OSchip
/
OpenBLAS
Not watched
1
0
Fork 0
Code
Releases 66 Wiki evaluate Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain HPC
Browse Source

Small Matrix: skylakex: sgemm_nn: optimize for M <= 8

tags/v0.3.18
Wangyang Guo 4 years ago
parent
f88470323b
commit
49b61a3f30
1 changed files with 301 additions and 1 deletions
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +301
    -1
      kernel/x86_64/sgemm_small_kernel_nn_skylakex.c

+ 301
- 1
kernel/x86_64/sgemm_small_kernel_nn_skylakex.c View File

@@ -28,6 +28,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include <immintrin.h>
#include "common.h"
#include <stdio.h>
#include <memory.h>

#define DECLARE_RESULT_512(M, N) __m512 result##M##N = _mm512_setzero_ps()
#define LOAD_A_512(M, N) __m512 Aval##M = _mm512_loadu_ps(&A[lda * k + i + (M*16)])
@@ -52,6 +53,18 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
_mm512_mask_storeu_ps(&C[offset##M##N], mask, result##M##N)
#endif

#define LOAD_KA_512(M, N) __m512 Aval##M = _mm512_loadu_ps(&mbuf[(mi + M)*K + k]);
#define LOAD_KB_512(M, N) __m512 Bval##N = _mm512_loadu_ps(&B[(j + N)*ldb + k])
#define MASK_LOAD_KA_512(M, N) __m512 Aval##M = _mm512_maskz_loadu_ps(mask, &mbuf[(mi + M)*K + k])
#define MASK_LOAD_KB_512(M, N) __m512 Bval##N = _mm512_maskz_loadu_ps(mask, &B[(j + N)*ldb + k])
#if defined(B0)
#define STORE_REDUCE(M, N) C[(j+N)*ldc + i + M] = alpha * _mm512_reduce_add_ps(result##M##N);
#else
#define STORE_REDUCE(M, N) C[(j+N)*ldc + i + M] = alpha * _mm512_reduce_add_ps(result##M##N) + beta * C[(j+N)*ldc + i + M];
#endif



#define DECLARE_RESULT_256(M, N) __m256 result##M##N = _mm256_setzero_ps()
#define LOAD_A_256(M, N) __m256 Aval##M = _mm256_loadu_ps(&A[lda * k + i + (M*8)])
#define BROADCAST_LOAD_B_256(M, N) __m256 Bval##N = _mm256_broadcastss_ps(_mm_load_ss(&B[k + ldb * (j+N)]))
@@ -249,7 +262,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT alp
STORE_512(0, 0);
}
}
if (M - i > 0) {
if (M - i > 8) {
register __mmask16 mask asm("k1") = (1UL << (M - i)) - 1;
for (j = 0; j < n4; j += 4) {
DECLARE_RESULT_512(0, 0);
@@ -294,6 +307,293 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT alp
}
return;
}
int mm = M - i;
if (mm) {
FLOAT *mbuf = (FLOAT *) malloc(sizeof(FLOAT)*mm*K);
__mmask8 mask8 = (1UL << mm) - 1;
__mmask16 mask;
BLASLONG k16 = K & ~15;
BLASLONG k8 = K & ~7;
for (k = 0; k < k8; k += 8) {
__m256 r0, r1, r2, r3, r4, r5, r6, r7;
__m256 t0, t1, t2, t3, t4, t5, t6, t7;
r0 = _mm256_maskz_loadu_ps(mask8, &A[i + lda*(0 + k)]);
r1 = _mm256_maskz_loadu_ps(mask8, &A[i + lda*(1 + k)]);
r2 = _mm256_maskz_loadu_ps(mask8, &A[i + lda*(2 + k)]);
r3 = _mm256_maskz_loadu_ps(mask8, &A[i + lda*(3 + k)]);
r4 = _mm256_maskz_loadu_ps(mask8, &A[i + lda*(4 + k)]);
r5 = _mm256_maskz_loadu_ps(mask8, &A[i + lda*(5 + k)]);
r6 = _mm256_maskz_loadu_ps(mask8, &A[i + lda*(6 + k)]);
r7 = _mm256_maskz_loadu_ps(mask8, &A[i + lda*(7 + k)]);

t0 = _mm256_unpacklo_ps(r0, r1);
t1 = _mm256_unpackhi_ps(r0, r1);
t2 = _mm256_unpacklo_ps(r2, r3);
t3 = _mm256_unpackhi_ps(r2, r3);
t4 = _mm256_unpacklo_ps(r4, r5);
t5 = _mm256_unpackhi_ps(r4, r5);
t6 = _mm256_unpacklo_ps(r6, r7);
t7 = _mm256_unpackhi_ps(r6, r7);

r0 = _mm256_shuffle_ps(t0,t2,_MM_SHUFFLE(1,0,1,0));
r1 = _mm256_shuffle_ps(t0,t2,_MM_SHUFFLE(3,2,3,2));
r2 = _mm256_shuffle_ps(t1,t3,_MM_SHUFFLE(1,0,1,0));
r3 = _mm256_shuffle_ps(t1,t3,_MM_SHUFFLE(3,2,3,2));
r4 = _mm256_shuffle_ps(t4,t6,_MM_SHUFFLE(1,0,1,0));
r5 = _mm256_shuffle_ps(t4,t6,_MM_SHUFFLE(3,2,3,2));
r6 = _mm256_shuffle_ps(t5,t7,_MM_SHUFFLE(1,0,1,0));
r7 = _mm256_shuffle_ps(t5,t7,_MM_SHUFFLE(3,2,3,2));

t0 = _mm256_permute2f128_ps(r0, r4, 0x20);
t1 = _mm256_permute2f128_ps(r1, r5, 0x20);
t2 = _mm256_permute2f128_ps(r2, r6, 0x20);
t3 = _mm256_permute2f128_ps(r3, r7, 0x20);
t4 = _mm256_permute2f128_ps(r0, r4, 0x31);
t5 = _mm256_permute2f128_ps(r1, r5, 0x31);
t6 = _mm256_permute2f128_ps(r2, r6, 0x31);
t7 = _mm256_permute2f128_ps(r3, r7, 0x31);

switch (mm) {
case 8: _mm256_storeu_ps(&mbuf[k + 7*K], t7);
case 7: _mm256_storeu_ps(&mbuf[k + 6*K], t6);
case 6: _mm256_storeu_ps(&mbuf[k + 5*K], t5);
case 5: _mm256_storeu_ps(&mbuf[k + 4*K], t4);
case 4: _mm256_storeu_ps(&mbuf[k + 3*K], t3);
case 3: _mm256_storeu_ps(&mbuf[k + 2*K], t2);
case 2: _mm256_storeu_ps(&mbuf[k + 1*K], t1);
case 1: _mm256_storeu_ps(&mbuf[k + 0*K], t0);
}
}
for (; k < K; k++) {
for (int ii = 0; ii < mm; ii++) {
mbuf[k + ii*K] = A[i + lda*k + ii];
}
}
int mi = 0;
for (; i < m4; i += 4, mi += 4) {
for (j = 0; j < n4; j += 4) {
DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0); DECLARE_RESULT_512(2, 0); DECLARE_RESULT_512(3, 0);
DECLARE_RESULT_512(0, 1); DECLARE_RESULT_512(1, 1); DECLARE_RESULT_512(2, 1); DECLARE_RESULT_512(3, 1);
DECLARE_RESULT_512(0, 2); DECLARE_RESULT_512(1, 2); DECLARE_RESULT_512(2, 2); DECLARE_RESULT_512(3, 2);
DECLARE_RESULT_512(0, 3); DECLARE_RESULT_512(1, 3); DECLARE_RESULT_512(2, 3); DECLARE_RESULT_512(3, 3);
for (k = 0; k < k16; k += 16) {
LOAD_KA_512(0, x); LOAD_KA_512(1, x); LOAD_KA_512(2, x); LOAD_KA_512(3, x);
LOAD_KB_512(x, 0); LOAD_KB_512(x, 1); LOAD_KB_512(x, 2); LOAD_KB_512(x, 3);

MATMUL_512(0, 0); MATMUL_512(1, 0); MATMUL_512(2, 0); MATMUL_512(3, 0);
MATMUL_512(0, 1); MATMUL_512(1, 1); MATMUL_512(2, 1); MATMUL_512(3, 1);
MATMUL_512(0, 2); MATMUL_512(1, 2); MATMUL_512(2, 2); MATMUL_512(3, 2);
MATMUL_512(0, 3); MATMUL_512(1, 3); MATMUL_512(2, 3); MATMUL_512(3, 3);
}
int remains = K - k;
if (remains) {
mask = (1UL << remains) - 1;
MASK_LOAD_KA_512(0, x); MASK_LOAD_KA_512(1, x); MASK_LOAD_KA_512(2, x); MASK_LOAD_KA_512(3, x);
MASK_LOAD_KB_512(x, 0); MASK_LOAD_KB_512(x, 1); MASK_LOAD_KB_512(x, 2); MASK_LOAD_KB_512(x, 3);

MATMUL_512(0, 0); MATMUL_512(1, 0); MATMUL_512(2, 0); MATMUL_512(3, 0);
MATMUL_512(0, 1); MATMUL_512(1, 1); MATMUL_512(2, 1); MATMUL_512(3, 1);
MATMUL_512(0, 2); MATMUL_512(1, 2); MATMUL_512(2, 2); MATMUL_512(3, 2);
MATMUL_512(0, 3); MATMUL_512(1, 3); MATMUL_512(2, 3); MATMUL_512(3, 3);
}
STORE_REDUCE(0, 0); STORE_REDUCE(1, 0); STORE_REDUCE(2, 0); STORE_REDUCE(3, 0);
STORE_REDUCE(0, 1); STORE_REDUCE(1, 1); STORE_REDUCE(2, 1); STORE_REDUCE(3, 1);
STORE_REDUCE(0, 2); STORE_REDUCE(1, 2); STORE_REDUCE(2, 2); STORE_REDUCE(3, 2);
STORE_REDUCE(0, 3); STORE_REDUCE(1, 3); STORE_REDUCE(2, 3); STORE_REDUCE(3, 3);
}
for (; j < n2; j += 2) {
DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0); DECLARE_RESULT_512(2, 0); DECLARE_RESULT_512(3, 0);
DECLARE_RESULT_512(0, 1); DECLARE_RESULT_512(1, 1); DECLARE_RESULT_512(2, 1); DECLARE_RESULT_512(3, 1);
for (k = 0; k < k16; k += 16) {
LOAD_KA_512(0, x); LOAD_KA_512(1, x); LOAD_KA_512(2, x); LOAD_KA_512(3, x);
LOAD_KB_512(x, 0); LOAD_KB_512(x, 1);

MATMUL_512(0, 0); MATMUL_512(1, 0); MATMUL_512(2, 0); MATMUL_512(3, 0);
MATMUL_512(0, 1); MATMUL_512(1, 1); MATMUL_512(2, 1); MATMUL_512(3, 1);
}
int remains = K - k;
if (remains) {
mask = (1UL << remains) - 1;
MASK_LOAD_KA_512(0, x); MASK_LOAD_KA_512(1, x); MASK_LOAD_KA_512(2, x); MASK_LOAD_KA_512(3, x);
MASK_LOAD_KB_512(x, 0); MASK_LOAD_KB_512(x, 1);

MATMUL_512(0, 0); MATMUL_512(1, 0); MATMUL_512(2, 0); MATMUL_512(3, 0);
MATMUL_512(0, 1); MATMUL_512(1, 1); MATMUL_512(2, 1); MATMUL_512(3, 1);
}
STORE_REDUCE(0, 0); STORE_REDUCE(1, 0); STORE_REDUCE(2, 0); STORE_REDUCE(3, 0);
STORE_REDUCE(0, 1); STORE_REDUCE(1, 1); STORE_REDUCE(2, 1); STORE_REDUCE(3, 1);

}
for (; j < N; j += 1) {
DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0); DECLARE_RESULT_512(2, 0); DECLARE_RESULT_512(3, 0);
for (k = 0; k < k16; k += 16) {
LOAD_KA_512(0, x); LOAD_KA_512(1, x); LOAD_KA_512(2, x); LOAD_KA_512(3, x);
LOAD_KB_512(x, 0);

MATMUL_512(0, 0); MATMUL_512(1, 0); MATMUL_512(2, 0); MATMUL_512(3, 0);
}
int remains = K - k;
if (remains) {
mask = (1UL << remains) - 1;
MASK_LOAD_KA_512(0, x); MASK_LOAD_KA_512(1, x); MASK_LOAD_KA_512(2, x); MASK_LOAD_KA_512(3, x);
MASK_LOAD_KB_512(x, 0);

MATMUL_512(0, 0); MATMUL_512(1, 0); MATMUL_512(2, 0); MATMUL_512(3, 0);
}
STORE_REDUCE(0, 0); STORE_REDUCE(1, 0); STORE_REDUCE(2, 0); STORE_REDUCE(3, 0);
}

}
for (; i < m2; i += 2, mi += 2) {
for (j = 0; j < n4; j += 4) {
DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0);
DECLARE_RESULT_512(0, 1); DECLARE_RESULT_512(1, 1);
DECLARE_RESULT_512(0, 2); DECLARE_RESULT_512(1, 2);
DECLARE_RESULT_512(0, 3); DECLARE_RESULT_512(1, 3);
for (k = 0; k < k16; k += 16) {
LOAD_KA_512(0, x); LOAD_KA_512(1, x);
LOAD_KB_512(x, 0); LOAD_KB_512(x, 1); LOAD_KB_512(x, 2); LOAD_KB_512(x, 3);

MATMUL_512(0, 0); MATMUL_512(1, 0);
MATMUL_512(0, 1); MATMUL_512(1, 1);
MATMUL_512(0, 2); MATMUL_512(1, 2);
MATMUL_512(0, 3); MATMUL_512(1, 3);
}
int remains = K - k;
if (remains) {
mask = (1UL << remains) - 1;
MASK_LOAD_KA_512(0, x); MASK_LOAD_KA_512(1, x);
MASK_LOAD_KB_512(x, 0); MASK_LOAD_KB_512(x, 1); MASK_LOAD_KB_512(x, 2); MASK_LOAD_KB_512(x, 3);

MATMUL_512(0, 0); MATMUL_512(1, 0);
MATMUL_512(0, 1); MATMUL_512(1, 1);
MATMUL_512(0, 2); MATMUL_512(1, 2);
MATMUL_512(0, 3); MATMUL_512(1, 3);
}
STORE_REDUCE(0, 0); STORE_REDUCE(1, 0);
STORE_REDUCE(0, 1); STORE_REDUCE(1, 1);
STORE_REDUCE(0, 2); STORE_REDUCE(1, 2);
STORE_REDUCE(0, 3); STORE_REDUCE(1, 3);
}
for (; j < n2; j += 2) {
DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0);
DECLARE_RESULT_512(0, 1); DECLARE_RESULT_512(1, 1);
for (k = 0; k < k16; k += 16) {
LOAD_KA_512(0, x); LOAD_KA_512(1, x);
LOAD_KB_512(x, 0); LOAD_KB_512(x, 1);

MATMUL_512(0, 0); MATMUL_512(1, 0);
MATMUL_512(0, 1); MATMUL_512(1, 1);
}
int remains = K - k;
if (remains) {
mask = (1UL << remains) - 1;
MASK_LOAD_KA_512(0, x); MASK_LOAD_KA_512(1, x);
MASK_LOAD_KB_512(x, 0); MASK_LOAD_KB_512(x, 1);

MATMUL_512(0, 0); MATMUL_512(1, 0);
MATMUL_512(0, 1); MATMUL_512(1, 1);
}
STORE_REDUCE(0, 0); STORE_REDUCE(1, 0);
STORE_REDUCE(0, 1); STORE_REDUCE(1, 1);

}
for (; j < N; j += 1) {
DECLARE_RESULT_512(0, 0); DECLARE_RESULT_512(1, 0);
for (k = 0; k < k16; k += 16) {
LOAD_KA_512(0, x); LOAD_KA_512(1, x);
LOAD_KB_512(x, 0);

MATMUL_512(0, 0); MATMUL_512(1, 0);
}
int remains = K - k;
if (remains) {
mask = (1UL << remains) - 1;
MASK_LOAD_KA_512(0, x); MASK_LOAD_KA_512(1, x);
MASK_LOAD_KB_512(x, 0);

MATMUL_512(0, 0); MATMUL_512(1, 0);
}
STORE_REDUCE(0, 0); STORE_REDUCE(1, 0);
}
}
for (; i < M; i += 1, mi += 1) {
for (j = 0; j < n4; j += 4) {
DECLARE_RESULT_512(0, 0);
DECLARE_RESULT_512(0, 1);
DECLARE_RESULT_512(0, 2);
DECLARE_RESULT_512(0, 3);
for (k = 0; k < k16; k += 16) {
LOAD_KA_512(0, x);
LOAD_KB_512(x, 0); LOAD_KB_512(x, 1); LOAD_KB_512(x, 2); LOAD_KB_512(x, 3);

MATMUL_512(0, 0);
MATMUL_512(0, 1);
MATMUL_512(0, 2);
MATMUL_512(0, 3);
}
int remains = K - k;
if (remains) {
mask = (1UL << remains) - 1;
MASK_LOAD_KA_512(0, x);
MASK_LOAD_KB_512(x, 0); MASK_LOAD_KB_512(x, 1); MASK_LOAD_KB_512(x, 2); MASK_LOAD_KB_512(x, 3);


MATMUL_512(0, 0);
MATMUL_512(0, 1);
MATMUL_512(0, 2);
MATMUL_512(0, 3);
}
STORE_REDUCE(0, 0);
STORE_REDUCE(0, 1);
STORE_REDUCE(0, 2);
STORE_REDUCE(0, 3);
}
for (; j < n2; j += 2) {
DECLARE_RESULT_512(0, 0);
DECLARE_RESULT_512(0, 1);
for (k = 0; k < k16; k += 16) {
LOAD_KA_512(0, x);
LOAD_KB_512(x, 0); LOAD_KB_512(x, 1);

MATMUL_512(0, 0);
MATMUL_512(0, 1);
}
int remains = K - k;
if (remains) {
mask = (1UL << remains) - 1;
MASK_LOAD_KA_512(0, x);
MASK_LOAD_KB_512(x, 0); MASK_LOAD_KB_512(x, 1);

MATMUL_512(0, 0);
MATMUL_512(0, 1);
}
STORE_REDUCE(0, 0);
STORE_REDUCE(0, 1);

}
for (; j < N; j += 1) {
DECLARE_RESULT_512(0, 0);
for (k = 0; k < k16; k += 16) {
LOAD_KA_512(0, x);
LOAD_KB_512(x, 0);

MATMUL_512(0, 0);
}
int remains = K - k;
if (remains) {
mask = (1UL << remains) - 1;
MASK_LOAD_KA_512(0, x);
MASK_LOAD_KB_512(x, 0);

MATMUL_512(0, 0);
}
STORE_REDUCE(0, 0);
}
}
free(mbuf);
return;
}
__m256 alpha_256 = _mm256_broadcastss_ps(_mm_load_ss(&alpha));
#if !defined(B0)
__m256 beta_256 = _mm256_broadcastss_ps(_mm_load_ss(&beta));


Write Preview
Loading…
Cancel
Save