Skip to content

Commit 70bcaa6

Browse files
brendandahlbrendandahl
committed
Add some more interesting test values and test O3 instead of O2.
1 parent b2e2f11 commit 70bcaa6

File tree

2 files changed

+118
-61
lines changed

2 files changed

+118
-61
lines changed

‎test/test_fp16.c‎

Lines changed: 116 additions & 60 deletions
Original file line numberDiff line numberDiff line change
@@ -4,6 +4,10 @@
44
#include <stdio.h>
55
#include <wasm_simd128.h>
66

7+
// SIMD True/False lane values.
8+
#define TRUE 0xFFFF
9+
#define FALSE 0
10+
711
// Not used below, but helpful for debugging.
812
void print_f16x8(v128_t v1) {
913
printf("[%g, %g, %g, %g, %g, %g, %g, %g]\n",
@@ -45,24 +49,51 @@ void assert_lanes_eq_float(v128_t v,
4549
float l5,
4650
float l6,
4751
float l7) {
48-
assert(wasm_f16x8_extract_lane(v, 0) == l0);
49-
assert(wasm_f16x8_extract_lane(v, 1) == l1);
50-
assert(wasm_f16x8_extract_lane(v, 2) == l2);
51-
assert(wasm_f16x8_extract_lane(v, 3) == l3);
52-
assert(wasm_f16x8_extract_lane(v, 4) == l4);
53-
assert(wasm_f16x8_extract_lane(v, 5) == l5);
54-
assert(wasm_f16x8_extract_lane(v, 6) == l6);
55-
assert(wasm_f16x8_extract_lane(v, 7) == l7);
52+
// We can't do a direct comparision since there may be quiet or signaling nan
53+
// values, so if the expected value is nan check that the actual value is nan
54+
// too.
55+
float actual;
56+
actual = wasm_f16x8_extract_lane(v, 0);
57+
assert(isnan(l0) ? isnan(actual) : actual == l0);
58+
actual = wasm_f16x8_extract_lane(v, 1);
59+
assert(isnan(l1) ? isnan(actual) : actual == l1);
60+
actual = wasm_f16x8_extract_lane(v, 2);
61+
assert(isnan(l2) ? isnan(actual) : actual == l2);
62+
actual = wasm_f16x8_extract_lane(v, 3);
63+
assert(isnan(l3) ? isnan(actual) : actual == l3);
64+
actual = wasm_f16x8_extract_lane(v, 4);
65+
assert(isnan(l4) ? isnan(actual) : actual == l4);
66+
actual = wasm_f16x8_extract_lane(v, 5);
67+
assert(isnan(l5) ? isnan(actual) : actual == l5);
68+
actual = wasm_f16x8_extract_lane(v, 6);
69+
assert(isnan(l6) ? isnan(actual) : actual == l6);
70+
actual = wasm_f16x8_extract_lane(v, 7);
71+
assert(isnan(l7) ? isnan(actual) : actual == l7);
5672
}
5773

5874
void assert_all_lanes_eq(v128_t v, float value) {
5975
assert_lanes_eq_float(
6076
v, value, value, value, value, value, value, value, value);
6177
}
6278

63-
void assert_all_lanes_eq_uint16(v128_t v, uint16_t value) {
64-
assert_lanes_eq_uint16(
65-
v, value, value, value, value, value, value, value, value);
79+
v128_t create_f16x8(float l0,
80+
float l1,
81+
float l2,
82+
float l3,
83+
float l4,
84+
float l5,
85+
float l6,
86+
float l7) {
87+
v128_t v;
88+
v = wasm_f16x8_replace_lane(v, 0, l0);
89+
v = wasm_f16x8_replace_lane(v, 1, l1);
90+
v = wasm_f16x8_replace_lane(v, 2, l2);
91+
v = wasm_f16x8_replace_lane(v, 3, l3);
92+
v = wasm_f16x8_replace_lane(v, 4, l4);
93+
v = wasm_f16x8_replace_lane(v, 5, l5);
94+
v = wasm_f16x8_replace_lane(v, 6, l6);
95+
v = wasm_f16x8_replace_lane(v, 7, l7);
96+
return v;
6697
}
6798

6899
int main() {
@@ -75,80 +106,105 @@ int main() {
75106
a = wasm_f16x8_replace_lane(a, 7, 99.0f);
76107
assert_lanes_eq_float(a, 99.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 99.0f);
77108

78-
a = wasm_f16x8_abs(wasm_f16x8_splat(-2.0f));
79-
assert_all_lanes_eq(a, 2.0f);
109+
// The following tests are not exhaustive, but try to hit some interesting
110+
// floats and how they interact.
80111

81-
a = wasm_f16x8_neg(wasm_f16x8_splat(2.0f));
82-
assert_all_lanes_eq(a, -2.0f);
112+
a = wasm_f16x8_abs(create_f16x8(-1.0f, 1.0f, 0.0f, -0.0f, INFINITY, -INFINITY, NAN, -NAN));
113+
assert_lanes_eq_float(a, 1.0f, 1.0f, 0.0f, 0.0f, INFINITY, INFINITY, NAN, NAN);
83114

84-
a = wasm_f16x8_sqrt(wasm_f16x8_splat(4.0f));
85-
assert_all_lanes_eq(a, 2.0f);
115+
a = wasm_f16x8_neg(create_f16x8(-1.0f, 1.0f, 0.0f, -0.0f, INFINITY, -INFINITY, NAN, -NAN));
116+
assert_lanes_eq_float(a, 1.0f, -1.0f, -0.0f, 0.0f, -INFINITY, INFINITY, NAN, NAN);
86117

87-
a = wasm_f16x8_ceil(wasm_f16x8_splat(1.2f));
88-
assert_all_lanes_eq(a, 2.0f);
118+
a = wasm_f16x8_sqrt(create_f16x8(-1.0f, 1.0f, 0.0f, 4.0f, INFINITY, -INFINITY, NAN, -NAN));
119+
assert_lanes_eq_float(a, NAN, 1.0f, 0.0f, 2.0f, INFINITY, NAN, NAN, NAN);
89120

90-
a = wasm_f16x8_floor(wasm_f16x8_splat(1.2f));
91-
assert_all_lanes_eq(a, 1.0f);
121+
a = wasm_f16x8_ceil(create_f16x8(-1.0f, 1.0f, -1.2f, 1.2f, -1.7f, 1.7f, INFINITY, -INFINITY));
122+
assert_lanes_eq_float(a, -1.0f, 1.0f, -1.0f, 2.0f, -1.0f, 2.0f, INFINITY, -INFINITY);
92123

93-
a = wasm_f16x8_trunc(wasm_f16x8_splat(1.2f));
94-
assert_all_lanes_eq(a, 1.0f);
124+
a = wasm_f16x8_floor(create_f16x8(-1.0f, 1.0f, -1.2f, 1.2f, -1.7f, 1.7f, INFINITY, -INFINITY));
125+
assert_lanes_eq_float(a, -1.0f, 1.0f, -2.0f, 1.0f, -2.0f, 1.0f, INFINITY, -INFINITY);
95126

96-
a = wasm_f16x8_nearest(wasm_f16x8_splat(1.5f));
97-
assert_all_lanes_eq(a, 2.0f);
127+
a = wasm_f16x8_trunc(create_f16x8(-1.0f, 1.0f, -1.2f, 1.2f, -1.7f, 1.7f, INFINITY, -INFINITY));
128+
assert_lanes_eq_float(a, -1.0f, 1.0f, -1.0f, 1.0f, -1.0f, 1.0f, INFINITY, -INFINITY);
98129

99-
a = wasm_f16x8_eq(wasm_f16x8_splat(2.0f), wasm_f16x8_splat(3.0f));
100-
assert_all_lanes_eq_uint16(a, 0);
130+
a = wasm_f16x8_nearest(create_f16x8(-1.0f, 1.0f, -1.2f, 1.2f, -1.7f, 1.7f, INFINITY, -INFINITY));
131+
assert_lanes_eq_float(a, -1.0f, 1.0f, -1.0f, 1.0f, -2.0f, 2.0f, INFINITY, -INFINITY);
101132

102-
a = wasm_f16x8_ne(wasm_f16x8_splat(2.0f), wasm_f16x8_splat(3.0f));
103-
assert_all_lanes_eq_uint16(a, 0xFFFF);
133+
a = wasm_f16x8_eq(create_f16x8(-1.0f, 1.0f, 1.0f, -0.0f, INFINITY, -INFINITY, NAN, -NAN),
134+
create_f16x8( 1.0f, -1.0f, 1.0f, -0.0f, INFINITY, -INFINITY, NAN, -NAN));
135+
assert_lanes_eq_uint16(a, FALSE, FALSE, TRUE, TRUE, TRUE, TRUE, FALSE, FALSE);
104136

105-
a = wasm_f16x8_lt(wasm_f16x8_splat(2.0f), wasm_f16x8_splat(3.0f));
106-
assert_all_lanes_eq_uint16(a, 0xFFFF);
137+
a = wasm_f16x8_ne(create_f16x8(-1.0f, 1.0f, 1.0f, -0.0f, INFINITY, -INFINITY, NAN, -NAN),
138+
create_f16x8( 1.0f, -1.0f, 1.0f, -0.0f, INFINITY, -INFINITY, NAN, -NAN));
139+
assert_lanes_eq_uint16(a, TRUE, TRUE, FALSE, FALSE, FALSE, FALSE, TRUE, TRUE);
107140

108-
a = wasm_f16x8_gt(wasm_f16x8_splat(2.0f), wasm_f16x8_splat(3.0f));
109-
assert_all_lanes_eq_uint16(a, 0);
141+
a = wasm_f16x8_lt(create_f16x8(-1.0f, 1.0f, 1.0f, -0.0f, INFINITY, -INFINITY, NAN, -NAN),
142+
create_f16x8( 1.0f, -1.0f, 1.0f, -0.0f, INFINITY, -INFINITY, NAN, -NAN));
143+
assert_lanes_eq_uint16(a, TRUE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE);
110144

111-
a = wasm_f16x8_le(wasm_f16x8_splat(2.0f), wasm_f16x8_splat(3.0f));
112-
assert_all_lanes_eq_uint16(a, 0xFFFF);
145+
a = wasm_f16x8_gt(create_f16x8( -1.0f, 1.0f, 1.0f, -0.0f, INFINITY, -INFINITY, NAN, -NAN),
146+
create_f16x8( 1.0f, -1.0f, 1.0f, -0.0f, INFINITY, -INFINITY, NAN, -NAN));
147+
assert_lanes_eq_uint16(a, FALSE, TRUE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE);
113148

114-
a = wasm_f16x8_ge(wasm_f16x8_splat(2.0f), wasm_f16x8_splat(3.0f));
115-
assert_all_lanes_eq_uint16(a, 0);
149+
a = wasm_f16x8_le(create_f16x8(-1.0f, 1.0f, 1.0f, -0.0f, INFINITY, -INFINITY, NAN, -NAN),
150+
create_f16x8( 1.0f, -1.0f, 1.0f, -0.0f, INFINITY, -INFINITY, NAN, -NAN));
151+
assert_lanes_eq_uint16(a, TRUE, FALSE, TRUE, TRUE, TRUE, TRUE, FALSE, FALSE);
116152

117-
a = wasm_f16x8_add(wasm_f16x8_splat(2.0f), wasm_f16x8_splat(3.0f));
118-
assert_all_lanes_eq(a, 5.0f);
153+
a = wasm_f16x8_ge(create_f16x8(-1.0f, 1.0f, 1.0f, -0.0f, INFINITY, -INFINITY, NAN, -NAN),
154+
create_f16x8( 1.0f, -1.0f, 1.0f, -0.0f, INFINITY, -INFINITY, NAN, -NAN));
155+
assert_lanes_eq_uint16(a, FALSE, TRUE, TRUE, TRUE, TRUE, TRUE, FALSE, FALSE);
119156

120-
a = wasm_f16x8_sub(wasm_f16x8_splat(5.0f), wasm_f16x8_splat(3.0f));
121-
assert_all_lanes_eq(a, 2.0f);
157+
a = wasm_f16x8_add(create_f16x8(-1.0f, 1.0f, 1.0f, 1.0f, INFINITY, INFINITY, NAN, NAN),
158+
create_f16x8( 1.0f, -1.0f, 1.0f, 0.0f, 1.0f, INFINITY, 1.0f, NAN));
159+
assert_lanes_eq_float(a, 0.0f, 0.0f, 2.0f, 1.0f, INFINITY, INFINITY, NAN, NAN);
122160

123-
a = wasm_f16x8_mul(wasm_f16x8_splat(5.0f), wasm_f16x8_splat(3.0f));
124-
assert_all_lanes_eq(a, 15.0f);
161+
a = wasm_f16x8_sub(create_f16x8(-1.0f, 1.0f, 1.0f, 1.0f, INFINITY, INFINITY, NAN, NAN),
162+
create_f16x8( 1.0f, -1.0f, 1.0f, 0.0f, 1.0f, INFINITY, 1.0f, NAN));
163+
assert_lanes_eq_float(a, -2.0f, 2.0f, 0.0f, 1.0f, INFINITY, NAN, NAN, NAN);
125164

126-
a = wasm_f16x8_div(wasm_f16x8_splat(10.0f), wasm_f16x8_splat(5.0f));
127-
assert_all_lanes_eq(a, 2.0f);
165+
a = wasm_f16x8_mul(create_f16x8(-1.0f, 1.0f, 1.0f, 1.0f, INFINITY, INFINITY, NAN, NAN),
166+
create_f16x8( 1.0f, -1.0f, 1.0f, 0.0f, 1.0f, INFINITY, 1.0f, NAN));
167+
assert_lanes_eq_float(a, -1.0f, -1.0f, 1.0f, 0.0f, INFINITY, INFINITY, NAN, NAN);
128168

129-
a = wasm_f16x8_min(wasm_f16x8_splat(10.0f), wasm_f16x8_splat(5.0f));
130-
assert_all_lanes_eq(a, 5.0f);
169+
a = wasm_f16x8_div(create_f16x8(-1.0f, 1.0f, 1.0f, 1.0f, INFINITY, INFINITY, NAN, NAN),
170+
create_f16x8( 1.0f, -1.0f, 1.0f, 0.0f, 1.0f, INFINITY, 1.0f, NAN));
171+
assert_lanes_eq_float(a, -1.0f, -1.0f, 1.0f, INFINITY, INFINITY, NAN, NAN, NAN);
131172

132-
a = wasm_f16x8_max(wasm_f16x8_splat(10.0f), wasm_f16x8_splat(5.0f));
133-
assert_all_lanes_eq(a, 10.0f);
173+
a = wasm_f16x8_min(create_f16x8(-1.0f, 1.0f, 1.0f, 1.0f, INFINITY, INFINITY, NAN, NAN),
174+
create_f16x8( 1.0f, -1.0f, 1.0f, 0.0f, 1.0f, INFINITY, 1.0f, NAN));
175+
assert_lanes_eq_float(a, -1.0f, -1.0f, 1.0f, 0.0f, 1.0f, INFINITY, NAN, NAN);
134176

135-
a = wasm_f16x8_pmin(wasm_f16x8_splat(10.0f), wasm_f16x8_splat(5.0f));
136-
assert_all_lanes_eq(a, 5.0f);
177+
a = wasm_f16x8_max(create_f16x8(-1.0f, 1.0f, 1.0f, 1.0f, INFINITY, INFINITY, NAN, NAN),
178+
create_f16x8( 1.0f, -1.0f, 1.0f, 0.0f, 1.0f, INFINITY, 1.0f, NAN));
179+
assert_lanes_eq_float(a, 1.0f, 1.0f, 1.0f, 1.0f, INFINITY, INFINITY, NAN, NAN);
137180

138-
a = wasm_f16x8_pmax(wasm_f16x8_splat(10.0f), wasm_f16x8_splat(5.0f));
139-
assert_all_lanes_eq(a, 10.0f);
181+
a = wasm_f16x8_pmin(create_f16x8(-1.0f, 1.0f, 1.0f, 1.0f, INFINITY, INFINITY, NAN, NAN),
182+
create_f16x8( 1.0f, -1.0f, 1.0f, 0.0f, 1.0f, INFINITY, 1.0f, NAN));
183+
assert_lanes_eq_float(a, -1.0f, -1.0f, 1.0f, 0.0f, 1.0f, INFINITY, NAN, NAN);
184+
185+
a = wasm_f16x8_pmax(create_f16x8(-1.0f, 1.0f, 1.0f, 1.0f, INFINITY, INFINITY, NAN, NAN),
186+
create_f16x8( 1.0f, -1.0f, 1.0f, 0.0f, 1.0f, INFINITY, 1.0f, NAN));
187+
assert_lanes_eq_float(a, 1.0f, 1.0f, 1.0f, 1.0f, INFINITY, INFINITY, NAN, NAN);
140188

141189
// TODO needed in binaryen
142190
// wasm_i16x8_trunc_sat_f16x8;
143191
// wasm_u16x8_trunc_sat_f16x8
144192
// wasm_f16x8_convert_i16x8
145193
// wasm_f16x8_convert_u16x8
146194

147-
a = wasm_f16x8_relaxed_madd(
148-
wasm_f16x8_splat(2), wasm_f16x8_splat(3), wasm_f16x8_splat(4));
149-
assert_all_lanes_eq(a, 10.0f);
150-
151-
a = wasm_f16x8_relaxed_nmadd(
152-
wasm_f16x8_splat(2), wasm_f16x8_splat(3), wasm_f16x8_splat(4));
153-
assert_all_lanes_eq(a, -2.0f);
195+
// Lane 0 illustrates the difference between fused/unfused. e.g.
196+
// fused: (positive overflow) + -inf = -inf
197+
// unfused: (inf) + -inf = NaN
198+
a = wasm_f16x8_relaxed_madd(create_f16x8( 1e4, INFINITY, -1.0f, 0.0f, 1.0f, 1.5f, -2.0f, 1.0f),
199+
create_f16x8( 1e4, INFINITY, -1.0f, 0.0f, 1.0f, 1.5f, 4.0f, 1.0f),
200+
create_f16x8(-INFINITY, INFINITY, -1.0f, 0.0f, 1.0f, 2.0f, 1.0f, -1.0f));
201+
assert_lanes_eq_float(a, -INFINITY, INFINITY, 0.0f, 0.0f, 2.0f, 4.25f, -7.0f, 0.0f);
202+
203+
// Lane 0 illustrates the difference between fused/unfused. e.g.
204+
// fused: -(positive overflow) + inf = inf
205+
// unfused: (-inf) + -inf = NaN
206+
a = wasm_f16x8_relaxed_nmadd(create_f16x8( 1e4, -INFINITY, -1.0f, 0.0f, 1.0f, 1.5f, -2.0f, 1.0f),
207+
create_f16x8( 1e4, INFINITY, -1.0f, 0.0f, 1.0f, 1.5f, 4.0f, 1.0f),
208+
create_f16x8(INFINITY, INFINITY, -1.0f, 0.0f, 1.0f, 2.0f, 1.0f, -1.0f));
209+
assert_lanes_eq_float(a, INFINITY, INFINITY, -2.0f, 0.0f, 0.0f, -0.25f, 9.0f, -2.0f);
154210
}

‎test/test_other.py‎

Lines changed: 2 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -15145,11 +15145,12 @@ def test_fs_writev_partial_write(self):
1514515145
@requires_v8
1514615146
@parameterized({
1514715147
'': [[]],
15148-
'O2': [['-O2']]
15148+
'O3': [['-O3']]
1514915149
})
1515015150
def test_fp16(self, opts):
1515115151
self.v8_args += ['--experimental-wasm-fp16']
1515215152
# TODO Remove this. Liftoff is currently broken for this test.
15153+
# https://chromium-review.googlesource.com/c/v8/v8/+/5842546
1515315154
self.v8_args += ['--no-liftoff']
1515415155
self.emcc_args = ['-msimd128', '-mfp16', '-sENVIRONMENT=shell'] + opts
1515515156
self.do_runf('test_fp16.c')

0 commit comments

Comments
 (0)