Skip to content

kernel/riscv64: fold the complex MAC into the accumulators in ZGEMM #6051

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
6eanut wants to merge 1 commit into
base: develop
Choose a base branch
from
+57 −52
Open

kernel/riscv64: fold the complex MAC into the accumulators in ZGEMM #6051

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
109 changes: 57 additions & 52 deletions kernel/riscv64/zgemm_kernel_8x4_zvl256b.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -49,6 +49,8 @@ AUTOGENERATED KERNEL
#define S3 1
#define VFMACC_RR __riscv_vfmsac
#define VFMACC_RI __riscv_vfmacc
#define FOLD_OPA __riscv_vfmacc
#define FOLD_OPB __riscv_vfmacc
#endif
#if defined(NR) || defined(NC) || defined(TR) || defined(TC)
#define S0 1
Expand All @@ -57,6 +59,8 @@ AUTOGENERATED KERNEL
#define S3 -1
#define VFMACC_RR __riscv_vfmacc
#define VFMACC_RI __riscv_vfmsac
#define FOLD_OPA __riscv_vfmacc
#define FOLD_OPB __riscv_vfnmsac
#endif
#if defined(RN) || defined(RT) || defined(CN) || defined(CT)
#define S0 1
Expand All @@ -65,6 +69,8 @@ AUTOGENERATED KERNEL
#define S3 1
#define VFMACC_RR __riscv_vfmacc
#define VFMACC_RI __riscv_vfnmsac
#define FOLD_OPA __riscv_vfnmsac
#define FOLD_OPB __riscv_vfmacc
#endif
#if defined(RR) || defined(RC) || defined(CR) || defined(CC)
#define S0 1
Expand All @@ -73,6 +79,8 @@ AUTOGENERATED KERNEL
#define S3 -1
#define VFMACC_RR __riscv_vfmsac
#define VFMACC_RI __riscv_vfnmacc
#define FOLD_OPA __riscv_vfnmsac
#define FOLD_OPB __riscv_vfnmsac
#endif

int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alphar, FLOAT alphai, FLOAT* A, FLOAT* B, FLOAT* C, BLASLONG ldc)
Expand Down Expand Up @@ -172,60 +180,57 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alphar, FLOAT alphai, FLOAT*
B3i = B[bi+3*2+1];
bi += 4*2;

A0r = __riscv_vlse64_v_f64m1( &A[ai+0*gvl*2], sizeof(FLOAT)*2, gvl );
A0i = __riscv_vlse64_v_f64m1( &A[ai+0*gvl*2+1], sizeof(FLOAT)*2, gvl );
A1r = __riscv_vlse64_v_f64m1( &A[ai+1*gvl*2], sizeof(FLOAT)*2, gvl );
A1i = __riscv_vlse64_v_f64m1( &A[ai+1*gvl*2+1], sizeof(FLOAT)*2, gvl );
{
vfloat64m1x2_t Aseg0 = __riscv_vlseg2e64_v_f64m1x2( &A[ai+0*gvl*2], gvl );
A0r = __riscv_vget_v_f64m1x2_f64m1( Aseg0, 0 );
A0i = __riscv_vget_v_f64m1x2_f64m1( Aseg0, 1 );
vfloat64m1x2_t Aseg1 = __riscv_vlseg2e64_v_f64m1x2( &A[ai+1*gvl*2], gvl );
A1r = __riscv_vget_v_f64m1x2_f64m1( Aseg1, 0 );
A1i = __riscv_vget_v_f64m1x2_f64m1( Aseg1, 1 );
}
ai += 8*2;

tmp0r = __riscv_vfmul_vf_f64m1( A0i, B0i, gvl);
tmp0i = __riscv_vfmul_vf_f64m1( A0r, B0i, gvl);
tmp1r = __riscv_vfmul_vf_f64m1( A1i, B0i, gvl);
tmp1i = __riscv_vfmul_vf_f64m1( A1r, B0i, gvl);
tmp2r = __riscv_vfmul_vf_f64m1( A0i, B1i, gvl);
tmp2i = __riscv_vfmul_vf_f64m1( A0r, B1i, gvl);
tmp3r = __riscv_vfmul_vf_f64m1( A1i, B1i, gvl);
tmp3i = __riscv_vfmul_vf_f64m1( A1r, B1i, gvl);
tmp0r = VFMACC_RR( tmp0r, B0r, A0r, gvl);
tmp0i = VFMACC_RI( tmp0i, B0r, A0i, gvl);
tmp1r = VFMACC_RR( tmp1r, B0r, A1r, gvl);
tmp1i = VFMACC_RI( tmp1i, B0r, A1i, gvl);
tmp2r = VFMACC_RR( tmp2r, B1r, A0r, gvl);
tmp2i = VFMACC_RI( tmp2i, B1r, A0i, gvl);
tmp3r = VFMACC_RR( tmp3r, B1r, A1r, gvl);
tmp3i = VFMACC_RI( tmp3i, B1r, A1i, 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_f64m1( A0i, B2i, gvl);
tmp0i = __riscv_vfmul_vf_f64m1( A0r, B2i, gvl);
tmp1r = __riscv_vfmul_vf_f64m1( A1i, B2i, gvl);
tmp1i = __riscv_vfmul_vf_f64m1( A1r, B2i, gvl);
tmp2r = __riscv_vfmul_vf_f64m1( A0i, B3i, gvl);
tmp2i = __riscv_vfmul_vf_f64m1( A0r, B3i, gvl);
tmp3r = __riscv_vfmul_vf_f64m1( A1i, B3i, gvl);
tmp3i = __riscv_vfmul_vf_f64m1( A1r, B3i, gvl);
tmp0r = VFMACC_RR( tmp0r, B2r, A0r, gvl);
tmp0i = VFMACC_RI( tmp0i, B2r, A0i, gvl);
tmp1r = VFMACC_RR( tmp1r, B2r, A1r, gvl);
tmp1i = VFMACC_RI( tmp1i, B2r, A1i, gvl);
tmp2r = VFMACC_RR( tmp2r, B3r, A0r, gvl);
tmp2i = VFMACC_RI( tmp2i, B3r, A0i, gvl);
tmp3r = VFMACC_RR( tmp3r, B3r, A1r, gvl);
tmp3i = VFMACC_RI( tmp3i, B3r, A1i, 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);
/* Fold the complex MAC into the accumulators: two signed FMAs
* per accumulator replace the tmp materialisation
* (vfmul + VFMACC_RR/RI) followed by vfadd. Applying the
* same op twice composes to a plain accumulation because
* vfmsac(x,u,v)=u*v-x and vfmsac(vfmsac(a,u,v),w,y)=a+w*y-u*v
* (and likewise for vfmacc / vfnmsac), so the k-step value
* added to each accumulator is identical to the original
* tmp += vfadd sequence, and the per-column accumulation
* order is unchanged. */
ACC0r = VFMACC_RR( ACC0r, B0i, A0i, gvl);
ACC0r = VFMACC_RR( ACC0r, B0r, A0r, gvl);
ACC0i = FOLD_OPA( ACC0i, B0r, A0i, gvl);
ACC0i = FOLD_OPB( ACC0i, B0i, A0r, gvl);
ACC1r = VFMACC_RR( ACC1r, B0i, A1i, gvl);
ACC1r = VFMACC_RR( ACC1r, B0r, A1r, gvl);
ACC1i = FOLD_OPA( ACC1i, B0r, A1i, gvl);
ACC1i = FOLD_OPB( ACC1i, B0i, A1r, gvl);
ACC2r = VFMACC_RR( ACC2r, B1i, A0i, gvl);
ACC2r = VFMACC_RR( ACC2r, B1r, A0r, gvl);
ACC2i = FOLD_OPA( ACC2i, B1r, A0i, gvl);
ACC2i = FOLD_OPB( ACC2i, B1i, A0r, gvl);
ACC3r = VFMACC_RR( ACC3r, B1i, A1i, gvl);
ACC3r = VFMACC_RR( ACC3r, B1r, A1r, gvl);
ACC3i = FOLD_OPA( ACC3i, B1r, A1i, gvl);
ACC3i = FOLD_OPB( ACC3i, B1i, A1r, gvl);
ACC4r = VFMACC_RR( ACC4r, B2i, A0i, gvl);
ACC4r = VFMACC_RR( ACC4r, B2r, A0r, gvl);
ACC4i = FOLD_OPA( ACC4i, B2r, A0i, gvl);
ACC4i = FOLD_OPB( ACC4i, B2i, A0r, gvl);
ACC5r = VFMACC_RR( ACC5r, B2i, A1i, gvl);
ACC5r = VFMACC_RR( ACC5r, B2r, A1r, gvl);
ACC5i = FOLD_OPA( ACC5i, B2r, A1i, gvl);
ACC5i = FOLD_OPB( ACC5i, B2i, A1r, gvl);
ACC6r = VFMACC_RR( ACC6r, B3i, A0i, gvl);
ACC6r = VFMACC_RR( ACC6r, B3r, A0r, gvl);
ACC6i = FOLD_OPA( ACC6i, B3r, A0i, gvl);
ACC6i = FOLD_OPB( ACC6i, B3i, A0r, gvl);
ACC7r = VFMACC_RR( ACC7r, B3i, A1i, gvl);
ACC7r = VFMACC_RR( ACC7r, B3r, A1r, gvl);
ACC7i = FOLD_OPA( ACC7i, B3r, A1i, gvl);
ACC7i = FOLD_OPB( ACC7i, B3i, A1r, gvl);
}


Expand Down
Loading