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_1x4c2__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_1x4c2__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 <= 1);
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
44 do {
45 v128_t vacc0x0123 = wasm_v128_load(w);
46 w = (const void*) ((const int32_t*) w + 4);
47
48 size_t k = kc;
49 const v128_t vb_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.kernel_zero_point);
50 while (k >= 8 * sizeof(uint8_t)) {
51 const v128_t vxa0 = wasm_u16x8_load8x8((const v128_t*) a0);
52 a0 += 8;
53
54 const v128_t vb01 = wasm_v128_load(w);
55 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_extend_low_u8x16(vb01), vb_zero_point);
56 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_extend_high_u8x16(vb01), vb_zero_point);
57
58 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
59 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
60
61 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
62 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
63 const v128_t vb23 = wasm_v128_load((const uint8_t*) w + 16);
64 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_extend_low_u8x16(vb23), vb_zero_point);
65 const v128_t vxb3 = wasm_i16x8_sub(wasm_u16x8_extend_high_u8x16(vb23), vb_zero_point);
66
67 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
68 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
69
70 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
71 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 3, 3, 3, 3), vxb3));
72
73 w = (const void*) ((const uint8_t*) w + 32);
74 k -= 8 * sizeof(uint8_t);
75 }
76 if (k != 0) {
77 const v128_t vxa0 = wasm_u16x8_load8x8(a0);
78 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
79
80 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
81 w = (const void*) ((const uint8_t*) w + 8);
82
83 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
84 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
85
86 if (k > 2 * sizeof(uint8_t)) {
87 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
88 w = (const void*) ((const uint8_t*) w + 8);
89
90 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
91 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
92
93 if (k > 4 * sizeof(uint8_t)) {
94 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
95 w = (const void*) ((const uint8_t*) w + 8);
96
97 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
98 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
99 }
100 }
101 }
102
103 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
104
105 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
106 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale);
107
108 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
109 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
110
111 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
112 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
113
114 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
115 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
116
117 v128_t vacc00x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc0x0123);
118
119 v128_t vout = wasm_u8x16_narrow_i16x8(vacc00x0123, vacc00x0123);
120
121 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
122 vout = wasm_u8x16_min(vout, voutput_max);
123
124 if (nc >= 4) {
125 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
126
127 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
128
129 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
130
131 nc -= 4;
132 } else {
133 uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
134 if (nc & 2) {
135 *((uint16_t*) c0) = (uint16_t) vout0;
136 vout0 >>= 16;
137 c0 += 2;
138 }
139 if (nc & 1) {
140 *c0 = (uint8_t) vout0;
141 }
142
143 nc = 0;
144 }
145 } while (nc != 0);
146 }
147