Polyester -> PolyesterWeave; closes #113 (#114)

* Polyester -> PolyesterWeave; closes #113

* Bump version

Author: chriselrod

Project.toml

@@ -1,14 +1,14 @@
 name = "Octavian"
 uuid = "6fd5a793-0b7e-452c-907f-f8bfe9c57db4"
 authors = ["Mason Protter", "Chris Elrod", "Dilum Aluthge", "contributors"]
-version = "0.3.2"
+version = "0.3.3"
 
 [deps]
 ArrayInterface = "4fba245c-0d91-5ea0-9b3e-6abc04ee57a9"
 IfElse = "615f187c-cbe4-4ef1-ba3b-2fcf58d6d173"
 LoopVectorization = "bdcacae8-1622-11e9-2a5c-532679323890"
 ManualMemory = "d125e4d3-2237-4719-b19c-fa641b8a4667"
-Polyester = "f517fe37-dbe3-4b94-8317-1923a5111588"
+PolyesterWeave = "1d0040c9-8b98-4ee7-8388-3f51789ca0ad"
 Static = "aedffcd0-7271-4cad-89d0-dc628f76c6d3"
 ThreadingUtilities = "8290d209-cae3-49c0-8002-c8c24d57dab5"
 VectorizationBase = "3d5dd08c-fd9d-11e8-17fa-ed2836048c2f"
@@ -18,7 +18,7 @@ ArrayInterface = "3.1.14"
 IfElse = "0.1"
 LoopVectorization = "0.12.34"
 ManualMemory = "0.1.1"
-Polyester = "0.4"
+PolyesterWeave = "0.1"
 Static = "0.2, 0.3"
 ThreadingUtilities = "0.4.6"
 VectorizationBase = "0.21.5"

src/Octavian.jl

@@ -8,7 +8,7 @@ using VectorizationBase: align, AbstractStridedPointer, zstridedpointer, vsub_ns
 using LoopVectorization: preserve_buffer, CloseOpen, UpperBoundedInteger
 using ArrayInterface: size, strides, offsets, indices, axes, StrideIndex
 using IfElse: ifelse
-using Polyester
+using PolyesterWeave
 using Static: StaticInt, Zero, One, StaticBool, True, False, gt, eq, StaticFloat64,
     roundtostaticint, floortostaticint
 using ManualMemory: MemoryBuffer, load, store!

src/matmul.jl

@@ -308,20 +308,20 @@ function matmulsplitn!(C::AbstractStridedPointer{T}, A, B, α, β, ::StaticInt{M
   _nspawn = Mblocks * Nblocks
   Mbsize_Mrem, Mbsize_ = promote(Mbsize +     W, Mbsize)
   Nbsize_Nrem, Nbsize_ = promote(Nbsize + One(), Nbsize)
-  (tnum, tuu) = Polyester.initial_state(threads)
+  (tnum, tuu) = PolyesterWeave.initial_state(threads)
   let _A = A, _B = B, _C = C, n = 0, Nrc = Nblocks - Nrem, Mrc = Mblocks - Mrem, __Mblocks = Mblocks - One()
     while true
       nsize = ifelse(Nblocks > Nrc, Nbsize_Nrem, Nbsize_); Nblocks -= 1
       let _A = _A, _C = _C, __Mblocks = __Mblocks
         while __Mblocks != 0
           msize = ifelse(__Mblocks ≥ Mrc, Mbsize_Mrem, Mbsize_); __Mblocks -= 1
-          (tnum, tuu) = Polyester.iter(tnum, tuu)
+          (tnum, tuu) = PolyesterWeave.iter(tnum, tuu)
           launch_thread_mul!(_C, _A, _B, α, β, msize, K, nsize, tnum, Val{PACK}())
           _A = gesp(_A, (msize, Zero()))
           _C = gesp(_C, (msize, Zero()))
         end
         if Nblocks != 0
-          (tnum, tuu) = Polyester.iter(tnum, tuu)
+          (tnum, tuu) = PolyesterWeave.iter(tnum, tuu)
           launch_thread_mul!(_C, _A, _B, α, β, Mremfinal, K, nsize, tnum, Val{PACK}())
         else
           call_loopmul!(_C, _A, _B, α, β, Mremfinal, K, nsize, Val{PACK}())
@@ -360,8 +360,8 @@ 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(_nrequest % UInt32, Polyester.worker_pointer())
-  # _threads, _torelease = Polyester.request_threads(Threads.threadid()%UInt32, _nrequest)
+  threads, torelease = PolyesterWeave.__request_threads(_nrequest % UInt32, PolyesterWeave.worker_pointer())
+  # _threads, _torelease = PolyesterWeave.request_threads(Threads.threadid()%UInt32, _nrequest)
 
   nrequest = threads.i
   iszero(nrequest) && @goto SINGLETHREAD
@@ -388,7 +388,7 @@ function __matmul!(
   else # TODO: Allow splitting along `N` for `matmul_pack_A_and_B!`
     matmul_pack_A_and_B!(C, A, B, α, β, M, K, N, threads, W₁Default(), W₂Default(), R₁Default(), R₂Default())
   end
-  Polyester.free_threads!(torelease)
+  PolyesterWeave.free_threads!(torelease)
   nothing
 end
 
@@ -398,9 +398,9 @@ function waitonmultasks(threads, nthread)
   # for (_,tid) ∈ threads
   #   wait(tid)
   # end
-  (tnum, tuu) = Polyester.initial_state(threads)
+  (tnum, tuu) = PolyesterWeave.initial_state(threads)
   for _ ∈ CloseOpen(One(), nthread)
-    (tnum, tuu) = Polyester.iter(tnum, tuu)
+    (tnum, tuu) = PolyesterWeave.iter(tnum, tuu)
     wait(tnum)
   end
 end
@@ -423,13 +423,13 @@ function matmul_pack_A_and_B!(
     end
     Mblock_Mrem, Mblock_ = promote(Mbsize + W, Mbsize)
     u_to_spawn = _to_spawn % UInt
-    (tnum, tuu) = Polyester.initial_state(threads)
+    (tnum, tuu) = PolyesterWeave.initial_state(threads)
     bc = _use_bcache()
     bc_ptr = Base.unsafe_convert(typeof(pointer(C)), pointer(bc))
     last_id = _to_spawn - One()
     for m ∈ CloseOpen(last_id) # ...thus the fact that `CloseOpen()` iterates at least once is okay.
       Mblock = ifelse(m < Mrem, Mblock_Mrem, Mblock_)
-      (tnum, tuu) = Polyester.iter(tnum, tuu)
+      (tnum, tuu) = PolyesterWeave.iter(tnum, tuu)
       launch_thread_mul!(C, A, B, α, β, Mblock, K, N, p, bc_ptr, tnum, m % UInt, u_to_spawn, StaticFloat64{W₁}(),StaticFloat64{W₂}(),StaticFloat64{R₁}(),StaticFloat64{R₂}())
       A = gesp(A, (Mblock, Zero()))
       C = gesp(C, (Mblock, Zero()))

test/matmul_coverage.jl

@@ -8,13 +8,13 @@ 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((nspawn-1)%UInt32, Octavian.Polyester.worker_pointer())
+      threads, torelease = Octavian.PolyesterWeave.__request_threads((nspawn-1)%UInt32, Octavian.PolyesterWeave.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,
         Octavian.W₁Default(), Octavian.W₂Default(), Octavian.R₁Default(), Octavian.R₂Default()
       )
-      Octavian.Polyester.free_threads!(torelease)
+      Octavian.PolyesterWeave.free_threads!(torelease)
     else
       Octavian.matmul_st_pack_A_and_B!(
         zc, za, zb, Octavian.StaticInt{1}(), Octavian.StaticInt{0}(), M, K, N,