1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c2-wasmsimd-dot16x2.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 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/igemm.h>
15 #include <xnnpack/math.h>
16
17
xnn_qu8_igemm_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2_ld64(size_t mr,size_t nc,size_t kc,size_t ks,const uint8_t ** restrict a,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_igemm_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2_ld64(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const uint8_t** restrict a,
24 const void* restrict w,
25 uint8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const uint8_t* zero,
30 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
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(uint8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 2);
44 uint8_t* c0 = c;
45 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr != 2) {
47 c1 = c0;
48 }
49
50 do {
51 v128_t vacc0x0123 = wasm_v128_load(w);
52 v128_t vacc1x0123 = vacc0x0123;
53 w = (const void*) ((const int32_t*) w + 4);
54
55 size_t p = ks;
56 do {
57 const uint8_t* restrict a0 = a[0];
58 if XNN_UNPREDICTABLE(a0 != zero) {
59 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
60 }
61 const uint8_t* restrict a1 = a[1];
62 if XNN_UNPREDICTABLE(a1 != zero) {
63 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
64 }
65 a += 2;
66
67 size_t k = kc;
68 const v128_t vb_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.kernel_zero_point);
69 while (k >= 8 * sizeof(uint8_t)) {
70 const v128_t vxa0 = wasm_u16x8_load8x8(a0);
71 a0 += 8;
72 const v128_t vxa1 = wasm_u16x8_load8x8(a1);
73 a1 += 8;
74
75 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
76
77 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
78 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
79 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
80 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
81 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 8), vb_zero_point);
82
83 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
84 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
85 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
86 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
87 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 16), vb_zero_point);
88
89 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
90 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
91 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
92 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
93 const v128_t vxb3 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 24), vb_zero_point);
94
95 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
96 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 3, 3, 3, 3), vxb3));
97 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
98 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 3, 3, 3, 3), vxb3));
99
100 w = (const void*) ((const uint8_t*) w + 32);
101 k -= 8 * sizeof(uint8_t);
102 }
103 if (k != 0) {
104 const v128_t vxa0 = wasm_u16x8_load8x8(a0);
105 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
106 const v128_t vxa1 = wasm_u16x8_load8x8(a1);
107 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
108
109 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
110 w = (const void*) ((const uint8_t*) w + 8);
111
112 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
113 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
114 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
115 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
116
117 if (k > 2 * sizeof(uint8_t)) {
118 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
119 w = (const void*) ((const uint8_t*) w + 8);
120
121 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
122 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
123 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
124 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
125
126 if (k > 4 * sizeof(uint8_t)) {
127 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
128 w = (const void*) ((const uint8_t*) w + 8);
129
130 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
131 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
132 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
133 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
134 }
135 }
136 }
137 p -= 2 * sizeof(void*);
138 } while (p != 0);
139
140 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
141 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
142
143 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
144 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale);
145 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale);
146
147 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
148 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
149 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
150
151 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
152 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
153 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
154
155 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
156 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
157 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
158
159 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
160
161 v128_t vout = wasm_u8x16_narrow_i16x8(vacc01x0123, vacc01x0123);
162
163 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
164 vout = wasm_u8x16_min(vout, voutput_max);
165
166 if (nc >= 4) {
167 *((float*) c1) = (float) wasm_f32x4_extract_lane(vout, 1);
168 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
169
170 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
171 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
172
173 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
174
175 nc -= 4;
176 } else {
177 uint32_t vout1 = wasm_i32x4_extract_lane(vout, 1);
178 uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
179 if (nc & 2) {
180 *((uint16_t*) c1) = (uint16_t) vout1;
181 vout1 >>= 16;
182 c1 += 2;
183 *((uint16_t*) c0) = (uint16_t) vout0;
184 vout0 >>= 16;
185 c0 += 2;
186 }
187 if (nc & 1) {
188 *c1 = (uint8_t) vout1;
189 *c0 = (uint8_t) vout0;
190 }
191
192 nc = 0;
193 }
194 } while (nc != 0);
195 }
196