annotate test/test_benchmark_vec.c @ 41:c6e0df09b86f default tip

*: performance improvements with old GCC, reimplement altivec
author Paper <paper@tflc.us>
date Mon, 28 Apr 2025 16:31:59 -0400
parents 4b5a557aa64f
children
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
37
4b5a557aa64f *: turns out extern is a practical joke. rewrite to be always inline again
Paper <paper@tflc.us>
parents:
diff changeset
1 #include "vec/vec.h"
4b5a557aa64f *: turns out extern is a practical joke. rewrite to be always inline again
Paper <paper@tflc.us>
parents:
diff changeset
2
4b5a557aa64f *: turns out extern is a practical joke. rewrite to be always inline again
Paper <paper@tflc.us>
parents:
diff changeset
3 extern void test_benchmark_sample_minmax_vec_impl(int16_t *smpl,
4b5a557aa64f *: turns out extern is a practical joke. rewrite to be always inline again
Paper <paper@tflc.us>
parents:
diff changeset
4 uint32_t length, int32_t *pmin, int32_t *pmax)
4b5a557aa64f *: turns out extern is a practical joke. rewrite to be always inline again
Paper <paper@tflc.us>
parents:
diff changeset
5 {
4b5a557aa64f *: turns out extern is a practical joke. rewrite to be always inline again
Paper <paper@tflc.us>
parents:
diff changeset
6 int32_t smin = INT32_MAX, smax = INT32_MIN;
4b5a557aa64f *: turns out extern is a practical joke. rewrite to be always inline again
Paper <paper@tflc.us>
parents:
diff changeset
7 uint32_t len32;
4b5a557aa64f *: turns out extern is a practical joke. rewrite to be always inline again
Paper <paper@tflc.us>
parents:
diff changeset
8 int i;
41
c6e0df09b86f *: performance improvements with old GCC, reimplement altivec
Paper <paper@tflc.us>
parents: 37
diff changeset
9 vint16x8 min = vint16x8_splat(*pmin);
c6e0df09b86f *: performance improvements with old GCC, reimplement altivec
Paper <paper@tflc.us>
parents: 37
diff changeset
10 vint16x8 max = vint16x8_splat(*pmax);
c6e0df09b86f *: performance improvements with old GCC, reimplement altivec
Paper <paper@tflc.us>
parents: 37
diff changeset
11 VINT16x8_ALIGNED_ARRAY(mins);
c6e0df09b86f *: performance improvements with old GCC, reimplement altivec
Paper <paper@tflc.us>
parents: 37
diff changeset
12 VINT16x8_ALIGNED_ARRAY(maxs);
37
4b5a557aa64f *: turns out extern is a practical joke. rewrite to be always inline again
Paper <paper@tflc.us>
parents:
diff changeset
13
41
c6e0df09b86f *: performance improvements with old GCC, reimplement altivec
Paper <paper@tflc.us>
parents: 37
diff changeset
14 len32 = length / 8;
37
4b5a557aa64f *: turns out extern is a practical joke. rewrite to be always inline again
Paper <paper@tflc.us>
parents:
diff changeset
15 while (len32--) {
41
c6e0df09b86f *: performance improvements with old GCC, reimplement altivec
Paper <paper@tflc.us>
parents: 37
diff changeset
16 vint16x8 vec = vint16x8_load_aligned(smpl);
37
4b5a557aa64f *: turns out extern is a practical joke. rewrite to be always inline again
Paper <paper@tflc.us>
parents:
diff changeset
17
41
c6e0df09b86f *: performance improvements with old GCC, reimplement altivec
Paper <paper@tflc.us>
parents: 37
diff changeset
18 min = vint16x8_min(vec, min);
c6e0df09b86f *: performance improvements with old GCC, reimplement altivec
Paper <paper@tflc.us>
parents: 37
diff changeset
19 max = vint16x8_max(vec, max);
37
4b5a557aa64f *: turns out extern is a practical joke. rewrite to be always inline again
Paper <paper@tflc.us>
parents:
diff changeset
20
41
c6e0df09b86f *: performance improvements with old GCC, reimplement altivec
Paper <paper@tflc.us>
parents: 37
diff changeset
21 smpl += 8;
37
4b5a557aa64f *: turns out extern is a practical joke. rewrite to be always inline again
Paper <paper@tflc.us>
parents:
diff changeset
22 }
4b5a557aa64f *: turns out extern is a practical joke. rewrite to be always inline again
Paper <paper@tflc.us>
parents:
diff changeset
23
41
c6e0df09b86f *: performance improvements with old GCC, reimplement altivec
Paper <paper@tflc.us>
parents: 37
diff changeset
24 vint16x8_store_aligned(min, mins);
c6e0df09b86f *: performance improvements with old GCC, reimplement altivec
Paper <paper@tflc.us>
parents: 37
diff changeset
25 vint16x8_store_aligned(max, maxs);
37
4b5a557aa64f *: turns out extern is a practical joke. rewrite to be always inline again
Paper <paper@tflc.us>
parents:
diff changeset
26
4b5a557aa64f *: turns out extern is a practical joke. rewrite to be always inline again
Paper <paper@tflc.us>
parents:
diff changeset
27 /* get the lowest minimum of what we have left */
41
c6e0df09b86f *: performance improvements with old GCC, reimplement altivec
Paper <paper@tflc.us>
parents: 37
diff changeset
28 for (i = 0; i < 8; i++) {
37
4b5a557aa64f *: turns out extern is a practical joke. rewrite to be always inline again
Paper <paper@tflc.us>
parents:
diff changeset
29 if (mins[i] < smin) smin = mins[i];
4b5a557aa64f *: turns out extern is a practical joke. rewrite to be always inline again
Paper <paper@tflc.us>
parents:
diff changeset
30 if (maxs[i] > smax) smax = maxs[i];
4b5a557aa64f *: turns out extern is a practical joke. rewrite to be always inline again
Paper <paper@tflc.us>
parents:
diff changeset
31 }
4b5a557aa64f *: turns out extern is a practical joke. rewrite to be always inline again
Paper <paper@tflc.us>
parents:
diff changeset
32
41
c6e0df09b86f *: performance improvements with old GCC, reimplement altivec
Paper <paper@tflc.us>
parents: 37
diff changeset
33 len32 = length % 8;
37
4b5a557aa64f *: turns out extern is a practical joke. rewrite to be always inline again
Paper <paper@tflc.us>
parents:
diff changeset
34 while (len32--) {
4b5a557aa64f *: turns out extern is a practical joke. rewrite to be always inline again
Paper <paper@tflc.us>
parents:
diff changeset
35 if (*smpl < smin) smin = *smpl;
4b5a557aa64f *: turns out extern is a practical joke. rewrite to be always inline again
Paper <paper@tflc.us>
parents:
diff changeset
36 if (*smpl > smax) smax = *smpl;
4b5a557aa64f *: turns out extern is a practical joke. rewrite to be always inline again
Paper <paper@tflc.us>
parents:
diff changeset
37
4b5a557aa64f *: turns out extern is a practical joke. rewrite to be always inline again
Paper <paper@tflc.us>
parents:
diff changeset
38 smpl++;
4b5a557aa64f *: turns out extern is a practical joke. rewrite to be always inline again
Paper <paper@tflc.us>
parents:
diff changeset
39 }
4b5a557aa64f *: turns out extern is a practical joke. rewrite to be always inline again
Paper <paper@tflc.us>
parents:
diff changeset
40
4b5a557aa64f *: turns out extern is a practical joke. rewrite to be always inline again
Paper <paper@tflc.us>
parents:
diff changeset
41 *pmin = smin;
4b5a557aa64f *: turns out extern is a practical joke. rewrite to be always inline again
Paper <paper@tflc.us>
parents:
diff changeset
42 *pmax = smax;
4b5a557aa64f *: turns out extern is a practical joke. rewrite to be always inline again
Paper <paper@tflc.us>
parents:
diff changeset
43 }