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 #include <math.h>
25
26 #include "nir/nir_builtin_builder.h"
27
28 #include "vtn_private.h"
29 #include "GLSL.std.450.h"
30
31 #ifndef M_PIf
32 #define M_PIf ((float) M_PI)
33 #endif
34 #ifndef M_PI_2f
35 #define M_PI_2f ((float) M_PI_2)
36 #endif
37 #ifndef M_PI_4f
38 #define M_PI_4f ((float) M_PI_4)
39 #endif
40
41 static nir_def *build_det(nir_builder *b, nir_def **col, unsigned cols);
42
43 /* Computes the determinate of the submatrix given by taking src and
44 * removing the specified row and column.
45 */
46 static nir_def *
build_mat_subdet(struct nir_builder * b,struct nir_def ** src,unsigned size,unsigned row,unsigned col)47 build_mat_subdet(struct nir_builder *b, struct nir_def **src,
48 unsigned size, unsigned row, unsigned col)
49 {
50 assert(row < size && col < size);
51 if (size == 2) {
52 return nir_channel(b, src[1 - col], 1 - row);
53 } else {
54 /* Swizzle to get all but the specified row */
55 unsigned swiz[NIR_MAX_VEC_COMPONENTS] = {0};
56 for (unsigned j = 0; j < 3; j++)
57 swiz[j] = j + (j >= row);
58
59 /* Grab all but the specified column */
60 nir_def *subcol[3];
61 for (unsigned j = 0; j < size; j++) {
62 if (j != col) {
63 subcol[j - (j > col)] = nir_swizzle(b, src[j], swiz, size - 1);
64 }
65 }
66
67 return build_det(b, subcol, size - 1);
68 }
69 }
70
71 static nir_def *
build_det(nir_builder * b,nir_def ** col,unsigned size)72 build_det(nir_builder *b, nir_def **col, unsigned size)
73 {
74 assert(size <= 4);
75 nir_def *subdet[4];
76 for (unsigned i = 0; i < size; i++)
77 subdet[i] = build_mat_subdet(b, col, size, i, 0);
78
79 nir_def *prod = nir_fmul(b, col[0], nir_vec(b, subdet, size));
80
81 nir_def *result = NULL;
82 for (unsigned i = 0; i < size; i += 2) {
83 nir_def *term;
84 if (i + 1 < size) {
85 term = nir_fsub(b, nir_channel(b, prod, i),
86 nir_channel(b, prod, i + 1));
87 } else {
88 term = nir_channel(b, prod, i);
89 }
90
91 result = result ? nir_fadd(b, result, term) : term;
92 }
93
94 return result;
95 }
96
97 static nir_def *
build_mat_det(struct vtn_builder * b,struct vtn_ssa_value * src)98 build_mat_det(struct vtn_builder *b, struct vtn_ssa_value *src)
99 {
100 unsigned size = glsl_get_vector_elements(src->type);
101
102 nir_def *cols[4];
103 for (unsigned i = 0; i < size; i++)
104 cols[i] = src->elems[i]->def;
105
106 return build_det(&b->nb, cols, size);
107 }
108
109 static struct vtn_ssa_value *
matrix_inverse(struct vtn_builder * b,struct vtn_ssa_value * src)110 matrix_inverse(struct vtn_builder *b, struct vtn_ssa_value *src)
111 {
112 nir_def *adj_col[4];
113 unsigned size = glsl_get_vector_elements(src->type);
114
115 nir_def *cols[4];
116 for (unsigned i = 0; i < size; i++)
117 cols[i] = src->elems[i]->def;
118
119 /* Build up an adjugate matrix */
120 for (unsigned c = 0; c < size; c++) {
121 nir_def *elem[4];
122 for (unsigned r = 0; r < size; r++) {
123 elem[r] = build_mat_subdet(&b->nb, cols, size, c, r);
124
125 if ((r + c) % 2)
126 elem[r] = nir_fneg(&b->nb, elem[r]);
127 }
128
129 adj_col[c] = nir_vec(&b->nb, elem, size);
130 }
131
132 nir_def *det_inv = nir_frcp(&b->nb, build_det(&b->nb, cols, size));
133
134 struct vtn_ssa_value *val = vtn_create_ssa_value(b, src->type);
135 for (unsigned i = 0; i < size; i++)
136 val->elems[i]->def = nir_fmul(&b->nb, adj_col[i], det_inv);
137
138 return val;
139 }
140
141 /**
142 * Approximate asin(x) by the piecewise formula:
143 * for |x| < 0.5, asin~(x) = x * (1 + x²(pS0 + x²(pS1 + x²*pS2)) / (1 + x²*qS1))
144 * for |x| ≥ 0.5, asin~(x) = sign(x) * (π/2 - sqrt(1 - |x|) * (π/2 + |x|(π/4 - 1 + |x|(p0 + |x|p1))))
145 *
146 * The latter is correct to first order at x=0 and x=±1 regardless of the p
147 * coefficients but can be made second-order correct at both ends by selecting
148 * the fit coefficients appropriately. Different p coefficients can be used
149 * in the asin and acos implementation to minimize some relative error metric
150 * in each case.
151 */
152 static nir_def *
build_asin(nir_builder * b,nir_def * x,float p0,float p1,bool piecewise)153 build_asin(nir_builder *b, nir_def *x, float p0, float p1, bool piecewise)
154 {
155 if (x->bit_size == 16) {
156 /* The polynomial approximation isn't precise enough to meet half-float
157 * precision requirements. Alternatively, we could implement this using
158 * the formula:
159 *
160 * asin(x) = atan2(x, sqrt(1 - x*x))
161 *
162 * But that is very expensive, so instead we just do the polynomial
163 * approximation in 32-bit math and then we convert the result back to
164 * 16-bit.
165 */
166 return nir_f2f16(b, build_asin(b, nir_f2f32(b, x), p0, p1, piecewise));
167 }
168 nir_def *one = nir_imm_floatN_t(b, 1.0f, x->bit_size);
169 nir_def *half = nir_imm_floatN_t(b, 0.5f, x->bit_size);
170 nir_def *abs_x = nir_fabs(b, x);
171
172 nir_def *p0_plus_xp1 = nir_ffma_imm12(b, abs_x, p1, p0);
173
174 nir_def *expr_tail =
175 nir_ffma_imm2(b, abs_x,
176 nir_ffma_imm2(b, abs_x, p0_plus_xp1, M_PI_4f - 1.0f),
177 M_PI_2f);
178
179 nir_def *result0 = nir_fmul(b, nir_fsign(b, x),
180 nir_a_minus_bc(b, nir_imm_floatN_t(b, M_PI_2f, x->bit_size),
181 nir_fsqrt(b, nir_fsub(b, one, abs_x)),
182 expr_tail));
183 if (piecewise) {
184 /* approximation for |x| < 0.5 */
185 const float pS0 = 1.6666586697e-01f;
186 const float pS1 = -4.2743422091e-02f;
187 const float pS2 = -8.6563630030e-03f;
188 const float qS1 = -7.0662963390e-01f;
189
190 nir_def *x2 = nir_fmul(b, x, x);
191 nir_def *p = nir_fmul(b,
192 x2,
193 nir_ffma_imm2(b, x2,
194 nir_ffma_imm12(b, x2, pS2, pS1),
195 pS0));
196
197 nir_def *q = nir_ffma_imm1(b, x2, qS1, one);
198 nir_def *result1 = nir_ffma(b, x, nir_fdiv(b, p, q), x);
199 return nir_bcsel(b, nir_flt(b, abs_x, half), result1, result0);
200 } else {
201 return result0;
202 }
203 }
204
205 static nir_op
vtn_nir_alu_op_for_spirv_glsl_opcode(struct vtn_builder * b,enum GLSLstd450 opcode,unsigned execution_mode,bool * exact)206 vtn_nir_alu_op_for_spirv_glsl_opcode(struct vtn_builder *b,
207 enum GLSLstd450 opcode,
208 unsigned execution_mode,
209 bool *exact)
210 {
211 *exact = false;
212 switch (opcode) {
213 case GLSLstd450Round: return nir_op_fround_even;
214 case GLSLstd450RoundEven: return nir_op_fround_even;
215 case GLSLstd450Trunc: return nir_op_ftrunc;
216 case GLSLstd450FAbs: return nir_op_fabs;
217 case GLSLstd450SAbs: return nir_op_iabs;
218 case GLSLstd450FSign: return nir_op_fsign;
219 case GLSLstd450SSign: return nir_op_isign;
220 case GLSLstd450Floor: return nir_op_ffloor;
221 case GLSLstd450Ceil: return nir_op_fceil;
222 case GLSLstd450Fract: return nir_op_ffract;
223 case GLSLstd450Sin: return nir_op_fsin;
224 case GLSLstd450Cos: return nir_op_fcos;
225 case GLSLstd450Pow: return nir_op_fpow;
226 case GLSLstd450Exp2: return nir_op_fexp2;
227 case GLSLstd450Log2: return nir_op_flog2;
228 case GLSLstd450Sqrt: return nir_op_fsqrt;
229 case GLSLstd450InverseSqrt: return nir_op_frsq;
230 case GLSLstd450NMin: *exact = true; return nir_op_fmin;
231 case GLSLstd450FMin: return nir_op_fmin;
232 case GLSLstd450UMin: return nir_op_umin;
233 case GLSLstd450SMin: return nir_op_imin;
234 case GLSLstd450NMax: *exact = true; return nir_op_fmax;
235 case GLSLstd450FMax: return nir_op_fmax;
236 case GLSLstd450UMax: return nir_op_umax;
237 case GLSLstd450SMax: return nir_op_imax;
238 case GLSLstd450FMix: return nir_op_flrp;
239 case GLSLstd450Fma: return nir_op_ffma;
240 case GLSLstd450Ldexp: return nir_op_ldexp;
241 case GLSLstd450FindILsb: return nir_op_find_lsb;
242 case GLSLstd450FindSMsb: return nir_op_ifind_msb;
243 case GLSLstd450FindUMsb: return nir_op_ufind_msb;
244
245 /* Packing/Unpacking functions */
246 case GLSLstd450PackSnorm4x8: return nir_op_pack_snorm_4x8;
247 case GLSLstd450PackUnorm4x8: return nir_op_pack_unorm_4x8;
248 case GLSLstd450PackSnorm2x16: return nir_op_pack_snorm_2x16;
249 case GLSLstd450PackUnorm2x16: return nir_op_pack_unorm_2x16;
250 case GLSLstd450PackHalf2x16: return nir_op_pack_half_2x16;
251 case GLSLstd450PackDouble2x32: return nir_op_pack_64_2x32;
252 case GLSLstd450UnpackSnorm4x8: return nir_op_unpack_snorm_4x8;
253 case GLSLstd450UnpackUnorm4x8: return nir_op_unpack_unorm_4x8;
254 case GLSLstd450UnpackSnorm2x16: return nir_op_unpack_snorm_2x16;
255 case GLSLstd450UnpackUnorm2x16: return nir_op_unpack_unorm_2x16;
256 case GLSLstd450UnpackHalf2x16:
257 if (execution_mode & FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP16)
258 return nir_op_unpack_half_2x16_flush_to_zero;
259 else
260 return nir_op_unpack_half_2x16;
261 case GLSLstd450UnpackDouble2x32: return nir_op_unpack_64_2x32;
262
263 default:
264 vtn_fail("No NIR equivalent");
265 }
266 }
267
268 #define NIR_IMM_FP(n, v) (nir_imm_floatN_t(n, v, src[0]->bit_size))
269
270 static void
handle_glsl450_alu(struct vtn_builder * b,enum GLSLstd450 entrypoint,const uint32_t * w,unsigned count)271 handle_glsl450_alu(struct vtn_builder *b, enum GLSLstd450 entrypoint,
272 const uint32_t *w, unsigned count)
273 {
274 struct nir_builder *nb = &b->nb;
275 const struct glsl_type *dest_type = vtn_get_type(b, w[1])->type;
276 struct vtn_value *dest_val = vtn_untyped_value(b, w[2]);
277
278 bool mediump_16bit;
279 switch (entrypoint) {
280 case GLSLstd450PackSnorm4x8:
281 case GLSLstd450PackUnorm4x8:
282 case GLSLstd450PackSnorm2x16:
283 case GLSLstd450PackUnorm2x16:
284 case GLSLstd450PackHalf2x16:
285 case GLSLstd450PackDouble2x32:
286 case GLSLstd450UnpackSnorm4x8:
287 case GLSLstd450UnpackUnorm4x8:
288 case GLSLstd450UnpackSnorm2x16:
289 case GLSLstd450UnpackUnorm2x16:
290 case GLSLstd450UnpackHalf2x16:
291 case GLSLstd450UnpackDouble2x32:
292 /* Asking for relaxed precision snorm 4x8 pack results (for example)
293 * doesn't even make sense. The NIR opcodes have a fixed output size, so
294 * no trying to reduce precision.
295 */
296 mediump_16bit = false;
297 break;
298
299 case GLSLstd450Frexp:
300 case GLSLstd450FrexpStruct:
301 case GLSLstd450Modf:
302 case GLSLstd450ModfStruct:
303 /* Not sure how to detect the ->elems[i] destinations on these in vtn_upconvert_value(). */
304 mediump_16bit = false;
305 break;
306
307 default:
308 mediump_16bit = b->options->mediump_16bit_alu && vtn_value_is_relaxed_precision(b, dest_val);
309 break;
310 }
311
312 /* Collect the various SSA sources */
313 unsigned num_inputs = count - 5;
314 nir_def *src[3] = { NULL, };
315 for (unsigned i = 0; i < num_inputs; i++) {
316 /* These are handled specially below */
317 if (vtn_untyped_value(b, w[i + 5])->value_type == vtn_value_type_pointer)
318 continue;
319
320 src[i] = vtn_get_nir_ssa(b, w[i + 5]);
321 if (mediump_16bit) {
322 struct vtn_ssa_value *vtn_src = vtn_ssa_value(b, w[i + 5]);
323 src[i] = vtn_mediump_downconvert(b, glsl_get_base_type(vtn_src->type), src[i]);
324 }
325 }
326
327 struct vtn_ssa_value *dest = vtn_create_ssa_value(b, dest_type);
328
329 vtn_handle_no_contraction(b, vtn_untyped_value(b, w[2]));
330 switch (entrypoint) {
331 case GLSLstd450Radians:
332 dest->def = nir_radians(nb, src[0]);
333 break;
334 case GLSLstd450Degrees:
335 dest->def = nir_degrees(nb, src[0]);
336 break;
337 case GLSLstd450Tan:
338 dest->def = nir_ftan(nb, src[0]);
339 break;
340
341 case GLSLstd450Modf: {
342 nir_def *inf = nir_imm_floatN_t(&b->nb, INFINITY, src[0]->bit_size);
343 nir_def *sign_bit =
344 nir_imm_intN_t(&b->nb, (uint64_t)1 << (src[0]->bit_size - 1),
345 src[0]->bit_size);
346 nir_def *sign = nir_fsign(nb, src[0]);
347 nir_def *abs = nir_fabs(nb, src[0]);
348
349 /* NaN input should produce a NaN results, and ±Inf input should provide
350 * ±0 result. The fmul(sign(x), ffract(x)) calculation will already
351 * produce the expected NaN. To get ±0, directly compare for equality
352 * with Inf instead of using fisfinite (which is false for NaN).
353 */
354 dest->def = nir_bcsel(nb,
355 nir_ieq(nb, abs, inf),
356 nir_iand(nb, src[0], sign_bit),
357 nir_fmul(nb, sign, nir_ffract(nb, abs)));
358
359 struct vtn_pointer *i_ptr = vtn_value(b, w[6], vtn_value_type_pointer)->pointer;
360 struct vtn_ssa_value *whole = vtn_create_ssa_value(b, i_ptr->type->type);
361 whole->def = nir_fmul(nb, sign, nir_ffloor(nb, abs));
362 vtn_variable_store(b, whole, i_ptr, 0);
363 break;
364 }
365
366 case GLSLstd450ModfStruct: {
367 nir_def *inf = nir_imm_floatN_t(&b->nb, INFINITY, src[0]->bit_size);
368 nir_def *sign_bit =
369 nir_imm_intN_t(&b->nb, (uint64_t)1 << (src[0]->bit_size - 1),
370 src[0]->bit_size);
371 nir_def *sign = nir_fsign(nb, src[0]);
372 nir_def *abs = nir_fabs(nb, src[0]);
373 vtn_assert(glsl_type_is_struct_or_ifc(dest_type));
374
375 /* See GLSLstd450Modf for explanation of the Inf and NaN handling. */
376 dest->elems[0]->def = nir_bcsel(nb,
377 nir_ieq(nb, abs, inf),
378 nir_iand(nb, src[0], sign_bit),
379 nir_fmul(nb, sign, nir_ffract(nb, abs)));
380 dest->elems[1]->def = nir_fmul(nb, sign, nir_ffloor(nb, abs));
381 break;
382 }
383
384 case GLSLstd450Step: {
385 /* The SPIR-V Extended Instructions for GLSL spec says:
386 *
387 * Result is 0.0 if x < edge; otherwise result is 1.0.
388 *
389 * Here src[1] is x, and src[0] is edge. The direct implementation is
390 *
391 * bcsel(src[1] < src[0], 0.0, 1.0)
392 *
393 * This is effectively b2f(!(src1 < src0)). Previously this was
394 * implemented using sge(src1, src0), but that produces incorrect
395 * results for NaN. Instead, we use the identity b2f(!x) = 1 - b2f(x).
396 */
397 const bool exact = nb->exact;
398 nb->exact = true;
399
400 nir_def *cmp = nir_slt(nb, src[1], src[0]);
401
402 nb->exact = exact;
403 dest->def = nir_fsub_imm(nb, 1.0f, cmp);
404 break;
405 }
406
407 case GLSLstd450Length:
408 dest->def = nir_fast_length(nb, src[0]);
409 break;
410 case GLSLstd450Distance:
411 dest->def = nir_fast_distance(nb, src[0], src[1]);
412 break;
413 case GLSLstd450Normalize:
414 dest->def = nir_fast_normalize(nb, src[0]);
415 break;
416
417 case GLSLstd450Exp:
418 dest->def = nir_fexp(nb, src[0]);
419 break;
420
421 case GLSLstd450Log:
422 dest->def = nir_flog(nb, src[0]);
423 break;
424
425 case GLSLstd450FClamp:
426 dest->def = nir_fclamp(nb, src[0], src[1], src[2]);
427 break;
428 case GLSLstd450NClamp:
429 nb->exact = true;
430 dest->def = nir_fclamp(nb, src[0], src[1], src[2]);
431 nb->exact = false;
432 break;
433 case GLSLstd450UClamp:
434 dest->def = nir_uclamp(nb, src[0], src[1], src[2]);
435 break;
436 case GLSLstd450SClamp:
437 dest->def = nir_iclamp(nb, src[0], src[1], src[2]);
438 break;
439
440 case GLSLstd450Cross: {
441 dest->def = nir_cross3(nb, src[0], src[1]);
442 break;
443 }
444
445 case GLSLstd450SmoothStep: {
446 dest->def = nir_smoothstep(nb, src[0], src[1], src[2]);
447 break;
448 }
449
450 case GLSLstd450FaceForward:
451 dest->def =
452 nir_bcsel(nb, nir_flt(nb, nir_fdot(nb, src[2], src[1]),
453 NIR_IMM_FP(nb, 0.0)),
454 src[0], nir_fneg(nb, src[0]));
455 break;
456
457 case GLSLstd450Reflect:
458 /* I - 2 * dot(N, I) * N */
459 dest->def =
460 nir_a_minus_bc(nb, src[0],
461 src[1],
462 nir_fmul(nb, nir_fdot(nb, src[0], src[1]),
463 NIR_IMM_FP(nb, 2.0)));
464 break;
465
466 case GLSLstd450Refract: {
467 nir_def *I = src[0];
468 nir_def *N = src[1];
469 nir_def *eta = src[2];
470 nir_def *n_dot_i = nir_fdot(nb, N, I);
471 nir_def *one = NIR_IMM_FP(nb, 1.0);
472 nir_def *zero = NIR_IMM_FP(nb, 0.0);
473 /* According to the SPIR-V and GLSL specs, eta is always a float
474 * regardless of the type of the other operands. However in practice it
475 * seems that if you try to pass it a float then glslang will just
476 * promote it to a double and generate invalid SPIR-V. In order to
477 * support a hypothetical fixed version of glslang we’ll promote eta to
478 * double if the other operands are double also.
479 */
480 if (I->bit_size != eta->bit_size) {
481 eta = nir_type_convert(nb, eta, nir_type_float,
482 nir_type_float | I->bit_size,
483 nir_rounding_mode_undef);
484 }
485 /* k = 1.0 - eta * eta * (1.0 - dot(N, I) * dot(N, I)) */
486 nir_def *k =
487 nir_a_minus_bc(nb, one, eta,
488 nir_fmul(nb, eta, nir_a_minus_bc(nb, one, n_dot_i, n_dot_i)));
489 nir_def *result =
490 nir_a_minus_bc(nb, nir_fmul(nb, eta, I),
491 nir_ffma(nb, eta, n_dot_i, nir_fsqrt(nb, k)),
492 N);
493 /* XXX: bcsel, or if statement? */
494 dest->def = nir_bcsel(nb, nir_flt(nb, k, zero), zero, result);
495 break;
496 }
497
498 case GLSLstd450Sinh:
499 /* 0.5 * (e^x - e^(-x)) */
500 dest->def =
501 nir_fmul_imm(nb, nir_fsub(nb, nir_fexp(nb, src[0]),
502 nir_fexp(nb, nir_fneg(nb, src[0]))),
503 0.5f);
504 break;
505
506 case GLSLstd450Cosh:
507 /* 0.5 * (e^x + e^(-x)) */
508 dest->def =
509 nir_fmul_imm(nb, nir_fadd(nb, nir_fexp(nb, src[0]),
510 nir_fexp(nb, nir_fneg(nb, src[0]))),
511 0.5f);
512 break;
513
514 case GLSLstd450Tanh: {
515 /* tanh(x) := (e^x - e^(-x)) / (e^x + e^(-x))
516 *
517 * We clamp x to [-10, +10] to avoid precision problems. When x > 10,
518 * e^x dominates the sum, e^(-x) is lost and tanh(x) is 1.0 for 32 bit
519 * floating point.
520 *
521 * For 16-bit precision this we clamp x to [-4.2, +4.2].
522 */
523 const uint32_t bit_size = src[0]->bit_size;
524 const double clamped_x = bit_size > 16 ? 10.0 : 4.2;
525 nir_def *x = nir_fclamp(nb, src[0],
526 nir_imm_floatN_t(nb, -clamped_x, bit_size),
527 nir_imm_floatN_t(nb, clamped_x, bit_size));
528
529 /* The clamping will filter out NaN values causing an incorrect result.
530 * The comparison is carefully structured to get NaN result for NaN and
531 * get -0 for -0.
532 *
533 * result = abs(s) > 0.0 ? ... : s;
534 */
535 const bool exact = nb->exact;
536
537 nb->exact = true;
538 nir_def *is_regular = nir_flt(nb,
539 nir_imm_floatN_t(nb, 0, bit_size),
540 nir_fabs(nb, src[0]));
541
542 /* The extra 1.0*s ensures that subnormal inputs are flushed to zero
543 * when that is selected by the shader.
544 */
545 nir_def *flushed = nir_fmul(nb,
546 src[0],
547 nir_imm_floatN_t(nb, 1.0, bit_size));
548 nb->exact = exact;
549
550 dest->def = nir_bcsel(nb,
551 is_regular,
552 nir_fdiv(nb, nir_fsub(nb, nir_fexp(nb, x),
553 nir_fexp(nb, nir_fneg(nb, x))),
554 nir_fadd(nb, nir_fexp(nb, x),
555 nir_fexp(nb, nir_fneg(nb, x)))),
556 flushed);
557 break;
558 }
559
560 case GLSLstd450Asinh:
561 dest->def = nir_fmul(nb, nir_fsign(nb, src[0]),
562 nir_flog(nb, nir_fadd(nb, nir_fabs(nb, src[0]),
563 nir_fsqrt(nb, nir_ffma_imm2(nb, src[0], src[0], 1.0f)))));
564 break;
565 case GLSLstd450Acosh:
566 dest->def = nir_flog(nb, nir_fadd(nb, src[0],
567 nir_fsqrt(nb, nir_ffma_imm2(nb, src[0], src[0], -1.0f))));
568 break;
569 case GLSLstd450Atanh: {
570 dest->def =
571 nir_fmul_imm(nb, nir_flog(nb, nir_fdiv(nb, nir_fadd_imm(nb, src[0], 1.0),
572 nir_fsub_imm(nb, 1.0, src[0]))),
573 0.5f);
574 break;
575 }
576
577 case GLSLstd450Asin:
578 dest->def = build_asin(nb, src[0], 0.086566724, -0.03102955, true);
579 break;
580
581 case GLSLstd450Acos:
582 dest->def =
583 nir_fsub_imm(nb, M_PI_2f,
584 build_asin(nb, src[0], 0.08132463, -0.02363318, false));
585 break;
586
587 case GLSLstd450Atan:
588 dest->def = nir_atan(nb, src[0]);
589 break;
590
591 case GLSLstd450Atan2:
592 dest->def = nir_atan2(nb, src[0], src[1]);
593 break;
594
595 case GLSLstd450Frexp: {
596 dest->def = nir_frexp_sig(nb, src[0]);
597
598 struct vtn_pointer *i_ptr = vtn_value(b, w[6], vtn_value_type_pointer)->pointer;
599 struct vtn_ssa_value *exp = vtn_create_ssa_value(b, i_ptr->type->type);
600 exp->def = nir_frexp_exp(nb, src[0]);
601 vtn_variable_store(b, exp, i_ptr, 0);
602 break;
603 }
604
605 case GLSLstd450FrexpStruct: {
606 vtn_assert(glsl_type_is_struct_or_ifc(dest_type));
607 dest->elems[0]->def = nir_frexp_sig(nb, src[0]);
608 dest->elems[1]->def = nir_frexp_exp(nb, src[0]);
609 break;
610 }
611
612 default: {
613 unsigned execution_mode =
614 b->shader->info.float_controls_execution_mode;
615 bool exact;
616 nir_op op = vtn_nir_alu_op_for_spirv_glsl_opcode(b, entrypoint, execution_mode, &exact);
617 /* don't override explicit decoration */
618 b->nb.exact |= exact;
619 dest->def = nir_build_alu(&b->nb, op, src[0], src[1], src[2], NULL);
620 break;
621 }
622 }
623 b->nb.exact = false;
624
625 if (mediump_16bit)
626 vtn_mediump_upconvert_value(b, dest);
627
628 vtn_push_ssa_value(b, w[2], dest);
629 }
630
631 static void
handle_glsl450_interpolation(struct vtn_builder * b,enum GLSLstd450 opcode,const uint32_t * w,unsigned count)632 handle_glsl450_interpolation(struct vtn_builder *b, enum GLSLstd450 opcode,
633 const uint32_t *w, unsigned count)
634 {
635 nir_intrinsic_op op;
636 switch (opcode) {
637 case GLSLstd450InterpolateAtCentroid:
638 op = nir_intrinsic_interp_deref_at_centroid;
639 break;
640 case GLSLstd450InterpolateAtSample:
641 op = nir_intrinsic_interp_deref_at_sample;
642 break;
643 case GLSLstd450InterpolateAtOffset:
644 op = nir_intrinsic_interp_deref_at_offset;
645 break;
646 default:
647 vtn_fail("Invalid opcode");
648 }
649
650 nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(b->nb.shader, op);
651
652 struct vtn_pointer *ptr =
653 vtn_value(b, w[5], vtn_value_type_pointer)->pointer;
654 nir_deref_instr *deref = vtn_pointer_to_deref(b, ptr);
655
656 /* If the value we are interpolating has an index into a vector then
657 * interpolate the vector and index the result of that instead. This is
658 * necessary because the index will get generated as a series of nir_bcsel
659 * instructions so it would no longer be an input variable.
660 */
661 const bool vec_array_deref = deref->deref_type == nir_deref_type_array &&
662 glsl_type_is_vector(nir_deref_instr_parent(deref)->type);
663
664 nir_deref_instr *vec_deref = NULL;
665 if (vec_array_deref) {
666 vec_deref = deref;
667 deref = nir_deref_instr_parent(deref);
668 }
669 intrin->src[0] = nir_src_for_ssa(&deref->def);
670
671 switch (opcode) {
672 case GLSLstd450InterpolateAtCentroid:
673 break;
674 case GLSLstd450InterpolateAtSample:
675 case GLSLstd450InterpolateAtOffset:
676 intrin->src[1] = nir_src_for_ssa(vtn_get_nir_ssa(b, w[6]));
677 break;
678 default:
679 vtn_fail("Invalid opcode");
680 }
681
682 intrin->num_components = glsl_get_vector_elements(deref->type);
683 nir_def_init(&intrin->instr, &intrin->def,
684 glsl_get_vector_elements(deref->type),
685 glsl_get_bit_size(deref->type));
686
687 nir_builder_instr_insert(&b->nb, &intrin->instr);
688
689 nir_def *def = &intrin->def;
690 if (vec_array_deref)
691 def = nir_vector_extract(&b->nb, def, vec_deref->arr.index.ssa);
692
693 vtn_push_nir_ssa(b, w[2], def);
694 }
695
696 bool
vtn_handle_glsl450_instruction(struct vtn_builder * b,SpvOp ext_opcode,const uint32_t * w,unsigned count)697 vtn_handle_glsl450_instruction(struct vtn_builder *b, SpvOp ext_opcode,
698 const uint32_t *w, unsigned count)
699 {
700 switch ((enum GLSLstd450)ext_opcode) {
701 case GLSLstd450Determinant: {
702 vtn_push_nir_ssa(b, w[2], build_mat_det(b, vtn_ssa_value(b, w[5])));
703 break;
704 }
705
706 case GLSLstd450MatrixInverse: {
707 vtn_push_ssa_value(b, w[2], matrix_inverse(b, vtn_ssa_value(b, w[5])));
708 break;
709 }
710
711 case GLSLstd450InterpolateAtCentroid:
712 case GLSLstd450InterpolateAtSample:
713 case GLSLstd450InterpolateAtOffset:
714 handle_glsl450_interpolation(b, (enum GLSLstd450)ext_opcode, w, count);
715 break;
716
717 default:
718 handle_glsl450_alu(b, (enum GLSLstd450)ext_opcode, w, count);
719 }
720
721 return true;
722 }
723