def GenerateMulCols(emitter, result_type, lhs_add, rhs_add, aligned, cols,
leftovers):
"""Emits code responsible for multiplication of one horizontal lhs strip."""
emitter.EmitOpenBracket('for (int i = 0; i < col_chunks; ++i)')
emitter.EmitCall(
zip_Nx8_neon.BuildName(4, leftovers, aligned),
['rhs_chunk', 'k', 'k', 'zipped_rhs_1', 'lhs_offset', 'const_offset'])
emitter.EmitAssignIncrement('rhs_chunk', 'chunk_size')
emitter.EmitCall(
zip_Nx8_neon.BuildName(4, leftovers, aligned),
['rhs_chunk', 'k', 'k', 'zipped_rhs_2', 'lhs_offset', 'const_offset'])
emitter.EmitAssignIncrement('rhs_chunk', 'chunk_size')
emitter.EmitCall(
mul_1x8_Mx8_neon.BuildName(result_type, lhs_add, rhs_add, 8),
GetMul2Params(result_type))
emitter.EmitAssignIncrement('mul_result_chunk', 8)
emitter.EmitCloseBracket()
if cols > 4:
emitter.EmitCall(zip_Nx8_neon.BuildName(4, leftovers, aligned), [
'rhs_chunk', 'k', 'k', 'zipped_rhs_1', 'lhs_offset', 'const_offset'
])
emitter.EmitAssignIncrement('rhs_chunk', 'chunk_size')
emitter.EmitCall(zip_Nx8_neon.BuildName(cols - 4, leftovers, aligned),
[
'rhs_chunk', 'k', 'k', 'zipped_rhs_2',
'lhs_offset', 'const_offset'
])
emitter.EmitCall(
mul_1x8_Mx8_neon.BuildName(result_type, lhs_add, rhs_add, cols),
GetMul2Params(result_type))
elif cols > 0:
emitter.EmitCall(zip_Nx8_neon.BuildName(cols, leftovers, aligned), [
'rhs_chunk', 'k', 'k', 'zipped_rhs_1', 'lhs_offset', 'const_offset'
])
emitter.EmitCall(
mul_Nx8_Mx8_neon.BuildName(result_type, lhs_add, rhs_add, 1, cols),
GetMulParams(result_type))
def ZipName(rows, leftovers, aligned):
return zip_Nx8_neon.BuildName(rows, leftovers, aligned)
def GenerateZipVector(emitter, aligned, leftovers):
emitter.EmitCall(zip_Nx8_neon.BuildName(1, leftovers, aligned),
['lhs', 'k', 'k', 'zipped_lhs', 'rhs_offset', 0])