1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx4c8-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
xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_dot16x2_ld128(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_dot16x2_ld128(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(mr != 0);
31 assert(mr <= 2);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(int8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 kc = round_up_po2(kc, 8);
40 const int8_t* a0 = a;
41 int8_t* c0 = c;
42 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
43 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr != 2) {
45 a1 = a0;
46 c1 = c0;
47 }
48
49 do {
50 v128_t vacc0x0 = wasm_v128_load32_zero(w);
51 v128_t vacc0x1 = wasm_v128_load32_zero((const int32_t*) w + 1);
52 v128_t vacc0x2 = wasm_v128_load32_zero((const int32_t*) w + 2);
53 v128_t vacc0x3 = wasm_v128_load32_zero((const int32_t*) w + 3);
54 v128_t vacc1x0 = vacc0x0;
55 v128_t vacc1x1 = vacc0x1;
56 v128_t vacc1x2 = vacc0x2;
57 v128_t vacc1x3 = vacc0x3;
58 w = (const void*) ((const int32_t*) w + 4);
59
60 size_t k = 0;
61 while (k < kc) {
62 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
63 a0 += 8;
64 const v128_t vxa1 = wasm_i16x8_load8x8(a1);
65 a1 += 8;
66
67 const v128_t vb01 = wasm_v128_load(w);
68 const v128_t vxb0 = wasm_i16x8_extend_low_i8x16(vb01);
69 const v128_t vxb1 = wasm_i16x8_extend_high_i8x16(vb01);
70
71 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_dot_i16x8(vxa0, vxb0));
72 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_dot_i16x8(vxa0, vxb1));
73 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_dot_i16x8(vxa1, vxb0));
74 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_dot_i16x8(vxa1, vxb1));
75 const v128_t vb23 = wasm_v128_load((const int8_t*) w + 16);
76 const v128_t vxb2 = wasm_i16x8_extend_low_i8x16(vb23);
77 const v128_t vxb3 = wasm_i16x8_extend_high_i8x16(vb23);
78
79 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_dot_i16x8(vxa0, vxb2));
80 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_dot_i16x8(vxa0, vxb3));
81 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_dot_i16x8(vxa1, vxb2));
82 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_dot_i16x8(vxa1, vxb3));
83
84 w = (const void*) ((const int8_t*) w + 32);
85 k += 8 * sizeof(int8_t);
86 }
87
88 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));
89 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));
90 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));
91 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));
92
93 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
94 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
95
96 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
97 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
98
99 const v128_t vscale0123 = wasm_v128_load(w);
100 w = (const void*) ((const float*) w + 4);
101 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
102 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale0123);
103
104 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
105 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
106 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
107
108 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
109 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
110 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
111
112 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
113 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
114 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
115
116 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
117
118 v128_t vout = wasm_i8x16_narrow_i16x8(vacc01x0123, vacc01x0123);
119
120 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
121 vout = wasm_i8x16_min(vout, voutput_max);
122
123 if (nc >= 4) {
124 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
125 *((float*) c1) = (float) wasm_f32x4_extract_lane(vout, 1);
126
127 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
128 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
129
130 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
131 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
132
133 nc -= 4;
134 } else {
135 uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
136 uint32_t vout1 = wasm_i32x4_extract_lane(vout, 1);
137 if (nc & 2) {
138 *((uint16_t*) c0) = (uint16_t) vout0;
139 vout0 >>= 16;
140 c0 += 2;
141 *((uint16_t*) c1) = (uint16_t) vout1;
142 vout1 >>= 16;
143 c1 += 2;
144 }
145 if (nc & 1) {
146 *c0 = (int8_t) vout0;
147 *c1 = (int8_t) vout1;
148 }
149
150 nc = 0;
151 }
152 } while (nc != 0);
153 }
154