node-gemmlowp

High performance low-precision (8-bit) GEMM kernel and 8-bit quantization/dequantization helpers for JavaScript.

gemmlowp.gemm8(transpose_a, transpose_b,
               M, N, K,
               a, transpose_a ? M : K,
               b, transpose_b ? K : N,
               c, N,
               a_offset, b_offset, c_offset, c_mult, c_shift);

See https://github.com/google/gemmlowp for details.

In addition to the 8-bit GEMM kernel we expose some vectorized primitivs to accelerate quantization and dequantization of floating point data into unsigned 8-bit arrays.

min = gemmlowp.min(float32Array)
max = gemmlowp.max(float32Array)
gemmlowp.clamp(float32Array, min, max)
gemmlowp.bias(float32Array)
gemmlowp.scale(float32Array)
gemmlowp.roundToUint8(float32Array, Uint8Array)
gemmlowp.convertUint8ToFloat(Uint8Array, float32Array)