Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-7x16c4-minmax-avx512skx-madd-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx16c4-avx512vnni.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 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 <immintrin.h>
13
14 #include "xnnpack/common.h"
15 #include "xnnpack/gemm.h"
16 #include "xnnpack/intrinsics-polyfill.h"
17 #include "xnnpack/math.h"
18 #include "xnnpack/unaligned.h"
19 #include "xnnpack/prefetch.h"
20
21
22 void xnn_qd8_f32_qc4w_gemm_minmax_ukernel_7x16c4__avx512skx_madd_prfm(
23 size_t mr,
24 size_t nc,
25 size_t kc,
26 const int8_t* restrict a,
27 size_t a_stride,
28 const void* restrict w,
29 float* restrict c,
30 size_t cm_stride,
31 size_t cn_stride,
32 const struct xnn_f32_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
33 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(mr != 0);
36 assert(mr <= 7);
37 assert(nc != 0);
38 assert(kc != 0);
39 assert(kc % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 4 * sizeof(int8_t));
45 const int8_t* a0 = a;
46 float* c0 = c;
47 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
48 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
49 if XNN_UNPREDICTABLE(mr < 2) {
50 a1 = a0;
51 c1 = c0;
52 }
53 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
54 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
55 if XNN_UNPREDICTABLE(mr <= 2) {
56 a2 = a1;
57 c2 = c1;
58 }
59 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
60 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
61 if XNN_UNPREDICTABLE(mr < 4) {
62 a3 = a2;
63 c3 = c2;
64 }
65 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
66 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
67 if XNN_UNPREDICTABLE(mr <= 4) {
68 a4 = a3;
69 c4 = c3;
70 }
71 const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
72 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
73 if XNN_UNPREDICTABLE(mr < 6) {
74 a5 = a4;
75 c5 = c4;
76 }
77 const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride);
78 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
79 if XNN_UNPREDICTABLE(mr <= 6) {
80 a6 = a5;
81 c6 = c5;
82 }
83
84 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
85 XNN_FORCE_REALIZATION(vsign_mask);
86 const __m512i vinput_zero_point0 = _mm512_set1_epi32((int) quantization_params[0].zero_point + 128);
87 const __m512i vinput_zero_point1 = _mm512_set1_epi32((int) quantization_params[1].zero_point + 128);
88 const __m512i vinput_zero_point2 = _mm512_set1_epi32((int) quantization_params[2].zero_point + 128);
89 const __m512i vinput_zero_point3 = _mm512_set1_epi32((int) quantization_params[3].zero_point + 128);
90 const __m512i vinput_zero_point4 = _mm512_set1_epi32((int) quantization_params[4].zero_point + 128);
91 const __m512i vinput_zero_point5 = _mm512_set1_epi32((int) quantization_params[5].zero_point + 128);
92 const __m512i vinput_zero_point6 = _mm512_set1_epi32((int) quantization_params[6].zero_point + 128);
93 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
94 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
95 // XNN_FORCE_REALIZATION(voutput_min);
96 // XNN_FORCE_REALIZATION(voutput_max);
97 const __m512i vmask = _mm512_set1_epi8(0x0F);
98 XNN_FORCE_REALIZATION(vmask);
99 do {
100 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w);
101 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point0);
102 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
103 __m512i vacc1x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point1);
104 __m512i vacc1x1x0123456789ABCDEF = _mm512_setzero_epi32();
105 __m512i vacc2x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point2);
106 __m512i vacc1x2x0123456789ABCDEF = _mm512_setzero_epi32();
107 __m512i vacc3x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point3);
108 __m512i vacc1x3x0123456789ABCDEF = _mm512_setzero_epi32();
109 __m512i vacc4x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point4);
110 __m512i vacc1x4x0123456789ABCDEF = _mm512_setzero_epi32();
111 __m512i vacc5x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point5);
112 __m512i vacc1x5x0123456789ABCDEF = _mm512_setzero_epi32();
113 __m512i vacc6x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point6);
114 __m512i vacc1x6x0123456789ABCDEF = _mm512_setzero_epi32();
115 w = (const int32_t*) w + 16;
116
117 size_t k = kc;
118 while (k >= 8 * sizeof(int8_t)) {
119 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
120 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
121 a0 += 8;
122 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
123 const __m512i va1x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1 + 4)), vsign_mask);
124 a1 += 8;
125 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
126 const __m512i va2x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2 + 4)), vsign_mask);
127 a2 += 8;
128 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
129 const __m512i va3x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3 + 4)), vsign_mask);
130 a3 += 8;
131 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
132 const __m512i va4x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4 + 4)), vsign_mask);
133 a4 += 8;
134 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
135 const __m512i va5x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5 + 4)), vsign_mask);
136 a5 += 8;
137 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
138 const __m512i va6x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6 + 4)), vsign_mask);
139 a6 += 8;
140
141 const __m512i vbb0123456789ABCDEFx01234567 = _mm512_load_si512(w);
142 const __m512i vbs0123456789ABCDEFx01234567 = _mm512_srli_epi32(vbb0123456789ABCDEFx01234567, 4);
143 const __m512i vb0123456789ABCDEFx0123 = _mm512_and_si512(vbb0123456789ABCDEFx01234567, vmask);
144 const __m512i vb0123456789ABCDEFx4567 = _mm512_and_si512(vbs0123456789ABCDEFx01234567, vmask);
145
146 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123);
147 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEFx0123);
148 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEFx0123);
149 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEFx0123);
150 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEFx0123);
151 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEFx0123);
152 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEFx0123);
153 xnn_prefetch_to_l1((const int8_t*) w + 960);
154 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
155 vacc1x1x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc1x1x0123456789ABCDEF, va1x4567, vb0123456789ABCDEFx4567);
156 vacc1x2x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc1x2x0123456789ABCDEF, va2x4567, vb0123456789ABCDEFx4567);
157 vacc1x3x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc1x3x0123456789ABCDEF, va3x4567, vb0123456789ABCDEFx4567);
158 vacc1x4x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc1x4x0123456789ABCDEF, va4x4567, vb0123456789ABCDEFx4567);
159 vacc1x5x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc1x5x0123456789ABCDEF, va5x4567, vb0123456789ABCDEFx4567);
160 vacc1x6x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc1x6x0123456789ABCDEF, va6x4567, vb0123456789ABCDEFx4567);
161
162 w = (const int8_t*) w + 64;
163 k -= 8 * sizeof(int8_t);
164 }
165 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
166 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, vacc1x1x0123456789ABCDEF);
167 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, vacc1x2x0123456789ABCDEF);
168 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, vacc1x3x0123456789ABCDEF);
169 vacc4x0123456789ABCDEF = _mm512_add_epi32(vacc4x0123456789ABCDEF, vacc1x4x0123456789ABCDEF);
170 vacc5x0123456789ABCDEF = _mm512_add_epi32(vacc5x0123456789ABCDEF, vacc1x5x0123456789ABCDEF);
171 vacc6x0123456789ABCDEF = _mm512_add_epi32(vacc6x0123456789ABCDEF, vacc1x6x0123456789ABCDEF);
172
173 if (k != 0) {
174 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
175 a0 += 4;
176 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
177 a1 += 4;
178 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
179 a2 += 4;
180 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
181 a3 += 4;
182 const __m512i va4x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a4)), vsign_mask);
183 a4 += 4;
184 const __m512i va5x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a5)), vsign_mask);
185 a5 += 4;
186 const __m512i va6x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a6)), vsign_mask);
187 a6 += 4;
188
189 const __m512i vbb0123456789ABCDEF = _mm512_load_si512(w);
190 const __m512i vb0123456789ABCDEF = _mm512_and_si512(vbb0123456789ABCDEF, vmask);
191 xnn_prefetch_to_l1((const int8_t*) w + 960);
192
193 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF);
194 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEF);
195 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEF);
196 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEF);
197 vacc4x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc4x0123456789ABCDEF, va4x0123, vb0123456789ABCDEF);
198 vacc5x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc5x0123456789ABCDEF, va5x0123, vb0123456789ABCDEF);
199 vacc6x0123456789ABCDEF = _mm512_dpbusd_epi32_madd(vacc6x0123456789ABCDEF, va6x0123, vb0123456789ABCDEF);
200
201 w = (const int8_t*) w + 64;
202 k -= 4 * sizeof(int8_t);
203 }
204
205 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
206 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
207 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
208 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
209 __m512 vscaled4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF);
210 __m512 vscaled5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF);
211 __m512 vscaled6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF);
212
213 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
214 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale));
215 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale));
216 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale));
217 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params[4].inv_scale));
218 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params[5].inv_scale));
219 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params[6].inv_scale));
220
221 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
222 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
223 w = (const float*) w + 32;
224
225 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
226 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
227 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
228 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
229 vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vscaled4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
230 vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vscaled5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
231 vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vscaled6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
232
233 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
234 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
235 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
236 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
237 vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min);
238 vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min);
239 vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min);
240
241 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
242 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
243 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
244 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
245 vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max);
246 vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max);
247 vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max);
248
249 if XNN_LIKELY(nc >= 16) {
250 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
251 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
252 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
253 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
254 _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF);
255 _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF);
256 _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF);
257
258 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
259 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
260 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
261 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
262 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
263 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
264 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
265
266 c0 = (float*) ((uintptr_t) c0 + cn_stride);
267 c1 = (float*) ((uintptr_t) c1 + cn_stride);
268 c2 = (float*) ((uintptr_t) c2 + cn_stride);
269 c3 = (float*) ((uintptr_t) c3 + cn_stride);
270 c4 = (float*) ((uintptr_t) c4 + cn_stride);
271 c5 = (float*) ((uintptr_t) c5 + cn_stride);
272 c6 = (float*) ((uintptr_t) c6 + cn_stride);
273
274 nc -= 16;
275 } else {
276 // Prepare mask for valid 32-bit elements (depends on nc).
277 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
278 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
279 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
280 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
281 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
282 _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF);
283 _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF);
284 _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF);
285 nc = 0;
286 }
287 } while (nc != 0);
288 }