|
53ea9cc6
|
2016-05-17T10:12:52
|
|
Replace rx::Renderer with rx::ContextImpl.
Previously Context had no Impl class, but had a special relationship
with the instanced Renderer class. Having a ContextImpl backing every
Context will allow new designs to enable things like multithreading
(where each ContextImpl stores a Context-specific device) or non-
virtual Contexts on Android or other platforms where it is more
efficient.
A large refactoring patch that touches every back-end.
BUG=angleproject:1363
Change-Id: Icb73a7d37447f08a664eeb499a310ba05d71a57e
Reviewed-on: https://chromium-review.googlesource.com/342052
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Commit-Queue: Jamie Madill <jmadill@chromium.org>
|
|
2b4ce80c
|
2016-04-28T13:34:50
|
|
Implement GL_CHROMIUM_sync_query for D3D9 and D3D11.
BUG=angleproject:1366
Change-Id: Iadde61968f45b969c76578a6dd9116a25d63fb4b
Reviewed-on: https://chromium-review.googlesource.com/341230
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Commit-Queue: Geoff Lang <geofflang@chromium.org>
|
|
|
9082b982
|
2016-04-27T15:21:51
|
|
Rename gl::Data to gl::ContextState.
Part of the new world order of renaming the Obj::Data classes to ObjState.
BUG=angleproject:1363
Change-Id: I15cf002b8b093d687f540b9e86f045874af24a7e
Reviewed-on: https://chromium-review.googlesource.com/340740
Reviewed-by: Geoff Lang <geofflang@chromium.org>
Commit-Queue: Jamie Madill <jmadill@chromium.org>
|
|
|
52b09c2f
|
2016-04-11T14:12:31
|
|
Re-re-land "D3D11: Implement dirty bits for VertexArray11.""
Translated attributes are now stored in the VertexArray11 in a cache,
and only updated when dirty bits change. Currently dynamic attributes
must be re-translated every call, so these are stored in a list and
processed repeatedly.
This skips doing a lot of the VertexDataManager work for vertex
attributes that don't change between draw calls.
Current value attributes, which correspond to disabled attributes that
the program will pulls vertex data from, are owned by the Context, so
these need to be handled outside of the VertexArray11.
Further changes will be necessary to reduce the redundant work we do in
the InputLayoutCache. We shouldn't need to re-check the cache if
nothing relevant changed.
This give about a 23% performance improvement on the draw call
benchmark on my machine.
Re-land with a fix for the start vertex offset.
Re-re-land with a fix for using XFB with deleted buffers.
BUG=angleproject:1327
Change-Id: I0fba49515375c149bbf54d933f8d1f747fbb8158
Reviewed-on: https://chromium-review.googlesource.com/338003
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Commit-Queue: Jamie Madill <jmadill@chromium.org>
|
|
|
53a36004
|
2016-04-08T19:03:18
|
|
Revert "Re-land "D3D11: Implement dirty bits for VertexArray11."""
Seems to make the following dEQP test flaky:
dEQP-GLES3.functional.lifetime.attach.deleted_output.buffer_transform_feedback
doesn't show up on every bot test, but run it a few times and it'll flake. Reverting while I investigate.
BUG=angleproject:1327
This reverts commit 3477f3a62dc139a253a0b361ee138116e9fa881f.
Change-Id: Ic23a392526f5f6e107cf0aa06448389804d6b208
Reviewed-on: https://chromium-review.googlesource.com/337961
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Commit-Queue: Jamie Madill <jmadill@chromium.org>
|
|
|
3477f3a6
|
2016-03-29T17:15:29
|
|
Re-land "D3D11: Implement dirty bits for VertexArray11.""
Translated attributes are now stored in the VertexArray11 in a cache,
and only updated when dirty bits change. Currently dynamic attributes
must be re-translated every call, so these are stored in a list and
processed repeatedly.
This skips doing a lot of the VertexDataManager work for vertex
attributes that don't change between draw calls.
Current value attributes, which correspond to disabled attributes that
the program will pulls vertex data from, are owned by the Context, so
these need to be handled outside of the VertexArray11.
Further changes will be necessary to reduce the redundant work we do in
the InputLayoutCache. We shouldn't need to re-check the cache if
nothing relevant changed.
This give about a 23% performance improvement on the draw call
benchmark on my machine.
Re-land with a fix for the start vertex offset.
BUG=angleproject:1327
Change-Id: Ic23e48fb18ed7f29c1999914a2f799ac04aa03e9
Reviewed-on: https://chromium-review.googlesource.com/334225
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Commit-Queue: Jamie Madill <jmadill@chromium.org>
|
|
00f394ec
|
2016-03-16T12:09:11
|
|
Revert Dirty bits for VertexArray11
This is a combination of two reverts:
Revert "D3D11: Remove unused mRenderer from VertexArray11."
Revert "D3D11: Implement dirty bits for VertexArray11."
Reverting only the first commit would trigger warnings on the Windows
clang bot.
BUG=594509
BUG=angleproject:1327
This reverts commit fc4712b5ed270436f2993bfda9e916d4f92684a4.
This reverts commit 7d8585b802b7eb741b380bd0d05769281d9507c9.
Change-Id: I612dbba0816d6144f71ce815701c13a798585bc7
Reviewed-on: https://chromium-review.googlesource.com/332989
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
|
|
|
7d8585b8
|
2016-03-09T15:53:12
|
|
D3D11: Implement dirty bits for VertexArray11.
Translated attributes are now stored in the VertexArray11 in a cache,
and only updated when dirty bits change. Currently dynamic attributes
must be re-translated every call, so these are stored in a list and
processed repeatedly.
This skips doing a lot of the VertexDataManager work for vertex
attributes that don't change between draw calls.
Current value attributes, which correspond to disabled attributes that
the program will pulls vertex data from, are owned by the Context, so
these need to be handled outside of the VertexArray11.
Further changes will be necessary to reduce the redundant work we do in
the InputLayoutCache. We shouldn't need to re-check the cache if
nothing relevant changed.
This give about a 23% performance improvement on the draw call
benchmark on my machine.
BUG=angleproject:1327
Change-Id: I7fb944d32ea7e6c78b9e478406bdb7e10a7fc05b
Reviewed-on: https://chromium-review.googlesource.com/330173
Reviewed-by: Zhenyao Mo <zmo@chromium.org>
Commit-Queue: Jamie Madill <jmadill@chromium.org>
|
|
|
e9e15349
|
2016-03-04T15:16:42
|
|
D3D11: Remove RTV and DSV cache vars from Renderer11.
These were only referenced by the StateManager11. Instead of using
these cache variables, ensure the StateManager11 marks the RT as
dirty and apply the RenderTarget state.
Also remove the RenderTarget setters from Renderer11, and don't
unapply RTs after a one-time set operation. The StateManager11
should take care of updating the necessary changes itself.
BUG=angleproject:1321
Change-Id: Idc8d22ea9e92ac8d07392ab10d067df6084403bb
Reviewed-on: https://chromium-review.googlesource.com/330462
Reviewed-by: John Bauman <jbauman@chromium.org>
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
|
|
|
1fbc59fe
|
2016-02-24T15:25:51
|
|
D3D11: Enable dirty bits for Framebuffer11.
This patch works using a notification scheme - whenever a Texture or
Renderbuffer changes in such a way as to recreate its RenderTarget, we
pass a signal to the Framebuffer to invalidate some internal state.
Everything is entirely tracked in the Renderer11 layer, and the GL
layer is left untouched.
A RenderTarget11 now tracks points to which it is bound, and the
Framebuffer11 is mostly responsible for managing those links.
The three locations where we notify a Framebuffer when its bound
RenderTargets might be dirty are:
1) RenderTarget11::~RenderTarget
2) EGLImageD3D::copyToLocalRendertarget
3) TextureStorage11_2D::useLevelZeroWorkaroundTexture
This patch gives about a 10% score increase in the D3D11 draw call
benchmark on my system.
BUG=angleproject:1260
Change-Id: Ide38aeadff4a2681bf5bd685e8ca3c9e2612a380
Reviewed-on: https://chromium-review.googlesource.com/327255
Reviewed-by: Geoff Lang <geofflang@chromium.org>
Commit-Queue: Jamie Madill <jmadill@chromium.org>
|
|
24a504ec
|
2016-02-18T13:53:00
|
|
Clear SRV cache on keyed mutex texture unbind from TextureStorage11
If the calling code releases the keyed mutex after the unbind then D3D will
unbind the SRV or RTV from the pipeline without notifying the cache. This
allows the SRV to be destroyed, which means that a subsequently-created SRV
could have the same address. This would cause the new SRV to not be correctly
bound to the pipeline if its slot is the same as the unbound SRV.
BUG=588000
Change-Id: I0ad1ba3706a6bcc1fee50745e6148c9b23e578f6
Reviewed-on: https://chromium-review.googlesource.com/328400
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Reviewed-by: Geoff Lang <geofflang@chromium.org>
Commit-Queue: John Bauman <jbauman@chromium.org>
|
|
|
b9ada71d
|
2016-02-17T11:45:58
|
|
Sync the D3D11 rasterizer state if pointDrawMode or multisample changes.
When drawing points, culling is disabled to make sure the generated point
sprites are not culled. Since there are no dirty bits for pointDrawMode or
multiSample, they were not being checked to ensure the rasterizer state was
correct.
BUG=586531
Change-Id: I8fe60dd8d5bbc79b1bce2c0aa62c40cee560fe24
Reviewed-on: https://chromium-review.googlesource.com/327862
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Commit-Queue: Geoff Lang <geofflang@chromium.org>
|
|
89f2845c
|
2016-02-10T11:33:07
|
|
Context virtualization for D3D11 queries
Queries in D3D11 are now virtualized such that they work as expected with
multiple contexts. Timer queries now only time the operations that their
context is responsible for and the operations of other contexts are
ignored.
BUG=angleproject:657
Change-Id: I667de594bdb5831d126d5801c0e692ded4c88bf4
Reviewed-on: https://chromium-review.googlesource.com/327150
Reviewed-by: Geoff Lang <geofflang@chromium.org>
Commit-Queue: Ian Ewell <ewell@google.com>
|
|
2a63b3f8
|
2016-02-08T12:29:08
|
|
Re-land "Implement EGL_experimental_present_path_angle"
- Re-land with clang fix.
This allows ANGLE to render directly onto a D3D swapchain in the correct
orientation when using the D3D11 renderer.
The trick is to add an extra uniform to each shader which takes either
the value +1.0 or -1.0. When rendering to a texture, ANGLE sets this
value to -1.0. When rendering to the default framebuffer, ANGLE sets
this value to +1.0. ANGLE multiplies vertex positions by this value in
the VS to invert rendering when appropriate. It also corrects other
state (e.g. viewport/scissor rect) and shader built-in values
(e.g. gl_FragCoord).
This saves a substantial amount of GPU time and lowers power
consumption. For example, the old method (where ANGLE renders all
content onto an offscreen texture, and then copies/inverts this onto the
swapchain at eglSwapBuffers() time) uses about 20% of the GPU each frame
on a Lumia 630.
Verification:
+ dEQP GL ES2 tests pass when "present path fast" is enabled
+ all ANGLE_end2end_tests pass when "present path fast" is enabled
BUG=angleproject:1219
Change-Id: I56b339897828753a616d7bae837a2f354dba9c63
Reviewed-on: https://chromium-review.googlesource.com/326730
Tryjob-Request: Austin Kinross <aukinros@microsoft.com>
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Reviewed-by: Geoff Lang <geofflang@chromium.org>
|
|
|
c9bdeff4
|
2016-02-08T12:36:55
|
|
D3D11: Refactor RenderTarget apply into StateManager11.
This is a refactoring patch only, and doesn't change the behaviour.
The intent is to make it easy to turn on the Framebuffer dirty bits
in a subsequent patch, once we can cleanly handle textures and
renderbuffers getting recreated.
BUG=angleproject:1260
Change-Id: Iaa5cfe222b020724e088eee5f1ae909b6f981a08
Reviewed-on: https://chromium-review.googlesource.com/325423
Reviewed-by: Geoff Lang <geofflang@chromium.org>
Commit-Queue: Jamie Madill <jmadill@chromium.org>
|
|
|
d1c46228
|
2016-02-08T14:51:18
|
|
Revert "Implement EGL_experimental_present_path_angle"
Compile failure on Clang/Win:
The reason for reverting is: FAILED: ninja -t msvc -e environment.x86 --
"..\..\third_party/llvm-build/Release+Asserts/bin/clang-cl" -m32 /nologo
/showIncludes /FC
@obj\third_party\angle\src\tests\egl_tests\angle_end2end_tests.EGLPresentPathD3D11Test.obj.rsp
/c ..\..\third_party\angle\src\tests\egl_tests\EGLPresentPathD3D11Test.cpp
/Foobj\third_party\angle\src\tests\egl_tests\angle_end2end_tests.EGLPresentPathD3D11Test.obj
/Fdobj\gpu\angle_end2end_tests.cc.pdb
In file included from
..\..\third_party\angle\src\tests\egl_tests\EGLPresentPathD3D11Test.cpp:7:
In file included from
..\..\third_party\angle\src\tests\test_utils/ANGLETest.h:13:
..\..\testing\gtest\include\gtest/gtest.h(1392,16) : error: comparison of
integers of different signs: 'const int' and 'const unsigned int'
[-Werror,-Wsign-compare]
if (expected == actual) {
~~~~~~~~ ^ ~~~~~~
..\..\testing\gtest\include\gtest/gtest.h(1422,12) : note: in instantiation of
function template specialization 'testing::internal::CmpHelperEQ<int, unsigned
int>' requested here
return CmpHelperEQ(expected_expression, actual_expression, expected,
^
..\..\third_party\angle\src\tests\egl_tests\EGLPresentPathD3D11Test.cpp(281,9) :
note: in instantiation of function template specialization
'testing::internal::EqHelper<false>::Compare<int, unsigned int>' requested here
ASSERT_EQ(mWindowWidth * 4, mappedSubresource.RowPitch);
^
..\..\testing\gtest\include\gtest/gtest.h(1960,32) : note: expanded from macro
'ASSERT_EQ'
# define ASSERT_EQ(val1, val2) GTEST_ASSERT_EQ(val1, val2)
^
..\..\testing\gtest\include\gtest/gtest.h(1943,67) : note: expanded from macro
'GTEST_ASSERT_EQ'
EqHelper<GTEST_IS_NULL_LITERAL_(expected)>::Compare, \
^
BUG=angleproject:1219
This reverts commit 6b3c1db5170450bbc4946d8f18ba0d8619da43a0.
Change-Id: Ia67ab82dd13295dc03235d57fa417c73f20a49e6
Reviewed-on: https://chromium-review.googlesource.com/326680
Reviewed-by: Jamie Madill <jmadill@chromium.org>
|
|
|
6b3c1db5
|
2015-12-18T14:01:46
|
|
Implement EGL_experimental_present_path_angle
This allows ANGLE to render directly onto a D3D swapchain in the correct
orientation when using the D3D11 renderer.
The trick is to add an extra uniform to each shader which takes either
the value +1.0 or -1.0. When rendering to a texture, ANGLE sets this
value to -1.0. When rendering to the default framebuffer, ANGLE sets
this value to +1.0. ANGLE multiplies vertex positions by this value in
the VS to invert rendering when appropriate. It also corrects other
state (e.g. viewport/scissor rect) and shader built-in values
(e.g. gl_FragCoord).
This saves a substantial amount of GPU time and lowers power
consumption. For example, the old method (where ANGLE renders all
content onto an offscreen texture, and then copies/inverts this onto the
swapchain at eglSwapBuffers() time) uses about 20% of the GPU each frame
on a Lumia 630.
Verification:
+ dEQP GL ES2 tests pass when "present path fast" is enabled
+ all ANGLE_end2end_tests pass when "present path fast" is enabled
BUG=angleproject:1219
Change-Id: Ib6eeea46bafa6ebce4adada0ae9db3a433b8fc4c
Reviewed-on: https://chromium-review.googlesource.com/321360
Reviewed-by: Geoff Lang <geofflang@chromium.org>
Tryjob-Request: Austin Kinross <aukinros@microsoft.com>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Commit-Queue: Geoff Lang <geofflang@chromium.org>
|
|
|
c9d442d7
|
2016-01-20T11:17:24
|
|
Sync dirty objects at the GL layer.
The dirty bit system currently puts dirty objects in line with the
rest of the dirty state. This means at the Renderer level, the sync
manager code would call back to the GL layer to sync the specific
objects that are dirty. This is a bit of a layering violation (impl
layer mutating top-level objects) and also a bit of repeated boiler-
plate code.
Fix this by treating dirty objects in a separate dirty bit set, called
the dirty objects. This also has the benefit of allowing us to re-
implement the dirty object set at a later date, if we want to store
them in a list, or other structure.
Also don't skip the state sync at the GL level if there are no GL
dirty bits. The Impl might have some dirty bits locally, and it's
better to call syncState and do the no-op check in the Impl than it
is to also sync local state at every sync point (draw call, read, etc)
in each Impl.
BUG=angleproject:1260
Change-Id: Id5d4beb2a1c4e3ad351edf54e3f32e828d5f5da3
Reviewed-on: https://chromium-review.googlesource.com/318790
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Tested-by: Jamie Madill <jmadill@chromium.org>
|
|
4028159e
|
2015-11-19T18:24:44
|
|
Using dirty bit notification for D3D11 viewport state
BUG=angleproject:1161
This is a continuation of the dirty bit refactor
Change-Id: I101f415094dd406f6b0a40cb260f1cbbfec7b62c
Reviewed-on: https://chromium-review.googlesource.com/313249
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Tryjob-Request: Dian Xiang <dianx@google.com>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Tested-by: Dian Xiang <dianx@google.com>
|
|
|
29a65546
|
2015-11-17T17:15:50
|
|
Using dirty bit notification for D3D11 scissor rectangle state
BUG=angleproject:1161
This is a continuation of the dirty bit refactor
Change-Id: I1a9f297835943eee5f756e4e4721b2db543184ec
Reviewed-on: https://chromium-review.googlesource.com/312987
Reviewed-by: Geoff Lang <geofflang@chromium.org>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Tested-by: Dian Xiang <dianx@google.com>
|
|
|
15ede106
|
2015-11-26T15:20:34
|
|
D3D11: Disable stencil buffer when not attached.
Similarly to the D24S8 attachment getting confused when we use only
the stencil portion, ANGLE could write to the stencil portion when
we were only using the depth. Fix this by internally disabling stencil
test and writes when using this type of configuration.
BUG=angleproject:1237
Change-Id: Ib7868d5e9f8aea73304a132005541dab947aa619
Reviewed-on: https://chromium-review.googlesource.com/314032
Tryjob-Request: Jamie Madill <jmadill@chromium.org>
Reviewed-by: Geoff Lang <geofflang@chromium.org>
Tested-by: Jamie Madill <jmadill@chromium.org>
|
|
|
0df8fe44
|
2015-11-24T16:10:24
|
|
D3D11: Don't read or write to the unused depth buffer.
When using STENCIL8, we emulate it on D3D11 with D24S8, since there
is no native stencil-only format. However in many places we would
write to the depth part of this format, and confuse the D3D runtime
when it would use the depth test. Fix this by never modifying the
depth portion of the buffer, or reading from it.
BUG=angleproject:1232
Change-Id: Ifd2e54eceae84e8deea85f439c132d07981b2286
Reviewed-on: https://chromium-review.googlesource.com/313996
Tryjob-Request: Jamie Madill <jmadill@chromium.org>
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Reviewed-by: Geoff Lang <geofflang@chromium.org>
Tested-by: Jamie Madill <jmadill@chromium.org>
|
|
|
0e4c8e73
|
2015-11-16T19:19:34
|
|
Using dirty bit notification for D3D11 rasterizer state
BUG=angleproject:1161
This is a continuation of the dirty bit refactor
Change-Id: I78ab046aaeb7de6dd288433c86e6524dcb2d6a66
Reviewed-on: https://chromium-review.googlesource.com/312837
Reviewed-by: Geoff Lang <geofflang@chromium.org>
Tryjob-Request: Dian Xiang <dianx@google.com>
Tested-by: Dian Xiang <dianx@google.com>
|
|
|
83af74f5
|
2015-11-13T16:39:53
|
|
Using dirty bit notification for D3D11 depth stencil state
BUG=angleproject:1161
This is a continuation of the dirty bit refactor
Change-Id: If7d523e6200fef631051cab0bde47fdec8385c17
Reviewed-on: https://chromium-review.googlesource.com/312722
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Tested-by: Dian Xiang <dianx@google.com>
|
|
|
451cb838
|
2015-11-11T10:45:57
|
|
Using dirty bit notification for d3d11 blend state
BUG=angleproject:1161
This is a re-land for commit ff2ab57. Dirty bits didn't check for
sampleMask, hence when sampleMask changed and nothing else, d3D
states didn't update. This commit also focuses on blend state for d3d11 to
make the cl change smaller and more manageable.
This commit also some performance regression issues that were occuring in
commit ff2ab57
Change-Id: I866066689911870a84125fa2150f9efbf1405ad0
Reviewed-on: https://chromium-review.googlesource.com/312468
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Tryjob-Request: Dian Xiang <dianx@google.com>
Tested-by: Dian Xiang <dianx@google.com>
|
|
|
63e4e724
|
2015-11-06T19:15:11
|
|
Revert "Using dirty bits notification for blend state"
Failing the dEQP-GLES3 tests:
http://build.chromium.org/p/chromium.gpu.fyi/builders/Win7%20Release%20dEQP%20%28NVIDIA%29/builds/4036/
3 tests failed:
dEQP_GLES3.Default/functional_multisample_fbo_4_samples_sample_coverage_invert (c:\b\build\slave\gpu_win_builder\build\src\third_party\angle\src\tests\deqp_support\angle_deqp_gtest.cpp:234)
dEQP_GLES3.Default/functional_multisample_fbo_8_samples_sample_coverage_invert (c:\b\build\slave\gpu_win_builder\build\src\third_party\angle\src\tests\deqp_support\angle_deqp_gtest.cpp:234)
dEQP_GLES3.Default/functional_multisample_fbo_max_samples_sample_coverage_invert (c:\b\build\slave\gpu_win_builder\build\src\third_party\angle\src\tests\deqp_support\angle_deqp_gtest.cpp:234)
BUG=angleproject:1161
This reverts commit ff2ab571b00023926ce39a91bfe20e13578d1281.
Change-Id: Ibf61fc2c64ffc31778645bfe0292516b1f107397
Reviewed-on: https://chromium-review.googlesource.com/311243
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Tested-by: Jamie Madill <jmadill@chromium.org>
|
|
|
ff2ab571
|
2015-10-05T16:48:06
|
|
Using dirty bits notification for blend state
BUG=angleproject:1161
Dirty bit notifications are used in GL for state tracking, but not D3D.
Before, D3D would use memcmp to check a change in state for every call.
This showed up as a hot spot in the perf test runs. Hence, switching D3D
over to the dirty bit system similar to GL backend will help.
Change-Id: I482edc852f1dcc888af3038ff3a61916496a02bc
Reviewed-on: https://chromium-review.googlesource.com/305295
Reviewed-by: Geoff Lang <geofflang@chromium.org>
Tryjob-Request: Dian Xiang <dianx@google.com>
Tested-by: Jamie Madill <jmadill@chromium.org>
|