MLBEDSW-1906: Extend IFM to full dimension
Extend IFM to full dimension for the performance
metrics calculation.
Change-Id: Iae923e37280ab0f22b7a272f28970973a5142534
Signed-off-by: Charles Xu <charles.xu@arm.com>
diff --git a/ethosu/vela/npu_performance.py b/ethosu/vela/npu_performance.py
index 32208c9..eda8e42 100644
--- a/ethosu/vela/npu_performance.py
+++ b/ethosu/vela/npu_performance.py
@@ -259,14 +259,14 @@
ifm_tensor, _, weight_tensor, ofm_tensor = ps.get_primary_op_ifm_ifm2_weights_ofm()
- npu_convolution_ops = set((NpuBlockType.ConvolutionMxN, NpuBlockType.ConvolutionDepthWise))
- if (npu_block_type == NpuBlockType.Pooling and len(ifm_tensor.shape) == 4) or (
- npu_block_type in npu_convolution_ops
- ):
+ if npu_block_type in set((NpuBlockType.ConvolutionMxN, NpuBlockType.ConvolutionDepthWise, NpuBlockType.Pooling)):
+ # extent the ifm to full dimension
+ ifm_tensor_brick_size = tuple(numeric_util.full_shape(4, list(ifm_tensor.brick_size), 1))
+ ifm_tensor_shape = numeric_util.full_shape(4, ifm_tensor.shape, 1)
+ ifm_tensor_bandwidth_shape = numeric_util.full_shape(4, ifm_tensor.bandwidth_shape, 1)
batch_size = ifm_tensor.shape[0]
- ifm_tensor_shape = list(ifm_tensor.shape)
- ifm_depth = ifm_tensor.bandwidth_shape[3]
+ ifm_depth = ifm_tensor_bandwidth_shape[3]
# add in padding
ifm_tensor_shape[1] += explicit_padding[0] + explicit_padding[2] # height += top and bottom
@@ -313,7 +313,7 @@
clamped_skirt[2] = min(clamped_skirt[2], sub_kernel_limits[0] - 1 - clamped_skirt[0])
clamped_skirt[3] = min(clamped_skirt[3], sub_kernel_limits[1] - 1 - clamped_skirt[1])
n_blocks, area, block_setup = get_n_blocks_and_area(
- ifm_tensor.brick_size,
+ ifm_tensor_brick_size,
ifm_tensor_shape[1:3],
skirt,
clamped_skirt,
diff --git a/ethosu/vela/numeric_util.py b/ethosu/vela/numeric_util.py
index 1722adc..9adf5ec 100644
--- a/ethosu/vela/numeric_util.py
+++ b/ethosu/vela/numeric_util.py
@@ -85,3 +85,6 @@
else:
y = 1 / (1 + math.exp(-x))
return y
+
+def full_shape(dim, shape, fill):
+ return ([fill] * (dim - len(shape))) + shape
\ No newline at end of file
diff --git a/ethosu/vela/register_command_stream_generator.py b/ethosu/vela/register_command_stream_generator.py
index 1489860..c0402b3 100644
--- a/ethosu/vela/register_command_stream_generator.py
+++ b/ethosu/vela/register_command_stream_generator.py
@@ -46,6 +46,7 @@
from .numeric_util import round_away_zero
from .numeric_util import round_up
from .numeric_util import round_up_to_int
+from .numeric_util import full_shape
from .operation import NpuBlockType
from .shared_buffer_allocation import SharedBufferAllocation
from .tensor import MemArea
@@ -266,10 +267,6 @@
return Kernel(k_w, k_h, strides[2], strides[1], dilation[2], dilation[1])
-def full_shape(shape, fill):
- return ([fill] * (4 - len(shape))) + shape
-
-
def has_prev_op_dependency(prev_cmd, cmd):
if prev_cmd is None:
return False
@@ -282,14 +279,14 @@
def get_op_ofm_rect(cmd):
- start = full_shape(cmd.ofm_box.start_coord, 0)
- end = full_shape(cmd.ofm_box.end_coord, 1)
+ start = full_shape(4, cmd.ofm_box.start_coord, 0)
+ end = full_shape(4, cmd.ofm_box.end_coord, 1)
return Rect(start[-2], start[-3], start[-1], end[-2] - 1, end[-3] - 1, end[-1] - 1)
def get_op_ifm_rect(cmd):
- start = full_shape(cmd.ifm_box.start_coord, 0)
- end = full_shape(cmd.ifm_box.end_coord, 1)
+ start = full_shape(4, cmd.ifm_box.start_coord, 0)
+ end = full_shape(4, cmd.ifm_box.end_coord, 1)
return Rect(start[-2], start[-3], start[-1], end[-2] - 1, end[-3] - 1, end[-1] - 1)