Skip to content

Commit

Permalink
Use the new tuple intrinsics to get rid of build errors in X280 BLIS …
Browse files Browse the repository at this point in the history 
…code.

However, it does not get correct results for complex BLIS routines
which use segment loads (or call those that do). The intrinsic types
check out and make sense, but it returns wrong answers. It's probably
something really simple.

For historical reference, see:

riscv-non-isa/riscv-c-api-doc#43

flame#737 (comment)

https://reviews.llvm.org/D152134

riscv-non-isa/rvv-intrinsic-doc#139
riscv-non-isa/rvv-intrinsic-doc#198

https://github.com/riscv-non-isa/rvv-intrinsic-doc/blob/master/rvv-intrinsic-rfc.md

https://github.com/riscv-non-isa/rvv-intrinsic-doc/blob/master/auto-generated/intrinsic_funcs/02_vector_unit-stride_segment_load_store_instructions_zvlsseg.md
https://github.com/riscv-non-isa/rvv-intrinsic-doc/blob/master/auto-generated/intrinsic_funcs/03_vector_stride_segment_load_store_instructions_zvlsseg.md
https://github.com/riscv-non-isa/rvv-intrinsic-doc/blob/master/auto-generated/intrinsic_funcs/04_vector_indexed_segment_load_store_instructions_zvlsseg.md
  • Loading branch information
@leekillough
leekillough committed Jul 6, 2023
1 parent 7ae57cf commit 9fc9359
Show file tree
Hide file tree
Showing 12 changed files with 226 additions and 123 deletions.
28 changes: 18 additions & 10 deletions kernels/sifive_x280/1/bli_addv_sifive_x280_intr/bli_addv_sifive_x280_intr_complex.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -47,29 +47,37 @@ ADDV(PRECISION_CHAR, void)
size_t avl = n;
while (avl) {
size_t vl = VSETVL(PREC, LMUL)(avl);
RVV_TYPE_F(PREC, LMUL) xvec_real, xvec_imag, yvec_real, yvec_imag;
RVV_TYPE_F_X2(PREC, LMUL) xvec, yvec;

if (incx == 1)
VLSEG2_V_F(PREC, LMUL)( &xvec_real, &xvec_imag, (BASE_DT*) x, vl);
xvec = VLSEG2_V_F(PREC, LMUL)( (BASE_DT*) x, vl);
else
VLSSEG2_V_F(PREC, LMUL)(&xvec_real, &xvec_imag, (BASE_DT*) x, 2*FLT_SIZE*incx, vl);
xvec = VLSSEG2_V_F(PREC, LMUL)( (BASE_DT*) x, 2*FLT_SIZE*incx, vl);

if (incy == 1)
VLSEG2_V_F(PREC, LMUL)( &yvec_real, &yvec_imag, (BASE_DT*) y, vl);
yvec = VLSEG2_V_F(PREC, LMUL)( (BASE_DT*) y, vl);
else
VLSSEG2_V_F(PREC, LMUL)(&yvec_real, &yvec_imag, (BASE_DT*) y, 2*FLT_SIZE*incy, vl);

yvec = VLSSEG2_V_F(PREC, LMUL)( (BASE_DT*) y, 2*FLT_SIZE*incy, vl);

RVV_TYPE_F(PREC, LMUL) xvec_real = RVV_GET_REAL(PREC, LMUL, xvec);
RVV_TYPE_F(PREC, LMUL) xvec_imag = RVV_GET_IMAG(PREC, LMUL, xvec);
RVV_TYPE_F(PREC, LMUL) yvec_real = RVV_GET_REAL(PREC, LMUL, yvec);
RVV_TYPE_F(PREC, LMUL) yvec_imag = RVV_GET_IMAG(PREC, LMUL, yvec);

yvec_real = VFADD_VV(PREC, LMUL)(yvec_real, xvec_real, vl);
if (conjx == BLIS_NO_CONJUGATE)
yvec_imag = VFADD_VV(PREC, LMUL)(yvec_imag, xvec_imag, vl);
else
yvec_imag = VFSUB_VV(PREC, LMUL)(yvec_imag, xvec_imag, vl);

RVV_SET_REAL(PREC, LMUL, yvec, yvec_real);
RVV_SET_IMAG(PREC, LMUL, yvec, yvec_imag);

if (incy == 1)
VSSEG2_V_F(PREC, LMUL)( (BASE_DT*) y, yvec_real, yvec_imag, vl);
VSSEG2_V_F(PREC, LMUL)( (BASE_DT*) y, yvec, vl);
else
VSSSEG2_V_F(PREC, LMUL)((BASE_DT*) y, 2*FLT_SIZE*incy, yvec_real, yvec_imag, vl);
VSSSEG2_V_F(PREC, LMUL)((BASE_DT*) y, 2*FLT_SIZE*incy, yvec, vl);

x += vl*incx;
y += vl*incy;
avl -= vl;
Expand Down
36 changes: 22 additions & 14 deletions kernels/sifive_x280/1/bli_axpbyv_sifive_x280_intr/bli_axpbyv_sifive_x280_intr_complex.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -38,7 +38,7 @@
AXPBYV(PRECISION_CHAR, void)
{
// Computes y := beta * y + alpha * conjx(x)

if (n <= 0) return;

const DATATYPE* restrict alpha = alpha_;
Expand All @@ -59,7 +59,7 @@ AXPBYV(PRECISION_CHAR, void)
return;
}

// Note: in the cases alpha = 0 && beta = 1, or alpha = 1 && beta = 0, we
// Note: in the cases alpha = 0 && beta = 1, or alpha = 1 && beta = 0, we
// will canonicalize NaNs whereas the reference code will propagate NaN payloads.

// TO DO (optimization): special cases for alpha = +-1, +-i, beta = +-1, +-i
Expand All @@ -68,23 +68,28 @@ AXPBYV(PRECISION_CHAR, void)
size_t avl = n;
while (avl) {
size_t vl = VSETVL(PREC, LMUL)(avl);
RVV_TYPE_F(PREC, LMUL) xvec_real, xvec_imag, yvec_real, yvec_imag, temp_real, temp_imag;
RVV_TYPE_F_X2(PREC, LMUL) xvec, yvec;

if (incx == 1)
VLSEG2_V_F(PREC, LMUL)( &xvec_real, &xvec_imag, (BASE_DT*) x, vl);
xvec = VLSEG2_V_F(PREC, LMUL)( (BASE_DT*) x, vl);
else
VLSSEG2_V_F(PREC, LMUL)(&xvec_real, &xvec_imag, (BASE_DT*) x, 2*FLT_SIZE*incx, vl);
xvec = VLSSEG2_V_F(PREC, LMUL)( (BASE_DT*) x, 2*FLT_SIZE*incx, vl);

if (incy == 1)
VLSEG2_V_F(PREC, LMUL)( &yvec_real, &yvec_imag, (BASE_DT*) y, vl);
yvec = VLSEG2_V_F(PREC, LMUL)( (BASE_DT*) y, vl);
else
VLSSEG2_V_F(PREC, LMUL)(&yvec_real, &yvec_imag, (BASE_DT*) y, 2*FLT_SIZE*incy, vl);

yvec = VLSSEG2_V_F(PREC, LMUL)( (BASE_DT*) y, 2*FLT_SIZE*incy, vl);

RVV_TYPE_F(PREC, LMUL) xvec_real = RVV_GET_REAL(PREC, LMUL, xvec);
RVV_TYPE_F(PREC, LMUL) xvec_imag = RVV_GET_IMAG(PREC, LMUL, xvec);
RVV_TYPE_F(PREC, LMUL) yvec_real = RVV_GET_REAL(PREC, LMUL, yvec);
RVV_TYPE_F(PREC, LMUL) yvec_imag = RVV_GET_IMAG(PREC, LMUL, yvec);

// Computed as:
// y.real = beta.real * y.real - beta.imag * y.imag + alpha.real * x.real - alpha.imag * conj(x.imag)
// y.imag = beta.real * y.imag + beta.imag * y.real + alpha.imag * x.real + alpha.real * conj(x.imag)
temp_real = VFMUL_VF(PREC, LMUL) (yvec_real, beta->real, vl);
temp_imag = VFMUL_VF(PREC, LMUL) (yvec_imag, beta->real, vl);
RVV_TYPE_F(PREC, LMUL) temp_real = VFMUL_VF(PREC, LMUL) (yvec_real, beta->real, vl);
RVV_TYPE_F(PREC, LMUL) temp_imag = VFMUL_VF(PREC, LMUL) (yvec_imag, beta->real, vl);
temp_real = VFNMSAC_VF(PREC, LMUL)(temp_real, beta->imag, yvec_imag, vl);
temp_imag = VFMACC_VF(PREC, LMUL) (temp_imag, beta->imag, yvec_real, vl);
yvec_real = VFMACC_VF(PREC, LMUL) (temp_real, alpha->real, xvec_real, vl);
Expand All @@ -97,11 +102,14 @@ AXPBYV(PRECISION_CHAR, void)
yvec_imag = VFNMSAC_VF(PREC, LMUL)(yvec_imag, alpha->real, xvec_imag, vl);
}

RVV_SET_REAL(PREC, LMUL, yvec, yvec_real);
RVV_SET_IMAG(PREC, LMUL, yvec, yvec_imag);

if (incy == 1)
VSSEG2_V_F(PREC, LMUL)( (BASE_DT*) y, yvec_real, yvec_imag, vl);
VSSEG2_V_F(PREC, LMUL)( (BASE_DT*) y, yvec, vl);
else
VSSSEG2_V_F(PREC, LMUL)((BASE_DT*) y, 2*FLT_SIZE*incy, yvec_real, yvec_imag, vl);
VSSSEG2_V_F(PREC, LMUL)((BASE_DT*) y, 2*FLT_SIZE*incy, yvec, vl);

x += vl*incx;
y += vl*incy;
avl -= vl;
Expand Down
31 changes: 20 additions & 11 deletions kernels/sifive_x280/1/bli_axpyv_sifive_x280_intr/bli_axpyv_sifive_x280_intr_complex.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -41,25 +41,30 @@ AXPYV(PRECISION_CHAR, void)
const DATATYPE* restrict alpha = alpha_;
const DATATYPE* restrict x = x_;
DATATYPE* restrict y = y_;

if (n <= 0) return;
if (alpha->real == 0 && alpha->imag == 0) return;

size_t avl = n;
while (avl) {
size_t vl = VSETVL(PREC, LMUL)(avl);
RVV_TYPE_F(PREC, LMUL) xvec_real, xvec_imag, yvec_real, yvec_imag;
RVV_TYPE_F_X2(PREC, LMUL) xvec, yvec;

if (incx == 1)
VLSEG2_V_F(PREC, LMUL)( &xvec_real, &xvec_imag, (BASE_DT*) x, vl);
xvec = VLSEG2_V_F(PREC, LMUL)( (BASE_DT*) x, vl);
else
VLSSEG2_V_F(PREC, LMUL)(&xvec_real, &xvec_imag, (BASE_DT*) x, 2*FLT_SIZE*incx, vl);
xvec = VLSSEG2_V_F(PREC, LMUL)( (BASE_DT*) x, 2*FLT_SIZE*incx, vl);

if (incy == 1)
VLSEG2_V_F(PREC, LMUL)( &yvec_real, &yvec_imag, (BASE_DT*) y, vl);
yvec = VLSEG2_V_F(PREC, LMUL)( (BASE_DT*) y, vl);
else
VLSSEG2_V_F(PREC, LMUL)(&yvec_real, &yvec_imag, (BASE_DT*) y, 2*FLT_SIZE*incy, vl);

yvec = VLSSEG2_V_F(PREC, LMUL)( (BASE_DT*) y, 2*FLT_SIZE*incy, vl);

RVV_TYPE_F(PREC, LMUL) xvec_real = RVV_GET_REAL(PREC, LMUL, xvec);
RVV_TYPE_F(PREC, LMUL) xvec_imag = RVV_GET_IMAG(PREC, LMUL, xvec);
RVV_TYPE_F(PREC, LMUL) yvec_real = RVV_GET_REAL(PREC, LMUL, yvec);
RVV_TYPE_F(PREC, LMUL) yvec_imag = RVV_GET_IMAG(PREC, LMUL, yvec);

yvec_real = VFMACC_VF(PREC, LMUL)( yvec_real, alpha->real, xvec_real, vl);
yvec_imag = VFMACC_VF(PREC, LMUL)( yvec_imag, alpha->imag, xvec_real, vl);
if (conjx == BLIS_NO_CONJUGATE){
Expand All @@ -70,11 +75,15 @@ AXPYV(PRECISION_CHAR, void)
yvec_imag = VFNMSAC_VF(PREC, LMUL)(yvec_imag, alpha->real, xvec_imag, vl);
}


RVV_SET_REAL(PREC, LMUL, yvec, yvec_real);
RVV_SET_IMAG(PREC, LMUL, yvec, yvec_imag);

if (incy == 1)
VSSEG2_V_F(PREC, LMUL)( (BASE_DT*) y, yvec_real, yvec_imag, vl);
VSSEG2_V_F(PREC, LMUL)( (BASE_DT*) y, yvec, vl);
else
VSSSEG2_V_F(PREC, LMUL)((BASE_DT*) y, 2*FLT_SIZE*incy, yvec_real, yvec_imag, vl);
VSSSEG2_V_F(PREC, LMUL)((BASE_DT*) y, 2*FLT_SIZE*incy, yvec, vl);

x += vl*incx;
y += vl*incy;
avl -= vl;
Expand Down
22 changes: 13 additions & 9 deletions kernels/sifive_x280/1/bli_dotv_sifive_x280_intr/bli_dotv_sifive_x280_intr_complex.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -42,7 +42,7 @@ DOTV(PRECISION_CHAR, void)
DATATYPE* restrict rho = rho_;
const DATATYPE* restrict x = x_;
const DATATYPE* restrict y = y_;

if (n <= 0) {
rho->real = 0;
rho->imag = 0;
Expand All @@ -60,18 +60,23 @@ DOTV(PRECISION_CHAR, void)
bool first = true;
while (avl) {
size_t vl = VSETVL(PREC, LMUL)(avl);
RVV_TYPE_F(PREC, LMUL) xvec_real, xvec_imag, yvec_real, yvec_imag;
RVV_TYPE_F_X2(PREC, LMUL) xvec, yvec;

if (incx == 1)
VLSEG2_V_F(PREC, LMUL)( &xvec_real, &xvec_imag, (BASE_DT*) x, vl);
xvec = VLSEG2_V_F(PREC, LMUL)( (BASE_DT*) x, vl);
else
VLSSEG2_V_F(PREC, LMUL)(&xvec_real, &xvec_imag, (BASE_DT*) x, 2*FLT_SIZE*incx, vl);
xvec = VLSSEG2_V_F(PREC, LMUL)( (BASE_DT*) x, 2*FLT_SIZE*incx, vl);

if (incy == 1)
VLSEG2_V_F(PREC, LMUL)( &yvec_real, &yvec_imag, (BASE_DT*) y, vl);
yvec = VLSEG2_V_F(PREC, LMUL)( (BASE_DT*) y, vl);
else
VLSSEG2_V_F(PREC, LMUL)(&yvec_real, &yvec_imag, (BASE_DT*) y, 2*FLT_SIZE*incy, vl);

yvec = VLSSEG2_V_F(PREC, LMUL)( (BASE_DT*) y, 2*FLT_SIZE*incy, vl);

RVV_TYPE_F(PREC, LMUL) xvec_real = RVV_GET_REAL(PREC, LMUL, xvec);
RVV_TYPE_F(PREC, LMUL) xvec_imag = RVV_GET_IMAG(PREC, LMUL, xvec);
RVV_TYPE_F(PREC, LMUL) yvec_real = RVV_GET_REAL(PREC, LMUL, yvec);
RVV_TYPE_F(PREC, LMUL) yvec_imag = RVV_GET_IMAG(PREC, LMUL, yvec);

if (first) {
acc_real = VFMUL_VV(PREC, LMUL)(xvec_real, yvec_real, vl);
acc_imag = VFMUL_VV(PREC, LMUL)(xvec_imag, yvec_real, vl);
Expand All @@ -93,7 +98,6 @@ DOTV(PRECISION_CHAR, void)
avl -= vl;
}


RVV_TYPE_F(PREC, m1) sum_real = VFMV_S_F(PREC, m1)(0.f, 1);
RVV_TYPE_F(PREC, m1) sum_imag = VFMV_S_F(PREC, m1)(0.f, 1);
sum_real = VF_REDUSUM_VS(PREC, LMUL)(acc_real, sum_real, n);
Expand Down
22 changes: 13 additions & 9 deletions kernels/sifive_x280/1/bli_dotxv_sifive_x280_intr/bli_dotxv_sifive_x280_intr_complex.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -44,7 +44,7 @@ DOTXV(PRECISION_CHAR, void)
DATATYPE* restrict rho = rho_;
const DATATYPE* restrict x = x_;
const DATATYPE* restrict y = y_;

if (beta->real == 0 && beta->imag == 0){
rho->real = 0;
rho->imag = 0;
Expand All @@ -69,18 +69,23 @@ DOTXV(PRECISION_CHAR, void)
bool first = true;
while (avl) {
size_t vl = VSETVL(PREC, LMUL)(avl);
RVV_TYPE_F(PREC, LMUL) xvec_real, xvec_imag, yvec_real, yvec_imag;
RVV_TYPE_F_X2(PREC, LMUL) xvec, yvec;

if (incx == 1)
VLSEG2_V_F(PREC, LMUL)( &xvec_real, &xvec_imag, (BASE_DT*) x, vl);
xvec = VLSEG2_V_F(PREC, LMUL)( (BASE_DT*) x, vl);
else
VLSSEG2_V_F(PREC, LMUL)(&xvec_real, &xvec_imag, (BASE_DT*) x, 2*FLT_SIZE*incx, vl);
xvec = VLSSEG2_V_F(PREC, LMUL)( (BASE_DT*) x, 2*FLT_SIZE*incx, vl);

if (incy == 1)
VLSEG2_V_F(PREC, LMUL)( &yvec_real, &yvec_imag, (BASE_DT*) y, vl);
yvec = VLSEG2_V_F(PREC, LMUL)( (BASE_DT*) y, vl);
else
VLSSEG2_V_F(PREC, LMUL)(&yvec_real, &yvec_imag, (BASE_DT*) y, 2*FLT_SIZE*incy, vl);

yvec = VLSSEG2_V_F(PREC, LMUL)( (BASE_DT*) y, 2*FLT_SIZE*incy, vl);

RVV_TYPE_F(PREC, LMUL) xvec_real = RVV_GET_REAL(PREC, LMUL, xvec);
RVV_TYPE_F(PREC, LMUL) xvec_imag = RVV_GET_IMAG(PREC, LMUL, xvec);
RVV_TYPE_F(PREC, LMUL) yvec_real = RVV_GET_REAL(PREC, LMUL, yvec);
RVV_TYPE_F(PREC, LMUL) yvec_imag = RVV_GET_IMAG(PREC, LMUL, yvec);

if (first) {
acc_real = VFMUL_VV(PREC, LMUL)(xvec_real, yvec_real, vl);
acc_imag = VFMUL_VV(PREC, LMUL)(xvec_imag, yvec_real, vl);
Expand All @@ -102,7 +107,6 @@ DOTXV(PRECISION_CHAR, void)
avl -= vl;
}


RVV_TYPE_F(PREC, m1) sum_real = VFMV_S_F(PREC, m1)(0.f, 1);
RVV_TYPE_F(PREC, m1) sum_imag = VFMV_S_F(PREC, m1)(0.f, 1);
sum_real = VF_REDUSUM_VS(PREC, LMUL)(acc_real, sum_real, n);
Expand Down
25 changes: 17 additions & 8 deletions kernels/sifive_x280/1/bli_scal2v_sifive_x280_intr/bli_scal2v_sifive_x280_intr_complex.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -41,7 +41,7 @@ SCAL2V(PRECISION_CHAR, void)
const DATATYPE* restrict alpha = alpha_;
const DATATYPE* restrict x = x_;
DATATYPE* restrict y = y_;

if (n <= 0) return;
if (alpha->real == 0 && alpha->imag == 0) {
SETV(PRECISION_CHAR)(BLIS_NO_CONJUGATE, n, alpha, y, incy, cntx);
Expand All @@ -56,13 +56,18 @@ SCAL2V(PRECISION_CHAR, void)
size_t avl = n;
while (avl) {
size_t vl = VSETVL(PREC, LMUL)(avl);
RVV_TYPE_F(PREC, LMUL) xvec_real, xvec_imag, yvec_real, yvec_imag;
RVV_TYPE_F_X2(PREC, LMUL) xvec, yvec;

if (incx == 1)
VLSEG2_V_F(PREC, LMUL)( &xvec_real, &xvec_imag, (BASE_DT*) x, vl);
xvec = VLSEG2_V_F(PREC, LMUL)( (BASE_DT*) x, vl);
else
VLSSEG2_V_F(PREC, LMUL)(&xvec_real, &xvec_imag, (BASE_DT*) x, 2*FLT_SIZE*incx, vl);

xvec = VLSSEG2_V_F(PREC, LMUL)( (BASE_DT*) x, 2*FLT_SIZE*incx, vl);

RVV_TYPE_F(PREC, LMUL) xvec_real = RVV_GET_REAL(PREC, LMUL, xvec);
RVV_TYPE_F(PREC, LMUL) xvec_imag = RVV_GET_IMAG(PREC, LMUL, xvec);
RVV_TYPE_F(PREC, LMUL) yvec_real = RVV_GET_REAL(PREC, LMUL, yvec);
RVV_TYPE_F(PREC, LMUL) yvec_imag = RVV_GET_IMAG(PREC, LMUL, yvec);

yvec_real = VFMUL_VF(PREC, LMUL)(xvec_real, alpha->real, vl);
yvec_imag = VFMUL_VF(PREC, LMUL)(xvec_real, alpha->imag, vl);
if (conjx == BLIS_NO_CONJUGATE) {
Expand All @@ -73,11 +78,15 @@ SCAL2V(PRECISION_CHAR, void)
yvec_imag = VFNMSAC_VF(PREC, LMUL)(yvec_imag, alpha->real, xvec_imag, vl);
}

RVV_SET_REAL(PREC, LMUL, yvec, yvec_real);
RVV_SET_IMAG(PREC, LMUL, yvec, yvec_imag);
#pragma GCC diagnostic ignored "-Wuninitialized"

if (incy == 1)
VSSEG2_V_F(PREC, LMUL)( (BASE_DT*) y, yvec_real, yvec_imag, vl);
VSSEG2_V_F(PREC, LMUL)( (BASE_DT*) y, yvec, vl);
else
VSSSEG2_V_F(PREC, LMUL)((BASE_DT*) y, 2*FLT_SIZE*incy, yvec_real, yvec_imag, vl);
VSSSEG2_V_F(PREC, LMUL)((BASE_DT*) y, 2*FLT_SIZE*incy, yvec, vl);

x += vl*incx;
y += vl*incy;
avl -= vl;
Expand Down
25 changes: 16 additions & 9 deletions kernels/sifive_x280/1/bli_scalv_sifive_x280_intr/bli_scalv_sifive_x280_intr_complex.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -40,7 +40,7 @@ SCALV(PRECISION_CHAR, void)
// Computes x = conjalpha(alpha) * x
const DATATYPE* restrict alpha = alpha_;
DATATYPE* restrict x = x_;

if (n <= 0 || (alpha->real == 1 && alpha->imag == 0)) return;

if (alpha->real == 0 && alpha->imag==0){
Expand All @@ -51,13 +51,16 @@ SCALV(PRECISION_CHAR, void)
size_t avl = n;
while (avl) {
size_t vl = VSETVL(PREC, LMUL)(avl);
RVV_TYPE_F(PREC, LMUL) xvec_real, xvec_imag;
RVV_TYPE_F_X2(PREC, LMUL) xvec;

if (incx == 1)
VLSEG2_V_F(PREC, LMUL)( &xvec_real, &xvec_imag, (BASE_DT*) x, vl);
xvec = VLSEG2_V_F(PREC, LMUL)( (BASE_DT*) x, vl);
else
VLSSEG2_V_F(PREC, LMUL)(&xvec_real, &xvec_imag, (BASE_DT*) x, 2*FLT_SIZE*incx, vl);

xvec = VLSSEG2_V_F(PREC, LMUL)( (BASE_DT*) x, 2*FLT_SIZE*incx, vl);

RVV_TYPE_F(PREC, LMUL) xvec_real = RVV_GET_REAL(PREC, LMUL, xvec);
RVV_TYPE_F(PREC, LMUL) xvec_imag = RVV_GET_IMAG(PREC, LMUL, xvec);

RVV_TYPE_F(PREC, LMUL) temp_real = VFMUL_VF(PREC, LMUL)(xvec_real, alpha->real, vl);
RVV_TYPE_F(PREC, LMUL) temp_imag = VFMUL_VF(PREC, LMUL)(xvec_imag, alpha->real, vl);
if (conjalpha == BLIS_NO_CONJUGATE) {
Expand All @@ -67,13 +70,17 @@ SCALV(PRECISION_CHAR, void)
temp_real = VFMACC_VF(PREC, LMUL) (temp_real, alpha->imag, xvec_imag, vl);
temp_imag = VFNMSAC_VF(PREC, LMUL)(temp_imag, alpha->imag, xvec_real, vl);
}


RVV_TYPE_F_X2(PREC, LMUL) temp;
RVV_SET_REAL(PREC, LMUL, temp, temp_real);
RVV_SET_IMAG(PREC, LMUL, temp, temp_imag);
#pragma GCC diagnostic ignored "-Wuninitialized"

if (incx == 1)
VSSEG2_V_F(PREC, LMUL)( (BASE_DT*) x, temp_real, temp_imag, vl);
VSSEG2_V_F(PREC, LMUL)( (BASE_DT*) x, temp, vl);
else
VSSSEG2_V_F(PREC, LMUL)((BASE_DT*) x, 2*FLT_SIZE*incx, temp_real, temp_imag, vl);
VSSSEG2_V_F(PREC, LMUL)((BASE_DT*) x, 2*FLT_SIZE*incx, temp, vl);

x += vl*incx;
avl -= vl;
}
Expand Down
Loading

0 comments on commit 9fc9359

Please sign in to comment.