morok-schedule 0.1.0-alpha.2

Optimization passes and pattern engine for the Morok ML compiler
Documentation 
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

//! Tests for do_expand (UNROLL propagation through operations).
//!
//! Ported from Tinygrad's TestExpander class (test_uop_graph.py:663-811).
//!
//! do_expand replicates operations that have UNROLL inputs:
//! - Broadcasts scalar operands
//! - Swizzles UNROLL operands with different axes
//! - Wraps results in UNROLL
//!
//! Value assertions match Tinygrad's exact test expectations.

use super::helpers::*;
use morok_dtype::DType;
use morok_ir::UOp;
use morok_ir::types::ConstValue;

// =============================================================================
// Broadcast Expansion Tests
// =============================================================================

/// Test: UNROLL + scalar broadcast
///
/// Tinygrad: test_expand_add_broadcast
/// ```python
/// e1 = UOp(Ops.UNROLL, dtypes.int, (UOp.const(dtypes.int.vec(4), tuple(range(4))),), ((1,4),))
/// sink = expander_rewrite(e1+3)
/// self.assertTupleEqual(sink.src[0].arg, (3,4,5,6))
/// ```
#[test]
fn test_expand_add_broadcast() {
    // Create UNROLL(VCONST([0,1,2,3]), [(1,4)])
    let unroll = create_unroll_iota(1, 4);

    // Add scalar constant 3
    let scalar = UOp::const_(DType::Int64, ConstValue::Int(3));
    let add = unroll.try_add(&scalar).unwrap();

    // Apply expander
    let result = phase2_only(&add);

    // Assert exact values like Tinygrad: (3, 4, 5, 6)
    assert_result_values(&result, &[3, 4, 5, 6]);

    // Also verify UNROLL structure
    let (_, axes) = unwrap_unroll(&result);
    assert_eq!(axes, vec![(1, 4)], "Should preserve axis");
}

// =============================================================================
// Same-Axis Expansion Tests
// =============================================================================

/// Test: Two UNROLLs with same axis
///
/// Tinygrad: test_expand_same_axis
/// ```python
/// e1 = UOp(Ops.UNROLL, dtypes.int, (UOp.const(dtypes.int.vec(4), tuple(range(4))),), ((1,4),))
/// e2 = UOp(Ops.UNROLL, dtypes.int, (UOp.const(dtypes.int.vec(4), tuple(x*4 for x in range(4))),), ((1,4),))
/// sink = expander_rewrite(e1+e2)
/// self.assertTupleEqual(sink.src[0].arg, (0, 5, 10, 15))
/// ```
#[test]
fn test_expand_same_axis() {
    // Create UNROLL(VCONST([0,1,2,3]), [(1,4)])
    let e1 = create_unroll_iota(1, 4);

    // Create UNROLL(VCONST([0,4,8,12]), [(1,4)])
    let e2 = create_unroll_scaled(1, 4, 4);

    // Add them
    let add = e1.try_add(&e2).unwrap();

    // Apply expander
    let result = phase2_only(&add);

    // Assert exact values: 0+0=0, 1+4=5, 2+8=10, 3+12=15
    assert_result_values(&result, &[0, 5, 10, 15]);

    // Verify UNROLL structure
    let (_, axes) = unwrap_unroll(&result);
    assert_eq!(axes, vec![(1, 4)], "Should preserve axis");
}

// =============================================================================
// Different-Axis Expansion Tests
// =============================================================================

/// Test: Two UNROLLs with different axes
///
/// Tinygrad: test_expand_different_axis
/// ```python
/// e1 = UOp(Ops.UNROLL, dtypes.int, (UOp.const(dtypes.int.vec(4), tuple(x*4 for x in range(4))),), ((1,4),))
/// e2 = UOp(Ops.UNROLL, dtypes.int, (UOp.const(dtypes.int.vec(4), tuple(range(4))),), ((2,4),))
/// sink = expander_rewrite(e1+e2)
/// self.assertTupleEqual(sink.arg, ((1, 4), (2, 4)))
/// self.assertTupleEqual(sink.src[0].arg, tuple(range(16)))
/// ```
#[test]
fn test_expand_different_axis() {
    // Create UNROLL(VCONST([0,4,8,12]), [(1,4)])
    let e1 = create_unroll_scaled(1, 4, 4);

    // Create UNROLL(VCONST([0,1,2,3]), [(2,4)])
    let e2 = create_unroll_iota(2, 4);

    // Add them
    let add = e1.try_add(&e2).unwrap();

    // Apply expander
    let result = phase2_only(&add);

    // When combining different axes, the result expands to 4*4=16 values.
    // Values follow the pattern: axis1_val + axis2_val
    // Row-major iteration: axis 1 is outer (slower), axis 2 is inner (faster)
    // (0,0)=0, (0,1)=1, (0,2)=2, (0,3)=3, (1,0)=4, (1,1)=5, ...
    // = 0+0, 0+1, 0+2, 0+3, 4+0, 4+1, 4+2, 4+3, 8+0, 8+1, 8+2, 8+3, 12+0, 12+1, 12+2, 12+3
    let expected: Vec<i64> = (0..16).collect();
    assert_result_values(&result, &expected);

    // Verify axes
    let (_, axes) = unwrap_unroll(&result);
    assert_eq!(axes, vec![(1, 4), (2, 4)], "Should have both axes");
}

/// Test: Two UNROLLs with different axes (operands flipped)
///
/// Tinygrad: test_expand_different_axis_flip
/// Same values but operands reversed.
#[test]
fn test_expand_different_axis_flip() {
    // Create UNROLL(VCONST([0,1,2,3]), [(2,4)])
    let e2 = create_unroll_iota(2, 4);

    // Create UNROLL(VCONST([0,4,8,12]), [(1,4)])
    let e1 = create_unroll_scaled(1, 4, 4);

    // Add them (flipped order)
    let add = e2.try_add(&e1).unwrap();

    // Apply expander
    let result = phase2_only(&add);

    // Same result as test_expand_different_axis (addition is commutative)
    let expected: Vec<i64> = (0..16).collect();
    assert_result_values(&result, &expected);

    // Verify axes
    let (_, axes) = unwrap_unroll(&result);
    assert_eq!(axes, vec![(1, 4), (2, 4)], "Should have both axes");
}

// =============================================================================
// Three-Axis Expansion Tests
// =============================================================================

/// Test: Three UNROLLs with different axes
///
/// This extends Tinygrad's pattern to verify 3D expansion.
#[test]
fn test_expand_three_axes() {
    // Create UNROLL with axis 1 (stride 4): [0, 4, 8, 12]
    let e1 = create_unroll_scaled(1, 4, 4);

    // Create UNROLL with axis 2 (stride 1): [0, 1, 2, 3]
    let e2 = create_unroll_iota(2, 4);

    // Create UNROLL with axis 3 (stride 16): [0, 16, 32, 48]
    let e3 = create_unroll_scaled(3, 4, 16);

    // Build: e1 + e2 + e3
    let sum = e1.try_add(&e2).unwrap().try_add(&e3).unwrap();

    // Apply expander
    let result = phase2_only(&sum);

    // Result should have 4*4*4=64 elements
    // Verify axes
    let (src, axes) = unwrap_unroll(&result);
    assert_eq!(axes, vec![(1, 4), (2, 4), (3, 4)], "Should have three axes");
    assert_eq!(src.dtype().vcount(), 64, "Inner should be vec64");
}

// =============================================================================
// Multiplication Expansion Tests
// =============================================================================

/// Test: UNROLL * scalar
#[test]
fn test_expand_mul_broadcast() {
    // Create UNROLL(VCONST([0,1,2,3]), [(1,4)])
    let unroll = create_unroll_iota(1, 4);

    // Multiply by scalar 2
    let scalar = UOp::const_(DType::Int64, ConstValue::Int(2));
    let mul = unroll.try_mul(&scalar).unwrap();

    // Apply expander
    let result = phase2_only(&mul);

    // Expected: [0*2, 1*2, 2*2, 3*2] = [0, 2, 4, 6]
    assert_result_values(&result, &[0, 2, 4, 6]);
}

/// Test: Two UNROLLs multiplied (same axis)
#[test]
fn test_expand_mul_same_axis() {
    // Create UNROLL(VCONST([1,2,3,4]), [(1,4)])
    let e1 = create_unroll_values(1, vec![1, 2, 3, 4]);

    // Create UNROLL(VCONST([1,2,3,4]), [(1,4)])
    let e2 = create_unroll_values(1, vec![1, 2, 3, 4]);

    // Multiply them
    let mul = e1.try_mul(&e2).unwrap();

    // Apply expander
    let result = phase2_only(&mul);

    // Expected: [1*1, 2*2, 3*3, 4*4] = [1, 4, 9, 16]
    assert_result_values(&result, &[1, 4, 9, 16]);
}

// =============================================================================
// Subtraction Expansion Tests
// =============================================================================

/// Test: UNROLL - scalar
#[test]
fn test_expand_sub_broadcast() {
    // Create UNROLL(VCONST([10,20,30,40]), [(1,4)])
    let unroll = create_unroll_values(1, vec![10, 20, 30, 40]);

    // Subtract scalar 5
    let scalar = UOp::const_(DType::Int64, ConstValue::Int(5));
    let sub = unroll.try_sub(&scalar).unwrap();

    // Apply expander
    let result = phase2_only(&sub);

    // Expected: [10-5, 20-5, 30-5, 40-5] = [5, 15, 25, 35]
    assert_result_values(&result, &[5, 15, 25, 35]);
}

// =============================================================================
// Mixed Operations Expansion Tests
// =============================================================================

/// Test: (UNROLL + scalar) * UNROLL
///
/// Verifies that compound expressions expand correctly.
#[test]
fn test_expand_compound_expression() {
    // Create UNROLL(VCONST([0,1,2,3]), [(1,4)])
    let e1 = create_unroll_iota(1, 4);

    // Create UNROLL(VCONST([2,2,2,2]), [(1,4)])
    let e2 = create_unroll_values(1, vec![2, 2, 2, 2]);

    // Build: (e1 + 1) * e2 = ([0,1,2,3] + 1) * [2,2,2,2]
    let scalar = UOp::const_(DType::Int64, ConstValue::Int(1));
    let sum = e1.try_add(&scalar).unwrap();
    let result = phase2_only(&sum.try_mul(&e2).unwrap());

    // Expected: [1*2, 2*2, 3*2, 4*2] = [2, 4, 6, 8]
    assert_result_values(&result, &[2, 4, 6, 8]);
}