[ref model] Change RescaleOp attrs to inputs
This patch implements changes required for RescaleOp's
multiplier and shift changing from attributes to inputs
Signed-off-by: Tai Ly <tai.ly@arm.com>
Change-Id: I178919727e3220c749dad0ebce141e695868fee0
diff --git a/examples/test_add_1x4x4x4_f32/model.pb b/examples/test_add_1x4x4x4_f32/model.pb
index f1af89b..026dcea 100644
--- a/examples/test_add_1x4x4x4_f32/model.pb
+++ b/examples/test_add_1x4x4x4_f32/model.pb
@@ -92,5 +92,5 @@
}
}
versions {
- producer: 1597
+ producer: 1581
}
diff --git a/examples/test_conv2d_1x1_1x32x32x8_f32_st11_padSAME_dilat11/flatbuffer-tf/test_conv2d_1x1_1x32x32x8_f32_st11_padSAME_dilat11.tosa b/examples/test_conv2d_1x1_1x32x32x8_f32_st11_padSAME_dilat11/flatbuffer-tf/test_conv2d_1x1_1x32x32x8_f32_st11_padSAME_dilat11.tosa
index e3bc565..20e1333 100644
--- a/examples/test_conv2d_1x1_1x32x32x8_f32_st11_padSAME_dilat11/flatbuffer-tf/test_conv2d_1x1_1x32x32x8_f32_st11_padSAME_dilat11.tosa
+++ b/examples/test_conv2d_1x1_1x32x32x8_f32_st11_padSAME_dilat11/flatbuffer-tf/test_conv2d_1x1_1x32x32x8_f32_st11_padSAME_dilat11.tosa
Binary files differ
diff --git a/examples/test_conv2d_1x1_1x32x32x8_f32_st11_padSAME_dilat11/flatbuffer-tflite/test_conv2d_1x1_1x32x32x8_f32_st11_padSAME_dilat11.tosa b/examples/test_conv2d_1x1_1x32x32x8_f32_st11_padSAME_dilat11/flatbuffer-tflite/test_conv2d_1x1_1x32x32x8_f32_st11_padSAME_dilat11.tosa
index 90e8e86..d55d5d6 100644
--- a/examples/test_conv2d_1x1_1x32x32x8_f32_st11_padSAME_dilat11/flatbuffer-tflite/test_conv2d_1x1_1x32x32x8_f32_st11_padSAME_dilat11.tosa
+++ b/examples/test_conv2d_1x1_1x32x32x8_f32_st11_padSAME_dilat11/flatbuffer-tflite/test_conv2d_1x1_1x32x32x8_f32_st11_padSAME_dilat11.tosa
Binary files differ
diff --git a/examples/test_conv2d_1x1_1x32x32x8_f32_st11_padSAME_dilat11/model.pb b/examples/test_conv2d_1x1_1x32x32x8_f32_st11_padSAME_dilat11/model.pb
index bf7ee3a..ce59e78 100644
--- a/examples/test_conv2d_1x1_1x32x32x8_f32_st11_padSAME_dilat11/model.pb
+++ b/examples/test_conv2d_1x1_1x32x32x8_f32_st11_padSAME_dilat11/model.pb
@@ -137,5 +137,5 @@
}
}
versions {
- producer: 1597
+ producer: 1581
}
diff --git a/examples/test_conv2d_1x1_1x32x32x8_qi8_st11_padSAME_dilat11/flatbuffer-tflite/test_conv2d_1x1_1x32x32x8_qi8_st11_padSAME_dilat11.tosa b/examples/test_conv2d_1x1_1x32x32x8_qi8_st11_padSAME_dilat11/flatbuffer-tflite/test_conv2d_1x1_1x32x32x8_qi8_st11_padSAME_dilat11.tosa
index e9fb643..87bafd1 100644
--- a/examples/test_conv2d_1x1_1x32x32x8_qi8_st11_padSAME_dilat11/flatbuffer-tflite/test_conv2d_1x1_1x32x32x8_qi8_st11_padSAME_dilat11.tosa
+++ b/examples/test_conv2d_1x1_1x32x32x8_qi8_st11_padSAME_dilat11/flatbuffer-tflite/test_conv2d_1x1_1x32x32x8_qi8_st11_padSAME_dilat11.tosa
Binary files differ
diff --git a/examples/test_lstm_stateful_13x21x3_f32/flatbuffer-tflite/test_lstm_stateful_13x21x3_f32.tosa b/examples/test_lstm_stateful_13x21x3_f32/flatbuffer-tflite/test_lstm_stateful_13x21x3_f32.tosa
index f9606a1..deaca6e 100644
--- a/examples/test_lstm_stateful_13x21x3_f32/flatbuffer-tflite/test_lstm_stateful_13x21x3_f32.tosa
+++ b/examples/test_lstm_stateful_13x21x3_f32/flatbuffer-tflite/test_lstm_stateful_13x21x3_f32.tosa
Binary files differ
diff --git a/reference_model/src/ops/type_conversion.cc b/reference_model/src/ops/type_conversion.cc
index 5dbc7bd..d58cfeb 100644
--- a/reference_model/src/ops/type_conversion.cc
+++ b/reference_model/src/ops/type_conversion.cc
@@ -35,7 +35,7 @@
OpRescale<Rank, InDtype, OutDtype>::OpRescale(SubgraphTraverser* sgt_, TosaAttributeBase* attribute_, uint64_t id_)
: GraphNode(sgt_, Op_RESCALE, id_)
{
- setRequiredOperands(1, 1);
+ setRequiredOperands(3, 1);
INIT_ATTRIBUTE(Rescale);
}
@@ -68,6 +68,20 @@
ASSERT_MEM(in && out);
+ multiplierI32 = dynamic_cast<TosaReference::TensorTemplate<TMultiplierI32>*>(inputs[1]);
+ multiplierI16 = dynamic_cast<TosaReference::TensorTemplate<TMultiplierI16>*>(inputs[1]);
+ shift = dynamic_cast<TosaReference::TensorTemplate<TShift>*>(inputs[2]);
+ ASSERT_MEM(shift);
+
+ if (attribute->scale32())
+ {
+ ASSERT_MEM(multiplierI32);
+ }
+ else
+ {
+ ASSERT_MEM(multiplierI16);
+ }
+
if ((InDtype != TOSA_REF_TYPE_INT8) && (InDtype != TOSA_REF_TYPE_UINT8) && (InDtype != TOSA_REF_TYPE_UINT16) &&
(attribute->input_zp() != 0))
{
@@ -124,15 +138,15 @@
template <int Rank, TOSA_REF_TYPE InDtype, TOSA_REF_TYPE OutDtype>
int OpRescale<Rank, InDtype, OutDtype>::eval()
{
- int32_t input_zp = attribute->input_zp();
- int32_t output_zp = attribute->output_zp();
- std::vector<int32_t> multiplier = attribute->multiplier();
- std::vector<int32_t> shift = attribute->shift();
- bool scale32 = attribute->scale32();
- bool double_round = attribute->double_round();
- bool per_channel = attribute->per_channel();
- bool input_unsigned = attribute->input_unsigned();
- bool output_unsigned = attribute->output_unsigned();
+ int32_t input_zp = attribute->input_zp();
+ int32_t output_zp = attribute->output_zp();
+ std::vector<int32_t> multiplier;
+ std::vector<int32_t> shift;
+ bool scale32 = attribute->scale32();
+ bool double_round = attribute->double_round();
+ bool per_channel = attribute->per_channel();
+ bool input_unsigned = attribute->input_unsigned();
+ bool output_unsigned = attribute->output_unsigned();
// reshape [d0, d1, ..., dn] into [d0 * d1 ..., dn]
Eigen::array<Eigen::Index, 2> shape_2d;
@@ -153,6 +167,28 @@
ETensor2<OutEigenType> output_2d(shape_2d);
+ if (scale32)
+ {
+ auto multiplier_val = this->multiplierI32->getTensor();
+ for (int i = 0; i < multiplier_val.size(); i++)
+ {
+ multiplier.push_back(static_cast<int32_t>(multiplier_val(i)));
+ }
+ }
+ else
+ {
+ auto multiplier_val = this->multiplierI16->getTensor();
+ for (int i = 0; i < multiplier_val.size(); i++)
+ {
+ multiplier.push_back(static_cast<int32_t>(multiplier_val(i)));
+ }
+ }
+ auto shift_val = this->shift->getTensor();
+ for (int i = 0; i < shift_val.size(); i++)
+ {
+ shift.push_back(static_cast<int32_t>(shift_val(i)));
+ }
+
if (per_channel)
{
ETensor2<InEigenType> curr_channel_slice_prescaled;
diff --git a/reference_model/src/ops/type_conversion.h b/reference_model/src/ops/type_conversion.h
index 75f244d..a06dccc 100644
--- a/reference_model/src/ops/type_conversion.h
+++ b/reference_model/src/ops/type_conversion.h
@@ -32,10 +32,16 @@
virtual int checkTensorAttributes() final;
virtual int eval() final;
- using InEigenType = typename GetEigenType<InDtype>::type;
- using OutEigenType = typename GetEigenType<OutDtype>::type;
- using TIn = Eigen::Tensor<InEigenType, Rank>;
- using TOut = Eigen::Tensor<OutEigenType, Rank>;
+ using InEigenType = typename GetEigenType<InDtype>::type;
+ using OutEigenType = typename GetEigenType<OutDtype>::type;
+ using TIn = Eigen::Tensor<InEigenType, Rank>;
+ using TOut = Eigen::Tensor<OutEigenType, Rank>;
+ using I8EigenType = typename GetEigenType<TOSA_REF_TYPE::TOSA_REF_TYPE_INT8>::type;
+ using I16EigenType = typename GetEigenType<TOSA_REF_TYPE::TOSA_REF_TYPE_INT16>::type;
+ using I32EigenType = typename GetEigenType<TOSA_REF_TYPE::TOSA_REF_TYPE_INT32>::type;
+ using TMultiplierI16 = Eigen::Tensor<I16EigenType, 1>;
+ using TMultiplierI32 = Eigen::Tensor<I32EigenType, 1>;
+ using TShift = Eigen::Tensor<I8EigenType, 1>;
static constexpr int32_t QMin = GetQMin<OutDtype>::value;
static constexpr int32_t QMax = GetQMax<OutDtype>::value;
@@ -44,6 +50,9 @@
TosaRescaleAttribute* attribute;
TosaReference::TensorTemplate<TIn>* in;
TosaReference::TensorTemplate<TOut>* out;
+ TosaReference::TensorTemplate<TMultiplierI16>* multiplierI16;
+ TosaReference::TensorTemplate<TMultiplierI32>* multiplierI32;
+ TosaReference::TensorTemplate<TShift>* shift;
};
template <TOSA_REF_TYPE InDtype, TOSA_REF_TYPE OutDtype>
diff --git a/verif/generator/tosa_arg_gen.py b/verif/generator/tosa_arg_gen.py
index 592c491..cbfffae 100644
--- a/verif/generator/tosa_arg_gen.py
+++ b/verif/generator/tosa_arg_gen.py
@@ -744,6 +744,10 @@
arr = np.int64(argsDict["fixed_data"][idx])
elif dtype == DType.INT8:
arr = np.int8(argsDict["fixed_data"][idx])
+ elif dtype == DType.INT16:
+ arr = np.int16(argsDict["fixed_data"][idx])
+ elif dtype == DType.INT32:
+ arr = np.int32(argsDict["fixed_data"][idx])
else:
assert False, "Unsupported fixed_data type"
else:
@@ -1060,6 +1064,26 @@
)
@staticmethod
+ def tvgRescale(testGen, opName, dtypeList, shapeList, argsDict, error_name=None):
+ scale32 = argsDict["scale"]
+ multiplier_arr = argsDict["multiplier"]
+ shift_arr = argsDict["shift"]
+
+ if scale32:
+ dtypeList[1] = DType.INT32
+ else:
+ dtypeList[1] = DType.INT16
+ shapeList[1] = [len(multiplier_arr)]
+ dtypeList[2] = DType.INT8
+ shapeList[2] = [len(shift_arr)]
+ # Create a new list for the pre-generated data in argsDict["fixed_data"]
+ argsDict["fixed_data"] = [None, multiplier_arr, shift_arr]
+
+ return TosaTensorValuesGen.tvgLazyGenDefault(
+ testGen, opName, dtypeList, shapeList, argsDict, error_name
+ )
+
+ @staticmethod
def tvgPad(testGen, opName, dtypeList, shapeList, argsDict, error_name=None):
# argsDict["pad"] is 2D array, need to flatten it to get list of values
pad_values = argsDict["pad"].flatten()
@@ -2842,6 +2866,43 @@
# Illegal condition. ERROR_IF(!scale32 && double_round)
continue
+ if per_channel:
+ nc = shapeList[0][-1]
+ else:
+ nc = 1
+
+ in_type_width = gtu.dtypeWidth(inDtype)
+ out_type_width = gtu.dtypeWidth(outDtype)
+
+ # Calculate scale based on:
+ # scale = a *(2^output_width)/(2^input_width))
+
+ a = np.float32(testGen.rng.random(size=[nc]))
+ scale_arr = a * np.float32(
+ (1 << out_type_width) / (1 << in_type_width)
+ )
+
+ if scale32:
+ # Cap the scaling at 2^31 - 1 for scale32
+ scale_arr = np.clip(
+ scale_arr, 1.0 / (1 << 31), (1 << 31) - 1
+ )
+ else:
+ # Cap the scaling at 2^15 - 1 for scale16
+ scale_arr = np.clip(scale_arr, 1.0 / (1 << 31), 32767.0)
+
+ # print('{} {} -> {}'.format(out_type_width, in_type_width, scale_arr))
+
+ multiplier_arr = np.int32(np.zeros(shape=[nc]))
+ shift_arr = np.int32(np.zeros(shape=[nc]))
+ for i in range(nc):
+ (
+ multiplier_arr[i],
+ shift_arr[i],
+ ) = TosaQuantGen.computeMultiplierAndShift(
+ scale_arr[i], scale32
+ )
+
arg_list.append(
(
"out{}_sc{}_dr{}_pc{}".format(
@@ -2855,6 +2916,8 @@
"scale": scale32,
"double_round": double_round,
"per_channel": per_channel,
+ "multiplier": multiplier_arr,
+ "shift": shift_arr,
},
)
)
diff --git a/verif/generator/tosa_test_gen.py b/verif/generator/tosa_test_gen.py
index 978e735..415858c 100644
--- a/verif/generator/tosa_test_gen.py
+++ b/verif/generator/tosa_test_gen.py
@@ -317,13 +317,6 @@
"Unknown dtype, cannot convert to string: {}".format(dtype)
)
- def typeWidth(self, dtype):
- """Get the datatype width for data types"""
- if dtype in gtu.DTYPE_ATTRIBUTES:
- return gtu.DTYPE_ATTRIBUTES[dtype]["width"]
- else:
- raise Exception(f"Unknown dtype, cannot determine width: {dtype}")
-
def constrictBatchSize(self, shape):
# Limit the batch size unless an explicit target shape set
if self.args.max_batch_size and not self.args.target_shapes:
@@ -2130,12 +2123,15 @@
error_name=None,
qinfo=None,
):
- assert len(inputs) == 1
+ assert len(inputs) == 3
val = inputs[0]
+ multiplier_val = inputs[1]
+ shift_val = inputs[2]
out_dtype = args_dict["output_dtype"]
scale32 = args_dict["scale"]
double_round = args_dict["double_round"]
per_channel = args_dict["per_channel"]
+ shift_arr = args_dict["shift"]
result_tensor = OutputShaper.typeConversionOp(
self.ser, self.rng, val, out_dtype, error_name
@@ -2146,8 +2142,8 @@
else:
nc = 1
- in_type_width = self.typeWidth(val.dtype)
- out_type_width = self.typeWidth(out_dtype)
+ in_type_width = gtu.dtypeWidth(val.dtype)
+ out_type_width = gtu.dtypeWidth(out_dtype)
input_unsigned = False
output_unsigned = False
@@ -2198,31 +2194,10 @@
else:
output_zp = 0
- # Calculate scale based on:
- # scale = a *(2^output_width)/(2^input_width))
-
- a = np.float32(self.rng.random(size=[nc]))
- scale_arr = a * np.float32((1 << out_type_width) / (1 << in_type_width))
-
- if scale32:
- pass
- # Cap the scaling at 2^31 - 1 for scale32
- scale_arr = np.clip(scale_arr, 1.0 / (1 << 31), (1 << 31) - 1)
- else:
- # Cap the scaling at 2^15 - 1 for scale16
- scale_arr = np.clip(scale_arr, 1.0 / (1 << 31), 32767.0)
-
- # print('{} {} -> {}'.format(out_type_width, in_type_width, scale_arr))
-
- multiplier_arr = np.int32(np.zeros(shape=[nc]))
- shift_arr = np.int32(np.zeros(shape=[nc]))
min_shift_value_arr = np.int64(np.zeros(shape=[nc]))
max_shift_value_arr = np.int64(np.zeros(shape=[nc]))
for i in range(nc):
- multiplier_arr[i], shift_arr[i] = TosaQuantGen.computeMultiplierAndShift(
- scale_arr[i], scale32
- )
min_shift_value_arr[i] = -1 << (shift_arr[i] - 1)
max_shift_value_arr[i] = (1 << (shift_arr[i] - 1)) - 1
@@ -2256,7 +2231,7 @@
)
# Invalidate Input/Output list for error if checks.
- input_list = [val.name]
+ input_list = [val.name, multiplier_val.name, shift_val.name]
output_list = [result_tensor.name]
pCount, cCount = op["operands"]
num_operands = pCount + cCount
@@ -2287,8 +2262,8 @@
attr.RescaleAttribute(
input_zp,
output_zp,
- multiplier_arr,
- shift_arr,
+ [],
+ [],
scale32,
double_round,
per_channel,
@@ -4809,11 +4784,11 @@
},
"rescale": {
"op": Op.RESCALE,
- "operands": (1, 0),
+ "operands": (3, 0),
"build_fcn": (
build_rescale,
TosaTensorGen.tgBasic,
- TosaTensorValuesGen.tvgLazyGenDefault,
+ TosaTensorValuesGen.tvgRescale,
TosaArgGen.agRescale,
),
"types": [
diff --git a/verif/generator/tosa_utils.py b/verif/generator/tosa_utils.py
index 31a0ff0..384463f 100644
--- a/verif/generator/tosa_utils.py
+++ b/verif/generator/tosa_utils.py
@@ -55,6 +55,14 @@
FIXED_DATA = 5
+def dtypeWidth(dtype):
+ """Get the datatype width for data types"""
+ if dtype in DTYPE_ATTRIBUTES:
+ return DTYPE_ATTRIBUTES[dtype]["width"]
+ else:
+ raise Exception(f"Unknown dtype, cannot determine width: {dtype}")
+
+
def dtypeIsSupportedByCompliance(dtype):
"""Types supported by the new data generation and compliance flow."""
if isinstance(dtype, list) or isinstance(dtype, tuple):