vendor agnostic

Project.toml

@@ -4,8 +4,8 @@ authors = ["Ben Arthur <[email protected]>"]
 version = "0.2.3"
 
 [deps]
-CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
+GPUArrays = "0c68f7d7-f131-5f86-a1c3-88cf8149b2d7"
+KernelAbstractions = "63c18a36-062a-441e-b654-da1e3ab1ce7c"
 
 [compat]
-CUDA = "3, 4"
 julia = "1.6"

bench/runbench.jl

@@ -5,6 +5,15 @@ macro belapsed_median(args...)
     esc(:(time(median(@benchmark $(args...))) / 1e9))
 end
 
+macro sync(ex)
+    quote
+        local ret = $(esc(ex))
+        KernelAbstractions.synchronize(CUDABackend())
+        ret
+
+    end
+end
+
 
 function doit(L,N)
     x2 = CuArray(rand(L,N));
@@ -14,8 +23,8 @@ function doit(L,N)
     o1 = CuArray(rand(N));
     o3 = CuArray(rand(1,1,N));
 
-    tbgemm = @belapsed_median CUDA.@sync batched_mul!($o3, batched_transpose($x3), $y3)
-    tbdot = @belapsed_median CUDA.@sync batched_dot!($o1, $x2, $y2)
+    tbgemm = @belapsed_median @sync batched_mul!($o3, batched_transpose($x3), $y3)
+    tbdot = @belapsed_median @sync batched_dot!($o1, $x2, $y2)
 
     CUDA.unsafe_free!.((x2, x3, y2, y3, o1, o3))
     CUDA.memory_status()
@@ -51,18 +60,18 @@ function doit(L,N)
     y2 = CuArray(rand(L,N));
     y3 = CuArray(rand(L,1,N));
 
-    tbgemm = @belapsed_median CUDA.@sync batched_mul!($y3, $A3, $x3)
+    tbgemm = @belapsed_median @sync batched_mul!($y3, $A3, $x3)
 
-    tbgemvn = @belapsed_median CUDA.@sync batched_gemv!('N', 1.0, $A3, $x2, 0.0, $y2)
-    tbgemvt = @belapsed_median CUDA.@sync batched_gemv!('T', 1.0, $A3, $x2, 0.0, $y2)
+    tbgemvn = @belapsed_median @sync batched_gemv!('N', 1.0, $A3, $x2, 0.0, $y2)
+    tbgemvt = @belapsed_median @sync batched_gemv!('T', 1.0, $A3, $x2, 0.0, $y2)
 
-    tbsymvu = @belapsed_median CUDA.@sync batched_symv!('U', 1.0, $A3, $x2, 0.0, $y2)
-    tbsymvl = @belapsed_median CUDA.@sync batched_symv!('L', 1.0, $A3, $x2, 0.0, $y2)
+    tbsymvu = @belapsed_median @sync batched_symv!('U', 1.0, $A3, $x2, 0.0, $y2)
+    tbsymvl = @belapsed_median @sync batched_symv!('L', 1.0, $A3, $x2, 0.0, $y2)
 
     AP = CuArray(hcat([SymmetricPacked(x, :U).tri for x in eachslice(_A, dims=3)]...));
-    tbspmvu = @belapsed_median CUDA.@sync batched_spmv!('U', 1.0, $AP, $x2, 0.0, $y2)
+    tbspmvu = @belapsed_median @sync batched_spmv!('U', 1.0, $AP, $x2, 0.0, $y2)
     AP = CuArray(hcat([SymmetricPacked(x, :L).tri for x in eachslice(_A, dims=3)]...));
-    tbspmvl = @belapsed_median CUDA.@sync batched_spmv!('L', 1.0, $AP, $x2, 0.0, $y2)
+    tbspmvl = @belapsed_median @sync batched_spmv!('L', 1.0, $AP, $x2, 0.0, $y2)
 
     CUDA.unsafe_free!.((A3, AP, x2, x3, y2, y3))
     CUDA.memory_status()
@@ -108,17 +117,17 @@ function doit(L,N)
     y2 = CuArray(rand(L,N));
     y3 = CuArray(rand(L,1,N));
 
-    tbgemm = @belapsed_median CUDA.@sync batched_mul!($A3, $x3, batched_transpose($x3), -1.0, 1.0)
+    tbgemm = @belapsed_median @sync batched_mul!($A3, $x3, batched_transpose($x3), -1.0, 1.0)
 
-    tbger = @belapsed_median CUDA.@sync batched_ger!(-1.0, $x2, $y2, $A3)
+    tbger = @belapsed_median @sync batched_ger!(-1.0, $x2, $y2, $A3)
 
-    tbsyru = @belapsed_median CUDA.@sync batched_syr!('U', -1.0, $x2, $A3)
-    tbsyrl = @belapsed_median CUDA.@sync batched_syr!('L', -1.0, $x2, $A3)
+    tbsyru = @belapsed_median @sync batched_syr!('U', -1.0, $x2, $A3)
+    tbsyrl = @belapsed_median @sync batched_syr!('L', -1.0, $x2, $A3)
 
     AP = CuArray(hcat([SymmetricPacked(x, :U).tri for x in eachslice(_A, dims=3)]...));
-    tbspru = @belapsed_median CUDA.@sync batched_spr!('U', -1.0, $x2, $AP)
+    tbspru = @belapsed_median @sync batched_spr!('U', -1.0, $x2, $AP)
     AP = CuArray(hcat([SymmetricPacked(x, :L).tri for x in eachslice(_A, dims=3)]...));
-    tbsprl = @belapsed_median CUDA.@sync batched_spr!('L', -1.0, $x2, $AP)
+    tbsprl = @belapsed_median @sync batched_spr!('L', -1.0, $x2, $AP)
 
     CUDA.unsafe_free!.((A3, AP, x2, x3, y2, y3))
     CUDA.memory_status()