Add new gpu sink for DDLs
Having this as a sink rather than a Via allows us to do more aggressive things with threads and shared contexts. Change-Id: I3ca1076686fa4f53387c12a9506e01910c1bc3e4 Reviewed-on: https://skia-review.googlesource.com/c/skia/+/272016 Reviewed-by: Greg Daniel <egdaniel@google.com> Commit-Queue: Robert Phillips <robertphillips@google.com>
This commit is contained in:
Robert Phillips
Skia Commit-Bot
parent
9d4e31d6cd
commit
291f3405f3
@ -944,6 +944,8 @@ static Sink* create_sink(const GrContextOptions& grCtxOptions, const SkCommandLi
|
||||
return new GPUPersistentCacheTestingSink(gpuConfig, grCtxOptions);
|
||||
} else if (gpuConfig->getTestPrecompile()) {
|
||||
return new GPUPrecompileTestingSink(gpuConfig, grCtxOptions);
|
||||
} else if (gpuConfig->getUseDDLSink()) {
|
||||
return new GPUDDLSink(gpuConfig, grCtxOptions);
|
||||
} else {
|
||||
return new GPUSink(gpuConfig, grCtxOptions);
|
||||
}
|
||||
|
||||
143
dm/DMSrcSink.cpp
143
dm/DMSrcSink.cpp
@ -79,6 +79,7 @@ static DEFINE_bool(RAW_threading, true, "Allow RAW decodes to run on multiple th
|
||||
DECLARE_int(gpuThreads);
|
||||
|
||||
using sk_gpu_test::GrContextFactory;
|
||||
using sk_gpu_test::ContextInfo;
|
||||
|
||||
namespace DM {
|
||||
|
||||
@ -1615,6 +1616,148 @@ Result GPUPrecompileTestingSink::draw(const Src& src, SkBitmap* dst, SkWStream*
|
||||
return compare_bitmaps(reference, *dst);
|
||||
}
|
||||
|
||||
GPUDDLSink::GPUDDLSink(const SkCommandLineConfigGpu* config, const GrContextOptions& grCtxOptions)
|
||||
: INHERITED(config, grCtxOptions)
|
||||
, fRecordingThreadPool(SkExecutor::MakeFIFOThreadPool(2))
|
||||
, fGPUThread(SkExecutor::MakeFIFOThreadPool(1)) {
|
||||
}
|
||||
|
||||
Result GPUDDLSink::ddlDraw(const Src& src,
|
||||
sk_sp<SkSurface> dstSurface,
|
||||
SkTaskGroup* recordingTaskGroup,
|
||||
SkTaskGroup* gpuTaskGroup,
|
||||
GrContext* gpuThreadCtx) const {
|
||||
auto size = src.size();
|
||||
SkPictureRecorder recorder;
|
||||
Result result = src.draw(recorder.beginRecording(SkIntToScalar(size.width()),
|
||||
SkIntToScalar(size.height())));
|
||||
if (!result.isOk()) {
|
||||
return result;
|
||||
}
|
||||
sk_sp<SkPicture> inputPicture(recorder.finishRecordingAsPicture());
|
||||
|
||||
// this is our ultimate final drawing area/rect
|
||||
SkIRect viewport = SkIRect::MakeWH(size.fWidth, size.fHeight);
|
||||
|
||||
DDLPromiseImageHelper promiseImageHelper;
|
||||
sk_sp<SkData> compressedPictureData = promiseImageHelper.deflateSKP(inputPicture.get());
|
||||
if (!compressedPictureData) {
|
||||
return Result::Fatal("GPUDDLSink: Couldn't deflate SkPicture");
|
||||
}
|
||||
|
||||
promiseImageHelper.createCallbackContexts(gpuThreadCtx);
|
||||
|
||||
// TODO: move the image upload to the utility thread
|
||||
promiseImageHelper.uploadAllToGPU(gpuTaskGroup, gpuThreadCtx);
|
||||
|
||||
constexpr int kNumDivisions = 3;
|
||||
DDLTileHelper tiles(dstSurface, viewport, kNumDivisions);
|
||||
|
||||
// Reinflate the compressed picture individually for each thread.
|
||||
tiles.createSKPPerTile(compressedPictureData.get(), promiseImageHelper);
|
||||
|
||||
tiles.kickOffThreadedWork(recordingTaskGroup, gpuTaskGroup, gpuThreadCtx);
|
||||
|
||||
// This should be the only explicit flush for the entire DDL draw
|
||||
gpuTaskGroup->add([gpuThreadCtx]() { gpuThreadCtx->flush(); });
|
||||
|
||||
// All the work is schedule we just need to wait
|
||||
recordingTaskGroup->wait(); // This should be a no-op at this point
|
||||
gpuTaskGroup->wait();
|
||||
|
||||
return Result::Ok();
|
||||
}
|
||||
|
||||
Result GPUDDLSink::draw(const Src& src, SkBitmap* dst, SkWStream* stream, SkString* log) const {
|
||||
GrContextOptions contextOptions = this->baseContextOptions();
|
||||
src.modifyGrContextOptions(&contextOptions);
|
||||
contextOptions.fPersistentCache = nullptr;
|
||||
contextOptions.fExecutor = nullptr;
|
||||
|
||||
GrContextFactory factory(contextOptions);
|
||||
|
||||
// This captures the context destined to be the main gpu context
|
||||
ContextInfo mainCtxInfo = factory.getContextInfo(this->contextType(), this->contextOverrides());
|
||||
sk_gpu_test::TestContext* mainTestCtx = mainCtxInfo.testContext();
|
||||
GrContext* mainCtx = mainCtxInfo.grContext();
|
||||
if (!mainCtx) {
|
||||
return Result::Fatal("Could not create context.");
|
||||
}
|
||||
|
||||
SkASSERT(mainCtx->priv().getGpu());
|
||||
|
||||
// TODO: make use of 'otherCtx' for uploads & compilation
|
||||
#if 0
|
||||
// This captures the context destined to be the utility context. It is in a share group
|
||||
// with the main context
|
||||
ContextInfo otherCtxInfo = factory.getSharedContextInfo(mainCtx);
|
||||
sk_gpu_test::TestContext* otherTestCtx = otherCtxInfo.testContext();
|
||||
GrContext* otherCtx = otherCtxInfo.grContext();
|
||||
if (!otherCtx) {
|
||||
return Result::Fatal("Cound not create shared context.");
|
||||
}
|
||||
|
||||
SkASSERT(otherCtx->priv().getGpu());
|
||||
#endif
|
||||
|
||||
SkTaskGroup recordingTaskGroup(*fRecordingThreadPool);
|
||||
SkTaskGroup gpuTaskGroup(*fGPUThread);
|
||||
|
||||
// Make sure 'mainCtx' is current
|
||||
mainTestCtx->makeCurrent();
|
||||
|
||||
GrBackendTexture backendTexture;
|
||||
GrBackendRenderTarget backendRT;
|
||||
sk_sp<SkSurface> surface = this->createDstSurface(mainCtx, src.size(),
|
||||
&backendTexture, &backendRT);
|
||||
if (!surface) {
|
||||
return Result::Fatal("Could not create a surface.");
|
||||
}
|
||||
|
||||
// 'mainCtx' is being shifted to the gpuThread. Leave the main thread w/o
|
||||
// a context.
|
||||
mainTestCtx->makeNotCurrent();
|
||||
|
||||
// Job one for the GPU thread is to make 'mainCtx' current!
|
||||
gpuTaskGroup.add([mainTestCtx] { mainTestCtx->makeCurrent(); });
|
||||
|
||||
Result result = this->ddlDraw(src, surface, &recordingTaskGroup, &gpuTaskGroup, mainCtx);
|
||||
|
||||
// ddlDraw schedules a flush on the gpu thread and waits so it is safe to make 'mainCtx'
|
||||
// current here.
|
||||
gpuTaskGroup.add([mainTestCtx] { mainTestCtx->makeNotCurrent(); });
|
||||
|
||||
if (!result.isOk()) {
|
||||
return result;
|
||||
}
|
||||
|
||||
mainTestCtx->makeCurrent();
|
||||
|
||||
if (FLAGS_gpuStats) {
|
||||
mainCtx->priv().dumpCacheStats(log);
|
||||
mainCtx->priv().dumpGpuStats(log);
|
||||
|
||||
#if 0
|
||||
otherCtx->priv().dumpCacheStats(log);
|
||||
otherCtx->priv().dumpGpuStats(log);
|
||||
#endif
|
||||
}
|
||||
|
||||
if (!this->readBack(surface.get(), dst)) {
|
||||
return Result::Fatal("Could not readback from surface.");
|
||||
}
|
||||
|
||||
surface.reset();
|
||||
if (backendTexture.isValid()) {
|
||||
mainCtx->deleteBackendTexture(backendTexture);
|
||||
}
|
||||
if (backendRT.isValid()) {
|
||||
mainCtx->priv().getGpu()->deleteTestingOnlyBackendRenderTarget(backendRT);
|
||||
}
|
||||
|
||||
return Result::Ok();
|
||||
}
|
||||
|
||||
/*~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~*/
|
||||
static Result draw_skdocument(const Src& src, SkDocument* doc, SkWStream* dst) {
|
||||
if (src.size().isEmpty()) {
|
||||
|
||||
@ -450,6 +450,33 @@ private:
|
||||
typedef GPUSink INHERITED;
|
||||
};
|
||||
|
||||
// This sink attempts to better simulate the Chrome DDL use-case. It:
|
||||
// creates the DDLs on separate recording threads
|
||||
// performs all the GPU work on a separate GPU thread
|
||||
// In the future this should be expanded to:
|
||||
// upload on a utility thread w/ access to a shared context
|
||||
// compile the programs on the utility thread
|
||||
// perform fine grained scheduling of gpu tasks based on their image and program prerequisites
|
||||
// create a single "compositing" DDL that is replayed last
|
||||
class GPUDDLSink : public GPUSink {
|
||||
public:
|
||||
GPUDDLSink(const SkCommandLineConfigGpu*, const GrContextOptions&);
|
||||
|
||||
Result draw(const Src&, SkBitmap*, SkWStream*, SkString*) const override;
|
||||
|
||||
private:
|
||||
Result ddlDraw(const Src&,
|
||||
sk_sp<SkSurface> dstSurface,
|
||||
SkTaskGroup* recordingTaskGroup,
|
||||
SkTaskGroup* gpuTaskGroup,
|
||||
GrContext* gpuCtx) const;
|
||||
|
||||
std::unique_ptr<SkExecutor> fRecordingThreadPool;
|
||||
std::unique_ptr<SkExecutor> fGPUThread;
|
||||
|
||||
typedef GPUSink INHERITED;
|
||||
};
|
||||
|
||||
class PDFSink : public Sink {
|
||||
public:
|
||||
PDFSink(bool pdfa, SkScalar rasterDpi) : fPDFA(pdfa), fRasterDpi(rasterDpi) {}
|
||||
|
||||
@ -71,6 +71,7 @@ static const struct {
|
||||
{ "gltestglslcache", "gpu", "api=gl,testPersistentCache=2" },
|
||||
{ "gltestprecompile", "gpu", "api=gl,testPrecompile=true" },
|
||||
{ "glestestprecompile", "gpu", "api=gles,testPrecompile=true" },
|
||||
{ "glddl", "gpu", "api=gl,useDDLSink=true" },
|
||||
{ "angle_d3d11_es2", "gpu", "api=angle_d3d11_es2" },
|
||||
{ "angle_d3d11_es3", "gpu", "api=angle_d3d11_es3" },
|
||||
{ "angle_d3d9_es2", "gpu", "api=angle_d3d9_es2" },
|
||||
@ -101,12 +102,14 @@ static const struct {
|
||||
{ "vkbetex", "gpu", "api=vulkan,surf=betex" },
|
||||
{ "vkbert", "gpu", "api=vulkan,surf=bert" },
|
||||
{ "vktestpersistentcache", "gpu", "api=vulkan,testPersistentCache=1" },
|
||||
{ "vkddl", "gpu", "api=vulkan,useDDLSink=true" },
|
||||
#endif
|
||||
#ifdef SK_METAL
|
||||
{ "mtl", "gpu", "api=metal" },
|
||||
{ "mtl1010102", "gpu", "api=metal,color=1010102" },
|
||||
{ "mtlmsaa4", "gpu", "api=metal,samples=4" },
|
||||
{ "mtlmsaa8", "gpu", "api=metal,samples=8" },
|
||||
{ "mtlddl", "gpu", "api=metal,useDDLSink=true" },
|
||||
#endif
|
||||
#ifdef SK_DIRECT3D
|
||||
{ "d3d", "gpu", "api=direct3d" },
|
||||
@ -455,6 +458,7 @@ SkCommandLineConfigGpu::SkCommandLineConfigGpu(const SkString& tag,
|
||||
bool testThreading,
|
||||
int testPersistentCache,
|
||||
bool testPrecompile,
|
||||
bool useDDLSink,
|
||||
SurfType surfType)
|
||||
: SkCommandLineConfig(tag, SkString("gpu"), viaParts)
|
||||
, fContextType(contextType)
|
||||
@ -467,6 +471,7 @@ SkCommandLineConfigGpu::SkCommandLineConfigGpu(const SkString& tag,
|
||||
, fTestThreading(testThreading)
|
||||
, fTestPersistentCache(testPersistentCache)
|
||||
, fTestPrecompile(testPrecompile)
|
||||
, fUseDDLSink(useDDLSink)
|
||||
, fSurfType(surfType) {
|
||||
if (!useStencilBuffers) {
|
||||
fContextOverrides |= ContextOverrides::kAvoidStencilBuffers;
|
||||
@ -487,6 +492,7 @@ SkCommandLineConfigGpu* parse_command_line_config_gpu(const SkString&
|
||||
bool testThreading = false;
|
||||
int testPersistentCache = 0;
|
||||
bool testPrecompile = false;
|
||||
bool useDDLs = false;
|
||||
SkCommandLineConfigGpu::SurfType surfType = SkCommandLineConfigGpu::SurfType::kDefault;
|
||||
|
||||
bool parseSucceeded = false;
|
||||
@ -504,6 +510,7 @@ SkCommandLineConfigGpu* parse_command_line_config_gpu(const SkString&
|
||||
extendedOptions.get_option_bool("testThreading", &testThreading) &&
|
||||
extendedOptions.get_option_int("testPersistentCache", &testPersistentCache) &&
|
||||
extendedOptions.get_option_bool("testPrecompile", &testPrecompile) &&
|
||||
extendedOptions.get_option_bool("useDDLs", &useDDLs) &&
|
||||
extendedOptions.get_option_gpu_surf_type("surf", &surfType);
|
||||
|
||||
// testing threading and the persistent cache are mutually exclusive.
|
||||
@ -523,6 +530,7 @@ SkCommandLineConfigGpu* parse_command_line_config_gpu(const SkString&
|
||||
testThreading,
|
||||
testPersistentCache,
|
||||
testPrecompile,
|
||||
useDDLs,
|
||||
surfType);
|
||||
}
|
||||
|
||||
|
||||
@ -63,6 +63,7 @@ public:
|
||||
bool testThreading,
|
||||
int testPersistentCache,
|
||||
bool testPrecompile,
|
||||
bool useDDLSink,
|
||||
SurfType);
|
||||
|
||||
const SkCommandLineConfigGpu* asConfigGpu() const override { return this; }
|
||||
@ -76,6 +77,7 @@ public:
|
||||
bool getTestThreading() const { return fTestThreading; }
|
||||
int getTestPersistentCache() const { return fTestPersistentCache; }
|
||||
bool getTestPrecompile() const { return fTestPrecompile; }
|
||||
bool getUseDDLSink() const { return fUseDDLSink; }
|
||||
SurfType getSurfType() const { return fSurfType; }
|
||||
|
||||
private:
|
||||
@ -89,6 +91,7 @@ private:
|
||||
bool fTestThreading;
|
||||
int fTestPersistentCache;
|
||||
bool fTestPrecompile;
|
||||
bool fUseDDLSink;
|
||||
SurfType fSurfType;
|
||||
};
|
||||
|
||||
|
||||
Loading…
Reference in New Issue
Block a user