Skip to content

Commit

Permalink
make scalar op
Browse files Browse the repository at this point in the history
  • Loading branch information
@mblum94 @uecker
mblum94 authored and uecker committed Dec 9, 2024
1 parent de26a78 commit 44513e7
Showing 1 changed file with 15 additions and 67 deletions.
82 changes: 15 additions & 67 deletions src/num/flpmath.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -395,58 +395,6 @@ static void* gpu_constant(const void* vp, size_t size)
#endif

static void make_z3op_scalar(md_z3op_t fun, int D, const long dims[D], const long ostr[D], complex float* optr, const long istr[D], const complex float* iptr, complex float val)
{
size_t size = CFL_SIZE;
unsigned long flags = 0;

int ND = MIN(md_calc_blockdim(D, dims, istr, size), md_calc_blockdim(D, dims, ostr, size));

for (int i = 0; i < ND; i++)
if (1024 * 1024 >= md_calc_size(i, dims))
flags = MD_SET(flags, i);


long tdims[D];
long strs1[D];

md_select_dims(D, flags, tdims, dims);
md_calc_strides(D, strs1, tdims, size);

complex float* valp = md_alloc_sameplace(D, tdims, size, optr);
md_zfill(D, tdims, valp, val);

fun(D, dims, ostr, optr, istr, iptr, strs1, valp);

md_free(valp);
}

static void make_3op_scalar(md_3op_t fun, int D, const long dims[D], const long ostr[D], float* optr, const long istr[D], const float* iptr, float val)
{
size_t size = FL_SIZE;
unsigned long flags = 0;

int ND = MIN(md_calc_blockdim(D, dims, istr, size), md_calc_blockdim(D, dims, ostr, size));

for (int i = 0; i < ND; i++)
if (1024 * 1024 >= md_calc_size(i, dims))
flags = MD_SET(flags, i);


long tdims[D];
long strs1[D];

md_select_dims(D, flags, tdims, dims);
md_calc_strides(D, strs1, tdims, size);

float* valp = md_alloc_sameplace(D, tdims, size, optr);
md_fill(D, tdims, valp, &val, size);

fun(D, dims, ostr, optr, istr, iptr, strs1, valp);

md_free(valp);
}

static void make_z3op_scalar_direct(md_z3op_t fun, int D, const long dims[D], const long ostr[D], complex float* optr, const long istr[D], const complex float* iptr, complex float val)
{
complex float* valp = md_alloc_sameplace(D, MD_SINGLETON_DIMS(D), CFL_SIZE, optr);
md_copy(D, MD_SINGLETON_DIMS(D), valp, &val, CFL_SIZE);
Expand All @@ -460,7 +408,7 @@ static void make_z3op_scalar_direct(md_z3op_t fun, int D, const long dims[D], co
}


static void make_3op_scalar_direct(md_3op_t fun, int D, const long dims[D], const long ostr[D], float* optr, const long istr[D], const float* iptr, float val)
static void make_3op_scalar(md_3op_t fun, int D, const long dims[D], const long ostr[D], float* optr, const long istr[D], const float* iptr, float val)
{
float* valp = md_alloc_sameplace(D, MD_SINGLETON_DIMS(D), FL_SIZE, optr);
md_copy(D, MD_SINGLETON_DIMS(D), valp, &val, FL_SIZE);
Expand Down Expand Up @@ -682,7 +630,7 @@ void md_zsmul2(int D, const long dims[D], const long ostr[D], complex float* opt
return;
}

make_z3op_scalar_direct(md_zmul2, D, dims, ostr, optr, istr, iptr, val);
make_z3op_scalar(md_zmul2, D, dims, ostr, optr, istr, iptr, val);
}


Expand All @@ -709,7 +657,7 @@ void md_zsmul(int D, const long dims[D], complex float* optr, const complex floa
*/
void md_smul2(int D, const long dims[D], const long ostr[D], float* optr, const long istr[D], const float* iptr, float var)
{
make_3op_scalar_direct(md_mul2, D, dims, ostr, optr, istr, iptr, var);
make_3op_scalar(md_mul2, D, dims, ostr, optr, istr, iptr, var);
}


Expand Down Expand Up @@ -1719,7 +1667,7 @@ void md_zaxpy2(int D, const long dims[D], const long ostr[D], complex float* opt
return;
}

make_z3op_scalar_direct(md_zfmac2, D, dims, ostr, optr, istr, iptr, val);
make_z3op_scalar(md_zfmac2, D, dims, ostr, optr, istr, iptr, val);
}


Expand Down Expand Up @@ -1910,10 +1858,11 @@ void md_zadd(int D, const long dims[D], complex float* optr, const complex float
*/
void md_zsadd2(int D, const long dims[D], const long ostr[D], complex float* optr, const long istr[D], const complex float* iptr, complex float val)
{
#if 0
make_z3op_scalar(md_zadd2, D, dims, ostr, optr, istr, iptr, val);
#else
// FIXME: we should rather optimize md_zadd2 for this case
if (is_vptr(optr) || is_vptr(iptr)) {

make_z3op_scalar(md_zadd2, D, dims, ostr, optr, istr, iptr, val);
return;
}

NESTED(void, nary_zsadd, (struct nary_opt_data_s* data, void* ptr[]))
{
Expand All @@ -1922,7 +1871,6 @@ void md_zsadd2(int D, const long dims[D], const long ostr[D], complex float* opt

optimized_twoop_oi(D, dims, ostr, optr, istr, iptr,
(size_t[2]){ CFL_SIZE, CFL_SIZE }, nary_zsadd);
#endif
}


Expand Down Expand Up @@ -2000,10 +1948,11 @@ void md_add(int D, const long dims[D], float* optr, const float* iptr1, const fl
*/
void md_sadd2(int D, const long dims[D], const long ostr[D], float* optr, const long istr[D], const float* iptr, float val)
{
#if 0
make_3op_scalar(md_add2, D, dims, ostr, optr, istr, iptr, val);
#else
// FIXME: we should rather optimize md_zadd2 for this case
if (is_vptr(optr) || is_vptr(iptr)) {

make_3op_scalar(md_add2, D, dims, ostr, optr, istr, iptr, val);
return;
}

NESTED(void, nary_sadd, (struct nary_opt_data_s* data, void* ptr[]))
{
Expand All @@ -2012,7 +1961,6 @@ void md_sadd2(int D, const long dims[D], const long ostr[D], float* optr, const

optimized_twoop_oi(D, dims, ostr, optr, istr, iptr,
(size_t[2]){ FL_SIZE, FL_SIZE }, nary_sadd);
#endif
}


Expand Down Expand Up @@ -2100,7 +2048,7 @@ void md_zreal2(int D, const long dim[D], const long ostr[D], complex float* optr
void md_zreal(int D, const long dims[D], complex float* optr, const complex float* iptr)
{
#ifdef USE_CUDA
if (cuda_ondevice(iptr)) {
if (cuda_ondevice(iptr) && !is_vptr(iptr)) {

assert(cuda_ondevice(optr));

Expand Down

0 comments on commit 44513e7

Please sign in to comment.