Changes to NvCmdList and NvGpfifo, see details:

- NvCmdList:
  - Added 'offset' parameter: start of the cmdlist within the buffer
  - Added nvCmdListReset for resetting a cmdlist (offset/num_cmds cleared)
- NvGpfifo:
  - nvGpfifoCreate: disabled fence wait
  - nvGpfifoSubmit -> nvGpfifoSubmitCmdList:
    - Added fence_incr parameter
    - Cleaned up flags and gpfifo entry creation
    - Now advances cmdlist offset by the number of cmds submitted
      instead of going back to the beginning

nx/include/switch/nvidia/gpu/cmd_list.h

@@ -4,6 +4,7 @@ typedef struct NvGpu NvGpu;
 
 typedef struct {
     NvBuffer buffer;
+    size_t   offset;
     size_t   num_cmds;
     size_t   max_cmds;
     NvGpu*   parent;
@@ -15,4 +16,5 @@ void   nvCmdListClose(NvCmdList* c);
 iova_t nvCmdListGetGpuAddr(NvCmdList* c);
 u64    nvCmdListGetListSize(NvCmdList* c);
 
+void   nvCmdListReset(NvCmdList* c);
 u32*   nvCmdListInsert(NvCmdList* c, size_t num_cmds);

nx/include/switch/nvidia/gpu/gpfifo.h

@@ -10,4 +10,4 @@ void   nvGpfifoClose(NvGpfifo* f);
 #define NV_MAKE_GPFIFO_ENTRY(iova, size) \
     ((iova) | (((u64)(size)) << 42))
 
-Result nvGpfifoSubmit(NvGpfifo* f, NvCmdList* cmd_list, NvFence* fence_out);
+Result nvGpfifoSubmitCmdList(NvGpfifo* f, NvCmdList* cmd_list, u32 fence_incr, NvFence* fence_out);

nx/source/nvidia/gpu/cmd_list.c

@@ -28,6 +28,7 @@ Result nvCmdListCreate(NvCmdList* c, NvGpu* parent, size_t max_cmds)
     if (R_SUCCEEDED(rc)) {
         nvBufferMakeCpuUncached(&c->buffer);
 
+        c->offset = 0;
         c->num_cmds = 0;
         c->max_cmds = max_cmds;
         c->parent = parent;
@@ -48,14 +49,19 @@ u64 nvCmdListGetListSize(NvCmdList* c) {
     return c->num_cmds;
 }
 
+void nvCmdListReset(NvCmdList* c) {
+    c->offset = 0;
+    c->num_cmds = 0;
+}
+
 u32* nvCmdListInsert(NvCmdList* c, size_t num_cmds)
 {
     // Has enough space?
-    if ((c->num_cmds + num_cmds) > c->max_cmds)
+    if ((c->offset + c->num_cmds + num_cmds) > c->max_cmds)
         return NULL;
 
     c->num_cmds += num_cmds;
 
     u32* list = (u32*) nvBufferGetCpuAddr(&c->buffer);
-    return &list[c->num_cmds - num_cmds];
+    return &list[c->offset + c->num_cmds - num_cmds];
 }

nx/source/nvidia/gpu/gpfifo.c

@@ -26,8 +26,9 @@ Result nvGpfifoCreate(NvGpfifo* f, NvChannel* parent)
 
     NvFence fence;
     Result res = nvioctlChannel_AllocGpfifoEx2(parent->fd, DEFAULT_FIFO_ENTRIES, 1, 0, 0, 0, 0, &fence);
-    if (R_SUCCEEDED(res) && (s32)fence.id >= 0)
-        nvFenceWait(&fence, -1);
+    //__builtin_printf("nvGpfifoCreate initial fence: %d %u\n", (int)fence.id, fence.value);
+    //if (R_SUCCEEDED(res) && (s32)fence.id >= 0)
+    //    nvFenceWait(&fence, -1);
     return res;
 }
 
@@ -35,23 +36,29 @@ void nvGpfifoClose(NvGpfifo* f) {
     /**/
 }
 
-Result nvGpfifoSubmit(NvGpfifo* f, NvCmdList* cmd_list, NvFence* fence_out)
+Result nvGpfifoSubmitCmdList(NvGpfifo* f, NvCmdList* cmd_list, u32 fence_incr, NvFence* fence_out)
 {
     Result rc;
     nvioctl_gpfifo_entry ent;
+    NvFence fence;
 
-    u64 a =
-        nvCmdListGetGpuAddr(cmd_list) | (nvCmdListGetListSize(cmd_list) << 42);
+    ent.desc = nvCmdListGetGpuAddr(cmd_list) + 4*cmd_list->offset;
+    ent.desc32[1] |= (2 << 8) | (nvCmdListGetListSize(cmd_list) << 10);
 
-    ent.desc32[0] = a;
-    ent.desc32[1] = a >> 32;
+    fence.id = 0;
+    fence.value = fence_incr;
 
-    fence_out->id = -1;
-    fence_out->value = 0;
+    u32 flags = BIT(2);
+    if (fence_incr)
+        flags |= BIT(8);
 
     rc = nvioctlChannel_SubmitGpfifo(
-        f->parent->fd, &ent, 1, /*0x104*/0x104/*flags*/, fence_out);
+        f->parent->fd, &ent, 1, flags, &fence);
 
+    if (R_SUCCEEDED(rc) && fence_out)
+        *fence_out = fence;
+
+    cmd_list->offset += cmd_list->num_cmds;
     cmd_list->num_cmds = 0;
 
     return rc;