[memory strides] update check functions

[Change]
- Add support for memory strides to the check functions.
- Check that memory strides provided are compatible with
  the given tensor shape.
- Check that memory strides are compatible with the kernel.
- Check that the inner most mem stride equals 1
- Check that the output tensor has enough memory allocated and
  take into consideration the memory strides provided in
  the output tensor.
- All memory strides set to zero is supported, which means
  that each kernel calculates its own memory strides
  such that all data is contiguous.

Signed-off-by: R. Hilkens <[email protected]>

Author: rhilkens

lib/src/kernels/common/mli_krn_common.h

@@ -150,10 +150,11 @@ static void __attribute__ ((always_inline))  basic_rnn_cell_prepare_and_run_fx(
     const MLI_PTR (io_T) in_ptr = (const MLI_PTR (io_T)) in->data;
     MLI_PTR (io_T) state_ptr = (MLI_PTR (io_T)) prev_out->data;
 
-    mli_tensor dense_out;
+    mli_tensor dense_out = { 0 };
     dense_out.data = (cfg->mode != RNN_BATCH_TO_LAST) ? out->data : cfg->ir_tsr->data;
     dense_out.capacity = (cfg->mode != RNN_BATCH_TO_LAST) ? out->capacity : cfg->ir_tsr->capacity;
     dense_out.shape[0] = out_elements;
+    dense_out.mem_stride[0] = 0;
     dense_out.rank = 1;
     dense_out.el_type = in->el_type;
     // 1sign and 3 integer bits for typical rnn nonlinearity (TANH/SIGM) is enough
@@ -269,6 +270,8 @@ static void __attribute__ ((always_inline)) lstm_cell_prepare_and_run_fx(
     ir_tensor->rank = bias->rank;
     ir_tensor->shape[0] = bias->shape[0];
     ir_tensor->shape[1] = bias->shape[1];
+    ir_tensor->mem_stride[0] = 0;
+    ir_tensor->mem_stride[1] = 0;
     ir_tensor->el_type = in->el_type;
     // 1sign and 3 integer bits for TANH/SIGM input is enough
     ir_tensor->el_params.fx.frac_bits = (sizeof(io_T) * 8) - 1 - 3;
@@ -278,9 +281,9 @@ static void __attribute__ ((always_inline)) lstm_cell_prepare_and_run_fx(
     const int in_to_state_dif = in->el_params.fx.frac_bits - prev_out->el_params.fx.frac_bits;
     const int dense_out_shift = mli_prv_calc_shift(prev_out, weights, ir_tensor);
 
-    // Paricular subtensors of intermediate tensor
-    mli_tensor in_gate, forget_gate, out_gate;  // Various gates to controll info flow
-    mli_tensor g_tsr;       // Information tensors
+    // Paricular subtensors of intermediate tensor (mli_tensor.mem_stride[] should be zero and cannot be left uninitialized)
+    mli_tensor in_gate = { 0 }, forget_gate = { 0 }, out_gate = { 0 }; // Various gates to controll info flow
+    mli_tensor g_tsr = { 0 }; // Information tensors
 
     // Init subtensors
     mli_point_to_subtsr_cfg iterator = {.start_coord = {0}, .coord_num=1, .first_out_dim_size=1};

lib/src/private/mli_prv_tensor.h

@@ -198,8 +198,10 @@ static inline mli_status __attribute__ ((always_inline)) mli_prv_copy_tensor_for
     if (check != MLI_STATUS_OK)
           return check;
 
-    for (int idx = 0; idx < src->rank; idx++)
+    for (int idx = 0; idx < src->rank; idx++) {
         dst->shape[idx] = src->shape[idx];
+        dst->mem_stride[idx] = src->mem_stride[idx];
+    }
 
     dst->rank = src->rank;
     dst->el_type = src->el_type;