AuroraMiddleware/skia2
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

eliminate the need for a tmp ymm register

Browse Source 
Like any other Instruction the store*s are assigned
a destination register d, which doesn't really make
sense, but works perfectly as a temporary register.

This means store8 doesn't need to reserve xmm/ymm15
as a temporary... it already has one naturally.   As
you might expect, the examples we have so far assign
the consumed input x register as the d register, so
things that used to look like

    vpackusdw %ymm6 ,%ymm6 ,%ymm15
    vpermq    $0xd8 ,%ymm15,%ymm15
    vpackuswb %ymm15,%ymm15,%ymm15
    vmoq      %xmm15,(%rdx)

now look more like

    vpackusdw %ymm6,%ymm6,%ymm6
    vpermq    $0xd8,%ymm6,%ymm6
    vpackuswb %ymm6,%ymm6,%ymm6
    vmoq      %xmm6,(%rdx)

Should be no perf difference, just simplified register bookkeeping.

This may suggest splitting load8/store8 into finer instructions,
two to do the physical loads and stores, and two for the 8->32
and 32->8 widen and narrow?  On the other hand load8 really is
just one vpmovzxbd instruction, so it'd be a shame to split it.
I suspect this will become more clear as I add 16-bit support.

Change-Id: I7c2b4d6b1689d40b50382f65fc00c01c54529c8a
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/220543
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
This commit is contained in:
Mike Klein 2019-06-12 10:04:37 -05:00 committed by Skia Commit-Bot
parent 072e6fc374
commit d3cc16c8bb
1 changed files with 9 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
src/core/SkVM.cpp
View File

@ -419,9 +419,8 @@ namespace skvm {
// All 16 ymm registers are available as scratch.
Xbyak::Ymm r[] = {
ymm0, ymm1, ymm2 , ymm3 , ymm4 , ymm5 , ymm6 , ymm7 ,
ymm8, ymm9, ymm10, ymm11, ymm12, ymm13, ymm14,
}, tmp = ymm15;
Xbyak::Xmm tmplo = xmm15;
ymm8, ymm9, ymm10, ymm11, ymm12, ymm13, ymm14, ymm15,
};
#endif
// Label / 4-byte values we need to write after ret.
@ -441,10 +440,13 @@ namespace skvm {
z = inst.z;
switch (op) {
case Op::store8:
vpackusdw(tmp, r[x], r[x]); // pack 32-bit -> 16-bit
vpermq (tmp, tmp, 0xd8); // u64 tmp[0,1,2,3] = tmp[0,2,1,3]
vpackuswb(tmp, tmp, tmp); // pack 16-bit -> 8-bit
vmovq(ptr[arg[y.imm]], tmplo); // store low 8 bytes
// Like any other instruction, store8 has been assigned
// a "destination" register we can use as a temporary scratch.
vpackusdw(r[d], r[x], r[x]); // pack 32-bit -> 16-bit
vpermq (r[d], r[d], 0xd8); // u64 tmp[0,1,2,3] = tmp[0,2,1,3]
vpackuswb(r[d], r[d], r[d]); // pack 16-bit -> 8-bit
vmovq(ptr[arg[y.imm]], // store low 8 bytes
Xbyak::Xmm{r[d].getIdx()}); // (arg must be an xmm register)
break;
case Op::store32: vmovups(ptr[arg[y.imm]], r[x]); break;