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_2x4c8__wasmsimd_ld128(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_2x4c8__wasmsimd_ld128(
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 <= 2);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (2 * 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 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr != 2) {
47 c1 = c0;
48 }
49
50 const v128_t vzero = wasm_f64x2_splat(0.0);
51 do {
52 v128_t vacc0x0 = wasm_f32x4_replace_lane(vzero, 0, ((const float*) w)[0]);
53 v128_t vacc0x1 = wasm_f32x4_replace_lane(vzero, 0, ((const float*) w)[1]);
54 v128_t vacc0x2 = wasm_f32x4_replace_lane(vzero, 0, ((const float*) w)[2]);
55 v128_t vacc0x3 = wasm_f32x4_replace_lane(vzero, 0, ((const float*) w)[3]);
56 v128_t vacc1x0 = vacc0x0;
57 v128_t vacc1x1 = vacc0x1;
58 v128_t vacc1x2 = vacc0x2;
59 v128_t vacc1x3 = vacc0x3;
60 w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
61
62 size_t p = ks;
63 do {
64 const int8_t* restrict a0 = a[0];
65 if XNN_UNPREDICTABLE(a0 != zero) {
66 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
67 }
68 const int8_t* restrict a1 = a[1];
69 if XNN_UNPREDICTABLE(a1 != zero) {
70 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
71 }
72 a += 2;
73
74 size_t k = 0;
75 while (k < kc) {
76 const v128_t vxa0 = wasm_i16x8_load_8x8(a0);
77 a0 += 8;
78 const v128_t vxa1 = wasm_i16x8_load_8x8(a1);
79 a1 += 8;
80
81 const v128_t vb01 = wasm_v128_load(w);
82 const v128_t vxb0 = wasm_i16x8_widen_low_i8x16(vb01);
83 const v128_t vxb1 = wasm_i16x8_widen_high_i8x16(vb01);
84
85 const v128_t vprod0x0 = wasm_i16x8_mul(vxb0, vxa0);
86 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_widen_low_i16x8(vprod0x0));
87 const v128_t vprod1x0 = wasm_i16x8_mul(vxb0, vxa1);
88 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_widen_low_i16x8(vprod1x0));
89
90 const v128_t vprod0x1 = wasm_i16x8_mul(vxb1, vxa0);
91 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_widen_low_i16x8(vprod0x1));
92 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_widen_high_i16x8(vprod0x0));
93 const v128_t vprod1x1 = wasm_i16x8_mul(vxb1, vxa1);
94 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_widen_low_i16x8(vprod1x1));
95 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_widen_high_i16x8(vprod1x0));
96
97 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_widen_high_i16x8(vprod0x1));
98 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_widen_high_i16x8(vprod1x1));
99 const v128_t vb23 = wasm_v128_load((const void*) ((uintptr_t) w + 16 * sizeof(int8_t)));
100 const v128_t vxb2 = wasm_i16x8_widen_low_i8x16(vb23);
101 const v128_t vxb3 = wasm_i16x8_widen_high_i8x16(vb23);
102
103 const v128_t vprod0x2 = wasm_i16x8_mul(vxb2, vxa0);
104 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_widen_low_i16x8(vprod0x2));
105 const v128_t vprod1x2 = wasm_i16x8_mul(vxb2, vxa1);
106 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_widen_low_i16x8(vprod1x2));
107
108 const v128_t vprod0x3 = wasm_i16x8_mul(vxb3, vxa0);
109 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_widen_low_i16x8(vprod0x3));
110 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_widen_high_i16x8(vprod0x2));
111 const v128_t vprod1x3 = wasm_i16x8_mul(vxb3, vxa1);
112 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_widen_low_i16x8(vprod1x3));
113 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_widen_high_i16x8(vprod1x2));
114
115 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_widen_high_i16x8(vprod0x3));
116 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_widen_high_i16x8(vprod1x3));
117
118 w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
119 k += 8 * sizeof(int8_t);
120 }
121 p -= 2 * sizeof(void*);
122 } while (p != 0);
123
124 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));
125 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));
126 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));
127 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));
128
129 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
130 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
131
132 const v128_t vsign0x0123 = wasm_i32x4_lt(vacc0x0123, vzero);
133 const v128_t vsign1x0123 = wasm_i32x4_lt(vacc1x0123, vzero);
134
135 const v128_t vacc0x01 = wasm_v32x4_shuffle(vacc0x0123, vsign0x0123, 0, 4, 1, 5);
136 const v128_t vacc1x01 = wasm_v32x4_shuffle(vacc1x0123, vsign1x0123, 0, 4, 1, 5);
137
138 const v128_t vmultiplier = wasm_v128_load(params->wasmsimd.multiplier);
139 const v128_t vrounding = wasm_v128_load(params->wasmsimd.rounding);
140 const v128_t vprod0x01 = wasm_i64x2_add(wasm_i64x2_mul(vacc0x01, vmultiplier), vrounding);
141 const v128_t vacc0x23 = wasm_v32x4_shuffle(vacc0x0123, vsign0x0123, 2, 6, 3, 7);
142 const v128_t vprod1x01 = wasm_i64x2_add(wasm_i64x2_mul(vacc1x01, vmultiplier), vrounding);
143 const v128_t vacc1x23 = wasm_v32x4_shuffle(vacc1x0123, vsign1x0123, 2, 6, 3, 7);
144
145 const v128_t vprod0x23 = wasm_i64x2_add(wasm_i64x2_mul(vacc0x23, vmultiplier), vrounding);
146 const v128_t vprod1x23 = wasm_i64x2_add(wasm_i64x2_mul(vacc1x23, vmultiplier), vrounding);
147
148 const v128_t vq31prod0x0123 = wasm_v32x4_shuffle(vprod0x01, vprod0x23, 1, 3, 5, 7);
149 const v128_t vq31prod1x0123 = wasm_v32x4_shuffle(vprod1x01, vprod1x23, 1, 3, 5, 7);
150
151 const v128_t vremainder_mask = wasm_v128_load(params->wasmsimd.remainder_mask);
152 const v128_t vrem0x0123 = wasm_i32x4_add(wasm_v128_and(vq31prod0x0123, vremainder_mask), wasm_i32x4_lt(vq31prod0x0123, vzero));
153 const v128_t vrem1x0123 = wasm_i32x4_add(wasm_v128_and(vq31prod1x0123, vremainder_mask), wasm_i32x4_lt(vq31prod1x0123, vzero));
154
155 const v128_t vthreshold = wasm_v128_load(params->wasmsimd.remainder_threshold);
156 const int32_t vshift = params->wasmsimd.shift;
157 vacc0x0123 = wasm_i32x4_sub(wasm_i32x4_shr(vq31prod0x0123, vshift), wasm_i32x4_gt(vrem0x0123, vthreshold));
158 vacc1x0123 = wasm_i32x4_sub(wasm_i32x4_shr(vq31prod1x0123, vshift), wasm_i32x4_gt(vrem1x0123, vthreshold));
159
160 const v128_t voutput_zero_point = wasm_v128_load(params->wasmsimd.output_zero_point);
161 v128_t vacc01x0123 = wasm_i16x8_add_saturate(wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123), voutput_zero_point);
162
163 v128_t vout = wasm_i8x16_narrow_i16x8(vacc01x0123, vacc01x0123);
164
165 const v128_t voutput_min = wasm_v128_load(params->wasmsimd.output_min);
166 vout = wasm_i8x16_max(vout, voutput_min);
167
168 const v128_t voutput_max = wasm_v128_load(params->wasmsimd.output_max);
169 vout = wasm_i8x16_min(vout, voutput_max);
170
171 if (nc >= 4) {
172 *((float*) c1) = (float) wasm_f32x4_extract_lane(vout, 1);
173 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
174
175 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
176 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
177
178 a = (const int8_t**restrict) ((uintptr_t) a - ks);
179
180 nc -= 4;
181 } else {
182 if (nc & 2) {
183 *((uint16_t*) c1) = (uint16_t) wasm_i16x8_extract_lane(vout, 2);
184 c1 += 2;
185 *((uint16_t*) c0) = (uint16_t) wasm_i16x8_extract_lane(vout, 0);
186 c0 += 2;
187 vout = wasm_u32x4_shr(vout, 16);
188 }
189 if (nc & 1) {
190 *c1 = (int8_t) wasm_i8x16_extract_lane(vout, 4);
191 *c0 = (int8_t) wasm_i8x16_extract_lane(vout, 0);
192 }
193
194 nc = 0;
195 }
196 } while (nc != 0);
197 }
198