Started on GPU initialization

nx/include/switch/nvidia/cmds/3d.h

@@ -1,8 +1,6 @@
 enum {
-    NvReg3D_ClearColor = 0x360,
+    NvReg3D_LineWidthSeparate = 0x83,
-    NvReg3D_ScreenScissorHorizontal = 0x3fd,
+    NvReg3D_LocalBase = 0x1df,
-    NvReg3D_ScreenScissorVertical = 0x3fe,
-    NvReg3D_RenderTargetControl = 0x487,
     NvReg3D_RenderTargetNAddr = 0x200,
     NvReg3D_RenderTargetNHorizontal = 0x202,
     NvReg3D_RenderTargetNVertical = 0x203,
@@ -11,8 +9,46 @@ enum {
     NvReg3D_RenderTargetNArrayMode = 0x206,
     NvReg3D_RenderTargetNLayerStride = 0x207,
     NvReg3D_RenderTargetNBaseLayer = 0x208,
+    NvReg3D_ClipRectNHorizontal = 0x340,
+    NvReg3D_ClipRectNVertical = 0x341,
+    NvReg3D_CallLimitLog = 0x359,
+    NvReg3D_ClearColor = 0x360,
+    NvReg3D_PrimRestartWithDrawArrays = 0x37a,
+    NvReg3D_ScissorNEnable = 0x380,
+    NvReg3D_ScissorNHorizontal = 0x381,
+    NvReg3D_ScissorNVertical = 0x382,
+    NvReg3D_VertexRunoutAddr = 0x3e1,
+    NvReg3D_ScreenScissorHorizontal = 0x3fd,
+    NvReg3D_ScreenScissorVertical = 0x3fe,
+    NvReg3D_ClearFlags = 0x43e,
+    NvReg3D_RenderTargetControl = 0x487, 
+    NvReg3D_LinkedTsc = 0x48d,
+    NvReg3D_ZcullStatCtrsEnable = 0x547,
     NvReg3D_MultisampleEnable = 0x54d,
+    NvReg3D_MultisampleControl = 0x54f,
+    NvReg3D_ZcullRegion = 0x564,
+    NvReg3D_StencilTwoSideEnable = 0x565,
+    NvReg3D_MultisampleCsaaEnable = 0x56d,
+    NvReg3D_Layer = 0x573,
+    NvReg3D_MultisampleMode = 0x574,
+    NvReg3D_VertexIdGenMode = 0x593,
+    NvReg3D_PointRasterRules = 0x597,
+    NvReg3D_ProvokingVertexLast = 0x5a1,
+    NvReg3D_VertexStreamNEnableDivisor = 0x620,
+    NvReg3D_ZcullTestMask = 0x65b,
     NvReg3D_ClearBufferTrigger = 0x674,
+    NvReg3D_TextureConstBufferIndex = 0x982,
 };
 
+#define NvReg3D_ClipRectHorizontal(n) \
+    ((NvReg3D_ClipRectNHorizontal) + 2*(n))
+#define NvReg3D_ClipRectVertical(n) \
+    ((NvReg3D_ClipRectNVertical) + 2*(n))
+
+#define NvReg3D_ScissorEnable(n) \
+    ((NvReg3D_ScissorNEnable) + 4*(n))
+
+#define NvReg3D_VertexStreamEnableDivisor(n) \
+    ((NvReg3D_VertexStreamNEnableDivisor) + (n))
+
 void nvCmdsClearColor(NvCmdList* cmds, float colors[4]);

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

@@ -22,13 +22,13 @@ u32*   nvCmdListInsert(NvCmdList* c, size_t num_cmds);
     (0x80000000 | (reg) | ((subc) << 13) | ((val) << 16))
 
 #define NvRep(subc, reg, ...) \
-    (0x60000000 | ((reg) | ((subc) << 13) | ((sizeof((u32[]) { __VA_ARGS__ })) << 16))), __VA_ARGS__
+    (0x60000000 | ((reg) | ((subc) << 13) | ((sizeof((u32[]) { __VA_ARGS__ }) / 4) << 16))), __VA_ARGS__
 
 #define NvIncr(subc, reg, ...) \
-    (0x20000000 | ((reg) | ((subc) << 13) | ((sizeof((u32[]) { __VA_ARGS__ })) << 16))), __VA_ARGS__
+    (0x20000000 | ((reg) | ((subc) << 13) | ((sizeof((u32[]) { __VA_ARGS__ }) / 4) << 16))), __VA_ARGS__
 
 #define NvIncrOnce(subc, reg, ...) \
-    (0xA0000000 | ((reg) | ((subc) << 13) | ((sizeof((u32[]) { __VA_ARGS__ })) << 16))), __VA_ARGS__
+    (0xA0000000 | ((reg) | ((subc) << 13) | ((sizeof((u32[]) { __VA_ARGS__ }) / 4) << 16))), __VA_ARGS__
 
 static inline u32 f2i(float f) {
     #pragma GCC diagnostic ignored "-Wstrict-aliasing"

nx/source/nvidia/cmds/3d.c

@@ -1,6 +1,105 @@
 #include <switch.h>
 #include <string.h>
 
+void nvCmds3DInitialize(NvCmdList* cmds) {
+    NvCmd(
+        cmds,
+        // ???
+        NvIncr(0, 0xd1a, 0, 0xffffffff),
+        NvImm(0, 0xd19, 0),
+        // Reset multisampling
+        NvImm(0, NvReg3D_MultisampleEnable, 0),
+        NvImm(0, NvReg3D_MultisampleCsaaEnable, 0),
+        NvImm(0, NvReg3D_MultisampleMode, 0),
+        NvImm(0, NvReg3D_MultisampleControl, 0),
+        // ???
+        NvImm(0, 0x433, 4),
+        NvImm(0, 0x438, 0xff),
+        NvImm(0, 0x439, 0xff),
+        NvImm(0, 0x43b, 0xff),
+        NvImm(0, 0x43c, 4),
+        NvImm(0, 0x1d3, 0x3f),
+        //
+        NvIncr(0, NvReg3D_ClipRectNHorizontal, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0),
+        NvImm(0, NvReg3D_ClearFlags, 0x101));
+
+    size_t i;
+    for (i=0; i<16; i++)
+        NvCmd(cmds, NvImm(0, NvReg3D_ScissorEnable(i), 1));
+
+    NvCmd(cmds, NvImm(0, NvReg3D_PrimRestartWithDrawArrays, 1),
+          NvImm(0, NvReg3D_PointRasterRules, 0),
+          NvImm(0, NvReg3D_LinkedTsc, 0),
+          NvImm(0, NvReg3D_ProvokingVertexLast, 1),
+          // ???
+          NvImm(0, 0x54a, 0),
+          NvImm(0, 0x400, 0x10),
+          NvImm(0, 0x86, 0x10),
+          NvImm(0, 0x43f, 0x10),
+          NvImm(0, 0x4a4, 0x10),
+          NvImm(0, 0x4b6, 0x10),
+          NvImm(0, 0x4b7, 0x10),
+          //
+          NvImm(0, NvReg3D_CallLimitLog, 8),
+          // ???
+          NvImm(0, 0x450, 0x10),
+          NvImm(0, 0x584, 0xe));
+
+    for (i=0; i<16; i++) {
+        NvCmd(cmds, NvImm(0, NvReg3D_VertexStreamEnableDivisor(i), 0));
+    }
+
+    NvCmd(
+        cmds,
+        NvImm(0, NvReg3D_VertexIdGenMode, 0),
+        NvImm(0, NvReg3D_ZcullStatCtrsEnable, 1),
+        NvImm(0, NvReg3D_LineWidthSeparate, 1),
+        // ???
+        NvImm(0, 0xc3, 0),
+        NvImm(0, 0xc0, 3),
+        NvImm(0, 0x3f7, 1),
+        NvImm(0, 0x670, 1),
+        NvImm(0, 0x3e3, 0),
+        NvImm(0, NvReg3D_StencilTwoSideEnable, 1),
+        NvImm(0, NvReg3D_TextureConstBufferIndex, 2),
+        NvImm(0, 0xc4, 0x503),
+        NvIncr(0, NvReg3D_LocalBase, 0x01000000),
+        NvImm(0, 0x44c, 0x13),
+        NvImm(0, 0xdd, 0),
+        NvIncr(0, NvReg3D_Layer, 0x10000),
+        NvImm(0, 0x488, 5),
+        NvIncr(0, 0x514, 0x00800008),
+        NvImm(0, 0xab, 3),
+        NvImm(0, 0xa4, 0),
+        NvImm(0, 0x221, 0x3f));
+
+    // TODO: Call some macro shit (0xe16).
+
+    NvCmd(
+        cmds,
+        // Reset Zcull.
+        NvImm(0, NvReg3D_ZcullTestMask, 0),
+        NvImm(0, 0x65a, 0x11),
+        NvImm(0, NvReg3D_ZcullRegion, 0),
+        NvIncr(0, 0x054, 0x49000000, 0x49000001),
+        NvIncr(0, 0xd18, 0x05000500),
+        );
+
+    // TODO: Call some macro shit (0xe34)
+
+    // TODO: Fill in NvReg3D_VertexRunoutAddr with a valid addr.
+
+    // TODO: Call some macro shit (0xe2a)
+
+    // TODO: CB_DATA stuff
+
+    // TODO: Call some macro shit (0xe32)
+
+    // TODO: CB_DATA stuff
+
+    // TODO: CB_BIND stuff
+}
+
 void nvCmdsClearBuffer(
     NvCmdList* cmds, NvBuffer* buf, u32 width, u32 height, float colors[4])
 {

nx/source/nvidia/cmds/common.c

@@ -2,9 +2,10 @@
 #include <string.h>
 
 void nvCmdsFifoInit(NvCmdList* cmds) {
-    NvCmd(cmds, NvIncr(0, NvCmdCommon_BindObject, NvClassNumber_3D));
+    NvCmd(cmds,
-    NvCmd(cmds, NvIncr(1, NvCmdCommon_BindObject, NvClassNumber_Compute));
+        NvIncr(0,  NvCmdCommon_BindObject, NvClassNumber_3D),
-    NvCmd(cmds, NvIncr(2, NvCmdCommon_BindObject, NvClassNumber_Kepler));
+        NvIncr(1, NvCmdCommon_BindObject, NvClassNumber_Compute),
-    NvCmd(cmds, NvIncr(3, NvCmdCommon_BindObject, NvClassNumber_2D));
+        NvIncr(2, NvCmdCommon_BindObject, NvClassNumber_Kepler),
-    NvCmd(cmds, NvIncr(4, NvCmdCommon_BindObject, NvClassNumber_DMA));
+        NvIncr(3, NvCmdCommon_BindObject, NvClassNumber_2D),
+        NvIncr(4, NvCmdCommon_BindObject, NvClassNumber_DMA));
 }