Skip to content

Commit

Permalink
Support per-axis quantization (#2390)
Browse files Browse the repository at this point in the history 
Reworked the simplify_qdq pass to support:

Per-axis quantization (ie. allow 1D scales and zero points)
Allow broadcast and transpose ops between dq and quant_op
  • Loading branch information
@shivadbhavsar
shivadbhavsar authored Nov 15, 2023
1 parent b2a40ea commit 0039b11
Show file tree
Hide file tree
Showing 4 changed files with 555 additions and 175 deletions.
19 changes: 19 additions & 0 deletions src/include/migraphx/matcher.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -591,6 +591,19 @@ MIGRAPHX_PRED_MATCHER(same_input_shapes, instruction_ref ins)
ins->inputs().begin(), ins->inputs().end(), [&](auto x) { return x->get_shape() == s; });
}

MIGRAPHX_PRED_MATCHER(has_same_value, instruction_ref ins)
{
if(ins->name() != "@literal")
return false;
bool all_same = false;
ins->get_literal().visit([&](auto s) {
all_same = std::all_of(s.begin() + 1, s.end(), [&](const auto& scale) {
return float_equal(scale, s.front());
});
});
return all_same;
}

MIGRAPHX_BASIC_MATCHER(output, const matcher_context&, instruction_ref ins)
{
if(ins->outputs().size() == 1)
Expand Down Expand Up @@ -844,6 +857,12 @@ auto skip_broadcasts_converts(Ms... ms)
return skip(name("broadcast", "multibroadcast", "contiguous", "convert"))(ms...);
}

template <class... Ms>
auto skip_broadcasts_transposes_contiguous(Ms... ms)
{
return skip(name("broadcast", "multibroadcast", "contiguous", "transpose"))(ms...);
}

template <class T>
inline auto has_value(T x, float tolerance = 1e-6)
{
Expand Down
138 changes: 103 additions & 35 deletions src/simplify_qdq.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -45,77 +45,145 @@ std::unordered_set<std::string> get_quantizable_op_names()
return s;
}

MIGRAPHX_PRED_MATCHER(has_same_value, instruction_ref ins)
struct match_find_quantizable_ops
{
if(ins->name() != "@literal")
return false;
bool all_same = false;
ins->get_literal().visit([&](auto s) {
all_same = std::all_of(s.begin() + 1, s.end(), [&](const auto& scale) {
return float_equal(scale, s.front());
static bool
is_valid_scale(instruction_ref scale, std::vector<std::size_t> lens, std::size_t axis)
{
return scale->get_shape().scalar() or scale->get_shape().elements() == lens.at(axis);
}

static bool is_valid_zero_point(instruction_ref zp)
{
if(not zp->can_eval())
return false;

bool all_zeros = false;
zp->eval().visit([&](auto z) {
all_zeros =
std::all_of(z.begin(), z.end(), [&](auto val) { return float_equal(val, 0); });
});
});
return all_same;
}
return all_zeros;
}

struct match_find_quantizable_ops
{
static auto
scale_broadcast_op(instruction_ref scale, std::vector<std::size_t> lens, std::size_t axis)
{
if(scale->get_shape().scalar())
{
return migraphx::make_op("multibroadcast", {{"out_lens", lens}});
}
else
{
return migraphx::make_op("broadcast", {{"out_lens", lens}, {"axis", axis}});
}
}

static auto dequantizelinear_op(const std::string& name, const std::string& scale)
// Helper function to insert quantized versions of any broadcasts and transpose ops that
// occur between dequantizelinear and the quantized op
static auto
propagate_quantized_ins(module& m, const instruction_ref dqins, const instruction_ref qop)
{
auto qinp = dqins->inputs().front();
auto next_ins = dqins;

while(next_ins != qop)
{
if(next_ins->name() != "dequantizelinear")
{
qinp = m.insert_instruction(qop, next_ins->get_operator(), qinp);
}
next_ins = next_ins->outputs().front();
}
return qinp;
}

static auto dequantizelinear_op(const std::string& scale, const std::string& zp)
{
return match::name("dequantizelinear")(
match::arg(0)(match::skip(match::name("quantizelinear"))(match::any().bind(name))),
match::arg(1)(match::skip_broadcasts(has_same_value().bind(scale))),
match::arg(2)(match::skip_broadcasts(match::all_of(match::has_value(0)))));
match::arg(0)(match::skip(match::name("quantizelinear"))(match::any())),
match::arg(1)(match::skip_broadcasts(match::is_constant().bind(scale))),
match::arg(2)(match::skip_broadcasts(match::is_constant().bind(zp))));
}

auto matcher() const
{
return match::name(get_quantizable_op_names())(
match::arg(0)(dequantizelinear_op("x1", "scale1")),
match::arg(1)(dequantizelinear_op("x2", "scale2")));
match::arg(0)(match::skip_broadcasts_transposes_contiguous(
dequantizelinear_op("scale1", "zp1").bind("dq1"))),
match::arg(1)(match::skip_broadcasts_transposes_contiguous(
dequantizelinear_op("scale2", "zp2").bind("dq2"))));
}

void apply(module& m, const match::matcher_result& r) const
{
auto qop = r.result;
auto q1 = r.instructions["x1"];
auto q2 = r.instructions["x2"];
auto dq1 = r.instructions["dq1"];
auto dq2 = r.instructions["dq2"];
auto scale1 = r.instructions["scale1"];
auto scale2 = r.instructions["scale2"];
auto zp1 = r.instructions["zp1"];
auto zp2 = r.instructions["zp2"];

// Only INT8 type currently supported
if(q1->get_shape().type() != migraphx::shape::int8_type or
q2->get_shape().type() != migraphx::shape::int8_type)
if(dq1->inputs().front()->get_shape().type() != migraphx::shape::int8_type or
dq2->inputs().front()->get_shape().type() != migraphx::shape::int8_type)
return;

double scale;
visit_all(scale1->get_literal(), scale2->get_literal())(
[&](const auto s1, const auto s2) { scale = s1.front() * s2.front(); });
// Only symmetric quantization supported (ie. non-zero zero_points not allowed)
if(not(is_valid_zero_point(zp1) and is_valid_zero_point(zp2)))
return;

// Only support scalar and 1D scales
if(scale1->get_shape().lens().size() != 1 or scale2->get_shape().lens().size() != 1)
return;

// Propagate q1 and q2 through any broadcasts and transposes before qop
auto qop_args = qop->inputs();
qop_args.at(0) = q1;
qop_args.at(1) = q2;
qop_args.at(0) = propagate_quantized_ins(m, dq1, qop);
qop_args.at(1) = propagate_quantized_ins(m, dq2, qop);
instruction_ref dq;
instruction_ref dq_scale;
instruction_ref out_scale;
instruction_ref zero_point;
if(qop->name() == "convolution")
{
auto conv_val = qop->get_operator().to_value();
dq = m.insert_instruction(
qop, migraphx::make_op("quant_convolution", conv_val), qop_args);
auto out_lens = dq->get_shape().lens();

// Input scale should always be scalar and weight scale can be scalar or 1D of the
// same lens as the output channel dim (dim 1 in the output)
if(not(is_valid_scale(scale1, out_lens, 1) and is_valid_scale(scale2, out_lens, 1)))
return;

auto s1_bcast =
m.insert_instruction(qop, scale_broadcast_op(scale1, out_lens, 1), scale1);
auto s2_bcast =
m.insert_instruction(qop, scale_broadcast_op(scale2, out_lens, 1), scale2);

out_scale = m.insert_instruction(qop, migraphx::make_op("mul"), s1_bcast, s2_bcast);
}
else if(qop->name() == "dot")
{
dq = m.insert_instruction(qop, migraphx::make_op("quant_dot"), qop_args);
dq = m.insert_instruction(qop, migraphx::make_op("quant_dot"), qop_args);
auto out_lens = dq->get_shape().lens();

// For (..., M, N) x (..., N, K) dot, only support cases where quantization axis is M
// for input1 and K for input 2
if(not(is_valid_scale(scale1, out_lens, out_lens.size() - 2) and
is_valid_scale(scale2, out_lens, out_lens.size() - 1)))
return;

auto s1_bcast = m.insert_instruction(
qop, scale_broadcast_op(scale1, out_lens, out_lens.size() - 2), scale1);
auto s2_bcast = m.insert_instruction(
qop, scale_broadcast_op(scale2, out_lens, out_lens.size() - 1), scale2);

out_scale = m.insert_instruction(qop, migraphx::make_op("mul"), s1_bcast, s2_bcast);
}
auto ins_type = qop->get_shape().type();
dq_scale = m.add_literal(literal({ins_type}, {scale}));

auto lens = dq->get_shape().lens();
auto scale_mb =
m.insert_instruction(qop, make_op("multibroadcast", {{"out_lens", lens}}), dq_scale);
dq = m.insert_instruction(qop, make_op("dequantizelinear"), dq, scale_mb);
dq = m.insert_instruction(qop, make_op("dequantizelinear"), dq, out_scale);
m.replace_instruction(qop, dq);
}
};
Expand Down
Loading

0 comments on commit 0039b11

Please sign in to comment.