1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c8-wasmsimd.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <wasm_simd128.h>
13
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/math.h>
16
17
xnn_qs8_igemm_minmax_ukernel_1x4c8__wasmsimd_ld64(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_ukernel_1x4c8__wasmsimd_ld64(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
31 {
32 assert(mr != 0);
33 assert(mr <= 1);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (1 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 8);
44 int8_t* c0 = c;
45
46 const v128_t vzero = wasm_f64x2_splat(0.0);
47 do {
48 v128_t vacc0x0 = wasm_f32x4_replace_lane(vzero, 0, ((const float*) w)[0]);
49 v128_t vacc0x1 = wasm_f32x4_replace_lane(vzero, 0, ((const float*) w)[1]);
50 v128_t vacc0x2 = wasm_f32x4_replace_lane(vzero, 0, ((const float*) w)[2]);
51 v128_t vacc0x3 = wasm_f32x4_replace_lane(vzero, 0, ((const float*) w)[3]);
52 w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
53
54 size_t p = ks;
55 do {
56 const int8_t* restrict a0 = a[0];
57 if XNN_UNPREDICTABLE(a0 != zero) {
58 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
59 }
60 a += 1;
61
62 size_t k = 0;
63 while (k < kc) {
64 const v128_t vxa0 = wasm_i16x8_load_8x8(a0);
65 a0 += 8;
66
67 const v128_t vxb0 = wasm_i16x8_load_8x8(w);
68
69 const v128_t vprod0x0 = wasm_i16x8_mul(vxa0, vxb0);
70 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_widen_low_i16x8(vprod0x0));
71 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_widen_high_i16x8(vprod0x0));
72 const v128_t vxb1 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 8 * sizeof(int8_t)));
73
74 const v128_t vprod0x1 = wasm_i16x8_mul(vxa0, vxb1);
75 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_widen_low_i16x8(vprod0x1));
76 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_widen_high_i16x8(vprod0x1));
77 const v128_t vxb2 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 16 * sizeof(int8_t)));
78
79 const v128_t vprod0x2 = wasm_i16x8_mul(vxa0, vxb2);
80 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_widen_low_i16x8(vprod0x2));
81 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_widen_high_i16x8(vprod0x2));
82 const v128_t vxb3 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 24 * sizeof(int8_t)));
83
84 const v128_t vprod0x3 = wasm_i16x8_mul(vxa0, vxb3);
85 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_widen_low_i16x8(vprod0x3));
86 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_widen_high_i16x8(vprod0x3));
87
88 w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
89 k += 8 * sizeof(int8_t);
90 }
91 p -= 1 * sizeof(void*);
92 } while (p != 0);
93
94 const v128_t vacc0x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x0, vacc0x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x0, vacc0x2, 2, 6, 3, 7));
95 const v128_t vacc0x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x1, vacc0x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x1, vacc0x3, 2, 6, 3, 7));
96
97 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
98
99 const v128_t vsign0x0123 = wasm_i32x4_lt(vacc0x0123, vzero);
100
101 const v128_t vacc0x01 = wasm_v32x4_shuffle(vacc0x0123, vsign0x0123, 0, 4, 1, 5);
102
103 const v128_t vmultiplier = wasm_v128_load(params->wasmsimd.multiplier);
104 const v128_t vrounding = wasm_v128_load(params->wasmsimd.rounding);
105 const v128_t vprod0x01 = wasm_i64x2_add(wasm_i64x2_mul(vacc0x01, vmultiplier), vrounding);
106 const v128_t vacc0x23 = wasm_v32x4_shuffle(vacc0x0123, vsign0x0123, 2, 6, 3, 7);
107
108 const v128_t vprod0x23 = wasm_i64x2_add(wasm_i64x2_mul(vacc0x23, vmultiplier), vrounding);
109
110 const v128_t vq31prod0x0123 = wasm_v32x4_shuffle(vprod0x01, vprod0x23, 1, 3, 5, 7);
111
112 const v128_t vremainder_mask = wasm_v128_load(params->wasmsimd.remainder_mask);
113 const v128_t vrem0x0123 = wasm_i32x4_add(wasm_v128_and(vq31prod0x0123, vremainder_mask), wasm_i32x4_lt(vq31prod0x0123, vzero));
114
115 const v128_t vthreshold = wasm_v128_load(params->wasmsimd.remainder_threshold);
116 const int32_t vshift = params->wasmsimd.shift;
117 vacc0x0123 = wasm_i32x4_sub(wasm_i32x4_shr(vq31prod0x0123, vshift), wasm_i32x4_gt(vrem0x0123, vthreshold));
118
119 const v128_t voutput_zero_point = wasm_v128_load(params->wasmsimd.output_zero_point);
120 v128_t vacc00x0123 = wasm_i16x8_add_saturate(wasm_i16x8_narrow_i32x4(vacc0x0123, vacc0x0123), voutput_zero_point);
121
122 v128_t vout = wasm_i8x16_narrow_i16x8(vacc00x0123, vacc00x0123);
123
124 const v128_t voutput_min = wasm_v128_load(params->wasmsimd.output_min);
125 vout = wasm_i8x16_max(vout, voutput_min);
126
127 const v128_t voutput_max = wasm_v128_load(params->wasmsimd.output_max);
128 vout = wasm_i8x16_min(vout, voutput_max);
129
130 if (nc >= 4) {
131 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
132
133 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
134
135 a = (const int8_t**restrict) ((uintptr_t) a - ks);
136
137 nc -= 4;
138 } else {
139 if (nc & 2) {
140 *((uint16_t*) c0) = (uint16_t) wasm_i16x8_extract_lane(vout, 0);
141 c0 += 2;
142 vout = wasm_u32x4_shr(vout, 16);
143 }
144 if (nc & 1) {
145 *c0 = (int8_t) wasm_i8x16_extract_lane(vout, 0);
146 }
147
148 nc = 0;
149 }
150 } while (nc != 0);
151 }
152