assert isinstance(algebraic_properties, str)
assert isinstance(const_expr, str)
assert len(input_sizes) == len(input_types)
- assert 0 <= output_size <= 4
+ assert 0 <= output_size <= 4 or (output_size == 8) or (output_size == 16)
for size in input_sizes:
- assert 0 <= size <= 4
+ assert 0 <= size <= 4 or (size == 8) or (size == 16)
if output_size != 0:
assert size != 0
self.name = name
tint = "int"
tbool = "bool"
tbool1 = "bool1"
+tbool8 = "bool8"
+tbool16 = "bool16"
tbool32 = "bool32"
tuint = "uint"
+tuint8 = "uint8"
+tint16 = "int16"
tuint16 = "uint16"
+tfloat16 = "float16"
tfloat32 = "float32"
tint32 = "int32"
tuint32 = "uint32"
if type_has_size(type_):
return [type_size(type_)]
elif type_ == 'bool':
- return [1, 32]
+ return [1, 8, 16, 32]
elif type_ == 'float':
return [16, 32, 64]
else:
unop("isign", tint, "(src0 == 0) ? 0 : ((src0 > 0) ? 1 : -1)")
unop("iabs", tint, "(src0 < 0) ? -src0 : src0")
unop("fabs", tfloat, "fabs(src0)")
-unop("fsat", tfloat, ("bit_size == 64 ? " +
- "((src0 > 1.0) ? 1.0 : ((src0 <= 0.0) ? 0.0 : src0)) : " +
- "((src0 > 1.0f) ? 1.0f : ((src0 <= 0.0f) ? 0.0f : src0))"))
+unop("fsat", tfloat, ("fmin(fmax(src0, 0.0), 1.0)"))
+unop("fsat_signed", tfloat, ("fmin(fmax(src0, -1.0), 1.0)"))
+unop("fclamp_pos", tfloat, ("fmax(src0, 0.0)"))
unop("frcp", tfloat, "bit_size == 64 ? 1.0 / src0 : 1.0f / src0")
unop("frsq", tfloat, "bit_size == 64 ? 1.0 / sqrt(src0) : 1.0f / sqrtf(src0)")
unop("fsqrt", tfloat, "bit_size == 64 ? sqrt(src0) : sqrtf(src0)")
# Generate all of the numeric conversion opcodes
for src_t in [tint, tuint, tfloat, tbool]:
if src_t == tbool:
- dst_types = [tfloat, tint]
+ dst_types = [tfloat, tint, tbool]
elif src_t == tint:
dst_types = [tfloat, tint, tbool]
elif src_t == tuint:
dst_bit_size),
dst_t + str(dst_bit_size), src_t, conv_expr)
+# Special opcode that is the same as f2f16, i2i16, u2u16 except that it is safe
+# to remove it if the result is immediately converted back to 32 bits again.
+# This is generated as part of the precision lowering pass. mp stands for medium
+# precision.
+unop_numeric_convert("f2fmp", tfloat16, tfloat, opcodes["f2f16"].const_expr)
+unop_numeric_convert("i2imp", tint16, tint, opcodes["i2i16"].const_expr)
+unop_numeric_convert("u2ump", tuint16, tuint, opcodes["u2u16"].const_expr)
# Unary floating-point rounding operations.
(src0.w << 24);
""")
+unop_horiz("pack_32_4x8", 1, tuint32, 4, tuint8,
+ "dst.x = src0.x | ((uint32_t)src0.y << 8) | ((uint32_t)src0.z << 16) | ((uint32_t)src0.w << 24);")
+
unop_horiz("pack_32_2x16", 1, tuint32, 2, tuint16,
"dst.x = src0.x | ((uint32_t)src0.y << 16);")
unop_horiz("unpack_32_2x16", 2, tuint16, 1, tuint32,
"dst.x = src0.x; dst.y = src0.x >> 16;")
+unop_horiz("unpack_32_4x8", 4, tuint8, 1, tuint32,
+ "dst.x = src0.x; dst.y = src0.x >> 8; dst.z = src0.x >> 16; dst.w = src0.x >> 24;")
+
unop_horiz("unpack_half_2x16_flush_to_zero", 2, tfloat32, 1, tuint32, """
dst.x = unpack_half_1x16_flush_to_zero((uint16_t)(src0.x & 0xffff));
dst.y = unpack_half_1x16_flush_to_zero((uint16_t)(src0.x << 16));
}
""")
+unop("uclz", tuint32, """
+int bit;
+for (bit = bit_size - 1; bit >= 0; bit--) {
+ if ((src0 & (1u << bit)) != 0)
+ break;
+}
+dst = (unsigned)(31 - bit);
+""")
+
unop("ifind_msb", tint32, """
dst = -1;
for (int bit = 31; bit >= 0; bit--) {
}
""")
-
-for i in range(1, 5):
- for j in range(1, 5):
- unop_horiz("fnoise{0}_{1}".format(i, j), i, tfloat, j, tfloat, "0.0f")
-
-
# AMD_gcn_shader extended instructions
unop_horiz("cube_face_coord", 2, tfloat32, 3, tfloat32, """
dst.x = dst.y = 0.0;
-float absX = fabs(src0.x);
-float absY = fabs(src0.y);
-float absZ = fabs(src0.z);
+float absX = fabsf(src0.x);
+float absY = fabsf(src0.y);
+float absZ = fabsf(src0.z);
float ma = 0.0;
if (absX >= absY && absX >= absZ) { ma = 2 * src0.x; }
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;
+dst.x = dst.x * (1.0f / ma) + 0.5f;
+dst.y = dst.y * (1.0f / ma) + 0.5f;
""")
unop_horiz("cube_face_index", 1, tfloat32, 3, tfloat32, """
-float absX = fabs(src0.x);
-float absY = fabs(src0.y);
-float absZ = fabs(src0.z);
+float absX = fabsf(src0.x);
+float absY = fabsf(src0.y);
+float absZ = fabsf(src0.z);
if (src0.x >= 0 && absX >= absY && absX >= absZ) dst.x = 0;
if (src0.x < 0 && absX >= absY && absX >= absZ) dst.x = 1;
if (src0.y >= 0 && absY >= absX && absY >= absZ) dst.x = 2;
def binop_compare(name, ty, alg_props, const_expr):
binop_convert(name, tbool1, ty, alg_props, const_expr)
+def binop_compare8(name, ty, alg_props, const_expr):
+ binop_convert(name, tbool8, ty, alg_props, const_expr)
+
+def binop_compare16(name, ty, alg_props, const_expr):
+ binop_convert(name, tbool16, 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_compare_all_sizes(name, ty, alg_props, const_expr):
binop_compare(name, ty, alg_props, const_expr)
+ binop_compare8(name + "8", ty, alg_props, const_expr)
+ binop_compare16(name + "16", ty, alg_props, const_expr)
binop_compare32(name + "32", ty, alg_props, const_expr)
def binop_horiz(name, out_size, out_type, src1_size, src1_type, src2_size,
return reduce_expr.format(src0=src0, src1=src1)
def prereduce(src0, src1):
return "(" + prereduce_expr.format(src0=src0, src1=src1) + ")"
- src0 = prereduce("src0.x", "src1.x")
- src1 = prereduce("src0.y", "src1.y")
- 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], False, _2src_commutative,
- final(reduce_(src0, src1)))
+ srcs = [prereduce("src0." + letter, "src1." + letter) for letter in "xyzwefghijklmnop"]
+ def pairwise_reduce(start, size):
+ if (size == 1):
+ return srcs[start]
+ return reduce_(pairwise_reduce(start, size // 2), pairwise_reduce(start + size // 2, size // 2))
+ for size in [2, 4, 8, 16]:
+ opcode(name + str(size), output_size, output_type,
+ [size, size], [src_type, src_type], False, _2src_commutative,
+ final(pairwise_reduce(0, size)))
opcode(name + "3", output_size, output_type,
[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], False, _2src_commutative,
- final(reduce_(reduce_(src0, src1), reduce_(src2, src3))))
+ final(reduce_(reduce_(srcs[0], srcs[1]), srcs[2])))
+
+def binop_reduce_all_sizes(name, output_size, src_type, prereduce_expr,
+ reduce_expr, final_expr):
+ binop_reduce(name, output_size, tbool1, src_type,
+ prereduce_expr, reduce_expr, final_expr)
+ binop_reduce("b8" + name[1:], output_size, tbool8, src_type,
+ prereduce_expr, reduce_expr, final_expr)
+ binop_reduce("b16" + name[1:], output_size, tbool16, src_type,
+ prereduce_expr, reduce_expr, final_expr)
+ binop_reduce("b32" + name[1:], output_size, tbool32, src_type,
+ prereduce_expr, reduce_expr, final_expr)
binop("fadd", tfloat, _2src_commutative + associative,"""
if (nir_is_rounding_mode_rtz(execution_mode, bit_size)) {
}
""")
binop("isub", tint, "", "src0 - src1")
+binop_convert("uabs_isub", tuint, tint, "", """
+ src1 > src0 ? (uint64_t) src1 - (uint64_t) src0
+ : (uint64_t) src0 - (uint64_t) src1
+""")
+binop("uabs_usub", tuint, "", "(src1 > src0) ? (src1 - src0) : (src0 - src1)")
binop("fmul", tfloat, _2src_commutative + associative, """
if (nir_is_rounding_mode_rtz(execution_mode, bit_size)) {
dst = ((uint64_t)src0 & mask) * ((uint64_t)src1 & mask);
""")
+# Multiply 32-bits with low 16-bits.
+binop("imul_32x16", tint32, "", "src0 * (int16_t) src1")
+binop("umul_32x16", tuint32, "", "src0 * (uint16_t) src1")
binop("fdiv", tfloat, "", "src0 / src1")
binop("idiv", tint, "", "src1 == 0 ? 0 : (src0 / src1)")
binop_compare_all_sizes("flt", tfloat, "", "src0 < src1")
binop_compare_all_sizes("fge", tfloat, "", "src0 >= src1")
binop_compare_all_sizes("feq", tfloat, _2src_commutative, "src0 == src1")
-binop_compare_all_sizes("fne", tfloat, _2src_commutative, "src0 != src1")
+binop_compare_all_sizes("fneu", tfloat, _2src_commutative, "src0 != src1")
binop_compare_all_sizes("ilt", tint, "", "src0 < src1")
binop_compare_all_sizes("ige", tint, "", "src0 >= src1")
binop_compare_all_sizes("ieq", tint, _2src_commutative, "src0 == src1")
# integer-aware GLSL-style comparisons that compare floats and ints
-binop_reduce("ball_fequal", 1, tbool1, tfloat, "{src0} == {src1}",
- "{src0} && {src1}", "{src}")
-binop_reduce("bany_fnequal", 1, tbool1, tfloat, "{src0} != {src1}",
- "{src0} || {src1}", "{src}")
-binop_reduce("ball_iequal", 1, tbool1, tint, "{src0} == {src1}",
- "{src0} && {src1}", "{src}")
-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}")
+binop_reduce_all_sizes("ball_fequal", 1, tfloat, "{src0} == {src1}",
+ "{src0} && {src1}", "{src}")
+binop_reduce_all_sizes("bany_fnequal", 1, tfloat, "{src0} != {src1}",
+ "{src0} || {src1}", "{src}")
+binop_reduce_all_sizes("ball_iequal", 1, tint, "{src0} == {src1}",
+ "{src0} && {src1}", "{src}")
+binop_reduce_all_sizes("bany_inequal", 1, tint, "{src0} != {src1}",
+ "{src0} || {src1}", "{src}")
# non-integer-aware GLSL-style comparisons that return 0.0 or 1.0
# 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))")
+ "(uint64_t)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, "",
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, "", "fmin(src0, src1)")
+binop("fmin", tfloat, _2src_commutative + associative, "fmin(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, "", "fmax(src0, src1)")
+binop("fmax", tfloat, _2src_commutative + associative, "fmax(src0, src1)")
binop("imax", tint, _2src_commutative + associative, "src1 > src0 ? src1 : src0")
binop("umax", tuint, _2src_commutative + associative, "src1 > src0 ? src1 : src0")
# component on vectors). There are two versions, one for floating point
# bools (0.0 vs 1.0) and one for integer bools (0 vs ~0).
-
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("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))")
-
opcode("bcsel", 0, tuint, [0, 0, 0],
- [tbool1, tuint, tuint], False, "", "src0 ? src1 : src2")
+ [tbool1, tuint, tuint], False, "", "src0 ? src1 : src2")
+opcode("b8csel", 0, tuint, [0, 0, 0],
+ [tbool8, tuint, tuint], False, "", "src0 ? src1 : src2")
+opcode("b16csel", 0, tuint, [0, 0, 0],
+ [tbool16, tuint, tuint], False, "", "src0 ? src1 : src2")
opcode("b32csel", 0, tuint, [0, 0, 0],
[tbool32, tuint, tuint], False, "", "src0 ? src1 : src2")
dst.w = src3.x;
""")
+opcode("vec8", 8, tuint,
+ [1] * 8, [tuint] * 8,
+ False, "", """
+dst.x = src0.x;
+dst.y = src1.x;
+dst.z = src2.x;
+dst.w = src3.x;
+dst.e = src4.x;
+dst.f = src5.x;
+dst.g = src6.x;
+dst.h = src7.x;
+""")
+
+opcode("vec16", 16, tuint,
+ [1] * 16, [tuint] * 16,
+ False, "", """
+dst.x = src0.x;
+dst.y = src1.x;
+dst.z = src2.x;
+dst.w = src3.x;
+dst.e = src4.x;
+dst.f = src5.x;
+dst.g = src6.x;
+dst.h = src7.x;
+dst.i = src8.x;
+dst.j = src9.x;
+dst.k = src10.x;
+dst.l = src11.x;
+dst.m = src12.x;
+dst.n = src13.x;
+dst.o = src14.x;
+dst.p = src15.x;
+""")
+
# An integer multiply instruction for address calculation. This is
# similar to imul, except that the results are undefined in case of
# overflow. Overflow is defined according to the size of the variable
# 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;
+opcode("imadsh_mix16", 0, tint32,
+ [0, 0, 0], [tint32, tint32, tint32], False, "", """
+dst = ((((src0 & 0xffff0000) >> 16) * (src1 & 0x0000ffff)) << 16) + src2;
""")
# ir3-specific instruction that maps directly to ir3 mad.s24.
# 24b multiply into 32b result (with sign extension)
binop("imul24", tint32, _2src_commutative + associative,
"(((int32_t)src0 << 8) >> 8) * (((int32_t)src1 << 8) >> 8)")
+
+# unsigned 24b multiply into 32b result plus 32b int
+triop("umad24", tuint32, _2src_commutative,
+ "(((uint32_t)src0 << 8) >> 8) * (((uint32_t)src1 << 8) >> 8) + src2")
+
+# unsigned 24b multiply into 32b result uint
+binop("umul24", tint32, _2src_commutative + associative,
+ "(((uint32_t)src0 << 8) >> 8) * (((uint32_t)src1 << 8) >> 8)")
+
+unop_convert("fisnormal", tbool1, tfloat, "isnormal(src0)")
+unop_convert("fisfinite", tbool1, tfloat, "isfinite(src0)")