Why does adding local variables cause a delay in the method?

Question

I recently started reading about tests and writing them for Android (in Java). I know about issues like warm-ups, garbage collector, and compiler optimizations, but I don’t know if the problem I am facing can be caused by any of them.

In my test application, I create an array of 10,000 floating-point variables and initialize it with random values. When running the control code:

private void runMinorBenchmarkFloat (float[] array) { float sum = 0; long startTime; long endTime; /* Fast warm-up */ startTime = System.nanoTime(); for(int i=0; i<SMALL_LOOP_ITERATION_COUNT; i++) for(int j=0; j<TAB_SIZE; j++) sum += array[j]; endTime = System.nanoTime() - startTime; postMessage("Warm-up for FLOAT finished in: " + endTime/1000000 + "ms.\n"); /* Main benchmark loop */ startTime = System.nanoTime(); for(int i=0; i<BIG_LOOP_ITERATION_COUNT; i++) { sum = 0; for(int j=0; j<TAB_SIZE; j++) sum += array[j]; } endTime = System.nanoTime() - startTime; postMessage("Benchmark for FLOAT finished in: " + endTime/1000000 + "ms.\n"); postMessage("Final value: " + sum + "\n\n"); } 

on my phone I get about 2 seconds for a warm up and 20 seconds for a “real” cycle.

Now, when I add two more floating variables (sum2 and sum3 - are never used inside the method):

 private void runMinorBenchmarkFloat (float[] array) { float sum = 0, sum2 = 0, sum3 = 0; // <------- the only code change here!!! long startTime; long endTime; /* Fast warm-up */ startTime = System.nanoTime(); for(int i=0; i<SMALL_LOOP_ITERATION_COUNT; i++) for(int j=0; j<TAB_SIZE; j++) sum += array[j]; endTime = System.nanoTime() - startTime; postMessage("Warm-up for FLOAT finished in: " + endTime/1000000 + "ms.\n"); /* Main benchmark loop */ startTime = System.nanoTime(); for(int i=0; i<BIG_LOOP_ITERATION_COUNT; i++) { sum = 0; for(int j=0; j<TAB_SIZE; j++) sum += array[j]; } endTime = System.nanoTime() - startTime; postMessage("Benchmark for FLOAT finished in: " + endTime/1000000 + "ms.\n"); postMessage("Final value: " + sum + "\n\n"); } 

runtime jumps from 2 seconds to warm up to 5 seconds and from 20 seconds for a real cycle to 50 seconds.

Constants:

 SMALL_LOOP_ITERATION_COUNT = 100,000 BIG_LOOP_ITERATION_COUNT = 1,000,000 

Do you think that such a difference can be caused by alignment problems (just losing the idea)?

Thanks in advance for any answers.

EDIT:

It seems that this error does not appear on every device. I can play it on Samsung Galaxy S5. The main goal of the program was to make a small benchmark. I performed four almost identical functions (runMinorBenchmark____, where _ was: int, short, float, double), which differed only in a variable of type sum. In the main control function, I called these functions. Since the indicated error occurred, I decided to combine this function into one big one. Now ... When I run the test, I have such moments: 1. 37640 ms. (for int) 2. 46728ms. (for short) 3. 60589ms. (for the float) 4. 34467ms. (for double)

I know that short means that it is slower due to type casting. I also thought that float should be slower if you double it (maybe FPU picks up casting every time (?)). But when I change the variable type for sumFloat from float to double the time for float, it is identical to double the time. I also did this “test” on another device that did not seem to suffer from this strange behavior, and the times for each test were almost the same: ~ 45000 ms. (in fact there are no visible differences).

Dalvik VM Error (?)

Answer 1

I refuse to believe that this is the cause of your problem. Of course, the compiler just clamps unused variables? Are you sure the input array is not changing, or your constants, or TAB_SIZE ?

If you're still sure, prove it by doing something like this and pasting the result here:

 public void proveIt() { float[] inputArray = new float[10000]; for (int i = 0; i < 10000; i++) { inputArray[i] = 1; } postMessage("Without declaration:"); runMinorBenchmarkFloatA(inputArray); postMessage("With declaration:"); runMinorBenchmarkFloatB(inputArray); postMessage("And again just to make sure..."); postMessage("Without declaration:"); runMinorBenchmarkFloatA(inputArray); postMessage("With declaration:"); runMinorBenchmarkFloatB(inputArray); } long TAB_SIZE = 10000; long SMALL_LOOP_ITERATION_COUNT = 100000; long BIG_LOOP_ITERATION_COUNT = 1000000; private void runMinorBenchmarkFloatA(float[] array) { float sum = 0; long startTime; long endTime; /* Fast warm-up */ startTime = System.nanoTime(); for (int i = 0; i < SMALL_LOOP_ITERATION_COUNT; i++) for (int j = 0; j < TAB_SIZE; j++) sum += array[j]; endTime = System.nanoTime() - startTime; postMessage("Warm-up for FLOAT finished in: " + endTime / 1000000 + "ms.\n"); /* Main benchmark loop */ startTime = System.nanoTime(); for (int i = 0; i < BIG_LOOP_ITERATION_COUNT; i++) { sum = 0; for (int j = 0; j < TAB_SIZE; j++) sum += array[j]; } endTime = System.nanoTime() - startTime; postMessage("Benchmark for FLOAT finished in: " + endTime / 1000000 + "ms.\n"); postMessage("Final value: " + sum + "\n\n"); } private void runMinorBenchmarkFloatB(float[] array) { float sum = 0, sum2 = 0, sum3 = 0; long startTime; long endTime; /* Fast warm-up */ startTime = System.nanoTime(); for (int i = 0; i < SMALL_LOOP_ITERATION_COUNT; i++) for (int j = 0; j < TAB_SIZE; j++) sum += array[j]; endTime = System.nanoTime() - startTime; postMessage("Warm-up for FLOAT finished in: " + endTime / 1000000 + "ms.\n"); /* Main benchmark loop */ startTime = System.nanoTime(); for (int i = 0; i < BIG_LOOP_ITERATION_COUNT; i++) { sum = 0; for (int j = 0; j < TAB_SIZE; j++) sum += array[j]; } endTime = System.nanoTime() - startTime; postMessage("Benchmark for FLOAT finished in: " + endTime / 1000000 + "ms.\n"); postMessage("Final value: " + sum + "\n\n"); } 

Answer 2

Now, when I add two more floating variables (sum2 and sum3 - are never used inside the method):

 float sum = 0, sum2 = 0, sum3 = 0; // <------- the only code change here!!! 

But sum2 and sum3 are used inside the method. They are initialized to zero. It is worth the time.

If they were not initialized, the generated bytecode would be identical with them and without them, except for the size of the allocated stack frame, which does not affect the time.