Update parallelRDP and MoltenVK (#1634)

ares-emulator · Sep 3, 2024 · 7135646 · 7135646
1 parent 4ab8472
commit 7135646
Show file tree

Hide file tree

Showing 23 changed files with 49,458 additions and 54,244 deletions.
diff --git a/ares/n64/vulkan/parallel-rdp/COMMIT b/ares/n64/vulkan/parallel-rdp/COMMIT
@@ -1 +1 @@
-43bc31642cc70d04adb828a285e68cdbde7110a9
+fe5becd13638873db90d46e7ba7d48255971f82a
diff --git a/ares/n64/vulkan/parallel-rdp/parallel-rdp/rdp_renderer.cpp b/ares/n64/vulkan/parallel-rdp/parallel-rdp/rdp_renderer.cpp
@@ -716,7 +716,7 @@ static int normalize_dzpix(int dz)
 	else if (dz == 0)
 		return 1;
 
-	unsigned bit = 31 - leading_zeroes(dz);
+	unsigned bit = 31 - Util::leading_zeroes(dz);
 	return 1 << (bit + 1);
 }
 
@@ -1680,7 +1680,7 @@ void Renderer::submit_span_setup_jobs(Vulkan::CommandBuffer &cmd, bool upscale)
 	cmd.set_buffer_view(1, 0, *instance.gpu.span_info_jobs_view);
 	cmd.set_specialization_constant_mask(3);
 	cmd.set_specialization_constant(0, (upscale ? caps.upscaling : 1) * ImplementationConstants::DefaultWorkgroupSize);
-	cmd.set_specialization_constant(1, upscale ? trailing_zeroes(caps.upscaling) : 0u);
+	cmd.set_specialization_constant(1, upscale ? Util::trailing_zeroes(caps.upscaling) : 0u);
 
 	Vulkan::QueryPoolHandle begin_ts, end_ts;
 	if (caps.timestamp >= 2)
@@ -1780,7 +1780,7 @@ void Renderer::submit_rasterization(Vulkan::CommandBuffer &cmd, Vulkan::Buffer &
 	if (caps.timestamp >= 2)
 		start_ts = cmd.write_timestamp(VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT);
 
-	uint32_t scale_log2_bit = (upscaling ? trailing_zeroes(caps.upscaling) : 0u) << RASTERIZATION_UPSCALING_LOG2_BIT_OFFSET;
+	uint32_t scale_log2_bit = (upscaling ? Util::trailing_zeroes(caps.upscaling) : 0u) << RASTERIZATION_UPSCALING_LOG2_BIT_OFFSET;
 
 	for (size_t i = 0; i < stream.static_raster_state_cache.size(); i++)
 	{
@@ -1891,7 +1891,7 @@ void Renderer::submit_tile_binning_combined(Vulkan::CommandBuffer &cmd, bool ups
 		if (supports_subgroup_size_control(32, subgroup_size))
 		{
 			cmd.enable_subgroup_size_control(true);
-			cmd.set_subgroup_size_log2(true, 5, trailing_zeroes(subgroup_size));
+			cmd.set_subgroup_size_log2(true, 5, Util::trailing_zeroes(subgroup_size));
 		}
 	}
 	else
@@ -2092,7 +2092,7 @@ void Renderer::submit_depth_blend(Vulkan::CommandBuffer &cmd, Vulkan::Buffer &tm
 	cmd.set_specialization_constant(5, Limits::MaxPrimitives);
 	cmd.set_specialization_constant(6, upscaled ? caps.max_width : Limits::MaxWidth);
 	cmd.set_specialization_constant(7, uint32_t(force_write_mask || (!is_host_coherent && !upscaled)) |
-	                                   ((upscaled ? trailing_zeroes(caps.upscaling) : 0u) << 1u));
+	                                   ((upscaled ? Util::trailing_zeroes(caps.upscaling) : 0u) << 1u));
 
 	if (upscaled)
 		cmd.set_storage_buffer(0, 0, *upscaling_multisampled_rdram);

diff --git a/ares/n64/vulkan/parallel-rdp/parallel-rdp/shaders/slangmosh.hpp b/ares/n64/vulkan/parallel-rdp/parallel-rdp/shaders/slangmosh.hpp
diff --git a/ares/n64/vulkan/parallel-rdp/parallel-rdp/video_interface.cpp b/ares/n64/vulkan/parallel-rdp/parallel-rdp/video_interface.cpp
@@ -484,7 +484,7 @@ Vulkan::ImageHandle VideoInterface::vram_fetch_stage(const Registers &regs, unsi
 	async_cmd->set_specialization_constant_mask(7);
 	async_cmd->set_specialization_constant(0, uint32_t(rdram_size));
 	async_cmd->set_specialization_constant(1, regs.status & (VI_CONTROL_TYPE_MASK | VI_CONTROL_META_AA_BIT));
-	async_cmd->set_specialization_constant(2, trailing_zeroes(scaling_factor));
+	async_cmd->set_specialization_constant(2, Util::trailing_zeroes(scaling_factor));
 
 	async_cmd->push_constants(&push, 0, sizeof(push));
 	async_cmd->dispatch((extract_width + 15) / 16,

diff --git a/ares/n64/vulkan/parallel-rdp/util/aligned_alloc.hpp b/ares/n64/vulkan/parallel-rdp/util/aligned_alloc.hpp
@@ -40,14 +40,18 @@ struct AlignedAllocation
     static void *operator new(size_t size)
     {
         void *ret = ::Util::memalign_alloc(alignof(T), size);
+#ifdef __EXCEPTIONS
         if (!ret) throw std::bad_alloc();
+#endif
         return ret;
     }
 
     static void *operator new[](size_t size)
     {
         void *ret = ::Util::memalign_alloc(alignof(T), size);
+#ifdef __EXCEPTIONS
         if (!ret) throw std::bad_alloc();
+#endif
         return ret;
     }
 

diff --git a/ares/n64/vulkan/parallel-rdp/util/bitops.hpp b/ares/n64/vulkan/parallel-rdp/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

diff --git a/ares/n64/vulkan/parallel-rdp/util/timer.cpp b/ares/n64/vulkan/parallel-rdp/util/timer.cpp
@@ -107,7 +107,7 @@ int64_t get_current_time_nsecs()
 	return int64_t(double(li.QuadPart) * static_qpc_freq.inv_freq);
 #else
 	struct timespec ts = {};
-#ifdef ANDROID
+#if defined(ANDROID) || defined(__FreeBSD__)
 	constexpr auto timebase = CLOCK_MONOTONIC;
 #else
 	constexpr auto timebase = CLOCK_MONOTONIC_RAW;
@@ -128,4 +128,4 @@ double Timer::end()
 	auto nt = get_current_time_nsecs();
 	return double(nt - t) * 1e-9;
 }
-}
+}