ICODE-ADC/lammps
4
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Explicitly check for subgroup support instead of CL version

Browse Source
This commit is contained in:
Richard Berger
2021-04-22 12:54:13 -04:00
parent 0f1f49afa7
commit 0632922a9b
6 changed files with 6 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
lib/gpu/lal_base_atomic.cpp
View File

@ -335,7 +335,7 @@ void BaseAtomicT::compile_kernels(UCL_Device &dev, const void *pair_str,
_compiled=true; _compiled=true;
#if defined(USE_OPENCL) && (defined(CL_VERSION_2_1) || defined(CL_VERSION_3_0)) #if defined(USE_OPENCL) && (defined(CL_VERSION_2_1) || defined(CL_VERSION_3_0))
if (dev.cl_device_version() >= 210) { if (dev.has_subgroup_support()) {
size_t mx_subgroup_sz = k_pair_fast.max_subgroup_size(_block_size); size_t mx_subgroup_sz = k_pair_fast.max_subgroup_size(_block_size);
#if defined(LAL_OCL_EV_JIT) #if defined(LAL_OCL_EV_JIT)
mx_subgroup_sz = std::min(mx_subgroup_sz, k_pair_noev.max_subgroup_size(_block_size)); mx_subgroup_sz = std::min(mx_subgroup_sz, k_pair_noev.max_subgroup_size(_block_size));

2
lib/gpu/lal_base_charge.cpp
View File

@ -348,7 +348,7 @@ void BaseChargeT::compile_kernels(UCL_Device &dev, const void *pair_str,
_compiled=true; _compiled=true;
#if defined(USE_OPENCL) && (defined(CL_VERSION_2_1) || defined(CL_VERSION_3_0)) #if defined(USE_OPENCL) && (defined(CL_VERSION_2_1) || defined(CL_VERSION_3_0))
if (dev.cl_device_version() >= 210) { if (dev.has_subgroup_support()) {
size_t mx_subgroup_sz = k_pair_fast.max_subgroup_size(_block_size); size_t mx_subgroup_sz = k_pair_fast.max_subgroup_size(_block_size);
#if defined(LAL_OCL_EV_JIT) #if defined(LAL_OCL_EV_JIT)
mx_subgroup_sz = std::min(mx_subgroup_sz, k_pair_noev.max_subgroup_size(_block_size)); mx_subgroup_sz = std::min(mx_subgroup_sz, k_pair_noev.max_subgroup_size(_block_size));

2
lib/gpu/lal_base_dipole.cpp
View File

@ -356,7 +356,7 @@ void BaseDipoleT::compile_kernels(UCL_Device &dev, const void *pair_str,
_compiled=true; _compiled=true;
#if defined(USE_OPENCL) && (defined(CL_VERSION_2_1) || defined(CL_VERSION_3_0)) #if defined(USE_OPENCL) && (defined(CL_VERSION_2_1) || defined(CL_VERSION_3_0))
if (dev.cl_device_version() >= 210) { if (dev.has_subgroup_support()) {
size_t mx_subgroup_sz = k_pair_fast.max_subgroup_size(_block_size); size_t mx_subgroup_sz = k_pair_fast.max_subgroup_size(_block_size);
#if defined(LAL_OCL_EV_JIT) #if defined(LAL_OCL_EV_JIT)
mx_subgroup_sz = std::min(mx_subgroup_sz, k_pair_noev.max_subgroup_size(_block_size)); mx_subgroup_sz = std::min(mx_subgroup_sz, k_pair_noev.max_subgroup_size(_block_size));

2
lib/gpu/lal_base_dpd.cpp
View File

@ -356,7 +356,7 @@ void BaseDPDT::compile_kernels(UCL_Device &dev, const void *pair_str,
_compiled=true; _compiled=true;
#if defined(USE_OPENCL) && (defined(CL_VERSION_2_1) || defined(CL_VERSION_3_0)) #if defined(USE_OPENCL) && (defined(CL_VERSION_2_1) || defined(CL_VERSION_3_0))
if (dev.cl_device_version() >= 210) { if (dev.has_subgroup_support()) {
size_t mx_subgroup_sz = k_pair_fast.max_subgroup_size(_block_size); size_t mx_subgroup_sz = k_pair_fast.max_subgroup_size(_block_size);
#if defined(LAL_OCL_EV_JIT) #if defined(LAL_OCL_EV_JIT)
mx_subgroup_sz = std::min(mx_subgroup_sz, k_pair_noev.max_subgroup_size(_block_size)); mx_subgroup_sz = std::min(mx_subgroup_sz, k_pair_noev.max_subgroup_size(_block_size));

2
lib/gpu/lal_base_ellipsoid.cpp
View File

@ -554,7 +554,7 @@ void BaseEllipsoidT::compile_kernels(UCL_Device &dev,
_compiled=true; _compiled=true;
#if defined(USE_OPENCL) && (defined(CL_VERSION_2_1) || defined(CL_VERSION_3_0)) #if defined(USE_OPENCL) && (defined(CL_VERSION_2_1) || defined(CL_VERSION_3_0))
if (dev.cl_device_version() >= 210) { if (dev.has_subgroup_support()) {
size_t mx_subgroup_sz = k_lj_fast.max_subgroup_size(_block_size); size_t mx_subgroup_sz = k_lj_fast.max_subgroup_size(_block_size);
mx_subgroup_sz = std::min(mx_subgroup_sz, k_ellipsoid.max_subgroup_size(_block_size)); mx_subgroup_sz = std::min(mx_subgroup_sz, k_ellipsoid.max_subgroup_size(_block_size));
mx_subgroup_sz = std::min(mx_subgroup_sz, k_sphere_ellipsoid.max_subgroup_size(_block_size)); mx_subgroup_sz = std::min(mx_subgroup_sz, k_sphere_ellipsoid.max_subgroup_size(_block_size));

2
lib/gpu/lal_base_three.cpp
View File

@ -461,7 +461,7 @@ void BaseThreeT::compile_kernels(UCL_Device &dev, const void *pair_str,
_compiled=true; _compiled=true;
#if defined(USE_OPENCL) && (defined(CL_VERSION_2_1) || defined(CL_VERSION_3_0)) #if defined(USE_OPENCL) && (defined(CL_VERSION_2_1) || defined(CL_VERSION_3_0))
if (dev.cl_device_version() >= 210) { if (dev.has_subgroup_support()) {
size_t mx_subgroup_sz = k_pair.max_subgroup_size(_block_size); size_t mx_subgroup_sz = k_pair.max_subgroup_size(_block_size);
mx_subgroup_sz = std::min(mx_subgroup_sz, k_three_center.max_subgroup_size(_block_size)); mx_subgroup_sz = std::min(mx_subgroup_sz, k_three_center.max_subgroup_size(_block_size));
mx_subgroup_sz = std::min(mx_subgroup_sz, k_three_end.max_subgroup_size(_block_size)); mx_subgroup_sz = std::min(mx_subgroup_sz, k_three_end.max_subgroup_size(_block_size));