Skip to content

Align ConvertPrecision with LSB-first nibble packing for u4/i4 types #32144

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 4 commits into from
Nov 6, 2025
Merged

Align ConvertPrecision with LSB-first nibble packing for u4/i4 types #32144

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
787f8bc
Use LSB-first nibble order in ConvertPrecision for u4/i4 types
evkotov Oct 22, 2025
af618a5
clang fixes
evkotov Oct 22, 2025
bc7e691
Refactor convert_low_precisions_int to use Convert::evaluate for con…
evkotov Nov 1, 2025
cb2f896
clang fixes
evkotov Nov 1, 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
21 changes: 16 additions & 5 deletions src/common/transformations/src/transformations/convert_precision.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,6 +10,7 @@
#include "itt.hpp"
#include "openvino/core/graph_util.hpp"
#include "openvino/core/rt_info/weightless_caching_attributes.hpp"
#include "openvino/core/type/element_iterator.hpp"
#include "openvino/op/ops.hpp"
#include "openvino/pass/constant_folding.hpp"
#include "openvino/pass/manager.hpp"
Expand Down Expand Up @@ -1290,7 +1291,7 @@ std::shared_ptr<Node> convert_low_precisions_int(std::shared_ptr<ov::op::v0::Con
const auto size = shape_size(constant->get_shape());
size_t src_idx(0), dst_idx(0), dst_off(0), src_off(0);
if (src_type.bitwidth() < 8) {
src_off = 8 - src_type.bitwidth();
src_off = element::is_lsb_packed(src_type) ? 0 : (8 - src_type.bitwidth());
}

if (to.bitwidth() < 8) {
Expand Down Expand Up @@ -1342,11 +1343,21 @@ std::shared_ptr<Node> convert_low_precisions_int(std::shared_ptr<ov::op::v0::Con
}
// Calculate offsets and indexes
if (src_type.bitwidth() < 8) {
if (src_off == 0) {
src_off = 8;
src_idx++;
if (element::is_lsb_packed(src_type)) {
// LSB-first: offset goes up 0 → 4 → next byte
src_off += src_type.bitwidth();
if (src_off >= 8) {
src_off = 0;
src_idx++;
}
} else {
// MSB-first: offset goes down 4 → 0 → next byte
if (src_off == 0) {
src_off = 8;
src_idx++;
}
src_off -= src_type.bitwidth();
}
src_off -= src_type.bitwidth();
} else {
src_idx++;
}
Expand Down
52 changes: 26 additions & 26 deletions src/common/transformations/tests/utils/convert_precision.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1598,110 +1598,110 @@ TEST(TransformationTests, ConvertPrecision_ConstantConversion_BoolToU8) {
}

TEST(TransformationTests, ConvertPrecision_ConstantConversion_U4ToI8) {
constant_convert_test<uint8_t, int8_t>(element::u4, element::i8, std::vector<uint8_t>{171}, {10, 11});
constant_convert_test<uint8_t, int8_t>(element::u4, element::i8, std::vector<uint8_t>{171}, {11, 10});
}

TEST(TransformationTests, ConvertPrecision_ConstantConversion_U4ToU8) {
constant_convert_test<uint8_t, uint8_t>(element::u4, element::u8, {171}, {10, 11});
constant_convert_test<uint8_t, uint8_t>(element::u4, element::u8, {171}, {11, 10});
}

TEST(TransformationTests, ConvertPrecision_ConstantConversion_U4ToI8_2) {
constant_convert_test<uint8_t, int8_t>(element::u4, element::i8, {96}, {6, 0});
constant_convert_test<uint8_t, int8_t>(element::u4, element::i8, {96}, {0, 6});
}

TEST(TransformationTests, ConvertPrecision_ConstantConversion_U4ToU8_96) {
constant_convert_test<uint8_t, uint8_t>(element::u4, element::u8, {96}, {6, 0});
constant_convert_test<uint8_t, uint8_t>(element::u4, element::u8, {96}, {0, 6});
}

TEST(TransformationTests, ConvertPrecision_ConstantConversion_I4ToU8) {
constant_convert_test<uint8_t, uint8_t>(element::i4, element::u8, {96}, {6, 0});
constant_convert_test<uint8_t, uint8_t>(element::i4, element::u8, {96}, {0, 6});
}

TEST(TransformationTests, ConvertPrecision_ConstantConversion_I4ToI8) {
constant_convert_test<uint8_t, int8_t>(element::i4, element::i8, {96}, {6, 0});
constant_convert_test<uint8_t, int8_t>(element::i4, element::i8, {96}, {0, 6});
}

TEST(TransformationTests, ConvertPrecision_ConstantConversion_I4ToU8_neg) {
constant_convert_test<uint8_t, uint8_t>(element::i4, element::u8, {171}, {250, 251});
constant_convert_test<uint8_t, uint8_t>(element::i4, element::u8, {171}, {251, 250});
}

TEST(TransformationTests, ConvertPrecision_ConstantConversion_I4ToI8_neg) {
constant_convert_test<uint8_t, int8_t>(element::i4, element::i8, {171}, {-6, -5});
constant_convert_test<uint8_t, int8_t>(element::i4, element::i8, {171}, {-5, -6});
}

TEST(TransformationTests, ConvertPrecision_ConstantConversion_U4ToI32) {
constant_convert_test<uint8_t, int32_t>(element::u4, element::i32, {171}, {10, 11});
constant_convert_test<uint8_t, int32_t>(element::u4, element::i32, {171}, {11, 10});
}

TEST(TransformationTests, ConvertPrecision_ConstantConversion_U4ToU32) {
constant_convert_test<uint8_t, uint32_t>(element::u4, element::u32, {171}, {10, 11});
constant_convert_test<uint8_t, uint32_t>(element::u4, element::u32, {171}, {11, 10});
}

TEST(TransformationTests, ConvertPrecision_ConstantConversion_I4ToU32) {
constant_convert_test<uint8_t, uint32_t>(element::i4, element::u32, {96}, {6, 0});
constant_convert_test<uint8_t, uint32_t>(element::i4, element::u32, {96}, {0, 6});
}

TEST(TransformationTests, ConvertPrecision_ConstantConversion_I4ToI32) {
constant_convert_test<uint8_t, int32_t>(element::i4, element::i32, {96}, {6, 0});
constant_convert_test<uint8_t, int32_t>(element::i4, element::i32, {96}, {0, 6});
}

TEST(TransformationTests, ConvertPrecision_ConstantConversion_I4ToU32_neg) {
constant_convert_test<uint8_t, uint32_t>(element::i4, element::u32, {171}, {4294967290, 4294967291});
constant_convert_test<uint8_t, uint32_t>(element::i4, element::u32, {171}, {4294967291, 4294967290});
}

TEST(TransformationTests, ConvertPrecision_ConstantConversion_I4ToI32_neg) {
constant_convert_test<uint8_t, int32_t>(element::i4, element::i32, {171}, {-6, -5});
constant_convert_test<uint8_t, int32_t>(element::i4, element::i32, {171}, {-5, -6});
}

TEST(TransformationTests, ConvertPrecision_ConstantConversion_U4ToI16) {
constant_convert_test<uint8_t, int16_t>(element::u4, element::i16, {171}, {10, 11});
constant_convert_test<uint8_t, int16_t>(element::u4, element::i16, {171}, {11, 10});
}

TEST(TransformationTests, ConvertPrecision_ConstantConversion_U4ToU16) {
constant_convert_test<uint8_t, uint16_t>(element::u4, element::u16, {171}, {10, 11});
constant_convert_test<uint8_t, uint16_t>(element::u4, element::u16, {171}, {11, 10});
}

TEST(TransformationTests, ConvertPrecision_ConstantConversion_I4ToU16) {
constant_convert_test<uint8_t, uint16_t>(element::i4, element::u16, {96}, {6, 0});
constant_convert_test<uint8_t, uint16_t>(element::i4, element::u16, {96}, {0, 6});
}

TEST(TransformationTests, ConvertPrecision_ConstantConversion_I4ToI16) {
constant_convert_test<uint8_t, int16_t>(element::i4, element::i16, {96}, {6, 0});
constant_convert_test<uint8_t, int16_t>(element::i4, element::i16, {96}, {0, 6});
}

TEST(TransformationTests, ConvertPrecision_ConstantConversion_I4ToU16_neg) {
constant_convert_test<uint8_t, uint16_t>(element::i4, element::u16, {171}, {65530, 65531});
constant_convert_test<uint8_t, uint16_t>(element::i4, element::u16, {171}, {65531, 65530});
}

TEST(TransformationTests, ConvertPrecision_ConstantConversion_I4ToI16_neg) {
constant_convert_test<uint8_t, int16_t>(element::i4, element::i16, {171}, {-6, -5});
constant_convert_test<uint8_t, int16_t>(element::i4, element::i16, {171}, {-5, -6});
}

TEST(TransformationTests, ConvertPrecision_ConstantConversion_U4ToI64) {
constant_convert_test<uint8_t, int64_t>(element::u4, element::i64, {171}, {10, 11});
constant_convert_test<uint8_t, int64_t>(element::u4, element::i64, {171}, {11, 10});
}

TEST(TransformationTests, ConvertPrecision_ConstantConversion_U4ToU64) {
constant_convert_test<uint8_t, int64_t>(element::u4, element::u64, {171}, {10, 11});
constant_convert_test<uint8_t, int64_t>(element::u4, element::u64, {171}, {11, 10});
}

TEST(TransformationTests, ConvertPrecision_ConstantConversion_I4ToU64) {
constant_convert_test<uint8_t, uint64_t>(element::i4, element::u64, {96}, {6, 0});
constant_convert_test<uint8_t, uint64_t>(element::i4, element::u64, {96}, {0, 6});
}

TEST(TransformationTests, ConvertPrecision_ConstantConversion_I4ToI64) {
constant_convert_test<uint8_t, int64_t>(element::i4, element::i64, {96}, {6, 0});
constant_convert_test<uint8_t, int64_t>(element::i4, element::i64, {96}, {0, 6});
}

TEST(TransformationTests, ConvertPrecision_ConstantConversion_I4ToU64_neg) {
constant_convert_test<uint8_t, uint64_t>(element::i4,
element::u64,
{171},
{18446744073709551610u, 18446744073709551611u});
{18446744073709551611u, 18446744073709551610u});
}

TEST(TransformationTests, ConvertPrecision_ConstantConversion_I4ToI64_neg) {
constant_convert_test<uint8_t, int64_t>(element::i4, element::i64, {171}, {-6, -5});
constant_convert_test<uint8_t, int64_t>(element::i4, element::i64, {171}, {-5, -6});
}

TEST(TransformationTests, ConvertPrecision_ConstantConversion_U1ToU8) {
Expand Down
Loading