Skip to content

Commit 3cd47cf

Browse files
sbc100sbc100
authored
Remove tabs and fix formatting in test/sse/. NFC (#18245)
1 parent 2d77437 commit 3cd47cf

12 files changed

+1724
-1786
lines changed

‎test/sse/benchmark_sse.h

Lines changed: 106 additions & 109 deletions
Original file line numberDiff line numberDiff line change
@@ -27,11 +27,10 @@
2727
#endif
2828

2929
// Scalar horizonal max across four lanes.
30-
float hmax(__m128 m)
31-
{
32-
float f[4];
33-
_mm_storeu_ps(f, m);
34-
return fmax(fmax(f[0], f[1]), fmax(f[2], f[3]));
30+
float hmax(__m128 m) {
31+
float f[4];
32+
_mm_storeu_ps(f, m);
33+
return fmax(fmax(f[0], f[1]), fmax(f[2], f[3]));
3534
}
3635

3736
#include "../tick.h"
@@ -45,33 +44,33 @@ const char *chartName = "";
4544
#define SETCHART(x) chartName = (x);
4645

4746
#define START() \
48-
do { \
49-
tick_t start = tick();
47+
do { \
48+
tick_t start = tick();
5049

5150
bool comma=false;
5251
#define END(result, name) \
53-
tick_t end = tick(); \
54-
tick_t ticks = end - start; \
55-
scalarTotalTicks += scalarTicks; \
56-
simdTotalTicks += ticks; \
57-
double nsecs = (double)ticks * 1000.0 * 1000.0 * 1000.0 / ticks_per_sec() / N; \
58-
printf("%s{ \"chart\": \"%s\", \"category\": \"%s\", \"scalar\": %f, \"simd\": %f }\n", comma?",":"", chartName, name, scalarTime, nsecs); \
59-
comma = true; \
60-
printf("%s", (result) != 0 ? "Error!" : ""); \
61-
} while(0)
52+
tick_t end = tick(); \
53+
tick_t ticks = end - start; \
54+
scalarTotalTicks += scalarTicks; \
55+
simdTotalTicks += ticks; \
56+
double nsecs = (double)ticks * 1000.0 * 1000.0 * 1000.0 / ticks_per_sec() / N; \
57+
printf("%s{ \"chart\": \"%s\", \"category\": \"%s\", \"scalar\": %f, \"simd\": %f }\n", comma?",":"", chartName, name, scalarTime, nsecs); \
58+
comma = true; \
59+
printf("%s", (result) != 0 ? "Error!" : ""); \
60+
} while(0)
6261

6362
#define ENDSCALAR(result, name) \
64-
tick_t end = tick(); \
65-
scalarTicks = end - start; \
66-
scalarTime = (double)scalarTicks * 1000.0 * 1000.0 * 1000.0 / ticks_per_sec() / N; \
67-
printf("%s", (result) != 0 ? "Error!" : ""); \
68-
} while(0)
63+
tick_t end = tick(); \
64+
scalarTicks = end - start; \
65+
scalarTime = (double)scalarTicks * 1000.0 * 1000.0 * 1000.0 / ticks_per_sec() / N; \
66+
printf("%s", (result) != 0 ? "Error!" : ""); \
67+
} while(0)
6968

7069
void Print(__m128 m)
7170
{
72-
float val[4];
73-
_mm_storeu_ps(val, m);
74-
fprintf(stderr, "[%g, %g, %g, %g]\n", val[3], val[2], val[1], val[0]);
71+
float val[4];
72+
_mm_storeu_ps(val, m);
73+
fprintf(stderr, "[%g, %g, %g, %g]\n", val[3], val[2], val[1], val[0]);
7574
}
7675

7776
bool always_true() { return time(NULL) != 0; } // This function always returns true, but the compiler should not know this.
@@ -90,14 +89,13 @@ float NOINLINE *alloc_float_buffer() { return always_true() ? (float*)aligned_al
9089
double NOINLINE *alloc_double_buffer() { return always_true() ? (double*)aligned_alloc(16, (N+16)*sizeof(double)) : 0; }
9190

9291
template<typename T>
93-
T checksum_dst(T *dst)
94-
{
95-
if (always_true()) return 0.f;
96-
else
97-
{
98-
T s = 0.f; for(int i = 0; i < N; ++i) s += dst[i];
99-
return s;
100-
}
92+
T checksum_dst(T *dst) {
93+
if (always_true()) {
94+
return 0.f;
95+
} else {
96+
T s = 0.f; for(int i = 0; i < N; ++i) s += dst[i];
97+
return s;
98+
}
10199
}
102100

103101
uint32_t fcastu(float f) { return *(uint32_t*)&f; }
@@ -106,115 +104,114 @@ float ucastf(uint32_t t) { return *(float*)&t; }
106104
double ucastd(uint64_t t) { return *(double*)&t; }
107105

108106
#define LOAD_STORE_F(msg, load_instr, load_offset, store_instr, store_ptr_type, store_offset, num_elems_stride) \
109-
START(); \
110-
for(int i = 0; i < N; i += num_elems_stride) \
111-
store_instr((store_ptr_type)dst_flt+store_offset+i, load_instr(src_flt+load_offset+i)); \
112-
END(checksum_dst(dst_flt), msg);
107+
START(); \
108+
for(int i = 0; i < N; i += num_elems_stride) \
109+
store_instr((store_ptr_type)dst_flt+store_offset+i, load_instr(src_flt+load_offset+i)); \
110+
END(checksum_dst(dst_flt), msg);
113111

114112
#define LOAD_STORE_D(msg, load_instr, load_offset, store_instr, store_ptr_type, store_offset, num_elems_stride) \
115-
START(); \
116-
for(int i = 0; i < N; i += num_elems_stride) \
117-
store_instr((store_ptr_type)dst_dbl+store_offset+i, load_instr(src_dbl+load_offset+i)); \
118-
END(checksum_dst(dst_dbl), msg);
113+
START(); \
114+
for(int i = 0; i < N; i += num_elems_stride) \
115+
store_instr((store_ptr_type)dst_dbl+store_offset+i, load_instr(src_dbl+load_offset+i)); \
116+
END(checksum_dst(dst_dbl), msg);
119117

120118
#define LOAD_STORE_I(msg, load_instr, load_offset, store_instr, store_offset, num_elems_stride) \
121-
START(); \
122-
for(int i = 0; i < N; i += num_elems_stride) \
123-
store_instr((__m128i*)(dst_int+store_offset+i), load_instr((__m128i*)(src_int+load_offset+i))); \
124-
END(checksum_dst(dst_int), msg);
119+
START(); \
120+
for(int i = 0; i < N; i += num_elems_stride) \
121+
store_instr((__m128i*)(dst_int+store_offset+i), load_instr((__m128i*)(src_int+load_offset+i))); \
122+
END(checksum_dst(dst_int), msg);
125123

126124
// load M64*, store M128
127125
#define LOAD_STORE_M64(msg, reg, load_instr, load_ptr_type, load_offset, store_instr, store_ptr_type, store_offset, num_elems_stride) \
128-
START(); \
129-
for(int i = 0; i < N; i += num_elems_stride) \
130-
store_instr((store_ptr_type)dst_flt+store_offset+i, load_instr(reg, (load_ptr_type)(src_flt+load_offset+i))); \
131-
END(checksum_dst(dst_flt), msg);
126+
START(); \
127+
for(int i = 0; i < N; i += num_elems_stride) \
128+
store_instr((store_ptr_type)dst_flt+store_offset+i, load_instr(reg, (load_ptr_type)(src_flt+load_offset+i))); \
129+
END(checksum_dst(dst_flt), msg);
132130

133131
#define LOAD_STORE_64_F(msg, load_instr, load_offset, store_instr, store_offset, num_elems_stride) \
134-
START(); \
135-
for(int i = 0; i < N; i += num_elems_stride) \
136-
store_instr((__m64*)(dst_flt+store_offset+i), load_instr(src_flt+load_offset+i)); \
137-
END(checksum_dst(dst_flt), msg);
132+
START(); \
133+
for(int i = 0; i < N; i += num_elems_stride) \
134+
store_instr((__m64*)(dst_flt+store_offset+i), load_instr(src_flt+load_offset+i)); \
135+
END(checksum_dst(dst_flt), msg);
138136

139137
#define LOAD_STORE_64_D(msg, load_instr, load_offset, store_instr, store_offset, num_elems_stride) \
140-
START(); \
141-
for(int i = 0; i < N; i += num_elems_stride) \
142-
store_instr((__m64*)(dst_dbl+store_offset+i), load_instr(src_dbl+load_offset+i)); \
143-
END(checksum_dst(dst_dbl), msg);
138+
START(); \
139+
for(int i = 0; i < N; i += num_elems_stride) \
140+
store_instr((__m64*)(dst_dbl+store_offset+i), load_instr(src_dbl+load_offset+i)); \
141+
END(checksum_dst(dst_dbl), msg);
144142

145143
#define SET_STORE_F(msg, set_instr) \
146-
START(); \
147-
for(int i = 0; i < N; i += 4) \
148-
_mm_store_ps(dst_flt+i, set_instr); \
149-
END(checksum_dst(dst_flt), msg);
144+
START(); \
145+
for(int i = 0; i < N; i += 4) \
146+
_mm_store_ps(dst_flt+i, set_instr); \
147+
END(checksum_dst(dst_flt), msg);
150148

151149
#define SET_STORE_D(msg, set_instr) \
152-
START(); \
153-
for(int i = 0; i < N; i += 4) \
154-
_mm_store_pd(dst_dbl+i, set_instr); \
155-
END(checksum_dst(dst_dbl), msg);
150+
START(); \
151+
for(int i = 0; i < N; i += 4) \
152+
_mm_store_pd(dst_dbl+i, set_instr); \
153+
END(checksum_dst(dst_dbl), msg);
156154

157155
#define UNARYOP_F_F(msg, instr, op0) \
158-
START(); \
159-
__m128 o = op0; \
160-
for(int i = 0; i < N; i += 4) \
161-
o = instr(o); \
162-
_mm_store_ps(dst_flt, o); \
163-
END(checksum_dst(dst_flt), msg);
156+
START(); \
157+
__m128 o = op0; \
158+
for(int i = 0; i < N; i += 4) \
159+
o = instr(o); \
160+
_mm_store_ps(dst_flt, o); \
161+
END(checksum_dst(dst_flt), msg);
164162

165163
#define UNARYOP_I_I(msg, instr, op0) \
166-
START(); \
167-
__m128 o = op0; \
168-
for(int i = 0; i < N; i += 4) \
169-
o = instr(o); \
170-
_mm_store_si128((__m128i*)dst_int, o); \
171-
END(checksum_dst(dst_int), msg);
164+
START(); \
165+
__m128 o = op0; \
166+
for(int i = 0; i < N; i += 4) \
167+
o = instr(o); \
168+
_mm_store_si128((__m128i*)dst_int, o); \
169+
END(checksum_dst(dst_int), msg);
172170

173171
#define UNARYOP_i_F(msg, instr) \
174-
START(); \
175-
for(int i = 0; i < N; i += 4) \
176-
dst_int_scalar += instr; \
177-
END(dst_int_scalar, msg);
172+
START(); \
173+
for(int i = 0; i < N; i += 4) \
174+
dst_int_scalar += instr; \
175+
END(dst_int_scalar, msg);
178176

179177
#define UNARYOP_D_D(msg, instr, op0) \
180-
START(); \
181-
__m128d o = op0; \
182-
for(int i = 0; i < N; i += 2) \
183-
o = instr(o); \
184-
_mm_store_pd(dst_dbl, o); \
185-
END(checksum_dst(dst_dbl), msg);
178+
START(); \
179+
__m128d o = op0; \
180+
for(int i = 0; i < N; i += 2) \
181+
o = instr(o); \
182+
_mm_store_pd(dst_dbl, o); \
183+
END(checksum_dst(dst_dbl), msg);
186184

187185
#define BINARYOP_F_FF(msg, instr, op0, op1) \
188-
START(); \
189-
__m128 o0 = op0; \
190-
__m128 o1 = op1; \
191-
for(int i = 0; i < N; i += 4) \
192-
o0 = instr(o0, o1); \
193-
_mm_store_ps(dst_flt, o0); \
194-
END(checksum_dst(dst_flt), msg);
186+
START(); \
187+
__m128 o0 = op0; \
188+
__m128 o1 = op1; \
189+
for(int i = 0; i < N; i += 4) \
190+
o0 = instr(o0, o1); \
191+
_mm_store_ps(dst_flt, o0); \
192+
END(checksum_dst(dst_flt), msg);
195193

196194
#define BINARYOP_I_II(msg, instr, op0, op1) \
197-
START(); \
198-
__m128 o0 = op0; \
199-
__m128 o1 = op1; \
200-
for(int i = 0; i < N; i += 4) \
201-
o0 = instr(o0, o1); \
202-
_mm_store_si128((__m128i*)dst_int, o0); \
203-
END(checksum_dst(dst_int), msg);
195+
START(); \
196+
__m128 o0 = op0; \
197+
__m128 o1 = op1; \
198+
for(int i = 0; i < N; i += 4) \
199+
o0 = instr(o0, o1); \
200+
_mm_store_si128((__m128i*)dst_int, o0); \
201+
END(checksum_dst(dst_int), msg);
204202

205203
#define BINARYOP_D_DD(msg, instr, op0, op1) \
206-
START(); \
207-
__m128d o0 = op0; \
208-
__m128d o1 = op1; \
209-
for(int i = 0; i < N; i += 2) \
210-
o0 = instr(o0, o1); \
211-
_mm_store_pd(dst_dbl, o0); \
212-
END(checksum_dst(dst_dbl), msg);
204+
START(); \
205+
__m128d o0 = op0; \
206+
__m128d o1 = op1; \
207+
for(int i = 0; i < N; i += 2) \
208+
o0 = instr(o0, o1); \
209+
_mm_store_pd(dst_dbl, o0); \
210+
END(checksum_dst(dst_dbl), msg);
213211

214212
#define Max(a,b) ((a) >= (b) ? (a) : (b))
215213
#define Min(a,b) ((a) <= (b) ? (a) : (b))
216214

217-
static INLINE int Isnan(float __f)
218-
{
215+
static INLINE int Isnan(float __f) {
219216
return (*(unsigned int*)&__f << 1) > 0xFF000000u;
220217
}

0 commit comments

Comments
 (0)