MLBEDSW-6909: Use int32 acc for the Mean op
Changed acc type from int16 to int32. This will solve
saturation problems and the constraint added in
commit "MLBEDSW-5029: Output diff for Mean op"
can be removed.
Signed-off-by: Johan Alfven <johan.alfven@arm.com>
Change-Id: I05ec8835b43313b1a264d61a2b147fa62da123fe
diff --git a/ethosu/vela/test/test_tflite_supported_operators.py b/ethosu/vela/test/test_tflite_supported_operators.py
index cc8b3d2..89c2799 100644
--- a/ethosu/vela/test/test_tflite_supported_operators.py
+++ b/ethosu/vela/test/test_tflite_supported_operators.py
@@ -623,22 +623,6 @@
op = create_mean([1, 16, 17, 16], [1, 1, 1, 16], [1, 2], DataType.int8, {"keep_dims": True})
assert not support.is_operator_supported(op)
- # Create OP that will not saturate the accumulator
- op = create_mean([1, 5, 14, 16], [1, 1, 1, 16], [1, 2], DataType.int8, {"keep_dims": True})
- op.ifm.quantization.scale_f32 = 2.0
- op.ifm.quantization.zero_point = 95
- op.ofm.quantization.scale_f32 = 1.0
- op.ofm.quantization.zero_point = 95
- assert support.is_operator_supported(op)
-
- # Create OP that can saturate the accumulator
- op = create_mean([1, 6, 14, 16], [1, 1, 1, 16], [1, 2], DataType.int8, {"keep_dims": True})
- op.ifm.quantization.scale_f32 = 2.0
- op.ifm.quantization.zero_point = 95
- op.ofm.quantization.scale_f32 = 1.0
- op.ofm.quantization.zero_point = 95
- assert not support.is_operator_supported(op)
-
def test_mean_hw_product_avgpool():
op = create_mean([1, 200, 200, 16], [1, 16], [1, 2], DataType.uint8, {"keep_dims": False})
diff --git a/ethosu/vela/tflite_graph_optimiser.py b/ethosu/vela/tflite_graph_optimiser.py
index aaa778e..0f199de 100644
--- a/ethosu/vela/tflite_graph_optimiser.py
+++ b/ethosu/vela/tflite_graph_optimiser.py
@@ -1476,9 +1476,10 @@
# followed by a multiplication with 1/N to get the MEAN
weight_scale = 1
intermediate = op.ofm.clone(suffix="_intermediate", set_unique=True)
- intermediate.dtype = DataType.int16
+ intermediate.dtype = DataType.int32
mul_op = Operation(Op.Mul, op.name + "_mul")
mul_op.add_input_tensor(intermediate)
+ mul_op.set_output_tensor(op.ofm)
# Create scalar containing 1/N
quant = QuantizationParameters()
quant.zero_point = 0
@@ -1492,11 +1493,23 @@
n = int(h * w)
eps = 1 / (256 * (n + 1)) if n % 2 == 0 else 0
quant.scale_f32 = 1 / (n - eps)
+
+ # For int8/int16 we could use IFM/OFM scaling to do the division
+ # intermediate * 1 -> scale > round and shift.
+ #
+ # For int32 scaling is not supported so instead multiply with the scale
+ # intermediate * scale -> round and shift.
+ #
+ # Calculate the scale and shift value. const Tensor must be created
+ # with correct quantization since the scale and shift is calculated later
+ # in the command stream generator.
+ mul_scale, _ = scaling.elementwise_mul_scale(
+ mul_op.ifm.quantization.scale_f32, quant.scale_f32, mul_op.ofm.quantization.scale_f32
+ )
scalar = create_const_tensor(
- op.name + "_scalar", [1, 1, 1, 1], DataType.uint8, [1], np.uint8, quantization=quant
+ op.name + "_scalar", [1, 1, 1, 1], DataType.int32, [mul_scale], np.int32, quantization=quant
)
mul_op.add_input_tensor(scalar)
- mul_op.set_output_tensor(op.ofm)
mul_op.set_ifm_ofm_shapes()
mul_op.rounding_mode = NpuRoundingMode.NATURAL
mul_op.activation = op.activation
diff --git a/ethosu/vela/tflite_supported_operators.py b/ethosu/vela/tflite_supported_operators.py
index f01a669..24cc26e 100644
--- a/ethosu/vela/tflite_supported_operators.py
+++ b/ethosu/vela/tflite_supported_operators.py
@@ -796,9 +796,10 @@
max_prod = cls.mean_kernel_product
return h * w <= max_prod, f"Product of height and width is {h * w}"
- @staticmethod
- def constraint_mean_height_width_product_int8(op):
- """Number of IFM height and width elements might cause accumulator saturation when;
+ @classmethod
+ @docstring_format_args([mean_kernel_product_int8])
+ def constraint_mean_height_width_product_int8(cls, op):
+ """Product of IFM height and width must be no greater than {} when:
The IFM shape has 4 dimensions; and
The axis indices specify reduction across 2 dimensions; and
The axis indices correspond to the width and height dimensions of the IFM; and
@@ -817,43 +818,8 @@
return True, ""
h = shape[-3]
w = shape[-2]
-
- ifmq, ofmq = op.ifm.quantization, op.ofm.quantization
-
- # Scale factor
- real_scale = ifmq.scale_f32 / ofmq.scale_f32
-
- # Min and max value
- ifm_min_val = np.iinfo(np.int8).min - ifmq.zero_point
- ifm_max_val = np.iinfo(np.int8).max - ifmq.zero_point
-
- # Accumulator limits
- min_acc_limit = np.iinfo(np.int16).min
- max_acc_limit = np.iinfo(np.int16).max
-
- # Theoretical max/min value that accumulator need to store
- min_acc_sum = h * w * ifm_min_val * real_scale + ofmq.zero_point
- max_acc_sum = h * w * ifm_max_val * real_scale + ofmq.zero_point
-
- # Max product of heigth and width that will not saturate the accumulator
- ifm_min_val = 1 if ifm_min_val == 0 else ifm_min_val
- ifm_max_val = 1 if ifm_max_val == 0 else ifm_max_val
- if max_acc_sum > abs(min_acc_sum):
- max_hw = int((max_acc_limit - ofmq.zero_point) / real_scale / ifm_max_val)
- else:
- max_hw = int((min_acc_limit - ofmq.zero_point) / real_scale / ifm_min_val)
-
- extra = []
-
- extra.append(f" Possible accumulator range is ({min_acc_sum} - {max_acc_sum})\n")
- extra.append(f" Maximum accumulator range is ({min_acc_limit} - {max_acc_limit})\n")
- extra.append(
- f" Based on the IFM and OFM quantization the IFM height and width must be no greater than {max_hw}"
- )
-
- extra = "".join(extra)
-
- return (min_acc_sum >= min_acc_limit and max_acc_sum <= max_acc_limit, f"\n{extra}")
+ max_prod = cls.mean_kernel_product_int8
+ return h * w <= max_prod, f"Product of height and width is {h * w}"
@classmethod
@docstring_format_args([filter_height_range[1], dilated_height_range[1]])