Updates to improve performance for loops where LLVM compiler needs aliasing information to eliminate loads.

Author: chriselrod

Manifest.toml

@@ -76,6 +76,6 @@ uuid = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 
 [[VectorizationBase]]
 deps = ["CpuId", "LinearAlgebra"]
-git-tree-sha1 = "1e8a90888ec61405ea345c1ac2bdc7d86b99bd69"
+git-tree-sha1 = "b68b3234127d7839280f39bd668fd0025633aa01"
 uuid = "3d5dd08c-fd9d-11e8-17fa-ed2836048c2f"
-version = "0.8.2"
+version = "0.8.5"

Project.toml

@@ -16,7 +16,7 @@ OffsetArrays = "1"
 Parameters = "0"
 SIMDPirates = "0.7"
 SLEEFPirates = "0.4"
-VectorizationBase = "0.8"
+VectorizationBase = "0.8.5"
 julia = "1.1"
 
 [extras]

benchmark/driver.jl

@@ -23,9 +23,9 @@ end
 # sizes = 23:23
 sizes = 256:-1:2
 
-filter2d_dynamic_bench = benchmark_filter2ddynamic(512:-1:2)
-filter2d_3x3_bench = benchmark_filter2d3x3(512:-1:2)
-filter2d_unrolled_bench = benchmark_filter2dunrolled(512:-1:2)
+filter2d_dynamic_bench = benchmark_filter2ddynamic(sizes)#512:-1:2)
+filter2d_3x3_bench = benchmark_filter2d3x3(sizes)#512:-1:2)
+filter2d_unrolled_bench = benchmark_filter2dunrolled(sizes)#512:-1:2)
 
 AmulB_bench = benchmark_AmulB(sizes)
 AmulBt_bench = benchmark_AmulBt(sizes)