1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c8-wasmsimd-mul16.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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_igemm_minmax_fp32_ukernel_2x4c8__wasmsimd_mul16_ld128(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_igemm_minmax_fp32_ukernel_2x4c8__wasmsimd_mul16_ld128(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 2);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (2 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 8);
44 int8_t* c0 = c;
45 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr != 2) {
47 c1 = c0;
48 }
49
50 do {
51 v128_t vacc0x0 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[0]);
52 v128_t vacc0x1 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[1]);
53 v128_t vacc0x2 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[2]);
54 v128_t vacc0x3 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[3]);
55 v128_t vacc1x0 = vacc0x0;
56 v128_t vacc1x1 = vacc0x1;
57 v128_t vacc1x2 = vacc0x2;
58 v128_t vacc1x3 = vacc0x3;
59 w = (const void*) ((const int32_t*) w + 4);
60
61 size_t p = ks;
62 do {
63 const int8_t* restrict a0 = a[0];
64 if XNN_UNPREDICTABLE(a0 != zero) {
65 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
66 }
67 const int8_t* restrict a1 = a[1];
68 if XNN_UNPREDICTABLE(a1 != zero) {
69 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
70 }
71 a += 2;
72
73 size_t k = 0;
74 while (k < kc) {
75 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
76 a0 += 8;
77 const v128_t vxa1 = wasm_i16x8_load8x8(a1);
78 a1 += 8;
79
80 const v128_t vb01 = wasm_v128_load(w);
81 const v128_t vxb0 = wasm_i16x8_extend_low_i8x16(vb01);
82 const v128_t vxb1 = wasm_i16x8_extend_high_i8x16(vb01);
83
84 const v128_t vprod0x0 = wasm_i16x8_mul(vxb0, vxa0);
85 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_extend_low_i16x8(vprod0x0));
86 const v128_t vprod1x0 = wasm_i16x8_mul(vxb0, vxa1);
87 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_extend_low_i16x8(vprod1x0));
88
89 const v128_t vprod0x1 = wasm_i16x8_mul(vxb1, vxa0);
90 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_extend_low_i16x8(vprod0x1));
91 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_extend_high_i16x8(vprod0x0));
92 const v128_t vprod1x1 = wasm_i16x8_mul(vxb1, vxa1);
93 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_extend_low_i16x8(vprod1x1));
94 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_extend_high_i16x8(vprod1x0));
95
96 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_extend_high_i16x8(vprod0x1));
97 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_extend_high_i16x8(vprod1x1));
98 const v128_t vb23 = wasm_v128_load((const int8_t*) w + 16);
99 const v128_t vxb2 = wasm_i16x8_extend_low_i8x16(vb23);
100 const v128_t vxb3 = wasm_i16x8_extend_high_i8x16(vb23);
101
102 const v128_t vprod0x2 = wasm_i16x8_mul(vxb2, vxa0);
103 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_extend_low_i16x8(vprod0x2));
104 const v128_t vprod1x2 = wasm_i16x8_mul(vxb2, vxa1);
105 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_extend_low_i16x8(vprod1x2));
106
107 const v128_t vprod0x3 = wasm_i16x8_mul(vxb3, vxa0);
108 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_extend_low_i16x8(vprod0x3));
109 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_extend_high_i16x8(vprod0x2));
110 const v128_t vprod1x3 = wasm_i16x8_mul(vxb3, vxa1);
111 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_extend_low_i16x8(vprod1x3));
112 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_extend_high_i16x8(vprod1x2));
113
114 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_extend_high_i16x8(vprod0x3));
115 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_extend_high_i16x8(vprod1x3));
116
117 w = (const void*) ((const int8_t*) w + 32);
118 k += 8 * sizeof(int8_t);
119 }
120 p -= 2 * sizeof(void*);
121 } while (p != 0);
122
123 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));
124 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));
125 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));
126 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));
127
128 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
129 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
130
131 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
132 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
133
134 const v128_t vscale0123 = wasm_v128_load(w);
135 w = (const void*) ((const float*) w + 4);
136 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
137 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale0123);
138
139 const v128_t vmagic_bias = wasm_v128_load64_splat(params->wasmsimd.magic_bias);
140 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
141 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
142
143 const v128_t vmagic_min = wasm_v128_load64_splat(params->wasmsimd.magic_min);
144 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
145 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
146
147 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->wasmsimd.magic_bias_less_output_zero_point);
148 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
149 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
150
151 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
152
153 v128_t vout = wasm_i8x16_narrow_i16x8(vacc01x0123, vacc01x0123);
154
155 const v128_t voutput_max = wasm_v128_load64_splat(params->wasmsimd.output_max);
156 vout = wasm_i8x16_min(vout, voutput_max);
157
158 if (nc >= 4) {
159 *((float*) c1) = (float) wasm_f32x4_extract_lane(vout, 1);
160 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
161
162 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
163 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
164
165 a = (const int8_t**restrict) ((uintptr_t) a - ks);
166
167 nc -= 4;
168 } else {
169 uint32_t vout1 = wasm_i32x4_extract_lane(vout, 1);
170 uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
171 if (nc & 2) {
172 *((uint16_t*) c1) = (uint16_t) vout1;
173 vout1 >>= 16;
174 c1 += 2;
175 *((uint16_t*) c0) = (uint16_t) vout0;
176 vout0 >>= 16;
177 c0 += 2;
178 }
179 if (nc & 1) {
180 *c1 = (int8_t) vout1;
181 *c0 = (int8_t) vout0;
182 }
183
184 nc = 0;
185 }
186 } while (nc != 0);
187 }
188