mirror of
https://github.com/matrix-construct/construct
synced 2024-11-04 21:08:57 +01:00
297 lines
7.2 KiB
C++
297 lines
7.2 KiB
C++
// The Construct
|
|
//
|
|
// Copyright (C) The Construct Developers, Authors & Contributors
|
|
// Copyright (C) 2016-2020 Jason Volk <jason@zemos.net>
|
|
//
|
|
// Permission to use, copy, modify, and/or distribute this software for any
|
|
// purpose with or without fee is hereby granted, provided that the above
|
|
// copyright notice and this permission notice is present in all copies. The
|
|
// full license for this software is available in the LICENSE file.
|
|
|
|
#pragma once
|
|
#define HAVE_IRCD_SIMD_TRANSFORM_H
|
|
|
|
namespace ircd::simd
|
|
{
|
|
template<class block_t>
|
|
using transform_fixed_proto = void (block_t &, block_t mask);
|
|
|
|
template<class block_t>
|
|
using transform_variable_proto = u64x2 (block_t &, block_t mask);
|
|
|
|
template<class block_t,
|
|
class lambda>
|
|
using transform_is_fixed_stride = std::is_same
|
|
<
|
|
std::invoke_result_t<lambda, block_t &, block_t>, void
|
|
>;
|
|
|
|
template<class block_t,
|
|
class lambda>
|
|
using transform_is_variable_stride = std::is_same
|
|
<
|
|
std::invoke_result_t<lambda, block_t &, block_t>, u64x2
|
|
>;
|
|
|
|
template<class block_t,
|
|
class lambda>
|
|
using transform_fixed_stride = std::enable_if
|
|
<
|
|
transform_is_fixed_stride<block_t, lambda>::value, u64x2
|
|
>;
|
|
|
|
template<class block_t,
|
|
class lambda>
|
|
using transform_variable_stride = std::enable_if
|
|
<
|
|
transform_is_variable_stride<block_t, lambda>::value, u64x2
|
|
>;
|
|
|
|
template<class block_t,
|
|
class lambda>
|
|
typename transform_fixed_stride<block_t, lambda>::type
|
|
transform(char *, const char *, const u64x2, lambda&&) noexcept;
|
|
|
|
template<class block_t,
|
|
class lambda>
|
|
typename transform_variable_stride<block_t, lambda>::type
|
|
transform(char *, const char *, const u64x2, lambda&&) noexcept;
|
|
|
|
template<class block_t,
|
|
class lambda>
|
|
pair<mutable_buffer, const_buffer>
|
|
transform(const pair<mutable_buffer, const_buffer> &, lambda&&) noexcept;
|
|
}
|
|
|
|
/// Streaming transform
|
|
///
|
|
/// Convenience wrapper using ircd::buffer. This will forward to the
|
|
/// appropriate overload. The return buffers are views on the output and input
|
|
/// buffers with a size of the respective resulting counter values. Unless
|
|
/// the closure broke the loop early the result buffers will be the same as
|
|
/// the input (contents having transformed of course).
|
|
///
|
|
template<class block_t,
|
|
class lambda>
|
|
inline std::pair<ircd::mutable_buffer, ircd::const_buffer>
|
|
ircd::simd::transform(const std::pair<mutable_buffer, const_buffer> &buf,
|
|
lambda&& closure)
|
|
noexcept
|
|
{
|
|
const auto &[output, input]
|
|
{
|
|
buf
|
|
};
|
|
|
|
const u64x2 max
|
|
{
|
|
size(output), size(input),
|
|
};
|
|
|
|
const auto res
|
|
{
|
|
transform(data(output), data(input), max, std::forward<lambda>(closure))
|
|
};
|
|
|
|
return std::pair<mutable_buffer, const_buffer>
|
|
{
|
|
{ data(output), res[0] },
|
|
{ data(input), res[1] },
|
|
};
|
|
}
|
|
|
|
/// Streaming transform
|
|
///
|
|
/// This template performs the loop boiler-plate for the developer who can
|
|
/// simply supply a conforming closure. Characteristics:
|
|
///
|
|
/// * byte-aligned (unaligned): the input and output buffers do not have to
|
|
/// be aligned and can be any size.
|
|
///
|
|
/// * full-duplex: the operation involves both input and output and there are
|
|
/// separate pointers for progress across the input and output buffers which
|
|
/// are incremented independently.
|
|
///
|
|
/// * variable-stride: progress for each iteration of the loop across the input
|
|
/// and output buffers is not fixed; the transform function may advance either
|
|
/// pointer zero to sizeof(block_t) bytes each iteration. Due to these
|
|
/// characteristics, unaligned bytes may be redundantly loaded or stored and
|
|
/// non-temporal features are not used to optimize the operation.
|
|
///
|
|
/// u64x2 counter lanes = { output_length, input_length }; The argument `max`
|
|
/// gives the buffer size in that format. The return value is the consumed
|
|
/// bytes (final counter value) in that format.
|
|
///
|
|
template<class block_t,
|
|
class lambda>
|
|
inline typename ircd::simd::transform_variable_stride<block_t, lambda>::type
|
|
ircd::simd::transform(char *const __restrict__ out,
|
|
const char *const __restrict__ in,
|
|
const u64x2 max,
|
|
lambda&& closure)
|
|
noexcept
|
|
{
|
|
using block_t_u = unaligned<block_t>;
|
|
|
|
u64x2 count
|
|
{
|
|
0, // output pos
|
|
0, // input pos
|
|
};
|
|
|
|
// primary broadband loop
|
|
while(count[1] + sizeof(block_t) <= max[1] && count[0] + sizeof(block_t) <= max[0])
|
|
{
|
|
static const auto mask
|
|
{
|
|
~block_t{0}
|
|
};
|
|
|
|
const auto di
|
|
{
|
|
reinterpret_cast<block_t_u *>(out + count[0])
|
|
};
|
|
|
|
const auto si
|
|
{
|
|
reinterpret_cast<const block_t_u *>(in + count[1])
|
|
};
|
|
|
|
block_t block
|
|
(
|
|
*si
|
|
);
|
|
|
|
const auto consume
|
|
{
|
|
closure(block, mask)
|
|
};
|
|
|
|
count += consume;
|
|
*di = block;
|
|
}
|
|
|
|
// trailing narrowband loop
|
|
while(count[1] < max[1])
|
|
{
|
|
block_t block {0}, mask {0};
|
|
for(size_t i(0); count[1] + i < max[1] && i < sizeof(block_t); ++i)
|
|
{
|
|
block[i] = in[count[1] + i];
|
|
mask[i] = 0xff;
|
|
}
|
|
|
|
const auto consume
|
|
{
|
|
closure(block, mask)
|
|
};
|
|
|
|
for(size_t i(0); i < consume[0] && count[0] + i < max[0]; ++i)
|
|
out[count[0] + i] = block[i];
|
|
|
|
count += consume;
|
|
}
|
|
|
|
return u64x2
|
|
{
|
|
std::min(count[0], max[0]),
|
|
std::min(count[1], max[1]),
|
|
};
|
|
}
|
|
|
|
/// Streaming transform
|
|
///
|
|
/// This template performs the loop boiler-plate for the developer who can
|
|
/// simply supply a conforming closure. Characteristics:
|
|
///
|
|
/// * byte-aligned (unaligned): the input and output buffers do not have to
|
|
/// be aligned and can be any size.
|
|
///
|
|
/// * full-duplex: the operation involves both input and output and there are
|
|
/// separate pointers for progress across the input and output buffers which
|
|
/// are incremented independently.
|
|
///
|
|
/// * fixed-stride: progress for each iteration of the loop across the input
|
|
/// and output buffers is fixed.
|
|
///
|
|
/// u64x2 counter lanes = { output_length, input_length }; The argument `max`
|
|
/// gives the buffer size in that format. The return value is the consumed
|
|
/// bytes (final counter value) in that format.
|
|
///
|
|
template<class block_t,
|
|
class lambda>
|
|
inline typename ircd::simd::transform_fixed_stride<block_t, lambda>::type
|
|
ircd::simd::transform(char *const __restrict__ out,
|
|
const char *const __restrict__ in,
|
|
const u64x2 max,
|
|
lambda&& closure)
|
|
noexcept
|
|
{
|
|
using block_t_u = unaligned<block_t>;
|
|
|
|
u64x2 count
|
|
{
|
|
0, // output pos
|
|
0, // input pos
|
|
};
|
|
|
|
// primary broadband loop
|
|
while(count[1] + sizeof(block_t) <= max[1] && count[0] + sizeof(block_t) <= max[0])
|
|
{
|
|
static const u64x2 consume
|
|
{
|
|
sizeof(block_t),
|
|
sizeof(block_t),
|
|
};
|
|
|
|
static const auto mask
|
|
{
|
|
~block_t{0}
|
|
};
|
|
|
|
const auto di
|
|
{
|
|
reinterpret_cast<block_t_u *>(out + count[0])
|
|
};
|
|
|
|
const auto si
|
|
{
|
|
reinterpret_cast<const block_t_u *>(in + count[1])
|
|
};
|
|
|
|
block_t block
|
|
(
|
|
*si
|
|
);
|
|
|
|
closure(block, mask);
|
|
count += consume;
|
|
*di = block;
|
|
}
|
|
|
|
// trailing narrowband loop
|
|
assert(count[1] + sizeof(block_t) > max[1]);
|
|
if(likely(count[1] < max[1]))
|
|
{
|
|
u64 i[2] {0};
|
|
block_t block {0}, mask {0};
|
|
for(; count[1] + i[1] < max[1]; ++i[1])
|
|
{
|
|
block[i[1]] = in[count[1] + i[1]];
|
|
mask[i[1]] = 0xff;
|
|
}
|
|
|
|
closure(block, mask);
|
|
for(; i[0] < i[1] && count[0] + i[0] < max[0]; ++i[0])
|
|
out[count[0] + i[0]] = block[i[0]];
|
|
|
|
count += u64x2
|
|
{
|
|
i[0], i[1]
|
|
};
|
|
}
|
|
|
|
assert(count[0] == max[0]);
|
|
assert(count[1] == max[1]);
|
|
return count;
|
|
}
|