1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker // Template: src/f32-spmm/neon-pipelined.c.in
3*4bdc9457SAndroid Build Coastguard Worker // Generator: tools/xngen
4*4bdc9457SAndroid Build Coastguard Worker //
5*4bdc9457SAndroid Build Coastguard Worker // Copyright 2019 Google LLC
6*4bdc9457SAndroid Build Coastguard Worker //
7*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
8*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
9*4bdc9457SAndroid Build Coastguard Worker
10*4bdc9457SAndroid Build Coastguard Worker #include <assert.h>
11*4bdc9457SAndroid Build Coastguard Worker
12*4bdc9457SAndroid Build Coastguard Worker #include <arm_neon.h>
13*4bdc9457SAndroid Build Coastguard Worker
14*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/spmm.h>
15*4bdc9457SAndroid Build Coastguard Worker
16*4bdc9457SAndroid Build Coastguard Worker
xnn_f32_spmm_minmax_ukernel_16x1__neon_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*4bdc9457SAndroid Build Coastguard Worker void xnn_f32_spmm_minmax_ukernel_16x1__neon_pipelined(
18*4bdc9457SAndroid Build Coastguard Worker size_t mc,
19*4bdc9457SAndroid Build Coastguard Worker size_t nc,
20*4bdc9457SAndroid Build Coastguard Worker const float*restrict input,
21*4bdc9457SAndroid Build Coastguard Worker const float*restrict weights,
22*4bdc9457SAndroid Build Coastguard Worker const int32_t*restrict widx_dmap,
23*4bdc9457SAndroid Build Coastguard Worker const uint32_t*restrict nidx_nnzmap,
24*4bdc9457SAndroid Build Coastguard Worker float*restrict output,
25*4bdc9457SAndroid Build Coastguard Worker size_t output_stride,
26*4bdc9457SAndroid Build Coastguard Worker const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
27*4bdc9457SAndroid Build Coastguard Worker {
28*4bdc9457SAndroid Build Coastguard Worker assert(mc != 0);
29*4bdc9457SAndroid Build Coastguard Worker assert(mc % sizeof(float) == 0);
30*4bdc9457SAndroid Build Coastguard Worker assert(nc != 0);
31*4bdc9457SAndroid Build Coastguard Worker
32*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
33*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max);
34*4bdc9457SAndroid Build Coastguard Worker size_t output_decrement = output_stride * nc - 16 * sizeof(float);
35*4bdc9457SAndroid Build Coastguard Worker while XNN_LIKELY(mc >= 16 * sizeof(float)) {
36*4bdc9457SAndroid Build Coastguard Worker const float*restrict w = weights;
37*4bdc9457SAndroid Build Coastguard Worker const int32_t* dmap = widx_dmap;
38*4bdc9457SAndroid Build Coastguard Worker const uint32_t* nnzmap = nidx_nnzmap;
39*4bdc9457SAndroid Build Coastguard Worker float32x4_t vw = vld1q_dup_f32(w); w += 1;
40*4bdc9457SAndroid Build Coastguard Worker intptr_t diff = *dmap++;
41*4bdc9457SAndroid Build Coastguard Worker float32x4_t vi0123 = vld1q_f32(input);
42*4bdc9457SAndroid Build Coastguard Worker float32x4_t vi4567 = vld1q_f32(input + 4);
43*4bdc9457SAndroid Build Coastguard Worker float32x4_t vi89AB = vld1q_f32(input + 8);
44*4bdc9457SAndroid Build Coastguard Worker float32x4_t viCDEF = vld1q_f32(input + 12);
45*4bdc9457SAndroid Build Coastguard Worker size_t n = nc;
46*4bdc9457SAndroid Build Coastguard Worker do {
47*4bdc9457SAndroid Build Coastguard Worker uint32_t nnz = *nnzmap++;
48*4bdc9457SAndroid Build Coastguard Worker float32x4_t vacc0123 = vw;
49*4bdc9457SAndroid Build Coastguard Worker float32x4_t vacc4567 = vw;
50*4bdc9457SAndroid Build Coastguard Worker float32x4_t vacc89AB = vw;
51*4bdc9457SAndroid Build Coastguard Worker float32x4_t vaccCDEF = vw;
52*4bdc9457SAndroid Build Coastguard Worker vw = vld1q_dup_f32(w); w += 1;
53*4bdc9457SAndroid Build Coastguard Worker if XNN_LIKELY(nnz != 0) {
54*4bdc9457SAndroid Build Coastguard Worker do {
55*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlaq_f32(vacc0123, vi0123, vw);
56*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlaq_f32(vacc4567, vi4567, vw);
57*4bdc9457SAndroid Build Coastguard Worker vacc89AB = vmlaq_f32(vacc89AB, vi89AB, vw);
58*4bdc9457SAndroid Build Coastguard Worker vaccCDEF = vmlaq_f32(vaccCDEF, viCDEF, vw);
59*4bdc9457SAndroid Build Coastguard Worker input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
60*4bdc9457SAndroid Build Coastguard Worker __builtin_prefetch(input + 16);
61*4bdc9457SAndroid Build Coastguard Worker diff = *dmap++;
62*4bdc9457SAndroid Build Coastguard Worker vw = vld1q_dup_f32(w); w += 1;
63*4bdc9457SAndroid Build Coastguard Worker __builtin_prefetch(w + 32);
64*4bdc9457SAndroid Build Coastguard Worker vi0123 = vld1q_f32(input);
65*4bdc9457SAndroid Build Coastguard Worker vi4567 = vld1q_f32(input + 4);
66*4bdc9457SAndroid Build Coastguard Worker vi89AB = vld1q_f32(input + 8);
67*4bdc9457SAndroid Build Coastguard Worker viCDEF = vld1q_f32(input + 12);
68*4bdc9457SAndroid Build Coastguard Worker } while (--nnz != 0);
69*4bdc9457SAndroid Build Coastguard Worker }
70*4bdc9457SAndroid Build Coastguard Worker float32x4_t vout0123 = vminq_f32(vacc0123, vmax);
71*4bdc9457SAndroid Build Coastguard Worker float32x4_t vout4567 = vminq_f32(vacc4567, vmax);
72*4bdc9457SAndroid Build Coastguard Worker float32x4_t vout89AB = vminq_f32(vacc89AB, vmax);
73*4bdc9457SAndroid Build Coastguard Worker float32x4_t voutCDEF = vminq_f32(vaccCDEF, vmax);
74*4bdc9457SAndroid Build Coastguard Worker vout0123 = vmaxq_f32(vout0123, vmin);
75*4bdc9457SAndroid Build Coastguard Worker vout4567 = vmaxq_f32(vout4567, vmin);
76*4bdc9457SAndroid Build Coastguard Worker vout89AB = vmaxq_f32(vout89AB, vmin);
77*4bdc9457SAndroid Build Coastguard Worker voutCDEF = vmaxq_f32(voutCDEF, vmin);
78*4bdc9457SAndroid Build Coastguard Worker vst1q_f32(output, vout0123);
79*4bdc9457SAndroid Build Coastguard Worker vst1q_f32(output + 4, vout4567);
80*4bdc9457SAndroid Build Coastguard Worker vst1q_f32(output + 8, vout89AB);
81*4bdc9457SAndroid Build Coastguard Worker vst1q_f32(output + 12, voutCDEF);
82*4bdc9457SAndroid Build Coastguard Worker output = (float*restrict) ((uintptr_t) output + output_stride);
83*4bdc9457SAndroid Build Coastguard Worker } while (--n != 0);
84*4bdc9457SAndroid Build Coastguard Worker output = (float*restrict) ((uintptr_t) output - output_decrement);
85*4bdc9457SAndroid Build Coastguard Worker input += 16;
86*4bdc9457SAndroid Build Coastguard Worker mc -= 16 * sizeof(float);
87*4bdc9457SAndroid Build Coastguard Worker }
88*4bdc9457SAndroid Build Coastguard Worker if XNN_UNLIKELY(mc != 0) {
89*4bdc9457SAndroid Build Coastguard Worker output_decrement += 8 * sizeof(float);
90*4bdc9457SAndroid Build Coastguard Worker if (mc & (8 * sizeof(float))) {
91*4bdc9457SAndroid Build Coastguard Worker const float*restrict w = weights;
92*4bdc9457SAndroid Build Coastguard Worker const int32_t* dmap = widx_dmap;
93*4bdc9457SAndroid Build Coastguard Worker const uint32_t* nnzmap = nidx_nnzmap;
94*4bdc9457SAndroid Build Coastguard Worker size_t n = nc;
95*4bdc9457SAndroid Build Coastguard Worker do {
96*4bdc9457SAndroid Build Coastguard Worker uint32_t nnz = *nnzmap++;
97*4bdc9457SAndroid Build Coastguard Worker float32x4_t vacc0123 = vld1q_dup_f32(w); w += 1;
98*4bdc9457SAndroid Build Coastguard Worker float32x4_t vacc4567 = vacc0123;
99*4bdc9457SAndroid Build Coastguard Worker if XNN_LIKELY(nnz != 0) {
100*4bdc9457SAndroid Build Coastguard Worker do {
101*4bdc9457SAndroid Build Coastguard Worker const intptr_t diff = *dmap++;
102*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vi0123 = vld1q_f32(input);
103*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vi4567 = vld1q_f32(input + 4);
104*4bdc9457SAndroid Build Coastguard Worker input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
105*4bdc9457SAndroid Build Coastguard Worker __builtin_prefetch(input + 16);
106*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vb = vld1q_dup_f32(w); w += 1;
107*4bdc9457SAndroid Build Coastguard Worker __builtin_prefetch(w + 32);
108*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlaq_f32(vacc0123, vi0123, vb);
109*4bdc9457SAndroid Build Coastguard Worker vacc4567 = vmlaq_f32(vacc4567, vi4567, vb);
110*4bdc9457SAndroid Build Coastguard Worker } while (--nnz != 0);
111*4bdc9457SAndroid Build Coastguard Worker }
112*4bdc9457SAndroid Build Coastguard Worker float32x4_t vout0123 = vminq_f32(vacc0123, vmax);
113*4bdc9457SAndroid Build Coastguard Worker float32x4_t vout4567 = vminq_f32(vacc4567, vmax);
114*4bdc9457SAndroid Build Coastguard Worker vout0123 = vmaxq_f32(vout0123, vmin);
115*4bdc9457SAndroid Build Coastguard Worker vout4567 = vmaxq_f32(vout4567, vmin);
116*4bdc9457SAndroid Build Coastguard Worker vst1q_f32(output, vout0123);
117*4bdc9457SAndroid Build Coastguard Worker vst1q_f32(output + 4, vout4567);
118*4bdc9457SAndroid Build Coastguard Worker output = (float*restrict) ((uintptr_t) output + output_stride);
119*4bdc9457SAndroid Build Coastguard Worker } while (--n != 0);
120*4bdc9457SAndroid Build Coastguard Worker output = (float*restrict) ((uintptr_t) output - output_decrement);
121*4bdc9457SAndroid Build Coastguard Worker input += 8;
122*4bdc9457SAndroid Build Coastguard Worker }
123*4bdc9457SAndroid Build Coastguard Worker output_decrement += 4 * sizeof(float);
124*4bdc9457SAndroid Build Coastguard Worker if (mc & (4 * sizeof(float))) {
125*4bdc9457SAndroid Build Coastguard Worker const float*restrict w = weights;
126*4bdc9457SAndroid Build Coastguard Worker const int32_t* dmap = widx_dmap;
127*4bdc9457SAndroid Build Coastguard Worker const uint32_t* nnzmap = nidx_nnzmap;
128*4bdc9457SAndroid Build Coastguard Worker size_t n = nc;
129*4bdc9457SAndroid Build Coastguard Worker do {
130*4bdc9457SAndroid Build Coastguard Worker uint32_t nnz = *nnzmap++;
131*4bdc9457SAndroid Build Coastguard Worker float32x4_t vacc0123 = vld1q_dup_f32(w); w += 1;
132*4bdc9457SAndroid Build Coastguard Worker if XNN_LIKELY(nnz != 0) {
133*4bdc9457SAndroid Build Coastguard Worker do {
134*4bdc9457SAndroid Build Coastguard Worker const intptr_t diff = *dmap++;
135*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vi0123 = vld1q_f32(input);
136*4bdc9457SAndroid Build Coastguard Worker input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
137*4bdc9457SAndroid Build Coastguard Worker __builtin_prefetch(input + 16);
138*4bdc9457SAndroid Build Coastguard Worker const float32x4_t vb = vld1q_dup_f32(w); w += 1;
139*4bdc9457SAndroid Build Coastguard Worker __builtin_prefetch(w + 32);
140*4bdc9457SAndroid Build Coastguard Worker vacc0123 = vmlaq_f32(vacc0123, vi0123, vb);
141*4bdc9457SAndroid Build Coastguard Worker } while (--nnz != 0);
142*4bdc9457SAndroid Build Coastguard Worker }
143*4bdc9457SAndroid Build Coastguard Worker float32x4_t vout0123 = vminq_f32(vacc0123, vmax);
144*4bdc9457SAndroid Build Coastguard Worker vout0123 = vmaxq_f32(vout0123, vmin);
145*4bdc9457SAndroid Build Coastguard Worker vst1q_f32(output, vout0123);
146*4bdc9457SAndroid Build Coastguard Worker output = (float*restrict) ((uintptr_t) output + output_stride);
147*4bdc9457SAndroid Build Coastguard Worker } while (--n != 0);
148*4bdc9457SAndroid Build Coastguard Worker output = (float*restrict) ((uintptr_t) output - output_decrement);
149*4bdc9457SAndroid Build Coastguard Worker input += 4;
150*4bdc9457SAndroid Build Coastguard Worker }
151*4bdc9457SAndroid Build Coastguard Worker output_decrement += 2 * sizeof(float);
152*4bdc9457SAndroid Build Coastguard Worker if (mc & (2 * sizeof(float))) {
153*4bdc9457SAndroid Build Coastguard Worker const float*restrict w = weights;
154*4bdc9457SAndroid Build Coastguard Worker const int32_t* dmap = widx_dmap;
155*4bdc9457SAndroid Build Coastguard Worker const uint32_t* nnzmap = nidx_nnzmap;
156*4bdc9457SAndroid Build Coastguard Worker size_t n = nc;
157*4bdc9457SAndroid Build Coastguard Worker do {
158*4bdc9457SAndroid Build Coastguard Worker uint32_t nnz = *nnzmap++;
159*4bdc9457SAndroid Build Coastguard Worker float32x2_t vacc01 = vld1_dup_f32(w); w += 1;
160*4bdc9457SAndroid Build Coastguard Worker if XNN_LIKELY(nnz != 0) {
161*4bdc9457SAndroid Build Coastguard Worker do {
162*4bdc9457SAndroid Build Coastguard Worker const intptr_t diff = *dmap++;
163*4bdc9457SAndroid Build Coastguard Worker const float32x2_t vi01 = vld1_f32(input);
164*4bdc9457SAndroid Build Coastguard Worker input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
165*4bdc9457SAndroid Build Coastguard Worker __builtin_prefetch(input + 16);
166*4bdc9457SAndroid Build Coastguard Worker const float32x2_t vb = vld1_dup_f32(w); w += 1;
167*4bdc9457SAndroid Build Coastguard Worker __builtin_prefetch(w + 32);
168*4bdc9457SAndroid Build Coastguard Worker vacc01 = vmla_f32(vacc01, vi01, vb);
169*4bdc9457SAndroid Build Coastguard Worker } while (--nnz != 0);
170*4bdc9457SAndroid Build Coastguard Worker }
171*4bdc9457SAndroid Build Coastguard Worker float32x2_t vout01 = vmin_f32(vacc01, vget_low_f32(vmax));
172*4bdc9457SAndroid Build Coastguard Worker vout01 = vmax_f32(vout01, vget_low_f32(vmin));
173*4bdc9457SAndroid Build Coastguard Worker vst1_f32(output, vout01);
174*4bdc9457SAndroid Build Coastguard Worker output = (float*restrict) ((uintptr_t) output + output_stride);
175*4bdc9457SAndroid Build Coastguard Worker } while (--n != 0);
176*4bdc9457SAndroid Build Coastguard Worker output = (float*restrict) ((uintptr_t) output - output_decrement);
177*4bdc9457SAndroid Build Coastguard Worker input += 2;
178*4bdc9457SAndroid Build Coastguard Worker }
179*4bdc9457SAndroid Build Coastguard Worker output_decrement += 1 * sizeof(float);
180*4bdc9457SAndroid Build Coastguard Worker if (mc & (1 * sizeof(float))) {
181*4bdc9457SAndroid Build Coastguard Worker const float*restrict w = weights;
182*4bdc9457SAndroid Build Coastguard Worker const int32_t* dmap = widx_dmap;
183*4bdc9457SAndroid Build Coastguard Worker const uint32_t* nnzmap = nidx_nnzmap;
184*4bdc9457SAndroid Build Coastguard Worker size_t n = nc;
185*4bdc9457SAndroid Build Coastguard Worker do {
186*4bdc9457SAndroid Build Coastguard Worker uint32_t nnz = *nnzmap++;
187*4bdc9457SAndroid Build Coastguard Worker float32x2_t vacc0 = vld1_dup_f32(w); w += 1;
188*4bdc9457SAndroid Build Coastguard Worker if XNN_LIKELY(nnz != 0) {
189*4bdc9457SAndroid Build Coastguard Worker do {
190*4bdc9457SAndroid Build Coastguard Worker const intptr_t diff = *dmap++;
191*4bdc9457SAndroid Build Coastguard Worker const float32x2_t vi0 = vld1_dup_f32(input);
192*4bdc9457SAndroid Build Coastguard Worker input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
193*4bdc9457SAndroid Build Coastguard Worker __builtin_prefetch(input + 16);
194*4bdc9457SAndroid Build Coastguard Worker const float32x2_t vb = vld1_dup_f32(w); w += 1;
195*4bdc9457SAndroid Build Coastguard Worker __builtin_prefetch(w + 32);
196*4bdc9457SAndroid Build Coastguard Worker vacc0 = vmla_f32(vacc0, vi0, vb);
197*4bdc9457SAndroid Build Coastguard Worker } while (--nnz != 0);
198*4bdc9457SAndroid Build Coastguard Worker }
199*4bdc9457SAndroid Build Coastguard Worker float32x2_t vout0 = vmin_f32(vacc0, vget_low_f32(vmax));
200*4bdc9457SAndroid Build Coastguard Worker vout0 = vmax_f32(vout0, vget_low_f32(vmin));
201*4bdc9457SAndroid Build Coastguard Worker vst1_lane_f32(output, vout0, 0);
202*4bdc9457SAndroid Build Coastguard Worker output = (float*restrict) ((uintptr_t) output + output_stride);
203*4bdc9457SAndroid Build Coastguard Worker } while (--n != 0);
204*4bdc9457SAndroid Build Coastguard Worker output = (float*restrict) ((uintptr_t) output - output_decrement);
205*4bdc9457SAndroid Build Coastguard Worker input += 1;
206*4bdc9457SAndroid Build Coastguard Worker }
207*4bdc9457SAndroid Build Coastguard Worker }
208*4bdc9457SAndroid Build Coastguard Worker }
209