• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1// Copyright 2019 Google LLC
2//
3// This source code is licensed under the BSD-style license found in the
4// LICENSE file in the root directory of this source tree.
5
6$assert BATCH_TILE >= 1
7$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
8#include <assert.h>
9#include <math.h>
10
11#include <xnnpack/common.h>
12#include <xnnpack/vunary.h>
13
14#include <fp16/bitcasts.h>
15
16
17// Note redefine as uint32[] to avoid redundant bitcasts.
18extern XNN_INTERNAL const uint32_t xnn_table_exp2_k_over_2048[2048];
19
20void xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x${BATCH_TILE}(
21    size_t n,
22    const float* x,
23    float* y,
24    const void* params)
25{
26  assert(n % sizeof(float) == 0);
27
28  const float vmagic_bias = 0x1.800000p23f;
29  // The largest z for which sigmoidf(-z) is normalized.
30  // This number is also the largest z for which expf(-z) is normalized.
31  const float vdenorm_cutoff = 0x1.5D589Ep+6f;
32  const float vminus_log2e_x2048 = -0x1.715476p11f;
33  // Last 18 bits are zeroes
34  const float vln2_o2048_hi = 0x1.600000p-12f;
35  const float vln2_o2048_lo = 0x1.7217F8p-19f;
36  const float vone = 1.0f;
37
38  const float vc1 = -0x1.FFFFFEp-1f;
39
40  const uint32_t vindex_mask = UINT32_C(0x7FF);
41
42  $if BATCH_TILE > 1:
43    for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
44      $for N in range(BATCH_TILE):
45        const float vx${N} = x[${N}];
46      x += ${BATCH_TILE};
47
48      // General structure of the algorithm:
49      //           / exp(x) / (1 + exp(x)) if x <= 0
50      //   f[x] :=
51      //           \ 1 - f[-x] if x >= 0
52      //
53      // First we compute f[-z] := exp(-z) / (1 + exp(-z)) where z = abs(x),
54      // then replace result with 1 - f[-z] if x >= 0.
55      $for N in range(BATCH_TILE):
56        const float vz${N} = fabsf(vx${N});
57
58      // Compute reduced argument n := round(-z * 2048 / log(2)).
59      // We do it by adding a large number (magic bias), which cause rounding of the result to an integer, then subtracing
60      // the large number back. The first addition is combined with multiplication by log2e into a single FMA instruction.
61      // The trick with adding large number is valid only within certain bounds (|z * 2048 / log(2)| <= 2**22, i.e.
62      // |z| <= 0x1.62E43p+10 = 1419.5654296875), but that is acceptable, because inputs x outside of
63      // [-87.336544, 17.328678] (i.e. z outsize [0, 87.336544]) underflow or saturate sigmoidf(x). We fixup the result
64      // for such inputs at the very end of the algorithm.
65      $for N in range(BATCH_TILE):
66        float vn${N} = vz${N} * vminus_log2e_x2048 + vmagic_bias;
67
68      // Create a floating-point number s (scale) such that s := 2**(n / 2048) for such inputs that sigmoidf(-z) is
69      // normalized, i.e. 0 <= z <= 87.33642. As n has 11 fractional bits, we split s == 2**(n / 2048) =
70      // = 2**e * 2**(n / 2048 - e), where e := int(n / 2048). We create s in two steps:
71      // 1. Fetch 2**(n / 2048 - e) = 2**(n % 2048) from table using the 6 low bits of n, as integer.
72      //    Note that the fetched values are in the [1.0, 2.0) range, i.e. their floating-point exponent is 0.
73      // 2. Adjust fecthed value by addition of e to its floating-point exponent. The result is always a normalized
74      //    number, because for 0 <= z <= 87.33642 (inputs for which sigmoidf(-z) is normalized) we have -126 <= e <= 0,
75      //    and thus the adjusted exponent is not lower than -126.
76      //
77      // Extract e from bits 11:19 of n and shift it into bits 23:31 (position of floating-point exponent).
78      $for N in range(BATCH_TILE):
79        const uint32_t ve${N} = (fp32_to_bits(vn${N}) & ~vindex_mask) << 12;
80
81      // Use bits 0:11 bits of n, as integer, as an index for table lookup of l := 2**(n % 2048).
82      $for N in range(BATCH_TILE):
83        const uint32_t vidx${N} = fp32_to_bits(vn${N}) & vindex_mask;
84      // Adjust exponent of the value l fetched from the table to get the final s value.
85      $for N in range(BATCH_TILE):
86        const float vs${N} = fp32_from_bits(xnn_table_exp2_k_over_2048[vidx${N}] + ve${N});
87
88      // Subtract the large number back to get the final n := round(-z * 2048 / log(2)) as a floating-point number.
89      $for N in range(BATCH_TILE):
90        vn${N} -= vmagic_bias;
91
92      // Compute reduced argument t := (z + n * log(2) / 2048). Note that -t = -z - n * log(2) / 2048.
93      // Use Cody-Waite range reduction method (note two constants to represent log(2) / 2048) to improve accuracy.
94      $for N in range(BATCH_TILE):
95        float vt${N} = vn${N} * vln2_o2048_hi + vz${N};
96
97      $for N in range(BATCH_TILE):
98        vt${N} = vn${N} * vln2_o2048_lo + vt${N};
99
100      // Compute degree-1 polynomial approximation for exp(-t) on [-log(2)/4096, log(2)/4096]:
101      //   P1(t) = 1 + t * c1
102      $for N in range(BATCH_TILE):
103        const float vp${N} = vt${N} * vc1;
104
105      // Reconstruct the exp(-z) value:
106      //   y = s * (1 + t * c1)
107      //     = s + s * (t * c1))
108      //     = s + s * p
109      $for N in range(BATCH_TILE):
110        const float vy${N} = vp${N} * vs${N} + vs${N};
111
112      // Reconstruct sigmoid(-z) = exp(-z) / (1.0 + exp(-z))
113      $for N in range(BATCH_TILE):
114        float vf${N} = vy${N} / (vy${N} + vone);
115
116      // For inputs above denormal cutoff, replace output with +0.0f.
117      // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
118      $for N in range(BATCH_TILE):
119        if XNN_UNPREDICTABLE(vz${N} > vdenorm_cutoff) {
120          vf${N} = 0.0f;
121        }
122
123      // Reconstruct sigmoid(x) = x < 0 ? sigmoid(-z) : 1.0 - sigmoid(-z)
124      $for N in range(BATCH_TILE):
125        if XNN_UNPREDICTABLE(vx${N} > 0.0f) {
126          vf${N} = vone - vf${N};
127        }
128
129      $for N in range(BATCH_TILE):
130        y[${N}] = vf${N};
131      y += ${BATCH_TILE};
132    }
133  $if BATCH_TILE == 1:
134    do {
135      const float vx = *x++;
136
137      // General structure of the algorithm:
138      //           / exp(x) / (1 + exp(x)) if x <= 0
139      //   f[x] :=
140      //           \ 1 - f[-x] if x >= 0
141      //
142      // First we compute f[-z] := exp(-z) / (1 + exp(-z)) where z = abs(x),
143      // then replace result with 1 - f[-z] if x >= 0.
144      const float vz = fabsf(vx);
145
146      // Compute reduced argument n := round(-z * 2048 / log(2)).
147      // We do it by adding a large number (magic bias), which cause rounding of the result to an integer, then subtracing
148      // the large number back. The first addition is combined with multiplication by log2e into a single FMA instruction.
149      // The trick with adding large number is valid only within certain bounds (|z * 2048 / log(2)| <= 2**22, i.e.
150      // |z| <= 0x1.62E43p+10 = 1419.5654296875), but that is acceptable, because inputs x outside of
151      // [-87.336544, 17.328678] (i.e. z outsize [0, 87.336544]) underflow or saturate sigmoidf(x). We fixup the result
152      // for such inputs at the very end of the algorithm.
153      float vn = vz * vminus_log2e_x2048 + vmagic_bias;
154
155      // Create a floating-point number s (scale) such that s := 2**(n / 2048) for such inputs that sigmoidf(-z) is
156      // normalized, i.e. 0 <= z <= 87.33642. As n has 11 fractional bits, we split s == 2**(n / 2048) =
157      // = 2**e * 2**(n / 2048 - e), where e := int(n / 2048). We create s in two steps:
158      // 1. Fetch 2**(n / 2048 - e) = 2**(n % 2048) from table using the 6 low bits of n, as integer.
159      //    Note that the fetched values are in the [1.0, 2.0) range, i.e. their floating-point exponent is 0.
160      // 2. Adjust fecthed value by addition of e to its floating-point exponent. The result is always a normalized
161      //    number, because for 0 <= z <= 87.33642 (inputs for which sigmoidf(-z) is normalized) we have -126 <= e <= 0,
162      //    and thus the adjusted exponent is not lower than -126.
163      //
164      // Extract e from bits 11:19 of n and shift it into bits 23:31 (position of floating-point exponent).
165      const uint32_t ve = (fp32_to_bits(vn) & ~vindex_mask) << 12;
166
167      // Use bits 0:11 bits of n, as integer, as an index for table lookup of l := 2**(n % 2048).
168      const uint32_t vidx = fp32_to_bits(vn) & vindex_mask;
169      // Adjust exponent of the value l fetched from the table to get the final s value.
170      const float vs = fp32_from_bits(xnn_table_exp2_k_over_2048[vidx] + ve);
171
172      // Subtract the large number back to get the final n := round(-z * 2048 / log(2)) as a floating-point number.
173      vn -= vmagic_bias;
174
175      // Compute reduced argument t := (z + n * log(2) / 2048). Note that -t = -z - n * log(2) / 2048.
176      // Use Cody-Waite range reduction method (note two constants to represent log(2) / 2048) to improve accuracy.
177      float vt = vn * vln2_o2048_hi + vz;
178      vt = vn * vln2_o2048_lo + vt;
179
180      // Compute degree-1 polynomial approximation for exp(-t) on [-log(2)/4096, log(2)/4096]:
181      //   P1(t) = 1 + t * c1
182      const float vp = vt * vc1;
183
184      // Reconstruct the exp(-z) value:
185      //   y = s * (1 + t * c1)
186      //     = s + s * (t * c1))
187      //     = s + s * p
188      const float vy = vp * vs + vs;
189
190      // Reconstruct sigmoid(-z) = exp(-z) / (1.0 + exp(-z))
191      float vf = vy / (vy + vone);
192
193      // For inputs above denormal cutoff, replace output with +0.0f.
194      // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
195      if XNN_UNPREDICTABLE(vz > vdenorm_cutoff) {
196        vf = 0.0f;
197      }
198
199      // Reconstruct sigmoid(x) = x < 0 ? sigmoid(-z) : 1.0 - sigmoid(-z)
200      if XNN_UNPREDICTABLE(vx > 0.0f) {
201        vf = vone - vf;
202      }
203
204      *y++ = vf;
205
206      n -= sizeof(float);
207    } while (n != 0);
208  $elif BATCH_TILE == 2:
209    if XNN_UNLIKELY(n != 0) {
210      const float vx = *x;
211
212      // General structure of the algorithm:
213      //           / exp(x) / (1 + exp(x)) if x <= 0
214      //   f[x] :=
215      //           \ 1 - f[-x] if x >= 0
216      //
217      // First we compute f[-z] := exp(-z) / (1 + exp(-z)) where z = abs(x),
218      // then replace result with 1 - f[-z] if x >= 0.
219      const float vz = fabsf(vx);
220
221      // Compute reduced argument n := round(-z * 2048 / log(2)).
222      // We do it by adding a large number (magic bias), which cause rounding of the result to an integer, then subtracing
223      // the large number back. The first addition is combined with multiplication by log2e into a single FMA instruction.
224      // The trick with adding large number is valid only within certain bounds (|z * 2048 / log(2)| <= 2**22, i.e.
225      // |z| <= 0x1.62E43p+10 = 1419.5654296875), but that is acceptable, because inputs x outside of
226      // [-87.336544, 17.328678] (i.e. z outsize [0, 87.336544]) underflow or saturate sigmoidf(x). We fixup the result
227      // for such inputs at the very end of the algorithm.
228      float vn = vz * vminus_log2e_x2048 + vmagic_bias;
229
230      // Create a floating-point number s (scale) such that s := 2**(n / 2048) for such inputs that sigmoidf(-z) is
231      // normalized, i.e. 0 <= z <= 87.33642. As n has 11 fractional bits, we split s == 2**(n / 2048) =
232      // = 2**e * 2**(n / 2048 - e), where e := int(n / 2048). We create s in two steps:
233      // 1. Fetch 2**(n / 2048 - e) = 2**(n % 2048) from table using the 6 low bits of n, as integer.
234      //    Note that the fetched values are in the [1.0, 2.0) range, i.e. their floating-point exponent is 0.
235      // 2. Adjust fecthed value by addition of e to its floating-point exponent. The result is always a normalized
236      //    number, because for 0 <= z <= 87.33642 (inputs for which sigmoidf(-z) is normalized) we have -126 <= e <= 0,
237      //    and thus the adjusted exponent is not lower than -126.
238      //
239      // Extract e from bits 11:19 of n and shift it into bits 23:31 (position of floating-point exponent).
240      const uint32_t ve = (fp32_to_bits(vn) & ~vindex_mask) << 12;
241
242      // Use bits 0:11 bits of n, as integer, as an index for table lookup of l := 2**(n % 2048).
243      const uint32_t vidx = fp32_to_bits(vn) & vindex_mask;
244      // Adjust exponent of the value l fetched from the table to get the final s value.
245      const float vs = fp32_from_bits(xnn_table_exp2_k_over_2048[vidx] + ve);
246
247      // Subtract the large number back to get the final n := round(-z * 2048 / log(2)) as a floating-point number.
248      vn -= vmagic_bias;
249
250      // Compute reduced argument t := (z + n * log(2) / 2048). Note that -t = -z - n * log(2) / 2048.
251      // Use Cody-Waite range reduction method (note two constants to represent log(2) / 2048) to improve accuracy.
252      float vt = vn * vln2_o2048_hi + vz;
253      vt = vn * vln2_o2048_lo + vt;
254
255      // Compute degree-1 polynomial approximation for exp(-t) on [-log(2)/4096, log(2)/4096]:
256      //   P1(t) = 1 + t * c1
257      const float vp = vt * vc1;
258
259      // Reconstruct the exp(-z) value:
260      //   y = s * (1 + t * c1)
261      //     = s + s * (t * c1))
262      //     = s + s * p
263      const float vy = vp * vs + vs;
264
265      // Reconstruct sigmoid(-z) = exp(-z) / (1.0 + exp(-z))
266      float vf = vy / (vy + vone);
267
268      // For inputs above denormal cutoff, replace output with +0.0f.
269      // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
270      if XNN_UNPREDICTABLE(vz > vdenorm_cutoff) {
271        vf = 0.0f;
272      }
273
274      // Reconstruct sigmoid(x) = x < 0 ? sigmoid(-z) : 1.0 - sigmoid(-z)
275      if XNN_UNPREDICTABLE(vx > 0.0f) {
276        vf = vone - vf;
277      }
278
279      *y = vf;
280    }
281  $else:
282    if XNN_UNLIKELY(n != 0) {
283      do {
284        const float vx = *x++;
285
286        // General structure of the algorithm:
287        //           / exp(x) / (1 + exp(x)) if x <= 0
288        //   f[x] :=
289        //           \ 1 - f[-x] if x >= 0
290        //
291        // First we compute f[-z] := exp(-z) / (1 + exp(-z)) where z = abs(x),
292        // then replace result with 1 - f[-z] if x >= 0.
293        const float vz = fabsf(vx);
294
295        // Compute reduced argument n := round(-z * 2048 / log(2)).
296        // We do it by adding a large number (magic bias), which cause rounding of the result to an integer, then subtracing
297        // the large number back. The first addition is combined with multiplication by log2e into a single FMA instruction.
298        // The trick with adding large number is valid only within certain bounds (|z * 2048 / log(2)| <= 2**22, i.e.
299        // |z| <= 0x1.62E43p+10 = 1419.5654296875), but that is acceptable, because inputs x outside of
300        // [-87.336544, 17.328678] (i.e. z outsize [0, 87.336544]) underflow or saturate sigmoidf(x). We fixup the result
301        // for such inputs at the very end of the algorithm.
302        float vn = vz * vminus_log2e_x2048 + vmagic_bias;
303
304        // Create a floating-point number s (scale) such that s := 2**(n / 2048) for such inputs that sigmoidf(-z) is
305        // normalized, i.e. 0 <= z <= 87.33642. As n has 11 fractional bits, we split s == 2**(n / 2048) =
306        // = 2**e * 2**(n / 2048 - e), where e := int(n / 2048). We create s in two steps:
307        // 1. Fetch 2**(n / 2048 - e) = 2**(n % 2048) from table using the 6 low bits of n, as integer.
308        //    Note that the fetched values are in the [1.0, 2.0) range, i.e. their floating-point exponent is 0.
309        // 2. Adjust fecthed value by addition of e to its floating-point exponent. The result is always a normalized
310        //    number, because for 0 <= z <= 87.33642 (inputs for which sigmoidf(-z) is normalized) we have -126 <= e <= 0,
311        //    and thus the adjusted exponent is not lower than -126.
312        //
313        // Extract e from bits 11:19 of n and shift it into bits 23:31 (position of floating-point exponent).
314        const uint32_t ve = (fp32_to_bits(vn) & ~vindex_mask) << 12;
315
316        // Use bits 0:11 bits of n, as integer, as an index for table lookup of l := 2**(n % 2048).
317        const uint32_t vidx = fp32_to_bits(vn) & vindex_mask;
318        // Adjust exponent of the value l fetched from the table to get the final s value.
319        const float vs = fp32_from_bits(xnn_table_exp2_k_over_2048[vidx] + ve);
320
321        // Subtract the large number back to get the final n := round(-z * 2048 / log(2)) as a floating-point number.
322        vn -= vmagic_bias;
323
324        // Compute reduced argument t := (z + n * log(2) / 2048). Note that -t = -z - n * log(2) / 2048.
325        // Use Cody-Waite range reduction method (note two constants to represent log(2) / 2048) to improve accuracy.
326        float vt = vn * vln2_o2048_hi + vz;
327        vt = vn * vln2_o2048_lo + vt;
328
329        // Compute degree-1 polynomial approximation for exp(-t) on [-log(2)/4096, log(2)/4096]:
330        //   P1(t) = 1 + t * c1
331        const float vp = vt * vc1;
332
333        // Reconstruct the exp(-z) value:
334        //   y = s * (1 + t * c1)
335        //     = s + s * (t * c1))
336        //     = s + s * p
337        const float vy = vp * vs + vs;
338
339        // Reconstruct sigmoid(-z) = exp(-z) / (1.0 + exp(-z))
340        float vf = vy / (vy + vone);
341
342        // For inputs above denormal cutoff, replace output with +0.0f.
343        // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
344        if XNN_UNPREDICTABLE(vz > vdenorm_cutoff) {
345          vf = 0.0f;
346        }
347
348        // Reconstruct sigmoid(x) = x < 0 ? sigmoid(-z) : 1.0 - sigmoid(-z)
349        if XNN_UNPREDICTABLE(vx > 0.0f) {
350          vf = vone - vf;
351        }
352
353        *y++ = vf;
354
355        n -= sizeof(float);
356      } while (n != 0);
357    }
358}
359