ahmed/OpenCL-CTS
1
0
Fork 0
mirror of https://github.com/KhronosGroup/OpenCL-CTS.git synced 2026-03-21 06:49:02 +00:00
Code Packages Projects Releases Wiki Activity

Synchronise with Khronos-private Gitlab branch

Browse Source 
The maintenance of the conformance tests is moving to Github.

This commit contains all the changes that have been done in
Gitlab since the first public release of the conformance tests.

Signed-off-by: Kevin Petit <kevin.petit@arm.com>
This commit is contained in:
Kevin Petit
2019-02-20 16:10:04 +00:00
committed by Kévin Petit
parent b1603eb6ba
commit 53db6e7f9f
115 changed files with 2632 additions and 1304 deletions
Download Patch File Download Diff File Expand all files Collapse all files

115
test_conformance/math_brute_force/binaryOperator.c
View File

@@ -207,6 +207,7 @@ typedef struct TestInfo
cl_kernel *k[VECTOR_SIZE_COUNT ]; // arrays of thread-specific kernels for each worker thread: k[vector_size][thread_id]
ThreadInfo *tinfo; // An array of thread specific information for each worker thread
cl_uint threadCount; // Number of worker threads
cl_uint jobCount; // Number of jobs
cl_uint step; // step between each chunk and the next.
cl_uint scale; // stride between individual test values
float ulps; // max_allowed ulps
@@ -260,6 +261,16 @@ int TestFunc_Float_Float_Float_Operator(const Func *f, MTdata d)
}
test_info.step = test_info.subBufferSize * test_info.scale;
if (test_info.step / test_info.subBufferSize != test_info.scale)
{
//there was overflow
test_info.jobCount = 1;
}
else
{
test_info.jobCount = (cl_uint)((1ULL << 32) / test_info.step);
}
test_info.f = f;
test_info.ulps = gIsEmbedded ? f->float_embedded_ulps : f->float_ulps;
test_info.ftz = f->ftz || gForceFTZ || 0 == (CL_FP_DENORM & gFloatCapabilities);
@@ -329,7 +340,7 @@ int TestFunc_Float_Float_Float_Operator(const Func *f, MTdata d)
if( !gSkipCorrectnessTesting )
{
error = ThreadPool_Do( TestFloat, (cl_uint) ((1ULL<<32) / test_info.step), &test_info );
error = ThreadPool_Do( TestFloat, test_info.jobCount, &test_info );
// Accumulate the arithmetic errors
for( i = 0; i < test_info.threadCount; i++ )
@@ -501,63 +512,51 @@ static cl_int TestFloat( cl_uint job_id, cl_uint thread_id, void *data )
int totalSpecialValueCount = specialValuesFloatCount * specialValuesFloatCount;
int indx = (totalSpecialValueCount - 1) / buffer_elements;
if( job_id <= (cl_uint)indx )
{ // test edge cases
float *fp = (float *)p;
float *fp2 = (float *)p2;
if( job_id <= (cl_uint)indx ) {
// Insert special values
uint32_t x, y;
x = (job_id * buffer_elements) % specialValuesFloatCount;
y = (job_id * buffer_elements) / specialValuesFloatCount;
x = (job_id * buffer_elements) % specialValuesFloatCount;
y = (job_id * buffer_elements) / specialValuesFloatCount;
for( ; j < buffer_elements; j++ )
{
fp[j] = specialValuesFloat[x];
fp2[j] = specialValuesFloat[y];
if( ++x >= specialValuesFloatCount )
{
for( ; j < buffer_elements; j++ ) {
p[j] = ((cl_uint *)specialValuesFloat)[x];
p2[j] = ((cl_uint *)specialValuesFloat)[y];
++x;
if (x >= specialValuesFloatCount) {
x = 0;
y++;
if( y >= specialValuesFloatCount )
if (y >= specialValuesFloatCount)
break;
}
if(gTestFastRelaxed && strcmp(name,"divide") == 0 )
{
float fpj = *(float*)&fp[j];
float fpj2 = *(float*)&fp2[j];
if(fabs(fpj) > 0x5E800000 ) //[2^-62,2^62]
{
fp[j] = NAN;
}
if( fabs(fpj2) > 0x5E800000 ) //[2^-62,2^62]
{
fp2[j] = NAN;
}
if (gTestFastRelaxed && strcmp(name,"divide") == 0) {
cl_uint pj = p[j] & 0x7fffffff;
cl_uint p2j = p2[j] & 0x7fffffff;
// Replace values outside [2^-62, 2^62] with QNaN
if (pj < 0x20800000 || pj > 0x5e800000)
p[j] = 0x7fc00000;
if (p2j < 0x20800000 || p2j > 0x5e800000)
p2[j] = 0x7fc00000;
}
}
}
}
//Init any remaining values.
// Init any remaining values.
for( ; j < buffer_elements; j++ )
{
p[j] = genrand_int32(d);
p2[j] = genrand_int32(d);
if(gTestFastRelaxed)
{
if( strcmp(name,"divide")==0){
float pj = *(float*)&p[j];
float pj2 = *(float*)&p2[j];
if(fabs(pj) > 0x5E800000 ) //[2^-62,2^62]
{
p[j] = NAN;
}
if( fabs(pj2) > 0x5E800000 ) //[2^-62,2^62]
{
p2[j] = NAN;
}
}
}
if (gTestFastRelaxed && strcmp(name,"divide") == 0) {
cl_uint pj = p[j] & 0x7fffffff;
cl_uint p2j = p2[j] & 0x7fffffff;
// Replace values outside [2^-62, 2^62] with QNaN
if (pj < 0x20800000 || pj > 0x5e800000)
p[j] = 0x7fc00000;
if (p2j < 0x20800000 || p2j > 0x5e800000)
p2[j] = 0x7fc00000;
}
}
if( (error = clEnqueueWriteBuffer( tinfo->tQueue, tinfo->inBuf, CL_FALSE, 0, buffer_size, p, 0, NULL, NULL) ))
@@ -950,6 +949,16 @@ int TestFunc_Double_Double_Double_Operator(const Func *f, MTdata d)
}
test_info.step = (cl_uint) test_info.subBufferSize * test_info.scale;
if (test_info.step / test_info.subBufferSize != test_info.scale)
{
//there was overflow
test_info.jobCount = 1;
}
else
{
test_info.jobCount = (cl_uint)((1ULL << 32) / test_info.step);
}
test_info.f = f;
test_info.ulps = f->double_ulps;
test_info.ftz = f->ftz || gForceFTZ;
@@ -1020,7 +1029,7 @@ int TestFunc_Double_Double_Double_Operator(const Func *f, MTdata d)
if( !gSkipCorrectnessTesting )
{
error = ThreadPool_Do( TestDouble, (cl_uint) ((1ULL<<32) / test_info.step), &test_info );
error = ThreadPool_Do( TestDouble, test_info.jobCount, &test_info );
// Accumulate the arithmetic errors
for( i = 0; i < test_info.threadCount; i++ )
@@ -1315,7 +1324,7 @@ static cl_int TestDouble( cl_uint job_id, cl_uint thread_id, void *data )
{
cl_double test = ((cl_double*) q)[j];
long double correct = func.f_ff( s[j], s2[j] );
float err = Ulp_Error_Double( test, correct );
float err = Bruteforce_Ulp_Error_Double( test, correct );
int fail = ! (fabsf(err) <= ulps);
if( fail && ftz )
@@ -1334,8 +1343,8 @@ static cl_int TestDouble( cl_uint job_id, cl_uint thread_id, void *data )
{
long double correct2 = func.f_ff( 0.0, s2[j] );
long double correct3 = func.f_ff( -0.0, s2[j] );
float err2 = Ulp_Error_Double( test, correct2 );
float err3 = Ulp_Error_Double( test, correct3 );
float err2 = Bruteforce_Ulp_Error_Double( test, correct2 );
float err3 = Bruteforce_Ulp_Error_Double( test, correct3 );
fail = fail && ((!(fabsf(err2) <= ulps)) && (!(fabsf(err3) <= ulps)));
if( fabsf( err2 ) < fabsf(err ) )
err = err2;
@@ -1357,10 +1366,10 @@ static cl_int TestDouble( cl_uint job_id, cl_uint thread_id, void *data )
correct3 = func.f_ff( -0.0, 0.0 );
long double correct4 = func.f_ff( 0.0, -0.0 );
long double correct5 = func.f_ff( -0.0, -0.0 );
err2 = Ulp_Error_Double( test, correct2 );
err3 = Ulp_Error_Double( test, correct3 );
float err4 = Ulp_Error_Double( test, correct4 );
float err5 = Ulp_Error_Double( test, correct5 );
err2 = Bruteforce_Ulp_Error_Double( test, correct2 );
err3 = Bruteforce_Ulp_Error_Double( test, correct3 );
float err4 = Bruteforce_Ulp_Error_Double( test, correct4 );
float err5 = Bruteforce_Ulp_Error_Double( test, correct5 );
fail = fail && ((!(fabsf(err2) <= ulps)) && (!(fabsf(err3) <= ulps)) &&
(!(fabsf(err4) <= ulps)) && (!(fabsf(err5) <= ulps)));
if( fabsf( err2 ) < fabsf(err ) )
@@ -1386,8 +1395,8 @@ static cl_int TestDouble( cl_uint job_id, cl_uint thread_id, void *data )
{
long double correct2 = func.f_ff( s[j], 0.0 );
long double correct3 = func.f_ff( s[j], -0.0 );
float err2 = Ulp_Error_Double( test, correct2 );
float err3 = Ulp_Error_Double( test, correct3 );
float err2 = Bruteforce_Ulp_Error_Double( test, correct2 );
float err3 = Bruteforce_Ulp_Error_Double( test, correct3 );
fail = fail && ((!(fabsf(err2) <= ulps)) && (!(fabsf(err3) <= ulps)));
if( fabsf( err2 ) < fabsf(err ) )
err = err2;