Merge pull request #19 from JuliaGPU/vc/uniform

mark values used in localmem initialization as uniform

Author: anicusan

src/accumulate/accumulate_1d.jl

@@ -19,7 +19,7 @@ end
     # NOTE: shmem_size MUST be greater than 2 * block_size
     # NOTE: block_size MUST be a power of 2
     len = length(v)
-    block_size = @groupsize()[1]
+    @uniform block_size = @groupsize()[1]
     temp = @localmem eltype(v) (0x2 * block_size + conflict_free_offset(0x2 * block_size),)
 
     # NOTE: for many index calculations in this library, computation using zero-indexing leads to

src/accumulate/accumulate_nd.jl

@@ -73,7 +73,7 @@ end
     length_dims = vsizes[dims]
     length_outer = length(v) ÷ length_dims
 
-    block_size = @groupsize()[1]
+    @uniform block_size = @groupsize()[1]
 
     temp = @localmem eltype(v) (0x2 * block_size + conflict_free_offset(0x2 * block_size),)
     running_prefix = @localmem eltype(v) (1,)

src/reduce/mapreduce_1d.jl

@@ -1,6 +1,6 @@
 @kernel inbounds=true cpu=false function _mapreduce_block!(@Const(src), dst, f, op, init)
 
-    N = @groupsize()[1]
+    @uniform N = @groupsize()[1]
     sdata = @localmem eltype(dst) (N,)
 
     len = length(src)

src/reduce/mapreduce_nd.jl

@@ -78,7 +78,7 @@ end
 
     ndims = length(src_sizes)
 
-    N = @groupsize()[1]
+    @uniform N = @groupsize()[1]
     sdata = @localmem eltype(dst) (N,)
 
     # NOTE: for many index calculations in this library, computation using zero-indexing leads to

src/reduce/reduce_1d.jl

@@ -1,6 +1,6 @@
 @kernel inbounds=true cpu=false function _reduce_block!(@Const(src), dst, op, init)
 
-    N = @groupsize()[1]
+    @uniform N = @groupsize()[1]
     sdata = @localmem eltype(dst) (N,)
 
     len = length(src)

src/reduce/reduce_nd.jl

@@ -78,7 +78,7 @@ end
 
     ndims = length(src_sizes)
 
-    N = @groupsize()[1]
+    @uniform N = @groupsize()[1]
     sdata = @localmem eltype(dst) (N,)
 
     # NOTE: for many index calculations in this library, computation using zero-indexing leads to

src/sort/merge_sort.jl

@@ -1,6 +1,6 @@
 @kernel inbounds=true function _merge_sort_block!(vec, comp)
 
-    N = @groupsize()[1]
+    @uniform N = @groupsize()[1]
     s_buf = @localmem eltype(vec) (N * 0x2,)
 
     T = eltype(vec)

src/sort/merge_sort_by_key.jl

@@ -1,6 +1,6 @@
 @kernel inbounds=true function _merge_sort_by_key_block!(keys, values, comp)
 
-    N = @groupsize()[1]
+    @uniform N = @groupsize()[1]
     s_keys = @localmem eltype(keys) (N * 0x2,)
     s_values = @localmem eltype(values) (N * 0x2,)