Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-1x8c8-minmax-wasmusdot-u2.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_qd8_f32_qc8w_igemm_minmax_ukernel_1x8c8__wasmusdot_u2(
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 float* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const int8_t* zero_data,
31 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 1);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (1 * sizeof(void*)) == 0);
40 assert(a_offset % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 8 * sizeof(int8_t));
46 float* c0 = c;
47
48 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
49 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
50 XNN_FORCE_REALIZATION(vmin);
51 XNN_FORCE_REALIZATION(vmax);
52
53 const v128_t vsign_mask = wasm_u8x16_const_splat(UINT8_C(0x80));
54 XNN_FORCE_REALIZATION(vsign_mask);
55 do {
56 const v128_t vinput_zero_point = wasm_i32x4_splat((int32_t) quantization_params->zero_point + 128);
57 v128_t vksum0123 = wasm_v128_load(w); w = (const int32_t*) w + 4;
58 v128_t vsum0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
59 v128_t vacc0x01 = wasm_u64x2_extend_low_u32x4(vsum0x0123);
60 v128_t vacc0x23 = wasm_u64x2_extend_high_u32x4(vsum0x0123);
61 v128_t vksum4567 = wasm_v128_load(w); w = (const int32_t*) w + 4;
62 v128_t vsum0x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point);
63 v128_t vacc0x45 = wasm_u64x2_extend_low_u32x4(vsum0x4567);
64 v128_t vacc0x67 = wasm_u64x2_extend_high_u32x4(vsum0x4567);
65
66 size_t p = ks;
67 do {
68 const int8_t* restrict a0 = a[0];
69 if XNN_UNPREDICTABLE(a0 != zero) {
70 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
71 } else {
72 a0 = zero_data;
73 }
74 a += 1;
75
76 size_t k = kc;
77 while (k >= 16 * sizeof(int8_t)) {
78 const v128_t va0x01 = wasm_v128_xor(wasm_v128_load64_splat(a0), vsign_mask);
79 const v128_t va0x23 = wasm_v128_xor(wasm_v128_load64_splat((const int8_t*) a0 + 8), vsign_mask);
80 a0 += 16;
81
82 const v128_t vb01x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
83 const v128_t vb23x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
84 const v128_t vb45x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
85 const v128_t vb67x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
86 const v128_t vb01x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
87 const v128_t vb23x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
88 const v128_t vb45x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
89 const v128_t vb67x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
90
91 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va0x01, vacc0x01);
92 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va0x01, vacc0x23);
93 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va0x01, vacc0x45);
94 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va0x01, vacc0x67);
95 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va0x23, vacc0x01);
96 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va0x23, vacc0x23);
97 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va0x23, vacc0x45);
98 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va0x23, vacc0x67);
99
100 k -= 16 * sizeof(int8_t);
101 }
102
103 if (k != 0) {
104 const v128_t va0x01 = wasm_v128_xor(wasm_v128_load64_splat(a0), vsign_mask);
105 a0 += 8;
106
107 const v128_t vb01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
108 const v128_t vb23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
109 const v128_t vb45 = wasm_v128_load(w); w = (const int8_t*) w + 16;
110 const v128_t vb67 = wasm_v128_load(w); w = (const int8_t*) w + 16;
111
112 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va0x01, vacc0x01);
113 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va0x01, vacc0x23);
114 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va0x01, vacc0x45);
115 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va0x01, vacc0x67);
116
117 k -= 8 * sizeof(int8_t);
118 }
119 p -= 1 * sizeof(void*);
120 } while (p != 0);
121
122 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x01, vacc0x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x01, vacc0x23, 1, 3, 5, 7));
123 v128_t vacc0x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x45, vacc0x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x45, vacc0x67, 1, 3, 5, 7));
124
125 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
126 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
127
128 const v128_t vinput_scale = wasm_v128_load32_splat(&quantization_params->inv_scale);
129
130 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale);
131 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vinput_scale);
132
133 const v128_t vfilter_output_scale0123 = wasm_v128_load(w); w = (const float*) w + 4;
134 const v128_t vfilter_output_scale4567 = wasm_v128_load(w); w = (const float*) w + 4;
135
136 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
137 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vfilter_output_scale4567);
138
139 const v128_t vbias0123 = wasm_v128_load(w); w = (const float*) w + 4;
140 const v128_t vbias4567 = wasm_v128_load(w); w = (const float*) w + 4;
141 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
142 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vbias4567);
143
144 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
145 vacc0x4567 = wasm_f32x4_pmax(vacc0x4567, vmin);
146
147 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
148 vacc0x4567 = wasm_f32x4_pmin(vacc0x4567, vmax);
149
150 if XNN_LIKELY(nc >= 8) {
151 wasm_v128_store(c0, vacc0x0123);
152 wasm_v128_store(c0 + 4, vacc0x4567);
153
154 c0 = (float*) ((uintptr_t) c0 + cn_stride);
155
156 a = (const int8_t**restrict) ((uintptr_t) a - ks);
157
158 nc -= 8;
159 } else {
160 if (nc & 4) {
161 wasm_v128_store(c0, vacc0x0123);
162 vacc0x0123 = vacc0x4567;
163 c0 += 4;
164 }
165 if (nc & 2) {
166 wasm_v128_store64_lane(c0, vacc0x0123, 0);
167 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
168 c0 += 2;
169 }
170 if (nc & 1) {
171 wasm_v128_store32_lane(c0, vacc0x0123, 0);
172 }
173 nc = 0;
174 }
175 } while (nc != 0);
176 }