Skip to content

Commit 1a039f1

Browse files
at0m741at0m741
committed
increasing SGEMM performances (still need to detect CPU features and memory topology at compile time to fine tune the GEMM model)
1 parent 69273fa commit 1a039f1

File tree

3 files changed

+55
-43
lines changed

3 files changed

+55
-43
lines changed

‎Tests/CMakeLists.txt‎

Lines changed: 13 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -30,6 +30,19 @@ if(USE_KNL)
3030
target_compile_definitions(TensoriumTests PRIVATE USE_KNL)
3131
target_link_libraries(TensoriumTests PRIVATE memkind)
3232
endif()
33+
# ─────────────────────────────── ASM / LLVM dump options ───────────────────────────────
34+
option(DUMP_ASM "Dump generated assembly (.s files) for inspection" OFF)
35+
option(DUMP_LLVM "Dump LLVM IR (.ll files) for inspection" OFF)
36+
37+
if(DUMP_ASM)
38+
message(STATUS "Assembly dump enabled (-S -masm=intel -fverbose-asm)")
39+
add_compile_options(-S -masm=intel -fverbose-asm)
40+
endif()
41+
42+
if(DUMP_LLVM)
43+
message(STATUS "LLVM IR dump enabled (-emit-llvm -S)")
44+
add_compile_options(-emit-llvm -S)
45+
endif()
3346

3447
find_library(OPENBLAS_LIB
3548
NAMES openblasp openblas

‎Tests/main.cpp‎

Lines changed: 6 additions & 16 deletions
Original file line numberDiff line numberDiff line change
@@ -2,26 +2,16 @@
22

33
using namespace tensorium;
44

5-
void test_lexer(const std::string& input) {
6-
Lexer lexer(input);
7-
auto tokens = lexer.tokenize();
85

9-
std::cout << "Input: " << input << "\n";
10-
for (const auto& token : tokens) {
11-
std::cout << "Token: \"" << token.value << "\" — type: " << static_cast<int>(token.type) << "\n";
12-
}
13-
std::cout << "------\n";
14-
}
156

167
int main() {
17-
#pragma tensorium dispatch
18-
float __attribute__((aligned(8))) low_align[32];
8+
// #pragma tensorium dispatch
199

20-
deriv_test();
21-
linear_solver_test();
10+
// deriv_test();
11+
// linear_solver_test();
2212
matrix_tests();
23-
tensor_test();
24-
vector_tests();
25-
deriv_test_spectral_fft();
13+
// tensor_test();
14+
// vector_tests();
15+
// deriv_test_spectral_fft();
2616
return 0;
2717
}

‎includes/Tensorium/Core/MatrixKernels/GemmKernel_bigger.hpp‎

Lines changed: 36 additions & 27 deletions
Original file line numberDiff line numberDiff line change
@@ -5,6 +5,7 @@
55
#include <algorithm>
66
#include <cstdlib>
77
#include <cstring>
8+
#include <omp.h>
89
/*
910
* this Gemm kernel is based on Aman Salykov version. Improvment of the OMP schedulding and Block
1011
* sizes
@@ -20,7 +21,7 @@ template <typename T> class GemmKernelBigger {
2021
static constexpr int SimdWidth = Simd::width;
2122
static constexpr int TileRows = SimdWidth * 4;
2223
static constexpr int TileCols = 6;
23-
static constexpr int NThreads = 72;
24+
static constexpr int NThreads = 36;
2425

2526
// static constexpr int BlockDepth = 256;
2627
// static constexpr int BlockRows = 384;
@@ -196,7 +197,7 @@ template <typename T> class GemmKernelBigger {
196197
}
197198

198199
inline static void build_masks(__m256i *packed_mask_0, __m256i *packed_mask_1, int mr) {
199-
#if defined(__AVX512F__)
200+
# if defined(__AVX512F__)
200201
__m128i m0 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(&mask[32 - mr]));
201202
__m128i m1 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(&mask[32 - mr + 16]));
202203

@@ -206,15 +207,15 @@ template <typename T> class GemmKernelBigger {
206207
*packed_mask_0 = _mm512_castsi512_si256(p0);
207208
*packed_mask_1 = _mm512_castsi512_si256(p1);
208209

209-
#elif defined(__AVX2__)
210+
# elif defined(__AVX2__)
210211
__m128i m0 = _mm_loadl_epi64(reinterpret_cast<const __m128i *>(&mask[16 - mr]));
211212
__m128i m1 = _mm_loadl_epi64(reinterpret_cast<const __m128i *>(&mask[16 - mr + 8]));
212213

213214
*packed_mask_0 = _mm256_cvtepi8_epi32(m0);
214215
*packed_mask_1 = _mm256_cvtepi8_epi32(m1);
215-
#else
216-
# error "AVX2 or AVX-512 required"
217-
#endif
216+
# else
217+
# error "AVX2 or AVX-512 required"
218+
# endif
218219
}
219220

220221
inline void maskload_accum_00(T *C, reg *C_accum_00, reg *C_accum_01, __m256i packed_mask_0,
@@ -507,8 +508,8 @@ template <typename T> class GemmKernelBigger {
507508
Simd::maskstore(&C[5 * M + 8], packed_mask_1, *C_accum_51);
508509
}
509510

510-
void kernel_16x6_load_accum(T* __restrict blockA_packed, T* __restrict blockB_packed, T* __restrict C, int mr, int nr, int kc,
511-
int M) {
511+
inline void kernel_16x6_load_accum(T *__restrict blockA_packed, T *__restrict blockB_packed,
512+
T *__restrict C, int mr, int nr, int kc, int M) {
512513
reg C_accum_00 = {};
513514
reg C_accum_01 = {};
514515
reg C_accum_10 = {};
@@ -651,8 +652,9 @@ template <typename T> class GemmKernelBigger {
651652
}
652653
}
653654

654-
void kernel_16x6_zero_init_accum(T* __restrict blockA_packed, T* __restrict blockB_packed, T* __restrict C, int mr, int nr,
655-
int kc, int M) {
655+
inline void kernel_16x6_zero_init_accum(T *__restrict blockA_packed,
656+
T *__restrict blockB_packed, T *__restrict C, int mr,
657+
int nr, int kc, int M) {
656658
reg C_accum_00 = {};
657659
reg C_accum_01 = {};
658660
reg C_accum_10 = {};
@@ -769,25 +771,25 @@ template <typename T> class GemmKernelBigger {
769771
}
770772
}
771773

772-
#ifndef NTHREADS
773-
# define NTHREADS 36
774-
#endif
774+
# ifndef NTHREADS
775+
# define NTHREADS 36
776+
# endif
775777

776-
#define MC (16 * (40 / NTHREADS) * NTHREADS)
777-
#define NC (6 * (800 / NTHREADS) * NTHREADS)
778-
#define KC 500
778+
# define KC 512
779+
# define MC 384
780+
# define NC 4096
779781

780-
#ifndef OMP_SCHEDULE
781-
# define OMP_SCHEDULE auto
782-
#endif
783-
#define _min(x, y) ((x) < (y) ? (x) : (y))
784-
#define PRAGMA_OMP_PARALLEL_FOR \
785-
_Pragma("omp parallel for schedule(OMP_SCHEDULE) num_threads(NTHREADS)")
782+
# ifndef OMP_SCHEDULE
783+
# define OMP_SCHEDULE dynamic
784+
# endif
785+
# define _min(x, y) ((x) < (y) ? (x) : (y))
786+
# define PRAGMA_OMP_PARALLEL_FOR \
787+
_Pragma("omp parallel for schedule(OMP_SCHEDULE) num_threads(NTHREADS)")
786788

787789
static T blockA_packed[MC * KC] __attribute__((aligned(64)));
788790
static T blockB_packed[NC * KC] __attribute__((aligned(64)));
789791

790-
void pack_panelB(T *B, T *blockB_packed, int nr, int kc, int K) {
792+
inline void pack_panelB(T *B, T *blockB_packed, int nr, int kc, int K) {
791793
for (int p = 0; p < kc; p++) {
792794
for (int j = 0; j < nr; j++) {
793795
*blockB_packed++ = B[j * K + p];
@@ -799,14 +801,14 @@ template <typename T> class GemmKernelBigger {
799801
}
800802

801803
void pack_blockB(T *B, T *blockB_packed, int nc, int kc, int K) {
802-
#pragma omp for schedule(dynamic)
804+
# pragma omp for schedule(dynamic)
803805
for (int j = 0; j < nc; j += 6) {
804806
int nr = _min(6, nc - j);
805807
pack_panelB(&B[j * K], &blockB_packed[j * kc], nr, kc, K);
806808
}
807809
}
808810

809-
void pack_panelA(T *A, T *blockA_packed, int mr, int kc, int M) {
811+
inline void pack_panelA(T *A, T *blockA_packed, int mr, int kc, int M) {
810812
for (int p = 0; p < kc; p++) {
811813
for (int i = 0; i < mr; i++) {
812814
*blockA_packed++ = A[p * M + i];
@@ -817,15 +819,22 @@ template <typename T> class GemmKernelBigger {
817819
}
818820
}
819821

820-
void pack_blockA(T *A, T *blockA_packed, int mc, int kc, int M) {
822+
inline void pack_blockA(T *A, T *blockA_packed, int mc, int kc, int M) {
821823
PRAGMA_OMP_PARALLEL_FOR
822824
for (int i = 0; i < mc; i += 16) {
823825
int mr = _min(16, mc - i);
824826
pack_panelA(&A[i], &blockA_packed[i * kc], mr, kc, M);
825827
}
826828
}
827-
828829
void matmul(T *A, T *B, T *C, int M, int N, int K) {
830+
# pragma omp parallel
831+
{
832+
int tid = omp_get_thread_num();
833+
cpu_set_t cpuset;
834+
CPU_ZERO(&cpuset);
835+
CPU_SET(tid % 36, &cpuset);
836+
sched_setaffinity(0, sizeof(cpuset), &cpuset);
837+
}
829838
for (int j = 0; j < N; j += NC) {
830839
int nc = _min(NC, N - j);
831840
int kc = _min(KC, K);

0 commit comments

Comments
 (0)