1 // Auto-generated file. Do not edit!
2 // Template: src/f32-spmm/wasmsimd-pipelined.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_pipelined(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_pipelined(
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(¶ms->scalar.min);
33 const v128_t vmax = wasm_v32x4_load_splat(¶ms->scalar.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 v128_t vw = wasm_v32x4_load_splat(w); w += 1;
40 intptr_t diff = *dmap++;
41 v128_t vi0123 = wasm_v128_load(input + 0);
42 size_t n = nc;
43 do {
44 uint32_t nnz = *nnzmap++;
45 v128_t vacc0123 = vw;
46 vw = wasm_v32x4_load_splat(w); w += 1;
47
48
49 if XNN_LIKELY(nnz != 0) {
50 do {
51 vacc0123 = wasm_f32x4_add(vacc0123, wasm_f32x4_mul(vi0123, vw));
52 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
53
54 diff = *dmap++;
55 vw = wasm_v32x4_load_splat(w); w += 1;
56 vi0123 = wasm_v128_load(input + 0);
57 } while (--nnz != 0);
58 }
59 v128_t vout0123 = wasm_v128_bitselect(vacc0123, vmax, wasm_f32x4_le(vacc0123, vmax));
60 vout0123 = wasm_v128_bitselect(vmin, vout0123, wasm_f32x4_lt(vout0123, vmin));
61 wasm_v128_store(output, vout0123);
62 output = (float*restrict) ((uintptr_t) output + output_stride);
63 } while (--n != 0);
64 output = (float*restrict) ((uintptr_t) output - output_decrement);
65 input += 4;
66 mc -= 4 * sizeof(float);
67 }
68 if XNN_UNLIKELY(mc != 0) {
69 output_decrement += 2 * sizeof(float);
70 if (mc & (2 * sizeof(float))) {
71 const float*restrict w = weights;
72 const int32_t* dmap = widx_dmap;
73 const uint32_t* nnzmap = nidx_nnzmap;
74 size_t n = nc;
75 do {
76 uint32_t nnz = *nnzmap++;
77 v128_t vacc01 = wasm_v32x4_load_splat(w); w += 1;
78 if XNN_LIKELY(nnz != 0) {
79 do {
80 const intptr_t diff = *dmap++;
81 const v128_t vi01 = wasm_v64x2_load_splat(input);
82 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
83 const v128_t vw = wasm_v32x4_load_splat(w); w += 1;
84 vacc01 = wasm_f32x4_add(vacc01, wasm_f32x4_mul(vi01, vw));
85 } while (--nnz != 0);
86 }
87 v128_t vout01 = wasm_v128_bitselect(vacc01, vmax, wasm_f32x4_le(vacc01, vmax));
88 vout01 = wasm_v128_bitselect(vmin, vout01, wasm_f32x4_lt(vout01, vmin));
89 *((double*) output) = wasm_f64x2_extract_lane(vout01, 0);
90
91 output = (float*restrict) ((uintptr_t) output + output_stride);
92 } while (--n != 0);
93 output = (float*restrict) ((uintptr_t) output - output_decrement);
94 input += 2;
95 }
96 output_decrement += 1 * sizeof(float);
97 if (mc & (1 * sizeof(float))) {
98 const float*restrict w = weights;
99 const int32_t* dmap = widx_dmap;
100 const uint32_t* nnzmap = nidx_nnzmap;
101 size_t n = nc;
102 do {
103 uint32_t nnz = *nnzmap++;
104 v128_t vacc0 = wasm_v32x4_load_splat(w); w += 1;
105 if XNN_LIKELY(nnz != 0) {
106 do {
107 const intptr_t diff = *dmap++;
108 const v128_t vi0 = wasm_v32x4_load_splat(input);
109 input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
110 const v128_t vw = wasm_v32x4_load_splat(w); w += 1;
111 vacc0 = wasm_f32x4_add(vacc0, wasm_f32x4_mul(vi0, vw));
112 } while (--nnz != 0);
113 }
114 v128_t vout0 = wasm_v128_bitselect(vacc0, vmax, wasm_f32x4_le(vacc0, vmax));
115 vout0 = wasm_v128_bitselect(vmin, vout0, wasm_f32x4_lt(vout0, vmin));
116 *output = wasm_f32x4_extract_lane(vout0, 0);
117
118 output = (float*restrict) ((uintptr_t) output + output_stride);
119 } while (--n != 0);
120 output = (float*restrict) ((uintptr_t) output - output_decrement);
121 input += 1;
122 }
123 }
124 }
125