Polyester 0.4 (#112)

* hacky 0.4 support (support only up to 64 threads)

* VectorizationBase 0.21

Author: chriselrod

Project.toml

@@ -1,7 +1,7 @@
 name = "Octavian"
 uuid = "6fd5a793-0b7e-452c-907f-f8bfe9c57db4"
 authors = ["Mason Protter", "Chris Elrod", "Dilum Aluthge", "contributors"]
-version = "0.3.1"
+version = "0.3.2"
 
 [deps]
 ArrayInterface = "4fba245c-0d91-5ea0-9b3e-6abc04ee57a9"
@@ -18,10 +18,10 @@ ArrayInterface = "3.1.14"
 IfElse = "0.1"
 LoopVectorization = "0.12.34"
 ManualMemory = "0.1.1"
-Polyester = "0.3.5"
+Polyester = "0.4"
 Static = "0.2, 0.3"
 ThreadingUtilities = "0.4.6"
-VectorizationBase = "0.20.16"
+VectorizationBase = "0.21.5"
 julia = "1.6"
 
 [extras]

src/Octavian.jl

@@ -6,7 +6,7 @@ using VectorizationBase: align, AbstractStridedPointer, zstridedpointer, vsub_ns
     static_sizeof, StridedPointer, gesp, pause, pick_vector_width, has_feature,
     cache_size, num_cores, num_cores, cache_inclusive, cache_linesize
 using LoopVectorization: preserve_buffer, CloseOpen, UpperBoundedInteger
-using ArrayInterface: size, strides, offsets, indices, axes
+using ArrayInterface: size, strides, offsets, indices, axes, StrideIndex
 using IfElse: ifelse
 using Polyester
 using Static: StaticInt, Zero, One, StaticBool, True, False, gt, eq, StaticFloat64,

src/matmul.jl

@@ -360,7 +360,9 @@ function __matmul!(
     clamp(div_fast(M * N, StaticInt{256}() * W), 0, _nthread-1)
   end
   # nkern = cld_fast(M * N,  MᵣW * Nᵣ)
-  threads, torelease = Polyester.request_threads(Threads.threadid()%UInt32, _nrequest)
+  threads, torelease = Polyester.__request_threads(_nrequest % UInt32, Polyester.worker_pointer())
+  # _threads, _torelease = Polyester.request_threads(Threads.threadid()%UInt32, _nrequest)
+
   nrequest = threads.i
   iszero(nrequest) && @goto SINGLETHREAD
   nspawn = nrequest + 1

src/utils.jl

@@ -38,7 +38,8 @@ function default_stridedpointer_quote(::Type{T}, N, Ot) where {T}
   quote
     $(Expr(:meta,:inline))
     st = $st
-    StridedPointer{$T,$N,$C,$B,$R}(ptr, $xt, $o)
+    si = StrideIndex{$N,$R,$C}($xt, $o)
+    stridedpointer(ptr, si, StaticInt{$B}())
   end
 end
 
@@ -66,9 +67,10 @@ end
   Bn = Core.ifelse(B > 1, B+1, B)
   quote
     $(Expr(:meta,:inline))
-    x = $gf(sp, :strd)
+    x = strides(sp)
     x0 = $gf(x, 1, false)
-    StridedPointer{$T,$(N+1),$Cn,$Bn,$Rn}($gf(sp,:p), $xt, $ot)
+    si = StrideIndex{$(N+1),$Rn,$Cn}($xt, $ot)
+    stridedpointer($gf(sp,:p), si, StaticInt{$Bn}())
   end
 end
 @generated function droplastdim(sp::StridedPointer{T,N,C,B,R}) where {T,N,C,B,R}
@@ -85,9 +87,10 @@ end
   end
   quote
     $(Expr(:meta,:inline))
-    x = $gf(sp, :strd)
-    o = $gf(sp, :offsets)
-    StridedPointer{$T,$(N-1),$Cn,$Bn,$rt}($gf(sp,:p), $xt, $ot)
+    x = strides(sp)
+    o = offsets(sp)
+    si = StrideIndex{$(N-1),$rt,$Cn}($xt, $ot)
+    stridedpointer($gf(sp,:p), si, StaticInt{$Bn}())
   end
 end
 

test/matmul_coverage.jl

@@ -8,7 +8,7 @@ function matmul_pack_ab!(C, A, B)
   nspawn = min(Threads.nthreads(), Octavian.num_cores())
   GC.@preserve C A B begin
     if nspawn > 1
-      threads, torelease = Octavian.Polyester.request_threads(Threads.threadid(), nspawn-1)
+      threads, torelease = Octavian.Polyester.__request_threads((nspawn-1)%UInt32, Octavian.Polyester.worker_pointer())
       @assert threads.i < Threads.nthreads()
       Octavian.matmul_pack_A_and_B!(
         zc, za, zb, Octavian.StaticInt{1}(), Octavian.StaticInt{0}(), M, K, N, threads,