffset, width, height, format, type,
           buffer->shm_id(), buffer->shm_offset() + offset, false);
-      buffer->set_transfer_ready_token(helper_->InsertToken());
+      MarkPixelTransferBufferLastUsage(buffer);
       CheckGLError();
     }
     return;
   }
 
   // compute the advance bytes per row for the src pixels
   uint32 src_padded_row_size;
   if (unpack_row_length_ > 0) {
     if (!GLES2Util::ComputeImagePaddedRowSize(
         unpack_row_length_, format, type, unpack_alignment_,
@@ skipping 562 matching lines @@
 }
 
 // NOTE #1: On old versions of OpenGL, calling glBindXXX with an unused id
 // generates a new resource. On newer versions of OpenGL they don't. The code
 // related to binding below will need to change if we switch to the new OpenGL
 // model. Specifically it assumes a bind will succeed which is always true in
 // the old model but possibly not true in the new model if another context has
 // deleted the resource.
 
 bool GLES2Implementation::BindBufferHelper(
-    GLenum target, GLuint buffer) {
+    GLenum target, GLuint buffer_id) {
   // TODO(gman): See note #1 above.
   bool changed = false;
   switch (target) {
     case GL_ARRAY_BUFFER:
-      if (bound_array_buffer_id_ != buffer) {
-        bound_array_buffer_id_ = buffer;
+      if (bound_array_buffer_id_ != buffer_id) {
+        bound_array_buffer_id_ = buffer_id;
         changed = true;
       }
       break;
     case GL_ELEMENT_ARRAY_BUFFER:
-      changed = vertex_array_object_manager_->BindElementArray(buffer);
+      changed = vertex_array_object_manager_->BindElementArray(buffer_id);
       break;
     case GL_PIXEL_PACK_TRANSFER_BUFFER_CHROMIUM:
-      bound_pixel_pack_transfer_buffer_id_ = buffer;
+      bound_pixel_pack_transfer_buffer_id_ = buffer_id;
       break;
     case GL_PIXEL_UNPACK_TRANSFER_BUFFER_CHROMIUM:
-      bound_pixel_unpack_transfer_buffer_id_ = buffer;
+      bound_pixel_unpack_transfer_buffer_id_ = buffer_id;
       break;
     default:
       changed = true;
       break;
   }
   // TODO(gman): There's a bug here. If the target is invalid the ID will not be
   // used even though it's marked it as used here.
-  GetIdHandler(id_namespaces::kBuffers)->MarkAsUsedForBind(buffer);
+  GetIdHandler(id_namespaces::kBuffers)->MarkAsUsedForBind(buffer_id);
   return changed;
 }
 
 bool GLES2Implementation::BindFramebufferHelper(
     GLenum target, GLuint framebuffer) {
   // TODO(gman): See note #1 above.
   bool changed = false;
   switch (target) {
     case GL_FRAMEBUFFER:
       if (bound_framebuffer_ != framebuffer ||
@@ skipping 113 matching lines @@
     SetGLError(
         GL_INVALID_VALUE,
         "glDeleteBuffers", "id not created by this context.");
     return;
   }
   for (GLsizei ii = 0; ii < n; ++ii) {
     if (buffers[ii] == bound_array_buffer_id_) {
       bound_array_buffer_id_ = 0;
     }
     vertex_array_object_manager_->UnbindBuffer(buffers[ii]);
+
     BufferTracker::Buffer* buffer = buffer_tracker_->GetBuffer(buffers[ii]);
     if (buffer) {
-      // Free buffer memory, pending the passage of a token.
-      buffer_tracker_->FreePendingToken(buffer, helper_->InsertToken());
-      // Remove buffer.
-      buffer_tracker_->RemoveBuffer(buffers[ii]);
+      FreeTransferBuffer(buffer);
+      buffer_tracker_->RemoveBuffer(buffer->id());
     }
+
     if (buffers[ii] == bound_pixel_unpack_transfer_buffer_id_) {
       bound_pixel_unpack_transfer_buffer_id_ = 0;
     }
   }
 }
 
 void GLES2Implementation::DeleteBuffersStub(
     GLsizei n, const GLuint* buffers) {
   helper_->DeleteBuffersImmediate(n, buffers);
 }
@@ skipping 699 matching lines @@
   // queries across contexts?
   return query_tracker_->GetQuery(id) != NULL;
 }
 
 void GLES2Implementation::BeginQueryEXT(GLenum target, GLuint id) {
   GPU_CLIENT_SINGLE_THREAD_CHECK();
   GPU_CLIENT_LOG("[" << GetLogPrefix() << "] BeginQueryEXT("
                  << GLES2Util::GetStringQueryTarget(target)
                  << ", " << id << ")");
 
+  QueryKey key = std::make_pair(target, false);
+
   // if any outstanding queries INV_OP
-  QueryMap::iterator it = current_queries_.find(target);
+  QueryMap::iterator it = current_queries_.find(key);
   if (it != current_queries_.end()) {
     SetGLError(
         GL_INVALID_OPERATION, "glBeginQueryEXT", "query already in progress");
     return;
   }
 
   // id = 0 INV_OP
   if (id == 0) {
     SetGLError(GL_INVALID_OPERATION, "glBeginQueryEXT", "id is 0");
     return;
   }
 
   // TODO(gman) if id not GENned INV_OPERATION
 
   // if id does not have an object
   QueryTracker::Query* query = query_tracker_->GetQuery(id);
   if (!query) {
     query = query_tracker_->CreateQuery(id, target);
     if (!query) {
       MustBeContextLost();
       return;
     }
   } else if (query->target() != target) {
     SetGLError(
         GL_INVALID_OPERATION, "glBeginQueryEXT", "target does not match");
     return;
   }
 
-  current_queries_[target] = query;
+  current_queries_[key] = query;
 
   query->Begin(this);
   CheckGLError();
 }
 
 void GLES2Implementation::EndQueryEXT(GLenum target) {
   GPU_CLIENT_SINGLE_THREAD_CHECK();
   GPU_CLIENT_LOG("[" << GetLogPrefix() << "] EndQueryEXT("
                  << GLES2Util::GetStringQueryTarget(target) << ")");
   // Don't do anything if the context is lost.
   if (helper_->IsContextLost()) {
     return;
   }
 
-  QueryMap::iterator it = current_queries_.find(target);
+  QueryKey key = std::make_pair(target, false);
+  QueryMap::iterator it = current_queries_.find(key);
   if (it == current_queries_.end()) {
     SetGLError(GL_INVALID_OPERATION, "glEndQueryEXT", "no active query");
     return;
   }
 
   QueryTracker::Query* query = it->second;
   query->End(this);
   current_queries_.erase(it);
   CheckGLError();
 }
 
 void GLES2Implementation::GetQueryivEXT(
     GLenum target, GLenum pname, GLint* params) {
   GPU_CLIENT_SINGLE_THREAD_CHECK();
   GPU_CLIENT_LOG("[" << GetLogPrefix() << "] GetQueryivEXT("
                  << GLES2Util::GetStringQueryTarget(target) << ", "
                  << GLES2Util::GetStringQueryParameter(pname) << ", "
                  << static_cast<const void*>(params) << ")");
 
   if (pname != GL_CURRENT_QUERY_EXT) {
     SetGLErrorInvalidEnum("glGetQueryivEXT", pname, "pname");
     return;
   }
-  QueryMap::iterator it = current_queries_.find(target);
+  QueryKey key = std::make_pair(target, false);
+  QueryMap::iterator it = current_queries_.find(key);
   if (it != current_queries_.end()) {
     QueryTracker::Query* query = it->second;
     *params = query->id();
   } else {
     *params = 0;
   }
   GPU_CLIENT_LOG("  " << *params);
   CheckGLError();
 }
 
 void GLES2Implementation::GetQueryObjectuivEXT(
     GLuint id, GLenum pname, GLuint* params) {
   GPU_CLIENT_SINGLE_THREAD_CHECK();
   GPU_CLIENT_LOG("[" << GetLogPrefix() << "] GetQueryivEXT(" << id << ", "
                  << GLES2Util::GetStringQueryObjectParameter(pname) << ", "
                  << static_cast<const void*>(params) << ")");
 
   QueryTracker::Query* query = query_tracker_->GetQuery(id);
   if (!query) {
     SetGLError(GL_INVALID_OPERATION, "glQueryObjectuivEXT", "unknown query id");
     return;
   }
 
-  QueryMap::iterator it = current_queries_.find(query->target());
+  QueryKey key = std::make_pair(query->target(), false);
+  QueryMap::iterator it = current_queries_.find(key);
   if (it != current_queries_.end()) {
     SetGLError(
         GL_INVALID_OPERATION,
         "glQueryObjectuivEXT", "query active. Did you to call glEndQueryEXT?");
     return;
   }
 
   if (query->NeverUsed()) {
     SetGLError(
         GL_INVALID_OPERATION,
@@ skipping 209 matching lines @@
   }
   if (buffer->mapped()) {
     SetGLError(GL_INVALID_OPERATION, "glMapBufferCHROMIUM", "already mapped");
     return NULL;
   }
   // Here we wait for previous transfer operations to be finished.
   // TODO(hubbe): AsyncTex(Sub)Image2dCHROMIUM does not currently work
   // with this method of synchronization. Until this is fixed,
   // MapBufferCHROMIUM will not block even if the transfer is not ready
   // for these calls.
-  if (buffer->transfer_ready_token()) {
-    helper_->WaitForToken(buffer->transfer_ready_token());
-    buffer->set_transfer_ready_token(0);
+  if (buffer->last_usage_token()) {
+    helper_->WaitForToken(buffer->last_usage_token());
+    buffer->set_last_usage_token(0);
   }
   buffer->set_mapped(true);
 
   GPU_CLIENT_LOG("  returned " << buffer->address());
   CheckGLError();
   return buffer->address();
 }
 
 GLboolean GLES2Implementation::UnmapBufferCHROMIUM(GLuint target) {
   GPU_CLIENT_SINGLE_THREAD_CHECK();
@@ skipping 13 matching lines @@
   }
   if (!buffer->mapped()) {
     SetGLError(GL_INVALID_OPERATION, "glUnmapBufferCHROMIUM", "not mapped");
     return false;
   }
   buffer->set_mapped(false);
   CheckGLError();
   return true;
 }
 
+void GLES2Implementation::BeginInternalQueryAsyncPixelUnpackCompleted(
+    BufferTracker::Buffer *buffer) {
+  QueryKey key = std::make_pair(GL_ASYNC_PIXEL_UNPACK_COMPLETED_CHROMIUM, true);
+
+  DCHECK(current_queries_.find(key) == current_queries_.end());
+
+  GLuint id;
+  GetIdHandler(id_namespaces::kQueries)->MakeIds(this, 0, 1, &id);
+  QueryTracker::Query* query = query_tracker_->CreateInternalQuery(
+      id, GL_ASYNC_PIXEL_UNPACK_COMPLETED_CHROMIUM);
+  if (!query) {
+    MustBeContextLost();
+    return;
+  }
+
+  buffer->set_async_query_id(id);
+  query->Begin(this);
+  CheckGLError();
+
+  current_queries_[key] = query;
+}
+
+void GLES2Implementation::EndInternalQueryAsyncPixelUnpackCompleted() {
+  QueryKey key = std::make_pair(GL_ASYNC_PIXEL_UNPACK_COMPLETED_CHROMIUM, true);
+
+  DCHECK(current_queries_.find(key) != current_queries_.end());
+
+  QueryMap::iterator it = current_queries_.find(key);
+  QueryTracker::Query* query = it->second;
+  query->End(this);
+  current_queries_.erase(it);
+  CheckGLError();
+}
+
+void GLES2Implementation::MarkPixelTransferBufferLastUsage(
+    BufferTracker::Buffer* buffer) {
+  buffer->set_last_usage_token(helper_->InsertToken());
+}
+
 void GLES2Implementation::AsyncTexImage2DCHROMIUM(
     GLenum target, GLint level, GLint internalformat, GLsizei width,
     GLsizei height, GLint border, GLenum format, GLenum type,
     const void* pixels) {
   GPU_CLIENT_SINGLE_THREAD_CHECK();
   GPU_CLIENT_LOG("[" << GetLogPrefix() << "] glTexImage2D("
       << GLES2Util::GetStringTextureTarget(target) << ", "
       << level << ", "
       << GLES2Util::GetStringTextureInternalFormat(internalformat) << ", "
       << width << ", " << height << ", " << border << ", "
@@ skipping 24 matching lines @@
 
   // Otherwise, async uploads require a transfer buffer to be bound.
   // TODO(hubbe): Make MapBufferCHROMIUM block if someone tries to re-use
   // the buffer before the transfer is finished. (Currently such
   // synchronization has to be handled manually.)
   GLuint offset = ToGLuint(pixels);
   BufferTracker::Buffer* buffer = GetBoundPixelUnpackTransferBufferIfValid(
       bound_pixel_unpack_transfer_buffer_id_,
       "glAsyncTexImage2DCHROMIUM", offset, size);
   if (buffer && buffer->shm_id() != -1) {
+    BeginInternalQueryAsyncPixelUnpackCompleted(buffer);
     helper_->AsyncTexImage2DCHROMIUM(
         target, level, internalformat, width, height, border, format, type,
         buffer->shm_id(), buffer->shm_offset() + offset);
+    EndInternalQueryAsyncPixelUnpackCompleted();
   }
 }
 
 void GLES2Implementation::AsyncTexSubImage2DCHROMIUM(
     GLenum target, GLint level, GLint xoffset, GLint yoffset, GLsizei width,
     GLsizei height, GLenum format, GLenum type, const void* pixels) {
   GPU_CLIENT_SINGLE_THREAD_CHECK();
   GPU_CLIENT_LOG("[" << GetLogPrefix() << "] glAsyncTexSubImage2DCHROMIUM("
       << GLES2Util::GetStringTextureTarget(target) << ", "
       << level << ", "
@@ skipping 21 matching lines @@
 
   // Async uploads require a transfer buffer to be bound.
   // TODO(hubbe): Make MapBufferCHROMIUM block if someone tries to re-use
   // the buffer before the transfer is finished. (Currently such
   // synchronization has to be handled manually.)
   GLuint offset = ToGLuint(pixels);
   BufferTracker::Buffer* buffer = GetBoundPixelUnpackTransferBufferIfValid(
       bound_pixel_unpack_transfer_buffer_id_,
       "glAsyncTexSubImage2DCHROMIUM", offset, size);
   if (buffer && buffer->shm_id() != -1) {
+    BeginInternalQueryAsyncPixelUnpackCompleted(buffer);
     helper_->AsyncTexSubImage2DCHROMIUM(
         target, level, xoffset, yoffset, width, height, format, type,
         buffer->shm_id(), buffer->shm_offset() + offset);
+    EndInternalQueryAsyncPixelUnpackCompleted();
   }
 }
 
 void GLES2Implementation::WaitAsyncTexImage2DCHROMIUM(GLenum target) {
   GPU_CLIENT_SINGLE_THREAD_CHECK();
   GPU_CLIENT_LOG("[" << GetLogPrefix() << "] glWaitAsyncTexImage2DCHROMIUM("
       << GLES2Util::GetStringTextureTarget(target) << ")");
   helper_->WaitAsyncTexImage2DCHROMIUM(target);
   CheckGLError();
 }
@@ skipping 166 matching lines @@
   CheckGLError();
 }
 
 // Include the auto-generated part of this file. We split this because it means
 // we can easily edit the non-auto generated parts right here in this file
 // instead of having to edit some template or the code generator.
 #include "gpu/command_buffer/client/gles2_implementation_impl_autogen.h"
 
 }  // namespace gles2
 }  // namespace gpu