ug 0.3.0

Micro compiler for tensor operations.
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
267
268
269
270
271
272
273

pub mod ssa {
    use crate::lang::ssa;
    use crate::lang::ssa::{BinaryOp, Const, DType, Instr as I, Kernel, VarId, A};
    use crate::Result;

    fn arg(index: usize, dtype: DType) -> I {
        I::DefineGlobal { index, dtype }
    }

    pub fn simple_add(vec_len: usize) -> Result<Kernel> {
        let v = VarId::new;
        let a = |i| A::Var(VarId::new(i));
        let i32 = |i| A::Const(Const::I32(i));
        let dtype = DType::I32;
        let instrs = vec![
            /* 0 */ arg(0, DType::I32),
            /* 1 */ arg(1, DType::I32),
            /* 2 */ arg(2, DType::I32),
            /* 3 */ I::Range { lo: i32(0), up: i32(vec_len as i32), end_idx: v(8), step: 1 },
            /* 4 */ I::Load { src: v(1), offset: a(3), dtype },
            /* 5 */ I::Load { src: v(2), offset: a(3), dtype },
            /* 6 */ I::Binary { op: self::BinaryOp::Add, lhs: a(4), rhs: a(5), dtype },
            /* 7 */ I::Store { dst: v(0), offset: a(3), value: a(6), dtype },
            /* 8 */ I::EndRange { start_idx: v(3) },
        ];
        Kernel::from_instrs(instrs)
    }

    pub fn simple_dotprod(vec_len: usize) -> Result<Kernel> {
        let v = VarId::new;
        let a = |i| A::Var(VarId::new(i));
        let dtype = DType::F32;
        let instrs = vec![
            /* 0 */ arg(0, DType::F32),
            /* 1 */ arg(1, DType::F32),
            /* 2 */ arg(2, DType::F32),
            /* 3 */ I::Const(Const::I32(0)),
            /* 4 */ I::Const(Const::I32(vec_len as i32)),
            /* 5 */ I::DefineAcc(0f32.try_into()?),
            /* 6 */ I::Range { lo: a(3), up: a(4), end_idx: v(12), step: 1 },
            /* 7 */ I::Load { src: v(1), offset: a(6), dtype },
            /* 8 */ I::Load { src: v(2), offset: a(6), dtype },
            /* 9 */ I::Binary { op: self::BinaryOp::Mul, lhs: a(7), rhs: a(8), dtype },
            /* 10*/ I::Binary { op: self::BinaryOp::Add, lhs: a(9), rhs: a(5), dtype },
            /* 11*/ I::Assign { dst: v(5), src: a(10) },
            /* 12*/ I::EndRange { start_idx: v(6) },
            /* 13*/ I::Store { dst: v(0), offset: a(3), value: a(5), dtype },
        ];
        Kernel::from_instrs(instrs)
    }

    pub fn exp(block_size: usize) -> Result<Kernel> {
        let mut b = crate::block::Block::empty();
        let dtype = DType::F32;
        let src_i = b.push(arg(0, DType::F32));
        let dst_i = b.push(arg(1, DType::F32));
        let g_i = b.push(I::Special(ssa::Special::BlockIdx));
        let l_i = b.push(I::Special(ssa::Special::ThreadIdx));
        let off_i = b.mul(g_i, block_size as i32);
        let off_i = b.binary(BinaryOp::Add, off_i, l_i, DType::I32);
        let load_i = b.push(I::Load { src: src_i.to_varid(), offset: off_i.to_a(), dtype });
        let value_i = b.unary(ssa::UnaryOp::Exp, load_i, dtype);
        b.push(I::Store {
            dst: dst_i.to_varid(),
            offset: off_i.to_a(),
            value: value_i.to_a(),
            dtype,
        });
        let instrs = b.relocate()?;
        Kernel::from_instrs(instrs)
    }

    pub fn exp_block(dim2: usize, block_size: usize) -> Result<Kernel> {
        if dim2 % block_size != 0 {
            crate::bail!("last-dim {dim2} must be divisible by block size {block_size}")
        }

        let mut b = crate::block::Block::empty();
        let dtype = DType::F32;
        let src_i = b.push(arg(0, DType::F32));
        let dst_i = b.push(arg(1, DType::F32));
        let g_i = b.push(I::Special(ssa::Special::BlockIdx));
        let l_i = b.push(I::Special(ssa::Special::ThreadIdx));
        let off_i = b.mul(g_i, dim2 as i32);
        let off_i = b.binary(BinaryOp::Add, off_i, l_i, DType::I32);

        for i in (0..dim2).step_by(block_size) {
            let off_i = b.add(off_i, i as i32);
            let load_i = b.push(I::Load { src: src_i.to_varid(), offset: off_i.to_a(), dtype });
            let value_i = b.unary(ssa::UnaryOp::Exp, load_i, dtype);
            b.push(I::Store {
                dst: dst_i.to_varid(),
                offset: off_i.to_a(),
                value: value_i.to_a(),
                dtype,
            });
        }
        Kernel::from_instrs(b.relocate()?)
    }

    pub fn softmax_barrier(_dim1: usize, dim2: usize) -> Result<Kernel> {
        let mut b = crate::block::Block::empty();
        let dtype = DType::F32;
        let src_i = b.push(arg(0, DType::F32));
        let dst_i = b.push(arg(1, DType::F32));
        let sto_i = b.push(I::DefineLocal { size: (2 * dim2), dtype: DType::F32 }).to_varid();
        let g_i = b.push(I::Special(ssa::Special::BlockIdx));
        let l_i = b.push(I::Special(ssa::Special::ThreadIdx));
        let base_off_i = b.mul(g_i, dim2 as i32);
        let global_off_i = b.binary(BinaryOp::Add, base_off_i, l_i, DType::I32);
        let load_i = b.push(I::Load { src: src_i.to_varid(), offset: global_off_i.to_a(), dtype });

        // Compute the max value over dim2.
        // This implementation uses shared memory which is likely slower than using warp-reduce
        // primitives but these are harder to scale to more than 32 threads.
        b.push(I::Store { dst: sto_i, offset: l_i.to_a(), value: load_i.to_a(), dtype });
        b.push(I::Barrier);
        let mut offset = 1;
        let mut v_i = load_i;
        // This is only ok if dim2 is a power of 2.
        while offset < dim2 {
            let off_i = b.add(l_i, offset as i32);
            let m_i = b.push(I::Load { src: sto_i, offset: off_i.to_a(), dtype });
            v_i = b.binary(BinaryOp::Max, m_i, v_i, dtype);
            b.push(I::Store { dst: sto_i, offset: l_i.to_a(), value: v_i.to_a(), dtype });
            b.push(I::Barrier);
            offset *= 2
        }
        let max_i = b.push(I::Load { src: sto_i, offset: 0.into(), dtype });

        let value_i = b.binary(BinaryOp::Sub, load_i, max_i, dtype);
        let value_i = b.unary(ssa::UnaryOp::Exp, value_i, dtype);

        // Compute the sum of the exps
        b.push(I::Store { dst: sto_i, offset: l_i.to_a(), value: value_i.to_a(), dtype });
        b.push(I::Barrier);
        let mut offset = 1;
        let mut v_i = value_i;
        while offset < dim2 {
            let off_i = b.add(l_i, offset as i32);
            let m_i = b.push(I::Load { src: sto_i, offset: off_i.to_a(), dtype });
            v_i = b.binary(BinaryOp::Add, m_i, v_i, dtype);
            b.push(I::Store { dst: sto_i, offset: l_i.to_a(), value: v_i.to_a(), dtype });
            b.push(I::Barrier);
            offset *= 2
        }
        let sum_i = b.push(I::Load { src: sto_i, offset: 0.into(), dtype });

        // Normalize by sum_exp
        let value_i = b.binary(BinaryOp::Div, value_i, sum_i, dtype);
        b.push(I::Store {
            dst: dst_i.to_varid(),
            offset: global_off_i.to_a(),
            value: value_i.to_a(),
            dtype,
        });
        Kernel::from_instrs(b.relocate()?)
    }

    pub fn softmax_reduce(_dim1: usize, dim2: usize) -> Result<Kernel> {
        let mut b = crate::block::Block::empty();
        let dtype = DType::F32;
        let src_i = b.push(arg(0, dtype));
        let dst_i = b.push(arg(1, dtype));
        let g_i = b.push(I::Special(ssa::Special::BlockIdx));
        let l_i = b.push(I::Special(ssa::Special::ThreadIdx));
        let base_off_i = b.mul(g_i, dim2 as i32);
        let global_off_i = b.binary(BinaryOp::Add, base_off_i, l_i, DType::I32);
        let load_i = b.push(I::Load { src: src_i.to_varid(), offset: global_off_i.to_a(), dtype });
        let max_i = b.push(I::ReduceLocal { op: ssa::ReduceOp::Max, arg: load_i.to_a(), dtype });
        let value_i = b.binary(BinaryOp::Sub, load_i, max_i, dtype);
        let value_i = b.unary(ssa::UnaryOp::Exp, value_i, dtype);
        let sum_i = b.push(I::ReduceLocal { op: ssa::ReduceOp::Sum, arg: value_i.to_a(), dtype });
        // Normalize by sum_exp
        let value_i = b.binary(BinaryOp::Div, value_i, sum_i, dtype);
        b.push(I::Store {
            dst: dst_i.to_varid(),
            offset: global_off_i.to_a(),
            value: value_i.to_a(),
            dtype,
        });
        Kernel::from_instrs(b.relocate()?)
    }

    pub fn softmax_block(_dim1: usize, dim2: usize, block_size: usize) -> Result<Kernel> {
        if dim2 % block_size != 0 {
            crate::bail!("last-dim {dim2} must be divisible by block size {block_size}")
        }
        let per_block = dim2 / block_size;
        let mut b = crate::block::Block::empty();
        let dtype = DType::F32;
        let src_i = b.push(arg(0, dtype));
        let dst_i = b.push(arg(1, dtype));
        let g_i = b.push(I::Special(ssa::Special::BlockIdx));
        let l_i = b.push(I::Special(ssa::Special::ThreadIdx));
        let base_off_i = b.mul(g_i, dim2 as i32);
        let global_off_i = b.binary(BinaryOp::Add, base_off_i, l_i, DType::I32);

        let mut load_is = Vec::with_capacity(per_block);

        let mut max_i = b.push(I::Const(f32::NEG_INFINITY.try_into()?));
        for i in (0..dim2).step_by(block_size) {
            let offset = b.add(global_off_i, i as i32).to_a();
            let load_i = b.push(I::Load { src: src_i.to_varid(), offset, dtype });
            max_i = b.binary(BinaryOp::Max, max_i, load_i, dtype);
            load_is.push(load_i)
        }
        let max_i = b.push(I::ReduceLocal { op: ssa::ReduceOp::Max, arg: max_i.to_a(), dtype });

        let mut value_is = Vec::with_capacity(per_block);
        let mut sum_i = b.push(I::Const(Const::I32(0)));
        for load_i in load_is.into_iter() {
            let value_i = b.binary(BinaryOp::Sub, load_i, max_i, dtype);
            let value_i = b.unary(ssa::UnaryOp::Exp, value_i, dtype);
            sum_i = b.binary(BinaryOp::Add, value_i, sum_i, dtype);
            value_is.push(value_i);
        }
        let sum_i = b.push(I::ReduceLocal { op: ssa::ReduceOp::Sum, arg: sum_i.to_a(), dtype });

        for (i, value_i) in value_is.into_iter().enumerate() {
            let i = i * block_size;
            let offset = b.add(global_off_i, i as i32).to_a();
            // Normalize by sum_exp
            let value_i = b.binary(BinaryOp::Div, value_i, sum_i, dtype);
            b.push(I::Store { dst: dst_i.to_varid(), offset, value: value_i.to_a(), dtype });
        }
        Kernel::from_instrs(b.relocate()?)
    }

    pub fn softmax(dim1: usize, dim2: usize) -> Result<Kernel> {
        softmax_barrier(dim1, dim2)
    }
}

pub mod op {
    use crate::lang::op::{self, Arg, DType, Kernel, Layout};
    use crate::Result;

    pub fn softmax(dim1: usize, dim2: usize) -> Result<Kernel> {
        let layout = Layout::from_shape(&[dim1, dim2]);
        let src_ptr = Arg::ptr(DType::F32);
        let dst_ptr = Arg::ptr(DType::F32);
        let src = op::load(src_ptr.id(), layout.clone(), DType::F32)?;
        let src_max = op::reduce(op::ReduceOp::Max, src.clone(), 1)?;
        let src_max = op::broadcast(src_max, (dim1, dim2))?;
        let diff = op::binary(op::BinaryOp::Sub, src, src_max)?;
        let exp = op::unary(op::UnaryOp::Exp, diff)?;
        let sum_exp = op::reduce(op::ReduceOp::Sum, exp.clone(), 1)?;
        let sum_exp = op::broadcast(sum_exp, (dim1, dim2))?;
        let sm = op::binary(op::BinaryOp::Div, exp, sum_exp)?;
        let st = op::store(dst_ptr.id(), layout, sm)?;
        let kernel =
            Kernel::new(format!("softmax_{dim1}_{dim2}"), vec![src_ptr, dst_ptr], vec![st]);
        Ok(kernel)
    }
}

use crate::lang::{Arg, DType, ExprNode as E, IndexExprNode as I, Kernel, Ops};

pub fn simple_add(block_size: usize) -> crate::Result<Kernel> {
    let lhs_ptr = Arg::ptr(DType::F32);
    let rhs_ptr = Arg::ptr(DType::F32);
    let dst_ptr = Arg::ptr(DType::F32);
    let offset = I::mul(&I::program_id(), &I::cst(block_size));
    let stride = I::cst(1);
    let len = I::cst(block_size);
    let lhs = E::load(&lhs_ptr, &offset, &len, &stride)?;
    let rhs = E::load(&rhs_ptr, &offset, &len, &stride)?;
    let op = Ops::store(&dst_ptr, &offset, &len, &stride, &lhs.add(&rhs));
    let k =
        Kernel::new(format!("simple_add_{block_size}"), vec![lhs_ptr, rhs_ptr, dst_ptr], vec![op]);
    Ok(k)
}