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_2x4c8__wasmsimd_ld128(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_2x4c8__wasmsimd_ld128(
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 <= 2);
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 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
43 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr != 2) {
45 a1 = a0;
46 c1 = c0;
47 }
48
49 const v128_t vzero = wasm_f64x2_splat(0.0);
50 do {
51 v128_t vacc0x0 = wasm_f32x4_replace_lane(vzero, 0, ((const float*) w)[0]);
52 v128_t vacc0x1 = wasm_f32x4_replace_lane(vzero, 0, ((const float*) w)[1]);
53 v128_t vacc0x2 = wasm_f32x4_replace_lane(vzero, 0, ((const float*) w)[2]);
54 v128_t vacc0x3 = wasm_f32x4_replace_lane(vzero, 0, ((const float*) w)[3]);
55 v128_t vacc1x0 = vacc0x0;
56 v128_t vacc1x1 = vacc0x1;
57 v128_t vacc1x2 = vacc0x2;
58 v128_t vacc1x3 = vacc0x3;
59 w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
60
61 size_t k = 0;
62 while (k < kc) {
63 const v128_t vxa0 = wasm_i16x8_load_8x8(a0);
64 a0 += 8;
65 const v128_t vxa1 = wasm_i16x8_load_8x8(a1);
66 a1 += 8;
67
68 const v128_t vb01 = wasm_v128_load(w);
69 const v128_t vxb0 = wasm_i16x8_widen_low_i8x16(vb01);
70 const v128_t vxb1 = wasm_i16x8_widen_high_i8x16(vb01);
71
72 const v128_t vprod0x0 = wasm_i16x8_mul(vxb0, vxa0);
73 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_widen_low_i16x8(vprod0x0));
74 const v128_t vprod1x0 = wasm_i16x8_mul(vxb0, vxa1);
75 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_widen_low_i16x8(vprod1x0));
76
77 const v128_t vprod0x1 = wasm_i16x8_mul(vxb1, vxa0);
78 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_widen_low_i16x8(vprod0x1));
79 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_widen_high_i16x8(vprod0x0));
80 const v128_t vprod1x1 = wasm_i16x8_mul(vxb1, vxa1);
81 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_widen_low_i16x8(vprod1x1));
82 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_widen_high_i16x8(vprod1x0));
83
84 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_widen_high_i16x8(vprod0x1));
85 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_widen_high_i16x8(vprod1x1));
86 const v128_t vb23 = wasm_v128_load((const void*) ((uintptr_t) w + 16 * sizeof(int8_t)));
87 const v128_t vxb2 = wasm_i16x8_widen_low_i8x16(vb23);
88 const v128_t vxb3 = wasm_i16x8_widen_high_i8x16(vb23);
89
90 const v128_t vprod0x2 = wasm_i16x8_mul(vxb2, vxa0);
91 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_widen_low_i16x8(vprod0x2));
92 const v128_t vprod1x2 = wasm_i16x8_mul(vxb2, vxa1);
93 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_widen_low_i16x8(vprod1x2));
94
95 const v128_t vprod0x3 = wasm_i16x8_mul(vxb3, vxa0);
96 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_widen_low_i16x8(vprod0x3));
97 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_widen_high_i16x8(vprod0x2));
98 const v128_t vprod1x3 = wasm_i16x8_mul(vxb3, vxa1);
99 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_widen_low_i16x8(vprod1x3));
100 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_widen_high_i16x8(vprod1x2));
101
102 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_widen_high_i16x8(vprod0x3));
103 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_widen_high_i16x8(vprod1x3));
104
105 w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
106 k += 8 * sizeof(int8_t);
107 }
108
109 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));
110 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));
111 const v128_t vacc1x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x0, vacc1x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x0, vacc1x2, 2, 6, 3, 7));
112 const v128_t vacc1x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x1, vacc1x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x1, vacc1x3, 2, 6, 3, 7));
113
114 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
115 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
116
117 const v128_t vsign0x0123 = wasm_i32x4_lt(vacc0x0123, vzero);
118 const v128_t vsign1x0123 = wasm_i32x4_lt(vacc1x0123, vzero);
119
120 const v128_t vacc0x01 = wasm_v32x4_shuffle(vacc0x0123, vsign0x0123, 0, 4, 1, 5);
121 const v128_t vacc1x01 = wasm_v32x4_shuffle(vacc1x0123, vsign1x0123, 0, 4, 1, 5);
122
123 const v128_t vmultiplier = wasm_v128_load(params->wasmsimd.multiplier);
124 const v128_t vrounding = wasm_v128_load(params->wasmsimd.rounding);
125 const v128_t vprod0x01 = wasm_i64x2_add(wasm_i64x2_mul(vacc0x01, vmultiplier), vrounding);
126 const v128_t vacc0x23 = wasm_v32x4_shuffle(vacc0x0123, vsign0x0123, 2, 6, 3, 7);
127 const v128_t vprod1x01 = wasm_i64x2_add(wasm_i64x2_mul(vacc1x01, vmultiplier), vrounding);
128 const v128_t vacc1x23 = wasm_v32x4_shuffle(vacc1x0123, vsign1x0123, 2, 6, 3, 7);
129
130 const v128_t vprod0x23 = wasm_i64x2_add(wasm_i64x2_mul(vacc0x23, vmultiplier), vrounding);
131 const v128_t vprod1x23 = wasm_i64x2_add(wasm_i64x2_mul(vacc1x23, vmultiplier), vrounding);
132
133 const v128_t vq31prod0x0123 = wasm_v32x4_shuffle(vprod0x01, vprod0x23, 1, 3, 5, 7);
134 const v128_t vq31prod1x0123 = wasm_v32x4_shuffle(vprod1x01, vprod1x23, 1, 3, 5, 7);
135
136 const v128_t vremainder_mask = wasm_v128_load(params->wasmsimd.remainder_mask);
137 const v128_t vrem0x0123 = wasm_i32x4_add(wasm_v128_and(vq31prod0x0123, vremainder_mask), wasm_i32x4_lt(vq31prod0x0123, vzero));
138 const v128_t vrem1x0123 = wasm_i32x4_add(wasm_v128_and(vq31prod1x0123, vremainder_mask), wasm_i32x4_lt(vq31prod1x0123, vzero));
139
140 const v128_t vthreshold = wasm_v128_load(params->wasmsimd.remainder_threshold);
141 const int32_t vshift = params->wasmsimd.shift;
142 vacc0x0123 = wasm_i32x4_sub(wasm_i32x4_shr(vq31prod0x0123, vshift), wasm_i32x4_gt(vrem0x0123, vthreshold));
143 vacc1x0123 = wasm_i32x4_sub(wasm_i32x4_shr(vq31prod1x0123, vshift), wasm_i32x4_gt(vrem1x0123, vthreshold));
144
145 const v128_t voutput_zero_point = wasm_v128_load(params->wasmsimd.output_zero_point);
146 v128_t vacc01x0123 = wasm_i16x8_add_saturate(wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123), voutput_zero_point);
147
148 v128_t vout = wasm_i8x16_narrow_i16x8(vacc01x0123, vacc01x0123);
149
150 const v128_t voutput_min = wasm_v128_load(params->wasmsimd.output_min);
151 vout = wasm_i8x16_max(vout, voutput_min);
152
153 const v128_t voutput_max = wasm_v128_load(params->wasmsimd.output_max);
154 vout = wasm_i8x16_min(vout, voutput_max);
155
156 if (nc >= 4) {
157 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
158 *((float*) c1) = (float) wasm_f32x4_extract_lane(vout, 1);
159
160 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
161 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
162
163 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
164 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
165
166 nc -= 4;
167 } else {
168 if (nc & 2) {
169 *((uint16_t*) c0) = (uint16_t) wasm_i16x8_extract_lane(vout, 0);
170 c0 += 2;
171 *((uint16_t*) c1) = (uint16_t) wasm_i16x8_extract_lane(vout, 2);
172 c1 += 2;
173 vout = wasm_u32x4_shr(vout, 16);
174 }
175 if (nc & 1) {
176 *c0 = (int8_t) wasm_i8x16_extract_lane(vout, 0);
177 *c1 = (int8_t) wasm_i8x16_extract_lane(vout, 4);
178 }
179
180 nc = 0;
181 }
182 } while (nc != 0);
183 }
184