def _create_npu_op_conv2d(
serial_2d_convolution: spec.Serial2DConvolution,
) -> Tuple[vapi.NpuConv2DOperation, int]:
"""This is a helper function to capture a list
of arguments to create Vela NpuConv2DOperation object.
"""
has_two_weights = serial_2d_convolution.weight2.address != -1
has_two_biases = serial_2d_convolution.scale_bias2.address != -1
npu_conv2d_op = vapi.NpuConv2DOperation()
npu_conv2d_op.ifm = _create_npu_feature_map(serial_2d_convolution.ifm)
npu_conv2d_op.ofm = _create_npu_feature_map(serial_2d_convolution.ofm)
npu_conv2d_op.kernel = _create_npu_kernel(serial_2d_convolution.kernel)
npu_conv2d_op.weights = ([
_create_npu_address_range(serial_2d_convolution.weight),
_create_npu_address_range(serial_2d_convolution.weight2),
] if has_two_weights else [
_create_npu_address_range(serial_2d_convolution.weight)
])
weights_zero_point = np.int64(
serial_2d_convolution.weight_zero_point.value)
npu_conv2d_op.biases = ([
_create_npu_address_range(serial_2d_convolution.scale_bias),
_create_npu_address_range(serial_2d_convolution.scale_bias2),
] if has_two_biases else [
_create_npu_address_range(serial_2d_convolution.scale_bias)
])
npu_conv2d_op.padding = _create_npu_padding(serial_2d_convolution.padding)
npu_conv2d_op.activation = _create_npu_activation(
serial_2d_convolution.activation)
if (npu_conv2d_op.activation and npu_conv2d_op.activation.op_type
== vapi.NpuActivationOp.NONE_OR_RELU):
_convert_clip_bounds(npu_conv2d_op)
npu_conv2d_op.rounding_mode = _create_npu_rounding_mode(
serial_2d_convolution.rounding_mode)
npu_conv2d_op.ifm_upscale = _create_npu_resampling_mode(
serial_2d_convolution.upscale)
weights_shape_ohwi = [
npu_conv2d_op.ofm.shape.depth,
npu_conv2d_op.kernel.height,
npu_conv2d_op.kernel.width,
npu_conv2d_op.ifm.shape.depth,
]
npu_conv2d_op.block_traversal = vela_api.calculate_block_traversal_mode(
is_depthwise=False,
weights_shape_ohwi=weights_shape_ohwi,
ifm_bitdepth=npu_conv2d_op.ifm.data_type.size_in_bits(),
)
npu_conv2d_op.block_config = _create_npu_block_config(
serial_2d_convolution.block_config)
if not npu_conv2d_op.block_config:
target_accel_config = vela_api.get_accelerator_config()
block_config = vela_api.get_optimal_block_config(
npu_conv2d_op, target_accel_config)
npu_conv2d_op.block_config = block_config
return npu_conv2d_op, weights_zero_point
def _create_npu_op_conv2d(serial_2d_convolution):
"""This is a helper function to capture a list
of arguments to create Vela NpuConv2DOperation object
"""
npu_conv2d_op = vapi.NpuConv2DOperation()
npu_conv2d_op.ifm = _create_npu_feature_map(serial_2d_convolution.ifm)
npu_conv2d_op.ofm = _create_npu_feature_map(serial_2d_convolution.ofm)
npu_conv2d_op.kernel = _create_npu_kernel(serial_2d_convolution.kernel)
npu_conv2d_op.weights = [
_create_npu_address_range(serial_2d_convolution.weight)
]
weights_zero_point = np.int64(
serial_2d_convolution.weight_zero_point.value)
npu_conv2d_op.biases = [
_create_npu_address_range(serial_2d_convolution.scale_bias)
]
npu_conv2d_op.padding = _create_npu_padding(serial_2d_convolution.padding)
npu_conv2d_op.activation = _create_npu_activation(
serial_2d_convolution.activation)
if (npu_conv2d_op.activation and npu_conv2d_op.activation.op_type
== vapi.NpuActivationOp.NONE_OR_RELU):
_convert_clip_bounds(npu_conv2d_op)
npu_conv2d_op.upscale = _create_npu_resampling_mode(
serial_2d_convolution.upscale)
target_accel_type = vela_api.get_target_accel_type() # type: ignore
block_config = vela_api.get_optimal_block_config(npu_conv2d_op,
target_accel_type)
npu_conv2d_op.block_config = block_config
weights_shape_ohwi = [
npu_conv2d_op.ofm.shape.depth,
npu_conv2d_op.kernel.height,
npu_conv2d_op.kernel.width,
npu_conv2d_op.ifm.shape.depth,
]
npu_conv2d_op.block_traversal = vela_api.calculate_block_traversal_mode(
is_depthwise=False,
weights_shape_ohwi=weights_shape_ohwi,
ifm_bitdepth=npu_conv2d_op.ifm.data_type.size_in_bits(),
)
return npu_conv2d_op, weights_zero_point