include/ling3/quantization.h

#pragma once

#include <cstddef>
#include <cstdint>
#include <span>

namespace ling3 {

struct DynamicQuantization {
    float scale = 1.0F;
    std::size_t clipped = 0;
};

DynamicQuantization QuantizeSymmetricInt8(
    std::span<const float> input,
    std::span<std::int8_t> output) noexcept;

DynamicQuantization QuantizeSymmetricInt4(
    std::span<const float> input,
    std::span<std::int8_t> output) noexcept;

std::int8_t DecodeInt4LowFirst(
    std::span<const std::byte> packed,
    std::size_t index);

void SplitInt8ForInt4Matmul(
    std::span<const std::int8_t> input,
    std::span<std::int8_t> high,
    std::span<std::int8_t> low);

void ReferenceW4Linear(
    std::span<const std::int8_t> input,
    std::span<const std::byte> packed_weights,
    std::size_t output_channels,
    std::span<std::int32_t> output);

void DequantizePerChannel(
    std::span<const std::int32_t> input,
    float input_scale,
    std::span<const float> weight_scales,
    std::span<float> output) noexcept;

} // namespace ling3