1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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_gemm_minmax_ukernel_1x4c8__wasmsimd_ld64(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gemm_minmax_ukernel_1x4c8__wasmsimd_ld64(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
29 {
30 assert(mr != 0);
31 assert(mr <= 1);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(int8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 kc = round_up_po2(kc, 8);
40 const int8_t* a0 = a;
41 int8_t* c0 = c;
42
43 const v128_t vzero = wasm_f64x2_splat(0.0);
44 do {
45 v128_t vacc0x0 = wasm_f32x4_replace_lane(vzero, 0, ((const float*) w)[0]);
46 v128_t vacc0x1 = wasm_f32x4_replace_lane(vzero, 0, ((const float*) w)[1]);
47 v128_t vacc0x2 = wasm_f32x4_replace_lane(vzero, 0, ((const float*) w)[2]);
48 v128_t vacc0x3 = wasm_f32x4_replace_lane(vzero, 0, ((const float*) w)[3]);
49 w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
50
51 size_t k = 0;
52 while (k < kc) {
53 const v128_t vxa0 = wasm_i16x8_load_8x8(a0);
54 a0 += 8;
55
56 const v128_t vxb0 = wasm_i16x8_load_8x8(w);
57
58 const v128_t vprod0x0 = wasm_i16x8_mul(vxa0, vxb0);
59 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_widen_low_i16x8(vprod0x0));
60 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_widen_high_i16x8(vprod0x0));
61 const v128_t vxb1 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 8 * sizeof(int8_t)));
62
63 const v128_t vprod0x1 = wasm_i16x8_mul(vxa0, vxb1);
64 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_widen_low_i16x8(vprod0x1));
65 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_widen_high_i16x8(vprod0x1));
66 const v128_t vxb2 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 16 * sizeof(int8_t)));
67
68 const v128_t vprod0x2 = wasm_i16x8_mul(vxa0, vxb2);
69 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_widen_low_i16x8(vprod0x2));
70 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_widen_high_i16x8(vprod0x2));
71 const v128_t vxb3 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 24 * sizeof(int8_t)));
72
73 const v128_t vprod0x3 = wasm_i16x8_mul(vxa0, vxb3);
74 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_widen_low_i16x8(vprod0x3));
75 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_widen_high_i16x8(vprod0x3));
76
77 w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
78 k += 8 * sizeof(int8_t);
79 }
80
81 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));
82 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));
83
84 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
85
86 const v128_t vsign0x0123 = wasm_i32x4_lt(vacc0x0123, vzero);
87
88 const v128_t vacc0x01 = wasm_v32x4_shuffle(vacc0x0123, vsign0x0123, 0, 4, 1, 5);
89
90 const v128_t vmultiplier = wasm_v128_load(params->wasmsimd.multiplier);
91 const v128_t vrounding = wasm_v128_load(params->wasmsimd.rounding);
92 const v128_t vprod0x01 = wasm_i64x2_add(wasm_i64x2_mul(vacc0x01, vmultiplier), vrounding);
93 const v128_t vacc0x23 = wasm_v32x4_shuffle(vacc0x0123, vsign0x0123, 2, 6, 3, 7);
94
95 const v128_t vprod0x23 = wasm_i64x2_add(wasm_i64x2_mul(vacc0x23, vmultiplier), vrounding);
96
97 const v128_t vq31prod0x0123 = wasm_v32x4_shuffle(vprod0x01, vprod0x23, 1, 3, 5, 7);
98
99 const v128_t vremainder_mask = wasm_v128_load(params->wasmsimd.remainder_mask);
100 const v128_t vrem0x0123 = wasm_i32x4_add(wasm_v128_and(vq31prod0x0123, vremainder_mask), wasm_i32x4_lt(vq31prod0x0123, vzero));
101
102 const v128_t vthreshold = wasm_v128_load(params->wasmsimd.remainder_threshold);
103 const int32_t vshift = params->wasmsimd.shift;
104 vacc0x0123 = wasm_i32x4_sub(wasm_i32x4_shr(vq31prod0x0123, vshift), wasm_i32x4_gt(vrem0x0123, vthreshold));
105
106 const v128_t voutput_zero_point = wasm_v128_load(params->wasmsimd.output_zero_point);
107 v128_t vacc00x0123 = wasm_i16x8_add_saturate(wasm_i16x8_narrow_i32x4(vacc0x0123, vacc0x0123), voutput_zero_point);
108
109 v128_t vout = wasm_i8x16_narrow_i16x8(vacc00x0123, vacc00x0123);
110
111 const v128_t voutput_min = wasm_v128_load(params->wasmsimd.output_min);
112 vout = wasm_i8x16_max(vout, voutput_min);
113
114 const v128_t voutput_max = wasm_v128_load(params->wasmsimd.output_max);
115 vout = wasm_i8x16_min(vout, voutput_max);
116
117 if (nc >= 4) {
118 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
119
120 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
121
122 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
123
124 nc -= 4;
125 } else {
126 if (nc & 2) {
127 *((uint16_t*) c0) = (uint16_t) wasm_i16x8_extract_lane(vout, 0);
128 c0 += 2;
129 vout = wasm_u32x4_shr(vout, 16);
130 }
131 if (nc & 1) {
132 *c0 = (int8_t) wasm_i8x16_extract_lane(vout, 0);
133 }
134
135 nc = 0;
136 }
137 } while (nc != 0);
138 }
139