static void
calcNrThreads(
size_t threads[2],
const SubproblemDim *subdims,
const PGranularity *pgran,
const void *args,
const void *_extra)
{
DUMMY_ARG_USAGE(subdims);
const CLBLASKernExtra *extra = ( CLBLASKernExtra *)_extra;
CLBlasKargs *kargs = (CLBlasKargs *)args;
SolutionStep *step = container_of(kargs, args, SolutionStep);
TargetDevice *kDevice = &(step->device);
cl_int err;
unsigned int numComputeUnits = deviceComputeUnits( (kDevice->id), &err );
if(err != CL_SUCCESS) {
numComputeUnits = 1;
}
unsigned int vecLen = extra->vecLenA;
unsigned int blockSize = pgran->wgSize[0] * pgran->wgSize[1];
unsigned int wgToSpawn = ((kargs->N - 1)/ (blockSize*vecLen)) + 1;
wgToSpawn = min( wgToSpawn, (numComputeUnits * WORKGROUPS_PER_CU) );
threads[0] = wgToSpawn * blockSize;
threads[1] = 1;
}
// unrolling generator for the f4zero function
static int
f4zeroSingle(struct KgenContext *ctx, void *priv)
{
DUMMY_ARG_USAGE(priv);
return kgenAddStmt(ctx, "*data++ = 0;\n");
}
static bool
checkCalcDecompDedicated(
PGranularity *pgran,
SubproblemDim *subdims,
unsigned int subdimsNum,
DataType dtype,
int check)
{
bool ret = true;
DUMMY_ARG_USAGE(subdimsNum);
if (check == PGRAN_CHECK) {
unsigned int minSize, maxSize;
maxSize = (dtype == TYPE_COMPLEX_DOUBLE) ? 4 : 8;
minSize = (dtype == TYPE_COMPLEX_DOUBLE) ? 1 : 2;
ret = decompSanityCheck(subdims, minSize, maxSize, 24, dtype, true);
ret = ret && (subdims[0].bwidth == subdims[1].bwidth);
ret = ret && (pgran->wgSize[0] == 64);
}
else {
calcPgranDedicated(pgran, subdims, -1, 3);
}
return ret;
}
static int trmmGetDefaultDecomp( PGranularity *pgran,
SubproblemDim *subdims,
unsigned int subdimsNum,
void *pArgs)
{
DUMMY_ARG_USAGE(subdimsNum);
if ( NULL == pArgs ) {
return -EINVAL;
}
subdims[1].bwidth = 2;
subdims[1].x = subdims[1].itemX = 8;
subdims[1].y = subdims[1].itemY = 8;
subdims[0].bwidth = 2;
subdims[0].x = subdims[0].itemX = 32;
subdims[0].y = 128;
subdims[0].itemY = -1;
pgran->wgDim = 1;
pgran->wgSize[0] = 64;
pgran->wgSize[1] = 1;
return 0;
}
/** The purpose of this function is to add an work-group size indicator in
kernelKey, so that a different kernel is generated when work-group size is changed.
Reduction loop is unrolled in kprintf based on work-group size.
Member of SubproblemDim- bwidth, will be used to store work-group size of the current kernel
this will become a kernelKey, and kernel cache will be accordingly managed.
Note -- SubproblemDim is a member of kernelKey
**/
static void
fixupArgs(void *args, SubproblemDim *subdims, void *extra)
{
DUMMY_ARG_USAGE(extra);
CLBlasKargs *kargs = (CLBlasKargs*)args;
SolutionStep *step = container_of(kargs, args, SolutionStep);
subdims->bwidth = (step->pgran.wgSize[0]) * (step->pgran.wgSize[1]);
}
static int
copyMemPreUnroll(struct KgenContext *ctx, void *priv)
{
DUMMY_ARG_USAGE(priv);
kgenAddStmt(ctx, "src1 = src;\n");
return kgenAddStmt(ctx, "dst1 = dst;\n\n");
}
//
// FIXME: Report correct return value - Needs change in KPRINTF
//
static ssize_t
generator(
char *buf,
size_t buflen,
const struct SubproblemDim *subdims,
const struct PGranularity *pgran,
void *extra)
{
DUMMY_ARG_USAGE(subdims);
size_t BLOCKSIZE = pgran->wgSize[0];
char tempTemplate[32*1024];
if ( buf == NULL) // return buffer size
{
buflen = (32 * 1024 * sizeof(char));
return (ssize_t)buflen;
}
CLBLASKernExtra *extraFlags = ( CLBLASKernExtra *)extra;
#ifdef DEBUG_ASUM
printf("ASUM GENERATOR called....\n");
printf("dataType : %c\n", Prefix[extraFlags->dtype]);
#endif
unsigned int vecLenA = extraFlags->vecLenA;
#ifdef DEBUG_ASUM
printf("Vector length used : %d\n\n", vecLenA);
#endif
bool doVLOAD = false;
if( extraFlags->flags & KEXTRA_NO_COPY_VEC_A )
{
doVLOAD = true;
#ifdef DEBUG_ASUM
printf("DOing VLOAD as Aligned Data Pointer not Availabe\n");
#endif
}
else
{
#ifdef DEBUG_ASUM
printf("Using Aligned Data Pointer \n");
#endif
}
strcpy( tempTemplate, (char*)asum_kernel );
kprintf kobj( Prefix[extraFlags->dtype], vecLenA, doVLOAD, doVLOAD, BLOCKSIZE);
kobj.spit((char*)buf, tempTemplate);
return (32 * 1024 * sizeof(char));
}
static int
subgGetPerf( unsigned int kflags,
const void *args )
{
DUMMY_ARG_USAGE(args);
if( !isMatrixAccessColMaj( CLBLAS_TRMM, kflags, MATRIX_A ) &&
!isMatrixAccessColMaj( CLBLAS_TRMM, kflags, MATRIX_B ) ){
return PPERF_GOOD;
}
return PPERF_NOT_SUPPORTED;
}
static int trmmSubgGetDefaultDecomp( PGranularity *pgran,
SubproblemDim *subdims,
unsigned int subdimsNum,
void *pArgs)
{
int itemsPerSubg = 4;
int subgA = 8;
int subgB = 2;
int bw1 = 8;
int x1 = 4;
int y1 = 4;
CLBlasKargs *kargs;
DUMMY_ARG_USAGE(subdimsNum);
if ( NULL == pArgs ) {
return -EINVAL;
}
kargs = (CLBlasKargs *)pArgs;
if( isComplexType(kargs->dtype) ){
bw1 /= 2;
}
if( isDoubleBasedType(kargs->dtype) ){
bw1 /= 2;
}
subdims[1].bwidth = bw1;
subdims[1].x = subdims[1].itemX = x1;
subdims[1].y = subdims[1].itemY = y1;
subdims[0].bwidth = bw1 * itemsPerSubg;
subdims[0].itemX = x1 * subgB;
subdims[0].x = x1*subgB;
subdims[0].itemY = y1*subgA;
subdims[0].y = y1*subgA;
pgran->wgDim = 1;
pgran->wgSize[0] = 64;
pgran->wgSize[1] = 1;
return 0;
}
static int
symvSubgGetDefaultDecomp(
PGranularity *pgran,
SubproblemDim *subdims,
unsigned int subdimsNum,
void * pArgs)
{
(void)subdimsNum;
DUMMY_ARG_USAGE(pArgs);
pgran->wgDim = 1;
pgran->wgSize[0] = 64;
pgran->wgSize[1] = 1;
subdims[1].bwidth = 4;
subdims[1].itemX = subdims[1].x = 1;
subdims[1].itemY = subdims[1].y = 4;
subdims[0].bwidth = 8 * subdims[1].bwidth;
subdims[0].itemX = subdims[0].x = 1;
subdims[0].itemY = subdims[0].y = 8 * subdims[1].y;
return 0;
}
void initialize_scalars(double alpha, double beta)
{
DUMMY_ARG_USAGE(beta);
buffer_.alpha_ = makeScalar<T>(alpha);
}
//-----------------------------------------------------------------------------
// TODO: reimplement via new validation API
static bool
subgCheckCalcDecomp( PGranularity *pgran,
SubproblemDim *subdims,
unsigned int subdimsNum,
DataType dtype,
int check )
{
unsigned int subgA = 0;
unsigned int subgB = 0;
unsigned int regUse = 0;
unsigned int itemsPerSubg = 0;
DUMMY_ARG_USAGE(subdimsNum);
if( 0 == subdims[0].x ||
0 == subdims[0].y ||
0 == subdims[0].bwidth ||
0 == subdims[1].x ||
0 == subdims[1].y ||
0 == subdims[1].bwidth ){
return false;
}
subgA = subdims[0].y/subdims[1].y;
subgB = subdims[0].x/subdims[1].x;
itemsPerSubg = subdims[0].bwidth/subdims[1].bwidth;
if( itemsPerSubg < 4 ){
return false;
}
if( subdims[1].y < 4 ||
subdims[1].x < 4 ||
subdims[1].bwidth < 4 ){
return false;
}
if( subdims[1].x != subdims[1].itemX ||
subdims[1].y != subdims[1].itemY ){
return false;
}
// the group block must consist of integer number of subgroup blocks
if( subdims[0].x % subdims[1].x ||
subdims[0].y % subdims[1].y ||
subdims[0].bwidth % subdims[1].bwidth ){
return false;
}
//check fitting of bw to common vector sizes
if( isComplexType(dtype) ){
if( 2*subdims[1].bwidth > 16 ){
return false;
}
}
// check dimensions
if( subdims[1].bwidth > 16 ||
subdims[1].x > 16 ||
subdims[1].y > 16 ){
return false;
}
// estimate register usage, drop
// inevitably slowed decompositions
regUse =
( subdims[1].bwidth * subdims[1].x +
subdims[1].bwidth * subdims[1].y +
subdims[1].x * subdims[1].y ) *
dtypeSize(dtype);
regUse /= 16; // 16 bytes per register
if( regUse >= 64 ){
return false;
}
// passed PGranularity should be checked
if( PGRAN_CHECK == check ){
if( pgran->wgDim != 1 ){
return false;
}
if( pgran->wgSize[0] != 64 ){
return false;
}
if( pgran->wgSize[0] != subgA*subgB*itemsPerSubg ){
return false;
}
}
// PGranularity should be calculated
else{
pgran->wgDim = 1;
pgran->wgSize[0] = subgA * subgB * itemsPerSubg;
}
return true;
}
