cubecl_cpp/shared/
unary.rs

1use super::{Component, Dialect, Elem, FmtLeft, Variable};
2use std::fmt::Display;
3
4pub trait Unary<D: Dialect> {
5    fn format(
6        f: &mut std::fmt::Formatter<'_>,
7        input: &Variable<D>,
8        out: &Variable<D>,
9    ) -> std::fmt::Result {
10        let out_item = out.item();
11
12        if out_item.vectorization == 1 {
13            write!(f, "{} = ", out.fmt_left())?;
14            Self::format_scalar(f, *input, out_item.elem)?;
15            f.write_str(";\n")
16        } else {
17            Self::unroll_vec(f, input, out, out_item.elem, out_item.vectorization)
18        }
19    }
20
21    fn format_scalar<Input: Component<D>>(
22        f: &mut std::fmt::Formatter<'_>,
23        input: Input,
24        out_elem: Elem<D>,
25    ) -> std::fmt::Result;
26
27    fn unroll_vec(
28        f: &mut std::fmt::Formatter<'_>,
29        input: &Variable<D>,
30        out: &Variable<D>,
31        out_elem: Elem<D>,
32        index: usize,
33    ) -> std::fmt::Result {
34        let mut write_op = |index, out_elem, input: &Variable<D>, out: &Variable<D>| {
35            let out_item = out.item();
36            let out = out.fmt_left();
37            writeln!(f, "{out} = {out_item}{{")?;
38
39            for i in 0..index {
40                let inputi = input.index(i);
41
42                Self::format_scalar(f, inputi, out_elem)?;
43                f.write_str(",")?;
44            }
45
46            f.write_str("};\n")
47        };
48
49        if Self::can_optimize() {
50            let optimized = Variable::optimized_args([*input, *out]);
51            let [input, out_optimized] = optimized.args;
52
53            let item_out_original = out.item();
54            let item_out_optimized = out_optimized.item();
55
56            let (index, out_elem) = match optimized.optimization_factor {
57                Some(factor) => (index / factor, out_optimized.elem()),
58                None => (index, out_elem),
59            };
60
61            if item_out_original != item_out_optimized {
62                let out_tmp = Variable::tmp(item_out_optimized);
63
64                write_op(index, out_elem, &input, &out_tmp)?;
65                let qualifier = out.const_qualifier();
66                let addr_space = D::address_space_for_variable(out);
67                let out_fmt = out.fmt_left();
68                writeln!(
69                    f,
70                    "{out_fmt} = reinterpret_cast<{addr_space}{item_out_original}{qualifier}&>({out_tmp});\n"
71                )
72            } else {
73                write_op(index, out_elem, &input, &out_optimized)
74            }
75        } else {
76            write_op(index, out_elem, input, out)
77        }
78    }
79
80    fn can_optimize() -> bool {
81        true
82    }
83}
84
85pub trait FunctionFmt<D: Dialect> {
86    fn base_function_name() -> &'static str;
87    fn function_name(elem: Elem<D>) -> String {
88        if Self::half_support() {
89            let prefix = match elem {
90                Elem::F16 | Elem::BF16 => D::compile_instruction_half_function_name_prefix(),
91                Elem::F16x2 | Elem::BF16x2 => D::compile_instruction_half2_function_name_prefix(),
92                _ => "",
93            };
94            format!("{prefix}{}", Self::base_function_name())
95        } else {
96            Self::base_function_name().into()
97        }
98    }
99    fn format_unary<Input: Display>(
100        f: &mut std::fmt::Formatter<'_>,
101        input: Input,
102        elem: Elem<D>,
103    ) -> std::fmt::Result {
104        if Self::half_support() {
105            write!(f, "{}({input})", Self::function_name(elem))
106        } else {
107            match elem {
108                Elem::F16 | Elem::F16x2 | Elem::BF16 | Elem::BF16x2 => {
109                    write!(f, "{}({}(float({input})))", elem, Self::function_name(elem))
110                }
111                _ => write!(f, "{}({input})", Self::function_name(elem)),
112            }
113        }
114    }
115
116    fn half_support() -> bool;
117}
118
119macro_rules! function {
120    ($name:ident, $func:expr) => {
121        function!($name, $func, true);
122    };
123    ($name:ident, $func:expr, $half_support:expr) => {
124        pub struct $name;
125
126        impl<D: Dialect> FunctionFmt<D> for $name {
127            fn base_function_name() -> &'static str {
128                $func
129            }
130            fn half_support() -> bool {
131                $half_support
132            }
133        }
134
135        impl<D: Dialect> Unary<D> for $name {
136            fn format_scalar<Input: Display>(
137                f: &mut std::fmt::Formatter<'_>,
138                input: Input,
139                elem: Elem<D>,
140            ) -> std::fmt::Result {
141                Self::format_unary(f, input, elem)
142            }
143
144            fn can_optimize() -> bool {
145                $half_support
146            }
147        }
148    };
149}
150
151function!(Log, "log");
152function!(FastLog, "__logf", false);
153function!(Cos, "cos");
154function!(FastCos, "__cosf", false);
155function!(Sin, "sin");
156function!(FastSin, "__sinf", false);
157function!(Sqrt, "sqrt");
158function!(InverseSqrt, "rsqrt");
159function!(FastSqrt, "__fsqrt_rn", false);
160function!(FastInverseSqrt, "__frsqrt_rn", false);
161function!(Exp, "exp");
162function!(FastExp, "__expf", false);
163function!(Ceil, "ceil");
164function!(Trunc, "trunc");
165function!(Floor, "floor");
166function!(Round, "rint");
167function!(FastRecip, "__frcp_rn", false);
168function!(FastTanh, "__tanhf", false);
169
170function!(Erf, "erf", false);
171function!(Abs, "abs", false);
172
173pub struct Log1p;
174
175impl<D: Dialect> Unary<D> for Log1p {
176    fn format_scalar<Input: Component<D>>(
177        f: &mut std::fmt::Formatter<'_>,
178        input: Input,
179        _out_elem: Elem<D>,
180    ) -> std::fmt::Result {
181        D::compile_instruction_log1p_scalar(f, input)
182    }
183
184    fn can_optimize() -> bool {
185        false
186    }
187}
188
189pub struct Tanh;
190
191impl<D: Dialect> Unary<D> for Tanh {
192    fn format_scalar<Input: Component<D>>(
193        f: &mut std::fmt::Formatter<'_>,
194        input: Input,
195        _out_elem: Elem<D>,
196    ) -> std::fmt::Result {
197        D::compile_instruction_tanh_scalar(f, input)
198    }
199
200    fn can_optimize() -> bool {
201        false
202    }
203}
204
205pub fn zero_extend<D: Dialect>(input: impl Component<D>) -> String {
206    match input.elem() {
207        Elem::I8 => format!("{}({}({input}))", Elem::<D>::U32, Elem::<D>::U8),
208        Elem::I16 => format!("{}({}({input}))", Elem::<D>::U32, Elem::<D>::U16),
209        Elem::U8 => format!("{}({input})", Elem::<D>::U32),
210        Elem::U16 => format!("{}({input})", Elem::<D>::U32),
211        _ => unreachable!("zero extend only supports integer < 32 bits"),
212    }
213}
214
215pub struct CountBits;
216
217impl<D: Dialect> Unary<D> for CountBits {
218    fn format_scalar<Input: Component<D>>(
219        f: &mut std::fmt::Formatter<'_>,
220        input: Input,
221        elem: Elem<D>,
222    ) -> std::fmt::Result {
223        D::compile_instruction_popcount_scalar(f, input, elem)
224    }
225}
226
227pub struct ReverseBits;
228
229impl<D: Dialect> Unary<D> for ReverseBits {
230    fn format_scalar<Input: Component<D>>(
231        f: &mut std::fmt::Formatter<'_>,
232        input: Input,
233        elem: Elem<D>,
234    ) -> std::fmt::Result {
235        D::compile_instruction_reverse_bits_scalar(f, input, elem)
236    }
237}
238
239pub struct LeadingZeros;
240
241impl<D: Dialect> Unary<D> for LeadingZeros {
242    fn format_scalar<Input: Component<D>>(
243        f: &mut std::fmt::Formatter<'_>,
244        input: Input,
245        elem: Elem<D>,
246    ) -> std::fmt::Result {
247        D::compile_instruction_leading_zeros_scalar(f, input, elem)
248    }
249}
250
251pub struct FindFirstSet;
252
253impl<D: Dialect> Unary<D> for FindFirstSet {
254    fn format_scalar<Input: Component<D>>(
255        f: &mut std::fmt::Formatter<'_>,
256        input: Input,
257        out_elem: Elem<D>,
258    ) -> std::fmt::Result {
259        D::compile_instruction_find_first_set(f, input, out_elem)
260    }
261}
262
263pub struct BitwiseNot;
264
265impl<D: Dialect> Unary<D> for BitwiseNot {
266    fn format_scalar<Input>(
267        f: &mut std::fmt::Formatter<'_>,
268        input: Input,
269        _out_elem: Elem<D>,
270    ) -> std::fmt::Result
271    where
272        Input: Component<D>,
273    {
274        write!(f, "~{input}")
275    }
276}
277
278pub struct Not;
279
280impl<D: Dialect> Unary<D> for Not {
281    fn format_scalar<Input>(
282        f: &mut std::fmt::Formatter<'_>,
283        input: Input,
284        _out_elem: Elem<D>,
285    ) -> std::fmt::Result
286    where
287        Input: Component<D>,
288    {
289        write!(f, "!{input}")
290    }
291}
292
293pub struct Assign;
294
295impl<D: Dialect> Unary<D> for Assign {
296    fn format(
297        f: &mut std::fmt::Formatter<'_>,
298        input: &Variable<D>,
299        out: &Variable<D>,
300    ) -> std::fmt::Result {
301        let item = out.item();
302
303        if item.vectorization == 1 || input.item() == item {
304            write!(f, "{} = ", out.fmt_left())?;
305            Self::format_scalar(f, *input, item.elem)?;
306            f.write_str(";\n")
307        } else {
308            Self::unroll_vec(f, input, out, item.elem, item.vectorization)
309        }
310    }
311
312    fn format_scalar<Input>(
313        f: &mut std::fmt::Formatter<'_>,
314        input: Input,
315        elem: Elem<D>,
316    ) -> std::fmt::Result
317    where
318        Input: Component<D>,
319    {
320        // Cast only when necessary.
321        if elem != input.elem() {
322            match elem {
323                Elem::TF32 => write!(f, "nvcuda::wmma::__float_to_tf32({input})"),
324                elem => write!(f, "{elem}({input})"),
325            }
326        } else {
327            write!(f, "{input}")
328        }
329    }
330}
331
332fn elem_function_name<D: Dialect>(base_name: &'static str, elem: Elem<D>) -> String {
333    // Math functions prefix (no leading underscores)
334    let prefix = match elem {
335        Elem::F16 | Elem::BF16 => D::compile_instruction_half_function_name_prefix(),
336        Elem::F16x2 | Elem::BF16x2 => D::compile_instruction_half2_function_name_prefix(),
337        _ => "",
338    };
339    if prefix.is_empty() {
340        base_name.to_string()
341    } else if prefix == "h" || prefix == "h2" {
342        format!("__{prefix}{base_name}")
343    } else {
344        panic!("Unknown prefix '{prefix}'");
345    }
346}
347
348// `isnan` / `isinf` are defined for cuda/hip/metal with same prefixes for half/bf16 on cuda/hip
349pub struct IsNan;
350
351impl<D: Dialect> Unary<D> for IsNan {
352    fn format_scalar<Input: Component<D>>(
353        f: &mut std::fmt::Formatter<'_>,
354        input: Input,
355        _elem: Elem<D>,
356    ) -> std::fmt::Result {
357        // Format unary function name based on *input* elem dtype
358        let elem = input.elem();
359        write!(f, "{}({input})", elem_function_name("isnan", elem))
360    }
361
362    fn can_optimize() -> bool {
363        true
364    }
365}
366
367pub struct IsInf;
368
369impl<D: Dialect> Unary<D> for IsInf {
370    fn format_scalar<Input: Component<D>>(
371        f: &mut std::fmt::Formatter<'_>,
372        input: Input,
373        _elem: Elem<D>,
374    ) -> std::fmt::Result {
375        // Format unary function name based on *input* elem dtype
376        let elem = input.elem();
377        write!(f, "{}({input})", elem_function_name("isinf", elem))
378    }
379
380    fn can_optimize() -> bool {
381        true
382    }
383}
cubecl_cpp/shared/unary.rs

cubecl_cpp/shared/
unary.rs
