gnmdriver: added support for gpu context reset

CMakeLists.txt

@@ -472,6 +472,7 @@ set(VIDEO_CORE src/video_core/amdgpu/liverpool.cpp
                src/video_core/amdgpu/pm4_cmds.h
                src/video_core/amdgpu/pm4_opcodes.h
                src/video_core/amdgpu/resource.h
+               src/video_core/amdgpu/default_context.cpp
                src/video_core/buffer_cache/buffer.cpp
                src/video_core/buffer_cache/buffer.h
                src/video_core/buffer_cache/buffer_cache.cpp

src/core/libraries/gnmdriver/gnmdriver.cpp

@@ -55,6 +55,10 @@ static constexpr auto HwInitPacketSize = 0x100u;
 
 // clang-format off
 static constexpr std::array InitSequence{
+    // A fake preamble to mimic context reset sent by FW
+    0xc0001200u, 0u, // IT_CLEAR_STATE
+
+    // Actual init state sequence
     0xc0017600u, 0x216u, 0xffffffffu,
     0xc0017600u, 0x217u, 0xffffffffu,
     0xc0017600u, 0x215u, 0u,
@@ -94,9 +98,13 @@ static constexpr std::array InitSequence{
     0xc0036900u, 0x295u, 0x100u, 0x100u, 4u,
     0xc0017900u, 0x200u, 0xe0000000u,
 };
-static_assert(InitSequence.size() == 0x73);
+static_assert(InitSequence.size() == 0x73 + 2);
 
 static constexpr std::array InitSequence175{
+    // A fake preamble to mimic context reset sent by FW
+    0xc0001200u, 0u, // IT_CLEAR_STATE
+
+    // Actual init state sequence 
     0xc0017600u, 0x216u, 0xffffffffu,
     0xc0017600u, 0x217u, 0xffffffffu,
     0xc0017600u, 0x215u, 0u,
@@ -136,9 +144,13 @@ static constexpr std::array InitSequence175{
     0xc0036900u, 0x295u, 0x100u, 0x100u, 4u,
     0xc0017900u, 0x200u, 0xe0000000u,
 };
-static_assert(InitSequence175.size() == 0x73);
+static_assert(InitSequence175.size() == 0x73 + 2);
 
 static constexpr std::array InitSequence200{
+    // A fake preamble to mimic context reset sent by FW
+    0xc0001200u, 0u, // IT_CLEAR_STATE
+
+    // Actual init state sequence    
     0xc0017600u, 0x216u, 0xffffffffu,
     0xc0017600u, 0x217u, 0xffffffffu,
     0xc0017600u, 0x215u, 0u,
@@ -179,9 +191,13 @@ static constexpr std::array InitSequence200{
     0xc0036900u, 0x295u, 0x100u, 0x100u, 4u,
     0xc0017900u, 0x200u, 0xe0000000u,
 };
-static_assert(InitSequence200.size() == 0x76);
+static_assert(InitSequence200.size() == 0x76 + 2);
 
 static constexpr std::array InitSequence350{
+    // A fake preamble to mimic context reset sent by FW
+    0xc0001200u, 0u, // IT_CLEAR_STATE
+
+    // Actual init state sequence    
     0xc0017600u, 0x216u, 0xffffffffu,
     0xc0017600u, 0x217u, 0xffffffffu,
     0xc0017600u, 0x215u, 0u,
@@ -224,7 +240,7 @@ static constexpr std::array InitSequence350{
     0xc0017900u, 0x200u, 0xe0000000u,
     0xc0016900u, 0x2aau, 0xffu,
 };
-static_assert(InitSequence350.size() == 0x7c);
+static_assert(InitSequence350.size() == 0x7c + 2);
 
 static constexpr std::array CtxInitSequence{
     0xc0012800u, 0x80000000u, 0x80000000u,
@@ -735,11 +751,11 @@ u32 PS4_SYSV_ABI sceGnmDrawInitDefaultHardwareState(u32* cmdbuf, u32 size) {
             cmdbuf = ClearContextState(cmdbuf);
         }
 
-        std::memcpy(cmdbuf, InitSequence.data(), InitSequence.size() * 4);
+        std::memcpy(cmdbuf, &InitSequence[2], (InitSequence.size() - 2) * 4);
-        cmdbuf += InitSequence.size();
+        cmdbuf += InitSequence.size() - 2;
 
         const auto cmdbuf_left =
-            HwInitPacketSize - InitSequence.size() - (clear_state ? 0xc : 0) - 1;
+            HwInitPacketSize - (InitSequence.size() - 2) - (clear_state ? 0xc : 0) - 1;
         cmdbuf = WriteHeader<PM4ItOpcode::Nop>(cmdbuf, cmdbuf_left);
         cmdbuf = WriteBody(cmdbuf, 0u);
 
@@ -757,10 +773,10 @@ u32 PS4_SYSV_ABI sceGnmDrawInitDefaultHardwareState175(u32* cmdbuf, u32 size) {
     }
 
     cmdbuf = ClearContextState(cmdbuf);
-    std::memcpy(cmdbuf, InitSequence175.data(), InitSequence175.size() * 4);
+    std::memcpy(cmdbuf, &InitSequence175[2], (InitSequence175.size() - 2) * 4);
-    cmdbuf += InitSequence175.size();
+    cmdbuf += InitSequence175.size() - 2;
 
-    constexpr auto cmdbuf_left = HwInitPacketSize - InitSequence175.size() - 0xc - 1;
+    constexpr auto cmdbuf_left = HwInitPacketSize - (InitSequence175.size() - 2) - 0xc - 1;
     WriteTrailingNop<cmdbuf_left>(cmdbuf);
 
     return HwInitPacketSize;
@@ -778,11 +794,11 @@ u32 PS4_SYSV_ABI sceGnmDrawInitDefaultHardwareState200(u32* cmdbuf, u32 size) {
             cmdbuf = ClearContextState(cmdbuf);
         }
 
-        std::memcpy(cmdbuf, InitSequence200.data(), InitSequence200.size() * 4);
+        std::memcpy(cmdbuf, &InitSequence200[2], (InitSequence200.size() - 2) * 4);
-        cmdbuf += InitSequence200.size();
+        cmdbuf += InitSequence200.size() - 2;
 
         const auto cmdbuf_left =
-            HwInitPacketSize - InitSequence200.size() - (clear_state ? 0xc : 0) - 1;
+            HwInitPacketSize - (InitSequence200.size() - 2) - (clear_state ? 0xc : 0) - 1;
         cmdbuf = WriteHeader<PM4ItOpcode::Nop>(cmdbuf, cmdbuf_left);
         cmdbuf = WriteBody(cmdbuf, 0u);
 
@@ -804,11 +820,11 @@ u32 PS4_SYSV_ABI sceGnmDrawInitDefaultHardwareState350(u32* cmdbuf, u32 size) {
             cmdbuf = ClearContextState(cmdbuf);
         }
 
-        std::memcpy(cmdbuf, InitSequence350.data(), InitSequence350.size() * 4);
+        std::memcpy(cmdbuf, &InitSequence350[2], (InitSequence350.size() - 2) * 4);
-        cmdbuf += InitSequence350.size();
+        cmdbuf += InitSequence350.size() - 2;
 
         const auto cmdbuf_left =
-            HwInitPacketSize - InitSequence350.size() - (clear_state ? 0xc : 0) - 1;
+            HwInitPacketSize - (InitSequence350.size() - 2) - (clear_state ? 0xc : 0) - 1;
         cmdbuf = WriteHeader<PM4ItOpcode::Nop>(cmdbuf, cmdbuf_left);
         cmdbuf = WriteBody(cmdbuf, 0u);
 
@@ -1743,7 +1759,7 @@ s32 PS4_SYSV_ABI sceGnmSetVsShader(u32* cmdbuf, u32 size, const u32* vs_regs, u3
         return -1;
     }
 
-    const u32 var = shader_modifier == 0 ? vs_regs[2] : (vs_regs[2] & 0xfcfffc3f | shader_modifier);
+    const u32 var = shader_modifier == 0 ? vs_regs[2] : (vs_regs[2] & 0xfcfffc3f) | shader_modifier;
     cmdbuf = PM4CmdSetData::SetShReg(cmdbuf, 0x48u, vs_regs[0], 0u);   // SPI_SHADER_PGM_LO_VS
     cmdbuf = PM4CmdSetData::SetShReg(cmdbuf, 0x4au, var, vs_regs[3]);  // SPI_SHADER_PGM_RSRC1_VS
     cmdbuf = PM4CmdSetData::SetContextReg(cmdbuf, 0x207u, vs_regs[6]); // PA_CL_VS_OUT_CNTL

src/video_core/amdgpu/default_context.cpp

@@ -0,0 +1,82 @@
+// SPDX-FileCopyrightText: Copyright 2024 shadPS4 Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#include "common/types.h"
+#include "video_core/amdgpu/liverpool.h"
+
+#include <array>
+
+namespace AmdGpu {
+
+// The following values are taken from fpPS4:
+// https://github.com/red-prig/fpPS4/blob/436b43064be4c78229500f3d3c054fc76639247d/chip/pm4_pfp.pas#L410
+//
+const std::array<u32, Liverpool::NumRegs> Liverpool::reg_array_default{
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x40004000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x80000000u, 0x40004000u, 0x00000000u, 0x40004000u, 0x00000000u,
+    0x40004000u, 0x00000000u, 0x40004000u, 0x00000000u, 0x40004000u, 0xaa99aaaau, 0x00000000u,
+    0x80000000u, 0x40004000u, 0x00000000u, 0x00000000u, 0x80000000u, 0x40004000u, 0x80000000u,
+    0x40004000u, 0x80000000u, 0x40004000u, 0x80000000u, 0x40004000u, 0x80000000u, 0x40004000u,
+    0x80000000u, 0x40004000u, 0x80000000u, 0x40004000u, 0x80000000u, 0x40004000u, 0x80000000u,
+    0x40004000u, 0x80000000u, 0x40004000u, 0x80000000u, 0x40004000u, 0x80000000u, 0x40004000u,
+    0x80000000u, 0x40004000u, 0x80000000u, 0x40004000u, 0x80000000u, 0x40004000u, 0x80000000u,
+    0x40004000u, 0x00000000u, 0x3f800000u, 0x00000000u, 0x3f800000u, 0x00000000u, 0x3f800000u,
+    0x00000000u, 0x3f800000u, 0x00000000u, 0x3f800000u, 0x00000000u, 0x3f800000u, 0x00000000u,
+    0x3f800000u, 0x00000000u, 0x3f800000u, 0x00000000u, 0x3f800000u, 0x00000000u, 0x3f800000u,
+    0x00000000u, 0x3f800000u, 0x00000000u, 0x3f800000u, 0x00000000u, 0x3f800000u, 0x00000000u,
+    0x3f800000u, 0x00000000u, 0x3f800000u, 0x00000000u, 0x3f800000u, 0x2a00161au, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000002u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00090000u,
+    0x00000004u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000100u, 0x00000080u, 0x00000002u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x000000ffu, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00001000u, 0x00000000u, 0x00000005u, 0x3f800000u,
+    0x3f800000u, 0x3f800000u, 0x3f800000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x0000000eu, 0x00000010u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+    0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u, 0x00000000u,
+};
+
+} // namespace AmdGpu

src/video_core/amdgpu/liverpool.cpp

@@ -216,6 +216,7 @@ Liverpool::Task Liverpool::ProcessGraphics(std::span<const u32> dcb, std::span<c
             break;
         }
         case PM4ItOpcode::ClearState: {
+            std::memcpy(regs.reg_array.data(), reg_array_default.data(), NumRegs * sizeof(u32));
             break;
         }
         case PM4ItOpcode::SetConfigReg: {

src/video_core/amdgpu/liverpool.h

@@ -1020,6 +1020,7 @@ struct Liverpool {
     };
 
     Regs regs{};
+    static const std::array<u32, NumRegs> reg_array_default;
 
     // See for a comment in context reg parsing code
     union CbDbExtent {