From 8c35c8d4de621bf130865456c447815f9cb6f75c Mon Sep 17 00:00:00 2001
From: Christopher Taylor <christopher.taylor@oculus.com>
Date: Wed, 31 May 2017 01:20:17 -0700
Subject: [PATCH] Speed up decoder with new DIT-IFFT

---
 LeopardFF8.cpp      | 26 +++++++++-----------------
 tests/benchmark.cpp |  4 ++--
 2 files changed, 11 insertions(+), 19 deletions(-)

diff --git a/LeopardFF8.cpp b/LeopardFF8.cpp
index 9fc1650..78ca61a 100644
--- a/LeopardFF8.cpp
+++ b/LeopardFF8.cpp
@@ -1345,23 +1345,14 @@ void ReedSolomonDecode(
 
     // work <- IFFT(work, n, 0)
 
-    const unsigned input_count = m + original_count;
-    unsigned mip_level = 0;
-
-    for (unsigned width = 1; width < n; width <<= 1, ++mip_level)
-    {
-        const unsigned range = width << 1;
-
-        for (unsigned j = width; j < n; j += range)
-        {
-            VectorIFFTButterfly(
-                buffer_bytes,
-                width,
-                work + j - width,
-                work + j,
-                FFTSkew[j - 1]);
-        }
-    }
+    IFFT_DIT(
+        buffer_bytes,
+        nullptr,
+        n,
+        work,
+        nullptr,
+        n,
+        FFTSkew - 1);
 
     // work <- FormalDerivative(work, n)
 
@@ -1378,6 +1369,7 @@ void ReedSolomonDecode(
 
     // work <- FFT(work, n, 0) truncated to m + original_count
 
+    unsigned mip_level = LastNonzeroBit32(n);
     const unsigned output_count = m + original_count;
     for (unsigned width = (n >> 1); width > 0; width >>= 1, --mip_level)
     {
diff --git a/tests/benchmark.cpp b/tests/benchmark.cpp
index 7b05b05..aaf00d1 100644
--- a/tests/benchmark.cpp
+++ b/tests/benchmark.cpp
@@ -48,7 +48,7 @@ struct TestParameters
     unsigned original_count = 128; // under 65536
     unsigned recovery_count = 128; // under 65536 - original_count
 #endif
-    unsigned buffer_bytes = 64; // multiple of 64 bytes
+    unsigned buffer_bytes = 64000; // multiple of 64 bytes
     unsigned loss_count = 32768; // some fraction of original_count
     unsigned seed = 2;
     bool multithreaded = true;
@@ -399,7 +399,7 @@ static LEO_FORCE_INLINE void SIMDSafeFree(void* ptr)
 
 static bool BasicTest(const TestParameters& params)
 {
-    const unsigned kTrials = params.original_count > 8000 ? 1 : 100000;
+    const unsigned kTrials = params.original_count > 8000 ? 1 : 100;
 
     std::vector<uint8_t*> original_data(params.original_count);