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_2x4c8__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_2x4c8__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 <= 2);
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
49 const v128_t vb_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.kernel_zero_point);
50 do {
51 v128_t vacc0x0 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[0]);
52 v128_t vacc0x1 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[1]);
53 v128_t vacc0x2 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[2]);
54 v128_t vacc0x3 = wasm_f32x4_replace_lane(wasm_f32x4_const_splat(0.0f), 0, ((const float*) w)[3]);
55 v128_t vacc1x0 = vacc0x0;
56 v128_t vacc1x1 = vacc0x1;
57 v128_t vacc1x2 = vacc0x2;
58 v128_t vacc1x3 = vacc0x3;
59 w = (const void*) ((const int32_t*) w + 4);
60
61 size_t k = 0;
62 while (k < kc) {
63 const v128_t vxa0 = wasm_u16x8_load8x8(a0);
64 a0 += 8;
65 const v128_t vxa1 = wasm_u16x8_load8x8(a1);
66 a1 += 8;
67
68 const v128_t vb01 = wasm_v128_load(w);
69 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_extend_low_u8x16(vb01), vb_zero_point);
70 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_extend_high_u8x16(vb01), vb_zero_point);
71
72 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa0), wasm_i32x4_extend_low_i16x8(vxb0)));
73 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa0), wasm_i32x4_extend_high_i16x8(vxb0)));
74 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa0), wasm_i32x4_extend_low_i16x8(vxb1)));
75 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa0), wasm_i32x4_extend_high_i16x8(vxb1)));
76 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa1), wasm_i32x4_extend_low_i16x8(vxb0)));
77 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa1), wasm_i32x4_extend_high_i16x8(vxb0)));
78 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa1), wasm_i32x4_extend_low_i16x8(vxb1)));
79 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa1), wasm_i32x4_extend_high_i16x8(vxb1)));
80 const v128_t vb23 = wasm_v128_load((const uint8_t*) w + 16);
81 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_extend_low_u8x16(vb23), vb_zero_point);
82 const v128_t vxb3 = wasm_i16x8_sub(wasm_u16x8_extend_high_u8x16(vb23), vb_zero_point);
83
84 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa0), wasm_i32x4_extend_low_i16x8(vxb2)));
85 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa0), wasm_i32x4_extend_high_i16x8(vxb2)));
86 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa0), wasm_i32x4_extend_low_i16x8(vxb3)));
87 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa0), wasm_i32x4_extend_high_i16x8(vxb3)));
88 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa1), wasm_i32x4_extend_low_i16x8(vxb2)));
89 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa1), wasm_i32x4_extend_high_i16x8(vxb2)));
90 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_mul(wasm_i32x4_extend_low_i16x8(vxa1), wasm_i32x4_extend_low_i16x8(vxb3)));
91 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(vxa1), wasm_i32x4_extend_high_i16x8(vxb3)));
92
93 w = (const void*) ((const uint8_t*) w + 32);
94 k += 8 * sizeof(uint8_t);
95 }
96
97 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));
98 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));
99 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));
100 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));
101
102 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
103 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
104
105 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
106 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
107
108 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
109 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale);
110 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale);
111
112 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
113 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
114 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
115
116 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
117 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
118 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
119
120 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
121 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
122 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
123
124 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
125
126 v128_t vout = wasm_u8x16_narrow_i16x8(vacc01x0123, vacc01x0123);
127
128 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
129 vout = wasm_u8x16_min(vout, voutput_max);
130
131 if (nc >= 4) {
132 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
133 *((float*) c1) = (float) wasm_f32x4_extract_lane(vout, 1);
134
135 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
136 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
137
138 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
139 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
140
141 nc -= 4;
142 } else {
143 uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
144 uint32_t vout1 = wasm_i32x4_extract_lane(vout, 1);
145 if (nc & 2) {
146 *((uint16_t*) c0) = (uint16_t) vout0;
147 vout0 >>= 16;
148 c0 += 2;
149 *((uint16_t*) c1) = (uint16_t) vout1;
150 vout1 >>= 16;
151 c1 += 2;
152 }
153 if (nc & 1) {
154 *c0 = (uint8_t) vout0;
155 *c1 = (uint8_t) vout1;
156 }
157
158 nc = 0;
159 }
160 } while (nc != 0);
161 }
162