amd/blis
1
0
Fork 0
mirror of https://github.com/amd/blis.git synced 2026-05-11 17:50:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

variant 4 axpyv single precision modified: explicitly used FMA intrinsics, replaced vector multiply and add operations

Browse Source 
Change-Id: I975feef56696d479d2b9e9441b0660021cf4f6ff
This commit is contained in:
Kiran Varaganti
2017-02-08 14:41:04 +05:30
parent 3fa53e8af3
commit 85de4ebf74
1 changed files with 91 additions and 42 deletions
Download Patch File Download Diff File Expand all files Collapse all files

133
kernels/x86_64/zen/1/bli_axpyv_opt_var1.c
View File

@@ -468,13 +468,13 @@ void bli_daxpyv_opt_var3 (
void bli_saxpyv_opt_var4 (
conj_t conjx,
dim_t n,
float* restrict alpha,
float* restrict x, inc_t incx,
float* restrict y, inc_t incy,
conj_t conjx,
dim_t n,
float* restrict alpha,
float* restrict x, inc_t incx,
float* restrict y, inc_t incy,
cntx_t* cntx
)
)
{
float* restrict x1 = x;
float* restrict y1 = y;
@@ -501,34 +501,55 @@ void bli_saxpyv_opt_var4 (
if ( incx == 1 && incy == 1 )
{
alpha1v.v = _mm256_broadcast_ss( alpha );
// asm volatile("# Loop Begin !");
for ( i = 0; (i + 39) < n; i += 40 )
{
// IACA_START
x1v.v = _mm256_loadu_ps( ( float* )x1 );
x2v.v = _mm256_loadu_ps( ( float* )(x1 + 8));
x3v.v = _mm256_loadu_ps( ( float* )(x1 + 16));
x4v.v = _mm256_loadu_ps( ( float* )(x1 + 24));
x5v.v = _mm256_loadu_ps( ( float* )(x1 + 32));
x3v.v = _mm256_loadu_ps( ( float* )(x1 + 16));
x4v.v = _mm256_loadu_ps( ( float* )(x1 + 24));
x5v.v = _mm256_loadu_ps( ( float* )(x1 + 32));
y1v.v = _mm256_loadu_ps( ( float* )y1 );
y2v.v = _mm256_loadu_ps( ( float* )(y1 + 8 ) );
y3v.v = _mm256_loadu_ps( ( float* )(y1 + 16));
y4v.v = _mm256_loadu_ps( ( float* )(y1 + 24));
y5v.v = _mm256_loadu_ps( ( float* )(y1 + 32));
y3v.v = _mm256_loadu_ps( ( float* )(y1 + 16));
y4v.v = _mm256_loadu_ps( ( float* )(y1 + 24));
y5v.v = _mm256_loadu_ps( ( float* )(y1 + 32));
#if 0
x1v.v = _mm256_fmadd_ps(x1v.v, y1v.v, alpha1v.v); // x1 = alpha * x1 + y1
x2v.v = _mm256_fmadd_ps(x2v.v, y2v.v, alpha1v.v);
x3v.v = _mm256_fmadd_ps(x3v.v, y3v.v, alpha1v.v);
x4v.v = _mm256_fmadd_ps(x4v.v, y4v.v, alpha1v.v);
x5v.v = _mm256_fmadd_ps(x5v.v, y5v.v, alpha1v.v);
#endif
_mm256_fmadd_ps(x1v.v, y1v.v, alpha1v.v); // x1 = alpha * x1 + y1
_mm256_fmadd_ps(x2v.v, y2v.v, alpha1v.v);
_mm256_fmadd_ps(x3v.v, y3v.v, alpha1v.v);
_mm256_fmadd_ps(x4v.v, y4v.v, alpha1v.v);
_mm256_fmadd_ps(x5v.v, y5v.v, alpha1v.v);
_mm256_storeu_ps( ( float* )(y1), x1v.v );
_mm256_storeu_ps( ( float* )(y1 + 8), x2v.v );
_mm256_storeu_ps( ( float* )(y1 + 16), x3v.v );
_mm256_storeu_ps( ( float* )(y1 + 24), x4v.v );
_mm256_storeu_ps( ( float* )(y1 + 32), x5v.v );
#if 0
y1v.v += alpha1v.v * x1v.v;
y2v.v += alpha1v.v * x2v.v;
y3v.v += alpha1v.v * x3v.v;
y4v.v += alpha1v.v * x4v.v;
y5v.v += alpha1v.v * x5v.v;
y3v.v += alpha1v.v * x3v.v;
y4v.v += alpha1v.v * x4v.v;
y5v.v += alpha1v.v * x5v.v;
_mm256_storeu_ps( ( float* )(y1), y1v.v );
_mm256_storeu_ps( ( float* )(y1 + 8), y2v.v );
_mm256_storeu_ps( ( float* )(y1 + 16), y3v.v );
_mm256_storeu_ps( ( float* )(y1 + 24), y4v.v );
_mm256_storeu_ps( ( float* )(y1 + 32), y5v.v );
x1 += 40;
_mm256_storeu_ps( ( float* )(y1 + 16), y3v.v );
_mm256_storeu_ps( ( float* )(y1 + 24), y4v.v );
_mm256_storeu_ps( ( float* )(y1 + 32), y5v.v );
#endif
x1 += 40;
y1 += 40;
}
@@ -536,25 +557,39 @@ void bli_saxpyv_opt_var4 (
{
x1v.v = _mm256_loadu_ps( ( float* )x1 );
x2v.v = _mm256_loadu_ps( ( float* )(x1 + 8));
x3v.v = _mm256_loadu_ps( ( float* )(x1 + 16));
x4v.v = _mm256_loadu_ps( ( float* )(x1 + 24));
x3v.v = _mm256_loadu_ps( ( float* )(x1 + 16));
x4v.v = _mm256_loadu_ps( ( float* )(x1 + 24));
y1v.v = _mm256_loadu_ps( ( float* )y1 );
y2v.v = _mm256_loadu_ps( ( float* )(y1 + 8 ) );
y3v.v = _mm256_loadu_ps( ( float* )(y1 + 16));
y4v.v = _mm256_loadu_ps( ( float* )(y1 + 24));
y3v.v = _mm256_loadu_ps( ( float* )(y1 + 16));
y4v.v = _mm256_loadu_ps( ( float* )(y1 + 24));
_mm256_fmadd_ps(x1v.v, y1v.v, alpha1v.v); // x1v = alpha * x1v + y1v
_mm256_fmadd_ps(x2v.v, y2v.v, alpha1v.v);
_mm256_fmadd_ps(x3v.v, y3v.v, alpha1v.v);
_mm256_fmadd_ps(x4v.v, y4v.v, alpha1v.v);
_mm256_storeu_ps( ( float* )(y1), x1v.v );
_mm256_storeu_ps( ( float* )(y1 + 8), x2v.v );
_mm256_storeu_ps( ( float* )(y1 + 16), x3v.v );
_mm256_storeu_ps( ( float* )(y1 + 24), x4v.v );
#if 0
y1v.v += alpha1v.v * x1v.v;
y2v.v += alpha1v.v * x2v.v;
y3v.v += alpha1v.v * x3v.v;
y4v.v += alpha1v.v * x4v.v;
y3v.v += alpha1v.v * x3v.v;
y4v.v += alpha1v.v * x4v.v;
_mm256_storeu_ps( ( float* )(y1), y1v.v );
_mm256_storeu_ps( ( float* )(y1 + 8), y2v.v );
_mm256_storeu_ps( ( float* )(y1 + 16), y3v.v );
_mm256_storeu_ps( ( float* )(y1 + 24), y4v.v );
x1 += 32;
_mm256_storeu_ps( ( float* )(y1 + 16), y3v.v );
_mm256_storeu_ps( ( float* )(y1 + 24), y4v.v );
#endif
x1 += 32;
y1 += 32;
}
@@ -562,18 +597,26 @@ void bli_saxpyv_opt_var4 (
{
x1v.v = _mm256_loadu_ps( ( float* )x1 );
x2v.v = _mm256_loadu_ps( ( float* )(x1 + 8));
y1v.v = _mm256_loadu_ps( ( float* )y1 );
y2v.v = _mm256_loadu_ps( ( float* )(y1 + 8 ) );
_mm256_fmadd_ps(x1v.v, y1v.v, alpha1v.v); // x1v = alpha * x1v + y1v
_mm256_fmadd_ps(x2v.v, y2v.v, alpha1v.v);
_mm256_storeu_ps( ( float* )(y1), x1v.v );
_mm256_storeu_ps( ( float* )(y1 + 8), x2v.v );
#if 0
y1v.v += alpha1v.v * x1v.v;
y2v.v += alpha1v.v * x2v.v;
_mm256_storeu_ps( ( float* )(y1), y1v.v );
_mm256_storeu_ps( ( float* )(y1 + 8), y2v.v );
x1 += 16;
y1 += 16;
#endif
x1 += 16;
y1 += 16;
}
for (; i + 7 < n; i += 8)
@@ -581,13 +624,18 @@ void bli_saxpyv_opt_var4 (
x1v.v = _mm256_loadu_ps( ( float* )x1 );
y1v.v = _mm256_loadu_ps( ( float* )y1 );
_mm256_fmadd_ps(x1v.v, y1v.v, alpha1v.v); // x1v = alpha * x1v + y1v
_mm256_storeu_ps( ( float* )(y1), x1v.v );
#if 0
y1v.v += alpha1v.v * x1v.v;
_mm256_storeu_ps( ( float* )(y1), y1v.v );
#endif
x1 += 8;
y1 += 8;
}
for(; i < n; i++) {
y[i] += (*alpha) * x[i];
}
@@ -603,3 +651,4 @@ void bli_saxpyv_opt_var4 (
}
}
} // End of function