DM: make GPU tasks multithreaded again. Big refactor.

The main meat of things is in SkThreadPool.  We can now give SkThreadPool a
type for each thread to create and destroy on its local stack.  It's TLS
without going through SkTLS.

I've split the DM tasks into CpuTasks that run on threads with no TLS, and
GpuTasks that run on threads with a thread local GrContextFactory.

The old CpuTask and GpuTask have been renamed to CpuGMTask and GpuGMTask.

Upshot: default run of out/Debug/dm goes from ~45 seconds to ~20 seconds.

BUG=skia:
R=bsalomon@google.com, mtklein@google.com, reed@google.com

Author: mtklein@chromium.org

Review URL: https://codereview.chromium.org/179233005

git-svn-id: http://skia.googlecode.com/svn/trunk@13632 2bbb7eff-a529-9590-31e7-b0007b416f81

25 files changed, 224 insertions, 215 deletions

diff --git a/dm/DM.cpp b/dm/DM.cpp
index bf37d20a80..9c5c25371c 100644
--- a/dm/DM.cpp
+++ b/dm/DM.cpp
@@ -12,8 +12,8 @@
 #include "gm.h"
 
 #include "DMBenchTask.h"
-#include "DMCpuTask.h"
-#include "DMGpuTask.h"
+#include "DMCpuGMTask.h"
+#include "DMGpuGMTask.h"
 #include "DMReporter.h"
 #include "DMTask.h"
 #include "DMTaskRunner.h"
@@ -28,6 +28,7 @@ using skiatest::Test;
 using skiatest::TestRegistry;
 
 DEFINE_int32(threads, -1, "Threads for CPU work. Default NUM_CPUS.");
+DEFINE_int32(gpuThreads, 1, "Threads for GPU work.");
 DEFINE_string2(expectations, r, "",
                "If a directory, compare generated images against images under this path. "
                "If a file, compare generated images against JSON expectations at this path.");
@@ -86,15 +87,15 @@ static void kick_off_gms(const SkTDArray<GMRegistry::Factory>& gms,
     }
     for (int i = 0; i < gms.count(); i++) {
         for (int j = 0; j < configs.count(); j++) {
-            START("565",      CpuTask, kRGB_565_SkColorType);
-            START("8888",     CpuTask, kPMColor_SkColorType);
-            START("gpu",      GpuTask, native, 0);
-            START("msaa4",    GpuTask, native, 4);
-            START("msaa16",   GpuTask, native, 16);
-            START("gpunull",  GpuTask, null,   0);
-            START("gpudebug", GpuTask, debug,  0);
-            START("angle",    GpuTask, angle,  0);
-            START("mesa",     GpuTask, mesa,   0);
+            START("565",      CpuGMTask, kRGB_565_SkColorType);
+            START("8888",     CpuGMTask, kPMColor_SkColorType);
+            START("gpu",      GpuGMTask, native, 0);
+            START("msaa4",    GpuGMTask, native, 4);
+            START("msaa16",   GpuGMTask, native, 16);
+            START("gpunull",  GpuGMTask, null,   0);
+            START("gpudebug", GpuGMTask, debug,  0);
+            START("angle",    GpuGMTask, angle,  0);
+            START("mesa",     GpuGMTask, mesa,   0);
         }
     }
 #undef START
@@ -129,7 +130,12 @@ static void kick_off_tests(const SkTDArray<TestRegistry::Factory>& tests,
                            DM::Reporter* reporter,
                            DM::TaskRunner* tasks) {
     for (int i = 0; i < tests.count(); i++) {
-        tasks->add(SkNEW_ARGS(DM::TestTask, (reporter, tasks, tests[i])));
+        SkAutoTDelete<Test> test(tests[i](NULL));
+        if (test->isGPUTest()) {
+            tasks->add(SkNEW_ARGS(DM::GpuTestTask, (reporter, tasks, tests[i])));
+        } else {
+            tasks->add(SkNEW_ARGS(DM::CpuTestTask, (reporter, tasks, tests[i])));
+        }
     }
 }
 
@@ -201,7 +207,7 @@ int tool_main(int argc, char** argv) {
     SkDebugf("(%d GMs, %d benches) x %d configs, %d tests\n",
              gms.count(), benches.count(), configs.count(), tests.count());
     DM::Reporter reporter;
-    DM::TaskRunner tasks(FLAGS_threads);
+    DM::TaskRunner tasks(FLAGS_threads, FLAGS_gpuThreads);
     kick_off_gms(gms, configs, *expectations, &reporter, &tasks);
     kick_off_benches(benches, configs, &reporter, &tasks);
     kick_off_tests(tests, &reporter, &tasks);
diff --git a/dm/DMBenchTask.cpp b/dm/DMBenchTask.cpp
index 4e251de2e9..30561a407d 100644
--- a/dm/DMBenchTask.cpp
+++ b/dm/DMBenchTask.cpp
@@ -14,7 +14,7 @@ NonRenderingBenchTask::NonRenderingBenchTask(const char* config,
                                              Reporter* reporter,
                                              TaskRunner* tasks,
                                              BenchRegistry::Factory factory)
-    : Task(reporter, tasks)
+    : CpuTask(reporter, tasks)
     , fBench(factory(NULL))
     , fName(bench_name(fBench->getName(), config)) {}
 
@@ -23,7 +23,7 @@ CpuBenchTask::CpuBenchTask(const char* config,
                            TaskRunner* tasks,
                            BenchRegistry::Factory factory,
                            SkColorType colorType)
-    : Task(reporter, tasks)
+    : CpuTask(reporter, tasks)
     , fBench(factory(NULL))
     , fName(bench_name(fBench->getName(), config))
     , fColorType(colorType) {}
@@ -34,7 +34,7 @@ GpuBenchTask::GpuBenchTask(const char* config,
                            BenchRegistry::Factory factory,
                            GrContextFactory::GLContextType contextType,
                            int sampleCount)
-    : Task(reporter, tasks)
+    : GpuTask(reporter, tasks)
     , fBench(factory(NULL))
     , fName(bench_name(fBench->getName(), config))
     , fContextType(contextType)
@@ -70,13 +70,13 @@ void CpuBenchTask::draw() {
     draw_raster(fBench.get(), fColorType);
 }
 
-void GpuBenchTask::draw() {
+void GpuBenchTask::draw(GrContextFactory* grFactory) {
     SkImageInfo info = SkImageInfo::Make(fBench->getSize().x(),
                                          fBench->getSize().y(),
                                          kPMColor_SkColorType,
                                          kPremul_SkAlphaType);
     SkAutoTUnref<SkSurface> surface(SkSurface::NewRenderTarget(
-            this->getGrContextFactory()->get(fContextType), info, fSampleCount));
+            grFactory->get(fContextType), info, fSampleCount));
 
     fBench->preDraw();
     fBench->draw(1, surface->getCanvas());
diff --git a/dm/DMBenchTask.h b/dm/DMBenchTask.h
index 1e9bc9912e..ac7030b5f0 100644
--- a/dm/DMBenchTask.h
+++ b/dm/DMBenchTask.h
@@ -12,12 +12,11 @@
 
 namespace DM {
 
-class NonRenderingBenchTask : public Task {
+class NonRenderingBenchTask : public CpuTask {
 public:
     NonRenderingBenchTask(const char* config, Reporter*, TaskRunner*, BenchRegistry::Factory);
 
     virtual void draw() SK_OVERRIDE;
-    virtual bool usesGpu() const SK_OVERRIDE { return false; }
     virtual bool shouldSkip() const SK_OVERRIDE;
     virtual SkString name() const SK_OVERRIDE { return fName; }
 
@@ -26,12 +25,11 @@ private:
     const SkString fName;
 };
 
-class CpuBenchTask : public Task {
+class CpuBenchTask : public CpuTask {
 public:
     CpuBenchTask(const char* config, Reporter*, TaskRunner*, BenchRegistry::Factory, SkColorType);
 
     virtual void draw() SK_OVERRIDE;
-    virtual bool usesGpu() const SK_OVERRIDE { return false; }
     virtual bool shouldSkip() const SK_OVERRIDE;
     virtual SkString name() const SK_OVERRIDE { return fName; }
 
@@ -41,7 +39,7 @@ private:
     const SkColorType fColorType;
 };
 
-class GpuBenchTask : public Task {
+class GpuBenchTask : public GpuTask {
 public:
     GpuBenchTask(const char* config,
                  Reporter*,
@@ -50,8 +48,7 @@ public:
                  GrContextFactory::GLContextType,
                  int sampleCount);
 
-    virtual void draw() SK_OVERRIDE;
-    virtual bool usesGpu() const SK_OVERRIDE { return true; }
+    virtual void draw(GrContextFactory*) SK_OVERRIDE;
     virtual bool shouldSkip() const SK_OVERRIDE;
     virtual SkString name() const SK_OVERRIDE { return fName; }
 
diff --git a/dm/DMCpuTask.cpp b/dm/DMCpuGMTask.cpp
index acbe8d2185..6ab0014fd5 100644
--- a/dm/DMCpuTask.cpp
+++ b/dm/DMCpuGMTask.cpp
@@ -1,4 +1,4 @@
-#include "DMCpuTask.h"
+#include "DMCpuGMTask.h"
 #include "DMExpectationsTask.h"
 #include "DMPipeTask.h"
 #include "DMReplayTask.h"
@@ -9,13 +9,13 @@
 
 namespace DM {
 
-CpuTask::CpuTask(const char* config,
-                 Reporter* reporter,
-                 TaskRunner* taskRunner,
-                 const Expectations& expectations,
-                 skiagm::GMRegistry::Factory gmFactory,
-                 SkColorType colorType)
-    : Task(reporter, taskRunner)
+CpuGMTask::CpuGMTask(const char* config,
+                     Reporter* reporter,
+                     TaskRunner* taskRunner,
+                     const Expectations& expectations,
+                     skiagm::GMRegistry::Factory gmFactory,
+                     SkColorType colorType)
+    : CpuTask(reporter, taskRunner)
     , fGMFactory(gmFactory)
     , fGM(fGMFactory(NULL))
     , fName(UnderJoin(fGM->getName(), config))
@@ -23,7 +23,7 @@ CpuTask::CpuTask(const char* config,
     , fColorType(colorType)
     {}
 
-void CpuTask::draw() {
+void CpuGMTask::draw() {
     SkBitmap bitmap;
     SetupBitmap(fColorType, fGM.get(), &bitmap);
 
@@ -47,7 +47,7 @@ void CpuTask::draw() {
 #undef SPAWN
 }
 
-bool CpuTask::shouldSkip() const {
+bool CpuGMTask::shouldSkip() const {
     if (kRGB_565_SkColorType == fColorType && (fGM->getFlags() & skiagm::GM::kSkip565_Flag)) {
         return true;
     }
diff --git a/dm/DMCpuTask.h b/dm/DMCpuGMTask.h
index 0ae112f754..7712da88d0 100644
--- a/dm/DMCpuTask.h
+++ b/dm/DMCpuGMTask.h
@@ -1,5 +1,5 @@
-#ifndef DMCpuTask_DEFINED
-#define DMCpuTask_DEFINED
+#ifndef DMCpuGMTask_DEFINED
+#define DMCpuGMTask_DEFINED
 
 #include "DMExpectations.h"
 #include "DMReporter.h"
@@ -15,17 +15,16 @@
 
 namespace DM {
 
-class CpuTask : public Task {
+class CpuGMTask : public CpuTask {
 public:
-    CpuTask(const char* config,
-            Reporter*,
-            TaskRunner*,
-            const Expectations&,
-            skiagm::GMRegistry::Factory,
-            SkColorType);
+    CpuGMTask(const char* config,
+              Reporter*,
+              TaskRunner*,
+              const Expectations&,
+              skiagm::GMRegistry::Factory,
+              SkColorType);
 
     virtual void draw() SK_OVERRIDE;
-    virtual bool usesGpu() const SK_OVERRIDE { return false; }
     virtual bool shouldSkip() const SK_OVERRIDE;
     virtual SkString name() const SK_OVERRIDE { return fName; }
 
@@ -39,4 +38,4 @@ private:
 
 }  // namespace DM
 
-#endif // DMCpuTask_DEFINED
+#endif // DMCpuGMTask_DEFINED
diff --git a/dm/DMExpectationsTask.cpp b/dm/DMExpectationsTask.cpp
index cb92486269..e29257afbd 100644
--- a/dm/DMExpectationsTask.cpp
+++ b/dm/DMExpectationsTask.cpp
@@ -6,7 +6,7 @@ namespace DM {
 ExpectationsTask::ExpectationsTask(const Task& parent,
                                    const Expectations& expectations,
                                    SkBitmap bitmap)
-    : Task(parent)
+    : CpuTask(parent)
     , fName(parent.name())  // Masquerade as parent so failures are attributed to it.
     , fExpectations(expectations)
     , fBitmap(bitmap)
diff --git a/dm/DMExpectationsTask.h b/dm/DMExpectationsTask.h
index cf76fc8bdf..7000de4b68 100644
--- a/dm/DMExpectationsTask.h
+++ b/dm/DMExpectationsTask.h
@@ -10,12 +10,11 @@ namespace DM {
 
 // ExpectationsTask compares an SkBitmap against some Expectations.
 // Moving this off the GPU threadpool is a nice (~30%) runtime win.
-class ExpectationsTask : public Task {
+class ExpectationsTask : public CpuTask {
 public:
     ExpectationsTask(const Task& parent, const Expectations&, SkBitmap);
 
     virtual void draw() SK_OVERRIDE;
-    virtual bool usesGpu() const SK_OVERRIDE { return false; }
     virtual bool shouldSkip() const SK_OVERRIDE { return false; }
     virtual SkString name() const SK_OVERRIDE { return fName; }
 
diff --git a/dm/DMGpuTask.cpp b/dm/DMGpuGMTask.cpp
index c285d88fd7..cffa2291c5 100644
--- a/dm/DMGpuTask.cpp
+++ b/dm/DMGpuGMTask.cpp
@@ -1,4 +1,4 @@
-#include "DMGpuTask.h"
+#include "DMGpuGMTask.h"
 
 #include "DMExpectationsTask.h"
 #include "DMUtil.h"
@@ -9,14 +9,14 @@
 
 namespace DM {
 
-GpuTask::GpuTask(const char* config,
-                 Reporter* reporter,
-                 TaskRunner* taskRunner,
-                 const Expectations& expectations,
-                 skiagm::GMRegistry::Factory gmFactory,
-                 GrContextFactory::GLContextType contextType,
-                 int sampleCount)
-    : Task(reporter, taskRunner)
+GpuGMTask::GpuGMTask(const char* config,
+                     Reporter* reporter,
+                     TaskRunner* taskRunner,
+                     const Expectations& expectations,
+                     skiagm::GMRegistry::Factory gmFactory,
+                     GrContextFactory::GLContextType contextType,
+                     int sampleCount)
+    : GpuTask(reporter, taskRunner)
     , fGM(gmFactory(NULL))
     , fName(UnderJoin(fGM->getName(), config))
     , fExpectations(expectations)
@@ -24,13 +24,13 @@ GpuTask::GpuTask(const char* config,
     , fSampleCount(sampleCount)
     {}
 
-void GpuTask::draw() {
+void GpuGMTask::draw(GrContextFactory* grFactory) {
     SkImageInfo info = SkImageInfo::Make(SkScalarCeilToInt(fGM->width()),
                                          SkScalarCeilToInt(fGM->height()),
                                          kPMColor_SkColorType,
                                          kPremul_SkAlphaType);
     SkAutoTUnref<SkSurface> surface(SkSurface::NewRenderTarget(
-            this->getGrContextFactory()->get(fContextType), info, fSampleCount));
+            grFactory->get(fContextType), info, fSampleCount));
     SkCanvas* canvas = surface->getCanvas();
 
     canvas->concat(fGM->getInitialTransform());
@@ -49,7 +49,7 @@ void GpuTask::draw() {
     this->spawnChild(SkNEW_ARGS(WriteTask, (*this, bitmap)));
 }
 
-bool GpuTask::shouldSkip() const {
+bool GpuGMTask::shouldSkip() const {
     return SkToBool(fGM->getFlags() & skiagm::GM::kSkipGPU_Flag);
 }
 
diff --git a/dm/DMGpuTask.h b/dm/DMGpuGMTask.h
index f74950c088..141994e6d3 100644
--- a/dm/DMGpuTask.h
+++ b/dm/DMGpuGMTask.h
@@ -1,5 +1,5 @@
-#ifndef DMGpuTask_DEFINED
-#define DMGpuTask_DEFINED
+#ifndef DMGpuGMTask_DEFINED
+#define DMGpuGMTask_DEFINED
 
 #include "DMExpectations.h"
 #include "DMReporter.h"
@@ -15,18 +15,17 @@
 
 namespace DM {
 
-class GpuTask : public Task {
+class GpuGMTask : public GpuTask {
 public:
-    GpuTask(const char* config,
-            Reporter*,
-            TaskRunner*,
-            const Expectations&,
-            skiagm::GMRegistry::Factory,
-            GrContextFactory::GLContextType,
-            int sampleCount);
-
-    virtual void draw() SK_OVERRIDE;
-    virtual bool usesGpu() const SK_OVERRIDE { return true; }
+    GpuGMTask(const char* config,
+              Reporter*,
+              TaskRunner*,
+              const Expectations&,
+              skiagm::GMRegistry::Factory,
+              GrContextFactory::GLContextType,
+              int sampleCount);
+
+    virtual void draw(GrContextFactory*) SK_OVERRIDE;
     virtual bool shouldSkip() const SK_OVERRIDE;
     virtual SkString name() const SK_OVERRIDE { return fName; }
 
@@ -40,4 +39,4 @@ private:
 
 }  // namespace DM
 
-#endif  // DMGpuTask_DEFINED
+#endif  // DMGpuGMTask_DEFINED
diff --git a/dm/DMPipeTask.cpp b/dm/DMPipeTask.cpp
index 163f1e64db..513594e0ea 100644
--- a/dm/DMPipeTask.cpp
+++ b/dm/DMPipeTask.cpp
@@ -38,7 +38,7 @@ PipeTask::PipeTask(const Task& parent,
                    SkBitmap reference,
                    bool crossProcess,
                    bool sharedAddressSpace)
-    : Task(parent)
+    : CpuTask(parent)
     , fFlags(get_flags(crossProcess, sharedAddressSpace))
     , fName(UnderJoin(parent.name().c_str(), get_name(fFlags)))
     , fGM(gm)
diff --git a/dm/DMPipeTask.h b/dm/DMPipeTask.h
index 23bbfef114..c251d08328 100644
--- a/dm/DMPipeTask.h
+++ b/dm/DMPipeTask.h
@@ -11,7 +11,7 @@
 
 namespace DM {
 
-class PipeTask : public Task {
+class PipeTask : public CpuTask {
 
 public:
     PipeTask(const Task& parent,        // PipeTask must be a child task.  Pass its parent here.
@@ -21,7 +21,6 @@ public:
              bool sharedAddressSpace);  // If cross process, should it assume shared address space?
 
     virtual void draw() SK_OVERRIDE;
-    virtual bool usesGpu() const SK_OVERRIDE { return false; }
     virtual bool shouldSkip() const SK_OVERRIDE;
     virtual SkString name() const SK_OVERRIDE { return fName; }
 
diff --git a/dm/DMReplayTask.cpp b/dm/DMReplayTask.cpp
index 3d010358c2..c915be20d2 100644
--- a/dm/DMReplayTask.cpp
+++ b/dm/DMReplayTask.cpp
@@ -14,7 +14,7 @@ ReplayTask::ReplayTask(const Task& parent,
                        skiagm::GM* gm,
                        SkBitmap reference,
                        bool useRTree)
-    : Task(parent)
+    : CpuTask(parent)
     , fName(UnderJoin(parent.name().c_str(), useRTree ? "rtree" : "replay"))
     , fGM(gm)
     , fReference(reference)
diff --git a/dm/DMReplayTask.h b/dm/DMReplayTask.h
index 1245009923..78bef0dfd8 100644
--- a/dm/DMReplayTask.h
+++ b/dm/DMReplayTask.h
@@ -11,7 +11,7 @@
 
 namespace DM {
 
-class ReplayTask : public Task {
+class ReplayTask : public CpuTask {
 
 public:
     ReplayTask(const Task& parent,  // ReplayTask must be a child task.  Pass its parent here.
@@ -20,7 +20,6 @@ public:
                bool useRTree);      // Record with an RTree?
 
     virtual void draw() SK_OVERRIDE;
-    virtual bool usesGpu() const SK_OVERRIDE { return false; }
     virtual bool shouldSkip() const SK_OVERRIDE;
     virtual SkString name() const SK_OVERRIDE { return fName; }
 
diff --git a/dm/DMSerializeTask.cpp b/dm/DMSerializeTask.cpp
index 7e45d49be5..4f55de57d5 100644
--- a/dm/DMSerializeTask.cpp
+++ b/dm/DMSerializeTask.cpp
@@ -13,7 +13,7 @@ namespace DM {
 SerializeTask::SerializeTask(const Task& parent,
                              skiagm::GM* gm,
                              SkBitmap reference)
-    : Task(parent)
+    : CpuTask(parent)
     , fName(UnderJoin(parent.name().c_str(), "serialize"))
     , fGM(gm)
     , fReference(reference)
diff --git a/dm/DMSerializeTask.h b/dm/DMSerializeTask.h
index 30a4303b06..1f8b83632b 100644
--- a/dm/DMSerializeTask.h
+++ b/dm/DMSerializeTask.h
@@ -11,7 +11,7 @@
 
 namespace DM {
 
-class SerializeTask : public Task {
+class SerializeTask : public CpuTask {
 
 public:
     SerializeTask(const Task& parent,
@@ -19,7 +19,6 @@ public:
                   SkBitmap reference);
 
     virtual void draw() SK_OVERRIDE;
-    virtual bool usesGpu() const SK_OVERRIDE { return false; }
     virtual bool shouldSkip() const SK_OVERRIDE;
     virtual SkString name() const SK_OVERRIDE { return fName; }
 
diff --git a/dm/DMTask.cpp b/dm/DMTask.cpp
index d26971c890..1c4cc25693 100644
--- a/dm/DMTask.cpp
+++ b/dm/DMTask.cpp
@@ -1,53 +1,59 @@
 #include "DMTask.h"
-
 #include "DMTaskRunner.h"
-#include "DMUtil.h"
-#include "SkBitmap.h"
-#include "SkCommandLineFlags.h"
 
 namespace DM {
 
 Task::Task(Reporter* reporter, TaskRunner* taskRunner)
-    : fReporter(reporter), fTaskRunner(taskRunner), fDepth(0) {
+    : fReporter(reporter)
+    , fTaskRunner(taskRunner)
+    , fDepth(0) {
     fReporter->start();
 }
 
 Task::Task(const Task& parent)
-    : INHERITED(parent)
-    , fReporter(parent.fReporter)
+    : fReporter(parent.fReporter)
     , fTaskRunner(parent.fTaskRunner)
-    , fDepth(parent.depth()+1) {
+    , fDepth(parent.depth() + 1) {
     fReporter->start();
 }
 
-Task::~Task() {}
-
-void Task::run() {
-    if (!this->shouldSkip()) {
-        this->draw();
+void Task::fail(const char* msg) {
+    SkString failure(this->name());
+    if (msg) {
+        failure.appendf(": %s", msg);
     }
+    fReporter->fail(failure);
+}
+
+void Task::finish() {
     fReporter->finish(this->name());
-    delete this;
 }
 
-void Task::spawnChild(Task* task) {
-    if (!task->usesGpu()) {
-        fTaskRunner->add(task);
-    } else {
-        SkDEBUGFAIL("Sorry, we can't spawn GPU tasks. :(  See comment in TaskRunner::wait().");
-    }
+void Task::spawnChild(CpuTask* task) {
+    fTaskRunner->add(task);
 }
 
-void Task::fail(const char* msg) {
-    SkString failure(this->name());
-    if (msg) {
-        failure.appendf(": %s", msg);
+CpuTask::CpuTask(Reporter* reporter, TaskRunner* taskRunner) : Task(reporter, taskRunner) {}
+CpuTask::CpuTask(const Task& parent) : Task(parent) {}
+
+void CpuTask::run() {
+    if (!this->shouldSkip()) {
+        this->draw();
     }
-    fReporter->fail(failure);
+    this->finish();
+    SkDELETE(this);
 }
 
-GrContextFactory* Task::getGrContextFactory() const {
-    return fTaskRunner->getGrContextFactory();
+GpuTask::GpuTask(Reporter* reporter, TaskRunner* taskRunner) : Task(reporter, taskRunner) {}
+
+void GpuTask::run(GrContextFactory& factory) {
+    if (!this->shouldSkip()) {
+        this->draw(&factory);
+    }
+    this->finish();
+    SkDELETE(this);
 }
 
+
+
 }  // namespace DM
diff --git a/dm/DMTask.h b/dm/DMTask.h
index e8598df85e..cad8234c05 100644
--- a/dm/DMTask.h
+++ b/dm/DMTask.h
@@ -4,28 +4,21 @@
 #include "DMReporter.h"
 #include "GrContextFactory.h"
 #include "SkRunnable.h"
-#include "SkThreadPool.h"
 
-// DM will run() these tasks on one of two threadpools, depending on the result
-// of usesGpu().  The subclasses can call fail() to mark this task as failed,
-// or make any number of spawnChild() calls to kick off dependent tasks.
+// DM will run() these tasks on one of two threadpools.
+// Subclasses can call fail() to mark this task as failed, or make any number of spawnChild() calls
+// to kick off dependent tasks.
 //
-// Task deletes itself when run.
+// Tasks delete themselves when run.
 
 namespace DM {
 
 class TaskRunner;
 
-class Task : public SkRunnable {
-public:
-    Task(Reporter* reporter, TaskRunner* taskRunner);
-    Task(const Task& parent);
-    virtual ~Task();
-
-    void run() SK_OVERRIDE;
+class CpuTask;
 
-    virtual void draw() = 0;
-    virtual bool usesGpu() const = 0;
+class Task {
+public:
     virtual bool shouldSkip() const = 0;
     virtual SkString name() const = 0;
 
@@ -34,19 +27,37 @@ public:
     int depth() const { return fDepth; }
 
 protected:
-    void spawnChild(Task* task);
-    void fail(const char* msg = NULL);
+    Task(Reporter* reporter, TaskRunner* taskRunner);
+    Task(const Task& parent);
+    virtual ~Task() {}
 
-    // This can only be safely called from a GPU task's draw() method.
-    GrContextFactory* getGrContextFactory() const;
+    void fail(const char* msg = NULL);
+    void finish();
+    void spawnChild(CpuTask* task);  // For now we don't allow GPU child tasks.
 
 private:
-    // Both unowned.
-    Reporter* fReporter;
-    TaskRunner* fTaskRunner;
+    Reporter* fReporter;      // Unowned.
+    TaskRunner* fTaskRunner;  // Unowned.
     int fDepth;
+};
+
+class CpuTask : public Task, public SkRunnable {
+public:
+    CpuTask(Reporter* reporter, TaskRunner* taskRunner);
+    CpuTask(const Task& parent);
+    virtual ~CpuTask() {}
+
+    void run() SK_OVERRIDE;
+    virtual void draw() = 0;
+};
+
+class GpuTask : public Task, public SkTRunnable<GrContextFactory> {
+ public:
+    GpuTask(Reporter* reporter, TaskRunner* taskRunner);
+    virtual ~GpuTask() {}
 
-    typedef SkRunnable INHERITED;
+    void run(GrContextFactory&) SK_OVERRIDE;
+    virtual void draw(GrContextFactory*) = 0;
 };
 
 }  // namespace DM
diff --git a/dm/DMTaskRunner.cpp b/dm/DMTaskRunner.cpp
index bd53ce615a..e0bd977288 100644
--- a/dm/DMTaskRunner.cpp
+++ b/dm/DMTaskRunner.cpp
@@ -3,48 +3,19 @@
 
 namespace DM {
 
+TaskRunner::TaskRunner(int cpuThreads, int gpuThreads) : fCpu(cpuThreads), fGpu(gpuThreads) {}
 
-TaskRunner::TaskRunner(int cputhreads)
-    : fMain(cputhreads)
-    , fGpu(1) {
-    // Enqueue a task on the GPU thread to create a GrContextFactory.
-    struct Create : public SkRunnable {
-        Create(GrContextFactory** ptr) : fPtr(ptr) {}
-        void run() SK_OVERRIDE {
-            *fPtr = SkNEW(GrContextFactory);
-            delete this;
-        }
-        GrContextFactory** fPtr;
-    };
-    fGpu.add(SkNEW_ARGS(Create, (&fGrContextFactory)));
-}
+void TaskRunner::add(CpuTask* task) { fCpu.add(task); }
 
-void TaskRunner::add(Task* task) {
-    if (task->usesGpu()) {
-        fGpu.add(task);
-    } else {
-        fMain.add(task);
-    }
-}
+void TaskRunner::add(GpuTask* task) { fGpu.add(task); }
 
 void TaskRunner::wait() {
-    // Enqueue a task on the GPU thread to destroy the GrContextFactory.
-    struct Delete : public SkRunnable {
-        Delete(GrContextFactory* ptr) : fPtr(ptr) {}
-        void run() SK_OVERRIDE {
-            delete fPtr;
-            delete this;
-        }
-        GrContextFactory* fPtr;
-    };
-    fGpu.add(SkNEW_ARGS(Delete, (fGrContextFactory)));
-
-    // These wait calls block until the threadpool is done.  We don't allow
-    // children to spawn new GPU tasks so we can wait for that first knowing
-    // we'll never try to add to it later.  Same can't be said of fMain: fGpu
-    // and fMain can both add tasks to fMain, so we have to wait for that last.
+    // These wait calls block until each threadpool is done.  We don't allow
+    // spawning new child GPU tasks, so we can wait for that first knowing
+    // we'll never try to add to it later.  Same can't be said of the CPU pool:
+    // both CPU and GPU tasks can spawn off new CPU work, so we wait for that last.
     fGpu.wait();
-    fMain.wait();
+    fCpu.wait();
 }
 
 }  // namespace DM
diff --git a/dm/DMTaskRunner.h b/dm/DMTaskRunner.h
index 8af1b63719..c7b40588e9 100644
--- a/dm/DMTaskRunner.h
+++ b/dm/DMTaskRunner.h
@@ -5,26 +5,25 @@
 #include "SkThreadPool.h"
 #include "SkTypes.h"
 
-// TaskRunner runs Tasks on one of two threadpools depending on the Task's usesGpu() method.  This
-// lets us drive the GPU from a single thread while parallelizing CPU-bound work.
+// TaskRunner runs Tasks on one of two threadpools depending on the need for a GrContextFactory.
+// It's typically a good idea to run fewer GPU threads than CPU threads (go nuts with those).
 
 namespace DM {
 
-class Task;
+class CpuTask;
+class GpuTask;
 
 class TaskRunner : SkNoncopyable {
 public:
-    explicit TaskRunner(int cputhreads);
+    explicit TaskRunner(int cpuThreads, int gpuThreads);
 
-    void add(Task* task);
+    void add(CpuTask* task);
+    void add(GpuTask* task);
     void wait();
 
-    // This can only be safely called from a GPU task's draw() method.
-    GrContextFactory* getGrContextFactory() const { return fGrContextFactory; }
-
 private:
-    SkThreadPool fMain, fGpu;
-    GrContextFactory* fGrContextFactory;  // Created and destroyed on fGpu threadpool.
+    SkTThreadPool<void> fCpu;
+    SkTThreadPool<GrContextFactory> fGpu;
 };
 
 }  // namespace DM
diff --git a/dm/DMTestTask.cpp b/dm/DMTestTask.cpp
index 32a698c673..6c3fcedf54 100644
--- a/dm/DMTestTask.cpp
+++ b/dm/DMTestTask.cpp
@@ -8,23 +8,32 @@ DEFINE_bool2(pathOpsVerbose,      V, false, "Tell pathOps tests to be verbose.")
 
 namespace DM {
 
+bool TestReporter::allowExtendedTest() const { return FLAGS_pathOpsExtended; }
+bool TestReporter::allowThreaded()     const { return !FLAGS_pathOpsSingleThread; }
+bool TestReporter::verbose()           const { return FLAGS_pathOpsVerbose; }
+
 static SkString test_name(const char* name) {
     SkString result("test ");
     result.append(name);
     return result;
 }
 
-TestTask::TestTask(Reporter* reporter,
-                   TaskRunner* taskRunner,
-                   skiatest::TestRegistry::Factory factory)
-    : Task(reporter, taskRunner)
+CpuTestTask::CpuTestTask(Reporter* reporter,
+                         TaskRunner* taskRunner,
+                         skiatest::TestRegistry::Factory factory)
+    : CpuTask(reporter, taskRunner)
     , fTest(factory(NULL))
     , fName(test_name(fTest->getName())) {}
 
-void TestTask::draw() {
-    if (this->usesGpu()) {
-        fTest->setGrContextFactory(this->getGrContextFactory());
-    }
+GpuTestTask::GpuTestTask(Reporter* reporter,
+                         TaskRunner* taskRunner,
+                         skiatest::TestRegistry::Factory factory)
+    : GpuTask(reporter, taskRunner)
+    , fTest(factory(NULL))
+    , fName(test_name(fTest->getName())) {}
+
+
+void CpuTestTask::draw() {
     fTest->setReporter(&fTestReporter);
     fTest->run();
     if (!fTest->passed()) {
@@ -32,8 +41,13 @@ void TestTask::draw() {
     }
 }
 
-bool TestTask::TestReporter::allowExtendedTest() const { return FLAGS_pathOpsExtended; }
-bool TestTask::TestReporter::allowThreaded()     const { return !FLAGS_pathOpsSingleThread; }
-bool TestTask::TestReporter::verbose()           const { return FLAGS_pathOpsVerbose; }
+void GpuTestTask::draw(GrContextFactory* grFactory) {
+    fTest->setGrContextFactory(grFactory);
+    fTest->setReporter(&fTestReporter);
+    fTest->run();
+    if (!fTest->passed()) {
+        this->fail(fTestReporter.failure());
+    }
+}
 
 }  // namespace DM
diff --git a/dm/DMTestTask.h b/dm/DMTestTask.h
index 49a8e773b6..87f59209b2 100644
--- a/dm/DMTestTask.h
+++ b/dm/DMTestTask.h
@@ -11,34 +11,47 @@
 // Runs a unit test.
 namespace DM {
 
-class TestTask : public Task {
+class TestReporter : public skiatest::Reporter {
 public:
-    TestTask(Reporter*, TaskRunner*, skiatest::TestRegistry::Factory);
+  TestReporter() {}
+
+  const char* failure() const { return fFailure.c_str(); }
+
+private:
+  virtual bool allowExtendedTest() const SK_OVERRIDE;
+  virtual bool allowThreaded()     const SK_OVERRIDE;
+  virtual bool verbose()           const SK_OVERRIDE;
+
+  virtual void onReportFailed(const SkString& desc) SK_OVERRIDE {
+      fFailure = desc;
+  }
+
+  SkString fFailure;
+};
+
+class CpuTestTask : public CpuTask {
+public:
+    CpuTestTask(Reporter*, TaskRunner*, skiatest::TestRegistry::Factory);
 
     virtual void draw() SK_OVERRIDE;
-    virtual bool usesGpu() const SK_OVERRIDE { return fTest->isGPUTest(); }
     virtual bool shouldSkip() const SK_OVERRIDE { return false; }
     virtual SkString name() const SK_OVERRIDE { return fName; }
 
 private:
-    class TestReporter : public skiatest::Reporter {
-    public:
-      TestReporter() {}
-
-      const char* failure() const { return fFailure.c_str(); }
-
-    private:
-      virtual bool allowExtendedTest() const SK_OVERRIDE;
-      virtual bool allowThreaded()     const SK_OVERRIDE;
-      virtual bool verbose()           const SK_OVERRIDE;
+    TestReporter fTestReporter;
+    SkAutoTDelete<skiatest::Test> fTest;
+    const SkString fName;
+};
 
-      virtual void onReportFailed(const SkString& desc) SK_OVERRIDE {
-          fFailure = desc;
-      }
+class GpuTestTask : public GpuTask {
+public:
+    GpuTestTask(Reporter*, TaskRunner*, skiatest::TestRegistry::Factory);
 
-      SkString fFailure;
-    };
+    virtual void draw(GrContextFactory*) SK_OVERRIDE;
+    virtual bool shouldSkip() const SK_OVERRIDE { return false; }
+    virtual SkString name() const SK_OVERRIDE { return fName; }
 
+private:
     TestReporter fTestReporter;
     SkAutoTDelete<skiatest::Test> fTest;
     const SkString fName;
diff --git a/dm/DMTileGridTask.cpp b/dm/DMTileGridTask.cpp
index beffbb0371..f9cac07de3 100644
--- a/dm/DMTileGridTask.cpp
+++ b/dm/DMTileGridTask.cpp
@@ -12,7 +12,7 @@ DEFINE_bool(tileGrid, false, "If true, run picture replay tests with a tile grid
 namespace DM {
 
 TileGridTask::TileGridTask(const Task& parent, skiagm::GM* gm, SkBitmap reference, SkISize tileSize)
-    : Task(parent)
+    : CpuTask(parent)
     , fName(UnderJoin(parent.name().c_str(), "tilegrid"))
     , fGM(gm)
     , fReference(reference)
diff --git a/dm/DMTileGridTask.h b/dm/DMTileGridTask.h
index 4a522b956c..911a1c52a1 100644
--- a/dm/DMTileGridTask.h
+++ b/dm/DMTileGridTask.h
@@ -11,7 +11,7 @@
 
 namespace DM {
 
-class TileGridTask : public Task {
+class TileGridTask : public CpuTask {
 
 public:
     TileGridTask(const Task& parent,  // TileGridTask must be a child task.  Pass its parent here.
@@ -20,7 +20,6 @@ public:
                  SkISize tileSize);   // Tile size to use.
 
     virtual void draw() SK_OVERRIDE;
-    virtual bool usesGpu() const SK_OVERRIDE { return false; }
     virtual bool shouldSkip() const SK_OVERRIDE;
     virtual SkString name() const SK_OVERRIDE { return fName; }
 
diff --git a/dm/DMWriteTask.cpp b/dm/DMWriteTask.cpp
index 5adb1d0437..e30cbdbf85 100644
--- a/dm/DMWriteTask.cpp
+++ b/dm/DMWriteTask.cpp
@@ -26,7 +26,7 @@ static int split_suffixes(int N, const char* name, SkTArray<SkString>* out) {
     return consumed;
 }
 
-WriteTask::WriteTask(const Task& parent, SkBitmap bitmap) : Task(parent), fBitmap(bitmap) {
+WriteTask::WriteTask(const Task& parent, SkBitmap bitmap) : CpuTask(parent), fBitmap(bitmap) {
     const int suffixes = parent.depth() + 1;
     const SkString& name = parent.name();
     const int totalSuffixLength = split_suffixes(suffixes, name.c_str(), &fSuffixes);
diff --git a/dm/DMWriteTask.h b/dm/DMWriteTask.h
index 49a5c746a6..839abd7ef1 100644
--- a/dm/DMWriteTask.h
+++ b/dm/DMWriteTask.h
@@ -12,14 +12,13 @@
 
 namespace DM {
 
-class WriteTask : public Task {
+class WriteTask : public CpuTask {
 
 public:
     WriteTask(const Task& parent,  // WriteTask must be a child Task.  Pass its parent here.
               SkBitmap bitmap);    // Bitmap to write.
 
     virtual void draw() SK_OVERRIDE;
-    virtual bool usesGpu() const SK_OVERRIDE { return false; }
     virtual bool shouldSkip() const SK_OVERRIDE;
     virtual SkString name() const SK_OVERRIDE;