Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-3x4c16-minmax-fp32-wasmsdot.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c16-wasmdot.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 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 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_3x4c16__wasmsdot(
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_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 3);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (3 * 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, 16 * sizeof(int8_t));
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 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53
54 const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
55 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
56 const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point));
57 const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point);
58 const v128_t voutput_max = wasm_i8x16_splat(params->fp32_scalar.output_max);
59 XNN_FORCE_REALIZATION(vmagic_bias);
60 XNN_FORCE_REALIZATION(vmagic_min);
61 XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point);
62 XNN_FORCE_REALIZATION(voutput_max);
63
64 do {
65 v128_t vacc0x0 = wasm_v128_load32_zero(w);
66 v128_t vacc0x1 = wasm_v128_load32_zero((const int32_t*) w + 1);
67 v128_t vacc0x2 = wasm_v128_load32_zero((const int32_t*) w + 2);
68 v128_t vacc0x3 = wasm_v128_load32_zero((const int32_t*) w + 3);
69 v128_t vacc1x0 = vacc0x0;
70 v128_t vacc1x1 = vacc0x1;
71 v128_t vacc1x2 = vacc0x2;
72 v128_t vacc1x3 = vacc0x3;
73 v128_t vacc2x0 = vacc0x0;
74 v128_t vacc2x1 = vacc0x1;
75 v128_t vacc2x2 = vacc0x2;
76 v128_t vacc2x3 = vacc0x3;
77 w = (const int32_t*) w + 4;
78
79 size_t p = ks;
80 do {
81 const int8_t* restrict a0 = a[0];
82 if XNN_UNPREDICTABLE(a0 != zero) {
83 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
84 }
85 const int8_t* restrict a1 = a[1];
86 if XNN_UNPREDICTABLE(a1 != zero) {
87 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
88 }
89 const int8_t* restrict a2 = a[2];
90 if XNN_UNPREDICTABLE(a2 != zero) {
91 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
92 }
93 a += 3;
94
95 size_t k = kc;
96 do {
97 const v128_t va0 = wasm_v128_load(a0);
98 a0 += 16;
99 const v128_t va1 = wasm_v128_load(a1);
100 a1 += 16;
101 const v128_t va2 = wasm_v128_load(a2);
102 a2 += 16;
103
104 const v128_t vb0 = wasm_v128_load(w);
105 vacc0x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va0, vacc0x0);
106 vacc1x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va1, vacc1x0);
107 vacc2x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va2, vacc2x0);
108 const v128_t vb1 = wasm_v128_load((const int8_t*) w + 16);
109 vacc0x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va0, vacc0x1);
110 vacc1x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va1, vacc1x1);
111 vacc2x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va2, vacc2x1);
112 const v128_t vb2 = wasm_v128_load((const int8_t*) w + 32);
113 vacc0x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va0, vacc0x2);
114 vacc1x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va1, vacc1x2);
115 vacc2x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va2, vacc2x2);
116 const v128_t vb3 = wasm_v128_load((const int8_t*) w + 48);
117 vacc0x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va0, vacc0x3);
118 vacc1x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va1, vacc1x3);
119 vacc2x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va2, vacc2x3);
120
121 w = (const int8_t*) w + 64;
122 k -= 16 * sizeof(int8_t);
123 } while (k != 0);
124 p -= 3 * sizeof(void*);
125 } while (p != 0);
126
127 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));
128 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));
129 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));
130 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));
131 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));
132 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));
133
134 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
135 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
136 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x02, vacc2x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x02, vacc2x13, 2, 6, 3, 7));
137
138 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
139 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
140 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
141
142 const v128_t vscale0123 = wasm_v128_load(w);
143 w = (const float*) w + 4;
144 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
145 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale0123);
146 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale0123);
147
148 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
149 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
150 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);
151
152 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
153 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
154 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vmagic_min);
155
156 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
157 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
158 vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);
159
160 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
161 v128_t vacc22x0123 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc2x0123);
162
163 v128_t vout = wasm_i8x16_narrow_i16x8(vacc01x0123, vacc22x0123);
164
165 vout = wasm_i8x16_min(vout, voutput_max);
166
167 if (nc >= 4) {
168 wasm_v128_store32_lane(c2, vout, 2);
169 wasm_v128_store32_lane(c1, vout, 1);
170 wasm_v128_store32_lane(c0, vout, 0);
171
172 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
173 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
174 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
175
176 a = (const int8_t**restrict) ((uintptr_t) a - ks);
177
178 nc -= 4;
179 } else {
180 if (nc & 2) {
181 wasm_v128_store16_lane(c2, vout, 4);
182 c2 += 2;
183 wasm_v128_store16_lane(c1, vout, 2);
184 c1 += 2;
185 wasm_v128_store16_lane(c0, vout, 0);
186 c0 += 2;
187
188 vout = wasm_u32x4_shr(vout, 16);
189 }
190 if (nc & 1) {
191 wasm_v128_store8_lane(c2, vout, 8);
192 wasm_v128_store8_lane(c1, vout, 4);
193 wasm_v128_store8_lane(c0, vout, 0);
194 }
195
196 nc = 0;
197 }
198 } while (nc != 0);
199 }