ahmed/OpenCL-CTS
1
0
Fork 0
mirror of https://github.com/KhronosGroup/OpenCL-CTS.git synced 2026-03-19 06:09:01 +00:00
Code Issues 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 16:25:19 +00:00
committed by Kévin Petit
parent de6c3db41b
commit 95b040bec2
87 changed files with 1358 additions and 589 deletions
Download Patch File Download Diff File Expand all files Collapse all files

25
test_conformance/math_brute_force/binary.c
View File

@@ -233,6 +233,7 @@ 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
@@ -268,6 +269,16 @@ int TestFunc_Float_Float_Float_common(const Func *f, MTdata d, int isNextafter)
test_info.scale = (cl_uint) sizeof(cl_float) * 2 * gWimpyReductionFactor;
}
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 = gIsEmbedded ? f->float_embedded_ulps : f->float_ulps;
test_info.ftz = f->ftz || gForceFTZ || 0 == (CL_FP_DENORM & gFloatCapabilities);
@@ -341,7 +352,7 @@ int TestFunc_Float_Float_Float_common(const Func *f, MTdata d, int isNextafter)
// Run the kernels
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++ )
@@ -991,6 +1002,16 @@ int TestFunc_Double_Double_Double_common(const Func *f, MTdata d, int isNextafte
test_info.scale = (cl_uint) sizeof(cl_double) * 2 * gWimpyReductionFactor;
}
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;
@@ -1063,7 +1084,7 @@ int TestFunc_Double_Double_Double_common(const Func *f, MTdata d, int isNextafte
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++ )

97
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++ )

25
test_conformance/math_brute_force/binary_i.c
View File

@@ -230,6 +230,7 @@ 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
@@ -262,6 +263,16 @@ int TestFunc_Float_Float_Int(const Func *f, MTdata d)
test_info.scale = (cl_uint) sizeof(cl_float) * 2 * gWimpyReductionFactor;
}
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 = gIsEmbedded ? f->float_embedded_ulps : f->float_ulps;
test_info.ftz = f->ftz || gForceFTZ || 0 == (CL_FP_DENORM & gFloatCapabilities);
@@ -330,7 +341,7 @@ int TestFunc_Float_Float_Int(const Func *f, MTdata d)
}
// Run the kernels
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
@@ -758,6 +769,16 @@ int TestFunc_Double_Double_Int(const Func *f, MTdata d)
test_info.scale = (cl_uint) sizeof(cl_double) * 2 * gWimpyReductionFactor;
}
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;
@@ -831,7 +852,7 @@ int TestFunc_Double_Double_Int(const Func *f, MTdata d)
// Run the kernels
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

25
test_conformance/math_brute_force/macro_binary.c
View File

@@ -222,6 +222,7 @@ 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
int ftz; // non-zero if running in flush to zero mode
@@ -249,6 +250,16 @@ int TestMacro_Int_Float_Float(const Func *f, MTdata d)
test_info.scale = (cl_uint) sizeof(cl_float) * 2 * gWimpyReductionFactor;
}
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.ftz = f->ftz || gForceFTZ || 0 == (CL_FP_DENORM & gFloatCapabilities);
@@ -319,7 +330,7 @@ int TestMacro_Int_Float_Float(const Func *f, MTdata d)
// Run the kernels
if( !gSkipCorrectnessTesting )
{
error = ThreadPool_Do( TestFloat, (cl_uint) ((1ULL<<32) / test_info.step), &test_info );
error = ThreadPool_Do( TestFloat, test_info.jobCount, &test_info );
if( error )
goto exit;
@@ -749,6 +760,16 @@ int TestMacro_Int_Double_Double(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.ftz = f->ftz || gForceFTZ;
@@ -820,7 +841,7 @@ int TestMacro_Int_Double_Double(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 );
if( error )
goto exit;

25
test_conformance/math_brute_force/macro_unary.c
View File

@@ -193,6 +193,7 @@ 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
int ftz; // non-zero if running in flush to zero mode
@@ -220,6 +221,16 @@ int TestMacro_Int_Float(const Func *f, MTdata d)
test_info.scale = (cl_uint) sizeof(cl_float) * 2 * gWimpyReductionFactor;
}
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.ftz = f->ftz || gForceFTZ || 0 == (CL_FP_DENORM & gFloatCapabilities);
// cl_kernels aren't thread safe, so we make one for each vector size for every thread
@@ -279,7 +290,7 @@ int TestMacro_Int_Float(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 );
if( error )
goto exit;
@@ -602,6 +613,16 @@ int TestMacro_Int_Double(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.ftz = f->ftz || gForceFTZ;
@@ -664,7 +685,7 @@ int TestMacro_Int_Double(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 );
if( error )
goto exit;

2
test_conformance/math_brute_force/main.c
View File

@@ -699,6 +699,8 @@ static void PrintArch( void )
vlog( "\tARCH:\tx86_64\n" );
#elif defined( __arm__ )
vlog( "\tARCH:\tarm\n" );
#elif defined( __aarch64__ )
vlog( "\tARCH:\taarch64\n" );
#else
vlog( "\tARCH:\tunknown\n" );
#endif

6
test_conformance/math_brute_force/reference_math.c
View File

@@ -1790,7 +1790,7 @@ static const double //two52 = 4.50359962737049600000e+15, /* 0x43300000, 0x00000
// *signgamp = 1;
ix = hx&0x7fffffff;
if(ix>=0x7ff00000) return x*x;
if((ix|lx)==0) return (double)(one)/(double)(zero);
if((ix|lx)==0) return INFINITY;
if(ix<0x3b900000) { /* |x|<2**-70, return -log(|x|) */
if(hx<0) {
// *signgamp = -1;
@@ -1799,9 +1799,9 @@ static const double //two52 = 4.50359962737049600000e+15, /* 0x43300000, 0x00000
}
if(hx<0) {
if(ix>=0x43300000) /* |x|>=2**52, must be -integer */
return (double)(one)/(double)(zero);
return INFINITY;
t = reference_sinpi(x);
if(t==zero) (double)(one)/(double)(zero); /* -integer */
if(t==zero) return INFINITY; /* -integer */
nadj = reference_log(pi/reference_fabs(t*x));
// if(t<zero) *signgamp = -1;
x = -x;

0
test_conformance/math_brute_force/run_math_brute_force_in_parallel.py Normal file → Executable file
View File

26
test_conformance/math_brute_force/unary.c
View File

@@ -200,6 +200,7 @@ 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
@@ -234,6 +235,16 @@ int TestFunc_Float_Float(const Func *f, MTdata d)
test_info.scale = (cl_uint) sizeof(cl_float) * 2 * gWimpyReductionFactor;
}
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 = gIsEmbedded ? f->float_embedded_ulps : f->float_ulps;
test_info.ftz = f->ftz || gForceFTZ || 0 == (CL_FP_DENORM & gFloatCapabilities);
@@ -309,7 +320,7 @@ int TestFunc_Float_Float(const Func *f, MTdata d)
if( !gSkipCorrectnessTesting || skipTestingRelaxed)
{
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++ )
@@ -997,7 +1008,16 @@ int TestFunc_Double_Double(const Func *f, MTdata d)
test_info.subBufferSize = gWimpyBufferSize / (sizeof( cl_double) * RoundUpToNextPowerOfTwo(test_info.threadCount));
test_info.scale = (cl_uint) sizeof(cl_double) * 2 * gWimpyReductionFactor;
}
test_info.step = (cl_uint) test_info.subBufferSize * test_info.scale;
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;
@@ -1062,7 +1082,7 @@ int TestFunc_Double_Double(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++ )