nir: Add support for 8 and 16-bit types

Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Reviewed-by: Eduardo Lima Mitev <elima@igalia.com>
This commit is contained in:
Jason Ekstrand
2017-03-08 20:34:28 -08:00
parent 28e41506a6
commit fbcf92a278
3 changed files with 24 additions and 2 deletions

View File

@@ -106,6 +106,10 @@ typedef enum {
typedef union { typedef union {
float f32[4]; float f32[4];
double f64[4]; double f64[4];
int8_t i8[4];
uint8_t u8[4];
int16_t i16[4];
uint16_t u16[4];
int32_t i32[4]; int32_t i32[4];
uint32_t u32[4]; uint32_t u32[4];
int64_t i64[4]; int64_t i64[4];

View File

@@ -13,8 +13,10 @@ def type_size(type_):
def type_sizes(type_): def type_sizes(type_):
if type_has_size(type_): if type_has_size(type_):
return [type_size(type_)] return [type_size(type_)]
elif type_ == 'float':
return [16, 32, 64]
else: else:
return [32, 64] return [8, 16, 32, 64]
def type_add_size(type_, size): def type_add_size(type_, size):
if type_has_size(type_): if type_has_size(type_):
@@ -38,6 +40,8 @@ def op_bit_sizes(op):
def get_const_field(type_): def get_const_field(type_):
if type_ == "bool32": if type_ == "bool32":
return "u32" return "u32"
elif type_ == "float16":
return "u16"
else: else:
m = type_split_re.match(type_) m = type_split_re.match(type_)
if not m: if not m:
@@ -250,6 +254,7 @@ unpack_half_1x16(uint16_t u)
} }
/* Some typed vector structures to make things like src0.y work */ /* Some typed vector structures to make things like src0.y work */
typedef float float16_t;
typedef float float32_t; typedef float float32_t;
typedef double float64_t; typedef double float64_t;
typedef bool bool32_t; typedef bool bool32_t;
@@ -292,6 +297,8 @@ struct bool32_vec {
% for k in range(op.input_sizes[j]): % for k in range(op.input_sizes[j]):
% if input_types[j] == "bool32": % if input_types[j] == "bool32":
_src[${j}].u32[${k}] != 0, _src[${j}].u32[${k}] != 0,
% elif input_types[j] == "float16":
_mesa_half_to_float(_src[${j}].u16[${k}]),
% else: % else:
_src[${j}].${get_const_field(input_types[j])}[${k}], _src[${j}].${get_const_field(input_types[j])}[${k}],
% endif % endif
@@ -317,6 +324,9 @@ struct bool32_vec {
<% continue %> <% continue %>
% elif input_types[j] == "bool32": % elif input_types[j] == "bool32":
const bool src${j} = _src[${j}].u32[_i] != 0; const bool src${j} = _src[${j}].u32[_i] != 0;
% elif input_types[j] == "float16":
const float src${j} =
_mesa_half_to_float(_src[${j}].u16[_i]);
% else: % else:
const ${input_types[j]}_t src${j} = const ${input_types[j]}_t src${j} =
_src[${j}].${get_const_field(input_types[j])}[_i]; _src[${j}].${get_const_field(input_types[j])}[_i];
@@ -339,6 +349,8 @@ struct bool32_vec {
% if output_type == "bool32": % if output_type == "bool32":
## Sanitize the C value to a proper NIR bool ## Sanitize the C value to a proper NIR bool
_dst_val.u32[_i] = dst ? NIR_TRUE : NIR_FALSE; _dst_val.u32[_i] = dst ? NIR_TRUE : NIR_FALSE;
% elif output_type == "float16":
_dst_val.u16[_i] = _mesa_float_to_half(dst);
% else: % else:
_dst_val.${get_const_field(output_type)}[_i] = dst; _dst_val.${get_const_field(output_type)}[_i] = dst;
% endif % endif
@@ -366,6 +378,8 @@ struct bool32_vec {
% if output_type == "bool32": % if output_type == "bool32":
## Sanitize the C value to a proper NIR bool ## Sanitize the C value to a proper NIR bool
_dst_val.u32[${k}] = dst.${"xyzw"[k]} ? NIR_TRUE : NIR_FALSE; _dst_val.u32[${k}] = dst.${"xyzw"[k]} ? NIR_TRUE : NIR_FALSE;
% elif output_type == "float16":
_dst_val.u16[${k}] = _mesa_float_to_half(dst.${"xyzw"[k]});
% else: % else:
_dst_val.${get_const_field(output_type)}[${k}] = dst.${"xyzw"[k]}; _dst_val.${get_const_field(output_type)}[${k}] = dst.${"xyzw"[k]};
% endif % endif

View File

@@ -174,7 +174,11 @@ for src_t in [tint, tuint, tfloat]:
dst_types = [tint, tuint, tfloat] dst_types = [tint, tuint, tfloat]
for dst_t in dst_types: for dst_t in dst_types:
for bit_size in [32, 64]: if dst_t == tfloat:
bit_sizes = [16, 32, 64]
else:
bit_sizes = [8, 16, 32, 64]
for bit_size in bit_sizes:
unop_convert("{0}2{1}{2}".format(src_t[0], dst_t[0], bit_size), unop_convert("{0}2{1}{2}".format(src_t[0], dst_t[0], bit_size),
dst_t + str(bit_size), src_t, "src0") dst_t + str(bit_size), src_t, "src0")