From 9324d137bcb28a03905ba44b54ae7febebaf7e7a Mon Sep 17 00:00:00 2001
From: paulfd <paulfd@outlook.fr>
Date: Sun, 15 Sep 2019 20:16:42 +0200
Subject: [PATCH] Added a "scalar add" SIMD helper

---
 benchmarks/BM_add.cpp | 31 +++++++++++++++++++++++++++++++
 sfizz/SIMDDummy.cpp   |  6 ++++++
 sfizz/SIMDHelpers.h   | 19 +++++++++++++++++++
 3 files changed, 56 insertions(+)

diff --git a/benchmarks/BM_add.cpp b/benchmarks/BM_add.cpp
index 20ab8386..8221e997 100644
--- a/benchmarks/BM_add.cpp
+++ b/benchmarks/BM_add.cpp
@@ -49,6 +49,33 @@ public:
   std::vector<float> output;
 };
 
+BENCHMARK_DEFINE_F(AddArray, Value_Scalar)(benchmark::State& state) {
+    for (auto _ : state)
+    {
+        add<float, false>(1.1f, absl::MakeSpan(output));
+    }
+}
+
+BENCHMARK_DEFINE_F(AddArray, Value_SIMD)(benchmark::State& state) {
+    for (auto _ : state)
+    {
+        add<float, true>(1.1f, absl::MakeSpan(output));
+    }
+}
+
+BENCHMARK_DEFINE_F(AddArray, Value_Scalar_Unaligned)(benchmark::State& state) {
+    for (auto _ : state)
+    {
+        add<float, false>(1.1f, absl::MakeSpan(output).subspan(1));
+    }
+}
+
+BENCHMARK_DEFINE_F(AddArray, Value_SIMD_Unaligned)(benchmark::State& state) {
+    for (auto _ : state)
+    {
+        add<float, true>(1.1f, absl::MakeSpan(output).subspan(1));
+    }
+}
 
 BENCHMARK_DEFINE_F(AddArray, Scalar)(benchmark::State& state) {
     for (auto _ : state)
@@ -78,6 +105,10 @@ BENCHMARK_DEFINE_F(AddArray, SIMD_Unaligned)(benchmark::State& state) {
     }
 }
 
+BENCHMARK_REGISTER_F(AddArray, Value_Scalar)->RangeMultiplier(4)->Range(1 << 2, 1 << 12);
+BENCHMARK_REGISTER_F(AddArray, Value_SIMD)->RangeMultiplier(4)->Range(1 << 2, 1 << 12);
+BENCHMARK_REGISTER_F(AddArray, Value_Scalar_Unaligned)->RangeMultiplier(4)->Range(1 << 2, 1 << 12);
+BENCHMARK_REGISTER_F(AddArray, Value_SIMD_Unaligned)->RangeMultiplier(4)->Range(1 << 2, 1 << 12);
 BENCHMARK_REGISTER_F(AddArray, Scalar)->RangeMultiplier(4)->Range(1 << 2, 1 << 12);
 BENCHMARK_REGISTER_F(AddArray, SIMD)->RangeMultiplier(4)->Range(1 << 2, 1 << 12);
 BENCHMARK_REGISTER_F(AddArray, Scalar_Unaligned)->RangeMultiplier(4)->Range(1 << 2, 1 << 12);
diff --git a/sfizz/SIMDDummy.cpp b/sfizz/SIMDDummy.cpp
index cbb98be6..d57ead41 100644
--- a/sfizz/SIMDDummy.cpp
+++ b/sfizz/SIMDDummy.cpp
@@ -114,6 +114,12 @@ void add<float, true>(absl::Span<const float> input, absl::Span<float> output) n
     add<float, false>(input, output);
 }
 
+template <>
+void add<float, true>(float value, absl::Span<float> output) noexcept
+{
+    add<float, false>(value, output);
+}
+
 template <>
 void subtract<float, true>(absl::Span<const float> input, absl::Span<float> output) noexcept
 {
diff --git a/sfizz/SIMDHelpers.h b/sfizz/SIMDHelpers.h
index 2647e709..064999e4 100644
--- a/sfizz/SIMDHelpers.h
+++ b/sfizz/SIMDHelpers.h
@@ -349,6 +349,25 @@ void add(absl::Span<const T> input, absl::Span<T> output) noexcept
 template <>
 void add<float, true>(absl::Span<const float> input, absl::Span<float> output) noexcept;
 
+template <class T>
+inline void snippetAdd(const T value, T*& output)
+{
+    *output++ += value;
+}
+
+template <class T, bool SIMD = SIMDConfig::add>
+void add(T value, absl::Span<T> output) noexcept
+{
+    auto* out = output.begin();
+    auto* sentinel = output.end();
+    while (out < sentinel)
+        snippetAdd(value, out);
+}
+
+template <>
+void add<float, true>(float value, absl::Span<float> output) noexcept;
+
+
 template <class T>
 inline void snippetSubtract(const T*& input, T*& output)
 {