notes.txt

A PARAMETER is a tuple of:
  name   -- string identifier,                            eg "x"
  type   -- C++ or cuda type as string,                   eg "double"
  arity  -- number of elements per set element,           eg 3
  set    -- per-particle(P) or per-neighbor(N),           eg P
  access -- read-only(RO), read-write(RW) or reduce(SUM), eg RO
  reload -- mutates between kernel invocations,           eg True
  id     -- additive identity,                            eg {0.0f, 0.0f, 0.0f}

(set, access) = (N, SUM) is not defined

pair_interaction device kernel, parameter list:
  if set, access = P, RO:
    <type> <name>i[arity]
    <type> <name>j[arity]
  elif set, access = P, RW:
    if arity > 1:
      <type> <name>i[arity]
    else:
      <type> *<name>i
  elif set, access = P, SUM:
    <type> <name>i_delta[arity]
  elif set, access = N, RO:
    <type> <name>[arity]
  elif set, access = N, RW:
    if arity > 1:
      <type> <name>[arity]
    else:
      <type> *<name>

struct particle declaration:
  if set, access = P, RO:
    <type> <name>[arity]
  else:
    error

compute kernel, parameter list:
  -- set, access = P, RO passed through struct particle *neigh
  if set, access = P, RW:
    <type> *<name>i_list -- list of len nparticles * arity
  if set, access = P, SUM:
    <type> *<name>i_list -- list of len nparticles * arity
  elif set, access = N, RO:
    <type> <name>_list   -- list of len nneighbors * arity
  elif set, access = N, RW:
    <type> <name>_list   -- list of len nneighbors * arity
  else:
    error

compute kernel, register declaration:
  -- same as pair_interaction device kernel, parameter list

compute kernel, particle(i) load:
  if set, access = P, RO:
    if arity > 1:
      for k in range(arity):
        <name>i[k] = particle_aos[idx].<name>[k]
    else:
      <name>i = particle_aos[idx].<name>
  elif set, access = P, RW:
    if arity > 1:
      for k in range(arity):
        <name>i[k] = <name>_list[(idx*arity)+k]
    else:
      <name>i = <name>_list[idx]
  elif set, access = P, SUM:
    <name>i_delta = id
  -- set = N variables are loaded at same time as particle(j)

compute kernel, particle(j) load:
  if set, access = P, RO:
    if arity > 1:
      for k in range(arity):
        <name>j[k] = neigh[neigh_idx].<name>[k]
    else:
      <name>j = neigh[neigh_idx].<name>
  elif set, access = P, RW|SUM:
    error
  elif set, access = N, RO|RW:
    if arity > 1:
      for k in range(arity):
        <name>[k] = <name>_list[(neigh_idx*arity)+k]
    else:
      <name> = <name>_list[neigh_idx]

compute kernel, pairwise call parameter list:
  if set, access = P, RO:
    <name>i
    <name>j
  elif set, access = P, RW:
    if arity > 1:
      <name>i
    else:
      &<name>i
  elif set, access = P, SUM:
    <name>i_delta
  elif set, access = N, RO|RW:
    if arity > 1:
      <name>
    else:
      &<name>

compute kernel, post pairwise call (still looping through jj):
  -- update N,RW data
  if set, access = N, RW:
    if arity > 1:
      for k in range(arity):
        <name>_list[(neigh_idx*arity)+k] = <name>[k]
    else:
      <name>_list[neigh_idx] = <name>

compute kernel, completed jj loop:
  -- update P,RW|SUM data
  if set, access = P, RW:
    if arity > 1:
      for k in range(arity):
        <name>_list[(idx*arity)+k] = <name>i[k]
    else:
      <name>_list[idx] = <type> <name>i
  elif set, access = P, SUM:
    if arity > 1:
      for k in range(arity):
        <name>_list[(idx*arity)+k] = <name>i_delta[k]
    else:
      <name>_list[idx] = <name>i_delta