1 // Auto-generated file. Do not edit!
2 // Template: src/f32-spmm/wasmsimd.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/spmm.h>
15
16
xnn_f32_spmm_minmax_ukernel_4x1__wasmsimd_x86_x4(size_t mc,size_t nc,const float * restrict input,const float * restrict weights,const int32_t * restrict widx_dmap,const uint32_t * restrict nidx_nnzmap,float * restrict output,size_t output_stride,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_spmm_minmax_ukernel_4x1__wasmsimd_x86_x4(
18 size_t mc,
19 size_t nc,
20 const float*restrict input,
21 const float*restrict weights,
22 const int32_t*restrict widx_dmap,
23 const uint32_t*restrict nidx_nnzmap,
24 float*restrict output,
25 size_t output_stride,
26 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
27 {
28 assert(mc != 0);
29 assert(mc % sizeof(float) == 0);
30 assert(nc != 0);
31
32 const v128_t vmin = wasm_v128_load64_splat(params->wasmsimd.min);
33 const v128_t vmax = wasm_v128_load64_splat(params->wasmsimd.max);
34 size_t output_decrement = output_stride * nc - 4 * sizeof(float);
35 while XNN_LIKELY(mc >= 4 * sizeof(float)) {
36 const float*restrict w = weights;
37 const int32_t* dmap = widx_dmap;
38 const uint32_t* nnzmap = nidx_nnzmap;
39 size_t n = nc;
40 do {
41 uint32_t nnz = *nnzmap++;
42 v128_t vacc0123x0 = wasm_v128_load32_splat(w);
43 w += 1;
44 v128_t vacc0123x1 = wasm_f32x4_const_splat(0.0f);
45 v128_t vacc0123x2 = wasm_f32x4_const_splat(0.0f);
46 v128_t vacc0123x3 = wasm_f32x4_const_splat(0.0f);
47 for (; nnz >= 4; nnz -= 4) {
48 const intptr_t diff0 = dmap[0];
49 const intptr_t diff1 = dmap[1];
50 const intptr_t diff2 = dmap[2];
51 const intptr_t diff3 = dmap[3];
52 dmap += 4;
53 const v128_t vi0123x0 = wasm_v128_load(input);
54 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff0);
55 const v128_t vw0 = wasm_v128_load32_splat(w);
56 w += 1;
57 vacc0123x0 = wasm_f32x4_add(vacc0123x0, wasm_f32x4_mul(vi0123x0, vw0));
58 const v128_t vi0123x1 = wasm_v128_load(input);
59 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff1);
60 const v128_t vw1 = wasm_v128_load32_splat(w);
61 w += 1;
62 vacc0123x1 = wasm_f32x4_add(vacc0123x1, wasm_f32x4_mul(vi0123x1, vw1));
63 const v128_t vi0123x2 = wasm_v128_load(input);
64 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff2);
65 const v128_t vw2 = wasm_v128_load32_splat(w);
66 w += 1;
67 vacc0123x2 = wasm_f32x4_add(vacc0123x2, wasm_f32x4_mul(vi0123x2, vw2));
68 const v128_t vi0123x3 = wasm_v128_load(input);
69 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff3);
70 const v128_t vw3 = wasm_v128_load32_splat(w);
71 w += 1;
72 vacc0123x3 = wasm_f32x4_add(vacc0123x3, wasm_f32x4_mul(vi0123x3, vw3));
73 }
74 v128_t vacc0123 = vacc0123x0;
75 vacc0123 = wasm_f32x4_add(vacc0123, vacc0123x1);
76 vacc0123 = wasm_f32x4_add(vacc0123, vacc0123x2);
77 vacc0123 = wasm_f32x4_add(vacc0123, vacc0123x3);
78 if XNN_LIKELY(nnz != 0) {
79 do {
80 const intptr_t diff = *dmap++;
81 const v128_t vi0123 = wasm_v128_load(input);
82 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
83 const v128_t vw = wasm_v128_load32_splat(w); w += 1;
84 vacc0123 = wasm_f32x4_add(vacc0123, wasm_f32x4_mul(vi0123, vw));
85 } while (--nnz != 0);
86 }
87 v128_t vout0123 = wasm_f32x4_pmin(vmax, vacc0123);
88 vout0123 = wasm_f32x4_pmax(vmin, vout0123);
89 wasm_v128_store(output, vout0123);
90 output = (float*restrict) ((uintptr_t) output + output_stride);
91 } while (--n != 0);
92 output = (float*restrict) ((uintptr_t) output - output_decrement);
93 input += 4;
94 mc -= 4 * sizeof(float);
95 }
96 if XNN_UNLIKELY(mc != 0) {
97 output_decrement += 2 * sizeof(float);
98 if (mc & (2 * sizeof(float))) {
99 const float*restrict w = weights;
100 const int32_t* dmap = widx_dmap;
101 const uint32_t* nnzmap = nidx_nnzmap;
102 size_t n = nc;
103 do {
104 uint32_t nnz = *nnzmap++;
105 v128_t vacc01 = wasm_v128_load32_splat(w); w += 1;
106 if XNN_LIKELY(nnz != 0) {
107 do {
108 const intptr_t diff = *dmap++;
109 const v128_t vi01 = wasm_v128_load64_splat(input);
110 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
111 const v128_t vw = wasm_v128_load32_splat(w); w += 1;
112 vacc01 = wasm_f32x4_add(vacc01, wasm_f32x4_mul(vi01, vw));
113 } while (--nnz != 0);
114 }
115 v128_t vout01 = wasm_f32x4_pmin(vmax, vacc01);
116 vout01 = wasm_f32x4_pmax(vmin, vout01);
117 *((double*) output) = wasm_f64x2_extract_lane(vout01, 0);
118
119 output = (float*restrict) ((uintptr_t) output + output_stride);
120 } while (--n != 0);
121 output = (float*restrict) ((uintptr_t) output - output_decrement);
122 input += 2;
123 }
124 output_decrement += 1 * sizeof(float);
125 if (mc & (1 * sizeof(float))) {
126 const float*restrict w = weights;
127 const int32_t* dmap = widx_dmap;
128 const uint32_t* nnzmap = nidx_nnzmap;
129 size_t n = nc;
130 do {
131 uint32_t nnz = *nnzmap++;
132 v128_t vacc0 = wasm_v128_load32_splat(w); w += 1;
133 if XNN_LIKELY(nnz != 0) {
134 do {
135 const intptr_t diff = *dmap++;
136 const v128_t vi0 = wasm_v128_load32_splat(input);
137 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
138 const v128_t vw = wasm_v128_load32_splat(w); w += 1;
139 vacc0 = wasm_f32x4_add(vacc0, wasm_f32x4_mul(vi0, vw));
140 } while (--nnz != 0);
141 }
142 v128_t vout0 = wasm_f32x4_pmin(vmax, vacc0);
143 vout0 = wasm_f32x4_pmax(vmin, vout0);
144 *output = wasm_f32x4_extract_lane(vout0, 0);
145
146 output = (float*restrict) ((uintptr_t) output + output_stride);
147 } while (--n != 0);
148 output = (float*restrict) ((uintptr_t) output - output_decrement);
149 input += 1;
150 }
151 }
152 }
153