1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c8-wasmsimd-mul32.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_qu8_igemm_minmax_fp32_ukernel_3x4c8__wasmsimd_mul32_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_3x4c8__wasmsimd_mul32_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 <= 3);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (3 * 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, 8);
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 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53
54 const v128_t vb_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.kernel_zero_point);
55 do {
56 v128_t vacc0x0 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[0]);
57 v128_t vacc0x1 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[1]);
58 v128_t vacc0x2 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[2]);
59 v128_t vacc0x3 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[3]);
60 v128_t vacc1x0 = vacc0x0;
61 v128_t vacc1x1 = vacc0x1;
62 v128_t vacc1x2 = vacc0x2;
63 v128_t vacc1x3 = vacc0x3;
64 v128_t vacc2x0 = vacc0x0;
65 v128_t vacc2x1 = vacc0x1;
66 v128_t vacc2x2 = vacc0x2;
67 v128_t vacc2x3 = vacc0x3;
68 w = (const void*) ((const int32_t*) w + 4);
69
70 size_t p = ks;
71 do {
72 const uint8_t* restrict a0 = a[0];
73 if XNN_UNPREDICTABLE(a0 != zero) {
74 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
75 }
76 const uint8_t* restrict a1 = a[1];
77 if XNN_UNPREDICTABLE(a1 != zero) {
78 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
79 }
80 const uint8_t* restrict a2 = a[2];
81 if XNN_UNPREDICTABLE(a2 != zero) {
82 a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
83 }
84 a += 3;
85
86 size_t k = 0;
87 while (k < kc) {
88 const v128_t vxa0 = wasm_u16x8_load8x8(a0);
89 a0 += 8;
90 const v128_t vxa1 = wasm_u16x8_load8x8(a1);
91 a1 += 8;
92 const v128_t vxa2 = wasm_u16x8_load8x8(a2);
93 a2 += 8;
94
95 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
96
97 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa0), wasm_i32x4_extend_low_i16x8(vxb0)));
98 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa0), wasm_i32x4_extend_high_i16x8(vxb0)));
99 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa1), wasm_i32x4_extend_low_i16x8(vxb0)));
100 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa1), wasm_i32x4_extend_high_i16x8(vxb0)));
101 vacc2x0 = wasm_i32x4_add(vacc2x0, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa2), wasm_i32x4_extend_low_i16x8(vxb0)));
102 vacc2x0 = wasm_i32x4_add(vacc2x0, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa2), wasm_i32x4_extend_high_i16x8(vxb0)));
103 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 8), vb_zero_point);
104
105 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa0), wasm_i32x4_extend_low_i16x8(vxb1)));
106 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa0), wasm_i32x4_extend_high_i16x8(vxb1)));
107 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa1), wasm_i32x4_extend_low_i16x8(vxb1)));
108 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa1), wasm_i32x4_extend_high_i16x8(vxb1)));
109 vacc2x1 = wasm_i32x4_add(vacc2x1, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa2), wasm_i32x4_extend_low_i16x8(vxb1)));
110 vacc2x1 = wasm_i32x4_add(vacc2x1, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa2), wasm_i32x4_extend_high_i16x8(vxb1)));
111 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 16), vb_zero_point);
112
113 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa0), wasm_i32x4_extend_low_i16x8(vxb2)));
114 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa0), wasm_i32x4_extend_high_i16x8(vxb2)));
115 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa1), wasm_i32x4_extend_low_i16x8(vxb2)));
116 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa1), wasm_i32x4_extend_high_i16x8(vxb2)));
117 vacc2x2 = wasm_i32x4_add(vacc2x2, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa2), wasm_i32x4_extend_low_i16x8(vxb2)));
118 vacc2x2 = wasm_i32x4_add(vacc2x2, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa2), wasm_i32x4_extend_high_i16x8(vxb2)));
119 const v128_t vxb3 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 24), vb_zero_point);
120
121 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa0), wasm_i32x4_extend_low_i16x8(vxb3)));
122 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa0), wasm_i32x4_extend_high_i16x8(vxb3)));
123 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa1), wasm_i32x4_extend_low_i16x8(vxb3)));
124 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa1), wasm_i32x4_extend_high_i16x8(vxb3)));
125 vacc2x3 = wasm_i32x4_add(vacc2x3, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa2), wasm_i32x4_extend_low_i16x8(vxb3)));
126 vacc2x3 = wasm_i32x4_add(vacc2x3, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa2), wasm_i32x4_extend_high_i16x8(vxb3)));
127
128 w = (const void*) ((const uint8_t*) w + 32);
129 k += 8 * sizeof(uint8_t);
130 }
131 p -= 3 * sizeof(void*);
132 } while (p != 0);
133
134 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));
135 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));
136 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));
137 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));
138 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));
139 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));
140
141 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
142 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
143 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x02, vacc2x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x02, vacc2x13, 2, 6, 3, 7));
144
145 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
146 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
147 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
148
149 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
150 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale);
151 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale);
152 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale);
153
154 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
155 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
156 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
157 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);
158
159 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
160 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
161 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
162 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vmagic_min);
163
164 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
165 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
166 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
167 vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);
168
169 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
170 v128_t vacc22x0123 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc2x0123);
171
172 v128_t vout = wasm_u8x16_narrow_i16x8(vacc01x0123, vacc22x0123);
173
174 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
175 vout = wasm_u8x16_min(vout, voutput_max);
176
177 if (nc >= 4) {
178 *((float*) c2) = (float) wasm_f32x4_extract_lane(vout, 2);
179 *((float*) c1) = (float) wasm_f32x4_extract_lane(vout, 1);
180 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
181
182 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
183 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
184 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
185
186 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
187
188 nc -= 4;
189 } else {
190 uint32_t vout2 = wasm_i32x4_extract_lane(vout, 2);
191 uint32_t vout1 = wasm_i32x4_extract_lane(vout, 1);
192 uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
193 if (nc & 2) {
194 *((uint16_t*) c2) = (uint16_t) vout2;
195 vout2 >>= 16;
196 c2 += 2;
197 *((uint16_t*) c1) = (uint16_t) vout1;
198 vout1 >>= 16;
199 c1 += 2;
200 *((uint16_t*) c0) = (uint16_t) vout0;
201 vout0 >>= 16;
202 c0 += 2;
203 }
204 if (nc & 1) {
205 *c2 = (uint8_t) vout2;
206 *c1 = (uint8_t) vout1;
207 *c0 = (uint8_t) vout0;
208 }
209
210 nc = 0;
211 }
212 } while (nc != 0);
213 }
214