Skip to content

Add 16bit support for unpack, transpose and comparisons operators #3028

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 17 commits into from
Mar 17, 2025
Merged

Add 16bit support for unpack, transpose and comparisons operators #3028

Show file tree
Hide file tree
Changes from 6 commits
Commits
Show all changes
17 commits
Select commit Hold shift + click to select a range
5e0b839
Add 16bit support for unpack, transpose and comparisons operators
rameshkunasi Dec 24, 2024
611d82b
Added support for FC16x8 per channel quantization
rameshkunasi Dec 24, 2024
6d6e3ee
indentation correction
rameshkunasi Jan 9, 2025
733e523
indentation correction
rameshkunasi Jan 9, 2025
2ce81e8
indentation correction
rameshkunasi Jan 27, 2025
b1f00e7
Merge branch 'main' into main
suleshahid Feb 10, 2025
fe90463
Added unit tests for transpose, greater and fully_connected per chann…
rameshkunasi Mar 13, 2025
31a2318
Indentations and function name changes
rameshkunasi Mar 13, 2025
33d617b
Merge branch 'main' of github.com:rameshkunasi/tflite-micro
rameshkunasi Mar 13, 2025
529d8fe
Merge branch 'main' into main
rameshkunasi Mar 13, 2025
d4db629
Revert "Indentations and function name changes"
rameshkunasi Mar 13, 2025
fae534c
clean up
rameshkunasi Mar 13, 2025
6c37801
modified as per the suggested changes
rameshkunasi Mar 14, 2025
615c34a
Resolved unused variable error for xtensa
rameshkunasi Mar 14, 2025
2393651
Removed CMSIS_NN not supported flag in FC INT8 per channel quantizati…
rameshkunasi Mar 14, 2025
ae015fa
Moved FC Per channel INT8 quantization unit test under not supported …
rameshkunasi Mar 15, 2025
e890dbe
Fixed formatting error and compilation error for HEXAGON platform
rameshkunasi Mar 16, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
13 changes: 13 additions & 0 deletions tensorflow/lite/micro/kernels/comparisons.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -286,6 +286,19 @@ TfLiteStatus GreaterEval(TfLiteContext* context, TfLiteNode* node) {
tflite::micro::GetTensorData<int8_t>(input2), output_shape,
output_data);
break;
case kTfLiteInt16:
requires_broadcast
? reference_ops::Broadcast4DSlowGreaterWithScaling(
data->params, input1_shape,
tflite::micro::GetTensorData<int16_t>(input1), input2_shape,
tflite::micro::GetTensorData<int16_t>(input2), output_shape,
output_data)
: reference_ops::GreaterWithScaling(
data->params, input1_shape,
tflite::micro::GetTensorData<int16_t>(input1), input2_shape,
tflite::micro::GetTensorData<int16_t>(input2), output_shape,
output_data);
break;
default:
MicroPrintf("Type %s (%d) not supported.",
TfLiteTypeGetName(input1->type), input1->type);
Expand Down
104 changes: 88 additions & 16 deletions tensorflow/lite/micro/kernels/fully_connected.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -238,25 +238,97 @@ TfLiteStatus FullyConnectedEval(TfLiteContext* context, TfLiteNode* node) {
case kTfLiteInt16: {
switch (filter->type) {
case kTfLiteInt8: {
tflite::reference_integer_ops::FullyConnected(
FullyConnectedParamsQuantized(data),
tflite::micro::GetTensorShape(input),
tflite::micro::GetTensorData<int16_t>(input),
tflite::micro::GetTensorShape(filter),
if (bias == nullptr || bias->type == kTfLiteInt32) {
data.is_per_channel
? tflite::reference_integer_ops::FullyConnectedPerChannel(
FullyConnectedParamsQuantized(data),
data.per_channel_output_multiplier,
reinterpret_cast<const int*>(
data.per_channel_output_shift),
tflite::micro::GetTensorShape(input),
tflite::micro::GetTensorData<int16_t>(input),
tflite::micro::GetTensorShape(filter),
#ifdef USE_TFLM_COMPRESSION
tflite::micro::GetTensorData<int8_t>(micro_context, filter,
weights_comp_td,
data.weights_scratch_index),
tflite::micro::GetTensorShape(bias),
tflite::micro::GetOptionalTensorData<int64_t>(
micro_context, bias, bias_comp_td, data.bias_scratch_index),
tflite::micro::GetTensorData<int8_t>(
micro_context, filter, weights_comp_td,
data.weights_scratch_index),
tflite::micro::GetTensorShape(bias),
tflite::micro::GetOptionalTensorData<int32_t>(
micro_context, bias, bias_comp_td,
data.bias_scratch_index),
#else // USE_TFLM_COMPRESSION
tflite::micro::GetTensorData<int8_t>(filter),
tflite::micro::GetTensorShape(bias),
tflite::micro::GetOptionalTensorData<int64_t>(bias),
tflite::micro::GetTensorData<int8_t>(filter),
tflite::micro::GetTensorShape(bias),
tflite::micro::GetOptionalTensorData<int32_t>(bias),
#endif // USE_TFLM_COMPRESSION
tflite::micro::GetTensorShape(output),
tflite::micro::GetTensorData<int16_t>(output));
tflite::micro::GetTensorShape(output),
tflite::micro::GetTensorData<int16_t>(output))
: tflite::reference_integer_ops::FullyConnected(
FullyConnectedParamsQuantized(data),
tflite::micro::GetTensorShape(input),
tflite::micro::GetTensorData<int16_t>(input),
tflite::micro::GetTensorShape(filter),
#ifdef USE_TFLM_COMPRESSION
tflite::micro::GetTensorData<int8_t>(
micro_context, filter, weights_comp_td,
data.weights_scratch_index),
tflite::micro::GetTensorShape(bias),
tflite::micro::GetOptionalTensorData<int32_t>(
micro_context, bias, bias_comp_td,
data.bias_scratch_index),
#else // USE_TFLM_COMPRESSION
tflite::micro::GetTensorData<int8_t>(filter),
tflite::micro::GetTensorShape(bias),
tflite::micro::GetOptionalTensorData<int32_t>(bias),
#endif // USE_TFLM_COMPRESSION
tflite::micro::GetTensorShape(output),
tflite::micro::GetTensorData<int16_t>(output));
} else if (bias->type == kTfLiteInt64) {
data.is_per_channel
? tflite::reference_integer_ops::FullyConnectedPerChannel(
FullyConnectedParamsQuantized(data),
data.per_channel_output_multiplier,
reinterpret_cast<const int*>(
data.per_channel_output_shift),
tflite::micro::GetTensorShape(input),
tflite::micro::GetTensorData<int16_t>(input),
tflite::micro::GetTensorShape(filter),
#ifdef USE_TFLM_COMPRESSION
tflite::micro::GetTensorData<int8_t>(
micro_context, filter, weights_comp_td,
data.weights_scratch_index),
tflite::micro::GetTensorShape(bias),
tflite::micro::GetOptionalTensorData<int64_t>(
micro_context, bias, bias_comp_td,
data.bias_scratch_index),
#else // USE_TFLM_COMPRESSION
tflite::micro::GetTensorData<int8_t>(filter),
tflite::micro::GetTensorShape(bias),
tflite::micro::GetOptionalTensorData<int64_t>(bias),
#endif // USE_TFLM_COMPRESSION
tflite::micro::GetTensorShape(output),
tflite::micro::GetTensorData<int16_t>(output))
: tflite::reference_integer_ops::FullyConnected(
FullyConnectedParamsQuantized(data),
tflite::micro::GetTensorShape(input),
tflite::micro::GetTensorData<int16_t>(input),
tflite::micro::GetTensorShape(filter),
#ifdef USE_TFLM_COMPRESSION
tflite::micro::GetTensorData<int8_t>(
micro_context, filter, weights_comp_td,
data.weights_scratch_index),
tflite::micro::GetTensorShape(bias),
tflite::micro::GetOptionalTensorData<int64_t>(
micro_context, bias, bias_comp_td,
data.bias_scratch_index),
#else // USE_TFLM_COMPRESSION
tflite::micro::GetTensorData<int8_t>(filter),
tflite::micro::GetTensorShape(bias),
tflite::micro::GetOptionalTensorData<int64_t>(bias),
#endif // USE_TFLM_COMPRESSION
tflite::micro::GetTensorShape(output),
tflite::micro::GetTensorData<int16_t>(output));
}
break;
}
default: {
Expand Down
11 changes: 8 additions & 3 deletions tensorflow/lite/micro/kernels/fully_connected_common.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -95,9 +95,14 @@ TfLiteStatus CalculateOpDataFullyConnected(
filter->quantization.params);
const int per_channel_quantization_size = affine_quantization->scale->size;

// Currently only Int8 is supported for per channel quantization.
TF_LITE_ENSURE(context,
input->type == kTfLiteInt8 && filter->type != kTfLiteInt4);
// Currently only Int8/Int16 are supported for per channel quantization.
TF_LITE_ENSURE(
context,
(input->type == kTfLiteInt8 && filter->type != kTfLiteInt4) ||
(input->type == kTfLiteInt16 && filter->type != kTfLiteInt4));

TF_LITE_ENSURE_EQ(context, affine_quantization->scale->size,
per_channel_quantization_size);

TF_LITE_ENSURE_EQ(
context, per_channel_quantization_size,
Expand Down
8 changes: 7 additions & 1 deletion tensorflow/lite/micro/kernels/transpose.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -103,10 +103,16 @@ TfLiteStatus TransposeEval(TfLiteContext* context, TfLiteNode* node) {
tflite::micro::GetTensorShape(output),
tflite::micro::GetTensorData<int8_t>(output));
break;
case kTfLiteInt16:
reference_ops::Transpose(params, tflite::micro::GetTensorShape(input),
tflite::micro::GetTensorData<int16_t>(input),
tflite::micro::GetTensorShape(output),
tflite::micro::GetTensorData<int16_t>(output));
break;
default:
MicroPrintf(
"Type %s is currently not supported by Transpose. "
"Only float32 and int8 is supported",
"Only float32, int8 and int16 are supported",
TfLiteTypeGetName(input->type));
return kTfLiteError;
}
Expand Down
3 changes: 3 additions & 0 deletions tensorflow/lite/micro/kernels/unpack.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -89,6 +89,9 @@ TfLiteStatus UnpackEval(TfLiteContext* context, TfLiteNode* node) {
case kTfLiteInt8: {
return UnpackImpl<int8_t>(context, node, input, data->num, data->axis);
}
case kTfLiteInt16: {
return UnpackImpl<int16_t>(context, node, input, data->num, data->axis);
}
default: {
MicroPrintf("Type '%s' is not supported by unpack.",
TfLiteTypeGetName(input->type));
Expand Down
Loading