X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fcompiler%2Fnir%2Fnir_opcodes.py;h=3020da98264e8253bb3855579f9374b7c682c21e;hb=7659c6197f08587f57f101a88a7e477337ce363c;hp=55b2ff0f678e01986a5a56713ca64dd5d067af73;hpb=dca6cd9ce65100896976e913bf72c2c68ea4e1a7;p=mesa.git diff --git a/src/compiler/nir/nir_opcodes.py b/src/compiler/nir/nir_opcodes.py index 55b2ff0f678..3020da98264 100644 --- a/src/compiler/nir/nir_opcodes.py +++ b/src/compiler/nir/nir_opcodes.py @@ -33,12 +33,13 @@ class Opcode(object): NOTE: this must be kept in sync with nir_op_info """ 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: - 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 - 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 prepended before each entry - 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_types, list) assert isinstance(input_types[0], str) + assert isinstance(is_conversion, bool) assert isinstance(algebraic_properties, str) assert isinstance(const_expr, str) assert len(input_sizes) == len(input_types) @@ -84,6 +86,7 @@ class Opcode(object): self.output_type = output_type self.input_sizes = input_sizes self.input_types = input_types + self.is_conversion = is_conversion self.algebraic_properties = algebraic_properties self.const_expr = const_expr @@ -91,6 +94,7 @@ class Opcode(object): tfloat = "float" tint = "int" tbool = "bool" +tbool1 = "bool1" tbool32 = "bool32" tuint = "uint" tuint16 = "uint16" @@ -119,39 +123,45 @@ def type_sizes(type_): if type_has_size(type_): return [type_size(type_)] elif type_ == 'bool': - return [32] + return [1, 32] elif type_ == 'float': return [16, 32, 64] else: - return [8, 16, 32, 64] + return [1, 8, 16, 32, 64] def type_base_type(type_): m = _TYPE_SPLIT_RE.match(type_) assert m is not None, 'Invalid NIR type string: "{}"'.format(type_) return m.group('type') -commutative = "commutative " +# Operation where the first two sources are commutative. +# +# For 2-source operations, this just mathematical commutativity. Some +# 3-source operations, like ffma, are only commutative in the first two +# sources. +_2src_commutative = "2src_commutative " associative = "associative " # global dictionary of opcodes opcodes = {} 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 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): - 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): - 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, const_expr): - opcode(name, output_size, output_type, [input_size], [input_type], "", - const_expr) + opcode(name, output_size, output_type, [input_size], [input_type], + False, "", const_expr) def unop_reduce(name, output_size, output_type, input_type, prereduce_expr, reduce_expr, final_expr): @@ -172,17 +182,14 @@ def unop_reduce(name, output_size, output_type, input_type, prereduce_expr, unop_horiz(name + "4", output_size, output_type, 4, input_type, 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 -# the negate modifier means. Otherwise, they are identical. -unop("fmov", tfloat, "src0") -unop("imov", tint, "src0") +unop("mov", tuint, "src0") unop("ineg", tint, "-src0") unop("fneg", tfloat, "-src0") unop("inot", tint, "~src0") # invert every bit of the integer -unop("fnot", tfloat, ("bit_size == 64 ? ((src0 == 0.0) ? 1.0 : 0.0f) : " + - "((src0 == 0.0f) ? 1.0f : 0.0f)")) unop("fsign", tfloat, ("bit_size == 64 ? " + "((src0 == 0.0) ? 0.0 : ((src0 > 0.0) ? 1.0 : -1.0)) : " + "((src0 == 0.0f) ? 0.0f : ((src0 > 0.0f) ? 1.0f : -1.0f))")) @@ -214,13 +221,13 @@ for src_t in [tint, tuint, tfloat, tbool]: if bit_size == 16 and dst_t == tfloat and src_t == tfloat: rnd_modes = ['_rtne', '_rtz', ''] for rnd_mode in rnd_modes: - unop_convert("{0}2{1}{2}{3}".format(src_t[0], dst_t[0], - bit_size, rnd_mode), - dst_t + str(bit_size), src_t, "src0") + unop_numeric_convert("{0}2{1}{2}{3}".format(src_t[0], dst_t[0], + bit_size, rnd_mode), + dst_t + str(bit_size), src_t, "src0") else: 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), - dst_t + str(bit_size), src_t, conv_expr) + 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) # Unary floating-point rounding operations. @@ -241,8 +248,8 @@ unop("fsin", tfloat, "bit_size == 64 ? sin(src0) : sinf(src0)") unop("fcos", tfloat, "bit_size == 64 ? cos(src0) : cosf(src0)") # dfrexp -unop_convert("frexp_exp", tint32, tfloat64, "frexp(src0, &dst);") -unop_convert("frexp_sig", tfloat64, tfloat64, "int n; dst = frexp(src0, &n);") +unop_convert("frexp_exp", tint32, tfloat, "frexp(src0, &dst);") +unop_convert("frexp_sig", tfloat, tfloat, "int n; dst = frexp(src0, &n);") # Partial derivatives. @@ -403,12 +410,21 @@ dst.x = dst.y = 0.0; float absX = fabs(src0.x); float absY = fabs(src0.y); float absZ = fabs(src0.z); -if (src0.x >= 0 && absX >= absY && absX >= absZ) { dst.x = -src0.y; dst.y = -src0.z; } -if (src0.x < 0 && absX >= absY && absX >= absZ) { dst.x = -src0.y; dst.y = src0.z; } -if (src0.y >= 0 && absY >= absX && absY >= absZ) { dst.x = src0.z; dst.y = src0.x; } -if (src0.y < 0 && absY >= absX && absY >= absZ) { dst.x = -src0.z; dst.y = src0.x; } -if (src0.z >= 0 && absZ >= absX && absZ >= absY) { dst.x = -src0.y; dst.y = src0.x; } -if (src0.z < 0 && absZ >= absX && absZ >= absY) { dst.x = -src0.y; dst.y = -src0.x; } + +float ma = 0.0; +if (absX >= absY && absX >= absZ) { ma = 2 * src0.x; } +if (absY >= absX && absY >= absZ) { ma = 2 * src0.y; } +if (absZ >= absX && absZ >= absY) { ma = 2 * src0.z; } + +if (src0.x >= 0 && absX >= absY && absX >= absZ) { dst.x = -src0.z; dst.y = -src0.y; } +if (src0.x < 0 && absX >= absY && absX >= absZ) { dst.x = src0.z; dst.y = -src0.y; } +if (src0.y >= 0 && absY >= absX && absY >= absZ) { dst.x = src0.x; dst.y = src0.z; } +if (src0.y < 0 && absY >= absX && absY >= absZ) { dst.x = src0.x; dst.y = -src0.z; } +if (src0.z >= 0 && absZ >= absX && absZ >= absY) { dst.x = src0.x; dst.y = -src0.y; } +if (src0.z < 0 && absZ >= absX && absZ >= absY) { dst.x = -src0.x; dst.y = -src0.y; } + +dst.x = dst.x / ma + 0.5; +dst.y = dst.y / ma + 0.5; """) unop_horiz("cube_face_index", 1, tfloat32, 3, tfloat32, """ @@ -423,20 +439,26 @@ if (src0.z >= 0 && absZ >= absX && absZ >= absY) dst.x = 4; if (src0.z < 0 && absZ >= absX && absZ >= absY) dst.x = 5; """) +# Sum of vector components +unop_reduce("fsum", 1, tfloat, tfloat, "{src}", "{src0} + {src1}", "{src}") 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): binop_convert(name, ty, ty, alg_props, const_expr) def binop_compare(name, ty, alg_props, const_expr): + binop_convert(name, tbool1, ty, alg_props, const_expr) + +def binop_compare32(name, ty, alg_props, const_expr): binop_convert(name, tbool32, ty, alg_props, const_expr) def binop_horiz(name, out_size, out_type, src1_size, src1_type, src2_size, src2_type, const_expr): 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, reduce_expr, final_expr): @@ -451,29 +473,91 @@ def binop_reduce(name, output_size, output_type, src_type, prereduce_expr, src2 = prereduce("src0.z", "src1.z") src3 = prereduce("src0.w", "src1.w") opcode(name + "2", output_size, output_type, - [2, 2], [src_type, src_type], commutative, + [2, 2], [src_type, src_type], False, _2src_commutative, final(reduce_(src0, src1))) opcode(name + "3", output_size, output_type, - [3, 3], [src_type, src_type], commutative, + [3, 3], [src_type, src_type], False, _2src_commutative, final(reduce_(reduce_(src0, src1), src2))) opcode(name + "4", output_size, output_type, - [4, 4], [src_type, src_type], commutative, + [4, 4], [src_type, src_type], False, _2src_commutative, final(reduce_(reduce_(src0, src1), reduce_(src2, src3)))) -binop("fadd", tfloat, commutative + associative, "src0 + src1") -binop("iadd", tint, commutative + associative, "src0 + src1") +binop("fadd", tfloat, _2src_commutative + associative, "src0 + src1") +binop("iadd", tint, _2src_commutative + associative, "src0 + src1") +binop("iadd_sat", tint, _2src_commutative, """ + src1 > 0 ? + (src0 + src1 < src0 ? (1ull << (bit_size - 1)) - 1 : src0 + src1) : + (src0 < src0 + src1 ? (1ull << (bit_size - 1)) : src0 + src1) +""") +binop("uadd_sat", tuint, _2src_commutative, + "(src0 + src1) < src0 ? MAX_UINT_FOR_SIZE(sizeof(src0) * 8) : (src0 + src1)") +binop("isub_sat", tint, "", """ + src1 < 0 ? + (src0 - src1 < src0 ? (1ull << (bit_size - 1)) - 1 : src0 - src1) : + (src0 < src0 - src1 ? (1ull << (bit_size - 1)) : src0 - src1) +""") +binop("usub_sat", tuint, "", "src0 < src1 ? 0 : src0 - src1") + binop("fsub", tfloat, "", "src0 - src1") binop("isub", tint, "", "src0 - src1") -binop("fmul", tfloat, commutative + associative, "src0 * src1") +binop("fmul", tfloat, _2src_commutative + associative, "src0 * src1") # low 32-bits of signed/unsigned integer multiply -binop("imul", tint, commutative + associative, "src0 * src1") +binop("imul", tint, _2src_commutative + associative, "src0 * src1") + +# Generate 64 bit result from 2 32 bits quantity +binop_convert("imul_2x32_64", tint64, tint32, _2src_commutative, + "(int64_t)src0 * (int64_t)src1") +binop_convert("umul_2x32_64", tuint64, tuint32, _2src_commutative, + "(uint64_t)src0 * (uint64_t)src1") + # high 32-bits of signed integer multiply -binop("imul_high", tint32, commutative, - "(int32_t)(((int64_t) src0 * (int64_t) src1) >> 32)") +binop("imul_high", tint, _2src_commutative, """ +if (bit_size == 64) { + /* We need to do a full 128-bit x 128-bit multiply in order for the sign + * extension to work properly. The casts are kind-of annoying but needed + * to prevent compiler warnings. + */ + uint32_t src0_u32[4] = { + src0, + (int64_t)src0 >> 32, + (int64_t)src0 >> 63, + (int64_t)src0 >> 63, + }; + uint32_t src1_u32[4] = { + src1, + (int64_t)src1 >> 32, + (int64_t)src1 >> 63, + (int64_t)src1 >> 63, + }; + uint32_t prod_u32[4]; + ubm_mul_u32arr(prod_u32, src0_u32, src1_u32); + dst = (uint64_t)prod_u32[2] | ((uint64_t)prod_u32[3] << 32); +} else { + dst = ((int64_t)src0 * (int64_t)src1) >> bit_size; +} +""") + # high 32-bits of unsigned integer multiply -binop("umul_high", tuint32, commutative, - "(uint32_t)(((uint64_t) src0 * (uint64_t) src1) >> 32)") +binop("umul_high", tuint, _2src_commutative, """ +if (bit_size == 64) { + /* The casts are kind-of annoying but needed to prevent compiler warnings. */ + uint32_t src0_u32[2] = { src0, (uint64_t)src0 >> 32 }; + uint32_t src1_u32[2] = { src1, (uint64_t)src1 >> 32 }; + uint32_t prod_u32[4]; + ubm_mul_u32arr(prod_u32, src0_u32, src1_u32); + dst = (uint64_t)prod_u32[2] | ((uint64_t)prod_u32[3] << 32); +} else { + dst = ((uint64_t)src0 * (uint64_t)src1) >> bit_size; +} +""") + +# low 32-bits of unsigned integer multiply +binop("umul_low", tuint32, _2src_commutative, """ +uint64_t mask = (1 << (bit_size / 2)) - 1; +dst = ((uint64_t)src0 & mask) * ((uint64_t)src1 & mask); +""") + binop("fdiv", tfloat, "", "src0 / src1") binop("idiv", tint, "", "src1 == 0 ? 0 : (src0 / src1)") @@ -482,13 +566,39 @@ binop("udiv", tuint, "", "src1 == 0 ? 0 : (src0 / src1)") # returns a boolean representing the carry resulting from the addition of # the two unsigned arguments. -binop_convert("uadd_carry", tuint, tuint, commutative, "src0 + src1 < src0") +binop_convert("uadd_carry", tuint, tuint, _2src_commutative, "src0 + src1 < src0") # returns a boolean representing the borrow resulting from the subtraction # of the two unsigned arguments. binop_convert("usub_borrow", tuint, tuint, "", "src0 < src1") +# hadd: (a + b) >> 1 (without overflow) +# x + y = x - (x & ~y) + (x & ~y) + y - (~x & y) + (~x & y) +# = (x & y) + (x & ~y) + (x & y) + (~x & y) +# = 2 * (x & y) + (x & ~y) + (~x & y) +# = ((x & y) << 1) + (x ^ y) +# +# Since we know that the bottom bit of (x & y) << 1 is zero, +# +# (x + y) >> 1 = (((x & y) << 1) + (x ^ y)) >> 1 +# = (x & y) + ((x ^ y) >> 1) +binop("ihadd", tint, _2src_commutative, "(src0 & src1) + ((src0 ^ src1) >> 1)") +binop("uhadd", tuint, _2src_commutative, "(src0 & src1) + ((src0 ^ src1) >> 1)") + +# rhadd: (a + b + 1) >> 1 (without overflow) +# x + y + 1 = x + (~x & y) - (~x & y) + y + (x & ~y) - (x & ~y) + 1 +# = (x | y) - (~x & y) + (x | y) - (x & ~y) + 1 +# = 2 * (x | y) - ((~x & y) + (x & ~y)) + 1 +# = ((x | y) << 1) - (x ^ y) + 1 +# +# Since we know that the bottom bit of (x & y) << 1 is zero, +# +# (x + y + 1) >> 1 = (x | y) + (-(x ^ y) + 1) >> 1) +# = (x | y) - ((x ^ y) >> 1) +binop("irhadd", tint, _2src_commutative, "(src0 | src1) + ((src0 ^ src1) >> 1)") +binop("urhadd", tuint, _2src_commutative, "(src0 | src1) + ((src0 ^ src1) >> 1)") + binop("umod", tuint, "", "src1 == 0 ? 0 : src0 % src1") # For signed integers, there are several different possible definitions of @@ -515,24 +625,43 @@ binop("frem", tfloat, "", "src0 - src1 * truncf(src0 / src1)") binop_compare("flt", tfloat, "", "src0 < src1") binop_compare("fge", tfloat, "", "src0 >= src1") -binop_compare("feq", tfloat, commutative, "src0 == src1") -binop_compare("fne", tfloat, commutative, "src0 != src1") +binop_compare("feq", tfloat, _2src_commutative, "src0 == src1") +binop_compare("fne", tfloat, _2src_commutative, "src0 != src1") binop_compare("ilt", tint, "", "src0 < src1") binop_compare("ige", tint, "", "src0 >= src1") -binop_compare("ieq", tint, commutative, "src0 == src1") -binop_compare("ine", tint, commutative, "src0 != src1") +binop_compare("ieq", tint, _2src_commutative, "src0 == src1") +binop_compare("ine", tint, _2src_commutative, "src0 != src1") binop_compare("ult", tuint, "", "src0 < src1") binop_compare("uge", tuint, "", "src0 >= src1") +binop_compare32("flt32", tfloat, "", "src0 < src1") +binop_compare32("fge32", tfloat, "", "src0 >= src1") +binop_compare32("feq32", tfloat, _2src_commutative, "src0 == src1") +binop_compare32("fne32", tfloat, _2src_commutative, "src0 != src1") +binop_compare32("ilt32", tint, "", "src0 < src1") +binop_compare32("ige32", tint, "", "src0 >= src1") +binop_compare32("ieq32", tint, _2src_commutative, "src0 == src1") +binop_compare32("ine32", tint, _2src_commutative, "src0 != src1") +binop_compare32("ult32", tuint, "", "src0 < src1") +binop_compare32("uge32", tuint, "", "src0 >= src1") # integer-aware GLSL-style comparisons that compare floats and ints -binop_reduce("ball_fequal", 1, tbool32, tfloat, "{src0} == {src1}", +binop_reduce("ball_fequal", 1, tbool1, tfloat, "{src0} == {src1}", "{src0} && {src1}", "{src}") -binop_reduce("bany_fnequal", 1, tbool32, tfloat, "{src0} != {src1}", +binop_reduce("bany_fnequal", 1, tbool1, tfloat, "{src0} != {src1}", "{src0} || {src1}", "{src}") -binop_reduce("ball_iequal", 1, tbool32, tint, "{src0} == {src1}", +binop_reduce("ball_iequal", 1, tbool1, tint, "{src0} == {src1}", "{src0} && {src1}", "{src}") -binop_reduce("bany_inequal", 1, tbool32, tint, "{src0} != {src1}", +binop_reduce("bany_inequal", 1, tbool1, tint, "{src0} != {src1}", + "{src0} || {src1}", "{src}") + +binop_reduce("b32all_fequal", 1, tbool32, tfloat, "{src0} == {src1}", + "{src0} && {src1}", "{src}") +binop_reduce("b32any_fnequal", 1, tbool32, tfloat, "{src0} != {src1}", + "{src0} || {src1}", "{src}") +binop_reduce("b32all_iequal", 1, tbool32, tint, "{src0} == {src1}", + "{src0} && {src1}", "{src}") +binop_reduce("b32any_inequal", 1, tbool32, tint, "{src0} != {src1}", "{src0} || {src1}", "{src}") # non-integer-aware GLSL-style comparisons that return 0.0 or 1.0 @@ -547,13 +676,29 @@ binop_reduce("fany_nequal", 1, tfloat32, tfloat32, "{src0} != {src1}", binop("slt", tfloat32, "", "(src0 < src1) ? 1.0f : 0.0f") # Set on Less Than binop("sge", tfloat, "", "(src0 >= src1) ? 1.0f : 0.0f") # Set on Greater or Equal -binop("seq", tfloat32, commutative, "(src0 == src1) ? 1.0f : 0.0f") # Set on Equal -binop("sne", tfloat32, commutative, "(src0 != src1) ? 1.0f : 0.0f") # Set on Not Equal - - -opcode("ishl", 0, tint, [0, 0], [tint, tuint32], "", "src0 << src1") -opcode("ishr", 0, tint, [0, 0], [tint, tuint32], "", "src0 >> src1") -opcode("ushr", 0, tuint, [0, 0], [tuint, tuint32], "", "src0 >> src1") +binop("seq", tfloat32, _2src_commutative, "(src0 == src1) ? 1.0f : 0.0f") # Set on Equal +binop("sne", tfloat32, _2src_commutative, "(src0 != src1) ? 1.0f : 0.0f") # Set on Not Equal + +# SPIRV shifts are undefined for shift-operands >= bitsize, +# but SM5 shifts are defined to use the least significant bits, only +# The NIR definition is according to the SM5 specification. +opcode("ishl", 0, tint, [0, 0], [tint, tuint32], False, "", + "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))") + +opcode("urol", 0, tuint, [0, 0], [tuint, tuint32], False, "", """ + uint32_t rotate_mask = sizeof(src0) * 8 - 1; + dst = (src0 << (src1 & rotate_mask)) | + (src0 >> (-src1 & rotate_mask)); +""") +opcode("uror", 0, tuint, [0, 0], [tuint, tuint32], False, "", """ + uint32_t rotate_mask = sizeof(src0) * 8 - 1; + dst = (src0 >> (src1 & rotate_mask)) | + (src0 << (-src1 & rotate_mask)); +""") # bitwise logic operators # @@ -561,22 +706,10 @@ opcode("ushr", 0, tuint, [0, 0], [tuint, tuint32], "", "src0 >> src1") # integers. -binop("iand", tuint, commutative + associative, "src0 & src1") -binop("ior", tuint, commutative + associative, "src0 | src1") -binop("ixor", tuint, commutative + associative, "src0 ^ src1") - +binop("iand", tuint, _2src_commutative + associative, "src0 & src1") +binop("ior", tuint, _2src_commutative + associative, "src0 | src1") +binop("ixor", tuint, _2src_commutative + associative, "src0 ^ src1") -# floating point logic operators -# -# These use (src != 0.0) for testing the truth of the input, and output 1.0 -# for true and 0.0 for false - -binop("fand", tfloat32, commutative, - "((src0 != 0.0f) && (src1 != 0.0f)) ? 1.0f : 0.0f") -binop("for", tfloat32, commutative, - "((src0 != 0.0f) || (src1 != 0.0f)) ? 1.0f : 0.0f") -binop("fxor", tfloat32, commutative, - "(src0 != 0.0f && src1 == 0.0f) || (src0 == 0.0f && src1 != 0.0f) ? 1.0f : 0.0f") binop_reduce("fdot", 1, tfloat, tfloat, "{src0} * {src1}", "{src0} + {src1}", "{src}") @@ -584,20 +717,20 @@ binop_reduce("fdot", 1, tfloat, tfloat, "{src0} * {src1}", "{src0} + {src1}", binop_reduce("fdot_replicated", 4, tfloat, tfloat, "{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") -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") binop("fmin", tfloat, "", "fminf(src0, src1)") -binop("imin", tint, commutative + associative, "src1 > src0 ? src0 : src1") -binop("umin", tuint, commutative + associative, "src1 > src0 ? src0 : src1") +binop("imin", tint, _2src_commutative + associative, "src1 > src0 ? src0 : src1") +binop("umin", tuint, _2src_commutative + associative, "src1 > src0 ? src0 : src1") binop("fmax", tfloat, "", "fmaxf(src0, src1)") -binop("imax", tint, commutative + associative, "src1 > src0 ? src1 : src0") -binop("umax", tuint, commutative + associative, "src1 > src0 ? src1 : src0") +binop("imax", tint, _2src_commutative + associative, "src1 > src0 ? src1 : src0") +binop("umax", tuint, _2src_commutative + associative, "src1 > src0 ? src1 : src0") # Saturated vector add for 4 8bit ints. -binop("usadd_4x8", tint32, commutative + associative, """ +binop("usadd_4x8", tint32, _2src_commutative + associative, """ dst = 0; for (int i = 0; i < 32; i += 8) { dst |= MIN2(((src0 >> i) & 0xff) + ((src1 >> i) & 0xff), 0xff) << i; @@ -616,7 +749,7 @@ for (int i = 0; i < 32; i += 8) { """) # vector min for 4 8bit ints. -binop("umin_4x8", tint32, commutative + associative, """ +binop("umin_4x8", tint32, _2src_commutative + associative, """ dst = 0; for (int i = 0; i < 32; i += 8) { dst |= MIN2((src0 >> i) & 0xff, (src1 >> i) & 0xff) << i; @@ -624,7 +757,7 @@ for (int i = 0; i < 32; i += 8) { """) # vector max for 4 8bit ints. -binop("umax_4x8", tint32, commutative + associative, """ +binop("umax_4x8", tint32, _2src_commutative + associative, """ dst = 0; for (int i = 0; i < 32; i += 8) { dst |= MAX2((src0 >> i) & 0xff, (src1 >> i) & 0xff) << i; @@ -632,7 +765,7 @@ for (int i = 0; i < 32; i += 8) { """) # unorm multiply: (a * b) / 255. -binop("umul_unorm_4x8", tint32, commutative + associative, """ +binop("umul_unorm_4x8", tint32, _2src_commutative + associative, """ dst = 0; for (int i = 0; i < 32; i += 8) { int src0_chan = (src0 >> i) & 0xff; @@ -653,17 +786,15 @@ binop_convert("pack_32_2x16_split", tuint32, tuint16, "", "src0 | ((uint32_t)src1 << 16)") # bfm implements the behavior of the first operation of the SM5 "bfi" assembly -# and that of the "bfi1" i965 instruction. That is, it has undefined behavior -# if either of its arguments are 32. +# and that of the "bfi1" i965 instruction. That is, the bits and offset values +# are from the low five bits of src0 and src1, respectively. binop_convert("bfm", tuint32, tint32, "", """ -int bits = src0, offset = src1; -if (offset < 0 || bits < 0 || offset > 31 || bits > 31 || offset + bits > 32) - dst = 0; /* undefined */ -else - dst = ((1u << bits) - 1) << offset; +int bits = src0 & 0x1F; +int offset = src1 & 0x1F; +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); /* flush denormals to zero. */ if (!isnormal(dst)) @@ -686,16 +817,16 @@ binop("extract_u16", tuint, "", "(uint16_t)(src0 >> (src1 * 16))") binop("extract_i16", tint, "", "(int16_t)(src0 >> (src1 * 16))") -def triop(name, ty, const_expr): - opcode(name, 0, ty, [0, 0, 0], [ty, ty, ty], "", const_expr) +def triop(name, ty, alg_props, const_expr): + opcode(name, 0, ty, [0, 0, 0], [ty, ty, ty], False, alg_props, const_expr) def triop_horiz(name, output_size, src1_size, src2_size, src3_size, const_expr): opcode(name, output_size, tuint, [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, _2src_commutative, "src0 * src1 + src2") -triop("flrp", tfloat, "src0 * (1 - src2) + src1 * src2") +triop("flrp", tfloat, "", "src0 * (1 - src2) + src1 * src2") # Conditional Select # @@ -704,26 +835,28 @@ triop("flrp", tfloat, "src0 * (1 - src2) + src1 * src2") # bools (0.0 vs 1.0) and one for integer bools (0 vs ~0). -triop("fcsel", tfloat32, "(src0 != 0.0f) ? src1 : src2") +triop("fcsel", tfloat32, "", "(src0 != 0.0f) ? src1 : src2") # 3 way min/max/med -triop("fmin3", tfloat, "fminf(src0, fminf(src1, src2))") -triop("imin3", tint, "MIN2(src0, MIN2(src1, src2))") -triop("umin3", tuint, "MIN2(src0, MIN2(src1, src2))") +triop("fmin3", tfloat, "", "fminf(src0, fminf(src1, src2))") +triop("imin3", tint, "", "MIN2(src0, MIN2(src1, src2))") +triop("umin3", tuint, "", "MIN2(src0, MIN2(src1, src2))") -triop("fmax3", tfloat, "fmaxf(src0, fmaxf(src1, src2))") -triop("imax3", tint, "MAX2(src0, MAX2(src1, src2))") -triop("umax3", tuint, "MAX2(src0, MAX2(src1, src2))") +triop("fmax3", tfloat, "", "fmaxf(src0, fmaxf(src1, src2))") +triop("imax3", tint, "", "MAX2(src0, MAX2(src1, src2))") +triop("umax3", tuint, "", "MAX2(src0, MAX2(src1, src2))") -triop("fmed3", tfloat, "fmaxf(fminf(fmaxf(src0, src1), src2), fminf(src0, src1))") -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("fmed3", tfloat, "", "fmaxf(fminf(fmaxf(src0, src1), src2), fminf(src0, src1))") +triop("imed3", tint, "", "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], - [tbool32, tuint, tuint], "", "src0 ? src1 : src2") + [tbool1, tuint, tuint], False, "", "src0 ? src1 : src2") +opcode("b32csel", 0, tuint, [0, 0, 0], + [tbool32, tuint, tuint], False, "", "src0 ? src1 : src2") # SM5 bfi assembly -triop("bfi", tuint32, """ +triop("bfi", tuint32, "", """ unsigned mask = src0, insert = src1, base = src2; if (mask == 0) { dst = base; @@ -737,15 +870,17 @@ if (mask == 0) { } """) -# SM5 ubfe/ibfe assembly + +triop("bitfield_select", tuint, "", "(src0 & src1) | (~src0 & src2)") + +# SM5 ubfe/ibfe assembly: only the 5 least significant bits of offset and bits are used. opcode("ubfe", 0, tuint32, - [0, 0, 0], [tuint32, tint32, tint32], "", """ + [0, 0, 0], [tuint32, tuint32, tuint32], False, "", """ unsigned base = src0; -int offset = src1, bits = src2; +unsigned offset = src1 & 0x1F; +unsigned bits = src2 & 0x1F; if (bits == 0) { dst = 0; -} else if (bits < 0 || offset < 0) { - dst = 0; /* undefined */ } else if (offset + bits < 32) { dst = (base << (32 - bits - offset)) >> (32 - bits); } else { @@ -753,13 +888,12 @@ if (bits == 0) { } """) opcode("ibfe", 0, tint32, - [0, 0, 0], [tint32, tint32, tint32], "", """ + [0, 0, 0], [tint32, tuint32, tuint32], False, "", """ int base = src0; -int offset = src1, bits = src2; +unsigned offset = src1 & 0x1F; +unsigned bits = src2 & 0x1F; if (bits == 0) { dst = 0; -} else if (bits < 0 || offset < 0) { - dst = 0; /* undefined */ } else if (offset + bits < 32) { dst = (base << (32 - bits - offset)) >> (32 - bits); } else { @@ -769,7 +903,7 @@ if (bits == 0) { # GLSL bitfieldExtract() opcode("ubitfield_extract", 0, tuint32, - [0, 0, 0], [tuint32, tint32, tint32], "", """ + [0, 0, 0], [tuint32, tint32, tint32], False, "", """ unsigned base = src0; int offset = src1, bits = src2; if (bits == 0) { @@ -781,7 +915,7 @@ if (bits == 0) { } """) opcode("ibitfield_extract", 0, tint32, - [0, 0, 0], [tint32, tint32, tint32], "", """ + [0, 0, 0], [tint32, tint32, tint32], False, "", """ int base = src0; int offset = src1, bits = src2; if (bits == 0) { @@ -806,10 +940,10 @@ def quadop_horiz(name, output_size, src1_size, src2_size, src3_size, opcode(name, output_size, tuint, [src1_size, src2_size, src3_size, src4_size], [tuint, tuint, tuint, tuint], - "", const_expr) + False, "", const_expr) opcode("bitfield_insert", 0, tuint32, [0, 0, 0, 0], - [tuint32, tuint32, tint32, tint32], "", """ + [tuint32, tuint32, tint32, tint32], False, "", """ unsigned base = src0, insert = src1; int offset = src2, bits = src3; if (bits == 0) { @@ -829,4 +963,10 @@ dst.z = src2.x; dst.w = src3.x; """) - +# ir3-specific instruction that maps directly to mul-add shift high mix, +# (IMADSH_MIX16 i.e. ah * bl << 16 + c). It is used for lowering integer +# multiplication (imul) on Freedreno backend.. +opcode("imadsh_mix16", 1, tint32, + [1, 1, 1], [tint32, tint32, tint32], False, "", """ +dst.x = ((((src0.x & 0xffff0000) >> 16) * (src1.x & 0x0000ffff)) << 16) + src2.x; +""")