Skip to content

GitLab

"src/propagate_constant.cpp" did not exist on "ae47918310cbdcafacead1e3126f370ced0032c1"
Commit 3855c6af authored by Shucai Xiao's avatar Shucai Xiao
Browse files

Merge branch 'opt_log_softmax_new_device_code' into argmax_min

parents b222af2f 93eae2df
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 1020 additions and 654 deletions
src/include/migraphx/array.hpp 0 → 100644
View file @ 3855c6af
#ifndef MIGRAPHX_GUARD_RTGLIB_ARRAY_HPP
#define MIGRAPHX_GUARD_RTGLIB_ARRAY_HPP
#include <migraphx/config.hpp>
#include <migraphx/functional.hpp>
#include <migraphx/requires.hpp>
#include <type_traits>
#include <array>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace detail {
template <class R, class...>
struct array_type
{
using type = R;
};
template <class... Ts>
struct array_type<void, Ts...> : std::common_type<Ts...>
{
};
template <class R, class... Ts>
using array_type_t = typename array_type<R, Ts...>::type;
template <class T, std::size_t N, std::size_t... I>
constexpr std::array<std::remove_cv_t<T>, N> to_array_impl(T (&a)[N], seq<I...>)
{
return {{a[I]...}};
}
} // namespace detail
template <class Result = void, class... Ts, MIGRAPHX_REQUIRES((sizeof...(Ts) > 0))>
constexpr std::array<detail::array_type_t<Result, Ts...>, sizeof...(Ts)> make_array(Ts&&... xs)
{
return {static_cast<detail::array_type_t<Result, Ts...>>(std::forward<Ts>(xs))...};
}
constexpr std::array<int, 0> make_array() { return {}; }
template <class T, std::size_t N>
constexpr auto to_array(T (&a)[N])
{
return detail::to_array_impl(a, detail::gens<N>{});
}
namespace detail {
template <std::size_t Offset = 0, class Array, std::size_t... I>
constexpr auto rearray_impl(Array a, seq<I...>)
{
return make_array(a[I + Offset]...);
}
} // namespace detail
template <class T, std::size_t N>
constexpr auto pop_front(std::array<T, N> a)
{
return detail::rearray_impl(a, detail::gens<N - 1>{});
}
template <class T, std::size_t N>
constexpr auto pop_back(std::array<T, N> a)
{
return detail::rearray_impl<1>(a, detail::gens<N - 1>{});
}
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif
src/include/migraphx/functional.hpp
View file @ 3855c6af
......@@ -15,6 +15,12 @@ struct swallow
}
};
template <class T>
auto tuple_size(const T&)
{
return typename std::tuple_size<T>::type{};
}
namespace detail {
template <class R, class F>
......@@ -83,6 +89,12 @@ constexpr auto sequence_c(F&& f)
return detail::sequence_c_impl(f, detail::gens<N>{});
}
template <class IntegerConstant, class F>
constexpr auto sequence(IntegerConstant ic, F&& f)
{
return sequence_c<ic>(f);
}
template <class F, class... Ts>
constexpr void each_args(F f, Ts&&... xs)
{
......@@ -95,9 +107,9 @@ constexpr void each_args(F)
}
template <class F, class T>
auto unpack(F f, T& x)
auto unpack(F f, T&& x)
{
return sequence_c<std::tuple_size<T>{}>([&](auto... is) { f(std::get<is>(x)...); });
return sequence(tuple_size(x), [&](auto... is) { f(std::get<is>(static_cast<T&&>(x))...); });
}
/// Implements a fix-point combinator
......@@ -149,6 +161,35 @@ auto index_of(T& x)
return [&](auto&& y) { return x[y]; };
}
template <class T, class... Ts>
decltype(auto) front_args(T&& x, Ts&&...)
{
return static_cast<T&&>(x);
}
template <class... Ts>
decltype(auto) back_args(Ts&&... xs)
{
return std::get<sizeof...(Ts) - 1>(std::tuple<Ts&&...>(static_cast<Ts&&>(xs)...));
}
template <class T, class... Ts>
auto pop_front_args(T&&, Ts&&... xs)
{
return [&](auto f) { f(static_cast<Ts&&>(xs)...); };
}
template <class... Ts>
auto pop_back_args(Ts&&... xs)
{
return [&](auto f) {
using tuple_type = std::tuple<Ts&&...>;
auto t = tuple_type(static_cast<Ts&&>(xs)...);
return sequence_c<sizeof...(Ts) - 1>(
[&](auto... is) { return f(std::get<is>(static_cast<tuple_type&&>(t))...); });
};
}
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
......
src/include/migraphx/ranges.hpp
View file @ 3855c6af
......@@ -33,6 +33,10 @@ auto generic_find_impl(rank<0>, C&& c, const T& x)
return std::find(c.begin(), c.end(), x);
}
struct empty
{
};
} // namespace detail
template <class C, class T>
......@@ -71,6 +75,12 @@ bool all_of(const std::initializer_list<T>& c, const Predicate& p)
return std::all_of(c.begin(), c.end(), p);
}
template <class Predicate>
bool all_of(detail::empty, const Predicate&)
{
return true;
}
template <class C, class Predicate>
bool any_of(const C& c, const Predicate& p)
{
......@@ -83,6 +93,12 @@ bool any_of(const std::initializer_list<T>& c, const Predicate& p)
return std::any_of(c.begin(), c.end(), p);
}
template <class Predicate>
bool any_of(detail::empty, const Predicate&)
{
return false;
}
template <class C, class Predicate>
bool none_of(const C& c, const Predicate& p)
{
......@@ -95,6 +111,12 @@ bool none_of(const std::initializer_list<T>& c, const Predicate& p)
return std::none_of(c.begin(), c.end(), p);
}
template <class Predicate>
bool none_of(detail::empty, const Predicate&)
{
return true;
}
template <class Range, class Iterator>
void copy(Range&& r, Iterator it)
{
......
src/include/migraphx/raw_data.hpp
View file @ 3855c6af
......@@ -212,6 +212,25 @@ auto visit_all(T&& x, Ts&&... xs)
};
}
template <class T>
auto visit_all(const std::vector<T>& x)
{
auto&& s = x.front().get_shape();
if(!std::all_of(
x.begin(), x.end(), [&](const T& y) { return y.get_shape().type() == s.type(); }))
MIGRAPHX_THROW("Types must be the same");
return [&](auto v) {
s.visit_type([&](auto as) {
using type = typename decltype(as)::type;
std::vector<tensor_view<type>> result;
std::transform(x.begin(), x.end(), std::back_inserter(result), [&](const auto& y) {
return make_view(y.get_shape(), as.from(y.data()));
});
v(result);
});
};
}
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
......
src/targets/cpu/lowering.cpp
View file @ 3855c6af
......@@ -530,17 +530,26 @@ struct cpu_softmax
std::string name() const { return "cpu::softmax"; }
shape compute_shape(const std::vector<shape>& inputs) const { return op.compute_shape(inputs); }
template <typename T>
std::size_t compute_batch_index(T idx, shape& batch_shape, int axis) const
std::vector<size_t> compute_batch_indices(size_t idx, const shape& s) const
{
idx[axis] = 0;
return batch_shape.index(idx);
std::vector<std::size_t> indices(s.lens().size());
std::transform(s.strides().begin(),
s.strides().end(),
s.lens().begin(),
indices.begin(),
[&](std::size_t stride, std::size_t len) {
assert(len > 0 and stride > 0);
return (idx / stride) % len;
});
return indices;
}
argument compute(context&, const shape& output_shape, std::vector<argument> args) const
{
argument result{output_shape};
auto batch_lens = output_shape.lens();
size_t n_dims = batch_lens[op.axis];
batch_lens[op.axis] = 1;
shape batch_shape{shape::int32_type, batch_lens};
......@@ -548,26 +557,34 @@ struct cpu_softmax
using value_type = typename decltype(input)::value_type;
std::vector<value_type> batch_max(batch_shape.elements(),
std::numeric_limits<value_type>::lowest());
shape_for_each(output_shape, [&](auto idx) {
auto index = this->compute_batch_index(idx, batch_shape, op.axis);
batch_max[index] = std::max(batch_max[index], input(idx.begin(), idx.end()));
});
std::vector<value_type> batch_sum(batch_shape.elements(), value_type(0));
par_for(batch_shape.elements(), [&](auto i) {
auto idx = this->compute_batch_indices(i, batch_shape);
shape_for_each(output_shape, [&](auto idx) {
auto index = this->compute_batch_index(idx, batch_shape, op.axis);
output(idx.begin(), idx.end()) =
std::exp(input(idx.begin(), idx.end()) - batch_max[index]);
});
for(size_t j = 0; j < n_dims; ++j)
{
idx[op.axis] = j;
batch_max[i] = std::max(batch_max[i], input(idx.begin(), idx.end()));
}
std::vector<value_type> batch_sum(batch_shape.elements(), value_type(0));
shape_for_each(output_shape, [&](auto idx) {
auto index = this->compute_batch_index(idx, batch_shape, op.axis);
batch_sum[index] += output(idx.begin(), idx.end());
});
for(size_t j = 0; j < n_dims; ++j)
{
idx[op.axis] = j;
size_t index = output_shape.index(idx);
output[index] = std::exp(input[index] - batch_max[i]);
}
for(size_t j = 0; j < n_dims; ++j)
{
idx[op.axis] = j;
batch_sum[i] += output(idx.begin(), idx.end());
}
shape_for_each(output_shape, [&](auto idx) {
auto index = this->compute_batch_index(idx, batch_shape, op.axis);
output(idx.begin(), idx.end()) /= batch_sum[index];
for(size_t j = 0; j < n_dims; ++j)
{
idx[op.axis] = j;
output(idx.begin(), idx.end()) /= batch_sum[i];
}
});
});
......@@ -588,48 +605,65 @@ struct cpu_logsoftmax
std::string name() const { return "cpu::logsoftmax"; }
shape compute_shape(const std::vector<shape>& inputs) const { return op.compute_shape(inputs); }
template <typename T>
std::size_t compute_batch_index(T idx, const shape& batch_shape, int axis) const
std::vector<size_t> compute_batch_indices(size_t idx, const shape& s) const
{
idx[axis] = 0;
return batch_shape.index(idx);
std::vector<std::size_t> indices(s.lens().size());
std::transform(s.strides().begin(),
s.strides().end(),
s.lens().begin(),
indices.begin(),
[&](std::size_t stride, std::size_t len) {
assert(len > 0 and stride > 0);
return (idx / stride) % len;
});
return indices;
}
argument compute(context&, const shape& output_shape, std::vector<argument> args) const
{
argument result{output_shape};
auto batch_lens = output_shape.lens();
size_t n_dims = batch_lens[op.axis];
batch_lens[op.axis] = 1;
shape batch_shape{shape::int32_type, batch_lens};
// use a parallel implementation to acheive better performance
// one thread for one batch
visit_all(result, args[0])([&](auto output, auto input) {
using value_type = typename decltype(input)::value_type;
std::vector<value_type> batch_max(batch_shape.elements(),
std::numeric_limits<value_type>::lowest());
shape_for_each(output_shape, [&](auto idx) {
auto index = this->compute_batch_index(idx, batch_shape, op.axis);
batch_max[index] = std::max(batch_max[index], input(idx.begin(), idx.end()));
});
std::vector<value_type> batch_sum(batch_shape.elements(), value_type(0));
shape_for_each(output_shape, [&](auto idx) {
auto index = this->compute_batch_index(idx, batch_shape, op.axis);
output(idx.begin(), idx.end()) = input(idx.begin(), idx.end()) - batch_max[index];
});
par_for(batch_shape.elements(), [&](auto i) {
auto idx = this->compute_batch_indices(i, batch_shape);
for(size_t j = 0; j < n_dims; ++j)
{
idx[op.axis] = j;
batch_max[i] = std::max(batch_max[i], input(idx.begin(), idx.end()));
}
std::vector<value_type> batch_sum(batch_shape.elements(), value_type(0));
shape_for_each(output_shape, [&](auto idx) {
auto index = this->compute_batch_index(idx, batch_shape, op.axis);
batch_sum[index] += std::exp(output(idx.begin(), idx.end()));
});
for(size_t j = 0; j < n_dims; ++j)
{
idx[op.axis] = j;
size_t index = output_shape.index(idx);
output[index] = input[index] - batch_max[i];
}
for(size_t j = 0; j < n_dims; ++j)
{
idx[op.axis] = j;
batch_sum[i] += std::exp(output(idx.begin(), idx.end()));
}
for(std::size_t i = 0; i < batch_sum.size(); ++i)
{
batch_sum[i] = std::log(batch_sum[i]);
}
shape_for_each(output_shape, [&](auto idx) {
auto index = this->compute_batch_index(idx, batch_shape, op.axis);
output(idx.begin(), idx.end()) -= batch_sum[index];
for(size_t j = 0; j < n_dims; ++j)
{
idx[op.axis] = j;
output(idx.begin(), idx.end()) -= batch_sum[i];
}
});
});
......
src/targets/gpu/device/concat.cpp
View file @ 3855c6af
......@@ -10,22 +10,20 @@ namespace gpu {
namespace device {
argument concat(hipStream_t stream,
const migraphx::shape& output_shape,
const migraphx::shape&,
std::vector<migraphx::argument> args,
std::vector<std::size_t> offsets)
{
for(std::size_t l = 0; l < args.size() - 1; l++)
auto ninputs = args.size() - 1;
for(std::size_t j = 0; j < ninputs; j++)
{
auto argl = args[l];
std::size_t nelements = argl.get_shape().elements();
visit_all(args.back(), argl)([&](auto output, auto input) {
visit_tensor_size(output_shape.lens().size(), [&](auto ndim) {
auto* outptr = output.data() + offsets[l];
const auto* inptr = input.data();
hip_tensor_descriptor<ndim> desc_input(input.get_shape());
hip_tensor_descriptor<ndim> desc_output(output.get_shape());
gs_launch(stream, nelements)(
[=](auto i) { outptr[desc_output.linear(desc_input.multi(i))] = inptr[i]; });
auto&& arg = args[j];
std::size_t nelements = arg.get_shape().elements();
auto offset = offsets[j];
hip_visit_all(args.back(), arg)([&](auto output, auto input) {
gs_launch(stream, nelements)([=](auto i) {
auto idx = output.get_shape().index(input.get_shape().multi(i));
output.data()[idx + offset] = input.data()[i];
});
});
}
......
src/targets/gpu/device/gather.cpp
View file @ 3855c6af
......@@ -11,35 +11,30 @@ inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
namespace device {
argument gather(hipStream_t stream,
const migraphx::shape& output_shape,
std::vector<migraphx::argument> args,
int axis)
argument gather(hipStream_t stream, argument result, argument arg1, argument arg2, int axis)
{
auto axis_index = (axis < 0) ? (axis + args[0].get_shape().lens().size()) : axis;
visit_all(args.back(), args[0])([&](auto output, auto input) {
std::size_t nelements = output_shape.elements();
args[1].visit([&](auto indices) {
const auto* indices_ptr = device_cast(indices.data());
auto* out_ptr = device_cast(output.data());
const auto* in_ptr = device_cast(input.data());
auto& input_shape = args[0].get_shape();
auto lens = input_shape.lens();
lens[axis_index] = args[1].get_shape().elements();
migraphx::shape out_comp_shape{output_shape.type(), lens};
visit_tensor_size(out_comp_shape.lens().size(), [&](auto n_out_dim) {
hip_tensor_descriptor<n_out_dim> desc_input(input_shape);
hip_tensor_descriptor<n_out_dim> desc_output(out_comp_shape);
gs_launch(stream, nelements)([=](auto ii) {
auto in_idx = desc_output.multi(ii);
in_idx[axis_index] = indices_ptr[in_idx[axis_index]];
out_ptr[ii] = in_ptr[desc_input.linear(in_idx)];
auto axis_index = (axis < 0) ? (axis + arg1.get_shape().lens().size()) : axis;
auto& input_shape = arg1.get_shape();
auto lens = input_shape.lens();
lens[axis_index] = arg2.get_shape().elements();
shape out_comp_shape{result.get_shape().type(), lens};
std::size_t nelements = result.get_shape().elements();
visit_all(result, arg1)([&](auto output, auto input_v) {
hip_visit_views(input_v, out_comp_shape)([&](auto input, auto out_comp) {
arg2.visit([&](auto indices) {
const auto* indices_ptr = device_cast(indices.data());
auto* output_ptr = device_cast(output.data());
gs_launch(stream, nelements)([=](auto i) {
auto idx = out_comp.multi(i);
idx[axis_index] = indices_ptr[idx[axis_index]];
output_ptr[i] = input[idx];
});
});
});
});
return args.back();
return result;
}
} // namespace device
......
src/targets/gpu/device/include/migraphx/gpu/device/array.hpp 0 → 100644
View file @ 3855c6af
#ifndef MIGRAPHX_GUARD_RTGLIB_DEVICE_ARRAY_HPP
#define MIGRAPHX_GUARD_RTGLIB_DEVICE_ARRAY_HPP
#include <migraphx/gpu/device/types.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
namespace device {
template <class T, std::size_t N>
struct hip_array
{
T d[N];
MIGRAPHX_DEVICE_CONSTEXPR T& operator[](std::size_t i) { return d[i]; }
MIGRAPHX_DEVICE_CONSTEXPR const T& operator[](std::size_t i) const { return d[i]; }
MIGRAPHX_DEVICE_CONSTEXPR T* data() { return d; }
MIGRAPHX_DEVICE_CONSTEXPR const T* data() const { return d; }
MIGRAPHX_DEVICE_CONSTEXPR std::integral_constant<std::size_t, N> size() const { return {}; }
MIGRAPHX_DEVICE_CONSTEXPR T* begin() { return d; }
MIGRAPHX_DEVICE_CONSTEXPR const T* begin() const { return d; }
MIGRAPHX_DEVICE_CONSTEXPR T* end() { return d + size(); }
MIGRAPHX_DEVICE_CONSTEXPR const T* end() const { return d + size(); }
MIGRAPHX_DEVICE_CONSTEXPR T dot(const hip_array& x) const
{
T result = 0;
for(std::size_t i = 0; i < N; i++)
result += x[i] * d[i];
return result;
}
MIGRAPHX_DEVICE_CONSTEXPR T product() const
{
T result = 1;
for(std::size_t i = 0; i < N; i++)
result *= d[i];
return result;
}
friend MIGRAPHX_DEVICE_CONSTEXPR hip_array operator*(const hip_array& x, const hip_array& y)
{
hip_array result;
for(std::size_t i = 0; i < N; i++)
result[i] = x[i] * y[i];
return result;
}
};
} // namespace device
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif
src/targets/gpu/device/include/migraphx/gpu/device/shape.hpp 0 → 100644
View file @ 3855c6af
#ifndef MIGRAPHX_GUARD_RTGLIB_DEVICE_SHAPE_HPP
#define MIGRAPHX_GUARD_RTGLIB_DEVICE_SHAPE_HPP
#include <migraphx/gpu/device/array.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
namespace device {
template <std::size_t N>
struct hip_shape
{
using hip_index = hip_array<std::size_t, N>;
hip_array<std::size_t, N> lens = {};
hip_array<std::size_t, N> strides = {};
bool standard = false;
__device__ __host__ hip_shape() = default;
hip_shape(const shape& s) : standard(s.standard())
{
assert(s.lens().size() == N);
assert(s.strides().size() == N);
std::copy(s.lens().begin(), s.lens().end(), lens.begin());
std::copy(s.strides().begin(), s.strides().end(), strides.begin());
}
MIGRAPHX_DEVICE_CONSTEXPR std::size_t elements() const { return lens.product(); }
MIGRAPHX_DEVICE_CONSTEXPR std::size_t index(hip_index x) const { return x.dot(strides); }
MIGRAPHX_DEVICE_CONSTEXPR std::size_t index(std::initializer_list<std::size_t> x) const
{
std::size_t idx = 0;
for(std::size_t i = 0; i < x.size(); i++)
idx += *(x.begin() + i) * strides[i];
return idx;
}
MIGRAPHX_DEVICE_CONSTEXPR std::size_t index(std::size_t i) const
{
if(this->standard)
return i;
else
{
const std::size_t rank = this->lens.size();
std::size_t s = 1;
std::size_t result = 0;
for(std::size_t j = 0; j < this->lens.size(); j++)
{
const std::size_t k = rank - j - 1;
const std::size_t stride = this->strides[k];
const std::size_t len = this->lens[k];
const std::size_t slen = s * len;
const std::size_t idx = (i % slen) / s;
result += stride * idx;
s = slen;
}
return result;
}
}
MIGRAPHX_DEVICE_CONSTEXPR hip_index multi(std::size_t idx) const
{
hip_index result;
std::size_t tidx = idx;
for(std::size_t is = 0; is < result.size(); is++)
{
result[is] = tidx / strides[is];
tidx = tidx % strides[is];
}
return result;
}
};
template <std::size_t N>
hip_shape<N> make_hip_shape(const shape& x)
{
return x;
}
} // namespace device
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif
src/targets/gpu/device/include/migraphx/gpu/device/tensor.hpp
View file @ 3855c6af
#ifndef MIGRAPHX_GUARD_RTGLIB_DEAVICE_TENSOR_HPP
#define MIGRAPHX_GUARD_RTGLIB_DEAVICE_TENSOR_HPP
#include <hip/hip_runtime.h>
#include <migraphx/functional.hpp>
#include <migraphx/config.hpp>
#include <migraphx/gpu/device/visit.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
namespace device {
template <class F>
void visit_tensor_size(std::size_t n, F f)
{
switch(n)
{
case 1:
{
f(std::integral_constant<std::size_t, 1>{});
break;
}
case 2:
{
f(std::integral_constant<std::size_t, 2>{});
break;
}
case 3:
{
f(std::integral_constant<std::size_t, 3>{});
break;
}
case 4:
{
f(std::integral_constant<std::size_t, 4>{});
break;
}
case 5:
{
f(std::integral_constant<std::size_t, 5>{});
break;
}
default: throw std::runtime_error("Unknown tensor size");
}
}
template <size_t NDim>
struct hip_index
{
size_t d[NDim];
__device__ __host__ size_t& operator[](size_t i) { return d[i]; }
__device__ __host__ size_t operator[](size_t i) const { return d[i]; }
};
template <std::size_t NDim>
using hip_tensor_index = hip_array<std::size_t, NDim>;
template <size_t NDim>
template <std::size_t NDim>
struct hip_tensor_descriptor
{
__device__ __host__ hip_tensor_descriptor() = default;
......@@ -63,26 +22,26 @@ struct hip_tensor_descriptor
std::copy(s.strides().begin(), s.strides().end(), strides);
}
__device__ __host__ hip_index<NDim> multi(size_t idx) const
__device__ __host__ hip_tensor_index<NDim> multi(std::size_t idx) const
{
hip_index<NDim> result{};
size_t tidx = idx;
for(size_t is = 0; is < NDim; is++)
hip_tensor_index<NDim> result{};
std::size_t tidx = idx;
for(std::size_t is = 0; is < NDim; is++)
{
result[is] = tidx / strides[is];
tidx = tidx % strides[is];
}
return result;
}
__device__ __host__ size_t linear(hip_index<NDim> s) const
__device__ __host__ std::size_t linear(hip_tensor_index<NDim> s) const
{
size_t idx = 0;
for(size_t i = 0; i < NDim; i++)
std::size_t idx = 0;
for(std::size_t i = 0; i < NDim; i++)
idx += s[i] * strides[i];
return idx;
}
size_t lens[NDim] = {};
size_t strides[NDim] = {};
std::size_t lens[NDim] = {};
std::size_t strides[NDim] = {};
};
} // namespace device
......
src/targets/gpu/device/include/migraphx/gpu/device/tensor_view.hpp 0 → 100644
View file @ 3855c6af
#ifndef MIGRAPHX_GUARD_RTGLIB_DEVICE_TENSOR_VIEW_HPP
#define MIGRAPHX_GUARD_RTGLIB_DEVICE_TENSOR_VIEW_HPP
#include <migraphx/gpu/device/shape.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
namespace device {
template <class T, std::size_t N>
struct hip_tensor_view
{
using value_type = T;
using hip_index = typename hip_shape<N>::hip_index;
__device__ __host__ hip_tensor_view() = default;
__host__ hip_tensor_view(tensor_view<T> x) : d(x.data()), s(x.get_shape()) {}
__host__ hip_tensor_view(T* x, const shape& ss) : d(x), s(ss) {}
MIGRAPHX_DEVICE_CONSTEXPR const hip_shape<N>& get_shape() const { return s; }
MIGRAPHX_DEVICE_CONSTEXPR std::size_t size() const { return s.elements(); }
MIGRAPHX_DEVICE_CONSTEXPR value_type* data() const { return d; }
template <class U>
MIGRAPHX_DEVICE_CONSTEXPR value_type& operator[](U i) const
{
return d[s.index(i)];
}
MIGRAPHX_DEVICE_CONSTEXPR value_type* begin() const { return d; }
MIGRAPHX_DEVICE_CONSTEXPR value_type* end() const { return d + size(); }
private:
value_type* d = nullptr;
hip_shape<N> s{};
};
template <std::size_t N, class T>
hip_tensor_view<T, N> make_hip_view(const shape& s, T* x)
{
return {x, s};
}
template <std::size_t N, class T>
hip_tensor_view<T, N> make_hip_view(tensor_view<T> x)
{
return {x};
}
} // namespace device
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif
src/targets/gpu/device/include/migraphx/gpu/device/types.hpp
View file @ 3855c6af
......@@ -8,14 +8,45 @@
#ifndef MIGRAPHX_GUARD_RTGLIB_GPU_DEVICE_TYPES_HPP
#define MIGRAPHX_GUARD_RTGLIB_GPU_DEVICE_TYPES_HPP
#include <hip/hip_runtime.h>
#include <migraphx/half.hpp>
#include <migraphx/config.hpp>
#include <migraphx/tensor_view.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
namespace device {
#define MIGRAPHX_DEVICE_CONSTEXPR constexpr __device__ __host__ // NOLINT
template <class T, std::size_t N>
using vec = T __attribute__((ext_vector_type(N)));
template <std::size_t N, class T>
__device__ __host__ T* as_pointer(vec<T, N>* x)
{
return reinterpret_cast<T*>(x);
}
template <std::size_t N, class T>
__device__ __host__ vec<T, N>* as_vec(T* x)
{
return reinterpret_cast<vec<T, N>*>(x);
}
template <std::size_t N, class T>
tensor_view<vec<T, N>> as_vec(tensor_view<T> x)
{
return {x.get_shape(), as_vec<N>(x.data())};
}
template <std::size_t N, class... Ts>
auto pack_vec(Ts... xs)
{
return [=](auto f, std::size_t n) { return f(as_vec<N>(xs)[n]...); };
}
using gpu_half = __fp16;
namespace detail {
......@@ -25,6 +56,12 @@ struct device_type
using type = T;
};
template <class T, std::size_t N>
struct device_type<vec<T, N>>
{
using type = vec<typename device_type<T>::type, N>;
};
template <>
struct device_type<half>
{
......@@ -38,7 +75,7 @@ struct host_type
};
template <>
struct device_type<gpu_half>
struct host_type<gpu_half>
{
using type = half;
};
......@@ -75,6 +112,12 @@ device_type<T>* device_cast(T* x)
return reinterpret_cast<device_type<T>*>(x);
}
template <class T>
tensor_view<device_type<T>> device_cast(tensor_view<T> x)
{
return {x.get_shape(), reinterpret_cast<device_type<T>*>(x.data())};
}
template <class T>
__device__ __host__ T to_hip_type(T x)
{
......
src/targets/gpu/device/include/migraphx/gpu/device/vector.hpp 0 → 100644
View file @ 3855c6af
#ifndef MIGRAPHX_GUARD_RTGLIB_DEVICE_VECTOR_HPP
#define MIGRAPHX_GUARD_RTGLIB_DEVICE_VECTOR_HPP
#include <migraphx/gpu/device/types.hpp>
#include <vector>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
namespace device {
template <class T, std::size_t N>
struct hip_vector
{
MIGRAPHX_DEVICE_CONSTEXPR hip_vector() = default;
MIGRAPHX_DEVICE_CONSTEXPR hip_vector(std::size_t s) : len(s) {}
template <class Iterator>
__device__ __host__ hip_vector(Iterator start, Iterator last)
{
auto it = std::copy(start, last, d);
len = std::distance(d, it);
}
__device__ __host__ hip_vector(std::initializer_list<T> x)
{
std::copy(x.begin(), x.end(), d);
len = x.size();
}
MIGRAPHX_DEVICE_CONSTEXPR T& operator[](std::size_t i) { return d[i]; }
MIGRAPHX_DEVICE_CONSTEXPR const T& operator[](std::size_t i) const { return d[i]; }
MIGRAPHX_DEVICE_CONSTEXPR T& front() { return d[0]; }
MIGRAPHX_DEVICE_CONSTEXPR const T& front() const { return d[0]; }
MIGRAPHX_DEVICE_CONSTEXPR T& back() { return d[size() - 1]; }
MIGRAPHX_DEVICE_CONSTEXPR const T& back() const { return d[size() - 1]; }
MIGRAPHX_DEVICE_CONSTEXPR T* data() { return d; }
MIGRAPHX_DEVICE_CONSTEXPR const T* data() const { return d; }
MIGRAPHX_DEVICE_CONSTEXPR std::size_t size() const { return len; }
MIGRAPHX_DEVICE_CONSTEXPR T* begin() { return d; }
MIGRAPHX_DEVICE_CONSTEXPR const T* begin() const { return d; }
MIGRAPHX_DEVICE_CONSTEXPR T* end() { return d + size(); }
MIGRAPHX_DEVICE_CONSTEXPR const T* end() const { return d + size(); }
template <class U>
MIGRAPHX_DEVICE_CONSTEXPR void push_back(U&& x)
{
d[len] = static_cast<U&&>(x);
len++;
}
private:
T d[N] = {};
std::size_t len = 0;
};
template <std::size_t N, class T>
hip_vector<T, N> to_hip_vector(const std::vector<T>& x)
{
hip_vector<T, N> result(x.size());
std::copy(x.begin(), x.end(), result.begin());
return result;
}
} // namespace device
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif
src/targets/gpu/device/include/migraphx/gpu/device/visit.hpp 0 → 100644
View file @ 3855c6af
#ifndef MIGRAPHX_GUARD_RTGLIB_DEVICE_VISIT_HPP
#define MIGRAPHX_GUARD_RTGLIB_DEVICE_VISIT_HPP
#include <migraphx/gpu/device/tensor_view.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
namespace device {
template <class F>
void visit_tensor_size(std::size_t n, F f)
{
switch(n)
{
case 1:
{
f(std::integral_constant<std::size_t, 1>{});
break;
}
case 2:
{
f(std::integral_constant<std::size_t, 2>{});
break;
}
case 3:
{
f(std::integral_constant<std::size_t, 3>{});
break;
}
case 4:
{
f(std::integral_constant<std::size_t, 4>{});
break;
}
case 5:
{
f(std::integral_constant<std::size_t, 5>{});
break;
}
default: throw std::runtime_error("Unknown tensor size");
}
}
inline shape get_shape(const shape& x) { return x; }
template <class T>
auto get_shape(const T& x) -> decltype(x.get_shape())
{
return x.get_shape();
}
template <class V, class F, class... Ts>
void hip_visit_all_impl(const shape& s, F f, V&& v, Ts&&... xs)
{
std::initializer_list<migraphx::shape::type_t> types = {get_shape(xs).type()...};
if(!std::all_of(
types.begin(), types.end(), [&](migraphx::shape::type_t t) { return t == s.type(); }))
MIGRAPHX_THROW("Types must be the same");
std::initializer_list<std::size_t> ranks = {get_shape(xs).lens().size()...};
if(!std::all_of(
ranks.begin(), ranks.end(), [&](std::size_t r) { return r == s.lens().size(); }))
MIGRAPHX_THROW("Ranks must be the same");
visit_tensor_size(s.lens().size(),
[&](auto ndim) { s.visit_type([&](auto as) { v(f(xs, ndim, as)...); }); });
}
template <class V, class F, class... Ts>
void hip_visit_views_impl(const shape& s, F f, V&& v, Ts&&... xs)
{
std::initializer_list<std::size_t> ranks = {get_shape(xs).lens().size()...};
if(!std::all_of(
ranks.begin(), ranks.end(), [&](std::size_t r) { return r == s.lens().size(); }))
MIGRAPHX_THROW("Ranks must be the same");
visit_tensor_size(s.lens().size(), [&](auto ndim) { v(f(xs, ndim)...); });
}
template <class F>
struct hip_convert
{
F f;
template <class RawData, class N, class As>
auto operator()(RawData x, N ndim, As as) const
-> decltype(make_hip_view<ndim>(x.get_shape(), f(as.from(x.data()))))
{
return make_hip_view<ndim>(x.get_shape(), f(as.from(x.data())));
}
template <class N, class As>
auto operator()(const shape& s, N ndim, As) const
{
return make_hip_shape<ndim>(s);
}
};
template <class F>
hip_convert<F> make_hip_convert(F f)
{
return {f};
}
template <class F>
struct hip_convert_view
{
F f;
template <class T, class N>
auto operator()(tensor_view<T> x, N ndim) const
{
return make_hip_view<ndim>(f(x));
}
template <class N>
auto operator()(const shape& s, N ndim) const
{
return make_hip_shape<ndim>(s);
}
};
template <class F>
hip_convert_view<F> make_hip_convert_view(F f)
{
return {f};
}
template <class T, class... Ts>
auto hip_visit_all(T&& x, Ts&&... xs)
{
return [&](auto f) {
hip_visit_all_impl(
get_shape(x), make_hip_convert([](auto* p) { return device_cast(p); }), f, x, xs...);
};
}
template <std::size_t N, class T, class... Ts>
auto hip_vec_visit_all(T&& x, Ts&&... xs)
{
return [&](auto f) {
hip_visit_all_impl(get_shape(x),
make_hip_convert([](auto* p) { return as_vec<N>(device_cast(p)); }),
f,
x,
xs...);
};
}
template <class T, class... Ts>
auto hip_pointer_visit_all(T&& x, Ts&&... xs)
{
return [&](auto f) { visit_all(x, xs...)([&](auto... vs) { f(device_cast(vs.data())...); }); };
}
template <class T, class... Ts>
auto hip_visit_views(T&& x, Ts&&... xs)
{
return [&](auto f) {
hip_visit_views_impl(get_shape(x),
make_hip_convert_view([](auto v) { return device_cast(v); }),
f,
x,
xs...);
};
}
} // namespace device
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif
src/targets/gpu/device/logsoftmax.cpp
View file @ 3855c6af
#include <migraphx/shape.hpp>
#include <migraphx/argument.hpp>
#include <migraphx/gpu/device/logsoftmax.hpp>
#include <migraphx/gpu/device/reduce_opers.hpp>
#include <migraphx/gpu/device/tensor.hpp>
#include <migraphx/gpu/device/launch.hpp>
#include <migraphx/gpu/device/types.hpp>
......@@ -11,127 +12,82 @@ inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
namespace device {
argument logsoftmax(hipStream_t stream,
const migraphx::shape& output_shape,
std::vector<migraphx::argument> args,
int axis)
void logsoftmax(hipStream_t stream, argument result, argument arg, int axis)
{
auto lens = output_shape.lens();
auto num_in_batch = lens[axis];
auto batch_lens = lens;
batch_lens[axis] = 1;
migraphx::shape batch_shape{output_shape.type(), batch_lens};
visit_all(args.back(), args.front())([&](auto output, auto input) {
const auto* input_ptr = device_cast(input.data());
auto* output_ptr = device_cast(output.data());
visit_tensor_size(batch_shape.lens().size(), [&](auto n_dim) {
hip_tensor_descriptor<n_dim> desc_batch(batch_shape);
hip_tensor_descriptor<n_dim> desc_data(output_shape);
// use one block for items in one batch.
// opt 1, load all data to lds then use the same approach as
// the current optimization
const size_t block_size = 1024;
launch(
stream, batch_shape.elements() * block_size, block_size)([=](auto idx) __device__ {
size_t thr_idx = idx.local;
size_t blk_idx = idx.group;
using type = device_type<std::remove_cv_t<typename decltype(output)::value_type>>;
// all data can be loaded to the lds once, so all operations are
// done in lds
MIGRAPHX_DEVICE_SHARED type lds_data[block_size + 2];
auto batch_idx = desc_batch.multi(blk_idx);
auto data_idx = batch_idx;
// load data to lds and compute the batch max
size_t item_num = num_in_batch;
lds_data[block_size] = input_ptr[0];
for(size_t i = thr_idx; i < num_in_batch; i += block_size)
auto lens = result.get_shape().lens();
auto batch_item_num = lens[axis];
auto batch_lens = lens;
batch_lens[axis] = 1;
migraphx::shape batch_shape{result.get_shape().type(), batch_lens};
hip_visit_all(result, arg, batch_shape)([&](auto output, auto input, auto batch) {
// use one block for items in one batch.
const size_t max_block_size = 1024;
size_t block_size = 1;
while(block_size < max_block_size and block_size < batch_item_num)
{
block_size *= 2;
}
launch(stream, batch_shape.elements() * block_size, block_size)([=](auto idx) __device__ {
size_t thr_idx = idx.local;
size_t blk_idx = idx.group;
using type = device_type<std::remove_cv_t<typename decltype(output)::value_type>>;
MIGRAPHX_DEVICE_SHARED type lds_data[max_block_size + 1];
auto batch_idx = batch.multi(blk_idx);
auto data_idx = batch_idx;
// load data to lds and compute the batch max
size_t remaining_item_num = batch_item_num;
size_t round_item_num = (batch_item_num + block_size - 1) / block_size * block_size;
lds_data[block_size] = input[0];
for(size_t i = thr_idx; i < round_item_num; i += block_size)
{
if(i < batch_item_num)
{
data_idx[axis] = i;
lds_data[i] = input_ptr[desc_data.linear(data_idx)];
__syncthreads();
auto size = (item_num > block_size) ? block_size : item_num;
auto stride = (size + 1) / 2;
while(true)
{
if(thr_idx + stride < size)
{
lds_data[thr_idx] = ::max(to_hip_type(lds_data[thr_idx]),
to_hip_type(lds_data[thr_idx + stride]));
}
__syncthreads();
size = stride;
stride = (stride + 1) / 2;
if(size == 1)
break;
}
if(thr_idx == 0)
{
lds_data[block_size] = (lds_data[0] < lds_data[block_size])
? lds_data[block_size]
: lds_data[0];
}
__syncthreads();
item_num -= block_size;
data_idx[axis] = i;
lds_data[thr_idx] = input[data_idx];
}
__syncthreads();
const size_t block_size1 = block_size + 1;
lds_data[block_size1] = 0;
item_num = num_in_batch;
for(size_t i = thr_idx; i < num_in_batch; i += block_size)
{
data_idx[axis] = i;
lds_data[i] = input_ptr[desc_data.linear(data_idx)] - lds_data[block_size];
lds_data[i] = ::exp(to_hip_type(lds_data[i]));
__syncthreads();
auto size = (item_num > block_size) ? block_size : item_num;
auto stride = (size + 1) / 2;
while(true)
{
if(thr_idx + stride < size)
{
lds_data[thr_idx] += lds_data[thr_idx + stride];
}
__syncthreads();
size = stride;
stride = (stride + 1) / 2;
if(size == 1)
break;
}
if(thr_idx == 0)
{
lds_data[block_size1] += lds_data[0];
}
__syncthreads();
item_num -= block_size;
}
auto item_num = (remaining_item_num > block_size) ? block_size : remaining_item_num;
reduce_max(lds_data, block_size, thr_idx, item_num);
auto log_batch_sum =
::log(to_hip_type(lds_data[block_size1])) + lds_data[block_size];
item_num = num_in_batch;
for(size_t i = thr_idx; i < num_in_batch; i += block_size)
remaining_item_num -= block_size;
}
auto batch_max = lds_data[block_size];
__syncthreads();
lds_data[block_size] = 0;
remaining_item_num = batch_item_num;
for(size_t i = thr_idx; i < round_item_num; i += block_size)
{
if(i < batch_item_num)
{
data_idx[axis] = i;
size_t index = desc_data.linear(data_idx);
output_ptr[index] = input_ptr[index] - log_batch_sum;
lds_data[thr_idx] = input[data_idx] - batch_max;
lds_data[thr_idx] = ::exp(to_hip_type(lds_data[thr_idx]));
}
});
__syncthreads();
auto item_num = (remaining_item_num > block_size) ? block_size : remaining_item_num;
reduce_sum(lds_data, block_size, thr_idx, item_num);
remaining_item_num -= block_size;
}
auto log_batch_sum = ::log(to_hip_type(lds_data[block_size])) + batch_max;
for(size_t i = thr_idx; i < batch_item_num; i += block_size)
{
data_idx[axis] = i;
output[data_idx] = input[data_idx] - log_batch_sum;
}
});
});
return args.back();
}
} // namespace device
......
src/targets/gpu/device/pad.cpp
View file @ 3855c6af
......@@ -15,33 +15,26 @@ argument
pad(hipStream_t stream, argument result, argument arg1, float value, std::vector<std::int64_t> pads)
{
std::size_t nelements = arg1.get_shape().elements();
visit_all(result)([&](auto output) {
auto* outptr = device_cast(output.data());
using type = typename decltype(output)::value_type;
device_type<type> device_val = value;
hip_visit_all(result, arg1)([&](auto output, auto input) {
using type = typename decltype(output)::value_type;
using hip_index = typename decltype(output)::hip_index;
type device_val = value;
if(float_equal(value, std::numeric_limits<float>::lowest()))
{
device_val = device_cast(std::numeric_limits<type>::lowest());
}
gs_launch(stream, result.get_shape().elements())([=](auto i) { outptr[i] = device_val; });
});
gs_launch(stream,
result.get_shape().elements())([=](auto i) { output.data()[i] = device_val; });
visit_all(result, arg1)([&](auto output, auto input) {
visit_tensor_size(result.get_shape().lens().size(), [&](auto ndim) {
std::size_t offsets[ndim];
std::copy(pads.begin(), pads.begin() + ndim, offsets);
auto* outptr = output.data();
const auto* inptr = input.data();
hip_tensor_descriptor<ndim> desc_input(input.get_shape());
hip_tensor_descriptor<ndim> desc_output(output.get_shape());
gs_launch(stream, nelements)([=](auto i) {
auto idx = desc_input.multi(i);
for(std::size_t j = 0; j < ndim; j++)
{
idx[j] += offsets[j];
}
outptr[desc_output.linear(idx)] = inptr[i];
});
hip_index offsets;
std::copy(pads.begin(), pads.begin() + offsets.size(), offsets.begin());
gs_launch(stream, nelements)([=](auto i) {
auto idx = input.get_shape().multi(i);
for(std::size_t j = 0; j < offsets.size(); j++)
{
idx[j] += offsets[j];
}
output[idx] = input.data()[i];
});
});
return result;
......
src/targets/gpu/device/softmax.cpp
View file @ 3855c6af
......@@ -2,6 +2,7 @@
#include <migraphx/argument.hpp>
#include <migraphx/dfor.hpp>
#include <migraphx/gpu/device/softmax.hpp>
#include <migraphx/gpu/device/reduce_opers.hpp>
#include <migraphx/gpu/device/tensor.hpp>
#include <migraphx/gpu/device/launch.hpp>
#include <migraphx/gpu/device/types.hpp>
......@@ -12,122 +13,82 @@ inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
namespace device {
argument softmax(hipStream_t stream,
const migraphx::shape& output_shape,
std::vector<migraphx::argument> args,
int axis)
void softmax(hipStream_t stream, argument result, argument arg, int axis)
{
auto lens = output_shape.lens();
auto batch_lens = lens;
size_t n_dims = lens[axis];
batch_lens[axis] = 1;
migraphx::shape batch_shape{shape::int32_type, batch_lens};
visit_all(args.back(), args.front())([&](auto output, auto input) {
const auto* input_ptr = device_cast(input.data());
auto* output_ptr = device_cast(output.data());
visit_tensor_size(batch_shape.lens().size(), [&](auto n_dim) {
hip_tensor_descriptor<n_dim> desc_batch(batch_shape);
hip_tensor_descriptor<n_dim> desc_data(output_shape);
// use one block for items in one batch.
const size_t block_size = 1024;
launch(
stream, batch_shape.elements() * block_size, block_size)([=](auto idx) __device__ {
size_t thr_idx = idx.local;
size_t blk_idx = idx.group;
using type = device_type<std::remove_cv_t<typename decltype(output)::value_type>>;
// all data can be loaded to the lds once, so all operations are
// done in lds
MIGRAPHX_DEVICE_SHARED type lds_data[block_size + 2];
auto batch_idx = desc_batch.multi(blk_idx);
auto data_idx = batch_idx;
// load data to lds and compute the batch max
size_t item_num = n_dims;
lds_data[block_size] = input_ptr[0];
for(size_t i = thr_idx; i < n_dims; i += block_size)
auto lens = result.get_shape().lens();
auto batch_lens = lens;
size_t batch_item_num = lens[axis];
batch_lens[axis] = 1;
migraphx::shape batch_shape{result.get_shape().type(), batch_lens};
hip_visit_all(result, arg, batch_shape)([&](auto output, auto input, auto batch) {
// use one block for items in one batch.
const size_t max_block_size = 1024;
size_t block_size = 1;
while(block_size < max_block_size and block_size < batch_item_num)
{
block_size *= 2;
}
launch(stream, batch_shape.elements() * block_size, block_size)([=](auto idx) __device__ {
size_t thr_idx = idx.local;
size_t blk_idx = idx.group;
using type = device_type<std::remove_cv_t<typename decltype(output)::value_type>>;
MIGRAPHX_DEVICE_SHARED type lds_data[max_block_size + 1];
auto batch_idx = batch.multi(blk_idx);
auto data_idx = batch_idx;
// load data to lds and compute the batch max
size_t remaining_item_num = batch_item_num;
size_t round_item_num = (batch_item_num + block_size - 1) / block_size * block_size;
lds_data[block_size] = input[0];
for(size_t i = thr_idx; i < round_item_num; i += block_size)
{
if(i < batch_item_num)
{
data_idx[axis] = i;
lds_data[i] = input_ptr[desc_data.linear(data_idx)];
__syncthreads();
data_idx[axis] = i;
lds_data[thr_idx] = input[data_idx];
}
auto size = (item_num > block_size) ? block_size : item_num;
auto stride = (size + 1) / 2;
while(true)
{
if(thr_idx + stride < size)
{
lds_data[thr_idx] = ::max(to_hip_type(lds_data[thr_idx]),
to_hip_type(lds_data[thr_idx + stride]));
}
__syncthreads();
size = stride;
stride = (stride + 1) / 2;
__syncthreads();
if(size == 1)
break;
}
auto item_num = (remaining_item_num > block_size) ? block_size : remaining_item_num;
reduce_max(lds_data, block_size, thr_idx, item_num);
if(thr_idx == 0)
{
lds_data[block_size] = (lds_data[0] < lds_data[block_size])
? lds_data[block_size]
: lds_data[0];
}
__syncthreads();
remaining_item_num -= block_size;
}
item_num -= block_size;
}
auto batch_max = lds_data[block_size];
__syncthreads();
const size_t block_size1 = block_size + 1;
lds_data[block_size1] = 0;
item_num = n_dims;
for(size_t i = thr_idx; i < n_dims; i += block_size)
lds_data[block_size] = 0;
remaining_item_num = batch_item_num;
for(size_t i = thr_idx; i < round_item_num; i += block_size)
{
if(i < batch_item_num)
{
data_idx[axis] = i;
lds_data[i] = input_ptr[desc_data.linear(data_idx)] - lds_data[block_size];
lds_data[i] = ::exp(to_hip_type(lds_data[i]));
__syncthreads();
data_idx[axis] = i;
lds_data[thr_idx] = input[data_idx] - batch_max;
lds_data[thr_idx] = ::exp(to_hip_type(lds_data[thr_idx]));
}
auto size = (item_num > block_size) ? block_size : item_num;
auto stride = (size + 1) / 2;
while(true)
{
if(thr_idx + stride < size)
{
lds_data[thr_idx] += lds_data[thr_idx + stride];
}
__syncthreads();
size = stride;
stride = (stride + 1) / 2;
if(size == 1)
break;
}
__syncthreads();
if(thr_idx == 0)
{
lds_data[block_size1] += lds_data[0];
}
__syncthreads();
auto item_num = (remaining_item_num > block_size) ? block_size : remaining_item_num;
reduce_sum(lds_data, block_size, thr_idx, item_num);
item_num -= block_size;
}
remaining_item_num -= block_size;
}
auto batch_sum = lds_data[block_size];
for(size_t i = thr_idx; i < n_dims; i += block_size)
{
data_idx[axis] = i;
size_t index = desc_data.linear(data_idx);
auto val = input_ptr[index] - lds_data[block_size];
output_ptr[index] = ::exp(to_hip_type(val)) / lds_data[block_size1];
}
});
for(size_t i = thr_idx; i < batch_item_num; i += block_size)
{
data_idx[axis] = i;
auto val = input[data_idx] - batch_max;
output[data_idx] = ::exp(to_hip_type(val)) / batch_sum;
}
});
});
return args.back();
}
} // namespace device
......
src/targets/gpu/fuse_ops.cpp
View file @ 3855c6af
......@@ -5,6 +5,7 @@
#include <migraphx/gpu/device/add_relu.hpp>
#include <migraphx/gpu/device/add.hpp>
#include <migraphx/instruction.hpp>
#include <migraphx/array.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
......@@ -122,13 +123,6 @@ MIGRAPHX_PRED_MATCHER(bias_shape, instruction_ref ins)
s.strides()[1] != 0 and s.strides()[2] == 0 and s.strides()[3] == 0;
}
// TODO: Move to another header
template <class T, class... Ts>
std::array<T, sizeof...(Ts) + 1> make_array(T x, Ts... xs)
{
return {std::move(x), std::move(static_cast<T>(xs))...};
}
MIGRAPHX_PRED_MATCHER(fusable_conv, instruction_ref ins)
{
if(ins->name() != "gpu::convolution")
......@@ -408,8 +402,8 @@ void fuse_ops::apply(program& p) const
// clang-format off
match::find_matches(p, find_triadd{});
match::find_matches(p,
find_conv_bias_relu{ctx},
find_conv_bias{ctx},
// find_conv_bias_relu{ctx},
// find_conv_bias{ctx},
find_add_relu{}
);
// clang-format on
......
src/targets/gpu/gather.cpp
View file @ 3855c6af
......@@ -12,11 +12,9 @@ shape hip_gather::compute_shape(std::vector<shape> inputs) const
return op.compute_shape(inputs);
}
argument hip_gather::compute(context& ctx,
const shape& output_shape,
const std::vector<argument>& args) const
argument hip_gather::compute(context& ctx, const shape&, const std::vector<argument>& args) const
{
return device::gather(ctx.get_stream().get(), output_shape, args, op.axis);
return device::gather(ctx.get_stream().get(), args.back(), args[0], args[1], op.axis);
}
} // namespace gpu
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment