xref: /aosp_15_r20/external/XNNPACK/src/qs8-requantization/rndnu-sse4-sra.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Copyright 2021 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 <stdint.h>
8 #include <stddef.h>
9 
10 #include <smmintrin.h>
11 
12 #include <xnnpack/math.h>
13 #include <xnnpack/requantization-stubs.h>
14 
15 
xnn_qs8_requantize_rndnu__sse4_sra(size_t n,const int32_t * input,float scale,int8_t zero_point,int8_t qmin,int8_t qmax,int8_t * output)16 void xnn_qs8_requantize_rndnu__sse4_sra(
17     size_t n,
18     const int32_t* input,
19     float scale,
20     int8_t zero_point,
21     int8_t qmin,
22     int8_t qmax,
23     int8_t* output)
24 {
25   assert(n % 16 == 0);
26   assert(scale < 1.0f);
27   assert(scale >= 0x1.0p-32f);
28 
29   const uint32_t scale_bits = float_as_uint32(scale);
30   const int32_t multiplier = ((int32_t) (scale_bits << 7) & INT32_C(0x3FFFFF80)) | INT32_C(0x40000000);
31   const uint32_t shift = 127 + 30 - (scale_bits >> 23);
32   assert(shift >= 31);
33   assert(shift < 63);
34   const int64_t rounding = INT64_C(1) << (shift - 1);
35 
36   const __m128i vmultiplier = _mm_set1_epi32(multiplier);
37   const __m128i vzero_point = _mm_set1_epi16((short) zero_point);
38   const __m128i vqmin = _mm_set1_epi8((char) qmin);
39   const __m128i vqmax = _mm_set1_epi8((char) qmax);
40   const __m128i vshift = _mm_cvtsi32_si128((int) (shift - 31));
41   const __m128i vrounding = _mm_set1_epi64x(rounding);
42   for (; n != 0; n -= 16) {
43     const __m128i x = _mm_loadu_si128((const __m128i*) input);
44     const __m128i y = _mm_loadu_si128((const __m128i*) (input + 4));
45     const __m128i z = _mm_loadu_si128((const __m128i*) (input + 8));
46     const __m128i w = _mm_loadu_si128((const __m128i*) (input + 12));
47     input += 16;
48 
49     const __m128i x_odd = _mm_shuffle_epi32(x, _MM_SHUFFLE(3, 3, 1, 1));
50     const __m128i y_odd = _mm_shuffle_epi32(y, _MM_SHUFFLE(3, 3, 1, 1));
51     const __m128i z_odd = _mm_shuffle_epi32(z, _MM_SHUFFLE(3, 3, 1, 1));
52     const __m128i w_odd = _mm_shuffle_epi32(w, _MM_SHUFFLE(3, 3, 1, 1));
53 
54     const __m128i x_product02 = _mm_mul_epi32(x, vmultiplier);
55     const __m128i y_product02 = _mm_mul_epi32(y, vmultiplier);
56     const __m128i z_product02 = _mm_mul_epi32(z, vmultiplier);
57     const __m128i w_product02 = _mm_mul_epi32(w, vmultiplier);
58 
59     const __m128i x_product13 = _mm_mul_epi32(x_odd, vmultiplier);
60     const __m128i y_product13 = _mm_mul_epi32(y_odd, vmultiplier);
61     const __m128i z_product13 = _mm_mul_epi32(z_odd, vmultiplier);
62     const __m128i w_product13 = _mm_mul_epi32(w_odd, vmultiplier);
63 
64     const __m128i x_prescaled02 = _mm_srli_epi64(_mm_add_epi64(x_product02, vrounding), 31);
65     const __m128i x_prescaled13 = _mm_slli_epi64(_mm_add_epi64(x_product13, vrounding), 1);
66     const __m128i y_prescaled02 = _mm_srli_epi64(_mm_add_epi64(y_product02, vrounding), 31);
67     const __m128i y_prescaled13 = _mm_slli_epi64(_mm_add_epi64(y_product13, vrounding), 1);
68     const __m128i z_prescaled02 = _mm_srli_epi64(_mm_add_epi64(z_product02, vrounding), 31);
69     const __m128i z_prescaled13 = _mm_slli_epi64(_mm_add_epi64(z_product13, vrounding), 1);
70     const __m128i w_prescaled02 = _mm_srli_epi64(_mm_add_epi64(w_product02, vrounding), 31);
71     const __m128i w_prescaled13 = _mm_slli_epi64(_mm_add_epi64(w_product13, vrounding), 1);
72 
73     const __m128i x_prescaled = _mm_blend_epi16(x_prescaled02, x_prescaled13, 0xCC);
74     const __m128i y_prescaled = _mm_blend_epi16(y_prescaled02, y_prescaled13, 0xCC);
75     const __m128i z_prescaled = _mm_blend_epi16(z_prescaled02, z_prescaled13, 0xCC);
76     const __m128i w_prescaled = _mm_blend_epi16(w_prescaled02, w_prescaled13, 0xCC);
77 
78     const __m128i x_scaled = _mm_sra_epi32(x_prescaled, vshift);
79     const __m128i y_scaled = _mm_sra_epi32(y_prescaled, vshift);
80     const __m128i z_scaled = _mm_sra_epi32(z_prescaled, vshift);
81     const __m128i w_scaled = _mm_sra_epi32(w_prescaled, vshift);
82 
83     const __m128i xy_packed = _mm_adds_epi16(_mm_packs_epi32(x_scaled, y_scaled), vzero_point);
84     const __m128i zw_packed = _mm_adds_epi16(_mm_packs_epi32(z_scaled, w_scaled), vzero_point);
85     const __m128i xyzw_packed = _mm_packs_epi16(xy_packed, zw_packed);
86     const __m128i xyzw_clamped = _mm_max_epi8(_mm_min_epi8(xyzw_packed, vqmax), vqmin);
87 
88     // 4x PSHUFD
89     // 8x PMULDQ
90     // 8x PADDQ
91     // 4x PSRLQ
92     // 4x PSLLQ
93     // 4x PBLENDW
94     // 4x PSRAD
95     // 2x PACKSSDW
96     // 2x PADDSW
97     // 1x PACKSSWB
98     // 1x PMAXSB
99     // 1x PMINSB
100     // ---------------------
101     // 43 instructions total
102 
103     _mm_storeu_si128((__m128i*) output, xyzw_clamped);
104     output += 16;
105   }
106 }
107