Skip to content

feat(bench): add vectorized + unrolled optimal version #63

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
dssgabriel wants to merge 2 commits into
base: main
Choose a base branch
from
Open

feat(bench): add vectorized + unrolled optimal version #63

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
bad4ff0
feat(bench): add vectorized + unrolled optimal version
dssgabriel Sep 5, 2025
0c5ccc5
feat(bench): add version w/ vec+unrll over all inner loops
dssgabriel Sep 9, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
187 changes: 174 additions & 13 deletions benchmarks/gemm.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,7 +1,8 @@
// benchmarks/gemm.cpp -*-C++-*-
// SPDX-License-Identifier: Apache-2.0

#include <getopt.h>
#include <immintrin.h>

#include <cassert>
#include <chrono>
Expand Down Expand Up @@ -192,9 +193,9 @@
assert(A.extent(0) == C.extent(0));
assert(B.extent(1) == C.extent(1));

using ATile = std::mdspan<Type, std::extents<size_t, TM, TO>, std::layout_right>;
using BTile = std::mdspan<Type, std::extents<size_t, TO, TN>, std::layout_right>;
using CTile = std::mdspan<Type, std::extents<size_t, TM, TN>, std::layout_right>;
using ATile = std::mdspan<Type, std::extents<size_t, TM, TO>, typename MdSpanA::mapping_type::inner_layout_type>;
using BTile = std::mdspan<Type, std::extents<size_t, TO, TN>, typename MdSpanB::mapping_type::inner_layout_type>;
using CTile = std::mdspan<Type, std::extents<size_t, TM, TN>, typename MdSpanC::mapping_type::inner_layout_type>;

const size_t M = A.extent(0);
const size_t N = B.extent(1);
Expand Down Expand Up @@ -228,14 +229,14 @@

template <size_t TM, size_t TN, size_t TO, class MdSpanA, class MdSpanB, class MdSpanC>
requires IsMdSpan2D<MdSpanA> && IsMdSpan2D<MdSpanB> && IsMdSpan2D<MdSpanC>
auto manual_blocked_tiled_subview_BLTL(const MdSpanA A, const MdSpanB B, MdSpanC C) -> void {
auto manual_blocked_tiled_subview_vec_unroll_k(const MdSpanA A, const MdSpanB B, MdSpanC C) -> void {
assert(A.extent(1) == B.extent(0));
assert(A.extent(0) == C.extent(0));
assert(B.extent(1) == C.extent(1));

using ATile = std::mdspan<Type, std::extents<size_t, TM, TO>, std::layout_right>;
using BTile = std::mdspan<Type, std::extents<size_t, TO, TN>, std::layout_left>;
using CTile = std::mdspan<Type, std::extents<size_t, TM, TN>, std::layout_right>;
using ATile = std::mdspan<Type, std::extents<size_t, TM, TO>, typename MdSpanA::mapping_type::inner_layout_type>;
using BTile = std::mdspan<Type, std::extents<size_t, TO, TN>, typename MdSpanB::mapping_type::inner_layout_type>;
using CTile = std::mdspan<Type, std::extents<size_t, TM, TN>, typename MdSpanC::mapping_type::inner_layout_type>;

const size_t M = A.extent(0);
const size_t N = B.extent(1);
Expand All @@ -253,20 +254,147 @@
for (size_t i = 0; i < std::min(TM, M - x); ++i) {
for (size_t j = 0; j < std::min(TN, N - y); ++j) {
Type acc = 0.0;
#if defined(__AVX__) && defined(__FMA__)
// 8x unrolled and vectorized kernel (processing 32 elements/iter).
// If TO == 32, this effectively fully unrolls the `k` loop.
//
// On Zen3, `vfmaddXXXpd` is 4 lat / 0.5 tp (CPI) => at least 2 FMAs needed to achieve max throughput.
// Unrolling more allows to amortize the cost of the final horizontal reduction.
//
// TODO(dssgabriel): Attempt to vectorize the `j` loop to remove the need for the horizontal reduction.
if constexpr (TO >= 32 and TO % 32 == 0) {
__m256d vacc = _mm256_setzero_pd();
for (size_t k = 0; k < std::min(TO, O - z); k += 32) {
__m256d va0 = _mm256_loadu_pd(&tA[i, k]);
__m256d vb0 = _mm256_loadu_pd(&tB[k, j]);
vacc = _mm256_fmadd_pd(va0, vb0, vacc);
__m256d va1 = _mm256_loadu_pd(&tA[i, k + 4]);
__m256d vb1 = _mm256_loadu_pd(&tB[k + 4, j]);
vacc = _mm256_fmadd_pd(va1, vb1, vacc);
__m256d va2 = _mm256_loadu_pd(&tA[i, k + 8]);
__m256d vb2 = _mm256_loadu_pd(&tB[k + 8, j]);
vacc = _mm256_fmadd_pd(va2, vb2, vacc);
__m256d va3 = _mm256_loadu_pd(&tA[i, k + 12]);
__m256d vb3 = _mm256_loadu_pd(&tB[k + 12, j]);
vacc = _mm256_fmadd_pd(va3, vb3, vacc);
__m256d va4 = _mm256_loadu_pd(&tA[i, k + 16]);
__m256d vb4 = _mm256_loadu_pd(&tB[k + 16, j]);
vacc = _mm256_fmadd_pd(va4, vb4, vacc);
__m256d va5 = _mm256_loadu_pd(&tA[i, k + 20]);
__m256d vb5 = _mm256_loadu_pd(&tB[k + 20, j]);
vacc = _mm256_fmadd_pd(va5, vb5, vacc);
__m256d va6 = _mm256_loadu_pd(&tA[i, k + 24]);
__m256d vb6 = _mm256_loadu_pd(&tB[k + 24, j]);
vacc = _mm256_fmadd_pd(va6, vb6, vacc);
__m256d va7 = _mm256_loadu_pd(&tA[i, k + 28]);
__m256d vb7 = _mm256_loadu_pd(&tB[k + 28, j]);
vacc = _mm256_fmadd_pd(va7, vb7, vacc);
}
__m128d hsum = _mm_add_pd(_mm256_castpd256_pd128(vacc), _mm256_extractf128_pd(vacc, 1));

Check warning on line 293 in benchmarks/gemm.cpp

View workflow job for this annotation

GitHub Actions / lint

benchmarks/gemm.cpp:293:30 [portability-simd-intrinsics] 

'_mm_add_pd' is a non-portable x86_64 intrinsic function

@github-actions github-actions Bot Sep 5, 2025

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

clang-tidy diagnostic

benchmarks/gemm.cpp:294:30: warning: [portability-simd-intrinsics]

'_mm_add_pd' is a non-portable x86_64 intrinsic function

  294 |               __m128d hsum = _mm_add_pd(_mm256_castpd256_pd128(vacc), _mm256_extractf128_pd(vacc, 1));
      |                              ^
/home/runner/work/layout-tiled/layout-tiled/benchmarks/gemm.cpp:539:10: warning: do not declare C-style arrays, use 'std::array' instead [cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays]
  539 |   static struct option long_options[] = {
      |          ^
/home/runner/work/layout-tiled/layout-tiled/build/_deps/fmt-src/include/fmt/base.h:1683:5: error: 5 uninitialized fields at the end of the constructor call [clang-analyzer-optin.cplusplus.UninitializedObject,-warnings-as-errors]
 1683 |     ignore_unused(arg_index, named_arg_index);
      |     ^
/home/runner/work/layout-tiled/layout-tiled/build/_deps/fmt-src/include/fmt/base.h:1213:7: note: uninitialized field 'this->context_.num_args_'
 1213 |   int num_args_;
      |       ^~~~~~~~~
/home/runner/work/layout-tiled/layout-tiled/build/_deps/fmt-src/include/fmt/base.h:1214:15: note: uninitialized pointer 'this->context_.types_'
 1214 |   const type* types_;
      |               ^~~~~~
/home/runner/work/layout-tiled/layout-tiled/build/_deps/fmt-src/include/fmt/base.h:520:15: note: uninitialized pointer 'this->context_.parse_context::fmt_.data_'
  520 |   const Char* data_;
      |               ^~~~~
/home/runner/work/layout-tiled/layout-tiled/build/_deps/fmt-src/include/fmt/base.h:521:10: note: uninitialized field 'this->context_.parse_context::fmt_.size_'
  521 |   size_t size_;
      |          ^~~~~
/home/runner/work/layout-tiled/layout-tiled/build/_deps/fmt-src/include/fmt/base.h:870:7: note: uninitialized field 'this->context_.parse_context::next_arg_id_'
  870 |   int next_arg_id_;
      |       ^~~~~~~~~~~~
/home/runner/work/layout-tiled/layout-tiled/benchmarks/gemm.cpp:549:10: note: Assuming the condition is true
  549 |   while ((opt = getopt_long(argc, argv, "n:t:o:h", static_cast<const struct option*>(long_options), nullptr)) != -1) {
      |          ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/home/runner/work/layout-tiled/layout-tiled/benchmarks/gemm.cpp:549:3: note: Loop condition is true.  Entering loop body
  549 |   while ((opt = getopt_long(argc, argv, "n:t:o:h", static_cast<const struct option*>(long_options), nullptr)) != -1) {
      |   ^
/home/runner/work/layout-tiled/layout-tiled/benchmarks/gemm.cpp:550:5: note: Control jumps to the 'default' case at line 571
  550 |     switch (opt) {
      |     ^
/home/runner/work/layout-tiled/layout-tiled/benchmarks/gemm.cpp:572:18: note: Calling constructor for 'fstring<std::basic_string<char> &>'
  572 |       fmt::print("{}", HELP_MSG);
      |                  ^~~~
/home/runner/work/layout-tiled/layout-tiled/build/_deps/fmt-src/include/fmt/base.h:2709:5: note: Taking true branch
 2709 |     if (FMT_USE_CONSTEVAL) parse_format_string<char>(s, checker(s, arg_pack()));
      |     ^
/home/runner/work/layout-tiled/layout-tiled/build/_deps/fmt-src/include/fmt/base.h:2709:57: note: Calling constructor for 'format_string_checker<char, 1, 0, false>'
 2709 |     if (FMT_USE_CONSTEVAL) parse_format_string<char>(s, checker(s, arg_pack()));
      |                                                         ^~~~~~~~~~~~~~~~~~~~~~
/home/runner/work/layout-tiled/layout-tiled/build/_deps/fmt-src/include/fmt/base.h:1683:5: note: 5 uninitialized fields at the end of the constructor call
 1683 |     ignore_unused(arg_index, named_arg_index);
      |     ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@github-actions github-actions Bot Sep 9, 2025

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

clang-tidy diagnostic

benchmarks/gemm.cpp:293:30: warning: [portability-simd-intrinsics]

'_mm_add_pd' is a non-portable x86_64 intrinsic function

  293 |               __m128d hsum = _mm_add_pd(_mm256_castpd256_pd128(vacc), _mm256_extractf128_pd(vacc, 1));
      |                              ^
/home/runner/work/layout-tiled/layout-tiled/benchmarks/gemm.cpp:627:10: warning: do not declare C-style arrays, use 'std::array' instead [cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays]
  627 |   static struct option long_options[] = {
      |          ^
/home/runner/work/layout-tiled/layout-tiled/build/_deps/fmt-src/include/fmt/base.h:1683:5: error: 5 uninitialized fields at the end of the constructor call [clang-analyzer-optin.cplusplus.UninitializedObject,-warnings-as-errors]
 1683 |     ignore_unused(arg_index, named_arg_index);
      |     ^
/home/runner/work/layout-tiled/layout-tiled/build/_deps/fmt-src/include/fmt/base.h:520:15: note: uninitialized pointer 'this->context_.parse_context::fmt_.data_'
  520 |   const Char* data_;
      |               ^~~~~
/home/runner/work/layout-tiled/layout-tiled/build/_deps/fmt-src/include/fmt/base.h:521:10: note: uninitialized field 'this->context_.parse_context::fmt_.size_'
  521 |   size_t size_;
      |          ^~~~~
/home/runner/work/layout-tiled/layout-tiled/build/_deps/fmt-src/include/fmt/base.h:870:7: note: uninitialized field 'this->context_.parse_context::next_arg_id_'
  870 |   int next_arg_id_;
      |       ^~~~~~~~~~~~
/home/runner/work/layout-tiled/layout-tiled/build/_deps/fmt-src/include/fmt/base.h:1213:7: note: uninitialized field 'this->context_.num_args_'
 1213 |   int num_args_;
      |       ^~~~~~~~~
/home/runner/work/layout-tiled/layout-tiled/build/_deps/fmt-src/include/fmt/base.h:1214:15: note: uninitialized pointer 'this->context_.types_'
 1214 |   const type* types_;
      |               ^~~~~~
/home/runner/work/layout-tiled/layout-tiled/benchmarks/gemm.cpp:637:10: note: Assuming the condition is true
  637 |   while ((opt = getopt_long(argc, argv, "n:t:o:h", static_cast<const struct option*>(long_options), nullptr)) != -1) {
      |          ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/home/runner/work/layout-tiled/layout-tiled/benchmarks/gemm.cpp:637:3: note: Loop condition is true.  Entering loop body
  637 |   while ((opt = getopt_long(argc, argv, "n:t:o:h", static_cast<const struct option*>(long_options), nullptr)) != -1) {
      |   ^
/home/runner/work/layout-tiled/layout-tiled/benchmarks/gemm.cpp:638:5: note: Control jumps to the 'default' case at line 659
  638 |     switch (opt) {
      |     ^
/home/runner/work/layout-tiled/layout-tiled/benchmarks/gemm.cpp:660:18: note: Calling constructor for 'fstring<std::basic_string<char> &>'
  660 |       fmt::print("{}", HELP_MSG);
      |                  ^~~~
/home/runner/work/layout-tiled/layout-tiled/build/_deps/fmt-src/include/fmt/base.h:2709:5: note: Taking true branch
 2709 |     if (FMT_USE_CONSTEVAL) parse_format_string<char>(s, checker(s, arg_pack()));
      |     ^
/home/runner/work/layout-tiled/layout-tiled/build/_deps/fmt-src/include/fmt/base.h:2709:57: note: Calling constructor for 'format_string_checker<char, 1, 0, false>'
 2709 |     if (FMT_USE_CONSTEVAL) parse_format_string<char>(s, checker(s, arg_pack()));
      |                                                         ^~~~~~~~~~~~~~~~~~~~~~
/home/runner/work/layout-tiled/layout-tiled/build/_deps/fmt-src/include/fmt/base.h:1683:5: note: 5 uninitialized fields at the end of the constructor call
 1683 |     ignore_unused(arg_index, named_arg_index);
      |     ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

hsum = _mm_hadd_pd(hsum, hsum);
acc = _mm_cvtsd_f64(hsum);
} else {
for (size_t k = 0; k < std::min(TO, O - z); ++k) {
acc += tA[i, k] * tB[k, j];
}
}
tC[i, j] += acc;
#else
#if defined(LAYOUT_TILED_BENCH_ENABLE_OMP)
#pragma omp simd
#endif
for (size_t k = 0; k < std::min(TO, O - z); ++k) {
acc += tA[i, k] * tB[k, j];
}
tC[i, j] += acc;
#endif
}
}
}
}
}
}

#if defined(__AVX__) && defined(__FMA__)
template <size_t TM, size_t TN, size_t TO, class MdSpanA, class MdSpanB, class MdSpanC>
requires IsMdSpan2D<MdSpanA> && IsMdSpan2D<MdSpanB> && IsMdSpan2D<MdSpanC>
auto manual_blocked_tiled_subview_vec_unroll2(const MdSpanA A, const MdSpanB B, MdSpanC C) -> void {
assert(A.extent(1) == B.extent(0));
assert(A.extent(0) == C.extent(0));
assert(B.extent(1) == C.extent(1));

using ATile = std::mdspan<Type, std::extents<size_t, TM, TO>, typename MdSpanA::mapping_type::inner_layout_type>;
using BTile = std::mdspan<Type, std::extents<size_t, TO, TN>, typename MdSpanB::mapping_type::inner_layout_type>;
using CTile = std::mdspan<Type, std::extents<size_t, TM, TN>, typename MdSpanC::mapping_type::inner_layout_type>;

constexpr size_t MR = 4;
constexpr size_t NR = 8;

const size_t M = A.extent(0);
const size_t N = B.extent(1);
const size_t O = A.extent(1);

#if defined(LAYOUT_TILED_BENCH_ENABLE_OMP)
#pragma omp parallel for schedule(static)
#endif
for (size_t x = 0; x < M; x += TM) {
for (size_t y = 0; y < N; y += TN) {
auto tC = CTile(C.data_handle() + C.mapping().tile_offset_outer(x, y));
for (size_t z = 0; z < O; z += TO) {
const auto tA = ATile(A.data_handle() + A.mapping().tile_offset_outer(x, z));
const auto tB = BTile(B.data_handle() + B.mapping().tile_offset_outer(z, y));
// 4x8 register blocking
for (size_t i = 0; i < TM; i += MR) {
for (size_t j = 0; j < TN; j += NR) {
// 8 accumulators over 4 rows/2 columns
__m256d vc00 = _mm256_loadu_pd(&tC[i, j]);
__m256d vc01 = _mm256_loadu_pd(&tC[i, j + 4]);
__m256d vc10 = _mm256_loadu_pd(&tC[i + 1, j]);
__m256d vc11 = _mm256_loadu_pd(&tC[i + 1, j + 4]);
__m256d vc20 = _mm256_loadu_pd(&tC[i + 2, j]);
__m256d vc21 = _mm256_loadu_pd(&tC[i + 2, j + 4]);
__m256d vc30 = _mm256_loadu_pd(&tC[i + 3, j]);
__m256d vc31 = _mm256_loadu_pd(&tC[i + 3, j + 4]);

for (size_t k = 0; k < TO; ++k) {
// Broadcast 4 elements of A
__m256d sa0 = _mm256_broadcast_sd(&tA[i, k]);
__m256d sa1 = _mm256_broadcast_sd(&tA[i + 1, k]);
__m256d sa2 = _mm256_broadcast_sd(&tA[i + 2, k]);
__m256d sa3 = _mm256_broadcast_sd(&tA[i + 3, k]);

// Multiply broadcast A with 2 columns of B
__m256d vb0 = _mm256_loadu_pd(&tB[k, j]);
vc00 = _mm256_fmadd_pd(sa0, vb0, vc00);
vc10 = _mm256_fmadd_pd(sa1, vb0, vc10);
vc20 = _mm256_fmadd_pd(sa2, vb0, vc20);
vc30 = _mm256_fmadd_pd(sa3, vb0, vc30);
__m256d vb1 = _mm256_loadu_pd(&tB[k, j + 4]);
vc01 = _mm256_fmadd_pd(sa0, vb1, vc01);
vc11 = _mm256_fmadd_pd(sa1, vb1, vc11);
vc21 = _mm256_fmadd_pd(sa2, vb1, vc21);
vc31 = _mm256_fmadd_pd(sa3, vb1, vc31);
}

// Store accumulators
_mm256_storeu_pd(&tC[i, j], vc00);
_mm256_storeu_pd(&tC[i, j + 4], vc01);
_mm256_storeu_pd(&tC[i + 1, j], vc10);
_mm256_storeu_pd(&tC[i + 1, j + 4], vc11);
_mm256_storeu_pd(&tC[i + 2, j], vc20);
_mm256_storeu_pd(&tC[i + 2, j + 4], vc21);
_mm256_storeu_pd(&tC[i + 3, j], vc30);
_mm256_storeu_pd(&tC[i + 3, j + 4], vc31);
}
}
}
}
}
}
#else
#error unsupported impl on target arch
#endif

#if defined(LAYOUT_TILED_BENCH_ENABLE_BLAS)
template <class MdSpanA, class MdSpanB, class MdSpanC>
requires IsMdSpan2D<MdSpanA> && IsMdSpan2D<MdSpanB> && IsMdSpan2D<MdSpanC> && IsLayoutRight<MdSpanA>
Expand Down Expand Up @@ -339,10 +467,10 @@
auto mtlB_l2 = std::mdspan<Type, Exts, layout_tiled::layout_ltl<TILE_L2, TILE_L2>>(tlB_l2.get(), cfg.n_, cfg.n_);

// Matrix initializations
ankerl::nanobench::Rng rng_naive(42);
benchmark_utils::fill_2d_uniform01(mrA, rng_naive);
benchmark_utils::fill_2d_uniform01(mrB, rng_naive);
benchmark_utils::fill_2d_uniform01(mlB, rng_naive);
ankerl::nanobench::Rng rng(42);
benchmark_utils::fill_2d_uniform01(mrA, rng);
benchmark_utils::fill_2d_uniform01(mrB, rng);
benchmark_utils::fill_2d_copy(mlB, mrB);
benchmark_utils::fill_2d_copy(mtrA_l1, mrA);
benchmark_utils::fill_2d_copy(mtrB_l1, mrB);
benchmark_utils::fill_2d_copy(mtlB_l1, mlB);
Expand Down Expand Up @@ -388,6 +516,27 @@
res = benchmark_utils::check(ref, benchmark_utils::l2_norm(mtrC_l2), "RTR-L2_mbtsv", big_o);
benchmark_utils::fill_zeroes(mtrC_l2);

manual_blocked_tiled_subview<TILE_L1, TILE_L1, TILE_L1>(mtrA_l1, mtlB_l1, mtrC_l1);
res = benchmark_utils::check(ref, benchmark_utils::l2_norm(mtrC_l1), "RTR-BLTL-L1_mbtsv", big_o);
benchmark_utils::fill_zeroes(mtrC_l1);
manual_blocked_tiled_subview<TILE_L2, TILE_L2, TILE_L2>(mtrA_l2, mtlB_l2, mtrC_l2);
res = benchmark_utils::check(ref, benchmark_utils::l2_norm(mtrC_l2), "RTR-BLTL-L2_mbtsv", big_o);
benchmark_utils::fill_zeroes(mtrC_l2);

manual_blocked_tiled_subview_vec_unroll_k<TILE_L1, TILE_L1, TILE_L1>(mtrA_l1, mtlB_l1, mtrC_l1);
res = benchmark_utils::check(ref, benchmark_utils::l2_norm(mtrC_l1), "RTR-BLTL-L1_mbtvu", big_o);
benchmark_utils::fill_zeroes(mtrC_l1);
manual_blocked_tiled_subview_vec_unroll_k<TILE_L2, TILE_L2, TILE_L2>(mtrA_l2, mtlB_l2, mtrC_l2);
res = benchmark_utils::check(ref, benchmark_utils::l2_norm(mtrC_l2), "RTR-BLTL-L2_mbtvu", big_o);
benchmark_utils::fill_zeroes(mtrC_l2);

manual_blocked_tiled_subview_vec_unroll2<TILE_L1, TILE_L1, TILE_L1>(mtrA_l1, mtrB_l1, mtrC_l1);
res = benchmark_utils::check(ref, benchmark_utils::l2_norm(mtrC_l1), "RTR-BLTL-L1_mbt_vu2", big_o);
benchmark_utils::fill_zeroes(mtrC_l1);
manual_blocked_tiled_subview_vec_unroll2<TILE_L2, TILE_L2, TILE_L2>(mtrA_l2, mtrB_l2, mtrC_l2);
res = benchmark_utils::check(ref, benchmark_utils::l2_norm(mtrC_l2), "RTR-BLTL-L2_mbt_vu2", big_o);
benchmark_utils::fill_zeroes(mtrC_l2);

return res;
#else
// Benchmarks
Expand Down Expand Up @@ -438,10 +587,22 @@
manual_blocked_tiled_subview<TILE_L2, TILE_L2, TILE_L2>(mtrA_l2, mtrB_l2, mtrC_l2);
});
bs.run(fmt::format("RTR_BLTL-{0}x{0} mbtsv", TILE_L1), [&]() {
manual_blocked_tiled_subview_BLTL<TILE_L1, TILE_L1, TILE_L1>(mtrA_l1, mtlB_l1, mtrC_l1);
manual_blocked_tiled_subview<TILE_L1, TILE_L1, TILE_L1>(mtrA_l1, mtlB_l1, mtrC_l1);
});
bs.run(fmt::format("RTR_BLTL-{0}x{0} mbtsv", TILE_L2), [&]() {
manual_blocked_tiled_subview_BLTL<TILE_L2, TILE_L2, TILE_L2>(mtrA_l2, mtlB_l2, mtrC_l2);
manual_blocked_tiled_subview<TILE_L2, TILE_L2, TILE_L2>(mtrA_l2, mtlB_l2, mtrC_l2);
});
bs.run(fmt::format("RTR_BLTL-{0}x{0} mbt_vu", TILE_L1), [&]() {
manual_blocked_tiled_subview_vec_unroll_k<TILE_L1, TILE_L1, TILE_L1>(mtrA_l1, mtlB_l1, mtrC_l1);
});
bs.run(fmt::format("RTR_BLTL-{0}x{0} mbt_vu", TILE_L2), [&]() {
manual_blocked_tiled_subview_vec_unroll_k<TILE_L2, TILE_L2, TILE_L2>(mtrA_l2, mtlB_l2, mtrC_l2);
});
bs.run(fmt::format("RTR_BLTL-{0}x{0} mbt_vu2", TILE_L1), [&]() {
manual_blocked_tiled_subview_vec_unroll2<TILE_L1, TILE_L1, TILE_L1>(mtrA_l1, mtrB_l1, mtrC_l1);
});
bs.run(fmt::format("RTR_BLTL-{0}x{0} mbt_vu2", TILE_L2), [&]() {
manual_blocked_tiled_subview_vec_unroll2<TILE_L2, TILE_L2, TILE_L2>(mtrA_l2, mtrB_l2, mtrC_l2);
});

// Output
Expand Down
Loading