from patch provided by aarnez in #991tags/v0.3.6^2
| @@ -81,7 +81,7 @@ static FLOAT damax_kernel_32(BLASLONG n, FLOAT *x) | |||||
| "vfmaxdb %%v16,%%v16,%%v17,8 \n\t" | "vfmaxdb %%v16,%%v16,%%v17,8 \n\t" | ||||
| "vfmaxdb %%v0,%%v0,%%16,8 \n\t" | |||||
| "vfmaxdb %%v0,%%v0,%%v16,8 \n\t" | |||||
| "agfi %%r1, 256 \n\t" | "agfi %%r1, 256 \n\t" | ||||
| "brctg %%r0, 0b \n\t" | "brctg %%r0, 0b \n\t" | ||||
| @@ -81,7 +81,7 @@ static FLOAT damin_kernel_32(BLASLONG n, FLOAT *x) | |||||
| "vfmindb %%v16,%%v16,%%v17,8 \n\t" | "vfmindb %%v16,%%v16,%%v17,8 \n\t" | ||||
| "vfmindb %%v0,%%v0,%%16,8 \n\t" | |||||
| "vfmindb %%v0,%%v0,%%v16,8 \n\t" | |||||
| "agfi %%r1, 256 \n\t" | "agfi %%r1, 256 \n\t" | ||||
| "brctg %%r0, 0b \n\t" | "brctg %%r0, 0b \n\t" | ||||
| @@ -74,7 +74,7 @@ static FLOAT dmax_kernel_32(BLASLONG n, FLOAT *x) | |||||
| "vfmaxdb %%v16,%%v16,%%v17,0 \n\t" | "vfmaxdb %%v16,%%v16,%%v17,0 \n\t" | ||||
| "vfmaxdb %%v0,%%v0,%%16,0 \n\t" | |||||
| "vfmaxdb %%v0,%%v0,%%v16,0 \n\t" | |||||
| "agfi %%r1, 256 \n\t" | "agfi %%r1, 256 \n\t" | ||||
| "brctg %%r0, 0b \n\t" | "brctg %%r0, 0b \n\t" | ||||
| @@ -74,7 +74,7 @@ static FLOAT dmin_kernel_32(BLASLONG n, FLOAT *x) | |||||
| "vfmindb %%v16,%%v16,%%v17,0 \n\t" | "vfmindb %%v16,%%v16,%%v17,0 \n\t" | ||||
| "vfmindb %%v0,%%v0,%%16,0 \n\t" | |||||
| "vfmindb %%v0,%%v0,%%v16,0 \n\t" | |||||
| "agfi %%r1, 256 \n\t" | "agfi %%r1, 256 \n\t" | ||||
| "brctg %%r0, 0b \n\t" | "brctg %%r0, 0b \n\t" | ||||
| @@ -81,7 +81,7 @@ static FLOAT samax_kernel_64(BLASLONG n, FLOAT *x) | |||||
| "vfmaxsb %%v16,%%v16,%%v17,8 \n\t" | "vfmaxsb %%v16,%%v16,%%v17,8 \n\t" | ||||
| "vfmaxsb %%v0,%%v0,%%16,8 \n\t" | |||||
| "vfmaxsb %%v0,%%v0,%%v16,8 \n\t" | |||||
| "agfi %%r1, 256 \n\t" | "agfi %%r1, 256 \n\t" | ||||
| "brctg %%r0, 0b \n\t" | "brctg %%r0, 0b \n\t" | ||||
| @@ -81,7 +81,7 @@ static FLOAT samin_kernel_64(BLASLONG n, FLOAT *x) | |||||
| "vfminsb %%v16,%%v16,%%v17,8 \n\t" | "vfminsb %%v16,%%v16,%%v17,8 \n\t" | ||||
| "vfminsb %%v0,%%v0,%%16,8 \n\t" | |||||
| "vfminsb %%v0,%%v0,%%v16,8 \n\t" | |||||
| "agfi %%r1, 256 \n\t" | "agfi %%r1, 256 \n\t" | ||||
| "brctg %%r0, 0b \n\t" | "brctg %%r0, 0b \n\t" | ||||
| @@ -74,7 +74,7 @@ static FLOAT smax_kernel_64(BLASLONG n, FLOAT *x) | |||||
| "vfmaxsb %%v16,%%v16,%%v17,0 \n\t" | "vfmaxsb %%v16,%%v16,%%v17,0 \n\t" | ||||
| "vfmaxsb %%v0,%%v0,%%16,0 \n\t" | |||||
| "vfmaxsb %%v0,%%v0,%%v16,0 \n\t" | |||||
| "agfi %%r1, 256 \n\t" | "agfi %%r1, 256 \n\t" | ||||
| "brctg %%r0, 0b \n\t" | "brctg %%r0, 0b \n\t" | ||||
| @@ -74,7 +74,7 @@ static FLOAT smin_kernel_64(BLASLONG n, FLOAT *x) | |||||
| "vfminsb %%v16,%%v16,%%v17,0 \n\t" | "vfminsb %%v16,%%v16,%%v17,0 \n\t" | ||||
| "vfminsb %%v0,%%v0,%%16,0 \n\t" | |||||
| "vfminsb %%v0,%%v0,%%v16,0 \n\t" | |||||
| "agfi %%r1, 256 \n\t" | "agfi %%r1, 256 \n\t" | ||||
| "brctg %%r0, 0b \n\t" | "brctg %%r0, 0b \n\t" | ||||