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_ld128(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_ld128(
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 vb01 = wasm_v128_load(w);
64 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_extend_low_u8x16(vb01), vb_zero_point);
65 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_extend_high_u8x16(vb01), vb_zero_point);
66
67 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
68 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
69 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
70 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
71
72 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
73 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
74 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
75 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
76 const v128_t vb23 = wasm_v128_load((const uint8_t*) w + 16);
77 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_extend_low_u8x16(vb23), vb_zero_point);
78 const v128_t vxb3 = wasm_i16x8_sub(wasm_u16x8_extend_high_u8x16(vb23), vb_zero_point);
79
80 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
81 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
82 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
83 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
84
85 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
86 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 3, 3, 3, 3), vxb3));
87 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
88 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 3, 3, 3, 3), vxb3));
89
90 w = (const void*) ((const uint8_t*) w + 32);
91 k -= 8 * sizeof(uint8_t);
92 }
93 if (k != 0) {
94 const v128_t vxa0 = wasm_u16x8_load8x8(a0);
95 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
96 const v128_t vxa1 = wasm_u16x8_load8x8(a1);
97 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
98
99 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
100 w = (const void*) ((const uint8_t*) w + 8);
101
102 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
103 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
104 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
105 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
106
107 if (k > 2 * sizeof(uint8_t)) {
108 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
109 w = (const void*) ((const uint8_t*) w + 8);
110
111 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
112 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
113 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
114 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
115
116 if (k > 4 * sizeof(uint8_t)) {
117 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
118 w = (const void*) ((const uint8_t*) w + 8);
119
120 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
121 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
122 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
123 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
124 }
125 }
126 }
127
128 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
129 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
130
131 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
132 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale);
133 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale);
134
135 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
136 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
137 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
138
139 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
140 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
141 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
142
143 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
144 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
145 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
146
147 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
148
149 v128_t vout = wasm_u8x16_narrow_i16x8(vacc01x0123, vacc01x0123);
150
151 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
152 vout = wasm_u8x16_min(vout, voutput_max);
153
154 if (nc >= 4) {
155 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
156 *((float*) c1) = (float) wasm_f32x4_extract_lane(vout, 1);
157
158 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
159 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
160
161 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
162 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
163
164 nc -= 4;
165 } else {
166 uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
167 uint32_t vout1 = wasm_i32x4_extract_lane(vout, 1);
168 if (nc & 2) {
169 *((uint16_t*) c0) = (uint16_t) vout0;
170 vout0 >>= 16;
171 c0 += 2;
172 *((uint16_t*) c1) = (uint16_t) vout1;
173 vout1 >>= 16;
174 c1 += 2;
175 }
176 if (nc & 1) {
177 *c0 = (uint8_t) vout0;
178 *c1 = (uint8_t) vout1;
179 }
180
181 nc = 0;
182 }
183 } while (nc != 0);
184 }
185