hot-fix: amdgpu: use different indirect dispatch packet on ASC

src/core/libraries/gnmdriver/gnmdriver.cpp

@@ -383,10 +383,17 @@ s32 PS4_SYSV_ABI sceGnmDispatchIndirect(u32* cmdbuf, u32 size, u32 data_offset,
     return -1;
 }
 
-int PS4_SYSV_ABI sceGnmDispatchIndirectOnMec() {
+s32 PS4_SYSV_ABI sceGnmDispatchIndirectOnMec(u32* cmdbuf, u32 size, VAddr args, u32 modifier) {
-    LOG_ERROR(Lib_GnmDriver, "(STUBBED) called");
+    if (cmdbuf != nullptr && size == 8 && args != 0 && ((args & 3u) == 0)) {
+        cmdbuf[0] = 0xc0021602 | (modifier & 1u);
+        *(VAddr*)(&cmdbuf[1]) = args;
+        cmdbuf[3] = (modifier & 0x18) | 1u;
+        cmdbuf[4] = 0xc0021000;
+        cmdbuf[5] = 0;
         return ORBIS_OK;
     }
+    return ORBIS_FAIL;
+}
 
 u32 PS4_SYSV_ABI sceGnmDispatchInitDefaultHardwareState(u32* cmdbuf, u32 size) {
     LOG_TRACE(Lib_GnmDriver, "called");

src/core/libraries/gnmdriver/gnmdriver.h

@@ -39,7 +39,7 @@ int PS4_SYSV_ABI sceGnmDisableMipStatsReport();
 s32 PS4_SYSV_ABI sceGnmDispatchDirect(u32* cmdbuf, u32 size, u32 threads_x, u32 threads_y,
                                       u32 threads_z, u32 flags);
 s32 PS4_SYSV_ABI sceGnmDispatchIndirect(u32* cmdbuf, u32 size, u32 data_offset, u32 flags);
-int PS4_SYSV_ABI sceGnmDispatchIndirectOnMec();
+s32 PS4_SYSV_ABI sceGnmDispatchIndirectOnMec(u32* cmdbuf, u32 size, VAddr args, u32 modifier);
 u32 PS4_SYSV_ABI sceGnmDispatchInitDefaultHardwareState(u32* cmdbuf, u32 size);
 s32 PS4_SYSV_ABI sceGnmDrawIndex(u32* cmdbuf, u32 size, u32 index_count, uintptr_t index_addr,
                                  u32 flags, u32 type);

src/video_core/amdgpu/liverpool.cpp

@@ -823,10 +823,10 @@ Liverpool::Task Liverpool::ProcessCompute(std::span<const u32> acb, u32 vqid) {
             break;
         }
         case PM4ItOpcode::DispatchIndirect: {
-            const auto* dispatch_indirect = reinterpret_cast<const PM4CmdDispatchIndirect*>(header);
+            const auto* dispatch_indirect =
+                reinterpret_cast<const PM4CmdDispatchIndirectMec*>(header);
             auto& cs_program = GetCsRegs();
-            const auto offset = dispatch_indirect->data_offset;
+            const auto ib_address = dispatch_indirect->Address<VAddr>();
-            const auto ib_address = mapped_queues[vqid].indirect_args_addr;
             const auto size = sizeof(PM4CmdDispatchIndirect::GroupDimensions);
             if (DebugState.DumpingCurrentReg()) {
                 DebugState.PushRegsDumpCompute(base_addr, reinterpret_cast<uintptr_t>(header),
@@ -835,7 +835,7 @@ Liverpool::Task Liverpool::ProcessCompute(std::span<const u32> acb, u32 vqid) {
             if (rasterizer && (cs_program.dispatch_initiator & 1)) {
                 const auto cmd_address = reinterpret_cast<const void*>(header);
                 rasterizer->ScopeMarkerBegin(fmt::format("acb[{}]:{}:Dispatch", vqid, cmd_address));
-                rasterizer->DispatchIndirect(ib_address, offset, size);
+                rasterizer->DispatchIndirect(ib_address, 0, size);
                 rasterizer->ScopeMarkerEnd();
             }
             break;

src/video_core/amdgpu/pm4_cmds.h

@@ -796,6 +796,18 @@ struct PM4CmdDispatchIndirect {
     u32 dispatch_initiator; ///< Dispatch Initiator Register
 };
 
+struct PM4CmdDispatchIndirectMec {
+    PM4Type3Header header;
+    u32 address0;
+    u32 address1;
+    u32 dispatch_initiator; ///< Dispatch Initiator Register
+
+    template <typename T>
+    T Address() const {
+        return reinterpret_cast<T>(address0 | (u64(address1 & 0xffff) << 32u));
+    }
+};
+
 struct DrawIndirectArgs {
     u32 vertex_count_per_instance;
     u32 instance_count;