OpenMathLib/OpenBLAS
1
0
Fork 0
mirror of https://github.com/OpenMathLib/OpenBLAS synced 2026-05-31 00:45:48 +08:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity

Merge pull request #5125 from martin-frbg/issue5122

Browse Source 
Fix SGEMV on POWER8 by reverting to the non-vectorized earlier code
This commit is contained in:
Martin Kroeker
2025-02-12 12:50:44 +01:00
committed by GitHub
parent 9a3948df82 81eed868b6
commit e8b11a126b
2 changed files with 39 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
kernel/power/sgemv_t.c
View File

@@ -78,7 +78,17 @@ static void sgemv_kernel_4x8(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOA
temp7 += v_x[i] * va7[i];
}
#if defined(POWER8)
y[0] += alpha * (temp0[0] + temp0[1]+temp0[2] + temp0[3]);
y[1] += alpha * (temp1[0] + temp1[1]+temp1[2] + temp1[3]);
y[2] += alpha * (temp2[0] + temp2[1]+temp2[2] + temp2[3]);
y[3] += alpha * (temp3[0] + temp3[1]+temp3[2] + temp3[3]);
y[4] += alpha * (temp4[0] + temp4[1]+temp4[2] + temp4[3]);
y[5] += alpha * (temp5[0] + temp5[1]+temp5[2] + temp5[3]);
y[6] += alpha * (temp6[0] + temp6[1]+temp6[2] + temp6[3]);
y[7] += alpha * (temp7[0] + temp7[1]+temp7[2] + temp7[3]);
#else
register __vector float t0, t1, t2, t3;
register __vector float a = { alpha, alpha, alpha, alpha };
__vector float *v_y = (__vector float*) y;
@@ -105,7 +115,7 @@ static void sgemv_kernel_4x8(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOA
v_y[0] += a * temp0;
v_y[1] += a * temp4;
#endif
}
@@ -132,7 +142,12 @@ static void sgemv_kernel_4x4(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOA
temp2 += v_x[i] * va2[i];
temp3 += v_x[i] * va3[i];
}
#if defined(POWER8)
y[0] += alpha * (temp0[0] + temp0[1]+temp0[2] + temp0[3]);
y[1] += alpha * (temp1[0] + temp1[1]+temp1[2] + temp1[3]);
y[2] += alpha * (temp2[0] + temp2[1]+temp2[2] + temp2[3]);
y[3] += alpha * (temp3[0] + temp3[1]+temp3[2] + temp3[3]);
#else
register __vector float t0, t1, t2, t3;
register __vector float a = { alpha, alpha, alpha, alpha };
__vector float *v_y = (__vector float*) y;
@@ -148,7 +163,7 @@ static void sgemv_kernel_4x4(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOA
temp0 += temp1 + temp2 + temp3;
v_y[0] += a * temp0;
#endif
}

24
kernel/power/sgemv_t_8.c
View File

@@ -99,7 +99,17 @@ static void sgemv_kernel_8x8(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOA
temp7 += vx1* va7_1 + vx2 * va7_2;
}
#if defined(POWER8)
y[0] += alpha * (temp0[0] + temp0[1]+temp0[2] + temp0[3]);
y[1] += alpha * (temp1[0] + temp1[1]+temp1[2] + temp1[3]);
y[2] += alpha * (temp2[0] + temp2[1]+temp2[2] + temp2[3]);
y[3] += alpha * (temp3[0] + temp3[1]+temp3[2] + temp3[3]);
y[4] += alpha * (temp4[0] + temp4[1]+temp4[2] + temp4[3]);
y[5] += alpha * (temp5[0] + temp5[1]+temp5[2] + temp5[3]);
y[6] += alpha * (temp6[0] + temp6[1]+temp6[2] + temp6[3]);
y[7] += alpha * (temp7[0] + temp7[1]+temp7[2] + temp7[3]);
#else
register __vector float t0, t1, t2, t3;
register __vector float a = { alpha, alpha, alpha, alpha };
__vector float *v_y = (__vector float*) y;
@@ -126,7 +136,7 @@ static void sgemv_kernel_8x8(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOA
v_y[0] += a * temp0;
v_y[1] += a * temp4;
#endif
}
@@ -153,7 +163,13 @@ static void sgemv_kernel_8x4(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOA
temp2 += v_x[i] * va2[i] + v_x[i+1] * va2[i+1];
temp3 += v_x[i] * va3[i] + v_x[i+1] * va3[i+1];
}
#if defined(POWER8)
y[0] += alpha * (temp0[0] + temp0[1]+temp0[2] + temp0[3]);
y[1] += alpha * (temp1[0] + temp1[1]+temp1[2] + temp1[3]);
y[2] += alpha * (temp2[0] + temp2[1]+temp2[2] + temp2[3]);
y[3] += alpha * (temp3[0] + temp3[1]+temp3[2] + temp3[3]);
#else
register __vector float t0, t1, t2, t3;
register __vector float a = { alpha, alpha, alpha, alpha };
__vector float *v_y = (__vector float*) y;
@@ -169,7 +185,7 @@ static void sgemv_kernel_8x4(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOA
temp0 += temp1 + temp2 + temp3;
v_y[0] += a * temp0;
#endif
}