def __init__(self, lst_svobjs, objIdOffset):
self.m_size = vki.SVUInt32(len(lst_svobjs))
self.m_buf = vki.SVObjBuffer(lst_svobjs)
self.m_id_offset = vki.SVUInt32(objIdOffset)
self.m_cptr = SVCombine_Create(
{
'size': self.m_size,
'data': self.m_buf,
'id_offset': self.m_id_offset
}, '''
uint get_size(in Comb_#hash# vec)
{{
return vec.size;
}}
{0} get_value(in Comb_#hash# vec, in uint id)
{{
return vec.data[id].v;
}}
'''.format(self.name_elem_type()))
kernel = vki.Computer(['dst', 'src', 'n'],
'''
shared {0} s_buf[{1}];
void main()
{{
uint tid = gl_LocalInvocationID.x;
uint i = gl_GlobalInvocationID.x;
if (i<n) s_buf[tid] = get_value(src, i);
barrier();
for (uint s = {1}/2; s>0; s>>=1)
{{
if (tid < s && i+s<n)
s_buf[tid] += s_buf[tid + s];
barrier();
}}
if (tid==0) set_value(dst, gl_WorkGroupID.x, s_buf[tid]);
}}
'''.format('int',str(BLOCK_SIZE)))
dst = darr
while dst.size()>1:
src = dst
n = src.size()
blocks = int((n + BLOCK_SIZE - 1) / BLOCK_SIZE)
dst = vki.SVVector("int", blocks)
kernel.launch(blocks, BLOCK_SIZE, [dst, src, vki.SVUInt32(n)])
print(dst.to_host()[0])