cutile-compiler 0.0.0-alpha

Crate for compiling kernels authored in cuTile Rust to executable kernels.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
/*
 * SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 * SPDX-License-Identifier: Apache-2.0
 */

//! Binary operation compilation: handles `compile_binary_op` and `compile_binary_op_from_values`
//! within the CUDA Tile compiler. This module covers the translation of binary arithmetic,
//! comparison, and bitwise operations into MLIR operations.

use quote::ToTokens;
use syn::spanned::Spanned;

use crate::bounds::bounds_from_bop;
use crate::compiler::_function::CUDATileFunctionCompiler;
pub use crate::compiler::_type::*;
pub use crate::compiler::_value::*;
use crate::compiler::utils::{
    cuda_tile_tile_ty_from_type_instance, get_cmp_predicate_attr, get_tile_bop_from_rust_bop,
    TileBinaryOp,
};
use crate::error::JITError;
use crate::generics::GenericVars;
use melior::ir::operation::{OperationBuilder, OperationLike};
use melior::ir::{self, BlockLike, Value};
use std::collections::HashMap;
use syn::ExprBinary;

impl<'m, 'c> CUDATileFunctionCompiler<'m> {
    pub fn compile_binary_op(
        &'c self,
        builder: &'c ir::Block<'c>,
        bin_expr: &ExprBinary,
        generic_vars: &GenericVars,
        ctx: &mut CompilerContext<'c, 'c>,
        return_type: Option<TileRustType<'c>>,
    ) -> Result<Option<TileRustValue<'c, 'c>>, JITError> {
        let lhs = self.compile_expression(
            builder,
            &bin_expr.left,
            generic_vars,
            ctx,
            return_type.clone(),
        )?;
        if lhs.is_none() {
            return self.jit_error_result(
                &bin_expr.left.span(),
                "failed to compile the left-hand side of this binary operation",
            );
        }
        let lhs = lhs.unwrap();
        let rhs = self.compile_expression(
            builder,
            &bin_expr.right,
            generic_vars,
            ctx,
            return_type.clone(),
        )?;
        if rhs.is_none() {
            return self.jit_error_result(
                &bin_expr.right.span(),
                "failed to compile the right-hand side of this binary operation",
            );
        }
        let rhs = rhs.unwrap();
        Ok(Some(self.compile_binary_op_from_values(
            builder,
            lhs,
            rhs,
            &get_tile_bop_from_rust_bop(&bin_expr.op)?,
            generic_vars,
            ctx,
            return_type,
            &bin_expr.span(),
        )?))
    }

    pub fn compile_binary_op_from_values(
        &'c self,
        builder: &'c ir::Block<'c>,
        lhs: TileRustValue<'c, 'c>,
        rhs: TileRustValue<'c, 'c>,
        tile_rust_arithmetic_op: &TileBinaryOp,
        generic_vars: &GenericVars,
        _ctx: &mut CompilerContext<'c, 'c>,
        return_type: Option<TileRustType<'c>>,
        span: &proc_macro2::Span,
    ) -> Result<TileRustValue<'c, 'c>, JITError> {
        if lhs.ty.rust_ty != rhs.ty.rust_ty {
            return self.jit_error_result(
                span,
                &format!(
                    "binary `{:?}` requires operands of the same type, but got `{}` and `{}`",
                    tile_rust_arithmetic_op,
                    lhs.ty.rust_ty.to_token_stream().to_string(),
                    rhs.ty.rust_ty.to_token_stream().to_string()
                ),
            );
        }
        let lhs_value = lhs.value;
        if lhs_value.is_none() {
            return self.jit_error_result(
                span,
                "left-hand side of binary operation did not produce a value",
            );
        }
        let lhs_value = lhs_value.unwrap();
        let rhs_value = rhs.value;
        if rhs_value.is_none() {
            return self.jit_error_result(
                span,
                "right-hand side of binary operation did not produce a value",
            );
        }
        let rhs_value = rhs_value.unwrap();
        let operand_type = lhs.ty.clone();
        let operand_rust_ty = &operand_type.rust_ty;
        let Some(operand_rust_element_type) =
            operand_type.get_instantiated_rust_element_type(&self.modules.primitives)
        else {
            return self.jit_error_result(
                span,
                &format!(
                    "unable to determine element type for `{:?}` on `{}`",
                    tile_rust_arithmetic_op,
                    operand_type.rust_ty.to_token_stream().to_string()
                ),
            );
        };
        let Some(operand_cuda_tile_ty) = operand_type.cuda_tile_ty else {
            return self.jit_error_result(
                span,
                &format!(
                    "type `{}` cannot be used with binary `{:?}`",
                    operand_type.rust_ty.to_token_stream().to_string(),
                    tile_rust_arithmetic_op
                ),
            );
        };
        let Some(operand_cuda_tile_element_type) =
            operand_type.get_cuda_tile_element_type(&self.modules.primitives)?
        else {
            return self.jit_error_result(
                span,
                &format!(
                    "unable to determine compiled element type for `{:?}`",
                    tile_rust_arithmetic_op
                ),
            );
        };
        let mut is_cmp = false;
        let signedness_str = match operand_rust_element_type.as_str() {
            "bool" | "u32" | "u64" => "unsigned",
            _ => "signed",
        };
        let signedness_attr = self.parse_named_attr(
            "signedness",
            format!("#cuda_tile.signedness<{signedness_str}>").as_str(),
        )?;

        let op_builder = match operand_cuda_tile_element_type.as_ref() {
            "i1" | "i4" | "i8" | "i32" | "i64" => {
                // TODO (hme): Add i4, i8, i16 support, as needed.
                if let Some(comparison_predicate) =
                    get_cmp_predicate_attr(&self.context, tile_rust_arithmetic_op)?
                {
                    is_cmp = true;
                    let cuda_tile_bool_ty = cuda_tile_tile_ty_from_type_instance(
                        &self.context,
                        &lhs.ty.type_instance,
                        &self.modules.primitives,
                        Some("i1"),
                    )?;
                    OperationBuilder::new("cuda_tile.cmpi", self.function_location())
                        .add_attributes(&[comparison_predicate, signedness_attr])
                        .add_operands(&[lhs_value, rhs_value])
                        .add_results(&[cuda_tile_bool_ty])
                } else {
                    // If both operands have bounds, we can generate bounds on the output.
                    match tile_rust_arithmetic_op {
                        TileBinaryOp::Min => {
                            OperationBuilder::new("cuda_tile.mini", self.function_location())
                                .add_operands(&[lhs_value, rhs_value])
                                .add_attributes(&[signedness_attr])
                                .add_results(&[operand_cuda_tile_ty])
                        }
                        TileBinaryOp::Max => {
                            OperationBuilder::new("cuda_tile.maxi", self.function_location())
                                .add_operands(&[lhs_value, rhs_value])
                                .add_attributes(&[signedness_attr])
                                .add_results(&[operand_cuda_tile_ty])
                        }
                        TileBinaryOp::Add => {
                            OperationBuilder::new("cuda_tile.addi", self.function_location())
                                .add_operands(&[lhs_value, rhs_value])
                                .add_results(&[operand_cuda_tile_ty])
                        }
                        TileBinaryOp::Sub => {
                            OperationBuilder::new("cuda_tile.subi", self.function_location())
                                .add_operands(&[lhs_value, rhs_value])
                                .add_results(&[operand_cuda_tile_ty])
                        }
                        TileBinaryOp::Mul => {
                            OperationBuilder::new("cuda_tile.muli", self.function_location())
                                .add_operands(&[lhs_value, rhs_value])
                                .add_results(&[operand_cuda_tile_ty])
                        }
                        TileBinaryOp::Rem => {
                            OperationBuilder::new("cuda_tile.remi", self.function_location())
                                .add_operands(&[lhs_value, rhs_value])
                                .add_attributes(&[signedness_attr])
                                .add_results(&[operand_cuda_tile_ty])
                        }
                        TileBinaryOp::Div => {
                            let rounding_mode_attr = self.parse_named_attr(
                                "rounding_mode",
                                "#cuda_tile.rounding<negative_inf>",
                            )?;
                            OperationBuilder::new("cuda_tile.divi", self.function_location())
                                .add_operands(&[lhs_value, rhs_value])
                                .add_results(&[operand_cuda_tile_ty])
                                .add_attributes(&[signedness_attr, rounding_mode_attr])
                        }
                        TileBinaryOp::CeilDiv => {
                            let rounding_mode_attr = self.parse_named_attr(
                                "rounding_mode",
                                "#cuda_tile.rounding<positive_inf>",
                            )?;
                            OperationBuilder::new("cuda_tile.divi", self.function_location())
                                .add_operands(&[lhs_value, rhs_value])
                                .add_results(&[operand_cuda_tile_ty])
                                .add_attributes(&[signedness_attr, rounding_mode_attr])
                        }
                        TileBinaryOp::BitAnd => {
                            OperationBuilder::new("cuda_tile.andi", self.function_location())
                                .add_operands(&[lhs_value, rhs_value])
                                .add_results(&[operand_cuda_tile_ty])
                        }
                        TileBinaryOp::BitOr => {
                            OperationBuilder::new("cuda_tile.ori", self.function_location())
                                .add_operands(&[lhs_value, rhs_value])
                                .add_results(&[operand_cuda_tile_ty])
                        }
                        TileBinaryOp::BitXor => {
                            OperationBuilder::new("cuda_tile.xori", self.function_location())
                                .add_operands(&[lhs_value, rhs_value])
                                .add_results(&[operand_cuda_tile_ty])
                        }
                        _ => {
                            return self.jit_error_result(
                                span,
                                &format!("Unimplemented binary op {tile_rust_arithmetic_op:#?}"),
                            );
                        }
                    }
                }
            }
            "f16" | "f32" | "f64" => {
                if let Some(comparison_predicate) =
                    get_cmp_predicate_attr(&self.context, tile_rust_arithmetic_op)?
                {
                    let comparison_ordering = self.parse_named_attr(
                        "comparison_ordering",
                        format!("#cuda_tile.comparison_ordering<ordered>").as_str(),
                    )?;
                    is_cmp = true;
                    let cuda_tile_bool_ty = cuda_tile_tile_ty_from_type_instance(
                        &self.context,
                        &lhs.ty.type_instance,
                        &self.modules.primitives,
                        Some("i1"),
                    )?;
                    OperationBuilder::new("cuda_tile.cmpf", self.function_location())
                        .add_attributes(&[comparison_predicate, comparison_ordering])
                        .add_operands(&[lhs_value, rhs_value])
                        .add_results(&[cuda_tile_bool_ty])
                } else {
                    let default_rounding_mode_attr = self
                        .parse_named_attr("rounding_mode", "#cuda_tile.rounding<nearest_even>")?;
                    let attrs = vec![default_rounding_mode_attr];
                    match tile_rust_arithmetic_op {
                        TileBinaryOp::Min => {
                            OperationBuilder::new("cuda_tile.minf", self.function_location())
                                .add_attributes(&attrs)
                                .add_operands(&[lhs_value, rhs_value])
                                .add_results(&[operand_cuda_tile_ty])
                        }
                        TileBinaryOp::Max => {
                            OperationBuilder::new("cuda_tile.maxf", self.function_location())
                                .add_attributes(&attrs)
                                .add_operands(&[lhs_value, rhs_value])
                                .add_results(&[operand_cuda_tile_ty])
                        }
                        TileBinaryOp::Add => {
                            OperationBuilder::new("cuda_tile.addf", self.function_location())
                                .add_attributes(&attrs)
                                .add_operands(&[lhs_value, rhs_value])
                                .add_results(&[operand_cuda_tile_ty])
                        }
                        TileBinaryOp::Sub => {
                            OperationBuilder::new("cuda_tile.subf", self.function_location())
                                .add_attributes(&attrs)
                                .add_operands(&[lhs_value, rhs_value])
                                .add_results(&[operand_cuda_tile_ty])
                        }
                        TileBinaryOp::Mul => {
                            OperationBuilder::new("cuda_tile.mulf", self.function_location())
                                .add_attributes(&attrs)
                                .add_operands(&[lhs_value, rhs_value])
                                .add_results(&[operand_cuda_tile_ty])
                        }
                        TileBinaryOp::Rem => {
                            OperationBuilder::new("cuda_tile.remf", self.function_location())
                                .add_operands(&[lhs_value, rhs_value])
                                .add_results(&[operand_cuda_tile_ty])
                        }
                        TileBinaryOp::Div => {
                            OperationBuilder::new("cuda_tile.divf", self.function_location())
                                .add_attributes(&attrs)
                                .add_operands(&[lhs_value, rhs_value])
                                .add_results(&[operand_cuda_tile_ty])
                        }
                        TileBinaryOp::TrueDiv => {
                            let rounding_mode_attr = self
                                .parse_named_attr("rounding_mode", "#cuda_tile.rounding<approx>")?;
                            let attrs = match operand_cuda_tile_element_type.as_ref() {
                                "f32" => {
                                    let flush_to_zero = self.flag_attr("flush_to_zero");
                                    vec![rounding_mode_attr, flush_to_zero]
                                }
                                "f16" | "f64" => {
                                    vec![rounding_mode_attr]
                                }
                                _ => unreachable!("Impossible"),
                            };
                            OperationBuilder::new("cuda_tile.divf", self.function_location())
                                .add_attributes(&attrs)
                                .add_operands(&[lhs_value, rhs_value])
                                .add_results(&[operand_cuda_tile_ty])
                        }
                        _ => {
                            return self.jit_error_result(
                                span,
                                &format!("Unimplemented binary op {tile_rust_arithmetic_op:#?}"),
                            );
                        }
                    }
                }
            }
            _ => {
                return self.jit_error_result(
                    span,
                    &format!(
                        "Binary operation is not implemented for {}",
                        operand_rust_ty.to_token_stream().to_string()
                    ),
                );
            }
        };

        let return_type = match return_type {
            Some(rt) => rt,
            None => {
                // Try to infer from lhs/rhs.
                if is_cmp {
                    let bool_ty = syn::parse2::<syn::Type>("bool".parse().unwrap()).unwrap();
                    self.compile_type(&bool_ty, &generic_vars, &HashMap::new())?
                        .unwrap()
                } else {
                    operand_type
                }
            }
        };

        let op_bounds = if let (Some(a), Some(b)) = (lhs.bounds, rhs.bounds) {
            if !(lhs.kind == Kind::PrimitiveType && rhs.kind == Kind::PrimitiveType) {
                return self.jit_error_result(
                    span,
                    &format!(
                        "Expected PrimitiveType for binary op bounds, got lhs={:#?}, rhs={:#?}",
                        lhs.kind, rhs.kind
                    ),
                );
            }
            bounds_from_bop(tile_rust_arithmetic_op, &a, &b)
        } else {
            None
        };
        if let Some(bounds) = &op_bounds {
            if bounds.is_exact() {
                // The lower/upper bounds are equivalent.
                return Ok(self.compile_constant_from_exact_bounds(
                    builder,
                    bounds.clone(),
                    return_type,
                )?);
            }
        }

        let op = op_builder.build();
        if op.is_err() {
            return self.jit_error_result(
                span,
                &format!("Failed to compile {tile_rust_arithmetic_op:#?}"),
            );
        }
        let op_ref = builder.append_operation(op.unwrap().into());
        if !op_ref.verify() {
            return self.jit_error_result(
                span,
                &format!("Failed to compile {tile_rust_arithmetic_op:#?}"),
            );
        }
        let value: Value = op_ref.result(0).unwrap().into();
        let mut tr_value = TileRustValue::new_value_kind_like(value, return_type.clone());
        tr_value.bounds = op_bounds;
        Ok(tr_value)
    }
}