tract_core/ops/
quant.rs

1#![allow(clippy::unnecessary_cast)]
2
3use crate::internal::*;
4use crate::ops::element_wise::ElementWiseOp;
5use crate::ops::math::QScale;
6use num_traits::AsPrimitive;
7use tract_linalg::Scaler;
8use tract_linalg::lut::Lut;
9use tract_linalg::mmm::RoundingPolicy;
10
11use super::binary::TypedBinOp;
12use super::math::round_ties_to_even;
13
14pub fn quantize_linear_f32_u8(x: f32, scale: f32, zero_point: i32) -> u8 {
15    (((x * scale).round() as i32) + zero_point).clamp(u8::MIN as i32, u8::MAX as i32) as u8
16}
17
18pub fn quantize_linear_f32_i8(x: f32, scale: f32, zero_point: i32) -> i8 {
19    (((x * scale).round() as i32) + zero_point).clamp(i8::MIN as i32, i8::MAX as i32) as i8
20}
21
22element_wise_oop!(quantize_linear_u8,
23 QuantizeLinearU8 {
24     scale: f32,
25     zero_point: u8
26 },
27 [f16] => u8 |op, xs, ys| {
28     xs.iter().zip(ys.iter_mut()).for_each(|(x,y)|
29                                           *y = quantize_linear_f32_u8(x.to_f32(), op.scale, op.zero_point as i32)
30                                          );
31     Ok(())
32 },
33 [f32,i32] => u8 |op, xs, ys| {
34     xs.iter().zip(ys.iter_mut()).for_each(|(x,y)|
35                                           *y = quantize_linear_f32_u8(*x as f32, op.scale, op.zero_point as i32)
36                                          );
37     Ok(())
38 };
39 info: info_quantize_linear_u8
40);
41
42fn info_quantize_linear_u8(q: &QuantizeLinearU8) -> TractResult<Vec<String>> {
43    Ok(vec![format!(
44        "scale: {} zero_point: {} 1/scale: {}",
45        q.scale,
46        q.zero_point,
47        q.scale.recip()
48    )])
49}
50
51element_wise_oop!(quantize_linear_i8,
52 QuantizeLinearI8 {
53     scale: f32,
54     zero_point: i8
55 },
56 [f32,i32] => i8 |op, xs, ys| {
57     xs.iter().zip(ys.iter_mut()).for_each(|(x,y)|
58                                           *y = quantize_linear_f32_i8(*x as f32, op.scale, op.zero_point as i32)
59                                          );
60     Ok(())
61 };
62 info: info_quantize_linear_i8
63);
64
65fn info_quantize_linear_i8(q: &QuantizeLinearI8) -> TractResult<Vec<String>> {
66    Ok(vec![format!(
67        "scale: {} zero_point: {} 1/scale: {}",
68        q.scale,
69        q.zero_point,
70        q.scale.recip()
71    )])
72}
73
74#[derive(Clone, Debug, new)]
75pub struct DequantizeLinearF32 {
76    pub scale: f32,
77    pub zero_point: i32,
78}
79
80impl DequantizeLinearF32 {
81    fn eval_t<T: Datum + AsPrimitive<i32>>(&self, input: &Tensor) -> TractResult<Tensor> {
82        let mut output = unsafe { Tensor::uninitialized::<f32>(input.shape())? };
83        input
84            .try_as_dense()?
85            .as_slice::<T>()?
86            .iter()
87            .zip(output.try_as_dense_mut()?.as_slice_mut::<f32>()?.iter_mut())
88            .for_each(|(x, y)| *y = (x.as_() - self.zero_point) as f32 * self.scale);
89        Ok(output)
90    }
91}
92
93impl Op for DequantizeLinearF32 {
94    fn name(&self) -> StaticName {
95        "DequantizeLinearF32".into()
96    }
97
98    fn info(&self) -> TractResult<Vec<String>> {
99        Ok(vec![format!("scale: {} zero_point: {}", self.scale, self.zero_point)])
100    }
101
102    fn validation(&self) -> Validation {
103        Validation::Accurate
104    }
105
106    op_as_typed_op!();
107}
108
109impl EvalOp for DequantizeLinearF32 {
110    fn is_stateless(&self) -> bool {
111        true
112    }
113    fn eval(&self, inputs: TVec<TValue>) -> TractResult<TVec<TValue>> {
114        let output = match inputs[0].datum_type() {
115            DatumType::I8 => self.eval_t::<i8>(&inputs[0])?,
116            DatumType::I32 => self.eval_t::<i32>(&inputs[0])?,
117            DatumType::U8 => self.eval_t::<u8>(&inputs[0])?,
118            dt => bail!("Unsupported type {:?}", dt),
119        };
120        Ok(tvec!(output.into_tvalue()))
121    }
122}
123
124impl TypedOp for DequantizeLinearF32 {
125    fn output_facts(&self, inputs: &[&TypedFact]) -> TractResult<TVec<TypedFact>> {
126        let mut fact = inputs[0].clone();
127        fact.datum_type = f32::datum_type();
128        Ok(tvec!(fact))
129    }
130
131    fn axes_mapping(
132        &self,
133        inputs: &[&TypedFact],
134        outputs: &[&TypedFact],
135    ) -> TractResult<AxesMapping> {
136        AxesMapping::natural(inputs, outputs)
137    }
138
139    fn change_axes(
140        &self,
141        model: &TypedModel,
142        node: &TypedNode,
143        _io: InOut,
144        change: &AxisOp,
145    ) -> TractResult<Option<AxisChangeConsequence>> {
146        Ok(Some(AxisChangeConsequence::new(model, node, None, change)))
147    }
148
149    fn declutter(
150        &self,
151        model: &TypedModel,
152        dequant: &TypedNode,
153    ) -> TractResult<Option<TypedModelPatch>> {
154        let mut current = dequant;
155        let incoming_dt = model.node_input_facts(dequant.id)?[0].datum_type;
156        while let Some(quant) = model.single_succ(current.id)? {
157            let q_params = if let Some(op) = quant.op_as::<ElementWiseOp>() {
158                if let Some(mop) = op.0.downcast_ref::<QuantizeLinearU8>() {
159                    Some((mop.scale, mop.zero_point as i32, u8::datum_type()))
160                } else {
161                    op.0.downcast_ref::<QuantizeLinearI8>()
162                        .map(|mop| (mop.scale, mop.zero_point as i32, i8::datum_type()))
163                }
164            } else {
165                None
166            };
167            if let Some((scale, zero_point, dt)) = q_params {
168                // first, try Op::quantize() on all ops in the chain
169                let mut patch = TypedModelPatch::default();
170                let mut wire: OutletId = patch.tap_model(model, dequant.inputs[0])?;
171                let mut next = model.single_succ(dequant.id)?.unwrap();
172                loop {
173                    if let Some(op) = next
174                        .op
175                        .quantize(model, dequant, dt, scale, zero_point)
176                        .with_context(|| format!("Quantizing {next}"))?
177                    {
178                        wire = patch.wire_node(&*next.name, op, [wire].as_ref())?[0];
179                    } else {
180                        break;
181                    }
182                    if next.id == current.id {
183                        patch.shunt_outside(model, OutletId::new(quant.id, 0), wire)?;
184                        return Ok(Some(patch));
185                    } else {
186                        next = model.single_succ(next.id)?.unwrap();
187                    }
188                }
189                // or else make a lookup table
190                if incoming_dt == DatumType::I8 || incoming_dt == DatumType::U8 {
191                    let mut adhoc_model = TypedModel::default();
192                    let mut wire = adhoc_model.add_source("ad-hoc", dt.fact([256]))?;
193                    let mut next = model.single_succ(dequant.id)?.unwrap();
194                    let mut name = None;
195                    // plug in dequant
196                    wire = adhoc_model.wire_node(
197                        &*dequant.name,
198                        dequant.op.clone(),
199                        [wire].as_ref(),
200                    )?[0];
201                    while next.id != quant.id {
202                        name.get_or_insert(&*next.name);
203                        wire =
204                            adhoc_model.wire_node(&*next.name, next.op.clone(), [wire].as_ref())?
205                                [0];
206                        next = model.single_succ(next.id)?.unwrap();
207                    }
208                    // plug in quant
209                    wire =
210                        adhoc_model.wire_node(&*quant.name, quant.op.clone(), [wire].as_ref())?[0];
211                    adhoc_model.set_output_outlets(&[wire])?;
212                    let input = (0u8..=255).collect::<Vec<u8>>();
213                    let input = match dt {
214                        DatumType::I8 => unsafe {
215                            tensor1(std::mem::transmute::<&[u8], &[i8]>(&*input))
216                        },
217                        DatumType::U8 => tensor1(&input),
218                        _ => unreachable!(),
219                    };
220                    let output =
221                        SimplePlan::new(adhoc_model)?.run(tvec!(input.into_tvalue()))?.remove(0);
222                    let table: &[u8] = match dt {
223                        DatumType::I8 => unsafe {
224                            std::mem::transmute::<&[i8], &[u8]>(
225                                output.try_as_dense()?.as_slice::<i8>()?,
226                            )
227                        },
228                        DatumType::U8 => output.try_as_dense()?.as_slice::<u8>()?,
229                        _ => unreachable!(),
230                    };
231                    let op = lookup_table((tract_linalg::ops().lut_u8)(table));
232                    let mut patch = TypedModelPatch::default();
233                    let mut wire: OutletId = patch.tap_model(model, dequant.inputs[0])?;
234
235                    wire = patch.wire_node(name.unwrap_or(&*dequant.name), op, [wire].as_ref())?[0];
236                    patch.shunt_outside(model, OutletId::new(quant.id, 0), wire)?;
237                    return Ok(Some(patch));
238                }
239            }
240            let (input_facts, output_facts) = model.node_facts(quant.id)?;
241            let invariants = quant
242                .op
243                .axes_mapping(&input_facts, &output_facts)
244                .with_context(|| format!("Querying invariants for {quant}"))?;
245            if invariants.is_element_wise_unary() {
246                current = quant;
247            } else {
248                break;
249            }
250        }
251        Ok(None)
252    }
253
254    as_op!();
255}
256
257element_wise_oop!(lookup_table,
258 LookupTable {
259     table: Box<dyn Lut>
260 },
261 [i8] => i8 |op, xs, ys| {
262     ys.copy_from_slice(xs);
263     unsafe {
264         let casted = std::slice::from_raw_parts_mut(ys.as_mut_ptr() as *mut u8, ys.len());
265         op.table.run(casted);
266     }
267     Ok(())
268 },
269 [u8] => u8 |op, xs, ys| {
270     ys.copy_from_slice(xs);
271     op.table.run(ys);
272     Ok(())
273 }
274);
275
276#[derive(Debug, Clone, Hash)]
277pub struct Scale;
278
279impl crate::ops::binary::BinMiniOp for Scale {
280    fn name(&self) -> &'static str {
281        "Scale"
282    }
283    fn result_datum_type(&self, a: DatumType, b: DatumType) -> TractResult<DatumType> {
284        if !a.is_float() {
285            bail!("Scale left operand must be float, got {:?}", a);
286        }
287        Ok(b)
288    }
289
290    fn operating_datum_type(&self, a: DatumType, b: DatumType) -> TractResult<DatumType> {
291        if !a.is_float() {
292            bail!("Scale left operand must be float, got {:?}", a);
293        }
294        Ok(b)
295    }
296
297    fn eval_out_of_place(&self, c: &mut Tensor, a: &Tensor, b: &Tensor) -> TractResult<()> {
298        let a = a.cast_to::<f32>()?;
299        let a = a.to_dense_array_view::<f32>()?;
300        unsafe fn eval_out_of_place_t<T: Datum + AsPrimitive<f32>>(
301            c: &mut Tensor,
302            a: &ndarray::ArrayViewD<f32>,
303            b: &Tensor,
304        ) where
305            f32: AsPrimitive<T>,
306        {
307            let b = unsafe { b.to_array_view_unchecked::<T>() };
308            let mut c = unsafe { c.to_array_view_mut_unchecked::<T>() };
309            ndarray::Zip::from(&mut c)
310                .and_broadcast(a)
311                .and_broadcast(b)
312                .for_each(|c, a, b| *c = scale_by(*b, *a))
313        }
314        unsafe { dispatch_numbers!(eval_out_of_place_t(b.datum_type())(c, &a, b)) }
315        Ok(())
316    }
317
318    fn eval_in_a(&self, a: &mut Tensor, b: &Tensor) -> TractResult<()> {
319        let mut a_dense = a.try_as_dense_mut()?;
320        let a = a_dense.to_array_view_mut::<f32>()?;
321        let b = b.to_dense_array_view::<f32>()?;
322        ndarray::Zip::from(a).and_broadcast(b).for_each(|a, b| *a = scale_by(*b, *a));
323        Ok(())
324    }
325
326    fn is_commutative(&self) -> bool {
327        false
328    }
329
330    fn declutter(
331        &self,
332        model: &TypedModel,
333        node: &TypedNode,
334    ) -> TractResult<Option<TypedModelPatch>> {
335        let a = model.outlet_fact(node.inputs[0])?;
336        if let Some(a) = &a.uniform {
337            if a.cast_to_scalar::<f32>()? == 1. {
338                return Ok(Some(TypedModelPatch::rewire(
339                    model,
340                    &node.inputs[1..2],
341                    &[node.id.into()],
342                    &|_p, x| Ok(x.into()),
343                )?));
344            } else if node.outputs[0].fact.datum_type == DatumType::I32 {
345                let factor = a.cast_to_scalar::<f32>()?;
346                let scaler = Scaler::new(factor, RoundingPolicy::Even);
347
348                let op = ElementWiseOp(Box::new(QScale { scaler }), None);
349                let patch =
350                    TypedModelPatch::replace_single_op(model, node, &node.inputs[1..2], op)?;
351
352                return Ok(Some(patch));
353            }
354        }
355        Ok(None)
356    }
357}
358
359#[inline]
360pub(crate) fn scale_by<T: Datum + AsPrimitive<f32>>(b: T, a: f32) -> T
361where
362    f32: AsPrimitive<T>,
363{
364    let b = b.as_();
365    (round_ties_to_even(b.abs() * a) * b.signum()).as_()
366}
367
368pub fn scale() -> TypedBinOp {
369    TypedBinOp(Box::new(Scale), None)
370}
371
372/// Offsets i8 integers as u8 integers.
373pub(crate) fn offset_i8_as_u8_elementwise(x: i8) -> u8 {
374    (x as u8).wrapping_add(128)
375}
376
377#[derive(Debug, Clone)]
378pub struct OffsetI8asU8 {}
379impl ElementWiseMiniOp for OffsetI8asU8 {
380    fn name(&self) -> String {
381        format!("{}{}", self.prefix(), stringify!(OffsetI8asU8))
382    }
383    fn output_type(&self, input_type: DatumType) -> Option<DatumType> {
384        Some(if let DatumType::QI8(qp) = input_type {
385            let (zp, scale) = qp.zp_scale();
386            DatumType::QU8(QParams::ZpScale { zero_point: zp + 128, scale })
387        } else if input_type == DatumType::I8 {
388            DatumType::U8
389        } else {
390            input_type
391        })
392    }
393    fn eval_out_of_place(&self, t: &Tensor, out_dt: Option<DatumType>) -> TractResult<Tensor> {
394        let output_type = out_dt.unwrap_or(self.output_type(t.datum_type()).unwrap());
395        let mut dst = unsafe { Tensor::uninitialized_dt(output_type, t.shape())? };
396        if t.datum_type().unquantized() == i8::datum_type() {
397            t.try_as_dense()?
398                .as_slice::<i8>()?
399                .iter()
400                .zip(dst.try_as_dense_mut()?.as_slice_mut::<u8>()?.iter_mut())
401                .for_each(|(x, y)| *y = offset_i8_as_u8_elementwise(*x));
402            return Ok(dst);
403        }
404
405        bail!("{} does not support {:?}", self.name(), t.datum_type());
406    }
407}
408
409pub fn offset_i8_as_u8() -> ElementWiseOp {
410    ElementWiseOp(Box::new(OffsetI8asU8 {}), None)
411}
412
413/// Offsets u8 integers as i8 integers.
414pub(crate) fn offset_u8_as_i8_elementwise(x: u8) -> i8 {
415    x.wrapping_sub(128) as i8
416}
417
418#[derive(Debug, Clone)]
419pub struct OffsetU8asI8 {}
420impl ElementWiseMiniOp for OffsetU8asI8 {
421    fn name(&self) -> String {
422        format!("{}{}", self.prefix(), stringify!(OffsetU8asI8))
423    }
424    fn output_type(&self, input_type: DatumType) -> Option<DatumType> {
425        Some(if let DatumType::QU8(qp) = input_type {
426            let (zp, scale) = qp.zp_scale();
427            DatumType::QI8(QParams::ZpScale { zero_point: zp - 128, scale })
428        } else if input_type == DatumType::U8 {
429            DatumType::I8
430        } else {
431            input_type
432        })
433    }
434    fn eval_out_of_place(&self, t: &Tensor, out_dt: Option<DatumType>) -> TractResult<Tensor> {
435        let output_type = out_dt.unwrap_or(self.output_type(t.datum_type()).unwrap());
436        let mut dst = unsafe { Tensor::uninitialized_dt(output_type, t.shape())? };
437        if t.datum_type().unquantized() == u8::datum_type() {
438            t.try_as_dense()?
439                .as_slice::<u8>()?
440                .iter()
441                .zip(dst.try_as_dense_mut()?.as_slice_mut::<i8>()?.iter_mut())
442                .for_each(|(x, y)| *y = offset_u8_as_i8_elementwise(*x));
443            return Ok(dst);
444        }
445
446        bail!("{} does not support {:?}", self.name(), t.datum_type());
447    }
448}
449pub fn offset_u8_as_i8() -> ElementWiseOp {
450    ElementWiseOp(Box::new(OffsetU8asI8 {}), None)
451}
452
453#[cfg(test)]
454pub mod scale {
455    use crate::internal::*;
456    use crate::ops::einsum::EinSum;
457    use crate::ops::math::round_ties_to_even;
458    use proptest::prelude::*;
459
460    fn test_scale(a: i8, b: i8, scale: f32) {
461        let expected = (((a as i32) * (b as i32)) as f32) / scale;
462        let expected = round_ties_to_even(expected.abs()) * expected.signum();
463        let expected = (expected as i32).clamp(-128, 127);
464        let expected = tensor2(&[[expected as i8]]);
465
466        let input = tvec!(tensor2(&[[b]]).into_tvalue());
467        let mut model = TypedModel::default();
468        let a = model.add_const("a", tensor2(&[[a]])).unwrap();
469        let b = model.add_source("b", i8::fact([1, 1])).unwrap();
470        let bias = model.add_const("bias", tensor0(0i32)).unwrap();
471        let a0 = model.add_const("a0", tensor0(0i8)).unwrap();
472        let a_scale = model.add_const("a_scale", tensor0(1f32)).unwrap();
473        let b0 = model.add_const("b0", tensor0(0i8)).unwrap();
474        let b_scale = model.add_const("b_scale", tensor0(1f32)).unwrap();
475        let c0 = model.add_const("c0", tensor0(0i8)).unwrap();
476        let c_scale = model.add_const("c_scale", tensor0(scale)).unwrap();
477        let op = EinSum {
478            axes: "mk,kn,,,,,,,->mn".parse().unwrap(),
479            operating_dt: i32::datum_type(),
480            q_params: Some(i8::datum_type()),
481        };
482        let output = model
483            .wire_node("mmm", op, &[a, b, bias, a0, a_scale, b0, b_scale, c0, c_scale])
484            .unwrap();
485        model.set_output_outlets(&output).unwrap();
486
487        let plain = model.clone().into_runnable().unwrap().run(input.clone()).unwrap();
488        assert_eq!(*plain[0], expected);
489
490        let optim = model.into_optimized().unwrap().into_runnable().unwrap().run(input).unwrap();
491        assert_eq!(*optim[0], expected);
492    }
493
494    proptest! {
495        #[test]
496        fn prop(a in any::<i8>(), b in any::<i8>(), scale in 0.00001f32..1000.) {
497            test_scale(a, b, scale);
498        }
499    }
500
501    #[test]
502    fn t1() {
503        test_scale(-117, 15, 37.753822);
504    }
505
506    #[test]
507    fn t2() {
508        test_scale(-4, -60, 475.21674);
509    }
510}
tract_core/ops/quant.rs

tract_core/ops/
quant.rs
