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_3x4c8__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_3x4c8__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 <= 3);
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 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
49 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 a2 = a1;
52 c2 = c1;
53 }
54
55 const v128_t vzero = wasm_f64x2_splat(0.0);
56 do {
57 v128_t vacc0x0 = wasm_f32x4_replace_lane(vzero, 0, ((const float*) w)[0]);
58 v128_t vacc0x1 = wasm_f32x4_replace_lane(vzero, 0, ((const float*) w)[1]);
59 v128_t vacc0x2 = wasm_f32x4_replace_lane(vzero, 0, ((const float*) w)[2]);
60 v128_t vacc0x3 = wasm_f32x4_replace_lane(vzero, 0, ((const float*) w)[3]);
61 v128_t vacc1x0 = vacc0x0;
62 v128_t vacc1x1 = vacc0x1;
63 v128_t vacc1x2 = vacc0x2;
64 v128_t vacc1x3 = vacc0x3;
65 v128_t vacc2x0 = vacc0x0;
66 v128_t vacc2x1 = vacc0x1;
67 v128_t vacc2x2 = vacc0x2;
68 v128_t vacc2x3 = vacc0x3;
69 w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
70
71 size_t k = 0;
72 while (k < kc) {
73 const v128_t vxa0 = wasm_i16x8_load_8x8(a0);
74 a0 += 8;
75 const v128_t vxa1 = wasm_i16x8_load_8x8(a1);
76 a1 += 8;
77 const v128_t vxa2 = wasm_i16x8_load_8x8(a2);
78 a2 += 8;
79
80 const v128_t vxb0 = wasm_i16x8_load_8x8(w);
81
82 const v128_t vprod0x0 = wasm_i16x8_mul(vxa0, vxb0);
83 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_widen_low_i16x8(vprod0x0));
84 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_widen_high_i16x8(vprod0x0));
85 const v128_t vprod1x0 = wasm_i16x8_mul(vxa1, vxb0);
86 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_widen_low_i16x8(vprod1x0));
87 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_widen_high_i16x8(vprod1x0));
88 const v128_t vprod2x0 = wasm_i16x8_mul(vxa2, vxb0);
89 vacc2x0 = wasm_i32x4_add(vacc2x0, wasm_i32x4_widen_low_i16x8(vprod2x0));
90 vacc2x0 = wasm_i32x4_add(vacc2x0, wasm_i32x4_widen_high_i16x8(vprod2x0));
91 const v128_t vxb1 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 8 * sizeof(int8_t)));
92
93 const v128_t vprod0x1 = wasm_i16x8_mul(vxa0, vxb1);
94 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_widen_low_i16x8(vprod0x1));
95 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_widen_high_i16x8(vprod0x1));
96 const v128_t vprod1x1 = wasm_i16x8_mul(vxa1, vxb1);
97 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_widen_low_i16x8(vprod1x1));
98 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_widen_high_i16x8(vprod1x1));
99 const v128_t vprod2x1 = wasm_i16x8_mul(vxa2, vxb1);
100 vacc2x1 = wasm_i32x4_add(vacc2x1, wasm_i32x4_widen_low_i16x8(vprod2x1));
101 vacc2x1 = wasm_i32x4_add(vacc2x1, wasm_i32x4_widen_high_i16x8(vprod2x1));
102 const v128_t vxb2 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 16 * sizeof(int8_t)));
103
104 const v128_t vprod0x2 = wasm_i16x8_mul(vxa0, vxb2);
105 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_widen_low_i16x8(vprod0x2));
106 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_widen_high_i16x8(vprod0x2));
107 const v128_t vprod1x2 = wasm_i16x8_mul(vxa1, vxb2);
108 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_widen_low_i16x8(vprod1x2));
109 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_widen_high_i16x8(vprod1x2));
110 const v128_t vprod2x2 = wasm_i16x8_mul(vxa2, vxb2);
111 vacc2x2 = wasm_i32x4_add(vacc2x2, wasm_i32x4_widen_low_i16x8(vprod2x2));
112 vacc2x2 = wasm_i32x4_add(vacc2x2, wasm_i32x4_widen_high_i16x8(vprod2x2));
113 const v128_t vxb3 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 24 * sizeof(int8_t)));
114
115 const v128_t vprod0x3 = wasm_i16x8_mul(vxa0, vxb3);
116 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_widen_low_i16x8(vprod0x3));
117 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_widen_high_i16x8(vprod0x3));
118 const v128_t vprod1x3 = wasm_i16x8_mul(vxa1, vxb3);
119 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_widen_low_i16x8(vprod1x3));
120 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_widen_high_i16x8(vprod1x3));
121 const v128_t vprod2x3 = wasm_i16x8_mul(vxa2, vxb3);
122 vacc2x3 = wasm_i32x4_add(vacc2x3, wasm_i32x4_widen_low_i16x8(vprod2x3));
123 vacc2x3 = wasm_i32x4_add(vacc2x3, wasm_i32x4_widen_high_i16x8(vprod2x3));
124
125 w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
126 k += 8 * sizeof(int8_t);
127 }
128
129 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));
130 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));
131 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));
132 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));
133 const v128_t vacc2x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x0, vacc2x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x0, vacc2x2, 2, 6, 3, 7));
134 const v128_t vacc2x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x1, vacc2x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x1, vacc2x3, 2, 6, 3, 7));
135
136 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
137 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
138 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x02, vacc2x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x02, vacc2x13, 2, 6, 3, 7));
139
140 const v128_t vsign0x0123 = wasm_i32x4_lt(vacc0x0123, vzero);
141 const v128_t vsign1x0123 = wasm_i32x4_lt(vacc1x0123, vzero);
142 const v128_t vsign2x0123 = wasm_i32x4_lt(vacc2x0123, vzero);
143
144 const v128_t vacc0x01 = wasm_v32x4_shuffle(vacc0x0123, vsign0x0123, 0, 4, 1, 5);
145 const v128_t vacc1x01 = wasm_v32x4_shuffle(vacc1x0123, vsign1x0123, 0, 4, 1, 5);
146 const v128_t vacc2x01 = wasm_v32x4_shuffle(vacc2x0123, vsign2x0123, 0, 4, 1, 5);
147
148 const v128_t vmultiplier = wasm_v128_load(params->wasmsimd.multiplier);
149 const v128_t vrounding = wasm_v128_load(params->wasmsimd.rounding);
150 const v128_t vprod0x01 = wasm_i64x2_add(wasm_i64x2_mul(vacc0x01, vmultiplier), vrounding);
151 const v128_t vacc0x23 = wasm_v32x4_shuffle(vacc0x0123, vsign0x0123, 2, 6, 3, 7);
152 const v128_t vprod1x01 = wasm_i64x2_add(wasm_i64x2_mul(vacc1x01, vmultiplier), vrounding);
153 const v128_t vacc1x23 = wasm_v32x4_shuffle(vacc1x0123, vsign1x0123, 2, 6, 3, 7);
154 const v128_t vprod2x01 = wasm_i64x2_add(wasm_i64x2_mul(vacc2x01, vmultiplier), vrounding);
155 const v128_t vacc2x23 = wasm_v32x4_shuffle(vacc2x0123, vsign2x0123, 2, 6, 3, 7);
156
157 const v128_t vprod0x23 = wasm_i64x2_add(wasm_i64x2_mul(vacc0x23, vmultiplier), vrounding);
158 const v128_t vprod1x23 = wasm_i64x2_add(wasm_i64x2_mul(vacc1x23, vmultiplier), vrounding);
159 const v128_t vprod2x23 = wasm_i64x2_add(wasm_i64x2_mul(vacc2x23, vmultiplier), vrounding);
160
161 const v128_t vq31prod0x0123 = wasm_v32x4_shuffle(vprod0x01, vprod0x23, 1, 3, 5, 7);
162 const v128_t vq31prod1x0123 = wasm_v32x4_shuffle(vprod1x01, vprod1x23, 1, 3, 5, 7);
163 const v128_t vq31prod2x0123 = wasm_v32x4_shuffle(vprod2x01, vprod2x23, 1, 3, 5, 7);
164
165 const v128_t vremainder_mask = wasm_v128_load(params->wasmsimd.remainder_mask);
166 const v128_t vrem0x0123 = wasm_i32x4_add(wasm_v128_and(vq31prod0x0123, vremainder_mask), wasm_i32x4_lt(vq31prod0x0123, vzero));
167 const v128_t vrem1x0123 = wasm_i32x4_add(wasm_v128_and(vq31prod1x0123, vremainder_mask), wasm_i32x4_lt(vq31prod1x0123, vzero));
168 const v128_t vrem2x0123 = wasm_i32x4_add(wasm_v128_and(vq31prod2x0123, vremainder_mask), wasm_i32x4_lt(vq31prod2x0123, vzero));
169
170 const v128_t vthreshold = wasm_v128_load(params->wasmsimd.remainder_threshold);
171 const int32_t vshift = params->wasmsimd.shift;
172 vacc0x0123 = wasm_i32x4_sub(wasm_i32x4_shr(vq31prod0x0123, vshift), wasm_i32x4_gt(vrem0x0123, vthreshold));
173 vacc1x0123 = wasm_i32x4_sub(wasm_i32x4_shr(vq31prod1x0123, vshift), wasm_i32x4_gt(vrem1x0123, vthreshold));
174 vacc2x0123 = wasm_i32x4_sub(wasm_i32x4_shr(vq31prod2x0123, vshift), wasm_i32x4_gt(vrem2x0123, vthreshold));
175
176 const v128_t voutput_zero_point = wasm_v128_load(params->wasmsimd.output_zero_point);
177 v128_t vacc01x0123 = wasm_i16x8_add_saturate(wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123), voutput_zero_point);
178 v128_t vacc22x0123 = wasm_i16x8_add_saturate(wasm_i16x8_narrow_i32x4(vacc2x0123, vacc2x0123), voutput_zero_point);
179
180 v128_t vout = wasm_i8x16_narrow_i16x8(vacc01x0123, vacc22x0123);
181
182 const v128_t voutput_min = wasm_v128_load(params->wasmsimd.output_min);
183 vout = wasm_i8x16_max(vout, voutput_min);
184
185 const v128_t voutput_max = wasm_v128_load(params->wasmsimd.output_max);
186 vout = wasm_i8x16_min(vout, voutput_max);
187
188 if (nc >= 4) {
189 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
190 *((float*) c1) = (float) wasm_f32x4_extract_lane(vout, 1);
191 *((float*) c2) = (float) wasm_f32x4_extract_lane(vout, 2);
192
193 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
194 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
195 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
196
197 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
198 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
199 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
200
201 nc -= 4;
202 } else {
203 if (nc & 2) {
204 *((uint16_t*) c0) = (uint16_t) wasm_i16x8_extract_lane(vout, 0);
205 c0 += 2;
206 *((uint16_t*) c1) = (uint16_t) wasm_i16x8_extract_lane(vout, 2);
207 c1 += 2;
208 *((uint16_t*) c2) = (uint16_t) wasm_i16x8_extract_lane(vout, 4);
209 c2 += 2;
210 vout = wasm_u32x4_shr(vout, 16);
211 }
212 if (nc & 1) {
213 *c0 = (int8_t) wasm_i8x16_extract_lane(vout, 0);
214 *c1 = (int8_t) wasm_i8x16_extract_lane(vout, 4);
215 *c2 = (int8_t) wasm_i8x16_extract_lane(vout, 8);
216 }
217
218 nc = 0;
219 }
220 } while (nc != 0);
221 }
222