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

Fixes RVV masked intrinsics for izamax/izamin kernels

This commit is contained in:
Octavian Maghiar 2023-07-12 12:55:50 +01:00
parent e1958eb705
commit 1e4a3a2b5e
2 changed files with 17 additions and 17 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
kernel/riscv64/izamax_rvv.c
View File

@ -48,9 +48,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VFADDVV_FLOAT __riscv_vfadd_vv_f64m4
#define VFIRSTM __riscv_vfirst_m_b16
#define UINT_V_T vuint64m4_t
#define VIDV_MASK_UINT __riscv_vid_v_u64m4_m
#define VIDV_MASK_UINT __riscv_vid_v_u64m4_mu
#define VIDV_UINT __riscv_vid_v_u64m4
#define VADDVX_MASK_UINT __riscv_vadd_vx_u64m4_m
#define VADDVX_MASK_UINT __riscv_vadd_vx_u64m4_mu
#define VADDVX_UINT __riscv_vadd_vx_u64m4
#define VMVVX_UINT __riscv_vmv_v_x_u64m4
#define VFMVFS_FLOAT_M1 __riscv_vfmv_f_s_f64m1_f64
@ -77,9 +77,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VFADDVV_FLOAT __riscv_vfadd_vv_f32m4
#define VFIRSTM __riscv_vfirst_m_b8
#define UINT_V_T vuint32m4_t
#define VIDV_MASK_UINT __riscv_vid_v_u32m4_m
#define VIDV_MASK_UINT __riscv_vid_v_u32m4_mu
#define VIDV_UINT __riscv_vid_v_u32m4
#define VADDVX_MASK_UINT __riscv_vadd_vx_u32m4_m
#define VADDVX_MASK_UINT __riscv_vadd_vx_u32m4_mu
#define VADDVX_UINT __riscv_vadd_vx_u32m4
#define VMVVX_UINT __riscv_vmv_v_x_u32m4
#define VFMVFS_FLOAT_M1 __riscv_vfmv_f_s_f32m1_f32
@ -116,8 +116,8 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
//index where element greater than v_max
mask = VMFLTVV_FLOAT(v_max, vx0, vl);
v_max_index = VIDV_MASK_UINT(mask, vl);
v_max_index = VADDVX_MASK_UINT(mask, v_max_index, j, vl);
v_max_index = VIDV_MASK_UINT(mask, v_max_index, vl);
v_max_index = VADDVX_MASK_UINT(mask, v_max_index, v_max_index, j, vl);
//update v_max and start_index j
v_max = VFMAXVV_FLOAT(v_max, vx0, vl);
@ -138,9 +138,9 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
//index where element greater than v_max
mask = VMFLTVV_FLOAT(v_max, vx0, vl);
v_max_index = VIDV_MASK_UINT(mask, vl);
v_max_index = VADDVX_MASK_UINT(mask, v_max_index, j, vl);
v_max_index = VIDV_MASK_UINT(mask, v_max_index, vl);
v_max_index = VADDVX_MASK_UINT(mask, v_max_index, v_max_index, j, vl);
//update v_max and start_index j
v_max = VFMAXVV_FLOAT(v_max, vx0, vl);
}

16
kernel/riscv64/izamin_rvv.c
View File

@ -47,9 +47,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VFADDVV_FLOAT __riscv_vfadd_vv_f64m4
#define VFIRSTM __riscv_vfirst_m_b16
#define UINT_V_T vuint64m4_t
#define VIDV_MASK_UINT __riscv_vid_v_u64m4_m
#define VIDV_MASK_UINT __riscv_vid_v_u64m4_mu
#define VIDV_UINT __riscv_vid_v_u64m4
#define VADDVX_MASK_UINT __riscv_vadd_vx_u64m4_m
#define VADDVX_MASK_UINT __riscv_vadd_vx_u64m4_mu
#define VADDVX_UINT __riscv_vadd_vx_u64m4
#define VMVVX_UINT __riscv_vmv_v_x_u64m4
#define VFMVFS_FLOAT_M1 __riscv_vfmv_f_s_f64m1_f64
@ -74,9 +74,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VFADDVV_FLOAT __riscv_vfadd_vv_f32m4
#define VFIRSTM __riscv_vfirst_m_b8
#define UINT_V_T vuint32m4_t
#define VIDV_MASK_UINT __riscv_vid_v_u32m4_m
#define VIDV_MASK_UINT __riscv_vid_v_u32m4_mu
#define VIDV_UINT __riscv_vid_v_u32m4
#define VADDVX_MASK_UINT __riscv_vadd_vx_u32m4_m
#define VADDVX_MASK_UINT __riscv_vadd_vx_u32m4_mu
#define VADDVX_UINT __riscv_vadd_vx_u32m4
#define VMVVX_UINT __riscv_vmv_v_x_u32m4
#define VFMVFS_FLOAT_M1 __riscv_vfmv_f_s_f32m1_f32
@ -113,8 +113,8 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
// index where element less than v_min
mask = VMFLTVV_FLOAT(vx0, v_min, vl);
v_min_index = VIDV_MASK_UINT(mask, vl);
v_min_index = VADDVX_MASK_UINT(mask, v_min_index, j, vl);
v_min_index = VIDV_MASK_UINT(mask, v_min_index, vl);
v_min_index = VADDVX_MASK_UINT(mask, v_min_index, v_min_index, j, vl);
//update v_min and start_index j
v_min = VFMINVV_FLOAT(v_min, vx0, vl);
@ -136,8 +136,8 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
// index where element less than v_min
mask = VMFLTVV_FLOAT(vx0, v_min, vl);
v_min_index = VIDV_MASK_UINT(mask, vl);
v_min_index = VADDVX_MASK_UINT(mask, v_min_index, j, vl);
v_min_index = VIDV_MASK_UINT(mask, v_min_index, vl);
v_min_index = VADDVX_MASK_UINT(mask, v_min_index, v_min_index, j, vl);
//update v_min and start_index j
v_min = VFMINVV_FLOAT(v_min, vx0, vl);