xref: /aosp_15_r20/external/XNNPACK/src/qs8-requantization/gemmlowp-sse4.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Copyright (c) Facebook, Inc. and its affiliates.
2 // All rights reserved.
3 //
4 // Copyright 2019 Google LLC
5 //
6 // This source code is licensed under the BSD-style license found in the
7 // LICENSE file in the root directory of this source tree.
8 
9 #include <assert.h>
10 #include <stdint.h>
11 #include <stddef.h>
12 
13 #include <smmintrin.h>
14 
15 #include <xnnpack/math.h>
16 #include <xnnpack/requantization-stubs.h>
17 
18 
xnn_qs8_requantize_gemmlowp__sse4(size_t n,const int32_t * input,float scale,int8_t zero_point,int8_t qmin,int8_t qmax,int8_t * output)19 void xnn_qs8_requantize_gemmlowp__sse4(
20     size_t n,
21     const int32_t* input,
22     float scale,
23     int8_t zero_point,
24     int8_t qmin,
25     int8_t qmax,
26     int8_t* output)
27 {
28   assert(n % 16 == 0);
29   assert(scale < 1.0f);
30   assert(scale >= 0x1.0p-32f);
31 
32   // Compute requantization parameters.
33   const uint32_t scale_bits = float_as_uint32(scale);
34 
35   // Multiplier is in [0x40000000, 0x7FFFFF80] range.
36   const int32_t multiplier = (int32_t)(((scale_bits & UINT32_C(0x007FFFFF)) | UINT32_C(0x00800000)) << 7);
37   assert(multiplier >= INT32_C(0x40000000));
38   assert(multiplier <= INT32_C(0x7FFFFF80));
39 
40   // Shift is in [0, 31] range.
41   const int32_t shift = 127 + 31 - 32 - (float_as_uint32(scale) >> 23);
42   assert(shift >= 0);
43   assert(shift < 32);
44 
45   const __m128i vmultiplier = _mm_set1_epi32(multiplier);
46   const __m128i vzero_point = _mm_set1_epi16((short) zero_point);
47   const __m128i vqmin = _mm_set1_epi8((char) qmin);
48   const __m128i vqmax = _mm_set1_epi8((char) qmax);
49   const __m128i vshift = _mm_cvtsi32_si128((int) shift);
50   const uint32_t remainder_mask = (UINT32_C(1) << shift) - UINT32_C(1);
51   const __m128i vremainder_mask = _mm_set1_epi32((int) remainder_mask);
52   const __m128i vthreshold = _mm_set1_epi32((int) (remainder_mask >> 1));
53   const __m128i vq31rounding = _mm_set1_epi64x(UINT64_C(0x40000000));
54   for (; n != 0; n -= 16) {
55     const __m128i x = _mm_loadu_si128((const __m128i*) input);
56     const __m128i y = _mm_loadu_si128((const __m128i*) (input + 4));
57     const __m128i z = _mm_loadu_si128((const __m128i*) (input + 8));
58     const __m128i w = _mm_loadu_si128((const __m128i*) (input + 12));
59     input += 16;
60 
61     const __m128i x_rev = _mm_shuffle_epi32(x, _MM_SHUFFLE(2, 3, 0, 1));
62     const __m128i y_rev = _mm_shuffle_epi32(y, _MM_SHUFFLE(2, 3, 0, 1));
63     const __m128i z_rev = _mm_shuffle_epi32(z, _MM_SHUFFLE(2, 3, 0, 1));
64     const __m128i w_rev = _mm_shuffle_epi32(w, _MM_SHUFFLE(2, 3, 0, 1));
65 
66     const __m128i x_product_even = _mm_add_epi64(_mm_mul_epi32(x, vmultiplier), vq31rounding);
67     const __m128i y_product_even = _mm_add_epi64(_mm_mul_epi32(y, vmultiplier), vq31rounding);
68     const __m128i z_product_even = _mm_add_epi64(_mm_mul_epi32(z, vmultiplier), vq31rounding);
69     const __m128i w_product_even = _mm_add_epi64(_mm_mul_epi32(w, vmultiplier), vq31rounding);
70 
71     const __m128i x_product_odd = _mm_add_epi64(_mm_mul_epi32(x_rev, vmultiplier), vq31rounding);
72     const __m128i y_product_odd = _mm_add_epi64(_mm_mul_epi32(y_rev, vmultiplier), vq31rounding);
73     const __m128i z_product_odd = _mm_add_epi64(_mm_mul_epi32(z_rev, vmultiplier), vq31rounding);
74     const __m128i w_product_odd = _mm_add_epi64(_mm_mul_epi32(w_rev, vmultiplier), vq31rounding);
75 
76     const __m128i x_q31product_even = _mm_srli_epi64(x_product_even, 31);
77     const __m128i x_q31product_odd = _mm_add_epi64(x_product_odd, x_product_odd);
78     const __m128i y_q31product_even = _mm_srli_epi64(y_product_even, 31);
79     const __m128i y_q31product_odd = _mm_add_epi64(y_product_odd, y_product_odd);
80     const __m128i z_q31product_even = _mm_srli_epi64(z_product_even, 31);
81     const __m128i z_q31product_odd = _mm_add_epi64(z_product_odd, z_product_odd);
82     const __m128i w_q31product_even = _mm_srli_epi64(w_product_even, 31);
83     const __m128i w_q31product_odd = _mm_add_epi64(w_product_odd, w_product_odd);
84 
85     const __m128i x_q31product = _mm_blend_epi16(x_q31product_even, x_q31product_odd, 0xCC);
86     const __m128i y_q31product = _mm_blend_epi16(y_q31product_even, y_q31product_odd, 0xCC);
87     const __m128i z_q31product = _mm_blend_epi16(z_q31product_even, z_q31product_odd, 0xCC);
88     const __m128i w_q31product = _mm_blend_epi16(w_q31product_even, w_q31product_odd, 0xCC);
89 
90     const __m128i x_remainder =
91         _mm_add_epi32(_mm_and_si128(x_q31product, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), x_q31product));
92     const __m128i y_remainder =
93         _mm_add_epi32(_mm_and_si128(y_q31product, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), y_q31product));
94     const __m128i z_remainder =
95         _mm_add_epi32(_mm_and_si128(z_q31product, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), z_q31product));
96     const __m128i w_remainder =
97         _mm_add_epi32(_mm_and_si128(w_q31product, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), w_q31product));
98 
99     const __m128i x_scaled =
100         _mm_sub_epi32(_mm_sra_epi32(x_q31product, vshift), _mm_cmpgt_epi32(x_remainder, vthreshold));
101     const __m128i y_scaled =
102         _mm_sub_epi32(_mm_sra_epi32(y_q31product, vshift), _mm_cmpgt_epi32(y_remainder, vthreshold));
103     const __m128i z_scaled =
104         _mm_sub_epi32(_mm_sra_epi32(z_q31product, vshift), _mm_cmpgt_epi32(z_remainder, vthreshold));
105     const __m128i w_scaled =
106         _mm_sub_epi32(_mm_sra_epi32(w_q31product, vshift), _mm_cmpgt_epi32(w_remainder, vthreshold));
107 
108     const __m128i xy_packed = _mm_adds_epi16(_mm_packs_epi32(x_scaled, y_scaled), vzero_point);
109     const __m128i zw_packed = _mm_adds_epi16(_mm_packs_epi32(z_scaled, w_scaled), vzero_point);
110     const __m128i xyzw_packed = _mm_packs_epi16(xy_packed, zw_packed);
111     const __m128i xyzw_clamped = _mm_max_epi8(_mm_min_epi8(xyzw_packed, vqmax), vqmin);
112 
113     // 4x PSHUFD
114     // 8x PMULDQ
115     // 12x PADDQ
116     // 4x PADDD
117     // 4x PSUBD
118     // 4x PSLRQ (immediate)
119     // 4x PSRAD (register)
120     // 4x PBLENDW
121     // 4x PAND
122     // 4x PXOR (setzero)
123     // 8x PCMPGTD
124     // 2x PACKSSDW
125     // 2x PADDSW
126     // 1x PACKSSWB
127     // 1x PMAXSB
128     // 1x PMINSB
129     // ---------------------
130     // 67 instructions total
131 
132     _mm_storeu_si128((__m128i*) output, xyzw_clamped);
133     output += 16;
134   }
135 }
136