+from __future__ import print_function
import re
-
-type_split_re = re.compile(r'(?P<type>[a-z]+)(?P<bits>\d+)')
-
-def type_has_size(type_):
- return type_[-1:].isdigit()
-
-def type_size(type_):
- assert type_has_size(type_)
- return int(type_split_re.match(type_).group('bits'))
-
-def type_sizes(type_):
- if type_has_size(type_):
- return [type_size(type_)]
- else:
- return [32, 64]
+from nir_opcodes import opcodes
+from nir_opcodes import type_has_size, type_size, type_sizes, type_base_type
def type_add_size(type_, size):
if type_has_size(type_):
return type_ + str(size)
def op_bit_sizes(op):
- sizes = set([8, 16, 32, 64])
+ sizes = None
if not type_has_size(op.output_type):
- sizes = sizes.intersection(set(type_sizes(op.output_type)))
+ sizes = set(type_sizes(op.output_type))
+
for input_type in op.input_types:
if not type_has_size(input_type):
- sizes = sizes.intersection(set(type_sizes(input_type)))
- return sorted(list(sizes))
+ if sizes is None:
+ sizes = set(type_sizes(input_type))
+ else:
+ sizes = sizes.intersection(set(type_sizes(input_type)))
+
+ return sorted(list(sizes)) if sizes is not None else None
def get_const_field(type_):
- if type_ == "bool32":
- return "u32"
+ if type_size(type_) == 1:
+ return 'b'
+ elif type_base_type(type_) == 'bool':
+ return 'i' + str(type_size(type_))
+ elif type_ == "float16":
+ return "u16"
else:
- m = type_split_re.match(type_)
- if not m:
- raise Exception(str(type_))
- return m.group('type')[0] + m.group('bits')
+ return type_base_type(type_)[0] + str(type_size(type_))
template = """\
/*
*/
#include <math.h>
-#include "main/core.h"
#include "util/rounding.h" /* for _mesa_roundeven */
#include "util/half_float.h"
+#include "util/bigmath.h"
#include "nir_constant_expressions.h"
+#define MAX_UINT_FOR_SIZE(bits) (UINT64_MAX >> (64 - (bits)))
+
/**
* Evaluate one component of packSnorm4x8.
*/
}
/* Some typed vector structures to make things like src0.y work */
+typedef int8_t int1_t;
+typedef uint8_t uint1_t;
+typedef float float16_t;
typedef float float32_t;
typedef double float64_t;
+typedef bool bool1_t;
+typedef bool bool8_t;
+typedef bool bool16_t;
typedef bool bool32_t;
-% for type in ["float", "int", "uint"]:
+typedef bool bool64_t;
+% for type in ["float", "int", "uint", "bool"]:
% for width in type_sizes(type):
struct ${type}${width}_vec {
${type}${width}_t x;
% endfor
% endfor
-struct bool32_vec {
- bool x;
- bool y;
- bool z;
- bool w;
-};
-
-% for name, op in sorted(opcodes.iteritems()):
-static nir_const_value
-evaluate_${name}(MAYBE_UNUSED unsigned num_components, unsigned bit_size,
- MAYBE_UNUSED nir_const_value *_src)
-{
- nir_const_value _dst_val = { {0, } };
-
- switch (bit_size) {
- % for bit_size in op_bit_sizes(op):
- case ${bit_size}: {
- <%
- output_type = type_add_size(op.output_type, bit_size)
- input_types = [type_add_size(type_, bit_size) for type_ in op.input_types]
- %>
-
- ## For each non-per-component input, create a variable srcN that
- ## contains x, y, z, and w elements which are filled in with the
- ## appropriately-typed values.
- % for j in range(op.num_inputs):
- % if op.input_sizes[j] == 0:
- <% continue %>
- % elif "src" + str(j) not in op.const_expr:
- ## Avoid unused variable warnings
- <% continue %>
- %endif
-
- const struct ${input_types[j]}_vec src${j} = {
- % for k in range(op.input_sizes[j]):
- % if input_types[j] == "bool32":
- _src[${j}].u32[${k}] != 0,
- % else:
- _src[${j}].${get_const_field(input_types[j])}[${k}],
- % endif
- % endfor
- % for k in range(op.input_sizes[j], 4):
- 0,
- % endfor
- };
+<%def name="evaluate_op(op, bit_size)">
+ <%
+ output_type = type_add_size(op.output_type, bit_size)
+ input_types = [type_add_size(type_, bit_size) for type_ in op.input_types]
+ %>
+
+ ## For each non-per-component input, create a variable srcN that
+ ## contains x, y, z, and w elements which are filled in with the
+ ## appropriately-typed values.
+ % for j in range(op.num_inputs):
+ % if op.input_sizes[j] == 0:
+ <% continue %>
+ % elif "src" + str(j) not in op.const_expr:
+ ## Avoid unused variable warnings
+ <% continue %>
+ %endif
+
+ const struct ${input_types[j]}_vec src${j} = {
+ % for k in range(op.input_sizes[j]):
+ % if input_types[j] == "int1":
+ /* 1-bit integers use a 0/-1 convention */
+ -(int1_t)_src[${j}][${k}].b,
+ % elif input_types[j] == "float16":
+ _mesa_half_to_float(_src[${j}][${k}].u16),
+ % else:
+ _src[${j}][${k}].${get_const_field(input_types[j])},
+ % endif
% endfor
+ % for k in range(op.input_sizes[j], 4):
+ 0,
+ % endfor
+ };
+ % endfor
- % if op.output_size == 0:
- ## For per-component instructions, we need to iterate over the
- ## components and apply the constant expression one component
- ## at a time.
- for (unsigned _i = 0; _i < num_components; _i++) {
- ## For each per-component input, create a variable srcN that
- ## contains the value of the current (_i'th) component.
- % for j in range(op.num_inputs):
- % if op.input_sizes[j] != 0:
- <% continue %>
- % elif "src" + str(j) not in op.const_expr:
- ## Avoid unused variable warnings
- <% continue %>
- % elif input_types[j] == "bool32":
- const bool src${j} = _src[${j}].u32[_i] != 0;
- % else:
- const ${input_types[j]}_t src${j} =
- _src[${j}].${get_const_field(input_types[j])}[_i];
- % endif
- % endfor
-
- ## Create an appropriately-typed variable dst and assign the
- ## result of the const_expr to it. If const_expr already contains
- ## writes to dst, just include const_expr directly.
- % if "dst" in op.const_expr:
- ${output_type}_t dst;
-
- ${op.const_expr}
- % else:
- ${output_type}_t dst = ${op.const_expr};
- % endif
-
- ## Store the current component of the actual destination to the
- ## value of dst.
- % if output_type == "bool32":
- ## Sanitize the C value to a proper NIR bool
- _dst_val.u32[_i] = dst ? NIR_TRUE : NIR_FALSE;
+ % if op.output_size == 0:
+ ## For per-component instructions, we need to iterate over the
+ ## components and apply the constant expression one component
+ ## at a time.
+ for (unsigned _i = 0; _i < num_components; _i++) {
+ ## For each per-component input, create a variable srcN that
+ ## contains the value of the current (_i'th) component.
+ % for j in range(op.num_inputs):
+ % if op.input_sizes[j] != 0:
+ <% continue %>
+ % elif "src" + str(j) not in op.const_expr:
+ ## Avoid unused variable warnings
+ <% continue %>
+ % elif input_types[j] == "int1":
+ /* 1-bit integers use a 0/-1 convention */
+ const int1_t src${j} = -(int1_t)_src[${j}][_i].b;
+ % elif input_types[j] == "float16":
+ const float src${j} =
+ _mesa_half_to_float(_src[${j}][_i].u16);
% else:
- _dst_val.${get_const_field(output_type)}[_i] = dst;
+ const ${input_types[j]}_t src${j} =
+ _src[${j}][_i].${get_const_field(input_types[j])};
% endif
- }
- % else:
- ## In the non-per-component case, create a struct dst with
- ## appropriately-typed elements x, y, z, and w and assign the result
- ## of the const_expr to all components of dst, or include the
- ## const_expr directly if it writes to dst already.
- struct ${output_type}_vec dst;
+ % endfor
+ ## Create an appropriately-typed variable dst and assign the
+ ## result of the const_expr to it. If const_expr already contains
+ ## writes to dst, just include const_expr directly.
% if "dst" in op.const_expr:
+ ${output_type}_t dst;
+
${op.const_expr}
% else:
- ## Splat the value to all components. This way expressions which
- ## write the same value to all components don't need to explicitly
- ## write to dest. One such example is fnoise which has a
- ## const_expr of 0.0f.
- dst.x = dst.y = dst.z = dst.w = ${op.const_expr};
+ ${output_type}_t dst = ${op.const_expr};
% endif
- ## For each component in the destination, copy the value of dst to
- ## the actual destination.
- % for k in range(op.output_size):
- % if output_type == "bool32":
- ## Sanitize the C value to a proper NIR bool
- _dst_val.u32[${k}] = dst.${"xyzw"[k]} ? NIR_TRUE : NIR_FALSE;
- % else:
- _dst_val.${get_const_field(output_type)}[${k}] = dst.${"xyzw"[k]};
- % endif
- % endfor
+ ## Store the current component of the actual destination to the
+ ## value of dst.
+ % if output_type == "int1" or output_type == "uint1":
+ /* 1-bit integers get truncated */
+ _dst_val[_i].b = dst & 1;
+ % elif output_type.startswith("bool"):
+ ## Sanitize the C value to a proper NIR 0/-1 bool
+ _dst_val[_i].${get_const_field(output_type)} = -(int)dst;
+ % elif output_type == "float16":
+ _dst_val[_i].u16 = _mesa_float_to_half(dst);
+ % else:
+ _dst_val[_i].${get_const_field(output_type)} = dst;
+ % endif
+ }
+ % else:
+ ## In the non-per-component case, create a struct dst with
+ ## appropriately-typed elements x, y, z, and w and assign the result
+ ## of the const_expr to all components of dst, or include the
+ ## const_expr directly if it writes to dst already.
+ struct ${output_type}_vec dst;
+
+ % if "dst" in op.const_expr:
+ ${op.const_expr}
+ % else:
+ ## Splat the value to all components. This way expressions which
+ ## write the same value to all components don't need to explicitly
+ ## write to dest. One such example is fnoise which has a
+ ## const_expr of 0.0f.
+ dst.x = dst.y = dst.z = dst.w = ${op.const_expr};
% endif
- break;
- }
- % endfor
-
- default:
- unreachable("unknown bit width");
- }
+ ## For each component in the destination, copy the value of dst to
+ ## the actual destination.
+ % for k in range(op.output_size):
+ % if output_type == "int1" or output_type == "uint1":
+ /* 1-bit integers get truncated */
+ _dst_val[${k}].b = dst.${"xyzw"[k]} & 1;
+ % elif output_type.startswith("bool"):
+ ## Sanitize the C value to a proper NIR 0/-1 bool
+ _dst_val[${k}].${get_const_field(output_type)} = -(int)dst.${"xyzw"[k]};
+ % elif output_type == "float16":
+ _dst_val[${k}].u16 = _mesa_float_to_half(dst.${"xyzw"[k]});
+ % else:
+ _dst_val[${k}].${get_const_field(output_type)} = dst.${"xyzw"[k]};
+ % endif
+ % endfor
+ % endif
+</%def>
+
+% for name, op in sorted(opcodes.items()):
+static void
+evaluate_${name}(nir_const_value *_dst_val,
+ UNUSED unsigned num_components,
+ ${"UNUSED" if op_bit_sizes(op) is None else ""} unsigned bit_size,
+ UNUSED nir_const_value **_src)
+{
+ % if op_bit_sizes(op) is not None:
+ switch (bit_size) {
+ % for bit_size in op_bit_sizes(op):
+ case ${bit_size}: {
+ ${evaluate_op(op, bit_size)}
+ break;
+ }
+ % endfor
- return _dst_val;
+ default:
+ unreachable("unknown bit width");
+ }
+ % else:
+ ${evaluate_op(op, 0)}
+ % endif
}
% endfor
-nir_const_value
-nir_eval_const_opcode(nir_op op, unsigned num_components,
- unsigned bit_width, nir_const_value *src)
+void
+nir_eval_const_opcode(nir_op op, nir_const_value *dest,
+ unsigned num_components, unsigned bit_width,
+ nir_const_value **src)
{
switch (op) {
-% for name in sorted(opcodes.iterkeys()):
+% for name in sorted(opcodes.keys()):
case nir_op_${name}:
- return evaluate_${name}(num_components, bit_width, src);
+ evaluate_${name}(dest, num_components, bit_width, src);
+ return;
% endfor
default:
unreachable("shouldn't get here");
}
}"""
-from nir_opcodes import opcodes
from mako.template import Template
-print Template(template).render(opcodes=opcodes, type_sizes=type_sizes,
+print(Template(template).render(opcodes=opcodes, type_sizes=type_sizes,
type_has_size=type_has_size,
type_add_size=type_add_size,
op_bit_sizes=op_bit_sizes,
- get_const_field=get_const_field)
+ get_const_field=get_const_field))