CPUSummary 0.1.26 divides L3 cache size by the number of cores, so we…

… compensate here (#157)

.github/workflows/ci-julia-nightly.yml

@@ -19,7 +19,6 @@ jobs:
       matrix:
         arch:
           - x64
-          - x86
         os:
           - ubuntu-latest
           - windows-latest
@@ -29,9 +28,6 @@ jobs:
           - '3' # GitHub runners have 2 cores, so `NUM_CORES+1` is 3
         version:
           - 'nightly'
-        exclude:
-          - os: macOS-latest
-            arch: x86 # 32-bit Julia binaries are not available on macOS
     steps:
       - uses: actions/checkout@v2
       - uses: julia-actions/setup-julia@v1

.github/workflows/ci.yml

@@ -66,7 +66,6 @@ jobs:
       matrix:
         arch:
           - x64
-          - x86
         os:
           - ubuntu-latest
           - windows-latest
@@ -76,9 +75,6 @@ jobs:
           - '3' # GitHub runners have 2 cores, so `NUM_CORES+1` is 3
         version:
           - '1' # automatically expands to the latest stable 1.x release of Julia
-        exclude:
-          - os: macOS-latest
-            arch: x86 # 32-bit Julia binaries are not available on macOS
     steps:
       - uses: actions/checkout@v2
       - uses: julia-actions/setup-julia@v1

Project.toml

@@ -1,7 +1,7 @@
 name = "Octavian"
 uuid = "6fd5a793-0b7e-452c-907f-f8bfe9c57db4"
-authors = ["Mason Protter", "Chris Elrod", "Dilum Aluthge", "contributors"]
-version = "0.3.17"
+authors = ["Chris Elrod", "Dilum Aluthge", "Mason Protter", "contributors"]
+version = "0.3.18"
 
 [deps]
 ArrayInterface = "4fba245c-0d91-5ea0-9b3e-6abc04ee57a9"
@@ -17,13 +17,13 @@ VectorizationBase = "3d5dd08c-fd9d-11e8-17fa-ed2836048c2f"
 
 [compat]
 ArrayInterface = "3.1.14, 5.0.1, 6"
-CPUSummary = "0.1.1 - 0.1.8, 0.1.14 - 0.1.25"
+CPUSummary = "0.1.26"
 IfElse = "0.1"
 LoopVectorization = "0.12.86"
 ManualMemory = "0.1.1"
 PolyesterWeave = "0.1.1"
 Requires = "1"
-Static = "0.7.5"
+Static = "0.7.5, 0.8"
 ThreadingUtilities = "0.5"
 VectorizationBase = "0.21.15"
 julia = "1.6"

README.md

@@ -34,6 +34,8 @@ matrix multiplication on the CPU, built on top of
 Please see the
 [Octavian documentation](https://octavian.JuliaLinearAlgebra.org/stable).
 
+Octavian dropped 32bit Julia support. See [PR#157](https://github.com/JuliaLinearAlgebra/Octavian.jl/pull/157). If you're interested in restoring it, please file a PR to fix failing tests.
+
 ## Benchmarks
 
 You can run benchmarks using [BLASBenchmarksCPU.jl](https://github.com/JuliaLinearAlgebra/BLASBenchmarksCPU.jl):

benchmark/tilesearch.jl

@@ -66,8 +66,8 @@ end
 
 
 T = Float64
-min_size = round(Int, sqrt(0.65 * Octavian.VectorizationBase.cache_size(Val(3)) / sizeof(T)))
-max_size = round(Int, sqrt( 32  * Octavian.VectorizationBase.cache_size(Val(3)) / sizeof(T)))
+min_size = round(Int, sqrt((0.65/4) * Octavian.num_cores() * Octavian.VectorizationBase.cache_size(Val(3)) / sizeof(T)))
+max_size = round(Int, sqrt( (32/4)  * Octavian.num_cores() * Octavian.VectorizationBase.cache_size(Val(3)) / sizeof(T)))
 
 SR = size_range(max_size, min_size, 400);
 const CsConst, AsConst, BsConst = matrix_range(SR, T);
@@ -111,7 +111,7 @@ const CsConst, AsConst, BsConst = matrix_range(SR, T);
 
 
 function matmul_objective(params)
-    print("Params: ", params, "; ")
+    print("Params= ", params, "; ")
     W₁, W₂, R₁, R₂ = params
     gflop = bench_size(CsConst, AsConst, BsConst, Val{W₁}(), Val{W₂}(), Val{R₁}(), Val{R₂}())
     println(gflop)
@@ -125,12 +125,12 @@ lower = 0.75 .* init;
 upper = [0.9, 1.25init[2], 0.999, 0.999];
 # init = [0.001, 0.9754033943603924, 0.5711159869399494, 0.7547361860432168];
 
-#=
+
 opt = Optim.optimize(
     matmul_objective, init, ParticleSwarm(lower = lower, upper = upper),
-    Optim.Options(iterations = 10^6, time_limit = 8hours)
+    Optim.Options(iterations = 10^6, time_limit = 14*hours)
 );
-=#
+
 
 
 

src/Octavian.jl

@@ -6,7 +6,7 @@ using VectorizationBase, ArrayInterface, LoopVectorization
 
 using VectorizationBase: align, AbstractStridedPointer, zstridedpointer, vsub_nsw, assume,
   static_sizeof, StridedPointer, gesp, pause, pick_vector_width, has_feature
-using CPUSummary: cache_size, num_cores, cache_inclusive, cache_linesize
+using CPUSummary: cache_size, num_cores, num_threads, cache_inclusive, cache_linesize
 using LoopVectorization: preserve_buffer, CloseOpen, UpperBoundedInteger
 using ArrayInterface: size, strides, offsets, indices, axes, StrideIndex
 using IfElse: ifelse

src/global_constants.jl

@@ -64,7 +64,7 @@ _second_cache_size(scs::StaticInt, ::False) = scs
 _second_cache_size(::StaticInt{0}, ::Nothing) = StaticInt(3145728)
 function second_cache_size()
     sc = second_cache()
-    _second_cache_size(cache_size(sc), cache_inclusive(sc))
+    _second_cache_size(cache_size(sc), cache_inclusive(sc)) * min(num_cores(), num_threads())
 end
 
 first_cache_size(::Val{T}) where {T} = first_cache_size() ÷ static_sizeof(T)