real_types = (float_type, double_type)
# This template is for unary and binary operations that can only have operands
-# of a single type. ir_unop_logic_not is an example.
+# of a single type or the implementation for all types is identical.
+# ir_unop_logic_not is an example of the former, and ir_quadop_bitfield_insert
+# is an example of the latter..
constant_template0 = mako.template.Template("""\
case ${op.get_enum_name()}:
+ % if len(op.source_types) == 1:
assert(op[0]->type->base_type == ${op.source_types[0].glsl_type});
+ % endif
for (unsigned c = 0; c < op[0]->type->components(); c++)
- data.${op.source_types[0].union_field}[c] = ${op.get_c_expression(op.source_types)};
+ % for dst_type, src_types in op.signatures():
+ % if loop.index == 0:
+ data.${dst_type.union_field}[c] = ${op.get_c_expression(src_types)};
+ % endif
+ % endfor
break;""")
# This template is for unary operations that can have operands of a several
# of scalar and vector operands.
constant_template_vector_scalar = mako.template.Template("""\
case ${op.get_enum_name()}:
+ % if "mixed" in op.flags:
+ % for i in xrange(op.num_operands):
+ assert(op[${i}]->type->base_type == ${op.source_types[0].glsl_type} ||
+ % for src_type in op.source_types[1:-1]:
+ op[${i}]->type->base_type == ${src_type.glsl_type} ||
+ % endfor
+ op[${i}]->type->base_type == ${op.source_types[-1].glsl_type});
+ % endfor
+ % else:
assert(op[0]->type == op[1]->type || op0_scalar || op1_scalar);
+ % endif
for (unsigned c = 0, c0 = 0, c1 = 0;
c < components;
c0 += c0_inc, c1 += c1_inc, c++) {
switch (op[0]->type->base_type) {
% for dst_type, src_types in op.signatures():
case ${src_types[0].glsl_type}:
- data.${dst_type.union_field}[c] = ${op.get_c_expression(src_types, ("c0", "c1"))};
+ data.${dst_type.union_field}[c] = ${op.get_c_expression(src_types, ("c0", "c1", "c2"))};
break;
% endfor
default:
}
break;""")
+# This template is for multiplication. It is unique because it has to support
+# matrix * vector and matrix * matrix operations, and those are just different.
+constant_template_mul = mako.template.Template("""\
+ case ${op.get_enum_name()}:
+ /* Check for equal types, or unequal types involving scalars */
+ if ((op[0]->type == op[1]->type && !op[0]->type->is_matrix())
+ || op0_scalar || op1_scalar) {
+ for (unsigned c = 0, c0 = 0, c1 = 0;
+ c < components;
+ c0 += c0_inc, c1 += c1_inc, c++) {
+
+ switch (op[0]->type->base_type) {
+ % for dst_type, src_types in op.signatures():
+ case ${src_types[0].glsl_type}:
+ data.${dst_type.union_field}[c] = ${op.get_c_expression(src_types, ("c0", "c1", "c2"))};
+ break;
+ % endfor
+ default:
+ assert(0);
+ }
+ }
+ } else {
+ assert(op[0]->type->is_matrix() || op[1]->type->is_matrix());
+
+ /* Multiply an N-by-M matrix with an M-by-P matrix. Since either
+ * matrix can be a GLSL vector, either N or P can be 1.
+ *
+ * For vec*mat, the vector is treated as a row vector. This
+ * means the vector is a 1-row x M-column matrix.
+ *
+ * For mat*vec, the vector is treated as a column vector. Since
+ * matrix_columns is 1 for vectors, this just works.
+ */
+ const unsigned n = op[0]->type->is_vector()
+ ? 1 : op[0]->type->vector_elements;
+ const unsigned m = op[1]->type->vector_elements;
+ const unsigned p = op[1]->type->matrix_columns;
+ for (unsigned j = 0; j < p; j++) {
+ for (unsigned i = 0; i < n; i++) {
+ for (unsigned k = 0; k < m; k++) {
+ if (op[0]->type->base_type == GLSL_TYPE_DOUBLE)
+ data.d[i+n*j] += op[0]->value.d[i+n*k]*op[1]->value.d[k+m*j];
+ else
+ data.f[i+n*j] += op[0]->value.f[i+n*k]*op[1]->value.f[k+m*j];
+ }
+ }
+ }
+ }
+ break;""")
+
# This template is for operations that are horizontal and either have only a
# single type or the implementation for all types is identical. That is, the
# operation consumes a vector and produces a scalar.
data.${op.dest_type.union_field}[0] = ${op.c_expression['default']};
break;""")
+# This template is for operations that are horizontal and do not assign the
+# result. The various unpack operations are examples.
+constant_template_horizontal_nonassignment = mako.template.Template("""\
+ case ${op.get_enum_name()}:
+ ${op.c_expression['default']};
+ break;""")
+
+# This template is for binary operations that are horizontal. That is, the
+# operation consumes a vector and produces a scalar.
+constant_template_horizontal = mako.template.Template("""\
+ case ${op.get_enum_name()}:
+ switch (op[0]->type->base_type) {
+ % for dst_type, src_types in op.signatures():
+ case ${src_types[0].glsl_type}:
+ data.${dst_type.union_field}[0] = ${op.get_c_expression(src_types)};
+ break;
+ % endfor
+ default:
+ assert(0);
+ }
+ break;""")
+
+# This template is for ir_binop_vector_extract.
+constant_template_vector_extract = mako.template.Template("""\
+ case ${op.get_enum_name()}: {
+ const int c = CLAMP(op[1]->value.i[0], 0,
+ (int) op[0]->type->vector_elements - 1);
+
+ switch (op[0]->type->base_type) {
+ % for dst_type, src_types in op.signatures():
+ case ${src_types[0].glsl_type}:
+ data.${dst_type.union_field}[0] = op[0]->value.${src_types[0].union_field}[c];
+ break;
+ % endfor
+ default:
+ assert(0);
+ }
+ break;
+ }""")
+
+# This template is for ir_triop_vector_insert.
+constant_template_vector_insert = mako.template.Template("""\
+ case ${op.get_enum_name()}: {
+ const unsigned idx = op[2]->value.u[0];
+
+ memcpy(&data, &op[0]->value, sizeof(data));
+
+ switch (this->type->base_type) {
+ % for dst_type, src_types in op.signatures():
+ case ${src_types[0].glsl_type}:
+ data.${dst_type.union_field}[idx] = op[1]->value.${src_types[0].union_field}[0];
+ break;
+ % endfor
+ default:
+ assert(!"Should not get here.");
+ break;
+ }
+ break;
+ }""")
+
+# This template is for ir_quadop_vector.
+constant_template_vector = mako.template.Template("""\
+ case ${op.get_enum_name()}:
+ for (unsigned c = 0; c < this->type->vector_elements; c++) {
+ switch (this->type->base_type) {
+ % for dst_type, src_types in op.signatures():
+ case ${src_types[0].glsl_type}:
+ data.${dst_type.union_field}[c] = op[c]->value.${src_types[0].union_field}[0];
+ break;
+ % endfor
+ default:
+ assert(0);
+ }
+ }
+ break;""")
+
vector_scalar_operation = "vector-scalar"
horizontal_operation = "horizontal"
types_identical_operation = "identical"
+non_assign_operation = "nonassign"
+mixed_type_operation = "mixed"
class operation(object):
- def __init__(self, name, num_operands, printable_name = None, source_types = None, dest_type = None, c_expression = None, flags = None):
+ def __init__(self, name, num_operands, printable_name = None, source_types = None, dest_type = None, c_expression = None, flags = None, all_signatures = None):
self.name = name
self.num_operands = num_operands
else:
self.printable_name = printable_name
- self.source_types = source_types
+ self.all_signatures = all_signatures
+
+ if source_types is None:
+ self.source_types = tuple()
+ else:
+ self.source_types = source_types
+
self.dest_type = dest_type
if c_expression is None:
return None
if self.num_operands == 1:
- if horizontal_operation in self.flags:
+ if horizontal_operation in self.flags and non_assign_operation in self.flags:
+ return constant_template_horizontal_nonassignment.render(op=self)
+ elif horizontal_operation in self.flags:
return constant_template_horizontal_single_implementation.render(op=self)
elif self.dest_type is not None and len(self.source_types) == 1:
return constant_template2.render(op=self)
else:
return constant_template3.render(op=self)
elif self.num_operands == 2:
- if vector_scalar_operation in self.flags:
+ if self.name == "mul":
+ return constant_template_mul.render(op=self)
+ elif self.name == "vector_extract":
+ return constant_template_vector_extract.render(op=self)
+ elif vector_scalar_operation in self.flags:
return constant_template_vector_scalar.render(op=self)
elif horizontal_operation in self.flags and types_identical_operation in self.flags:
return constant_template_horizontal_single_implementation.render(op=self)
+ elif horizontal_operation in self.flags:
+ return constant_template_horizontal.render(op=self)
elif len(self.source_types) == 1:
return constant_template0.render(op=self)
elif self.dest_type is not None:
return constant_template5.render(op=self)
+ else:
+ return constant_template3.render(op=self)
+ elif self.num_operands == 3:
+ if self.name == "vector_insert":
+ return constant_template_vector_insert.render(op=self)
+ else:
+ return constant_template3.render(op=self)
+ elif self.num_operands == 4:
+ if self.name == "vector":
+ return constant_template_vector.render(op=self)
+ elif types_identical_operation in self.flags:
+ return constant_template0.render(op=self)
return None
- def get_c_expression(self, types, indices=("c", "c")):
+ def get_c_expression(self, types, indices=("c", "c", "c")):
src0 = "op[0]->value.{}[{}]".format(types[0].union_field, indices[0])
src1 = "op[1]->value.{}[{}]".format(types[1].union_field, indices[1]) if len(types) >= 2 else "ERROR"
+ src2 = "op[2]->value.{}[{}]".format(types[2].union_field, indices[2]) if len(types) >= 3 else "ERROR"
+ src3 = "op[3]->value.{}[c]".format(types[3].union_field) if len(types) >= 4 else "ERROR"
expr = self.c_expression[types[0].union_field] if types[0].union_field in self.c_expression else self.c_expression['default']
return expr.format(src0=src0,
- src1=src1)
+ src1=src1,
+ src2=src2,
+ src3=src3)
def signatures(self):
- return type_signature_iter(self.dest_type, self.source_types, self.num_operands)
+ if self.all_signatures is not None:
+ return self.all_signatures
+ else:
+ return type_signature_iter(self.dest_type, self.source_types, self.num_operands)
ir_expression_operation = [
operation("pack_unorm_2x16", 1, printable_name="packUnorm2x16", source_types=(float_type,), dest_type=uint_type, c_expression="pack_2x16(pack_unorm_1x16, op[0]->value.f[0], op[0]->value.f[1])", flags=horizontal_operation),
operation("pack_unorm_4x8", 1, printable_name="packUnorm4x8", source_types=(float_type,), dest_type=uint_type, c_expression="pack_4x8(pack_unorm_1x8, op[0]->value.f[0], op[0]->value.f[1], op[0]->value.f[2], op[0]->value.f[3])", flags=horizontal_operation),
operation("pack_half_2x16", 1, printable_name="packHalf2x16", source_types=(float_type,), dest_type=uint_type, c_expression="pack_2x16(pack_half_1x16, op[0]->value.f[0], op[0]->value.f[1])", flags=horizontal_operation),
- operation("unpack_snorm_2x16", 1, printable_name="unpackSnorm2x16"),
- operation("unpack_snorm_4x8", 1, printable_name="unpackSnorm4x8"),
- operation("unpack_unorm_2x16", 1, printable_name="unpackUnorm2x16"),
- operation("unpack_unorm_4x8", 1, printable_name="unpackUnorm4x8"),
- operation("unpack_half_2x16", 1, printable_name="unpackHalf2x16"),
+ operation("unpack_snorm_2x16", 1, printable_name="unpackSnorm2x16", source_types=(uint_type,), dest_type=float_type, c_expression="unpack_2x16(unpack_snorm_1x16, op[0]->value.u[0], &data.f[0], &data.f[1])", flags=frozenset((horizontal_operation, non_assign_operation))),
+ operation("unpack_snorm_4x8", 1, printable_name="unpackSnorm4x8", source_types=(uint_type,), dest_type=float_type, c_expression="unpack_4x8(unpack_snorm_1x8, op[0]->value.u[0], &data.f[0], &data.f[1], &data.f[2], &data.f[3])", flags=frozenset((horizontal_operation, non_assign_operation))),
+ operation("unpack_unorm_2x16", 1, printable_name="unpackUnorm2x16", source_types=(uint_type,), dest_type=float_type, c_expression="unpack_2x16(unpack_unorm_1x16, op[0]->value.u[0], &data.f[0], &data.f[1])", flags=frozenset((horizontal_operation, non_assign_operation))),
+ operation("unpack_unorm_4x8", 1, printable_name="unpackUnorm4x8", source_types=(uint_type,), dest_type=float_type, c_expression="unpack_4x8(unpack_unorm_1x8, op[0]->value.u[0], &data.f[0], &data.f[1], &data.f[2], &data.f[3])", flags=frozenset((horizontal_operation, non_assign_operation))),
+ operation("unpack_half_2x16", 1, printable_name="unpackHalf2x16", source_types=(uint_type,), dest_type=float_type, c_expression="unpack_2x16(unpack_half_1x16, op[0]->value.u[0], &data.f[0], &data.f[1])", flags=frozenset((horizontal_operation, non_assign_operation))),
# Bit operations, part of ARB_gpu_shader5.
operation("bitfield_reverse", 1, source_types=integer_types, c_expression="bitfield_reverse({src0})"),
operation("saturate", 1, printable_name="sat", source_types=(float_type,), c_expression="CLAMP({src0}, 0.0f, 1.0f)"),
# Double packing, part of ARB_gpu_shader_fp64.
- operation("pack_double_2x32", 1, printable_name="packDouble2x32"),
- operation("unpack_double_2x32", 1, printable_name="unpackDouble2x32"),
+ operation("pack_double_2x32", 1, printable_name="packDouble2x32", source_types=(uint_type,), dest_type=double_type, c_expression="memcpy(&data.d[0], &op[0]->value.u[0], sizeof(double))", flags=frozenset((horizontal_operation, non_assign_operation))),
+ operation("unpack_double_2x32", 1, printable_name="unpackDouble2x32", source_types=(double_type,), dest_type=uint_type, c_expression="memcpy(&data.u[0], &op[0]->value.d[0], sizeof(double))", flags=frozenset((horizontal_operation, non_assign_operation))),
operation("frexp_sig", 1),
operation("frexp_exp", 1),
operation("add", 2, printable_name="+", source_types=numeric_types, c_expression="{src0} + {src1}", flags=vector_scalar_operation),
operation("sub", 2, printable_name="-", source_types=numeric_types, c_expression="{src0} - {src1}", flags=vector_scalar_operation),
# "Floating-point or low 32-bit integer multiply."
- operation("mul", 2, printable_name="*"),
+ operation("mul", 2, printable_name="*", source_types=numeric_types, c_expression="{src0} * {src1}"),
operation("imul_high", 2), # Calculates the high 32-bits of a 64-bit multiply.
operation("div", 2, printable_name="/", source_types=numeric_types, c_expression={'u': "{src1} == 0 ? 0 : {src0} / {src1}", 'i': "{src1} == 0 ? 0 : {src0} / {src1}", 'default': "{src0} / {src1}"}, flags=vector_scalar_operation),
operation("any_nequal", 2, source_types=all_types, dest_type=bool_type, c_expression="!op[0]->has_value(op[1])", flags=frozenset((horizontal_operation, types_identical_operation))),
# Bit-wise binary operations.
- operation("lshift", 2, printable_name="<<"),
- operation("rshift", 2, printable_name=">>"),
+ operation("lshift", 2, printable_name="<<", source_types=integer_types, c_expression="{src0} << {src1}", flags=frozenset((vector_scalar_operation, mixed_type_operation))),
+ operation("rshift", 2, printable_name=">>", source_types=integer_types, c_expression="{src0} >> {src1}", flags=frozenset((vector_scalar_operation, mixed_type_operation))),
operation("bit_and", 2, printable_name="&", source_types=integer_types, c_expression="{src0} & {src1}", flags=vector_scalar_operation),
operation("bit_xor", 2, printable_name="^", source_types=integer_types, c_expression="{src0} ^ {src1}", flags=vector_scalar_operation),
operation("bit_or", 2, printable_name="|", source_types=integer_types, c_expression="{src0} | {src1}", flags=vector_scalar_operation),
operation("logic_xor", 2, printable_name="^^", source_types=(bool_type,), c_expression="{src0} != {src1}"),
operation("logic_or", 2, printable_name="||", source_types=(bool_type,), c_expression="{src0} || {src1}"),
- operation("dot", 2),
+ operation("dot", 2, source_types=real_types, c_expression={'f': "dot_f(op[0], op[1])", 'd': "dot_d(op[0], op[1])"}, flags=horizontal_operation),
operation("min", 2, source_types=numeric_types, c_expression="MIN2({src0}, {src1})", flags=vector_scalar_operation),
operation("max", 2, source_types=numeric_types, c_expression="MAX2({src0}, {src1})", flags=vector_scalar_operation),
operation("ubo_load", 2),
# Multiplies a number by two to a power, part of ARB_gpu_shader5.
- operation("ldexp", 2),
+ operation("ldexp", 2,
+ all_signatures=((float_type, (float_type, int_type)),
+ (double_type, (double_type, int_type))),
+ c_expression={'f': "ldexpf_flush_subnormal({src0}, {src1})",
+ 'd': "ldexp_flush_subnormal({src0}, {src1})"}),
# Extract a scalar from a vector
#
# operand0 is the vector
# operand1 is the index of the field to read from operand0
- operation("vector_extract", 2),
+ operation("vector_extract", 2, source_types=all_types, c_expression="anything-except-None"),
# Interpolate fs input at offset
#
operation("interpolate_at_sample", 2),
# Fused floating-point multiply-add, part of ARB_gpu_shader5.
- operation("fma", 3),
+ operation("fma", 3, source_types=real_types, c_expression="{src0} * {src1} + {src2}"),
operation("lrp", 3),
# See also lower_instructions_visitor::ldexp_to_arith
operation("csel", 3),
- operation("bitfield_extract", 3),
+ operation("bitfield_extract", 3,
+ all_signatures=((int_type, (uint_type, int_type, int_type)),
+ (int_type, (int_type, int_type, int_type))),
+ c_expression={'u': "bitfield_extract_uint({src0}, {src1}, {src2})",
+ 'i': "bitfield_extract_int({src0}, {src1}, {src2})"}),
# Generate a value with one field of a vector changed
#
# operand0 is the vector
# operand1 is the value to write into the vector result
# operand2 is the index in operand0 to be modified
- operation("vector_insert", 3),
+ operation("vector_insert", 3, source_types=all_types, c_expression="anything-except-None"),
- operation("bitfield_insert", 4),
+ operation("bitfield_insert", 4,
+ all_signatures=((uint_type, (uint_type, uint_type, int_type, int_type)),
+ (int_type, (int_type, int_type, int_type, int_type))),
+ c_expression="bitfield_insert({src0}, {src1}, {src2}, {src3})",
+ flags=types_identical_operation),
- operation("vector", 4),
+ operation("vector", 4, source_types=all_types, c_expression="anything-except-None"),
]