Android #174

README.md

@@ -1,7 +1,8 @@
 | Pull Request | Commit | Title | Author | Merged? |
 |----|----|----|----|----|
-| [12466](https://github.com/yuzu-emu/yuzu//pull/12466) | [`ddda76f9b`](https://github.com/yuzu-emu/yuzu//pull/12466/files) |  core: track separate heap allocation for linux  | [liamwhite](https://github.com/liamwhite/) | Yes |
+| [12466](https://github.com/yuzu-emu/yuzu//pull/12466) | [`adb2af0a2`](https://github.com/yuzu-emu/yuzu//pull/12466/files) |  core: track separate heap allocation for linux  | [liamwhite](https://github.com/liamwhite/) | Yes |
 | [12479](https://github.com/yuzu-emu/yuzu//pull/12479) | [`20e040723`](https://github.com/yuzu-emu/yuzu//pull/12479/files) | video_core: Fix buffer_row_length for linear compressed textures | [GPUCode](https://github.com/GPUCode/) | Yes |
+| [12487](https://github.com/yuzu-emu/yuzu//pull/12487) | [`d0c60605a`](https://github.com/yuzu-emu/yuzu//pull/12487/files) | shader_recompiler: use default value for clip distances array | [liamwhite](https://github.com/liamwhite/) | Yes |
 
 
 End of merge log. You can find the original README.md below the break.

src/common/heap_tracker.cpp

@@ -1,7 +1,7 @@
 // SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
-#include <algorithm>
+#include <fstream>
 #include <vector>
 
 #include "common/heap_tracker.h"
@@ -11,11 +11,25 @@ namespace Common {
 
 namespace {
 
-constexpr s64 MaxResidentMapCount = 0x8000;
+s64 GetMaxPermissibleResidentMapCount() {
+    // Default value.
+    s64 value = 65530;
+
+    // Try to read how many mappings we can make.
+    std::ifstream s("/proc/sys/vm/max_map_count");
+    s >> value;
+
+    // Print, for debug.
+    LOG_INFO(HW_Memory, "Current maximum map count: {}", value);
+
+    // Allow 20000 maps for other code and to account for split inaccuracy.
+    return std::max<s64>(value - 20000, 0);
+}
 
 } // namespace
 
-HeapTracker::HeapTracker(Common::HostMemory& buffer) : m_buffer(buffer) {}
+HeapTracker::HeapTracker(Common::HostMemory& buffer)
+    : m_buffer(buffer), m_max_resident_map_count(GetMaxPermissibleResidentMapCount()) {}
 HeapTracker::~HeapTracker() = default;
 
 void HeapTracker::Map(size_t virtual_offset, size_t host_offset, size_t length,
@@ -74,8 +88,8 @@ void HeapTracker::Unmap(size_t virtual_offset, size_t size, bool is_separate_hea
             }
 
             // Erase from map.
-            it = m_mappings.erase(it);
             ASSERT(--m_map_count >= 0);
+            it = m_mappings.erase(it);
 
             // Free the item.
             delete item;
@@ -94,8 +108,8 @@ void HeapTracker::Protect(size_t virtual_offset, size_t size, MemoryPermission p
     this->SplitHeapMap(virtual_offset, size);
 
     // Declare tracking variables.
+    const VAddr end = virtual_offset + size;
     VAddr cur = virtual_offset;
-    VAddr end = virtual_offset + size;
 
     while (cur < end) {
         VAddr next = cur;
@@ -167,7 +181,7 @@ bool HeapTracker::DeferredMapSeparateHeap(size_t virtual_offset) {
         it->tick = m_tick++;
 
         // Check if we need to rebuild.
-        if (m_resident_map_count > MaxResidentMapCount) {
+        if (m_resident_map_count > m_max_resident_map_count) {
             rebuild_required = true;
         }
 
@@ -193,8 +207,12 @@ void HeapTracker::RebuildSeparateHeapAddressSpace() {
 
     ASSERT(!m_resident_mappings.empty());
 
-    // Unmap so we have at least 4 maps available.
-    const size_t desired_count = std::min(m_resident_map_count, MaxResidentMapCount - 4);
+    // Dump half of the mappings.
+    //
+    // Despite being worse in theory, this has proven to be better in practice than more
+    // regularly dumping a smaller amount, because it significantly reduces average case
+    // lock contention.
+    const size_t desired_count = std::min(m_resident_map_count, m_max_resident_map_count) / 2;
     const size_t evict_count = m_resident_map_count - desired_count;
     auto it = m_resident_mappings.begin();
 
@@ -247,8 +265,8 @@ void HeapTracker::SplitHeapMapLocked(VAddr offset) {
 
     // If resident, also insert into resident map.
     if (right->is_resident) {
-        m_resident_mappings.insert(*right);
         m_resident_map_count++;
+        m_resident_mappings.insert(*right);
     }
 }
 

src/common/heap_tracker.h

@@ -86,6 +86,7 @@ private:
 
 private:
     Common::HostMemory& m_buffer;
+    const s64 m_max_resident_map_count;
 
     std::shared_mutex m_rebuild_lock{};
     std::mutex m_lock{};

src/shader_recompiler/backend/spirv/spirv_emit_context.cpp

@@ -96,9 +96,9 @@ Id ImageType(EmitContext& ctx, const ImageDescriptor& desc, Id sampled_type) {
 }
 
 Id DefineVariable(EmitContext& ctx, Id type, std::optional<spv::BuiltIn> builtin,
-                  spv::StorageClass storage_class) {
+                  spv::StorageClass storage_class, std::optional<Id> initializer = std::nullopt) {
     const Id pointer_type{ctx.TypePointer(storage_class, type)};
-    const Id id{ctx.AddGlobalVariable(pointer_type, storage_class)};
+    const Id id{ctx.AddGlobalVariable(pointer_type, storage_class, initializer)};
     if (builtin) {
         ctx.Decorate(id, spv::Decoration::BuiltIn, *builtin);
     }
@@ -144,11 +144,12 @@ Id DefineInput(EmitContext& ctx, Id type, bool per_invocation,
 }
 
 Id DefineOutput(EmitContext& ctx, Id type, std::optional<u32> invocations,
-                std::optional<spv::BuiltIn> builtin = std::nullopt) {
+                std::optional<spv::BuiltIn> builtin = std::nullopt,
+                std::optional<Id> initializer = std::nullopt) {
     if (invocations && ctx.stage == Stage::TessellationControl) {
         type = ctx.TypeArray(type, ctx.Const(*invocations));
     }
-    return DefineVariable(ctx, type, builtin, spv::StorageClass::Output);
+    return DefineVariable(ctx, type, builtin, spv::StorageClass::Output, initializer);
 }
 
 void DefineGenericOutput(EmitContext& ctx, size_t index, std::optional<u32> invocations) {
@@ -811,10 +812,14 @@ void EmitContext::DefineAttributeMemAccess(const Info& info) {
             labels.push_back(OpLabel());
         }
         if (info.stores.ClipDistances()) {
-            literals.push_back(static_cast<u32>(IR::Attribute::ClipDistance0) >> 2);
-            labels.push_back(OpLabel());
-            literals.push_back(static_cast<u32>(IR::Attribute::ClipDistance4) >> 2);
-            labels.push_back(OpLabel());
+            if (profile.max_user_clip_distances >= 4) {
+                literals.push_back(static_cast<u32>(IR::Attribute::ClipDistance0) >> 2);
+                labels.push_back(OpLabel());
+            }
+            if (profile.max_user_clip_distances >= 8) {
+                literals.push_back(static_cast<u32>(IR::Attribute::ClipDistance4) >> 2);
+                labels.push_back(OpLabel());
+            }
         }
         OpSelectionMerge(end_block, spv::SelectionControlMask::MaskNone);
         OpSwitch(compare_index, default_label, literals, labels);
@@ -843,17 +848,21 @@ void EmitContext::DefineAttributeMemAccess(const Info& info) {
             ++label_index;
         }
         if (info.stores.ClipDistances()) {
-            AddLabel(labels[label_index]);
-            const Id pointer{OpAccessChain(output_f32, clip_distances, masked_index)};
-            OpStore(pointer, store_value);
-            OpReturn();
-            ++label_index;
-            AddLabel(labels[label_index]);
-            const Id fixed_index{OpIAdd(U32[1], masked_index, Const(4U))};
-            const Id pointer2{OpAccessChain(output_f32, clip_distances, fixed_index)};
-            OpStore(pointer2, store_value);
-            OpReturn();
-            ++label_index;
+            if (profile.max_user_clip_distances >= 4) {
+                AddLabel(labels[label_index]);
+                const Id pointer{OpAccessChain(output_f32, clip_distances, masked_index)};
+                OpStore(pointer, store_value);
+                OpReturn();
+                ++label_index;
+            }
+            if (profile.max_user_clip_distances >= 8) {
+                AddLabel(labels[label_index]);
+                const Id fixed_index{OpIAdd(U32[1], masked_index, Const(4U))};
+                const Id pointer{OpAccessChain(output_f32, clip_distances, fixed_index)};
+                OpStore(pointer, store_value);
+                OpReturn();
+                ++label_index;
+            }
         }
         AddLabel(end_block);
         OpUnreachable();
@@ -1532,9 +1541,16 @@ void EmitContext::DefineOutputs(const IR::Program& program) {
         if (stage == Stage::Fragment) {
             throw NotImplementedException("Storing ClipDistance in fragment stage");
         }
-        const Id type{TypeArray(
-            F32[1], Const(std::min(info.used_clip_distances, profile.max_user_clip_distances)))};
-        clip_distances = DefineOutput(*this, type, invocations, spv::BuiltIn::ClipDistance);
+        if (profile.max_user_clip_distances > 0) {
+            const u32 used{std::min(profile.max_user_clip_distances, 8u)};
+            const std::array<Id, 8> zero{f32_zero_value, f32_zero_value, f32_zero_value,
+                                         f32_zero_value, f32_zero_value, f32_zero_value,
+                                         f32_zero_value, f32_zero_value};
+            const Id type{TypeArray(F32[1], Const(used))};
+            const Id initializer{ConstantComposite(type, std::span(zero).subspan(0, used))};
+            clip_distances =
+                DefineOutput(*this, type, invocations, spv::BuiltIn::ClipDistance, initializer);
+        }
     }
     if (info.stores[IR::Attribute::Layer] &&
         (profile.support_viewport_index_layer_non_geometry || stage == Stage::Geometry)) {