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
mergify 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 14 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
52 changes: 52 additions & 0 deletions tensorflow/lite/micro/kernels/comparisons_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -126,6 +126,29 @@ void TestComparisonQuantizedInt8(const TFLMRegistration& registration,
TestComparison(registration, tensors, expected_output_data, output_data);
}

void TestComparisonQuantizedInt16(const TFLMRegistration& registration,
int* input1_dims_data, float* input1_data,
int16_t* input1_quantized, float input1_scale,
int input1_zero_point, int* input2_dims_data,
float* input2_data, int16_t* input2_quantized,
float input2_scale, int input2_zero_point,
bool* expected_output_data,
int* output_dims_data, bool* output_data) {
TfLiteIntArray* input1_dims = IntArrayFromInts(input1_dims_data);
TfLiteIntArray* input2_dims = IntArrayFromInts(input2_dims_data);
TfLiteIntArray* output_dims = IntArrayFromInts(output_dims_data);

TfLiteTensor tensors[tensors_size] = {
CreateQuantizedTensor(input1_data, input1_quantized, input1_dims,
input1_scale, input1_zero_point),
CreateQuantizedTensor(input2_data, input2_quantized, input2_dims,
input2_scale, input2_zero_point),
CreateTensor(output_data, output_dims),
};

TestComparison(registration, tensors, expected_output_data, output_data);
}

} // namespace
} // namespace testing
} // namespace tflite
Expand Down Expand Up @@ -656,6 +679,35 @@ TF_LITE_MICRO_TEST(GreaterQuantizedInt8WithBroadcast) {
}
}

TF_LITE_MICRO_TEST(GreaterQuantizedInt16WithBroadcast) {
const int num_shapes = 4;
const int max_shape_size = 5;
int test_shapes[num_shapes][max_shape_size] = {
{1, 6}, {2, 2, 3}, {3, 2, 1, 3}, {4, 1, 3, 1, 2}};

for (int i = 0; i < num_shapes; ++i) {
int* input1_dim = test_shapes[i];
int input2_dim[] = {1, 1};
float input1_data[] = {20, -2, -71, 8, 11, 20};
float input2_data[] = {8};

bool expected_data[] = {true, false, false, false, true, true};
int* expected_dim = input1_dim;

const float input1_scale = 0.5;
const int input1_zero_point = -9;
int16_t input1_quantized[6];
int16_t input2_quantized[6];

bool output_data[6];
tflite::testing::TestComparisonQuantizedInt16(
tflite::Register_GREATER(), input1_dim, input1_data, input1_quantized,
input1_scale, input1_zero_point, input2_dim, input2_data,
input2_quantized, input1_scale, input1_zero_point, expected_data,
expected_dim, output_data);
}
}

TF_LITE_MICRO_TEST(GreaterEqualQuantizedInt8WithBroadcast) {
const int num_shapes = 4;
const int max_shape_size = 5;
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
129 changes: 127 additions & 2 deletions tensorflow/lite/micro/kernels/fully_connected_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -45,6 +45,14 @@ const float simple_weights_data[] = {
int simple_bias_dims[] = {1, 3};
const float simple_bias_data[] = {1, 2, 3};

#if (defined(USE_TFLM_COMPRESSION) || (!defined(XTENSA) && !defined(CMSIS_NN)))

constexpr size_t simple_bias_size =
std::extent<decltype(simple_bias_data)>::value;

#endif // (!defined(XTENSA) && !defined(CMSIS_NN)) ||
// (defined(USE_TFLM_COMPRESSION))

#ifdef USE_TFLM_COMPRESSION

// compressed filter data for kBinQuant scheme
Expand All @@ -60,8 +68,6 @@ constexpr int kBinQuantWeightBitWidth = 4;
// Align the tensor data the same as a Buffer in the schema
alignas(16) constexpr uint8_t kBinQuantBiasData[] = {0x18};
constexpr int kBinQuantBiasBitWidth = 2;
constexpr size_t simple_bias_size =
std::extent<decltype(simple_bias_data)>::value;

#endif // USE_TFLM_COMPRESSION

Expand Down Expand Up @@ -504,6 +510,58 @@ TfLiteStatus TestFullyConnectedQuantizedCompressed(

#endif // USE_TFLM_COMPRESSION

template <typename dataT, typename weightT, typename biasT>
TfLiteStatus TestFullyConnectedQuantizedPerChannel(
int* input_dims_data, const float* input_data, dataT* input_quantized,
const float input_scale, const int input_zero_point, int* weights_dims_data,
const float* weights_data, weightT* weights_quantized,
float* weights_scales, int* weights_zero_points, int* bias_dims_data,
const float* bias_data, biasT* bias_quantized, const float* golden,
dataT* golden_quantized, int* output_dims_data, const float output_scale,
const int output_zero_point, TfLiteFusedActivation activation,
dataT* output_data, TfLiteType weights_packed_type = kTfLiteNoType) {
TfLiteIntArray* input_dims = IntArrayFromInts(input_dims_data);
TfLiteIntArray* weights_dims = IntArrayFromInts(weights_dims_data);
TfLiteIntArray* bias_dims = IntArrayFromInts(bias_dims_data);
TfLiteIntArray* output_dims = IntArrayFromInts(output_dims_data);
const int output_dims_count = ElementCount(*output_dims);
bool null_bias = bias_data == nullptr ? true : false;

constexpr int array_size = 4; // Avoid variable length array warning.
const int inputs_size = null_bias ? 2 : 3;
constexpr int outputs_size = 1;
const int tensors_size = inputs_size + outputs_size;
TfLiteTensor tensors[array_size];
TfLiteAffineQuantization weights_quant, bias_quant;
float bias_scales[5];
int bias_zero_points[5];

tensors[0] = CreateQuantizedTensor(input_data, input_quantized, input_dims,
input_scale, input_zero_point);
tensors[1] = CreateSymmetricPerChannelQuantizedTensorWithoutScaleEstimation(
weights_data, weights_quantized, weights_dims, weights_scales,
weights_zero_points, &weights_quant, 0 /* quantized dimension */, false,
weights_packed_type);

if (null_bias) {
tensors[2] = CreateQuantizedTensor(output_data, output_dims, output_scale,
output_zero_point);
} else {
tensors[2] = CreatePerChannelQuantizedBiasTensor(
bias_data, bias_quantized, bias_dims, input_scale, &weights_scales[1],
bias_scales, bias_zero_points, &bias_quant,
0 /* quantized dimension */);
tensors[3] = CreateQuantizedTensor(output_data, output_dims, output_scale,
output_zero_point);
}

Quantize(golden, golden_quantized, output_dims_count, output_scale,
output_zero_point);
return ValidateFullyConnectedGoldens(
tensors, tensors_size, null_bias, activation, 1.0f /* tolerance */,
output_dims_count, golden_quantized, output_data);
}

} // namespace
} // namespace testing
} // namespace tflite
Expand Down Expand Up @@ -651,6 +709,40 @@ TF_LITE_MICRO_TEST(SimpleTestQuantizedInt8Compressed) {

#endif // USE_TFLM_COMPRESSION

#if !defined(XTENSA) && !defined(CMSIS_NN)

TF_LITE_MICRO_TEST(SimpleTestQuantizedPerChannelInt8) {
const float input_scale = 0.5f;
const int input_zero_point = -1;
const float output_scale = 1.0f;
const int output_zero_point = -1;
int weights_zero_points[tflite::testing::simple_bias_size + 1] = {
tflite::testing::simple_bias_size, 0, 0, 0};
float weights_scales[tflite::testing::simple_bias_size + 1] = {
tflite::testing::simple_bias_size, 0.2f, 0.25f, 0.5f};

int8_t input_quantized[tflite::testing::simple_input_size];
int8_t weights_quantized[tflite::testing::simple_weights_size];
int32_t bias_quantized[tflite::testing::simple_output_size];
int8_t golden_quantized[tflite::testing::simple_output_size];
int8_t output_data[tflite::testing::simple_output_size];

TF_LITE_MICRO_EXPECT_EQ(
tflite::testing::TestFullyConnectedQuantizedPerChannel(
tflite::testing::simple_input_dims,
tflite::testing::simple_input_data, input_quantized, input_scale,
input_zero_point, tflite::testing::simple_weights_dims,
tflite::testing::simple_weights_data, weights_quantized,
weights_scales, weights_zero_points,
tflite::testing::simple_bias_dims, tflite::testing::simple_bias_data,
bias_quantized, tflite::testing::simple_golden, golden_quantized,
tflite::testing::simple_output_dims, output_scale, output_zero_point,
kTfLiteActNone, output_data),
kTfLiteOk);
}

#endif // #if !defined(XTENSA) && !defined(CMSIS_NN)

#if !defined(HEXAGON)
TF_LITE_MICRO_TEST(SimpleTestQuantizedInt16) {
const float input_scale = 128.0 / 65536;
Expand Down Expand Up @@ -732,6 +824,39 @@ TF_LITE_MICRO_TEST(SimpleTestQuantizedInt16Compressed) {

#endif // USE_TFLM_COMPRESSION

#if !defined(XTENSA) && !defined(CMSIS_NN)

TF_LITE_MICRO_TEST(SimpleTestPerChannelQuantizedInt16) {
const float input_scale = 128.0 / 65536;
const int input_zero_point = 0;
const float output_scale = 128.0 / 65536;
const int output_zero_point = 0;
int weights_zero_points[tflite::testing::simple_bias_size + 1] = {
tflite::testing::simple_bias_size, 0, 0, 0};
float weights_scales[tflite::testing::simple_bias_size + 1] = {
tflite::testing::simple_bias_size, 0.2f, 0.25f, 0.5f};

int16_t input_quantized[tflite::testing::simple_input_size];
int8_t weights_quantized[tflite::testing::simple_weights_size];
int64_t bias_quantized[tflite::testing::simple_output_size];
int16_t golden_quantized[tflite::testing::simple_output_size];
int16_t output_data[tflite::testing::simple_output_size];

TF_LITE_MICRO_EXPECT_EQ(
tflite::testing::TestFullyConnectedQuantizedPerChannel(
tflite::testing::simple_input_dims,
tflite::testing::simple_input_data, input_quantized, input_scale,
input_zero_point, tflite::testing::simple_weights_dims,
tflite::testing::simple_weights_data, weights_quantized,
weights_scales, weights_zero_points,
tflite::testing::simple_bias_dims, tflite::testing::simple_bias_data,
bias_quantized, tflite::testing::simple_golden, golden_quantized,
tflite::testing::simple_output_dims, output_scale, output_zero_point,
kTfLiteActNone, output_data),
kTfLiteOk);
}

#endif // #if !defined(XTENSA) && !defined(CMSIS_NN)
#endif // !defined(HEXAGON)

TF_LITE_MICRO_TEST(SimpleTest4DInputQuantizedInt8) {
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
Loading
Loading