/*

AUTOGENERATED KERNEL
Settings:
 LMUL=1
 M=8
 M_tail_scalar_from=1
 N=8
 __riscv_='__riscv_'
 complex=True
 conjugate=False
 cpu='zvl256b'
 force_acc_double=False
 index_type='BLASLONG'
 op='trmm'
 param_precision='float'
 reg_width_bits=256
 tail_policy=''
 trace=False

Derived:
 ELEN_ACC=32
 ELEN_PARAM=32
 LMUL_ACC=1
 VFMACC='__riscv_vfmacc_vf_f32m1'
 VFMUL='__riscv_vfmul_vf_f32m1'
 VLEV='__riscv_vle32_v_f32m1'
 VLSEV='__riscv_vlse32_v_f32m1'
 VMACC_TO_ACC='__riscv_vfmacc_vf_f32m1'
 VMUL_TO_ACC='__riscv_vfmul_vf_f32m1'
 VSETVL='__riscv_vsetvl_e32m1'
 VSEV='__riscv_vse32_v_f32m1'
 VSSEV='__riscv_vsse32_v_f32m1'
 acc_vector_t='vfloat32m1_t'
 output='ctrmm_kernel_8x8_zvl256b.c'
 param_scalar_t='float'
 param_vector_t='vfloat32m1_t'

*/

#include "common.h"



#if   defined(NN) || defined(NT) || defined(TN) || defined(TT)
    #define S0  1
    #define S1 -1
    #define S2  1
    #define S3  1
    #define VFMACC_RR __riscv_vfmsac
    #define VFMACC_RI __riscv_vfmacc
#endif
#if   defined(NR) || defined(NC) || defined(TR) || defined(TC)
    #define S0  1
    #define S1  1
    #define S2  1
    #define S3 -1
    #define VFMACC_RR __riscv_vfmacc
    #define VFMACC_RI __riscv_vfmsac
#endif
#if   defined(RN) || defined(RT) || defined(CN) || defined(CT)
    #define S0  1
    #define S1  1
    #define S2 -1
    #define S3  1
    #define VFMACC_RR __riscv_vfmacc
    #define VFMACC_RI __riscv_vfnmsac
#endif
#if   defined(RR) || defined(RC) || defined(CR) || defined(CC)
    #define S0  1
    #define S1 -1
    #define S2 -1
    #define S3 -1
    #define VFMACC_RR __riscv_vfmsac
    #define VFMACC_RI __riscv_vfnmacc
#endif


#if defined(LEFT) != defined(TRANSA)
    #define BACKWARDS
#endif

int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alphar, FLOAT alphai, FLOAT* A, FLOAT* B, FLOAT* C, BLASLONG ldc, BLASLONG offset)

{
    BLASLONG gvl = 0;
    BLASLONG m_top = 0;
    BLASLONG n_top = 0;


    // -- MAIN PASS

    for (BLASLONG j=0; j<N/8; j+=1) {
        m_top = 0;
        BLASLONG gvl = __riscv_vsetvl_e32m1(8);


        for (BLASLONG i=0; i<M/8; i+=1) {
            BLASLONG ai=m_top*K*2;
            BLASLONG bi=n_top*K*2;
            BLASLONG pass_K = K;
            #ifdef LEFT
                BLASLONG off = offset + m_top;
            #else
                BLASLONG off = -offset + n_top;
            #endif
            #ifdef BACKWARDS
                ai += off*8*2;
                bi += off*8*2;
                pass_K -= off;
            #else
                #ifdef LEFT
                    pass_K = off + 8;
                #else
                    pass_K = off + 8;
                #endif
            #endif
            float B0r = B[bi+0*2+0];
            float B0i = B[bi+0*2+1];
            float B1r = B[bi+1*2+0];
            float B1i = B[bi+1*2+1];
            float B2r = B[bi+2*2+0];
            float B2i = B[bi+2*2+1];
            float B3r = B[bi+3*2+0];
            float B3i = B[bi+3*2+1];
            float B4r = B[bi+4*2+0];
            float B4i = B[bi+4*2+1];
            float B5r = B[bi+5*2+0];
            float B5i = B[bi+5*2+1];
            float B6r = B[bi+6*2+0];
            float B6i = B[bi+6*2+1];
            float B7r = B[bi+7*2+0];
            float B7i = B[bi+7*2+1];
            bi += 8*2;

            vfloat32m1_t A0r = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2], sizeof(FLOAT)*2, gvl );
            vfloat32m1_t A0i = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2+1], sizeof(FLOAT)*2, gvl );
            ai += 8*2;

            // 2 vector regs to hold A array contents, 16 regs to hold values accumulated over k
            // leaving 14 vector registers for temporaries
            // performing 4 operations between reuses of temporaries
            vfloat32m1_t tmp0r = __riscv_vfmul_vf_f32m1( A0i, B0i, gvl);
            vfloat32m1_t tmp0i = __riscv_vfmul_vf_f32m1( A0r, B0i, gvl);
            vfloat32m1_t tmp1r = __riscv_vfmul_vf_f32m1( A0i, B1i, gvl);
            vfloat32m1_t tmp1i = __riscv_vfmul_vf_f32m1( A0r, B1i, gvl);
            vfloat32m1_t tmp2r = __riscv_vfmul_vf_f32m1( A0i, B2i, gvl);
            vfloat32m1_t tmp2i = __riscv_vfmul_vf_f32m1( A0r, B2i, gvl);
            vfloat32m1_t tmp3r = __riscv_vfmul_vf_f32m1( A0i, B3i, gvl);
            vfloat32m1_t tmp3i = __riscv_vfmul_vf_f32m1( A0r, B3i, gvl);
            tmp0r = VFMACC_RR( tmp0r, B0r, A0r, gvl);
            tmp0i = VFMACC_RI( tmp0i, B0r, A0i, gvl);
            tmp1r = VFMACC_RR( tmp1r, B1r, A0r, gvl);
            tmp1i = VFMACC_RI( tmp1i, B1r, A0i, gvl);
            tmp2r = VFMACC_RR( tmp2r, B2r, A0r, gvl);
            tmp2i = VFMACC_RI( tmp2i, B2r, A0i, gvl);
            tmp3r = VFMACC_RR( tmp3r, B3r, A0r, gvl);
            tmp3i = VFMACC_RI( tmp3i, B3r, A0i, gvl);
            vfloat32m1_t ACC0r = tmp0r;
            vfloat32m1_t ACC0i = tmp0i;
            vfloat32m1_t ACC1r = tmp1r;
            vfloat32m1_t ACC1i = tmp1i;
            vfloat32m1_t ACC2r = tmp2r;
            vfloat32m1_t ACC2i = tmp2i;
            vfloat32m1_t ACC3r = tmp3r;
            vfloat32m1_t ACC3i = tmp3i;
            tmp0r = __riscv_vfmul_vf_f32m1( A0i, B4i, gvl);
            tmp0i = __riscv_vfmul_vf_f32m1( A0r, B4i, gvl);
            tmp1r = __riscv_vfmul_vf_f32m1( A0i, B5i, gvl);
            tmp1i = __riscv_vfmul_vf_f32m1( A0r, B5i, gvl);
            tmp2r = __riscv_vfmul_vf_f32m1( A0i, B6i, gvl);
            tmp2i = __riscv_vfmul_vf_f32m1( A0r, B6i, gvl);
            tmp3r = __riscv_vfmul_vf_f32m1( A0i, B7i, gvl);
            tmp3i = __riscv_vfmul_vf_f32m1( A0r, B7i, gvl);
            tmp0r = VFMACC_RR( tmp0r, B4r, A0r, gvl);
            tmp0i = VFMACC_RI( tmp0i, B4r, A0i, gvl);
            tmp1r = VFMACC_RR( tmp1r, B5r, A0r, gvl);
            tmp1i = VFMACC_RI( tmp1i, B5r, A0i, gvl);
            tmp2r = VFMACC_RR( tmp2r, B6r, A0r, gvl);
            tmp2i = VFMACC_RI( tmp2i, B6r, A0i, gvl);
            tmp3r = VFMACC_RR( tmp3r, B7r, A0r, gvl);
            tmp3i = VFMACC_RI( tmp3i, B7r, A0i, gvl);
            vfloat32m1_t ACC4r = tmp0r;
            vfloat32m1_t ACC4i = tmp0i;
            vfloat32m1_t ACC5r = tmp1r;
            vfloat32m1_t ACC5i = tmp1i;
            vfloat32m1_t ACC6r = tmp2r;
            vfloat32m1_t ACC6i = tmp2i;
            vfloat32m1_t ACC7r = tmp3r;
            vfloat32m1_t ACC7i = tmp3i;

            for(BLASLONG k=1; k<pass_K; k++) {
                B0r = B[bi+0*2+0];
                B0i = B[bi+0*2+1];
                B1r = B[bi+1*2+0];
                B1i = B[bi+1*2+1];
                B2r = B[bi+2*2+0];
                B2i = B[bi+2*2+1];
                B3r = B[bi+3*2+0];
                B3i = B[bi+3*2+1];
                B4r = B[bi+4*2+0];
                B4i = B[bi+4*2+1];
                B5r = B[bi+5*2+0];
                B5i = B[bi+5*2+1];
                B6r = B[bi+6*2+0];
                B6i = B[bi+6*2+1];
                B7r = B[bi+7*2+0];
                B7i = B[bi+7*2+1];
                bi += 8*2;

                A0r = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2], sizeof(FLOAT)*2, gvl );
                A0i = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2+1], sizeof(FLOAT)*2, gvl );
                ai += 8*2;

                tmp0r = __riscv_vfmul_vf_f32m1( A0i, B0i, gvl);
                tmp0i = __riscv_vfmul_vf_f32m1( A0r, B0i, gvl);
                tmp1r = __riscv_vfmul_vf_f32m1( A0i, B1i, gvl);
                tmp1i = __riscv_vfmul_vf_f32m1( A0r, B1i, gvl);
                tmp2r = __riscv_vfmul_vf_f32m1( A0i, B2i, gvl);
                tmp2i = __riscv_vfmul_vf_f32m1( A0r, B2i, gvl);
                tmp3r = __riscv_vfmul_vf_f32m1( A0i, B3i, gvl);
                tmp3i = __riscv_vfmul_vf_f32m1( A0r, B3i, gvl);
                tmp0r = VFMACC_RR( tmp0r, B0r, A0r, gvl);
                tmp0i = VFMACC_RI( tmp0i, B0r, A0i, gvl);
                tmp1r = VFMACC_RR( tmp1r, B1r, A0r, gvl);
                tmp1i = VFMACC_RI( tmp1i, B1r, A0i, gvl);
                tmp2r = VFMACC_RR( tmp2r, B2r, A0r, gvl);
                tmp2i = VFMACC_RI( tmp2i, B2r, A0i, gvl);
                tmp3r = VFMACC_RR( tmp3r, B3r, A0r, gvl);
                tmp3i = VFMACC_RI( tmp3i, B3r, A0i, gvl);
                ACC0r = __riscv_vfadd( ACC0r, tmp0r, gvl);
                ACC0i = __riscv_vfadd( ACC0i, tmp0i, gvl);
                ACC1r = __riscv_vfadd( ACC1r, tmp1r, gvl);
                ACC1i = __riscv_vfadd( ACC1i, tmp1i, gvl);
                ACC2r = __riscv_vfadd( ACC2r, tmp2r, gvl);
                ACC2i = __riscv_vfadd( ACC2i, tmp2i, gvl);
                ACC3r = __riscv_vfadd( ACC3r, tmp3r, gvl);
                ACC3i = __riscv_vfadd( ACC3i, tmp3i, gvl);
                tmp0r = __riscv_vfmul_vf_f32m1( A0i, B4i, gvl);
                tmp0i = __riscv_vfmul_vf_f32m1( A0r, B4i, gvl);
                tmp1r = __riscv_vfmul_vf_f32m1( A0i, B5i, gvl);
                tmp1i = __riscv_vfmul_vf_f32m1( A0r, B5i, gvl);
                tmp2r = __riscv_vfmul_vf_f32m1( A0i, B6i, gvl);
                tmp2i = __riscv_vfmul_vf_f32m1( A0r, B6i, gvl);
                tmp3r = __riscv_vfmul_vf_f32m1( A0i, B7i, gvl);
                tmp3i = __riscv_vfmul_vf_f32m1( A0r, B7i, gvl);
                tmp0r = VFMACC_RR( tmp0r, B4r, A0r, gvl);
                tmp0i = VFMACC_RI( tmp0i, B4r, A0i, gvl);
                tmp1r = VFMACC_RR( tmp1r, B5r, A0r, gvl);
                tmp1i = VFMACC_RI( tmp1i, B5r, A0i, gvl);
                tmp2r = VFMACC_RR( tmp2r, B6r, A0r, gvl);
                tmp2i = VFMACC_RI( tmp2i, B6r, A0i, gvl);
                tmp3r = VFMACC_RR( tmp3r, B7r, A0r, gvl);
                tmp3i = VFMACC_RI( tmp3i, B7r, A0i, gvl);
                ACC4r = __riscv_vfadd( ACC4r, tmp0r, gvl);
                ACC4i = __riscv_vfadd( ACC4i, tmp0i, gvl);
                ACC5r = __riscv_vfadd( ACC5r, tmp1r, gvl);
                ACC5i = __riscv_vfadd( ACC5i, tmp1i, gvl);
                ACC6r = __riscv_vfadd( ACC6r, tmp2r, gvl);
                ACC6i = __riscv_vfadd( ACC6i, tmp2i, gvl);
                ACC7r = __riscv_vfadd( ACC7r, tmp3r, gvl);
                ACC7i = __riscv_vfadd( ACC7i, tmp3i, gvl);
            }


            BLASLONG ci=n_top*ldc+m_top;

            vfloat32m1_t C0r = __riscv_vfmul( ACC0r, alphar, gvl );
            vfloat32m1_t C0i = __riscv_vfmul( ACC0i, alphar, gvl );
            vfloat32m1_t C1r = __riscv_vfmul( ACC1r, alphar, gvl );
            vfloat32m1_t C1i = __riscv_vfmul( ACC1i, alphar, gvl );
            vfloat32m1_t C2r = __riscv_vfmul( ACC2r, alphar, gvl );
            vfloat32m1_t C2i = __riscv_vfmul( ACC2i, alphar, gvl );
            vfloat32m1_t C3r = __riscv_vfmul( ACC3r, alphar, gvl );
            vfloat32m1_t C3i = __riscv_vfmul( ACC3i, alphar, gvl );
            vfloat32m1_t C4r = __riscv_vfmul( ACC4r, alphar, gvl );
            vfloat32m1_t C4i = __riscv_vfmul( ACC4i, alphar, gvl );
            vfloat32m1_t C5r = __riscv_vfmul( ACC5r, alphar, gvl );
            vfloat32m1_t C5i = __riscv_vfmul( ACC5i, alphar, gvl );
            vfloat32m1_t C6r = __riscv_vfmul( ACC6r, alphar, gvl );
            vfloat32m1_t C6i = __riscv_vfmul( ACC6i, alphar, gvl );
            vfloat32m1_t C7r = __riscv_vfmul( ACC7r, alphar, gvl );
            vfloat32m1_t C7i = __riscv_vfmul( ACC7i, alphar, gvl );
            C0r = __riscv_vfnmsac( C0r, alphai, ACC0i, gvl );
            C0i = __riscv_vfmacc ( C0i, alphai, ACC0r, gvl );
            C1r = __riscv_vfnmsac( C1r, alphai, ACC1i, gvl );
            C1i = __riscv_vfmacc ( C1i, alphai, ACC1r, gvl );
            C2r = __riscv_vfnmsac( C2r, alphai, ACC2i, gvl );
            C2i = __riscv_vfmacc ( C2i, alphai, ACC2r, gvl );
            C3r = __riscv_vfnmsac( C3r, alphai, ACC3i, gvl );
            C3i = __riscv_vfmacc ( C3i, alphai, ACC3r, gvl );
            C4r = __riscv_vfnmsac( C4r, alphai, ACC4i, gvl );
            C4i = __riscv_vfmacc ( C4i, alphai, ACC4r, gvl );
            C5r = __riscv_vfnmsac( C5r, alphai, ACC5i, gvl );
            C5i = __riscv_vfmacc ( C5i, alphai, ACC5r, gvl );
            C6r = __riscv_vfnmsac( C6r, alphai, ACC6i, gvl );
            C6i = __riscv_vfmacc ( C6i, alphai, ACC6r, gvl );
            C7r = __riscv_vfnmsac( C7r, alphai, ACC7i, gvl );
            C7i = __riscv_vfmacc ( C7i, alphai, ACC7r, gvl );
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C0r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C0i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C1r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C1i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C2r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C2i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C3r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C3i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C4r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C4i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C5r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C5i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C6r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C6i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C7r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C7i, gvl);
            
            m_top += 8;
        }



        // -- tails for main pass

        if( M & 4 ) {
            gvl = __riscv_vsetvl_e32m1(4);

            BLASLONG ai=m_top*K*2;
            BLASLONG bi=n_top*K*2;
            BLASLONG pass_K = K;
            #ifdef LEFT
                BLASLONG off = offset + m_top;
            #else
                BLASLONG off = -offset + n_top;
            #endif
            #ifdef BACKWARDS
                ai += off*4*2;
                bi += off*8*2;
                pass_K -= off;
            #else
                #ifdef LEFT
                    pass_K = off + 4;
                #else
                    pass_K = off + 8;
                #endif
            #endif
            float B0r = B[bi+0*2+0];
            float B0i = B[bi+0*2+1];
            float B1r = B[bi+1*2+0];
            float B1i = B[bi+1*2+1];
            float B2r = B[bi+2*2+0];
            float B2i = B[bi+2*2+1];
            float B3r = B[bi+3*2+0];
            float B3i = B[bi+3*2+1];
            float B4r = B[bi+4*2+0];
            float B4i = B[bi+4*2+1];
            float B5r = B[bi+5*2+0];
            float B5i = B[bi+5*2+1];
            float B6r = B[bi+6*2+0];
            float B6i = B[bi+6*2+1];
            float B7r = B[bi+7*2+0];
            float B7i = B[bi+7*2+1];
            bi += 8*2;

            vfloat32m1_t A0r = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2], sizeof(FLOAT)*2, gvl );
            vfloat32m1_t A0i = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2+1], sizeof(FLOAT)*2, gvl );
            ai += 4*2;

            // 2 vector regs to hold A array contents, 16 regs to hold values accumulated over k
            // leaving 14 vector registers for temporaries
            // performing 4 operations between reuses of temporaries
            vfloat32m1_t tmp0r = __riscv_vfmul_vf_f32m1( A0i, B0i, gvl);
            vfloat32m1_t tmp0i = __riscv_vfmul_vf_f32m1( A0r, B0i, gvl);
            vfloat32m1_t tmp1r = __riscv_vfmul_vf_f32m1( A0i, B1i, gvl);
            vfloat32m1_t tmp1i = __riscv_vfmul_vf_f32m1( A0r, B1i, gvl);
            vfloat32m1_t tmp2r = __riscv_vfmul_vf_f32m1( A0i, B2i, gvl);
            vfloat32m1_t tmp2i = __riscv_vfmul_vf_f32m1( A0r, B2i, gvl);
            vfloat32m1_t tmp3r = __riscv_vfmul_vf_f32m1( A0i, B3i, gvl);
            vfloat32m1_t tmp3i = __riscv_vfmul_vf_f32m1( A0r, B3i, gvl);
            tmp0r = VFMACC_RR( tmp0r, B0r, A0r, gvl);
            tmp0i = VFMACC_RI( tmp0i, B0r, A0i, gvl);
            tmp1r = VFMACC_RR( tmp1r, B1r, A0r, gvl);
            tmp1i = VFMACC_RI( tmp1i, B1r, A0i, gvl);
            tmp2r = VFMACC_RR( tmp2r, B2r, A0r, gvl);
            tmp2i = VFMACC_RI( tmp2i, B2r, A0i, gvl);
            tmp3r = VFMACC_RR( tmp3r, B3r, A0r, gvl);
            tmp3i = VFMACC_RI( tmp3i, B3r, A0i, gvl);
            vfloat32m1_t ACC0r = tmp0r;
            vfloat32m1_t ACC0i = tmp0i;
            vfloat32m1_t ACC1r = tmp1r;
            vfloat32m1_t ACC1i = tmp1i;
            vfloat32m1_t ACC2r = tmp2r;
            vfloat32m1_t ACC2i = tmp2i;
            vfloat32m1_t ACC3r = tmp3r;
            vfloat32m1_t ACC3i = tmp3i;
            tmp0r = __riscv_vfmul_vf_f32m1( A0i, B4i, gvl);
            tmp0i = __riscv_vfmul_vf_f32m1( A0r, B4i, gvl);
            tmp1r = __riscv_vfmul_vf_f32m1( A0i, B5i, gvl);
            tmp1i = __riscv_vfmul_vf_f32m1( A0r, B5i, gvl);
            tmp2r = __riscv_vfmul_vf_f32m1( A0i, B6i, gvl);
            tmp2i = __riscv_vfmul_vf_f32m1( A0r, B6i, gvl);
            tmp3r = __riscv_vfmul_vf_f32m1( A0i, B7i, gvl);
            tmp3i = __riscv_vfmul_vf_f32m1( A0r, B7i, gvl);
            tmp0r = VFMACC_RR( tmp0r, B4r, A0r, gvl);
            tmp0i = VFMACC_RI( tmp0i, B4r, A0i, gvl);
            tmp1r = VFMACC_RR( tmp1r, B5r, A0r, gvl);
            tmp1i = VFMACC_RI( tmp1i, B5r, A0i, gvl);
            tmp2r = VFMACC_RR( tmp2r, B6r, A0r, gvl);
            tmp2i = VFMACC_RI( tmp2i, B6r, A0i, gvl);
            tmp3r = VFMACC_RR( tmp3r, B7r, A0r, gvl);
            tmp3i = VFMACC_RI( tmp3i, B7r, A0i, gvl);
            vfloat32m1_t ACC4r = tmp0r;
            vfloat32m1_t ACC4i = tmp0i;
            vfloat32m1_t ACC5r = tmp1r;
            vfloat32m1_t ACC5i = tmp1i;
            vfloat32m1_t ACC6r = tmp2r;
            vfloat32m1_t ACC6i = tmp2i;
            vfloat32m1_t ACC7r = tmp3r;
            vfloat32m1_t ACC7i = tmp3i;

            for(BLASLONG k=1; k<pass_K; k++) {
                B0r = B[bi+0*2+0];
                B0i = B[bi+0*2+1];
                B1r = B[bi+1*2+0];
                B1i = B[bi+1*2+1];
                B2r = B[bi+2*2+0];
                B2i = B[bi+2*2+1];
                B3r = B[bi+3*2+0];
                B3i = B[bi+3*2+1];
                B4r = B[bi+4*2+0];
                B4i = B[bi+4*2+1];
                B5r = B[bi+5*2+0];
                B5i = B[bi+5*2+1];
                B6r = B[bi+6*2+0];
                B6i = B[bi+6*2+1];
                B7r = B[bi+7*2+0];
                B7i = B[bi+7*2+1];
                bi += 8*2;

                A0r = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2], sizeof(FLOAT)*2, gvl );
                A0i = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2+1], sizeof(FLOAT)*2, gvl );
                ai += 4*2;

                tmp0r = __riscv_vfmul_vf_f32m1( A0i, B0i, gvl);
                tmp0i = __riscv_vfmul_vf_f32m1( A0r, B0i, gvl);
                tmp1r = __riscv_vfmul_vf_f32m1( A0i, B1i, gvl);
                tmp1i = __riscv_vfmul_vf_f32m1( A0r, B1i, gvl);
                tmp2r = __riscv_vfmul_vf_f32m1( A0i, B2i, gvl);
                tmp2i = __riscv_vfmul_vf_f32m1( A0r, B2i, gvl);
                tmp3r = __riscv_vfmul_vf_f32m1( A0i, B3i, gvl);
                tmp3i = __riscv_vfmul_vf_f32m1( A0r, B3i, gvl);
                tmp0r = VFMACC_RR( tmp0r, B0r, A0r, gvl);
                tmp0i = VFMACC_RI( tmp0i, B0r, A0i, gvl);
                tmp1r = VFMACC_RR( tmp1r, B1r, A0r, gvl);
                tmp1i = VFMACC_RI( tmp1i, B1r, A0i, gvl);
                tmp2r = VFMACC_RR( tmp2r, B2r, A0r, gvl);
                tmp2i = VFMACC_RI( tmp2i, B2r, A0i, gvl);
                tmp3r = VFMACC_RR( tmp3r, B3r, A0r, gvl);
                tmp3i = VFMACC_RI( tmp3i, B3r, A0i, gvl);
                ACC0r = __riscv_vfadd( ACC0r, tmp0r, gvl);
                ACC0i = __riscv_vfadd( ACC0i, tmp0i, gvl);
                ACC1r = __riscv_vfadd( ACC1r, tmp1r, gvl);
                ACC1i = __riscv_vfadd( ACC1i, tmp1i, gvl);
                ACC2r = __riscv_vfadd( ACC2r, tmp2r, gvl);
                ACC2i = __riscv_vfadd( ACC2i, tmp2i, gvl);
                ACC3r = __riscv_vfadd( ACC3r, tmp3r, gvl);
                ACC3i = __riscv_vfadd( ACC3i, tmp3i, gvl);
                tmp0r = __riscv_vfmul_vf_f32m1( A0i, B4i, gvl);
                tmp0i = __riscv_vfmul_vf_f32m1( A0r, B4i, gvl);
                tmp1r = __riscv_vfmul_vf_f32m1( A0i, B5i, gvl);
                tmp1i = __riscv_vfmul_vf_f32m1( A0r, B5i, gvl);
                tmp2r = __riscv_vfmul_vf_f32m1( A0i, B6i, gvl);
                tmp2i = __riscv_vfmul_vf_f32m1( A0r, B6i, gvl);
                tmp3r = __riscv_vfmul_vf_f32m1( A0i, B7i, gvl);
                tmp3i = __riscv_vfmul_vf_f32m1( A0r, B7i, gvl);
                tmp0r = VFMACC_RR( tmp0r, B4r, A0r, gvl);
                tmp0i = VFMACC_RI( tmp0i, B4r, A0i, gvl);
                tmp1r = VFMACC_RR( tmp1r, B5r, A0r, gvl);
                tmp1i = VFMACC_RI( tmp1i, B5r, A0i, gvl);
                tmp2r = VFMACC_RR( tmp2r, B6r, A0r, gvl);
                tmp2i = VFMACC_RI( tmp2i, B6r, A0i, gvl);
                tmp3r = VFMACC_RR( tmp3r, B7r, A0r, gvl);
                tmp3i = VFMACC_RI( tmp3i, B7r, A0i, gvl);
                ACC4r = __riscv_vfadd( ACC4r, tmp0r, gvl);
                ACC4i = __riscv_vfadd( ACC4i, tmp0i, gvl);
                ACC5r = __riscv_vfadd( ACC5r, tmp1r, gvl);
                ACC5i = __riscv_vfadd( ACC5i, tmp1i, gvl);
                ACC6r = __riscv_vfadd( ACC6r, tmp2r, gvl);
                ACC6i = __riscv_vfadd( ACC6i, tmp2i, gvl);
                ACC7r = __riscv_vfadd( ACC7r, tmp3r, gvl);
                ACC7i = __riscv_vfadd( ACC7i, tmp3i, gvl);
            }


            BLASLONG ci=n_top*ldc+m_top;

            vfloat32m1_t C0r = __riscv_vfmul( ACC0r, alphar, gvl );
            vfloat32m1_t C0i = __riscv_vfmul( ACC0i, alphar, gvl );
            vfloat32m1_t C1r = __riscv_vfmul( ACC1r, alphar, gvl );
            vfloat32m1_t C1i = __riscv_vfmul( ACC1i, alphar, gvl );
            vfloat32m1_t C2r = __riscv_vfmul( ACC2r, alphar, gvl );
            vfloat32m1_t C2i = __riscv_vfmul( ACC2i, alphar, gvl );
            vfloat32m1_t C3r = __riscv_vfmul( ACC3r, alphar, gvl );
            vfloat32m1_t C3i = __riscv_vfmul( ACC3i, alphar, gvl );
            vfloat32m1_t C4r = __riscv_vfmul( ACC4r, alphar, gvl );
            vfloat32m1_t C4i = __riscv_vfmul( ACC4i, alphar, gvl );
            vfloat32m1_t C5r = __riscv_vfmul( ACC5r, alphar, gvl );
            vfloat32m1_t C5i = __riscv_vfmul( ACC5i, alphar, gvl );
            vfloat32m1_t C6r = __riscv_vfmul( ACC6r, alphar, gvl );
            vfloat32m1_t C6i = __riscv_vfmul( ACC6i, alphar, gvl );
            vfloat32m1_t C7r = __riscv_vfmul( ACC7r, alphar, gvl );
            vfloat32m1_t C7i = __riscv_vfmul( ACC7i, alphar, gvl );
            C0r = __riscv_vfnmsac( C0r, alphai, ACC0i, gvl );
            C0i = __riscv_vfmacc ( C0i, alphai, ACC0r, gvl );
            C1r = __riscv_vfnmsac( C1r, alphai, ACC1i, gvl );
            C1i = __riscv_vfmacc ( C1i, alphai, ACC1r, gvl );
            C2r = __riscv_vfnmsac( C2r, alphai, ACC2i, gvl );
            C2i = __riscv_vfmacc ( C2i, alphai, ACC2r, gvl );
            C3r = __riscv_vfnmsac( C3r, alphai, ACC3i, gvl );
            C3i = __riscv_vfmacc ( C3i, alphai, ACC3r, gvl );
            C4r = __riscv_vfnmsac( C4r, alphai, ACC4i, gvl );
            C4i = __riscv_vfmacc ( C4i, alphai, ACC4r, gvl );
            C5r = __riscv_vfnmsac( C5r, alphai, ACC5i, gvl );
            C5i = __riscv_vfmacc ( C5i, alphai, ACC5r, gvl );
            C6r = __riscv_vfnmsac( C6r, alphai, ACC6i, gvl );
            C6i = __riscv_vfmacc ( C6i, alphai, ACC6r, gvl );
            C7r = __riscv_vfnmsac( C7r, alphai, ACC7i, gvl );
            C7i = __riscv_vfmacc ( C7i, alphai, ACC7r, gvl );
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C0r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C0i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C1r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C1i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C2r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C2i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C3r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C3i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C4r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C4i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C5r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C5i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C6r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C6i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C7r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C7i, gvl);
            
            m_top += 4;
        }


        if( M & 2 ) {
            gvl = __riscv_vsetvl_e32m1(2);

            BLASLONG ai=m_top*K*2;
            BLASLONG bi=n_top*K*2;
            BLASLONG pass_K = K;
            #ifdef LEFT
                BLASLONG off = offset + m_top;
            #else
                BLASLONG off = -offset + n_top;
            #endif
            #ifdef BACKWARDS
                ai += off*2*2;
                bi += off*8*2;
                pass_K -= off;
            #else
                #ifdef LEFT
                    pass_K = off + 2;
                #else
                    pass_K = off + 8;
                #endif
            #endif
            float B0r = B[bi+0*2+0];
            float B0i = B[bi+0*2+1];
            float B1r = B[bi+1*2+0];
            float B1i = B[bi+1*2+1];
            float B2r = B[bi+2*2+0];
            float B2i = B[bi+2*2+1];
            float B3r = B[bi+3*2+0];
            float B3i = B[bi+3*2+1];
            float B4r = B[bi+4*2+0];
            float B4i = B[bi+4*2+1];
            float B5r = B[bi+5*2+0];
            float B5i = B[bi+5*2+1];
            float B6r = B[bi+6*2+0];
            float B6i = B[bi+6*2+1];
            float B7r = B[bi+7*2+0];
            float B7i = B[bi+7*2+1];
            bi += 8*2;

            vfloat32m1_t A0r = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2], sizeof(FLOAT)*2, gvl );
            vfloat32m1_t A0i = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2+1], sizeof(FLOAT)*2, gvl );
            ai += 2*2;

            // 2 vector regs to hold A array contents, 16 regs to hold values accumulated over k
            // leaving 14 vector registers for temporaries
            // performing 4 operations between reuses of temporaries
            vfloat32m1_t tmp0r = __riscv_vfmul_vf_f32m1( A0i, B0i, gvl);
            vfloat32m1_t tmp0i = __riscv_vfmul_vf_f32m1( A0r, B0i, gvl);
            vfloat32m1_t tmp1r = __riscv_vfmul_vf_f32m1( A0i, B1i, gvl);
            vfloat32m1_t tmp1i = __riscv_vfmul_vf_f32m1( A0r, B1i, gvl);
            vfloat32m1_t tmp2r = __riscv_vfmul_vf_f32m1( A0i, B2i, gvl);
            vfloat32m1_t tmp2i = __riscv_vfmul_vf_f32m1( A0r, B2i, gvl);
            vfloat32m1_t tmp3r = __riscv_vfmul_vf_f32m1( A0i, B3i, gvl);
            vfloat32m1_t tmp3i = __riscv_vfmul_vf_f32m1( A0r, B3i, gvl);
            tmp0r = VFMACC_RR( tmp0r, B0r, A0r, gvl);
            tmp0i = VFMACC_RI( tmp0i, B0r, A0i, gvl);
            tmp1r = VFMACC_RR( tmp1r, B1r, A0r, gvl);
            tmp1i = VFMACC_RI( tmp1i, B1r, A0i, gvl);
            tmp2r = VFMACC_RR( tmp2r, B2r, A0r, gvl);
            tmp2i = VFMACC_RI( tmp2i, B2r, A0i, gvl);
            tmp3r = VFMACC_RR( tmp3r, B3r, A0r, gvl);
            tmp3i = VFMACC_RI( tmp3i, B3r, A0i, gvl);
            vfloat32m1_t ACC0r = tmp0r;
            vfloat32m1_t ACC0i = tmp0i;
            vfloat32m1_t ACC1r = tmp1r;
            vfloat32m1_t ACC1i = tmp1i;
            vfloat32m1_t ACC2r = tmp2r;
            vfloat32m1_t ACC2i = tmp2i;
            vfloat32m1_t ACC3r = tmp3r;
            vfloat32m1_t ACC3i = tmp3i;
            tmp0r = __riscv_vfmul_vf_f32m1( A0i, B4i, gvl);
            tmp0i = __riscv_vfmul_vf_f32m1( A0r, B4i, gvl);
            tmp1r = __riscv_vfmul_vf_f32m1( A0i, B5i, gvl);
            tmp1i = __riscv_vfmul_vf_f32m1( A0r, B5i, gvl);
            tmp2r = __riscv_vfmul_vf_f32m1( A0i, B6i, gvl);
            tmp2i = __riscv_vfmul_vf_f32m1( A0r, B6i, gvl);
            tmp3r = __riscv_vfmul_vf_f32m1( A0i, B7i, gvl);
            tmp3i = __riscv_vfmul_vf_f32m1( A0r, B7i, gvl);
            tmp0r = VFMACC_RR( tmp0r, B4r, A0r, gvl);
            tmp0i = VFMACC_RI( tmp0i, B4r, A0i, gvl);
            tmp1r = VFMACC_RR( tmp1r, B5r, A0r, gvl);
            tmp1i = VFMACC_RI( tmp1i, B5r, A0i, gvl);
            tmp2r = VFMACC_RR( tmp2r, B6r, A0r, gvl);
            tmp2i = VFMACC_RI( tmp2i, B6r, A0i, gvl);
            tmp3r = VFMACC_RR( tmp3r, B7r, A0r, gvl);
            tmp3i = VFMACC_RI( tmp3i, B7r, A0i, gvl);
            vfloat32m1_t ACC4r = tmp0r;
            vfloat32m1_t ACC4i = tmp0i;
            vfloat32m1_t ACC5r = tmp1r;
            vfloat32m1_t ACC5i = tmp1i;
            vfloat32m1_t ACC6r = tmp2r;
            vfloat32m1_t ACC6i = tmp2i;
            vfloat32m1_t ACC7r = tmp3r;
            vfloat32m1_t ACC7i = tmp3i;

            for(BLASLONG k=1; k<pass_K; k++) {
                B0r = B[bi+0*2+0];
                B0i = B[bi+0*2+1];
                B1r = B[bi+1*2+0];
                B1i = B[bi+1*2+1];
                B2r = B[bi+2*2+0];
                B2i = B[bi+2*2+1];
                B3r = B[bi+3*2+0];
                B3i = B[bi+3*2+1];
                B4r = B[bi+4*2+0];
                B4i = B[bi+4*2+1];
                B5r = B[bi+5*2+0];
                B5i = B[bi+5*2+1];
                B6r = B[bi+6*2+0];
                B6i = B[bi+6*2+1];
                B7r = B[bi+7*2+0];
                B7i = B[bi+7*2+1];
                bi += 8*2;

                A0r = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2], sizeof(FLOAT)*2, gvl );
                A0i = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2+1], sizeof(FLOAT)*2, gvl );
                ai += 2*2;

                tmp0r = __riscv_vfmul_vf_f32m1( A0i, B0i, gvl);
                tmp0i = __riscv_vfmul_vf_f32m1( A0r, B0i, gvl);
                tmp1r = __riscv_vfmul_vf_f32m1( A0i, B1i, gvl);
                tmp1i = __riscv_vfmul_vf_f32m1( A0r, B1i, gvl);
                tmp2r = __riscv_vfmul_vf_f32m1( A0i, B2i, gvl);
                tmp2i = __riscv_vfmul_vf_f32m1( A0r, B2i, gvl);
                tmp3r = __riscv_vfmul_vf_f32m1( A0i, B3i, gvl);
                tmp3i = __riscv_vfmul_vf_f32m1( A0r, B3i, gvl);
                tmp0r = VFMACC_RR( tmp0r, B0r, A0r, gvl);
                tmp0i = VFMACC_RI( tmp0i, B0r, A0i, gvl);
                tmp1r = VFMACC_RR( tmp1r, B1r, A0r, gvl);
                tmp1i = VFMACC_RI( tmp1i, B1r, A0i, gvl);
                tmp2r = VFMACC_RR( tmp2r, B2r, A0r, gvl);
                tmp2i = VFMACC_RI( tmp2i, B2r, A0i, gvl);
                tmp3r = VFMACC_RR( tmp3r, B3r, A0r, gvl);
                tmp3i = VFMACC_RI( tmp3i, B3r, A0i, gvl);
                ACC0r = __riscv_vfadd( ACC0r, tmp0r, gvl);
                ACC0i = __riscv_vfadd( ACC0i, tmp0i, gvl);
                ACC1r = __riscv_vfadd( ACC1r, tmp1r, gvl);
                ACC1i = __riscv_vfadd( ACC1i, tmp1i, gvl);
                ACC2r = __riscv_vfadd( ACC2r, tmp2r, gvl);
                ACC2i = __riscv_vfadd( ACC2i, tmp2i, gvl);
                ACC3r = __riscv_vfadd( ACC3r, tmp3r, gvl);
                ACC3i = __riscv_vfadd( ACC3i, tmp3i, gvl);
                tmp0r = __riscv_vfmul_vf_f32m1( A0i, B4i, gvl);
                tmp0i = __riscv_vfmul_vf_f32m1( A0r, B4i, gvl);
                tmp1r = __riscv_vfmul_vf_f32m1( A0i, B5i, gvl);
                tmp1i = __riscv_vfmul_vf_f32m1( A0r, B5i, gvl);
                tmp2r = __riscv_vfmul_vf_f32m1( A0i, B6i, gvl);
                tmp2i = __riscv_vfmul_vf_f32m1( A0r, B6i, gvl);
                tmp3r = __riscv_vfmul_vf_f32m1( A0i, B7i, gvl);
                tmp3i = __riscv_vfmul_vf_f32m1( A0r, B7i, gvl);
                tmp0r = VFMACC_RR( tmp0r, B4r, A0r, gvl);
                tmp0i = VFMACC_RI( tmp0i, B4r, A0i, gvl);
                tmp1r = VFMACC_RR( tmp1r, B5r, A0r, gvl);
                tmp1i = VFMACC_RI( tmp1i, B5r, A0i, gvl);
                tmp2r = VFMACC_RR( tmp2r, B6r, A0r, gvl);
                tmp2i = VFMACC_RI( tmp2i, B6r, A0i, gvl);
                tmp3r = VFMACC_RR( tmp3r, B7r, A0r, gvl);
                tmp3i = VFMACC_RI( tmp3i, B7r, A0i, gvl);
                ACC4r = __riscv_vfadd( ACC4r, tmp0r, gvl);
                ACC4i = __riscv_vfadd( ACC4i, tmp0i, gvl);
                ACC5r = __riscv_vfadd( ACC5r, tmp1r, gvl);
                ACC5i = __riscv_vfadd( ACC5i, tmp1i, gvl);
                ACC6r = __riscv_vfadd( ACC6r, tmp2r, gvl);
                ACC6i = __riscv_vfadd( ACC6i, tmp2i, gvl);
                ACC7r = __riscv_vfadd( ACC7r, tmp3r, gvl);
                ACC7i = __riscv_vfadd( ACC7i, tmp3i, gvl);
            }


            BLASLONG ci=n_top*ldc+m_top;

            vfloat32m1_t C0r = __riscv_vfmul( ACC0r, alphar, gvl );
            vfloat32m1_t C0i = __riscv_vfmul( ACC0i, alphar, gvl );
            vfloat32m1_t C1r = __riscv_vfmul( ACC1r, alphar, gvl );
            vfloat32m1_t C1i = __riscv_vfmul( ACC1i, alphar, gvl );
            vfloat32m1_t C2r = __riscv_vfmul( ACC2r, alphar, gvl );
            vfloat32m1_t C2i = __riscv_vfmul( ACC2i, alphar, gvl );
            vfloat32m1_t C3r = __riscv_vfmul( ACC3r, alphar, gvl );
            vfloat32m1_t C3i = __riscv_vfmul( ACC3i, alphar, gvl );
            vfloat32m1_t C4r = __riscv_vfmul( ACC4r, alphar, gvl );
            vfloat32m1_t C4i = __riscv_vfmul( ACC4i, alphar, gvl );
            vfloat32m1_t C5r = __riscv_vfmul( ACC5r, alphar, gvl );
            vfloat32m1_t C5i = __riscv_vfmul( ACC5i, alphar, gvl );
            vfloat32m1_t C6r = __riscv_vfmul( ACC6r, alphar, gvl );
            vfloat32m1_t C6i = __riscv_vfmul( ACC6i, alphar, gvl );
            vfloat32m1_t C7r = __riscv_vfmul( ACC7r, alphar, gvl );
            vfloat32m1_t C7i = __riscv_vfmul( ACC7i, alphar, gvl );
            C0r = __riscv_vfnmsac( C0r, alphai, ACC0i, gvl );
            C0i = __riscv_vfmacc ( C0i, alphai, ACC0r, gvl );
            C1r = __riscv_vfnmsac( C1r, alphai, ACC1i, gvl );
            C1i = __riscv_vfmacc ( C1i, alphai, ACC1r, gvl );
            C2r = __riscv_vfnmsac( C2r, alphai, ACC2i, gvl );
            C2i = __riscv_vfmacc ( C2i, alphai, ACC2r, gvl );
            C3r = __riscv_vfnmsac( C3r, alphai, ACC3i, gvl );
            C3i = __riscv_vfmacc ( C3i, alphai, ACC3r, gvl );
            C4r = __riscv_vfnmsac( C4r, alphai, ACC4i, gvl );
            C4i = __riscv_vfmacc ( C4i, alphai, ACC4r, gvl );
            C5r = __riscv_vfnmsac( C5r, alphai, ACC5i, gvl );
            C5i = __riscv_vfmacc ( C5i, alphai, ACC5r, gvl );
            C6r = __riscv_vfnmsac( C6r, alphai, ACC6i, gvl );
            C6i = __riscv_vfmacc ( C6i, alphai, ACC6r, gvl );
            C7r = __riscv_vfnmsac( C7r, alphai, ACC7i, gvl );
            C7i = __riscv_vfmacc ( C7i, alphai, ACC7r, gvl );
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C0r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C0i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C1r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C1i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C2r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C2i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C3r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C3i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C4r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C4i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C5r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C5i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C6r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C6i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C7r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C7i, gvl);
            
            m_top += 2;
        }


        if( M & 1 ) {
            float result0 = 0;
            float result1 = 0;
            float result2 = 0;
            float result3 = 0;
            float result4 = 0;
            float result5 = 0;
            float result6 = 0;
            float result7 = 0;
            float result8 = 0;
            float result9 = 0;
            float result10 = 0;
            float result11 = 0;
            float result12 = 0;
            float result13 = 0;
            float result14 = 0;
            float result15 = 0;
            BLASLONG ai=m_top*K*2;
            BLASLONG bi=n_top*K*2;
            BLASLONG pass_K = K;
            #ifdef LEFT
                BLASLONG off = offset + m_top;
            #else
                BLASLONG off = -offset + n_top;
            #endif
            #ifdef BACKWARDS
                ai += off*1*2;
                bi += off*8*2;
                pass_K -= off;
            #else
                #ifdef LEFT
                    pass_K = off + 1;
                #else
                    pass_K = off + 8;
                #endif
            #endif

            for(BLASLONG k=0; k<pass_K; k++) {
                result0+=S0*A[ai+0+0]*B[bi+0+0] + S1*A[ai+0+1]*B[bi+0+1];
                result1+=S2*A[ai+0+1]*B[bi+0+0] + S3*A[ai+0+0]*B[bi+0+1];
                result2+=S0*A[ai+0+0]*B[bi+2+0] + S1*A[ai+0+1]*B[bi+2+1];
                result3+=S2*A[ai+0+1]*B[bi+2+0] + S3*A[ai+0+0]*B[bi+2+1];
                result4+=S0*A[ai+0+0]*B[bi+4+0] + S1*A[ai+0+1]*B[bi+4+1];
                result5+=S2*A[ai+0+1]*B[bi+4+0] + S3*A[ai+0+0]*B[bi+4+1];
                result6+=S0*A[ai+0+0]*B[bi+6+0] + S1*A[ai+0+1]*B[bi+6+1];
                result7+=S2*A[ai+0+1]*B[bi+6+0] + S3*A[ai+0+0]*B[bi+6+1];
                result8+=S0*A[ai+0+0]*B[bi+8+0] + S1*A[ai+0+1]*B[bi+8+1];
                result9+=S2*A[ai+0+1]*B[bi+8+0] + S3*A[ai+0+0]*B[bi+8+1];
                result10+=S0*A[ai+0+0]*B[bi+10+0] + S1*A[ai+0+1]*B[bi+10+1];
                result11+=S2*A[ai+0+1]*B[bi+10+0] + S3*A[ai+0+0]*B[bi+10+1];
                result12+=S0*A[ai+0+0]*B[bi+12+0] + S1*A[ai+0+1]*B[bi+12+1];
                result13+=S2*A[ai+0+1]*B[bi+12+0] + S3*A[ai+0+0]*B[bi+12+1];
                result14+=S0*A[ai+0+0]*B[bi+14+0] + S1*A[ai+0+1]*B[bi+14+1];
                result15+=S2*A[ai+0+1]*B[bi+14+0] + S3*A[ai+0+0]*B[bi+14+1];
                ai+=1*2;
                bi+=8*2;
            }

            BLASLONG ci=n_top*ldc+m_top;
            float Cr, Ci;
            Cr = result0*alphar;
            Ci = result1*alphar;
            Cr -= result1*alphai;
            Ci += result0*alphai;
            C[(ci+0*ldc+0)*2+0] = Cr;
            C[(ci+0*ldc+0)*2+1] = Ci;
            Cr = result2*alphar;
            Ci = result3*alphar;
            Cr -= result3*alphai;
            Ci += result2*alphai;
            C[(ci+1*ldc+0)*2+0] = Cr;
            C[(ci+1*ldc+0)*2+1] = Ci;
            Cr = result4*alphar;
            Ci = result5*alphar;
            Cr -= result5*alphai;
            Ci += result4*alphai;
            C[(ci+2*ldc+0)*2+0] = Cr;
            C[(ci+2*ldc+0)*2+1] = Ci;
            Cr = result6*alphar;
            Ci = result7*alphar;
            Cr -= result7*alphai;
            Ci += result6*alphai;
            C[(ci+3*ldc+0)*2+0] = Cr;
            C[(ci+3*ldc+0)*2+1] = Ci;
            Cr = result8*alphar;
            Ci = result9*alphar;
            Cr -= result9*alphai;
            Ci += result8*alphai;
            C[(ci+4*ldc+0)*2+0] = Cr;
            C[(ci+4*ldc+0)*2+1] = Ci;
            Cr = result10*alphar;
            Ci = result11*alphar;
            Cr -= result11*alphai;
            Ci += result10*alphai;
            C[(ci+5*ldc+0)*2+0] = Cr;
            C[(ci+5*ldc+0)*2+1] = Ci;
            Cr = result12*alphar;
            Ci = result13*alphar;
            Cr -= result13*alphai;
            Ci += result12*alphai;
            C[(ci+6*ldc+0)*2+0] = Cr;
            C[(ci+6*ldc+0)*2+1] = Ci;
            Cr = result14*alphar;
            Ci = result15*alphar;
            Cr -= result15*alphai;
            Ci += result14*alphai;
            C[(ci+7*ldc+0)*2+0] = Cr;
            C[(ci+7*ldc+0)*2+1] = Ci;
            m_top+=1;
        }

        n_top += 8;
    }



    // -- tails for N=4

    if( N & 4 ) {
        gvl = __riscv_vsetvl_e32m1(8);
        m_top = 0;

        for (BLASLONG i=0; i<M/8; i+=1) {
            BLASLONG ai=m_top*K*2;
            BLASLONG bi=n_top*K*2;
            BLASLONG pass_K = K;
            #ifdef LEFT
                BLASLONG off = offset + m_top;
            #else
                BLASLONG off = -offset + n_top;
            #endif
            #ifdef BACKWARDS
                ai += off*8*2;
                bi += off*4*2;
                pass_K -= off;
            #else
                #ifdef LEFT
                    pass_K = off + 8;
                #else
                    pass_K = off + 4;
                #endif
            #endif
            float B0r = B[bi+0*2+0];
            float B0i = B[bi+0*2+1];
            float B1r = B[bi+1*2+0];
            float B1i = B[bi+1*2+1];
            float B2r = B[bi+2*2+0];
            float B2i = B[bi+2*2+1];
            float B3r = B[bi+3*2+0];
            float B3i = B[bi+3*2+1];
            bi += 4*2;

            vfloat32m1_t A0r = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2], sizeof(FLOAT)*2, gvl );
            vfloat32m1_t A0i = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2+1], sizeof(FLOAT)*2, gvl );
            ai += 8*2;

            // 2 vector regs to hold A array contents, 8 regs to hold values accumulated over k
            // leaving 22 vector registers for temporaries
            vfloat32m1_t tmp0r = __riscv_vfmul_vf_f32m1( A0i, B0i, gvl);
            vfloat32m1_t tmp0i = __riscv_vfmul_vf_f32m1( A0r, B0i, gvl);
            vfloat32m1_t tmp1r = __riscv_vfmul_vf_f32m1( A0i, B1i, gvl);
            vfloat32m1_t tmp1i = __riscv_vfmul_vf_f32m1( A0r, B1i, gvl);
            vfloat32m1_t tmp2r = __riscv_vfmul_vf_f32m1( A0i, B2i, gvl);
            vfloat32m1_t tmp2i = __riscv_vfmul_vf_f32m1( A0r, B2i, gvl);
            vfloat32m1_t tmp3r = __riscv_vfmul_vf_f32m1( A0i, B3i, gvl);
            vfloat32m1_t tmp3i = __riscv_vfmul_vf_f32m1( A0r, B3i, gvl);
            tmp0r = VFMACC_RR( tmp0r, B0r, A0r, gvl);
            tmp0i = VFMACC_RI( tmp0i, B0r, A0i, gvl);
            tmp1r = VFMACC_RR( tmp1r, B1r, A0r, gvl);
            tmp1i = VFMACC_RI( tmp1i, B1r, A0i, gvl);
            tmp2r = VFMACC_RR( tmp2r, B2r, A0r, gvl);
            tmp2i = VFMACC_RI( tmp2i, B2r, A0i, gvl);
            tmp3r = VFMACC_RR( tmp3r, B3r, A0r, gvl);
            tmp3i = VFMACC_RI( tmp3i, B3r, A0i, gvl);
            vfloat32m1_t ACC0r = tmp0r;
            vfloat32m1_t ACC0i = tmp0i;
            vfloat32m1_t ACC1r = tmp1r;
            vfloat32m1_t ACC1i = tmp1i;
            vfloat32m1_t ACC2r = tmp2r;
            vfloat32m1_t ACC2i = tmp2i;
            vfloat32m1_t ACC3r = tmp3r;
            vfloat32m1_t ACC3i = tmp3i;

            for(BLASLONG k=1; k<pass_K; k++) {
                B0r = B[bi+0*2+0];
                B0i = B[bi+0*2+1];
                B1r = B[bi+1*2+0];
                B1i = B[bi+1*2+1];
                B2r = B[bi+2*2+0];
                B2i = B[bi+2*2+1];
                B3r = B[bi+3*2+0];
                B3i = B[bi+3*2+1];
                bi += 4*2;

                A0r = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2], sizeof(FLOAT)*2, gvl );
                A0i = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2+1], sizeof(FLOAT)*2, gvl );
                ai += 8*2;

                tmp0r = __riscv_vfmul_vf_f32m1( A0i, B0i, gvl);
                tmp0i = __riscv_vfmul_vf_f32m1( A0r, B0i, gvl);
                tmp1r = __riscv_vfmul_vf_f32m1( A0i, B1i, gvl);
                tmp1i = __riscv_vfmul_vf_f32m1( A0r, B1i, gvl);
                tmp2r = __riscv_vfmul_vf_f32m1( A0i, B2i, gvl);
                tmp2i = __riscv_vfmul_vf_f32m1( A0r, B2i, gvl);
                tmp3r = __riscv_vfmul_vf_f32m1( A0i, B3i, gvl);
                tmp3i = __riscv_vfmul_vf_f32m1( A0r, B3i, gvl);
                tmp0r = VFMACC_RR( tmp0r, B0r, A0r, gvl);
                tmp0i = VFMACC_RI( tmp0i, B0r, A0i, gvl);
                tmp1r = VFMACC_RR( tmp1r, B1r, A0r, gvl);
                tmp1i = VFMACC_RI( tmp1i, B1r, A0i, gvl);
                tmp2r = VFMACC_RR( tmp2r, B2r, A0r, gvl);
                tmp2i = VFMACC_RI( tmp2i, B2r, A0i, gvl);
                tmp3r = VFMACC_RR( tmp3r, B3r, A0r, gvl);
                tmp3i = VFMACC_RI( tmp3i, B3r, A0i, gvl);
                ACC0r = __riscv_vfadd( ACC0r, tmp0r, gvl);
                ACC0i = __riscv_vfadd( ACC0i, tmp0i, gvl);
                ACC1r = __riscv_vfadd( ACC1r, tmp1r, gvl);
                ACC1i = __riscv_vfadd( ACC1i, tmp1i, gvl);
                ACC2r = __riscv_vfadd( ACC2r, tmp2r, gvl);
                ACC2i = __riscv_vfadd( ACC2i, tmp2i, gvl);
                ACC3r = __riscv_vfadd( ACC3r, tmp3r, gvl);
                ACC3i = __riscv_vfadd( ACC3i, tmp3i, gvl);
            }


            BLASLONG ci=n_top*ldc+m_top;

            vfloat32m1_t C0r = __riscv_vfmul( ACC0r, alphar, gvl );
            vfloat32m1_t C0i = __riscv_vfmul( ACC0i, alphar, gvl );
            vfloat32m1_t C1r = __riscv_vfmul( ACC1r, alphar, gvl );
            vfloat32m1_t C1i = __riscv_vfmul( ACC1i, alphar, gvl );
            vfloat32m1_t C2r = __riscv_vfmul( ACC2r, alphar, gvl );
            vfloat32m1_t C2i = __riscv_vfmul( ACC2i, alphar, gvl );
            vfloat32m1_t C3r = __riscv_vfmul( ACC3r, alphar, gvl );
            vfloat32m1_t C3i = __riscv_vfmul( ACC3i, alphar, gvl );
            C0r = __riscv_vfnmsac( C0r, alphai, ACC0i, gvl );
            C0i = __riscv_vfmacc ( C0i, alphai, ACC0r, gvl );
            C1r = __riscv_vfnmsac( C1r, alphai, ACC1i, gvl );
            C1i = __riscv_vfmacc ( C1i, alphai, ACC1r, gvl );
            C2r = __riscv_vfnmsac( C2r, alphai, ACC2i, gvl );
            C2i = __riscv_vfmacc ( C2i, alphai, ACC2r, gvl );
            C3r = __riscv_vfnmsac( C3r, alphai, ACC3i, gvl );
            C3i = __riscv_vfmacc ( C3i, alphai, ACC3r, gvl );
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C0r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C0i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C1r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C1i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C2r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C2i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C3r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C3i, gvl);
            
            m_top += 8;
        }


        if( M & 4 ) {
            gvl = __riscv_vsetvl_e32m1(4);

            BLASLONG ai=m_top*K*2;
            BLASLONG bi=n_top*K*2;
            BLASLONG pass_K = K;
            #ifdef LEFT
                BLASLONG off = offset + m_top;
            #else
                BLASLONG off = -offset + n_top;
            #endif
            #ifdef BACKWARDS
                ai += off*4*2;
                bi += off*4*2;
                pass_K -= off;
            #else
                #ifdef LEFT
                    pass_K = off + 4;
                #else
                    pass_K = off + 4;
                #endif
            #endif
            float B0r = B[bi+0*2+0];
            float B0i = B[bi+0*2+1];
            float B1r = B[bi+1*2+0];
            float B1i = B[bi+1*2+1];
            float B2r = B[bi+2*2+0];
            float B2i = B[bi+2*2+1];
            float B3r = B[bi+3*2+0];
            float B3i = B[bi+3*2+1];
            bi += 4*2;

            vfloat32m1_t A0r = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2], sizeof(FLOAT)*2, gvl );
            vfloat32m1_t A0i = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2+1], sizeof(FLOAT)*2, gvl );
            ai += 4*2;

            // 2 vector regs to hold A array contents, 8 regs to hold values accumulated over k
            // leaving 22 vector registers for temporaries
            vfloat32m1_t tmp0r = __riscv_vfmul_vf_f32m1( A0i, B0i, gvl);
            vfloat32m1_t tmp0i = __riscv_vfmul_vf_f32m1( A0r, B0i, gvl);
            vfloat32m1_t tmp1r = __riscv_vfmul_vf_f32m1( A0i, B1i, gvl);
            vfloat32m1_t tmp1i = __riscv_vfmul_vf_f32m1( A0r, B1i, gvl);
            vfloat32m1_t tmp2r = __riscv_vfmul_vf_f32m1( A0i, B2i, gvl);
            vfloat32m1_t tmp2i = __riscv_vfmul_vf_f32m1( A0r, B2i, gvl);
            vfloat32m1_t tmp3r = __riscv_vfmul_vf_f32m1( A0i, B3i, gvl);
            vfloat32m1_t tmp3i = __riscv_vfmul_vf_f32m1( A0r, B3i, gvl);
            tmp0r = VFMACC_RR( tmp0r, B0r, A0r, gvl);
            tmp0i = VFMACC_RI( tmp0i, B0r, A0i, gvl);
            tmp1r = VFMACC_RR( tmp1r, B1r, A0r, gvl);
            tmp1i = VFMACC_RI( tmp1i, B1r, A0i, gvl);
            tmp2r = VFMACC_RR( tmp2r, B2r, A0r, gvl);
            tmp2i = VFMACC_RI( tmp2i, B2r, A0i, gvl);
            tmp3r = VFMACC_RR( tmp3r, B3r, A0r, gvl);
            tmp3i = VFMACC_RI( tmp3i, B3r, A0i, gvl);
            vfloat32m1_t ACC0r = tmp0r;
            vfloat32m1_t ACC0i = tmp0i;
            vfloat32m1_t ACC1r = tmp1r;
            vfloat32m1_t ACC1i = tmp1i;
            vfloat32m1_t ACC2r = tmp2r;
            vfloat32m1_t ACC2i = tmp2i;
            vfloat32m1_t ACC3r = tmp3r;
            vfloat32m1_t ACC3i = tmp3i;

            for(BLASLONG k=1; k<pass_K; k++) {
                B0r = B[bi+0*2+0];
                B0i = B[bi+0*2+1];
                B1r = B[bi+1*2+0];
                B1i = B[bi+1*2+1];
                B2r = B[bi+2*2+0];
                B2i = B[bi+2*2+1];
                B3r = B[bi+3*2+0];
                B3i = B[bi+3*2+1];
                bi += 4*2;

                A0r = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2], sizeof(FLOAT)*2, gvl );
                A0i = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2+1], sizeof(FLOAT)*2, gvl );
                ai += 4*2;

                tmp0r = __riscv_vfmul_vf_f32m1( A0i, B0i, gvl);
                tmp0i = __riscv_vfmul_vf_f32m1( A0r, B0i, gvl);
                tmp1r = __riscv_vfmul_vf_f32m1( A0i, B1i, gvl);
                tmp1i = __riscv_vfmul_vf_f32m1( A0r, B1i, gvl);
                tmp2r = __riscv_vfmul_vf_f32m1( A0i, B2i, gvl);
                tmp2i = __riscv_vfmul_vf_f32m1( A0r, B2i, gvl);
                tmp3r = __riscv_vfmul_vf_f32m1( A0i, B3i, gvl);
                tmp3i = __riscv_vfmul_vf_f32m1( A0r, B3i, gvl);
                tmp0r = VFMACC_RR( tmp0r, B0r, A0r, gvl);
                tmp0i = VFMACC_RI( tmp0i, B0r, A0i, gvl);
                tmp1r = VFMACC_RR( tmp1r, B1r, A0r, gvl);
                tmp1i = VFMACC_RI( tmp1i, B1r, A0i, gvl);
                tmp2r = VFMACC_RR( tmp2r, B2r, A0r, gvl);
                tmp2i = VFMACC_RI( tmp2i, B2r, A0i, gvl);
                tmp3r = VFMACC_RR( tmp3r, B3r, A0r, gvl);
                tmp3i = VFMACC_RI( tmp3i, B3r, A0i, gvl);
                ACC0r = __riscv_vfadd( ACC0r, tmp0r, gvl);
                ACC0i = __riscv_vfadd( ACC0i, tmp0i, gvl);
                ACC1r = __riscv_vfadd( ACC1r, tmp1r, gvl);
                ACC1i = __riscv_vfadd( ACC1i, tmp1i, gvl);
                ACC2r = __riscv_vfadd( ACC2r, tmp2r, gvl);
                ACC2i = __riscv_vfadd( ACC2i, tmp2i, gvl);
                ACC3r = __riscv_vfadd( ACC3r, tmp3r, gvl);
                ACC3i = __riscv_vfadd( ACC3i, tmp3i, gvl);
            }


            BLASLONG ci=n_top*ldc+m_top;

            vfloat32m1_t C0r = __riscv_vfmul( ACC0r, alphar, gvl );
            vfloat32m1_t C0i = __riscv_vfmul( ACC0i, alphar, gvl );
            vfloat32m1_t C1r = __riscv_vfmul( ACC1r, alphar, gvl );
            vfloat32m1_t C1i = __riscv_vfmul( ACC1i, alphar, gvl );
            vfloat32m1_t C2r = __riscv_vfmul( ACC2r, alphar, gvl );
            vfloat32m1_t C2i = __riscv_vfmul( ACC2i, alphar, gvl );
            vfloat32m1_t C3r = __riscv_vfmul( ACC3r, alphar, gvl );
            vfloat32m1_t C3i = __riscv_vfmul( ACC3i, alphar, gvl );
            C0r = __riscv_vfnmsac( C0r, alphai, ACC0i, gvl );
            C0i = __riscv_vfmacc ( C0i, alphai, ACC0r, gvl );
            C1r = __riscv_vfnmsac( C1r, alphai, ACC1i, gvl );
            C1i = __riscv_vfmacc ( C1i, alphai, ACC1r, gvl );
            C2r = __riscv_vfnmsac( C2r, alphai, ACC2i, gvl );
            C2i = __riscv_vfmacc ( C2i, alphai, ACC2r, gvl );
            C3r = __riscv_vfnmsac( C3r, alphai, ACC3i, gvl );
            C3i = __riscv_vfmacc ( C3i, alphai, ACC3r, gvl );
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C0r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C0i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C1r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C1i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C2r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C2i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C3r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C3i, gvl);
            
            m_top += 4;
        }


        if( M & 2 ) {
            gvl = __riscv_vsetvl_e32m1(2);

            BLASLONG ai=m_top*K*2;
            BLASLONG bi=n_top*K*2;
            BLASLONG pass_K = K;
            #ifdef LEFT
                BLASLONG off = offset + m_top;
            #else
                BLASLONG off = -offset + n_top;
            #endif
            #ifdef BACKWARDS
                ai += off*2*2;
                bi += off*4*2;
                pass_K -= off;
            #else
                #ifdef LEFT
                    pass_K = off + 2;
                #else
                    pass_K = off + 4;
                #endif
            #endif
            float B0r = B[bi+0*2+0];
            float B0i = B[bi+0*2+1];
            float B1r = B[bi+1*2+0];
            float B1i = B[bi+1*2+1];
            float B2r = B[bi+2*2+0];
            float B2i = B[bi+2*2+1];
            float B3r = B[bi+3*2+0];
            float B3i = B[bi+3*2+1];
            bi += 4*2;

            vfloat32m1_t A0r = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2], sizeof(FLOAT)*2, gvl );
            vfloat32m1_t A0i = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2+1], sizeof(FLOAT)*2, gvl );
            ai += 2*2;

            // 2 vector regs to hold A array contents, 8 regs to hold values accumulated over k
            // leaving 22 vector registers for temporaries
            vfloat32m1_t tmp0r = __riscv_vfmul_vf_f32m1( A0i, B0i, gvl);
            vfloat32m1_t tmp0i = __riscv_vfmul_vf_f32m1( A0r, B0i, gvl);
            vfloat32m1_t tmp1r = __riscv_vfmul_vf_f32m1( A0i, B1i, gvl);
            vfloat32m1_t tmp1i = __riscv_vfmul_vf_f32m1( A0r, B1i, gvl);
            vfloat32m1_t tmp2r = __riscv_vfmul_vf_f32m1( A0i, B2i, gvl);
            vfloat32m1_t tmp2i = __riscv_vfmul_vf_f32m1( A0r, B2i, gvl);
            vfloat32m1_t tmp3r = __riscv_vfmul_vf_f32m1( A0i, B3i, gvl);
            vfloat32m1_t tmp3i = __riscv_vfmul_vf_f32m1( A0r, B3i, gvl);
            tmp0r = VFMACC_RR( tmp0r, B0r, A0r, gvl);
            tmp0i = VFMACC_RI( tmp0i, B0r, A0i, gvl);
            tmp1r = VFMACC_RR( tmp1r, B1r, A0r, gvl);
            tmp1i = VFMACC_RI( tmp1i, B1r, A0i, gvl);
            tmp2r = VFMACC_RR( tmp2r, B2r, A0r, gvl);
            tmp2i = VFMACC_RI( tmp2i, B2r, A0i, gvl);
            tmp3r = VFMACC_RR( tmp3r, B3r, A0r, gvl);
            tmp3i = VFMACC_RI( tmp3i, B3r, A0i, gvl);
            vfloat32m1_t ACC0r = tmp0r;
            vfloat32m1_t ACC0i = tmp0i;
            vfloat32m1_t ACC1r = tmp1r;
            vfloat32m1_t ACC1i = tmp1i;
            vfloat32m1_t ACC2r = tmp2r;
            vfloat32m1_t ACC2i = tmp2i;
            vfloat32m1_t ACC3r = tmp3r;
            vfloat32m1_t ACC3i = tmp3i;

            for(BLASLONG k=1; k<pass_K; k++) {
                B0r = B[bi+0*2+0];
                B0i = B[bi+0*2+1];
                B1r = B[bi+1*2+0];
                B1i = B[bi+1*2+1];
                B2r = B[bi+2*2+0];
                B2i = B[bi+2*2+1];
                B3r = B[bi+3*2+0];
                B3i = B[bi+3*2+1];
                bi += 4*2;

                A0r = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2], sizeof(FLOAT)*2, gvl );
                A0i = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2+1], sizeof(FLOAT)*2, gvl );
                ai += 2*2;

                tmp0r = __riscv_vfmul_vf_f32m1( A0i, B0i, gvl);
                tmp0i = __riscv_vfmul_vf_f32m1( A0r, B0i, gvl);
                tmp1r = __riscv_vfmul_vf_f32m1( A0i, B1i, gvl);
                tmp1i = __riscv_vfmul_vf_f32m1( A0r, B1i, gvl);
                tmp2r = __riscv_vfmul_vf_f32m1( A0i, B2i, gvl);
                tmp2i = __riscv_vfmul_vf_f32m1( A0r, B2i, gvl);
                tmp3r = __riscv_vfmul_vf_f32m1( A0i, B3i, gvl);
                tmp3i = __riscv_vfmul_vf_f32m1( A0r, B3i, gvl);
                tmp0r = VFMACC_RR( tmp0r, B0r, A0r, gvl);
                tmp0i = VFMACC_RI( tmp0i, B0r, A0i, gvl);
                tmp1r = VFMACC_RR( tmp1r, B1r, A0r, gvl);
                tmp1i = VFMACC_RI( tmp1i, B1r, A0i, gvl);
                tmp2r = VFMACC_RR( tmp2r, B2r, A0r, gvl);
                tmp2i = VFMACC_RI( tmp2i, B2r, A0i, gvl);
                tmp3r = VFMACC_RR( tmp3r, B3r, A0r, gvl);
                tmp3i = VFMACC_RI( tmp3i, B3r, A0i, gvl);
                ACC0r = __riscv_vfadd( ACC0r, tmp0r, gvl);
                ACC0i = __riscv_vfadd( ACC0i, tmp0i, gvl);
                ACC1r = __riscv_vfadd( ACC1r, tmp1r, gvl);
                ACC1i = __riscv_vfadd( ACC1i, tmp1i, gvl);
                ACC2r = __riscv_vfadd( ACC2r, tmp2r, gvl);
                ACC2i = __riscv_vfadd( ACC2i, tmp2i, gvl);
                ACC3r = __riscv_vfadd( ACC3r, tmp3r, gvl);
                ACC3i = __riscv_vfadd( ACC3i, tmp3i, gvl);
            }


            BLASLONG ci=n_top*ldc+m_top;

            vfloat32m1_t C0r = __riscv_vfmul( ACC0r, alphar, gvl );
            vfloat32m1_t C0i = __riscv_vfmul( ACC0i, alphar, gvl );
            vfloat32m1_t C1r = __riscv_vfmul( ACC1r, alphar, gvl );
            vfloat32m1_t C1i = __riscv_vfmul( ACC1i, alphar, gvl );
            vfloat32m1_t C2r = __riscv_vfmul( ACC2r, alphar, gvl );
            vfloat32m1_t C2i = __riscv_vfmul( ACC2i, alphar, gvl );
            vfloat32m1_t C3r = __riscv_vfmul( ACC3r, alphar, gvl );
            vfloat32m1_t C3i = __riscv_vfmul( ACC3i, alphar, gvl );
            C0r = __riscv_vfnmsac( C0r, alphai, ACC0i, gvl );
            C0i = __riscv_vfmacc ( C0i, alphai, ACC0r, gvl );
            C1r = __riscv_vfnmsac( C1r, alphai, ACC1i, gvl );
            C1i = __riscv_vfmacc ( C1i, alphai, ACC1r, gvl );
            C2r = __riscv_vfnmsac( C2r, alphai, ACC2i, gvl );
            C2i = __riscv_vfmacc ( C2i, alphai, ACC2r, gvl );
            C3r = __riscv_vfnmsac( C3r, alphai, ACC3i, gvl );
            C3i = __riscv_vfmacc ( C3i, alphai, ACC3r, gvl );
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C0r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C0i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C1r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C1i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C2r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C2i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C3r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C3i, gvl);
            
            m_top += 2;
        }


        if( M & 1 ) {
            float result0 = 0;
            float result1 = 0;
            float result2 = 0;
            float result3 = 0;
            float result4 = 0;
            float result5 = 0;
            float result6 = 0;
            float result7 = 0;
            BLASLONG ai=m_top*K*2;
            BLASLONG bi=n_top*K*2;
            BLASLONG pass_K = K;
            #ifdef LEFT
                BLASLONG off = offset + m_top;
            #else
                BLASLONG off = -offset + n_top;
            #endif
            #ifdef BACKWARDS
                ai += off*1*2;
                bi += off*4*2;
                pass_K -= off;
            #else
                #ifdef LEFT
                    pass_K = off + 1;
                #else
                    pass_K = off + 4;
                #endif
            #endif

            for(BLASLONG k=0; k<pass_K; k++) {
                result0+=S0*A[ai+0+0]*B[bi+0+0] + S1*A[ai+0+1]*B[bi+0+1];
                result1+=S2*A[ai+0+1]*B[bi+0+0] + S3*A[ai+0+0]*B[bi+0+1];
                result2+=S0*A[ai+0+0]*B[bi+2+0] + S1*A[ai+0+1]*B[bi+2+1];
                result3+=S2*A[ai+0+1]*B[bi+2+0] + S3*A[ai+0+0]*B[bi+2+1];
                result4+=S0*A[ai+0+0]*B[bi+4+0] + S1*A[ai+0+1]*B[bi+4+1];
                result5+=S2*A[ai+0+1]*B[bi+4+0] + S3*A[ai+0+0]*B[bi+4+1];
                result6+=S0*A[ai+0+0]*B[bi+6+0] + S1*A[ai+0+1]*B[bi+6+1];
                result7+=S2*A[ai+0+1]*B[bi+6+0] + S3*A[ai+0+0]*B[bi+6+1];
                ai+=1*2;
                bi+=4*2;
            }

            BLASLONG ci=n_top*ldc+m_top;
            float Cr, Ci;
            Cr = result0*alphar;
            Ci = result1*alphar;
            Cr -= result1*alphai;
            Ci += result0*alphai;
            C[(ci+0*ldc+0)*2+0] = Cr;
            C[(ci+0*ldc+0)*2+1] = Ci;
            Cr = result2*alphar;
            Ci = result3*alphar;
            Cr -= result3*alphai;
            Ci += result2*alphai;
            C[(ci+1*ldc+0)*2+0] = Cr;
            C[(ci+1*ldc+0)*2+1] = Ci;
            Cr = result4*alphar;
            Ci = result5*alphar;
            Cr -= result5*alphai;
            Ci += result4*alphai;
            C[(ci+2*ldc+0)*2+0] = Cr;
            C[(ci+2*ldc+0)*2+1] = Ci;
            Cr = result6*alphar;
            Ci = result7*alphar;
            Cr -= result7*alphai;
            Ci += result6*alphai;
            C[(ci+3*ldc+0)*2+0] = Cr;
            C[(ci+3*ldc+0)*2+1] = Ci;
            m_top+=1;
        }

        n_top += 4;
    }



    // -- tails for N=2

    if( N & 2 ) {
        gvl = __riscv_vsetvl_e32m1(8);
        m_top = 0;

        for (BLASLONG i=0; i<M/8; i+=1) {
            BLASLONG ai=m_top*K*2;
            BLASLONG bi=n_top*K*2;
            BLASLONG pass_K = K;
            #ifdef LEFT
                BLASLONG off = offset + m_top;
            #else
                BLASLONG off = -offset + n_top;
            #endif
            #ifdef BACKWARDS
                ai += off*8*2;
                bi += off*2*2;
                pass_K -= off;
            #else
                #ifdef LEFT
                    pass_K = off + 8;
                #else
                    pass_K = off + 2;
                #endif
            #endif
            float B0r = B[bi+0*2+0];
            float B0i = B[bi+0*2+1];
            float B1r = B[bi+1*2+0];
            float B1i = B[bi+1*2+1];
            bi += 2*2;

            vfloat32m1_t A0r = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2], sizeof(FLOAT)*2, gvl );
            vfloat32m1_t A0i = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2+1], sizeof(FLOAT)*2, gvl );
            ai += 8*2;

            // 2 vector regs to hold A array contents, 4 regs to hold values accumulated over k
            // leaving 26 vector registers for temporaries
            vfloat32m1_t tmp0r = __riscv_vfmul_vf_f32m1( A0i, B0i, gvl);
            vfloat32m1_t tmp0i = __riscv_vfmul_vf_f32m1( A0r, B0i, gvl);
            vfloat32m1_t tmp1r = __riscv_vfmul_vf_f32m1( A0i, B1i, gvl);
            vfloat32m1_t tmp1i = __riscv_vfmul_vf_f32m1( A0r, B1i, gvl);
            tmp0r = VFMACC_RR( tmp0r, B0r, A0r, gvl);
            tmp0i = VFMACC_RI( tmp0i, B0r, A0i, gvl);
            tmp1r = VFMACC_RR( tmp1r, B1r, A0r, gvl);
            tmp1i = VFMACC_RI( tmp1i, B1r, A0i, gvl);
            vfloat32m1_t ACC0r = tmp0r;
            vfloat32m1_t ACC0i = tmp0i;
            vfloat32m1_t ACC1r = tmp1r;
            vfloat32m1_t ACC1i = tmp1i;

            for(BLASLONG k=1; k<pass_K; k++) {
                B0r = B[bi+0*2+0];
                B0i = B[bi+0*2+1];
                B1r = B[bi+1*2+0];
                B1i = B[bi+1*2+1];
                bi += 2*2;

                A0r = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2], sizeof(FLOAT)*2, gvl );
                A0i = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2+1], sizeof(FLOAT)*2, gvl );
                ai += 8*2;

                tmp0r = __riscv_vfmul_vf_f32m1( A0i, B0i, gvl);
                tmp0i = __riscv_vfmul_vf_f32m1( A0r, B0i, gvl);
                tmp1r = __riscv_vfmul_vf_f32m1( A0i, B1i, gvl);
                tmp1i = __riscv_vfmul_vf_f32m1( A0r, B1i, gvl);
                tmp0r = VFMACC_RR( tmp0r, B0r, A0r, gvl);
                tmp0i = VFMACC_RI( tmp0i, B0r, A0i, gvl);
                tmp1r = VFMACC_RR( tmp1r, B1r, A0r, gvl);
                tmp1i = VFMACC_RI( tmp1i, B1r, A0i, gvl);
                ACC0r = __riscv_vfadd( ACC0r, tmp0r, gvl);
                ACC0i = __riscv_vfadd( ACC0i, tmp0i, gvl);
                ACC1r = __riscv_vfadd( ACC1r, tmp1r, gvl);
                ACC1i = __riscv_vfadd( ACC1i, tmp1i, gvl);
            }


            BLASLONG ci=n_top*ldc+m_top;

            vfloat32m1_t C0r = __riscv_vfmul( ACC0r, alphar, gvl );
            vfloat32m1_t C0i = __riscv_vfmul( ACC0i, alphar, gvl );
            vfloat32m1_t C1r = __riscv_vfmul( ACC1r, alphar, gvl );
            vfloat32m1_t C1i = __riscv_vfmul( ACC1i, alphar, gvl );
            C0r = __riscv_vfnmsac( C0r, alphai, ACC0i, gvl );
            C0i = __riscv_vfmacc ( C0i, alphai, ACC0r, gvl );
            C1r = __riscv_vfnmsac( C1r, alphai, ACC1i, gvl );
            C1i = __riscv_vfmacc ( C1i, alphai, ACC1r, gvl );
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C0r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C0i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C1r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C1i, gvl);
            
            m_top += 8;
        }


        if( M & 4 ) {
            gvl = __riscv_vsetvl_e32m1(4);

            BLASLONG ai=m_top*K*2;
            BLASLONG bi=n_top*K*2;
            BLASLONG pass_K = K;
            #ifdef LEFT
                BLASLONG off = offset + m_top;
            #else
                BLASLONG off = -offset + n_top;
            #endif
            #ifdef BACKWARDS
                ai += off*4*2;
                bi += off*2*2;
                pass_K -= off;
            #else
                #ifdef LEFT
                    pass_K = off + 4;
                #else
                    pass_K = off + 2;
                #endif
            #endif
            float B0r = B[bi+0*2+0];
            float B0i = B[bi+0*2+1];
            float B1r = B[bi+1*2+0];
            float B1i = B[bi+1*2+1];
            bi += 2*2;

            vfloat32m1_t A0r = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2], sizeof(FLOAT)*2, gvl );
            vfloat32m1_t A0i = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2+1], sizeof(FLOAT)*2, gvl );
            ai += 4*2;

            // 2 vector regs to hold A array contents, 4 regs to hold values accumulated over k
            // leaving 26 vector registers for temporaries
            vfloat32m1_t tmp0r = __riscv_vfmul_vf_f32m1( A0i, B0i, gvl);
            vfloat32m1_t tmp0i = __riscv_vfmul_vf_f32m1( A0r, B0i, gvl);
            vfloat32m1_t tmp1r = __riscv_vfmul_vf_f32m1( A0i, B1i, gvl);
            vfloat32m1_t tmp1i = __riscv_vfmul_vf_f32m1( A0r, B1i, gvl);
            tmp0r = VFMACC_RR( tmp0r, B0r, A0r, gvl);
            tmp0i = VFMACC_RI( tmp0i, B0r, A0i, gvl);
            tmp1r = VFMACC_RR( tmp1r, B1r, A0r, gvl);
            tmp1i = VFMACC_RI( tmp1i, B1r, A0i, gvl);
            vfloat32m1_t ACC0r = tmp0r;
            vfloat32m1_t ACC0i = tmp0i;
            vfloat32m1_t ACC1r = tmp1r;
            vfloat32m1_t ACC1i = tmp1i;

            for(BLASLONG k=1; k<pass_K; k++) {
                B0r = B[bi+0*2+0];
                B0i = B[bi+0*2+1];
                B1r = B[bi+1*2+0];
                B1i = B[bi+1*2+1];
                bi += 2*2;

                A0r = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2], sizeof(FLOAT)*2, gvl );
                A0i = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2+1], sizeof(FLOAT)*2, gvl );
                ai += 4*2;

                tmp0r = __riscv_vfmul_vf_f32m1( A0i, B0i, gvl);
                tmp0i = __riscv_vfmul_vf_f32m1( A0r, B0i, gvl);
                tmp1r = __riscv_vfmul_vf_f32m1( A0i, B1i, gvl);
                tmp1i = __riscv_vfmul_vf_f32m1( A0r, B1i, gvl);
                tmp0r = VFMACC_RR( tmp0r, B0r, A0r, gvl);
                tmp0i = VFMACC_RI( tmp0i, B0r, A0i, gvl);
                tmp1r = VFMACC_RR( tmp1r, B1r, A0r, gvl);
                tmp1i = VFMACC_RI( tmp1i, B1r, A0i, gvl);
                ACC0r = __riscv_vfadd( ACC0r, tmp0r, gvl);
                ACC0i = __riscv_vfadd( ACC0i, tmp0i, gvl);
                ACC1r = __riscv_vfadd( ACC1r, tmp1r, gvl);
                ACC1i = __riscv_vfadd( ACC1i, tmp1i, gvl);
            }


            BLASLONG ci=n_top*ldc+m_top;

            vfloat32m1_t C0r = __riscv_vfmul( ACC0r, alphar, gvl );
            vfloat32m1_t C0i = __riscv_vfmul( ACC0i, alphar, gvl );
            vfloat32m1_t C1r = __riscv_vfmul( ACC1r, alphar, gvl );
            vfloat32m1_t C1i = __riscv_vfmul( ACC1i, alphar, gvl );
            C0r = __riscv_vfnmsac( C0r, alphai, ACC0i, gvl );
            C0i = __riscv_vfmacc ( C0i, alphai, ACC0r, gvl );
            C1r = __riscv_vfnmsac( C1r, alphai, ACC1i, gvl );
            C1i = __riscv_vfmacc ( C1i, alphai, ACC1r, gvl );
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C0r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C0i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C1r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C1i, gvl);
            
            m_top += 4;
        }


        if( M & 2 ) {
            gvl = __riscv_vsetvl_e32m1(2);

            BLASLONG ai=m_top*K*2;
            BLASLONG bi=n_top*K*2;
            BLASLONG pass_K = K;
            #ifdef LEFT
                BLASLONG off = offset + m_top;
            #else
                BLASLONG off = -offset + n_top;
            #endif
            #ifdef BACKWARDS
                ai += off*2*2;
                bi += off*2*2;
                pass_K -= off;
            #else
                #ifdef LEFT
                    pass_K = off + 2;
                #else
                    pass_K = off + 2;
                #endif
            #endif
            float B0r = B[bi+0*2+0];
            float B0i = B[bi+0*2+1];
            float B1r = B[bi+1*2+0];
            float B1i = B[bi+1*2+1];
            bi += 2*2;

            vfloat32m1_t A0r = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2], sizeof(FLOAT)*2, gvl );
            vfloat32m1_t A0i = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2+1], sizeof(FLOAT)*2, gvl );
            ai += 2*2;

            // 2 vector regs to hold A array contents, 4 regs to hold values accumulated over k
            // leaving 26 vector registers for temporaries
            vfloat32m1_t tmp0r = __riscv_vfmul_vf_f32m1( A0i, B0i, gvl);
            vfloat32m1_t tmp0i = __riscv_vfmul_vf_f32m1( A0r, B0i, gvl);
            vfloat32m1_t tmp1r = __riscv_vfmul_vf_f32m1( A0i, B1i, gvl);
            vfloat32m1_t tmp1i = __riscv_vfmul_vf_f32m1( A0r, B1i, gvl);
            tmp0r = VFMACC_RR( tmp0r, B0r, A0r, gvl);
            tmp0i = VFMACC_RI( tmp0i, B0r, A0i, gvl);
            tmp1r = VFMACC_RR( tmp1r, B1r, A0r, gvl);
            tmp1i = VFMACC_RI( tmp1i, B1r, A0i, gvl);
            vfloat32m1_t ACC0r = tmp0r;
            vfloat32m1_t ACC0i = tmp0i;
            vfloat32m1_t ACC1r = tmp1r;
            vfloat32m1_t ACC1i = tmp1i;

            for(BLASLONG k=1; k<pass_K; k++) {
                B0r = B[bi+0*2+0];
                B0i = B[bi+0*2+1];
                B1r = B[bi+1*2+0];
                B1i = B[bi+1*2+1];
                bi += 2*2;

                A0r = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2], sizeof(FLOAT)*2, gvl );
                A0i = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2+1], sizeof(FLOAT)*2, gvl );
                ai += 2*2;

                tmp0r = __riscv_vfmul_vf_f32m1( A0i, B0i, gvl);
                tmp0i = __riscv_vfmul_vf_f32m1( A0r, B0i, gvl);
                tmp1r = __riscv_vfmul_vf_f32m1( A0i, B1i, gvl);
                tmp1i = __riscv_vfmul_vf_f32m1( A0r, B1i, gvl);
                tmp0r = VFMACC_RR( tmp0r, B0r, A0r, gvl);
                tmp0i = VFMACC_RI( tmp0i, B0r, A0i, gvl);
                tmp1r = VFMACC_RR( tmp1r, B1r, A0r, gvl);
                tmp1i = VFMACC_RI( tmp1i, B1r, A0i, gvl);
                ACC0r = __riscv_vfadd( ACC0r, tmp0r, gvl);
                ACC0i = __riscv_vfadd( ACC0i, tmp0i, gvl);
                ACC1r = __riscv_vfadd( ACC1r, tmp1r, gvl);
                ACC1i = __riscv_vfadd( ACC1i, tmp1i, gvl);
            }


            BLASLONG ci=n_top*ldc+m_top;

            vfloat32m1_t C0r = __riscv_vfmul( ACC0r, alphar, gvl );
            vfloat32m1_t C0i = __riscv_vfmul( ACC0i, alphar, gvl );
            vfloat32m1_t C1r = __riscv_vfmul( ACC1r, alphar, gvl );
            vfloat32m1_t C1i = __riscv_vfmul( ACC1i, alphar, gvl );
            C0r = __riscv_vfnmsac( C0r, alphai, ACC0i, gvl );
            C0i = __riscv_vfmacc ( C0i, alphai, ACC0r, gvl );
            C1r = __riscv_vfnmsac( C1r, alphai, ACC1i, gvl );
            C1i = __riscv_vfmacc ( C1i, alphai, ACC1r, gvl );
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C0r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C0i, gvl);
            ci += ldc-gvl*0;
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C1r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C1i, gvl);
            
            m_top += 2;
        }


        if( M & 1 ) {
            float result0 = 0;
            float result1 = 0;
            float result2 = 0;
            float result3 = 0;
            BLASLONG ai=m_top*K*2;
            BLASLONG bi=n_top*K*2;
            BLASLONG pass_K = K;
            #ifdef LEFT
                BLASLONG off = offset + m_top;
            #else
                BLASLONG off = -offset + n_top;
            #endif
            #ifdef BACKWARDS
                ai += off*1*2;
                bi += off*2*2;
                pass_K -= off;
            #else
                #ifdef LEFT
                    pass_K = off + 1;
                #else
                    pass_K = off + 2;
                #endif
            #endif

            for(BLASLONG k=0; k<pass_K; k++) {
                result0+=S0*A[ai+0+0]*B[bi+0+0] + S1*A[ai+0+1]*B[bi+0+1];
                result1+=S2*A[ai+0+1]*B[bi+0+0] + S3*A[ai+0+0]*B[bi+0+1];
                result2+=S0*A[ai+0+0]*B[bi+2+0] + S1*A[ai+0+1]*B[bi+2+1];
                result3+=S2*A[ai+0+1]*B[bi+2+0] + S3*A[ai+0+0]*B[bi+2+1];
                ai+=1*2;
                bi+=2*2;
            }

            BLASLONG ci=n_top*ldc+m_top;
            float Cr, Ci;
            Cr = result0*alphar;
            Ci = result1*alphar;
            Cr -= result1*alphai;
            Ci += result0*alphai;
            C[(ci+0*ldc+0)*2+0] = Cr;
            C[(ci+0*ldc+0)*2+1] = Ci;
            Cr = result2*alphar;
            Ci = result3*alphar;
            Cr -= result3*alphai;
            Ci += result2*alphai;
            C[(ci+1*ldc+0)*2+0] = Cr;
            C[(ci+1*ldc+0)*2+1] = Ci;
            m_top+=1;
        }

        n_top += 2;
    }



    // -- tails for N=1

    if( N & 1 ) {
        gvl = __riscv_vsetvl_e32m1(8);
        m_top = 0;

        for (BLASLONG i=0; i<M/8; i+=1) {
            BLASLONG ai=m_top*K*2;
            BLASLONG bi=n_top*K*2;
            BLASLONG pass_K = K;
            #ifdef LEFT
                BLASLONG off = offset + m_top;
            #else
                BLASLONG off = -offset + n_top;
            #endif
            #ifdef BACKWARDS
                ai += off*8*2;
                bi += off*1*2;
                pass_K -= off;
            #else
                #ifdef LEFT
                    pass_K = off + 8;
                #else
                    pass_K = off + 1;
                #endif
            #endif
            float B0r = B[bi+0*2+0];
            float B0i = B[bi+0*2+1];
            bi += 1*2;

            vfloat32m1_t A0r = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2], sizeof(FLOAT)*2, gvl );
            vfloat32m1_t A0i = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2+1], sizeof(FLOAT)*2, gvl );
            ai += 8*2;

            // 2 vector regs to hold A array contents, 2 regs to hold values accumulated over k
            // leaving 28 vector registers for temporaries
            vfloat32m1_t tmp0r = __riscv_vfmul_vf_f32m1( A0i, B0i, gvl);
            vfloat32m1_t tmp0i = __riscv_vfmul_vf_f32m1( A0r, B0i, gvl);
            tmp0r = VFMACC_RR( tmp0r, B0r, A0r, gvl);
            tmp0i = VFMACC_RI( tmp0i, B0r, A0i, gvl);
            vfloat32m1_t ACC0r = tmp0r;
            vfloat32m1_t ACC0i = tmp0i;

            for(BLASLONG k=1; k<pass_K; k++) {
                B0r = B[bi+0*2+0];
                B0i = B[bi+0*2+1];
                bi += 1*2;

                A0r = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2], sizeof(FLOAT)*2, gvl );
                A0i = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2+1], sizeof(FLOAT)*2, gvl );
                ai += 8*2;

                tmp0r = __riscv_vfmul_vf_f32m1( A0i, B0i, gvl);
                tmp0i = __riscv_vfmul_vf_f32m1( A0r, B0i, gvl);
                tmp0r = VFMACC_RR( tmp0r, B0r, A0r, gvl);
                tmp0i = VFMACC_RI( tmp0i, B0r, A0i, gvl);
                ACC0r = __riscv_vfadd( ACC0r, tmp0r, gvl);
                ACC0i = __riscv_vfadd( ACC0i, tmp0i, gvl);
            }


            BLASLONG ci=n_top*ldc+m_top;

            vfloat32m1_t C0r = __riscv_vfmul( ACC0r, alphar, gvl );
            vfloat32m1_t C0i = __riscv_vfmul( ACC0i, alphar, gvl );
            C0r = __riscv_vfnmsac( C0r, alphai, ACC0i, gvl );
            C0i = __riscv_vfmacc ( C0i, alphai, ACC0r, gvl );
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C0r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C0i, gvl);
            
            m_top += 8;
        }


        if( M & 4 ) {
            gvl = __riscv_vsetvl_e32m1(4);

            BLASLONG ai=m_top*K*2;
            BLASLONG bi=n_top*K*2;
            BLASLONG pass_K = K;
            #ifdef LEFT
                BLASLONG off = offset + m_top;
            #else
                BLASLONG off = -offset + n_top;
            #endif
            #ifdef BACKWARDS
                ai += off*4*2;
                bi += off*1*2;
                pass_K -= off;
            #else
                #ifdef LEFT
                    pass_K = off + 4;
                #else
                    pass_K = off + 1;
                #endif
            #endif
            float B0r = B[bi+0*2+0];
            float B0i = B[bi+0*2+1];
            bi += 1*2;

            vfloat32m1_t A0r = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2], sizeof(FLOAT)*2, gvl );
            vfloat32m1_t A0i = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2+1], sizeof(FLOAT)*2, gvl );
            ai += 4*2;

            // 2 vector regs to hold A array contents, 2 regs to hold values accumulated over k
            // leaving 28 vector registers for temporaries
            vfloat32m1_t tmp0r = __riscv_vfmul_vf_f32m1( A0i, B0i, gvl);
            vfloat32m1_t tmp0i = __riscv_vfmul_vf_f32m1( A0r, B0i, gvl);
            tmp0r = VFMACC_RR( tmp0r, B0r, A0r, gvl);
            tmp0i = VFMACC_RI( tmp0i, B0r, A0i, gvl);
            vfloat32m1_t ACC0r = tmp0r;
            vfloat32m1_t ACC0i = tmp0i;

            for(BLASLONG k=1; k<pass_K; k++) {
                B0r = B[bi+0*2+0];
                B0i = B[bi+0*2+1];
                bi += 1*2;

                A0r = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2], sizeof(FLOAT)*2, gvl );
                A0i = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2+1], sizeof(FLOAT)*2, gvl );
                ai += 4*2;

                tmp0r = __riscv_vfmul_vf_f32m1( A0i, B0i, gvl);
                tmp0i = __riscv_vfmul_vf_f32m1( A0r, B0i, gvl);
                tmp0r = VFMACC_RR( tmp0r, B0r, A0r, gvl);
                tmp0i = VFMACC_RI( tmp0i, B0r, A0i, gvl);
                ACC0r = __riscv_vfadd( ACC0r, tmp0r, gvl);
                ACC0i = __riscv_vfadd( ACC0i, tmp0i, gvl);
            }


            BLASLONG ci=n_top*ldc+m_top;

            vfloat32m1_t C0r = __riscv_vfmul( ACC0r, alphar, gvl );
            vfloat32m1_t C0i = __riscv_vfmul( ACC0i, alphar, gvl );
            C0r = __riscv_vfnmsac( C0r, alphai, ACC0i, gvl );
            C0i = __riscv_vfmacc ( C0i, alphai, ACC0r, gvl );
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C0r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C0i, gvl);
            
            m_top += 4;
        }


        if( M & 2 ) {
            gvl = __riscv_vsetvl_e32m1(2);

            BLASLONG ai=m_top*K*2;
            BLASLONG bi=n_top*K*2;
            BLASLONG pass_K = K;
            #ifdef LEFT
                BLASLONG off = offset + m_top;
            #else
                BLASLONG off = -offset + n_top;
            #endif
            #ifdef BACKWARDS
                ai += off*2*2;
                bi += off*1*2;
                pass_K -= off;
            #else
                #ifdef LEFT
                    pass_K = off + 2;
                #else
                    pass_K = off + 1;
                #endif
            #endif
            float B0r = B[bi+0*2+0];
            float B0i = B[bi+0*2+1];
            bi += 1*2;

            vfloat32m1_t A0r = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2], sizeof(FLOAT)*2, gvl );
            vfloat32m1_t A0i = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2+1], sizeof(FLOAT)*2, gvl );
            ai += 2*2;

            // 2 vector regs to hold A array contents, 2 regs to hold values accumulated over k
            // leaving 28 vector registers for temporaries
            vfloat32m1_t tmp0r = __riscv_vfmul_vf_f32m1( A0i, B0i, gvl);
            vfloat32m1_t tmp0i = __riscv_vfmul_vf_f32m1( A0r, B0i, gvl);
            tmp0r = VFMACC_RR( tmp0r, B0r, A0r, gvl);
            tmp0i = VFMACC_RI( tmp0i, B0r, A0i, gvl);
            vfloat32m1_t ACC0r = tmp0r;
            vfloat32m1_t ACC0i = tmp0i;

            for(BLASLONG k=1; k<pass_K; k++) {
                B0r = B[bi+0*2+0];
                B0i = B[bi+0*2+1];
                bi += 1*2;

                A0r = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2], sizeof(FLOAT)*2, gvl );
                A0i = __riscv_vlse32_v_f32m1( &A[ai+0*gvl*2+1], sizeof(FLOAT)*2, gvl );
                ai += 2*2;

                tmp0r = __riscv_vfmul_vf_f32m1( A0i, B0i, gvl);
                tmp0i = __riscv_vfmul_vf_f32m1( A0r, B0i, gvl);
                tmp0r = VFMACC_RR( tmp0r, B0r, A0r, gvl);
                tmp0i = VFMACC_RI( tmp0i, B0r, A0i, gvl);
                ACC0r = __riscv_vfadd( ACC0r, tmp0r, gvl);
                ACC0i = __riscv_vfadd( ACC0i, tmp0i, gvl);
            }


            BLASLONG ci=n_top*ldc+m_top;

            vfloat32m1_t C0r = __riscv_vfmul( ACC0r, alphar, gvl );
            vfloat32m1_t C0i = __riscv_vfmul( ACC0i, alphar, gvl );
            C0r = __riscv_vfnmsac( C0r, alphai, ACC0i, gvl );
            C0i = __riscv_vfmacc ( C0i, alphai, ACC0r, gvl );
            __riscv_vsse32_v_f32m1( &C[ci*2+0], sizeof(FLOAT)*2, C0r, gvl);
            __riscv_vsse32_v_f32m1( &C[ci*2+1], sizeof(FLOAT)*2, C0i, gvl);
            
            m_top += 2;
        }


        if( M & 1 ) {
            float result0 = 0;
            float result1 = 0;
            BLASLONG ai=m_top*K*2;
            BLASLONG bi=n_top*K*2;
            BLASLONG pass_K = K;
            #ifdef LEFT
                BLASLONG off = offset + m_top;
            #else
                BLASLONG off = -offset + n_top;
            #endif
            #ifdef BACKWARDS
                ai += off*1*2;
                bi += off*1*2;
                pass_K -= off;
            #else
                #ifdef LEFT
                    pass_K = off + 1;
                #else
                    pass_K = off + 1;
                #endif
            #endif

            for(BLASLONG k=0; k<pass_K; k++) {
                result0+=S0*A[ai+0+0]*B[bi+0+0] + S1*A[ai+0+1]*B[bi+0+1];
                result1+=S2*A[ai+0+1]*B[bi+0+0] + S3*A[ai+0+0]*B[bi+0+1];
                ai+=1*2;
                bi+=1*2;
            }

            BLASLONG ci=n_top*ldc+m_top;
            float Cr, Ci;
            Cr = result0*alphar;
            Ci = result1*alphar;
            Cr -= result1*alphai;
            Ci += result0*alphai;
            C[(ci+0*ldc+0)*2+0] = Cr;
            C[(ci+0*ldc+0)*2+1] = Ci;
            m_top+=1;
        }

        n_top += 1;
    }

    return 0;
}