Merge pull request #17808 from hrydgard/frustum-cull-small-draws

Frustum-cull small draws (experiment)

CMakeLists.txt

@@ -737,6 +737,7 @@ add_library(Common STATIC
 	Common/Input/InputState.cpp
 	Common/Input/InputState.h
 	Common/Math/fast/fast_matrix.c
+	Common/Math/CrossSIMD.h
 	Common/Math/curves.cpp
 	Common/Math/curves.h
 	Common/Math/expression_parser.cpp

Common/Common.vcxproj

@@ -484,6 +484,7 @@
     <ClInclude Include="Input\GestureDetector.h" />
     <ClInclude Include="Input\InputState.h" />
     <ClInclude Include="Input\KeyCodes.h" />
+    <ClInclude Include="Math\CrossSIMD.h" />
     <ClInclude Include="Math\curves.h" />
     <ClInclude Include="Math\expression_parser.h" />
     <ClInclude Include="Math\fast\fast_matrix.h" />

Common/Common.vcxproj.filters

@@ -518,6 +518,9 @@
     <ClInclude Include="GPU\Vulkan\VulkanDescSet.h">
       <Filter>GPU\Vulkan</Filter>
     </ClInclude>
+    <ClInclude Include="Math\CrossSIMD.h">
+      <Filter>Math</Filter>
+    </ClInclude>
   </ItemGroup>
   <ItemGroup>
     <ClCompile Include="ABI.cpp" />

Common/Math/CrossSIMD.h

@@ -0,0 +1,58 @@
+// CrossSIMD
+//
+// Compatibility wrappers for SIMD dialects.
+//
+// In the long run, might do a more general single-source-SIMD wrapper here consisting
+// of defines that translate to either NEON or SSE. It would be possible to write quite a lot of
+// our various color conversion functions and so on in a pretty generic manner.
+
+#include "ppsspp_config.h"
+
+#include <cstdint>
+
+#if PPSSPP_ARCH(SSE2)
+#include <emmintrin.h>
+#endif
+
+#if PPSSPP_ARCH(ARM_NEON)
+#if defined(_MSC_VER) && PPSSPP_ARCH(ARM64)
+#include <arm64_neon.h>
+#else
+#include <arm_neon.h>
+#endif
+#endif
+
+// Basic types
+
+#if PPSSPP_ARCH(ARM64_NEON)
+
+// No special ones here.
+
+#elif PPSSPP_ARCH(ARM_NEON)
+
+// Compatibility wrappers making ARM64 NEON code run on ARM32
+// With optimization on, these should compile down to the optimal code.
+
+inline float32x4_t vmulq_laneq_f32(float32x4_t a, float32x4_t b, int lane) {
+	switch (lane & 3) {
+	case 0: return vmulq_lane_f32(a, vget_low_f32(b), 0);
+	case 1: return vmulq_lane_f32(a, vget_low_f32(b), 1);
+	case 2: return vmulq_lane_f32(a, vget_high_f32(b), 0);
+	default: return vmulq_lane_f32(a, vget_high_f32(b), 1);
+	}
+}
+
+inline float32x4_t vmlaq_laneq_f32(float32x4_t a, float32x4_t b, float32x4_t c, int lane) {
+	switch (lane & 3) {
+	case 0: return vmlaq_lane_f32(a, b, vget_low_f32(c), 0);
+	case 1: return vmlaq_lane_f32(a, b, vget_low_f32(c), 1);
+	case 2: return vmlaq_lane_f32(a, b, vget_high_f32(c), 0);
+	default: return vmlaq_lane_f32(a, b, vget_high_f32(c), 1);
+	}
+}
+
+inline uint32x4_t vcgezq_f32(float32x4_t v) {
+	return vcgeq_f32(v, vdupq_n_f32(0.0f));
+}
+
+#endif

Common/UI/View.cpp

@@ -620,7 +620,6 @@ CollapsibleHeader::CollapsibleHeader(bool *toggle, const std::string &text, Layo
 
 void CollapsibleHeader::Draw(UIContext &dc) {
 	Style style = dc.theme->itemStyle;
-	style.background.color = 0;
 	if (HasFocus()) style = dc.theme->itemFocusedStyle;
 	if (down_) style = dc.theme->itemDownStyle;
 	if (!IsEnabled()) style = dc.theme->itemDisabledStyle;

Core/Config.cpp

@@ -1893,7 +1893,7 @@ void PlayTimeTracker::Load(const Section *section) {
 
 		// Parse the string.
 		PlayTime gameTime{};
-		if (2 == sscanf(value.c_str(), "%d,%llu", &gameTime.totalTimePlayed, &gameTime.lastTimePlayed)) {
+		if (2 == sscanf(value.c_str(), "%d,%llu", &gameTime.totalTimePlayed, (long long *)&gameTime.lastTimePlayed)) {
 			tracker_[key] = gameTime;
 		}
 	}