summaryrefslogtreecommitdiff
path: root/src
diff options
context:
space:
mode:
Diffstat (limited to 'src')
-rw-r--r--src/audio_core/renderer/command/effect/reverb.cpp3
-rw-r--r--src/audio_core/renderer/effect/i3dl2.h3
-rw-r--r--src/audio_core/renderer/effect/reverb.h8
-rw-r--r--src/common/settings.h8
-rw-r--r--src/core/constants.cpp27
-rw-r--r--src/core/constants.h2
-rw-r--r--src/core/hle/kernel/k_scoped_lock.h1
-rw-r--r--src/core/hle/service/psc/psc.cpp18
-rw-r--r--src/core/hle/service/ssl/ssl.cpp76
-rw-r--r--src/core/hle/service/usb/usb.cpp70
-rw-r--r--src/input_common/helpers/joycon_protocol/joycon_types.h12
-rw-r--r--src/input_common/helpers/joycon_protocol/poller.cpp95
-rw-r--r--src/input_common/helpers/joycon_protocol/poller.h9
-rw-r--r--src/video_core/buffer_cache/buffer_cache.h53
-rw-r--r--src/video_core/engines/maxwell_dma.cpp107
-rw-r--r--src/video_core/engines/maxwell_dma.h88
-rw-r--r--src/video_core/renderer_null/null_rasterizer.h8
-rw-r--r--src/video_core/renderer_opengl/gl_rasterizer.h10
-rw-r--r--src/video_core/renderer_vulkan/vk_rasterizer.cpp234
-rw-r--r--src/video_core/renderer_vulkan/vk_rasterizer.h11
-rw-r--r--src/video_core/renderer_vulkan/vk_texture_cache.cpp14
-rw-r--r--src/video_core/texture_cache/image_info.cpp45
-rw-r--r--src/video_core/texture_cache/image_info.h2
-rw-r--r--src/video_core/texture_cache/texture_cache.h69
-rw-r--r--src/video_core/texture_cache/texture_cache_base.h5
-rw-r--r--src/video_core/texture_cache/types.h1
-rw-r--r--src/video_core/texture_cache/util.cpp98
-rw-r--r--src/video_core/texture_cache/util.h10
28 files changed, 877 insertions, 210 deletions
diff --git a/src/audio_core/renderer/command/effect/reverb.cpp b/src/audio_core/renderer/command/effect/reverb.cpp
index 6fe844ff0..8b9b65214 100644
--- a/src/audio_core/renderer/command/effect/reverb.cpp
+++ b/src/audio_core/renderer/command/effect/reverb.cpp
@@ -308,7 +308,8 @@ static void ApplyReverbEffect(const ReverbInfo::ParameterVersion2& params, Rever
308 } 308 }
309 309
310 Common::FixedPoint<50, 14> pre_delay_sample{ 310 Common::FixedPoint<50, 14> pre_delay_sample{
311 state.pre_delay_line.Read() * Common::FixedPoint<50, 14>::from_base(params.late_gain)}; 311 state.pre_delay_line.TapOut(state.pre_delay_time) *
312 Common::FixedPoint<50, 14>::from_base(params.late_gain)};
312 313
313 std::array<Common::FixedPoint<50, 14>, ReverbInfo::MaxDelayLines> mix_matrix{ 314 std::array<Common::FixedPoint<50, 14>, ReverbInfo::MaxDelayLines> mix_matrix{
314 state.prev_feedback_output[2] + state.prev_feedback_output[1] + pre_delay_sample, 315 state.prev_feedback_output[2] + state.prev_feedback_output[1] + pre_delay_sample,
diff --git a/src/audio_core/renderer/effect/i3dl2.h b/src/audio_core/renderer/effect/i3dl2.h
index 1ebbc5c4c..6e3ffd1d4 100644
--- a/src/audio_core/renderer/effect/i3dl2.h
+++ b/src/audio_core/renderer/effect/i3dl2.h
@@ -104,7 +104,8 @@ public:
104 } 104 }
105 105
106 void Write(const Common::FixedPoint<50, 14> sample) { 106 void Write(const Common::FixedPoint<50, 14> sample) {
107 *(input++) = sample; 107 *input = sample;
108 input++;
108 if (input >= buffer_end) { 109 if (input >= buffer_end) {
109 input = buffer.data(); 110 input = buffer.data();
110 } 111 }
diff --git a/src/audio_core/renderer/effect/reverb.h b/src/audio_core/renderer/effect/reverb.h
index a72475c3c..6cc345ef6 100644
--- a/src/audio_core/renderer/effect/reverb.h
+++ b/src/audio_core/renderer/effect/reverb.h
@@ -79,12 +79,10 @@ public:
79 return; 79 return;
80 } 80 }
81 sample_count = delay_time; 81 sample_count = delay_time;
82 input = &buffer[(output - buffer.data() + sample_count) % (sample_count_max + 1)]; 82 input = &buffer[0];
83 } 83 }
84 84
85 Common::FixedPoint<50, 14> Tick(const Common::FixedPoint<50, 14> sample) { 85 Common::FixedPoint<50, 14> Tick(const Common::FixedPoint<50, 14> sample) {
86 Write(sample);
87
88 auto out_sample{Read()}; 86 auto out_sample{Read()};
89 87
90 output++; 88 output++;
@@ -92,6 +90,7 @@ public:
92 output = buffer.data(); 90 output = buffer.data();
93 } 91 }
94 92
93 Write(sample);
95 return out_sample; 94 return out_sample;
96 } 95 }
97 96
@@ -100,7 +99,8 @@ public:
100 } 99 }
101 100
102 void Write(const Common::FixedPoint<50, 14> sample) { 101 void Write(const Common::FixedPoint<50, 14> sample) {
103 *(input++) = sample; 102 *input = sample;
103 input++;
104 if (input >= buffer_end) { 104 if (input >= buffer_end) {
105 input = buffer.data(); 105 input = buffer.data();
106 } 106 }
diff --git a/src/common/settings.h b/src/common/settings.h
index 512ecff69..1ae28ce93 100644
--- a/src/common/settings.h
+++ b/src/common/settings.h
@@ -128,7 +128,7 @@ public:
128 /** 128 /**
129 * Sets a default value, label, and setting value. 129 * Sets a default value, label, and setting value.
130 * 130 *
131 * @param default_val Intial value of the setting, and default value of the setting 131 * @param default_val Initial value of the setting, and default value of the setting
132 * @param name Label for the setting 132 * @param name Label for the setting
133 */ 133 */
134 explicit Setting(const Type& default_val, const std::string& name) 134 explicit Setting(const Type& default_val, const std::string& name)
@@ -139,7 +139,7 @@ public:
139 /** 139 /**
140 * Sets a default value, minimum value, maximum value, and label. 140 * Sets a default value, minimum value, maximum value, and label.
141 * 141 *
142 * @param default_val Intial value of the setting, and default value of the setting 142 * @param default_val Initial value of the setting, and default value of the setting
143 * @param min_val Sets the minimum allowed value of the setting 143 * @param min_val Sets the minimum allowed value of the setting
144 * @param max_val Sets the maximum allowed value of the setting 144 * @param max_val Sets the maximum allowed value of the setting
145 * @param name Label for the setting 145 * @param name Label for the setting
@@ -231,7 +231,7 @@ public:
231 /** 231 /**
232 * Sets a default value, label, and setting value. 232 * Sets a default value, label, and setting value.
233 * 233 *
234 * @param default_val Intial value of the setting, and default value of the setting 234 * @param default_val Initial value of the setting, and default value of the setting
235 * @param name Label for the setting 235 * @param name Label for the setting
236 */ 236 */
237 explicit SwitchableSetting(const Type& default_val, const std::string& name) 237 explicit SwitchableSetting(const Type& default_val, const std::string& name)
@@ -242,7 +242,7 @@ public:
242 /** 242 /**
243 * Sets a default value, minimum value, maximum value, and label. 243 * Sets a default value, minimum value, maximum value, and label.
244 * 244 *
245 * @param default_val Intial value of the setting, and default value of the setting 245 * @param default_val Initial value of the setting, and default value of the setting
246 * @param min_val Sets the minimum allowed value of the setting 246 * @param min_val Sets the minimum allowed value of the setting
247 * @param max_val Sets the maximum allowed value of the setting 247 * @param max_val Sets the maximum allowed value of the setting
248 * @param name Label for the setting 248 * @param name Label for the setting
diff --git a/src/core/constants.cpp b/src/core/constants.cpp
index 4430173ef..760dc5f23 100644
--- a/src/core/constants.cpp
+++ b/src/core/constants.cpp
@@ -4,13 +4,24 @@
4#include "core/constants.h" 4#include "core/constants.h"
5 5
6namespace Core::Constants { 6namespace Core::Constants {
7const std::array<u8, 107> ACCOUNT_BACKUP_JPEG{{ 7const std::array<u8, 287> ACCOUNT_BACKUP_JPEG{{
8 0xff, 0xd8, 0xff, 0xdb, 0x00, 0x43, 0x00, 0x03, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x02, 0x02, 8 0xff, 0xd8, 0xff, 0xe0, 0x00, 0x10, 0x4a, 0x46, 0x49, 0x46, 0x00, 0x01, 0x01, 0x01, 0x00, 0x48,
9 0x02, 0x03, 0x03, 0x03, 0x03, 0x04, 0x06, 0x04, 0x04, 0x04, 0x04, 0x04, 0x08, 0x06, 0x06, 0x05, 9 0x00, 0x48, 0x00, 0x00, 0xff, 0xdb, 0x00, 0x43, 0x00, 0x06, 0x04, 0x04, 0x04, 0x05, 0x04, 0x06,
10 0x06, 0x09, 0x08, 0x0a, 0x0a, 0x09, 0x08, 0x09, 0x09, 0x0a, 0x0c, 0x0f, 0x0c, 0x0a, 0x0b, 0x0e, 10 0x05, 0x05, 0x06, 0x09, 0x06, 0x05, 0x06, 0x09, 0x0b, 0x08, 0x06, 0x06, 0x08, 0x0b, 0x0c, 0x0a,
11 0x0b, 0x09, 0x09, 0x0d, 0x11, 0x0d, 0x0e, 0x0f, 0x10, 0x10, 0x11, 0x10, 0x0a, 0x0c, 0x12, 0x13, 11 0x0a, 0x0b, 0x0a, 0x0a, 0x0c, 0x10, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x10, 0x0c, 0x0e, 0x0f,
12 0x12, 0x10, 0x13, 0x0f, 0x10, 0x10, 0x10, 0xff, 0xc9, 0x00, 0x0b, 0x08, 0x00, 0x01, 0x00, 0x01, 12 0x10, 0x0f, 0x0e, 0x0c, 0x13, 0x13, 0x14, 0x14, 0x13, 0x13, 0x1c, 0x1b, 0x1b, 0x1b, 0x1c, 0x20,
13 0x01, 0x01, 0x11, 0x00, 0xff, 0xcc, 0x00, 0x06, 0x00, 0x10, 0x10, 0x05, 0xff, 0xda, 0x00, 0x08, 13 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0xff, 0xdb, 0x00, 0x43, 0x01, 0x07, 0x07,
14 0x01, 0x01, 0x00, 0x00, 0x3f, 0x00, 0xd2, 0xcf, 0x20, 0xff, 0xd9, 14 0x07, 0x0d, 0x0c, 0x0d, 0x18, 0x10, 0x10, 0x18, 0x1a, 0x15, 0x11, 0x15, 0x1a, 0x20, 0x20, 0x20,
15 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
16 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
17 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0xff, 0xc0,
18 0x00, 0x11, 0x08, 0x00, 0x20, 0x00, 0x20, 0x03, 0x01, 0x22, 0x00, 0x02, 0x11, 0x01, 0x03, 0x11,
19 0x01, 0xff, 0xc4, 0x00, 0x14, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
20 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xff, 0xc4, 0x00, 0x14, 0x10, 0x01, 0x00, 0x00, 0x00,
21 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xff, 0xc4, 0x00,
22 0x14, 0x01, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
23 0x00, 0x00, 0x00, 0xff, 0xc4, 0x00, 0x14, 0x11, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
24 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xff, 0xda, 0x00, 0x0c, 0x03, 0x01, 0x00,
25 0x02, 0x11, 0x03, 0x11, 0x00, 0x3f, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xff, 0xd9,
15}}; 26}};
16} 27}
diff --git a/src/core/constants.h b/src/core/constants.h
index f916ce0b6..f1f67d3b8 100644
--- a/src/core/constants.h
+++ b/src/core/constants.h
@@ -12,6 +12,6 @@
12namespace Core::Constants { 12namespace Core::Constants {
13 13
14// ACC Service - Blank JPEG used as user icon in absentia of real one. 14// ACC Service - Blank JPEG used as user icon in absentia of real one.
15extern const std::array<u8, 107> ACCOUNT_BACKUP_JPEG; 15extern const std::array<u8, 287> ACCOUNT_BACKUP_JPEG;
16 16
17} // namespace Core::Constants 17} // namespace Core::Constants
diff --git a/src/core/hle/kernel/k_scoped_lock.h b/src/core/hle/kernel/k_scoped_lock.h
index 59b3e32ae..a15640fd2 100644
--- a/src/core/hle/kernel/k_scoped_lock.h
+++ b/src/core/hle/kernel/k_scoped_lock.h
@@ -4,6 +4,7 @@
4#pragma once 4#pragma once
5 5
6#include <concepts> 6#include <concepts>
7#include <memory>
7#include <type_traits> 8#include <type_traits>
8 9
9namespace Kernel { 10namespace Kernel {
diff --git a/src/core/hle/service/psc/psc.cpp b/src/core/hle/service/psc/psc.cpp
index 25702703e..cd0cc9287 100644
--- a/src/core/hle/service/psc/psc.cpp
+++ b/src/core/hle/service/psc/psc.cpp
@@ -11,9 +11,9 @@
11 11
12namespace Service::PSC { 12namespace Service::PSC {
13 13
14class PSC_C final : public ServiceFramework<PSC_C> { 14class IPmControl final : public ServiceFramework<IPmControl> {
15public: 15public:
16 explicit PSC_C(Core::System& system_) : ServiceFramework{system_, "psc:c"} { 16 explicit IPmControl(Core::System& system_) : ServiceFramework{system_, "psc:c"} {
17 // clang-format off 17 // clang-format off
18 static const FunctionInfo functions[] = { 18 static const FunctionInfo functions[] = {
19 {0, nullptr, "Initialize"}, 19 {0, nullptr, "Initialize"},
@@ -23,8 +23,8 @@ public:
23 {4, nullptr, "Cancel"}, 23 {4, nullptr, "Cancel"},
24 {5, nullptr, "PrintModuleInformation"}, 24 {5, nullptr, "PrintModuleInformation"},
25 {6, nullptr, "GetModuleInformation"}, 25 {6, nullptr, "GetModuleInformation"},
26 {10, nullptr, "Unknown10"}, 26 {10, nullptr, "AcquireStateLock"},
27 {11, nullptr, "Unknown11"}, 27 {11, nullptr, "HasStateLock"},
28 }; 28 };
29 // clang-format on 29 // clang-format on
30 30
@@ -49,12 +49,12 @@ public:
49 } 49 }
50}; 50};
51 51
52class PSC_M final : public ServiceFramework<PSC_M> { 52class IPmService final : public ServiceFramework<IPmService> {
53public: 53public:
54 explicit PSC_M(Core::System& system_) : ServiceFramework{system_, "psc:m"} { 54 explicit IPmService(Core::System& system_) : ServiceFramework{system_, "psc:m"} {
55 // clang-format off 55 // clang-format off
56 static const FunctionInfo functions[] = { 56 static const FunctionInfo functions[] = {
57 {0, &PSC_M::GetPmModule, "GetPmModule"}, 57 {0, &IPmService::GetPmModule, "GetPmModule"},
58 }; 58 };
59 // clang-format on 59 // clang-format on
60 60
@@ -74,8 +74,8 @@ private:
74void LoopProcess(Core::System& system) { 74void LoopProcess(Core::System& system) {
75 auto server_manager = std::make_unique<ServerManager>(system); 75 auto server_manager = std::make_unique<ServerManager>(system);
76 76
77 server_manager->RegisterNamedService("psc:c", std::make_shared<PSC_C>(system)); 77 server_manager->RegisterNamedService("psc:c", std::make_shared<IPmControl>(system));
78 server_manager->RegisterNamedService("psc:m", std::make_shared<PSC_M>(system)); 78 server_manager->RegisterNamedService("psc:m", std::make_shared<IPmService>(system));
79 ServerManager::RunServer(std::move(server_manager)); 79 ServerManager::RunServer(std::move(server_manager));
80} 80}
81 81
diff --git a/src/core/hle/service/ssl/ssl.cpp b/src/core/hle/service/ssl/ssl.cpp
index b19bc1b3e..2b99dd7ac 100644
--- a/src/core/hle/service/ssl/ssl.cpp
+++ b/src/core/hle/service/ssl/ssl.cpp
@@ -8,14 +8,36 @@
8 8
9namespace Service::SSL { 9namespace Service::SSL {
10 10
11// This is nn::ssl::sf::CertificateFormat
11enum class CertificateFormat : u32 { 12enum class CertificateFormat : u32 {
12 Pem = 1, 13 Pem = 1,
13 Der = 2, 14 Der = 2,
14}; 15};
15 16
17// This is nn::ssl::sf::ContextOption
18enum class ContextOption : u32 {
19 None = 0,
20 CrlImportDateCheckEnable = 1,
21};
22
23// This is nn::ssl::sf::SslVersion
24struct SslVersion {
25 union {
26 u32 raw{};
27
28 BitField<0, 1, u32> tls_auto;
29 BitField<3, 1, u32> tls_v10;
30 BitField<4, 1, u32> tls_v11;
31 BitField<5, 1, u32> tls_v12;
32 BitField<6, 1, u32> tls_v13;
33 BitField<24, 7, u32> api_version;
34 };
35};
36
16class ISslConnection final : public ServiceFramework<ISslConnection> { 37class ISslConnection final : public ServiceFramework<ISslConnection> {
17public: 38public:
18 explicit ISslConnection(Core::System& system_) : ServiceFramework{system_, "ISslConnection"} { 39 explicit ISslConnection(Core::System& system_, SslVersion version)
40 : ServiceFramework{system_, "ISslConnection"}, ssl_version{version} {
19 // clang-format off 41 // clang-format off
20 static const FunctionInfo functions[] = { 42 static const FunctionInfo functions[] = {
21 {0, nullptr, "SetSocketDescriptor"}, 43 {0, nullptr, "SetSocketDescriptor"},
@@ -59,11 +81,15 @@ public:
59 81
60 RegisterHandlers(functions); 82 RegisterHandlers(functions);
61 } 83 }
84
85private:
86 SslVersion ssl_version;
62}; 87};
63 88
64class ISslContext final : public ServiceFramework<ISslContext> { 89class ISslContext final : public ServiceFramework<ISslContext> {
65public: 90public:
66 explicit ISslContext(Core::System& system_) : ServiceFramework{system_, "ISslContext"} { 91 explicit ISslContext(Core::System& system_, SslVersion version)
92 : ServiceFramework{system_, "ISslContext"}, ssl_version{version} {
67 static const FunctionInfo functions[] = { 93 static const FunctionInfo functions[] = {
68 {0, &ISslContext::SetOption, "SetOption"}, 94 {0, &ISslContext::SetOption, "SetOption"},
69 {1, nullptr, "GetOption"}, 95 {1, nullptr, "GetOption"},
@@ -84,17 +110,20 @@ public:
84 } 110 }
85 111
86private: 112private:
113 SslVersion ssl_version;
114
87 void SetOption(HLERequestContext& ctx) { 115 void SetOption(HLERequestContext& ctx) {
88 struct Parameters { 116 struct Parameters {
89 u8 enable; 117 ContextOption option;
90 u32 option; 118 s32 value;
91 }; 119 };
120 static_assert(sizeof(Parameters) == 0x8, "Parameters is an invalid size");
92 121
93 IPC::RequestParser rp{ctx}; 122 IPC::RequestParser rp{ctx};
94 const auto parameters = rp.PopRaw<Parameters>(); 123 const auto parameters = rp.PopRaw<Parameters>();
95 124
96 LOG_WARNING(Service_SSL, "(STUBBED) called. enable={}, option={}", parameters.enable, 125 LOG_WARNING(Service_SSL, "(STUBBED) called. option={}, value={}", parameters.option,
97 parameters.option); 126 parameters.value);
98 127
99 IPC::ResponseBuilder rb{ctx, 2}; 128 IPC::ResponseBuilder rb{ctx, 2};
100 rb.Push(ResultSuccess); 129 rb.Push(ResultSuccess);
@@ -105,7 +134,7 @@ private:
105 134
106 IPC::ResponseBuilder rb{ctx, 2, 0, 1}; 135 IPC::ResponseBuilder rb{ctx, 2, 0, 1};
107 rb.Push(ResultSuccess); 136 rb.Push(ResultSuccess);
108 rb.PushIpcInterface<ISslConnection>(system); 137 rb.PushIpcInterface<ISslConnection>(system, ssl_version);
109 } 138 }
110 139
111 void ImportServerPki(HLERequestContext& ctx) { 140 void ImportServerPki(HLERequestContext& ctx) {
@@ -142,20 +171,21 @@ private:
142 } 171 }
143}; 172};
144 173
145class SSL final : public ServiceFramework<SSL> { 174class ISslService final : public ServiceFramework<ISslService> {
146public: 175public:
147 explicit SSL(Core::System& system_) : ServiceFramework{system_, "ssl"} { 176 explicit ISslService(Core::System& system_) : ServiceFramework{system_, "ssl"} {
148 // clang-format off 177 // clang-format off
149 static const FunctionInfo functions[] = { 178 static const FunctionInfo functions[] = {
150 {0, &SSL::CreateContext, "CreateContext"}, 179 {0, &ISslService::CreateContext, "CreateContext"},
151 {1, nullptr, "GetContextCount"}, 180 {1, nullptr, "GetContextCount"},
152 {2, nullptr, "GetCertificates"}, 181 {2, nullptr, "GetCertificates"},
153 {3, nullptr, "GetCertificateBufSize"}, 182 {3, nullptr, "GetCertificateBufSize"},
154 {4, nullptr, "DebugIoctl"}, 183 {4, nullptr, "DebugIoctl"},
155 {5, &SSL::SetInterfaceVersion, "SetInterfaceVersion"}, 184 {5, &ISslService::SetInterfaceVersion, "SetInterfaceVersion"},
156 {6, nullptr, "FlushSessionCache"}, 185 {6, nullptr, "FlushSessionCache"},
157 {7, nullptr, "SetDebugOption"}, 186 {7, nullptr, "SetDebugOption"},
158 {8, nullptr, "GetDebugOption"}, 187 {8, nullptr, "GetDebugOption"},
188 {8, nullptr, "ClearTls12FallbackFlag"},
159 }; 189 };
160 // clang-format on 190 // clang-format on
161 191
@@ -163,20 +193,30 @@ public:
163 } 193 }
164 194
165private: 195private:
166 u32 ssl_version{};
167 void CreateContext(HLERequestContext& ctx) { 196 void CreateContext(HLERequestContext& ctx) {
168 LOG_WARNING(Service_SSL, "(STUBBED) called"); 197 struct Parameters {
198 SslVersion ssl_version;
199 INSERT_PADDING_BYTES(0x4);
200 u64 pid_placeholder;
201 };
202 static_assert(sizeof(Parameters) == 0x10, "Parameters is an invalid size");
203
204 IPC::RequestParser rp{ctx};
205 const auto parameters = rp.PopRaw<Parameters>();
206
207 LOG_WARNING(Service_SSL, "(STUBBED) called, api_version={}, pid_placeholder={}",
208 parameters.ssl_version.api_version, parameters.pid_placeholder);
169 209
170 IPC::ResponseBuilder rb{ctx, 2, 0, 1}; 210 IPC::ResponseBuilder rb{ctx, 2, 0, 1};
171 rb.Push(ResultSuccess); 211 rb.Push(ResultSuccess);
172 rb.PushIpcInterface<ISslContext>(system); 212 rb.PushIpcInterface<ISslContext>(system, parameters.ssl_version);
173 } 213 }
174 214
175 void SetInterfaceVersion(HLERequestContext& ctx) { 215 void SetInterfaceVersion(HLERequestContext& ctx) {
176 LOG_DEBUG(Service_SSL, "called");
177
178 IPC::RequestParser rp{ctx}; 216 IPC::RequestParser rp{ctx};
179 ssl_version = rp.Pop<u32>(); 217 u32 ssl_version = rp.Pop<u32>();
218
219 LOG_DEBUG(Service_SSL, "called, ssl_version={}", ssl_version);
180 220
181 IPC::ResponseBuilder rb{ctx, 2}; 221 IPC::ResponseBuilder rb{ctx, 2};
182 rb.Push(ResultSuccess); 222 rb.Push(ResultSuccess);
@@ -186,7 +226,7 @@ private:
186void LoopProcess(Core::System& system) { 226void LoopProcess(Core::System& system) {
187 auto server_manager = std::make_unique<ServerManager>(system); 227 auto server_manager = std::make_unique<ServerManager>(system);
188 228
189 server_manager->RegisterNamedService("ssl", std::make_shared<SSL>(system)); 229 server_manager->RegisterNamedService("ssl", std::make_shared<ISslService>(system));
190 ServerManager::RunServer(std::move(server_manager)); 230 ServerManager::RunServer(std::move(server_manager));
191} 231}
192 232
diff --git a/src/core/hle/service/usb/usb.cpp b/src/core/hle/service/usb/usb.cpp
index a2855e783..f29fff1dd 100644
--- a/src/core/hle/service/usb/usb.cpp
+++ b/src/core/hle/service/usb/usb.cpp
@@ -16,19 +16,19 @@ public:
16 explicit IDsInterface(Core::System& system_) : ServiceFramework{system_, "IDsInterface"} { 16 explicit IDsInterface(Core::System& system_) : ServiceFramework{system_, "IDsInterface"} {
17 // clang-format off 17 // clang-format off
18 static const FunctionInfo functions[] = { 18 static const FunctionInfo functions[] = {
19 {0, nullptr, "BindDevice"}, 19 {0, nullptr, "AddEndpoint"},
20 {1, nullptr, "BindClientProcess"}, 20 {1, nullptr, "GetSetupEvent"},
21 {2, nullptr, "AddInterface"}, 21 {2, nullptr, "GetSetupPacket"},
22 {3, nullptr, "GetStateChangeEvent"}, 22 {3, nullptr, "Enable"},
23 {4, nullptr, "GetState"}, 23 {4, nullptr, "Disable"},
24 {5, nullptr, "ClearDeviceData"}, 24 {5, nullptr, "CtrlIn"},
25 {6, nullptr, "AddUsbStringDescriptor"}, 25 {6, nullptr, "CtrlOut"},
26 {7, nullptr, "DeleteUsbStringDescriptor"}, 26 {7, nullptr, "GetCtrlInCompletionEvent"},
27 {8, nullptr, "SetUsbDeviceDescriptor"}, 27 {8, nullptr, "GetCtrlInUrbReport"},
28 {9, nullptr, "SetBinaryObjectStore"}, 28 {9, nullptr, "GetCtrlOutCompletionEvent"},
29 {10, nullptr, "Enable"}, 29 {10, nullptr, "GetCtrlOutUrbReport"},
30 {11, nullptr, "Disable"}, 30 {11, nullptr, "CtrlStall"},
31 {12, nullptr, "Unknown12"}, 31 {12, nullptr, "AppendConfigurationData"},
32 }; 32 };
33 // clang-format on 33 // clang-format on
34 34
@@ -36,9 +36,9 @@ public:
36 } 36 }
37}; 37};
38 38
39class USB_DS final : public ServiceFramework<USB_DS> { 39class IDsRootSession final : public ServiceFramework<IDsRootSession> {
40public: 40public:
41 explicit USB_DS(Core::System& system_) : ServiceFramework{system_, "usb:ds"} { 41 explicit IDsRootSession(Core::System& system_) : ServiceFramework{system_, "usb:ds"} {
42 // clang-format off 42 // clang-format off
43 static const FunctionInfo functions[] = { 43 static const FunctionInfo functions[] = {
44 {0, nullptr, "OpenDsService"}, 44 {0, nullptr, "OpenDsService"},
@@ -94,9 +94,9 @@ public:
94 } 94 }
95}; 95};
96 96
97class USB_HS final : public ServiceFramework<USB_HS> { 97class IClientRootSession final : public ServiceFramework<IClientRootSession> {
98public: 98public:
99 explicit USB_HS(Core::System& system_) : ServiceFramework{system_, "usb:hs"} { 99 explicit IClientRootSession(Core::System& system_) : ServiceFramework{system_, "usb:hs"} {
100 // clang-format off 100 // clang-format off
101 static const FunctionInfo functions[] = { 101 static const FunctionInfo functions[] = {
102 {0, nullptr, "BindClientProcess"}, 102 {0, nullptr, "BindClientProcess"},
@@ -107,7 +107,7 @@ public:
107 {5, nullptr, "DestroyInterfaceAvailableEvent"}, 107 {5, nullptr, "DestroyInterfaceAvailableEvent"},
108 {6, nullptr, "GetInterfaceStateChangeEvent"}, 108 {6, nullptr, "GetInterfaceStateChangeEvent"},
109 {7, nullptr, "AcquireUsbIf"}, 109 {7, nullptr, "AcquireUsbIf"},
110 {8, nullptr, "ResetDevice"}, 110 {8, nullptr, "SetTestMode"},
111 }; 111 };
112 // clang-format on 112 // clang-format on
113 113
@@ -134,12 +134,12 @@ public:
134 } 134 }
135}; 135};
136 136
137class USB_PD final : public ServiceFramework<USB_PD> { 137class IPdManager final : public ServiceFramework<IPdManager> {
138public: 138public:
139 explicit USB_PD(Core::System& system_) : ServiceFramework{system_, "usb:pd"} { 139 explicit IPdManager(Core::System& system_) : ServiceFramework{system_, "usb:pd"} {
140 // clang-format off 140 // clang-format off
141 static const FunctionInfo functions[] = { 141 static const FunctionInfo functions[] = {
142 {0, &USB_PD::GetPdSession, "GetPdSession"}, 142 {0, &IPdManager::OpenSession, "OpenSession"},
143 }; 143 };
144 // clang-format on 144 // clang-format on
145 145
@@ -147,7 +147,7 @@ public:
147 } 147 }
148 148
149private: 149private:
150 void GetPdSession(HLERequestContext& ctx) { 150 void OpenSession(HLERequestContext& ctx) {
151 LOG_DEBUG(Service_USB, "called"); 151 LOG_DEBUG(Service_USB, "called");
152 152
153 IPC::ResponseBuilder rb{ctx, 2, 0, 1}; 153 IPC::ResponseBuilder rb{ctx, 2, 0, 1};
@@ -178,12 +178,12 @@ public:
178 } 178 }
179}; 179};
180 180
181class USB_PD_C final : public ServiceFramework<USB_PD_C> { 181class IPdCradleManager final : public ServiceFramework<IPdCradleManager> {
182public: 182public:
183 explicit USB_PD_C(Core::System& system_) : ServiceFramework{system_, "usb:pd:c"} { 183 explicit IPdCradleManager(Core::System& system_) : ServiceFramework{system_, "usb:pd:c"} {
184 // clang-format off 184 // clang-format off
185 static const FunctionInfo functions[] = { 185 static const FunctionInfo functions[] = {
186 {0, &USB_PD_C::GetPdCradleSession, "GetPdCradleSession"}, 186 {0, &IPdCradleManager::OpenCradleSession, "OpenCradleSession"},
187 }; 187 };
188 // clang-format on 188 // clang-format on
189 189
@@ -191,18 +191,18 @@ public:
191 } 191 }
192 192
193private: 193private:
194 void GetPdCradleSession(HLERequestContext& ctx) { 194 void OpenCradleSession(HLERequestContext& ctx) {
195 LOG_DEBUG(Service_USB, "called");
196
195 IPC::ResponseBuilder rb{ctx, 2, 0, 1}; 197 IPC::ResponseBuilder rb{ctx, 2, 0, 1};
196 rb.Push(ResultSuccess); 198 rb.Push(ResultSuccess);
197 rb.PushIpcInterface<IPdCradleSession>(system); 199 rb.PushIpcInterface<IPdCradleSession>(system);
198
199 LOG_DEBUG(Service_USB, "called");
200 } 200 }
201}; 201};
202 202
203class USB_PM final : public ServiceFramework<USB_PM> { 203class IPmMainService final : public ServiceFramework<IPmMainService> {
204public: 204public:
205 explicit USB_PM(Core::System& system_) : ServiceFramework{system_, "usb:pm"} { 205 explicit IPmMainService(Core::System& system_) : ServiceFramework{system_, "usb:pm"} {
206 // clang-format off 206 // clang-format off
207 static const FunctionInfo functions[] = { 207 static const FunctionInfo functions[] = {
208 {0, nullptr, "GetPowerEvent"}, 208 {0, nullptr, "GetPowerEvent"},
@@ -221,11 +221,11 @@ public:
221void LoopProcess(Core::System& system) { 221void LoopProcess(Core::System& system) {
222 auto server_manager = std::make_unique<ServerManager>(system); 222 auto server_manager = std::make_unique<ServerManager>(system);
223 223
224 server_manager->RegisterNamedService("usb:ds", std::make_shared<USB_DS>(system)); 224 server_manager->RegisterNamedService("usb:ds", std::make_shared<IDsRootSession>(system));
225 server_manager->RegisterNamedService("usb:hs", std::make_shared<USB_HS>(system)); 225 server_manager->RegisterNamedService("usb:hs", std::make_shared<IClientRootSession>(system));
226 server_manager->RegisterNamedService("usb:pd", std::make_shared<USB_PD>(system)); 226 server_manager->RegisterNamedService("usb:pd", std::make_shared<IPdManager>(system));
227 server_manager->RegisterNamedService("usb:pd:c", std::make_shared<USB_PD_C>(system)); 227 server_manager->RegisterNamedService("usb:pd:c", std::make_shared<IPdCradleManager>(system));
228 server_manager->RegisterNamedService("usb:pm", std::make_shared<USB_PM>(system)); 228 server_manager->RegisterNamedService("usb:pm", std::make_shared<IPmMainService>(system));
229 ServerManager::RunServer(std::move(server_manager)); 229 ServerManager::RunServer(std::move(server_manager));
230} 230}
231 231
diff --git a/src/input_common/helpers/joycon_protocol/joycon_types.h b/src/input_common/helpers/joycon_protocol/joycon_types.h
index b91934990..2e50a99a8 100644
--- a/src/input_common/helpers/joycon_protocol/joycon_types.h
+++ b/src/input_common/helpers/joycon_protocol/joycon_types.h
@@ -95,6 +95,18 @@ enum class PasivePadButton : u32 {
95 ZL_ZR = 0x8000, 95 ZL_ZR = 0x8000,
96}; 96};
97 97
98enum class PasivePadStick : u8 {
99 Right = 0x00,
100 RightDown = 0x01,
101 Down = 0x02,
102 DownLeft = 0x03,
103 Left = 0x04,
104 LeftUp = 0x05,
105 Up = 0x06,
106 UpRight = 0x07,
107 Neutral = 0x08,
108};
109
98enum class OutputReport : u8 { 110enum class OutputReport : u8 {
99 RUMBLE_AND_SUBCMD = 0x01, 111 RUMBLE_AND_SUBCMD = 0x01,
100 FW_UPDATE_PKT = 0x03, 112 FW_UPDATE_PKT = 0x03,
diff --git a/src/input_common/helpers/joycon_protocol/poller.cpp b/src/input_common/helpers/joycon_protocol/poller.cpp
index 9bb15e935..ab48352b8 100644
--- a/src/input_common/helpers/joycon_protocol/poller.cpp
+++ b/src/input_common/helpers/joycon_protocol/poller.cpp
@@ -12,7 +12,7 @@ JoyconPoller::JoyconPoller(ControllerType device_type_, JoyStickCalibration left
12 : device_type{device_type_}, left_stick_calibration{left_stick_calibration_}, 12 : device_type{device_type_}, left_stick_calibration{left_stick_calibration_},
13 right_stick_calibration{right_stick_calibration_}, motion_calibration{motion_calibration_} {} 13 right_stick_calibration{right_stick_calibration_}, motion_calibration{motion_calibration_} {}
14 14
15void JoyconPoller::SetCallbacks(const Joycon::JoyconCallbacks& callbacks_) { 15void JoyconPoller::SetCallbacks(const JoyconCallbacks& callbacks_) {
16 callbacks = std::move(callbacks_); 16 callbacks = std::move(callbacks_);
17} 17}
18 18
@@ -22,13 +22,13 @@ void JoyconPoller::ReadActiveMode(std::span<u8> buffer, const MotionStatus& moti
22 memcpy(&data, buffer.data(), sizeof(InputReportActive)); 22 memcpy(&data, buffer.data(), sizeof(InputReportActive));
23 23
24 switch (device_type) { 24 switch (device_type) {
25 case Joycon::ControllerType::Left: 25 case ControllerType::Left:
26 UpdateActiveLeftPadInput(data, motion_status); 26 UpdateActiveLeftPadInput(data, motion_status);
27 break; 27 break;
28 case Joycon::ControllerType::Right: 28 case ControllerType::Right:
29 UpdateActiveRightPadInput(data, motion_status); 29 UpdateActiveRightPadInput(data, motion_status);
30 break; 30 break;
31 case Joycon::ControllerType::Pro: 31 case ControllerType::Pro:
32 UpdateActiveProPadInput(data, motion_status); 32 UpdateActiveProPadInput(data, motion_status);
33 break; 33 break;
34 default: 34 default:
@@ -47,13 +47,13 @@ void JoyconPoller::ReadPassiveMode(std::span<u8> buffer) {
47 memcpy(&data, buffer.data(), sizeof(InputReportPassive)); 47 memcpy(&data, buffer.data(), sizeof(InputReportPassive));
48 48
49 switch (device_type) { 49 switch (device_type) {
50 case Joycon::ControllerType::Left: 50 case ControllerType::Left:
51 UpdatePasiveLeftPadInput(data); 51 UpdatePasiveLeftPadInput(data);
52 break; 52 break;
53 case Joycon::ControllerType::Right: 53 case ControllerType::Right:
54 UpdatePasiveRightPadInput(data); 54 UpdatePasiveRightPadInput(data);
55 break; 55 break;
56 case Joycon::ControllerType::Pro: 56 case ControllerType::Pro:
57 UpdatePasiveProPadInput(data); 57 UpdatePasiveProPadInput(data);
58 break; 58 break;
59 default: 59 default:
@@ -211,13 +211,11 @@ void JoyconPoller::UpdateActiveProPadInput(const InputReportActive& input,
211} 211}
212 212
213void JoyconPoller::UpdatePasiveLeftPadInput(const InputReportPassive& input) { 213void JoyconPoller::UpdatePasiveLeftPadInput(const InputReportPassive& input) {
214 static constexpr std::array<Joycon::PasivePadButton, 11> left_buttons{ 214 static constexpr std::array<PasivePadButton, 11> left_buttons{
215 Joycon::PasivePadButton::Down_A, Joycon::PasivePadButton::Right_X, 215 PasivePadButton::Down_A, PasivePadButton::Right_X, PasivePadButton::Left_B,
216 Joycon::PasivePadButton::Left_B, Joycon::PasivePadButton::Up_Y, 216 PasivePadButton::Up_Y, PasivePadButton::SL, PasivePadButton::SR,
217 Joycon::PasivePadButton::SL, Joycon::PasivePadButton::SR, 217 PasivePadButton::L_R, PasivePadButton::ZL_ZR, PasivePadButton::Minus,
218 Joycon::PasivePadButton::L_R, Joycon::PasivePadButton::ZL_ZR, 218 PasivePadButton::Capture, PasivePadButton::StickL,
219 Joycon::PasivePadButton::Minus, Joycon::PasivePadButton::Capture,
220 Joycon::PasivePadButton::StickL,
221 }; 219 };
222 220
223 for (auto left_button : left_buttons) { 221 for (auto left_button : left_buttons) {
@@ -225,16 +223,19 @@ void JoyconPoller::UpdatePasiveLeftPadInput(const InputReportPassive& input) {
225 const int button = static_cast<int>(left_button); 223 const int button = static_cast<int>(left_button);
226 callbacks.on_button_data(button, button_status); 224 callbacks.on_button_data(button, button_status);
227 } 225 }
226
227 const auto [left_axis_x, left_axis_y] =
228 GetPassiveAxisValue(static_cast<PasivePadStick>(input.stick_state));
229 callbacks.on_stick_data(static_cast<int>(PadAxes::LeftStickX), left_axis_x);
230 callbacks.on_stick_data(static_cast<int>(PadAxes::LeftStickY), left_axis_y);
228} 231}
229 232
230void JoyconPoller::UpdatePasiveRightPadInput(const InputReportPassive& input) { 233void JoyconPoller::UpdatePasiveRightPadInput(const InputReportPassive& input) {
231 static constexpr std::array<Joycon::PasivePadButton, 11> right_buttons{ 234 static constexpr std::array<PasivePadButton, 11> right_buttons{
232 Joycon::PasivePadButton::Down_A, Joycon::PasivePadButton::Right_X, 235 PasivePadButton::Down_A, PasivePadButton::Right_X, PasivePadButton::Left_B,
233 Joycon::PasivePadButton::Left_B, Joycon::PasivePadButton::Up_Y, 236 PasivePadButton::Up_Y, PasivePadButton::SL, PasivePadButton::SR,
234 Joycon::PasivePadButton::SL, Joycon::PasivePadButton::SR, 237 PasivePadButton::L_R, PasivePadButton::ZL_ZR, PasivePadButton::Plus,
235 Joycon::PasivePadButton::L_R, Joycon::PasivePadButton::ZL_ZR, 238 PasivePadButton::Home, PasivePadButton::StickR,
236 Joycon::PasivePadButton::Plus, Joycon::PasivePadButton::Home,
237 Joycon::PasivePadButton::StickR,
238 }; 239 };
239 240
240 for (auto right_button : right_buttons) { 241 for (auto right_button : right_buttons) {
@@ -242,17 +243,20 @@ void JoyconPoller::UpdatePasiveRightPadInput(const InputReportPassive& input) {
242 const int button = static_cast<int>(right_button); 243 const int button = static_cast<int>(right_button);
243 callbacks.on_button_data(button, button_status); 244 callbacks.on_button_data(button, button_status);
244 } 245 }
246
247 const auto [right_axis_x, right_axis_y] =
248 GetPassiveAxisValue(static_cast<PasivePadStick>(input.stick_state));
249 callbacks.on_stick_data(static_cast<int>(PadAxes::RightStickX), right_axis_x);
250 callbacks.on_stick_data(static_cast<int>(PadAxes::RightStickY), right_axis_y);
245} 251}
246 252
247void JoyconPoller::UpdatePasiveProPadInput(const InputReportPassive& input) { 253void JoyconPoller::UpdatePasiveProPadInput(const InputReportPassive& input) {
248 static constexpr std::array<Joycon::PasivePadButton, 14> pro_buttons{ 254 static constexpr std::array<PasivePadButton, 14> pro_buttons{
249 Joycon::PasivePadButton::Down_A, Joycon::PasivePadButton::Right_X, 255 PasivePadButton::Down_A, PasivePadButton::Right_X, PasivePadButton::Left_B,
250 Joycon::PasivePadButton::Left_B, Joycon::PasivePadButton::Up_Y, 256 PasivePadButton::Up_Y, PasivePadButton::SL, PasivePadButton::SR,
251 Joycon::PasivePadButton::SL, Joycon::PasivePadButton::SR, 257 PasivePadButton::L_R, PasivePadButton::ZL_ZR, PasivePadButton::Minus,
252 Joycon::PasivePadButton::L_R, Joycon::PasivePadButton::ZL_ZR, 258 PasivePadButton::Plus, PasivePadButton::Capture, PasivePadButton::Home,
253 Joycon::PasivePadButton::Minus, Joycon::PasivePadButton::Plus, 259 PasivePadButton::StickL, PasivePadButton::StickR,
254 Joycon::PasivePadButton::Capture, Joycon::PasivePadButton::Home,
255 Joycon::PasivePadButton::StickL, Joycon::PasivePadButton::StickR,
256 }; 260 };
257 261
258 for (auto pro_button : pro_buttons) { 262 for (auto pro_button : pro_buttons) {
@@ -260,6 +264,15 @@ void JoyconPoller::UpdatePasiveProPadInput(const InputReportPassive& input) {
260 const int button = static_cast<int>(pro_button); 264 const int button = static_cast<int>(pro_button);
261 callbacks.on_button_data(button, button_status); 265 callbacks.on_button_data(button, button_status);
262 } 266 }
267
268 const auto [left_axis_x, left_axis_y] =
269 GetPassiveAxisValue(static_cast<PasivePadStick>(input.stick_state && 0xf));
270 const auto [right_axis_x, right_axis_y] =
271 GetPassiveAxisValue(static_cast<PasivePadStick>(input.stick_state >> 4));
272 callbacks.on_stick_data(static_cast<int>(PadAxes::LeftStickX), left_axis_x);
273 callbacks.on_stick_data(static_cast<int>(PadAxes::LeftStickY), left_axis_y);
274 callbacks.on_stick_data(static_cast<int>(PadAxes::RightStickX), right_axis_x);
275 callbacks.on_stick_data(static_cast<int>(PadAxes::RightStickY), right_axis_y);
263} 276}
264 277
265f32 JoyconPoller::GetAxisValue(u16 raw_value, Joycon::JoyStickAxisCalibration calibration) const { 278f32 JoyconPoller::GetAxisValue(u16 raw_value, Joycon::JoyStickAxisCalibration calibration) const {
@@ -270,6 +283,30 @@ f32 JoyconPoller::GetAxisValue(u16 raw_value, Joycon::JoyStickAxisCalibration ca
270 return value / calibration.min; 283 return value / calibration.min;
271} 284}
272 285
286std::pair<f32, f32> JoyconPoller::GetPassiveAxisValue(PasivePadStick raw_value) const {
287 switch (raw_value) {
288 case PasivePadStick::Right:
289 return {1.0f, 0.0f};
290 case PasivePadStick::RightDown:
291 return {1.0f, -1.0f};
292 case PasivePadStick::Down:
293 return {0.0f, -1.0f};
294 case PasivePadStick::DownLeft:
295 return {-1.0f, -1.0f};
296 case PasivePadStick::Left:
297 return {-1.0f, 0.0f};
298 case PasivePadStick::LeftUp:
299 return {-1.0f, 1.0f};
300 case PasivePadStick::Up:
301 return {0.0f, 1.0f};
302 case PasivePadStick::UpRight:
303 return {1.0f, 1.0f};
304 case PasivePadStick::Neutral:
305 default:
306 return {0.0f, 0.0f};
307 }
308}
309
273f32 JoyconPoller::GetAccelerometerValue(s16 raw, const MotionSensorCalibration& cal, 310f32 JoyconPoller::GetAccelerometerValue(s16 raw, const MotionSensorCalibration& cal,
274 AccelerometerSensitivity sensitivity) const { 311 AccelerometerSensitivity sensitivity) const {
275 const f32 value = raw * (1.0f / (cal.scale - cal.offset)) * 4; 312 const f32 value = raw * (1.0f / (cal.scale - cal.offset)) * 4;
diff --git a/src/input_common/helpers/joycon_protocol/poller.h b/src/input_common/helpers/joycon_protocol/poller.h
index 354d41dad..5c897f070 100644
--- a/src/input_common/helpers/joycon_protocol/poller.h
+++ b/src/input_common/helpers/joycon_protocol/poller.h
@@ -22,7 +22,7 @@ public:
22 JoyStickCalibration right_stick_calibration_, 22 JoyStickCalibration right_stick_calibration_,
23 MotionCalibration motion_calibration_); 23 MotionCalibration motion_calibration_);
24 24
25 void SetCallbacks(const Joycon::JoyconCallbacks& callbacks_); 25 void SetCallbacks(const JoyconCallbacks& callbacks_);
26 26
27 /// Handles data from passive packages 27 /// Handles data from passive packages
28 void ReadPassiveMode(std::span<u8> buffer); 28 void ReadPassiveMode(std::span<u8> buffer);
@@ -51,7 +51,10 @@ private:
51 void UpdatePasiveProPadInput(const InputReportPassive& buffer); 51 void UpdatePasiveProPadInput(const InputReportPassive& buffer);
52 52
53 /// Returns a calibrated joystick axis from raw axis data 53 /// Returns a calibrated joystick axis from raw axis data
54 f32 GetAxisValue(u16 raw_value, Joycon::JoyStickAxisCalibration calibration) const; 54 f32 GetAxisValue(u16 raw_value, JoyStickAxisCalibration calibration) const;
55
56 /// Returns a digital joystick axis from passive axis data
57 std::pair<f32, f32> GetPassiveAxisValue(PasivePadStick raw_value) const;
55 58
56 /// Returns a calibrated accelerometer axis from raw motion data 59 /// Returns a calibrated accelerometer axis from raw motion data
57 f32 GetAccelerometerValue(s16 raw, const MotionSensorCalibration& cal, 60 f32 GetAccelerometerValue(s16 raw, const MotionSensorCalibration& cal,
@@ -75,7 +78,7 @@ private:
75 JoyStickCalibration right_stick_calibration{}; 78 JoyStickCalibration right_stick_calibration{};
76 MotionCalibration motion_calibration{}; 79 MotionCalibration motion_calibration{};
77 80
78 Joycon::JoyconCallbacks callbacks{}; 81 JoyconCallbacks callbacks{};
79}; 82};
80 83
81} // namespace InputCommon::Joycon 84} // namespace InputCommon::Joycon
diff --git a/src/video_core/buffer_cache/buffer_cache.h b/src/video_core/buffer_cache/buffer_cache.h
index 06fd40851..2a150ccdc 100644
--- a/src/video_core/buffer_cache/buffer_cache.h
+++ b/src/video_core/buffer_cache/buffer_cache.h
@@ -55,6 +55,19 @@ constexpr u32 NUM_STORAGE_BUFFERS = 16;
55constexpr u32 NUM_TEXTURE_BUFFERS = 16; 55constexpr u32 NUM_TEXTURE_BUFFERS = 16;
56constexpr u32 NUM_STAGES = 5; 56constexpr u32 NUM_STAGES = 5;
57 57
58enum class ObtainBufferSynchronize : u32 {
59 NoSynchronize = 0,
60 FullSynchronize = 1,
61 SynchronizeNoDirty = 2,
62};
63
64enum class ObtainBufferOperation : u32 {
65 DoNothing = 0,
66 MarkAsWritten = 1,
67 DiscardWrite = 2,
68 MarkQuery = 3,
69};
70
58using UniformBufferSizes = std::array<std::array<u32, NUM_GRAPHICS_UNIFORM_BUFFERS>, NUM_STAGES>; 71using UniformBufferSizes = std::array<std::array<u32, NUM_GRAPHICS_UNIFORM_BUFFERS>, NUM_STAGES>;
59using ComputeUniformBufferSizes = std::array<u32, NUM_COMPUTE_UNIFORM_BUFFERS>; 72using ComputeUniformBufferSizes = std::array<u32, NUM_COMPUTE_UNIFORM_BUFFERS>;
60 73
@@ -191,6 +204,10 @@ public:
191 204
192 bool DMAClear(GPUVAddr src_address, u64 amount, u32 value); 205 bool DMAClear(GPUVAddr src_address, u64 amount, u32 value);
193 206
207 [[nodiscard]] std::pair<Buffer*, u32> ObtainBuffer(GPUVAddr gpu_addr, u32 size,
208 ObtainBufferSynchronize sync_info,
209 ObtainBufferOperation post_op);
210
194 /// Return true when a CPU region is modified from the GPU 211 /// Return true when a CPU region is modified from the GPU
195 [[nodiscard]] bool IsRegionGpuModified(VAddr addr, size_t size); 212 [[nodiscard]] bool IsRegionGpuModified(VAddr addr, size_t size);
196 213
@@ -642,6 +659,42 @@ bool BufferCache<P>::DMAClear(GPUVAddr dst_address, u64 amount, u32 value) {
642} 659}
643 660
644template <class P> 661template <class P>
662std::pair<typename P::Buffer*, u32> BufferCache<P>::ObtainBuffer(GPUVAddr gpu_addr, u32 size,
663 ObtainBufferSynchronize sync_info,
664 ObtainBufferOperation post_op) {
665 const std::optional<VAddr> cpu_addr = gpu_memory->GpuToCpuAddress(gpu_addr);
666 if (!cpu_addr) {
667 return {&slot_buffers[NULL_BUFFER_ID], 0};
668 }
669 const BufferId buffer_id = FindBuffer(*cpu_addr, size);
670 Buffer& buffer = slot_buffers[buffer_id];
671
672 // synchronize op
673 switch (sync_info) {
674 case ObtainBufferSynchronize::FullSynchronize:
675 SynchronizeBuffer(buffer, *cpu_addr, size);
676 break;
677 default:
678 break;
679 }
680
681 switch (post_op) {
682 case ObtainBufferOperation::MarkAsWritten:
683 MarkWrittenBuffer(buffer_id, *cpu_addr, size);
684 break;
685 case ObtainBufferOperation::DiscardWrite: {
686 IntervalType interval{*cpu_addr, size};
687 ClearDownload(interval);
688 break;
689 }
690 default:
691 break;
692 }
693
694 return {&buffer, buffer.Offset(*cpu_addr)};
695}
696
697template <class P>
645void BufferCache<P>::BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr, 698void BufferCache<P>::BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr,
646 u32 size) { 699 u32 size) {
647 const std::optional<VAddr> cpu_addr = gpu_memory->GpuToCpuAddress(gpu_addr); 700 const std::optional<VAddr> cpu_addr = gpu_memory->GpuToCpuAddress(gpu_addr);
diff --git a/src/video_core/engines/maxwell_dma.cpp b/src/video_core/engines/maxwell_dma.cpp
index 7762c7d96..e68850dc5 100644
--- a/src/video_core/engines/maxwell_dma.cpp
+++ b/src/video_core/engines/maxwell_dma.cpp
@@ -14,7 +14,13 @@
14#include "video_core/textures/decoders.h" 14#include "video_core/textures/decoders.h"
15 15
16MICROPROFILE_DECLARE(GPU_DMAEngine); 16MICROPROFILE_DECLARE(GPU_DMAEngine);
17MICROPROFILE_DECLARE(GPU_DMAEngineBL);
18MICROPROFILE_DECLARE(GPU_DMAEngineLB);
19MICROPROFILE_DECLARE(GPU_DMAEngineBB);
17MICROPROFILE_DEFINE(GPU_DMAEngine, "GPU", "DMA Engine", MP_RGB(224, 224, 128)); 20MICROPROFILE_DEFINE(GPU_DMAEngine, "GPU", "DMA Engine", MP_RGB(224, 224, 128));
21MICROPROFILE_DEFINE(GPU_DMAEngineBL, "GPU", "DMA Engine Block - Linear", MP_RGB(224, 224, 128));
22MICROPROFILE_DEFINE(GPU_DMAEngineLB, "GPU", "DMA Engine Linear - Block", MP_RGB(224, 224, 128));
23MICROPROFILE_DEFINE(GPU_DMAEngineBB, "GPU", "DMA Engine Block - Block", MP_RGB(224, 224, 128));
18 24
19namespace Tegra::Engines { 25namespace Tegra::Engines {
20 26
@@ -72,6 +78,7 @@ void MaxwellDMA::Launch() {
72 memory_manager.FlushCaching(); 78 memory_manager.FlushCaching();
73 if (!is_src_pitch && !is_dst_pitch) { 79 if (!is_src_pitch && !is_dst_pitch) {
74 // If both the source and the destination are in block layout, assert. 80 // If both the source and the destination are in block layout, assert.
81 MICROPROFILE_SCOPE(GPU_DMAEngineBB);
75 CopyBlockLinearToBlockLinear(); 82 CopyBlockLinearToBlockLinear();
76 ReleaseSemaphore(); 83 ReleaseSemaphore();
77 return; 84 return;
@@ -87,8 +94,10 @@ void MaxwellDMA::Launch() {
87 } 94 }
88 } else { 95 } else {
89 if (!is_src_pitch && is_dst_pitch) { 96 if (!is_src_pitch && is_dst_pitch) {
97 MICROPROFILE_SCOPE(GPU_DMAEngineBL);
90 CopyBlockLinearToPitch(); 98 CopyBlockLinearToPitch();
91 } else { 99 } else {
100 MICROPROFILE_SCOPE(GPU_DMAEngineLB);
92 CopyPitchToBlockLinear(); 101 CopyPitchToBlockLinear();
93 } 102 }
94 } 103 }
@@ -153,21 +162,35 @@ void MaxwellDMA::Launch() {
153} 162}
154 163
155void MaxwellDMA::CopyBlockLinearToPitch() { 164void MaxwellDMA::CopyBlockLinearToPitch() {
156 UNIMPLEMENTED_IF(regs.src_params.block_size.width != 0); 165 UNIMPLEMENTED_IF(regs.launch_dma.remap_enable != 0);
157 UNIMPLEMENTED_IF(regs.src_params.layer != 0); 166
158 167 u32 bytes_per_pixel = 1;
159 const bool is_remapping = regs.launch_dma.remap_enable != 0; 168 DMA::ImageOperand src_operand;
160 169 src_operand.bytes_per_pixel = bytes_per_pixel;
161 // Optimized path for micro copies. 170 src_operand.params = regs.src_params;
162 const size_t dst_size = static_cast<size_t>(regs.pitch_out) * regs.line_count; 171 src_operand.address = regs.offset_in;
163 if (!is_remapping && dst_size < GOB_SIZE && regs.pitch_out <= GOB_SIZE_X && 172
164 regs.src_params.height > GOB_SIZE_Y) { 173 DMA::BufferOperand dst_operand;
165 FastCopyBlockLinearToPitch(); 174 dst_operand.pitch = regs.pitch_out;
175 dst_operand.width = regs.line_length_in;
176 dst_operand.height = regs.line_count;
177 dst_operand.address = regs.offset_out;
178 DMA::ImageCopy copy_info{};
179 copy_info.length_x = regs.line_length_in;
180 copy_info.length_y = regs.line_count;
181 auto& accelerate = rasterizer->AccessAccelerateDMA();
182 if (accelerate.ImageToBuffer(copy_info, src_operand, dst_operand)) {
166 return; 183 return;
167 } 184 }
168 185
186 UNIMPLEMENTED_IF(regs.src_params.block_size.width != 0);
187 UNIMPLEMENTED_IF(regs.src_params.block_size.depth != 0);
188 UNIMPLEMENTED_IF(regs.src_params.block_size.depth == 0 && regs.src_params.depth != 1);
189
169 // Deswizzle the input and copy it over. 190 // Deswizzle the input and copy it over.
170 const Parameters& src_params = regs.src_params; 191 const DMA::Parameters& src_params = regs.src_params;
192
193 const bool is_remapping = regs.launch_dma.remap_enable != 0;
171 194
172 const u32 num_remap_components = regs.remap_const.num_dst_components_minus_one + 1; 195 const u32 num_remap_components = regs.remap_const.num_dst_components_minus_one + 1;
173 const u32 remap_components_size = regs.remap_const.component_size_minus_one + 1; 196 const u32 remap_components_size = regs.remap_const.component_size_minus_one + 1;
@@ -187,7 +210,7 @@ void MaxwellDMA::CopyBlockLinearToPitch() {
187 x_offset >>= bpp_shift; 210 x_offset >>= bpp_shift;
188 } 211 }
189 212
190 const u32 bytes_per_pixel = base_bpp << bpp_shift; 213 bytes_per_pixel = base_bpp << bpp_shift;
191 const u32 height = src_params.height; 214 const u32 height = src_params.height;
192 const u32 depth = src_params.depth; 215 const u32 depth = src_params.depth;
193 const u32 block_height = src_params.block_size.height; 216 const u32 block_height = src_params.block_size.height;
@@ -195,11 +218,12 @@ void MaxwellDMA::CopyBlockLinearToPitch() {
195 const size_t src_size = 218 const size_t src_size =
196 CalculateSize(true, bytes_per_pixel, width, height, depth, block_height, block_depth); 219 CalculateSize(true, bytes_per_pixel, width, height, depth, block_height, block_depth);
197 220
221 const size_t dst_size = static_cast<size_t>(regs.pitch_out) * regs.line_count;
198 read_buffer.resize_destructive(src_size); 222 read_buffer.resize_destructive(src_size);
199 write_buffer.resize_destructive(dst_size); 223 write_buffer.resize_destructive(dst_size);
200 224
201 memory_manager.ReadBlock(regs.offset_in, read_buffer.data(), src_size); 225 memory_manager.ReadBlock(src_operand.address, read_buffer.data(), src_size);
202 memory_manager.ReadBlock(regs.offset_out, write_buffer.data(), dst_size); 226 memory_manager.ReadBlockUnsafe(dst_operand.address, write_buffer.data(), dst_size);
203 227
204 UnswizzleSubrect(write_buffer, read_buffer, bytes_per_pixel, width, height, depth, x_offset, 228 UnswizzleSubrect(write_buffer, read_buffer, bytes_per_pixel, width, height, depth, x_offset,
205 src_params.origin.y, x_elements, regs.line_count, block_height, block_depth, 229 src_params.origin.y, x_elements, regs.line_count, block_height, block_depth,
@@ -216,6 +240,24 @@ void MaxwellDMA::CopyPitchToBlockLinear() {
216 const u32 num_remap_components = regs.remap_const.num_dst_components_minus_one + 1; 240 const u32 num_remap_components = regs.remap_const.num_dst_components_minus_one + 1;
217 const u32 remap_components_size = regs.remap_const.component_size_minus_one + 1; 241 const u32 remap_components_size = regs.remap_const.component_size_minus_one + 1;
218 242
243 u32 bytes_per_pixel = 1;
244 DMA::ImageOperand dst_operand;
245 dst_operand.bytes_per_pixel = bytes_per_pixel;
246 dst_operand.params = regs.dst_params;
247 dst_operand.address = regs.offset_out;
248 DMA::BufferOperand src_operand;
249 src_operand.pitch = regs.pitch_in;
250 src_operand.width = regs.line_length_in;
251 src_operand.height = regs.line_count;
252 src_operand.address = regs.offset_in;
253 DMA::ImageCopy copy_info{};
254 copy_info.length_x = regs.line_length_in;
255 copy_info.length_y = regs.line_count;
256 auto& accelerate = rasterizer->AccessAccelerateDMA();
257 if (accelerate.BufferToImage(copy_info, src_operand, dst_operand)) {
258 return;
259 }
260
219 const auto& dst_params = regs.dst_params; 261 const auto& dst_params = regs.dst_params;
220 262
221 const u32 base_bpp = !is_remapping ? 1U : num_remap_components * remap_components_size; 263 const u32 base_bpp = !is_remapping ? 1U : num_remap_components * remap_components_size;
@@ -233,7 +275,7 @@ void MaxwellDMA::CopyPitchToBlockLinear() {
233 x_offset >>= bpp_shift; 275 x_offset >>= bpp_shift;
234 } 276 }
235 277
236 const u32 bytes_per_pixel = base_bpp << bpp_shift; 278 bytes_per_pixel = base_bpp << bpp_shift;
237 const u32 height = dst_params.height; 279 const u32 height = dst_params.height;
238 const u32 depth = dst_params.depth; 280 const u32 depth = dst_params.depth;
239 const u32 block_height = dst_params.block_size.height; 281 const u32 block_height = dst_params.block_size.height;
@@ -260,45 +302,14 @@ void MaxwellDMA::CopyPitchToBlockLinear() {
260 memory_manager.WriteBlockCached(regs.offset_out, write_buffer.data(), dst_size); 302 memory_manager.WriteBlockCached(regs.offset_out, write_buffer.data(), dst_size);
261} 303}
262 304
263void MaxwellDMA::FastCopyBlockLinearToPitch() {
264 const u32 bytes_per_pixel = 1U;
265 const size_t src_size = GOB_SIZE;
266 const size_t dst_size = static_cast<size_t>(regs.pitch_out) * regs.line_count;
267 u32 pos_x = regs.src_params.origin.x;
268 u32 pos_y = regs.src_params.origin.y;
269 const u64 offset = GetGOBOffset(regs.src_params.width, regs.src_params.height, pos_x, pos_y,
270 regs.src_params.block_size.height, bytes_per_pixel);
271 const u32 x_in_gob = 64 / bytes_per_pixel;
272 pos_x = pos_x % x_in_gob;
273 pos_y = pos_y % 8;
274
275 read_buffer.resize_destructive(src_size);
276 write_buffer.resize_destructive(dst_size);
277
278 if (Settings::IsGPULevelExtreme()) {
279 memory_manager.ReadBlock(regs.offset_in + offset, read_buffer.data(), src_size);
280 memory_manager.ReadBlock(regs.offset_out, write_buffer.data(), dst_size);
281 } else {
282 memory_manager.ReadBlockUnsafe(regs.offset_in + offset, read_buffer.data(), src_size);
283 memory_manager.ReadBlockUnsafe(regs.offset_out, write_buffer.data(), dst_size);
284 }
285
286 UnswizzleSubrect(write_buffer, read_buffer, bytes_per_pixel, regs.src_params.width,
287 regs.src_params.height, 1, pos_x, pos_y, regs.line_length_in, regs.line_count,
288 regs.src_params.block_size.height, regs.src_params.block_size.depth,
289 regs.pitch_out);
290
291 memory_manager.WriteBlockCached(regs.offset_out, write_buffer.data(), dst_size);
292}
293
294void MaxwellDMA::CopyBlockLinearToBlockLinear() { 305void MaxwellDMA::CopyBlockLinearToBlockLinear() {
295 UNIMPLEMENTED_IF(regs.src_params.block_size.width != 0); 306 UNIMPLEMENTED_IF(regs.src_params.block_size.width != 0);
296 307
297 const bool is_remapping = regs.launch_dma.remap_enable != 0; 308 const bool is_remapping = regs.launch_dma.remap_enable != 0;
298 309
299 // Deswizzle the input and copy it over. 310 // Deswizzle the input and copy it over.
300 const Parameters& src = regs.src_params; 311 const DMA::Parameters& src = regs.src_params;
301 const Parameters& dst = regs.dst_params; 312 const DMA::Parameters& dst = regs.dst_params;
302 313
303 const u32 num_remap_components = regs.remap_const.num_dst_components_minus_one + 1; 314 const u32 num_remap_components = regs.remap_const.num_dst_components_minus_one + 1;
304 const u32 remap_components_size = regs.remap_const.component_size_minus_one + 1; 315 const u32 remap_components_size = regs.remap_const.component_size_minus_one + 1;
diff --git a/src/video_core/engines/maxwell_dma.h b/src/video_core/engines/maxwell_dma.h
index 0e594fa74..69e26cb32 100644
--- a/src/video_core/engines/maxwell_dma.h
+++ b/src/video_core/engines/maxwell_dma.h
@@ -24,6 +24,54 @@ namespace VideoCore {
24class RasterizerInterface; 24class RasterizerInterface;
25} 25}
26 26
27namespace Tegra {
28namespace DMA {
29
30union Origin {
31 BitField<0, 16, u32> x;
32 BitField<16, 16, u32> y;
33};
34static_assert(sizeof(Origin) == 4);
35
36struct ImageCopy {
37 u32 length_x{};
38 u32 length_y{};
39};
40
41union BlockSize {
42 BitField<0, 4, u32> width;
43 BitField<4, 4, u32> height;
44 BitField<8, 4, u32> depth;
45 BitField<12, 4, u32> gob_height;
46};
47static_assert(sizeof(BlockSize) == 4);
48
49struct Parameters {
50 BlockSize block_size;
51 u32 width;
52 u32 height;
53 u32 depth;
54 u32 layer;
55 Origin origin;
56};
57static_assert(sizeof(Parameters) == 24);
58
59struct ImageOperand {
60 u32 bytes_per_pixel;
61 Parameters params;
62 GPUVAddr address;
63};
64
65struct BufferOperand {
66 u32 pitch;
67 u32 width;
68 u32 height;
69 GPUVAddr address;
70};
71
72} // namespace DMA
73} // namespace Tegra
74
27namespace Tegra::Engines { 75namespace Tegra::Engines {
28 76
29class AccelerateDMAInterface { 77class AccelerateDMAInterface {
@@ -32,6 +80,12 @@ public:
32 virtual bool BufferCopy(GPUVAddr src_address, GPUVAddr dest_address, u64 amount) = 0; 80 virtual bool BufferCopy(GPUVAddr src_address, GPUVAddr dest_address, u64 amount) = 0;
33 81
34 virtual bool BufferClear(GPUVAddr src_address, u64 amount, u32 value) = 0; 82 virtual bool BufferClear(GPUVAddr src_address, u64 amount, u32 value) = 0;
83
84 virtual bool ImageToBuffer(const DMA::ImageCopy& copy_info, const DMA::ImageOperand& src,
85 const DMA::BufferOperand& dst) = 0;
86
87 virtual bool BufferToImage(const DMA::ImageCopy& copy_info, const DMA::BufferOperand& src,
88 const DMA::ImageOperand& dst) = 0;
35}; 89};
36 90
37/** 91/**
@@ -51,30 +105,6 @@ public:
51 } 105 }
52 }; 106 };
53 107
54 union BlockSize {
55 BitField<0, 4, u32> width;
56 BitField<4, 4, u32> height;
57 BitField<8, 4, u32> depth;
58 BitField<12, 4, u32> gob_height;
59 };
60 static_assert(sizeof(BlockSize) == 4);
61
62 union Origin {
63 BitField<0, 16, u32> x;
64 BitField<16, 16, u32> y;
65 };
66 static_assert(sizeof(Origin) == 4);
67
68 struct Parameters {
69 BlockSize block_size;
70 u32 width;
71 u32 height;
72 u32 depth;
73 u32 layer;
74 Origin origin;
75 };
76 static_assert(sizeof(Parameters) == 24);
77
78 struct Semaphore { 108 struct Semaphore {
79 PackedGPUVAddr address; 109 PackedGPUVAddr address;
80 u32 payload; 110 u32 payload;
@@ -227,8 +257,6 @@ private:
227 257
228 void CopyBlockLinearToBlockLinear(); 258 void CopyBlockLinearToBlockLinear();
229 259
230 void FastCopyBlockLinearToPitch();
231
232 void ReleaseSemaphore(); 260 void ReleaseSemaphore();
233 261
234 void ConsumeSinkImpl() override; 262 void ConsumeSinkImpl() override;
@@ -261,17 +289,17 @@ private:
261 u32 reserved05[0x3f]; 289 u32 reserved05[0x3f];
262 PackedGPUVAddr offset_in; 290 PackedGPUVAddr offset_in;
263 PackedGPUVAddr offset_out; 291 PackedGPUVAddr offset_out;
264 u32 pitch_in; 292 s32 pitch_in;
265 u32 pitch_out; 293 s32 pitch_out;
266 u32 line_length_in; 294 u32 line_length_in;
267 u32 line_count; 295 u32 line_count;
268 u32 reserved06[0xb6]; 296 u32 reserved06[0xb6];
269 u32 remap_consta_value; 297 u32 remap_consta_value;
270 u32 remap_constb_value; 298 u32 remap_constb_value;
271 RemapConst remap_const; 299 RemapConst remap_const;
272 Parameters dst_params; 300 DMA::Parameters dst_params;
273 u32 reserved07[0x1]; 301 u32 reserved07[0x1];
274 Parameters src_params; 302 DMA::Parameters src_params;
275 u32 reserved08[0x275]; 303 u32 reserved08[0x275];
276 u32 pm_trigger_end; 304 u32 pm_trigger_end;
277 u32 reserved09[0x3ba]; 305 u32 reserved09[0x3ba];
diff --git a/src/video_core/renderer_null/null_rasterizer.h b/src/video_core/renderer_null/null_rasterizer.h
index 51f896e43..0c59e6a1f 100644
--- a/src/video_core/renderer_null/null_rasterizer.h
+++ b/src/video_core/renderer_null/null_rasterizer.h
@@ -22,6 +22,14 @@ public:
22 explicit AccelerateDMA(); 22 explicit AccelerateDMA();
23 bool BufferCopy(GPUVAddr start_address, GPUVAddr end_address, u64 amount) override; 23 bool BufferCopy(GPUVAddr start_address, GPUVAddr end_address, u64 amount) override;
24 bool BufferClear(GPUVAddr src_address, u64 amount, u32 value) override; 24 bool BufferClear(GPUVAddr src_address, u64 amount, u32 value) override;
25 bool ImageToBuffer(const Tegra::DMA::ImageCopy& copy_info, const Tegra::DMA::ImageOperand& src,
26 const Tegra::DMA::BufferOperand& dst) override {
27 return false;
28 }
29 bool BufferToImage(const Tegra::DMA::ImageCopy& copy_info, const Tegra::DMA::BufferOperand& src,
30 const Tegra::DMA::ImageOperand& dst) override {
31 return false;
32 }
25}; 33};
26 34
27class RasterizerNull final : public VideoCore::RasterizerAccelerated, 35class RasterizerNull final : public VideoCore::RasterizerAccelerated,
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.h b/src/video_core/renderer_opengl/gl_rasterizer.h
index 0c45832ae..7e21fc43d 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.h
+++ b/src/video_core/renderer_opengl/gl_rasterizer.h
@@ -56,6 +56,16 @@ public:
56 56
57 bool BufferClear(GPUVAddr src_address, u64 amount, u32 value) override; 57 bool BufferClear(GPUVAddr src_address, u64 amount, u32 value) override;
58 58
59 bool ImageToBuffer(const Tegra::DMA::ImageCopy& copy_info, const Tegra::DMA::ImageOperand& src,
60 const Tegra::DMA::BufferOperand& dst) override {
61 return false;
62 }
63
64 bool BufferToImage(const Tegra::DMA::ImageCopy& copy_info, const Tegra::DMA::BufferOperand& src,
65 const Tegra::DMA::ImageOperand& dst) override {
66 return false;
67 }
68
59private: 69private:
60 BufferCache& buffer_cache; 70 BufferCache& buffer_cache;
61}; 71};
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
index 719edbcfb..f085d53a1 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
@@ -172,7 +172,7 @@ RasterizerVulkan::RasterizerVulkan(Core::Frontend::EmuWindow& emu_window_, Tegra
172 buffer_cache(*this, cpu_memory_, buffer_cache_runtime), 172 buffer_cache(*this, cpu_memory_, buffer_cache_runtime),
173 pipeline_cache(*this, device, scheduler, descriptor_pool, update_descriptor_queue, 173 pipeline_cache(*this, device, scheduler, descriptor_pool, update_descriptor_queue,
174 render_pass_cache, buffer_cache, texture_cache, gpu.ShaderNotify()), 174 render_pass_cache, buffer_cache, texture_cache, gpu.ShaderNotify()),
175 query_cache{*this, device, scheduler}, accelerate_dma{buffer_cache}, 175 query_cache{*this, device, scheduler}, accelerate_dma(buffer_cache, texture_cache, scheduler),
176 fence_manager(*this, gpu, texture_cache, buffer_cache, query_cache, device, scheduler), 176 fence_manager(*this, gpu, texture_cache, buffer_cache, query_cache, device, scheduler),
177 wfi_event(device.GetLogical().CreateEvent()) { 177 wfi_event(device.GetLogical().CreateEvent()) {
178 scheduler.SetQueryCache(query_cache); 178 scheduler.SetQueryCache(query_cache);
@@ -756,7 +756,9 @@ void RasterizerVulkan::FlushWork() {
756 draw_counter = 0; 756 draw_counter = 0;
757} 757}
758 758
759AccelerateDMA::AccelerateDMA(BufferCache& buffer_cache_) : buffer_cache{buffer_cache_} {} 759AccelerateDMA::AccelerateDMA(BufferCache& buffer_cache_, TextureCache& texture_cache_,
760 Scheduler& scheduler_)
761 : buffer_cache{buffer_cache_}, texture_cache{texture_cache_}, scheduler{scheduler_} {}
760 762
761bool AccelerateDMA::BufferClear(GPUVAddr src_address, u64 amount, u32 value) { 763bool AccelerateDMA::BufferClear(GPUVAddr src_address, u64 amount, u32 value) {
762 std::scoped_lock lock{buffer_cache.mutex}; 764 std::scoped_lock lock{buffer_cache.mutex};
@@ -768,6 +770,234 @@ bool AccelerateDMA::BufferCopy(GPUVAddr src_address, GPUVAddr dest_address, u64
768 return buffer_cache.DMACopy(src_address, dest_address, amount); 770 return buffer_cache.DMACopy(src_address, dest_address, amount);
769} 771}
770 772
773bool AccelerateDMA::ImageToBuffer(const Tegra::DMA::ImageCopy& copy_info,
774 const Tegra::DMA::ImageOperand& src,
775 const Tegra::DMA::BufferOperand& dst) {
776 std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
777 auto query_image = texture_cache.ObtainImage(src, false);
778 if (!query_image) {
779 return false;
780 }
781 auto* image = query_image->first;
782 auto [level, base] = query_image->second;
783 const u32 buffer_size = static_cast<u32>(dst.pitch * dst.height);
784 const auto [buffer, offset] = buffer_cache.ObtainBuffer(
785 dst.address, buffer_size, VideoCommon::ObtainBufferSynchronize::FullSynchronize,
786 VideoCommon::ObtainBufferOperation::MarkAsWritten);
787
788 const bool is_rescaled = image->IsRescaled();
789 if (is_rescaled) {
790 image->ScaleDown();
791 }
792 VkImageSubresourceLayers subresources{
793 .aspectMask = image->AspectMask(),
794 .mipLevel = level,
795 .baseArrayLayer = base,
796 .layerCount = 1,
797 };
798 const u32 bpp = VideoCore::Surface::BytesPerBlock(image->info.format);
799 const auto convert = [old_bpp = src.bytes_per_pixel, bpp](u32 value) {
800 return (old_bpp * value) / bpp;
801 };
802 const u32 base_x = convert(src.params.origin.x.Value());
803 const u32 base_y = src.params.origin.y.Value();
804 const u32 length_x = convert(copy_info.length_x);
805 const u32 length_y = copy_info.length_y;
806 VkOffset3D image_offset{
807 .x = static_cast<s32>(base_x),
808 .y = static_cast<s32>(base_y),
809 .z = 0,
810 };
811 VkExtent3D image_extent{
812 .width = length_x,
813 .height = length_y,
814 .depth = 1,
815 };
816 auto buff_info(dst);
817 buff_info.pitch = convert(dst.pitch);
818 scheduler.RequestOutsideRenderPassOperationContext();
819 scheduler.Record([src_image = image->Handle(), dst_buffer = buffer->Handle(),
820 buffer_offset = offset, subresources, image_offset, image_extent,
821 buff_info](vk::CommandBuffer cmdbuf) {
822 const std::array buffer_copy_info{
823 VkBufferImageCopy{
824 .bufferOffset = buffer_offset,
825 .bufferRowLength = buff_info.pitch,
826 .bufferImageHeight = buff_info.height,
827 .imageSubresource = subresources,
828 .imageOffset = image_offset,
829 .imageExtent = image_extent,
830 },
831 };
832 const VkImageSubresourceRange range{
833 .aspectMask = subresources.aspectMask,
834 .baseMipLevel = subresources.mipLevel,
835 .levelCount = 1,
836 .baseArrayLayer = subresources.baseArrayLayer,
837 .layerCount = 1,
838 };
839 static constexpr VkMemoryBarrier WRITE_BARRIER{
840 .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
841 .pNext = nullptr,
842 .srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
843 .dstAccessMask = VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT,
844 };
845 const std::array pre_barriers{
846 VkImageMemoryBarrier{
847 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
848 .pNext = nullptr,
849 .srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT | VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT |
850 VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT |
851 VK_ACCESS_TRANSFER_WRITE_BIT,
852 .dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT,
853 .oldLayout = VK_IMAGE_LAYOUT_GENERAL,
854 .newLayout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
855 .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
856 .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
857 .image = src_image,
858 .subresourceRange = range,
859 },
860 };
861 const std::array post_barriers{
862 VkImageMemoryBarrier{
863 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
864 .pNext = nullptr,
865 .srcAccessMask = 0,
866 .dstAccessMask = 0,
867 .oldLayout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
868 .newLayout = VK_IMAGE_LAYOUT_GENERAL,
869 .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
870 .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
871 .image = src_image,
872 .subresourceRange = range,
873 },
874 };
875 cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT,
876 0, {}, {}, pre_barriers);
877 cmdbuf.CopyImageToBuffer(src_image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, dst_buffer,
878 buffer_copy_info);
879 cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
880 0, WRITE_BARRIER, nullptr, post_barriers);
881 });
882 if (is_rescaled) {
883 image->ScaleUp(true);
884 }
885 return true;
886}
887
888bool AccelerateDMA::BufferToImage(const Tegra::DMA::ImageCopy& copy_info,
889 const Tegra::DMA::BufferOperand& src,
890 const Tegra::DMA::ImageOperand& dst) {
891 std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
892 auto query_image = texture_cache.ObtainImage(dst, true);
893 if (!query_image) {
894 return false;
895 }
896 auto* image = query_image->first;
897 auto [level, base] = query_image->second;
898 const u32 buffer_size = static_cast<u32>(src.pitch * src.height);
899 const auto [buffer, offset] = buffer_cache.ObtainBuffer(
900 src.address, buffer_size, VideoCommon::ObtainBufferSynchronize::FullSynchronize,
901 VideoCommon::ObtainBufferOperation::DoNothing);
902 const bool is_rescaled = image->IsRescaled();
903 if (is_rescaled) {
904 image->ScaleDown(true);
905 }
906 VkImageSubresourceLayers subresources{
907 .aspectMask = image->AspectMask(),
908 .mipLevel = level,
909 .baseArrayLayer = base,
910 .layerCount = 1,
911 };
912 const u32 bpp = VideoCore::Surface::BytesPerBlock(image->info.format);
913 const auto convert = [old_bpp = dst.bytes_per_pixel, bpp](u32 value) {
914 return (old_bpp * value) / bpp;
915 };
916 const u32 base_x = convert(dst.params.origin.x.Value());
917 const u32 base_y = dst.params.origin.y.Value();
918 const u32 length_x = convert(copy_info.length_x);
919 const u32 length_y = copy_info.length_y;
920 VkOffset3D image_offset{
921 .x = static_cast<s32>(base_x),
922 .y = static_cast<s32>(base_y),
923 .z = 0,
924 };
925 VkExtent3D image_extent{
926 .width = length_x,
927 .height = length_y,
928 .depth = 1,
929 };
930 auto buff_info(src);
931 buff_info.pitch = convert(src.pitch);
932 scheduler.RequestOutsideRenderPassOperationContext();
933 scheduler.Record([dst_image = image->Handle(), src_buffer = buffer->Handle(),
934 buffer_offset = offset, subresources, image_offset, image_extent,
935 buff_info](vk::CommandBuffer cmdbuf) {
936 const std::array buffer_copy_info{
937 VkBufferImageCopy{
938 .bufferOffset = buffer_offset,
939 .bufferRowLength = buff_info.pitch,
940 .bufferImageHeight = buff_info.height,
941 .imageSubresource = subresources,
942 .imageOffset = image_offset,
943 .imageExtent = image_extent,
944 },
945 };
946 const VkImageSubresourceRange range{
947 .aspectMask = subresources.aspectMask,
948 .baseMipLevel = subresources.mipLevel,
949 .levelCount = 1,
950 .baseArrayLayer = subresources.baseArrayLayer,
951 .layerCount = 1,
952 };
953 static constexpr VkMemoryBarrier READ_BARRIER{
954 .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
955 .pNext = nullptr,
956 .srcAccessMask = VK_ACCESS_MEMORY_WRITE_BIT,
957 .dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT | VK_ACCESS_TRANSFER_WRITE_BIT,
958 };
959 const std::array pre_barriers{
960 VkImageMemoryBarrier{
961 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
962 .pNext = nullptr,
963 .srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT | VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT |
964 VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT |
965 VK_ACCESS_TRANSFER_WRITE_BIT,
966 .dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT,
967 .oldLayout = VK_IMAGE_LAYOUT_GENERAL,
968 .newLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
969 .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
970 .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
971 .image = dst_image,
972 .subresourceRange = range,
973 },
974 };
975 const std::array post_barriers{
976 VkImageMemoryBarrier{
977 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
978 .pNext = nullptr,
979 .srcAccessMask = 0,
980 .dstAccessMask = 0,
981 .oldLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
982 .newLayout = VK_IMAGE_LAYOUT_GENERAL,
983 .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
984 .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
985 .image = dst_image,
986 .subresourceRange = range,
987 },
988 };
989 cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT,
990 0, READ_BARRIER, {}, pre_barriers);
991 cmdbuf.CopyBufferToImage(src_buffer, dst_image, VK_IMAGE_LAYOUT_GENERAL, buffer_copy_info);
992 cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
993 0, nullptr, nullptr, post_barriers);
994 });
995 if (is_rescaled) {
996 image->ScaleUp();
997 }
998 return true;
999}
1000
771void RasterizerVulkan::UpdateDynamicStates() { 1001void RasterizerVulkan::UpdateDynamicStates() {
772 auto& regs = maxwell3d->regs; 1002 auto& regs = maxwell3d->regs;
773 UpdateViewportsState(regs); 1003 UpdateViewportsState(regs);
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.h b/src/video_core/renderer_vulkan/vk_rasterizer.h
index a0508b57c..7746c5434 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.h
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.h
@@ -45,14 +45,23 @@ class StateTracker;
45 45
46class AccelerateDMA : public Tegra::Engines::AccelerateDMAInterface { 46class AccelerateDMA : public Tegra::Engines::AccelerateDMAInterface {
47public: 47public:
48 explicit AccelerateDMA(BufferCache& buffer_cache); 48 explicit AccelerateDMA(BufferCache& buffer_cache, TextureCache& texture_cache,
49 Scheduler& scheduler);
49 50
50 bool BufferCopy(GPUVAddr start_address, GPUVAddr end_address, u64 amount) override; 51 bool BufferCopy(GPUVAddr start_address, GPUVAddr end_address, u64 amount) override;
51 52
52 bool BufferClear(GPUVAddr src_address, u64 amount, u32 value) override; 53 bool BufferClear(GPUVAddr src_address, u64 amount, u32 value) override;
53 54
55 bool ImageToBuffer(const Tegra::DMA::ImageCopy& copy_info, const Tegra::DMA::ImageOperand& src,
56 const Tegra::DMA::BufferOperand& dst) override;
57
58 bool BufferToImage(const Tegra::DMA::ImageCopy& copy_info, const Tegra::DMA::BufferOperand& src,
59 const Tegra::DMA::ImageOperand& dst) override;
60
54private: 61private:
55 BufferCache& buffer_cache; 62 BufferCache& buffer_cache;
63 TextureCache& texture_cache;
64 Scheduler& scheduler;
56}; 65};
57 66
58class RasterizerVulkan final : public VideoCore::RasterizerAccelerated, 67class RasterizerVulkan final : public VideoCore::RasterizerAccelerated,
diff --git a/src/video_core/renderer_vulkan/vk_texture_cache.cpp b/src/video_core/renderer_vulkan/vk_texture_cache.cpp
index 80adb70eb..8a204f93f 100644
--- a/src/video_core/renderer_vulkan/vk_texture_cache.cpp
+++ b/src/video_core/renderer_vulkan/vk_texture_cache.cpp
@@ -864,13 +864,19 @@ void TextureCacheRuntime::ReinterpretImage(Image& dst, Image& src,
864 const VkImageAspectFlags src_aspect_mask = src.AspectMask(); 864 const VkImageAspectFlags src_aspect_mask = src.AspectMask();
865 const VkImageAspectFlags dst_aspect_mask = dst.AspectMask(); 865 const VkImageAspectFlags dst_aspect_mask = dst.AspectMask();
866 866
867 std::ranges::transform(copies, vk_in_copies.begin(), [src_aspect_mask](const auto& copy) { 867 const auto bpp_in = BytesPerBlock(src.info.format) / DefaultBlockWidth(src.info.format);
868 return MakeBufferImageCopy(copy, true, src_aspect_mask); 868 const auto bpp_out = BytesPerBlock(dst.info.format) / DefaultBlockWidth(dst.info.format);
869 }); 869 std::ranges::transform(copies, vk_in_copies.begin(),
870 [src_aspect_mask, bpp_in, bpp_out](const auto& copy) {
871 auto copy2 = copy;
872 copy2.src_offset.x = (bpp_out * copy.src_offset.x) / bpp_in;
873 copy2.extent.width = (bpp_out * copy.extent.width) / bpp_in;
874 return MakeBufferImageCopy(copy2, true, src_aspect_mask);
875 });
870 std::ranges::transform(copies, vk_out_copies.begin(), [dst_aspect_mask](const auto& copy) { 876 std::ranges::transform(copies, vk_out_copies.begin(), [dst_aspect_mask](const auto& copy) {
871 return MakeBufferImageCopy(copy, false, dst_aspect_mask); 877 return MakeBufferImageCopy(copy, false, dst_aspect_mask);
872 }); 878 });
873 const u32 img_bpp = BytesPerBlock(src.info.format); 879 const u32 img_bpp = BytesPerBlock(dst.info.format);
874 size_t total_size = 0; 880 size_t total_size = 0;
875 for (const auto& copy : copies) { 881 for (const auto& copy : copies) {
876 total_size += copy.extent.width * copy.extent.height * copy.extent.depth * img_bpp; 882 total_size += copy.extent.width * copy.extent.height * copy.extent.depth * img_bpp;
diff --git a/src/video_core/texture_cache/image_info.cpp b/src/video_core/texture_cache/image_info.cpp
index e9100091e..a1296b574 100644
--- a/src/video_core/texture_cache/image_info.cpp
+++ b/src/video_core/texture_cache/image_info.cpp
@@ -216,10 +216,51 @@ ImageInfo::ImageInfo(const Tegra::Engines::Fermi2D::Surface& config) noexcept {
216 .height = config.height, 216 .height = config.height,
217 .depth = 1, 217 .depth = 1,
218 }; 218 };
219 rescaleable = block.depth == 0; 219 rescaleable = block.depth == 0 && size.height > 256;
220 rescaleable &= size.height > 256;
221 downscaleable = size.height > 512; 220 downscaleable = size.height > 512;
222 } 221 }
223} 222}
224 223
224static PixelFormat ByteSizeToFormat(u32 bytes_per_pixel) {
225 switch (bytes_per_pixel) {
226 case 1:
227 return PixelFormat::R8_UINT;
228 case 2:
229 return PixelFormat::R8G8_UINT;
230 case 4:
231 return PixelFormat::A8B8G8R8_UINT;
232 case 8:
233 return PixelFormat::R16G16B16A16_UINT;
234 case 16:
235 return PixelFormat::R32G32B32A32_UINT;
236 default:
237 UNIMPLEMENTED();
238 return PixelFormat::Invalid;
239 }
240}
241
242ImageInfo::ImageInfo(const Tegra::DMA::ImageOperand& config) noexcept {
243 const u32 bytes_per_pixel = config.bytes_per_pixel;
244 format = ByteSizeToFormat(bytes_per_pixel);
245 type = config.params.block_size.depth > 0 ? ImageType::e3D : ImageType::e2D;
246 num_samples = 1;
247 block = Extent3D{
248 .width = config.params.block_size.width,
249 .height = config.params.block_size.height,
250 .depth = config.params.block_size.depth,
251 };
252 size = Extent3D{
253 .width = config.params.width,
254 .height = config.params.height,
255 .depth = config.params.depth,
256 };
257 tile_width_spacing = 0;
258 resources.levels = 1;
259 resources.layers = 1;
260 layer_stride = CalculateLayerStride(*this);
261 maybe_unaligned_layer_stride = CalculateLayerSize(*this);
262 rescaleable = block.depth == 0 && size.height > 256;
263 downscaleable = size.height > 512;
264}
265
225} // namespace VideoCommon 266} // namespace VideoCommon
diff --git a/src/video_core/texture_cache/image_info.h b/src/video_core/texture_cache/image_info.h
index 93755e15e..a12f5b44f 100644
--- a/src/video_core/texture_cache/image_info.h
+++ b/src/video_core/texture_cache/image_info.h
@@ -5,6 +5,7 @@
5 5
6#include "video_core/engines/fermi_2d.h" 6#include "video_core/engines/fermi_2d.h"
7#include "video_core/engines/maxwell_3d.h" 7#include "video_core/engines/maxwell_3d.h"
8#include "video_core/engines/maxwell_dma.h"
8#include "video_core/surface.h" 9#include "video_core/surface.h"
9#include "video_core/texture_cache/types.h" 10#include "video_core/texture_cache/types.h"
10 11
@@ -19,6 +20,7 @@ struct ImageInfo {
19 explicit ImageInfo(const Tegra::Engines::Maxwell3D::Regs& regs, size_t index) noexcept; 20 explicit ImageInfo(const Tegra::Engines::Maxwell3D::Regs& regs, size_t index) noexcept;
20 explicit ImageInfo(const Tegra::Engines::Maxwell3D::Regs& regs) noexcept; 21 explicit ImageInfo(const Tegra::Engines::Maxwell3D::Regs& regs) noexcept;
21 explicit ImageInfo(const Tegra::Engines::Fermi2D::Surface& config) noexcept; 22 explicit ImageInfo(const Tegra::Engines::Fermi2D::Surface& config) noexcept;
23 explicit ImageInfo(const Tegra::DMA::ImageOperand& config) noexcept;
22 24
23 PixelFormat format = PixelFormat::Invalid; 25 PixelFormat format = PixelFormat::Invalid;
24 ImageType type = ImageType::e1D; 26 ImageType type = ImageType::e1D;
diff --git a/src/video_core/texture_cache/texture_cache.h b/src/video_core/texture_cache/texture_cache.h
index 9dd152fbe..335338434 100644
--- a/src/video_core/texture_cache/texture_cache.h
+++ b/src/video_core/texture_cache/texture_cache.h
@@ -1359,6 +1359,75 @@ std::optional<typename TextureCache<P>::BlitImages> TextureCache<P>::GetBlitImag
1359} 1359}
1360 1360
1361template <class P> 1361template <class P>
1362ImageId TextureCache<P>::FindDMAImage(const ImageInfo& info, GPUVAddr gpu_addr) {
1363 std::optional<VAddr> cpu_addr = gpu_memory->GpuToCpuAddress(gpu_addr);
1364 if (!cpu_addr) {
1365 cpu_addr = gpu_memory->GpuToCpuAddress(gpu_addr, CalculateGuestSizeInBytes(info));
1366 if (!cpu_addr) {
1367 return ImageId{};
1368 }
1369 }
1370 ImageId image_id{};
1371 boost::container::small_vector<ImageId, 1> image_ids;
1372 const auto lambda = [&](ImageId existing_image_id, ImageBase& existing_image) {
1373 if (True(existing_image.flags & ImageFlagBits::Remapped)) {
1374 return false;
1375 }
1376 if (info.type == ImageType::Linear || existing_image.info.type == ImageType::Linear)
1377 [[unlikely]] {
1378 const bool strict_size = True(existing_image.flags & ImageFlagBits::Strong);
1379 const ImageInfo& existing = existing_image.info;
1380 if (existing_image.gpu_addr == gpu_addr && existing.type == info.type &&
1381 existing.pitch == info.pitch &&
1382 IsPitchLinearSameSize(existing, info, strict_size) &&
1383 IsViewCompatible(existing.format, info.format, false, true)) {
1384 image_id = existing_image_id;
1385 image_ids.push_back(existing_image_id);
1386 return true;
1387 }
1388 } else if (IsSubCopy(info, existing_image, gpu_addr)) {
1389 image_id = existing_image_id;
1390 image_ids.push_back(existing_image_id);
1391 return true;
1392 }
1393 return false;
1394 };
1395 ForEachImageInRegion(*cpu_addr, CalculateGuestSizeInBytes(info), lambda);
1396 if (image_ids.size() <= 1) [[likely]] {
1397 return image_id;
1398 }
1399 auto image_ids_compare = [this](ImageId a, ImageId b) {
1400 auto& image_a = slot_images[a];
1401 auto& image_b = slot_images[b];
1402 return image_a.modification_tick < image_b.modification_tick;
1403 };
1404 return *std::ranges::max_element(image_ids, image_ids_compare);
1405}
1406
1407template <class P>
1408std::optional<std::pair<typename TextureCache<P>::Image*, std::pair<u32, u32>>>
1409TextureCache<P>::ObtainImage(const Tegra::DMA::ImageOperand& operand, bool mark_as_modified) {
1410 ImageInfo dst_info(operand);
1411 ImageId dst_id = FindDMAImage(dst_info, operand.address);
1412 if (!dst_id) {
1413 return std::nullopt;
1414 }
1415 auto& image = slot_images[dst_id];
1416 auto base = image.TryFindBase(operand.address);
1417 if (!base) {
1418 return std::nullopt;
1419 }
1420 if (False(image.flags & ImageFlagBits::GpuModified)) {
1421 // No need to waste time on an image that's synced with guest
1422 return std::nullopt;
1423 }
1424 PrepareImage(dst_id, mark_as_modified, false);
1425 auto& new_image = slot_images[dst_id];
1426 lru_cache.Touch(new_image.lru_index, frame_tick);
1427 return std::make_pair(&new_image, std::make_pair(base->level, base->layer));
1428}
1429
1430template <class P>
1362SamplerId TextureCache<P>::FindSampler(const TSCEntry& config) { 1431SamplerId TextureCache<P>::FindSampler(const TSCEntry& config) {
1363 if (std::ranges::all_of(config.raw, [](u64 value) { return value == 0; })) { 1432 if (std::ranges::all_of(config.raw, [](u64 value) { return value == 0; })) {
1364 return NULL_SAMPLER_ID; 1433 return NULL_SAMPLER_ID;
diff --git a/src/video_core/texture_cache/texture_cache_base.h b/src/video_core/texture_cache/texture_cache_base.h
index 013836933..848a5d9ea 100644
--- a/src/video_core/texture_cache/texture_cache_base.h
+++ b/src/video_core/texture_cache/texture_cache_base.h
@@ -209,6 +209,9 @@ public:
209 /// Pop asynchronous downloads 209 /// Pop asynchronous downloads
210 void PopAsyncFlushes(); 210 void PopAsyncFlushes();
211 211
212 [[nodiscard]] std::optional<std::pair<Image*, std::pair<u32, u32>>> ObtainImage(
213 const Tegra::DMA::ImageOperand& operand, bool mark_as_modified);
214
212 /// Return true when a CPU region is modified from the GPU 215 /// Return true when a CPU region is modified from the GPU
213 [[nodiscard]] bool IsRegionGpuModified(VAddr addr, size_t size); 216 [[nodiscard]] bool IsRegionGpuModified(VAddr addr, size_t size);
214 217
@@ -300,6 +303,8 @@ private:
300 /// Remove joined images from the cache 303 /// Remove joined images from the cache
301 [[nodiscard]] ImageId JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VAddr cpu_addr); 304 [[nodiscard]] ImageId JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VAddr cpu_addr);
302 305
306 [[nodiscard]] ImageId FindDMAImage(const ImageInfo& info, GPUVAddr gpu_addr);
307
303 /// Return a blit image pair from the given guest blit parameters 308 /// Return a blit image pair from the given guest blit parameters
304 [[nodiscard]] std::optional<BlitImages> GetBlitImages( 309 [[nodiscard]] std::optional<BlitImages> GetBlitImages(
305 const Tegra::Engines::Fermi2D::Surface& dst, const Tegra::Engines::Fermi2D::Surface& src, 310 const Tegra::Engines::Fermi2D::Surface& dst, const Tegra::Engines::Fermi2D::Surface& src,
diff --git a/src/video_core/texture_cache/types.h b/src/video_core/texture_cache/types.h
index 0453456b4..a0e10643f 100644
--- a/src/video_core/texture_cache/types.h
+++ b/src/video_core/texture_cache/types.h
@@ -54,6 +54,7 @@ enum class RelaxedOptions : u32 {
54 Format = 1 << 1, 54 Format = 1 << 1,
55 Samples = 1 << 2, 55 Samples = 1 << 2,
56 ForceBrokenViews = 1 << 3, 56 ForceBrokenViews = 1 << 3,
57 FormatBpp = 1 << 4,
57}; 58};
58DECLARE_ENUM_FLAG_OPERATORS(RelaxedOptions) 59DECLARE_ENUM_FLAG_OPERATORS(RelaxedOptions)
59 60
diff --git a/src/video_core/texture_cache/util.cpp b/src/video_core/texture_cache/util.cpp
index 697f86641..de37db684 100644
--- a/src/video_core/texture_cache/util.cpp
+++ b/src/video_core/texture_cache/util.cpp
@@ -743,6 +743,44 @@ std::vector<ImageCopy> MakeShrinkImageCopies(const ImageInfo& dst, const ImageIn
743 return copies; 743 return copies;
744} 744}
745 745
746std::vector<ImageCopy> MakeReinterpretImageCopies(const ImageInfo& src, u32 up_scale,
747 u32 down_shift) {
748 std::vector<ImageCopy> copies;
749 copies.reserve(src.resources.levels);
750 const bool is_3d = src.type == ImageType::e3D;
751 for (s32 level = 0; level < src.resources.levels; ++level) {
752 ImageCopy& copy = copies.emplace_back();
753 copy.src_subresource = SubresourceLayers{
754 .base_level = level,
755 .base_layer = 0,
756 .num_layers = src.resources.layers,
757 };
758 copy.dst_subresource = SubresourceLayers{
759 .base_level = level,
760 .base_layer = 0,
761 .num_layers = src.resources.layers,
762 };
763 copy.src_offset = Offset3D{
764 .x = 0,
765 .y = 0,
766 .z = 0,
767 };
768 copy.dst_offset = Offset3D{
769 .x = 0,
770 .y = 0,
771 .z = 0,
772 };
773 const Extent3D mip_size = AdjustMipSize(src.size, level);
774 copy.extent = AdjustSamplesSize(mip_size, src.num_samples);
775 if (is_3d) {
776 copy.extent.depth = src.size.depth;
777 }
778 copy.extent.width = std::max<u32>((copy.extent.width * up_scale) >> down_shift, 1);
779 copy.extent.height = std::max<u32>((copy.extent.height * up_scale) >> down_shift, 1);
780 }
781 return copies;
782}
783
746bool IsValidEntry(const Tegra::MemoryManager& gpu_memory, const TICEntry& config) { 784bool IsValidEntry(const Tegra::MemoryManager& gpu_memory, const TICEntry& config) {
747 const GPUVAddr address = config.Address(); 785 const GPUVAddr address = config.Address();
748 if (address == 0) { 786 if (address == 0) {
@@ -999,6 +1037,20 @@ bool IsBlockLinearSizeCompatible(const ImageInfo& lhs, const ImageInfo& rhs, u32
999 } 1037 }
1000} 1038}
1001 1039
1040bool IsBlockLinearSizeCompatibleBPPRelaxed(const ImageInfo& lhs, const ImageInfo& rhs,
1041 u32 lhs_level, u32 rhs_level) noexcept {
1042 ASSERT(lhs.type != ImageType::Linear);
1043 ASSERT(rhs.type != ImageType::Linear);
1044 const auto lhs_bpp = BytesPerBlock(lhs.format);
1045 const auto rhs_bpp = BytesPerBlock(rhs.format);
1046 const Extent3D lhs_size = AdjustMipSize(lhs.size, lhs_level);
1047 const Extent3D rhs_size = AdjustMipSize(rhs.size, rhs_level);
1048 return Common::AlignUpLog2(lhs_size.width * lhs_bpp, GOB_SIZE_X_SHIFT) ==
1049 Common::AlignUpLog2(rhs_size.width * rhs_bpp, GOB_SIZE_X_SHIFT) &&
1050 Common::AlignUpLog2(lhs_size.height, GOB_SIZE_Y_SHIFT) ==
1051 Common::AlignUpLog2(rhs_size.height, GOB_SIZE_Y_SHIFT);
1052}
1053
1002bool IsPitchLinearSameSize(const ImageInfo& lhs, const ImageInfo& rhs, bool strict_size) noexcept { 1054bool IsPitchLinearSameSize(const ImageInfo& lhs, const ImageInfo& rhs, bool strict_size) noexcept {
1003 ASSERT(lhs.type == ImageType::Linear); 1055 ASSERT(lhs.type == ImageType::Linear);
1004 ASSERT(rhs.type == ImageType::Linear); 1056 ASSERT(rhs.type == ImageType::Linear);
@@ -1073,7 +1125,8 @@ std::optional<SubresourceBase> FindSubresource(const ImageInfo& candidate, const
1073 // Format checking is relaxed, but we still have to check for matching bytes per block. 1125 // Format checking is relaxed, but we still have to check for matching bytes per block.
1074 // This avoids creating a view for blits on UE4 titles where formats with different bytes 1126 // This avoids creating a view for blits on UE4 titles where formats with different bytes
1075 // per block are aliased. 1127 // per block are aliased.
1076 if (BytesPerBlock(existing.format) != BytesPerBlock(candidate.format)) { 1128 if (BytesPerBlock(existing.format) != BytesPerBlock(candidate.format) &&
1129 False(options & RelaxedOptions::FormatBpp)) {
1077 return std::nullopt; 1130 return std::nullopt;
1078 } 1131 }
1079 } else { 1132 } else {
@@ -1088,10 +1141,8 @@ std::optional<SubresourceBase> FindSubresource(const ImageInfo& candidate, const
1088 if (existing.type != candidate.type) { 1141 if (existing.type != candidate.type) {
1089 return std::nullopt; 1142 return std::nullopt;
1090 } 1143 }
1091 if (False(options & RelaxedOptions::Samples)) { 1144 if (False(options & RelaxedOptions::Samples) && existing.num_samples != candidate.num_samples) {
1092 if (existing.num_samples != candidate.num_samples) { 1145 return std::nullopt;
1093 return std::nullopt;
1094 }
1095 } 1146 }
1096 if (existing.resources.levels < candidate.resources.levels + base->level) { 1147 if (existing.resources.levels < candidate.resources.levels + base->level) {
1097 return std::nullopt; 1148 return std::nullopt;
@@ -1101,14 +1152,16 @@ std::optional<SubresourceBase> FindSubresource(const ImageInfo& candidate, const
1101 if (mip_depth < candidate.size.depth + base->layer) { 1152 if (mip_depth < candidate.size.depth + base->layer) {
1102 return std::nullopt; 1153 return std::nullopt;
1103 } 1154 }
1104 } else { 1155 } else if (existing.resources.layers < candidate.resources.layers + base->layer) {
1105 if (existing.resources.layers < candidate.resources.layers + base->layer) { 1156 return std::nullopt;
1106 return std::nullopt;
1107 }
1108 } 1157 }
1109 const bool strict_size = False(options & RelaxedOptions::Size); 1158 const bool strict_size = False(options & RelaxedOptions::Size);
1110 if (!IsBlockLinearSizeCompatible(existing, candidate, base->level, 0, strict_size)) { 1159 if (!IsBlockLinearSizeCompatible(existing, candidate, base->level, 0, strict_size)) {
1111 return std::nullopt; 1160 if (False(options & RelaxedOptions::FormatBpp)) {
1161 return std::nullopt;
1162 } else if (!IsBlockLinearSizeCompatibleBPPRelaxed(existing, candidate, base->level, 0)) {
1163 return std::nullopt;
1164 }
1112 } 1165 }
1113 // TODO: compare block sizes 1166 // TODO: compare block sizes
1114 return base; 1167 return base;
@@ -1120,6 +1173,31 @@ bool IsSubresource(const ImageInfo& candidate, const ImageBase& image, GPUVAddr
1120 .has_value(); 1173 .has_value();
1121} 1174}
1122 1175
1176bool IsSubCopy(const ImageInfo& candidate, const ImageBase& image, GPUVAddr candidate_addr) {
1177 const std::optional<SubresourceBase> base = image.TryFindBase(candidate_addr);
1178 if (!base) {
1179 return false;
1180 }
1181 const ImageInfo& existing = image.info;
1182 if (existing.resources.levels < candidate.resources.levels + base->level) {
1183 return false;
1184 }
1185 if (existing.type == ImageType::e3D) {
1186 const u32 mip_depth = std::max(1U, existing.size.depth << base->level);
1187 if (mip_depth < candidate.size.depth + base->layer) {
1188 return false;
1189 }
1190 } else {
1191 if (existing.resources.layers < candidate.resources.layers + base->layer) {
1192 return false;
1193 }
1194 }
1195 if (!IsBlockLinearSizeCompatibleBPPRelaxed(existing, candidate, base->level, 0)) {
1196 return false;
1197 }
1198 return true;
1199}
1200
1123void DeduceBlitImages(ImageInfo& dst_info, ImageInfo& src_info, const ImageBase* dst, 1201void DeduceBlitImages(ImageInfo& dst_info, ImageInfo& src_info, const ImageBase* dst,
1124 const ImageBase* src) { 1202 const ImageBase* src) {
1125 const auto original_dst_format = dst_info.format; 1203 const auto original_dst_format = dst_info.format;
diff --git a/src/video_core/texture_cache/util.h b/src/video_core/texture_cache/util.h
index d103db8ae..84aa6880d 100644
--- a/src/video_core/texture_cache/util.h
+++ b/src/video_core/texture_cache/util.h
@@ -56,6 +56,10 @@ struct OverlapResult {
56 SubresourceBase base, u32 up_scale = 1, 56 SubresourceBase base, u32 up_scale = 1,
57 u32 down_shift = 0); 57 u32 down_shift = 0);
58 58
59[[nodiscard]] std::vector<ImageCopy> MakeReinterpretImageCopies(const ImageInfo& src,
60 u32 up_scale = 1,
61 u32 down_shift = 0);
62
59[[nodiscard]] bool IsValidEntry(const Tegra::MemoryManager& gpu_memory, const TICEntry& config); 63[[nodiscard]] bool IsValidEntry(const Tegra::MemoryManager& gpu_memory, const TICEntry& config);
60 64
61[[nodiscard]] std::vector<BufferImageCopy> UnswizzleImage(Tegra::MemoryManager& gpu_memory, 65[[nodiscard]] std::vector<BufferImageCopy> UnswizzleImage(Tegra::MemoryManager& gpu_memory,
@@ -88,6 +92,9 @@ void SwizzleImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr, const Ima
88[[nodiscard]] bool IsPitchLinearSameSize(const ImageInfo& lhs, const ImageInfo& rhs, 92[[nodiscard]] bool IsPitchLinearSameSize(const ImageInfo& lhs, const ImageInfo& rhs,
89 bool strict_size) noexcept; 93 bool strict_size) noexcept;
90 94
95[[nodiscard]] bool IsBlockLinearSizeCompatibleBPPRelaxed(const ImageInfo& lhs, const ImageInfo& rhs,
96 u32 lhs_level, u32 rhs_level) noexcept;
97
91[[nodiscard]] std::optional<OverlapResult> ResolveOverlap(const ImageInfo& new_info, 98[[nodiscard]] std::optional<OverlapResult> ResolveOverlap(const ImageInfo& new_info,
92 GPUVAddr gpu_addr, VAddr cpu_addr, 99 GPUVAddr gpu_addr, VAddr cpu_addr,
93 const ImageBase& overlap, 100 const ImageBase& overlap,
@@ -106,6 +113,9 @@ void SwizzleImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr, const Ima
106 GPUVAddr candidate_addr, RelaxedOptions options, bool broken_views, 113 GPUVAddr candidate_addr, RelaxedOptions options, bool broken_views,
107 bool native_bgr); 114 bool native_bgr);
108 115
116[[nodiscard]] bool IsSubCopy(const ImageInfo& candidate, const ImageBase& image,
117 GPUVAddr candidate_addr);
118
109void DeduceBlitImages(ImageInfo& dst_info, ImageInfo& src_info, const ImageBase* dst, 119void DeduceBlitImages(ImageInfo& dst_info, ImageInfo& src_info, const ImageBase* dst,
110 const ImageBase* src); 120 const ImageBase* src);
111 121
HIC SVNT DRACONES