Added nvioctlChannel_SubmitGPFIFO().

nx/include/switch/gfx/nvioctl.h

@@ -53,6 +53,11 @@ typedef struct {
     u32 value;
 } nvioctl_fence;
 
+typedef struct {
+    u32 entry0;
+    u32 entry1;
+} nvioctl_gpfifo_entry;
+
 //Used with nvioctlChannel_AllocObjCtx().
 enum nvioctl_channel_obj_classnum {
     NVIOCTL_CHANNEL_OBJ_CLASSNUM_2d = 0x902D,
@@ -86,6 +91,7 @@ Result nvioctlNvmap_Create(u32 fd, u32 size, u32 *nvmap_handle);
 Result nvioctlNvmap_Alloc(u32 fd, u32 nvmap_handle, u32 heapmask, u32 flags, u32 align, u8 kind, void* addr);
 
 Result nvioctlChannel_SetNvmapFd(u32 fd, u32 nvmap_fd);
+Result nvioctlChannel_SubmitGPFIFO(u32 fd, nvioctl_gpfifo_entry *entries, u32 num_entries, u32 flags, nvioctl_fence *fence_out);
 Result nvioctlChannel_AllocObjCtx(u32 fd, u32 class_num, u32 flags);
 Result nvioctlChannel_ZCullBind(u32 fd, u32 in[4]);
 Result nvioctlChannel_SetErrorNotifier(u32 fd, u64 offset, u64 size, u32 nvmap_handle);

nx/source/gfx/nvgfx.c

@@ -23,6 +23,7 @@ static u32 g_nvgfx_zcullinfo[40>>2];
 static nvioctl_va_region g_nvgfx_nvhostasgpu_varegions[2];
 static nvioctl_l2_state g_nvgfx_l2state;
 static nvioctl_fence g_nvgfx_nvhost_fence;
+static nvioctl_fence g_nvgfx_nvhostgpu_gpfifo_fence;
 static u8 *g_nvgfx_nvhost_userdata;
 static size_t g_nvgfx_nvhost_userdata_size;
 
@@ -78,6 +79,8 @@ Result nvgfxInitialize(void) {
 
     u32 zcullbind_data[4] = {0x58000, 0x5, 0x2, 0x0};
 
+    nvioctl_gpfifo_entry gpfifo_entries[2] = {{0x00030000, 0x00177a05}, {0x00031778, 0x80002e05}};
+
     g_nvgfx_fd_nvhostctrlgpu = 0;
     g_nvgfx_fd_nvhostasgpu = 0;
     g_nvgfx_fd_nvmap = 0;
@@ -91,6 +94,7 @@ Result nvgfxInitialize(void) {
     memset(g_nvgfx_nvhostasgpu_varegions, 0, sizeof(g_nvgfx_nvhostasgpu_varegions));
     memset(&g_nvgfx_l2state, 0, sizeof(nvioctl_l2_state));
     memset(&g_nvgfx_nvhost_fence, 0, sizeof(g_nvgfx_nvhost_fence));
+    memset(&g_nvgfx_nvhostgpu_gpfifo_fence, 0, sizeof(g_nvgfx_nvhostgpu_gpfifo_fence));
     g_nvgfx_nvhostasgpu_allocspace_offset = 0;
     g_nvgfx_zcullctxsize = 0;
 
@@ -171,7 +175,10 @@ Result nvgfxInitialize(void) {
 
     if (R_SUCCEEDED(rc)) rc = nvioctlChannel_ZCullBind(g_nvgfx_fd_nvhostgpu, zcullbind_data);
 
-    //Officially, ipcQueryPointerBufferSize and NVGPU_IOCTL_CHANNEL_SUBMIT_GPFIFO are used here with the duplicate session setup during nv serv init.
+    //Officially, ipcQueryPointerBufferSize and NVGPU_IOCTL_CHANNEL_SUBMIT_GPFIFO(nvioctlChannel_SubmitGPFIFO) are used here with the duplicate service session setup during nv serv init.
+    //TODO: This is probably used for GPU rendering? Is this really needed when not doing actual GPU rendering?
+    //Skip this since this causes a white-screen hang.
+    //if (R_SUCCEEDED(rc)) rc = nvioctlChannel_SubmitGPFIFO(g_nvgfx_fd_nvhostgpu, gpfifo_entries, 2, 0x104, &g_nvgfx_nvhostgpu_gpfifo_fence);
 
     //if (R_SUCCEEDED(rc)) rc = -1;
 

nx/source/gfx/nvioctl.c

@@ -251,6 +251,34 @@ Result nvioctlChannel_SetNvmapFd(u32 fd, u32 nvmap_fd) {
     return nvIoctl(fd, _IOW(0x48, 0x01, data), &data);
 }
 
+Result nvioctlChannel_SubmitGPFIFO(u32 fd, nvioctl_gpfifo_entry *entries, u32 num_entries, u32 flags, nvioctl_fence *fence_out) {
+    Result rc=0;
+
+    if(num_entries > 0x200) return MAKERESULT(MODULE_LIBNX, LIBNX_OUTOFMEM);//Make sure stack data doesn't get very large.
+
+    struct {
+        u64 gpfifo;                      // in (ignored) pointer to gpfifo fence structs
+        u32 num_entries;                 // in number of fence objects being submitted
+        u32 flags;                       // in
+        nvioctl_fence fence_out;          // out returned new fence object for others to wait on
+        nvioctl_gpfifo_entry entries[num_entries];   // in depends on num_entries
+    } data;
+
+
+    memset(&data, 0, sizeof(data));
+    data.gpfifo = 1;
+    data.num_entries = num_entries;
+    data.flags = flags;
+    memcpy(data.entries, entries, sizeof(data.entries));
+
+    rc = nvIoctl(fd, _IOWR(0x48, 0x08, data), &data);
+    if (R_FAILED(rc)) return rc;
+
+    if(fence_out) memcpy(fence_out, &data.fence_out, sizeof(data.fence_out));
+
+    return rc;
+}
+
 Result nvioctlChannel_AllocObjCtx(u32 fd, u32 class_num, u32 flags) {
     struct {
         u32 class_num;    // 0x902D=2d, 0xB197=3d, 0xB1C0=compute, 0xA140=kepler, 0xB0B5=DMA, 0xB06F=channel_gpfifo