floraachy
/
OpenBLAS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

replace the hasell dgemv_kernel_4x4 kernel with a the same code written in intrinsics 

using intrinsics is a bit easier to read (at least for the non-math part of the code)
and also allows the compiler to be better about register allocation and optimizing the
non-math (loop/setup) code.

It also allows the code to honor the "no fma" flag if the user so desires.

The result of this change is (measured for a size of 16) a 15% performance increase.

And it is a step towards being able to add an AVX512 version of the code.
This commit is contained in:
Arjan van de Ven 2018-08-04 17:25:54 +00:00
parent 350531e76a
commit 4a8ae8b8aa
1 changed files with 25 additions and 71 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

96
kernel/x86_64/dgemv_n_microk_haswell-4.c
View File

@ -28,88 +28,42 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define HAVE_KERNEL_4x4 1 #define HAVE_KERNEL_4x4 1
static void dgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha) __attribute__ ((noinline));
#include <immintrin.h>
static void dgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha) static void dgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
{ {
BLASLONG register i = 0; int i = 0;
__asm__ __volatile__ __m256d x0, x1, x2, x3;
( __m256d __alpha;
"vbroadcastsd (%[x]), %%ymm12 \n\t" // x0
"vbroadcastsd 8(%[x]), %%ymm13 \n\t" // x1
"vbroadcastsd 16(%[x]), %%ymm14 \n\t" // x2
"vbroadcastsd 24(%[x]), %%ymm15 \n\t" // x3
"vmovups (%[ap0],%[i],8), %%ymm0 \n\t" x0 = _mm256_broadcastsd_pd(_mm_load_sd(&x[0]));
"vmovups (%[ap1],%[i],8), %%ymm1 \n\t" x1 = _mm256_broadcastsd_pd(_mm_load_sd(&x[1]));
"vmovups (%[ap2],%[i],8), %%ymm2 \n\t" x2 = _mm256_broadcastsd_pd(_mm_load_sd(&x[2]));
"vmovups (%[ap3],%[i],8), %%ymm3 \n\t" x3 = _mm256_broadcastsd_pd(_mm_load_sd(&x[3]));
"vbroadcastsd (%[alpha]), %%ymm6 \n\t" // alpha
"addq $4 , %[i] \n\t" __alpha = _mm256_broadcastsd_pd(_mm_load_sd(alpha));
"subq $4 , %[n] \n\t"
"jz 2f \n\t"
// ".align 16 \n\t"
"1: \n\t"
"vmulpd %%ymm0 , %%ymm12, %%ymm4 \n\t"
"vmulpd %%ymm1 , %%ymm13, %%ymm5 \n\t"
"vmovups (%[ap0],%[i],8), %%ymm0 \n\t"
"vmovups (%[ap1],%[i],8), %%ymm1 \n\t"
"vfmadd231pd %%ymm2 , %%ymm14, %%ymm4 \n\t"
"vfmadd231pd %%ymm3 , %%ymm15, %%ymm5 \n\t"
"vmovups (%[ap2],%[i],8), %%ymm2 \n\t"
"vmovups (%[ap3],%[i],8), %%ymm3 \n\t"
"vmovups -32(%[y],%[i],8), %%ymm8 \n\t" // 4 * y
"vaddpd %%ymm4 , %%ymm5 , %%ymm4 \n\t"
"vfmadd231pd %%ymm6 , %%ymm4 , %%ymm8 \n\t"
"vmovups %%ymm8, -32(%[y],%[i],8) \n\t" // 4 * y
"addq $4 , %[i] \n\t"
"subq $4 , %[n] \n\t"
"jnz 1b \n\t"
"2: \n\t" for (i = 0; i < n; i+= 4) {
__m256d tempY;
__m256d sum;
"vmulpd %%ymm0 , %%ymm12, %%ymm4 \n\t" sum = _mm256_add_pd(
"vmulpd %%ymm1 , %%ymm13, %%ymm5 \n\t" _mm256_add_pd(
"vfmadd231pd %%ymm2 , %%ymm14, %%ymm4 \n\t" _mm256_mul_pd(_mm256_loadu_pd(&ap[0][i]), x0),
"vfmadd231pd %%ymm3 , %%ymm15, %%ymm5 \n\t" _mm256_mul_pd(_mm256_loadu_pd(&ap[1][i]), x1)),
_mm256_add_pd(
_mm256_mul_pd(_mm256_loadu_pd(&ap[2][i]), x2),
_mm256_mul_pd(_mm256_loadu_pd(&ap[3][i]), x3))
);
tempY = _mm256_loadu_pd(&y[i]);
"vmovups -32(%[y],%[i],8), %%ymm8 \n\t" // 4 * y tempY = _mm256_add_pd(tempY, _mm256_mul_pd(sum, __alpha));
"vaddpd %%ymm4 , %%ymm5 , %%ymm4 \n\t" _mm256_storeu_pd(&y[i], tempY);
"vfmadd231pd %%ymm6 , %%ymm4 , %%ymm8 \n\t" }
"vmovups %%ymm8, -32(%[y],%[i],8) \n\t" // 4 * y
"vzeroupper \n\t"
:
[i] "+r" (i), // 0
[n] "+r" (n) // 1
:
[x] "r" (x), // 2
[y] "r" (y), // 3
[ap0] "r" (ap[0]), // 4
[ap1] "r" (ap[1]), // 5
[ap2] "r" (ap[2]), // 6
[ap3] "r" (ap[3]), // 7
[alpha] "r" (alpha) // 8
: "cc",
"%xmm4", "%xmm5",
"%xmm6", "%xmm7",
"%xmm8", "%xmm9",
"%xmm12", "%xmm13", "%xmm14", "%xmm15",
"memory"
);
} }