Fix incorrect group/batch size computation for qp8 fully-connected …

…ops.

PiperOrigin-RevId: 707518051

src/subgraph/unary.c

@@ -149,13 +149,11 @@ static enum xnn_status reshape_convert_operator(
       break;
     }
     case xnn_operator_type_convert_nc_f32_qp8: {
-      const size_t num_groups =
-          xnn_shape_multiply_batch_dims(&input_value->shape, 2);
-      const size_t batch_size = input_value->shape.dim[num_input_dims - 2];
-      const size_t channels = input_value->shape.dim[num_input_dims - 1];
       status = xnn_reshape_convert_nc_f32_qp8(
-          opdata->operator_objects[0], num_groups, batch_size, channels,
-          /*input_stride=*/channels, threadpool);
+          opdata->operator_objects[0], /*num_groups=*/1,
+          /*batch_size=*/dq_batch_size,
+          /*channels=*/dq_channel_stride,
+          /*input_stride=*/dq_channel_stride, threadpool);
       break;
     }
     default:

src/tensor.c

@@ -613,11 +613,10 @@ size_t xnn_tensor_get_size(const struct xnn_value* value)
   // Special handling for packed quantized types.
   if (value->datatype == xnn_datatype_qpint8) {
     assert(value->gemm_config != NULL);
-    const size_t num_groups = xnn_shape_multiply_batch_dims(&value->shape, 2);
-    const size_t m = value->shape.dim[value->shape.num_dims - 2];
+    const size_t batch_dims = xnn_shape_multiply_batch_dims(&value->shape, 1);
     const size_t k = value->shape.dim[value->shape.num_dims - 1];
-    return num_groups *
-           xnn_x8_packq_f32qp8_gemm_packed_size(value->gemm_config, m, k);
+    return xnn_x8_packq_f32qp8_gemm_packed_size(value->gemm_config, batch_dims,
+                                                k);
   }
 
   uint64_t size_bits = xnn_datatype_size_bits(value->datatype);