def gen_calc_planarity_inlined(code, block_size, args):
''' Optimised version by manually inlining all code. '''
filterbank = args['filterbank']
f = filterbank.filters[0]
rows, cols = block_size
frows, fcols = f.size()
hfrow, hfcol = [x//2 for x in f.size()]
out_ptr = args['out_ptr'] if 'out_ptr' in args else rows*cols
in_ptr = args['in_ptr'] if 'in_ptr' in args else 0
buffer_ptr = rows*cols*2
assert buffer_ptr != in_ptr
assert buffer_ptr != out_ptr
for filter_nr, f in enumerate(filterbank.filters):
# convolution + abs
for i in xrange(rows):
for j in xrange(cols):
with scoped_alloc(code, 1) as acc:
# convolution
yield Xor(acc, acc, acc)
for x, y, coeff in f.coefficients:
ii = i + y - hfrow
jj = j + x - hfcol
with scoped_alloc(code, 2) as (coeff_reg, v):
yield Imm(coeff_reg, coeff)
for instr in load_mem_value(code, in_ptr, (jj, ii), v, block_size):
yield instr
yield Mul(v, v, coeff_reg)
yield Add(acc, acc, v)
# take max
with scoped_alloc(code, 1) as const0:
yield Imm(const0, 0)
yield Cmp(acc, const0)
yield Neg(acc, acc, cond='LT')
yield MemWImm(buffer_ptr+i*cols + j, acc)
# gather
for i in xrange(rows):
for j in xrange(cols):
with scoped_alloc(code, 1) as max_v:
# local max
yield Imm(max_v, -float('inf'))
for ii in xrange(frows):
for jj in xrange(fcols):
if not f.mask[ii][jj]: continue # skip if not enabled
iii = i + ii - hfrow
jjj = j + jj - hfcol
with scoped_alloc(code, 1) as v:
for instr in load_mem_value(code, buffer_ptr, (jjj, iii), v, block_size):
yield instr
yield Cmp(v, max_v)
yield Mov(max_v, v, cond='GT')
# global max
if filter_nr != 0:
with scoped_alloc(code, 1) as old_v:
yield MemRImm(old_v, out_ptr+i*cols+j)
yield Cmp(old_v, max_v)
yield Mov(max_v, old_v, cond='GT')
yield MemWImm(out_ptr+i*cols + j, max_v)
def gen_gather_local_max(code, block_size, args):
''' Gather local maximum from mask code generation. '''
f = args['filter']
rows, cols = block_size
frows, fcols = f.size()
hfrow, hfcol = [x//2 for x in f.size()]
out_ptr = args['out_ptr'] if 'out_ptr' in args else rows*cols
in_ptr = args['in_ptr'] if 'in_ptr' in args else 0
for i in xrange(rows):
for j in xrange(cols):
with scoped_alloc(code, 1) as max_v:
yield Imm(max_v, -float('inf'))
for ii in xrange(frows):
for jj in xrange(fcols):
if not f.mask[ii][jj]: continue # skip if not enabled
iii = i + ii - hfrow
jjj = j + jj - hfcol
with scoped_alloc(code, 1) as v:
for instr in load_mem_value(code, in_ptr, (jjj, iii), v, block_size):
yield instr
yield Cmp(v, max_v)
yield Mov(max_v, v, cond='GT')
yield MemWImm(out_ptr+i*cols + j, max_v)
def gen_apply_sparse_filter(code, block_size, args): ''' Apply sparse filter code generation. ''' f = args['filter'] rows, cols = block_size hfrow, hfcol = [x//2 for x in f.size()] out_ptr = args['out_ptr'] if 'out_ptr' in args else rows*cols in_ptr = args['in_ptr'] if 'in_ptr' in args else 0 for i in xrange(rows): for j in xrange(cols): with scoped_alloc(code, 1) as acc: yield Xor(acc, acc, acc) for x, y, coeff in f.coefficients: ii = i + y - hfrow jj = j + x - hfcol with scoped_alloc(code, 2) as (coeff_reg, v): yield Imm(coeff_reg, coeff) for instr in load_mem_value(code, in_ptr, (jj, ii), v, block_size): yield instr yield Mul(v, v, coeff_reg) yield Add(acc, acc, v) yield MemWImm(out_ptr+i*cols + j, acc)
