From 95ca0ad63126c5bc70ada1eaf7856506c1a19d60 Mon Sep 17 00:00:00 2001
From: Mike Oliphant <oliphant@nostatic.org>
Date: Fri, 24 Mar 2023 11:00:51 -0700
Subject: [PATCH] Add floating point denormal code

---
 CMakeLists.txt                 |   1 +
 deps/denormal/architecture.hpp | 102 +++++++++++++++++++++++++++++++++
 src/nam_lv2.cpp                |  10 ++++
 src/nam_plugin.cpp             |   6 +-
 4 files changed, 116 insertions(+), 3 deletions(-)
 create mode 100644 deps/denormal/architecture.hpp

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 6e18eda..ab6a9ab 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -23,6 +23,7 @@ include_directories(SYSTEM deps/eigen)
 include_directories(SYSTEM deps/lv2/include)
 include_directories(SYSTEM deps/NeuralAmpModelerCore/NAM)
 include_directories(SYSTEM deps/json)
+include_directories(SYSTEM deps/denormal)
 
 add_subdirectory(src)
 
diff --git a/deps/denormal/architecture.hpp b/deps/denormal/architecture.hpp
new file mode 100644
index 0000000..53be234
--- /dev/null
+++ b/deps/denormal/architecture.hpp
@@ -0,0 +1,102 @@
+// From https://github.com/Dougal-s/Aether
+
+#ifndef ARCHITECTURE_HPP
+#define ARCHITECTURE_HPP
+
+// check cpu architecture
+
+#if /* x86_64 */ \
+	/* clang & gcc */ defined(__x86_64__) || \
+	/* msvc        */ defined(_M_AMD64) \
+
+	#define ARCH_X86
+	#define ARCH_X86_64
+
+#elif /* i386 */ \
+	/* clang & gcc */ defined(__i386__) || \
+	/* msvc        */ defined(_M_IX86) \
+
+	#define ARCH_X86
+	#define ARCH_I386
+
+#elif /* Arm64 */ \
+	/* clang & gcc */ defined(__aarch64__) || \
+	/* msvc        */ defined(_M_ARM64) \
+
+	#define ARCH_ARM
+	#define ARCH_ARM64
+
+#elif /* Arm */ \
+	/* clang & gcc */ defined(__arm__) || \
+	/* msvc        */ defined(_M_ARM) \
+
+	#define ARCH_ARM
+	#define ARCH_ARM32
+
+#else
+	#define ARCH_UNKNOWN
+#endif
+
+
+// check cpu extensions
+
+/* clang & gcc */
+#ifdef __SSE__
+	#define ARCH_EXT_SSE
+#endif
+
+#ifdef __SSE2__
+	#define ARCH_EXT_SSE2
+#endif
+
+#ifdef __SSE3__
+	#define ARCH_EXT_SSE3
+#endif
+
+/* msvc */
+#if defined(ARCH_X86_64)
+	#define ARCH_EXT_SSE
+	#define ARCH_EXT_SSE2
+
+	// msvc doesn't seem to have anything for sse3 so I am just assuming
+	// it is supported
+	#define ARCH_EXT_SSE3
+#elif defined(ARCH_I386)
+	#if _M_IX86_FP > 0
+		#define ARCH_EXT_SSE
+	#elif _M_IX86_FP > 1
+		#define ARCH_EXT_SSE3
+		#define ARCH_EXT_SSE2
+		#define ARCH_EXT_SSE
+	#endif
+#endif
+
+
+// misc functions
+
+#ifdef ARCH_EXT_SSE
+
+	#include <cfenv>
+	#ifndef FE_DFL_DISABLE_SSE_DENORMS_ENV
+		#include <immintrin.h>
+	#endif
+
+#endif
+
+inline void disable_denormals() noexcept {
+
+	#if defined(ARCH_EXT_SSE)
+		#ifdef FE_DFL_DISABLE_SSE_DENORMS_ENV
+			std::fesetenv(FE_DFL_DISABLE_SSE_DENORMS_ENV);
+		#else
+			_mm_setcsr(_mm_getcsr() | 0x8040);
+		#endif
+	#elif defined(ARCH_ARM)
+		#if __has_builtin(__builtin_arm_set_fpscr) && __has_builtin(__builtin_arm_get_fpscr)
+			__builtin_arm_set_fpscr(__builtin_arm_get_fpscr() | (1 << 24));
+		#endif
+	#endif
+
+}
+
+#endif
diff --git a/src/nam_lv2.cpp b/src/nam_lv2.cpp
index 6deda30..f01e6c8 100644
--- a/src/nam_lv2.cpp
+++ b/src/nam_lv2.cpp
@@ -9,6 +9,8 @@
 #include <lv2/urid/urid.h>
 #include <lv2/worker/worker.h>
 
+#include "architecture.hpp"
+
 #include "nam_plugin.h"
 
 // LV2 Functions
@@ -41,7 +43,15 @@ static void activate(LV2_Handle) {}
 
 static void run(LV2_Handle instance, uint32_t n_samples)
 {
+	// Disable floating point denormals
+	std::fenv_t fe_state;
+	std::feholdexcept(&fe_state);
+	disable_denormals();
+
 	static_cast<NAM::Plugin*>(instance)->process(n_samples);
+
+	// restore previous floating point state
+	std::feupdateenv(&fe_state);
 }
 
 static void deactivate(LV2_Handle) {}
diff --git a/src/nam_plugin.cpp b/src/nam_plugin.cpp
index dd8b445..0b3567d 100644
--- a/src/nam_plugin.cpp
+++ b/src/nam_plugin.cpp
@@ -155,8 +155,8 @@ namespace NAM {
 		if (dblData.size() != n_samples)
 			dblData.resize(n_samples);
 
-		float inputLevel = pow(10, *(ports.input_level) * 0.05);
-		float outputLevel = pow(10, *(ports.output_level) * 0.05);
+		float inputLevel = powf(10, *(ports.input_level) * 0.05f);
+		float outputLevel = powf(10, *(ports.output_level) * 0.05f);
 
 		for (unsigned int i = 0; i < n_samples; i++)
 		{
@@ -176,7 +176,7 @@ namespace NAM {
 
 		for (unsigned int i = 0; i < n_samples; i++)
 		{
-			ports.audio_out[i] = dblData[i] * outputLevel;
+			ports.audio_out[i] = (float)(dblData[i] * outputLevel);
 		}
 	}
 }