dgemmjit looks identical to regular MKL, suggesting (a) the JIT isn't working (compiler warnings about ignored preprocessor directives reinforce this), and (b) a way to benchmark both OpenBLAS and MKL.

Author: chriselrod

docs/src/assets/bench_AmulB_v1.svg

@@ -8,7 +8,7 @@ using VectorizationBase: REGISTER_SIZE, REGISTER_COUNT, extract_data, num_vector
     PackedStridedPointer, SparseStridedPointer, RowMajorStridedPointer, StaticStridedPointer, StaticStridedStruct
 using SIMDPirates: VECTOR_SYMBOLS, evadd, evmul, vrange, reduced_add, reduced_prod, reduce_to_add, reduce_to_prod,
     sizeequivalentfloat, sizeequivalentint, vadd!, vsub!, vmul!, vfdiv!, vfmadd!, vfnmadd!, vfmsub!, vfnmsub!,
-    vfmadd231, vfmsub231, vfnmadd231, vfnmsub231,
+    vfmadd231, vfmsub231, vfnmadd231, vfnmsub231, #prefetch,
     vmullog2, vmullog10, vdivlog2, vdivlog10, vmullog2add!, vmullog10add!, vdivlog2add!, vdivlog10add!, vfmaddaddone
 using Base.Broadcast: Broadcasted, DefaultArrayStyle
 using LinearAlgebra: Adjoint, Transpose

docs/src/assets/bench_AmulBt_v1.svg

@@ -194,8 +194,8 @@ const COST = Dict{Instruction,InstructionCost}(
     Instruction(:sincospi_fast) => InstructionCost(25,22.0,70.0,26),
     Instruction(:identity) => InstructionCost(0,0.0,0.0,0),
     Instruction(:adjoint) => InstructionCost(0,0.0,0.0,0),
-    Instruction(:transpose) => InstructionCost(0,0.0,0.0,0)
-    # Symbol("##CONSTANT##") => InstructionCost(0,0.0)
+    Instruction(:transpose) => InstructionCost(0,0.0,0.0,0),
+    Instruction(:prefetch) => InstructionCost(0,0.0,0.0,0)
 )
 
 # const KNOWNINSTRUCTIONS = keys(COST)