diff --git a/bench/BitmapBench.cpp b/bench/BitmapBench.cpp
index a9b10f466..982c917c0 100644
--- a/bench/BitmapBench.cpp
+++ b/bench/BitmapBench.cpp
@@ -363,7 +363,6 @@ DEF_BENCH( return new BitmapBench(true, SkBitmap::kIndex8_Config); )
 DEF_BENCH( return new BitmapBench(true, SkBitmap::kARGB_8888_Config, true, true); )
 DEF_BENCH( return new BitmapBench(true, SkBitmap::kARGB_8888_Config, true, false); )
 
-/* TODO(mtklein): figure out why these are broken (suspect caching is interfering) and restore
 // scale filter -> S32_opaque_D32_filter_DX_{SSE2,SSSE3} and Fact9 is also for S32_D16_filter_DX_SSE2
 DEF_BENCH( return new FilterBitmapBench(false, SkBitmap::kARGB_8888_Config, false, false, kScale_Flag | kBilerp_Flag); )
 DEF_BENCH( return new FilterBitmapBench(true, SkBitmap::kARGB_8888_Config, false, false, kScale_Flag | kBilerp_Flag); )
@@ -378,7 +377,6 @@ DEF_BENCH( return new FilterBitmapBench(true, SkBitmap::kARGB_8888_Config, true,
 
 DEF_BENCH( return new FilterBitmapBench(false, SkBitmap::kARGB_8888_Config, false, false, kScale_Flag | kBilerp_Flag | kBicubic_Flag); )
 DEF_BENCH( return new FilterBitmapBench(false, SkBitmap::kARGB_8888_Config, false, false, kScale_Flag | kRotate_Flag | kBilerp_Flag | kBicubic_Flag); )
-*/
 
 // source alpha tests -> S32A_Opaque_BlitRow32_{arm,neon}
 DEF_BENCH( return new SourceAlphaBitmapBench(SourceAlphaBitmapBench::kOpaque_SourceAlpha, SkBitmap::kARGB_8888_Config); )
diff --git a/bench/benchmain.cpp b/bench/benchmain.cpp
index b970d09bc..6b5250816 100644
--- a/bench/benchmain.cpp
+++ b/bench/benchmain.cpp
@@ -593,15 +593,14 @@ int tool_main(int argc, char** argv) {
                 }
 
                 timer.start();
-                if (NULL != canvas) {
-                    canvas->save();
-                }
-
                 // Inner loop that allows us to break the run into smaller
                 // chunks (e.g. frames). This is especially useful for the GPU
                 // as we can flush and/or swap buffers to keep the GPU from
                 // queuing up too much work.
                 for (int loopCount = loopsPerIter; loopCount > 0; ) {
+                    if (NULL != canvas) {
+                        canvas->save();
+                    }
                     if (frameIntervalComputed && loopCount > loopsPerFrame) {
                         bench->setLoops(loopsPerFrame);
                         loopCount -= loopsPerFrame;
@@ -629,11 +628,11 @@ int tool_main(int argc, char** argv) {
                         glContext->swapBuffers();
                     }
 #endif
+                    if (NULL != canvas) {
+                        canvas->restore();
+                    }
                 }
 
-                if (NULL != canvas) {
-                    canvas->restore();
-                }
 
 
                 // Stop truncated timers before GL calls complete, and stop the full timers after.