• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright © 2015 Intel Corporation
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice (including the next
12  * paragraph) shall be included in all copies or substantial portions of the
13  * Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21  * IN THE SOFTWARE.
22  *
23  */
24 
25 #include "nir.h"
26 #include "nir_builder.h"
27 #include "c99_math.h"
28 
29 /*
30  * Lowers some unsupported double operations, using only:
31  *
32  * - pack/unpackDouble2x32
33  * - conversion to/from single-precision
34  * - double add, mul, and fma
35  * - conditional select
36  * - 32-bit integer and floating point arithmetic
37  */
38 
39 /* Creates a double with the exponent bits set to a given integer value */
40 static nir_ssa_def *
set_exponent(nir_builder * b,nir_ssa_def * src,nir_ssa_def * exp)41 set_exponent(nir_builder *b, nir_ssa_def *src, nir_ssa_def *exp)
42 {
43    /* Split into bits 0-31 and 32-63 */
44    nir_ssa_def *lo = nir_unpack_64_2x32_split_x(b, src);
45    nir_ssa_def *hi = nir_unpack_64_2x32_split_y(b, src);
46 
47    /* The exponent is bits 52-62, or 20-30 of the high word, so set the exponent
48     * to 1023
49     */
50    nir_ssa_def *new_hi = nir_bfi(b, nir_imm_int(b, 0x7ff00000), exp, hi);
51    /* recombine */
52    return nir_pack_64_2x32_split(b, lo, new_hi);
53 }
54 
55 static nir_ssa_def *
get_exponent(nir_builder * b,nir_ssa_def * src)56 get_exponent(nir_builder *b, nir_ssa_def *src)
57 {
58    /* get bits 32-63 */
59    nir_ssa_def *hi = nir_unpack_64_2x32_split_y(b, src);
60 
61    /* extract bits 20-30 of the high word */
62    return nir_ubitfield_extract(b, hi, nir_imm_int(b, 20), nir_imm_int(b, 11));
63 }
64 
65 /* Return infinity with the sign of the given source which is +/-0 */
66 
67 static nir_ssa_def *
get_signed_inf(nir_builder * b,nir_ssa_def * zero)68 get_signed_inf(nir_builder *b, nir_ssa_def *zero)
69 {
70    nir_ssa_def *zero_hi = nir_unpack_64_2x32_split_y(b, zero);
71 
72    /* The bit pattern for infinity is 0x7ff0000000000000, where the sign bit
73     * is the highest bit. Only the sign bit can be non-zero in the passed in
74     * source. So we essentially need to OR the infinity and the zero, except
75     * the low 32 bits are always 0 so we can construct the correct high 32
76     * bits and then pack it together with zero low 32 bits.
77     */
78    nir_ssa_def *inf_hi = nir_ior(b, nir_imm_int(b, 0x7ff00000), zero_hi);
79    return nir_pack_64_2x32_split(b, nir_imm_int(b, 0), inf_hi);
80 }
81 
82 /*
83  * Generates the correctly-signed infinity if the source was zero, and flushes
84  * the result to 0 if the source was infinity or the calculated exponent was
85  * too small to be representable.
86  */
87 
88 static nir_ssa_def *
fix_inv_result(nir_builder * b,nir_ssa_def * res,nir_ssa_def * src,nir_ssa_def * exp)89 fix_inv_result(nir_builder *b, nir_ssa_def *res, nir_ssa_def *src,
90                nir_ssa_def *exp)
91 {
92    /* If the exponent is too small or the original input was infinity/NaN,
93     * force the result to 0 (flush denorms) to avoid the work of handling
94     * denorms properly. Note that this doesn't preserve positive/negative
95     * zeros, but GLSL doesn't require it.
96     */
97    res = nir_bcsel(b, nir_ior(b, nir_ige(b, nir_imm_int(b, 0), exp),
98                               nir_feq(b, nir_fabs(b, src),
99                                       nir_imm_double(b, INFINITY))),
100                    nir_imm_double(b, 0.0f), res);
101 
102    /* If the original input was 0, generate the correctly-signed infinity */
103    res = nir_bcsel(b, nir_fne(b, src, nir_imm_double(b, 0.0f)),
104                    res, get_signed_inf(b, src));
105 
106    return res;
107 
108 }
109 
110 static nir_ssa_def *
lower_rcp(nir_builder * b,nir_ssa_def * src)111 lower_rcp(nir_builder *b, nir_ssa_def *src)
112 {
113    /* normalize the input to avoid range issues */
114    nir_ssa_def *src_norm = set_exponent(b, src, nir_imm_int(b, 1023));
115 
116    /* cast to float, do an rcp, and then cast back to get an approximate
117     * result
118     */
119    nir_ssa_def *ra = nir_f2f64(b, nir_frcp(b, nir_f2f32(b, src_norm)));
120 
121    /* Fixup the exponent of the result - note that we check if this is too
122     * small below.
123     */
124    nir_ssa_def *new_exp = nir_isub(b, get_exponent(b, ra),
125                                    nir_isub(b, get_exponent(b, src),
126                                             nir_imm_int(b, 1023)));
127 
128    ra = set_exponent(b, ra, new_exp);
129 
130    /* Do a few Newton-Raphson steps to improve precision.
131     *
132     * Each step doubles the precision, and we started off with around 24 bits,
133     * so we only need to do 2 steps to get to full precision. The step is:
134     *
135     * x_new = x * (2 - x*src)
136     *
137     * But we can re-arrange this to improve precision by using another fused
138     * multiply-add:
139     *
140     * x_new = x + x * (1 - x*src)
141     *
142     * See https://en.wikipedia.org/wiki/Division_algorithm for more details.
143     */
144 
145    ra = nir_ffma(b, ra, nir_ffma(b, ra, src, nir_imm_double(b, -1)), ra);
146    ra = nir_ffma(b, ra, nir_ffma(b, ra, src, nir_imm_double(b, -1)), ra);
147 
148    return fix_inv_result(b, ra, src, new_exp);
149 }
150 
151 static nir_ssa_def *
lower_sqrt_rsq(nir_builder * b,nir_ssa_def * src,bool sqrt)152 lower_sqrt_rsq(nir_builder *b, nir_ssa_def *src, bool sqrt)
153 {
154    /* We want to compute:
155     *
156     * 1/sqrt(m * 2^e)
157     *
158     * When the exponent is even, this is equivalent to:
159     *
160     * 1/sqrt(m) * 2^(-e/2)
161     *
162     * and then the exponent is odd, this is equal to:
163     *
164     * 1/sqrt(m * 2) * 2^(-(e - 1)/2)
165     *
166     * where the m * 2 is absorbed into the exponent. So we want the exponent
167     * inside the square root to be 1 if e is odd and 0 if e is even, and we
168     * want to subtract off e/2 from the final exponent, rounded to negative
169     * infinity. We can do the former by first computing the unbiased exponent,
170     * and then AND'ing it with 1 to get 0 or 1, and we can do the latter by
171     * shifting right by 1.
172     */
173 
174    nir_ssa_def *unbiased_exp = nir_isub(b, get_exponent(b, src),
175                                         nir_imm_int(b, 1023));
176    nir_ssa_def *even = nir_iand(b, unbiased_exp, nir_imm_int(b, 1));
177    nir_ssa_def *half = nir_ishr(b, unbiased_exp, nir_imm_int(b, 1));
178 
179    nir_ssa_def *src_norm = set_exponent(b, src,
180                                         nir_iadd(b, nir_imm_int(b, 1023),
181                                                  even));
182 
183    nir_ssa_def *ra = nir_f2f64(b, nir_frsq(b, nir_f2f32(b, src_norm)));
184    nir_ssa_def *new_exp = nir_isub(b, get_exponent(b, ra), half);
185    ra = set_exponent(b, ra, new_exp);
186 
187    /*
188     * The following implements an iterative algorithm that's very similar
189     * between sqrt and rsqrt. We start with an iteration of Goldschmit's
190     * algorithm, which looks like:
191     *
192     * a = the source
193     * y_0 = initial (single-precision) rsqrt estimate
194     *
195     * h_0 = .5 * y_0
196     * g_0 = a * y_0
197     * r_0 = .5 - h_0 * g_0
198     * g_1 = g_0 * r_0 + g_0
199     * h_1 = h_0 * r_0 + h_0
200     *
201     * Now g_1 ~= sqrt(a), and h_1 ~= 1/(2 * sqrt(a)). We could continue
202     * applying another round of Goldschmit, but since we would never refer
203     * back to a (the original source), we would add too much rounding error.
204     * So instead, we do one last round of Newton-Raphson, which has better
205     * rounding characteristics, to get the final rounding correct. This is
206     * split into two cases:
207     *
208     * 1. sqrt
209     *
210     * Normally, doing a round of Newton-Raphson for sqrt involves taking a
211     * reciprocal of the original estimate, which is slow since it isn't
212     * supported in HW. But we can take advantage of the fact that we already
213     * computed a good estimate of 1/(2 * g_1) by rearranging it like so:
214     *
215     * g_2 = .5 * (g_1 + a / g_1)
216     *     = g_1 + .5 * (a / g_1 - g_1)
217     *     = g_1 + (.5 / g_1) * (a - g_1^2)
218     *     = g_1 + h_1 * (a - g_1^2)
219     *
220     * The second term represents the error, and by splitting it out we can get
221     * better precision by computing it as part of a fused multiply-add. Since
222     * both Newton-Raphson and Goldschmit approximately double the precision of
223     * the result, these two steps should be enough.
224     *
225     * 2. rsqrt
226     *
227     * First off, note that the first round of the Goldschmit algorithm is
228     * really just a Newton-Raphson step in disguise:
229     *
230     * h_1 = h_0 * (.5 - h_0 * g_0) + h_0
231     *     = h_0 * (1.5 - h_0 * g_0)
232     *     = h_0 * (1.5 - .5 * a * y_0^2)
233     *     = (.5 * y_0) * (1.5 - .5 * a * y_0^2)
234     *
235     * which is the standard formula multiplied by .5. Unlike in the sqrt case,
236     * we don't need the inverse to do a Newton-Raphson step; we just need h_1,
237     * so we can skip the calculation of g_1. Instead, we simply do another
238     * Newton-Raphson step:
239     *
240     * y_1 = 2 * h_1
241     * r_1 = .5 - h_1 * y_1 * a
242     * y_2 = y_1 * r_1 + y_1
243     *
244     * Where the difference from Goldschmit is that we calculate y_1 * a
245     * instead of using g_1. Doing it this way should be as fast as computing
246     * y_1 up front instead of h_1, and it lets us share the code for the
247     * initial Goldschmit step with the sqrt case.
248     *
249     * Putting it together, the computations are:
250     *
251     * h_0 = .5 * y_0
252     * g_0 = a * y_0
253     * r_0 = .5 - h_0 * g_0
254     * h_1 = h_0 * r_0 + h_0
255     * if sqrt:
256     *    g_1 = g_0 * r_0 + g_0
257     *    r_1 = a - g_1 * g_1
258     *    g_2 = h_1 * r_1 + g_1
259     * else:
260     *    y_1 = 2 * h_1
261     *    r_1 = .5 - y_1 * (h_1 * a)
262     *    y_2 = y_1 * r_1 + y_1
263     *
264     * For more on the ideas behind this, see "Software Division and Square
265     * Root Using Goldschmit's Algorithms" by Markstein and the Wikipedia page
266     * on square roots
267     * (https://en.wikipedia.org/wiki/Methods_of_computing_square_roots).
268     */
269 
270    nir_ssa_def *one_half = nir_imm_double(b, 0.5);
271    nir_ssa_def *h_0 = nir_fmul(b, one_half, ra);
272    nir_ssa_def *g_0 = nir_fmul(b, src, ra);
273    nir_ssa_def *r_0 = nir_ffma(b, nir_fneg(b, h_0), g_0, one_half);
274    nir_ssa_def *h_1 = nir_ffma(b, h_0, r_0, h_0);
275    nir_ssa_def *res;
276    if (sqrt) {
277       nir_ssa_def *g_1 = nir_ffma(b, g_0, r_0, g_0);
278       nir_ssa_def *r_1 = nir_ffma(b, nir_fneg(b, g_1), g_1, src);
279       res = nir_ffma(b, h_1, r_1, g_1);
280    } else {
281       nir_ssa_def *y_1 = nir_fmul(b, nir_imm_double(b, 2.0), h_1);
282       nir_ssa_def *r_1 = nir_ffma(b, nir_fneg(b, y_1), nir_fmul(b, h_1, src),
283                                   one_half);
284       res = nir_ffma(b, y_1, r_1, y_1);
285    }
286 
287    if (sqrt) {
288       /* Here, the special cases we need to handle are
289        * 0 -> 0 and
290        * +inf -> +inf
291        */
292       res = nir_bcsel(b, nir_ior(b, nir_feq(b, src, nir_imm_double(b, 0.0)),
293                                  nir_feq(b, src, nir_imm_double(b, INFINITY))),
294                                  src, res);
295    } else {
296       res = fix_inv_result(b, res, src, new_exp);
297    }
298 
299    return res;
300 }
301 
302 static nir_ssa_def *
lower_trunc(nir_builder * b,nir_ssa_def * src)303 lower_trunc(nir_builder *b, nir_ssa_def *src)
304 {
305    nir_ssa_def *unbiased_exp = nir_isub(b, get_exponent(b, src),
306                                         nir_imm_int(b, 1023));
307 
308    nir_ssa_def *frac_bits = nir_isub(b, nir_imm_int(b, 52), unbiased_exp);
309 
310    /*
311     * Decide the operation to apply depending on the unbiased exponent:
312     *
313     * if (unbiased_exp < 0)
314     *    return 0
315     * else if (unbiased_exp > 52)
316     *    return src
317     * else
318     *    return src & (~0 << frac_bits)
319     *
320     * Notice that the else branch is a 64-bit integer operation that we need
321     * to implement in terms of 32-bit integer arithmetics (at least until we
322     * support 64-bit integer arithmetics).
323     */
324 
325    /* Compute "~0 << frac_bits" in terms of hi/lo 32-bit integer math */
326    nir_ssa_def *mask_lo =
327       nir_bcsel(b,
328                 nir_ige(b, frac_bits, nir_imm_int(b, 32)),
329                 nir_imm_int(b, 0),
330                 nir_ishl(b, nir_imm_int(b, ~0), frac_bits));
331 
332    nir_ssa_def *mask_hi =
333       nir_bcsel(b,
334                 nir_ilt(b, frac_bits, nir_imm_int(b, 33)),
335                 nir_imm_int(b, ~0),
336                 nir_ishl(b,
337                          nir_imm_int(b, ~0),
338                          nir_isub(b, frac_bits, nir_imm_int(b, 32))));
339 
340    nir_ssa_def *src_lo = nir_unpack_64_2x32_split_x(b, src);
341    nir_ssa_def *src_hi = nir_unpack_64_2x32_split_y(b, src);
342 
343    return
344       nir_bcsel(b,
345                 nir_ilt(b, unbiased_exp, nir_imm_int(b, 0)),
346                 nir_imm_double(b, 0.0),
347                 nir_bcsel(b, nir_ige(b, unbiased_exp, nir_imm_int(b, 53)),
348                           src,
349                           nir_pack_64_2x32_split(b,
350                                                  nir_iand(b, mask_lo, src_lo),
351                                                  nir_iand(b, mask_hi, src_hi))));
352 }
353 
354 static nir_ssa_def *
lower_floor(nir_builder * b,nir_ssa_def * src)355 lower_floor(nir_builder *b, nir_ssa_def *src)
356 {
357    /*
358     * For x >= 0, floor(x) = trunc(x)
359     * For x < 0,
360     *    - if x is integer, floor(x) = x
361     *    - otherwise, floor(x) = trunc(x) - 1
362     */
363    nir_ssa_def *tr = nir_ftrunc(b, src);
364    nir_ssa_def *positive = nir_fge(b, src, nir_imm_double(b, 0.0));
365    return nir_bcsel(b,
366                     nir_ior(b, positive, nir_feq(b, src, tr)),
367                     tr,
368                     nir_fsub(b, tr, nir_imm_double(b, 1.0)));
369 }
370 
371 static nir_ssa_def *
lower_ceil(nir_builder * b,nir_ssa_def * src)372 lower_ceil(nir_builder *b, nir_ssa_def *src)
373 {
374    /* if x < 0,                    ceil(x) = trunc(x)
375     * else if (x - trunc(x) == 0), ceil(x) = x
376     * else,                        ceil(x) = trunc(x) + 1
377     */
378    nir_ssa_def *tr = nir_ftrunc(b, src);
379    nir_ssa_def *negative = nir_flt(b, src, nir_imm_double(b, 0.0));
380    return nir_bcsel(b,
381                     nir_ior(b, negative, nir_feq(b, src, tr)),
382                     tr,
383                     nir_fadd(b, tr, nir_imm_double(b, 1.0)));
384 }
385 
386 static nir_ssa_def *
lower_fract(nir_builder * b,nir_ssa_def * src)387 lower_fract(nir_builder *b, nir_ssa_def *src)
388 {
389    return nir_fsub(b, src, nir_ffloor(b, src));
390 }
391 
392 static nir_ssa_def *
lower_round_even(nir_builder * b,nir_ssa_def * src)393 lower_round_even(nir_builder *b, nir_ssa_def *src)
394 {
395    /* If fract(src) == 0.5, then we will have to decide the rounding direction.
396     * We will do this by computing the mod(abs(src), 2) and testing if it
397     * is < 1 or not.
398     *
399     * We compute mod(abs(src), 2) as:
400     * abs(src) - 2.0 * floor(abs(src) / 2.0)
401     */
402    nir_ssa_def *two = nir_imm_double(b, 2.0);
403    nir_ssa_def *abs_src = nir_fabs(b, src);
404    nir_ssa_def *mod =
405       nir_fsub(b,
406                abs_src,
407                nir_fmul(b,
408                         two,
409                         nir_ffloor(b,
410                                    nir_fmul(b,
411                                             abs_src,
412                                             nir_imm_double(b, 0.5)))));
413 
414    /*
415     * If fract(src) != 0.5, then we round as floor(src + 0.5)
416     *
417     * If fract(src) == 0.5, then we have to check the modulo:
418     *
419     *   if it is < 1 we need a trunc operation so we get:
420     *      0.5 -> 0,   -0.5 -> -0
421     *      2.5 -> 2,   -2.5 -> -2
422     *
423     *   otherwise we need to check if src >= 0, in which case we need to round
424     *   upwards, or not, in which case we need to round downwards so we get:
425     *      1.5 -> 2,   -1.5 -> -2
426     *      3.5 -> 4,   -3.5 -> -4
427     */
428    nir_ssa_def *fract = nir_ffract(b, src);
429    return nir_bcsel(b,
430                     nir_fne(b, fract, nir_imm_double(b, 0.5)),
431                     nir_ffloor(b, nir_fadd(b, src, nir_imm_double(b, 0.5))),
432                     nir_bcsel(b,
433                               nir_flt(b, mod, nir_imm_double(b, 1.0)),
434                               nir_ftrunc(b, src),
435                               nir_bcsel(b,
436                                         nir_fge(b, src, nir_imm_double(b, 0.0)),
437                                         nir_fadd(b, src, nir_imm_double(b, 0.5)),
438                                         nir_fsub(b, src, nir_imm_double(b, 0.5)))));
439 }
440 
441 static nir_ssa_def *
lower_mod(nir_builder * b,nir_ssa_def * src0,nir_ssa_def * src1)442 lower_mod(nir_builder *b, nir_ssa_def *src0, nir_ssa_def *src1)
443 {
444    /* mod(x,y) = x - y * floor(x/y)
445     *
446     * If the division is lowered, it could add some rounding errors that make
447     * floor() to return the quotient minus one when x = N * y. If this is the
448     * case, we return zero because mod(x, y) output value is [0, y).
449     */
450    nir_ssa_def *floor = nir_ffloor(b, nir_fdiv(b, src0, src1));
451    nir_ssa_def *mod = nir_fsub(b, src0, nir_fmul(b, src1, floor));
452 
453    return nir_bcsel(b,
454                     nir_fne(b, mod, src1),
455                     mod,
456                     nir_imm_double(b, 0.0));
457 }
458 
459 static bool
lower_doubles_instr(nir_alu_instr * instr,nir_lower_doubles_options options)460 lower_doubles_instr(nir_alu_instr *instr, nir_lower_doubles_options options)
461 {
462    assert(instr->dest.dest.is_ssa);
463    if (instr->dest.dest.ssa.bit_size != 64)
464       return false;
465 
466    switch (instr->op) {
467    case nir_op_frcp:
468       if (!(options & nir_lower_drcp))
469          return false;
470       break;
471 
472    case nir_op_fsqrt:
473       if (!(options & nir_lower_dsqrt))
474          return false;
475       break;
476 
477    case nir_op_frsq:
478       if (!(options & nir_lower_drsq))
479          return false;
480       break;
481 
482    case nir_op_ftrunc:
483       if (!(options & nir_lower_dtrunc))
484          return false;
485       break;
486 
487    case nir_op_ffloor:
488       if (!(options & nir_lower_dfloor))
489          return false;
490       break;
491 
492    case nir_op_fceil:
493       if (!(options & nir_lower_dceil))
494          return false;
495       break;
496 
497    case nir_op_ffract:
498       if (!(options & nir_lower_dfract))
499          return false;
500       break;
501 
502    case nir_op_fround_even:
503       if (!(options & nir_lower_dround_even))
504          return false;
505       break;
506 
507    case nir_op_fmod:
508       if (!(options & nir_lower_dmod))
509          return false;
510       break;
511 
512    default:
513       return false;
514    }
515 
516    nir_builder bld;
517    nir_builder_init(&bld, nir_cf_node_get_function(&instr->instr.block->cf_node));
518    bld.cursor = nir_before_instr(&instr->instr);
519 
520    nir_ssa_def *src = nir_fmov_alu(&bld, instr->src[0],
521                                    instr->dest.dest.ssa.num_components);
522 
523    nir_ssa_def *result;
524 
525    switch (instr->op) {
526    case nir_op_frcp:
527       result = lower_rcp(&bld, src);
528       break;
529    case nir_op_fsqrt:
530       result = lower_sqrt_rsq(&bld, src, true);
531       break;
532    case nir_op_frsq:
533       result = lower_sqrt_rsq(&bld, src, false);
534       break;
535    case nir_op_ftrunc:
536       result = lower_trunc(&bld, src);
537       break;
538    case nir_op_ffloor:
539       result = lower_floor(&bld, src);
540       break;
541    case nir_op_fceil:
542       result = lower_ceil(&bld, src);
543       break;
544    case nir_op_ffract:
545       result = lower_fract(&bld, src);
546       break;
547    case nir_op_fround_even:
548       result = lower_round_even(&bld, src);
549       break;
550 
551    case nir_op_fmod: {
552       nir_ssa_def *src1 = nir_fmov_alu(&bld, instr->src[1],
553                                       instr->dest.dest.ssa.num_components);
554       result = lower_mod(&bld, src, src1);
555    }
556       break;
557    default:
558       unreachable("unhandled opcode");
559    }
560 
561    nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(result));
562    nir_instr_remove(&instr->instr);
563    return true;
564 }
565 
566 static bool
nir_lower_doubles_impl(nir_function_impl * impl,nir_lower_doubles_options options)567 nir_lower_doubles_impl(nir_function_impl *impl,
568                        nir_lower_doubles_options options)
569 {
570    bool progress = false;
571 
572    nir_foreach_block(block, impl) {
573       nir_foreach_instr_safe(instr, block) {
574          if (instr->type == nir_instr_type_alu)
575             progress |= lower_doubles_instr(nir_instr_as_alu(instr),
576                                             options);
577       }
578    }
579 
580    if (progress)
581       nir_metadata_preserve(impl, nir_metadata_block_index |
582                                   nir_metadata_dominance);
583 
584    return progress;
585 }
586 
587 bool
nir_lower_doubles(nir_shader * shader,nir_lower_doubles_options options)588 nir_lower_doubles(nir_shader *shader, nir_lower_doubles_options options)
589 {
590    bool progress = false;
591 
592    nir_foreach_function(function, shader) {
593       if (function->impl) {
594          progress |= nir_lower_doubles_impl(function->impl, options);
595       }
596    }
597 
598    return progress;
599 }
600