Merge pull request #100 from microsoft/dev/ritdas/fix_strides

Fix strides of randomly generated data

hatlib/arg_value.py

@@ -145,7 +145,7 @@ def get_dimension_arg_indices(array_arg: ArgInfo, all_arguments: List[ArgInfo])
     return indices
 
 
-def _gen_random_data(dtype, shape):
+def _gen_random_data(dtype, shape, strides=None):
     dtype = np.uint16 if dtype == "bfloat16" else dtype
     if isinstance(dtype, np.dtype):
         dtype = dtype.type
@@ -157,7 +157,7 @@ def _gen_random_data(dtype, shape):
     else:
         data = np.random.random(tuple(shape)).astype(dtype)
 
-    return data
+    return np.lib.stride_tricks.as_strided(data, strides=strides) if strides is not None else data
 
 
 def generate_arg_values(arguments: List[ArgInfo], dim_names_to_values={}) -> List[ArgValue]:
@@ -214,7 +214,7 @@ def generate_arg_values(arguments: List[ArgInfo], dim_names_to_values={}) -> Lis
                     arg.numpy_strides = list(map(lambda x: x * arg.element_num_bytes, arg.shape[1:] + [1]))
 
             if arg.usage != hat_file.UsageType.Output:
-                arg_data = _gen_random_data(arg.numpy_dtype, arg.shape)
+                arg_data = _gen_random_data(arg.numpy_dtype, arg.shape, arg.numpy_strides)
                 values.append(ArgValue(arg, arg_data))
             else:
                 values.append(ArgValue(arg))

hatlib/benchmark_hat_package.py

@@ -184,7 +184,7 @@ class Benchmark:
             else:
                 print_verbose(verbose, "[Benchmarking] Benchmarking device that does not need cache flushing, skipping generation of multiple datasets")
 
-                input_sets = [generate_arg_sets_for_func(func, dyn_func_shape_fn=dyn_func_shape_fn)]
+                input_sets = generate_arg_sets_for_func(func, dyn_func_shape_fn=dyn_func_shape_fn)
 
             if input_data_process_fn:
                 input_sets = input_data_process_fn(input_sets)

hatlib/cuda_loader.py

@@ -205,7 +205,7 @@ class CudaCallableFunc(CallableFunc):
         cuda.cuCtxDestroy(self.context)
 
     def init_batch(self, benchmark: bool, warmup_iters=0, device_id: int = 0, args=[]):
-        self.func_info.verify(args[0] if benchmark else args)
+        self.func_info.verify(args)
         self.device_mem = allocate_cuda_mem(self.func_info.arguments)
 
         if not benchmark:

hatlib/rocm_loader.py

@@ -123,7 +123,7 @@ class RocmCallableFunc(CallableFunc):
         pass
 
     def init_batch(self, benchmark: bool, warmup_iters=0, device_id: int = 0, args=[]):
-        self.func_info.verify(args[0] if benchmark else args)
+        self.func_info.verify(args)
         self.device_mem = allocate_rocm_mem(benchmark, self.func_info.arguments, device_id)
 
         if not benchmark: