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_qs8_gemm_minmax_fp32_ukernel_1x4c8__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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(mr != 0);
31 assert(mr <= 1);
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
43 do {
44 v128_t vacc0x0 = wasm_v128_load32_zero(w);
45 v128_t vacc0x1 = wasm_v128_load32_zero((const int32_t*) w + 1);
46 v128_t vacc0x2 = wasm_v128_load32_zero((const int32_t*) w + 2);
47 v128_t vacc0x3 = wasm_v128_load32_zero((const int32_t*) w + 3);
48 w = (const void*) ((const int32_t*) w + 4);
49
50 size_t k = 0;
51 while (k < kc) {
52 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
53 a0 += 8;
54
55 const v128_t vb01 = wasm_v128_load(w);
56 const v128_t vxb0 = wasm_i16x8_extend_low_i8x16(vb01);
57 const v128_t vxb1 = wasm_i16x8_extend_high_i8x16(vb01);
58
59 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_dot_i16x8(vxa0, vxb0));
60 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_dot_i16x8(vxa0, vxb1));
61 const v128_t vb23 = wasm_v128_load((const int8_t*) w + 16);
62 const v128_t vxb2 = wasm_i16x8_extend_low_i8x16(vb23);
63 const v128_t vxb3 = wasm_i16x8_extend_high_i8x16(vb23);
64
65 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_dot_i16x8(vxa0, vxb2));
66 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_dot_i16x8(vxa0, vxb3));
67
68 w = (const void*) ((const int8_t*) w + 32);
69 k += 8 * sizeof(int8_t);
70 }
71
72 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));
73 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));
74
75 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
76
77 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
78
79 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
80 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale);
81
82 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
83 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
84
85 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
86 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
87
88 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
89 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
90
91 v128_t vacc00x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc0x0123);
92
93 v128_t vout = wasm_i8x16_narrow_i16x8(vacc00x0123, vacc00x0123);
94
95 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
96 vout = wasm_i8x16_min(vout, voutput_max);
97
98 if (nc >= 4) {
99 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
100
101 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
102
103 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
104
105 nc -= 4;
106 } else {
107 uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
108 if (nc & 2) {
109 *((uint16_t*) c0) = (uint16_t) vout0;
110 vout0 >>= 16;
111 c0 += 2;
112 }
113 if (nc & 1) {
114 *c0 = (int8_t) vout0;
115 }
116
117 nc = 0;
118 }
119 } while (nc != 0);
120 }
121