Make bitops proper inline functions.

Avoid macro conflicts.
Themaister · Jul 12, 2024 · f788e2a · f788e2a
1 parent efd0eb5
commit f788e2a
Show file tree

Hide file tree

Showing 8 changed files with 33 additions and 24 deletions.
diff --git a/audio/audio_mixer.cpp b/audio/audio_mixer.cpp
@@ -322,7 +322,7 @@ StreamID Mixer::add_mixer_stream(MixerStream *stream, bool start_playing,
 		if (!vacant_mask)
 			continue;
 
-		uint32_t subindex = trailing_zeroes(vacant_mask);
+		uint32_t subindex = Util::trailing_zeroes(vacant_mask);
 		uint32_t index = i * 32 + subindex;
 
 		MixerStream *old_stream = mixer_streams[index];

diff --git a/renderer/fft/fft.cpp b/renderer/fft/fft.cpp
@@ -454,7 +454,7 @@ std::vector<unsigned> FFT::Impl::split_radices(unsigned N) const
 		return {};
 
 	std::vector<unsigned> splits;
-	unsigned N_log2 = trailing_zeroes(N);
+	unsigned N_log2 = Util::trailing_zeroes(N);
 
 	// This should be deduced based on limits.
 	constexpr unsigned max_split = 9;
@@ -733,4 +733,4 @@ unsigned FFT::get_num_iterations() const
 		return 0;
 	return unsigned(impl->iterations.size());
 }
-}
+}
diff --git a/renderer/lights/clusterer.cpp b/renderer/lights/clusterer.cpp
@@ -1978,7 +1978,7 @@ void LightClusterer::build_cluster(Vulkan::CommandBuffer &cmd, Vulkan::ImageView
 
 	Push push = {
 		inverse_cluster_transform,
-		uvec4(res_x, res_y, res_z, trailing_zeroes(res_z)),
+		uvec4(res_x, res_y, res_z, Util::trailing_zeroes(res_z)),
 		vec4(1.0f / res_x, 1.0f / res_y, 1.0f / ((ClusterHierarchies + 1) * res_z), 1.0f),
 		vec4(inv_res, radius),
 		legacy.spots.count, legacy.points.count,

diff --git a/renderer/render_graph.cpp b/renderer/render_graph.cpp
@@ -2434,7 +2434,7 @@ void RenderGraph::enqueue_swapchain_scale_pass(Vulkan::Device &device_)
 	physical_events[index].to_flush_access = 0;
 	for (auto &e : physical_events[index].invalidated_in_stage)
 		e = 0;
-	physical_events[index].invalidated_in_stage[trailing_zeroes(VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT)] = VK_ACCESS_2_SHADER_SAMPLED_READ_BIT;
+	physical_events[index].invalidated_in_stage[Util::trailing_zeroes(VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT)] = VK_ACCESS_2_SHADER_SAMPLED_READ_BIT;
 
 	cmd->end_region();
 	if (physical_dimensions[index].uses_semaphore())

diff --git a/renderer/scene.cpp b/renderer/scene.cpp
@@ -606,7 +606,7 @@ void Scene::update_transform_tree(TaskComposer *composer)
 			uint32_t mask = pending_hierarchy_level_mask.load(std::memory_order_relaxed);
 			if (!mask)
 				return;
-			uint32_t num_pending_levels = 32 - leading_zeroes(mask);
+			uint32_t num_pending_levels = 32 - Util::leading_zeroes(mask);
 
 			TaskComposer stage_composer(thread_group);
 			for (unsigned level = 0, count = num_pending_levels; level < count; level++)
@@ -624,7 +624,7 @@ void Scene::update_transform_tree(TaskComposer *composer)
 		uint32_t mask = pending_hierarchy_level_mask.load(std::memory_order_relaxed);
 		if (mask)
 		{
-			uint32_t num_pending_levels = 32 - leading_zeroes(mask);
+			uint32_t num_pending_levels = 32 - Util::leading_zeroes(mask);
 			for (unsigned level = 0, count = num_pending_levels; level < count; level++)
 				perform_per_level_updates(level, nullptr);
 		}

diff --git a/scene-export/meshlet_export.cpp b/scene-export/meshlet_export.cpp
@@ -101,7 +101,7 @@ static void adjust_quant(std::vector<T> &values, int &exp)
 	if (active_bits == 0)
 		return;
 
-	int extra_shift = trailing_zeroes(active_bits);
+	int extra_shift = Util::trailing_zeroes(active_bits);
 	for (auto &value : values)
 		for (auto &c : value.data)
 			c >>= extra_shift;
@@ -297,7 +297,7 @@ static std::vector<i16vec2> mesh_extract_uv_snorm_scale(const SceneFormats::Mesh
 // Analyze bits required to encode a delta.
 static uint32_t compute_required_bits_unsigned(uint32_t delta)
 {
-	return delta == 0 ? 0 : (32 - leading_zeroes(delta));
+	return delta == 0 ? 0 : (32 - Util::leading_zeroes(delta));
 }
 
 static vec3 decode_snorm_exp(i16vec3 p, int exp)

diff --git a/tests/meshlet_viewer.cpp b/tests/meshlet_viewer.cpp
@@ -1152,7 +1152,7 @@ struct MeshletViewerApplication : Granite::Application, Granite::EventHandler //
 					unsigned count = 0;
 					auto *m = static_cast<const uint32_t *>(device.map_host_buffer(buffer, MEMORY_ACCESS_READ_BIT));
 					for (uint32_t i = 0; i < buffer.get_create_info().size / 4; i++)
-						count += popcount32(m[i]);
+						count += Util::popcount32(m[i]);
 					return count;
 				};
 

diff --git a/util/bitops.hpp b/util/bitops.hpp
@@ -29,13 +29,22 @@
 namespace Util
 {
 #ifdef __GNUC__
-#define leading_zeroes(x) ((x) == 0 ? 32 : __builtin_clz(x))
-#define trailing_zeroes(x) ((x) == 0 ? 32 : __builtin_ctz(x))
-#define trailing_ones(x) __builtin_ctz(~uint32_t(x))
-#define leading_zeroes64(x) ((x) == 0 ? 64 : __builtin_clzll(x))
-#define trailing_zeroes64(x) ((x) == 0 ? 64 : __builtin_ctzll(x))
-#define trailing_ones64(x) __builtin_ctzll(~uint64_t(x))
-#define popcount32(x) __builtin_popcount(x)
+#define leading_zeroes_(x) ((x) == 0 ? 32 : __builtin_clz(x))
+#define trailing_zeroes_(x) ((x) == 0 ? 32 : __builtin_ctz(x))
+#define trailing_ones_(x) __builtin_ctz(~uint32_t(x))
+#define leading_zeroes64_(x) ((x) == 0 ? 64 : __builtin_clzll(x))
+#define trailing_zeroes64_(x) ((x) == 0 ? 64 : __builtin_ctzll(x))
+#define trailing_ones64_(x) __builtin_ctzll(~uint64_t(x))
+#define popcount32_(x) __builtin_popcount(x)
+
+static inline uint32_t leading_zeroes(uint32_t x) { return leading_zeroes_(x); }
+static inline uint32_t trailing_zeroes(uint32_t x) { return trailing_zeroes_(x); }
+static inline uint32_t trailing_ones(uint32_t x) { return trailing_ones_(x); }
+static inline uint32_t leading_zeroes64(uint64_t x) { return leading_zeroes64_(x); }
+static inline uint32_t trailing_zeroes64(uint64_t x) { return trailing_zeroes64_(x); }
+static inline uint32_t trailing_ones64(uint64_t x) { return trailing_ones64_(x); }
+static inline uint32_t popcount32(uint32_t x) { return popcount32_(x); }
+
 #elif defined(_MSC_VER)
 namespace Internal
 {
@@ -81,13 +90,13 @@ static inline uint32_t ctz64(uint64_t x)
 }
 }
 
-#define popcount32(x) ::Util::Internal::popcount32(x)
-#define leading_zeroes(x) ::Util::Internal::clz(x)
-#define trailing_zeroes(x) ::Util::Internal::ctz(x)
-#define trailing_ones(x) ::Util::Internal::ctz(~uint32_t(x))
-#define leading_zeroes64(x) ::Util::Internal::clz64(x)
-#define trailing_zeroes64(x) ::Util::Internal::ctz64(x)
-#define trailing_ones64(x) ::Util::Internal::ctz64(~uint64_t(x))
+static inline uint32_t leading_zeroes(uint32_t x) { return Internal::clz(x); }
+static inline uint32_t trailing_zeroes(uint32_t x) { return Internal::ctz(x); }
+static inline uint32_t trailing_ones(uint32_t x) { return Internal::ctz(~x); }
+static inline uint32_t leading_zeroes64(uint64_t x) { return Internal::clz64(x); }
+static inline uint32_t trailing_zeroes64(uint64_t x) { return Internal::ctz64(x); }
+static inline uint32_t trailing_ones64(uint64_t x) { return Internal::ctz64(~x); }
+static inline uint32_t popcount32(uint32_t x) { return Internal::popcount32(x); }
 #else
 #error "Implement me."
 #endif