Performance tests

include/gba/io_reg.h

@@ -664,6 +664,7 @@
 #define TIMER_64CLK       0x01
 #define TIMER_256CLK      0x02
 #define TIMER_1024CLK     0x03
+#define TIMER_COUNTUP     0x04
 #define TIMER_INTR_ENABLE 0x40
 #define TIMER_ENABLE      0x80
 

test/random.c

@@ -151,3 +151,60 @@ TEST("RandomElement generates a uniform distribution")
 
     EXPECT_LT(error, UQ_4_12(0.025));
 }
+
+TEST("RandomUniform mul-based faster than mod-based (compile-time)")
+{
+    u32 i;
+    struct Benchmark mulBenchmark, modBenchmark;
+    u32 mulSum = 0, modSum = 0;
+
+    BENCHMARK(&mulBenchmark)
+    {
+        mulSum += RandomUniformDefault(RNG_NONE, 0, 1);
+        mulSum += RandomUniformDefault(RNG_NONE, 0, 2);
+        mulSum += RandomUniformDefault(RNG_NONE, 0, 3);
+        mulSum += RandomUniformDefault(RNG_NONE, 0, 4);
+    }
+
+    BENCHMARK(&modBenchmark)
+    {
+        modSum += Random() % 2;
+        modSum += Random() % 3;
+        modSum += Random() % 4;
+        modSum += Random() % 5;
+    }
+
+    EXPECT_FASTER(mulBenchmark, modBenchmark);
+
+    // Reference mulSum/modSum to prevent optimization.
+    // These numbers are different because multiplication and modulus
+    // have subtly different biases (so subtle that it's irrelevant for
+    // our purposes).
+    EXPECT_EQ(mulSum, 3);
+    EXPECT_EQ(modSum, 4);
+}
+
+TEST("RandomUniform mul-based faster than mod-based (run-time)")
+{
+    u32 i;
+    struct Benchmark mulBenchmark, modBenchmark;
+    u32 mulSum = 0, modSum = 0;
+
+    BENCHMARK(&mulBenchmark)
+    {
+        for (i = 0; i < 32; i++)
+            mulSum += RandomUniformDefault(RNG_NONE, 0, i);
+    }
+
+    BENCHMARK(&modBenchmark)
+    {
+        for (i = 0; i < 32; i++)
+            modSum += Random() % (i + 1);
+    }
+
+    EXPECT_FASTER(mulBenchmark, modBenchmark);
+
+    // Reference mulSum/modSum to prevent optimization.
+    EXPECT_EQ(mulSum, 232);
+    EXPECT_EQ(modSum, 249);
+}

test/test.h

@@ -46,6 +46,7 @@ struct TestRunnerState
     u8 result;
     u8 expectedResult;
     bool8 expectLeaks:1;
+    bool8 inBenchmark:1;
     u32 timeoutSeconds;
 };
 
@@ -158,6 +159,49 @@ s32 MgbaPrintf_(const char *fmt, ...);
             Test_ExitWithResult(TEST_RESULT_FAIL, "%s:%d: EXPECT_GE(%d, %d) failed", gTestRunnerState.test->filename, __LINE__, _a, _b); \
     } while (0)
 
+struct Benchmark { s32 ticks; };
+
+static inline void BenchmarkStart(void)
+{
+    gTestRunnerState.inBenchmark = TRUE;
+    REG_TM3CNT = (TIMER_ENABLE | TIMER_64CLK) << 16;
+}
+
+static inline struct Benchmark BenchmarkStop(void)
+{
+    REG_TM3CNT_H = 0;
+    gTestRunnerState.inBenchmark = FALSE;
+    return (struct Benchmark) { REG_TM3CNT_L };
+}
+
+#define BENCHMARK(id) \
+    for (BenchmarkStart(); gTestRunnerState.inBenchmark; *(id) = BenchmarkStop())
+
+// An approximation of how much overhead benchmarks introduce.
+#define BENCHMARK_ABS 2
+
+// An approximation for what percentage faster a benchmark has to be for
+// us to be confident that it's faster than another.
+#define BENCHMARK_REL 95
+
+#define EXPECT_FASTER(a, b) \
+    do \
+    { \
+        u32 a_ = (a).ticks; u32 b_ = (b).ticks; \
+        MgbaPrintf_(#a ": %d ticks, " #b ": %d ticks", a_, b_); \
+        if (((a_ - BENCHMARK_ABS) * BENCHMARK_REL) >= (b_ * 100)) \
+            Test_ExitWithResult(TEST_RESULT_FAIL, "%s:%d: EXPECT_FASTER(" #a ", " #b ") failed", gTestRunnerState.test->filename, __LINE__); \
+    } while (0)
+
+#define EXPECT_SLOWER(a, b) \
+    do \
+    { \
+        u32 a_ = (a).ticks; u32 b_ = (b).ticks; \
+        MgbaPrintf_(#a ": %d ticks, " #b ": %d ticks", a_, b_); \
+        if ((a_ * 100) <= ((b_ - BENCHMARK_ABS) * BENCHMARK_REL)) \
+            Test_ExitWithResult(TEST_RESULT_FAIL, "%s:%d: EXPECT_SLOWER(" #a ", " #b ") failed", gTestRunnerState.test->filename, __LINE__); \
+    } while (0)
+
 #define KNOWN_FAILING \
     Test_ExpectedResult(TEST_RESULT_FAIL)