66 lines
2.5 KiB
C++
66 lines
2.5 KiB
C++
|
// nisps/ml/warm_start.hpp — copy overlapping weights between two MLPs of
|
|||
|
|
// (possibly) different shapes.
|
|||
|
|
//
|
|||
|
|
// Used by the runtime-reshape path (one-core-engine-refactor P2): reshape =
|
|||
|
|
// construct a NEW instance at the new dimensions, then warm-start it by
|
|||
|
|
// copying every weight/bias whose (layer, node, input) coordinate exists in
|
|||
|
|
// BOTH shapes. Weights outside the overlap keep the destination's fresh
|
|||
|
|
// initialisation. Deterministic, allocation-free, works across storage
|
|||
|
|
// policies (fixed→dynamic, dynamic→dynamic, fixed→fixed).
|
|||
|
|
//
|
|||
|
|
// Row-major layout per layer: w[node * fan_in + j]. The overlap is the
|
|||
|
|
// top-left submatrix min(fan_out) × min(fan_in) plus the bias prefix
|
|||
|
|
// min(fan_out).
|
|||
|
|
|
|||
|
|
#pragma once
|
|||
|
|
|
|||
|
|
#include <cstddef>
|
|||
|
|
#include <span>
|
|||
|
|
|
|||
|
|
#include "../core/perf.hpp"
|
|||
|
|
|
|||
|
|
namespace nisps::ml {
|
|||
|
|
|
|||
|
|
namespace detail {
|
|||
|
|
|
|||
|
|
template <std::size_t L, typename DstMLP, typename SrcMLP>
|
|||
|
|
NISPS_FORCE_INLINE void warm_start_copy_layer(DstMLP& dst, const SrcMLP& src) noexcept {
|
|||
|
|
const std::size_t src_in = src.template fan_in_l<L>();
|
|||
|
|
const std::size_t src_out = src.template fan_out_l<L>();
|
|||
|
|
const std::size_t dst_in = dst.template fan_in_l<L>();
|
|||
|
|
const std::size_t dst_out = dst.template fan_out_l<L>();
|
|||
|
|
const std::size_t n_in = (src_in < dst_in) ? src_in : dst_in;
|
|||
|
|
const std::size_t n_out = (src_out < dst_out) ? src_out : dst_out;
|
|||
|
|
|
|||
|
|
std::span<const float> sw = src.template weights_l<L>();
|
|||
|
|
std::span<float> dw = dst.template weights_l<L>();
|
|||
|
|
for (std::size_t node = 0; node < n_out; ++node) {
|
|||
|
|
const std::size_t src_row = node * src_in;
|
|||
|
|
const std::size_t dst_row = node * dst_in;
|
|||
|
|
for (std::size_t j = 0; j < n_in; ++j) {
|
|||
|
|
dw[dst_row + j] = sw[src_row + j];
|
|||
|
|
}
|
|||
|
|
}
|
|||
|
|
|
|||
|
|
std::span<const float> sb = src.template biases_l<L>();
|
|||
|
|
std::span<float> db = dst.template biases_l<L>();
|
|||
|
|
for (std::size_t node = 0; node < n_out; ++node) {
|
|||
|
|
db[node] = sb[node];
|
|||
|
|
}
|
|||
|
|
}
|
|||
|
|
|
|||
|
|
} // namespace detail
|
|||
|
|
|
|||
|
|
// Copy the overlapping region of every layer from `src` into `dst`. Both
|
|||
|
|
// must expose the MLP storage surface (fan_in_l/fan_out_l/weights_l/
|
|||
|
|
// biases_l) — i.e. any MLPCore instantiation.
|
|||
|
|
template <typename DstMLP, typename SrcMLP>
|
|||
|
|
inline void warm_start_copy(DstMLP& dst, const SrcMLP& src) noexcept {
|
|||
|
|
detail::warm_start_copy_layer<0u>(dst, src);
|
|||
|
|
detail::warm_start_copy_layer<1u>(dst, src);
|
|||
|
|
detail::warm_start_copy_layer<2u>(dst, src);
|
|||
|
|
detail::warm_start_copy_layer<3u>(dst, src);
|
|||
|
|
}
|
|||
|
|
|
|||
|
|
} // namespace nisps::ml
|