Skip to content

GitLab

Commit ee4e6cab authored by Andreas Marek's avatar Andreas Marek
Browse files

Check in AVX-512 kernels

parent 4ccc4ad5
Hide whitespace changes
Inline Side-by-side
Showing with 176 additions and 60 deletions
src/elpa2/kernels/complex_avx512_1hv_template.Xc
View file @ ee4e6cab
......@@ -51,6 +51,7 @@
#include <complex.h>
#include <x86intrin.h>
#include <stdio.h>
#include <stdlib.h>
#define __forceinline __attribute__((always_inline))
......@@ -269,7 +270,8 @@ void single_hh_trafo_complex_avx512_1hv_single(float complex* q, float complex*
#endif
if (worked_on != nq)
{
printf("Error in complex AVX512 BLOCK 1 kernel \n");
// printf("Error in complex AVX512 BLOCK 1 kernel \n");
// abort();
}
}
......
src/elpa2/kernels/complex_avx512_2hv_template.Xc
View file @ ee4e6cab
......@@ -49,6 +49,7 @@
#include <complex.h>
#include <x86intrin.h>
#include <stdio.h>
#include <stdlib.h>
#define __forceinline __attribute__((always_inline))
......@@ -193,25 +194,29 @@ void double_hh_trafo_complex_avx512_2hv_single(float complex* q, float complex*
return;
}
#ifdef DOUBLE_PRECISION_COMPLEX
if (nq-i == 12 ) {
if (nq-i == 12 )
{
hh_trafo_complex_kernel_12_AVX512_2hv_double(&q[i], hh, nb, ldq, ldh, s);
worked_on += 12;
}
}
#endif
#ifdef SINGLE_PRECISION_COMPLEX
if (nq-i == 24 ) {
if (nq-i == 24 )
{
hh_trafo_complex_kernel_24_AVX512_2hv_single(&q[i], hh, nb, ldq, ldh, s);
worked_on += 24;
}
#endif
#ifdef DOUBLE_PRECISION_COMPLEX
if (nq-i == 8 ) {
if (nq-i == 8 )
{
hh_trafo_complex_kernel_8_AVX512_2hv_double(&q[i], hh, nb, ldq, ldh, s);
worked_on += 8;
}
}
#endif
#ifdef SINGLE_PRECISION_COMPLEX
if (nq-i == 16 ) {
if (nq-i == 16 )
{
hh_trafo_complex_kernel_16_AVX512_2hv_single(&q[i], hh, nb, ldq, ldh, s);
worked_on += 16;
}
......@@ -235,7 +240,8 @@ void double_hh_trafo_complex_avx512_2hv_single(float complex* q, float complex*
if (worked_on != nq)
{
printf("Error in complex AVX512 BLOCK 2 kernel \n");
// printf("Error in complex AVX512 BLOCK 2 kernel \n");
// abort();
}
}
......
src/elpa2/kernels/real_avx512_2hv_template.Xc
View file @ ee4e6cab
......@@ -48,6 +48,8 @@
#include "config-f90.h"
#include <x86intrin.h>
#include <stdio.h>
#include <stdlib.h>
#define __forceinline __attribute__((always_inline)) static
......@@ -151,6 +153,9 @@ void double_hh_trafo_real_avx512_2hv_single(float* q, float* hh, int* pnb, int*
int nq = *pldq;
int ldq = *pldq;
int ldh = *pldh;
int worked_on;
worked_on = 0;
// calculating scalar product to compute
// 2 householder vectors simultaneously
......@@ -171,12 +176,14 @@ void double_hh_trafo_real_avx512_2hv_single(float* q, float* hh, int* pnb, int*
for (i = 0; i < nq-24; i+=32)
{
hh_trafo_kernel_32_AVX512_2hv_double(&q[i], hh, nb, ldq, ldh, s);
worked_on += i;
}
#endif
#ifdef SINGLE_PRECISION_REAL
for (i = 0; i < nq-48; i+=64)
{
hh_trafo_kernel_64_AVX512_2hv_single(&q[i], hh, nb, ldq, ldh, s);
worked_on += i;
}
#endif
if (nq == i)
......@@ -184,40 +191,52 @@ void double_hh_trafo_real_avx512_2hv_single(float* q, float* hh, int* pnb, int*
return;
}
#ifdef DOUBLE_PRECISION_REAL
if (nq-i == 24)
if (nq-i == 24)
{
hh_trafo_kernel_24_AVX512_2hv_double(&q[i], hh, nb, ldq, ldh, s);
worked_on += 24;
}
#endif
#ifdef SINGLE_PRECISION_REAL
if (nq-i == 48)
{
hh_trafo_kernel_48_AVX512_2hv_single(&q[i], hh, nb, ldq, ldh, s);
worked_on += 48;
}
#endif
#ifdef DOUBLE_PRECISION_REAL
else if (nq-i == 16)
if (nq-i == 16)
{
hh_trafo_kernel_16_AVX512_2hv_double(&q[i], hh, nb, ldq, ldh, s);
worked_on += 16;
}
#endif
#ifdef SINGLE_PRECISION_REAL
else if (nq-i == 32)
if (nq-i == 32)
{
hh_trafo_kernel_32_AVX512_2hv_single(&q[i], hh, nb, ldq, ldh, s);
worked_on += 32;
}
#endif
else
#ifdef DOUBLE_PRECISION_REAL
if (nq-i == 8)
{
hh_trafo_kernel_8_AVX512_2hv_double(&q[i], hh, nb, ldq, ldh, s);
worked_on += 8;
}
#endif
#ifdef SINGLE_PRECISION_REAL
if (nq-i == 16)
{
hh_trafo_kernel_16_AVX512_2hv_single(&q[i], hh, nb, ldq, ldh, s);
worked_on += 16;
}
#endif
if (worked_on != nq)
{
// printf("Error in AVX512 real BLOCK 2 kernel \n");
// abort();
}
}
/**
......
src/elpa2/kernels/real_avx512_4hv_template.Xc
View file @ ee4e6cab
......@@ -46,7 +46,8 @@
// --------------------------------------------------------------------------------------------------
#include "config-f90.h"
#include <x86intrin.h>
#include <stdio.h>
#include <stdlib.h>
#define __forceinline __attribute__((always_inline)) static
......@@ -104,10 +105,10 @@ void quad_hh_trafo_real_avx512_4hv_double(double* q, double* hh, int* pnb, int*
#endif
#ifdef SINGLE_PRECISION_REAL
//Forward declaration
__forceinline void hh_trafo_kernel_16_AVX_4hv_single(float* q, float* hh, int nb, int ldq, int ldh, float s_1_2, float s_1_3, float s_2_3, float s_1_4, float s_2_4, float s_3_4);
__forceinline void hh_trafo_kernel_32_AVX_4hv_single(float* q, float* hh, int nb, int ldq, int ldh, float s_1_2, float s_1_3, float s_2_3, float s_1_4, float s_2_4, float s_3_4);
__forceinline void hh_trafo_kernel_48_AVX_4hv_single(float* q, float* hh, int nb, int ldq, int ldh, float s_1_2, float s_1_3, float s_2_3, float s_1_4, float s_2_4, float s_3_4);
__forceinline void hh_trafo_kernel_16_AVX512_4hv_single(float* q, float* hh, int nb, int ldq, int ldh, float s_1_2, float s_1_3, float s_2_3, float s_1_4, float s_2_4, float s_3_4);
__forceinline void hh_trafo_kernel_32_AVX512_4hv_single(float* q, float* hh, int nb, int ldq, int ldh, float s_1_2, float s_1_3, float s_2_3, float s_1_4, float s_2_4, float s_3_4);
__forceinline void hh_trafo_kernel_48_AVX512_4hv_single(float* q, float* hh, int nb, int ldq, int ldh, float s_1_2, float s_1_3, float s_2_3, float s_1_4, float s_2_4, float s_3_4);
__forceinline void hh_trafo_kernel_64_AVX512_4hv_single(float* q, float* hh, int nb, int ldq, int ldh, float s_1_2, float s_1_3, float s_2_3, float s_1_4, float s_2_4, float s_3_4);
void quad_hh_trafo_real_avx_avx2_4hv_single(float* q, float* hh, int* pnb, int* pnq, int* pldq, int* pldh);
#endif
......@@ -157,6 +158,9 @@ void quad_hh_trafo_real_avx512_4hv_single(float* q, float* hh, int* pnb, int* pn
int nq = *pldq;
int ldq = *pldq;
int ldh = *pldh;
int worked_on;
worked_on = 0;
// calculating scalar products to compute
// 4 householder vectors simultaneously
......@@ -210,12 +214,14 @@ void quad_hh_trafo_real_avx512_4hv_single(float* q, float* hh, int* pnb, int* pn
for (i = 0; i < nq-24; i+=32)
{
hh_trafo_kernel_32_AVX512_4hv_double(&q[i], hh, nb, ldq, ldh, s_1_2, s_1_3, s_2_3, s_1_4, s_2_4, s_3_4);
worked_on += i;
}
#endif
#ifdef SINGLE_PRECISION_REAL
for (i = 0; i < nq-32; i+=48)
for (i = 0; i < nq-48; i+=64)
{
hh_trafo_kernel_48_AVX_4hv_single(&q[i], hh, nb, ldq, ldh, s_1_2, s_1_3, s_2_3, s_1_4, s_2_4, s_3_4);
hh_trafo_kernel_64_AVX512_4hv_single(&q[i], hh, nb, ldq, ldh, s_1_2, s_1_3, s_2_3, s_1_4, s_2_4, s_3_4);
worked_on += i;
}
#endif
if (nq == i)
......@@ -226,41 +232,77 @@ void quad_hh_trafo_real_avx512_4hv_single(float* q, float* hh, int* pnb, int* pn
if (nq-i == 24)
{
hh_trafo_kernel_24_AVX512_4hv_double(&q[i], hh, nb, ldq, ldh, s_1_2, s_1_3, s_2_3, s_1_4, s_2_4, s_3_4);
worked_on += 24;
}
#endif
#ifdef SINGLE_PRECISION_REAL
if (nq-i == 48)
{
hh_trafo_kernel_48_AVX512_4hv_single(&q[i], hh, nb, ldq, ldh, s_1_2, s_1_3, s_2_3, s_1_4, s_2_4, s_3_4);
worked_on += 48;
}
#endif
else
{
if (nq-i == 16)
{
#ifdef DOUBLE_PRECISION_REAL
if (nq-i == 16)
{
hh_trafo_kernel_16_AVX512_4hv_double(&q[i], hh, nb, ldq, ldh, s_1_2, s_1_3, s_2_3, s_1_4, s_2_4, s_3_4);
}
else
{
hh_trafo_kernel_8_AVX512_4hv_double(&q[i], hh, nb, ldq, ldh, s_1_2, s_1_3, s_2_3, s_1_4, s_2_4, s_3_4);
}
}
worked_on += 16;
}
#endif
#ifdef SINGLE_PRECISION_REAL
if (nq-i == 32)
if (nq-i == 32)
{
hh_trafo_kernel_32_AVX_4hv_single(&q[i], hh, nb, ldq, ldh, s_1_2, s_1_3, s_2_3, s_1_4, s_2_4, s_3_4);
}
else
hh_trafo_kernel_32_AVX512_4hv_single(&q[i], hh, nb, ldq, ldh, s_1_2, s_1_3, s_2_3, s_1_4, s_2_4, s_3_4);
worked_on += 32;
}
#endif
#ifdef DOUBLE_PRECISION_REAL
if (nq-i == 8)
{
hh_trafo_kernel_16_AVX_4hv_single(&q[i], hh, nb, ldq, ldh, s_1_2, s_1_3, s_2_3, s_1_4, s_2_4, s_3_4);
hh_trafo_kernel_8_AVX512_4hv_double(&q[i], hh, nb, ldq, ldh, s_1_2, s_1_3, s_2_3, s_1_4, s_2_4, s_3_4);
worked_on += 8;
}
#endif
#ifdef SINGLE_PRECISION_REAL
if (nq-i == 16)
{
hh_trafo_kernel_16_AVX512_4hv_single(&q[i], hh, nb, ldq, ldh, s_1_2, s_1_3, s_2_3, s_1_4, s_2_4, s_3_4);
worked_on += 16;
}
#endif
if (worked_on != nq)
{
// printf("Error in AVX512 real BLOCK 2 kernel \n");
// abort();
}
}
#ifdef DOUBLE_PRECISION_REAL
/**
* Unrolled kernel that computes
#ifdef DOUBLE_PRECISION_REAL
* 32 rows of Q simultaneously, a
#endif
#ifdef SINGLE_PRECISION_REAL
* 64 rows of Q simultaneously, a
#endif
* matrix Vector product with two householder
* vectors + a rank 1 update is performed
*/
#ifdef DOUBLE_PRECISION_REAL
__forceinline void hh_trafo_kernel_32_AVX512_4hv_double(double* q, double* hh, int nb, int ldq, int ldh, double s_1_2, double s_1_3, double s_2_3, double s_1_4, double s_2_4, double s_3_4)
#endif
#ifdef SINGLE_PRECISION_REAL
__forceinline void hh_trafo_kernel_64_AVX512_4hv_single(float* q, float* hh, int nb, int ldq, int ldh, float s_1_2, float s_1_3, float s_2_3, float s_1_4, float s_2_4, float s_3_4)
#endif
{
/////////////////////////////////////////////////////
// Matrix Vector Multiplication, Q [4 x nb+3] * hh
......@@ -691,7 +733,6 @@ __forceinline void hh_trafo_kernel_32_AVX512_4hv_double(double* q, double* hh, i
_AVX512_STORE(&q[((nb+2)*ldq)+3*offset],q4);
}
#endif /* DOUBLE_PRECISION_REAL */
/**
* Unrolled kernel that computes
......@@ -708,7 +749,7 @@ __forceinline void hh_trafo_kernel_32_AVX512_4hv_double(double* q, double* hh, i
__forceinline void hh_trafo_kernel_24_AVX512_4hv_double(double* q, double* hh, int nb, int ldq, int ldh, double s_1_2, double s_1_3, double s_2_3, double s_1_4, double s_2_4, double s_3_4)
#endif
#ifdef SINGLE_PRECISION_REAL
__forceinline void hh_trafo_kernel_48_AVX_4hv_single(float* q, float* hh, int nb, int ldq, int ldh, float s_1_2, float s_1_3, float s_2_3, float s_1_4, float s_2_4, float s_3_4)
__forceinline void hh_trafo_kernel_48_AVX512_4hv_single(float* q, float* hh, int nb, int ldq, int ldh, float s_1_2, float s_1_3, float s_2_3, float s_1_4, float s_2_4, float s_3_4)
#endif
{
/////////////////////////////////////////////////////
......@@ -1089,7 +1130,7 @@ __forceinline void hh_trafo_kernel_48_AVX_4hv_single(float* q, float* hh, int nb
__forceinline void hh_trafo_kernel_16_AVX512_4hv_double(double* q, double* hh, int nb, int ldq, int ldh, double s_1_2, double s_1_3, double s_2_3, double s_1_4, double s_2_4, double s_3_4)
#endif
#ifdef SINGLE_PRECISION_REAL
__forceinline void hh_trafo_kernel_32_AVX_4hv_single(float* q, float* hh, int nb, int ldq, int ldh, float s_1_2, float s_1_3, float s_2_3, float s_1_4, float s_2_4, float s_3_4)
__forceinline void hh_trafo_kernel_32_AVX512_4hv_single(float* q, float* hh, int nb, int ldq, int ldh, float s_1_2, float s_1_3, float s_2_3, float s_1_4, float s_2_4, float s_3_4)
#endif
{
/////////////////////////////////////////////////////
......@@ -1380,7 +1421,7 @@ __forceinline void hh_trafo_kernel_32_AVX_4hv_single(float* q, float* hh, int nb
__forceinline void hh_trafo_kernel_8_AVX512_4hv_double(double* q, double* hh, int nb, int ldq, int ldh, double s_1_2, double s_1_3, double s_2_3, double s_1_4, double s_2_4, double s_3_4)
#endif
#ifdef SINGLE_PRECISION_REAL
__forceinline void hh_trafo_kernel_16_AVX_4hv_single(float* q, float* hh, int nb, int ldq, int ldh, float s_1_2, float s_1_3, float s_2_3, float s_1_4, float s_2_4, float s_3_4)
__forceinline void hh_trafo_kernel_16_AVX512_4hv_single(float* q, float* hh, int nb, int ldq, int ldh, float s_1_2, float s_1_3, float s_2_3, float s_1_4, float s_2_4, float s_3_4)
#endif
{
/////////////////////////////////////////////////////
......
src/elpa2/kernels/real_avx512_6hv_template.Xc
View file @ ee4e6cab
......@@ -48,6 +48,7 @@
#include "config-f90.h"
#include <x86intrin.h>
#include <stdio.h>
#include <stdlib.h>
#define __forceinline __attribute__((always_inline)) static
......@@ -113,6 +114,8 @@ void hexa_hh_trafo_real_avx512_6hv_double(double* q, double* hh, int* pnb, int*
#ifdef SINGLE_PRECISION_REAL
static void hh_trafo_kernel_16_AVX512_6hv_single(float* q, float* hh, int nb, int ldq, int ldh, float* scalarprods);
static void hh_trafo_kernel_32_AVX512_6hv_single(float* q, float* hh, int nb, int ldq, int ldh, float* scalarprods);
static void hh_trafo_kernel_48_AVX512_6hv_single(float* q, float* hh, int nb, int ldq, int ldh, float* scalarprods);
static void hh_trafo_kernel_64_AVX512_6hv_single(float* q, float* hh, int nb, int ldq, int ldh, float* scalarprods);
void hexa_hh_trafo_real_avx512_6hv_single_(float* q, float* hh, int* pnb, int* pnq, int* pldq, int* pldh);
......@@ -162,6 +165,9 @@ void hexa_hh_trafo_real_avx512_6hv_single(float* q, float* hh, int* pnb, int* pn
int nq = *pldq;
int ldq = *pldq;
int ldh = *pldh;
int worked_on;
worked_on = 0;
// calculating scalar products to compute
// 6 householder vectors simultaneously
......@@ -273,12 +279,14 @@ void hexa_hh_trafo_real_avx512_6hv_single(float* q, float* hh, int* pnb, int* pn
for (i = 0; i < nq-24; i+=32)
{
hh_trafo_kernel_32_AVX512_6hv_double(&q[i], hh, nb, ldq, ldh, scalarprods);
worked_on += i;
}
#endif
#ifdef SINGLE_PRECISION_REAL
for (i = 0; i < nq-16; i+=32)
for (i = 0; i < nq-48; i+=64)
{
hh_trafo_kernel_32_AVX512_6hv_single(&q[i], hh, nb, ldq, ldh, scalarprods);
hh_trafo_kernel_64_AVX512_6hv_single(&q[i], hh, nb, ldq, ldh, scalarprods);
worked_on += i;
}
#endif
if (nq == i)
......@@ -289,26 +297,54 @@ void hexa_hh_trafo_real_avx512_6hv_single(float* q, float* hh, int* pnb, int* pn
if (nq-i == 24)
{
hh_trafo_kernel_24_AVX512_6hv_double(&q[i], hh, nb, ldq, ldh, scalarprods);
worked_on += 24;
}
#endif
#ifdef SINGLE_PRECISION_REAL
if (nq-i ==48)
{
hh_trafo_kernel_48_AVX512_6hv_single(&q[i], hh, nb, ldq, ldh, scalarprods);
worked_on += 48;
}
else
{
if (nq-i == 16)
{
#endif
#ifdef DOUBLE_PRECISION_REAL
if (nq-i == 16)
{
hh_trafo_kernel_16_AVX512_6hv_double(&q[i], hh, nb, ldq, ldh, scalarprods);
}
else
{
hh_trafo_kernel_8_AVX512_6hv_double(&q[i], hh, nb, ldq, ldh, scalarprods);
}
worked_on += 16;
}
#endif
#ifdef SINGLE_PRECISION_REAL
if (nq-i ==32)
{
hh_trafo_kernel_32_AVX512_6hv_single(&q[i], hh, nb, ldq, ldh, scalarprods);
worked_on += 32;
}
#endif
#ifdef DOUBLE_PRECISION_REAL
if (nq-i == 8)
{
hh_trafo_kernel_8_AVX512_6hv_double(&q[i], hh, nb, ldq, ldh, scalarprods);
worked_on += 8;
}
#endif
#ifdef SINGLE_PRECISION_REAL
if (nq-i == 16) {
if (nq-i == 16)
{
hh_trafo_kernel_16_AVX512_6hv_single(&q[i], hh, nb, ldq, ldh, scalarprods);
} else {
printf("ERROR in avx512 kernel\n");
}
worked_on += 16;
}
#endif
if (worked_on != nq)
{
// printf("ERROR in avx512 kernel\n");
// abort();
}
}
/**
......@@ -319,7 +355,6 @@ void hexa_hh_trafo_real_avx512_6hv_single(float* q, float* hh, int* pnb, int* pn
#ifdef SINGLE_PRECISION_REAL
* 16 rows of Q simultaneously, a
#endif
* matrix Vector product with two householder
* vectors + a rank 1 update is performed
*/
......@@ -1349,14 +1384,24 @@ __forceinline void hh_trafo_kernel_32_AVX512_6hv_single(float* q, float* hh, int
}
#ifdef DOUBLE_PRECISION_REAL
/**
* Unrolled kernel that computes
#ifdef DOUBLE_PRECISION_REAL
* 24 rows of Q simultaneously, a
#endif
#ifdef DOUBLE_PRECISION_REAL
* 48 rows of Q simultaneously, a
#endif
* matrix Vector product with two householder
* vectors + a rank 1 update is performed
*/
#ifdef DOUBLE_PRECISION_REAL
__forceinline void hh_trafo_kernel_24_AVX512_6hv_double(double* q, double* hh, int nb, int ldq, int ldh, double* scalarprods)
#endif
#ifdef SINGLE_PRECISION_REAL
__forceinline void hh_trafo_kernel_48_AVX512_6hv_single(float* q, float* hh, int nb, int ldq, int ldh, float* scalarprods)
#endif
{
/////////////////////////////////////////////////////
// Matrix Vector Multiplication, Q [8 x nb+3] * hh
......@@ -2074,20 +2119,24 @@ __forceinline void hh_trafo_kernel_24_AVX512_6hv_double(double* q, double* hh, i
_AVX512_STORE(&q[((nb+4)*ldq)+2*offset],q3);
}
#endif /* DOUBLE_PRECISION_REAL */
#ifdef DOUBLE_PRECISION_REAL
/**
* Unrolled kernel that computes
#ifdef DOUBLE_PRECISION_REAL
* 32 rows of Q simultaneously, a
#endif
#ifdef SINGLE_PRECISION_REAL
* 64 rows of Q simultaneously, a
#endif
* matrix Vector product with two householder
* vectors + a rank 1 update is performed
*/
#ifdef DOUBLE_PRECISION_REAL
__forceinline void hh_trafo_kernel_32_AVX512_6hv_double(double* q, double* hh, int nb, int ldq, int ldh, double* scalarprods)
#endif
#ifdef SINGLE_PRECISION_REAL
__forceinline void hh_trafo_kernel_64_AVX512_6hv_single(float* q, float* hh, int nb, int ldq, int ldh, float* scalarprods)
#endif
{
/////////////////////////////////////////////////////
// Matrix Vector Multiplication, Q [8 x nb+3] * hh
......@@ -2947,5 +2996,4 @@ __forceinline void hh_trafo_kernel_32_AVX512_6hv_double(double* q, double* hh, i
_AVX512_STORE(&q[((nb+4)*ldq)+3*offset],q4);
}
#endif /* DOUBLE_PRECISION_REAL */
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment