Merge pull request #225 from shadps4-emu/stabilization/10

Various fixes and improvements

src/core/libraries/gnmdriver/gnmdriver.cpp

@@ -8,6 +8,7 @@
 #include "common/slot_vector.h"
 #include "core/libraries/error_codes.h"
 #include "core/libraries/gnmdriver/gnmdriver.h"
+#include "core/libraries/kernel/libkernel.h"
 #include "core/libraries/libs.h"
 #include "core/libraries/videoout/video_out.h"
 #include "core/platform.h"
@@ -212,12 +213,61 @@ static constexpr std::array InitSequence350{
     0xc0016900u, 0x2aau, 0xffu,
 };
 static_assert(InitSequence350.size() == 0x7c);
+
+static constexpr std::array CtxInitSequence{
+    0xc0012800u, 0x80000000u, 0x80000000u,
+    0xc0001200u, 0u,
+    0xc0002f00u, 1u,
+    0xc0016900u, 0x102u, 0u,
+    0xc0016900u, 0x202u, 0xcc0010u,
+    0xc0111000u, 0u
+};
+static_assert(CtxInitSequence.size() == 0x0f);
+
+static constexpr std::array CtxInitSequence400{
+    0xc0012800u, 0x80000000u, 0x80000000u,
+    0xc0001200u, 0u,
+    0xc0016900u, 0x2f9u, 0x2du,
+    0xc0016900u, 0x282u, 8u,
+    0xc0016900u, 0x280u, 0x80008u,
+    0xc0016900u, 0x281u, 0xffff0000u,
+    0xc0016900u, 0x204u, 0u,
+    0xc0016900u, 0x206u, 0x43fu,
+    0xc0016900u, 0x83u,  0xffffu,
+    0xc0016900u, 0x317u, 0x10u,
+    0xc0016900u, 0x2fau, 0x3f800000u,
+    0xc0016900u, 0x2fcu, 0x3f800000u,
+    0xc0016900u, 0x2fbu, 0x3f800000u,
+    0xc0016900u, 0x2fdu, 0x3f800000u,
+    0xc0016900u, 0x202u, 0xcc0010u,
+    0xc0016900u, 0x30eu, 0xffffffffu,
+    0xc0016900u, 0x30fu, 0xffffffffu,
+    0xc0002f00u, 1u,
+    0xc0016900u, 0x1b1u, 2u,
+    0xc0016900u, 0x101u, 0u,
+    0xc0016900u, 0x100u, 0xffffffffu,
+    0xc0016900u, 0x103u, 0u,
+    0xc0016900u, 0x284u, 0u,
+    0xc0016900u, 0x290u, 0u,
+    0xc0016900u, 0x2aeu, 0u,
+    0xc0016900u, 0x102u, 0u,
+    0xc0016900u, 0x292u, 0u,
+    0xc0016900u, 0x293u, 0x6020000u,
+    0xc0016900u, 0x2f8u, 0u,
+    0xc0016900u, 0x2deu, 0x1e9u,
+    0xc0036900u, 0x295u, 0x100u, 0x100u, 4u,
+    0xc0016900u, 0x2aau, 0xffu,
+    0xc09e1000u,
+};
+static_assert(CtxInitSequence400.size() == 0x61);
 // clang-format on
 
 // In case if `submitDone` is issued we need to block submissions until GPU idle
 static u32 submission_lock{};
 static std::mutex m_submission{};
-static u64 frames_submitted{}; // frame counter
+static u64 frames_submitted{};      // frame counter
+static bool send_init_packet{true}; // initialize HW state before first game's submit in a frame
+static int sdk_version{0};
 
 struct AscQueueInfo {
     VAddr map_addr;
@@ -664,9 +714,10 @@ u32 PS4_SYSV_ABI sceGnmDrawInitDefaultHardwareState175(u32* cmdbuf, u32 size) {
 
     cmdbuf = ClearContextState(cmdbuf);
     std::memcpy(cmdbuf, InitSequence175.data(), InitSequence175.size() * 4);
+    cmdbuf += InitSequence175.size();
 
-    cmdbuf[0x7f] = 0xc07f1000;
-    cmdbuf[0x80] = 0;
+    constexpr auto cmdbuf_left = HwInitPacketSize - InitSequence175.size() - 0xc - 1;
+    WriteTrailingNop<cmdbuf_left>(cmdbuf);
 
     return HwInitPacketSize;
 }
@@ -723,14 +774,28 @@ u32 PS4_SYSV_ABI sceGnmDrawInitDefaultHardwareState350(u32* cmdbuf, u32 size) {
     return SetupContext350(cmdbuf, size, true);
 }
 
-int PS4_SYSV_ABI sceGnmDrawInitToDefaultContextState() {
-    LOG_ERROR(Lib_GnmDriver, "(STUBBED) called");
-    return ORBIS_OK;
+u32 PS4_SYSV_ABI sceGnmDrawInitToDefaultContextState(u32* cmdbuf, u32 size) {
+    LOG_TRACE(Lib_GnmDriver, "called");
+
+    constexpr auto CtxInitPacketSize = 0x20u;
+    if (size != CtxInitPacketSize) {
+        return 0;
+    }
+
+    std::memcpy(cmdbuf, CtxInitSequence.data(), CtxInitSequence.size() * 4);
+    return CtxInitPacketSize;
 }
 
-int PS4_SYSV_ABI sceGnmDrawInitToDefaultContextState400() {
-    LOG_ERROR(Lib_GnmDriver, "(STUBBED) called");
-    return ORBIS_OK;
+u32 PS4_SYSV_ABI sceGnmDrawInitToDefaultContextState400(u32* cmdbuf, u32 size) {
+    LOG_TRACE(Lib_GnmDriver, "called");
+
+    constexpr auto CtxInitPacketSize = 0x100u;
+    if (size != CtxInitPacketSize) {
+        return 0;
+    }
+
+    std::memcpy(cmdbuf, CtxInitSequence400.data(), CtxInitSequence400.size() * 4);
+    return CtxInitPacketSize;
 }
 
 int PS4_SYSV_ABI sceGnmDrawOpaqueAuto() {
@@ -1873,6 +1938,17 @@ s32 PS4_SYSV_ABI sceGnmSubmitCommandBuffers(u32 count, const u32* dcb_gpu_addrs[
         submission_lock = 0;
     }
 
+    if (send_init_packet) {
+        if (sdk_version <= 0x1ffffffu) {
+            liverpool->SubmitGfx(InitSequence, {});
+        } else if (sdk_version <= 0x3ffffffu) {
+            liverpool->SubmitGfx(InitSequence200, {});
+        } else {
+            liverpool->SubmitGfx(InitSequence350, {});
+        }
+        send_init_packet = false;
+    }
+
     for (auto cbpair = 0u; cbpair < count; ++cbpair) {
         const auto* ccb = ccb_gpu_addrs ? ccb_gpu_addrs[cbpair] : nullptr;
         const auto ccb_size_in_bytes = ccb_sizes_in_bytes ? ccb_sizes_in_bytes[cbpair] : 0;
@@ -1915,6 +1991,7 @@ int PS4_SYSV_ABI sceGnmSubmitDone() {
         submission_lock = true;
     }
     liverpool->NotifySubmitDone();
+    send_init_packet = true;
     ++frames_submitted;
     return ORBIS_OK;
 }
@@ -2388,6 +2465,11 @@ void RegisterlibSceGnmDriver(Core::Loader::SymbolsResolver* sym) {
     liverpool = std::make_unique<AmdGpu::Liverpool>();
     renderer = std::make_unique<Vulkan::RendererVulkan>(*g_window, liverpool.get());
 
+    const int result = sceKernelGetCompiledSdkVersion(&sdk_version);
+    if (result != ORBIS_OK) {
+        sdk_version = 0;
+    }
+
     LIB_FUNCTION("b0xyllnVY-I", "libSceGnmDriver", 1, "libSceGnmDriver", 1, 1, sceGnmAddEqEvent);
     LIB_FUNCTION("b08AgtPlHPg", "libSceGnmDriver", 1, "libSceGnmDriver", 1, 1,
                  sceGnmAreSubmitsAllowed);

src/core/libraries/gnmdriver/gnmdriver.h

@@ -60,8 +60,8 @@ u32 PS4_SYSV_ABI sceGnmDrawInitDefaultHardwareState(u32* cmdbuf, u32 size);
 u32 PS4_SYSV_ABI sceGnmDrawInitDefaultHardwareState175(u32* cmdbuf, u32 size);
 u32 PS4_SYSV_ABI sceGnmDrawInitDefaultHardwareState200(u32* cmdbuf, u32 size);
 u32 PS4_SYSV_ABI sceGnmDrawInitDefaultHardwareState350(u32* cmdbuf, u32 size);
-int PS4_SYSV_ABI sceGnmDrawInitToDefaultContextState();
-int PS4_SYSV_ABI sceGnmDrawInitToDefaultContextState400();
+u32 PS4_SYSV_ABI sceGnmDrawInitToDefaultContextState(u32* cmdbuf, u32 size);
+u32 PS4_SYSV_ABI sceGnmDrawInitToDefaultContextState400(u32* cmdbuf, u32 size);
 int PS4_SYSV_ABI sceGnmDrawOpaqueAuto();
 int PS4_SYSV_ABI sceGnmDriverCaptureInProgress();
 int PS4_SYSV_ABI sceGnmDriverInternalRetrieveGnmInterface();

src/core/libraries/kernel/libkernel.cpp

@@ -161,7 +161,7 @@ int PS4_SYSV_ABI sceKernelGetCompiledSdkVersion(int* ver) {
     int version = param_sfo->GetInteger("SYSTEM_VER");
     LOG_INFO(Kernel, "returned system version = {:#x}", version);
     *ver = version;
-    return ORBIS_OK;
+    return (version > 0) ? ORBIS_OK : ORBIS_KERNEL_ERROR_EINVAL;
 }
 
 s64 PS4_SYSV_ABI ps4__read(int d, void* buf, u64 nbytes) {

src/core/libraries/kernel/libkernel.h

@@ -30,6 +30,7 @@ typedef struct {
 } OrbisKernelUuid;
 
 int* PS4_SYSV_ABI __Error();
+int PS4_SYSV_ABI sceKernelGetCompiledSdkVersion(int* ver);
 
 void LibKernel_Register(Core::Loader::SymbolsResolver* sym);