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

initial RISC-V support (#2614)

Browse Source 
Unlike related PR #2344 that simply warns about unsupported FTZ, this PR
attempts to correctly handle FTZ on RISC-V.
RISC-V 'f' extension does not support any way to enable/disable flushing
subnormals to zero, implementations are required to always support
subnormals. Therefore this PR re-uses FTZ handling code from PPC, where
flushing also has to be explicitly performed.
This commit is contained in:
Michal Babej
2026-03-17 18:25:59 +02:00
committed by GitHub
parent 6506421614
commit 4e3f16b2b9
7 changed files with 32 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
test_conformance/contractions/contractions.cpp
View File

@@ -191,7 +191,7 @@ double sse_mul_sd(double x, double y)
}
#endif
#ifdef __PPC__
#if defined(__PPC__) || defined(__riscv)
float ppc_mul(float a, float b)
{
float p;
@@ -630,9 +630,11 @@ test_status InitCL( cl_device_id device )
// turn that off
f3[i] = sse_mul(q, q2);
f4[i] = sse_mul(-q, q2);
#elif defined(__PPC__)
// None of the current generation PPC processors support HW
// FTZ, emulate it in sw.
#elif (defined(__PPC__) || defined(__riscv))
// RISC-V CPUs with default 'f' fp32 extension do not support
// enabling/disabling FTZ mode, subnormals are always handled
// without FTZ. None of the current generation PPC processors
// support HW FTZ, emulate it in sw.
f3[i] = ppc_mul(q, q2);
f4[i] = ppc_mul(-q, q2);
#else
@@ -721,9 +723,10 @@ test_status InitCL( cl_device_id device )
skipTest[j][i] = (bufSkip[i] ||
(gSkipNanInf && (FE_OVERFLOW == (FE_OVERFLOW & fetestexcept(FE_OVERFLOW)))));
#if defined(__PPC__)
// Since the current Power processors don't emulate flush to zero in HW,
// it must be emulated in SW instead.
#if defined(__PPC__) || defined(__riscv)
// Since the current Power processors don't emulate flush to
// zero in HW, it must be emulated in SW instead. (same for
// RISC-V CPUs with 'f' extension)
if (gForceFTZ)
{
if ((fabsf(correct[j][i]) < FLT_MIN) && (correct[j][i] != 0.0f))
@@ -760,7 +763,6 @@ test_status InitCL( cl_device_id device )
}
}
double *f = (double*) buf1;
double *f2 = (double*) buf2;
double *f3 = (double*) buf3_double;