ahmed/OpenCL-CTS
1
0
Fork 0
mirror of https://github.com/KhronosGroup/OpenCL-CTS.git synced 2026-03-20 22:39:03 +00:00
Code Packages Projects Releases Wiki Activity

Added cl_khr_fp16 extension support for test_commonfns (#1695)

Browse Source 
* Added cl_khr_fp16 extension support for commonfns test (issue #142, commonfns)

* Added missing header due to presubmit check

* Corrected radians/degrees ulp calculations + cosmetic fixes

* Corrected presubmit code format

* Corrections related to code review

* Moved string format helper to test_common in separate header

* Added clang format for last commit

* Corrections related to code review

* Modified mix verification procedure for half type to only report max error

* Removed redundant condition for logging mix verification

* Corrected generator limits for half tests
This commit is contained in:
Marcin Hajder
2023-06-27 17:42:02 +02:00
committed by GitHub
parent 60f025a7da
commit 2495eca9fa
10 changed files with 527 additions and 259 deletions
Download Patch File Download Diff File Expand all files Collapse all files

120
test_conformance/commonfns/test_mix.cpp
View File

@@ -18,6 +18,8 @@
#include <sys/types.h>
#include <sys/stat.h>
#include "harness/stringHelpers.h"
#include "procs.h"
#include "test_base.h"
@@ -52,33 +54,42 @@ const char *mix_fn_code_pattern_v3_scalar =
" vstore3(mix(vload3(tid, x), vload3(tid, y), a[tid]), tid, dst);\n"
"}\n";
#define MAX_ERR 1e-3
namespace {
template <typename T>
int verify_mix(const T *const inptrX, const T *const inptrY,
const T *const inptrA, const T *const outptr, const int n,
const int veclen, const bool vecParam)
{
T r;
float delta = 0.0f;
double r, o;
float delta = 0.f, max_delta = 0.f;
int i;
if (vecParam)
{
for (i = 0; i < n * veclen; i++)
{
r = inptrX[i] + ((inptrY[i] - inptrX[i]) * inptrA[i]);
delta = fabs(double(r - outptr[i])) / r;
if (delta > MAX_ERR)
r = conv_to_dbl(inptrX[i])
+ ((conv_to_dbl(inptrY[i]) - conv_to_dbl(inptrX[i]))
* conv_to_dbl(inptrA[i]));
o = conv_to_dbl(outptr[i]);
delta = fabs(double(r - o)) / r;
if (!std::is_same<T, half>::value)
{
log_error(
"%d) verification error: mix(%a, %a, %a) = *%a vs. %a\n", i,
inptrX[i], inptrY[i], inptrA[i], r, outptr[i]);
return -1;
if (delta > MAX_ERR)
{
log_error("%d) verification error: mix(%a, %a, %a) = *%a "
"vs. %a\n",
i, inptrX[i], inptrY[i], inptrA[i], r, outptr[i]);
return -1;
}
}
else
{
max_delta = std::max(max_delta, delta);
}
}
}
@@ -90,25 +101,40 @@ int verify_mix(const T *const inptrX, const T *const inptrY,
int vi = i * veclen;
for (int j = 0; j < veclen; ++j, ++vi)
{
r = inptrX[vi] + ((inptrY[vi] - inptrX[vi]) * inptrA[i]);
delta = fabs(double(r - outptr[vi])) / r;
if (delta > MAX_ERR)
r = conv_to_dbl(inptrX[vi])
+ ((conv_to_dbl(inptrY[vi]) - conv_to_dbl(inptrX[vi]))
* conv_to_dbl(inptrA[i]));
delta = fabs(double(r - conv_to_dbl(outptr[vi]))) / r;
if (!std::is_same<T, half>::value)
{
log_error("{%d, element %d}) verification error: mix(%a, "
"%a, %a) = *%a vs. %a\n",
ii, j, inptrX[vi], inptrY[vi], inptrA[i], r,
outptr[vi]);
return -1;
if (delta > MAX_ERR)
{
log_error(
"{%d, element %d}) verification error: mix(%a, "
"%a, %a) = *%a vs. %a\n",
ii, j, inptrX[vi], inptrY[vi], inptrA[i], r,
outptr[vi]);
return -1;
}
}
else
{
max_delta = std::max(max_delta, delta);
}
}
}
}
// due to the fact that accuracy of mix for cl_khr_fp16 is implementation
// defined this test only reports maximum error without testing maximum
// error threshold
if (std::is_same<T, half>::value)
log_error("mix half verification result, max delta: %a\n", max_delta);
return 0;
}
} // namespace
template <typename T>
int test_mix_fn(cl_device_id device, cl_context context, cl_command_queue queue,
int n_elems, bool vecParam)
@@ -120,7 +146,7 @@ int test_mix_fn(cl_device_id device, cl_context context, cl_command_queue queue,
std::vector<clKernelWrapper> kernels;
int err, i;
MTdataHolder d = MTdataHolder(gRandomSeed);
MTdataHolder d(gRandomSeed);
assert(BaseFunctionTest::type2name.find(sizeof(T))
!= BaseFunctionTest::type2name.end());
@@ -142,19 +168,32 @@ int test_mix_fn(cl_device_id device, cl_context context, cl_command_queue queue,
test_error(err, "clCreateBuffer failed");
}
for (i = 0; i < num_elements; i++)
{
input_ptr[0][i] = (T)genrand_real1(d);
input_ptr[1][i] = (T)genrand_real1(d);
input_ptr[2][i] = (T)genrand_real1(d);
}
std::string pragma_str;
if (std::is_same<T, double>::value)
{
pragma_str = "#pragma OPENCL EXTENSION cl_khr_fp64 : enable\n";
}
if (std::is_same<T, half>::value)
{
pragma_str = "#pragma OPENCL EXTENSION cl_khr_fp16 : enable\n";
for (i = 0; i < num_elements; i++)
{
input_ptr[0][i] = conv_to_half((float)genrand_real1(d));
input_ptr[1][i] = conv_to_half((float)genrand_real1(d));
input_ptr[2][i] = conv_to_half((float)genrand_real1(d));
}
}
else
{
for (i = 0; i < num_elements; i++)
{
input_ptr[0][i] = (T)genrand_real1(d);
input_ptr[1][i] = (T)genrand_real1(d);
input_ptr[2][i] = (T)genrand_real1(d);
}
}
for (i = 0; i < 3; i++)
{
err = clEnqueueWriteBuffer(queue, streams[i], CL_TRUE, 0,
@@ -164,7 +203,6 @@ int test_mix_fn(cl_device_id device, cl_context context, cl_command_queue queue,
}
char vecSizeNames[][3] = { "", "2", "4", "8", "16", "3" };
for (i = 0; i < kTotalVecCount; i++)
{
std::string kernelSource;
@@ -174,15 +212,15 @@ int test_mix_fn(cl_device_id device, cl_context context, cl_command_queue queue,
{
std::string str = mix_fn_code_pattern_v3;
kernelSource =
string_format(str, pragma_str.c_str(), tname.c_str(),
tname.c_str(), tname.c_str(), tname.c_str());
str_sprintf(str, pragma_str.c_str(), tname.c_str(),
tname.c_str(), tname.c_str(), tname.c_str());
}
else
{
std::string str = mix_fn_code_pattern_v3_scalar;
kernelSource =
string_format(str, pragma_str.c_str(), tname.c_str(),
tname.c_str(), tname.c_str(), tname.c_str());
str_sprintf(str, pragma_str.c_str(), tname.c_str(),
tname.c_str(), tname.c_str(), tname.c_str());
}
}
else
@@ -190,10 +228,10 @@ int test_mix_fn(cl_device_id device, cl_context context, cl_command_queue queue,
// regular path
std::string str = mix_fn_code_pattern;
kernelSource =
string_format(str, pragma_str.c_str(), tname.c_str(),
vecSizeNames[i], tname.c_str(), vecSizeNames[i],
tname.c_str(), vecParam ? vecSizeNames[i] : "",
tname.c_str(), vecSizeNames[i]);
str_sprintf(str, pragma_str.c_str(), tname.c_str(),
vecSizeNames[i], tname.c_str(), vecSizeNames[i],
tname.c_str(), vecParam ? vecSizeNames[i] : "",
tname.c_str(), vecSizeNames[i]);
}
const char *programPtr = kernelSource.c_str();
err =
@@ -242,10 +280,14 @@ int test_mix_fn(cl_device_id device, cl_context context, cl_command_queue queue,
return err;
}
cl_int MixTest::Run()
{
cl_int error = CL_SUCCESS;
if (is_extension_available(device, "cl_khr_fp16"))
{
error = test_mix_fn<half>(device, context, queue, num_elems, vecParam);
test_error(error, "MixTest::Run<cl_half> failed");
}
error = test_mix_fn<float>(device, context, queue, num_elems, vecParam);
test_error(error, "MixTest::Run<float> failed");
@@ -260,7 +302,6 @@ cl_int MixTest::Run()
return error;
}
int test_mix(cl_device_id device, cl_context context, cl_command_queue queue,
int n_elems)
{
@@ -268,7 +309,6 @@ int test_mix(cl_device_id device, cl_context context, cl_command_queue queue,
true);
}
int test_mixf(cl_device_id device, cl_context context, cl_command_queue queue,
int n_elems)
{