Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-9p16c-minmax-fp32-wasmsimd-mul16-add16.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-wasmsimd-mul16.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/dwconv.h"
15 #include "xnnpack/math.h"
16
17
18 void xnn_qs8_dwconv_minmax_fp32_ukernel_9p16c__wasmsimd_mul16_add16(
19 size_t channels,
20 size_t output_width,
21 const int8_t** input,
22 const void* weights,
23 int8_t* output,
24 intptr_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const int8_t* zero,
28 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33
34 const v128_t vscale = wasm_v128_load32_splat(&params->fp32_scalar.scale);
35 XNN_FORCE_REALIZATION(vscale);
36 const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
37 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
38 const v128_t vmagic_min = wasm_i32x4_splat((int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point));
39 const v128_t vmagic_bias_less_output_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point);
40 const v128_t voutput_max = wasm_i8x16_splat(params->fp32_scalar.output_max);
41 XNN_FORCE_REALIZATION(vmagic_bias);
42 XNN_FORCE_REALIZATION(vmagic_min);
43 XNN_FORCE_REALIZATION(vmagic_bias_less_output_zero_point);
44 XNN_FORCE_REALIZATION(voutput_max);
45
46 do {
47 const int8_t* i0 = input[0];
48 assert(i0 != NULL);
49 if XNN_UNPREDICTABLE(i0 != zero) {
50 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
51 }
52 const int8_t* i1 = input[1];
53 assert(i1 != NULL);
54 if XNN_UNPREDICTABLE(i1 != zero) {
55 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
56 }
57 const int8_t* i2 = input[2];
58 assert(i2 != NULL);
59 if XNN_UNPREDICTABLE(i2 != zero) {
60 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
61 }
62 const int8_t* i3 = input[3];
63 assert(i3 != NULL);
64 if XNN_UNPREDICTABLE(i3 != zero) {
65 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
66 }
67 const int8_t* i4 = input[4];
68 assert(i4 != NULL);
69 if XNN_UNPREDICTABLE(i4 != zero) {
70 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
71 }
72 const int8_t* i5 = input[5];
73 assert(i5 != NULL);
74 if XNN_UNPREDICTABLE(i5 != zero) {
75 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
76 }
77 const int8_t* i6 = input[6];
78 assert(i6 != NULL);
79 if XNN_UNPREDICTABLE(i6 != zero) {
80 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
81 }
82 const int8_t* i7 = input[7];
83 assert(i7 != NULL);
84 if XNN_UNPREDICTABLE(i7 != zero) {
85 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
86 }
87 const int8_t* i8 = input[8];
88 assert(i8 != NULL);
89 if XNN_UNPREDICTABLE(i8 != zero) {
90 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
91 }
92 input = (const int8_t**) ((uintptr_t) input + input_stride);
93
94 size_t c = channels;
95 const void* w = weights;
96 for (; c >= 16; c -= 16) {
97 v128_t vacc0123 = wasm_v128_load(w);
98 v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
99 v128_t vacc89AB = wasm_v128_load((const void*) ((uintptr_t) w + 8 * sizeof(int32_t)));
100 v128_t vaccCDEF = wasm_v128_load((const void*) ((uintptr_t) w + 12 * sizeof(int32_t)));
101
102
103 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
104 const v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
105 const v128_t vi0x89ABCDEF = wasm_i16x8_load8x8(i0 + 8);
106 const v128_t vk0x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
107 i0 += 16;
108
109 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
110 v128_t vprod89ABCDEF = wasm_i16x8_mul(vi0x89ABCDEF, vk0x89ABCDEF);
111
112
113 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
114 const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
115 const v128_t vi1x89ABCDEF = wasm_i16x8_load8x8(i1 + 8);
116 const v128_t vk1x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
117 i1 += 16;
118
119 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
120 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi1x89ABCDEF, vk1x89ABCDEF));
121
122 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
123 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
124 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
125 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
126
127 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
128 const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
129 const v128_t vi2x89ABCDEF = wasm_i16x8_load8x8(i2 + 8);
130 const v128_t vk2x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
131 i2 += 16;
132
133 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
134 vprod89ABCDEF = wasm_i16x8_mul(vi2x89ABCDEF, vk2x89ABCDEF);
135
136
137 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
138 const v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
139 const v128_t vi3x89ABCDEF = wasm_i16x8_load8x8(i3 + 8);
140 const v128_t vk3x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
141 i3 += 16;
142
143 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi3x01234567, vk3x01234567));
144 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi3x89ABCDEF, vk3x89ABCDEF));
145
146 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
147 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
148 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
149 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
150
151 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
152 const v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
153 const v128_t vi4x89ABCDEF = wasm_i16x8_load8x8(i4 + 8);
154 const v128_t vk4x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
155 i4 += 16;
156
157 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
158 vprod89ABCDEF = wasm_i16x8_mul(vi4x89ABCDEF, vk4x89ABCDEF);
159
160
161 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
162 const v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)));
163 const v128_t vi5x89ABCDEF = wasm_i16x8_load8x8(i5 + 8);
164 const v128_t vk5x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)));
165 i5 += 16;
166
167 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi5x01234567, vk5x01234567));
168 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi5x89ABCDEF, vk5x89ABCDEF));
169
170 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
171 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
172 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
173 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
174
175 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
176 const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)));
177 const v128_t vi6x89ABCDEF = wasm_i16x8_load8x8(i6 + 8);
178 const v128_t vk6x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)));
179 i6 += 16;
180
181 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
182 vprod89ABCDEF = wasm_i16x8_mul(vi6x89ABCDEF, vk6x89ABCDEF);
183
184
185 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
186 const v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)));
187 const v128_t vi7x89ABCDEF = wasm_i16x8_load8x8(i7 + 8);
188 const v128_t vk7x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)));
189 i7 += 16;
190
191 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi7x01234567, vk7x01234567));
192 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi7x89ABCDEF, vk7x89ABCDEF));
193
194 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
195 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
196 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
197 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
198
199 const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
200 const v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)));
201 const v128_t vi8x89ABCDEF = wasm_i16x8_load8x8(i8 + 8);
202 const v128_t vk8x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)));
203 i8 += 16;
204
205 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
206 vprod89ABCDEF = wasm_i16x8_mul(vi8x89ABCDEF, vk8x89ABCDEF);
207
208 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
209 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
210 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
211 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
212
213
214 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t));
215
216 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
217 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
218 vacc89AB = wasm_f32x4_convert_i32x4(vacc89AB);
219 vaccCDEF = wasm_f32x4_convert_i32x4(vaccCDEF);
220
221 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
222 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
223 vacc89AB = wasm_f32x4_mul(vacc89AB, vscale);
224 vaccCDEF = wasm_f32x4_mul(vaccCDEF, vscale);
225
226 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
227 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
228 vacc89AB = wasm_f32x4_add(vacc89AB, vmagic_bias);
229 vaccCDEF = wasm_f32x4_add(vaccCDEF, vmagic_bias);
230
231 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
232 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
233 vacc89AB = wasm_i32x4_max(vacc89AB, vmagic_min);
234 vaccCDEF = wasm_i32x4_max(vaccCDEF, vmagic_min);
235
236 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
237 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
238 vacc89AB = wasm_i32x4_sub(vacc89AB, vmagic_bias_less_output_zero_point);
239 vaccCDEF = wasm_i32x4_sub(vaccCDEF, vmagic_bias_less_output_zero_point);
240
241 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
242 v128_t vout89ABCDEF = wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF);
243
244 v128_t vout0123456789ABCDEF = wasm_i8x16_narrow_i16x8(vout01234567, vout89ABCDEF);
245
246 vout0123456789ABCDEF = wasm_i8x16_min(vout0123456789ABCDEF, voutput_max);
247
248 wasm_v128_store(output, vout0123456789ABCDEF);
249 output += 16;
250 }
251 if XNN_UNLIKELY(c != 0) {
252 const int8_t* k = (const int8_t*) ((uintptr_t) w + 16 * sizeof(int32_t));
253 do {
254 v128_t vacc0123 = wasm_v128_load(w);
255 v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
256
257
258 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
259 const v128_t vk0x01234567 = wasm_i16x8_load8x8(k);
260 i0 += 8;
261
262 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
263
264
265 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
266 const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) (k + 16));
267 i1 += 8;
268
269 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
270
271 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
272 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
273
274 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
275 const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) (k + 32));
276 i2 += 8;
277
278 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
279
280
281 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
282 const v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) (k + 48));
283 i3 += 8;
284
285 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi3x01234567, vk3x01234567));
286
287 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
288 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
289
290 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
291 const v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) (k + 64));
292 i4 += 8;
293
294 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
295
296
297 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
298 const v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) (k + 80));
299 i5 += 8;
300
301 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi5x01234567, vk5x01234567));
302
303 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
304 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
305
306 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
307 const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) (k + 96));
308 i6 += 8;
309
310 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
311
312
313 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
314 const v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) (k + 112));
315 i7 += 8;
316
317 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi7x01234567, vk7x01234567));
318
319 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
320 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
321
322 const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
323 const v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) (k + 128));
324 i8 += 8;
325
326 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
327
328 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
329 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
330
331 k += 8;
332
333
334 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
335 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
336
337 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
338 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
339
340 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
341 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
342
343 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
344 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
345
346 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
347 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
348
349 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
350 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
351
352 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
353
354 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t));
355
356 if XNN_LIKELY(c >= 8) {
357 wasm_v128_store64_lane(output, vout0123456701234567, 0);
358 output += 8;
359 c -= 8;
360 } else {
361 if (c & 4) {
362 wasm_v128_store32_lane(output, vout0123456701234567, 0);
363 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
364 output += 4;
365 }
366 if (c & 2) {
367 wasm_v128_store16_lane(output, vout0123456701234567, 0);
368 vout0123456701234567 = wasm_u32x4_shr(vout0123456701234567, 16);
369 output += 2;
370 }
371 if (c & 1) {
372 wasm_v128_store8_lane(output, vout0123456701234567, 0);
373 output += 1;
374 }
375 c = 0;
376 }
377 } while (c != 0);
378 }
379
380 output = (int8_t*) ((uintptr_t) output + output_increment);
381 } while (--output_width != 0);
382 }