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_v32x4_load_splat(&params->scalar.min);
33   const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
34   const v128_t vzero = wasm_f64x2_splat(0.0);
35   size_t output_decrement = output_stride * nc - 4 * sizeof(float);
36   while XNN_LIKELY(mc >= 4 * sizeof(float)) {
37     const float*restrict w = weights;
38     const int32_t* dmap = widx_dmap;
39     const uint32_t* nnzmap = nidx_nnzmap;
40     size_t n = nc;
41     do {
42       uint32_t nnz = *nnzmap++;
43       v128_t vacc0123x0 = wasm_v32x4_load_splat(w);
44       w += 1;
45       v128_t vacc0123x1 = vzero;
46       v128_t vacc0123x2 = vzero;
47       v128_t vacc0123x3 = vzero;
48       for (; nnz >= 4; nnz -= 4) {
49         const intptr_t diff0 = dmap[0];
50         const intptr_t diff1 = dmap[1];
51         const intptr_t diff2 = dmap[2];
52         const intptr_t diff3 = dmap[3];
53         dmap += 4;
54         const v128_t vi0123x0 = wasm_v128_load(input);
55         input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff0);
56         const v128_t vw0 = wasm_v32x4_load_splat(w);
57         w += 1;
58         vacc0123x0 = wasm_f32x4_add(vacc0123x0, wasm_f32x4_mul(vi0123x0, vw0));
59         const v128_t vi0123x1 = wasm_v128_load(input);
60         input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff1);
61         const v128_t vw1 = wasm_v32x4_load_splat(w);
62         w += 1;
63         vacc0123x1 = wasm_f32x4_add(vacc0123x1, wasm_f32x4_mul(vi0123x1, vw1));
64         const v128_t vi0123x2 = wasm_v128_load(input);
65         input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff2);
66         const v128_t vw2 = wasm_v32x4_load_splat(w);
67         w += 1;
68         vacc0123x2 = wasm_f32x4_add(vacc0123x2, wasm_f32x4_mul(vi0123x2, vw2));
69         const v128_t vi0123x3 = wasm_v128_load(input);
70         input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff3);
71         const v128_t vw3 = wasm_v32x4_load_splat(w);
72         w += 1;
73         vacc0123x3 = wasm_f32x4_add(vacc0123x3, wasm_f32x4_mul(vi0123x3, vw3));
74       }
75       v128_t vacc0123 = vacc0123x0;
76       vacc0123 = wasm_f32x4_add(vacc0123, vacc0123x1);
77       vacc0123 = wasm_f32x4_add(vacc0123, vacc0123x2);
78       vacc0123 = wasm_f32x4_add(vacc0123, vacc0123x3);
79       if XNN_LIKELY(nnz != 0) {
80         do {
81           const intptr_t diff = *dmap++;
82           const v128_t vi0123 = wasm_v128_load(input);
83           input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
84           const v128_t vw = wasm_v32x4_load_splat(w); w += 1;
85           vacc0123 = wasm_f32x4_add(vacc0123, wasm_f32x4_mul(vi0123, vw));
86         } while (--nnz != 0);
87       }
88       v128_t vout0123 = wasm_v128_bitselect(vacc0123, vmax, wasm_f32x4_le(vacc0123, vmax));
89       vout0123 = wasm_v128_bitselect(vmin, vout0123, wasm_f32x4_lt(vout0123, vmin));
90       wasm_v128_store(output, vout0123);
91       output = (float*restrict) ((uintptr_t) output + output_stride);
92     } while (--n != 0);
93     output = (float*restrict) ((uintptr_t) output - output_decrement);
94     input += 4;
95     mc -= 4 * sizeof(float);
96   }
97   if XNN_UNLIKELY(mc != 0) {
98     output_decrement += 2 * sizeof(float);
99     if (mc & (2 * sizeof(float))) {
100       const float*restrict w = weights;
101       const int32_t* dmap = widx_dmap;
102       const uint32_t* nnzmap = nidx_nnzmap;
103       size_t n = nc;
104       do {
105         uint32_t nnz = *nnzmap++;
106         v128_t vacc01 = wasm_v32x4_load_splat(w); w += 1;
107         if XNN_LIKELY(nnz != 0) {
108           do {
109             const intptr_t diff = *dmap++;
110             const v128_t vi01 = wasm_v64x2_load_splat(input);
111             input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
112             const v128_t vw = wasm_v32x4_load_splat(w); w += 1;
113             vacc01 = wasm_f32x4_add(vacc01, wasm_f32x4_mul(vi01, vw));
114           } while (--nnz != 0);
115         }
116         v128_t vout01 = wasm_v128_bitselect(vacc01, vmax, wasm_f32x4_le(vacc01, vmax));
117         vout01 = wasm_v128_bitselect(vmin, vout01, wasm_f32x4_lt(vout01, vmin));
118         *((double*) output) = wasm_f64x2_extract_lane(vout01, 0);
119 
120         output = (float*restrict) ((uintptr_t) output + output_stride);
121       } while (--n != 0);
122       output = (float*restrict) ((uintptr_t) output - output_decrement);
123       input += 2;
124     }
125     output_decrement += 1 * sizeof(float);
126     if (mc & (1 * sizeof(float))) {
127       const float*restrict w = weights;
128       const int32_t* dmap = widx_dmap;
129       const uint32_t* nnzmap = nidx_nnzmap;
130       size_t n = nc;
131       do {
132         uint32_t nnz = *nnzmap++;
133         v128_t vacc0 = wasm_v32x4_load_splat(w); w += 1;
134         if XNN_LIKELY(nnz != 0) {
135           do {
136             const intptr_t diff = *dmap++;
137             const v128_t vi0 = wasm_v32x4_load_splat(input);
138             input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
139             const v128_t vw = wasm_v32x4_load_splat(w); w += 1;
140             vacc0 = wasm_f32x4_add(vacc0, wasm_f32x4_mul(vi0, vw));
141           } while (--nnz != 0);
142         }
143         v128_t vout0 = wasm_v128_bitselect(vacc0, vmax, wasm_f32x4_le(vacc0, vmax));
144         vout0 = wasm_v128_bitselect(vmin, vout0, wasm_f32x4_lt(vout0, vmin));
145         *output = wasm_f32x4_extract_lane(vout0, 0);
146 
147         output = (float*restrict) ((uintptr_t) output + output_stride);
148       } while (--n != 0);
149       output = (float*restrict) ((uintptr_t) output - output_decrement);
150       input += 1;
151     }
152   }
153 }
154