1 // Copyright 2020 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5
6 #include <assert.h>
7 #include <stddef.h>
8
9 #include <emmintrin.h>
10
11 #include <xnnpack/math-stubs.h>
12
13
xnn_math_f32_roundu__sse2_cvt(size_t n,const float * input,float * output)14 void xnn_math_f32_roundu__sse2_cvt(
15 size_t n,
16 const float* input,
17 float* output)
18 {
19 assert(n % (4 * sizeof(float)) == 0);
20
21 // This magic number serves two purposes:
22 // 1. Set the bit corresponding to the sign of a floating-point number in a bitmask.
23 // 2. Check if the input to CVTTPS2DQ (_mm_cvttps_epi32) is out-of-range, which results in 0x80000000 output.
24 const __m128i vmagic = _mm_set1_epi32(0x80000000);
25 // Unit constant to increment results rounded "wrong way" (i.e. down) in the round-towards-zero operation.
26 const __m128 vone = _mm_set1_ps(1.0f);
27
28 for (; n != 0; n -= 4 * sizeof(float)) {
29 const __m128 vx = _mm_load_ps(input);
30 input += 4;
31
32 // Convert floating-point value x to integer, with rounding towards zero.
33 // If x is beyond [-2**31, 2**31-1] range or x is NaN, the result is -2**31 (0x80000000).
34 const __m128i vintx = _mm_cvttps_epi32(vx);
35
36 // Compute bitmask for the bits we want to copy from the rounded x. Other bits will be copied from x.
37 // If x is out-of-range for CVTTPS2DQ, we want all bits from x.
38 // If x is in-range for CVTTPS2DQ, we want all but the sign bit from the rounded x and the sign bit from x.
39 const __m128 vrndmask = _mm_castsi128_ps(_mm_or_si128(vmagic, _mm_cmpeq_epi32(vintx, vmagic)));
40
41 // Convert integer back to floating-point.
42 // We binary OR the result with the sign of x to restore the sign of negative zero.
43 const __m128 vprerndx = _mm_cvtepi32_ps(vintx);
44
45 // Combine x rounded via conversion to integer and the initial x value.
46 // For -2**31 < x < 2**31, the result is x rounded via conversion to integer.
47 // Otherwise (including NaN inputs), the result is x itself.
48 const __m128 vrndx = _mm_or_ps(_mm_and_ps(vx, vrndmask), _mm_andnot_ps(vrndmask, vprerndx));
49
50 // Compute bitmask for the bits to copy from the rounded x. Other bits will be copied from the adjusted rounded x.
51 // If rounded x >= x, we want all bits from rounded x.
52 // If rounded x < x or rounded x is NaN (implies x is NaN), we want all but the sign bit from the adjusted rounded
53 // x and the sign bit from rounded x (same as the sign bit of x).
54 const __m128 vadjmask = _mm_or_ps(_mm_cmpge_ps(vrndx, vx), _mm_castsi128_ps(vmagic));
55 // Adjust the rounded x value.
56 // The adjusted value is a unit above the rounded-towards-zero x value, but is used only if the rounded value is
57 // below x. In these cases, the adjusted value is x rounded up.
58 // Note: addition implicitly converts SNaN inputs to QNaNs.
59 const __m128 vadjrndx = _mm_add_ps(vrndx, vone);
60
61 // Combine the adjusted rounded x and the original rounded towards zero x.
62 // For rounded x < x, the result is the absolute value of adjusted rounded-towards-zero x with the sign of
63 // rounded-towards x (same as sign of x). Propagating the sign of x is important to produce negative zero
64 // for -1.0 < x < -0.5 inputs, where otherwise we would get -1.0 (rounded x) + 1.0 (adjustment) = +0.0.
65 // For rounded x >= x, the result is the rounded-towards-zero x.
66 // For NaN inputs, the result is rounded x (same as x converted to QNaN as a side-effect of adjustment).
67 const __m128 vy = _mm_or_ps(_mm_and_ps(vrndx, vadjmask), _mm_andnot_ps(vadjmask, vadjrndx));
68
69 _mm_store_ps(output, vy);
70 output += 4;
71 }
72 }
73