compiler/nir: add an is_conversion field to nir_op_info

This is set to True only for numeric conversion opcodes.

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This commit is contained in:
Iago Toral Quiroga
2019-02-12 12:55:28 +01:00
committed by Jason Ekstrand
parent 55e6454d5e
commit ca2b5e9069
3 changed files with 47 additions and 33 deletions

View File

@@ -918,6 +918,9 @@ typedef struct {
nir_alu_type input_types[NIR_MAX_VEC_COMPONENTS]; nir_alu_type input_types[NIR_MAX_VEC_COMPONENTS];
nir_op_algebraic_property algebraic_properties; nir_op_algebraic_property algebraic_properties;
/* Whether this represents a numeric conversion opcode */
bool is_conversion;
} nir_op_info; } nir_op_info;
extern const nir_op_info nir_op_infos[nir_num_opcodes]; extern const nir_op_info nir_op_infos[nir_num_opcodes];

View File

@@ -33,12 +33,13 @@ class Opcode(object):
NOTE: this must be kept in sync with nir_op_info NOTE: this must be kept in sync with nir_op_info
""" """
def __init__(self, name, output_size, output_type, input_sizes, def __init__(self, name, output_size, output_type, input_sizes,
input_types, algebraic_properties, const_expr): input_types, is_conversion, algebraic_properties, const_expr):
"""Parameters: """Parameters:
- name is the name of the opcode (prepend nir_op_ for the enum name) - name is the name of the opcode (prepend nir_op_ for the enum name)
- all types are strings that get nir_type_ prepended to them - all types are strings that get nir_type_ prepended to them
- input_types is a list of types - input_types is a list of types
- is_conversion is true if this opcode represents a type conversion
- algebraic_properties is a space-seperated string, where nir_op_is_ is - algebraic_properties is a space-seperated string, where nir_op_is_ is
prepended before each entry prepended before each entry
- const_expr is an expression or series of statements that computes the - const_expr is an expression or series of statements that computes the
@@ -70,6 +71,7 @@ class Opcode(object):
assert isinstance(input_sizes[0], int) assert isinstance(input_sizes[0], int)
assert isinstance(input_types, list) assert isinstance(input_types, list)
assert isinstance(input_types[0], str) assert isinstance(input_types[0], str)
assert isinstance(is_conversion, bool)
assert isinstance(algebraic_properties, str) assert isinstance(algebraic_properties, str)
assert isinstance(const_expr, str) assert isinstance(const_expr, str)
assert len(input_sizes) == len(input_types) assert len(input_sizes) == len(input_types)
@@ -84,6 +86,7 @@ class Opcode(object):
self.output_type = output_type self.output_type = output_type
self.input_sizes = input_sizes self.input_sizes = input_sizes
self.input_types = input_types self.input_types = input_types
self.is_conversion = is_conversion
self.algebraic_properties = algebraic_properties self.algebraic_properties = algebraic_properties
self.const_expr = const_expr self.const_expr = const_expr
@@ -138,21 +141,22 @@ associative = "associative "
opcodes = {} opcodes = {}
def opcode(name, output_size, output_type, input_sizes, input_types, def opcode(name, output_size, output_type, input_sizes, input_types,
algebraic_properties, const_expr): is_conversion, algebraic_properties, const_expr):
assert name not in opcodes assert name not in opcodes
opcodes[name] = Opcode(name, output_size, output_type, input_sizes, opcodes[name] = Opcode(name, output_size, output_type, input_sizes,
input_types, algebraic_properties, const_expr) input_types, is_conversion, algebraic_properties,
const_expr)
def unop_convert(name, out_type, in_type, const_expr): def unop_convert(name, out_type, in_type, const_expr):
opcode(name, 0, out_type, [0], [in_type], "", const_expr) opcode(name, 0, out_type, [0], [in_type], False, "", const_expr)
def unop(name, ty, const_expr): def unop(name, ty, const_expr):
opcode(name, 0, ty, [0], [ty], "", const_expr) opcode(name, 0, ty, [0], [ty], False, "", const_expr)
def unop_horiz(name, output_size, output_type, input_size, input_type, def unop_horiz(name, output_size, output_type, input_size, input_type,
const_expr): const_expr):
opcode(name, output_size, output_type, [input_size], [input_type], "", opcode(name, output_size, output_type, [input_size], [input_type],
const_expr) False, "", const_expr)
def unop_reduce(name, output_size, output_type, input_type, prereduce_expr, def unop_reduce(name, output_size, output_type, input_type, prereduce_expr,
reduce_expr, final_expr): reduce_expr, final_expr):
@@ -173,6 +177,8 @@ def unop_reduce(name, output_size, output_type, input_type, prereduce_expr,
unop_horiz(name + "4", output_size, output_type, 4, input_type, unop_horiz(name + "4", output_size, output_type, 4, input_type,
final(reduce_(reduce_(src0, src1), reduce_(src2, src3)))) final(reduce_(reduce_(src0, src1), reduce_(src2, src3))))
def unop_numeric_convert(name, out_type, in_type, const_expr):
opcode(name, 0, out_type, [0], [in_type], True, "", const_expr)
# These two move instructions differ in what modifiers they support and what # These two move instructions differ in what modifiers they support and what
# the negate modifier means. Otherwise, they are identical. # the negate modifier means. Otherwise, they are identical.
@@ -215,13 +221,13 @@ for src_t in [tint, tuint, tfloat, tbool]:
if bit_size == 16 and dst_t == tfloat and src_t == tfloat: if bit_size == 16 and dst_t == tfloat and src_t == tfloat:
rnd_modes = ['_rtne', '_rtz', ''] rnd_modes = ['_rtne', '_rtz', '']
for rnd_mode in rnd_modes: for rnd_mode in rnd_modes:
unop_convert("{0}2{1}{2}{3}".format(src_t[0], dst_t[0], unop_numeric_convert("{0}2{1}{2}{3}".format(src_t[0], dst_t[0],
bit_size, rnd_mode), bit_size, rnd_mode),
dst_t + str(bit_size), src_t, "src0") dst_t + str(bit_size), src_t, "src0")
else: else:
conv_expr = "src0 != 0" if dst_t == tbool else "src0" conv_expr = "src0 != 0" if dst_t == tbool else "src0"
unop_convert("{0}2{1}{2}".format(src_t[0], dst_t[0], bit_size), unop_numeric_convert("{0}2{1}{2}".format(src_t[0], dst_t[0], bit_size),
dst_t + str(bit_size), src_t, conv_expr) dst_t + str(bit_size), src_t, conv_expr)
# Unary floating-point rounding operations. # Unary floating-point rounding operations.
@@ -426,7 +432,8 @@ if (src0.z < 0 && absZ >= absX && absZ >= absY) dst.x = 5;
def binop_convert(name, out_type, in_type, alg_props, const_expr): def binop_convert(name, out_type, in_type, alg_props, const_expr):
opcode(name, 0, out_type, [0, 0], [in_type, in_type], alg_props, const_expr) opcode(name, 0, out_type, [0, 0], [in_type, in_type],
False, alg_props, const_expr)
def binop(name, ty, alg_props, const_expr): def binop(name, ty, alg_props, const_expr):
binop_convert(name, ty, ty, alg_props, const_expr) binop_convert(name, ty, ty, alg_props, const_expr)
@@ -440,7 +447,7 @@ def binop_compare32(name, ty, alg_props, const_expr):
def binop_horiz(name, out_size, out_type, src1_size, src1_type, src2_size, def binop_horiz(name, out_size, out_type, src1_size, src1_type, src2_size,
src2_type, const_expr): src2_type, const_expr):
opcode(name, out_size, out_type, [src1_size, src2_size], [src1_type, src2_type], opcode(name, out_size, out_type, [src1_size, src2_size], [src1_type, src2_type],
"", const_expr) False, "", const_expr)
def binop_reduce(name, output_size, output_type, src_type, prereduce_expr, def binop_reduce(name, output_size, output_type, src_type, prereduce_expr,
reduce_expr, final_expr): reduce_expr, final_expr):
@@ -455,13 +462,13 @@ def binop_reduce(name, output_size, output_type, src_type, prereduce_expr,
src2 = prereduce("src0.z", "src1.z") src2 = prereduce("src0.z", "src1.z")
src3 = prereduce("src0.w", "src1.w") src3 = prereduce("src0.w", "src1.w")
opcode(name + "2", output_size, output_type, opcode(name + "2", output_size, output_type,
[2, 2], [src_type, src_type], commutative, [2, 2], [src_type, src_type], False, commutative,
final(reduce_(src0, src1))) final(reduce_(src0, src1)))
opcode(name + "3", output_size, output_type, opcode(name + "3", output_size, output_type,
[3, 3], [src_type, src_type], commutative, [3, 3], [src_type, src_type], False, commutative,
final(reduce_(reduce_(src0, src1), src2))) final(reduce_(reduce_(src0, src1), src2)))
opcode(name + "4", output_size, output_type, opcode(name + "4", output_size, output_type,
[4, 4], [src_type, src_type], commutative, [4, 4], [src_type, src_type], False, commutative,
final(reduce_(reduce_(src0, src1), reduce_(src2, src3)))) final(reduce_(reduce_(src0, src1), reduce_(src2, src3))))
binop("fadd", tfloat, commutative + associative, "src0 + src1") binop("fadd", tfloat, commutative + associative, "src0 + src1")
@@ -657,9 +664,12 @@ binop("sne", tfloat32, commutative, "(src0 != src1) ? 1.0f : 0.0f") # Set on Not
# SPIRV shifts are undefined for shift-operands >= bitsize, # SPIRV shifts are undefined for shift-operands >= bitsize,
# but SM5 shifts are defined to use the least significant bits, only # but SM5 shifts are defined to use the least significant bits, only
# The NIR definition is according to the SM5 specification. # The NIR definition is according to the SM5 specification.
opcode("ishl", 0, tint, [0, 0], [tint, tuint32], "", "src0 << (src1 & (sizeof(src0) * 8 - 1))") opcode("ishl", 0, tint, [0, 0], [tint, tuint32], False, "",
opcode("ishr", 0, tint, [0, 0], [tint, tuint32], "", "src0 >> (src1 & (sizeof(src0) * 8 - 1))") "src0 << (src1 & (sizeof(src0) * 8 - 1))")
opcode("ushr", 0, tuint, [0, 0], [tuint, tuint32], "", "src0 >> (src1 & (sizeof(src0) * 8 - 1))") opcode("ishr", 0, tint, [0, 0], [tint, tuint32], False, "",
"src0 >> (src1 & (sizeof(src0) * 8 - 1))")
opcode("ushr", 0, tuint, [0, 0], [tuint, tuint32], False, "",
"src0 >> (src1 & (sizeof(src0) * 8 - 1))")
# bitwise logic operators # bitwise logic operators
# #
@@ -690,9 +700,9 @@ binop_reduce("fdot", 1, tfloat, tfloat, "{src0} * {src1}", "{src0} + {src1}",
binop_reduce("fdot_replicated", 4, tfloat, tfloat, binop_reduce("fdot_replicated", 4, tfloat, tfloat,
"{src0} * {src1}", "{src0} + {src1}", "{src}") "{src0} * {src1}", "{src0} + {src1}", "{src}")
opcode("fdph", 1, tfloat, [3, 4], [tfloat, tfloat], "", opcode("fdph", 1, tfloat, [3, 4], [tfloat, tfloat], False, "",
"src0.x * src1.x + src0.y * src1.y + src0.z * src1.z + src1.w") "src0.x * src1.x + src0.y * src1.y + src0.z * src1.z + src1.w")
opcode("fdph_replicated", 4, tfloat, [3, 4], [tfloat, tfloat], "", opcode("fdph_replicated", 4, tfloat, [3, 4], [tfloat, tfloat], False, "",
"src0.x * src1.x + src0.y * src1.y + src0.z * src1.z + src1.w") "src0.x * src1.x + src0.y * src1.y + src0.z * src1.z + src1.w")
binop("fmin", tfloat, "", "fminf(src0, src1)") binop("fmin", tfloat, "", "fminf(src0, src1)")
@@ -769,7 +779,7 @@ else
dst = ((1u << bits) - 1) << offset; dst = ((1u << bits) - 1) << offset;
""") """)
opcode("ldexp", 0, tfloat, [0, 0], [tfloat, tint32], "", """ opcode("ldexp", 0, tfloat, [0, 0], [tfloat, tint32], False, "", """
dst = (bit_size == 64) ? ldexp(src0, src1) : ldexpf(src0, src1); dst = (bit_size == 64) ? ldexp(src0, src1) : ldexpf(src0, src1);
/* flush denormals to zero. */ /* flush denormals to zero. */
if (!isnormal(dst)) if (!isnormal(dst))
@@ -793,11 +803,11 @@ binop("extract_i16", tint, "", "(int16_t)(src0 >> (src1 * 16))")
def triop(name, ty, const_expr): def triop(name, ty, const_expr):
opcode(name, 0, ty, [0, 0, 0], [ty, ty, ty], "", const_expr) opcode(name, 0, ty, [0, 0, 0], [ty, ty, ty], False, "", const_expr)
def triop_horiz(name, output_size, src1_size, src2_size, src3_size, const_expr): def triop_horiz(name, output_size, src1_size, src2_size, src3_size, const_expr):
opcode(name, output_size, tuint, opcode(name, output_size, tuint,
[src1_size, src2_size, src3_size], [src1_size, src2_size, src3_size],
[tuint, tuint, tuint], "", const_expr) [tuint, tuint, tuint], False, "", const_expr)
triop("ffma", tfloat, "src0 * src1 + src2") triop("ffma", tfloat, "src0 * src1 + src2")
@@ -826,9 +836,9 @@ triop("imed3", tint, "MAX2(MIN2(MAX2(src0, src1), src2), MIN2(src0, src1))")
triop("umed3", tuint, "MAX2(MIN2(MAX2(src0, src1), src2), MIN2(src0, src1))") triop("umed3", tuint, "MAX2(MIN2(MAX2(src0, src1), src2), MIN2(src0, src1))")
opcode("bcsel", 0, tuint, [0, 0, 0], opcode("bcsel", 0, tuint, [0, 0, 0],
[tbool1, tuint, tuint], "", "src0 ? src1 : src2") [tbool1, tuint, tuint], False, "", "src0 ? src1 : src2")
opcode("b32csel", 0, tuint, [0, 0, 0], opcode("b32csel", 0, tuint, [0, 0, 0],
[tbool32, tuint, tuint], "", "src0 ? src1 : src2") [tbool32, tuint, tuint], False, "", "src0 ? src1 : src2")
# SM5 bfi assembly # SM5 bfi assembly
triop("bfi", tuint32, """ triop("bfi", tuint32, """
@@ -847,7 +857,7 @@ if (mask == 0) {
# SM5 ubfe/ibfe assembly # SM5 ubfe/ibfe assembly
opcode("ubfe", 0, tuint32, opcode("ubfe", 0, tuint32,
[0, 0, 0], [tuint32, tint32, tint32], "", """ [0, 0, 0], [tuint32, tint32, tint32], False, "", """
unsigned base = src0; unsigned base = src0;
int offset = src1, bits = src2; int offset = src1, bits = src2;
if (bits == 0) { if (bits == 0) {
@@ -861,7 +871,7 @@ if (bits == 0) {
} }
""") """)
opcode("ibfe", 0, tint32, opcode("ibfe", 0, tint32,
[0, 0, 0], [tint32, tint32, tint32], "", """ [0, 0, 0], [tint32, tint32, tint32], False, "", """
int base = src0; int base = src0;
int offset = src1, bits = src2; int offset = src1, bits = src2;
if (bits == 0) { if (bits == 0) {
@@ -877,7 +887,7 @@ if (bits == 0) {
# GLSL bitfieldExtract() # GLSL bitfieldExtract()
opcode("ubitfield_extract", 0, tuint32, opcode("ubitfield_extract", 0, tuint32,
[0, 0, 0], [tuint32, tint32, tint32], "", """ [0, 0, 0], [tuint32, tint32, tint32], False, "", """
unsigned base = src0; unsigned base = src0;
int offset = src1, bits = src2; int offset = src1, bits = src2;
if (bits == 0) { if (bits == 0) {
@@ -889,7 +899,7 @@ if (bits == 0) {
} }
""") """)
opcode("ibitfield_extract", 0, tint32, opcode("ibitfield_extract", 0, tint32,
[0, 0, 0], [tint32, tint32, tint32], "", """ [0, 0, 0], [tint32, tint32, tint32], False, "", """
int base = src0; int base = src0;
int offset = src1, bits = src2; int offset = src1, bits = src2;
if (bits == 0) { if (bits == 0) {
@@ -914,10 +924,10 @@ def quadop_horiz(name, output_size, src1_size, src2_size, src3_size,
opcode(name, output_size, tuint, opcode(name, output_size, tuint,
[src1_size, src2_size, src3_size, src4_size], [src1_size, src2_size, src3_size, src4_size],
[tuint, tuint, tuint, tuint], [tuint, tuint, tuint, tuint],
"", const_expr) False, "", const_expr)
opcode("bitfield_insert", 0, tuint32, [0, 0, 0, 0], opcode("bitfield_insert", 0, tuint32, [0, 0, 0, 0],
[tuint32, tuint32, tint32, tint32], "", """ [tuint32, tuint32, tint32, tint32], False, "", """
unsigned base = src0, insert = src1; unsigned base = src0, insert = src1;
int offset = src2, bits = src3; int offset = src2, bits = src3;
if (bits == 0) { if (bits == 0) {

View File

@@ -117,6 +117,7 @@ const nir_op_info nir_op_infos[nir_num_opcodes] = {
.input_types = { .input_types = {
${ ", ".join("nir_type_" + type for type in opcode.input_types) } ${ ", ".join("nir_type_" + type for type in opcode.input_types) }
}, },
.is_conversion = ${"true" if opcode.is_conversion else "false"},
.algebraic_properties = .algebraic_properties =
${ "0" if opcode.algebraic_properties == "" else " | ".join( ${ "0" if opcode.algebraic_properties == "" else " | ".join(
"NIR_OP_IS_" + prop.upper() for prop in "NIR_OP_IS_" + prop.upper() for prop in