1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx4c2s4-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_3x4c2s4__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_3x4c2s4__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 <= 3);
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 const uint8_t* a0 = a;
41 uint8_t* c0 = c;
42 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
43 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr < 2) {
45 a1 = a0;
46 c1 = c0;
47 }
48 const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
49 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 a2 = a1;
52 c2 = c1;
53 }
54
55 kc = round_up_po2(kc, 8 * sizeof(uint8_t));
56 do {
57 v128_t vacc0x0123 = wasm_v128_load(w);
58 v128_t vacc1x0123 = vacc0x0123;
59 v128_t vacc2x0123 = vacc0x0123;
60 w = (const void*) ((const int32_t*) w + 4);
61
62 const v128_t vb_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.kernel_zero_point);
63 size_t k = kc;
64 do {
65 v128_t vxa0 = wasm_u16x8_load8x8((const v128_t*) a0);
66 a0 += 8;
67 v128_t vxa1 = wasm_u16x8_load8x8((const v128_t*) a1);
68 a1 += 8;
69 v128_t vxa2 = wasm_u16x8_load8x8((const v128_t*) a2);
70 a2 += 8;
71
72 const v128_t vb01 = wasm_v128_load(w);
73 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_extend_low_u8x16(vb01), vb_zero_point);
74 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_extend_high_u8x16(vb01), vb_zero_point);
75
76 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb0));
77 vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
78 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb0));
79 vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
80 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb0));
81 vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
82
83 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb1));
84 vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
85 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb1));
86 vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
87 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb1));
88 vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
89 const v128_t vb23 = wasm_v128_load((const uint8_t*) w + 16);
90 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_extend_low_u8x16(vb23), vb_zero_point);
91 const v128_t vxb3 = wasm_i16x8_sub(wasm_u16x8_extend_high_u8x16(vb23), vb_zero_point);
92
93 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb2));
94 vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
95 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb2));
96 vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
97 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb2));
98 vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
99
100 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb3));
101 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb3));
102 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb3));
103
104 w = (const uint8_t*) w + 32;
105 k -= 8 * sizeof(uint8_t);
106 } while (k != 0);
107
108 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
109 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
110 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
111
112 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
113 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale);
114 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale);
115 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale);
116
117 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
118 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
119 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
120 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);
121
122 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
123 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
124 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
125 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vmagic_min);
126
127 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
128 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
129 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
130 vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);
131
132 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
133 v128_t vacc22x0123 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc2x0123);
134
135 v128_t vout = wasm_u8x16_narrow_i16x8(vacc01x0123, vacc22x0123);
136
137 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
138 vout = wasm_u8x16_min(vout, voutput_max);
139
140 if (nc >= 4) {
141 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
142 *((float*) c1) = (float) wasm_f32x4_extract_lane(vout, 1);
143 *((float*) c2) = (float) wasm_f32x4_extract_lane(vout, 2);
144
145 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
146 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
147 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
148
149 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
150 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
151 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
152
153 nc -= 4;
154 } else {
155 uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
156 uint32_t vout1 = wasm_i32x4_extract_lane(vout, 1);
157 uint32_t vout2 = wasm_i32x4_extract_lane(vout, 2);
158 if (nc & 2) {
159 *((uint16_t*) c0) = (uint16_t) vout0;
160 vout0 >>= 16;
161 c0 += 2;
162 *((uint16_t*) c1) = (uint16_t) vout1;
163 vout1 >>= 16;
164 c1 += 2;
165 *((uint16_t*) c2) = (uint16_t) vout2;
166 vout2 >>= 16;
167 c2 += 2;
168 }
169 if (nc & 1) {
170 *c0 = (uint8_t) vout0;
171 *c1 = (uint8_t) vout1;
172 *c2 = (uint8_t) vout2;
173 }
174
175 nc = 0;
176 }
177 } while (nc != 0);
178 }
179