1 /*
2 * Copyright © 2018 Red Hat Inc.
3 * Copyright © 2015 Intel Corporation
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
21 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
22 * IN THE SOFTWARE.
23 */
24
25 #include <math.h>
26
27 #include "nir.h"
28 #include "nir_builtin_builder.h"
29
30 nir_ssa_def*
nir_cross3(nir_builder * b,nir_ssa_def * x,nir_ssa_def * y)31 nir_cross3(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y)
32 {
33 unsigned yzx[3] = { 1, 2, 0 };
34 unsigned zxy[3] = { 2, 0, 1 };
35
36 return nir_ffma(b, nir_swizzle(b, x, yzx, 3),
37 nir_swizzle(b, y, zxy, 3),
38 nir_fneg(b, nir_fmul(b, nir_swizzle(b, x, zxy, 3),
39 nir_swizzle(b, y, yzx, 3))));
40 }
41
42 nir_ssa_def*
nir_cross4(nir_builder * b,nir_ssa_def * x,nir_ssa_def * y)43 nir_cross4(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y)
44 {
45 nir_ssa_def *cross = nir_cross3(b, x, y);
46
47 return nir_vec4(b,
48 nir_channel(b, cross, 0),
49 nir_channel(b, cross, 1),
50 nir_channel(b, cross, 2),
51 nir_imm_intN_t(b, 0, cross->bit_size));
52 }
53
54 nir_ssa_def*
nir_fast_length(nir_builder * b,nir_ssa_def * vec)55 nir_fast_length(nir_builder *b, nir_ssa_def *vec)
56 {
57 return nir_fsqrt(b, nir_fdot(b, vec, vec));
58 }
59
60 nir_ssa_def*
nir_nextafter(nir_builder * b,nir_ssa_def * x,nir_ssa_def * y)61 nir_nextafter(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y)
62 {
63 nir_ssa_def *zero = nir_imm_intN_t(b, 0, x->bit_size);
64 nir_ssa_def *one = nir_imm_intN_t(b, 1, x->bit_size);
65
66 nir_ssa_def *condeq = nir_feq(b, x, y);
67 nir_ssa_def *conddir = nir_flt(b, x, y);
68 nir_ssa_def *condzero = nir_feq(b, x, zero);
69
70 uint64_t sign_mask = 1ull << (x->bit_size - 1);
71 uint64_t min_abs = 1;
72
73 if (nir_is_denorm_flush_to_zero(b->shader->info.float_controls_execution_mode, x->bit_size)) {
74 switch (x->bit_size) {
75 case 16:
76 min_abs = 1 << 10;
77 break;
78 case 32:
79 min_abs = 1 << 23;
80 break;
81 case 64:
82 min_abs = 1ULL << 52;
83 break;
84 }
85
86 /* Flush denorm to zero to avoid returning a denorm when condeq is true. */
87 x = nir_fmul(b, x, nir_imm_floatN_t(b, 1.0, x->bit_size));
88 }
89
90 /* beware of: +/-0.0 - 1 == NaN */
91 nir_ssa_def *xn =
92 nir_bcsel(b,
93 condzero,
94 nir_imm_intN_t(b, sign_mask | min_abs, x->bit_size),
95 nir_isub(b, x, one));
96
97 /* beware of -0.0 + 1 == -0x1p-149 */
98 nir_ssa_def *xp = nir_bcsel(b, condzero,
99 nir_imm_intN_t(b, min_abs, x->bit_size),
100 nir_iadd(b, x, one));
101
102 /* nextafter can be implemented by just +/- 1 on the int value */
103 nir_ssa_def *res =
104 nir_bcsel(b, nir_ixor(b, conddir, nir_flt(b, x, zero)), xp, xn);
105
106 return nir_nan_check2(b, x, y, nir_bcsel(b, condeq, x, res));
107 }
108
109 nir_ssa_def*
nir_normalize(nir_builder * b,nir_ssa_def * vec)110 nir_normalize(nir_builder *b, nir_ssa_def *vec)
111 {
112 if (vec->num_components == 1)
113 return nir_fsign(b, vec);
114
115 nir_ssa_def *f0 = nir_imm_floatN_t(b, 0.0, vec->bit_size);
116 nir_ssa_def *f1 = nir_imm_floatN_t(b, 1.0, vec->bit_size);
117 nir_ssa_def *finf = nir_imm_floatN_t(b, INFINITY, vec->bit_size);
118
119 /* scale the input to increase precision */
120 nir_ssa_def *maxc = nir_fmax_abs_vec_comp(b, vec);
121 nir_ssa_def *svec = nir_fdiv(b, vec, maxc);
122 /* for inf */
123 nir_ssa_def *finfvec = nir_copysign(b, nir_bcsel(b, nir_feq(b, vec, finf), f1, f0), f1);
124
125 nir_ssa_def *temp = nir_bcsel(b, nir_feq(b, maxc, finf), finfvec, svec);
126 nir_ssa_def *res = nir_fmul(b, temp, nir_frsq(b, nir_fdot(b, temp, temp)));
127
128 return nir_bcsel(b, nir_feq(b, maxc, f0), vec, res);
129 }
130
131 nir_ssa_def*
nir_smoothstep(nir_builder * b,nir_ssa_def * edge0,nir_ssa_def * edge1,nir_ssa_def * x)132 nir_smoothstep(nir_builder *b, nir_ssa_def *edge0, nir_ssa_def *edge1, nir_ssa_def *x)
133 {
134 nir_ssa_def *f2 = nir_imm_floatN_t(b, 2.0, x->bit_size);
135 nir_ssa_def *f3 = nir_imm_floatN_t(b, 3.0, x->bit_size);
136
137 /* t = clamp((x - edge0) / (edge1 - edge0), 0, 1) */
138 nir_ssa_def *t =
139 nir_fsat(b, nir_fdiv(b, nir_fsub(b, x, edge0),
140 nir_fsub(b, edge1, edge0)));
141
142 /* result = t * t * (3 - 2 * t) */
143 return nir_fmul(b, t, nir_fmul(b, t, nir_a_minus_bc(b, f3, f2, t)));
144 }
145
146 nir_ssa_def*
nir_upsample(nir_builder * b,nir_ssa_def * hi,nir_ssa_def * lo)147 nir_upsample(nir_builder *b, nir_ssa_def *hi, nir_ssa_def *lo)
148 {
149 assert(lo->num_components == hi->num_components);
150 assert(lo->bit_size == hi->bit_size);
151
152 nir_ssa_def *res[NIR_MAX_VEC_COMPONENTS];
153 for (unsigned i = 0; i < lo->num_components; ++i) {
154 nir_ssa_def *vec = nir_vec2(b, nir_channel(b, lo, i), nir_channel(b, hi, i));
155 res[i] = nir_pack_bits(b, vec, vec->bit_size * 2);
156 }
157
158 return nir_vec(b, res, lo->num_components);
159 }
160
161 /**
162 * Compute xs[0] + xs[1] + xs[2] + ... using fadd.
163 */
164 static nir_ssa_def *
build_fsum(nir_builder * b,nir_ssa_def ** xs,int terms)165 build_fsum(nir_builder *b, nir_ssa_def **xs, int terms)
166 {
167 nir_ssa_def *accum = xs[0];
168
169 for (int i = 1; i < terms; i++)
170 accum = nir_fadd(b, accum, xs[i]);
171
172 return accum;
173 }
174
175 nir_ssa_def *
nir_atan(nir_builder * b,nir_ssa_def * y_over_x)176 nir_atan(nir_builder *b, nir_ssa_def *y_over_x)
177 {
178 const uint32_t bit_size = y_over_x->bit_size;
179
180 nir_ssa_def *abs_y_over_x = nir_fabs(b, y_over_x);
181 nir_ssa_def *one = nir_imm_floatN_t(b, 1.0f, bit_size);
182
183 /*
184 * range-reduction, first step:
185 *
186 * / y_over_x if |y_over_x| <= 1.0;
187 * x = <
188 * \ 1.0 / y_over_x otherwise
189 */
190 nir_ssa_def *x = nir_fdiv(b, nir_fmin(b, abs_y_over_x, one),
191 nir_fmax(b, abs_y_over_x, one));
192
193 /*
194 * approximate atan by evaluating polynomial:
195 *
196 * x * 0.9999793128310355 - x^3 * 0.3326756418091246 +
197 * x^5 * 0.1938924977115610 - x^7 * 0.1173503194786851 +
198 * x^9 * 0.0536813784310406 - x^11 * 0.0121323213173444
199 */
200 nir_ssa_def *x_2 = nir_fmul(b, x, x);
201 nir_ssa_def *x_3 = nir_fmul(b, x_2, x);
202 nir_ssa_def *x_5 = nir_fmul(b, x_3, x_2);
203 nir_ssa_def *x_7 = nir_fmul(b, x_5, x_2);
204 nir_ssa_def *x_9 = nir_fmul(b, x_7, x_2);
205 nir_ssa_def *x_11 = nir_fmul(b, x_9, x_2);
206
207 nir_ssa_def *polynomial_terms[] = {
208 nir_fmul_imm(b, x, 0.9999793128310355f),
209 nir_fmul_imm(b, x_3, -0.3326756418091246f),
210 nir_fmul_imm(b, x_5, 0.1938924977115610f),
211 nir_fmul_imm(b, x_7, -0.1173503194786851f),
212 nir_fmul_imm(b, x_9, 0.0536813784310406f),
213 nir_fmul_imm(b, x_11, -0.0121323213173444f),
214 };
215
216 nir_ssa_def *tmp =
217 build_fsum(b, polynomial_terms, ARRAY_SIZE(polynomial_terms));
218
219 /* range-reduction fixup */
220 tmp = nir_ffma(b,
221 nir_b2f(b, nir_flt(b, one, abs_y_over_x), bit_size),
222 nir_ffma_imm12(b, tmp, -2.0f, M_PI_2),
223 tmp);
224
225 /* sign fixup */
226 nir_ssa_def *result = nir_fmul(b, tmp, nir_fsign(b, y_over_x));
227
228 /* The fmin and fmax above will filter out NaN values. This leads to
229 * non-NaN results for NaN inputs. Work around this by doing
230 *
231 * !isnan(y_over_x) ? ... : y_over_x;
232 */
233 if (b->exact ||
234 nir_is_float_control_signed_zero_inf_nan_preserve(b->shader->info.float_controls_execution_mode, bit_size)) {
235 const bool exact = b->exact;
236
237 b->exact = true;
238 nir_ssa_def *is_not_nan = nir_feq(b, y_over_x, y_over_x);
239 b->exact = exact;
240
241 /* The extra 1.0*y_over_x ensures that subnormal results are flushed to
242 * zero.
243 */
244 result = nir_bcsel(b, is_not_nan, result, nir_fmul_imm(b, y_over_x, 1.0));
245 }
246
247 return result;
248 }
249
250 nir_ssa_def *
nir_atan2(nir_builder * b,nir_ssa_def * y,nir_ssa_def * x)251 nir_atan2(nir_builder *b, nir_ssa_def *y, nir_ssa_def *x)
252 {
253 assert(y->bit_size == x->bit_size);
254 const uint32_t bit_size = x->bit_size;
255
256 nir_ssa_def *zero = nir_imm_floatN_t(b, 0, bit_size);
257 nir_ssa_def *one = nir_imm_floatN_t(b, 1, bit_size);
258
259 /* If we're on the left half-plane rotate the coordinates π/2 clock-wise
260 * for the y=0 discontinuity to end up aligned with the vertical
261 * discontinuity of atan(s/t) along t=0. This also makes sure that we
262 * don't attempt to divide by zero along the vertical line, which may give
263 * unspecified results on non-GLSL 4.1-capable hardware.
264 */
265 nir_ssa_def *flip = nir_fge(b, zero, x);
266 nir_ssa_def *s = nir_bcsel(b, flip, nir_fabs(b, x), y);
267 nir_ssa_def *t = nir_bcsel(b, flip, y, nir_fabs(b, x));
268
269 /* If the magnitude of the denominator exceeds some huge value, scale down
270 * the arguments in order to prevent the reciprocal operation from flushing
271 * its result to zero, which would cause precision problems, and for s
272 * infinite would cause us to return a NaN instead of the correct finite
273 * value.
274 *
275 * If fmin and fmax are respectively the smallest and largest positive
276 * normalized floating point values representable by the implementation,
277 * the constants below should be in agreement with:
278 *
279 * huge <= 1 / fmin
280 * scale <= 1 / fmin / fmax (for |t| >= huge)
281 *
282 * In addition scale should be a negative power of two in order to avoid
283 * loss of precision. The values chosen below should work for most usual
284 * floating point representations with at least the dynamic range of ATI's
285 * 24-bit representation.
286 */
287 const double huge_val = bit_size >= 32 ? 1e18 : 16384;
288 nir_ssa_def *huge = nir_imm_floatN_t(b, huge_val, bit_size);
289 nir_ssa_def *scale = nir_bcsel(b, nir_fge(b, nir_fabs(b, t), huge),
290 nir_imm_floatN_t(b, 0.25, bit_size), one);
291 nir_ssa_def *rcp_scaled_t = nir_frcp(b, nir_fmul(b, t, scale));
292 nir_ssa_def *s_over_t = nir_fmul(b, nir_fmul(b, s, scale), rcp_scaled_t);
293
294 /* For |x| = |y| assume tan = 1 even if infinite (i.e. pretend momentarily
295 * that ∞/∞ = 1) in order to comply with the rather artificial rules
296 * inherited from IEEE 754-2008, namely:
297 *
298 * "atan2(±∞, −∞) is ±3π/4
299 * atan2(±∞, +∞) is ±π/4"
300 *
301 * Note that this is inconsistent with the rules for the neighborhood of
302 * zero that are based on iterated limits:
303 *
304 * "atan2(±0, −0) is ±π
305 * atan2(±0, +0) is ±0"
306 *
307 * but GLSL specifically allows implementations to deviate from IEEE rules
308 * at (0,0), so we take that license (i.e. pretend that 0/0 = 1 here as
309 * well).
310 */
311 nir_ssa_def *tan = nir_bcsel(b, nir_feq(b, nir_fabs(b, x), nir_fabs(b, y)),
312 one, nir_fabs(b, s_over_t));
313
314 /* Calculate the arctangent and fix up the result if we had flipped the
315 * coordinate system.
316 */
317 nir_ssa_def *arc =
318 nir_ffma_imm1(b, nir_b2f(b, flip, bit_size), M_PI_2, nir_atan(b, tan));
319
320 /* Rather convoluted calculation of the sign of the result. When x < 0 we
321 * cannot use fsign because we need to be able to distinguish between
322 * negative and positive zero. We don't use bitwise arithmetic tricks for
323 * consistency with the GLSL front-end. When x >= 0 rcp_scaled_t will
324 * always be non-negative so this won't be able to distinguish between
325 * negative and positive zero, but we don't care because atan2 is
326 * continuous along the whole positive y = 0 half-line, so it won't affect
327 * the result significantly.
328 */
329 return nir_bcsel(b, nir_flt(b, nir_fmin(b, y, rcp_scaled_t), zero),
330 nir_fneg(b, arc), arc);
331 }
332
333 nir_ssa_def *
nir_get_texture_size(nir_builder * b,nir_tex_instr * tex)334 nir_get_texture_size(nir_builder *b, nir_tex_instr *tex)
335 {
336 b->cursor = nir_before_instr(&tex->instr);
337
338 nir_tex_instr *txs;
339
340 unsigned num_srcs = 1; /* One for the LOD */
341 for (unsigned i = 0; i < tex->num_srcs; i++) {
342 if (tex->src[i].src_type == nir_tex_src_texture_deref ||
343 tex->src[i].src_type == nir_tex_src_sampler_deref ||
344 tex->src[i].src_type == nir_tex_src_texture_offset ||
345 tex->src[i].src_type == nir_tex_src_sampler_offset ||
346 tex->src[i].src_type == nir_tex_src_texture_handle ||
347 tex->src[i].src_type == nir_tex_src_sampler_handle)
348 num_srcs++;
349 }
350
351 txs = nir_tex_instr_create(b->shader, num_srcs);
352 txs->op = nir_texop_txs;
353 txs->sampler_dim = tex->sampler_dim;
354 txs->is_array = tex->is_array;
355 txs->is_shadow = tex->is_shadow;
356 txs->is_new_style_shadow = tex->is_new_style_shadow;
357 txs->texture_index = tex->texture_index;
358 txs->sampler_index = tex->sampler_index;
359 txs->dest_type = nir_type_int32;
360
361 unsigned idx = 0;
362 for (unsigned i = 0; i < tex->num_srcs; i++) {
363 if (tex->src[i].src_type == nir_tex_src_texture_deref ||
364 tex->src[i].src_type == nir_tex_src_sampler_deref ||
365 tex->src[i].src_type == nir_tex_src_texture_offset ||
366 tex->src[i].src_type == nir_tex_src_sampler_offset ||
367 tex->src[i].src_type == nir_tex_src_texture_handle ||
368 tex->src[i].src_type == nir_tex_src_sampler_handle) {
369 nir_src_copy(&txs->src[idx].src, &tex->src[i].src);
370 txs->src[idx].src_type = tex->src[i].src_type;
371 idx++;
372 }
373 }
374 /* Add in an LOD because some back-ends require it */
375 txs->src[idx].src = nir_src_for_ssa(nir_imm_int(b, 0));
376 txs->src[idx].src_type = nir_tex_src_lod;
377
378 nir_ssa_dest_init(&txs->instr, &txs->dest,
379 nir_tex_instr_dest_size(txs), 32, NULL);
380 nir_builder_instr_insert(b, &txs->instr);
381
382 return &txs->dest.ssa;
383 }
384
385 nir_ssa_def *
nir_get_texture_lod(nir_builder * b,nir_tex_instr * tex)386 nir_get_texture_lod(nir_builder *b, nir_tex_instr *tex)
387 {
388 b->cursor = nir_before_instr(&tex->instr);
389
390 nir_tex_instr *tql;
391
392 unsigned num_srcs = 0;
393 for (unsigned i = 0; i < tex->num_srcs; i++) {
394 if (tex->src[i].src_type == nir_tex_src_coord ||
395 tex->src[i].src_type == nir_tex_src_texture_deref ||
396 tex->src[i].src_type == nir_tex_src_sampler_deref ||
397 tex->src[i].src_type == nir_tex_src_texture_offset ||
398 tex->src[i].src_type == nir_tex_src_sampler_offset ||
399 tex->src[i].src_type == nir_tex_src_texture_handle ||
400 tex->src[i].src_type == nir_tex_src_sampler_handle)
401 num_srcs++;
402 }
403
404 tql = nir_tex_instr_create(b->shader, num_srcs);
405 tql->op = nir_texop_lod;
406 tql->coord_components = tex->coord_components;
407 tql->sampler_dim = tex->sampler_dim;
408 tql->is_array = tex->is_array;
409 tql->is_shadow = tex->is_shadow;
410 tql->is_new_style_shadow = tex->is_new_style_shadow;
411 tql->texture_index = tex->texture_index;
412 tql->sampler_index = tex->sampler_index;
413 tql->dest_type = nir_type_float32;
414
415 unsigned idx = 0;
416 for (unsigned i = 0; i < tex->num_srcs; i++) {
417 if (tex->src[i].src_type == nir_tex_src_coord ||
418 tex->src[i].src_type == nir_tex_src_texture_deref ||
419 tex->src[i].src_type == nir_tex_src_sampler_deref ||
420 tex->src[i].src_type == nir_tex_src_texture_offset ||
421 tex->src[i].src_type == nir_tex_src_sampler_offset ||
422 tex->src[i].src_type == nir_tex_src_texture_handle ||
423 tex->src[i].src_type == nir_tex_src_sampler_handle) {
424 nir_src_copy(&tql->src[idx].src, &tex->src[i].src);
425 tql->src[idx].src_type = tex->src[i].src_type;
426 idx++;
427 }
428 }
429
430 nir_ssa_dest_init(&tql->instr, &tql->dest, 2, 32, NULL);
431 nir_builder_instr_insert(b, &tql->instr);
432
433 /* The LOD is the y component of the result */
434 return nir_channel(b, &tql->dest.ssa, 1);
435 }
436