MLBEDSW-819: make int16 changes
Enabled int16 support quantization to match the reference.
Change-Id: Ib369640241a9a491f2b0bc52d7f6cb025e30344b
Signed-off-by: Fredrik Svedberg <fredrik.svedberg@arm.com>
diff --git a/ethosu/vela/register_command_stream_generator.py b/ethosu/vela/register_command_stream_generator.py
index faae2cf..5fa71aa 100644
--- a/ethosu/vela/register_command_stream_generator.py
+++ b/ethosu/vela/register_command_stream_generator.py
@@ -27,7 +27,7 @@
from .tensor import MemArea, TensorBlockTraversal
from .operation import NpuBlockType
from .numeric_util import quantise_float32, round_up, round_away_zero, round_up_to_int, clamp_sigmoid, clamp_tanh
-from .data_type import BaseType
+from .data_type import BaseType, DataType
import numpy as np
from .shared_buffer_allocation import SharedBufferAllocation
from .architecture_features import SharedBufferArea, SHRAMElements, ArchitectureFeatures
@@ -615,6 +615,9 @@
else: # Convolution
assert cmd.weight_tensor.block_traversal != TensorBlockTraversal.Default
+ # Reduced precision quantization and natural rounding used for int16
+ if cmd.ifm_tensor.dtype == DataType.int16:
+ rounding_mode = rounding.NATURAL
emit.cmd0_with_param(cmd0.NPU_SET_KERNEL_HEIGHT_M1, cmd.weight_tensor.shape[0] - 1)
emit.cmd0_with_param(cmd0.NPU_SET_KERNEL_WIDTH_M1, cmd.weight_tensor.shape[1] - 1)
if cmd.weight_tensor.block_traversal == TensorBlockTraversal.PartKernelFirst: