@@ -68,7 +68,7 @@ void copy_sweep_grid_shape(nvbench::state &state)
6868 state.add_global_memory_writes <nvbench::int32_t >(num_values);
6969
7070 // Allocate device memory:
71- thrust::device_vector<nvbench::int32_t > in (num_values, 0 );
71+ thrust::device_vector<nvbench::int32_t > in (num_values, 1 );
7272 thrust::device_vector<nvbench::int32_t > out (num_values, 0 );
7373
7474 state.exec ([block_size,
@@ -102,7 +102,7 @@ void copy_type_sweep(nvbench::state &state, nvbench::type_list<ValueType>)
102102 state.add_global_memory_writes <ValueType>(num_values);
103103
104104 // Allocate device memory:
105- thrust::device_vector<ValueType> in (num_values, 0 );
105+ thrust::device_vector<ValueType> in (num_values, ValueType{ 17 } );
106106 thrust::device_vector<ValueType> out (num_values, 0 );
107107
108108 state.exec ([num_values,
@@ -173,18 +173,19 @@ NVBENCH_BENCH_TYPES(copy_type_conversion_sweep, NVBENCH_TYPE_AXES(ctcs_types, ct
173173 .set_type_axes_names({" In" , " Out" });
174174
175175// ==================================================================================
176- // Passing `type_list` of typenames, and `enum_type_list` to build cartesian product
176+ // Passing list of typenames and `enum_type_list` to build cartesian product
177177// of typenames and integral constants
178178
179- template <typename ValueT, uint BLOCK_DIM>
179+ template <typename ValueT, unsigned BLOCK_DIM>
180180void copy_type_and_block_size_sweep (nvbench::state &state,
181181 nvbench::type_list<ValueT, nvbench::enum_type<BLOCK_DIM>>)
182182{
183183 const std::size_t nelems = 256 * 1024 * 1024 / sizeof (ValueT);
184- thrust::device_vector<ValueT> inp (nelems);
185- thrust::device_vector<ValueT> out (nelems);
184+ ValueT fill_value{42 };
185+ thrust::device_vector<ValueT> inp (nelems, fill_value);
186+ thrust::device_vector<ValueT> out (nelems, ValueT{});
186187
187- const uint gridSize = (nelems + BLOCK_DIM - 1 ) / BLOCK_DIM ;
188+ const uint gridSize = cuda::ceil_div (nelems, BLOCK_DIM) ;
188189
189190 const ValueT *inp_p = thrust::raw_pointer_cast (inp.data ());
190191 ValueT *out_p = thrust::raw_pointer_cast (out.data ());
@@ -198,8 +199,6 @@ void copy_type_and_block_size_sweep(nvbench::state &state,
198199 });
199200}
200201
201- using types_list =
202- nvbench::type_list<nvbench::int8_t , nvbench::int16_t , nvbench::int32_t , nvbench::int64_t >;
203202using block_sizes = nvbench::enum_type_list<64u , 128u , 196u , 256u , 320u , 512u >;
204- NVBENCH_BENCH_TYPES (copy_type_and_block_size_sweep, NVBENCH_TYPE_AXES(types_list , block_sizes))
203+ NVBENCH_BENCH_TYPES (copy_type_and_block_size_sweep, NVBENCH_TYPE_AXES(ctcs_types , block_sizes))
205204 .set_type_axes_names({" Type" , " BlockSize" });
0 commit comments