1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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/gemm.h>
15 #include <xnnpack/math.h>
16
17
18
xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2_ld64(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2_ld64(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const uint8_t* restrict a,
24 size_t a_stride,
25 const void* restrict w,
26 uint8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 2);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(uint8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 2);
41 const uint8_t* a0 = a;
42 uint8_t* c0 = c;
43 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
44 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr != 2) {
46 a1 = a0;
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 k = kc;
56 const v128_t vb_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.kernel_zero_point);
57 while (k >= 8 * sizeof(uint8_t)) {
58 const v128_t vxa0 = wasm_u16x8_load8x8((const v128_t*) a0);
59 a0 += 8;
60 const v128_t vxa1 = wasm_u16x8_load8x8((const v128_t*) a1);
61 a1 += 8;
62
63 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
64
65 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
66 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
67 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
68 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
69 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 8), vb_zero_point);
70
71 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
72 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
73 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
74 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
75 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 16), vb_zero_point);
76
77 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
78 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
79 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
80 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
81 const v128_t vxb3 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 24), vb_zero_point);
82
83 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
84 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 3, 3, 3, 3), vxb3));
85 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
86 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 3, 3, 3, 3), vxb3));
87
88 w = (const void*) ((const uint8_t*) w + 32);
89 k -= 8 * sizeof(uint8_t);
90 }
91 if (k != 0) {
92 const v128_t vxa0 = wasm_u16x8_load8x8(a0);
93 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
94 const v128_t vxa1 = wasm_u16x8_load8x8(a1);
95 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
96
97 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
98 w = (const void*) ((const uint8_t*) w + 8);
99
100 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
101 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
102 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
103 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
104
105 if (k > 2 * sizeof(uint8_t)) {
106 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
107 w = (const void*) ((const uint8_t*) w + 8);
108
109 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
110 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
111 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
112 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
113
114 if (k > 4 * sizeof(uint8_t)) {
115 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
116 w = (const void*) ((const uint8_t*) w + 8);
117
118 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
119 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
120 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
121 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
122 }
123 }
124 }
125
126 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
127 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
128
129 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
130 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale);
131 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale);
132
133 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
134 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
135 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
136
137 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
138 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
139 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
140
141 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
142 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
143 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
144
145 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
146
147 v128_t vout = wasm_u8x16_narrow_i16x8(vacc01x0123, vacc01x0123);
148
149 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
150 vout = wasm_u8x16_min(vout, voutput_max);
151
152 if (nc >= 4) {
153 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
154 *((float*) c1) = (float) wasm_f32x4_extract_lane(vout, 1);
155
156 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
157 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
158
159 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
160 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
161
162 nc -= 4;
163 } else {
164 uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
165 uint32_t vout1 = wasm_i32x4_extract_lane(vout, 1);
166 if (nc & 2) {
167 *((uint16_t*) c0) = (uint16_t) vout0;
168 vout0 >>= 16;
169 c0 += 2;
170 *((uint16_t*) c1) = (uint16_t) vout1;
171 vout1 >>= 16;
172 c1 += 2;
173 }
174 if (nc & 1) {
175 *c0 = (uint8_t) vout0;
176 *c1 = (uint8_t) vout1;
177 }
178
179 nc = 0;
180 }
181 } while (nc != 0);
182 }
183