Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-3x4c2s4-minmax-wasmsimd-dot16x2-ld128.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx4c2s4-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_qd8_f32_qc8w_gemm_minmax_ukernel_3x4c2s4__wasmsimd_dot16x2_ld128(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const int8_t* restrict a,
24 size_t a_stride,
25 const void* restrict w,
26 float* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
30 const struct xnn_qd8_quantization_params quantization_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(kc % sizeof(int8_t) == 0);
37 assert(a != NULL);
38 assert(w != NULL);
39 assert(c != NULL);
40
41 const int8_t* a0 = a;
42 float* c0 = c;
43 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
44 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
50 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 a2 = a1;
53 c2 = c1;
54 }
55
56 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
57 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
58 XNN_FORCE_REALIZATION(vmin);
59 XNN_FORCE_REALIZATION(vmax);
60
61 kc = round_up_po2(kc, 8 * sizeof(int8_t));
62 do {
63 const v128_t vksum0123 = wasm_v128_load(w);
64 const v128_t vinput_zero_point0 = wasm_v128_load32_splat(&quantization_params[0].zero_point);
65 const v128_t vinput_zero_point1 = wasm_v128_load32_splat(&quantization_params[1].zero_point);
66 const v128_t vinput_zero_point2 = wasm_v128_load32_splat(&quantization_params[2].zero_point);
67 v128_t vacc0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point0);
68 v128_t vacc1x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point1);
69 v128_t vacc2x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point2);
70 w = (const int32_t*) w + 4;
71
72 size_t k = kc;
73 do {
74 v128_t vxa0 = wasm_i16x8_load8x8((const v128_t*) a0);
75 a0 += 8;
76 v128_t vxa1 = wasm_i16x8_load8x8((const v128_t*) a1);
77 a1 += 8;
78 v128_t vxa2 = wasm_i16x8_load8x8((const v128_t*) a2);
79 a2 += 8;
80
81 const v128_t vb01 = wasm_v128_load(w);
82 const v128_t vxb0 = wasm_i16x8_extend_low_i8x16(vb01);
83 const v128_t vxb1 = wasm_i16x8_extend_high_i8x16(vb01);
84
85 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb0));
86 vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
87 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb0));
88 vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
89 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb0));
90 vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
91
92 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb1));
93 vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
94 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb1));
95 vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
96 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb1));
97 vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
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 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb2));
103 vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
104 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb2));
105 vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
106 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb2));
107 vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
108
109 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb3));
110 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb3));
111 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb3));
112
113 w = (const int8_t*) w + 32;
114 k -= 8 * sizeof(int8_t);
115 } while (k != 0);
116
117 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
118 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
119 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
120
121 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
122 const v128_t vinput_scale1 = wasm_v128_load32_splat(&quantization_params[1].inv_scale);
123 const v128_t vinput_scale2 = wasm_v128_load32_splat(&quantization_params[2].inv_scale);
124
125 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0);
126 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale1);
127 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale2);
128
129 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
130 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
131 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
132 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123);
133 w = (const float*) w + 4;
134
135 const v128_t vbias0123 = wasm_v128_load(w);
136 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
137 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
138 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123);
139 w = (const float*) w + 4;
140
141 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
142 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
143 vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin);
144
145 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
146 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
147 vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax);
148
149 if XNN_LIKELY(nc >= 4) {
150 wasm_v128_store(c0, vacc0x0123);
151 wasm_v128_store(c1, vacc1x0123);
152 wasm_v128_store(c2, vacc2x0123);
153
154 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
155 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
156 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
157
158 c0 = (float*) ((uintptr_t) c0 + cn_stride);
159 c1 = (float*) ((uintptr_t) c1 + cn_stride);
160 c2 = (float*) ((uintptr_t) c2 + cn_stride);
161
162 nc -= 4;
163 } else {
164 if (nc & 2) {
165 wasm_v128_store64_lane(c0, vacc0x0123, 0);
166 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
167 c0 += 2;
168 wasm_v128_store64_lane(c1, vacc1x0123, 0);
169 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
170 c1 += 2;
171 wasm_v128_store64_lane(c2, vacc2x0123, 0);
172 vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1);
173 c2 += 2;
174 }
175 if (nc & 1) {
176 wasm_v128_store32_lane(c0, vacc0x0123, 0);
177 wasm_v128_store32_lane(c1, vacc1x0123, 0);
178 wasm_v128_store32_lane(c2, vacc2x0123, 0);
179 }
180 nc = 0;
181 }
182 } while (nc != 0);
183 }