Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-3x4c8-minmax-fp32-wasmsimd-dot16x2-ld64.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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/common.h"
15 #include "xnnpack/gemm.h"
16 #include "xnnpack/math.h"
17
18
19 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_3x4c8__wasmsimd_dot16x2_ld64(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 size_t ks,
24 const int8_t** restrict a,
25 const void* restrict w,
26 int8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const int8_t* zero,
31 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 3);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(ks != 0);
38 assert(ks % (3 * sizeof(void*)) == 0);
39 assert(a_offset % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 8 * sizeof(int8_t));
45 int8_t* c0 = c;
46 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 c1 = c0;
49 }
50 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 c2 = c1;
53 }
54
55
56 const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
57 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
58 const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point));
59 const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point);
60 const v128_t voutput_max = wasm_i8x16_splat(params->fp32_scalar.output_max);
61 XNN_FORCE_REALIZATION(vmagic_bias);
62 XNN_FORCE_REALIZATION(vmagic_min);
63 XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point);
64 XNN_FORCE_REALIZATION(voutput_max);
65 do {
66 v128_t vacc0x0 = wasm_v128_load32_zero(w);
67 v128_t vacc0x1 = wasm_v128_load32_zero((const int32_t*) w + 1);
68 v128_t vacc0x2 = wasm_v128_load32_zero((const int32_t*) w + 2);
69 v128_t vacc0x3 = wasm_v128_load32_zero((const int32_t*) w + 3);
70 v128_t vacc1x0 = vacc0x0;
71 v128_t vacc1x1 = vacc0x1;
72 v128_t vacc1x2 = vacc0x2;
73 v128_t vacc1x3 = vacc0x3;
74 v128_t vacc2x0 = vacc0x0;
75 v128_t vacc2x1 = vacc0x1;
76 v128_t vacc2x2 = vacc0x2;
77 v128_t vacc2x3 = vacc0x3;
78 w = (const void*) ((const int32_t*) w + 4);
79
80 size_t p = ks;
81 do {
82 const int8_t* restrict a0 = a[0];
83 if XNN_UNPREDICTABLE(a0 != zero) {
84 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
85 }
86 const int8_t* restrict a1 = a[1];
87 if XNN_UNPREDICTABLE(a1 != zero) {
88 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
89 }
90 const int8_t* restrict a2 = a[2];
91 if XNN_UNPREDICTABLE(a2 != zero) {
92 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
93 }
94 a += 3;
95
96 size_t k = kc;
97 do {
98 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
99 a0 += 8;
100 const v128_t vxa1 = wasm_i16x8_load8x8(a1);
101 a1 += 8;
102 const v128_t vxa2 = wasm_i16x8_load8x8(a2);
103 a2 += 8;
104
105 const v128_t vxb0 = wasm_i16x8_load8x8(w);
106
107 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_dot_i16x8(vxa0, vxb0));
108 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_dot_i16x8(vxa1, vxb0));
109 vacc2x0 = wasm_i32x4_add(vacc2x0, wasm_i32x4_dot_i16x8(vxa2, vxb0));
110 const v128_t vxb1 = wasm_i16x8_load8x8((const int8_t*) w + 8);
111
112 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_dot_i16x8(vxa0, vxb1));
113 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_dot_i16x8(vxa1, vxb1));
114 vacc2x1 = wasm_i32x4_add(vacc2x1, wasm_i32x4_dot_i16x8(vxa2, vxb1));
115 const v128_t vxb2 = wasm_i16x8_load8x8((const int8_t*) w + 16);
116
117 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_dot_i16x8(vxa0, vxb2));
118 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_dot_i16x8(vxa1, vxb2));
119 vacc2x2 = wasm_i32x4_add(vacc2x2, wasm_i32x4_dot_i16x8(vxa2, vxb2));
120 const v128_t vxb3 = wasm_i16x8_load8x8((const int8_t*) w + 24);
121
122 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_dot_i16x8(vxa0, vxb3));
123 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_dot_i16x8(vxa1, vxb3));
124 vacc2x3 = wasm_i32x4_add(vacc2x3, wasm_i32x4_dot_i16x8(vxa2, vxb3));
125
126 w = (const void*) ((const int8_t*) w + 32);
127 k -= 8 * sizeof(int8_t);
128 } while (k != 0);
129 p -= 3 * sizeof(void*);
130 } while (p != 0);
131
132 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));
133 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));
134 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));
135 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));
136 const v128_t vacc2x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x0, vacc2x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x0, vacc2x2, 2, 6, 3, 7));
137 const v128_t vacc2x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x1, vacc2x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x1, vacc2x3, 2, 6, 3, 7));
138
139 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
140 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
141 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x02, vacc2x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x02, vacc2x13, 2, 6, 3, 7));
142
143 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
144 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
145 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
146
147
148 const v128_t vscale0123 = wasm_v128_load(w);
149 w = (const float*) w + 4;
150 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
151 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale0123);
152 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale0123);
153
154 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
155 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
156 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);
157
158 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
159 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
160 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vmagic_min);
161
162 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
163 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
164 vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);
165
166 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
167 v128_t vacc22x0123 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc2x0123);
168
169 v128_t vout = wasm_i8x16_narrow_i16x8(vacc01x0123, vacc22x0123);
170
171 vout = wasm_i8x16_min(vout, voutput_max);
172
173 if (nc >= 4) {
174 wasm_v128_store32_lane(c2, vout, 2);
175 wasm_v128_store32_lane(c1, vout, 1);
176 wasm_v128_store32_lane(c0, vout, 0);
177
178 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
179 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
180 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
181
182 a = (const int8_t**restrict) ((uintptr_t) a - ks);
183
184 nc -= 4;
185 } else {
186 if (nc & 2) {
187 wasm_v128_store16_lane(c2, vout, 4);
188 c2 += 2;
189 wasm_v128_store16_lane(c1, vout, 2);
190 c1 += 2;
191 wasm_v128_store16_lane(c0, vout, 0);
192 c0 += 2;
193
194 vout = wasm_u32x4_shr(vout, 16);
195 }
196 if (nc & 1) {
197 wasm_v128_store8_lane(c2, vout, 8);
198 wasm_v128_store8_lane(c1, vout, 4);
199 wasm_v128_store8_lane(c0, vout, 0);
200 }
201
202 nc = 0;
203 }
204 } while (nc != 0);
205 }