Skip to content

GitLab

Commit a045fb19 authored by Alan Turner's avatar Alan Turner
Browse files

Merge branch 'develop' into ck-flash-attn

parents 135eb63e 434a06cf
Hide whitespace changes
Inline Side-by-side
Showing with 453 additions and 63 deletions
src/include/migraphx/op/convolution_backwards.hpp
View file @ a045fb19
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
* Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
......@@ -164,7 +164,7 @@ struct convolution_backwards
shape win_shape{dyn_out.computed_shape.type(), win_size};
par_dfor(in_n, wei_c)([&](int o, int k) {
shape_for_each(win_shape, [&](auto idx_win) {
shape_for_each(win_shape, [&](const auto& idx_win) {
const int w = idx_win[0];
auto input_dims_start = idx_win.begin() + 1;
......
src/include/migraphx/op/fill.hpp 0 → 100644
View file @ a045fb19
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
* THE SOFTWARE.
*/
#ifndef MIGRAPHX_GUARD_OPERATORS_FILL_HPP
#define MIGRAPHX_GUARD_OPERATORS_FILL_HPP
#include <migraphx/check_shapes.hpp>
#include <migraphx/dyn_output.hpp>
#include <migraphx/par_for.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace op {
/**
* fill(default_value, output_buffer)
* Fill an output buffer with the given default_value.
* Note that if the default_value is a literal and the output_buffer
* has a static shape this operator can be replaced with a literal.
*/
struct fill
{
std::string name() const { return "fill"; }
shape compute_shape(std::vector<shape> inputs) const
{
check_shapes{inputs, *this, true}.has(2).same_type();
if(inputs.at(0).dynamic() or inputs.at(0).elements() != 1)
{
MIGRAPHX_THROW("FILL: default_value is dynamic or more than one element");
}
return inputs.back();
}
argument compute(const dyn_output& dyn_out, std::vector<argument> args) const
{
visit_all(args[0], args[1])([&](auto value, auto output) {
par_for(dyn_out.computed_shape.elements(), [&](auto i) { output[i] = value.front(); });
});
return args[1];
}
std::ptrdiff_t output_alias(const std::vector<shape>&) const { return 1; }
};
} // namespace op
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif
src/include/migraphx/op/gather.hpp
View file @ a045fb19
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
* Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
......@@ -125,13 +125,12 @@ struct gather
auto out_lens = data.get_shape().lens();
out_lens[axis] = indices.get_shape().elements();
migraphx::shape out_comp_shape{data.get_shape().type(), out_lens};
shape_for_each(out_comp_shape, [&](const auto& out_idx) {
auto data_idx = out_idx;
auto in_index = indices[data_idx[axis]];
in_index = (in_index < 0) ? in_index + axis_dim_size : in_index;
data_idx[axis] = in_index;
output[out_comp_shape.index(out_idx.begin(), out_idx.end())] =
data(data_idx.begin(), data_idx.end());
shape_for_each(out_comp_shape, [&](const auto& out_idx_v, size_t out_idx) {
auto data_idx = out_idx_v;
auto in_index = indices[data_idx[axis]];
in_index = (in_index < 0) ? in_index + axis_dim_size : in_index;
data_idx[axis] = in_index;
output[out_idx] = data(data_idx.begin(), data_idx.end());
});
}
});
......
src/include/migraphx/op/nonmaxsuppression.hpp
View file @ a045fb19
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
* Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
......@@ -258,7 +258,7 @@ struct nonmaxsuppression
selected_boxes_inside_class.reserve(max_output_shape.elements());
// iterate over batches and classes
shape comp_s{shape::double_type, {num_batches, num_classes}};
shape_for_each(comp_s, [&](auto idx) {
shape_for_each(comp_s, [&](const auto& idx) {
auto batch_idx = idx[0];
auto class_idx = idx[1];
// index offset for this class
......
src/include/migraphx/op/nonzero.hpp
View file @ a045fb19
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
* Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
......@@ -56,10 +56,10 @@ struct nonzero
std::vector<std::vector<std::size_t>> vec_idx;
auto s = args.front().get_shape();
args.front().visit([&](auto v) {
shape_for_each(s, [&](auto idx) {
if(not float_equal(v[s.index(idx)], 0))
shape_for_each(s, [&](const auto& idx_v, size_t idx) {
if(not float_equal(v[idx], 0))
{
vec_idx.push_back(idx);
vec_idx.push_back(idx_v);
}
});
});
......
src/include/migraphx/op/pooling.hpp
View file @ a045fb19
......@@ -29,6 +29,7 @@
#include <migraphx/config.hpp>
#include <migraphx/value.hpp>
#include <migraphx/argument.hpp>
#include <migraphx/pad_calc.hpp>
#include <migraphx/par_for.hpp>
#include <migraphx/shape_for_each.hpp>
#include <migraphx/dyn_output.hpp>
......@@ -40,10 +41,20 @@ namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace op {
// The Pooling operator mostly follows the specifications for the Onnx pooling op.
// It assumes an NCHW layout, extended to support any number of spatial dimensions
// from 1 on up; dimensions are <batch index, channels, spatial dimensions...>
//
struct pooling
{
// Class members mode, ceil_mode, padding_mode have similar names but refer to separate
// concepts.
pooling_mode mode = {pooling_mode::average};
// If the input has rank other than 4 then padding, stride, lengths must all be specified
// since the defaults have 2-dimensions. Exception: padding not required if
// padding_mode != default_
// Padding along each spatial input dimension
// Can be ndim or 2*ndim values where ndim is size of lengths
// ndim values means pad the same before and after each dimension
......@@ -63,13 +74,14 @@ struct pooling
// ceiling mode is a flag affecting output size
// or equivalently, placements of the pooling kernel.
// When true, round the size upwards, possibly
// including partial placements where the kernel extends beyond the edge
// of input and even padding. When false, round down so that all
// When true, round the size upwards. When false, round down so that all
// kernel placements fit but some input values may be dropped.
bool ceil_mode = false;
int lp_order = 2;
// Mode for auto padding. default_ indicates no auto padding.
padding_mode_t padding_mode = padding_mode_t::default_;
// Global pooling with dynamic shape input
bool dyn_global = false;
......@@ -84,6 +96,7 @@ struct pooling
{
return pack(f(self.mode, "mode"),
f(self.padding, "padding"),
f(self.padding_mode, "padding_mode"),
f(self.stride, "stride"),
f(self.lengths, "lengths"),
f(self.ceil_mode, "ceil_mode"),
......@@ -97,7 +110,8 @@ struct pooling
{
if(dyn_global)
return;
if((padding.size() != stride.size() and (padding.size()) != stride.size() * 2) or
if((padding_mode != default_ and padding.size() != stride.size() and
(padding.size()) != stride.size() * 2) or
stride.size() != lengths.size())
{
MIGRAPHX_THROW("POOLING: inconsistent attribute sizes");
......@@ -137,8 +151,19 @@ struct pooling
std::size_t padding_factor = 2 * padding[i];
if(padding.size() == 2 * kdims)
padding_factor = padding[i] + padding[i + kdims];
assert(input_lens[i + 2] + padding_factor >= lengths[i]);
std::size_t dim_size = input_lens[i + 2] + padding_factor - lengths[i];
std::size_t dim_size;
if(input_lens[i + 2] + padding_factor < lengths[i])
{
if(padding_mode == default_)
MIGRAPHX_THROW("POOLING: not enough padding for the given kernel size");
// lengths can be legitimately larger only if we're doing auto padding
// with a dynamic shape, in which case given padding is ignored. Set a dummy value.
dim_size = 2;
}
else
{
dim_size = input_lens[i + 2] + padding_factor - lengths[i];
}
std::size_t len =
(ceil_mode)
? dim_size / stride[i] +
......@@ -151,17 +176,13 @@ struct pooling
shape normalize_compute_shape(std::vector<shape> inputs) const
{
check_shapes{inputs, *this, true}.has(1);
check_shapes{inputs, *this, true}.has(1).min_ndims(3);
check_attribute_size();
const shape& input = inputs.at(0);
auto padding_size = padding.size();
auto stride_size = stride.size();
size_t kdims = input.ndim() - 2;
if(input.ndim() < 3)
{
MIGRAPHX_THROW("POOLING: input must have 3 or more dimensions and be nonempty");
}
if(input.ndim() * 2 != padding_size + 4 and input.ndim() != padding_size + 2)
if(input.ndim() != stride_size + 2)
{
MIGRAPHX_THROW("POOLING: input and attribute size mismatch!");
}
......@@ -179,6 +200,28 @@ struct pooling
}
return {input.type(), output_dyn_dims};
}
else if(padding_mode != default_)
{
const size_t num_spatial_dims = inputs[0].ndim() - 2;
const shape& x_shape = inputs[0];
// same as convolution::dynamic_compute_shape()
for(std::size_t i = 0; i < num_spatial_dims; ++i)
{
auto ceil_div = [](std::size_t x, std::size_t y) { return (x + y - 1) / y; };
auto s = stride[i];
auto x = x_shape.dyn_dims()[i + 2];
std::set<std::size_t> optimals{};
std::transform(x.optimals.begin(),
x.optimals.end(),
std::inserter(optimals, optimals.begin()),
[&](auto o) { return ceil_div(o, s); });
output_dyn_dims.push_back(
shape::dynamic_dimension{ceil_div(x.min, s), ceil_div(x.max, s), optimals});
}
return {input.type(), output_dyn_dims};
}
else
{
// does not compute optimals
......@@ -267,6 +310,7 @@ struct pooling
Out& output,
const In& input,
const std::vector<std::size_t>& kernel_dims,
const std::vector<std::size_t>& padding_vals,
Op op) const
{
auto in_s = input.get_shape();
......@@ -283,9 +327,9 @@ struct pooling
// For each spatial dimension, find starting and ending index of pooling kernel
for(std::size_t dim = 2; dim < n_dim; ++dim)
{
auto d_2 = dim - 2;
int start =
static_cast<int>(idx_o[dim] * stride[d_2]) - static_cast<int>(padding[d_2]);
auto d_2 = dim - 2;
int start = static_cast<int>(idx_o[dim] * stride[d_2]) -
static_cast<int>(padding_vals[d_2]);
int end;
// NOLINT
if(count_include_pad and ceil_mode and (mode != pooling_mode::max))
......@@ -297,7 +341,7 @@ struct pooling
// Check if this kernel extends beyond the padding at end of dimension
end = std::min(start + kernel_dims[d_2],
in_lens[dim] + static_cast<int>(padding[d_2]));
in_lens[dim] + static_cast<int>(padding_vals[d_2]));
}
else
{
......@@ -316,11 +360,12 @@ struct pooling
}
shape win_shape{output_shape.type(), win_size};
auto pool_size = win_shape.elements();
double output_val = op.template init<Type>();
// for each element in the window...
shape_for_each(win_shape, [&](auto idx_w) {
shape_for_each(win_shape, [&](const auto& idx_w) {
// the coordinates of this element
auto idx = idx_o;
......@@ -354,30 +399,65 @@ struct pooling
argument compute(const dyn_output& dyn_out, std::vector<argument> args) const
{
argument result{dyn_out.computed_shape};
argument result;
auto input_lens = args[0].get_shape().lens();
std::vector<std::size_t> kernel_dims;
shape output_shape;
// If we have to auto-calculate padding, it will be passed to calc_pooling() as an argument
// instead of the member variable padding.
std::vector<std::size_t> temp_padding(padding);
if(dyn_global)
{
// for dynamic GlobalPooling, there's no padding
kernel_dims.insert(kernel_dims.end(), input_lens.begin() + 2, input_lens.end());
output_shape = dyn_out.computed_shape;
result = dyn_out.computed_shape;
}
else
else if((padding_mode != op::padding_mode_t::default_))
{
// if padding_mode is set, input was a dynamic size. Calculate padded size now.
// kernel_lens is the same as kernel_dims, but prepended with the 2 non-
// spatial dimensions. For size computations, it's used like the weights
// tensor for convolutions.
std::vector<std::size_t> kernel_lens;
kernel_lens.insert(kernel_lens.end(), input_lens.begin(), input_lens.begin() + 2);
kernel_lens.insert(kernel_lens.end(), lengths.begin(), lengths.end());
kernel_dims = this->lengths;
auto type = args[0].get_shape().type();
// dilation not currently supported for pooling, so default to all 1's
temp_padding = calc_dyn_auto_pad(
input_lens, kernel_lens, stride, {1, 1}, bool(padding_mode == op::same_upper));
output_shape = compute_padded_pool_shape(
args[0].get_shape(), shape(type, kernel_dims), temp_padding, stride, {1, 1});
result = argument(output_shape);
}
else // fixed/static input
{
kernel_dims = this->lengths;
output_shape = dyn_out.computed_shape;
result = dyn_out.computed_shape;
}
// Perform the computation and populate result
visit_all(result, args[0])([&](auto output, auto input) {
using type = typename decltype(output)::value_type;
switch(mode)
{
case migraphx::op::pooling_mode::average:
calc_pooling<type>(dyn_out.computed_shape, output, input, kernel_dims, avg_pool{});
calc_pooling<type>(
output_shape, output, input, kernel_dims, temp_padding, avg_pool{});
break;
case migraphx::op::pooling_mode::max:
calc_pooling<type>(dyn_out.computed_shape, output, input, kernel_dims, max_pool{});
calc_pooling<type>(
output_shape, output, input, kernel_dims, temp_padding, max_pool{});
break;
case migraphx::op::pooling_mode::lpnorm:
calc_pooling<type>(
dyn_out.computed_shape, output, input, kernel_dims, lpnorm_pool{lp_order});
output_shape, output, input, kernel_dims, temp_padding, lpnorm_pool{lp_order});
break;
}
});
......
src/include/migraphx/op/random_seed.hpp 0 → 100644
View file @ a045fb19
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
* THE SOFTWARE.
*/
#ifndef MIGRAPHX_GUARD_OPERATORS_RANDOM_SEED_HPP
#define MIGRAPHX_GUARD_OPERATORS_RANDOM_SEED_HPP
#include <migraphx/check_shapes.hpp>
#include <migraphx/argument.hpp>
#include <random>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace op {
/**
* Generates a random seed for the use of random number generators. Generating the seed
* at runtime guarantees there will be a different random sequence on every execution.
* This operation has no inputs or attributes, and outputs an unsigned integer tensor with
* a single value.
*/
struct random_seed
{
shape::type_t dtype = shape::type_t::uint64_type;
template <class Self, class F>
static auto reflect(Self& self, F f)
{
return pack(f(self.dtype, "dtype"));
}
std::string name() const { return "random_seed"; }
shape compute_shape(const std::vector<shape>& inputs) const
{
check_shapes{inputs, *this}.has(0);
return shape{dtype};
}
argument compute(const shape& output_shape, const std::vector<argument>&) const
{
argument result(output_shape);
result.visit([&](auto output) { output.front() = std::random_device{}(); });
return result;
}
};
} // namespace op
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif
src/include/migraphx/op/random_uniform.hpp 0 → 100644
View file @ a045fb19
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
* THE SOFTWARE.
*/
/**
* Random Uniform distribution operator. Given a shape, populate it with random
* values. Calls to random_uniform using the same randomization seed as a
* literal input will
* always generate the same pseudo-random sequence.
*
* Inputs: (1) randomization seed (any type is allowed)
* (2) output buffer argument to be populated.
*
* Attributes: none
*
* Output: Returns the buffer from input #2.
*
*/
#ifndef MIGRAPHX_GUARD_OPERATORS_RANDOM_UNIFORM_HPP
#define MIGRAPHX_GUARD_OPERATORS_RANDOM_UNIFORM_HPP
#include <migraphx/check_shapes.hpp>
#include <migraphx/argument.hpp>
#include <random>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace op {
/**
* random_uniform populates the passed shape with random numbers, in a uniform
* distribution. Range for floating-point data types is (0, 1);
* for integer types it is [0, <max value for the type>]
*/
struct random_uniform
{
// The random_uniform operation needs the random number generator seed
// to be passed as a runtime input.
std::string name() const { return "random_uniform"; }
shape compute_shape(std::vector<shape> inputs) const
{
check_shapes{inputs, *this, true}.has(2);
return inputs.at(1);
}
argument compute(const shape&, std::vector<argument> args) const
{
// Output goes into the passed buffer, not the shape output.
auto result = args[1];
uint64_t local_seed = args[0].at<uint64_t>(0);
std::mt19937 gen(local_seed);
result.visit([&](auto output) {
using type = typename decltype(output)::value_type;
if constexpr(std::is_integral<type>{})
{
// default range for all integer types is
// (0, std::uniform_int_distribution<type>::max()).
// Todo: enable different ranges
std::uniform_int_distribution<type> dis;
std::generate(output.begin(), output.end(), [&] { return dis(gen); });
}
else
{
// default real distribution type is double with range (0, 1);
std::uniform_real_distribution<> dis;
std::generate(output.begin(), output.end(), [&] { return dis(gen); });
}
});
return result;
}
std::ptrdiff_t output_alias(const std::vector<shape>&) const { return 1; }
};
} // namespace op
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif
src/include/migraphx/op/reduce_op.hpp
View file @ a045fb19
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
* Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
......@@ -163,7 +163,7 @@ struct reduce_op : op_name<Derived>
auto& self = static_cast<const Derived&>(*this);
auto data_idx = out_idx;
accumulator val = self.init();
shape_for_each(batch_shape, [&](auto b_idx) {
shape_for_each(batch_shape, [&](const auto& b_idx) {
this->tune_dims(tuned_axes, b_idx, data_idx);
accumulator x = input(data_idx.begin(), data_idx.end());
val = self.op()(accumulator{self.input()(x)}, val);
......
src/include/migraphx/op/reverse.hpp
View file @ a045fb19
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
* Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
......@@ -70,13 +70,13 @@ struct reverse
argument result{s};
auto lens = s.lens();
visit_all(result, args.front())([&](auto output, auto input) {
shape_for_each(s, [&](const auto& out_idx) {
auto in_idx = out_idx;
shape_for_each(s, [&](const auto& out_idx_v, size_t out_idx) {
auto in_idx = out_idx_v;
for(const auto& axis : axes)
{
in_idx[axis] = lens[axis] - 1 - out_idx[axis];
in_idx[axis] = lens[axis] - 1 - out_idx_v[axis];
}
output[s.index(out_idx)] = input[s.index(in_idx)];
output[out_idx] = input[s.index(in_idx)];
});
});
......
src/include/migraphx/op/roialign.hpp
View file @ a045fb19
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
* Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
......@@ -113,10 +113,9 @@ struct roialign
{
std::vector<pos_weight> results(bin_grid_size[0] * bin_grid_size[1] * output_height *
output_width);
shape_for_each(comp_s, [&](auto idx) {
std::array<std::size_t, 2> p = {idx[0], idx[1]};
std::array<std::size_t, 2> i = {idx[2], idx[3]};
auto index = comp_s.index(idx);
shape_for_each(comp_s, [&](const auto& idx_v, size_t index) {
std::array<std::size_t, 2> p = {idx_v[0], idx_v[1]};
std::array<std::size_t, 2> i = {idx_v[2], idx_v[3]};
std::array<float, 2> xy{};
std::array<int64_t, 2> low{};
......@@ -125,7 +124,7 @@ struct roialign
{
xy[ii] = roi_start[ii] + p[ii] * bin_size[ii] +
(i[ii] + .5f) * bin_size[ii] / bin_grid_size[ii];
xy[ii] = (coord_trans_mode == "output_half_pixel") ? (xy[ii] - 0.5f) : xy[ii];
xy[ii] = (coord_trans_mode == "half_pixel") ? (xy[ii] - 0.5f) : xy[ii];
if(xy[ii] < -1.0 or xy[ii] > dims[ii])
{
results[index] = pos_weight{};
......@@ -255,7 +254,7 @@ struct roialign
std::vector<std::size_t> comp_lens1 = {channels, out_dims[0], out_dims[1]};
shape comp_s1{migraphx::shape::float_type, comp_lens1};
std::vector<int64_t> vec_index(channels, 0);
shape_for_each(comp_s1, [&](auto idx) {
shape_for_each(comp_s1, [&](const auto& idx) {
auto c = idx[0];
auto ph = idx[1];
auto pw = idx[2];
......
src/include/migraphx/op/scatter.hpp
View file @ a045fb19
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
* Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
......
src/include/migraphx/operators.hpp
View file @ a045fb19
......@@ -55,6 +55,7 @@
#include <migraphx/op/equal.hpp>
#include <migraphx/op/erf.hpp>
#include <migraphx/op/exp.hpp>
#include <migraphx/op/fill.hpp>
#include <migraphx/op/flatten.hpp>
#include <migraphx/op/floor.hpp>
#include <migraphx/op/fmod.hpp>
......
src/include/migraphx/pad_calc.hpp
View file @ a045fb19
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
* Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
......@@ -62,6 +62,14 @@ shape compute_padded_shape(const shape& input,
const std::vector<std::size_t>& stride,
const std::vector<std::size_t>& dilation);
// Used for dynamic auto padding of pooling operators where padding needs to be computed at
// evaulation time.
shape compute_padded_pool_shape(const shape& input,
const shape& kernel,
const std::vector<std::size_t>& padding,
const std::vector<std::size_t>& stride,
const std::vector<std::size_t>& dilation);
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
......
src/include/migraphx/ranges.hpp
View file @ a045fb19
......@@ -205,7 +205,7 @@ void transform(Range1&& r1, Range2&& r2, Iterator it, F f)
}
template <class Range>
auto reverse(Range& r)
auto reverse(Range&& r)
{
return range(std::make_reverse_iterator(r.end()), std::make_reverse_iterator(r.begin()));
}
......
src/include/migraphx/shape.hpp
View file @ a045fb19
......@@ -263,7 +263,7 @@ struct MIGRAPHX_EXPORT shape
/// no padding
bool packed() const;
/// Returns true is the shape has been transposed. That is the strides are not in descending
/// Returns true if the shape has been transposed. That is the strides are not in descending
/// order
bool transposed() const;
......
src/include/migraphx/shape_for_each.hpp
View file @ a045fb19
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
* Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
......@@ -37,11 +37,11 @@ inline namespace MIGRAPHX_INLINE_NS {
template <class F>
void shape_for_each(const migraphx::shape& s, F f)
{
// Ensure calls to f use const ref to vector
auto call = [&f](const std::vector<std::size_t>& i) { f(i); };
std::vector<std::size_t> indices(s.lens().size());
const auto& index_const_ref = indices;
shape ss{s.type(), s.lens()};
for(std::size_t i = 0; i < ss.elements(); i++)
size_t max = ss.elements();
for(std::size_t i = 0; i < max; i++)
{
std::transform(ss.strides().begin(),
ss.strides().end(),
......@@ -51,9 +51,13 @@ void shape_for_each(const migraphx::shape& s, F f)
assert(len > 0 and stride > 0);
return (i / stride) % len;
});
call(indices);
if constexpr(std::is_invocable<F, decltype(index_const_ref), decltype(i)>{})
f(index_const_ref, i);
else
f(index_const_ref);
}
}
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
......
src/include/migraphx/simplify_dyn_ops.hpp 0 → 100644
View file @ a045fb19
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
* THE SOFTWARE.
*/
#ifndef MIGRAPHX_GUARD_RTGLIB_SIMPLIFY_DYN_OPS_HPP
#define MIGRAPHX_GUARD_RTGLIB_SIMPLIFY_DYN_OPS_HPP
#include <string>
#include <migraphx/instruction_ref.hpp>
#include <migraphx/config.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
struct module;
/**
* Convert dynamic ops to their static version if possible.
* Should be run after the split_single_dyn_dims pass.
*/
struct MIGRAPHX_EXPORT simplify_dyn_ops
{
std::string name() const { return "simplify_dyn_ops"; }
void apply(module& m) const;
};
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif
src/include/migraphx/simplify_reshapes.hpp
View file @ a045fb19
......@@ -38,6 +38,7 @@ struct module;
*/
struct MIGRAPHX_EXPORT simplify_reshapes
{
size_t depth = 4;
std::string name() const { return "simplify_reshapes"; }
void apply(module& m) const;
};
......
src/include/migraphx/stringutils.hpp
View file @ a045fb19
......@@ -86,7 +86,7 @@ inline std::string join_strings(Strings strings, const std::string& delim)
inline std::vector<std::string> split_string(const std::string& s, char delim)
{
std::vector<std::string> elems;
std::stringstream ss(s + ' ');
std::stringstream ss(s + delim);
std::string item;
while(std::getline(ss, item, delim))
{
......@@ -149,6 +149,10 @@ interpolate_string(const std::string& input, F f, std::string start = "${", std:
result.append(it, next_start);
if(next_start == input.end())
break;
if(next_end == input.end())
{
throw std::runtime_error("Unbalanced brackets");
}
auto r = f(next_start + start.size(), next_end);
result.append(r.begin(), r.end());
it = next_end + end.size();
......
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