diff options
author | Adrian Kummerlaender | 2019-11-10 21:14:07 +0100 |
---|---|---|
committer | Adrian Kummerlaender | 2019-11-10 21:18:57 +0100 |
commit | 4a2885ad3ae0396486d288df94339d0c45e6db8b (patch) | |
tree | 1a0b5aa000bbcde65fa020381a02b19bb452e284 /boltzgen/kernel/template/pattern | |
parent | d136bb30bc8a9393372ec905aea500a0b61000e3 (diff) | |
download | boltzgen-4a2885ad3ae0396486d288df94339d0c45e6db8b.tar boltzgen-4a2885ad3ae0396486d288df94339d0c45e6db8b.tar.gz boltzgen-4a2885ad3ae0396486d288df94339d0c45e6db8b.tar.bz2 boltzgen-4a2885ad3ae0396486d288df94339d0c45e6db8b.tar.lz boltzgen-4a2885ad3ae0396486d288df94339d0c45e6db8b.tar.xz boltzgen-4a2885ad3ae0396486d288df94339d0c45e6db8b.tar.zst boltzgen-4a2885ad3ae0396486d288df94339d0c45e6db8b.zip |
Implement basic CUDA target
Currently only for the SSS streaming pattern.
CudaCodePrinter in `utility/printer.py` is required to add a 'f' suffix
to all single precision floating point literals. If this is not done
(when targeting single precision) most calculations happen in double
precision which destroys performance. (In OpenCL this is not necessary
as we can simply set the `-cl-single-precision-constant` flag. Sadly
such a flag doesn't seem to exist for nvcc.)
Diffstat (limited to 'boltzgen/kernel/template/pattern')
-rw-r--r-- | boltzgen/kernel/template/pattern/SSS.cuda.mako | 81 |
1 files changed, 81 insertions, 0 deletions
diff --git a/boltzgen/kernel/template/pattern/SSS.cuda.mako b/boltzgen/kernel/template/pattern/SSS.cuda.mako new file mode 100644 index 0000000..8f571c7 --- /dev/null +++ b/boltzgen/kernel/template/pattern/SSS.cuda.mako @@ -0,0 +1,81 @@ +<%def name="operator(name, params = None)"> +<% +if layout.__class__.__name__ != 'SOA': + raise Exception('SSS pattern only works for the AOS memory layout') +%> +__global__ void ${name}( + ${float_type}** f +% if 'cell_list_dispatch' in extras: + , std::size_t* cells + , std::size_t cell_count +% else: + , std::size_t gid +% endif +% if params is not None: +% for param_type, param_name in params: + , ${param_type} ${param_name} +% endfor +% endif +) { +% if 'cell_list_dispatch' in extras: + const std::size_t index = blockIdx.x * blockDim.x + threadIdx.x; + if (!(index < cell_count)) { + return; + } + const std::size_t gid = cells[index]; +% endif + +% for i, c_i in enumerate(descriptor.c): + ${float_type}* preshifted_f_${i} = f[${i}] + ${layout.cell_preshift('gid')}; +% endfor + +% for i, c_i in enumerate(descriptor.c): + const ${float_type} f_curr_${i} = *preshifted_f_${i}; +% endfor + + ${caller.body()} + +% for i, c_i in enumerate(descriptor.c): + *preshifted_f_${i} = f_next_${descriptor.c.index(-c_i)}; +% endfor +} +</%def> + +<%def name="functor(name, params = None)"> +<% +if layout.__class__.__name__ != 'SOA': + raise Exception('SSS pattern only works for the AOS memory layout') +%> +__global__ void ${name}( + ${float_type}** f +% if 'cell_list_dispatch' in extras: + , std::size_t* cells + , std::size_t cell_count +% else: + , std::size_t gid +% endif +% if params is not None: +% for param_type, param_name in params: + , ${param_type} ${param_name} +% endfor +% endif +) { +% if 'cell_list_dispatch' in extras: + const std::size_t index = blockIdx.x * blockDim.x + threadIdx.x; + if (!(index < cell_count)) { + return; + } + const std::size_t gid = cells[index]; +% endif + +% for i, c_i in enumerate(descriptor.c): + const ${float_type}* preshifted_f_${i} = f[${i}] + ${layout.cell_preshift('gid')}; +% endfor + +% for i, c_i in enumerate(descriptor.c): + const ${float_type} f_curr_${i} = *preshifted_f_${descriptor.c.index(-c_i)}; +% endfor + + ${caller.body()} +} +</%def> |