Benchmark setup

LeopardCommon.cpp

@@ -136,6 +136,13 @@ void InitializeCPUArch()
     CpuHasAVX2 = ((cpu_info[1] & CPUID_EBX_AVX2) != 0);
 #endif // LEO_TRY_AVX2
 
+#ifndef LEO_USE_SSSE3_OPT
+    CpuHasAVX2 = false;
+#endif // LEO_USE_SSSE3_OPT
+#ifndef LEO_USE_AVX2_OPT
+    CpuHasSSSE3 = false;
+#endif // LEO_USE_AVX2_OPT
+
 #endif // LEO_TARGET_MOBILE
 }
 

LeopardCommon.h

@@ -165,7 +165,11 @@
 
 // Enable 8-bit or 16-bit fields
 #define LEO_HAS_FF8
-#define LEO_HAS_FF16
+//#define LEO_HAS_FF16
+
+// Enable using SIMD instructions
+#define LEO_USE_SSSE3_OPT
+#define LEO_USE_AVX2_OPT
 
 // Avoid calculating final FFT values in decoder using bitfield
 #define LEO_ERROR_BITFIELD_OPT

LeopardFF16.cpp

@@ -32,6 +32,10 @@
 
 #include <string.h>
 
+#ifdef _MSC_VER
+    #pragma warning(disable: 4752) // found Intel(R) Advanced Vector Extensions; consider using /arch:AVX
+#endif
+
 namespace leopard { namespace ff16 {
 
 
@@ -871,6 +875,9 @@ static void IFFT_DIT(
     // If there is one layer left:
     if (dist < m)
     {
+        // Assuming that dist = m / 2
+        LEO_DEBUG_ASSERT(dist * 2 == m);
+
         const ffe_t log_m = skewLUT[dist];
 
         if (log_m == kModulus)

tests/benchmark.cpp

@@ -42,8 +42,8 @@ using namespace std;
 struct TestParameters
 {
 #ifdef LEO_HAS_FF16
-    unsigned original_count = 1000; // under 65536
-    unsigned recovery_count = 200; // under 65536 - original_count
+    unsigned original_count = 100; // under 65536
+    unsigned recovery_count = 20; // under 65536 - original_count
 #else
     unsigned original_count = 128; // under 65536
     unsigned recovery_count = 128; // under 65536 - original_count
@@ -370,7 +370,7 @@ static void ShuffleDeck16(PCGRandom &prng, uint16_t * LEO_RESTRICT deck, uint32_
 
 static bool Benchmark(const TestParameters& params)
 {
-    const unsigned kTrials = params.original_count > 8000 ? 1 : 10;
+    const unsigned kTrials = params.original_count > 8000 ? 1 : 100;
 
     std::vector<uint8_t*> original_data(params.original_count);