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
|
// Copyright Vespa.ai. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
#include "dense_matmul_function.h"
#include <vespa/vespalib/objects/objectvisitor.h>
#include <vespa/eval/eval/value.h>
#include <vespa/eval/eval/operation.h>
#include <cassert>
#include <cblas.h>
namespace vespalib::eval {
using namespace tensor_function;
using namespace operation;
namespace {
template <typename LCT, typename RCT, typename OCT, bool lhs_common_inner, bool rhs_common_inner>
OCT my_dot_product(const LCT *lhs, const RCT *rhs, size_t lhs_size, size_t common_size, size_t rhs_size) {
OCT result = 0.0;
for (size_t i = 0; i < common_size; ++i) {
result += ((*lhs) * (*rhs));
lhs += (lhs_common_inner ? 1 : lhs_size);
rhs += (rhs_common_inner ? 1 : rhs_size);
}
return result;
}
template <typename LCT, typename RCT, typename OCT, bool lhs_common_inner, bool rhs_common_inner>
void my_matmul_op(InterpretedFunction::State &state, uint64_t param) {
const DenseMatMulFunction::Self &self = unwrap_param<DenseMatMulFunction::Self>(param);
auto lhs_cells = state.peek(1).cells().typify<LCT>();
auto rhs_cells = state.peek(0).cells().typify<RCT>();
auto dst_cells = state.stash.create_uninitialized_array<OCT>(self.lhs_size * self.rhs_size);
OCT *dst = dst_cells.begin();
const LCT *lhs = lhs_cells.cbegin();
for (size_t i = 0; i < self.lhs_size; ++i) {
const RCT *rhs = rhs_cells.cbegin();
for (size_t j = 0; j < self.rhs_size; ++j) {
*dst++ = my_dot_product<LCT,RCT,OCT,lhs_common_inner,rhs_common_inner>(lhs, rhs,
self.lhs_size, self.common_size, self.rhs_size);
rhs += (rhs_common_inner ? self.common_size : 1);
}
lhs += (lhs_common_inner ? self.common_size : 1);
}
state.pop_pop_push(state.stash.create<DenseValueView>(self.result_type, TypedCells(dst_cells)));
}
template <bool lhs_common_inner, bool rhs_common_inner>
void my_cblas_double_matmul_op(InterpretedFunction::State &state, uint64_t param) {
const DenseMatMulFunction::Self &self = unwrap_param<DenseMatMulFunction::Self>(param);
auto lhs_cells = state.peek(1).cells().typify<double>();
auto rhs_cells = state.peek(0).cells().typify<double>();
auto dst_cells = state.stash.create_array<double>(self.lhs_size * self.rhs_size);
cblas_dgemm(CblasRowMajor, lhs_common_inner ? CblasNoTrans : CblasTrans, rhs_common_inner ? CblasTrans : CblasNoTrans,
self.lhs_size, self.rhs_size, self.common_size, 1.0,
lhs_cells.cbegin(), lhs_common_inner ? self.common_size : self.lhs_size,
rhs_cells.cbegin(), rhs_common_inner ? self.common_size : self.rhs_size,
0.0, dst_cells.begin(), self.rhs_size);
state.pop_pop_push(state.stash.create<DenseValueView>(self.result_type, TypedCells(dst_cells)));
}
template <bool lhs_common_inner, bool rhs_common_inner>
void my_cblas_float_matmul_op(InterpretedFunction::State &state, uint64_t param) {
const DenseMatMulFunction::Self &self = unwrap_param<DenseMatMulFunction::Self>(param);
auto lhs_cells = state.peek(1).cells().typify<float>();
auto rhs_cells = state.peek(0).cells().typify<float>();
auto dst_cells = state.stash.create_array<float>(self.lhs_size * self.rhs_size);
cblas_sgemm(CblasRowMajor, lhs_common_inner ? CblasNoTrans : CblasTrans, rhs_common_inner ? CblasTrans : CblasNoTrans,
self.lhs_size, self.rhs_size, self.common_size, 1.0,
lhs_cells.cbegin(), lhs_common_inner ? self.common_size : self.lhs_size,
rhs_cells.cbegin(), rhs_common_inner ? self.common_size : self.rhs_size,
0.0, dst_cells.begin(), self.rhs_size);
state.pop_pop_push(state.stash.create<DenseValueView>(self.result_type, TypedCells(dst_cells)));
}
bool is_matrix(const ValueType &type) {
return (type.is_dense() && (type.dimensions().size() == 2));
}
bool is_matmul(const ValueType &a, const ValueType &b,
const vespalib::string &reduce_dim, const ValueType &result_type)
{
size_t npos = ValueType::Dimension::npos;
return (is_matrix(a) && is_matrix(b) && is_matrix(result_type) &&
(a.dimension_index(reduce_dim) != npos) &&
(b.dimension_index(reduce_dim) != npos));
}
const ValueType::Dimension &dim(const TensorFunction &expr, size_t idx) {
return expr.result_type().dimensions()[idx];
}
size_t inv(size_t idx) { return (1 - idx); }
const TensorFunction &create_matmul(const TensorFunction &a, const TensorFunction &b,
const vespalib::string &reduce_dim, const ValueType &result_type, Stash &stash) {
size_t a_idx = a.result_type().dimension_index(reduce_dim);
size_t b_idx = b.result_type().dimension_index(reduce_dim);
assert(a_idx != ValueType::Dimension::npos);
assert(b_idx != ValueType::Dimension::npos);
assert(dim(a, a_idx).size == dim(b, b_idx).size);
bool a_common_inner = (a_idx == 1);
bool b_common_inner = (b_idx == 1);
size_t a_size = dim(a, inv(a_idx)).size;
size_t b_size = dim(b, inv(b_idx)).size;
size_t common_size = dim(a, a_idx).size;
bool a_is_lhs = (dim(a, inv(a_idx)).name < dim(b, inv(b_idx)).name);
if (a_is_lhs) {
return stash.create<DenseMatMulFunction>(result_type, a, b, a_size, common_size, b_size, a_common_inner, b_common_inner);
} else {
return stash.create<DenseMatMulFunction>(result_type, b, a, b_size, common_size, a_size, b_common_inner, a_common_inner);
}
}
struct SelectDenseMatmul {
template<typename LCM, typename RCM, typename LhsCommonInner, typename RhsCommonInner> static auto invoke() {
constexpr CellMeta ocm = CellMeta::join(LCM::value, RCM::value).reduce(false);
using LCT = CellValueType<LCM::value.cell_type>;
using RCT = CellValueType<RCM::value.cell_type>;
using OCT = CellValueType<ocm.cell_type>;
if (std::is_same_v<LCT,double> && std::is_same_v<RCT,double>) {
return my_cblas_double_matmul_op<LhsCommonInner::value, RhsCommonInner::value>;
} else if (std::is_same_v<LCT,float> && std::is_same_v<RCT,float>) {
return my_cblas_float_matmul_op<LhsCommonInner::value, RhsCommonInner::value>;
} else {
return my_matmul_op<LCT, RCT, OCT, LhsCommonInner::value, RhsCommonInner::value>;
}
}
};
} // namespace <unnamed>
DenseMatMulFunction::Self::Self(const ValueType &result_type_in,
size_t lhs_size_in,
size_t common_size_in,
size_t rhs_size_in)
: result_type(result_type_in),
lhs_size(lhs_size_in),
common_size(common_size_in),
rhs_size(rhs_size_in)
{
}
DenseMatMulFunction::Self::~Self() = default;
DenseMatMulFunction::DenseMatMulFunction(const ValueType &result_type,
const TensorFunction &lhs_in,
const TensorFunction &rhs_in,
size_t lhs_size,
size_t common_size,
size_t rhs_size,
bool lhs_common_inner,
bool rhs_common_inner)
: Super(result_type, lhs_in, rhs_in),
_lhs_size(lhs_size),
_common_size(common_size),
_rhs_size(rhs_size),
_lhs_common_inner(lhs_common_inner),
_rhs_common_inner(rhs_common_inner)
{
}
DenseMatMulFunction::~DenseMatMulFunction() = default;
InterpretedFunction::Instruction
DenseMatMulFunction::compile_self(const ValueBuilderFactory &, Stash &stash) const
{
using MyTypify = TypifyValue<TypifyCellMeta,TypifyBool>;
Self &self = stash.create<Self>(result_type(), _lhs_size, _common_size, _rhs_size);
auto op = typify_invoke<4,MyTypify,SelectDenseMatmul>(
lhs().result_type().cell_meta().not_scalar(),
rhs().result_type().cell_meta().not_scalar(),
_lhs_common_inner, _rhs_common_inner);
return InterpretedFunction::Instruction(op, wrap_param<DenseMatMulFunction::Self>(self));
}
void
DenseMatMulFunction::visit_self(vespalib::ObjectVisitor &visitor) const
{
Super::visit_self(visitor);
visitor.visitInt("lhs_size", _lhs_size);
visitor.visitInt("common_size", _common_size);
visitor.visitInt("rhs_size", _rhs_size);
visitor.visitBool("lhs_common_inner", _lhs_common_inner);
visitor.visitBool("rhs_common_inner", _rhs_common_inner);
}
const TensorFunction &
DenseMatMulFunction::optimize(const TensorFunction &expr, Stash &stash)
{
auto reduce = as<Reduce>(expr);
if (reduce && (reduce->aggr() == Aggr::SUM) && (reduce->dimensions().size() == 1)) {
auto join = as<Join>(reduce->child());
if (join && (join->function() == Mul::f)) {
const TensorFunction &a = join->lhs();
const TensorFunction &b = join->rhs();
if (is_matmul(a.result_type(), b.result_type(), reduce->dimensions()[0], expr.result_type())) {
return create_matmul(a, b, reduce->dimensions()[0], expr.result_type(), stash);
}
}
}
return expr;
}
} // namespace
|
