Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Corrected math_brute_force for cl_khr_fp16 tests sqrt and divide to use correctly rounded reference #2161

Closed
wants to merge 1 commit into from
Closed

Corrected math_brute_force for cl_khr_fp16 tests sqrt and divide to use correctly rounded reference #2161

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
74 changes: 50 additions & 24 deletions test_conformance/math_brute_force/binary_operator_half.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -102,6 +102,42 @@ const cl_half specialValuesHalf[] = {

constexpr size_t specialValuesHalfCount = ARRAY_SIZE(specialValuesHalf);

cl_float ComputeRefDouble(cl_float p0, cl_float p1, void *data)
{
TestInfo *job = (TestInfo *)data;
fptr func = job->f->func;
return func.f_ff(p0, p1);
}

cl_float ComputeRefSingle(cl_float p0, cl_float p1, void *data)
{
TestInfo *job = (TestInfo *)data;
fptr func = job->f->hfunc;
return func.f_ff_f(p0, p1);
}

struct CorrectRoundScopeGuard
{
CorrectRoundScopeGuard(int ftz): mFtz(ftz)
{
memset(&mOldMode, 0, sizeof(mOldMode));
if (ftz) ForceFTZ(&mOldMode);

// Set the rounding mode to match the device
if (gIsInRTZMode) mOldRoundMode = set_round(kRoundTowardZero, kfloat);
}

~CorrectRoundScopeGuard()
{
if (mFtz) RestoreFPState(&mOldMode);
if (gIsInRTZMode) (void)set_round(mOldRoundMode, kfloat);
}

int mFtz;
FPU_mode_type mOldMode;
RoundingMode mOldRoundMode = kRoundToNearestEven;
};

cl_int TestHalf(cl_uint job_id, cl_uint thread_id, void *data)
{
TestInfo *job = (TestInfo *)data;
Expand All @@ -110,19 +146,20 @@ cl_int TestHalf(cl_uint job_id, cl_uint thread_id, void *data)
cl_uint base = job_id * (cl_uint)job->step;
ThreadInfo *tinfo = &(job->tinfo[thread_id]);
float ulps = job->ulps;
fptr func = job->f->func;
int ftz = job->ftz;
MTdata d = tinfo->d;
cl_int error;

const char *name = job->f->name;
cl_half *r = 0;
std::vector<float> s(0), s2(0);
RoundingMode oldRoundMode;

cl_event e[VECTOR_SIZE_COUNT];
cl_half *out[VECTOR_SIZE_COUNT];

auto ref_fnptr = &ComputeRefDouble;
if (0 == strcmp(job->f->name, "divide")) ref_fnptr = &ComputeRefSingle;

if (gHostFill)
{
// start the map of the output arrays
Expand Down Expand Up @@ -270,13 +307,7 @@ cl_int TestHalf(cl_uint job_id, cl_uint thread_id, void *data)
}

// Calculate the correctly rounded reference result
FPU_mode_type oldMode;
memset(&oldMode, 0, sizeof(oldMode));
if (ftz) ForceFTZ(&oldMode);

// Set the rounding mode to match the device
oldRoundMode = kRoundToNearestEven;
if (gIsInRTZMode) oldRoundMode = set_round(kRoundTowardZero, kfloat);
CorrectRoundScopeGuard crsg(ftz);

// Calculate the correctly rounded reference result
r = (cl_half *)gOut_Ref + thread_id * buffer_elements;
Expand All @@ -287,11 +318,9 @@ cl_int TestHalf(cl_uint job_id, cl_uint thread_id, void *data)
{
s[j] = HTF(p[j]);
s2[j] = HTF(p2[j]);
r[j] = HFF(func.f_ff(s[j], s2[j]));
r[j] = HFF(ref_fnptr(s[j], s2[j], data));
}

if (ftz) RestoreFPState(&oldMode);

// Read the data back -- no need to wait for the first N-1 buffers but wait
// for the last buffer. This is an in order queue.
for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
Expand Down Expand Up @@ -320,7 +349,7 @@ cl_int TestHalf(cl_uint job_id, cl_uint thread_id, void *data)
if (r[j] != q[j])
{
float test = HTF(q[j]);
float correct = func.f_ff(s[j], s2[j]);
float correct = HTF(r[j]);

// Per section 10 paragraph 6, accept any result if an input or
// output is a infinity or NaN or overflow
Expand Down Expand Up @@ -353,8 +382,8 @@ cl_int TestHalf(cl_uint job_id, cl_uint thread_id, void *data)
double correct2, correct3;
float err2, err3;

correct2 = func.f_ff(0.0, s2[j]);
correct3 = func.f_ff(-0.0, s2[j]);
correct2 = ref_fnptr(0.0, s2[j], data);
correct3 = ref_fnptr(-0.0, s2[j], data);

// Per section 10 paragraph 6, accept any result if an
// input or output is a infinity or NaN or overflow
Expand Down Expand Up @@ -393,10 +422,10 @@ cl_int TestHalf(cl_uint job_id, cl_uint thread_id, void *data)
double correct4, correct5;
float err4, err5;

correct2 = func.f_ff(0.0, 0.0);
correct3 = func.f_ff(-0.0, 0.0);
correct4 = func.f_ff(0.0, -0.0);
correct5 = func.f_ff(-0.0, -0.0);
correct2 = ref_fnptr(0.0, 0.0, data);
correct3 = ref_fnptr(-0.0, 0.0, data);
correct4 = ref_fnptr(0.0, -0.0, data);
correct5 = ref_fnptr(-0.0, -0.0, data);

// Per section 10 paragraph 6, accept any result if
// an input or output is a infinity or NaN or
Expand Down Expand Up @@ -446,9 +475,8 @@ cl_int TestHalf(cl_uint job_id, cl_uint thread_id, void *data)
double correct2, correct3;
float err2, err3;

correct2 = func.f_ff(s[j], 0.0);
correct3 = func.f_ff(s[j], -0.0);

correct2 = ref_fnptr(s[j], 0.0, data);
correct3 = ref_fnptr(s[j], -0.0, data);

// Per section 10 paragraph 6, accept any result if an
// input or output is a infinity or NaN or overflow
Expand Down Expand Up @@ -499,8 +527,6 @@ cl_int TestHalf(cl_uint job_id, cl_uint thread_id, void *data)
}
}

if (gIsInRTZMode) (void)set_round(oldRoundMode, kfloat);

for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
{
if ((error = clEnqueueUnmapMemObject(tinfo->tQueue, tinfo->outBuf[j],
Expand Down
59 changes: 41 additions & 18 deletions test_conformance/math_brute_force/function_list.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,28 +32,35 @@
#define ENTRY(_name, _ulp, _embedded_ulp, _half_ulp, _rmode, _type) \
{ \
STRINGIFY(_name), STRINGIFY(_name), { NULL }, { NULL }, { NULL }, \
_ulp, _ulp, _half_ulp, _embedded_ulp, INFINITY, INFINITY, _rmode, \
RELAXED_OFF, _type \
{ NULL }, _ulp, _ulp, _half_ulp, _embedded_ulp, INFINITY, \
INFINITY, _rmode, RELAXED_OFF, _type \
}
#define ENTRY_EXT(_name, _ulp, _embedded_ulp, _half_ulp, _relaxed_ulp, _rmode, \
_type, _relaxed_embedded_ulp) \
{ \
STRINGIFY(_name), STRINGIFY(_name), { NULL }, { NULL }, { NULL }, \
_ulp, _ulp, _half_ulp, _embedded_ulp, _relaxed_ulp, \
{ NULL }, _ulp, _ulp, _half_ulp, _embedded_ulp, _relaxed_ulp, \
_relaxed_embedded_ulp, _rmode, RELAXED_ON, _type \
}
#define HALF_ENTRY(_name, _ulp, _embedded_ulp, _rmode, _type) \
{ \
"half_" STRINGIFY(_name), "half_" STRINGIFY(_name), { NULL }, \
{ NULL }, { NULL }, _ulp, _ulp, _ulp, _embedded_ulp, INFINITY, \
INFINITY, _rmode, RELAXED_OFF, _type \
{ NULL }, { NULL }, { NULL }, _ulp, _ulp, _ulp, _embedded_ulp, \
INFINITY, INFINITY, _rmode, RELAXED_OFF, _type \
}
#define OPERATOR_ENTRY(_name, _operator, _ulp, _embedded_ulp, _half_ulp, \
_rmode, _type) \
{ \
STRINGIFY(_name), _operator, { NULL }, { NULL }, { NULL }, _ulp, _ulp, \
_half_ulp, _embedded_ulp, INFINITY, INFINITY, _rmode, RELAXED_OFF, \
_type \
STRINGIFY(_name), _operator, { NULL }, { NULL }, { NULL }, { NULL }, \
_ulp, _ulp, _half_ulp, _embedded_ulp, INFINITY, INFINITY, _rmode, \
RELAXED_OFF, _type \
}
#define OPERATOR_ENTRY_EXT(_name, _operator, _ulp, _embedded_ulp, _half_ulp, \
_rmode, _type) \
{ \
STRINGIFY(_name), _operator, { NULL }, { NULL }, { NULL }, { NULL }, \
_ulp, _ulp, _half_ulp, _embedded_ulp, INFINITY, INFINITY, _rmode, \
RELAXED_OFF, _type \
}

#define unaryF NULL
Expand Down Expand Up @@ -81,37 +88,48 @@
#define reference_divide NULL
#define reference_dividel NULL
#define reference_relaxed_divide NULL
#define reference_sqrtf NULL
#define reference_dividef NULL

#else // FUNCTION_LIST_ULPS_ONLY

#define ENTRY(_name, _ulp, _embedded_ulp, _half_ulp, _rmode, _type) \
{ \
STRINGIFY(_name), STRINGIFY(_name), { (void*)reference_##_name }, \
{ (void*)reference_##_name##l }, { (void*)reference_##_name }, \
_ulp, _ulp, _half_ulp, _embedded_ulp, INFINITY, INFINITY, _rmode, \
RELAXED_OFF, _type \
{ (void*)reference_##_name }, _ulp, _ulp, _half_ulp, \
_embedded_ulp, INFINITY, INFINITY, _rmode, RELAXED_OFF, _type \
}
#define ENTRY_EXT(_name, _ulp, _embedded_ulp, _half_ulp, _relaxed_ulp, _rmode, \
_type, _relaxed_embedded_ulp) \
{ \
STRINGIFY(_name), STRINGIFY(_name), { (void*)reference_##_name }, \
{ (void*)reference_##_name##l }, \
{ (void*)reference_##relaxed_##_name }, _ulp, _ulp, _half_ulp, \
_embedded_ulp, _relaxed_ulp, _relaxed_embedded_ulp, _rmode, \
RELAXED_ON, _type \
{ (void*)reference_##relaxed_##_name }, { NULL }, _ulp, _ulp, \
_half_ulp, _embedded_ulp, _relaxed_ulp, _relaxed_embedded_ulp, \
_rmode, RELAXED_ON, _type \
}
#define HALF_ENTRY(_name, _ulp, _embedded_ulp, _rmode, _type) \
{ \
"half_" STRINGIFY(_name), "half_" STRINGIFY(_name), \
{ (void*)reference_##_name }, { NULL }, { NULL }, _ulp, _ulp, \
_ulp, _embedded_ulp, INFINITY, INFINITY, _rmode, RELAXED_OFF, \
_type \
{ (void*)reference_##_name }, { NULL }, { NULL }, { NULL }, _ulp, \
_ulp, _ulp, _embedded_ulp, INFINITY, INFINITY, _rmode, \
RELAXED_OFF, _type \
}
#define OPERATOR_ENTRY(_name, _operator, _ulp, _embedded_ulp, _half_ulp, \
_rmode, _type) \
{ \
STRINGIFY(_name), _operator, { (void*)reference_##_name }, \
{ (void*)reference_##_name##l }, { NULL }, _ulp, _ulp, _half_ulp, \
{ (void*)reference_##_name##l }, { NULL }, { NULL }, _ulp, _ulp, \
_half_ulp, _embedded_ulp, INFINITY, INFINITY, _rmode, RELAXED_OFF, \
_type \
}
#define OPERATOR_ENTRY_EXT(_name, _operator, _ulp, _embedded_ulp, _half_ulp, \
_rmode, _type) \
{ \
STRINGIFY(_name), _operator, { (void*)reference_##_name }, \
{ (void*)reference_##_name##l }, { NULL }, \
{ (void*)reference_##_name##f }, _ulp, _ulp, _half_ulp, \
_embedded_ulp, INFINITY, INFINITY, _rmode, RELAXED_OFF, _type \
}

Expand Down Expand Up @@ -266,6 +284,7 @@ const Func functionList[] = {
{ (void*)reference_copysignf },
{ (void*)reference_copysign },
{ (void*)reference_copysignf },
{ (void*)reference_copysignf },
0.0f,
0.0f,
0.0f,
Expand Down Expand Up @@ -367,6 +386,7 @@ const Func functionList[] = {
{ (void*)reference_sqrt },
{ (void*)reference_sqrtl },
{ NULL },
{ (void*)reference_sqrtf },
3.0f,
0.0f,
0.0f,
Expand All @@ -380,7 +400,8 @@ const Func functionList[] = {
"sqrt",
{ (void*)reference_sqrt },
{ nullptr },
{ NULL },
{ nullptr },
{ nullptr },
0.0f,
INFINITY,
INFINITY,
Expand Down Expand Up @@ -423,6 +444,7 @@ const Func functionList[] = {
{ (void*)reference_divide },
{ (void*)reference_dividel },
{ (void*)reference_relaxed_divide },
{ (void*)reference_dividef },
2.5f,
0.0f,
0.0f,
Expand All @@ -437,6 +459,7 @@ const Func functionList[] = {
{ (void*)reference_divide },
{ nullptr },
{ nullptr },
{ nullptr },
0.0f,
INFINITY,
INFINITY,
Expand Down
2 changes: 2 additions & 0 deletions test_conformance/math_brute_force/function_list.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,6 +33,7 @@
union fptr {
void *p;
double (*f_f)(double);
float (*f_f_f)(float);
double (*f_u)(cl_uint);
int (*i_f)(double);
int (*i_f_f)(float);
Expand Down Expand Up @@ -84,6 +85,7 @@ struct Func
fptr func;
dptr dfunc;
fptr rfunc;
fptr hfunc;
float float_ulps;
float double_ulps;
float half_ulps;
Expand Down
5 changes: 4 additions & 1 deletion test_conformance/math_brute_force/reference_math.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5771,4 +5771,7 @@ long double reference_erfcl(long double x) { return erfc(x); }
long double reference_erfl(long double x) { return erf(x); }

double reference_erfc(double x) { return erfc(x); }
double reference_erf(double x) { return erf(x); }
double reference_erf(double x) { return erf(x); }

float reference_sqrtf(float x) { return sqrtf(x); }
float reference_dividef(float x, float y) { return x / y; }
3 changes: 3 additions & 0 deletions test_conformance/math_brute_force/reference_math.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -240,4 +240,7 @@ long double reference_erfcl(long double x);
long double reference_erfl(long double x);
double reference_erfc(double x);
double reference_erf(double x);

float reference_sqrtf(float x);
float reference_dividef(float x, float y);
#endif
Loading
Loading