Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-2x8c8-minmax-fp32-wasmsdot.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c8-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_2x8c8__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 <= 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 * 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
50 const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
51 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
52 const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point));
53 const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point);
54 const v128_t voutput_max = wasm_i8x16_splat(params->fp32_scalar.output_max);
55 XNN_FORCE_REALIZATION(vmagic_bias);
56 XNN_FORCE_REALIZATION(vmagic_min);
57 XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point);
58 XNN_FORCE_REALIZATION(voutput_max);
59
60 do {
61 v128_t vacc0x01 = wasm_u64x2_load32x2(w); w = (const int32_t*) w + 2;
62 v128_t vacc0x23 = wasm_u64x2_load32x2(w); w = (const int32_t*) w + 2;
63 v128_t vacc0x45 = wasm_u64x2_load32x2(w); w = (const int32_t*) w + 2;
64 v128_t vacc0x67 = wasm_u64x2_load32x2(w); w = (const int32_t*) w + 2;
65 v128_t vacc1x01 = vacc0x01;
66 v128_t vacc1x23 = vacc0x23;
67 v128_t vacc1x45 = vacc0x45;
68 v128_t vacc1x67 = vacc0x67;
69
70 size_t p = ks;
71 do {
72 const int8_t* restrict a0 = a[0];
73 if XNN_UNPREDICTABLE(a0 != zero) {
74 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
75 }
76 const int8_t* restrict a1 = a[1];
77 if XNN_UNPREDICTABLE(a1 != zero) {
78 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
79 }
80 a += 2;
81
82 size_t k = kc;
83 while (k != 0) {
84 const v128_t va0x01 = wasm_v128_load64_splat(a0);
85 a0 += 8;
86 const v128_t va1x01 = wasm_v128_load64_splat(a1);
87 a1 += 8;
88
89 const v128_t vb01x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
90 const v128_t vb23x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
91 const v128_t vb45x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
92 const v128_t vb67x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
93
94 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va0x01, vacc0x01);
95 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va0x01, vacc0x23);
96 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va0x01, vacc0x45);
97 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va0x01, vacc0x67);
98 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va1x01, vacc1x01);
99 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va1x01, vacc1x23);
100 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va1x01, vacc1x45);
101 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va1x01, vacc1x67);
102
103 k -= 8 * sizeof(int8_t);
104 }
105
106 p -= 2 * sizeof(void*);
107 } while (p != 0);
108
109 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x01, vacc0x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x01, vacc0x23, 1, 3, 5, 7));
110 v128_t vacc0x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x45, vacc0x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x45, vacc0x67, 1, 3, 5, 7));
111 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x01, vacc1x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x01, vacc1x23, 1, 3, 5, 7));
112 v128_t vacc1x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x45, vacc1x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x45, vacc1x67, 1, 3, 5, 7));
113
114 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
115 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
116 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
117 vacc1x4567 = wasm_f32x4_convert_i32x4(vacc1x4567);
118
119 const v128_t vscale0123 = wasm_v128_load(w);
120 w = (const float*) w + 4;
121 const v128_t vscale4567 = wasm_v128_load(w);
122 w = (const float*) w + 4;
123 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale0123);
124 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vscale4567);
125 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale0123);
126 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vscale4567);
127
128 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
129 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vmagic_bias);
130 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
131 vacc1x4567 = wasm_f32x4_add(vacc1x4567, vmagic_bias);
132
133 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
134 vacc0x4567 = wasm_i32x4_max(vacc0x4567, vmagic_min);
135 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
136 vacc1x4567 = wasm_i32x4_max(vacc1x4567, vmagic_min);
137
138 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
139 vacc0x4567 = wasm_i32x4_sub(vacc0x4567, vmagic_bias_less_output_zero_point);
140 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
141 vacc1x4567 = wasm_i32x4_sub(vacc1x4567, vmagic_bias_less_output_zero_point);
142
143 v128_t vacc0x01234567 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc0x4567);
144 v128_t vacc1x01234567 = wasm_i16x8_narrow_i32x4(vacc1x0123, vacc1x4567);
145
146 v128_t vout0x01234567_1x01234567 = wasm_i8x16_narrow_i16x8(vacc0x01234567, vacc1x01234567);
147
148 vout0x01234567_1x01234567 = wasm_i8x16_min(vout0x01234567_1x01234567, voutput_max);
149
150 if (nc >= 8) {
151 wasm_v128_store64_lane(c1, vout0x01234567_1x01234567, 1);
152 wasm_v128_store64_lane(c0, vout0x01234567_1x01234567, 0);
153
154 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
155 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
156
157 a = (const int8_t**restrict) ((uintptr_t) a - ks);
158
159 nc -= 8;
160 } else {
161 if (nc & 4) {
162 wasm_v128_store32_lane(c1, vout0x01234567_1x01234567, 2);
163 c1 += 4;
164 wasm_v128_store32_lane(c0, vout0x01234567_1x01234567, 0);
165 c0 += 4;
166
167 vout0x01234567_1x01234567 = wasm_u64x2_shr(vout0x01234567_1x01234567, 32);
168 }
169 if (nc & 2) {
170 wasm_v128_store16_lane(c1, vout0x01234567_1x01234567, 4);
171 c1 += 2;
172 wasm_v128_store16_lane(c0, vout0x01234567_1x01234567, 0);
173 c0 += 2;
174
175 vout0x01234567_1x01234567 = wasm_u32x4_shr(vout0x01234567_1x01234567, 16);
176 }
177 if (nc & 1) {
178 wasm_v128_store8_lane(c1, vout0x01234567_1x01234567, 8);
179 wasm_v128_store8_lane(c0, vout0x01234567_1x01234567, 0);
180 }
181
182 nc = 0;
183 }
184 } while (nc != 0);
185 }