GPU: Move backend work off CPU thread

stenzek · Dec 19, 2024 · 2d659fc · 2d659fc
1 parent 831c982
commit 2d659fc
Show file tree

Hide file tree

Showing 59 changed files with 6,753 additions and 4,902 deletions.
diff --git a/src/common/intrin.h b/src/common/intrin.h
@@ -14,9 +14,9 @@
 #define CPU_ARCH_SIMD 1
 #define CPU_ARCH_SSE 1
 #include <emmintrin.h>
-#include <tmmintrin.h>
-#include <smmintrin.h>
 #include <immintrin.h>
+#include <smmintrin.h>
+#include <tmmintrin.h>
 
 #if defined(__AVX2__)
 #define CPU_ARCH_AVX 1
@@ -96,3 +96,40 @@ ALWAYS_INLINE_RELEASE static void MemsetPtrs(T* ptr, T value, u32 count)
   for (u32 i = 0; i < remaining_count; i++)
     *(dest++) = value;
 }
+
+ALWAYS_INLINE static void MultiPause()
+{
+#if defined(CPU_ARCH_X86) || defined(CPU_ARCH_X64)
+  _mm_pause();
+  _mm_pause();
+  _mm_pause();
+  _mm_pause();
+  _mm_pause();
+  _mm_pause();
+  _mm_pause();
+  _mm_pause();
+#elif defined(CPU_ARCH_ARM64) && defined(_MSC_VER)
+  __isb(_ARM64_BARRIER_SY);
+  __isb(_ARM64_BARRIER_SY);
+  __isb(_ARM64_BARRIER_SY);
+  __isb(_ARM64_BARRIER_SY);
+  __isb(_ARM64_BARRIER_SY);
+  __isb(_ARM64_BARRIER_SY);
+  __isb(_ARM64_BARRIER_SY);
+  __isb(_ARM64_BARRIER_SY);
+#elif defined(CPU_ARCH_ARM64) || defined(CPU_ARCH_ARM32)
+  __asm__ __volatile__("isb");
+  __asm__ __volatile__("isb");
+  __asm__ __volatile__("isb");
+  __asm__ __volatile__("isb");
+  __asm__ __volatile__("isb");
+  __asm__ __volatile__("isb");
+  __asm__ __volatile__("isb");
+  __asm__ __volatile__("isb");
+#elif defined(CPU_ARCH_RISCV64)
+  // Probably wrong... pause is optional :/
+  asm volatile("fence" ::: "memory");
+#else
+#pragma warning("Missing implementation")
+#endif
+}
diff --git a/src/common/log_channels.h b/src/common/log_channels.h
@@ -27,6 +27,7 @@
   X(GPU)                                                                                                               \
   X(GPUDevice)                                                                                                         \
   X(GPUDump)                                                                                                           \
+  X(GPUThread)                                                                                                         \
   X(GPU_SW)                                                                                                            \
   X(GPU_HW)                                                                                                            \
   X(GameDatabase)                                                                                                      \

diff --git a/src/core/CMakeLists.txt b/src/core/CMakeLists.txt
@@ -61,10 +61,11 @@ add_library(core
   gpu_shadergen.h
   gpu_sw.cpp
   gpu_sw.h
-  gpu_sw_backend.cpp
-  gpu_sw_backend.h
   gpu_sw_rasterizer.cpp
   gpu_sw_rasterizer.h
+  gpu_thread.cpp
+  gpu_thread.h
+  gpu_thread_commands.h
   gpu_types.h
   guncon.cpp
   guncon.h
@@ -73,8 +74,6 @@ add_library(core
   gte_types.h
   host.cpp
   host.h
-  host_interface_progress_callback.cpp
-  host_interface_progress_callback.h
   hotkeys.cpp
   input_types.h
   imgui_overlays.cpp