xref: /aosp_15_r20/external/XNNPACK/src/qs8-requantization/rndnu-neon-mull.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 <arm_neon.h>
11 
12 #include <xnnpack/math.h>
13 #include <xnnpack/requantization-stubs.h>
14 
15 
xnn_qs8_requantize_rndnu__neon_mull(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__neon_mull(
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 & INT32_C(0x007FFFFF)) | INT32_C(0x00800000);
31   const int32_t shift = 127 + 23 - (scale_bits >> 23);
32   assert(shift >= 24);
33   assert(shift < 56);
34 
35 #if defined(__aarch64__)
36   const int32x4_t vmultiplier = vdupq_n_s32(multiplier);
37 #else
38   const int32x2_t vmultiplier = vdup_n_s32(multiplier);
39 #endif
40   const int16x8_t vzero_point = vdupq_n_s16((int16_t) zero_point);
41   const int64x2_t vshift = vdupq_n_s64(-shift);
42   const int8x16_t vqmin = vdupq_n_s8(qmin);
43   const int8x16_t vqmax = vdupq_n_s8(qmax);
44   for (; n != 0; n -= 16) {
45     const int32x4_t x = vld1q_s32(input);
46     const int32x4_t y = vld1q_s32(input + 4);
47     const int32x4_t z = vld1q_s32(input + 8);
48     const int32x4_t w = vld1q_s32(input + 12);
49     input += 16;
50 
51 #if defined(__aarch64__)
52     const int64x2_t x01_product = vmull_s32(vget_low_s32(x), vget_low_s32(vmultiplier));
53     const int64x2_t x23_product = vmull_high_s32(x, vmultiplier);
54     const int64x2_t y01_product = vmull_s32(vget_low_s32(y), vget_low_s32(vmultiplier));
55     const int64x2_t y23_product = vmull_high_s32(y, vmultiplier);
56     const int64x2_t z01_product = vmull_s32(vget_low_s32(z), vget_low_s32(vmultiplier));
57     const int64x2_t z23_product = vmull_high_s32(z, vmultiplier);
58     const int64x2_t w01_product = vmull_s32(vget_low_s32(w), vget_low_s32(vmultiplier));
59     const int64x2_t w23_product = vmull_high_s32(w, vmultiplier);
60 #else
61     const int64x2_t x01_product = vmull_s32(vget_low_s32(x), vmultiplier);
62     const int64x2_t x23_product = vmull_s32(vget_high_s32(x), vmultiplier);
63     const int64x2_t y01_product = vmull_s32(vget_low_s32(y), vmultiplier);
64     const int64x2_t y23_product = vmull_s32(vget_high_s32(y), vmultiplier);
65     const int64x2_t z01_product = vmull_s32(vget_low_s32(z), vmultiplier);
66     const int64x2_t z23_product = vmull_s32(vget_high_s32(z), vmultiplier);
67     const int64x2_t w01_product = vmull_s32(vget_low_s32(w), vmultiplier);
68     const int64x2_t w23_product = vmull_s32(vget_high_s32(w), vmultiplier);
69 #endif
70 
71     const int64x2_t x01_scaled = vrshlq_s64(x01_product, vshift);
72     const int64x2_t x23_scaled = vrshlq_s64(x23_product, vshift);
73     const int64x2_t y01_scaled = vrshlq_s64(y01_product, vshift);
74     const int64x2_t y23_scaled = vrshlq_s64(y23_product, vshift);
75     const int64x2_t z01_scaled = vrshlq_s64(z01_product, vshift);
76     const int64x2_t z23_scaled = vrshlq_s64(z23_product, vshift);
77     const int64x2_t w01_scaled = vrshlq_s64(w01_product, vshift);
78     const int64x2_t w23_scaled = vrshlq_s64(w23_product, vshift);
79 
80 #ifdef __aarch64__
81     const int32x4_t x_scaled = vuzp1q_s32(vreinterpretq_s32_s64(x01_scaled), vreinterpretq_s32_s64(x23_scaled));
82     const int32x4_t y_scaled = vuzp1q_s32(vreinterpretq_s32_s64(y01_scaled), vreinterpretq_s32_s64(y23_scaled));
83     const int32x4_t z_scaled = vuzp1q_s32(vreinterpretq_s32_s64(z01_scaled), vreinterpretq_s32_s64(z23_scaled));
84     const int32x4_t w_scaled = vuzp1q_s32(vreinterpretq_s32_s64(w01_scaled), vreinterpretq_s32_s64(w23_scaled));
85 
86     const int16x8_t xy_packed = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(x_scaled), y_scaled), vzero_point);
87     const int16x8_t zw_packed = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(z_scaled), w_scaled), vzero_point);
88     const int8x16_t xyzw_packed = vqmovn_high_s16(vqmovn_s16(xy_packed), zw_packed);
89 #else
90     const int32x4_t x_scaled = vcombine_s32(vmovn_s64(x01_scaled), vmovn_s64(x23_scaled));
91     const int32x4_t y_scaled = vcombine_s32(vmovn_s64(y01_scaled), vmovn_s64(y23_scaled));
92     const int32x4_t z_scaled = vcombine_s32(vmovn_s64(z01_scaled), vmovn_s64(z23_scaled));
93     const int32x4_t w_scaled = vcombine_s32(vmovn_s64(w01_scaled), vmovn_s64(w23_scaled));
94 
95     const int16x8_t xy_packed = vqaddq_s16(vcombine_s16(vqmovn_s32(x_scaled), vqmovn_s32(y_scaled)), vzero_point);
96     const int16x8_t zw_packed = vqaddq_s16(vcombine_s16(vqmovn_s32(z_scaled), vqmovn_s32(w_scaled)), vzero_point);
97     const int8x16_t xyzw_packed = vcombine_s8(vqmovn_s16(xy_packed), vqmovn_s16(zw_packed));
98 #endif
99 
100     const int8x16_t xyzw_clamped = vmaxq_s8(vminq_s8(xyzw_packed, vqmax), vqmin);
101 
102     // AArch32 version:
103     //   8x VMULL.S32 Qd, Dm, Dn
104     //   8x VRSHL.S32 Qd, Qm, Qn
105     //   8x VMOVN.S64 Dd, Qm
106     //   4x VQMOVN.S32 Dd, Qm
107     //   2x VQADD.S16 Qd, Qm, Qn
108     //   2x VQMOVUN.S16 Dd, Qm
109     //   1x VMAX.U8 Qd, Qm, Qn
110     //   1x VMIN.U8 Qd, Qm, Qn
111     // ---------------------
112     // 34 instructions total
113     //
114     // AArch64 version:
115     //   4x SMULL Vd.2D, Vn.2S, Vm.2S
116     //   4x SMULL2 Vd.2D, Vn.4S, Vm.4S
117     //   8x SRSHL Vd.2D, Vn.2D, Vm.2D
118     //   4x UZP1 Vd.4S, Vn.4S, Vm.4S
119     //   2x SQXTN Vd.4H, Vn.4S
120     //   2x SQXTN2 Vd.8H, Vn.4S
121     //   2x SQADD Vd.8H, Vn.8H, Vm.8H
122     //   1x SQXTN Vd.8B, Vn.8H
123     //   1x SQXTN2 Vd.16B, Vn.8H
124     //   1x SMIN Vd.16B, Vn.16B, Vm.16B
125     //   1x SMAX Vd.16B, Vn.16B, Vm.16B
126     // ---------------------
127     // 30 instructions total
128 
129     vst1q_s8(output, xyzw_clamped);
130     output += 16;
131   }
132 }
133