From 3816e08f5b89cd1139a4c46a47bb1aef2b6b15f4 Mon Sep 17 00:00:00 2001
From: Julian Waller <git@julusian.co.uk>
Date: Thu, 5 Oct 2023 15:36:31 +0100
Subject: [PATCH] fix: replace tbbmalloc with os native implementations

---
 src/CMakeModules/Bootstrap_Windows.cmake            |  4 ----
 src/common/memshfl.h                                | 10 ++++++++++
 src/modules/decklink/consumer/decklink_consumer.cpp |  5 ++---
 src/modules/decklink/consumer/frame.cpp             |  4 ++--
 src/shell/CMakeLists.txt                            |  2 --
 src/shell/main.cpp                                  |  6 ------
 6 files changed, 14 insertions(+), 17 deletions(-)

diff --git a/src/CMakeModules/Bootstrap_Windows.cmake b/src/CMakeModules/Bootstrap_Windows.cmake
index 2f3fe5085a..14cb9509e6 100644
--- a/src/CMakeModules/Bootstrap_Windows.cmake
+++ b/src/CMakeModules/Bootstrap_Windows.cmake
@@ -113,10 +113,6 @@ set(TBB_BIN_PATH "${PACKAGES_FOLDER}/tbb/bin/intel64")
 link_directories("${PACKAGES_FOLDER}/tbb/lib/intel64")
 casparcg_add_runtime_dependency("${TBB_BIN_PATH}/tbb12.dll")
 casparcg_add_runtime_dependency("${TBB_BIN_PATH}/tbb12_debug.dll")
-casparcg_add_runtime_dependency("${TBB_BIN_PATH}/tbbmalloc.dll")
-casparcg_add_runtime_dependency("${TBB_BIN_PATH}/tbbmalloc_debug.dll")
-casparcg_add_runtime_dependency("${TBB_BIN_PATH}/tbbmalloc_proxy.dll")
-casparcg_add_runtime_dependency("${TBB_BIN_PATH}/tbbmalloc_proxy_debug.dll")
 
 # GLEW
 set(GLEW_INCLUDE_PATH "${PACKAGES_FOLDER}/glew/include")
diff --git a/src/common/memshfl.h b/src/common/memshfl.h
index e9b1e4ac10..71ab527d10 100644
--- a/src/common/memshfl.h
+++ b/src/common/memshfl.h
@@ -34,6 +34,16 @@
 
 namespace caspar {
 
+#ifdef _MSC_VER
+static std::shared_ptr<void> create_aligned_buffer(size_t size) {
+    return std::shared_ptr<void>(_aligned_malloc(size, 64), _aligned_free);
+}
+#else
+static std::shared_ptr<void> create_aligned_buffer(size_t size) {
+    return std::shared_ptr<void>(aligned_alloc(64, size), free);
+}
+#endif
+
 static void* aligned_memshfl(void* dest, const void* source, size_t count, int m1, int m2, int m3, int m4)
 {
     __m128i*       dest128   = reinterpret_cast<__m128i*>(dest);
diff --git a/src/modules/decklink/consumer/decklink_consumer.cpp b/src/modules/decklink/consumer/decklink_consumer.cpp
index 01f166814d..82d6889ea7 100644
--- a/src/modules/decklink/consumer/decklink_consumer.cpp
+++ b/src/modules/decklink/consumer/decklink_consumer.cpp
@@ -43,11 +43,11 @@
 #include <common/timer.h>
 
 #include <tbb/parallel_for.h>
-#include <tbb/scalable_allocator.h>
 
 #include <boost/circular_buffer.hpp>
 
 #include <atomic>
+#include <common/memshfl.h>
 #include <common/prec_timer.h>
 #include <condition_variable>
 #include <future>
@@ -520,8 +520,7 @@ struct decklink_consumer final : public IDeckLinkVideoOutputCallback
                                     nb_samples);
             }
 
-            std::shared_ptr<void> image_data(scalable_aligned_malloc(decklink_format_desc_.size, 64),
-                                             scalable_aligned_free);
+            std::shared_ptr<void> image_data = create_aligned_buffer(decklink_format_desc_.size);
 
             schedule_next_video(image_data, nb_samples, video_scheduled_);
             for (auto& context : secondary_port_contexts_) {
diff --git a/src/modules/decklink/consumer/frame.cpp b/src/modules/decklink/consumer/frame.cpp
index c771bbb693..89da192099 100644
--- a/src/modules/decklink/consumer/frame.cpp
+++ b/src/modules/decklink/consumer/frame.cpp
@@ -31,7 +31,7 @@ namespace caspar { namespace decklink {
 
 std::shared_ptr<void> convert_to_key_only(const std::shared_ptr<void>& image_data, std::size_t byte_count)
 {
-    auto key_data = std::shared_ptr<void>(scalable_aligned_malloc(byte_count, 64), scalable_aligned_free);
+    auto key_data = create_aligned_buffer(byte_count);
 
     aligned_memshfl(key_data.get(), image_data.get(), byte_count, 0x0F0F0F0F, 0x0B0B0B0B, 0x07070707, 0x03030303);
 
@@ -133,7 +133,7 @@ std::shared_ptr<void> convert_frame_for_port(const core::video_format_desc& chan
                                          const core::const_frame&       frame2,
                                          BMDFieldDominance              field_dominance)
 {
-    std::shared_ptr<void> image_data(scalable_aligned_malloc(decklink_format_desc.size, 64), scalable_aligned_free);
+    std::shared_ptr<void> image_data = create_aligned_buffer(decklink_format_desc.size);
 
     if (field_dominance != bmdProgressiveFrame) {
         convert_frame(channel_format_desc,
diff --git a/src/shell/CMakeLists.txt b/src/shell/CMakeLists.txt
index 519100f24a..edd9ca4fda 100644
--- a/src/shell/CMakeLists.txt
+++ b/src/shell/CMakeLists.txt
@@ -54,9 +54,7 @@ if (MSVC)
 		Winmm.lib
 		Ws2_32.lib
 		optimized tbb.lib
-		optimized tbbmalloc.lib
 		debug tbb_debug.lib
-		debug tbbmalloc_debug.lib
 		OpenGL32.lib
 		glew32.lib
 		openal32.lib
diff --git a/src/shell/main.cpp b/src/shell/main.cpp
index 3c0dafc094..8c9ebe679f 100644
--- a/src/shell/main.cpp
+++ b/src/shell/main.cpp
@@ -19,16 +19,10 @@
  * Author: Robert Nagy, ronag89@gmail.com
  */
 
-// tbbmalloc_proxy:
-// Replace the standard memory allocation routines in Microsoft* C/C++ RTL
-// (malloc/free, global new/delete, etc.) with the TBB memory allocator.
-
 #if defined _DEBUG && defined _MSC_VER
 #define _CRTDBG_MAP_ALLOC
 #include <crtdbg.h>
 #include <stdlib.h>
-#else
-#include <tbb/tbbmalloc_proxy.h>
 #endif
 
 #include "included_modules.h"