//
// Generated by Microsoft (R) HLSL Shader Compiler 10.1
//
//
// Buffer Definitions: 
//
// cbuffer XeTextureLoadConstants
// {
//
//   uint xe_texture_load_guest_base;   // Offset:    0 Size:     4
//   uint xe_texture_load_guest_pitch;  // Offset:    4 Size:     4
//   uint2 xe_texture_load_guest_storage_width_height;// Offset:    8 Size:     8
//   uint3 xe_texture_load_size_blocks; // Offset:   16 Size:    12
//   uint xe_texture_load_is_3d_endian; // Offset:   28 Size:     4
//   uint xe_texture_load_host_base;    // Offset:   32 Size:     4
//   uint xe_texture_load_host_pitch;   // Offset:   36 Size:     4
//   uint xe_texture_load_height_texels;// Offset:   40 Size:     4 [unused]
//
// }
//
//
// Resource Bindings:
//
// Name                                 Type  Format         Dim      ID      HLSL Bind  Count
// ------------------------------ ---------- ------- ----------- ------- -------------- ------
// xe_texture_load_source            texture   uint2         buf      T0             t0      1 
// xe_texture_load_dest                  UAV   uint2         buf      U0             u0      1 
// XeTextureLoadConstants            cbuffer      NA          NA     CB0            cb0      1 
//
//
//
// Input signature:
//
// Name                 Index   Mask Register SysValue  Format   Used
// -------------------- ----- ------ -------- -------- ------- ------
// no Input
//
// Output signature:
//
// Name                 Index   Mask Register SysValue  Format   Used
// -------------------- ----- ------ -------- -------- ------- ------
// no Output
cs_5_1
dcl_globalFlags refactoringAllowed
dcl_constantbuffer CB0[0:0][3], immediateIndexed, space=0
dcl_resource_buffer (uint,uint,uint,uint) T0[0:0], space=0
dcl_uav_typed_buffer (uint,uint,uint,uint) U0[0:0], space=0
dcl_input vThreadID.xyz
dcl_temps 11
dcl_thread_group 2, 32, 1
ishl r0.x, vThreadID.x, l(4)
mov r0.yz, vThreadID.yyzy
uge r1.xyz, r0.xyzx, CB0[0][1].xyzx
or r0.w, r1.y, r1.x
or r0.w, r1.z, r0.w
if_nz r0.w
  ret 
endif 
mov r1.x, l(3)
mov r1.z, CB0[0][1].y
imul null, r0.yzw, r0.xxyz, r1.xxxz
ishl r0.y, r0.y, l(1)
imad r0.z, l(3), r0.w, r0.z
imad r0.y, r0.z, CB0[0][2].y, r0.y
iadd r0.y, r0.y, CB0[0][2].x
ushr r0.z, CB0[0][2].y, l(3)
ieq r0.w, CB0[0][0].y, l(-1)
if_nz r0.w
  and r1.x, CB0[0][1].w, l(1)
  if_nz r1.x
    iadd r1.xy, CB0[0][0].wzww, l(31, 31, 0, 0)
    ishr r2.xyz, vThreadID.yzyy, l(4, 2, 3, 0)
    ushr r1.xy, r1.xyxx, l(4, 5, 0, 0)
    and r1.x, r1.x, l(0x0ffffffe)
    imad r1.x, r2.y, r1.x, r2.x
    ibfe r1.z, l(27), l(1), vThreadID.x
    imad r1.x, r1.x, r1.y, r1.z
    ishl r1.y, vThreadID.y, l(9)
    ishr r1.y, r1.y, l(6)
    iadd r1.z, r2.y, r2.z
    and r1.yw, r1.yyyz, l(0, 48, 0, 1)
    ishr r2.x, r0.x, l(3)
    bfi r1.z, l(1), l(1), r1.z, l(0)
    iadd r1.z, r1.z, r2.x
    bfi r1.z, l(2), l(1), r1.z, l(0)
    iadd r1.z, r1.z, r1.w
    bfi r1.xw, l(21, 0, 0, 21), l(9, 0, 0, 12), r1.xxxx, l(0, 0, 0, 0)
    imad r1.xy, r1.yyyy, l(2, 16, 0, 0), r1.xwxx
    bfi r1.xy, l(2, 2, 0, 0), l(7, 10, 0, 0), vThreadID.zzzz, r1.xyxx
    bfi r1.w, l(1), l(4), vThreadID.y, l(0)
    ubfe r2.x, l(3), l(6), r1.x
    and r2.y, r1.z, l(4)
    bfi r1.z, l(2), l(8), r1.z, l(0)
    imad r1.z, r2.x, l(32), r1.z
    imad r1.z, r2.y, l(4), r1.z
    bfi r1.xy, l(5, 5, 0, 0), l(0, 3, 0, 0), r1.wwww, r1.xyxx
    bfi r1.y, l(9), l(3), r1.z, r1.y
    bfi r1.x, l(6), l(0), r1.x, r1.y
  else 
    ibfe r1.y, l(27), l(1), vThreadID.x
    ishr r1.zw, vThreadID.yyyy, l(0, 0, 5, 2)
    iadd r2.x, CB0[0][0].z, l(31)
    ushr r2.x, r2.x, l(5)
    imad r1.y, r1.z, r2.x, r1.y
    bfi r2.xyz, l(4, 4, 4, 0), l(4, 7, 6, 0), vThreadID.yyyy, l(0, 0, 0, 0)
    bfi r2.xyz, l(24, 24, 24, 0), l(8, 11, 10, 0), r1.yyyy, r2.xyzx
    ishl r1.y, vThreadID.y, l(7)
    and r1.y, r1.y, l(2048)
    bfi r1.y, l(12), l(0), r1.y, r2.y
    and r1.z, r2.z, l(1792)
    iadd r1.y, r1.y, r1.z
    and r1.z, r1.w, l(2)
    ishr r1.w, r0.x, l(3)
    iadd r1.z, r1.w, r1.z
    bfi r1.z, l(2), l(6), r1.z, l(0)
    iadd r1.y, r1.y, r1.z
    bfi r1.x, l(6), l(0), r2.x, r1.y
  endif 
else 
  ishl r0.x, r0.x, l(1)
  iadd r1.y, CB0[0][1].y, l(31)
  and r1.y, r1.y, l(-32)
  imad r1.y, vThreadID.z, r1.y, vThreadID.y
  imad r1.x, r1.y, CB0[0][0].y, r0.x
endif 
iadd r0.x, r1.x, CB0[0][0].x
imul null, r0.x, r0.x, l(9)
ushr r0.xy, r0.xyxx, l(3, 3, 0, 0)
ushr r1.x, CB0[0][1].w, l(1)
ld r1.yz, r0.xxxx, T0[0].zxyw
ieq r1.x, r1.x, l(1)
if_nz r1.x
  ishl r2.xy, r1.yzyy, l(8, 8, 0, 0)
  ushr r2.zw, r1.yyyz, l(0, 0, 8, 8)
  and r2.xyzw, r2.xyzw, l(0xff00ff00, 0xff00ff00, 0x00ff00ff, 0x00ff00ff)
  iadd r1.yz, r2.zzwz, r2.xxyx
endif 
and r2.xy, r1.yzyy, l(0xf0f0f0f0, 0xf0f0f0f0, 0, 0)
ishl r2.zw, r1.yyyz, l(0, 0, 8, 8)
and r2.zw, r2.zzzw, l(0, 0, 0x0f000f00, 0x0f000f00)
iadd r2.xy, r2.zwzz, r2.xyxx
ushr r1.yz, r1.yyzy, l(0, 8, 8, 0)
and r1.yz, r1.yyzy, l(0, 0x000f000f, 0x000f000f, 0)
iadd r2.xy, r1.yzyy, r2.xyxx
iadd r3.xyzw, r0.xxxx, l(1, 2, 3, 4)
ld r1.yz, r3.xxxx, T0[0].zxyw
if_nz r1.x
  ishl r4.xy, r1.yzyy, l(8, 8, 0, 0)
  ushr r4.zw, r1.yyyz, l(0, 0, 8, 8)
  and r4.xyzw, r4.xyzw, l(0xff00ff00, 0xff00ff00, 0x00ff00ff, 0x00ff00ff)
  iadd r1.yz, r4.zzwz, r4.xxyx
endif 
and r4.xy, r1.yzyy, l(0xf0f0f0f0, 0xf0f0f0f0, 0, 0)
ishl r4.zw, r1.yyyz, l(0, 0, 8, 8)
and r4.zw, r4.zzzw, l(0, 0, 0x0f000f00, 0x0f000f00)
iadd r4.xy, r4.zwzz, r4.xyxx
ushr r1.yz, r1.yyzy, l(0, 8, 8, 0)
and r1.yz, r1.yyzy, l(0, 0x000f000f, 0x000f000f, 0)
iadd r4.xy, r1.yzyy, r4.xyxx
ld r1.yz, r3.yyyy, T0[0].zxyw
if_nz r1.x
  ishl r3.xy, r1.yzyy, l(8, 8, 0, 0)
  and r3.xy, r3.xyxx, l(0xff00ff00, 0xff00ff00, 0, 0)
  ushr r5.xy, r1.yzyy, l(8, 8, 0, 0)
  and r5.xy, r5.xyxx, l(0x00ff00ff, 0x00ff00ff, 0, 0)
  iadd r1.yz, r3.xxyx, r5.xxyx
endif 
and r3.xy, r1.yzyy, l(0xf0f0f0f0, 0xf0f0f0f0, 0, 0)
ishl r5.xy, r1.yzyy, l(8, 8, 0, 0)
and r5.xy, r5.xyxx, l(0x0f000f00, 0x0f000f00, 0, 0)
iadd r3.xy, r3.xyxx, r5.xyxx
ushr r1.yz, r1.yyzy, l(0, 8, 8, 0)
and r1.yz, r1.yyzy, l(0, 0x000f000f, 0x000f000f, 0)
iadd r5.xy, r1.yzyy, r3.xyxx
ld r1.yz, r3.zzzz, T0[0].zxyw
if_nz r1.x
  ishl r3.xy, r1.yzyy, l(8, 8, 0, 0)
  and r3.xy, r3.xyxx, l(0xff00ff00, 0xff00ff00, 0, 0)
  ushr r6.xy, r1.yzyy, l(8, 8, 0, 0)
  and r6.xy, r6.xyxx, l(0x00ff00ff, 0x00ff00ff, 0, 0)
  iadd r1.yz, r3.xxyx, r6.xxyx
endif 
and r3.xy, r1.yzyy, l(0xf0f0f0f0, 0xf0f0f0f0, 0, 0)
ishl r6.xy, r1.yzyy, l(8, 8, 0, 0)
and r6.xy, r6.xyxx, l(0x0f000f00, 0x0f000f00, 0, 0)
iadd r3.xy, r3.xyxx, r6.xyxx
ushr r1.yz, r1.yyzy, l(0, 8, 8, 0)
and r1.yz, r1.yyzy, l(0, 0x000f000f, 0x000f000f, 0)
iadd r1.yz, r1.yyzy, r3.xxyx
ld r3.xy, r3.wwww, T0[0].xyzw
if_nz r1.x
  ishl r3.zw, r3.xxxy, l(0, 0, 8, 8)
  and r3.zw, r3.zzzw, l(0, 0, 0xff00ff00, 0xff00ff00)
  ushr r6.xy, r3.xyxx, l(8, 8, 0, 0)
  and r6.xy, r6.xyxx, l(0x00ff00ff, 0x00ff00ff, 0, 0)
  iadd r3.xy, r3.zwzz, r6.xyxx
endif 
and r3.zw, r3.xxxy, l(0, 0, 0xf0f0f0f0, 0xf0f0f0f0)
ishl r6.xy, r3.xyxx, l(8, 8, 0, 0)
and r6.xy, r6.xyxx, l(0x0f000f00, 0x0f000f00, 0, 0)
iadd r3.zw, r3.zzzw, r6.xxxy
ushr r3.xy, r3.xyxx, l(8, 8, 0, 0)
and r3.xy, r3.xyxx, l(0x000f000f, 0x000f000f, 0, 0)
iadd r5.zw, r3.xxxy, r3.zzzw
iadd r3.xyzw, r0.xxxx, l(5, 6, 7, 8)
ld r6.xy, r3.xxxx, T0[0].xyzw
if_nz r1.x
  ishl r6.zw, r6.xxxy, l(0, 0, 8, 8)
  and r6.zw, r6.zzzw, l(0, 0, 0xff00ff00, 0xff00ff00)
  ushr r7.xy, r6.xyxx, l(8, 8, 0, 0)
  and r7.xy, r7.xyxx, l(0x00ff00ff, 0x00ff00ff, 0, 0)
  iadd r6.xy, r6.zwzz, r7.xyxx
endif 
and r6.zw, r6.xxxy, l(0, 0, 0xf0f0f0f0, 0xf0f0f0f0)
ishl r7.xy, r6.xyxx, l(8, 8, 0, 0)
and r7.xy, r7.xyxx, l(0x0f000f00, 0x0f000f00, 0, 0)
iadd r6.zw, r6.zzzw, r7.xxxy
ushr r6.xy, r6.xyxx, l(8, 8, 0, 0)
and r6.xy, r6.xyxx, l(0x000f000f, 0x000f000f, 0, 0)
iadd r6.xy, r6.xyxx, r6.zwzz
ld r3.xy, r3.yyyy, T0[0].xyzw
if_nz r1.x
  ishl r6.zw, r3.xxxy, l(0, 0, 8, 8)
  and r6.zw, r6.zzzw, l(0, 0, 0xff00ff00, 0xff00ff00)
  ushr r7.xy, r3.xyxx, l(8, 8, 0, 0)
  and r7.xy, r7.xyxx, l(0x00ff00ff, 0x00ff00ff, 0, 0)
  iadd r3.xy, r6.zwzz, r7.xyxx
endif 
and r6.zw, r3.xxxy, l(0, 0, 0xf0f0f0f0, 0xf0f0f0f0)
ishl r7.xy, r3.xyxx, l(8, 8, 0, 0)
and r7.xy, r7.xyxx, l(0x0f000f00, 0x0f000f00, 0, 0)
iadd r6.zw, r6.zzzw, r7.xxxy
ushr r3.xy, r3.xyxx, l(8, 8, 0, 0)
and r3.xy, r3.xyxx, l(0x000f000f, 0x000f000f, 0, 0)
iadd r7.xy, r3.xyxx, r6.zwzz
ld r3.xy, r3.zzzz, T0[0].xyzw
if_nz r1.x
  ishl r6.zw, r3.xxxy, l(0, 0, 8, 8)
  and r6.zw, r6.zzzw, l(0, 0, 0xff00ff00, 0xff00ff00)
  ushr r8.xy, r3.xyxx, l(8, 8, 0, 0)
  and r8.xy, r8.xyxx, l(0x00ff00ff, 0x00ff00ff, 0, 0)
  iadd r3.xy, r6.zwzz, r8.xyxx
endif 
and r6.zw, r3.xxxy, l(0, 0, 0xf0f0f0f0, 0xf0f0f0f0)
ishl r8.xy, r3.xyxx, l(8, 8, 0, 0)
and r8.xy, r8.xyxx, l(0x0f000f00, 0x0f000f00, 0, 0)
iadd r6.zw, r6.zzzw, r8.xxxy
ushr r3.xy, r3.xyxx, l(8, 8, 0, 0)
and r3.xy, r3.xyxx, l(0x000f000f, 0x000f000f, 0, 0)
iadd r3.yz, r3.xxyx, r6.zzwz
ld r6.zw, r3.wwww, T0[0].zwxy
if_nz r1.x
  ishl r8.xy, r6.zwzz, l(8, 8, 0, 0)
  ushr r8.zw, r6.zzzw, l(0, 0, 8, 8)
  and r8.xyzw, r8.xyzw, l(0xff00ff00, 0xff00ff00, 0x00ff00ff, 0x00ff00ff)
  iadd r6.zw, r8.zzzw, r8.xxxy
endif 
and r8.xy, r6.zwzz, l(0xf0f0f0f0, 0xf0f0f0f0, 0, 0)
ishl r8.zw, r6.zzzw, l(0, 0, 8, 8)
and r8.zw, r8.zzzw, l(0, 0, 0x0f000f00, 0x0f000f00)
iadd r8.xy, r8.zwzz, r8.xyxx
ushr r6.zw, r6.zzzw, l(0, 0, 8, 8)
and r6.zw, r6.zzzw, l(0, 0, 0x000f000f, 0x000f000f)
iadd r8.yz, r6.zzwz, r8.xxyx
bfi r2.z, l(16), l(0), r2.y, r5.x
store_uav_typed U0[0].xyzw, r0.yyyy, r2.xzxx
iadd r9.xyzw, r0.yyyy, l(1, 2, 3, 4)
store_uav_typed U0[0].xyzw, r9.xxxx, r5.ywyy
bfi r3.x, l(16), l(0), r6.x, r7.y
store_uav_typed U0[0].xyzw, r9.yyyy, r3.xyxx
iadd r1.w, r0.z, r0.y
ushr r2.x, r2.y, l(16)
bfi r2.xzw, l(16, 0, 16, 16), l(16, 0, 16, 16), r4.xxxx, r2.xxxx
ushr r3.x, r4.x, l(16)
bfi r2.y, l(16), l(16), r1.y, r3.x
store_uav_typed U0[0].xyzw, r1.wwww, r2.xyzw
iadd r2.xy, r1.wwww, l(1, 2, 0, 0)
ushr r1.y, r1.y, l(16)
bfi r10.xzw, l(16, 0, 16, 16), l(16, 0, 16, 16), r1.zzzz, r1.yyyy
ushr r2.zw, r6.xxxy, l(0, 0, 16, 16)
bfi r10.y, l(16), l(16), r6.y, r2.z
store_uav_typed U0[0].xyzw, r2.xxxx, r10.xyzw
bfi r6.xzw, l(16, 0, 16, 16), l(16, 0, 16, 16), r3.zzzz, r2.wwww
ushr r1.y, r3.z, l(16)
bfi r6.y, l(16), l(16), r8.y, r1.y
store_uav_typed U0[0].xyzw, r2.yyyy, r6.xyzw
iadd r1.y, r0.z, r1.w
bfi r4.z, l(16), l(0), r5.x, r1.z
store_uav_typed U0[0].xyzw, r1.yyyy, r4.yzyy
iadd r1.zw, r1.yyyy, l(0, 0, 1, 2)
mov r7.z, r5.z
store_uav_typed U0[0].xyzw, r1.zzzz, r7.zxzz
bfi r8.x, l(16), l(0), r7.y, r8.y
store_uav_typed U0[0].xyzw, r1.wwww, r8.xzxx
iadd r2.xyzw, r0.xxxx, l(9, 10, 11, 12)
ld r1.yz, r2.xxxx, T0[0].zxyw
if_nz r1.x
  ishl r3.xy, r1.yzyy, l(8, 8, 0, 0)
  ushr r3.zw, r1.yyyz, l(0, 0, 8, 8)
  and r3.xyzw, r3.xyzw, l(0xff00ff00, 0xff00ff00, 0x00ff00ff, 0x00ff00ff)
  iadd r1.yz, r3.zzwz, r3.xxyx
endif 
and r3.xy, r1.yzyy, l(0xf0f0f0f0, 0xf0f0f0f0, 0, 0)
ishl r3.zw, r1.yyyz, l(0, 0, 8, 8)
and r3.zw, r3.zzzw, l(0, 0, 0x0f000f00, 0x0f000f00)
iadd r3.xy, r3.zwzz, r3.xyxx
ushr r1.yz, r1.yyzy, l(0, 8, 8, 0)
and r1.yz, r1.yyzy, l(0, 0x000f000f, 0x000f000f, 0)
iadd r3.xy, r1.yzyy, r3.xyxx
ld r1.yz, r2.yyyy, T0[0].zxyw
if_nz r1.x
  ishl r2.xy, r1.yzyy, l(8, 8, 0, 0)
  and r2.xy, r2.xyxx, l(0xff00ff00, 0xff00ff00, 0, 0)
  ushr r4.xy, r1.yzyy, l(8, 8, 0, 0)
  and r4.xy, r4.xyxx, l(0x00ff00ff, 0x00ff00ff, 0, 0)
  iadd r1.yz, r2.xxyx, r4.xxyx
endif 
and r2.xy, r1.yzyy, l(0xf0f0f0f0, 0xf0f0f0f0, 0, 0)
ishl r4.xy, r1.yzyy, l(8, 8, 0, 0)
and r4.xy, r4.xyxx, l(0x0f000f00, 0x0f000f00, 0, 0)
iadd r2.xy, r2.xyxx, r4.xyxx
ushr r1.yz, r1.yyzy, l(0, 8, 8, 0)
and r1.yz, r1.yyzy, l(0, 0x000f000f, 0x000f000f, 0)
iadd r4.xy, r1.yzyy, r2.xyxx
ld r1.yz, r2.zzzz, T0[0].zxyw
if_nz r1.x
  ishl r2.xy, r1.yzyy, l(8, 8, 0, 0)
  and r2.xy, r2.xyxx, l(0xff00ff00, 0xff00ff00, 0, 0)
  ushr r5.xy, r1.yzyy, l(8, 8, 0, 0)
  and r5.xy, r5.xyxx, l(0x00ff00ff, 0x00ff00ff, 0, 0)
  iadd r1.yz, r2.xxyx, r5.xxyx
endif 
and r2.xy, r1.yzyy, l(0xf0f0f0f0, 0xf0f0f0f0, 0, 0)
ishl r5.xy, r1.yzyy, l(8, 8, 0, 0)
and r5.xy, r5.xyxx, l(0x0f000f00, 0x0f000f00, 0, 0)
iadd r2.xy, r2.xyxx, r5.xyxx
ushr r1.yz, r1.yyzy, l(0, 8, 8, 0)
and r1.yz, r1.yyzy, l(0, 0x000f000f, 0x000f000f, 0)
iadd r5.xy, r1.yzyy, r2.xyxx
ld r1.yz, r2.wwww, T0[0].zxyw
if_nz r1.x
  ishl r2.xy, r1.yzyy, l(8, 8, 0, 0)
  ushr r2.zw, r1.yyyz, l(0, 0, 8, 8)
  and r2.xyzw, r2.xyzw, l(0xff00ff00, 0xff00ff00, 0x00ff00ff, 0x00ff00ff)
  iadd r1.yz, r2.zzwz, r2.xxyx
endif 
and r2.xy, r1.yzyy, l(0xf0f0f0f0, 0xf0f0f0f0, 0, 0)
ishl r2.zw, r1.yyyz, l(0, 0, 8, 8)
and r2.zw, r2.zzzw, l(0, 0, 0x0f000f00, 0x0f000f00)
iadd r2.xy, r2.zwzz, r2.xyxx
ushr r1.yz, r1.yyzy, l(0, 8, 8, 0)
and r1.yz, r1.yyzy, l(0, 0x000f000f, 0x000f000f, 0)
iadd r1.yz, r1.yyzy, r2.xxyx
iadd r2.xyzw, r0.xxxx, l(13, 14, 15, 16)
ld r6.xy, r2.xxxx, T0[0].xyzw
if_nz r1.x
  ishl r6.zw, r6.xxxy, l(0, 0, 8, 8)
  and r6.zw, r6.zzzw, l(0, 0, 0xff00ff00, 0xff00ff00)
  ushr r7.xy, r6.xyxx, l(8, 8, 0, 0)
  and r7.xy, r7.xyxx, l(0x00ff00ff, 0x00ff00ff, 0, 0)
  iadd r6.xy, r6.zwzz, r7.xyxx
endif 
and r6.zw, r6.xxxy, l(0, 0, 0xf0f0f0f0, 0xf0f0f0f0)
ishl r7.xy, r6.xyxx, l(8, 8, 0, 0)
and r7.xy, r7.xyxx, l(0x0f000f00, 0x0f000f00, 0, 0)
iadd r6.zw, r6.zzzw, r7.xxxy
ushr r6.xy, r6.xyxx, l(8, 8, 0, 0)
and r6.xy, r6.xyxx, l(0x000f000f, 0x000f000f, 0, 0)
iadd r5.zw, r6.xxxy, r6.zzzw
ld r2.xy, r2.yyyy, T0[0].xyzw
if_nz r1.x
  ishl r6.xy, r2.xyxx, l(8, 8, 0, 0)
  ushr r6.zw, r2.xxxy, l(0, 0, 8, 8)
  and r6.xyzw, r6.xyzw, l(0xff00ff00, 0xff00ff00, 0x00ff00ff, 0x00ff00ff)
  iadd r2.xy, r6.zwzz, r6.xyxx
endif 
and r6.xy, r2.xyxx, l(0xf0f0f0f0, 0xf0f0f0f0, 0, 0)
ishl r6.zw, r2.xxxy, l(0, 0, 8, 8)
and r6.zw, r6.zzzw, l(0, 0, 0x0f000f00, 0x0f000f00)
iadd r6.xy, r6.zwzz, r6.xyxx
ushr r2.xy, r2.xyxx, l(8, 8, 0, 0)
and r2.xy, r2.xyxx, l(0x000f000f, 0x000f000f, 0, 0)
iadd r2.xy, r2.xyxx, r6.xyxx
ld r6.xy, r2.zzzz, T0[0].xyzw
if_nz r1.x
  ishl r6.zw, r6.xxxy, l(0, 0, 8, 8)
  and r6.zw, r6.zzzw, l(0, 0, 0xff00ff00, 0xff00ff00)
  ushr r7.xy, r6.xyxx, l(8, 8, 0, 0)
  and r7.xy, r7.xyxx, l(0x00ff00ff, 0x00ff00ff, 0, 0)
  iadd r6.xy, r6.zwzz, r7.xyxx
endif 
and r6.zw, r6.xxxy, l(0, 0, 0xf0f0f0f0, 0xf0f0f0f0)
ishl r7.xy, r6.xyxx, l(8, 8, 0, 0)
and r7.xy, r7.xyxx, l(0x0f000f00, 0x0f000f00, 0, 0)
iadd r6.zw, r6.zzzw, r7.xxxy
ushr r6.xy, r6.xyxx, l(8, 8, 0, 0)
and r6.xy, r6.xyxx, l(0x000f000f, 0x000f000f, 0, 0)
iadd r6.xy, r6.xyxx, r6.zwzz
iadd r1.w, r0.x, l(17)
ld r2.zw, r2.wwww, T0[0].zwxy
if_nz r1.x
  ishl r7.xy, r2.zwzz, l(8, 8, 0, 0)
  ushr r7.zw, r2.zzzw, l(0, 0, 8, 8)
  and r7.xyzw, r7.xyzw, l(0xff00ff00, 0xff00ff00, 0x00ff00ff, 0x00ff00ff)
  iadd r2.zw, r7.zzzw, r7.xxxy
endif 
and r7.xy, r2.zwzz, l(0xf0f0f0f0, 0xf0f0f0f0, 0, 0)
ishl r7.zw, r2.zzzw, l(0, 0, 8, 8)
and r7.zw, r7.zzzw, l(0, 0, 0x0f000f00, 0x0f000f00)
iadd r7.xy, r7.zwzz, r7.xyxx
ushr r2.zw, r2.zzzw, l(0, 0, 8, 8)
and r2.zw, r2.zzzw, l(0, 0, 0x000f000f, 0x000f000f)
iadd r7.yz, r2.zzwz, r7.xxyx
ld r2.zw, r1.wwww, T0[0].zwxy
if_nz r1.x
  ishl r8.xy, r2.zwzz, l(8, 8, 0, 0)
  ushr r8.zw, r2.zzzw, l(0, 0, 8, 8)
  and r8.xyzw, r8.xyzw, l(0xff00ff00, 0xff00ff00, 0x00ff00ff, 0x00ff00ff)
  iadd r2.zw, r8.zzzw, r8.xxxy
endif 
and r8.xy, r2.zwzz, l(0xf0f0f0f0, 0xf0f0f0f0, 0, 0)
ishl r8.zw, r2.zzzw, l(0, 0, 8, 8)
and r8.zw, r8.zzzw, l(0, 0, 0x0f000f00, 0x0f000f00)
iadd r8.xy, r8.zwzz, r8.xyxx
ushr r2.zw, r2.zzzw, l(0, 0, 8, 8)
and r2.zw, r2.zzzw, l(0, 0, 0x000f000f, 0x000f000f)
iadd r8.yz, r2.zzwz, r8.xxyx
bfi r3.z, l(16), l(0), r3.y, r5.x
store_uav_typed U0[0].xyzw, r9.zzzz, r3.xzxx
store_uav_typed U0[0].xyzw, r9.wwww, r5.ywyy
iadd r10.xyzw, r0.yyyy, l(5, 6, 7, 8)
bfi r7.x, l(16), l(0), r2.x, r6.y
store_uav_typed U0[0].xyzw, r10.xxxx, r7.xyxx
iadd r1.w, r0.z, r9.z
ushr r2.z, r3.y, l(16)
bfi r3.xzw, l(16, 0, 16, 16), l(16, 0, 16, 16), r4.xxxx, r2.zzzz
ushr r2.z, r4.x, l(16)
bfi r3.y, l(16), l(16), r1.y, r2.z
store_uav_typed U0[0].xyzw, r1.wwww, r3.xyzw
iadd r2.zw, r1.wwww, l(0, 0, 1, 2)
ushr r1.y, r1.y, l(16)
bfi r3.xzw, l(16, 0, 16, 16), l(16, 0, 16, 16), r1.zzzz, r1.yyyy
ushr r4.xw, r2.xxxy, l(16, 0, 0, 16)
bfi r3.y, l(16), l(16), r2.y, r4.x
store_uav_typed U0[0].xyzw, r2.zzzz, r3.xyzw
bfi r3.xzw, l(16, 0, 16, 16), l(16, 0, 16, 16), r7.zzzz, r4.wwww
ushr r1.y, r7.z, l(16)
bfi r3.y, l(16), l(16), r8.y, r1.y
store_uav_typed U0[0].xyzw, r2.wwww, r3.xyzw
iadd r1.y, r0.z, r1.w
bfi r4.z, l(16), l(0), r5.x, r1.z
store_uav_typed U0[0].xyzw, r1.yyyy, r4.yzyy
iadd r1.zw, r1.yyyy, l(0, 0, 1, 2)
mov r6.z, r5.z
store_uav_typed U0[0].xyzw, r1.zzzz, r6.zxzz
bfi r8.x, l(16), l(0), r6.y, r8.y
store_uav_typed U0[0].xyzw, r1.wwww, r8.xzxx
if_nz r0.w
  iadd r0.w, r0.x, l(72)
else 
  iadd r0.w, r0.x, l(18)
endif 
ld r1.yz, r0.wwww, T0[0].zxyw
if_nz r1.x
  ishl r2.xy, r1.yzyy, l(8, 8, 0, 0)
  ushr r2.zw, r1.yyyz, l(0, 0, 8, 8)
  and r2.xyzw, r2.xyzw, l(0xff00ff00, 0xff00ff00, 0x00ff00ff, 0x00ff00ff)
  iadd r1.yz, r2.zzwz, r2.xxyx
endif 
and r2.xy, r1.yzyy, l(0xf0f0f0f0, 0xf0f0f0f0, 0, 0)
ishl r2.zw, r1.yyyz, l(0, 0, 8, 8)
and r2.zw, r2.zzzw, l(0, 0, 0x0f000f00, 0x0f000f00)
iadd r2.xy, r2.zwzz, r2.xyxx
ushr r1.yz, r1.yyzy, l(0, 8, 8, 0)
and r1.yz, r1.yyzy, l(0, 0x000f000f, 0x000f000f, 0)
iadd r2.xy, r1.yzyy, r2.xyxx
iadd r3.xyzw, r0.wwww, l(1, 2, 3, 4)
ld r1.yz, r3.xxxx, T0[0].zxyw
if_nz r1.x
  ishl r4.xy, r1.yzyy, l(8, 8, 0, 0)
  ushr r4.zw, r1.yyyz, l(0, 0, 8, 8)
  and r4.xyzw, r4.xyzw, l(0xff00ff00, 0xff00ff00, 0x00ff00ff, 0x00ff00ff)
  iadd r1.yz, r4.zzwz, r4.xxyx
endif 
and r4.xy, r1.yzyy, l(0xf0f0f0f0, 0xf0f0f0f0, 0, 0)
ishl r4.zw, r1.yyyz, l(0, 0, 8, 8)
and r4.zw, r4.zzzw, l(0, 0, 0x0f000f00, 0x0f000f00)
iadd r4.xy, r4.zwzz, r4.xyxx
ushr r1.yz, r1.yyzy, l(0, 8, 8, 0)
and r1.yz, r1.yyzy, l(0, 0x000f000f, 0x000f000f, 0)
iadd r4.xy, r1.yzyy, r4.xyxx
ld r1.yz, r3.yyyy, T0[0].zxyw
if_nz r1.x
  ishl r3.xy, r1.yzyy, l(8, 8, 0, 0)
  and r3.xy, r3.xyxx, l(0xff00ff00, 0xff00ff00, 0, 0)
  ushr r5.xy, r1.yzyy, l(8, 8, 0, 0)
  and r5.xy, r5.xyxx, l(0x00ff00ff, 0x00ff00ff, 0, 0)
  iadd r1.yz, r3.xxyx, r5.xxyx
endif 
and r3.xy, r1.yzyy, l(0xf0f0f0f0, 0xf0f0f0f0, 0, 0)
ishl r5.xy, r1.yzyy, l(8, 8, 0, 0)
and r5.xy, r5.xyxx, l(0x0f000f00, 0x0f000f00, 0, 0)
iadd r3.xy, r3.xyxx, r5.xyxx
ushr r1.yz, r1.yyzy, l(0, 8, 8, 0)
and r1.yz, r1.yyzy, l(0, 0x000f000f, 0x000f000f, 0)
iadd r5.xy, r1.yzyy, r3.xyxx
ld r1.yz, r3.zzzz, T0[0].zxyw
if_nz r1.x
  ishl r3.xy, r1.yzyy, l(8, 8, 0, 0)
  and r3.xy, r3.xyxx, l(0xff00ff00, 0xff00ff00, 0, 0)
  ushr r6.xy, r1.yzyy, l(8, 8, 0, 0)
  and r6.xy, r6.xyxx, l(0x00ff00ff, 0x00ff00ff, 0, 0)
  iadd r1.yz, r3.xxyx, r6.xxyx
endif 
and r3.xy, r1.yzyy, l(0xf0f0f0f0, 0xf0f0f0f0, 0, 0)
ishl r6.xy, r1.yzyy, l(8, 8, 0, 0)
and r6.xy, r6.xyxx, l(0x0f000f00, 0x0f000f00, 0, 0)
iadd r3.xy, r3.xyxx, r6.xyxx
ushr r1.yz, r1.yyzy, l(0, 8, 8, 0)
and r1.yz, r1.yyzy, l(0, 0x000f000f, 0x000f000f, 0)
iadd r1.yz, r1.yyzy, r3.xxyx
ld r3.xy, r3.wwww, T0[0].xyzw
if_nz r1.x
  ishl r3.zw, r3.xxxy, l(0, 0, 8, 8)
  and r3.zw, r3.zzzw, l(0, 0, 0xff00ff00, 0xff00ff00)
  ushr r6.xy, r3.xyxx, l(8, 8, 0, 0)
  and r6.xy, r6.xyxx, l(0x00ff00ff, 0x00ff00ff, 0, 0)
  iadd r3.xy, r3.zwzz, r6.xyxx
endif 
and r3.zw, r3.xxxy, l(0, 0, 0xf0f0f0f0, 0xf0f0f0f0)
ishl r6.xy, r3.xyxx, l(8, 8, 0, 0)
and r6.xy, r6.xyxx, l(0x0f000f00, 0x0f000f00, 0, 0)
iadd r3.zw, r3.zzzw, r6.xxxy
ushr r3.xy, r3.xyxx, l(8, 8, 0, 0)
and r3.xy, r3.xyxx, l(0x000f000f, 0x000f000f, 0, 0)
iadd r5.zw, r3.xxxy, r3.zzzw
iadd r3.xyzw, r0.wwww, l(5, 6, 7, 8)
ld r6.xy, r3.xxxx, T0[0].xyzw
if_nz r1.x
  ishl r6.zw, r6.xxxy, l(0, 0, 8, 8)
  and r6.zw, r6.zzzw, l(0, 0, 0xff00ff00, 0xff00ff00)
  ushr r7.xy, r6.xyxx, l(8, 8, 0, 0)
  and r7.xy, r7.xyxx, l(0x00ff00ff, 0x00ff00ff, 0, 0)
  iadd r6.xy, r6.zwzz, r7.xyxx
endif 
and r6.zw, r6.xxxy, l(0, 0, 0xf0f0f0f0, 0xf0f0f0f0)
ishl r7.xy, r6.xyxx, l(8, 8, 0, 0)
and r7.xy, r7.xyxx, l(0x0f000f00, 0x0f000f00, 0, 0)
iadd r6.zw, r6.zzzw, r7.xxxy
ushr r6.xy, r6.xyxx, l(8, 8, 0, 0)
and r6.xy, r6.xyxx, l(0x000f000f, 0x000f000f, 0, 0)
iadd r6.xy, r6.xyxx, r6.zwzz
ld r3.xy, r3.yyyy, T0[0].xyzw
if_nz r1.x
  ishl r6.zw, r3.xxxy, l(0, 0, 8, 8)
  and r6.zw, r6.zzzw, l(0, 0, 0xff00ff00, 0xff00ff00)
  ushr r7.xy, r3.xyxx, l(8, 8, 0, 0)
  and r7.xy, r7.xyxx, l(0x00ff00ff, 0x00ff00ff, 0, 0)
  iadd r3.xy, r6.zwzz, r7.xyxx
endif 
and r6.zw, r3.xxxy, l(0, 0, 0xf0f0f0f0, 0xf0f0f0f0)
ishl r7.xy, r3.xyxx, l(8, 8, 0, 0)
and r7.xy, r7.xyxx, l(0x0f000f00, 0x0f000f00, 0, 0)
iadd r6.zw, r6.zzzw, r7.xxxy
ushr r3.xy, r3.xyxx, l(8, 8, 0, 0)
and r3.xy, r3.xyxx, l(0x000f000f, 0x000f000f, 0, 0)
iadd r7.xy, r3.xyxx, r6.zwzz
ld r3.xy, r3.zzzz, T0[0].xyzw
if_nz r1.x
  ishl r6.zw, r3.xxxy, l(0, 0, 8, 8)
  and r6.zw, r6.zzzw, l(0, 0, 0xff00ff00, 0xff00ff00)
  ushr r8.xy, r3.xyxx, l(8, 8, 0, 0)
  and r8.xy, r8.xyxx, l(0x00ff00ff, 0x00ff00ff, 0, 0)
  iadd r3.xy, r6.zwzz, r8.xyxx
endif 
and r6.zw, r3.xxxy, l(0, 0, 0xf0f0f0f0, 0xf0f0f0f0)
ishl r8.xy, r3.xyxx, l(8, 8, 0, 0)
and r8.xy, r8.xyxx, l(0x0f000f00, 0x0f000f00, 0, 0)
iadd r6.zw, r6.zzzw, r8.xxxy
ushr r3.xy, r3.xyxx, l(8, 8, 0, 0)
and r3.xy, r3.xyxx, l(0x000f000f, 0x000f000f, 0, 0)
iadd r3.yz, r3.xxyx, r6.zzwz
ld r6.zw, r3.wwww, T0[0].zwxy
if_nz r1.x
  ishl r8.xy, r6.zwzz, l(8, 8, 0, 0)
  ushr r8.zw, r6.zzzw, l(0, 0, 8, 8)
  and r8.xyzw, r8.xyzw, l(0xff00ff00, 0xff00ff00, 0x00ff00ff, 0x00ff00ff)
  iadd r6.zw, r8.zzzw, r8.xxxy
endif 
and r8.xy, r6.zwzz, l(0xf0f0f0f0, 0xf0f0f0f0, 0, 0)
ishl r8.zw, r6.zzzw, l(0, 0, 8, 8)
and r8.zw, r8.zzzw, l(0, 0, 0x0f000f00, 0x0f000f00)
iadd r8.xy, r8.zwzz, r8.xyxx
ushr r6.zw, r6.zzzw, l(0, 0, 8, 8)
and r6.zw, r6.zzzw, l(0, 0, 0x000f000f, 0x000f000f)
iadd r8.yz, r6.zzwz, r8.xxyx
bfi r2.z, l(16), l(0), r2.y, r5.x
store_uav_typed U0[0].xyzw, r10.yyyy, r2.xzxx
store_uav_typed U0[0].xyzw, r10.zzzz, r5.ywyy
bfi r3.x, l(16), l(0), r6.x, r7.y
store_uav_typed U0[0].xyzw, r10.wwww, r3.xyxx
iadd r0.x, r0.z, r10.y
ushr r1.w, r2.y, l(16)
bfi r2.xzw, l(16, 0, 16, 16), l(16, 0, 16, 16), r4.xxxx, r1.wwww
ushr r1.w, r4.x, l(16)
bfi r2.y, l(16), l(16), r1.y, r1.w
store_uav_typed U0[0].xyzw, r0.xxxx, r2.xyzw
iadd r2.xy, r0.xxxx, l(1, 2, 0, 0)
ushr r1.y, r1.y, l(16)
bfi r9.xzw, l(16, 0, 16, 16), l(16, 0, 16, 16), r1.zzzz, r1.yyyy
ushr r1.yw, r6.xxxy, l(0, 16, 0, 16)
bfi r9.y, l(16), l(16), r6.y, r1.y
store_uav_typed U0[0].xyzw, r2.xxxx, r9.xyzw
bfi r6.xzw, l(16, 0, 16, 16), l(16, 0, 16, 16), r3.zzzz, r1.wwww
ushr r1.y, r3.z, l(16)
bfi r6.y, l(16), l(16), r8.y, r1.y
store_uav_typed U0[0].xyzw, r2.yyyy, r6.xyzw
iadd r0.x, r0.z, r0.x
bfi r4.z, l(16), l(0), r5.x, r1.z
store_uav_typed U0[0].xyzw, r0.xxxx, r4.yzyy
iadd r1.yz, r0.xxxx, l(0, 1, 2, 0)
mov r7.z, r5.z
store_uav_typed U0[0].xyzw, r1.yyyy, r7.zxzz
bfi r8.x, l(16), l(0), r7.y, r8.y
store_uav_typed U0[0].xyzw, r1.zzzz, r8.xzxx
iadd r1.yzw, r0.yyyy, l(0, 9, 10, 11)
iadd r2.xyzw, r0.wwww, l(9, 10, 11, 12)
ld r3.xy, r2.xxxx, T0[0].xyzw
if_nz r1.x
  ishl r3.zw, r3.xxxy, l(0, 0, 8, 8)
  and r3.zw, r3.zzzw, l(0, 0, 0xff00ff00, 0xff00ff00)
  ushr r4.xy, r3.xyxx, l(8, 8, 0, 0)
  and r4.xy, r4.xyxx, l(0x00ff00ff, 0x00ff00ff, 0, 0)
  iadd r3.xy, r3.zwzz, r4.xyxx
endif 
and r3.zw, r3.xxxy, l(0, 0, 0xf0f0f0f0, 0xf0f0f0f0)
ishl r4.xy, r3.xyxx, l(8, 8, 0, 0)
and r4.xy, r4.xyxx, l(0x0f000f00, 0x0f000f00, 0, 0)
iadd r3.zw, r3.zzzw, r4.xxxy
ushr r3.xy, r3.xyxx, l(8, 8, 0, 0)
and r3.xy, r3.xyxx, l(0x000f000f, 0x000f000f, 0, 0)
iadd r3.xy, r3.xyxx, r3.zwzz
ld r2.xy, r2.yyyy, T0[0].xyzw
if_nz r1.x
  ishl r4.xy, r2.xyxx, l(8, 8, 0, 0)
  ushr r4.zw, r2.xxxy, l(0, 0, 8, 8)
  and r4.xyzw, r4.xyzw, l(0xff00ff00, 0xff00ff00, 0x00ff00ff, 0x00ff00ff)
  iadd r2.xy, r4.zwzz, r4.xyxx
endif 
and r4.xy, r2.xyxx, l(0xf0f0f0f0, 0xf0f0f0f0, 0, 0)
ishl r4.zw, r2.xxxy, l(0, 0, 8, 8)
and r4.zw, r4.zzzw, l(0, 0, 0x0f000f00, 0x0f000f00)
iadd r4.xy, r4.zwzz, r4.xyxx
ushr r2.xy, r2.xyxx, l(8, 8, 0, 0)
and r2.xy, r2.xyxx, l(0x000f000f, 0x000f000f, 0, 0)
iadd r4.xy, r2.xyxx, r4.xyxx
ld r2.xy, r2.zzzz, T0[0].xyzw
if_nz r1.x
  ishl r5.xy, r2.xyxx, l(8, 8, 0, 0)
  ushr r5.zw, r2.xxxy, l(0, 0, 8, 8)
  and r5.xyzw, r5.xyzw, l(0xff00ff00, 0xff00ff00, 0x00ff00ff, 0x00ff00ff)
  iadd r2.xy, r5.zwzz, r5.xyxx
endif 
and r5.xy, r2.xyxx, l(0xf0f0f0f0, 0xf0f0f0f0, 0, 0)
ishl r5.zw, r2.xxxy, l(0, 0, 8, 8)
and r5.zw, r5.zzzw, l(0, 0, 0x0f000f00, 0x0f000f00)
iadd r5.xy, r5.zwzz, r5.xyxx
ushr r2.xy, r2.xyxx, l(8, 8, 0, 0)
and r2.xy, r2.xyxx, l(0x000f000f, 0x000f000f, 0, 0)
iadd r5.xy, r2.xyxx, r5.xyxx
ld r2.xy, r2.wwww, T0[0].xyzw
if_nz r1.x
  ishl r2.zw, r2.xxxy, l(0, 0, 8, 8)
  and r2.zw, r2.zzzw, l(0, 0, 0xff00ff00, 0xff00ff00)
  ushr r6.xy, r2.xyxx, l(8, 8, 0, 0)
  and r6.xy, r6.xyxx, l(0x00ff00ff, 0x00ff00ff, 0, 0)
  iadd r2.xy, r2.zwzz, r6.xyxx
endif 
and r2.zw, r2.xxxy, l(0, 0, 0xf0f0f0f0, 0xf0f0f0f0)
ishl r6.xy, r2.xyxx, l(8, 8, 0, 0)
and r6.xy, r6.xyxx, l(0x0f000f00, 0x0f000f00, 0, 0)
iadd r2.zw, r2.zzzw, r6.xxxy
ushr r2.xy, r2.xyxx, l(8, 8, 0, 0)
and r2.xy, r2.xyxx, l(0x000f000f, 0x000f000f, 0, 0)
iadd r2.xy, r2.xyxx, r2.zwzz
iadd r6.xyzw, r0.wwww, l(13, 14, 15, 16)
ld r2.zw, r6.xxxx, T0[0].zwxy
if_nz r1.x
  ishl r7.xy, r2.zwzz, l(8, 8, 0, 0)
  ushr r7.zw, r2.zzzw, l(0, 0, 8, 8)
  and r7.xyzw, r7.xyzw, l(0xff00ff00, 0xff00ff00, 0x00ff00ff, 0x00ff00ff)
  iadd r2.zw, r7.zzzw, r7.xxxy
endif 
and r7.xy, r2.zwzz, l(0xf0f0f0f0, 0xf0f0f0f0, 0, 0)
ishl r7.zw, r2.zzzw, l(0, 0, 8, 8)
and r7.zw, r7.zzzw, l(0, 0, 0x0f000f00, 0x0f000f00)
iadd r7.xy, r7.zwzz, r7.xyxx
ushr r2.zw, r2.zzzw, l(0, 0, 8, 8)
and r2.zw, r2.zzzw, l(0, 0, 0x000f000f, 0x000f000f)
iadd r5.zw, r2.zzzw, r7.xxxy
ld r2.zw, r6.yyyy, T0[0].zwxy
if_nz r1.x
  ishl r6.xy, r2.zwzz, l(8, 8, 0, 0)
  and r6.xy, r6.xyxx, l(0xff00ff00, 0xff00ff00, 0, 0)
  ushr r7.xy, r2.zwzz, l(8, 8, 0, 0)
  and r7.xy, r7.xyxx, l(0x00ff00ff, 0x00ff00ff, 0, 0)
  iadd r2.zw, r6.xxxy, r7.xxxy
endif 
and r6.xy, r2.zwzz, l(0xf0f0f0f0, 0xf0f0f0f0, 0, 0)
ishl r7.xy, r2.zwzz, l(8, 8, 0, 0)
and r7.xy, r7.xyxx, l(0x0f000f00, 0x0f000f00, 0, 0)
iadd r6.xy, r6.xyxx, r7.xyxx
ushr r2.zw, r2.zzzw, l(0, 0, 8, 8)
and r2.zw, r2.zzzw, l(0, 0, 0x000f000f, 0x000f000f)
iadd r2.zw, r2.zzzw, r6.xxxy
ld r6.xy, r6.zzzz, T0[0].xyzw
if_nz r1.x
  ishl r7.xy, r6.xyxx, l(8, 8, 0, 0)
  ushr r7.zw, r6.xxxy, l(0, 0, 8, 8)
  and r7.xyzw, r7.xyzw, l(0xff00ff00, 0xff00ff00, 0x00ff00ff, 0x00ff00ff)
  iadd r6.xy, r7.zwzz, r7.xyxx
endif 
and r7.xy, r6.xyxx, l(0xf0f0f0f0, 0xf0f0f0f0, 0, 0)
ishl r7.zw, r6.xxxy, l(0, 0, 8, 8)
and r7.zw, r7.zzzw, l(0, 0, 0x0f000f00, 0x0f000f00)
iadd r7.xy, r7.zwzz, r7.xyxx
ushr r6.xy, r6.xyxx, l(8, 8, 0, 0)
and r6.xy, r6.xyxx, l(0x000f000f, 0x000f000f, 0, 0)
iadd r6.xy, r6.xyxx, r7.xyxx
iadd r0.x, r0.w, l(17)
ld r7.xy, r6.wwww, T0[0].xyzw
if_nz r1.x
  ishl r7.zw, r7.xxxy, l(0, 0, 8, 8)
  and r7.zw, r7.zzzw, l(0, 0, 0xff00ff00, 0xff00ff00)
  ushr r8.xy, r7.xyxx, l(8, 8, 0, 0)
  and r8.xy, r8.xyxx, l(0x00ff00ff, 0x00ff00ff, 0, 0)
  iadd r7.xy, r7.zwzz, r8.xyxx
endif 
and r7.zw, r7.xxxy, l(0, 0, 0xf0f0f0f0, 0xf0f0f0f0)
ishl r8.xy, r7.xyxx, l(8, 8, 0, 0)
and r8.xy, r8.xyxx, l(0x0f000f00, 0x0f000f00, 0, 0)
iadd r7.zw, r7.zzzw, r8.xxxy
ushr r7.xy, r7.xyxx, l(8, 8, 0, 0)
and r7.xy, r7.xyxx, l(0x000f000f, 0x000f000f, 0, 0)
iadd r7.yz, r7.xxyx, r7.zzwz
ld r0.xw, r0.xxxx, T0[0].xzwy
if_nz r1.x
  ishl r8.xy, r0.xwxx, l(8, 8, 0, 0)
  ushr r8.zw, r0.xxxw, l(0, 0, 8, 8)
  and r8.xyzw, r8.xyzw, l(0xff00ff00, 0xff00ff00, 0x00ff00ff, 0x00ff00ff)
  iadd r0.xw, r8.zzzw, r8.xxxy
endif 
and r8.xy, r0.xwxx, l(0xf0f0f0f0, 0xf0f0f0f0, 0, 0)
ishl r8.zw, r0.xxxw, l(0, 0, 8, 8)
and r8.zw, r8.zzzw, l(0, 0, 0x0f000f00, 0x0f000f00)
iadd r8.xy, r8.zwzz, r8.xyxx
ushr r0.xw, r0.xxxw, l(8, 0, 0, 8)
and r0.xw, r0.xxxw, l(0x000f000f, 0, 0, 0x000f000f)
iadd r8.yz, r0.xxwx, r8.xxyx
bfi r3.z, l(16), l(0), r3.y, r5.x
store_uav_typed U0[0].xyzw, r1.yyyy, r3.xzxx
store_uav_typed U0[0].xyzw, r1.zzzz, r5.ywyy
bfi r7.x, l(16), l(0), r2.z, r6.y
store_uav_typed U0[0].xyzw, r1.wwww, r7.xyxx
iadd r0.x, r0.z, r1.y
ushr r0.y, r3.y, l(16)
bfi r1.xzw, l(16, 0, 16, 16), l(16, 0, 16, 16), r4.xxxx, r0.yyyy
ushr r0.y, r4.x, l(16)
bfi r1.y, l(16), l(16), r2.x, r0.y
store_uav_typed U0[0].xyzw, r0.xxxx, r1.xyzw
iadd r0.yw, r0.xxxx, l(0, 1, 0, 2)
ushr r1.x, r2.x, l(16)
bfi r1.xzw, l(16, 0, 16, 16), l(16, 0, 16, 16), r2.yyyy, r1.xxxx
ushr r2.xz, r2.zzwz, l(16, 0, 16, 0)
bfi r1.y, l(16), l(16), r2.w, r2.x
store_uav_typed U0[0].xyzw, r0.yyyy, r1.xyzw
bfi r1.xzw, l(16, 0, 16, 16), l(16, 0, 16, 16), r7.zzzz, r2.zzzz
ushr r0.y, r7.z, l(16)
bfi r1.y, l(16), l(16), r8.y, r0.y
store_uav_typed U0[0].xyzw, r0.wwww, r1.xyzw
iadd r0.x, r0.z, r0.x
bfi r4.z, l(16), l(0), r5.x, r2.y
store_uav_typed U0[0].xyzw, r0.xxxx, r4.yzyy
iadd r0.yz, r0.xxxx, l(0, 1, 2, 0)
mov r6.z, r5.z
store_uav_typed U0[0].xyzw, r0.yyyy, r6.zxzz
bfi r8.x, l(16), l(0), r6.y, r8.y
store_uav_typed U0[0].xyzw, r0.zzzz, r8.xzxx
ret 
// Approximately 739 instruction slots used
