Tweak opus build flags and allocs

This gets us to ~40% of one core cpu usage during playback. Good enough
for a while I reckon!

Paid for the internal ram usage by reclaiming some stack size headroom.

lib/opusfile/CMakeLists.txt

@@ -11,6 +11,11 @@ set(CMAKE_POLICY_DEFAULT_CMP0077 NEW)
 
 set(OPUS_FIXED_POINT ON)
 set(OPUS_ENABLE_FLOAT_API OFF)
+
+set(OPUS_VAR_ARRAYS OFF)
+set(OPUS_USE_ALLOCA ON)
+set(OPUS_NONTHREADSAFE_PSEUDOSTACK OFF)
+
 set(OPUS_INSTALL_PKG_CONFIG_MODULE OFF)
 set(OPUS_INSTALL_CMAKE_CONFIG_MODULE OFF)
 set(OPUS_BUILD_TESTING OFF)
@@ -18,5 +23,6 @@ set(OPUS_BUILD_SHARED_LIBS OFF)
 
 add_subdirectory($ENV{PROJ_PATH}/lib/opus ${CMAKE_CURRENT_BINARY_DIR}/opus)
 target_compile_definitions(opus PRIVATE CUSTOM_SUPPORT)
+target_compile_options(opus PRIVATE -Os -DSMALL_FOOTPRINT -funroll-loops -ffast-math)
 target_include_directories(opus PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/include)
 target_link_libraries(${COMPONENT_LIB} PUBLIC opus)

lib/opusfile/include/custom_support.h

@@ -6,12 +6,10 @@
 #define OVERRIDE_OPUS_ALLOC
 #define OVERRIDE_OPUS_FREE
 
-static OPUS_INLINE void *opus_alloc (size_t size)
-{
-   return heap_caps_malloc(size, MALLOC_CAP_SPIRAM);
+static OPUS_INLINE void* opus_alloc(size_t size) {
+  return heap_caps_malloc(size, MALLOC_CAP_INTERNAL | MALLOC_CAP_8BIT);
 }
 
-static OPUS_INLINE void opus_free (void *ptr)
-{
-   heap_caps_free(ptr);
+static OPUS_INLINE void opus_free(void* ptr) {
+  heap_caps_free(ptr);
 }

src/dev_console/include/console.hpp

@@ -18,7 +18,7 @@ class Console {
   auto Launch() -> void;
 
  protected:
-  virtual auto GetStackSizeKiB() -> uint16_t { return 16; }
+  virtual auto GetStackSizeKiB() -> uint16_t { return 8; }
   virtual auto RegisterExtraComponents() -> void {}
 
  private:

src/tasks/tasks.cpp

@@ -39,22 +39,20 @@ auto AllocateStack() -> cpp::span<StackType_t>;
 // usually written with embedded use cases in mind.
 template <>
 auto AllocateStack<Type::kAudioDecoder>() -> cpp::span<StackType_t> {
-  constexpr std::size_t size = 24 * 1024;
+  constexpr std::size_t size = 20 * 1024;
   static StackType_t sStack[size];
   return {sStack, size};
 }
-// LVGL requires only a relatively small stack. However, it can be allocated in
-// PSRAM so we give it a bit of headroom for safety.
+// LVGL requires only a relatively small stack. Lua's stack is allocated
+// separately.
 template <>
 auto AllocateStack<Type::kUi>() -> cpp::span<StackType_t> {
-  constexpr std::size_t size = 16 * 1024;
+  constexpr std::size_t size = 14 * 1024;
   static StackType_t sStack[size];
   return {sStack, size};
 }
 template <>
-// PCM conversion and resampling uses a very small amount of stack. It works
-// entirely with PSRAM-allocated buffers, so no real speed gain from allocating
-// it internally.
+// PCM conversion and resampling uses a very small amount of stack.
 auto AllocateStack<Type::kAudioConverter>() -> cpp::span<StackType_t> {
   constexpr std::size_t size = 4 * 1024;
   static StackType_t sStack[size];