Source code
Revision control
Copy as Markdown
Other Tools
// Copyright 2021 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
//
//
// 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.
// clang-format off
#if defined(HIGHWAY_HWY_CONTRIB_DOT_DOT_INL_H_) == defined(HWY_TARGET_TOGGLE) // NOLINT
// clang-format on
#ifdef HIGHWAY_HWY_CONTRIB_DOT_DOT_INL_H_
#undef HIGHWAY_HWY_CONTRIB_DOT_DOT_INL_H_
#else
#define HIGHWAY_HWY_CONTRIB_DOT_DOT_INL_H_
#endif
#include <stddef.h>
#include "hwy/highway.h"
HWY_BEFORE_NAMESPACE();
namespace hwy {
namespace HWY_NAMESPACE {
struct Dot {
// Specify zero or more of these, ORed together, as the kAssumptions template
// argument to Compute. Each one may improve performance or reduce code size,
// at the cost of additional requirements on the arguments.
enum Assumptions {
// num_elements is at least N, which may be up to HWY_MAX_BYTES / sizeof(T).
kAtLeastOneVector = 1,
// num_elements is divisible by N (a power of two, so this can be used if
// the problem size is known to be a power of two >= HWY_MAX_BYTES /
// sizeof(T)).
kMultipleOfVector = 2,
// RoundUpTo(num_elements, N) elements are accessible; their value does not
// matter (will be treated as if they were zero).
kPaddedToVector = 4,
};
// Returns sum{pa[i] * pb[i]} for floating-point inputs, including float16_t
// and double if HWY_HAVE_FLOAT16/64. Aligning the
// pointers to a multiple of N elements is helpful but not required.
template <int kAssumptions, class D, typename T = TFromD<D>>
static HWY_INLINE T Compute(const D d, const T* const HWY_RESTRICT pa,
const T* const HWY_RESTRICT pb,
const size_t num_elements) {
static_assert(IsFloat<T>(), "MulAdd requires float type");
using V = decltype(Zero(d));
const size_t N = Lanes(d);
size_t i = 0;
constexpr bool kIsAtLeastOneVector =
(kAssumptions & kAtLeastOneVector) != 0;
constexpr bool kIsMultipleOfVector =
(kAssumptions & kMultipleOfVector) != 0;
constexpr bool kIsPaddedToVector = (kAssumptions & kPaddedToVector) != 0;
// Won't be able to do a full vector load without padding => scalar loop.
if (!kIsAtLeastOneVector && !kIsMultipleOfVector && !kIsPaddedToVector &&
HWY_UNLIKELY(num_elements < N)) {
// Only 2x unroll to avoid excessive code size.
T sum0 = ConvertScalarTo<T>(0);
T sum1 = ConvertScalarTo<T>(0);
for (; i + 2 <= num_elements; i += 2) {
// For reasons unknown, fp16 += does not compile on clang (Arm).
sum0 = ConvertScalarTo<T>(sum0 + pa[i + 0] * pb[i + 0]);
sum1 = ConvertScalarTo<T>(sum1 + pa[i + 1] * pb[i + 1]);
}
if (i < num_elements) {
sum1 = ConvertScalarTo<T>(sum1 + pa[i] * pb[i]);
}
return ConvertScalarTo<T>(sum0 + sum1);
}
// Compiler doesn't make independent sum* accumulators, so unroll manually.
// 2 FMA ports * 4 cycle latency = up to 8 in-flight, but that is excessive
// for unaligned inputs (each unaligned pointer halves the throughput
// because it occupies both L1 load ports for a cycle). We cannot have
// arrays of vectors on RVV/SVE, so always unroll 4x.
V sum0 = Zero(d);
V sum1 = Zero(d);
V sum2 = Zero(d);
V sum3 = Zero(d);
// Main loop: unrolled
for (; i + 4 * N <= num_elements; /* i += 4 * N */) { // incr in loop
const auto a0 = LoadU(d, pa + i);
const auto b0 = LoadU(d, pb + i);
i += N;
sum0 = MulAdd(a0, b0, sum0);
const auto a1 = LoadU(d, pa + i);
const auto b1 = LoadU(d, pb + i);
i += N;
sum1 = MulAdd(a1, b1, sum1);
const auto a2 = LoadU(d, pa + i);
const auto b2 = LoadU(d, pb + i);
i += N;
sum2 = MulAdd(a2, b2, sum2);
const auto a3 = LoadU(d, pa + i);
const auto b3 = LoadU(d, pb + i);
i += N;
sum3 = MulAdd(a3, b3, sum3);
}
// Up to 3 iterations of whole vectors
for (; i + N <= num_elements; i += N) {
const auto a = LoadU(d, pa + i);
const auto b = LoadU(d, pb + i);
sum0 = MulAdd(a, b, sum0);
}
if (!kIsMultipleOfVector) {
const size_t remaining = num_elements - i;
if (remaining != 0) {
if (kIsPaddedToVector) {
const auto mask = FirstN(d, remaining);
const auto a = LoadU(d, pa + i);
const auto b = LoadU(d, pb + i);
sum1 = MulAdd(IfThenElseZero(mask, a), IfThenElseZero(mask, b), sum1);
} else {
// Unaligned load such that the last element is in the highest lane -
// ensures we do not touch any elements outside the valid range.
// If we get here, then num_elements >= N.
HWY_DASSERT(i >= N);
i += remaining - N;
const auto skip = FirstN(d, N - remaining);
const auto a = LoadU(d, pa + i); // always unaligned
const auto b = LoadU(d, pb + i);
sum1 = MulAdd(IfThenZeroElse(skip, a), IfThenZeroElse(skip, b), sum1);
}
}
} // kMultipleOfVector
// Reduction tree: sum of all accumulators by pairs, then across lanes.
sum0 = Add(sum0, sum1);
sum2 = Add(sum2, sum3);
sum0 = Add(sum0, sum2);
return ReduceSum(d, sum0);
}
// f32 * bf16
template <int kAssumptions, class DF, HWY_IF_F32_D(DF)>
static HWY_INLINE float Compute(const DF df,
const float* const HWY_RESTRICT pa,
const hwy::bfloat16_t* const HWY_RESTRICT pb,
const size_t num_elements) {
#if HWY_TARGET == HWY_SCALAR
const Rebind<hwy::bfloat16_t, DF> dbf;
#else
const Repartition<hwy::bfloat16_t, DF> dbf;
using VBF = decltype(Zero(dbf));
#endif
const Half<decltype(dbf)> dbfh;
using VF = decltype(Zero(df));
const size_t NF = Lanes(df);
constexpr bool kIsAtLeastOneVector =
(kAssumptions & kAtLeastOneVector) != 0;
constexpr bool kIsMultipleOfVector =
(kAssumptions & kMultipleOfVector) != 0;
constexpr bool kIsPaddedToVector = (kAssumptions & kPaddedToVector) != 0;
// Won't be able to do a full vector load without padding => scalar loop.
if (!kIsAtLeastOneVector && !kIsMultipleOfVector && !kIsPaddedToVector &&
HWY_UNLIKELY(num_elements < NF)) {
// Only 2x unroll to avoid excessive code size.
float sum0 = 0.0f;
float sum1 = 0.0f;
size_t i = 0;
for (; i + 2 <= num_elements; i += 2) {
sum0 += pa[i + 0] * ConvertScalarTo<float>(pb[i + 0]);
sum1 += pa[i + 1] * ConvertScalarTo<float>(pb[i + 1]);
}
for (; i < num_elements; ++i) {
sum1 += pa[i] * ConvertScalarTo<float>(pb[i]);
}
return sum0 + sum1;
}
// Compiler doesn't make independent sum* accumulators, so unroll manually.
// 2 FMA ports * 4 cycle latency = up to 8 in-flight, but that is excessive
// for unaligned inputs (each unaligned pointer halves the throughput
// because it occupies both L1 load ports for a cycle). We cannot have
// arrays of vectors on RVV/SVE, so always unroll 4x.
VF sum0 = Zero(df);
VF sum1 = Zero(df);
VF sum2 = Zero(df);
VF sum3 = Zero(df);
size_t i = 0;
#if HWY_TARGET != HWY_SCALAR // PromoteUpperTo supported
// Main loop: unrolled
for (; i + 4 * NF <= num_elements; /* i += 4 * N */) { // incr in loop
const VF a0 = LoadU(df, pa + i);
const VBF b0 = LoadU(dbf, pb + i);
i += NF;
sum0 = MulAdd(a0, PromoteLowerTo(df, b0), sum0);
const VF a1 = LoadU(df, pa + i);
i += NF;
sum1 = MulAdd(a1, PromoteUpperTo(df, b0), sum1);
const VF a2 = LoadU(df, pa + i);
const VBF b2 = LoadU(dbf, pb + i);
i += NF;
sum2 = MulAdd(a2, PromoteLowerTo(df, b2), sum2);
const VF a3 = LoadU(df, pa + i);
i += NF;
sum3 = MulAdd(a3, PromoteUpperTo(df, b2), sum3);
}
#endif // HWY_TARGET == HWY_SCALAR
// Up to 3 iterations of whole vectors
for (; i + NF <= num_elements; i += NF) {
const VF a = LoadU(df, pa + i);
const VF b = PromoteTo(df, LoadU(dbfh, pb + i));
sum0 = MulAdd(a, b, sum0);
}
if (!kIsMultipleOfVector) {
const size_t remaining = num_elements - i;
if (remaining != 0) {
if (kIsPaddedToVector) {
const auto mask = FirstN(df, remaining);
const VF a = LoadU(df, pa + i);
const VF b = PromoteTo(df, LoadU(dbfh, pb + i));
sum1 = MulAdd(IfThenElseZero(mask, a), IfThenElseZero(mask, b), sum1);
} else {
// Unaligned load such that the last element is in the highest lane -
// ensures we do not touch any elements outside the valid range.
// If we get here, then num_elements >= N.
HWY_DASSERT(i >= NF);
i += remaining - NF;
const auto skip = FirstN(df, NF - remaining);
const VF a = LoadU(df, pa + i); // always unaligned
const VF b = PromoteTo(df, LoadU(dbfh, pb + i));
sum1 = MulAdd(IfThenZeroElse(skip, a), IfThenZeroElse(skip, b), sum1);
}
}
} // kMultipleOfVector
// Reduction tree: sum of all accumulators by pairs, then across lanes.
sum0 = Add(sum0, sum1);
sum2 = Add(sum2, sum3);
sum0 = Add(sum0, sum2);
return ReduceSum(df, sum0);
}
// Returns sum{pa[i] * pb[i]} for bfloat16 inputs. Aligning the pointers to a
// multiple of N elements is helpful but not required.
template <int kAssumptions, class D, HWY_IF_BF16_D(D)>
static HWY_INLINE float Compute(const D d,
const bfloat16_t* const HWY_RESTRICT pa,
const bfloat16_t* const HWY_RESTRICT pb,
const size_t num_elements) {
const RebindToUnsigned<D> du16;
const Repartition<float, D> df32;
using V = decltype(Zero(df32));
const size_t N = Lanes(d);
size_t i = 0;
constexpr bool kIsAtLeastOneVector =
(kAssumptions & kAtLeastOneVector) != 0;
constexpr bool kIsMultipleOfVector =
(kAssumptions & kMultipleOfVector) != 0;
constexpr bool kIsPaddedToVector = (kAssumptions & kPaddedToVector) != 0;
// Won't be able to do a full vector load without padding => scalar loop.
if (!kIsAtLeastOneVector && !kIsMultipleOfVector && !kIsPaddedToVector &&
HWY_UNLIKELY(num_elements < N)) {
float sum0 = 0.0f; // Only 2x unroll to avoid excessive code size for..
float sum1 = 0.0f; // this unlikely(?) case.
for (; i + 2 <= num_elements; i += 2) {
sum0 += F32FromBF16(pa[i + 0]) * F32FromBF16(pb[i + 0]);
sum1 += F32FromBF16(pa[i + 1]) * F32FromBF16(pb[i + 1]);
}
if (i < num_elements) {
sum1 += F32FromBF16(pa[i]) * F32FromBF16(pb[i]);
}
return sum0 + sum1;
}
// See comment in the other Compute() overload. Unroll 2x, but we need
// twice as many sums for ReorderWidenMulAccumulate.
V sum0 = Zero(df32);
V sum1 = Zero(df32);
V sum2 = Zero(df32);
V sum3 = Zero(df32);
// Main loop: unrolled
for (; i + 2 * N <= num_elements; /* i += 2 * N */) { // incr in loop
const auto a0 = LoadU(d, pa + i);
const auto b0 = LoadU(d, pb + i);
i += N;
sum0 = ReorderWidenMulAccumulate(df32, a0, b0, sum0, sum1);
const auto a1 = LoadU(d, pa + i);
const auto b1 = LoadU(d, pb + i);
i += N;
sum2 = ReorderWidenMulAccumulate(df32, a1, b1, sum2, sum3);
}
// Possibly one more iteration of whole vectors
if (i + N <= num_elements) {
const auto a0 = LoadU(d, pa + i);
const auto b0 = LoadU(d, pb + i);
i += N;
sum0 = ReorderWidenMulAccumulate(df32, a0, b0, sum0, sum1);
}
if (!kIsMultipleOfVector) {
const size_t remaining = num_elements - i;
if (remaining != 0) {
if (kIsPaddedToVector) {
const auto mask = FirstN(du16, remaining);
const auto va = LoadU(d, pa + i);
const auto vb = LoadU(d, pb + i);
const auto a16 = BitCast(d, IfThenElseZero(mask, BitCast(du16, va)));
const auto b16 = BitCast(d, IfThenElseZero(mask, BitCast(du16, vb)));
sum2 = ReorderWidenMulAccumulate(df32, a16, b16, sum2, sum3);
} else {
// Unaligned load such that the last element is in the highest lane -
// ensures we do not touch any elements outside the valid range.
// If we get here, then num_elements >= N.
HWY_DASSERT(i >= N);
i += remaining - N;
const auto skip = FirstN(du16, N - remaining);
const auto va = LoadU(d, pa + i); // always unaligned
const auto vb = LoadU(d, pb + i);
const auto a16 = BitCast(d, IfThenZeroElse(skip, BitCast(du16, va)));
const auto b16 = BitCast(d, IfThenZeroElse(skip, BitCast(du16, vb)));
sum2 = ReorderWidenMulAccumulate(df32, a16, b16, sum2, sum3);
}
}
} // kMultipleOfVector
// Reduction tree: sum of all accumulators by pairs, then across lanes.
sum0 = Add(sum0, sum1);
sum2 = Add(sum2, sum3);
sum0 = Add(sum0, sum2);
return ReduceSum(df32, sum0);
}
};
// NOLINTNEXTLINE(google-readability-namespace-comments)
} // namespace HWY_NAMESPACE
} // namespace hwy
HWY_AFTER_NAMESPACE();
#endif // HIGHWAY_HWY_CONTRIB_DOT_DOT_INL_H_