1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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_igemm_minmax_fp32_ukernel_2x4c8__wasmsimd_mul32_ld128(size_t mr,size_t nc,size_t kc,size_t ks,const uint8_t ** restrict a,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_igemm_minmax_fp32_ukernel_2x4c8__wasmsimd_mul32_ld128(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const uint8_t** restrict a,
24 const void* restrict w,
25 uint8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const uint8_t* zero,
30 const union xnn_qu8_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(uint8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 8);
44 uint8_t* c0 = c;
45 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr != 2) {
47 c1 = c0;
48 }
49
50 const v128_t vb_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.kernel_zero_point);
51 do {
52 v128_t vacc0x0 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[0]);
53 v128_t vacc0x1 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[1]);
54 v128_t vacc0x2 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[2]);
55 v128_t vacc0x3 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[3]);
56 v128_t vacc1x0 = vacc0x0;
57 v128_t vacc1x1 = vacc0x1;
58 v128_t vacc1x2 = vacc0x2;
59 v128_t vacc1x3 = vacc0x3;
60 w = (const void*) ((const int32_t*) w + 4);
61
62 size_t p = ks;
63 do {
64 const uint8_t* restrict a0 = a[0];
65 if XNN_UNPREDICTABLE(a0 != zero) {
66 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
67 }
68 const uint8_t* restrict a1 = a[1];
69 if XNN_UNPREDICTABLE(a1 != zero) {
70 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
71 }
72 a += 2;
73
74 size_t k = 0;
75 while (k < kc) {
76 const v128_t vxa0 = wasm_u16x8_load8x8(a0);
77 a0 += 8;
78 const v128_t vxa1 = wasm_u16x8_load8x8(a1);
79 a1 += 8;
80
81 const v128_t vb01 = wasm_v128_load(w);
82 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_extend_low_u8x16(vb01), vb_zero_point);
83 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_extend_high_u8x16(vb01), vb_zero_point);
84
85 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa0), wasm_i32x4_extend_low_i16x8(vxb0)));
86 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa0), wasm_i32x4_extend_high_i16x8(vxb0)));
87 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa0), wasm_i32x4_extend_low_i16x8(vxb1)));
88 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa0), wasm_i32x4_extend_high_i16x8(vxb1)));
89 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa1), wasm_i32x4_extend_low_i16x8(vxb0)));
90 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa1), wasm_i32x4_extend_high_i16x8(vxb0)));
91 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa1), wasm_i32x4_extend_low_i16x8(vxb1)));
92 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa1), wasm_i32x4_extend_high_i16x8(vxb1)));
93 const v128_t vb23 = wasm_v128_load((const uint8_t*) w + 16);
94 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_extend_low_u8x16(vb23), vb_zero_point);
95 const v128_t vxb3 = wasm_i16x8_sub(wasm_u16x8_extend_high_u8x16(vb23), vb_zero_point);
96
97 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa0), wasm_i32x4_extend_low_i16x8(vxb2)));
98 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa0), wasm_i32x4_extend_high_i16x8(vxb2)));
99 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa0), wasm_i32x4_extend_low_i16x8(vxb3)));
100 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa0), wasm_i32x4_extend_high_i16x8(vxb3)));
101 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa1), wasm_i32x4_extend_low_i16x8(vxb2)));
102 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa1), wasm_i32x4_extend_high_i16x8(vxb2)));
103 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa1), wasm_i32x4_extend_low_i16x8(vxb3)));
104 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa1), wasm_i32x4_extend_high_i16x8(vxb3)));
105
106 w = (const void*) ((const uint8_t*) w + 32);
107 k += 8 * sizeof(uint8_t);
108 }
109 p -= 2 * sizeof(void*);
110 } while (p != 0);
111
112 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));
113 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));
114 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));
115 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));
116
117 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
118 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
119
120 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
121 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
122
123 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
124 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale);
125 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale);
126
127 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
128 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
129 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
130
131 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
132 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
133 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
134
135 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
136 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
137 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
138
139 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
140
141 v128_t vout = wasm_u8x16_narrow_i16x8(vacc01x0123, vacc01x0123);
142
143 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
144 vout = wasm_u8x16_min(vout, voutput_max);
145
146 if (nc >= 4) {
147 *((float*) c1) = (float) wasm_f32x4_extract_lane(vout, 1);
148 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
149
150 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
151 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
152
153 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
154
155 nc -= 4;
156 } else {
157 uint32_t vout1 = wasm_i32x4_extract_lane(vout, 1);
158 uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
159 if (nc & 2) {
160 *((uint16_t*) c1) = (uint16_t) vout1;
161 vout1 >>= 16;
162 c1 += 2;
163 *((uint16_t*) c0) = (uint16_t) vout0;
164 vout0 >>= 16;
165 c0 += 2;
166 }
167 if (nc & 1) {
168 *c1 = (uint8_t) vout1;
169 *c0 = (uint8_t) vout0;
170 }
171
172 nc = 0;
173 }
174 } while (nc != 0);
175 }
176