// SPDX-License-Identifier: AGPL-3.0-or-later
// quantize_f64.wgsl - Convert FP32 to INT8/INT4 quantization (f64 canonical)
//
// Quantization: Convert floating point values to low-precision integers
// Used for model compression and efficient inference
//
// Algorithm: q = round((x - zero_point) * scale)
// For INT8: clamp to [-128, 127]
// For INT4: clamp to [-8, 7]
struct Params {
size: u32,
scale: f64,
zero_point: f64,
num_bits: u32, // 4 for INT4, 8 for INT8
_padding: u32,
}
@group(0) @binding(0) var<storage, read> input: array<f64>;
@group(0) @binding(1) var<storage, read_write> output: array<i32>; // Quantized integers
@group(0) @binding(2) var<uniform> params: Params;
@compute @workgroup_size(256)
fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {
let idx = global_id.x;
if (idx >= params.size) {
return;
}
let value = input[idx];
// Quantize: round((x - zero_point) * scale)
// Note: scale here is 1/quantization_scale (inverse scale)
let quantized = round((value - params.zero_point) * params.scale);
// Calculate quantization range based on bits
var qmin: f64;
var qmax: f64;
if (params.num_bits == 4u) {
qmin = -8.0;
qmax = 7.0;
} else {
// INT8 (default)
qmin = -128.0;
qmax = 127.0;
}
// Clamp to valid range
let clamped = clamp(quantized, qmin, qmax);
// Convert to integer
output[idx] = i32(clamped);
}