gnm: Implement sceGnmDrawIndexIndirectMulti (#2889)

src/core/libraries/gnmdriver/gnmdriver.cpp

@@ -505,9 +505,10 @@ s32 PS4_SYSV_ABI sceGnmDrawIndexIndirectCountMulti(u32* cmdbuf, u32 size, u32 da
                                                    u32 flags) {
     LOG_TRACE(Lib_GnmDriver, "called");
 
-    if ((!sceKernelIsNeoMode() || !UseNeoCompatSequences) && !cmdbuf && (size == 16) &&
+    if ((!sceKernelIsNeoMode() || !UseNeoCompatSequences) && cmdbuf && (size == 16) &&
-        (shader_stage < ShaderStages::Max) && (vertex_sgpr_offset < 0x10u) &&
+        (vertex_sgpr_offset < 0x10u) && (instance_sgpr_offset < 0x10u) &&
-        (instance_sgpr_offset < 0x10u)) {
+        (shader_stage == ShaderStages::Vs || shader_stage == ShaderStages::Es ||
+         shader_stage == ShaderStages::Ls)) {
 
         cmdbuf = WriteHeader<PM4ItOpcode::Nop>(cmdbuf, 2);
         cmdbuf = WriteBody(cmdbuf, 0u);
@@ -535,10 +536,33 @@ s32 PS4_SYSV_ABI sceGnmDrawIndexIndirectCountMulti(u32* cmdbuf, u32 size, u32 da
     return -1;
 }
 
-int PS4_SYSV_ABI sceGnmDrawIndexIndirectMulti() {
+int PS4_SYSV_ABI sceGnmDrawIndexIndirectMulti(u32* cmdbuf, u32 size, u32 data_offset, u32 max_count,
-    LOG_ERROR(Lib_GnmDriver, "(STUBBED) called");
+                                              u32 shader_stage, u32 vertex_sgpr_offset,
-    UNREACHABLE();
+                                              u32 instance_sgpr_offset, u32 flags) {
-    return ORBIS_OK;
+    LOG_TRACE(Lib_GnmDriver, "called");
+
+    if (cmdbuf && (size == 11) && (vertex_sgpr_offset < 0x10u) && (instance_sgpr_offset < 0x10u) &&
+        (shader_stage == ShaderStages::Vs || shader_stage == ShaderStages::Es ||
+         shader_stage == ShaderStages::Ls)) {
+
+        const auto predicate = flags & 1 ? PM4Predicate::PredEnable : PM4Predicate::PredDisable;
+        cmdbuf = WriteHeader<PM4ItOpcode::DrawIndexIndirectMulti>(
+            cmdbuf, 6, PM4ShaderType::ShaderGraphics, predicate);
+
+        const auto sgpr_offset = indirect_sgpr_offsets[shader_stage];
+
+        cmdbuf[0] = data_offset;
+        cmdbuf[1] = vertex_sgpr_offset == 0 ? 0 : (vertex_sgpr_offset & 0xffffu) + sgpr_offset;
+        cmdbuf[2] = instance_sgpr_offset == 0 ? 0 : (instance_sgpr_offset & 0xffffu) + sgpr_offset;
+        cmdbuf[3] = max_count;
+        cmdbuf[4] = sizeof(DrawIndexedIndirectArgs);
+        cmdbuf[5] = sceKernelIsNeoMode() ? flags & 0xe0000000u : 0;
+
+        cmdbuf += 6;
+        WriteTrailingNop<3>(cmdbuf);
+        return ORBIS_OK;
+    }
+    return -1;
 }
 
 int PS4_SYSV_ABI sceGnmDrawIndexMultiInstanced() {

src/core/libraries/gnmdriver/gnmdriver.h

@@ -51,7 +51,9 @@ s32 PS4_SYSV_ABI sceGnmDrawIndexIndirectCountMulti(u32* cmdbuf, u32 size, u32 da
                                                    u32 max_count, u64 count_addr, u32 shader_stage,
                                                    u32 vertex_sgpr_offset, u32 instance_sgpr_offset,
                                                    u32 flags);
-int PS4_SYSV_ABI sceGnmDrawIndexIndirectMulti();
+int PS4_SYSV_ABI sceGnmDrawIndexIndirectMulti(u32* cmdbuf, u32 size, u32 data_offset, u32 max_count,
+                                              u32 shader_stage, u32 vertex_sgpr_offset,
+                                              u32 instance_sgpr_offset, u32 flags);
 int PS4_SYSV_ABI sceGnmDrawIndexMultiInstanced();
 s32 PS4_SYSV_ABI sceGnmDrawIndexOffset(u32* cmdbuf, u32 size, u32 index_offset, u32 index_count,
                                        u32 flags);

src/video_core/amdgpu/liverpool.cpp

@@ -455,14 +455,14 @@ Liverpool::Task Liverpool::ProcessGraphics(std::span<const u32> dcb, std::span<c
             case PM4ItOpcode::DrawIndirect: {
                 const auto* draw_indirect = reinterpret_cast<const PM4CmdDrawIndirect*>(header);
                 const auto offset = draw_indirect->data_offset;
-                const auto size = sizeof(DrawIndirectArgs);
+                const auto stride = sizeof(DrawIndirectArgs);
                 if (DebugState.DumpingCurrentReg()) {
                     DebugState.PushRegsDump(base_addr, reinterpret_cast<uintptr_t>(header), regs);
                 }
                 if (rasterizer) {
                     const auto cmd_address = reinterpret_cast<const void*>(header);
                     rasterizer->ScopeMarkerBegin(fmt::format("gfx:{}:DrawIndirect", cmd_address));
-                    rasterizer->DrawIndirect(false, indirect_args_addr, offset, size, 1, 0);
+                    rasterizer->DrawIndirect(false, indirect_args_addr, offset, stride, 1, 0);
                     rasterizer->ScopeMarkerEnd();
                 }
                 break;
@@ -471,7 +471,7 @@ Liverpool::Task Liverpool::ProcessGraphics(std::span<const u32> dcb, std::span<c
                 const auto* draw_index_indirect =
                     reinterpret_cast<const PM4CmdDrawIndexIndirect*>(header);
                 const auto offset = draw_index_indirect->data_offset;
-                const auto size = sizeof(DrawIndexedIndirectArgs);
+                const auto stride = sizeof(DrawIndexedIndirectArgs);
                 if (DebugState.DumpingCurrentReg()) {
                     DebugState.PushRegsDump(base_addr, reinterpret_cast<uintptr_t>(header), regs);
                 }
@@ -479,25 +479,46 @@ Liverpool::Task Liverpool::ProcessGraphics(std::span<const u32> dcb, std::span<c
                     const auto cmd_address = reinterpret_cast<const void*>(header);
                     rasterizer->ScopeMarkerBegin(
                         fmt::format("gfx:{}:DrawIndexIndirect", cmd_address));
-                    rasterizer->DrawIndirect(true, indirect_args_addr, offset, size, 1, 0);
+                    rasterizer->DrawIndirect(true, indirect_args_addr, offset, stride, 1, 0);
                     rasterizer->ScopeMarkerEnd();
                 }
                 break;
             }
-            case PM4ItOpcode::DrawIndexIndirectCountMulti: {
+            case PM4ItOpcode::DrawIndexIndirectMulti: {
                 const auto* draw_index_indirect =
                     reinterpret_cast<const PM4CmdDrawIndexIndirectMulti*>(header);
                 const auto offset = draw_index_indirect->data_offset;
                 if (DebugState.DumpingCurrentReg()) {
                     DebugState.PushRegsDump(base_addr, reinterpret_cast<uintptr_t>(header), regs);
                 }
+                if (rasterizer) {
+                    const auto cmd_address = reinterpret_cast<const void*>(header);
+                    rasterizer->ScopeMarkerBegin(
+                        fmt::format("gfx:{}:DrawIndexIndirectMulti", cmd_address));
+                    rasterizer->DrawIndirect(true, indirect_args_addr, offset,
+                                             draw_index_indirect->stride,
+                                             draw_index_indirect->count, 0);
+                    rasterizer->ScopeMarkerEnd();
+                }
+                break;
+            }
+            case PM4ItOpcode::DrawIndexIndirectCountMulti: {
+                const auto* draw_index_indirect =
+                    reinterpret_cast<const PM4CmdDrawIndexIndirectCountMulti*>(header);
+                const auto offset = draw_index_indirect->data_offset;
+                if (DebugState.DumpingCurrentReg()) {
+                    DebugState.PushRegsDump(base_addr, reinterpret_cast<uintptr_t>(header), regs);
+                }
                 if (rasterizer) {
                     const auto cmd_address = reinterpret_cast<const void*>(header);
                     rasterizer->ScopeMarkerBegin(
                         fmt::format("gfx:{}:DrawIndexIndirectCountMulti", cmd_address));
-                    rasterizer->DrawIndirect(
+                    rasterizer->DrawIndirect(true, indirect_args_addr, offset,
-                        true, indirect_args_addr, offset, draw_index_indirect->stride,
+                                             draw_index_indirect->stride,
-                        draw_index_indirect->count, draw_index_indirect->countAddr);
+                                             draw_index_indirect->count,
+                                             draw_index_indirect->count_indirect_enable.Value()
+                                                 ? draw_index_indirect->count_addr
+                                                 : 0);
                     rasterizer->ScopeMarkerEnd();
                 }
                 break;

src/video_core/amdgpu/pm4_cmds.h

@@ -860,6 +860,24 @@ struct PM4CmdDrawIndexIndirect {
 };
 
 struct PM4CmdDrawIndexIndirectMulti {
+    PM4Type3Header header; ///< header
+    u32 data_offset;       ///< Byte aligned offset where the required data structure starts
+    union {
+        u32 dw2;
+        BitField<0, 16, u32> base_vtx_loc; ///< Offset where the CP will write the
+                                           ///< BaseVertexLocation it fetched from memory
+    };
+    union {
+        u32 dw3;
+        BitField<0, 16, u32> start_inst_loc; ///< Offset where the CP will write the
+                                             ///< StartInstanceLocation it fetched from memory
+    };
+    u32 count;          ///< Count of data structures to loop through before going to next packet
+    u32 stride;         ///< Stride in memory from one data structure to the next
+    u32 draw_initiator; ///< Draw Initiator Register
+};
+
+struct PM4CmdDrawIndexIndirectCountMulti {
     PM4Type3Header header; ///< header
     u32 data_offset;       ///< Byte aligned offset where the required data structure starts
     union {
@@ -874,14 +892,14 @@ struct PM4CmdDrawIndexIndirectMulti {
     };
     union {
         u32 dw4;
-        BitField<0, 16, u32> drawIndexLoc; ///< register offset to write the Draw Index count
+        BitField<0, 16, u32> draw_index_loc; ///< register offset to write the Draw Index count
         BitField<30, 1, u32>
-            countIndirectEnable; ///< Indicates the data structure count is in memory
+            count_indirect_enable; ///< Indicates the data structure count is in memory
         BitField<31, 1, u32>
-            drawIndexEnable; ///< Enables writing of Draw Index count to DRAW_INDEX_LOC
+            draw_index_enable; ///< Enables writing of Draw Index count to DRAW_INDEX_LOC
     };
     u32 count;          ///< Count of data structures to loop through before going to next packet
-    u64 countAddr;      ///< DWord aligned Address[31:2]; Valid if countIndirectEnable is set
+    u64 count_addr;     ///< DWord aligned Address[31:2]; Valid if countIndirectEnable is set
     u32 stride;         ///< Stride in memory from one data structure to the next
     u32 draw_initiator; ///< Draw Initiator Register
 };

src/video_core/renderer_vulkan/vk_instance.cpp

@@ -338,6 +338,7 @@ bool Instance::CreateDevice() {
                 .geometryShader = features.geometryShader,
                 .tessellationShader = features.tessellationShader,
                 .logicOp = features.logicOp,
+                .multiDrawIndirect = features.multiDrawIndirect,
                 .depthBiasClamp = features.depthBiasClamp,
                 .fillModeNonSolid = features.fillModeNonSolid,
                 .depthBounds = features.depthBounds,