xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxx
111
{SY}{SS}{JP}{NAME}{G}{L}{R}{W}
x1xxxx
0000
0001
{SY}{SS}{JP}{NAME}{TYPE}
xxxxxx
xx0
Short sleep
Long sleep, around 20x longer than short
Short/Long Sleep
TODO: how is it different from a bunch of nops?
{SY}{SS}{JP}{NAME}{DURATION}
xxxxxx
xxx
0010
Seem to be Instruction Cache Invalidate, supported by the fact
that it considerably slows shader execution compared to
data cache instructions.
{SY}{SS}{JP}{NAME}
xxxxxxxxxx
0011
Data (Cache?) Clean
0100
Data (Cache?) Invalidate
0101
Data (Cache?) Flush
0110
_Presumably_ invalidates workgroup-wide cache for image/buffer data access.
So while "fence" is enough to synchronize data access inside a workgroup,
for cross-workgroup synchronization we have to invalidate that cache.
{SY}{SS}{JP}{NAME}
x1xxxx
1010
0101
Are met at the end of compute shader:
(sy)(jp)lock;
unlock;
end ;
{SY}{SS}{JP}{NAME}
1000010000
0111
{SY}{SS}{JP}{NAME}
1001010000
0111
{SRC_TYPE} == 0 /* b16 */
h({IMMED})
{SRC_TYPE} == 1 /* b32 */
({IMMED})
{IMMED}
{HALF}{CONST}
({SRC_TYPE} == 0) /* b16 */
{HALF}{SRC}
({SRC_TYPE} == 0) /* b16 */
TODO: Yes, something is wrong here, needs to be tested.
These types are clearly used by the blob.
However, it seems that there may be f32/f16/i32/i16
types but they are interwind with the _scope_ bitfield.
TODO: Check if it does matter.
Add an entry to the scope-specific "alias table", when instruction
from that scope tries to access a source register it would search
its alias table first.
This allows to reduce the amount of data passed around when reading
immediates/constants and reduce register pressure. In addition,
the alias table could be populated in the preamble further reducing
the amount of instructions being run.
Used like this:
alias.tex.b32.1 r40.x, (-1.456763);
alias.tex.b32.0 r40.y, (0.056702);
gather4g.s2en.mode6.base0 (f32)(xyzw)r0.x, r40.x, 1;
Or this:
alias.tex.b32.0 r2.y, c1.w;
isam.s2en.mode6.base0.1d (f32)(xyzw)r46.z, r2.y, 0;
(sy)stib.f32.2d.4.mode4.base0 r46.z, r2.y, 1;
Notice the lack of nops between alias and the instruction
that uses it.
{SY}{SS}{JP}{NAME}.{SCOPE}.{SRC_TYPE}.{UNK} {DST}, {SRC}
{SRC_REG_TYPE} == 0
000000000000000000000000
{SRC_REG_TYPE} == 1
000000000000000000000
xx
1x
1000
111