summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat
-rw-r--r--src/common/polyfill_thread.h36
-rw-r--r--src/core/hle/kernel/k_condition_variable.cpp2
-rw-r--r--src/core/hle/kernel/k_light_lock.cpp2
-rw-r--r--src/core/hle/kernel/k_memory_layout.h6
-rw-r--r--src/core/hle/kernel/k_thread.cpp21
-rw-r--r--src/core/hle/kernel/k_thread.h24
-rw-r--r--src/core/hle/kernel/kernel.cpp39
-rw-r--r--src/input_common/drivers/gc_adapter.cpp4
-rw-r--r--src/input_common/drivers/joycon.cpp7
-rw-r--r--src/shader_recompiler/backend/glasm/emit_glasm.cpp2
-rw-r--r--src/shader_recompiler/backend/glasm/emit_glasm_bitwise_conversion.cpp4
-rw-r--r--src/shader_recompiler/backend/glasm/emit_glasm_image.cpp11
-rw-r--r--src/shader_recompiler/backend/glasm/emit_glasm_instructions.h1
-rw-r--r--src/shader_recompiler/backend/glsl/emit_glsl_bitwise_conversion.cpp4
-rw-r--r--src/shader_recompiler/backend/glsl/emit_glsl_image.cpp17
-rw-r--r--src/shader_recompiler/backend/glsl/emit_glsl_instructions.h1
-rw-r--r--src/shader_recompiler/backend/glsl/glsl_emit_context.cpp49
-rw-r--r--src/shader_recompiler/backend/spirv/emit_spirv_bitwise_conversion.cpp4
-rw-r--r--src/shader_recompiler/backend/spirv/emit_spirv_instructions.h1
-rw-r--r--src/shader_recompiler/frontend/ir/ir_emitter.cpp5
-rw-r--r--src/shader_recompiler/frontend/ir/opcodes.h1
-rw-r--r--src/shader_recompiler/frontend/ir/opcodes.inc1
-rw-r--r--src/shader_recompiler/frontend/ir/type.h31
-rw-r--r--src/shader_recompiler/frontend/ir/value.cpp3
-rw-r--r--src/shader_recompiler/frontend/ir/value.h1
-rw-r--r--src/shader_recompiler/ir_opt/texture_pass.cpp8
-rw-r--r--src/video_core/CMakeLists.txt4
-rw-r--r--src/video_core/fsr.cpp148
-rw-r--r--src/video_core/fsr.h19
-rw-r--r--src/video_core/host_shaders/CMakeLists.txt30
-rw-r--r--src/video_core/host_shaders/opengl_fidelityfx_fsr.frag108
-rw-r--r--src/video_core/host_shaders/opengl_fidelityfx_fsr_easu.frag9
-rw-r--r--src/video_core/host_shaders/opengl_fidelityfx_fsr_rcas.frag9
-rw-r--r--src/video_core/renderer_opengl/gl_fsr.cpp101
-rw-r--r--src/video_core/renderer_opengl/gl_fsr.h43
-rw-r--r--src/video_core/renderer_opengl/renderer_opengl.cpp90
-rw-r--r--src/video_core/renderer_opengl/renderer_opengl.h3
-rw-r--r--src/video_core/renderer_vulkan/vk_fsr.cpp144
-rw-r--r--src/yuzu/configuration/configure_graphics.ui2
-rw-r--r--src/yuzu/configuration/input_profiles.cpp7
-rw-r--r--src/yuzu/main.cpp9
-rw-r--r--src/yuzu/multiplayer/direct_connect.cpp21
-rw-r--r--src/yuzu/multiplayer/direct_connect.ui23
-rw-r--r--src/yuzu/multiplayer/validation.h25
44 files changed, 749 insertions, 331 deletions
diff --git a/src/common/polyfill_thread.h b/src/common/polyfill_thread.h
index 5a8d1ce08..b2c929d2f 100644
--- a/src/common/polyfill_thread.h
+++ b/src/common/polyfill_thread.h
@@ -11,6 +11,8 @@
11 11
12#ifdef __cpp_lib_jthread 12#ifdef __cpp_lib_jthread
13 13
14#include <chrono>
15#include <condition_variable>
14#include <stop_token> 16#include <stop_token>
15#include <thread> 17#include <thread>
16 18
@@ -21,11 +23,23 @@ void CondvarWait(Condvar& cv, Lock& lock, std::stop_token token, Pred&& pred) {
21 cv.wait(lock, token, std::move(pred)); 23 cv.wait(lock, token, std::move(pred));
22} 24}
23 25
26template <typename Rep, typename Period>
27bool StoppableTimedWait(std::stop_token token, const std::chrono::duration<Rep, Period>& rel_time) {
28 std::condition_variable_any cv;
29 std::mutex m;
30
31 // Perform the timed wait.
32 std::unique_lock lk{m};
33 return !cv.wait_for(lk, token, rel_time, [&] { return token.stop_requested(); });
34}
35
24} // namespace Common 36} // namespace Common
25 37
26#else 38#else
27 39
28#include <atomic> 40#include <atomic>
41#include <chrono>
42#include <condition_variable>
29#include <functional> 43#include <functional>
30#include <list> 44#include <list>
31#include <memory> 45#include <memory>
@@ -318,6 +332,28 @@ void CondvarWait(Condvar& cv, Lock& lock, std::stop_token token, Pred pred) {
318 cv.wait(lock, [&] { return pred() || token.stop_requested(); }); 332 cv.wait(lock, [&] { return pred() || token.stop_requested(); });
319} 333}
320 334
335template <typename Rep, typename Period>
336bool StoppableTimedWait(std::stop_token token, const std::chrono::duration<Rep, Period>& rel_time) {
337 if (token.stop_requested()) {
338 return false;
339 }
340
341 bool stop_requested = false;
342 std::condition_variable cv;
343 std::mutex m;
344
345 std::stop_callback cb(token, [&] {
346 // Wake up the waiting thread.
347 std::unique_lock lk{m};
348 stop_requested = true;
349 cv.notify_one();
350 });
351
352 // Perform the timed wait.
353 std::unique_lock lk{m};
354 return !cv.wait_for(lk, rel_time, [&] { return stop_requested; });
355}
356
321} // namespace Common 357} // namespace Common
322 358
323#endif 359#endif
diff --git a/src/core/hle/kernel/k_condition_variable.cpp b/src/core/hle/kernel/k_condition_variable.cpp
index 124149697..0c6b20db3 100644
--- a/src/core/hle/kernel/k_condition_variable.cpp
+++ b/src/core/hle/kernel/k_condition_variable.cpp
@@ -171,7 +171,7 @@ Result KConditionVariable::WaitForAddress(Handle handle, VAddr addr, u32 value)
171 R_UNLESS(owner_thread != nullptr, ResultInvalidHandle); 171 R_UNLESS(owner_thread != nullptr, ResultInvalidHandle);
172 172
173 // Update the lock. 173 // Update the lock.
174 cur_thread->SetAddressKey(addr, value); 174 cur_thread->SetUserAddressKey(addr, value);
175 owner_thread->AddWaiter(cur_thread); 175 owner_thread->AddWaiter(cur_thread);
176 176
177 // Begin waiting. 177 // Begin waiting.
diff --git a/src/core/hle/kernel/k_light_lock.cpp b/src/core/hle/kernel/k_light_lock.cpp
index 43185320d..d791acbe3 100644
--- a/src/core/hle/kernel/k_light_lock.cpp
+++ b/src/core/hle/kernel/k_light_lock.cpp
@@ -68,7 +68,7 @@ bool KLightLock::LockSlowPath(uintptr_t _owner, uintptr_t _cur_thread) {
68 68
69 // Add the current thread as a waiter on the owner. 69 // Add the current thread as a waiter on the owner.
70 KThread* owner_thread = reinterpret_cast<KThread*>(_owner & ~1ULL); 70 KThread* owner_thread = reinterpret_cast<KThread*>(_owner & ~1ULL);
71 cur_thread->SetAddressKey(reinterpret_cast<uintptr_t>(std::addressof(tag))); 71 cur_thread->SetKernelAddressKey(reinterpret_cast<uintptr_t>(std::addressof(tag)));
72 owner_thread->AddWaiter(cur_thread); 72 owner_thread->AddWaiter(cur_thread);
73 73
74 // Begin waiting to hold the lock. 74 // Begin waiting to hold the lock.
diff --git a/src/core/hle/kernel/k_memory_layout.h b/src/core/hle/kernel/k_memory_layout.h
index fd6e1d3e6..17fa1a6ed 100644
--- a/src/core/hle/kernel/k_memory_layout.h
+++ b/src/core/hle/kernel/k_memory_layout.h
@@ -67,9 +67,9 @@ constexpr size_t KernelPageBufferAdditionalSize = 0x33C000;
67constexpr std::size_t KernelResourceSize = KernelPageTableHeapSize + KernelInitialPageHeapSize + 67constexpr std::size_t KernelResourceSize = KernelPageTableHeapSize + KernelInitialPageHeapSize +
68 KernelSlabHeapSize + KernelPageBufferHeapSize; 68 KernelSlabHeapSize + KernelPageBufferHeapSize;
69 69
70constexpr bool IsKernelAddressKey(VAddr key) { 70//! NB: Use KThread::GetAddressKeyIsKernel().
71 return KernelVirtualAddressSpaceBase <= key && key <= KernelVirtualAddressSpaceLast; 71//! See explanation for deviation of GetAddressKey.
72} 72bool IsKernelAddressKey(VAddr key) = delete;
73 73
74constexpr bool IsKernelAddress(VAddr address) { 74constexpr bool IsKernelAddress(VAddr address) {
75 return KernelVirtualAddressSpaceBase <= address && address < KernelVirtualAddressSpaceEnd; 75 return KernelVirtualAddressSpaceBase <= address && address < KernelVirtualAddressSpaceEnd;
diff --git a/src/core/hle/kernel/k_thread.cpp b/src/core/hle/kernel/k_thread.cpp
index 21207fe99..84ff3c64b 100644
--- a/src/core/hle/kernel/k_thread.cpp
+++ b/src/core/hle/kernel/k_thread.cpp
@@ -330,7 +330,7 @@ void KThread::Finalize() {
330 KThread* const waiter = std::addressof(*it); 330 KThread* const waiter = std::addressof(*it);
331 331
332 // The thread shouldn't be a kernel waiter. 332 // The thread shouldn't be a kernel waiter.
333 ASSERT(!IsKernelAddressKey(waiter->GetAddressKey())); 333 ASSERT(!waiter->GetAddressKeyIsKernel());
334 334
335 // Clear the lock owner. 335 // Clear the lock owner.
336 waiter->SetLockOwner(nullptr); 336 waiter->SetLockOwner(nullptr);
@@ -763,19 +763,6 @@ void KThread::Continue() {
763 KScheduler::OnThreadStateChanged(kernel, this, old_state); 763 KScheduler::OnThreadStateChanged(kernel, this, old_state);
764} 764}
765 765
766void KThread::WaitUntilSuspended() {
767 // Make sure we have a suspend requested.
768 ASSERT(IsSuspendRequested());
769
770 // Loop until the thread is not executing on any core.
771 for (std::size_t i = 0; i < static_cast<std::size_t>(Core::Hardware::NUM_CPU_CORES); ++i) {
772 KThread* core_thread{};
773 do {
774 core_thread = kernel.Scheduler(i).GetSchedulerCurrentThread();
775 } while (core_thread == this);
776 }
777}
778
779Result KThread::SetActivity(Svc::ThreadActivity activity) { 766Result KThread::SetActivity(Svc::ThreadActivity activity) {
780 // Lock ourselves. 767 // Lock ourselves.
781 KScopedLightLock lk(activity_pause_lock); 768 KScopedLightLock lk(activity_pause_lock);
@@ -897,7 +884,7 @@ void KThread::AddWaiterImpl(KThread* thread) {
897 } 884 }
898 885
899 // Keep track of how many kernel waiters we have. 886 // Keep track of how many kernel waiters we have.
900 if (IsKernelAddressKey(thread->GetAddressKey())) { 887 if (thread->GetAddressKeyIsKernel()) {
901 ASSERT((num_kernel_waiters++) >= 0); 888 ASSERT((num_kernel_waiters++) >= 0);
902 KScheduler::SetSchedulerUpdateNeeded(kernel); 889 KScheduler::SetSchedulerUpdateNeeded(kernel);
903 } 890 }
@@ -911,7 +898,7 @@ void KThread::RemoveWaiterImpl(KThread* thread) {
911 ASSERT(kernel.GlobalSchedulerContext().IsLocked()); 898 ASSERT(kernel.GlobalSchedulerContext().IsLocked());
912 899
913 // Keep track of how many kernel waiters we have. 900 // Keep track of how many kernel waiters we have.
914 if (IsKernelAddressKey(thread->GetAddressKey())) { 901 if (thread->GetAddressKeyIsKernel()) {
915 ASSERT((num_kernel_waiters--) > 0); 902 ASSERT((num_kernel_waiters--) > 0);
916 KScheduler::SetSchedulerUpdateNeeded(kernel); 903 KScheduler::SetSchedulerUpdateNeeded(kernel);
917 } 904 }
@@ -987,7 +974,7 @@ KThread* KThread::RemoveWaiterByKey(s32* out_num_waiters, VAddr key) {
987 KThread* thread = std::addressof(*it); 974 KThread* thread = std::addressof(*it);
988 975
989 // Keep track of how many kernel waiters we have. 976 // Keep track of how many kernel waiters we have.
990 if (IsKernelAddressKey(thread->GetAddressKey())) { 977 if (thread->GetAddressKeyIsKernel()) {
991 ASSERT((num_kernel_waiters--) > 0); 978 ASSERT((num_kernel_waiters--) > 0);
992 KScheduler::SetSchedulerUpdateNeeded(kernel); 979 KScheduler::SetSchedulerUpdateNeeded(kernel);
993 } 980 }
diff --git a/src/core/hle/kernel/k_thread.h b/src/core/hle/kernel/k_thread.h
index 7cd94a340..9d771de0e 100644
--- a/src/core/hle/kernel/k_thread.h
+++ b/src/core/hle/kernel/k_thread.h
@@ -214,8 +214,6 @@ public:
214 214
215 void Continue(); 215 void Continue();
216 216
217 void WaitUntilSuspended();
218
219 constexpr void SetSyncedIndex(s32 index) { 217 constexpr void SetSyncedIndex(s32 index) {
220 synced_index = index; 218 synced_index = index;
221 } 219 }
@@ -607,13 +605,30 @@ public:
607 return address_key_value; 605 return address_key_value;
608 } 606 }
609 607
610 void SetAddressKey(VAddr key) { 608 [[nodiscard]] bool GetAddressKeyIsKernel() const {
609 return address_key_is_kernel;
610 }
611
612 //! NB: intentional deviation from official kernel.
613 //
614 // Separate SetAddressKey into user and kernel versions
615 // to cope with arbitrary host pointers making their way
616 // into things.
617
618 void SetUserAddressKey(VAddr key) {
611 address_key = key; 619 address_key = key;
620 address_key_is_kernel = false;
612 } 621 }
613 622
614 void SetAddressKey(VAddr key, u32 val) { 623 void SetUserAddressKey(VAddr key, u32 val) {
615 address_key = key; 624 address_key = key;
616 address_key_value = val; 625 address_key_value = val;
626 address_key_is_kernel = false;
627 }
628
629 void SetKernelAddressKey(VAddr key) {
630 address_key = key;
631 address_key_is_kernel = true;
617 } 632 }
618 633
619 void ClearWaitQueue() { 634 void ClearWaitQueue() {
@@ -772,6 +787,7 @@ private:
772 bool debug_attached{}; 787 bool debug_attached{};
773 s8 priority_inheritance_count{}; 788 s8 priority_inheritance_count{};
774 bool resource_limit_release_hint{}; 789 bool resource_limit_release_hint{};
790 bool address_key_is_kernel{};
775 StackParameters stack_parameters{}; 791 StackParameters stack_parameters{};
776 Common::SpinLock context_guard{}; 792 Common::SpinLock context_guard{};
777 793
diff --git a/src/core/hle/kernel/kernel.cpp b/src/core/hle/kernel/kernel.cpp
index 1fb25f221..d9eafe261 100644
--- a/src/core/hle/kernel/kernel.cpp
+++ b/src/core/hle/kernel/kernel.cpp
@@ -1198,28 +1198,35 @@ void KernelCore::Suspend(bool suspended) {
1198 const bool should_suspend{exception_exited || suspended}; 1198 const bool should_suspend{exception_exited || suspended};
1199 const auto activity = should_suspend ? ProcessActivity::Paused : ProcessActivity::Runnable; 1199 const auto activity = should_suspend ? ProcessActivity::Paused : ProcessActivity::Runnable;
1200 1200
1201 std::vector<KScopedAutoObject<KThread>> process_threads; 1201 //! This refers to the application process, not the current process.
1202 { 1202 KScopedAutoObject<KProcess> process = CurrentProcess();
1203 KScopedSchedulerLock sl{*this}; 1203 if (process.IsNull()) {
1204 return;
1205 }
1204 1206
1205 if (auto* process = CurrentProcess(); process != nullptr) { 1207 // Set the new activity.
1206 process->SetActivity(activity); 1208 process->SetActivity(activity);
1207 1209
1208 if (!should_suspend) { 1210 // Wait for process execution to stop.
1209 // Runnable now; no need to wait. 1211 bool must_wait{should_suspend};
1210 return; 1212
1211 } 1213 // KernelCore::Suspend must be called from locked context, or we
1214 // could race another call to SetActivity, interfering with waiting.
1215 while (must_wait) {
1216 KScopedSchedulerLock sl{*this};
1217
1218 // Assume that all threads have finished running.
1219 must_wait = false;
1212 1220
1213 for (auto* thread : process->GetThreadList()) { 1221 for (auto i = 0; i < static_cast<s32>(Core::Hardware::NUM_CPU_CORES); ++i) {
1214 process_threads.emplace_back(thread); 1222 if (Scheduler(i).GetSchedulerCurrentThread()->GetOwnerProcess() ==
1223 process.GetPointerUnsafe()) {
1224 // A thread has not finished running yet.
1225 // Continue waiting.
1226 must_wait = true;
1215 } 1227 }
1216 } 1228 }
1217 } 1229 }
1218
1219 // Wait for execution to stop.
1220 for (auto& thread : process_threads) {
1221 thread->WaitUntilSuspended();
1222 }
1223} 1230}
1224 1231
1225void KernelCore::ShutdownCores() { 1232void KernelCore::ShutdownCores() {
diff --git a/src/input_common/drivers/gc_adapter.cpp b/src/input_common/drivers/gc_adapter.cpp
index ecb3e9dc2..d09ff178b 100644
--- a/src/input_common/drivers/gc_adapter.cpp
+++ b/src/input_common/drivers/gc_adapter.cpp
@@ -6,6 +6,7 @@
6 6
7#include "common/logging/log.h" 7#include "common/logging/log.h"
8#include "common/param_package.h" 8#include "common/param_package.h"
9#include "common/polyfill_thread.h"
9#include "common/settings_input.h" 10#include "common/settings_input.h"
10#include "common/thread.h" 11#include "common/thread.h"
11#include "input_common/drivers/gc_adapter.h" 12#include "input_common/drivers/gc_adapter.h"
@@ -217,8 +218,7 @@ void GCAdapter::AdapterScanThread(std::stop_token stop_token) {
217 Common::SetCurrentThreadName("ScanGCAdapter"); 218 Common::SetCurrentThreadName("ScanGCAdapter");
218 usb_adapter_handle = nullptr; 219 usb_adapter_handle = nullptr;
219 pads = {}; 220 pads = {};
220 while (!stop_token.stop_requested() && !Setup()) { 221 while (!Setup() && Common::StoppableTimedWait(stop_token, std::chrono::seconds{2})) {
221 std::this_thread::sleep_for(std::chrono::seconds(2));
222 } 222 }
223} 223}
224 224
diff --git a/src/input_common/drivers/joycon.cpp b/src/input_common/drivers/joycon.cpp
index 40cda400d..cedc94e63 100644
--- a/src/input_common/drivers/joycon.cpp
+++ b/src/input_common/drivers/joycon.cpp
@@ -5,6 +5,7 @@
5 5
6#include "common/param_package.h" 6#include "common/param_package.h"
7#include "common/polyfill_ranges.h" 7#include "common/polyfill_ranges.h"
8#include "common/polyfill_thread.h"
8#include "common/settings.h" 9#include "common/settings.h"
9#include "common/thread.h" 10#include "common/thread.h"
10#include "input_common/drivers/joycon.h" 11#include "input_common/drivers/joycon.h"
@@ -67,7 +68,8 @@ void Joycons::Setup() {
67void Joycons::ScanThread(std::stop_token stop_token) { 68void Joycons::ScanThread(std::stop_token stop_token) {
68 constexpr u16 nintendo_vendor_id = 0x057e; 69 constexpr u16 nintendo_vendor_id = 0x057e;
69 Common::SetCurrentThreadName("JoyconScanThread"); 70 Common::SetCurrentThreadName("JoyconScanThread");
70 while (!stop_token.stop_requested()) { 71
72 do {
71 SDL_hid_device_info* devs = SDL_hid_enumerate(nintendo_vendor_id, 0x0); 73 SDL_hid_device_info* devs = SDL_hid_enumerate(nintendo_vendor_id, 0x0);
72 SDL_hid_device_info* cur_dev = devs; 74 SDL_hid_device_info* cur_dev = devs;
73 75
@@ -81,8 +83,7 @@ void Joycons::ScanThread(std::stop_token stop_token) {
81 } 83 }
82 84
83 SDL_hid_free_enumeration(devs); 85 SDL_hid_free_enumeration(devs);
84 std::this_thread::sleep_for(std::chrono::seconds(5)); 86 } while (Common::StoppableTimedWait(stop_token, std::chrono::seconds{5}));
85 }
86} 87}
87 88
88bool Joycons::IsDeviceNew(SDL_hid_device_info* device_info) const { 89bool Joycons::IsDeviceNew(SDL_hid_device_info* device_info) const {
diff --git a/src/shader_recompiler/backend/glasm/emit_glasm.cpp b/src/shader_recompiler/backend/glasm/emit_glasm.cpp
index 0cb1e193e..fd4a61a4d 100644
--- a/src/shader_recompiler/backend/glasm/emit_glasm.cpp
+++ b/src/shader_recompiler/backend/glasm/emit_glasm.cpp
@@ -279,6 +279,8 @@ void SetupOptions(const IR::Program& program, const Profile& profile,
279 header += "OPTION NV_internal;" 279 header += "OPTION NV_internal;"
280 "OPTION NV_shader_storage_buffer;" 280 "OPTION NV_shader_storage_buffer;"
281 "OPTION NV_gpu_program_fp64;"; 281 "OPTION NV_gpu_program_fp64;";
282 // TODO: Enable only when MS is used
283 header += "OPTION NV_texture_multisample;";
282 if (info.uses_int64_bit_atomics) { 284 if (info.uses_int64_bit_atomics) {
283 header += "OPTION NV_shader_atomic_int64;"; 285 header += "OPTION NV_shader_atomic_int64;";
284 } 286 }
diff --git a/src/shader_recompiler/backend/glasm/emit_glasm_bitwise_conversion.cpp b/src/shader_recompiler/backend/glasm/emit_glasm_bitwise_conversion.cpp
index 5bfdecc09..2fc2a0ac6 100644
--- a/src/shader_recompiler/backend/glasm/emit_glasm_bitwise_conversion.cpp
+++ b/src/shader_recompiler/backend/glasm/emit_glasm_bitwise_conversion.cpp
@@ -43,10 +43,6 @@ void EmitBitCastU64F64(EmitContext&, IR::Inst& inst, const IR::Value& value) {
43 Alias(inst, value); 43 Alias(inst, value);
44} 44}
45 45
46void EmitBitCastS32F32(EmitContext&, IR::Inst& inst, const IR::Value& value) {
47 Alias(inst, value);
48}
49
50void EmitBitCastF16U16(EmitContext&, IR::Inst& inst, const IR::Value& value) { 46void EmitBitCastF16U16(EmitContext&, IR::Inst& inst, const IR::Value& value) {
51 Alias(inst, value); 47 Alias(inst, value);
52} 48}
diff --git a/src/shader_recompiler/backend/glasm/emit_glasm_image.cpp b/src/shader_recompiler/backend/glasm/emit_glasm_image.cpp
index e67e80fac..cf6065208 100644
--- a/src/shader_recompiler/backend/glasm/emit_glasm_image.cpp
+++ b/src/shader_recompiler/backend/glasm/emit_glasm_image.cpp
@@ -59,7 +59,7 @@ std::string Image(EmitContext& ctx, IR::TextureInstInfo info,
59 } 59 }
60} 60}
61 61
62std::string_view TextureType(IR::TextureInstInfo info) { 62std::string_view TextureType(IR::TextureInstInfo info, bool is_ms = false) {
63 if (info.is_depth) { 63 if (info.is_depth) {
64 switch (info.type) { 64 switch (info.type) {
65 case TextureType::Color1D: 65 case TextureType::Color1D:
@@ -88,9 +88,9 @@ std::string_view TextureType(IR::TextureInstInfo info) {
88 return "ARRAY1D"; 88 return "ARRAY1D";
89 case TextureType::Color2D: 89 case TextureType::Color2D:
90 case TextureType::Color2DRect: 90 case TextureType::Color2DRect:
91 return "2D"; 91 return is_ms ? "2DMS" : "2D";
92 case TextureType::ColorArray2D: 92 case TextureType::ColorArray2D:
93 return "ARRAY2D"; 93 return is_ms ? "ARRAY2DMS" : "ARRAY2D";
94 case TextureType::Color3D: 94 case TextureType::Color3D:
95 return "3D"; 95 return "3D";
96 case TextureType::ColorCube: 96 case TextureType::ColorCube:
@@ -510,15 +510,16 @@ void EmitImageFetch(EmitContext& ctx, IR::Inst& inst, const IR::Value& index,
510 const IR::Value& coord, const IR::Value& offset, ScalarS32 lod, ScalarS32 ms) { 510 const IR::Value& coord, const IR::Value& offset, ScalarS32 lod, ScalarS32 ms) {
511 const auto info{inst.Flags<IR::TextureInstInfo>()}; 511 const auto info{inst.Flags<IR::TextureInstInfo>()};
512 const auto sparse_inst{PrepareSparse(inst)}; 512 const auto sparse_inst{PrepareSparse(inst)};
513 const bool is_multisample{ms.type != Type::Void};
513 const std::string_view sparse_mod{sparse_inst ? ".SPARSE" : ""}; 514 const std::string_view sparse_mod{sparse_inst ? ".SPARSE" : ""};
514 const std::string_view type{TextureType(info)}; 515 const std::string_view type{TextureType(info, is_multisample)};
515 const std::string texture{Texture(ctx, info, index)}; 516 const std::string texture{Texture(ctx, info, index)};
516 const std::string offset_vec{Offset(ctx, offset)}; 517 const std::string offset_vec{Offset(ctx, offset)};
517 const auto [coord_vec, coord_alloc]{Coord(ctx, coord)}; 518 const auto [coord_vec, coord_alloc]{Coord(ctx, coord)};
518 const Register ret{ctx.reg_alloc.Define(inst)}; 519 const Register ret{ctx.reg_alloc.Define(inst)};
519 if (info.type == TextureType::Buffer) { 520 if (info.type == TextureType::Buffer) {
520 ctx.Add("TXF.F{} {},{},{},{}{};", sparse_mod, ret, coord_vec, texture, type, offset_vec); 521 ctx.Add("TXF.F{} {},{},{},{}{};", sparse_mod, ret, coord_vec, texture, type, offset_vec);
521 } else if (ms.type != Type::Void) { 522 } else if (is_multisample) {
522 ctx.Add("MOV.S {}.w,{};" 523 ctx.Add("MOV.S {}.w,{};"
523 "TXFMS.F{} {},{},{},{}{};", 524 "TXFMS.F{} {},{},{},{}{};",
524 coord_vec, ms, sparse_mod, ret, coord_vec, texture, type, offset_vec); 525 coord_vec, ms, sparse_mod, ret, coord_vec, texture, type, offset_vec);
diff --git a/src/shader_recompiler/backend/glasm/emit_glasm_instructions.h b/src/shader_recompiler/backend/glasm/emit_glasm_instructions.h
index eaaf9ba39..415a249e4 100644
--- a/src/shader_recompiler/backend/glasm/emit_glasm_instructions.h
+++ b/src/shader_recompiler/backend/glasm/emit_glasm_instructions.h
@@ -197,7 +197,6 @@ void EmitSelectF64(EmitContext& ctx, ScalarS32 cond, Register true_value, Regist
197void EmitBitCastU16F16(EmitContext& ctx, IR::Inst& inst, const IR::Value& value); 197void EmitBitCastU16F16(EmitContext& ctx, IR::Inst& inst, const IR::Value& value);
198void EmitBitCastU32F32(EmitContext& ctx, IR::Inst& inst, const IR::Value& value); 198void EmitBitCastU32F32(EmitContext& ctx, IR::Inst& inst, const IR::Value& value);
199void EmitBitCastU64F64(EmitContext& ctx, IR::Inst& inst, const IR::Value& value); 199void EmitBitCastU64F64(EmitContext& ctx, IR::Inst& inst, const IR::Value& value);
200void EmitBitCastS32F32(EmitContext& ctx, IR::Inst& inst, const IR::Value& value);
201void EmitBitCastF16U16(EmitContext& ctx, IR::Inst& inst, const IR::Value& value); 200void EmitBitCastF16U16(EmitContext& ctx, IR::Inst& inst, const IR::Value& value);
202void EmitBitCastF32U32(EmitContext& ctx, IR::Inst& inst, const IR::Value& value); 201void EmitBitCastF32U32(EmitContext& ctx, IR::Inst& inst, const IR::Value& value);
203void EmitBitCastF64U64(EmitContext& ctx, IR::Inst& inst, const IR::Value& value); 202void EmitBitCastF64U64(EmitContext& ctx, IR::Inst& inst, const IR::Value& value);
diff --git a/src/shader_recompiler/backend/glsl/emit_glsl_bitwise_conversion.cpp b/src/shader_recompiler/backend/glsl/emit_glsl_bitwise_conversion.cpp
index 8e5e6cf1f..1be4a0f59 100644
--- a/src/shader_recompiler/backend/glsl/emit_glsl_bitwise_conversion.cpp
+++ b/src/shader_recompiler/backend/glsl/emit_glsl_bitwise_conversion.cpp
@@ -48,10 +48,6 @@ void EmitBitCastU64F64(EmitContext& ctx, IR::Inst& inst, std::string_view value)
48 ctx.AddU64("{}=doubleBitsToUint64({});", inst, value); 48 ctx.AddU64("{}=doubleBitsToUint64({});", inst, value);
49} 49}
50 50
51void EmitBitCastS32F32(EmitContext& ctx, IR::Inst& inst, std::string_view value) {
52 ctx.AddF32("{}=ftoi({});", inst, value);
53}
54
55void EmitBitCastF16U16([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& inst) { 51void EmitBitCastF16U16([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& inst) {
56 NotImplemented(); 52 NotImplemented();
57} 53}
diff --git a/src/shader_recompiler/backend/glsl/emit_glsl_image.cpp b/src/shader_recompiler/backend/glsl/emit_glsl_image.cpp
index cecdbb9d6..d8874b0cc 100644
--- a/src/shader_recompiler/backend/glsl/emit_glsl_image.cpp
+++ b/src/shader_recompiler/backend/glsl/emit_glsl_image.cpp
@@ -414,7 +414,7 @@ void EmitImageGatherDref(EmitContext& ctx, IR::Inst& inst, const IR::Value& inde
414 414
415void EmitImageFetch(EmitContext& ctx, IR::Inst& inst, const IR::Value& index, 415void EmitImageFetch(EmitContext& ctx, IR::Inst& inst, const IR::Value& index,
416 std::string_view coords, std::string_view offset, std::string_view lod, 416 std::string_view coords, std::string_view offset, std::string_view lod,
417 [[maybe_unused]] std::string_view ms) { 417 std::string_view ms) {
418 const auto info{inst.Flags<IR::TextureInstInfo>()}; 418 const auto info{inst.Flags<IR::TextureInstInfo>()};
419 if (info.has_bias) { 419 if (info.has_bias) {
420 throw NotImplementedException("EmitImageFetch Bias texture samples"); 420 throw NotImplementedException("EmitImageFetch Bias texture samples");
@@ -431,19 +431,24 @@ void EmitImageFetch(EmitContext& ctx, IR::Inst& inst, const IR::Value& index,
431 ctx.AddU1("{}=true;", *sparse_inst); 431 ctx.AddU1("{}=true;", *sparse_inst);
432 } 432 }
433 if (!sparse_inst || !supports_sparse) { 433 if (!sparse_inst || !supports_sparse) {
434 if (!offset.empty()) { 434 const auto int_coords{CoordsCastToInt(coords, info)};
435 ctx.Add("{}=texelFetchOffset({},{},int({}),{});", texel, texture, 435 if (!ms.empty()) {
436 CoordsCastToInt(coords, info), lod, CoordsCastToInt(offset, info)); 436 ctx.Add("{}=texelFetch({},{},int({}));", texel, texture, int_coords, ms);
437 } else if (!offset.empty()) {
438 ctx.Add("{}=texelFetchOffset({},{},int({}),{});", texel, texture, int_coords, lod,
439 CoordsCastToInt(offset, info));
437 } else { 440 } else {
438 if (info.type == TextureType::Buffer) { 441 if (info.type == TextureType::Buffer) {
439 ctx.Add("{}=texelFetch({},int({}));", texel, texture, coords); 442 ctx.Add("{}=texelFetch({},int({}));", texel, texture, coords);
440 } else { 443 } else {
441 ctx.Add("{}=texelFetch({},{},int({}));", texel, texture, 444 ctx.Add("{}=texelFetch({},{},int({}));", texel, texture, int_coords, lod);
442 CoordsCastToInt(coords, info), lod);
443 } 445 }
444 } 446 }
445 return; 447 return;
446 } 448 }
449 if (!ms.empty()) {
450 throw NotImplementedException("EmitImageFetch Sparse MSAA samples");
451 }
447 if (!offset.empty()) { 452 if (!offset.empty()) {
448 ctx.AddU1("{}=sparseTexelsResidentARB(sparseTexelFetchOffsetARB({},{},int({}),{},{}));", 453 ctx.AddU1("{}=sparseTexelsResidentARB(sparseTexelFetchOffsetARB({},{},int({}),{},{}));",
449 *sparse_inst, texture, CastToIntVec(coords, info), lod, 454 *sparse_inst, texture, CastToIntVec(coords, info), lod,
diff --git a/src/shader_recompiler/backend/glsl/emit_glsl_instructions.h b/src/shader_recompiler/backend/glsl/emit_glsl_instructions.h
index 4151c89de..c6df1dba7 100644
--- a/src/shader_recompiler/backend/glsl/emit_glsl_instructions.h
+++ b/src/shader_recompiler/backend/glsl/emit_glsl_instructions.h
@@ -231,7 +231,6 @@ void EmitSelectF64(EmitContext& ctx, IR::Inst& inst, std::string_view cond,
231void EmitBitCastU16F16(EmitContext& ctx, IR::Inst& inst); 231void EmitBitCastU16F16(EmitContext& ctx, IR::Inst& inst);
232void EmitBitCastU32F32(EmitContext& ctx, IR::Inst& inst, std::string_view value); 232void EmitBitCastU32F32(EmitContext& ctx, IR::Inst& inst, std::string_view value);
233void EmitBitCastU64F64(EmitContext& ctx, IR::Inst& inst, std::string_view value); 233void EmitBitCastU64F64(EmitContext& ctx, IR::Inst& inst, std::string_view value);
234void EmitBitCastS32F32(EmitContext& ctx, IR::Inst& inst, std::string_view value);
235void EmitBitCastF16U16(EmitContext& ctx, IR::Inst& inst); 234void EmitBitCastF16U16(EmitContext& ctx, IR::Inst& inst);
236void EmitBitCastF32U32(EmitContext& ctx, IR::Inst& inst, std::string_view value); 235void EmitBitCastF32U32(EmitContext& ctx, IR::Inst& inst, std::string_view value);
237void EmitBitCastF64U64(EmitContext& ctx, IR::Inst& inst, std::string_view value); 236void EmitBitCastF64U64(EmitContext& ctx, IR::Inst& inst, std::string_view value);
diff --git a/src/shader_recompiler/backend/glsl/glsl_emit_context.cpp b/src/shader_recompiler/backend/glsl/glsl_emit_context.cpp
index 5d01ec0cd..1b006e811 100644
--- a/src/shader_recompiler/backend/glsl/glsl_emit_context.cpp
+++ b/src/shader_recompiler/backend/glsl/glsl_emit_context.cpp
@@ -61,24 +61,28 @@ std::string OutputDecorator(Stage stage, u32 size) {
61 } 61 }
62} 62}
63 63
64std::string_view SamplerType(TextureType type, bool is_depth) { 64std::string_view DepthSamplerType(TextureType type) {
65 if (is_depth) { 65 switch (type) {
66 switch (type) { 66 case TextureType::Color1D:
67 case TextureType::Color1D: 67 return "sampler1DShadow";
68 return "sampler1DShadow"; 68 case TextureType::ColorArray1D:
69 case TextureType::ColorArray1D: 69 return "sampler1DArrayShadow";
70 return "sampler1DArrayShadow"; 70 case TextureType::Color2D:
71 case TextureType::Color2D: 71 return "sampler2DShadow";
72 return "sampler2DShadow"; 72 case TextureType::ColorArray2D:
73 case TextureType::ColorArray2D: 73 return "sampler2DArrayShadow";
74 return "sampler2DArrayShadow"; 74 case TextureType::ColorCube:
75 case TextureType::ColorCube: 75 return "samplerCubeShadow";
76 return "samplerCubeShadow"; 76 case TextureType::ColorArrayCube:
77 case TextureType::ColorArrayCube: 77 return "samplerCubeArrayShadow";
78 return "samplerCubeArrayShadow"; 78 default:
79 default: 79 throw NotImplementedException("Texture type: {}", type);
80 throw NotImplementedException("Texture type: {}", type); 80 }
81 } 81}
82
83std::string_view ColorSamplerType(TextureType type, bool is_multisample = false) {
84 if (is_multisample) {
85 ASSERT(type == TextureType::Color2D || type == TextureType::ColorArray2D);
82 } 86 }
83 switch (type) { 87 switch (type) {
84 case TextureType::Color1D: 88 case TextureType::Color1D:
@@ -87,9 +91,9 @@ std::string_view SamplerType(TextureType type, bool is_depth) {
87 return "sampler1DArray"; 91 return "sampler1DArray";
88 case TextureType::Color2D: 92 case TextureType::Color2D:
89 case TextureType::Color2DRect: 93 case TextureType::Color2DRect:
90 return "sampler2D"; 94 return is_multisample ? "sampler2DMS" : "sampler2D";
91 case TextureType::ColorArray2D: 95 case TextureType::ColorArray2D:
92 return "sampler2DArray"; 96 return is_multisample ? "sampler2DMSArray" : "sampler2DArray";
93 case TextureType::Color3D: 97 case TextureType::Color3D:
94 return "sampler3D"; 98 return "sampler3D";
95 case TextureType::ColorCube: 99 case TextureType::ColorCube:
@@ -677,7 +681,7 @@ void EmitContext::SetupTextures(Bindings& bindings) {
677 texture_buffers.reserve(info.texture_buffer_descriptors.size()); 681 texture_buffers.reserve(info.texture_buffer_descriptors.size());
678 for (const auto& desc : info.texture_buffer_descriptors) { 682 for (const auto& desc : info.texture_buffer_descriptors) {
679 texture_buffers.push_back({bindings.texture, desc.count}); 683 texture_buffers.push_back({bindings.texture, desc.count});
680 const auto sampler_type{SamplerType(TextureType::Buffer, false)}; 684 const auto sampler_type{ColorSamplerType(TextureType::Buffer)};
681 const auto array_decorator{desc.count > 1 ? fmt::format("[{}]", desc.count) : ""}; 685 const auto array_decorator{desc.count > 1 ? fmt::format("[{}]", desc.count) : ""};
682 header += fmt::format("layout(binding={}) uniform {} tex{}{};", bindings.texture, 686 header += fmt::format("layout(binding={}) uniform {} tex{}{};", bindings.texture,
683 sampler_type, bindings.texture, array_decorator); 687 sampler_type, bindings.texture, array_decorator);
@@ -686,7 +690,8 @@ void EmitContext::SetupTextures(Bindings& bindings) {
686 textures.reserve(info.texture_descriptors.size()); 690 textures.reserve(info.texture_descriptors.size());
687 for (const auto& desc : info.texture_descriptors) { 691 for (const auto& desc : info.texture_descriptors) {
688 textures.push_back({bindings.texture, desc.count}); 692 textures.push_back({bindings.texture, desc.count});
689 const auto sampler_type{SamplerType(desc.type, desc.is_depth)}; 693 const auto sampler_type{desc.is_depth ? DepthSamplerType(desc.type)
694 : ColorSamplerType(desc.type, desc.is_multisample)};
690 const auto array_decorator{desc.count > 1 ? fmt::format("[{}]", desc.count) : ""}; 695 const auto array_decorator{desc.count > 1 ? fmt::format("[{}]", desc.count) : ""};
691 header += fmt::format("layout(binding={}) uniform {} tex{}{};", bindings.texture, 696 header += fmt::format("layout(binding={}) uniform {} tex{}{};", bindings.texture,
692 sampler_type, bindings.texture, array_decorator); 697 sampler_type, bindings.texture, array_decorator);
diff --git a/src/shader_recompiler/backend/spirv/emit_spirv_bitwise_conversion.cpp b/src/shader_recompiler/backend/spirv/emit_spirv_bitwise_conversion.cpp
index 50daacd95..c4ca28d11 100644
--- a/src/shader_recompiler/backend/spirv/emit_spirv_bitwise_conversion.cpp
+++ b/src/shader_recompiler/backend/spirv/emit_spirv_bitwise_conversion.cpp
@@ -18,10 +18,6 @@ void EmitBitCastU64F64(EmitContext&) {
18 throw NotImplementedException("SPIR-V Instruction"); 18 throw NotImplementedException("SPIR-V Instruction");
19} 19}
20 20
21void EmitBitCastS32F32(EmitContext&) {
22 throw NotImplementedException("SPIR-V Instruction");
23}
24
25void EmitBitCastF16U16(EmitContext&) { 21void EmitBitCastF16U16(EmitContext&) {
26 throw NotImplementedException("SPIR-V Instruction"); 22 throw NotImplementedException("SPIR-V Instruction");
27} 23}
diff --git a/src/shader_recompiler/backend/spirv/emit_spirv_instructions.h b/src/shader_recompiler/backend/spirv/emit_spirv_instructions.h
index e31cdc5e8..db12e8176 100644
--- a/src/shader_recompiler/backend/spirv/emit_spirv_instructions.h
+++ b/src/shader_recompiler/backend/spirv/emit_spirv_instructions.h
@@ -179,7 +179,6 @@ Id EmitSelectF64(EmitContext& ctx, Id cond, Id true_value, Id false_value);
179void EmitBitCastU16F16(EmitContext& ctx); 179void EmitBitCastU16F16(EmitContext& ctx);
180Id EmitBitCastU32F32(EmitContext& ctx, Id value); 180Id EmitBitCastU32F32(EmitContext& ctx, Id value);
181void EmitBitCastU64F64(EmitContext& ctx); 181void EmitBitCastU64F64(EmitContext& ctx);
182void EmitBitCastS32F32(EmitContext& ctx);
183void EmitBitCastF16U16(EmitContext&); 182void EmitBitCastF16U16(EmitContext&);
184Id EmitBitCastF32U32(EmitContext& ctx, Id value); 183Id EmitBitCastF32U32(EmitContext& ctx, Id value);
185void EmitBitCastF64U64(EmitContext& ctx); 184void EmitBitCastF64U64(EmitContext& ctx);
diff --git a/src/shader_recompiler/frontend/ir/ir_emitter.cpp b/src/shader_recompiler/frontend/ir/ir_emitter.cpp
index eb2e49a68..430797d23 100644
--- a/src/shader_recompiler/frontend/ir/ir_emitter.cpp
+++ b/src/shader_recompiler/frontend/ir/ir_emitter.cpp
@@ -704,11 +704,6 @@ IR::U32 IREmitter::BitCast<IR::U32, IR::F32>(const IR::F32& value) {
704} 704}
705 705
706template <> 706template <>
707IR::S32 IREmitter::BitCast<IR::S32, IR::F32>(const IR::F32& value) {
708 return Inst<IR::S32>(Opcode::BitCastS32F32, value);
709}
710
711template <>
712IR::F32 IREmitter::BitCast<IR::F32, IR::U32>(const IR::U32& value) { 707IR::F32 IREmitter::BitCast<IR::F32, IR::U32>(const IR::U32& value) {
713 return Inst<IR::F32>(Opcode::BitCastF32U32, value); 708 return Inst<IR::F32>(Opcode::BitCastF32U32, value);
714} 709}
diff --git a/src/shader_recompiler/frontend/ir/opcodes.h b/src/shader_recompiler/frontend/ir/opcodes.h
index d155afd0f..e300714f3 100644
--- a/src/shader_recompiler/frontend/ir/opcodes.h
+++ b/src/shader_recompiler/frontend/ir/opcodes.h
@@ -38,7 +38,6 @@ constexpr Type U8{Type::U8};
38constexpr Type U16{Type::U16}; 38constexpr Type U16{Type::U16};
39constexpr Type U32{Type::U32}; 39constexpr Type U32{Type::U32};
40constexpr Type U64{Type::U64}; 40constexpr Type U64{Type::U64};
41constexpr Type S32{Type::S32};
42constexpr Type F16{Type::F16}; 41constexpr Type F16{Type::F16};
43constexpr Type F32{Type::F32}; 42constexpr Type F32{Type::F32};
44constexpr Type F64{Type::F64}; 43constexpr Type F64{Type::F64};
diff --git a/src/shader_recompiler/frontend/ir/opcodes.inc b/src/shader_recompiler/frontend/ir/opcodes.inc
index 1fe3749cc..24e82f802 100644
--- a/src/shader_recompiler/frontend/ir/opcodes.inc
+++ b/src/shader_recompiler/frontend/ir/opcodes.inc
@@ -175,7 +175,6 @@ OPCODE(SelectF64, F64, U1,
175OPCODE(BitCastU16F16, U16, F16, ) 175OPCODE(BitCastU16F16, U16, F16, )
176OPCODE(BitCastU32F32, U32, F32, ) 176OPCODE(BitCastU32F32, U32, F32, )
177OPCODE(BitCastU64F64, U64, F64, ) 177OPCODE(BitCastU64F64, U64, F64, )
178OPCODE(BitCastS32F32, S32, F32, )
179OPCODE(BitCastF16U16, F16, U16, ) 178OPCODE(BitCastF16U16, F16, U16, )
180OPCODE(BitCastF32U32, F32, U32, ) 179OPCODE(BitCastF32U32, F32, U32, )
181OPCODE(BitCastF64U64, F64, U64, ) 180OPCODE(BitCastF64U64, F64, U64, )
diff --git a/src/shader_recompiler/frontend/ir/type.h b/src/shader_recompiler/frontend/ir/type.h
index 5a7c706ad..04c8c4ddb 100644
--- a/src/shader_recompiler/frontend/ir/type.h
+++ b/src/shader_recompiler/frontend/ir/type.h
@@ -24,22 +24,21 @@ enum class Type {
24 U16 = 1 << 7, 24 U16 = 1 << 7,
25 U32 = 1 << 8, 25 U32 = 1 << 8,
26 U64 = 1 << 9, 26 U64 = 1 << 9,
27 S32 = 1 << 10, 27 F16 = 1 << 10,
28 F16 = 1 << 11, 28 F32 = 1 << 11,
29 F32 = 1 << 12, 29 F64 = 1 << 12,
30 F64 = 1 << 13, 30 U32x2 = 1 << 13,
31 U32x2 = 1 << 14, 31 U32x3 = 1 << 14,
32 U32x3 = 1 << 15, 32 U32x4 = 1 << 15,
33 U32x4 = 1 << 16, 33 F16x2 = 1 << 16,
34 F16x2 = 1 << 17, 34 F16x3 = 1 << 17,
35 F16x3 = 1 << 18, 35 F16x4 = 1 << 18,
36 F16x4 = 1 << 19, 36 F32x2 = 1 << 19,
37 F32x2 = 1 << 20, 37 F32x3 = 1 << 20,
38 F32x3 = 1 << 21, 38 F32x4 = 1 << 21,
39 F32x4 = 1 << 22, 39 F64x2 = 1 << 22,
40 F64x2 = 1 << 23, 40 F64x3 = 1 << 23,
41 F64x3 = 1 << 24, 41 F64x4 = 1 << 24,
42 F64x4 = 1 << 25,
43}; 42};
44DECLARE_ENUM_FLAG_OPERATORS(Type) 43DECLARE_ENUM_FLAG_OPERATORS(Type)
45 44
diff --git a/src/shader_recompiler/frontend/ir/value.cpp b/src/shader_recompiler/frontend/ir/value.cpp
index 30ba12316..346169328 100644
--- a/src/shader_recompiler/frontend/ir/value.cpp
+++ b/src/shader_recompiler/frontend/ir/value.cpp
@@ -23,8 +23,6 @@ Value::Value(u16 value) noexcept : type{Type::U16}, imm_u16{value} {}
23 23
24Value::Value(u32 value) noexcept : type{Type::U32}, imm_u32{value} {} 24Value::Value(u32 value) noexcept : type{Type::U32}, imm_u32{value} {}
25 25
26Value::Value(s32 value) noexcept : type{Type::S32}, imm_s32{value} {}
27
28Value::Value(f32 value) noexcept : type{Type::F32}, imm_f32{value} {} 26Value::Value(f32 value) noexcept : type{Type::F32}, imm_f32{value} {}
29 27
30Value::Value(u64 value) noexcept : type{Type::U64}, imm_u64{value} {} 28Value::Value(u64 value) noexcept : type{Type::U64}, imm_u64{value} {}
@@ -71,7 +69,6 @@ bool Value::operator==(const Value& other) const {
71 return imm_u16 == other.imm_u16; 69 return imm_u16 == other.imm_u16;
72 case Type::U32: 70 case Type::U32:
73 case Type::F32: 71 case Type::F32:
74 case Type::S32:
75 return imm_u32 == other.imm_u32; 72 return imm_u32 == other.imm_u32;
76 case Type::U64: 73 case Type::U64:
77 case Type::F64: 74 case Type::F64:
diff --git a/src/shader_recompiler/frontend/ir/value.h b/src/shader_recompiler/frontend/ir/value.h
index 8b34356fd..883dfa24e 100644
--- a/src/shader_recompiler/frontend/ir/value.h
+++ b/src/shader_recompiler/frontend/ir/value.h
@@ -268,7 +268,6 @@ using U8 = TypedValue<Type::U8>;
268using U16 = TypedValue<Type::U16>; 268using U16 = TypedValue<Type::U16>;
269using U32 = TypedValue<Type::U32>; 269using U32 = TypedValue<Type::U32>;
270using U64 = TypedValue<Type::U64>; 270using U64 = TypedValue<Type::U64>;
271using S32 = TypedValue<Type::S32>;
272using F16 = TypedValue<Type::F16>; 271using F16 = TypedValue<Type::F16>;
273using F32 = TypedValue<Type::F32>; 272using F32 = TypedValue<Type::F32>;
274using F64 = TypedValue<Type::F64>; 273using F64 = TypedValue<Type::F64>;
diff --git a/src/shader_recompiler/ir_opt/texture_pass.cpp b/src/shader_recompiler/ir_opt/texture_pass.cpp
index 9718c6921..70b620dcb 100644
--- a/src/shader_recompiler/ir_opt/texture_pass.cpp
+++ b/src/shader_recompiler/ir_opt/texture_pass.cpp
@@ -486,10 +486,10 @@ void PatchTexelFetch(IR::Block& block, IR::Inst& inst, TexturePixelFormat pixel_
486 const IR::F32 w(ir.CompositeExtract(new_inst, 3)); 486 const IR::F32 w(ir.CompositeExtract(new_inst, 3));
487 const IR::F16F32F64 max_value(ir.Imm32(get_max_value())); 487 const IR::F16F32F64 max_value(ir.Imm32(get_max_value()));
488 const IR::Value converted = 488 const IR::Value converted =
489 ir.CompositeConstruct(ir.FPMul(ir.ConvertSToF(32, 32, ir.BitCast<IR::S32>(x)), max_value), 489 ir.CompositeConstruct(ir.FPMul(ir.ConvertSToF(32, 32, ir.BitCast<IR::U32>(x)), max_value),
490 ir.FPMul(ir.ConvertSToF(32, 32, ir.BitCast<IR::S32>(y)), max_value), 490 ir.FPMul(ir.ConvertSToF(32, 32, ir.BitCast<IR::U32>(y)), max_value),
491 ir.FPMul(ir.ConvertSToF(32, 32, ir.BitCast<IR::S32>(z)), max_value), 491 ir.FPMul(ir.ConvertSToF(32, 32, ir.BitCast<IR::U32>(z)), max_value),
492 ir.FPMul(ir.ConvertSToF(32, 32, ir.BitCast<IR::S32>(w)), max_value)); 492 ir.FPMul(ir.ConvertSToF(32, 32, ir.BitCast<IR::U32>(w)), max_value));
493 inst.ReplaceUsesWith(converted); 493 inst.ReplaceUsesWith(converted);
494} 494}
495} // Anonymous namespace 495} // Anonymous namespace
diff --git a/src/video_core/CMakeLists.txt b/src/video_core/CMakeLists.txt
index b474eb363..4742bcbe9 100644
--- a/src/video_core/CMakeLists.txt
+++ b/src/video_core/CMakeLists.txt
@@ -52,6 +52,8 @@ add_library(video_core STATIC
52 engines/puller.cpp 52 engines/puller.cpp
53 engines/puller.h 53 engines/puller.h
54 framebuffer_config.h 54 framebuffer_config.h
55 fsr.cpp
56 fsr.h
55 host1x/codecs/codec.cpp 57 host1x/codecs/codec.cpp
56 host1x/codecs/codec.h 58 host1x/codecs/codec.h
57 host1x/codecs/h264.cpp 59 host1x/codecs/h264.cpp
@@ -110,6 +112,8 @@ add_library(video_core STATIC
110 renderer_opengl/gl_device.h 112 renderer_opengl/gl_device.h
111 renderer_opengl/gl_fence_manager.cpp 113 renderer_opengl/gl_fence_manager.cpp
112 renderer_opengl/gl_fence_manager.h 114 renderer_opengl/gl_fence_manager.h
115 renderer_opengl/gl_fsr.cpp
116 renderer_opengl/gl_fsr.h
113 renderer_opengl/gl_graphics_pipeline.cpp 117 renderer_opengl/gl_graphics_pipeline.cpp
114 renderer_opengl/gl_graphics_pipeline.h 118 renderer_opengl/gl_graphics_pipeline.h
115 renderer_opengl/gl_rasterizer.cpp 119 renderer_opengl/gl_rasterizer.cpp
diff --git a/src/video_core/fsr.cpp b/src/video_core/fsr.cpp
new file mode 100644
index 000000000..5653c64fc
--- /dev/null
+++ b/src/video_core/fsr.cpp
@@ -0,0 +1,148 @@
1// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
2// SPDX-License-Identifier: GPL-2.0-or-later
3
4#include <cmath>
5#include "video_core/fsr.h"
6
7namespace FSR {
8namespace {
9// Reimplementations of the constant generating functions in ffx_fsr1.h
10// GCC generated a lot of warnings when using the official header.
11u32 AU1_AH1_AF1(f32 f) {
12 static constexpr u32 base[512]{
13 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
14 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
15 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
16 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
17 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
18 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
19 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
20 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
21 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
22 0x0000, 0x0000, 0x0000, 0x0000, 0x0001, 0x0002, 0x0004, 0x0008, 0x0010, 0x0020, 0x0040,
23 0x0080, 0x0100, 0x0200, 0x0400, 0x0800, 0x0c00, 0x1000, 0x1400, 0x1800, 0x1c00, 0x2000,
24 0x2400, 0x2800, 0x2c00, 0x3000, 0x3400, 0x3800, 0x3c00, 0x4000, 0x4400, 0x4800, 0x4c00,
25 0x5000, 0x5400, 0x5800, 0x5c00, 0x6000, 0x6400, 0x6800, 0x6c00, 0x7000, 0x7400, 0x7800,
26 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
27 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
28 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
29 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
30 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
31 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
32 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
33 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
34 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
35 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
36 0x7bff, 0x7bff, 0x7bff, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
37 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
38 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
39 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
40 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
41 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
42 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
43 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
44 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
45 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8001, 0x8002, 0x8004, 0x8008,
46 0x8010, 0x8020, 0x8040, 0x8080, 0x8100, 0x8200, 0x8400, 0x8800, 0x8c00, 0x9000, 0x9400,
47 0x9800, 0x9c00, 0xa000, 0xa400, 0xa800, 0xac00, 0xb000, 0xb400, 0xb800, 0xbc00, 0xc000,
48 0xc400, 0xc800, 0xcc00, 0xd000, 0xd400, 0xd800, 0xdc00, 0xe000, 0xe400, 0xe800, 0xec00,
49 0xf000, 0xf400, 0xf800, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
50 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
51 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
52 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
53 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
54 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
55 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
56 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
57 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
58 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
59 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
60 };
61 static constexpr s8 shift[512]{
62 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
63 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
64 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
65 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
66 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
67 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
68 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x17, 0x16,
69 0x15, 0x14, 0x13, 0x12, 0x11, 0x10, 0x0f, 0x0e, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d,
70 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d,
71 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
72 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
73 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
74 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
75 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
76 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
77 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
78 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
79 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
80 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
81 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
82 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
83 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
84 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
85 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x17,
86 0x16, 0x15, 0x14, 0x13, 0x12, 0x11, 0x10, 0x0f, 0x0e, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d,
87 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d,
88 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
89 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
90 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
91 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
92 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
93 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
94 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
95 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
96 0x18, 0x18,
97 };
98 const u32 u = Common::BitCast<u32>(f);
99 const u32 i = u >> 23;
100 return base[i] + ((u & 0x7fffff) >> shift[i]);
101}
102
103u32 AU1_AH2_AF2(f32 a[2]) {
104 return AU1_AH1_AF1(a[0]) + (AU1_AH1_AF1(a[1]) << 16);
105}
106
107void FsrEasuCon(u32 con0[4], u32 con1[4], u32 con2[4], u32 con3[4], f32 inputViewportInPixelsX,
108 f32 inputViewportInPixelsY, f32 inputSizeInPixelsX, f32 inputSizeInPixelsY,
109 f32 outputSizeInPixelsX, f32 outputSizeInPixelsY) {
110 con0[0] = Common::BitCast<u32>(inputViewportInPixelsX / outputSizeInPixelsX);
111 con0[1] = Common::BitCast<u32>(inputViewportInPixelsY / outputSizeInPixelsY);
112 con0[2] = Common::BitCast<u32>(0.5f * inputViewportInPixelsX / outputSizeInPixelsX - 0.5f);
113 con0[3] = Common::BitCast<u32>(0.5f * inputViewportInPixelsY / outputSizeInPixelsY - 0.5f);
114 con1[0] = Common::BitCast<u32>(1.0f / inputSizeInPixelsX);
115 con1[1] = Common::BitCast<u32>(1.0f / inputSizeInPixelsY);
116 con1[2] = Common::BitCast<u32>(1.0f / inputSizeInPixelsX);
117 con1[3] = Common::BitCast<u32>(-1.0f / inputSizeInPixelsY);
118 con2[0] = Common::BitCast<u32>(-1.0f / inputSizeInPixelsX);
119 con2[1] = Common::BitCast<u32>(2.0f / inputSizeInPixelsY);
120 con2[2] = Common::BitCast<u32>(1.0f / inputSizeInPixelsX);
121 con2[3] = Common::BitCast<u32>(2.0f / inputSizeInPixelsY);
122 con3[0] = Common::BitCast<u32>(0.0f / inputSizeInPixelsX);
123 con3[1] = Common::BitCast<u32>(4.0f / inputSizeInPixelsY);
124 con3[2] = con3[3] = 0;
125}
126} // Anonymous namespace
127
128void FsrEasuConOffset(u32 con0[4], u32 con1[4], u32 con2[4], u32 con3[4],
129 f32 inputViewportInPixelsX, f32 inputViewportInPixelsY,
130 f32 inputSizeInPixelsX, f32 inputSizeInPixelsY, f32 outputSizeInPixelsX,
131 f32 outputSizeInPixelsY, f32 inputOffsetInPixelsX, f32 inputOffsetInPixelsY) {
132 FsrEasuCon(con0, con1, con2, con3, inputViewportInPixelsX, inputViewportInPixelsY,
133 inputSizeInPixelsX, inputSizeInPixelsY, outputSizeInPixelsX, outputSizeInPixelsY);
134 con0[2] = Common::BitCast<u32>(0.5f * inputViewportInPixelsX / outputSizeInPixelsX - 0.5f +
135 inputOffsetInPixelsX);
136 con0[3] = Common::BitCast<u32>(0.5f * inputViewportInPixelsY / outputSizeInPixelsY - 0.5f +
137 inputOffsetInPixelsY);
138}
139
140void FsrRcasCon(u32* con, f32 sharpness) {
141 sharpness = std::exp2f(-sharpness);
142 f32 hSharp[2]{sharpness, sharpness};
143 con[0] = Common::BitCast<u32>(sharpness);
144 con[1] = AU1_AH2_AF2(hSharp);
145 con[2] = 0;
146 con[3] = 0;
147}
148} // namespace FSR
diff --git a/src/video_core/fsr.h b/src/video_core/fsr.h
new file mode 100644
index 000000000..db0d4ec6f
--- /dev/null
+++ b/src/video_core/fsr.h
@@ -0,0 +1,19 @@
1// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
2// SPDX-License-Identifier: GPL-2.0-or-later
3
4#pragma once
5
6#include "common/bit_cast.h"
7#include "common/common_types.h"
8
9namespace FSR {
10// Reimplementations of the constant generating functions in ffx_fsr1.h
11// GCC generated a lot of warnings when using the official header.
12void FsrEasuConOffset(u32 con0[4], u32 con1[4], u32 con2[4], u32 con3[4],
13 f32 inputViewportInPixelsX, f32 inputViewportInPixelsY,
14 f32 inputSizeInPixelsX, f32 inputSizeInPixelsY, f32 outputSizeInPixelsX,
15 f32 outputSizeInPixelsY, f32 inputOffsetInPixelsX, f32 inputOffsetInPixelsY);
16
17void FsrRcasCon(u32* con, f32 sharpness);
18
19} // namespace FSR
diff --git a/src/video_core/host_shaders/CMakeLists.txt b/src/video_core/host_shaders/CMakeLists.txt
index e968ae220..dad7b07d4 100644
--- a/src/video_core/host_shaders/CMakeLists.txt
+++ b/src/video_core/host_shaders/CMakeLists.txt
@@ -3,12 +3,16 @@
3 3
4set(FIDELITYFX_INCLUDE_DIR ${CMAKE_SOURCE_DIR}/externals/FidelityFX-FSR/ffx-fsr) 4set(FIDELITYFX_INCLUDE_DIR ${CMAKE_SOURCE_DIR}/externals/FidelityFX-FSR/ffx-fsr)
5 5
6set(GLSL_INCLUDES 6set(FIDELITYFX_FILES
7 fidelityfx_fsr.comp
8 ${FIDELITYFX_INCLUDE_DIR}/ffx_a.h 7 ${FIDELITYFX_INCLUDE_DIR}/ffx_a.h
9 ${FIDELITYFX_INCLUDE_DIR}/ffx_fsr1.h 8 ${FIDELITYFX_INCLUDE_DIR}/ffx_fsr1.h
10) 9)
11 10
11set(GLSL_INCLUDES
12 fidelityfx_fsr.comp
13 ${FIDELITYFX_FILES}
14)
15
12set(SHADER_FILES 16set(SHADER_FILES
13 astc_decoder.comp 17 astc_decoder.comp
14 blit_color_float.frag 18 blit_color_float.frag
@@ -24,6 +28,9 @@ set(SHADER_FILES
24 fxaa.vert 28 fxaa.vert
25 opengl_convert_s8d24.comp 29 opengl_convert_s8d24.comp
26 opengl_copy_bc4.comp 30 opengl_copy_bc4.comp
31 opengl_fidelityfx_fsr.frag
32 opengl_fidelityfx_fsr_easu.frag
33 opengl_fidelityfx_fsr_rcas.frag
27 opengl_present.frag 34 opengl_present.frag
28 opengl_present.vert 35 opengl_present.vert
29 opengl_present_scaleforce.frag 36 opengl_present_scaleforce.frag
@@ -118,6 +125,25 @@ foreach(FILENAME IN ITEMS ${SHADER_FILES})
118 endif() 125 endif()
119endforeach() 126endforeach()
120 127
128foreach(FILEPATH IN ITEMS ${FIDELITYFX_FILES})
129 get_filename_component(FILENAME ${FILEPATH} NAME)
130 string(REPLACE "." "_" HEADER_NAME ${FILENAME})
131 set(SOURCE_FILE ${FILEPATH})
132 set(SOURCE_HEADER_FILE ${SHADER_DIR}/${HEADER_NAME}.h)
133 add_custom_command(
134 OUTPUT
135 ${SOURCE_HEADER_FILE}
136 COMMAND
137 ${CMAKE_COMMAND} -P ${HEADER_GENERATOR} ${SOURCE_FILE} ${SOURCE_HEADER_FILE} ${INPUT_FILE}
138 MAIN_DEPENDENCY
139 ${SOURCE_FILE}
140 DEPENDS
141 ${INPUT_FILE}
142 # HEADER_GENERATOR should be included here but msbuild seems to assume it's always modified
143 )
144 set(SHADER_HEADERS ${SHADER_HEADERS} ${SOURCE_HEADER_FILE})
145endforeach()
146
121set(SHADER_SOURCES ${SHADER_FILES}) 147set(SHADER_SOURCES ${SHADER_FILES})
122list(APPEND SHADER_SOURCES ${GLSL_INCLUDES}) 148list(APPEND SHADER_SOURCES ${GLSL_INCLUDES})
123 149
diff --git a/src/video_core/host_shaders/opengl_fidelityfx_fsr.frag b/src/video_core/host_shaders/opengl_fidelityfx_fsr.frag
new file mode 100644
index 000000000..16d22f58e
--- /dev/null
+++ b/src/video_core/host_shaders/opengl_fidelityfx_fsr.frag
@@ -0,0 +1,108 @@
1// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
2// SPDX-License-Identifier: GPL-2.0-or-later
3
4//!#version 460 core
5#extension GL_ARB_separate_shader_objects : enable
6#extension GL_ARB_shading_language_420pack : enable
7
8#extension GL_AMD_gpu_shader_half_float : enable
9#extension GL_NV_gpu_shader5 : enable
10
11// FidelityFX Super Resolution Sample
12//
13// Copyright (c) 2021 Advanced Micro Devices, Inc. All rights reserved.
14// Permission is hereby granted, free of charge, to any person obtaining a copy
15// of this software and associated documentation files(the "Software"), to deal
16// in the Software without restriction, including without limitation the rights
17// to use, copy, modify, merge, publish, distribute, sublicense, and / or sell
18// copies of the Software, and to permit persons to whom the Software is
19// furnished to do so, subject to the following conditions :
20// The above copyright notice and this permission notice shall be included in
21// all copies or substantial portions of the Software.
22// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
23// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
24// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.IN NO EVENT SHALL THE
25// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
26// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
27// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
28// THE SOFTWARE.
29
30layout (location = 0) uniform uvec4 constants[4];
31
32#define A_GPU 1
33#define A_GLSL 1
34
35#ifdef YUZU_USE_FP16
36 #define A_HALF
37#endif
38#include "ffx_a.h"
39
40#ifndef YUZU_USE_FP16
41 layout (binding=0) uniform sampler2D InputTexture;
42 #if USE_EASU
43 #define FSR_EASU_F 1
44 AF4 FsrEasuRF(AF2 p) { AF4 res = textureGather(InputTexture, p, 0); return res; }
45 AF4 FsrEasuGF(AF2 p) { AF4 res = textureGather(InputTexture, p, 1); return res; }
46 AF4 FsrEasuBF(AF2 p) { AF4 res = textureGather(InputTexture, p, 2); return res; }
47 #endif
48 #if USE_RCAS
49 #define FSR_RCAS_F
50 AF4 FsrRcasLoadF(ASU2 p) { return texelFetch(InputTexture, ASU2(p), 0); }
51 void FsrRcasInputF(inout AF1 r, inout AF1 g, inout AF1 b) {}
52 #endif
53#else
54 layout (binding=0) uniform sampler2D InputTexture;
55 #if USE_EASU
56 #define FSR_EASU_H 1
57 AH4 FsrEasuRH(AF2 p) { AH4 res = AH4(textureGather(InputTexture, p, 0)); return res; }
58 AH4 FsrEasuGH(AF2 p) { AH4 res = AH4(textureGather(InputTexture, p, 1)); return res; }
59 AH4 FsrEasuBH(AF2 p) { AH4 res = AH4(textureGather(InputTexture, p, 2)); return res; }
60 #endif
61 #if USE_RCAS
62 #define FSR_RCAS_H
63 AH4 FsrRcasLoadH(ASW2 p) { return AH4(texelFetch(InputTexture, ASU2(p), 0)); }
64 void FsrRcasInputH(inout AH1 r,inout AH1 g,inout AH1 b){}
65 #endif
66#endif
67
68#include "ffx_fsr1.h"
69
70#if USE_RCAS
71 layout(location = 0) in vec2 frag_texcoord;
72#endif
73layout (location = 0) out vec4 frag_color;
74
75void CurrFilter(AU2 pos)
76{
77#if USE_EASU
78 #ifndef YUZU_USE_FP16
79 AF3 c;
80 FsrEasuF(c, pos, constants[0], constants[1], constants[2], constants[3]);
81 frag_color = AF4(c, 1.0);
82 #else
83 AH3 c;
84 FsrEasuH(c, pos, constants[0], constants[1], constants[2], constants[3]);
85 frag_color = AH4(c, 1.0);
86 #endif
87#endif
88#if USE_RCAS
89 #ifndef YUZU_USE_FP16
90 AF3 c;
91 FsrRcasF(c.r, c.g, c.b, pos, constants[0]);
92 frag_color = AF4(c, 1.0);
93 #else
94 AH3 c;
95 FsrRcasH(c.r, c.g, c.b, pos, constants[0]);
96 frag_color = AH4(c, 1.0);
97 #endif
98#endif
99}
100
101void main()
102{
103#if USE_RCAS
104 CurrFilter(AU2(frag_texcoord * vec2(textureSize(InputTexture, 0))));
105#else
106 CurrFilter(AU2(gl_FragCoord.xy));
107#endif
108}
diff --git a/src/video_core/host_shaders/opengl_fidelityfx_fsr_easu.frag b/src/video_core/host_shaders/opengl_fidelityfx_fsr_easu.frag
new file mode 100644
index 000000000..d39f80ac1
--- /dev/null
+++ b/src/video_core/host_shaders/opengl_fidelityfx_fsr_easu.frag
@@ -0,0 +1,9 @@
1// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
2// SPDX-License-Identifier: GPL-2.0-or-later
3
4#version 460 core
5#extension GL_GOOGLE_include_directive : enable
6
7#define USE_EASU 1
8
9#include "opengl_fidelityfx_fsr.frag"
diff --git a/src/video_core/host_shaders/opengl_fidelityfx_fsr_rcas.frag b/src/video_core/host_shaders/opengl_fidelityfx_fsr_rcas.frag
new file mode 100644
index 000000000..cfa78ddc7
--- /dev/null
+++ b/src/video_core/host_shaders/opengl_fidelityfx_fsr_rcas.frag
@@ -0,0 +1,9 @@
1// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
2// SPDX-License-Identifier: GPL-2.0-or-later
3
4#version 460 core
5#extension GL_GOOGLE_include_directive : enable
6
7#define USE_RCAS 1
8
9#include "opengl_fidelityfx_fsr.frag"
diff --git a/src/video_core/renderer_opengl/gl_fsr.cpp b/src/video_core/renderer_opengl/gl_fsr.cpp
new file mode 100644
index 000000000..77262dcf1
--- /dev/null
+++ b/src/video_core/renderer_opengl/gl_fsr.cpp
@@ -0,0 +1,101 @@
1// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
2// SPDX-License-Identifier: GPL-2.0-or-later
3
4#include "common/settings.h"
5#include "video_core/fsr.h"
6#include "video_core/renderer_opengl/gl_fsr.h"
7#include "video_core/renderer_opengl/gl_shader_manager.h"
8#include "video_core/renderer_opengl/gl_shader_util.h"
9
10namespace OpenGL {
11using namespace FSR;
12
13using FsrConstants = std::array<u32, 4 * 4>;
14
15FSR::FSR(std::string_view fsr_vertex_source, std::string_view fsr_easu_source,
16 std::string_view fsr_rcas_source)
17 : fsr_vertex{CreateProgram(fsr_vertex_source, GL_VERTEX_SHADER)},
18 fsr_easu_frag{CreateProgram(fsr_easu_source, GL_FRAGMENT_SHADER)},
19 fsr_rcas_frag{CreateProgram(fsr_rcas_source, GL_FRAGMENT_SHADER)} {
20 glProgramUniform2f(fsr_vertex.handle, 0, 1.0f, 1.0f);
21 glProgramUniform2f(fsr_vertex.handle, 1, 0.0f, 0.0f);
22}
23
24FSR::~FSR() = default;
25
26void FSR::Draw(ProgramManager& program_manager, const Common::Rectangle<u32>& screen,
27 u32 input_image_width, u32 input_image_height,
28 const Common::Rectangle<int>& crop_rect) {
29
30 const auto output_image_width = screen.GetWidth();
31 const auto output_image_height = screen.GetHeight();
32
33 if (fsr_intermediate_tex.handle) {
34 GLint fsr_tex_width, fsr_tex_height;
35 glGetTextureLevelParameteriv(fsr_intermediate_tex.handle, 0, GL_TEXTURE_WIDTH,
36 &fsr_tex_width);
37 glGetTextureLevelParameteriv(fsr_intermediate_tex.handle, 0, GL_TEXTURE_HEIGHT,
38 &fsr_tex_height);
39 if (static_cast<u32>(fsr_tex_width) != output_image_width ||
40 static_cast<u32>(fsr_tex_height) != output_image_height) {
41 fsr_intermediate_tex.Release();
42 }
43 }
44 if (!fsr_intermediate_tex.handle) {
45 fsr_intermediate_tex.Create(GL_TEXTURE_2D);
46 glTextureStorage2D(fsr_intermediate_tex.handle, 1, GL_RGB16F, output_image_width,
47 output_image_height);
48 glNamedFramebufferTexture(fsr_framebuffer.handle, GL_COLOR_ATTACHMENT0,
49 fsr_intermediate_tex.handle, 0);
50 }
51
52 GLint old_draw_fb;
53 glGetIntegerv(GL_DRAW_FRAMEBUFFER_BINDING, &old_draw_fb);
54
55 glFrontFace(GL_CW);
56 glBindFramebuffer(GL_DRAW_FRAMEBUFFER, fsr_framebuffer.handle);
57 glViewportIndexedf(0, 0.0f, 0.0f, static_cast<GLfloat>(output_image_width),
58 static_cast<GLfloat>(output_image_height));
59
60 FsrConstants constants;
61 FsrEasuConOffset(
62 constants.data() + 0, constants.data() + 4, constants.data() + 8, constants.data() + 12,
63
64 static_cast<f32>(crop_rect.GetWidth()), static_cast<f32>(crop_rect.GetHeight()),
65 static_cast<f32>(input_image_width), static_cast<f32>(input_image_height),
66 static_cast<f32>(output_image_width), static_cast<f32>(output_image_height),
67 static_cast<f32>(crop_rect.left), static_cast<f32>(crop_rect.top));
68
69 glProgramUniform4uiv(fsr_easu_frag.handle, 0, sizeof(constants), std::data(constants));
70
71 program_manager.BindPresentPrograms(fsr_vertex.handle, fsr_easu_frag.handle);
72 glDrawArrays(GL_TRIANGLES, 0, 3);
73
74 glBindFramebuffer(GL_DRAW_FRAMEBUFFER, old_draw_fb);
75 glBindTextureUnit(0, fsr_intermediate_tex.handle);
76
77 const float sharpening =
78 static_cast<float>(Settings::values.fsr_sharpening_slider.GetValue()) / 100.0f;
79
80 FsrRcasCon(constants.data(), sharpening);
81 glProgramUniform4uiv(fsr_rcas_frag.handle, 0, sizeof(constants), std::data(constants));
82}
83
84void FSR::InitBuffers() {
85 fsr_framebuffer.Create();
86}
87
88void FSR::ReleaseBuffers() {
89 fsr_framebuffer.Release();
90 fsr_intermediate_tex.Release();
91}
92
93const OGLProgram& FSR::GetPresentFragmentProgram() const noexcept {
94 return fsr_rcas_frag;
95}
96
97bool FSR::AreBuffersInitialized() const noexcept {
98 return fsr_framebuffer.handle;
99}
100
101} // namespace OpenGL
diff --git a/src/video_core/renderer_opengl/gl_fsr.h b/src/video_core/renderer_opengl/gl_fsr.h
new file mode 100644
index 000000000..1f6ae3115
--- /dev/null
+++ b/src/video_core/renderer_opengl/gl_fsr.h
@@ -0,0 +1,43 @@
1// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
2// SPDX-License-Identifier: GPL-2.0-or-later
3
4#pragma once
5
6#include <string_view>
7
8#include "common/common_types.h"
9#include "common/math_util.h"
10#include "video_core/fsr.h"
11#include "video_core/renderer_opengl/gl_resource_manager.h"
12
13namespace OpenGL {
14
15class ProgramManager;
16
17class FSR {
18public:
19 explicit FSR(std::string_view fsr_vertex_source, std::string_view fsr_easu_source,
20 std::string_view fsr_rcas_source);
21 ~FSR();
22
23 void Draw(ProgramManager& program_manager, const Common::Rectangle<u32>& screen,
24 u32 input_image_width, u32 input_image_height,
25 const Common::Rectangle<int>& crop_rect);
26
27 void InitBuffers();
28
29 void ReleaseBuffers();
30
31 [[nodiscard]] const OGLProgram& GetPresentFragmentProgram() const noexcept;
32
33 [[nodiscard]] bool AreBuffersInitialized() const noexcept;
34
35private:
36 OGLFramebuffer fsr_framebuffer;
37 OGLProgram fsr_vertex;
38 OGLProgram fsr_easu_frag;
39 OGLProgram fsr_rcas_frag;
40 OGLTexture fsr_intermediate_tex;
41};
42
43} // namespace OpenGL
diff --git a/src/video_core/renderer_opengl/renderer_opengl.cpp b/src/video_core/renderer_opengl/renderer_opengl.cpp
index de95f2634..2a74c1d05 100644
--- a/src/video_core/renderer_opengl/renderer_opengl.cpp
+++ b/src/video_core/renderer_opengl/renderer_opengl.cpp
@@ -17,8 +17,14 @@
17#include "core/frontend/emu_window.h" 17#include "core/frontend/emu_window.h"
18#include "core/memory.h" 18#include "core/memory.h"
19#include "core/telemetry_session.h" 19#include "core/telemetry_session.h"
20#include "video_core/host_shaders/ffx_a_h.h"
21#include "video_core/host_shaders/ffx_fsr1_h.h"
22#include "video_core/host_shaders/full_screen_triangle_vert.h"
20#include "video_core/host_shaders/fxaa_frag.h" 23#include "video_core/host_shaders/fxaa_frag.h"
21#include "video_core/host_shaders/fxaa_vert.h" 24#include "video_core/host_shaders/fxaa_vert.h"
25#include "video_core/host_shaders/opengl_fidelityfx_fsr_easu_frag.h"
26#include "video_core/host_shaders/opengl_fidelityfx_fsr_frag.h"
27#include "video_core/host_shaders/opengl_fidelityfx_fsr_rcas_frag.h"
22#include "video_core/host_shaders/opengl_present_frag.h" 28#include "video_core/host_shaders/opengl_present_frag.h"
23#include "video_core/host_shaders/opengl_present_scaleforce_frag.h" 29#include "video_core/host_shaders/opengl_present_scaleforce_frag.h"
24#include "video_core/host_shaders/opengl_present_vert.h" 30#include "video_core/host_shaders/opengl_present_vert.h"
@@ -31,6 +37,7 @@
31#include "video_core/host_shaders/smaa_edge_detection_vert.h" 37#include "video_core/host_shaders/smaa_edge_detection_vert.h"
32#include "video_core/host_shaders/smaa_neighborhood_blending_frag.h" 38#include "video_core/host_shaders/smaa_neighborhood_blending_frag.h"
33#include "video_core/host_shaders/smaa_neighborhood_blending_vert.h" 39#include "video_core/host_shaders/smaa_neighborhood_blending_vert.h"
40#include "video_core/renderer_opengl/gl_fsr.h"
34#include "video_core/renderer_opengl/gl_rasterizer.h" 41#include "video_core/renderer_opengl/gl_rasterizer.h"
35#include "video_core/renderer_opengl/gl_shader_manager.h" 42#include "video_core/renderer_opengl/gl_shader_manager.h"
36#include "video_core/renderer_opengl/gl_shader_util.h" 43#include "video_core/renderer_opengl/gl_shader_util.h"
@@ -268,12 +275,17 @@ void RendererOpenGL::InitOpenGLObjects() {
268 fxaa_vertex = CreateProgram(HostShaders::FXAA_VERT, GL_VERTEX_SHADER); 275 fxaa_vertex = CreateProgram(HostShaders::FXAA_VERT, GL_VERTEX_SHADER);
269 fxaa_fragment = CreateProgram(HostShaders::FXAA_FRAG, GL_FRAGMENT_SHADER); 276 fxaa_fragment = CreateProgram(HostShaders::FXAA_FRAG, GL_FRAGMENT_SHADER);
270 277
271 const auto SmaaShader = [](std::string_view specialized_source, GLenum stage) { 278 const auto replace_include = [](std::string& shader_source, std::string_view include_name,
272 std::string shader_source{specialized_source}; 279 std::string_view include_content) {
273 constexpr std::string_view include_string = "#include \"opengl_smaa.glsl\""; 280 const std::string include_string = fmt::format("#include \"{}\"", include_name);
274 const std::size_t pos = shader_source.find(include_string); 281 const std::size_t pos = shader_source.find(include_string);
275 ASSERT(pos != std::string::npos); 282 ASSERT(pos != std::string::npos);
276 shader_source.replace(pos, include_string.size(), HostShaders::OPENGL_SMAA_GLSL); 283 shader_source.replace(pos, include_string.size(), include_content);
284 };
285
286 const auto SmaaShader = [&](std::string_view specialized_source, GLenum stage) {
287 std::string shader_source{specialized_source};
288 replace_include(shader_source, "opengl_smaa.glsl", HostShaders::OPENGL_SMAA_GLSL);
277 return CreateProgram(shader_source, stage); 289 return CreateProgram(shader_source, stage);
278 }; 290 };
279 291
@@ -298,14 +310,32 @@ void RendererOpenGL::InitOpenGLObjects() {
298 CreateProgram(fmt::format("#version 460\n{}", HostShaders::OPENGL_PRESENT_SCALEFORCE_FRAG), 310 CreateProgram(fmt::format("#version 460\n{}", HostShaders::OPENGL_PRESENT_SCALEFORCE_FRAG),
299 GL_FRAGMENT_SHADER); 311 GL_FRAGMENT_SHADER);
300 312
313 std::string fsr_source{HostShaders::OPENGL_FIDELITYFX_FSR_FRAG};
314 replace_include(fsr_source, "ffx_a.h", HostShaders::FFX_A_H);
315 replace_include(fsr_source, "ffx_fsr1.h", HostShaders::FFX_FSR1_H);
316
317 std::string fsr_easu_frag_source{HostShaders::OPENGL_FIDELITYFX_FSR_EASU_FRAG};
318 std::string fsr_rcas_frag_source{HostShaders::OPENGL_FIDELITYFX_FSR_RCAS_FRAG};
319 replace_include(fsr_easu_frag_source, "opengl_fidelityfx_fsr.frag", fsr_source);
320 replace_include(fsr_rcas_frag_source, "opengl_fidelityfx_fsr.frag", fsr_source);
321
322 fsr = std::make_unique<FSR>(HostShaders::FULL_SCREEN_TRIANGLE_VERT, fsr_easu_frag_source,
323 fsr_rcas_frag_source);
324
301 // Generate presentation sampler 325 // Generate presentation sampler
302 present_sampler.Create(); 326 present_sampler.Create();
303 glSamplerParameteri(present_sampler.handle, GL_TEXTURE_MIN_FILTER, GL_LINEAR); 327 glSamplerParameteri(present_sampler.handle, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
304 glSamplerParameteri(present_sampler.handle, GL_TEXTURE_MAG_FILTER, GL_LINEAR); 328 glSamplerParameteri(present_sampler.handle, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
329 glSamplerParameteri(present_sampler.handle, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
330 glSamplerParameteri(present_sampler.handle, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
331 glSamplerParameteri(present_sampler.handle, GL_TEXTURE_WRAP_R, GL_CLAMP_TO_EDGE);
305 332
306 present_sampler_nn.Create(); 333 present_sampler_nn.Create();
307 glSamplerParameteri(present_sampler_nn.handle, GL_TEXTURE_MIN_FILTER, GL_NEAREST); 334 glSamplerParameteri(present_sampler_nn.handle, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
308 glSamplerParameteri(present_sampler_nn.handle, GL_TEXTURE_MAG_FILTER, GL_NEAREST); 335 glSamplerParameteri(present_sampler_nn.handle, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
336 glSamplerParameteri(present_sampler_nn.handle, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
337 glSamplerParameteri(present_sampler_nn.handle, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
338 glSamplerParameteri(present_sampler_nn.handle, GL_TEXTURE_WRAP_R, GL_CLAMP_TO_EDGE);
309 339
310 // Generate VBO handle for drawing 340 // Generate VBO handle for drawing
311 vertex_buffer.Create(); 341 vertex_buffer.Create();
@@ -525,6 +555,31 @@ void RendererOpenGL::DrawScreen(const Layout::FramebufferLayout& layout) {
525 555
526 glBindTextureUnit(0, aa_texture.handle); 556 glBindTextureUnit(0, aa_texture.handle);
527 } 557 }
558 glDisablei(GL_SCISSOR_TEST, 0);
559
560 if (Settings::values.scaling_filter.GetValue() == Settings::ScalingFilter::Fsr) {
561 if (!fsr->AreBuffersInitialized()) {
562 fsr->InitBuffers();
563 }
564
565 auto crop_rect = framebuffer_crop_rect;
566 if (crop_rect.GetWidth() == 0) {
567 crop_rect.right = framebuffer_width;
568 }
569 if (crop_rect.GetHeight() == 0) {
570 crop_rect.bottom = framebuffer_height;
571 }
572 crop_rect = crop_rect.Scale(Settings::values.resolution_info.up_factor);
573 const auto fsr_input_width = Settings::values.resolution_info.ScaleUp(framebuffer_width);
574 const auto fsr_input_height = Settings::values.resolution_info.ScaleUp(framebuffer_height);
575 glBindSampler(0, present_sampler.handle);
576 fsr->Draw(program_manager, layout.screen, fsr_input_width, fsr_input_height, crop_rect);
577 } else {
578 if (fsr->AreBuffersInitialized()) {
579 fsr->ReleaseBuffers();
580 }
581 }
582
528 const std::array ortho_matrix = 583 const std::array ortho_matrix =
529 MakeOrthographicMatrix(static_cast<float>(layout.width), static_cast<float>(layout.height)); 584 MakeOrthographicMatrix(static_cast<float>(layout.width), static_cast<float>(layout.height));
530 585
@@ -540,10 +595,7 @@ void RendererOpenGL::DrawScreen(const Layout::FramebufferLayout& layout) {
540 case Settings::ScalingFilter::ScaleForce: 595 case Settings::ScalingFilter::ScaleForce:
541 return present_scaleforce_fragment.handle; 596 return present_scaleforce_fragment.handle;
542 case Settings::ScalingFilter::Fsr: 597 case Settings::ScalingFilter::Fsr:
543 LOG_WARNING( 598 return fsr->GetPresentFragmentProgram().handle;
544 Render_OpenGL,
545 "FidelityFX Super Resolution is not supported in OpenGL, changing to ScaleForce");
546 return present_scaleforce_fragment.handle;
547 default: 599 default:
548 return present_bilinear_fragment.handle; 600 return present_bilinear_fragment.handle;
549 } 601 }
@@ -578,15 +630,18 @@ void RendererOpenGL::DrawScreen(const Layout::FramebufferLayout& layout) {
578 f32 scale_u = static_cast<f32>(framebuffer_width) / static_cast<f32>(screen_info.texture.width); 630 f32 scale_u = static_cast<f32>(framebuffer_width) / static_cast<f32>(screen_info.texture.width);
579 f32 scale_v = 631 f32 scale_v =
580 static_cast<f32>(framebuffer_height) / static_cast<f32>(screen_info.texture.height); 632 static_cast<f32>(framebuffer_height) / static_cast<f32>(screen_info.texture.height);
581 // Scale the output by the crop width/height. This is commonly used with 1280x720 rendering 633
582 // (e.g. handheld mode) on a 1920x1080 framebuffer. 634 if (Settings::values.scaling_filter.GetValue() != Settings::ScalingFilter::Fsr) {
583 if (framebuffer_crop_rect.GetWidth() > 0) { 635 // Scale the output by the crop width/height. This is commonly used with 1280x720 rendering
584 scale_u = static_cast<f32>(framebuffer_crop_rect.GetWidth()) / 636 // (e.g. handheld mode) on a 1920x1080 framebuffer.
585 static_cast<f32>(screen_info.texture.width); 637 if (framebuffer_crop_rect.GetWidth() > 0) {
586 } 638 scale_u = static_cast<f32>(framebuffer_crop_rect.GetWidth()) /
587 if (framebuffer_crop_rect.GetHeight() > 0) { 639 static_cast<f32>(screen_info.texture.width);
588 scale_v = static_cast<f32>(framebuffer_crop_rect.GetHeight()) / 640 }
589 static_cast<f32>(screen_info.texture.height); 641 if (framebuffer_crop_rect.GetHeight() > 0) {
642 scale_v = static_cast<f32>(framebuffer_crop_rect.GetHeight()) /
643 static_cast<f32>(screen_info.texture.height);
644 }
590 } 645 }
591 if (Settings::values.anti_aliasing.GetValue() == Settings::AntiAliasing::Fxaa && 646 if (Settings::values.anti_aliasing.GetValue() == Settings::AntiAliasing::Fxaa &&
592 !screen_info.was_accelerated) { 647 !screen_info.was_accelerated) {
@@ -612,7 +667,6 @@ void RendererOpenGL::DrawScreen(const Layout::FramebufferLayout& layout) {
612 } else { 667 } else {
613 glDisable(GL_FRAMEBUFFER_SRGB); 668 glDisable(GL_FRAMEBUFFER_SRGB);
614 } 669 }
615 glDisablei(GL_SCISSOR_TEST, 0);
616 glViewportIndexedf(0, 0.0f, 0.0f, static_cast<GLfloat>(layout.width), 670 glViewportIndexedf(0, 0.0f, 0.0f, static_cast<GLfloat>(layout.width),
617 static_cast<GLfloat>(layout.height)); 671 static_cast<GLfloat>(layout.height));
618 672
diff --git a/src/video_core/renderer_opengl/renderer_opengl.h b/src/video_core/renderer_opengl/renderer_opengl.h
index cc97d7b26..f1d5fd954 100644
--- a/src/video_core/renderer_opengl/renderer_opengl.h
+++ b/src/video_core/renderer_opengl/renderer_opengl.h
@@ -10,6 +10,7 @@
10 10
11#include "video_core/renderer_base.h" 11#include "video_core/renderer_base.h"
12#include "video_core/renderer_opengl/gl_device.h" 12#include "video_core/renderer_opengl/gl_device.h"
13#include "video_core/renderer_opengl/gl_fsr.h"
13#include "video_core/renderer_opengl/gl_rasterizer.h" 14#include "video_core/renderer_opengl/gl_rasterizer.h"
14#include "video_core/renderer_opengl/gl_resource_manager.h" 15#include "video_core/renderer_opengl/gl_resource_manager.h"
15#include "video_core/renderer_opengl/gl_shader_manager.h" 16#include "video_core/renderer_opengl/gl_shader_manager.h"
@@ -141,6 +142,8 @@ private:
141 OGLTexture smaa_edges_tex; 142 OGLTexture smaa_edges_tex;
142 OGLTexture smaa_blend_tex; 143 OGLTexture smaa_blend_tex;
143 144
145 std::unique_ptr<FSR> fsr;
146
144 /// OpenGL framebuffer data 147 /// OpenGL framebuffer data
145 std::vector<u8> gl_framebuffer_data; 148 std::vector<u8> gl_framebuffer_data;
146 149
diff --git a/src/video_core/renderer_vulkan/vk_fsr.cpp b/src/video_core/renderer_vulkan/vk_fsr.cpp
index 33daa8c1c..df972cd54 100644
--- a/src/video_core/renderer_vulkan/vk_fsr.cpp
+++ b/src/video_core/renderer_vulkan/vk_fsr.cpp
@@ -1,12 +1,11 @@
1// SPDX-FileCopyrightText: Copyright 2021 yuzu Emulator Project 1// SPDX-FileCopyrightText: Copyright 2021 yuzu Emulator Project
2// SPDX-License-Identifier: GPL-2.0-or-later 2// SPDX-License-Identifier: GPL-2.0-or-later
3 3
4#include <cmath>
5#include "common/bit_cast.h"
6#include "common/common_types.h" 4#include "common/common_types.h"
7#include "common/div_ceil.h" 5#include "common/div_ceil.h"
8#include "common/settings.h" 6#include "common/settings.h"
9 7
8#include "video_core/fsr.h"
10#include "video_core/host_shaders/vulkan_fidelityfx_fsr_easu_fp16_comp_spv.h" 9#include "video_core/host_shaders/vulkan_fidelityfx_fsr_easu_fp16_comp_spv.h"
11#include "video_core/host_shaders/vulkan_fidelityfx_fsr_easu_fp32_comp_spv.h" 10#include "video_core/host_shaders/vulkan_fidelityfx_fsr_easu_fp32_comp_spv.h"
12#include "video_core/host_shaders/vulkan_fidelityfx_fsr_rcas_fp16_comp_spv.h" 11#include "video_core/host_shaders/vulkan_fidelityfx_fsr_rcas_fp16_comp_spv.h"
@@ -17,146 +16,7 @@
17#include "video_core/vulkan_common/vulkan_device.h" 16#include "video_core/vulkan_common/vulkan_device.h"
18 17
19namespace Vulkan { 18namespace Vulkan {
20namespace { 19using namespace FSR;
21// Reimplementations of the constant generating functions in ffx_fsr1.h
22// GCC generated a lot of warnings when using the official header.
23u32 AU1_AH1_AF1(f32 f) {
24 static constexpr u32 base[512]{
25 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
26 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
27 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
28 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
29 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
30 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
31 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
32 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
33 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
34 0x0000, 0x0000, 0x0000, 0x0000, 0x0001, 0x0002, 0x0004, 0x0008, 0x0010, 0x0020, 0x0040,
35 0x0080, 0x0100, 0x0200, 0x0400, 0x0800, 0x0c00, 0x1000, 0x1400, 0x1800, 0x1c00, 0x2000,
36 0x2400, 0x2800, 0x2c00, 0x3000, 0x3400, 0x3800, 0x3c00, 0x4000, 0x4400, 0x4800, 0x4c00,
37 0x5000, 0x5400, 0x5800, 0x5c00, 0x6000, 0x6400, 0x6800, 0x6c00, 0x7000, 0x7400, 0x7800,
38 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
39 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
40 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
41 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
42 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
43 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
44 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
45 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
46 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
47 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
48 0x7bff, 0x7bff, 0x7bff, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
49 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
50 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
51 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
52 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
53 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
54 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
55 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
56 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
57 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8001, 0x8002, 0x8004, 0x8008,
58 0x8010, 0x8020, 0x8040, 0x8080, 0x8100, 0x8200, 0x8400, 0x8800, 0x8c00, 0x9000, 0x9400,
59 0x9800, 0x9c00, 0xa000, 0xa400, 0xa800, 0xac00, 0xb000, 0xb400, 0xb800, 0xbc00, 0xc000,
60 0xc400, 0xc800, 0xcc00, 0xd000, 0xd400, 0xd800, 0xdc00, 0xe000, 0xe400, 0xe800, 0xec00,
61 0xf000, 0xf400, 0xf800, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
62 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
63 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
64 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
65 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
66 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
67 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
68 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
69 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
70 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
71 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
72 };
73 static constexpr s8 shift[512]{
74 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
75 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
76 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
77 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
78 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
79 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
80 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x17, 0x16,
81 0x15, 0x14, 0x13, 0x12, 0x11, 0x10, 0x0f, 0x0e, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d,
82 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d,
83 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
84 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
85 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
86 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
87 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
88 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
89 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
90 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
91 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
92 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
93 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
94 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
95 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
96 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
97 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x17,
98 0x16, 0x15, 0x14, 0x13, 0x12, 0x11, 0x10, 0x0f, 0x0e, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d,
99 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d,
100 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
101 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
102 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
103 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
104 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
105 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
106 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
107 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
108 0x18, 0x18,
109 };
110 const u32 u = Common::BitCast<u32>(f);
111 const u32 i = u >> 23;
112 return base[i] + ((u & 0x7fffff) >> shift[i]);
113}
114
115u32 AU1_AH2_AF2(f32 a[2]) {
116 return AU1_AH1_AF1(a[0]) + (AU1_AH1_AF1(a[1]) << 16);
117}
118
119void FsrEasuCon(u32 con0[4], u32 con1[4], u32 con2[4], u32 con3[4], f32 inputViewportInPixelsX,
120 f32 inputViewportInPixelsY, f32 inputSizeInPixelsX, f32 inputSizeInPixelsY,
121 f32 outputSizeInPixelsX, f32 outputSizeInPixelsY) {
122 con0[0] = Common::BitCast<u32>(inputViewportInPixelsX / outputSizeInPixelsX);
123 con0[1] = Common::BitCast<u32>(inputViewportInPixelsY / outputSizeInPixelsY);
124 con0[2] = Common::BitCast<u32>(0.5f * inputViewportInPixelsX / outputSizeInPixelsX - 0.5f);
125 con0[3] = Common::BitCast<u32>(0.5f * inputViewportInPixelsY / outputSizeInPixelsY - 0.5f);
126 con1[0] = Common::BitCast<u32>(1.0f / inputSizeInPixelsX);
127 con1[1] = Common::BitCast<u32>(1.0f / inputSizeInPixelsY);
128 con1[2] = Common::BitCast<u32>(1.0f / inputSizeInPixelsX);
129 con1[3] = Common::BitCast<u32>(-1.0f / inputSizeInPixelsY);
130 con2[0] = Common::BitCast<u32>(-1.0f / inputSizeInPixelsX);
131 con2[1] = Common::BitCast<u32>(2.0f / inputSizeInPixelsY);
132 con2[2] = Common::BitCast<u32>(1.0f / inputSizeInPixelsX);
133 con2[3] = Common::BitCast<u32>(2.0f / inputSizeInPixelsY);
134 con3[0] = Common::BitCast<u32>(0.0f / inputSizeInPixelsX);
135 con3[1] = Common::BitCast<u32>(4.0f / inputSizeInPixelsY);
136 con3[2] = con3[3] = 0;
137}
138
139void FsrEasuConOffset(u32 con0[4], u32 con1[4], u32 con2[4], u32 con3[4],
140 f32 inputViewportInPixelsX, f32 inputViewportInPixelsY,
141 f32 inputSizeInPixelsX, f32 inputSizeInPixelsY, f32 outputSizeInPixelsX,
142 f32 outputSizeInPixelsY, f32 inputOffsetInPixelsX, f32 inputOffsetInPixelsY) {
143 FsrEasuCon(con0, con1, con2, con3, inputViewportInPixelsX, inputViewportInPixelsY,
144 inputSizeInPixelsX, inputSizeInPixelsY, outputSizeInPixelsX, outputSizeInPixelsY);
145 con0[2] = Common::BitCast<u32>(0.5f * inputViewportInPixelsX / outputSizeInPixelsX - 0.5f +
146 inputOffsetInPixelsX);
147 con0[3] = Common::BitCast<u32>(0.5f * inputViewportInPixelsY / outputSizeInPixelsY - 0.5f +
148 inputOffsetInPixelsY);
149}
150
151void FsrRcasCon(u32* con, f32 sharpness) {
152 sharpness = std::exp2f(-sharpness);
153 f32 hSharp[2]{sharpness, sharpness};
154 con[0] = Common::BitCast<u32>(sharpness);
155 con[1] = AU1_AH2_AF2(hSharp);
156 con[2] = 0;
157 con[3] = 0;
158}
159} // Anonymous namespace
160 20
161FSR::FSR(const Device& device_, MemoryAllocator& memory_allocator_, size_t image_count_, 21FSR::FSR(const Device& device_, MemoryAllocator& memory_allocator_, size_t image_count_,
162 VkExtent2D output_size_) 22 VkExtent2D output_size_)
diff --git a/src/yuzu/configuration/configure_graphics.ui b/src/yuzu/configuration/configure_graphics.ui
index bb9910a53..a45ec69ec 100644
--- a/src/yuzu/configuration/configure_graphics.ui
+++ b/src/yuzu/configuration/configure_graphics.ui
@@ -460,7 +460,7 @@
460 </item> 460 </item>
461 <item> 461 <item>
462 <property name="text"> 462 <property name="text">
463 <string>AMD FidelityFX™️ Super Resolution (Vulkan Only)</string> 463 <string>AMD FidelityFX™️ Super Resolution</string>
464 </property> 464 </property>
465 </item> 465 </item>
466 </widget> 466 </widget>
diff --git a/src/yuzu/configuration/input_profiles.cpp b/src/yuzu/configuration/input_profiles.cpp
index 9bb69cab1..41ef4250a 100644
--- a/src/yuzu/configuration/input_profiles.cpp
+++ b/src/yuzu/configuration/input_profiles.cpp
@@ -58,13 +58,16 @@ std::vector<std::string> InputProfiles::GetInputProfileNames() {
58 std::vector<std::string> profile_names; 58 std::vector<std::string> profile_names;
59 profile_names.reserve(map_profiles.size()); 59 profile_names.reserve(map_profiles.size());
60 60
61 for (const auto& [profile_name, config] : map_profiles) { 61 auto it = map_profiles.cbegin();
62 while (it != map_profiles.cend()) {
63 const auto& [profile_name, config] = *it;
62 if (!ProfileExistsInFilesystem(profile_name)) { 64 if (!ProfileExistsInFilesystem(profile_name)) {
63 DeleteProfile(profile_name); 65 it = map_profiles.erase(it);
64 continue; 66 continue;
65 } 67 }
66 68
67 profile_names.push_back(profile_name); 69 profile_names.push_back(profile_name);
70 ++it;
68 } 71 }
69 72
70 std::stable_sort(profile_names.begin(), profile_names.end()); 73 std::stable_sort(profile_names.begin(), profile_names.end());
diff --git a/src/yuzu/main.cpp b/src/yuzu/main.cpp
index 62aaf41bf..42b7b64c8 100644
--- a/src/yuzu/main.cpp
+++ b/src/yuzu/main.cpp
@@ -983,11 +983,6 @@ void GMainWindow::InitializeWidgets() {
983 filter_status_button->setFocusPolicy(Qt::NoFocus); 983 filter_status_button->setFocusPolicy(Qt::NoFocus);
984 connect(filter_status_button, &QPushButton::clicked, this, 984 connect(filter_status_button, &QPushButton::clicked, this,
985 &GMainWindow::OnToggleAdaptingFilter); 985 &GMainWindow::OnToggleAdaptingFilter);
986 auto filter = Settings::values.scaling_filter.GetValue();
987 if (Settings::values.renderer_backend.GetValue() == Settings::RendererBackend::OpenGL &&
988 filter == Settings::ScalingFilter::Fsr) {
989 Settings::values.scaling_filter.SetValue(Settings::ScalingFilter::NearestNeighbor);
990 }
991 UpdateFilterText(); 986 UpdateFilterText();
992 filter_status_button->setCheckable(true); 987 filter_status_button->setCheckable(true);
993 filter_status_button->setChecked(true); 988 filter_status_button->setChecked(true);
@@ -3468,10 +3463,6 @@ void GMainWindow::OnToggleAdaptingFilter() {
3468 } else { 3463 } else {
3469 filter = static_cast<Settings::ScalingFilter>(static_cast<u32>(filter) + 1); 3464 filter = static_cast<Settings::ScalingFilter>(static_cast<u32>(filter) + 1);
3470 } 3465 }
3471 if (Settings::values.renderer_backend.GetValue() == Settings::RendererBackend::OpenGL &&
3472 filter == Settings::ScalingFilter::Fsr) {
3473 filter = Settings::ScalingFilter::NearestNeighbor;
3474 }
3475 Settings::values.scaling_filter.SetValue(filter); 3466 Settings::values.scaling_filter.SetValue(filter);
3476 filter_status_button->setChecked(true); 3467 filter_status_button->setChecked(true);
3477 UpdateFilterText(); 3468 UpdateFilterText();
diff --git a/src/yuzu/multiplayer/direct_connect.cpp b/src/yuzu/multiplayer/direct_connect.cpp
index cbd52da85..d71cc23a7 100644
--- a/src/yuzu/multiplayer/direct_connect.cpp
+++ b/src/yuzu/multiplayer/direct_connect.cpp
@@ -81,20 +81,13 @@ void DirectConnectWindow::Connect() {
81 } 81 }
82 } 82 }
83 } 83 }
84 switch (static_cast<ConnectionType>(ui->connection_type->currentIndex())) { 84 if (!ui->ip->hasAcceptableInput()) {
85 case ConnectionType::TraversalServer: 85 NetworkMessage::ErrorManager::ShowError(NetworkMessage::ErrorManager::IP_ADDRESS_NOT_VALID);
86 break; 86 return;
87 case ConnectionType::IP: 87 }
88 if (!ui->ip->hasAcceptableInput()) { 88 if (!ui->port->hasAcceptableInput()) {
89 NetworkMessage::ErrorManager::ShowError( 89 NetworkMessage::ErrorManager::ShowError(NetworkMessage::ErrorManager::PORT_NOT_VALID);
90 NetworkMessage::ErrorManager::IP_ADDRESS_NOT_VALID); 90 return;
91 return;
92 }
93 if (!ui->port->hasAcceptableInput()) {
94 NetworkMessage::ErrorManager::ShowError(NetworkMessage::ErrorManager::PORT_NOT_VALID);
95 return;
96 }
97 break;
98 } 91 }
99 92
100 // Store settings 93 // Store settings
diff --git a/src/yuzu/multiplayer/direct_connect.ui b/src/yuzu/multiplayer/direct_connect.ui
index 57d6ec25a..0dd4e6829 100644
--- a/src/yuzu/multiplayer/direct_connect.ui
+++ b/src/yuzu/multiplayer/direct_connect.ui
@@ -27,19 +27,10 @@
27 <number>0</number> 27 <number>0</number>
28 </property> 28 </property>
29 <item> 29 <item>
30 <widget class="QComboBox" name="connection_type">
31 <item>
32 <property name="text">
33 <string>IP Address</string>
34 </property>
35 </item>
36 </widget>
37 </item>
38 <item>
39 <widget class="QWidget" name="ip_container" native="true"> 30 <widget class="QWidget" name="ip_container" native="true">
40 <layout class="QHBoxLayout" name="ip_layout"> 31 <layout class="QHBoxLayout" name="ip_layout">
41 <property name="leftMargin"> 32 <property name="leftMargin">
42 <number>5</number> 33 <number>0</number>
43 </property> 34 </property>
44 <property name="topMargin"> 35 <property name="topMargin">
45 <number>0</number> 36 <number>0</number>
@@ -53,17 +44,17 @@
53 <item> 44 <item>
54 <widget class="QLabel" name="label_2"> 45 <widget class="QLabel" name="label_2">
55 <property name="text"> 46 <property name="text">
56 <string>IP</string> 47 <string>Server Address</string>
57 </property> 48 </property>
58 </widget> 49 </widget>
59 </item> 50 </item>
60 <item> 51 <item>
61 <widget class="QLineEdit" name="ip"> 52 <widget class="QLineEdit" name="ip">
62 <property name="toolTip"> 53 <property name="toolTip">
63 <string>&lt;html&gt;&lt;head/&gt;&lt;body&gt;&lt;p&gt;IPv4 address of the host&lt;/p&gt;&lt;/body&gt;&lt;/html&gt;</string> 54 <string>&lt;html&gt;&lt;head/&gt;&lt;body&gt;&lt;p&gt;Server address of the host&lt;/p&gt;&lt;/body&gt;&lt;/html&gt;</string>
64 </property> 55 </property>
65 <property name="maxLength"> 56 <property name="maxLength">
66 <number>16</number> 57 <number>253</number>
67 </property> 58 </property>
68 </widget> 59 </widget>
69 </item> 60 </item>
@@ -85,6 +76,12 @@
85 <property name="placeholderText"> 76 <property name="placeholderText">
86 <string notr="true" extracomment="placeholder string that tells user default port">24872</string> 77 <string notr="true" extracomment="placeholder string that tells user default port">24872</string>
87 </property> 78 </property>
79 <property name="maximumSize">
80 <size>
81 <width>65</width>
82 <height>50</height>
83 </size>
84 </property>
88 </widget> 85 </widget>
89 </item> 86 </item>
90 </layout> 87 </layout>
diff --git a/src/yuzu/multiplayer/validation.h b/src/yuzu/multiplayer/validation.h
index dd25af280..cbbe6757b 100644
--- a/src/yuzu/multiplayer/validation.h
+++ b/src/yuzu/multiplayer/validation.h
@@ -38,11 +38,28 @@ private:
38 QRegularExpression(QStringLiteral("^[a-zA-Z0-9._ -]{4,20}")); 38 QRegularExpression(QStringLiteral("^[a-zA-Z0-9._ -]{4,20}"));
39 QRegularExpressionValidator nickname; 39 QRegularExpressionValidator nickname;
40 40
41 /// ipv4 address only 41 /// ipv4 / ipv6 / hostnames
42 // TODO remove this when we support hostnames in direct connect
43 QRegularExpression ip_regex = QRegularExpression(QStringLiteral( 42 QRegularExpression ip_regex = QRegularExpression(QStringLiteral(
44 "(([0-9]|[1-9][0-9]|1[0-9]{2}|2[0-4][0-9]|25[0-5])\\.){3}([0-9]|[1-9][0-9]|1[0-9]{2}|" 43 // IPv4 regex
45 "2[0-4][0-9]|25[0-5])")); 44 "^((25[0-5]|2[0-4][0-9]|[01]?[0-9][0-9]?)\\.){3}(25[0-5]|2[0-4][0-9]|[01]?[0-9][0-9]?)$|"
45 // IPv6 regex
46 "^((([0-9A-Fa-f]{1,4}:){7}([0-9A-Fa-f]{1,4}|:))|"
47 "(([0-9A-Fa-f]{1,4}:){6}(:[0-9A-Fa-f]{1,4}|((25[0-5]|2[0-4]\\d|1\\d\\d|[1-9]?\\d)(\\.(25[0-"
48 "5]|2[0-4]\\d|1\\d\\d|[1-9]?\\d)){3})|:))|"
49 "(([0-9A-Fa-f]{1,4}:){5}(((:[0-9A-Fa-f]{1,4}){1,2})|:((25[0-5]|2[0-4]\\d|1\\d\\d|[1-9]?\\d)"
50 "(\\.(25[0-5]|2[0-4]\\d|1\\d\\d|[1-9]?\\d)){3})|:))|"
51 "(([0-9A-Fa-f]{1,4}:){4}(((:[0-9A-Fa-f]{1,4}){1,3})|((:[0-9A-Fa-f]{1,4})?:((25[0-5]|2[0-4]"
52 "\\d|1\\d\\d|[1-9]?\\d)(\\.(25[0-5]|2[0-4]\\d|1\\d\\d|[1-9]?\\d)){3}))|:))|"
53 "(([0-9A-Fa-f]{1,4}:){3}(((:[0-9A-Fa-f]{1,4}){1,4})|((:[0-9A-Fa-f]{1,4}){0,2}:((25[0-5]|2["
54 "0-4]\\d|1\\d\\d|[1-9]?\\d)(\\.(25[0-5]|2[0-4]\\d|1\\d\\d|[1-9]?\\d)){3}))|:))|"
55 "(([0-9A-Fa-f]{1,4}:){2}(((:[0-9A-Fa-f]{1,4}){1,5})|((:[0-9A-Fa-f]{1,4}){0,3}:((25[0-5]|2["
56 "0-4]\\d|1\\d\\d|[1-9]?\\d)(\\.(25[0-5]|2[0-4]\\d|1\\d\\d|[1-9]?\\d)){3}))|:))|"
57 "(([0-9A-Fa-f]{1,4}:){1}(((:[0-9A-Fa-f]{1,4}){1,6})|((:[0-9A-Fa-f]{1,4}){0,4}:((25[0-5]|2["
58 "0-4]\\d|1\\d\\d|[1-9]?\\d)(\\.(25[0-5]|2[0-4]\\d|1\\d\\d|[1-9]?\\d)){3}))|:))|"
59 "(:(((:[0-9A-Fa-f]{1,4}){1,7})|((:[0-9A-Fa-f]{1,4}){0,5}:((25[0-5]|2[0-4]\\d|1\\d\\d|[1-9]?"
60 "\\d)(\\.(25[0-5]|2[0-4]\\d|1\\d\\d|[1-9]?\\d)){3}))|:)))(%.+)?$|"
61 // Hostname regex
62 "^([a-zA-Z0-9]+(-[a-zA-Z0-9]+)*\\.)+[a-zA-Z]{2,}$"));
46 QRegularExpressionValidator ip; 63 QRegularExpressionValidator ip;
47 64
48 /// port must be between 0 and 65535 65 /// port must be between 0 and 65535
HIC SVNT DRACONES