Add sk_parallel_for()

src/core/SkTaskGroup.h

Index: src/core/SkTaskGroup.h
diff --git a/src/core/SkTaskGroup.h b/src/core/SkTaskGroup.h
index 14a52c6aac1aadd1d3393466200fec43f4bf0e45..9011a15f5449e3c6776a47e8fc6e696d75b31663 100644
--- a/src/core/SkTaskGroup.h
+++ b/src/core/SkTaskGroup.h
@@ -9,6 +9,7 @@
 #define SkTaskGroup_DEFINED
 
 #include "SkTypes.h"
+#include "SkTemplates.h"
 
 struct SkRunnable;
 
@@ -48,4 +49,42 @@ private:
     /*atomic*/ int32_t fPending;
 };
 
+// Returns best estimate of number of CPU cores available to use.
+int sk_num_cores();
+
+// Call f(i) for i in [0, end).
+template <typename Func>
+void sk_parallel_for(int end, const Func& f) {
+    if (end <= 0) { return; }
+
+    struct Chunk {
+        const Func* f;
+        int start, end;
+    };
+
+    // TODO(mtklein): this chunking strategy could probably use some tuning.
+    int max_chunks  = sk_num_cores() * 2,
+        stride      = (end + max_chunks - 1 ) / max_chunks,
+        nchunks     = (end + stride - 1 ) / stride;
+    SkASSERT(nchunks <= max_chunks);
+
+    // With the chunking strategy above this won't malloc until we have a machine with >512 cores.
+    SkAutoSTMalloc<1024, Chunk> chunks(nchunks);
+
+    for (int i = 0; i < nchunks; i++) {
+        Chunk& c = chunks[i];
+        c.f     = &f;
+        c.start = i * stride;
+        c.end   = SkTMin(c.start + stride, end);
+        SkASSERT(c.start < c.end);  // Nothing will break if start >= end, but it's a wasted chunk.
+    }
+
+    void(*run_chunk)(Chunk*) = [](Chunk* c) {
+        for (int i = c->start; i < c->end; i++) {
+            (*c->f)(i);
+        }
+    };
+    SkTaskGroup().batch(run_chunk, chunks.get(), nchunks);
+}
+
 #endif//SkTaskGroup_DEFINED