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_ld128(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_ld128(
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 vb01 = wasm_v128_load(w);
76 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_extend_low_u8x16(vb01), vb_zero_point);
77 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_extend_high_u8x16(vb01), vb_zero_point);
78
79 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
80 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
81 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
82 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
83
84 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
85 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
86 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
87 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
88 const v128_t vb23 = wasm_v128_load((const uint8_t*) w + 16);
89 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_extend_low_u8x16(vb23), vb_zero_point);
90 const v128_t vxb3 = wasm_i16x8_sub(wasm_u16x8_extend_high_u8x16(vb23), vb_zero_point);
91
92 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
93 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
94 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
95 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
96
97 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
98 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 3, 3, 3, 3), vxb3));
99 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
100 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 3, 3, 3, 3), vxb3));
101
102 w = (const void*) ((const uint8_t*) w + 32);
103 k -= 8 * sizeof(uint8_t);
104 }
105 if (k != 0) {
106 const v128_t vxa0 = wasm_u16x8_load8x8(a0);
107 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
108 const v128_t vxa1 = wasm_u16x8_load8x8(a1);
109 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
110
111 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
112 w = (const void*) ((const uint8_t*) w + 8);
113
114 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
115 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
116 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
117 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
118
119 if (k > 2 * sizeof(uint8_t)) {
120 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
121 w = (const void*) ((const uint8_t*) w + 8);
122
123 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
124 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
125 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
126 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
127
128 if (k > 4 * sizeof(uint8_t)) {
129 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
130 w = (const void*) ((const uint8_t*) w + 8);
131
132 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
133 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
134 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
135 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
136 }
137 }
138 }
139 p -= 2 * sizeof(void*);
140 } while (p != 0);
141
142 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
143 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
144
145 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
146 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale);
147 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale);
148
149 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
150 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
151 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
152
153 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
154 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
155 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
156
157 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
158 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
159 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
160
161 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
162
163 v128_t vout = wasm_u8x16_narrow_i16x8(vacc01x0123, vacc01x0123);
164
165 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
166 vout = wasm_u8x16_min(vout, voutput_max);
167
168 if (nc >= 4) {
169 *((float*) c1) = (float) wasm_f32x4_extract_lane(vout, 1);
170 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
171
172 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
173 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
174
175 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
176
177 nc -= 4;
178 } else {
179 uint32_t vout1 = wasm_i32x4_extract_lane(vout, 1);
180 uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
181 if (nc & 2) {
182 *((uint16_t*) c1) = (uint16_t) vout1;
183 vout1 >>= 16;
184 c1 += 2;
185 *((uint16_t*) c0) = (uint16_t) vout0;
186 vout0 >>= 16;
187 c0 += 2;
188 }
189 if (nc & 1) {
190 *c1 = (uint8_t) vout1;
191 *c0 = (uint8_t) vout0;
192 }
193
194 nc = 0;
195 }
196 } while (nc != 0);
197 }
198