Move the div operation out of the SIMD helper

benchmarks/BM_meanSquared.cpp

@@ -34,7 +34,7 @@ BENCHMARK_DEFINE_F(MeanSquaredArray, Scalar)
 (benchmark::State& state)
 {
     for (auto _ : state) {
-        sfz::setSIMDOpStatus<float>(sfz::SIMDOps::meanSquared, false);
+        sfz::setSIMDOpStatus<float>(sfz::SIMDOps::sumSquares, false);
         auto result = sfz::meanSquared<float>(input);
         benchmark::DoNotOptimize(result);
     }
@@ -44,7 +44,7 @@ BENCHMARK_DEFINE_F(MeanSquaredArray, SIMD)
 (benchmark::State& state)
 {
     for (auto _ : state) {
-        sfz::setSIMDOpStatus<float>(sfz::SIMDOps::meanSquared, true);
+        sfz::setSIMDOpStatus<float>(sfz::SIMDOps::sumSquares, true);
         auto result = sfz::meanSquared<float>(input);
         benchmark::DoNotOptimize(result);
     }
@@ -54,7 +54,7 @@ BENCHMARK_DEFINE_F(MeanSquaredArray, Scalar_Unaligned)
 (benchmark::State& state)
 {
     for (auto _ : state) {
-        sfz::setSIMDOpStatus<float>(sfz::SIMDOps::meanSquared, false);
+        sfz::setSIMDOpStatus<float>(sfz::SIMDOps::sumSquares, false);
         auto result = sfz::meanSquared<float>(absl::MakeSpan(input).subspan(1));
         benchmark::DoNotOptimize(result);
     }
@@ -64,7 +64,7 @@ BENCHMARK_DEFINE_F(MeanSquaredArray, SIMD_Unaligned)
 (benchmark::State& state)
 {
     for (auto _ : state) {
-        sfz::setSIMDOpStatus<float>(sfz::SIMDOps::meanSquared, true);
+        sfz::setSIMDOpStatus<float>(sfz::SIMDOps::sumSquares, true);
         auto result = sfz::meanSquared<float>(absl::MakeSpan(input).subspan(1));
         benchmark::DoNotOptimize(result);
     }

src/sfizz/SIMDHelpers.cpp

@@ -39,7 +39,7 @@ struct SIMDDispatch {
     decltype(&cumsumScalar<T>) cumsum = &cumsumScalar<T>;
     decltype(&diffScalar<T>) diff = &diffScalar<T>;
     decltype(&meanScalar<T>) mean = &meanScalar<T>;
-    decltype(&meanSquaredScalar<T>) meanSquared = &meanSquaredScalar<T>;
+    decltype(&sumSquaresScalar<T>) sumSquares = &sumSquaresScalar<T>;
     decltype(&clampAllScalar<T>) clampAll = &clampAllScalar<T>;
     decltype(&allWithinScalar<T>) allWithin = &allWithinScalar<T>;
 
@@ -85,7 +85,7 @@ void SIMDDispatch<float>::setStatus(SIMDOps op, bool enable)
             SIMD_OP(cumsum)
             SIMD_OP(diff)
             SIMD_OP(mean)
-            SIMD_OP(meanSquared)
+            SIMD_OP(sumSquares)
             SIMD_OP(clampAll)
             SIMD_OP(allWithin)
         }
@@ -122,7 +122,7 @@ void SIMDDispatch<float>::setStatus(SIMDOps op, bool enable)
             SIMD_OP(cumsum)
             SIMD_OP(diff)
             SIMD_OP(mean)
-            SIMD_OP(meanSquared)
+            SIMD_OP(sumSquares)
             SIMD_OP(clampAll)
             SIMD_OP(allWithin)
         }
@@ -163,7 +163,7 @@ void SIMDDispatch<float>::resetStatus()
     setStatus(SIMDOps::diff, false);
     setStatus(SIMDOps::sfzInterpolationCast, true);
     setStatus(SIMDOps::mean, false);
-    setStatus(SIMDOps::meanSquared, false);
+    setStatus(SIMDOps::sumSquares, false);
     setStatus(SIMDOps::upsampling, true);
     setStatus(SIMDOps::clampAll, false);
     setStatus(SIMDOps::allWithin, true);
@@ -292,9 +292,9 @@ float mean<float>(const float* vector, unsigned size) noexcept
 }
 
 template <>
-float meanSquared<float>(const float* vector, unsigned size) noexcept
+float sumSquares<float>(const float* vector, unsigned size) noexcept
 {
-    return simdDispatch<float>().meanSquared(vector, size);
+    return simdDispatch<float>().sumSquares(vector, size);
 }
 
 template <>

src/sfizz/SIMDHelpers.h

@@ -57,7 +57,7 @@ enum class SIMDOps {
     diff,
     sfzInterpolationCast,
     mean,
-    meanSquared,
+    sumSquares,
     upsampling,
     clampAll,
     allWithin,
@@ -515,7 +515,7 @@ T mean(absl::Span<const T> vector) noexcept
 }
 
 /**
- * @brief Computes the mean squared of a span
+ * @brief Computes the sum of squares of a span
  *
  * @tparam T the underlying type
  * @tparam SIMD use the SIMD version or the scalar version
@@ -523,13 +523,33 @@ T mean(absl::Span<const T> vector) noexcept
  * @return T
  */
 template <class T>
-T meanSquared(const T* vector, unsigned size) noexcept
+T sumSquares(const T* vector, unsigned size) noexcept
 {
-    meanSquaredScalar(vector, size);
+    return sumSquaresScalar(vector, size);
 }
 
 template <>
-float meanSquared<float>(const float* vector, unsigned size) noexcept;
+float sumSquares<float>(const float* vector, unsigned size) noexcept;
+
+template <class T>
+T sumSquares(absl::Span<const T> vector) noexcept
+{
+    return sumSquares(vector.data(), vector.size());
+}
+
+/**
+ * @brief Computes the mean squared of a span
+ *
+ * @tparam T the underlying type
+ * @param vector
+ * @return T
+ */
+template <class T>
+T meanSquared(const T* vector, unsigned size) noexcept
+{
+    T sum = sumSquares(vector, size);
+    return sum / size;
+}
 
 template <class T>
 T meanSquared(absl::Span<const T> vector) noexcept

src/sfizz/simd/HelpersSSE.cpp

@@ -370,7 +370,7 @@ float meanSSE(const float* vector, unsigned size) noexcept
     return result / static_cast<float>(size);
 }
 
-float meanSquaredSSE(const float* vector, unsigned size) noexcept
+float sumSquaresSSE(const float* vector, unsigned size) noexcept
 {
     const auto sentinel = vector + size;
 
@@ -404,7 +404,7 @@ float meanSquaredSSE(const float* vector, unsigned size) noexcept
         vector++;
     }
 
-    return result / static_cast<float>(size);
+    return result;
 }
 
 void cumsumSSE(const float* input, float* output, unsigned size) noexcept

src/sfizz/simd/HelpersSSE.h

@@ -22,7 +22,7 @@ void subtractSSE(const float* input, float* output, unsigned size) noexcept;
 void subtract1SSE(float value, float* output, unsigned size) noexcept;
 void copySSE(const float* input, float* output, unsigned size) noexcept;
 float meanSSE(const float* vector, unsigned size) noexcept;
-float meanSquaredSSE(const float* vector, unsigned size) noexcept;
+float sumSquaresSSE(const float* vector, unsigned size) noexcept;
 void cumsumSSE(const float* input, float* output, unsigned size) noexcept;
 void diffSSE(const float* input, float* output, unsigned size) noexcept;
 void clampAllSSE(float* input, float low, float high, unsigned size) noexcept;

src/sfizz/simd/HelpersScalar.h

@@ -142,7 +142,7 @@ T meanScalar(const T* vector, unsigned size) noexcept
 }
 
 template <class T>
-T meanSquaredScalar(const T* vector, unsigned size) noexcept
+T sumSquaresScalar(const T* vector, unsigned size) noexcept
 {
     T result{ 0.0 };
     if (size == 0)
@@ -154,7 +154,7 @@ T meanSquaredScalar(const T* vector, unsigned size) noexcept
         vector++;
     }
 
-    return result / static_cast<T>(size);
+    return result;
 }
 
 template <class T>

tests/SIMDHelpersT.cpp

@@ -690,9 +690,9 @@ TEST_CASE("[Helpers] Mean (SIMD vs scalar)")
 TEST_CASE("[Helpers] Mean Squared")
 {
     std::array<float, 10> input { 1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f, 7.0f, 8.0f, 9.0f, 10.0f };
-    sfz::setSIMDOpStatus<float>(sfz::SIMDOps::meanSquared, false);
+    sfz::setSIMDOpStatus<float>(sfz::SIMDOps::sumSquares, false);
     REQUIRE(sfz::meanSquared<float>(input) == 38.5f);
-    sfz::setSIMDOpStatus<float>(sfz::SIMDOps::meanSquared, true);
+    sfz::setSIMDOpStatus<float>(sfz::SIMDOps::sumSquares, true);
     REQUIRE(sfz::meanSquared<float>(input) == 38.5f);
 }
 
@@ -700,9 +700,9 @@ TEST_CASE("[Helpers] Mean Squared (SIMD vs scalar)")
 {
     std::vector<float> input(medBufferSize);
     absl::c_iota(input, 0.0f);
-    sfz::setSIMDOpStatus<float>(sfz::SIMDOps::meanSquared, false);
+    sfz::setSIMDOpStatus<float>(sfz::SIMDOps::sumSquares, false);
     auto scalarResult = sfz::meanSquared<float>(input);
-    sfz::setSIMDOpStatus<float>(sfz::SIMDOps::meanSquared, true);
+    sfz::setSIMDOpStatus<float>(sfz::SIMDOps::sumSquares, true);
     auto simdResult = sfz::meanSquared<float>(input);
     REQUIRE( scalarResult == Approx(simdResult).margin(1e-3) );
 }