happyz
/
gemma.cpp
mirror of https://github.com/google/gemma.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
gemma.cpp/compression/nuq-inl.h

991 lines
42 KiB
C++
Raw Blame History

// Copyright 2023 Google LLC
// SPDX-License-Identifier: Apache-2.0
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
// Normal include guard.
#ifndef THIRD_PARTY_GEMMA_CPP_COMPRESSION_NUQ_INL_H_
#define THIRD_PARTY_GEMMA_CPP_COMPRESSION_NUQ_INL_H_
#include <stddef.h>
#include <stdint.h>
#include <stdio.h>
#include <cstdio>
#include "compression/shared.h"
#include "util/basics.h"
#include "hwy/base.h"
#endif // THIRD_PARTY_GEMMA_CPP_COMPRESSION_NUQ_INL_H_
// Actual per-target include guard.
#if defined(THIRD_PARTY_GEMMA_CPP_COMPRESSION_NUQ_INL_TOGGLE) == \
defined(HWY_TARGET_TOGGLE)
#ifdef THIRD_PARTY_GEMMA_CPP_COMPRESSION_NUQ_INL_TOGGLE
#undef THIRD_PARTY_GEMMA_CPP_COMPRESSION_NUQ_INL_TOGGLE
#else
#define THIRD_PARTY_GEMMA_CPP_COMPRESSION_NUQ_INL_TOGGLE
#endif
#include "hwy/highway.h"
// After highway.h
#include "compression/sfp-inl.h"
#include "hwy/contrib/sort/vqsort-inl.h"
HWY_BEFORE_NAMESPACE();
namespace gcpp {
namespace HWY_NAMESPACE {
namespace hn = hwy::HWY_NAMESPACE;
// For internal use by NuqCodec.
class NuqClustering {
static constexpr size_t kGroupSize = NuqStream::kGroupSize;
// To go from sorted order back to the original order in O(1), we store the
// original index in the lower bits of the float32 mantissa, which means they
// are sorted alongside the value.
struct FloatPayload {
// Resets payload to zero; useful for displaying the actual value.
static HWY_INLINE float Clear(float f) {
const uint32_t binary32 = hwy::BitCastScalar<uint32_t>(f);
return hwy::BitCastScalar<float>(binary32 &
~static_cast<uint32_t>(kGroupSize - 1));
}
// Sets payload to `bits`.
static HWY_INLINE float Set(float f, size_t bits) {
HWY_DASSERT(bits < kGroupSize);
const uint32_t binary32 = hwy::BitCastScalar<uint32_t>(Clear(f));
return hwy::BitCastScalar<float>(static_cast<uint32_t>(binary32 | bits));
}
// Obtains the payload (index) previously set by `Set`.
static HWY_INLINE size_t Get(float f) {
return hwy::BitCastScalar<uint32_t>(f) &
static_cast<uint32_t>(kGroupSize - 1);
}
};
// Cumulative sums for O(1) mean and interval sums.
class ClusterCost {
// Ensures it is safe to load a vector from the last element.
static constexpr size_t kMaxLanes = hn::MaxLanes(hn::ScalableTag<float>());
// Initialization value for table elements where `valid` is false.
static constexpr float kSentinel = -1.0f;
public:
explicit ClusterCost(const float* HWY_RESTRICT sorted) {
double cumsum = 0.0;
double cumsum2 = 0.0;
dcumsum_[0] = 0.0;
cumsum_[0] = cumsum2_[0] = 0.0f;
for (size_t i = 0; i < kGroupSize; ++i) {
const float x = FloatPayload::Clear(sorted[i]);
cumsum += x;
cumsum2 += static_cast<double>(x) * x;
dcumsum_[1 + i] = cumsum;
cumsum_[1 + i] = static_cast<float>(cumsum);
cumsum2_[1 + i] = static_cast<float>(cumsum2);
}
const hn::ScalableTag<float> df;
using VF = hn::Vec<decltype(df)>;
const VF k1 = hn::Set(df, 1.0f);
const size_t N = hn::Lanes(df);
HWY_DASSERT(kGroupSize % N == 0);
// Precomputed length and reciprocal.
for (size_t len = 0; len < kGroupSize; len += N) {
const VF vlen = hn::Iota(df, static_cast<int32_t>(len));
hn::StoreU(vlen, df, len_ + kMaxLanes + len);
hn::StoreU(hn::Div(k1, vlen), df, inv_len_ + kMaxLanes + len);
}
// len = kGroupSize is legitimate, e.g., for all-equal weights.
len_[kMaxLanes + kGroupSize] = static_cast<float>(kGroupSize);
inv_len_[kMaxLanes + kGroupSize] = 1.0f / static_cast<float>(kGroupSize);
// len = 0 can happen, but valid is false for that lane.
len_[kMaxLanes + 0] = kSentinel;
inv_len_[kMaxLanes + 0] = kSentinel;
// Ensure it is safe to load a vector from the last element.
for (size_t i = 0; i < kMaxLanes; ++i) {
constexpr size_t kEnd = kGroupSize + 1;
cumsum_[kEnd + i] = cumsum_[kGroupSize];
cumsum2_[kEnd + i] = cumsum2_[kGroupSize];
len_[kMaxLanes + kEnd + i] = len_[kMaxLanes + kGroupSize];
inv_len_[kMaxLanes + kEnd + i] = inv_len_[kMaxLanes + kGroupSize];
}
// For inv_len_ we also prepend MaxLanes in case first > last.
for (size_t i = 0; i < kMaxLanes; ++i) {
len_[i] = kSentinel;
inv_len_[i] = kSentinel;
}
}
// Returns cost (L2 norm) for a single cluster, used for backtracking.
double SumOfSorted(size_t first, size_t last) const {
HWY_DASSERT(first < kGroupSize);
HWY_DASSERT(last < kGroupSize);
return dcumsum_[last + 1] - dcumsum_[first];
}
// Returns vector of costs of clustering first..last + i with their means.
// O(1) thanks to cumulative sums, which works for Lp-norms with p > 1; we
// choose p=2 for simplicity (fewer terms). Caller ignores lanes where
// `!valid[i]`, i.e. `first > last + i`.
template <class DF, class MF = hn::Mask<DF>, class VF = hn::Vec<DF>>
VF SumCosts(DF df, size_t first, size_t last, MF valid) const {
HWY_DASSERT(first < kGroupSize);
HWY_DASSERT(last < kGroupSize);
VF inv_len;
const VF vlen = Lengths(df, first, last, valid, inv_len);
const VF u_lo = hn::Set(df, cumsum_[first]);
const VF u_lo2 = hn::Set(df, cumsum2_[first]);
const VF hi = hn::LoadU(df, cumsum_ + last + 1);
const VF hi2 = hn::LoadU(df, cumsum2_ + last + 1);
const VF sum = hn::Sub(hi, u_lo);
const VF sum2 = hn::Sub(hi2, u_lo2);
// Sum of L2 over i in [first, last] = (x[i] - mu)^2. `sum` and `sum2` are
// the cumulative sums of x and x^2, so expand to `sum x^2 + sum x * -2 *
// mu + sum mu^2`. The last term is the sum of a constant, hence `mu^2 *
// len`. Thus we have: `sum2 + mu * (-2 * sum + mu * len)`. We avoid a
// (-)2 constant by adding.
const VF mu = hn::Mul(sum, inv_len); // mean := sum[i] / len[i]
const VF two_sum = hn::Add(sum, sum);
const VF l2 = hn::MulAdd(mu, hn::MulSub(mu, vlen, two_sum), sum2);
// mu can have some roundoff error. To avoid multiple redundant clusters,
// clamp to zero.
return hn::ZeroIfNegative(l2);
}
private:
// Returns precomputed lengths of [first, last + i] and their reciprocals.
template <class DF, class VF = hn::Vec<DF>, class MF = hn::Mask<DF>>
VF Lengths(DF df, size_t first, size_t last, MF valid, VF& inv_len) const {
const int len = static_cast<int>(last) - static_cast<int>(first) + 1;
HWY_DASSERT(kMaxLanes + len >= 0);
HWY_DASSERT(len <= static_cast<int>(kGroupSize));
// last + i are contiguous, hence single loads instead of gather.
const VF vlen = hn::LoadU(df, len_ + kMaxLanes + len);
inv_len = hn::LoadU(df, inv_len_ + kMaxLanes + len);
if constexpr (HWY_IS_DEBUG_BUILD) {
// Sanity check: no valid lanes are sentinels, all invalid are.
const VF sentinel = hn::Set(df, kSentinel);
const MF bad = hn::Eq(vlen, sentinel);
const MF inv_bad = hn::Eq(inv_len, sentinel);
HWY_DASSERT(hn::AllFalse(df, hn::And(valid, bad)));
HWY_DASSERT(hn::AllFalse(df, hn::And(valid, inv_bad)));
HWY_DASSERT(hn::AllTrue(df, hn::Or(valid, bad)));
HWY_DASSERT(hn::AllTrue(df, hn::Or(valid, inv_bad)));
}
return vlen;
}
// Float has enough precision for our relatively small kGroupSize (256).
// Element i = sums of [0..i-1].
float cumsum_[kGroupSize + 1 + kMaxLanes];
float cumsum2_[kGroupSize + 1 + kMaxLanes];
float len_[kMaxLanes + kGroupSize + 1 + kMaxLanes]; // = vlen[i]
float inv_len_[kMaxLanes + kGroupSize + 1 + kMaxLanes]; // = 1 / vlen[i]
double dcumsum_[kGroupSize + 1]; // for SumOfSorted
};
// Dynamic programming step: returns costs of clustering 0..last+i, where the
// rightmost clusters start at `first`. Called for each `idx_cluster`,
// `first`, and `last`; vectorized across `last`. `first` may be greater than
// `last`. `valid[i]` is `first <= last + i`.
template <class DF, class VF = hn::Vec<DF>, class MF = hn::Mask<DF>>
static HWY_INLINE VF
ClusterDynProg(DF df, const NuqStream::AlignedMatrix<float>& costs,
const ClusterCost& cc, const size_t idx_cluster,
const size_t first, const size_t last, const MF valid) {
HWY_DASSERT(idx_cluster != 0);
HWY_DASSERT(0 != first && first < kGroupSize);
HWY_DASSERT(last < kGroupSize);
HWY_DASSERT(last % hn::Lanes(df) == 0); // Called in steps of N
// Cost of clustering 0..first-1 with one fewer cluster than now.
const VF prev = hn::Set(df, costs(idx_cluster - 1, first - 1));
// Eq2: add to that the cost of another cluster from first..last.
return hn::Add(prev, cc.SumCosts(df, first, last, valid));
}
public:
// Clusters `num <= kGroupSize` values in `x`, which need not be sorted
// already nor aligned, by choosing and filling `centers` (size `kClusters`,
// ascending order, not necessarily equal to one of the `x`). Fills `indices`
// with the index of the cluster to which each `x` belongs (16-bit for
// bit-packing). `buf` is per-thread.
//
// Returns the number of unused clusters, i.e., the starting index within
// `centers`; prior centers are zero-initialized.
//
// O(kClusters * kGroupSize * kGroupSize), but the constant factors are so low
// that this is about 5 times as fast as the O(kClusters * kGroupSize) SMAWK
// as implemented in FAISS, for our kGroupSize of 256.
template <class DF>
static HWY_NOINLINE size_t ClusterExactL2(DF df, const float* HWY_RESTRICT x,
size_t num,
NuqStream::ClusterBuf& buf,
float* HWY_RESTRICT centers,
uint16_t* HWY_RESTRICT indices) {
HWY_DASSERT(num <= kGroupSize);
const hn::RebindToSigned<decltype(df)> di;
using VF = hn::Vec<decltype(df)>;
using MF = hn::Mask<decltype(df)>;
using VI = hn::Vec<decltype(di)>;
const VI k1 = hn::Set(di, 1);
const size_t N = hn::Lanes(df);
HWY_DASSERT(kGroupSize % N == 0);
HWY_ALIGN float sorted_and_i[kGroupSize];
for (size_t i = 0; i < num; ++i) {
sorted_and_i[i] = FloatPayload::Set(x[i], i);
}
if (num != kGroupSize) {
// Initialize the rest of the group. Use an existing value so we do not
// waste a cluster on a sentinel value. Arbitrarily choose the largest.
float max = -1E38f;
for (size_t i = 0; i < num; ++i) {
max = HWY_MAX(max, x[i]);
}
for (size_t i = num; i < kGroupSize; ++i) {
sorted_and_i[i] = FloatPayload::Set(max, i);
}
}
hn::VQSortStatic(sorted_and_i, kGroupSize, hwy::SortAscending());
ClusterCost cc(sorted_and_i); // ignores payload bits.
// Reference: https://arxiv.org/abs/1701.07204
// costs[k-1][m] is the lowest cost of clustering x1..m into k clusters.
NuqStream::AlignedMatrix<float>& costs = buf.costs;
// argmin[k][m] is the starting index within sorted_and_i[] of the k-th
// cluster.
NuqStream::AlignedMatrix<int32_t>& argmin = buf.argmin;
// Fill first row of `costs` and `argmin`: single cluster, iterate over all
// `last`.
{
const size_t cluster_idx = 0;
const size_t first = 0;
const VI vfirst = hn::Set(di, static_cast<int32_t>(first));
const MF all_valid = hn::FirstN(df, N); // first <= last is always true
for (size_t last = 0; last < kGroupSize; last += N) {
const VF vcosts = cc.SumCosts(df, first, last, all_valid);
hn::Store(vcosts, df, &costs(cluster_idx, last));
hn::Store(vfirst, di, &argmin(cluster_idx, last));
}
}
constexpr size_t kClusters = NuqStream::kClusters;
for (size_t cluster_idx = 1; cluster_idx < kClusters; ++cluster_idx) {
// For vectors of `last + i` with `i < N`:
for (size_t last = 0; last < kGroupSize; last += N) {
const VI vlast = hn::Iota(di, static_cast<int32_t>(last));
const VF prev_cost = hn::LoadU(df, &costs(cluster_idx - 1, last));
VF min = prev_cost;
VI arg = hn::LoadU(di, &argmin(cluster_idx - 1, last));
// For each `first` (j), which is the start of the rightmost of at least
// two clusters, hence never zero. `first` also continues past `last`
// because the last `vlast` lane is `last + N - 1`.
for (size_t first = 1; first < last + N; ++first) {
const VI vfirst = hn::Set(di, static_cast<int32_t>(first));
const MF valid = hn::RebindMask(df, hn::Le(vfirst, vlast));
const VF c =
ClusterDynProg(df, costs, cc, cluster_idx, first, last, valid);
// Retain the min cost and the `first` that caused it.
const MF less = hn::And(valid, hn::Lt(c, min));
min = hn::IfThenElse(less, c, min);
arg = hn::IfThenElse(RebindMask(di, less), vfirst, arg);
}
HWY_DASSERT(hn::AllTrue(df, hn::Le(min, prev_cost)));
hn::Store(min, df, &costs(cluster_idx, last));
hn::Store(arg, di, &argmin(cluster_idx, last));
}
}
// Backtrack to find centers. Clusters are [argmin(k, last), last].
size_t last = kGroupSize - 1;
size_t unused_clusters = 0;
for (size_t k = kClusters - 1; k < kClusters; --k) {
const size_t start = static_cast<size_t>(argmin(k, last));
// Center = mean, O(1) thanks to cumulative sums.
const double sum = cc.SumOfSorted(start, last);
const int size = static_cast<int>(last) - static_cast<int>(start) + 1;
HWY_DASSERT(0 < size && size <= static_cast<int>(kGroupSize));
centers[k] = static_cast<float>(sum / size);
// We know the range inside sorted_and_i[]; translate to original indices,
// which are stored inside each of the sorted_and_i mantissas.
for (size_t i = start; i <= last; ++i) {
const size_t idx_x = FloatPayload::Get(sorted_and_i[i]);
HWY_DASSERT(idx_x < kGroupSize);
indices[idx_x] = static_cast<uint16_t>(k);
}
// Not using all clusters. Avoid out of bounds accesses by stopping early.
if (start == 0) {
unused_clusters = k;
for (size_t cluster = 0; cluster < unused_clusters; ++cluster) {
centers[cluster] = 0.0f;
}
break;
}
last = start - 1;
HWY_DASSERT(last < kGroupSize);
}
if (HWY_IS_DEBUG_BUILD) {
// If centers are not in ascending order, print them.
for (size_t i = unused_clusters + 1; i < kClusters; ++i) {
if (centers[i] < centers[i - 1]) {
for (size_t i = 0; i < kClusters; ++i) {
fprintf(stderr, "%2zu: %.8f\n", i, centers[i]);
}
for (size_t i = 0; i < kGroupSize; ++i) {
fprintf(stderr, "%3zu: %.8f\n", i,
FloatPayload::Clear(sorted_and_i[i]));
}
for (size_t i = 0; i < num; ++i) {
fprintf(stderr, "%3zu: %.8f\n", i, x[i]);
}
HWY_ABORT("Centers not in ascending order at %zu; unused %zu\n", i,
unused_clusters);
}
}
}
MaybeCheckInitialized(centers, kClusters * sizeof(centers[0]));
return unused_clusters;
}
}; // NuqClustering
// Half-vector of u8 from u16/bf16.
template <class D16>
using D8HFromD16 = hn::Half<hn::Repartition<uint8_t, D16>>;
// Bit-packing 4-bit values is trivial if we have 2 or 4 independent vectors:
// simply shift+OR them together into a full vector of 8 or 16-bit lanes.
// However, the order then depends on the vector length, which is unacceptable
// because we may store the encoding to disk and decode on another CPU.
//
// The dependency on vector length could be removed by introducing fixed-size
// packets and loading the next vector from a fixed offset known to be at
// least the vector length. However, this may require packets that are larger
// than the seek granularity of the application (e.g. matrix rows).
//
// We instead choose a continuous stream layout, which seems to entail the
// nibbles being stored and decoded in-order. This involves nontrivial shuffle
// operations which benefit from special-casing for target and vector length.
class NibbleCodec {
public:
// Returns a byte vector whose nibbles are the lanes of four u16 vectors, in
// the same order.
template <class D16, class V16 = hn::Vec<D16>,
class D8 = hn::Repartition<uint8_t, D16>, class V8 = hn::Vec<D8>>
static HWY_INLINE V8 OrderedPackU16(D16 d16, V16 in0, V16 in1, V16 in2,
V16 in3) {
const D8 d8;
const hn::Repartition<uint32_t, D16> d32;
const hn::Repartition<uint64_t, D16> d64;
// Pairwise compaction of a single vector so nibbles are packed in-order.
// v16 lanes hold a 4-bit value; OR together adjacent pairs into the lower
// byte of *even* u16.
const auto combine_u16_pair_to_8 = [d16, d32](V16 v16) HWY_ATTR {
return hn::Xor(
v16, hn::BitCast(d16, hn::ShiftRight<12>(hn::BitCast(d32, v16))));
};
const V16 u8_0 = combine_u16_pair_to_8(in0);
const V16 u8_1 = combine_u16_pair_to_8(in1);
const V16 u8_2 = combine_u16_pair_to_8(in2);
const V16 u8_3 = combine_u16_pair_to_8(in3);
if constexpr (HWY_TARGET <= HWY_AVX3_DL || !HWY_ARCH_X86) {
// 8-bit ConcatEven is efficient. Let digits denote eight u8 lanes
// of u8_1/0: ?d?3 ?c?2 / ?b?1 ?a?0. 8-bit ConcatEven = d3c2 b1a0, and
// again with the second x2_1 gives 7654 3210.
const V8 x2_0 = hn::ConcatEven(d8, BitCast(d8, u8_1), BitCast(d8, u8_0));
const V8 x2_1 = hn::ConcatEven(d8, BitCast(d8, u8_3), BitCast(d8, u8_2));
return hn::ConcatEven(d8, x2_1, x2_0);
} else {
// To avoid expensive 8-bit ConcatEven, compact pairs of u32 into the
// lower 16 bits in each u64, with other bits undefined.
const auto combine_u32_pair_to_16 = [d16, d64](V16 v16) HWY_ATTR {
return hn::Xor(
v16, hn::BitCast(d16, hn::ShiftRight<24>(hn::BitCast(d64, v16))));
};
const V16 u16_0 = combine_u32_pair_to_16(u8_0);
const V16 u16_1 = combine_u32_pair_to_16(u8_1);
const V16 u16_2 = combine_u32_pair_to_16(u8_2);
const V16 u16_3 = combine_u32_pair_to_16(u8_3);
// In-order compaction of four vectors into one, keeping only the low
// u16 of every u64. This is the same as above but with 16-bit Concat.
const V16 x2_0 = hn::ConcatEven(d16, u16_1, u16_0);
const V16 x2_1 = hn::ConcatEven(d16, u16_3, u16_2);
return hn::BitCast(d8, hn::ConcatEven(d16, x2_1, x2_0));
}
}
// Unpacks nibbles from the `kHalf` (0 or 1) half of a half-vector of bytes.
// Thus we use a quarter of a vector of bytes and expand nibbles 4x into u16,
// which fills a whole vector. Its first lane comes from the low nibble of the
// first byte, the second from its high nibble, then the next low nibble, etc.
template <size_t kHalf, class D16, class V16 = hn::Vec<D16>,
class D8H = D8HFromD16<D16>, class V8H = hn::Vec<D8H>>
static HWY_INLINE V16 OrderedUnpackU16(D16 d16, const V8H packed) {
const hn::Twice<D8H> d8; // full vector
using V8 = hn::Vec<decltype(d8)>;
// Replicate each byte 4x, so that its two nibbles propagate to both u16
// lanes that they will initialize.
const V8 rep4 = Replicate4x<kHalf>(d8, hn::ResizeBitCast(d8, packed));
const V16 mask4 = hn::Set(d16, 0xF);
const V16 u16 = BitCast(d16, rep4);
// In-order unpack. Right-shift odd u16 by 4. Example with two packed
// bytes, one digit representing a nibble:
// 32 32 32 32 | 10 10 10 10 u16
// z3 23 32 32 | z1 01 10 10 OddEven+ShiftRight
// zz z3 zz z2 | zz z1 zz z0 And (unpacked result)
return hn::And(mask4, hn::OddEven(hn::ShiftRight<4>(u16), u16));
}
private:
// Returns `bytes[0 + kHalf * N/2]` in lanes 0..3, `bytes[1 + kHalf * N/2]` in
// lanes 4..7, etc. We fuse `kHalf` into the tables, which avoids the caller
// having to pass in `UpperHalf(bytes)`.
template <size_t kHalf, class D8, class V8 = hn::Vec<D8>>
static HWY_INLINE V8 Replicate4x(D8 d8, V8 bytes) {
static_assert(kHalf <= 1);
const size_t N = hn::Lanes(d8);
constexpr size_t kMaxN = hn::MaxLanes(d8);
// For kHalf=1 and 512-bit vectors, kAdd would be 16, which is out of
// bounds for TableLookupBytes. We instead BroadcastBlock<1> there.
constexpr uint8_t kAdd = kMaxN < 64 ? kHalf * kMaxN / 4 : 0;
// The only performance-portable op to replicate bytes is TableLookupBytes,
// but this only works if vectors are 128-bit or we first BroadcastBlock,
// which only works for <= 512-bit vectors. For scalable vectors, we
// instead synthesize this table via Iota+ShiftRight.
alignas(64) static constexpr uint8_t kRep4[64] = {
HWY_REP4(kAdd + 0), HWY_REP4(kAdd + 1), HWY_REP4(kAdd + 2),
HWY_REP4(kAdd + 3), HWY_REP4(kAdd + 4), HWY_REP4(kAdd + 5),
HWY_REP4(kAdd + 6), HWY_REP4(kAdd + 7), HWY_REP4(kAdd + 8),
HWY_REP4(kAdd + 9), HWY_REP4(kAdd + 10), HWY_REP4(kAdd + 11),
HWY_REP4(kAdd + 12), HWY_REP4(kAdd + 13), HWY_REP4(kAdd + 14),
HWY_REP4(kAdd + 15)};
if constexpr (HWY_HAVE_SCALABLE) {
// Replicate bytes 4x: lowest 4 = 0, next 4 = 1 etc. This works for up to
// 1024-bit vectors: Iota is [128, 256), and [32, 64) after shifting.
// For larger vectors, this would overflow and we should instead add kAdd.
HWY_DASSERT(N <= 128);
const V8 iota = hn::Iota(d8, static_cast<uint8_t>(kHalf * N));
const V8 idx = hn::ShiftRight<2>(iota);
return hn::TableLookupLanes(bytes, hn::IndicesFromVec(d8, idx));
} else if constexpr (kMaxN <= 16) { // <= 128-bit
// No BroadcastBlock, we anyway only have one block.
return hn::TableLookupBytes(bytes, hn::Load(d8, kRep4));
} else if constexpr (HWY_TARGET <= HWY_AVX3_DL || !HWY_ARCH_X86) {
// No BroadcastBlock, can directly permute across blocks.
return hn::TableLookupLanes(bytes, hn::SetTableIndices(d8, kRep4));
} else { // 256..512-bit, no efficient TableLookupLanes
static_assert(kMaxN <= 64); // Else BroadcastBlock does not work.
// See kAdd comment above.
constexpr size_t kBlock = (kMaxN == 64 && kHalf == 1) ? 1 : 0;
bytes = hn::BroadcastBlock<kBlock>(bytes);
return hn::TableLookupBytes(bytes, hn::Load(d8, kRep4));
}
}
};
// Encode/decode functions.
class NuqCodec {
static constexpr size_t kClusters = NuqStream::kClusters;
static constexpr size_t kGroupSize = NuqStream::kGroupSize;
// 256-bit vectors can hold 16 bf16, otherwise we require 2x128-bit.
template <class DU>
static constexpr size_t NumTables(DU du) {
return (!HWY_HAVE_SCALABLE && du.MaxBytes() >= 32) ? 1 : 2;
}
// Offset (in bytes) of a group's table for packed_ofs (in elements) within a
// set of groups.
static constexpr size_t TableByteOffset(size_t packed_ofs) {
const size_t kBytesPerGroup =
(kClusters * sizeof(SfpStream)) + kGroupSize / 2;
return (packed_ofs / kGroupSize) * kBytesPerGroup;
}
// Unpacks `centers` from SFP into bf16 and loads them into one or two vectors
// for use by [Two]TableLookups. Returns as u16 because TableLookupLanes might
// not be available for bf16.
template <class DU, HWY_IF_U16_D(DU)>
static HWY_INLINE hn::Vec<DU> LoadTable(DU du, const uint8_t* centers,
hn::Vec<DU>* HWY_RESTRICT tbl1) {
// Cap to the table size (kClusters) for decoding SFP - sufficient, and may
// be faster than a large vector.
const hn::CappedTag<BF16, kClusters> d_table;
// We ResizeCast tables to DU: if DU is bigger, table lookups will only
// access lanes < kClusters. If DU is smaller (128-bit), we have 2 tables.
HWY_DASSERT(hn::Lanes(du) >= hn::Lanes(d_table) || NumTables(du) == 2);
HWY_ALIGN BF16 table[kClusters];
SfpCodec::DecompressAndZeroPad(
d_table,
MakeSpan(reinterpret_cast<const SfpStream*>(centers), kClusters), 0,
table, kClusters);
// If we assume >= 128-bit vectors, we can use [Two]TableLookupLanes
// instead of TableLookupBytes, which requires extra interleaving of lo/hi.
HWY_DASSERT(hn::Lanes(du) >= 8);
if constexpr (NumTables(du) == 2) {
// Reduce cap for second half to avoid loading past the end of the table.
const hn::CappedTag<BF16, kClusters / 2> d_table2;
*tbl1 = hn::ResizeBitCast(du, hn::LoadU(d_table2, table + kClusters / 2));
}
return hn::ResizeBitCast(du, hn::Load(d_table, table));
}
// Unpacks a half-vector of nibbles into two vectors of u16 indices and sets
// c0/c1 to the corresponding bf16 (stored in u16) centers from tbl0/tbl1.
template <class DU, class VU = hn::Vec<DU>, class D8H = D8HFromD16<DU>,
class V8H = hn::Vec<D8H>>
static HWY_INLINE void TableLookups(DU du, VU tbl0, VU tbl1, const V8H packed,
VU& c0, VU& c1) {
const VU idx0 = NibbleCodec::OrderedUnpackU16<0>(du, packed);
const VU idx1 = NibbleCodec::OrderedUnpackU16<1>(du, packed);
const auto indices0 = hn::IndicesFromVec(du, idx0);
const auto indices1 = hn::IndicesFromVec(du, idx1);
if constexpr (NumTables(du) == 1) {
(void)tbl1;
c0 = hn::TableLookupLanes(tbl0, indices0);
c1 = hn::TableLookupLanes(tbl0, indices1);
}
if constexpr (NumTables(du) == 2) { // `else` is poorly formatted.
c0 = hn::TwoTablesLookupLanes(du, tbl0, tbl1, indices0);
c1 = hn::TwoTablesLookupLanes(du, tbl0, tbl1, indices1);
}
}
// As above, but returns a single 16-bit output vector for f32 Dec2, thus
// packed is only a quarter-vector.
template <class DU, class VU = hn::Vec<DU>,
class D8Q = hn::Half<D8HFromD16<DU>>, class V8Q = hn::Vec<D8Q>>
static HWY_INLINE VU TableLookups(DU du, VU tbl0, VU tbl1, const V8Q packed) {
const D8HFromD16<DU> d8h;
// OrderedUnpackU16 expects a half-vector, but will only use the lower half
// of it.
const hn::Vec<decltype(d8h)> packed_h = hn::ZeroExtendVector(d8h, packed);
const VU idx0 = NibbleCodec::OrderedUnpackU16<0>(du, packed_h);
const auto indices0 = hn::IndicesFromVec(du, idx0);
if constexpr (NumTables(du) == 1) {
(void)tbl1;
return hn::TableLookupLanes(tbl0, indices0);
}
if constexpr (NumTables(du) == 2) { // `else` is poorly formatted.
return hn::TwoTablesLookupLanes(du, tbl0, tbl1, indices0);
}
}
public:
// Encodes `num` floats from `raw` into `packed`. `packed` points to
// compressed storage and `packed_ofs` indicates the destination offset within
// it, in number of elements. Tables are interleaved with indices (clustered
// elements) to allow for easier unpacking. Returns the total number of
// unused clusters, which is typically zero.
template <class DF, HWY_IF_F32_D(DF)>
static HWY_INLINE size_t Enc(DF df, const float* HWY_RESTRICT raw,
const size_t num, NuqStream::ClusterBuf& buf,
const PackedSpan<NuqStream>& packed,
size_t packed_ofs) {
const hn::Repartition<uint16_t, DF> d16;
const hn::Repartition<uint8_t, DF> d8;
using V16 = hn::Vec<decltype(d16)>;
using V8 = hn::Vec<decltype(d8)>;
const size_t N16 = hn::Lanes(d16);
HWY_ASSERT(packed_ofs % kGroupSize == 0);
const size_t num_groups = hwy::DivCeil(num, kGroupSize);
// TODO: dynamic resize should be removed; it is no longer necessary as
// interleaved encoding uses only a single buffer of the same size.
buf.Resize(1);
size_t unused_clusters = 0;
size_t current_offset = packed_ofs;
for (size_t g = 0; g < num_groups; ++g) {
const size_t g_num = HWY_MIN(num - g * kGroupSize, kGroupSize);
const float* HWY_RESTRICT g_in = raw + g * kGroupSize;
float* HWY_RESTRICT g_centers = buf.centers.get();
uint16_t* HWY_RESTRICT g_idx = buf.idx.get();
unused_clusters +=
NuqClustering::ClusterExactL2(df, g_in, g_num, buf, g_centers, g_idx);
uint8_t* centers = &packed.ptr->byte + TableByteOffset(current_offset);
SfpCodec::Enc(df, buf.centers.get(), kClusters,
reinterpret_cast<SfpStream*>(centers));
uint8_t* packed_start = centers + kClusters;
current_offset += g_num;
size_t i = 0;
HWY_UNROLL(1)
for (; i < g_num; i += 4 * N16) {
const V16 idx0 = hn::LoadU(d16, g_idx + i + 0 * N16);
const V16 idx1 = hn::LoadU(d16, g_idx + i + 1 * N16);
const V16 idx2 = hn::LoadU(d16, g_idx + i + 2 * N16);
const V16 idx3 = hn::LoadU(d16, g_idx + i + 3 * N16);
const V8 nibbles =
NibbleCodec::OrderedPackU16(d16, idx0, idx1, idx2, idx3);
hn::StoreU(nibbles, d8, packed_start + i / 2);
}
const size_t remaining = g_num - i;
if (HWY_UNLIKELY(remaining != 0)) {
const V16 idx0 = hn::LoadU(d16, g_idx + i + 0 * N16);
const V16 idx1 = hn::LoadU(d16, g_idx + i + 1 * N16);
const V16 idx2 = hn::LoadU(d16, g_idx + i + 2 * N16);
const V16 idx3 = hn::LoadU(d16, g_idx + i + 3 * N16);
const V8 nibbles =
NibbleCodec::OrderedPackU16(d16, idx0, idx1, idx2, idx3);
// i is even, but remaining might not be.
hn::StoreN(nibbles, d8, packed_start + i / 2,
hwy::DivCeil(remaining, 2));
}
}
return unused_clusters;
}
// Decompresses to two bf16 vectors. `packed_ofs` must be a multiple of two
// vectors so that we only have to load one group's table.
template <class DBF, HWY_IF_BF16_D(DBF)>
static HWY_INLINE void Dec2(DBF dbf,
const PackedSpan<const NuqStream>& packed,
const size_t packed_ofs, hn::Vec<DBF>& raw0,
hn::Vec<DBF>& raw1) {
const hn::RebindToUnsigned<decltype(dbf)> d16;
const D8HFromD16<DBF> d8h;
using V16 = hn::Vec<decltype(d16)>;
using V8H = hn::Vec<decltype(d8h)>;
const size_t within_group = packed_ofs % kGroupSize;
HWY_DASSERT(within_group % (2 * hn::Lanes(d16)) == 0);
const uint8_t* table = &packed.ptr->byte + TableByteOffset(packed_ofs);
const uint8_t* indices = table + kClusters + hwy::DivCeil(within_group, 2);
V16 tbl1 = Zero(d16);
const V16 tbl0 = LoadTable(d16, table, &tbl1);
const V8H nibbles = hn::LoadU(d8h, indices);
V16 c0, c1;
TableLookups(d16, tbl0, tbl1, nibbles, c0, c1);
raw0 = BitCast(dbf, c0);
raw1 = BitCast(dbf, c1);
}
// Decompresses to two f32 vectors. `packed_ofs` must be a multiple of two
// vectors so that we only have to load one group's table.
template <class DF, HWY_IF_F32_D(DF)>
static HWY_INLINE void Dec2(DF df, const PackedSpan<const NuqStream>& packed,
const size_t packed_ofs, hn::Vec<DF>& raw0,
hn::Vec<DF>& raw1) {
const hn::Repartition<BF16, decltype(df)> dbf;
const hn::RebindToUnsigned<decltype(dbf)> d16;
const hn::Half<D8HFromD16<decltype(d16)>> d8q;
using V8Q = hn::Vec<decltype(d8q)>;
using V16 = hn::Vec<decltype(d16)>;
const size_t within_group = packed_ofs % kGroupSize;
HWY_DASSERT(within_group % (2 * hn::Lanes(df)) == 0);
const uint8_t* table = &packed.ptr->byte + TableByteOffset(packed_ofs);
const uint8_t* indices = table + kClusters + hwy::DivCeil(within_group, 2);
V16 tbl1 = Zero(d16);
const V16 tbl0 = LoadTable(d16, table, &tbl1);
// The single-vector TableLookups overload only calls OrderedUnpackU16<0>,
// which expects a quarter vector of bytes.
const V8Q nibbles = hn::LoadU(d8q, indices);
// TODO(janwas): From janwas: on AVX-512 I imagine we can get a
// bit more speed for this function by changing LoadTable to return floats,
// then we could have a single lookup here instead of PromoteUpperTo which
// is not cheap.
const V16 c0 = TableLookups(d16, tbl0, tbl1, nibbles);
raw0 = hn::PromoteLowerTo(df, BitCast(dbf, c0));
raw1 = hn::PromoteUpperTo(df, BitCast(dbf, c0));
}
template <class D, typename Raw = hn::TFromD<D>>
static HWY_INLINE void DecompressAndZeroPad(
D d, const PackedSpan<const NuqStream>& packed, size_t packed_ofs,
Raw* HWY_RESTRICT raw, size_t num) {
// If unaligned, load elements from the first group and update the args,
// from which we compute new tables/indices below.
size_t current_offset = packed_ofs;
if (size_t within_group = packed_ofs % kGroupSize; within_group != 0) {
const uint8_t* tables =
&packed.ptr->byte + TableByteOffset(current_offset);
const uint8_t* indices = tables + kClusters + within_group / 2;
const size_t remaining = HWY_MIN(num, kGroupSize - within_group);
DecPartialGroup(d, tables, indices, raw, remaining);
packed_ofs += remaining;
current_offset += remaining;
raw += remaining;
num -= remaining;
if (num == 0) return;
}
HWY_DASSERT(packed_ofs % kGroupSize == 0);
const size_t num_groups = hwy::DivCeil(num, kGroupSize);
HWY_UNROLL(1)
for (size_t g = 0; g < num_groups - 1; ++g) {
const uint8_t* tables =
&packed.ptr->byte + TableByteOffset(current_offset);
const uint8_t* indices = tables + kClusters;
DecWholeGroup(d, tables, indices, raw + g * kGroupSize);
current_offset += kGroupSize;
}
const size_t g = num_groups - 1;
const uint8_t* tables = &packed.ptr->byte + TableByteOffset(current_offset);
const uint8_t* indices = tables + kClusters;
DecPartialGroup(d, tables, indices, raw + g * kGroupSize,
num - g * kGroupSize);
}
private:
template <class DBF, HWY_IF_BF16_D(DBF)>
static HWY_INLINE void DecWholeGroup(DBF dbf,
const uint8_t* HWY_RESTRICT table,
const uint8_t* HWY_RESTRICT indices,
BF16* HWY_RESTRICT raw_bf) {
const hn::RebindToUnsigned<decltype(dbf)> d16;
const D8HFromD16<DBF> d8h;
using V16 = hn::Vec<decltype(d16)>;
using V8H = hn::Vec<decltype(d8h)>;
const size_t N16 = hn::Lanes(d16);
V16 tbl1 = Zero(d16);
const V16 tbl0 = LoadTable(d16, table, &tbl1);
HWY_UNROLL(1)
for (size_t i = 0; i < kGroupSize; i += 2 * N16) {
const V8H nibbles = hn::LoadU(d8h, indices + i / 2);
V16 c0, c1;
TableLookups(d16, tbl0, tbl1, nibbles, c0, c1);
hn::StoreU(BitCast(dbf, c0), dbf, raw_bf + i + 0 * N16);
hn::StoreU(BitCast(dbf, c1), dbf, raw_bf + i + 1 * N16);
}
}
// Called for first and last group.
template <class DBF, HWY_IF_BF16_D(DBF)>
static HWY_INLINE void DecPartialGroup(DBF dbf,
const uint8_t* HWY_RESTRICT table,
const uint8_t* HWY_RESTRICT indices,
BF16* HWY_RESTRICT raw_bf,
size_t num) {
HWY_DASSERT(num <= kGroupSize);
const hn::RebindToUnsigned<decltype(dbf)> d16;
const D8HFromD16<DBF> d8h;
using V16 = hn::Vec<decltype(d16)>;
using V8H = hn::Vec<decltype(d8h)>;
const size_t N16 = hn::Lanes(d16);
V16 tbl1 = Zero(d16);
const V16 tbl0 = LoadTable(d16, table, &tbl1);
size_t i = 0;
if (num >= 2 * N16) {
HWY_UNROLL(1)
for (; i <= num - 2 * N16; i += 2 * N16) {
const V8H nibbles = hn::LoadU(d8h, indices + i / 2);
V16 c0, c1;
TableLookups(d16, tbl0, tbl1, nibbles, c0, c1);
hn::StoreU(BitCast(dbf, c0), dbf, raw_bf + i + 0 * N16);
hn::StoreU(BitCast(dbf, c1), dbf, raw_bf + i + 1 * N16);
}
}
const size_t remaining = num - i;
HWY_DASSERT(remaining < 2 * N16);
if (HWY_UNLIKELY(remaining != 0)) {
// i is even, but remaining might not be.
const V8H nibbles =
hn::LoadN(d8h, indices + i / 2, hwy::DivCeil(remaining, 2));
V16 c0, c1;
TableLookups(d16, tbl0, tbl1, nibbles, c0, c1);
// Out of bounds `nibbles` are 0, but this does not yet guarantee
// c0/c1 are, because centers[0] might not be 0.
c0 = hn::IfThenElseZero(hn::FirstN(d16, remaining), c0);
hn::StoreU(BitCast(dbf, c0), dbf, raw_bf + i);
// Callers only pad to one vector, so check before storing the second.
if (remaining > N16) {
c1 = hn::IfThenElseZero(hn::FirstN(d16, remaining - N16), c1);
hn::StoreU(BitCast(dbf, c1), dbf, raw_bf + i + N16);
}
}
}
template <class DF, HWY_IF_F32_D(DF)>
static HWY_INLINE void DecWholeGroup(DF df, const uint8_t* HWY_RESTRICT table,
const uint8_t* HWY_RESTRICT indices,
float* HWY_RESTRICT raw_f) {
const hn::Repartition<BF16, decltype(df)> dbf;
const hn::RebindToUnsigned<decltype(dbf)> d16;
const D8HFromD16<decltype(d16)> d8h;
using V16 = hn::Vec<decltype(d16)>;
using V8H = hn::Vec<decltype(d8h)>;
using VF = hn::Vec<decltype(df)>;
const size_t NF = hn::Lanes(df);
V16 tbl1 = Zero(d16);
const V16 tbl0 = LoadTable(d16, table, &tbl1);
HWY_UNROLL(1)
for (size_t i = 0; i < kGroupSize; i += 4 * NF) {
const V8H nibbles = hn::LoadU(d8h, indices + i / 2);
V16 c0, c1;
TableLookups(d16, tbl0, tbl1, nibbles, c0, c1);
const VF f0 = hn::PromoteLowerTo(df, BitCast(dbf, c0));
const VF f1 = hn::PromoteUpperTo(df, BitCast(dbf, c0));
const VF f2 = hn::PromoteLowerTo(df, BitCast(dbf, c1));
const VF f3 = hn::PromoteUpperTo(df, BitCast(dbf, c1));
hn::StoreU(f0, df, raw_f + i + 0 * NF);
hn::StoreU(f1, df, raw_f + i + 1 * NF);
hn::StoreU(f2, df, raw_f + i + 2 * NF);
hn::StoreU(f3, df, raw_f + i + 3 * NF);
}
}
// Called for first and last group.
template <class DF, HWY_IF_F32_D(DF)>
static HWY_INLINE void DecPartialGroup(DF df,
const uint8_t* HWY_RESTRICT table,
const uint8_t* HWY_RESTRICT indices,
float* HWY_RESTRICT raw_f,
const size_t num) {
HWY_DASSERT(num <= kGroupSize);
const hn::Repartition<BF16, decltype(df)> dbf;
const hn::RebindToUnsigned<decltype(dbf)> d16;
const D8HFromD16<decltype(d16)> d8h;
using V16 = hn::Vec<decltype(d16)>;
using V8H = hn::Vec<decltype(d8h)>;
using VF = hn::Vec<decltype(df)>;
const size_t NF = hn::Lanes(df);
V16 tbl1 = Zero(d16);
const V16 tbl0 = LoadTable(d16, table, &tbl1);
size_t i = 0;
if (num >= 4 * NF) {
HWY_UNROLL(1)
for (; i <= num - 4 * NF; i += 4 * NF) {
const V8H nibbles = hn::LoadU(d8h, indices + i / 2);
V16 c0, c1;
TableLookups(d16, tbl0, tbl1, nibbles, c0, c1);
const VF f0 = hn::PromoteLowerTo(df, BitCast(dbf, c0));
const VF f1 = hn::PromoteUpperTo(df, BitCast(dbf, c0));
const VF f2 = hn::PromoteLowerTo(df, BitCast(dbf, c1));
const VF f3 = hn::PromoteUpperTo(df, BitCast(dbf, c1));
hn::StoreU(f0, df, raw_f + i + 0 * NF);
hn::StoreU(f1, df, raw_f + i + 1 * NF);
hn::StoreU(f2, df, raw_f + i + 2 * NF);
hn::StoreU(f3, df, raw_f + i + 3 * NF);
}
}
const size_t remaining = num - i;
HWY_DASSERT(remaining < 4 * NF);
if (HWY_UNLIKELY(remaining != 0)) {
// i is even, but remaining might not be.
const V8H nibbles =
hn::LoadN(d8h, indices + i / 2, hwy::DivCeil(remaining, 2));
V16 c0, c1;
TableLookups(d16, tbl0, tbl1, nibbles, c0, c1);
const VF f0 = hn::PromoteLowerTo(df, BitCast(dbf, c0));
const VF f1 = hn::PromoteUpperTo(df, BitCast(dbf, c0));
const VF f2 = hn::PromoteLowerTo(df, BitCast(dbf, c1));
const VF f3 = hn::PromoteUpperTo(df, BitCast(dbf, c1));
// `raw_f` is only guaranteed to padded to NF, hence we cannot store all
// four vectors. We could conditionally store vectors either to `raw_f`
// or a buffer. However, we still have to mask because only `nibbles`
// are guaranteed to be 0, not c0/c1. Copying also involves branches,
// so we fully unroll the copy loop to avoid a buffer. We could also
// change the contract to pad to four vectors, but it would anyway be
// better to decompress to bf16.
if (remaining <= 1 * NF) {
const hn::Mask<DF> mask = hn::FirstN(df, remaining);
hn::StoreU(hn::IfThenElseZero(mask, f0), df, raw_f + i + 0 * NF);
return;
}
hn::StoreU(f0, df, raw_f + i + 0 * NF);
if (remaining <= 2 * NF) {
const hn::Mask<DF> mask = hn::FirstN(df, remaining - NF);
hn::StoreU(hn::IfThenElseZero(mask, f1), df, raw_f + i + 1 * NF);
return;
}
hn::StoreU(f1, df, raw_f + i + 1 * NF);
if (remaining <= 3 * NF) {
const hn::Mask<DF> mask = hn::FirstN(df, remaining - 2 * NF);
hn::StoreU(hn::IfThenElseZero(mask, f2), df, raw_f + i + 2 * NF);
return;
}
hn::StoreU(f2, df, raw_f + i + 2 * NF);
{
const hn::Mask<DF> mask = hn::FirstN(df, remaining - 3 * NF);
hn::StoreU(hn::IfThenElseZero(mask, f3), df, raw_f + i + 3 * NF);
}
}
}
}; // NuqCodec
// NOLINTNEXTLINE(google-readability-namespace-comments)
} // namespace HWY_NAMESPACE
} // namespace gcpp
HWY_AFTER_NAMESPACE();
#endif // THIRD_PARTY_GEMMA_CPP_COMPRESSION_NUQ_INL_H_
Reference in New Issue View Git Blame Copy Permalink