Commit 8f72741d authored by Markus Holzer's avatar Markus Holzer
Browse files

Merge branch 'scattergather' into 'master'

Vector scatter/gather support

Closes #34

See merge request pycodegen/pystencils!241
parents 63536d22 d314ef10
......@@ -88,9 +88,16 @@ def get_vector_instruction_set_arm(data_type='double', instruction_set='neon'):
result['makeVecConstInt'] = f'svdup_s{bits["int"]}' + '({0})'
result['makeVecIndex'] = f'svindex_s{bits["int"]}' + '({0}, {1})'
vindex = f'svindex_u{bits[data_type]}(0, {{0}})'
result['scatter'] = f'svst1_scatter_u{bits[data_type]}index_f{bits[data_type]}({predicate}, {{0}}, ' + \
vindex.format("{2}") + ', {1})'
result['gather'] = f'svld1_gather_u{bits[data_type]}index_f{bits[data_type]}({predicate}, {{0}}, ' + \
vindex.format("{1}") + ')'
result['+int'] = f"svadd_s{bits['int']}_x({int_predicate}, " + "{0}, {1})"
result[data_type] = f'svfloat{bits[data_type]}_st'
result['float'] = 'svfloat32_st'
result['double'] = 'svfloat64_st'
result['int'] = f'svint{bits["int"]}_st'
result['bool'] = 'svbool_st'
......@@ -102,6 +109,10 @@ def get_vector_instruction_set_arm(data_type='double', instruction_set='neon'):
result['any'] = f'svptest_any({predicate}, {{0}})'
result['all'] = f'svcntp_b{bits[data_type]}({predicate}, {{0}}) == {width}'
result['maskStoreU'] = result['storeU'].replace(predicate, '{2}')
result['maskStoreA'] = result['storeA'].replace(predicate, '{2}')
result['maskScatter'] = result['scatter'].replace(predicate, '{3}')
result['compile_flags'] = [f'-msve-vector-bits={bitwidth}']
else:
result['makeVecConst'] = f'vdupq_n_f{bits[data_type]}' + '({0})'
......
......@@ -256,7 +256,7 @@ class CBackend:
lhs_type = get_type_of_expression(node.lhs)
printed_mask = ""
if type(lhs_type) is VectorType and isinstance(node.lhs, cast_func):
arg, data_type, aligned, nontemporal, mask = node.lhs.args
arg, data_type, aligned, nontemporal, mask, stride = node.lhs.args
instr = 'storeU'
if aligned:
instr = 'stream' if nontemporal and 'stream' in self._vector_instruction_set else 'storeA'
......@@ -285,6 +285,12 @@ class CBackend:
rhs = node.rhs
ptr = "&" + self.sympy_printer.doprint(node.lhs.args[0])
if stride != 1:
instr = 'maskScatter' if mask != True else 'scatter' # NOQA
return self._vector_instruction_set[instr].format(ptr, self.sympy_printer.doprint(rhs),
stride, printed_mask) + ';'
pre_code = ''
if nontemporal and 'cachelineZero' in self._vector_instruction_set:
pre_code = f"if (((uintptr_t) {ptr} & {CachelineSize.mask_symbol}) == 0) " + "{\n\t" + \
......@@ -609,7 +615,9 @@ class VectorizedCustomSympyPrinter(CustomSympyPrinter):
def _print_Function(self, expr):
if isinstance(expr, vector_memory_access):
arg, data_type, aligned, _, mask = expr.args
arg, data_type, aligned, _, mask, stride = expr.args
if stride != 1:
return self.instruction_set['gather'].format("& " + self._print(arg), stride)
instruction = self.instruction_set['loadA'] if aligned else self.instruction_set['loadU']
return instruction.format("& " + self._print(arg))
elif isinstance(expr, cast_func):
......
......@@ -130,7 +130,7 @@ def get_vector_instruction_set_x86(data_type='double', instruction_set='avx'):
result['all'] = f"{pre}_movemask_{suf}({{0}}) == {hex(2**result['width']-1)}"
if instruction_set == 'avx512':
size = 8 if data_type == 'double' else 16
size = result['width']
result['&'] = f'_kand_mask{size}({{0}}, {{1}})'
result['|'] = f'_kor_mask{size}({{0}}, {{1}})'
result['any'] = f'!_ktestz_mask{size}_u8({{0}}, {{0}})'
......@@ -145,6 +145,14 @@ def get_vector_instruction_set_x86(data_type='double', instruction_set='avx'):
params = " | ".join(["({{0}} ? {power} : 0)".format(power=2 ** i) for i in range(8)])
result['makeVecConstBool'] = f"__mmask8(({params}) )"
vindex = f'{pre}_set_epi{bit_width//size}(' + ', '.join([str(i) for i in range(result['width'])][::-1]) + ')'
vindex = f'{pre}_mullo_epi{bit_width//size}({vindex}, {pre}_set1_epi{bit_width//size}({{0}}))'
result['scatter'] = f'{pre}_i{bit_width//size}scatter_{suf}({{0}}, ' + vindex.format("{2}") + \
f', {{1}}, {64//size})'
result['maskScatter'] = f'{pre}_mask_i{bit_width//size}scatter_{suf}({{0}}, {{3}}, ' + vindex.format("{2}") + \
f', {{1}}, {64//size})'
result['gather'] = f'{pre}_i{bit_width//size}gather_{suf}(' + vindex.format("{1}") + f', {{0}}, {64//size})'
if instruction_set == 'avx' and data_type == 'float':
result['rsqrt'] = f"{pre}_rsqrt_{suf}({{0}})"
......
......@@ -80,8 +80,9 @@ def vectorize(kernel_ast: ast.KernelFunction, instruction_set: str = 'best',
kernel_ast.instruction_set = vector_is
vectorize_rng(kernel_ast, vector_width)
vectorize_inner_loops_and_adapt_load_stores(kernel_ast, vector_width, assume_aligned,
nontemporal, assume_sufficient_line_padding)
scattergather = 'scatter' in vector_is and 'gather' in vector_is
vectorize_inner_loops_and_adapt_load_stores(kernel_ast, vector_width, assume_aligned, nontemporal,
scattergather, assume_sufficient_line_padding)
insert_vector_casts(kernel_ast)
......@@ -104,7 +105,7 @@ def vectorize_rng(kernel_ast, vector_width):
def vectorize_inner_loops_and_adapt_load_stores(ast_node, vector_width, assume_aligned, nontemporal_fields,
assume_sufficient_line_padding):
scattergather, assume_sufficient_line_padding):
"""Goes over all innermost loops, changes increment to vector width and replaces field accesses by vector type."""
all_loops = filtered_tree_iteration(ast_node, ast.LoopOverCoordinate, stop_type=ast.SympyAssignment)
inner_loops = [n for n in all_loops if n.is_innermost_loop]
......@@ -135,7 +136,8 @@ def vectorize_inner_loops_and_adapt_load_stores(ast_node, vector_width, assume_a
if loop_counter_symbol in index.atoms(sp.Symbol):
loop_counter_is_offset = loop_counter_symbol not in (index - loop_counter_symbol).atoms()
aligned_access = (index - loop_counter_symbol).subs(zero_loop_counters) == 0
if not loop_counter_is_offset:
stride = sp.simplify(index.subs({loop_counter_symbol: loop_counter_symbol + 1}) - index)
if not loop_counter_is_offset and (not scattergather or loop_counter_symbol in stride.atoms()):
successful = False
break
typed_symbol = base.label
......@@ -147,7 +149,8 @@ def vectorize_inner_loops_and_adapt_load_stores(ast_node, vector_width, assume_a
nontemporal = False
if hasattr(indexed, 'field'):
nontemporal = (indexed.field in nontemporal_fields) or (indexed.field.name in nontemporal_fields)
substitutions[indexed] = vector_memory_access(indexed, vec_type, use_aligned_access, nontemporal, True)
substitutions[indexed] = vector_memory_access(indexed, vec_type, use_aligned_access, nontemporal, True,
stride if scattergather else 1)
if nontemporal:
# insert NontemporalFence after the outermost loop
parent = loop_node.parent
......@@ -188,7 +191,7 @@ def mask_conditionals(loop_body):
node.condition_expr = vec_any(node.condition_expr)
elif isinstance(node, ast.SympyAssignment):
if mask is not True:
s = {ma: vector_memory_access(ma.args[0], ma.args[1], ma.args[2], ma.args[3], sp.And(mask, ma.args[4]))
s = {ma: vector_memory_access(*ma.args[0:4], sp.And(mask, ma.args[4]), *ma.args[5:])
for ma in node.atoms(vector_memory_access)}
node.subs(s)
else:
......@@ -205,8 +208,7 @@ def insert_vector_casts(ast_node):
def visit_expr(expr):
if isinstance(expr, vector_memory_access):
return vector_memory_access(expr.args[0], expr.args[1], expr.args[2], expr.args[3],
visit_expr(expr.args[4]))
return vector_memory_access(*expr.args[0:4], visit_expr(expr.args[4]), *expr.args[5:])
elif isinstance(expr, cast_func):
return expr
elif expr.func is sp.Abs and 'abs' not in ast_node.instruction_set:
......
......@@ -195,8 +195,8 @@ class boolean_cast_func(cast_func, Boolean):
# noinspection PyPep8Naming
class vector_memory_access(cast_func):
# Arguments are: read/write expression, type, aligned, nontemporal, mask (or none)
nargs = (5,)
# Arguments are: read/write expression, type, aligned, nontemporal, mask (or none), stride
nargs = (6,)
# noinspection PyPep8Naming
......
......@@ -53,6 +53,31 @@ def test_vectorized_abs(instruction_set, dtype):
np.testing.assert_equal(np.sum(dst[1:-1, 1:-1]), 2 ** 2 * 2 ** 3)
@pytest.mark.parametrize('dtype', ('float', 'double'))
@pytest.mark.parametrize('instruction_set', supported_instruction_sets)
def test_scatter_gather(instruction_set, dtype):
f, g = ps.fields(f"f, g : float{64 if dtype == 'double' else 32}[2D]")
update_rule = [ps.Assignment(g[0, 0], f[0, 0] + f[-1, 0] + f[1, 0] + f[0, 1] + f[0, -1] + 42.0)]
if 'scatter' not in get_vector_instruction_set(dtype, instruction_set) and not instruction_set == 'avx512' and not instruction_set.startswith('sve'):
with pytest.warns(UserWarning) as warn:
ast = ps.create_kernel(update_rule, cpu_vectorize_info={'instruction_set': instruction_set})
assert 'Could not vectorize loop' in warn[0].message.args[0]
else:
with pytest.warns(None) as warn:
ast = ps.create_kernel(update_rule, cpu_vectorize_info={'instruction_set': instruction_set})
assert len(warn) == 0
func = ast.compile()
ref_func = ps.create_kernel(update_rule).compile()
arr = np.random.random((23 + 2, 17 + 2)).astype(np.float64 if dtype == 'double' else np.float32)
dst = np.zeros_like(arr, dtype=np.float64 if dtype == 'double' else np.float32)
ref = np.zeros_like(arr, dtype=np.float64 if dtype == 'double' else np.float32)
func(g=dst, f=arr)
ref_func(g=ref, f=arr)
np.testing.assert_almost_equal(dst, ref, 13 if dtype == 'double' else 5)
@pytest.mark.parametrize('dtype', ('float', 'double'))
@pytest.mark.parametrize('instruction_set', supported_instruction_sets)
@pytest.mark.parametrize('gl_field, gl_kernel', [(1, 0), (0, 1), (1, 1)])
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment