stick to 256-bit ymm on skx

I haven't yet seen it slower, and it ought to avoid the most egregious
downclocking from zmm use.  Still gets to use all the fancy AVX-512 new
instructions, so it does seem to run faster than just targeting AVX2.

Change-Id: I5738bc28c490619e0303813d69e00c8b6d7bb184
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/274592
Auto-Submit: Mike Klein <mtklein@google.com>
Commit-Queue: Herb Derby <herb@google.com>
Reviewed-by: Herb Derby <herb@google.com>

Author: Mike Klein

src/core/SkVM.cpp

@@ -1901,8 +1901,8 @@ namespace skvm {
         auto mod = std::make_unique<llvm::Module>("", ctx);
         // All the scary bare pointers from here on are owned by ctx or mod, I think.
 
-        const int K = (true && SkCpu::Supports(SkCpu::SKX)) ? 16 :
-                      (true && SkCpu::Supports(SkCpu::HSW)) ?  8 : 4;
+        // Everything I've tested runs faster at K=8 (using ymm) than K=16 (zmm) on SKX machines.
+        const int K = (true && SkCpu::Supports(SkCpu::HSW)) ? 8 : 4;
 
         llvm::Type *ptr = llvm::Type::getInt8Ty(ctx)->getPointerTo(),
                    *i32 = llvm::Type::getInt32Ty(ctx);