gtr

Author: christiangnrd

src/ROCKernels.jl

@@ -140,7 +140,7 @@ function (obj::KI.Kernel{ROCBackend})(args...; numworkgroups = 1, workgroupsize
     KI.check_launch_args(numworkgroups, workgroupsize)
 
     groupsize = workgroupsize
-    gridsize = groupsize .* numworkgroups
+    gridsize = numworkgroups
 
     obj.kern(args...; groupsize, gridsize)
     return nothing

test/ka_tests.jl

@@ -159,13 +159,13 @@ function part_ki_testsuite(backend, AT)
                 return
             end
             arr1d = AT(zeros(Float32, 4))
-            @roc gridsize = 4 groupsize = 2 launch_kernel1d(arr1d)
+            @roc gridsize = 2 groupsize = 2 launch_kernel1d(arr1d)
             KernelAbstractions.synchronize(backend())
             @test all(Array(arr1d) .== 1)
 
             # 1d tuple
             arr1dt = AT(zeros(Float32, 4))
-            @roc gridsize = (4,) groupsize = (2,) launch_kernel1d(arr1dt)
+            @roc gridsize = (2,) groupsize = (2,) launch_kernel1d(arr1dt)
             KernelAbstractions.synchronize(backend())
             @test all(Array(arr1dt) .== 1)
 
@@ -179,7 +179,7 @@ function part_ki_testsuite(backend, AT)
                 return
             end
             arr2d = AT(zeros(Float32, 4, 4))
-            @roc gridsize = (4, 4) groupsize = (2, 2) launch_kernel2d(arr2d)
+            @roc gridsize = (2, 2) groupsize = (2, 2) launch_kernel2d(arr2d)
             KernelAbstractions.synchronize(backend())
             @test all(Array(arr2d) .== 1)
 
@@ -193,7 +193,7 @@ function part_ki_testsuite(backend, AT)
                 return
             end
             arr3d = AT(zeros(Float32, 4, 4, 4))
-            @roc gridsize = (4, 4, 4) groupsize = (2, 2, 2) launch_kernel3d(arr3d)
+            @roc gridsize = (2, 2, 2) groupsize = (2, 2, 2) launch_kernel3d(arr3d)
             KernelAbstractions.synchronize(backend())
             @test all(Array(arr3d) .== 1)
 
@@ -261,13 +261,13 @@ function no_ki_testsuite(backend, AT)
                 return
             end
             arr1d = AT(zeros(Float32, 4))
-            @roc gridsize = 4 groupsize = 2 launch_kernel1d(arr1d)
+            @roc gridsize = 2 groupsize = 2 launch_kernel1d(arr1d)
             KernelAbstractions.synchronize(backend())
             @test all(Array(arr1d) .== 1)
 
             # 1d tuple
             arr1dt = AT(zeros(Float32, 4))
-            @roc gridsize = (4,) groupsize = (2,) launch_kernel1d(arr1dt)
+            @roc gridsize = (2,) groupsize = (2,) launch_kernel1d(arr1dt)
             KernelAbstractions.synchronize(backend())
             @test all(Array(arr1dt) .== 1)
 
@@ -281,7 +281,7 @@ function no_ki_testsuite(backend, AT)
                 return
             end
             arr2d = AT(zeros(Float32, 4, 4))
-            @roc gridsize = (4, 4) groupsize = (2, 2) launch_kernel2d(arr2d)
+            @roc gridsize = (2, 2) groupsize = (2, 2) launch_kernel2d(arr2d)
             KernelAbstractions.synchronize(backend())
             @test all(Array(arr2d) .== 1)
 
@@ -295,7 +295,7 @@ function no_ki_testsuite(backend, AT)
                 return
             end
             arr3d = AT(zeros(Float32, 4, 4, 4))
-            @roc gridsize = (4, 4, 4) groupsize = (2, 2, 2) launch_kernel3d(arr3d)
+            @roc gridsize = (2, 2, 2) groupsize = (2, 2, 2) launch_kernel3d(arr3d)
             KernelAbstractions.synchronize(backend())
             @test all(Array(arr3d) .== 1)
 
@@ -317,7 +317,7 @@ function no_ki_testsuite(backend, AT)
             # @test KI.kernel_max_work_group_size(kernel) isa Int
             # @test KI.kernel_max_work_group_size(kernel; max_work_items = 1) == 1
 
-            kernel(results; groupsize = 4, gridsize = 16)
+            kernel(results; groupsize = 4, gridsize = 4)
             KernelAbstractions.synchronize(backend())
 
             host_results = Array(results)