1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx4c8-wasmsimd-mul32.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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
xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_mul32_ld128(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_mul32_ld128(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const uint8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 uint8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(mr != 0);
31 assert(mr <= 3);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(uint8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 kc = round_up_po2(kc, 8);
40 const uint8_t* a0 = a;
41 uint8_t* c0 = c;
42 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
43 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr < 2) {
45 a1 = a0;
46 c1 = c0;
47 }
48 const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
49 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 a2 = a1;
52 c2 = c1;
53 }
54
55 const v128_t vb_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.kernel_zero_point);
56 do {
57 v128_t vacc0x0 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[0]);
58 v128_t vacc0x1 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[1]);
59 v128_t vacc0x2 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[2]);
60 v128_t vacc0x3 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[3]);
61 v128_t vacc1x0 = vacc0x0;
62 v128_t vacc1x1 = vacc0x1;
63 v128_t vacc1x2 = vacc0x2;
64 v128_t vacc1x3 = vacc0x3;
65 v128_t vacc2x0 = vacc0x0;
66 v128_t vacc2x1 = vacc0x1;
67 v128_t vacc2x2 = vacc0x2;
68 v128_t vacc2x3 = vacc0x3;
69 w = (const void*) ((const int32_t*) w + 4);
70
71 size_t k = 0;
72 while (k < kc) {
73 const v128_t vxa0 = wasm_u16x8_load8x8(a0);
74 a0 += 8;
75 const v128_t vxa1 = wasm_u16x8_load8x8(a1);
76 a1 += 8;
77 const v128_t vxa2 = wasm_u16x8_load8x8(a2);
78 a2 += 8;
79
80 const v128_t vb01 = wasm_v128_load(w);
81 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_extend_low_u8x16(vb01), vb_zero_point);
82 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_extend_high_u8x16(vb01), vb_zero_point);
83
84 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa0), wasm_i32x4_extend_low_i16x8(vxb0)));
85 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa0), wasm_i32x4_extend_high_i16x8(vxb0)));
86 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa0), wasm_i32x4_extend_low_i16x8(vxb1)));
87 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa0), wasm_i32x4_extend_high_i16x8(vxb1)));
88 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa1), wasm_i32x4_extend_low_i16x8(vxb0)));
89 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa1), wasm_i32x4_extend_high_i16x8(vxb0)));
90 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa1), wasm_i32x4_extend_low_i16x8(vxb1)));
91 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa1), wasm_i32x4_extend_high_i16x8(vxb1)));
92 vacc2x0 = wasm_i32x4_add(vacc2x0, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa2), wasm_i32x4_extend_low_i16x8(vxb0)));
93 vacc2x0 = wasm_i32x4_add(vacc2x0, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa2), wasm_i32x4_extend_high_i16x8(vxb0)));
94 vacc2x1 = wasm_i32x4_add(vacc2x1, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa2), wasm_i32x4_extend_low_i16x8(vxb1)));
95 vacc2x1 = wasm_i32x4_add(vacc2x1, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa2), wasm_i32x4_extend_high_i16x8(vxb1)));
96 const v128_t vb23 = wasm_v128_load((const uint8_t*) w + 16);
97 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_extend_low_u8x16(vb23), vb_zero_point);
98 const v128_t vxb3 = wasm_i16x8_sub(wasm_u16x8_extend_high_u8x16(vb23), vb_zero_point);
99
100 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa0), wasm_i32x4_extend_low_i16x8(vxb2)));
101 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa0), wasm_i32x4_extend_high_i16x8(vxb2)));
102 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa0), wasm_i32x4_extend_low_i16x8(vxb3)));
103 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa0), wasm_i32x4_extend_high_i16x8(vxb3)));
104 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa1), wasm_i32x4_extend_low_i16x8(vxb2)));
105 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa1), wasm_i32x4_extend_high_i16x8(vxb2)));
106 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa1), wasm_i32x4_extend_low_i16x8(vxb3)));
107 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa1), wasm_i32x4_extend_high_i16x8(vxb3)));
108 vacc2x2 = wasm_i32x4_add(vacc2x2, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa2), wasm_i32x4_extend_low_i16x8(vxb2)));
109 vacc2x2 = wasm_i32x4_add(vacc2x2, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa2), wasm_i32x4_extend_high_i16x8(vxb2)));
110 vacc2x3 = wasm_i32x4_add(vacc2x3, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa2), wasm_i32x4_extend_low_i16x8(vxb3)));
111 vacc2x3 = wasm_i32x4_add(vacc2x3, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa2), wasm_i32x4_extend_high_i16x8(vxb3)));
112
113 w = (const void*) ((const uint8_t*) w + 32);
114 k += 8 * sizeof(uint8_t);
115 }
116
117 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));
118 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));
119 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));
120 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));
121 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));
122 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));
123
124 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
125 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
126 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x02, vacc2x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x02, vacc2x13, 2, 6, 3, 7));
127
128 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
129 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
130 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
131
132 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
133 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale);
134 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale);
135 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale);
136
137 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
138 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
139 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
140 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);
141
142 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
143 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
144 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
145 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vmagic_min);
146
147 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
148 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
149 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
150 vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);
151
152 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
153 v128_t vacc22x0123 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc2x0123);
154
155 v128_t vout = wasm_u8x16_narrow_i16x8(vacc01x0123, vacc22x0123);
156
157 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
158 vout = wasm_u8x16_min(vout, voutput_max);
159
160 if (nc >= 4) {
161 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
162 *((float*) c1) = (float) wasm_f32x4_extract_lane(vout, 1);
163 *((float*) c2) = (float) wasm_f32x4_extract_lane(vout, 2);
164
165 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
166 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
167 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
168
169 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
170 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
171 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
172
173 nc -= 4;
174 } else {
175 uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
176 uint32_t vout1 = wasm_i32x4_extract_lane(vout, 1);
177 uint32_t vout2 = wasm_i32x4_extract_lane(vout, 2);
178 if (nc & 2) {
179 *((uint16_t*) c0) = (uint16_t) vout0;
180 vout0 >>= 16;
181 c0 += 2;
182 *((uint16_t*) c1) = (uint16_t) vout1;
183 vout1 >>= 16;
184 c1 += 2;
185 *((uint16_t*) c2) = (uint16_t) vout2;
186 vout2 >>= 16;
187 c2 += 2;
188 }
189 if (nc & 1) {
190 *c0 = (uint8_t) vout0;
191 *c1 = (uint8_t) vout1;
192 *c2 = (uint8_t) vout2;
193 }
194
195 nc = 0;
196 }
197 } while (nc != 0);
198 }
199