• 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  * Authors:
24  *    Jason Ekstrand (jason@jlekstrand.net)
25  *
26  */
27 
28 #include "vtn_private.h"
29 #include "GLSL.std.450.h"
30 
31 #define M_PIf   ((float) M_PI)
32 #define M_PI_2f ((float) M_PI_2)
33 #define M_PI_4f ((float) M_PI_4)
34 
35 static nir_ssa_def *
build_mat2_det(nir_builder * b,nir_ssa_def * col[2])36 build_mat2_det(nir_builder *b, nir_ssa_def *col[2])
37 {
38    unsigned swiz[4] = {1, 0, 0, 0};
39    nir_ssa_def *p = nir_fmul(b, col[0], nir_swizzle(b, col[1], swiz, 2, true));
40    return nir_fsub(b, nir_channel(b, p, 0), nir_channel(b, p, 1));
41 }
42 
43 static nir_ssa_def *
build_mat3_det(nir_builder * b,nir_ssa_def * col[3])44 build_mat3_det(nir_builder *b, nir_ssa_def *col[3])
45 {
46    unsigned yzx[4] = {1, 2, 0, 0};
47    unsigned zxy[4] = {2, 0, 1, 0};
48 
49    nir_ssa_def *prod0 =
50       nir_fmul(b, col[0],
51                nir_fmul(b, nir_swizzle(b, col[1], yzx, 3, true),
52                            nir_swizzle(b, col[2], zxy, 3, true)));
53    nir_ssa_def *prod1 =
54       nir_fmul(b, col[0],
55                nir_fmul(b, nir_swizzle(b, col[1], zxy, 3, true),
56                            nir_swizzle(b, col[2], yzx, 3, true)));
57 
58    nir_ssa_def *diff = nir_fsub(b, prod0, prod1);
59 
60    return nir_fadd(b, nir_channel(b, diff, 0),
61                       nir_fadd(b, nir_channel(b, diff, 1),
62                                   nir_channel(b, diff, 2)));
63 }
64 
65 static nir_ssa_def *
build_mat4_det(nir_builder * b,nir_ssa_def ** col)66 build_mat4_det(nir_builder *b, nir_ssa_def **col)
67 {
68    nir_ssa_def *subdet[4];
69    for (unsigned i = 0; i < 4; i++) {
70       unsigned swiz[3];
71       for (unsigned j = 0; j < 3; j++)
72          swiz[j] = j + (j >= i);
73 
74       nir_ssa_def *subcol[3];
75       subcol[0] = nir_swizzle(b, col[1], swiz, 3, true);
76       subcol[1] = nir_swizzle(b, col[2], swiz, 3, true);
77       subcol[2] = nir_swizzle(b, col[3], swiz, 3, true);
78 
79       subdet[i] = build_mat3_det(b, subcol);
80    }
81 
82    nir_ssa_def *prod = nir_fmul(b, col[0], nir_vec(b, subdet, 4));
83 
84    return nir_fadd(b, nir_fsub(b, nir_channel(b, prod, 0),
85                                   nir_channel(b, prod, 1)),
86                       nir_fsub(b, nir_channel(b, prod, 2),
87                                   nir_channel(b, prod, 3)));
88 }
89 
90 static nir_ssa_def *
build_mat_det(struct vtn_builder * b,struct vtn_ssa_value * src)91 build_mat_det(struct vtn_builder *b, struct vtn_ssa_value *src)
92 {
93    unsigned size = glsl_get_vector_elements(src->type);
94 
95    nir_ssa_def *cols[4];
96    for (unsigned i = 0; i < size; i++)
97       cols[i] = src->elems[i]->def;
98 
99    switch(size) {
100    case 2: return build_mat2_det(&b->nb, cols);
101    case 3: return build_mat3_det(&b->nb, cols);
102    case 4: return build_mat4_det(&b->nb, cols);
103    default:
104       unreachable("Invalid matrix size");
105    }
106 }
107 
108 /* Computes the determinate of the submatrix given by taking src and
109  * removing the specified row and column.
110  */
111 static nir_ssa_def *
build_mat_subdet(struct nir_builder * b,struct vtn_ssa_value * src,unsigned size,unsigned row,unsigned col)112 build_mat_subdet(struct nir_builder *b, struct vtn_ssa_value *src,
113                  unsigned size, unsigned row, unsigned col)
114 {
115    assert(row < size && col < size);
116    if (size == 2) {
117       return nir_channel(b, src->elems[1 - col]->def, 1 - row);
118    } else {
119       /* Swizzle to get all but the specified row */
120       unsigned swiz[3];
121       for (unsigned j = 0; j < 3; j++)
122          swiz[j] = j + (j >= row);
123 
124       /* Grab all but the specified column */
125       nir_ssa_def *subcol[3];
126       for (unsigned j = 0; j < size; j++) {
127          if (j != col) {
128             subcol[j - (j > col)] = nir_swizzle(b, src->elems[j]->def,
129                                                 swiz, size - 1, true);
130          }
131       }
132 
133       if (size == 3) {
134          return build_mat2_det(b, subcol);
135       } else {
136          assert(size == 4);
137          return build_mat3_det(b, subcol);
138       }
139    }
140 }
141 
142 static struct vtn_ssa_value *
matrix_inverse(struct vtn_builder * b,struct vtn_ssa_value * src)143 matrix_inverse(struct vtn_builder *b, struct vtn_ssa_value *src)
144 {
145    nir_ssa_def *adj_col[4];
146    unsigned size = glsl_get_vector_elements(src->type);
147 
148    /* Build up an adjugate matrix */
149    for (unsigned c = 0; c < size; c++) {
150       nir_ssa_def *elem[4];
151       for (unsigned r = 0; r < size; r++) {
152          elem[r] = build_mat_subdet(&b->nb, src, size, c, r);
153 
154          if ((r + c) % 2)
155             elem[r] = nir_fneg(&b->nb, elem[r]);
156       }
157 
158       adj_col[c] = nir_vec(&b->nb, elem, size);
159    }
160 
161    nir_ssa_def *det_inv = nir_frcp(&b->nb, build_mat_det(b, src));
162 
163    struct vtn_ssa_value *val = vtn_create_ssa_value(b, src->type);
164    for (unsigned i = 0; i < size; i++)
165       val->elems[i]->def = nir_fmul(&b->nb, adj_col[i], det_inv);
166 
167    return val;
168 }
169 
170 static nir_ssa_def*
build_length(nir_builder * b,nir_ssa_def * vec)171 build_length(nir_builder *b, nir_ssa_def *vec)
172 {
173    switch (vec->num_components) {
174    case 1: return nir_fsqrt(b, nir_fmul(b, vec, vec));
175    case 2: return nir_fsqrt(b, nir_fdot2(b, vec, vec));
176    case 3: return nir_fsqrt(b, nir_fdot3(b, vec, vec));
177    case 4: return nir_fsqrt(b, nir_fdot4(b, vec, vec));
178    default:
179       unreachable("Invalid number of components");
180    }
181 }
182 
183 static inline nir_ssa_def *
build_fclamp(nir_builder * b,nir_ssa_def * x,nir_ssa_def * min_val,nir_ssa_def * max_val)184 build_fclamp(nir_builder *b,
185              nir_ssa_def *x, nir_ssa_def *min_val, nir_ssa_def *max_val)
186 {
187    return nir_fmin(b, nir_fmax(b, x, min_val), max_val);
188 }
189 
190 /**
191  * Return e^x.
192  */
193 static nir_ssa_def *
build_exp(nir_builder * b,nir_ssa_def * x)194 build_exp(nir_builder *b, nir_ssa_def *x)
195 {
196    return nir_fexp2(b, nir_fmul(b, x, nir_imm_float(b, M_LOG2E)));
197 }
198 
199 /**
200  * Return ln(x) - the natural logarithm of x.
201  */
202 static nir_ssa_def *
build_log(nir_builder * b,nir_ssa_def * x)203 build_log(nir_builder *b, nir_ssa_def *x)
204 {
205    return nir_fmul(b, nir_flog2(b, x), nir_imm_float(b, 1.0 / M_LOG2E));
206 }
207 
208 /**
209  * Approximate asin(x) by the formula:
210  *    asin~(x) = sign(x) * (pi/2 - sqrt(1 - |x|) * (pi/2 + |x|(pi/4 - 1 + |x|(p0 + |x|p1))))
211  *
212  * which is correct to first order at x=0 and x=±1 regardless of the p
213  * coefficients but can be made second-order correct at both ends by selecting
214  * the fit coefficients appropriately.  Different p coefficients can be used
215  * in the asin and acos implementation to minimize some relative error metric
216  * in each case.
217  */
218 static nir_ssa_def *
build_asin(nir_builder * b,nir_ssa_def * x,float p0,float p1)219 build_asin(nir_builder *b, nir_ssa_def *x, float p0, float p1)
220 {
221    nir_ssa_def *abs_x = nir_fabs(b, x);
222    return nir_fmul(b, nir_fsign(b, x),
223                    nir_fsub(b, nir_imm_float(b, M_PI_2f),
224                             nir_fmul(b, nir_fsqrt(b, nir_fsub(b, nir_imm_float(b, 1.0f), abs_x)),
225                                      nir_fadd(b, nir_imm_float(b, M_PI_2f),
226                                               nir_fmul(b, abs_x,
227                                                        nir_fadd(b, nir_imm_float(b, M_PI_4f - 1.0f),
228                                                                 nir_fmul(b, abs_x,
229                                                                          nir_fadd(b, nir_imm_float(b, p0),
230                                                                                   nir_fmul(b, abs_x,
231                                                                                            nir_imm_float(b, p1))))))))));
232 }
233 
234 /**
235  * Compute xs[0] + xs[1] + xs[2] + ... using fadd.
236  */
237 static nir_ssa_def *
build_fsum(nir_builder * b,nir_ssa_def ** xs,int terms)238 build_fsum(nir_builder *b, nir_ssa_def **xs, int terms)
239 {
240    nir_ssa_def *accum = xs[0];
241 
242    for (int i = 1; i < terms; i++)
243       accum = nir_fadd(b, accum, xs[i]);
244 
245    return accum;
246 }
247 
248 static nir_ssa_def *
build_atan(nir_builder * b,nir_ssa_def * y_over_x)249 build_atan(nir_builder *b, nir_ssa_def *y_over_x)
250 {
251    nir_ssa_def *abs_y_over_x = nir_fabs(b, y_over_x);
252    nir_ssa_def *one = nir_imm_float(b, 1.0f);
253 
254    /*
255     * range-reduction, first step:
256     *
257     *      / y_over_x         if |y_over_x| <= 1.0;
258     * x = <
259     *      \ 1.0 / y_over_x   otherwise
260     */
261    nir_ssa_def *x = nir_fdiv(b, nir_fmin(b, abs_y_over_x, one),
262                                 nir_fmax(b, abs_y_over_x, one));
263 
264    /*
265     * approximate atan by evaluating polynomial:
266     *
267     * x   * 0.9999793128310355 - x^3  * 0.3326756418091246 +
268     * x^5 * 0.1938924977115610 - x^7  * 0.1173503194786851 +
269     * x^9 * 0.0536813784310406 - x^11 * 0.0121323213173444
270     */
271    nir_ssa_def *x_2  = nir_fmul(b, x,   x);
272    nir_ssa_def *x_3  = nir_fmul(b, x_2, x);
273    nir_ssa_def *x_5  = nir_fmul(b, x_3, x_2);
274    nir_ssa_def *x_7  = nir_fmul(b, x_5, x_2);
275    nir_ssa_def *x_9  = nir_fmul(b, x_7, x_2);
276    nir_ssa_def *x_11 = nir_fmul(b, x_9, x_2);
277 
278    nir_ssa_def *polynomial_terms[] = {
279       nir_fmul(b, x,    nir_imm_float(b,  0.9999793128310355f)),
280       nir_fmul(b, x_3,  nir_imm_float(b, -0.3326756418091246f)),
281       nir_fmul(b, x_5,  nir_imm_float(b,  0.1938924977115610f)),
282       nir_fmul(b, x_7,  nir_imm_float(b, -0.1173503194786851f)),
283       nir_fmul(b, x_9,  nir_imm_float(b,  0.0536813784310406f)),
284       nir_fmul(b, x_11, nir_imm_float(b, -0.0121323213173444f)),
285    };
286 
287    nir_ssa_def *tmp =
288       build_fsum(b, polynomial_terms, ARRAY_SIZE(polynomial_terms));
289 
290    /* range-reduction fixup */
291    tmp = nir_fadd(b, tmp,
292                   nir_fmul(b,
293                            nir_b2f(b, nir_flt(b, one, abs_y_over_x)),
294                            nir_fadd(b, nir_fmul(b, tmp,
295                                                 nir_imm_float(b, -2.0f)),
296                                        nir_imm_float(b, M_PI_2f))));
297 
298    /* sign fixup */
299    return nir_fmul(b, tmp, nir_fsign(b, y_over_x));
300 }
301 
302 static nir_ssa_def *
build_atan2(nir_builder * b,nir_ssa_def * y,nir_ssa_def * x)303 build_atan2(nir_builder *b, nir_ssa_def *y, nir_ssa_def *x)
304 {
305    nir_ssa_def *zero = nir_imm_float(b, 0.0f);
306 
307    /* If |x| >= 1.0e-8 * |y|: */
308    nir_ssa_def *condition =
309       nir_fge(b, nir_fabs(b, x),
310               nir_fmul(b, nir_imm_float(b, 1.0e-8f), nir_fabs(b, y)));
311 
312    /* Then...call atan(y/x) and fix it up: */
313    nir_ssa_def *atan1 = build_atan(b, nir_fdiv(b, y, x));
314    nir_ssa_def *r_then =
315       nir_bcsel(b, nir_flt(b, x, zero),
316                    nir_fadd(b, atan1,
317                                nir_bcsel(b, nir_fge(b, y, zero),
318                                             nir_imm_float(b, M_PIf),
319                                             nir_imm_float(b, -M_PIf))),
320                    atan1);
321 
322    /* Else... */
323    nir_ssa_def *r_else =
324       nir_fmul(b, nir_fsign(b, y), nir_imm_float(b, M_PI_2f));
325 
326    return nir_bcsel(b, condition, r_then, r_else);
327 }
328 
329 static nir_ssa_def *
build_frexp(nir_builder * b,nir_ssa_def * x,nir_ssa_def ** exponent)330 build_frexp(nir_builder *b, nir_ssa_def *x, nir_ssa_def **exponent)
331 {
332    nir_ssa_def *abs_x = nir_fabs(b, x);
333    nir_ssa_def *zero = nir_imm_float(b, 0.0f);
334 
335    /* Single-precision floating-point values are stored as
336     *   1 sign bit;
337     *   8 exponent bits;
338     *   23 mantissa bits.
339     *
340     * An exponent shift of 23 will shift the mantissa out, leaving only the
341     * exponent and sign bit (which itself may be zero, if the absolute value
342     * was taken before the bitcast and shift.
343     */
344    nir_ssa_def *exponent_shift = nir_imm_int(b, 23);
345    nir_ssa_def *exponent_bias = nir_imm_int(b, -126);
346 
347    nir_ssa_def *sign_mantissa_mask = nir_imm_int(b, 0x807fffffu);
348 
349    /* Exponent of floating-point values in the range [0.5, 1.0). */
350    nir_ssa_def *exponent_value = nir_imm_int(b, 0x3f000000u);
351 
352    nir_ssa_def *is_not_zero = nir_fne(b, abs_x, zero);
353 
354    *exponent =
355       nir_iadd(b, nir_ushr(b, abs_x, exponent_shift),
356                   nir_bcsel(b, is_not_zero, exponent_bias, zero));
357 
358    return nir_ior(b, nir_iand(b, x, sign_mantissa_mask),
359                      nir_bcsel(b, is_not_zero, exponent_value, zero));
360 }
361 
362 static nir_op
vtn_nir_alu_op_for_spirv_glsl_opcode(enum GLSLstd450 opcode)363 vtn_nir_alu_op_for_spirv_glsl_opcode(enum GLSLstd450 opcode)
364 {
365    switch (opcode) {
366    case GLSLstd450Round:         return nir_op_fround_even;
367    case GLSLstd450RoundEven:     return nir_op_fround_even;
368    case GLSLstd450Trunc:         return nir_op_ftrunc;
369    case GLSLstd450FAbs:          return nir_op_fabs;
370    case GLSLstd450SAbs:          return nir_op_iabs;
371    case GLSLstd450FSign:         return nir_op_fsign;
372    case GLSLstd450SSign:         return nir_op_isign;
373    case GLSLstd450Floor:         return nir_op_ffloor;
374    case GLSLstd450Ceil:          return nir_op_fceil;
375    case GLSLstd450Fract:         return nir_op_ffract;
376    case GLSLstd450Sin:           return nir_op_fsin;
377    case GLSLstd450Cos:           return nir_op_fcos;
378    case GLSLstd450Pow:           return nir_op_fpow;
379    case GLSLstd450Exp2:          return nir_op_fexp2;
380    case GLSLstd450Log2:          return nir_op_flog2;
381    case GLSLstd450Sqrt:          return nir_op_fsqrt;
382    case GLSLstd450InverseSqrt:   return nir_op_frsq;
383    case GLSLstd450FMin:          return nir_op_fmin;
384    case GLSLstd450UMin:          return nir_op_umin;
385    case GLSLstd450SMin:          return nir_op_imin;
386    case GLSLstd450FMax:          return nir_op_fmax;
387    case GLSLstd450UMax:          return nir_op_umax;
388    case GLSLstd450SMax:          return nir_op_imax;
389    case GLSLstd450FMix:          return nir_op_flrp;
390    case GLSLstd450Fma:           return nir_op_ffma;
391    case GLSLstd450Ldexp:         return nir_op_ldexp;
392    case GLSLstd450FindILsb:      return nir_op_find_lsb;
393    case GLSLstd450FindSMsb:      return nir_op_ifind_msb;
394    case GLSLstd450FindUMsb:      return nir_op_ufind_msb;
395 
396    /* Packing/Unpacking functions */
397    case GLSLstd450PackSnorm4x8:     return nir_op_pack_snorm_4x8;
398    case GLSLstd450PackUnorm4x8:     return nir_op_pack_unorm_4x8;
399    case GLSLstd450PackSnorm2x16:    return nir_op_pack_snorm_2x16;
400    case GLSLstd450PackUnorm2x16:    return nir_op_pack_unorm_2x16;
401    case GLSLstd450PackHalf2x16:     return nir_op_pack_half_2x16;
402    case GLSLstd450PackDouble2x32:   return nir_op_pack_double_2x32;
403    case GLSLstd450UnpackSnorm4x8:   return nir_op_unpack_snorm_4x8;
404    case GLSLstd450UnpackUnorm4x8:   return nir_op_unpack_unorm_4x8;
405    case GLSLstd450UnpackSnorm2x16:  return nir_op_unpack_snorm_2x16;
406    case GLSLstd450UnpackUnorm2x16:  return nir_op_unpack_unorm_2x16;
407    case GLSLstd450UnpackHalf2x16:   return nir_op_unpack_half_2x16;
408    case GLSLstd450UnpackDouble2x32: return nir_op_unpack_double_2x32;
409 
410    default:
411       unreachable("No NIR equivalent");
412    }
413 }
414 
415 static void
handle_glsl450_alu(struct vtn_builder * b,enum GLSLstd450 entrypoint,const uint32_t * w,unsigned count)416 handle_glsl450_alu(struct vtn_builder *b, enum GLSLstd450 entrypoint,
417                    const uint32_t *w, unsigned count)
418 {
419    struct nir_builder *nb = &b->nb;
420    const struct glsl_type *dest_type =
421       vtn_value(b, w[1], vtn_value_type_type)->type->type;
422 
423    struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
424    val->ssa = vtn_create_ssa_value(b, dest_type);
425 
426    /* Collect the various SSA sources */
427    unsigned num_inputs = count - 5;
428    nir_ssa_def *src[3] = { NULL, };
429    for (unsigned i = 0; i < num_inputs; i++)
430       src[i] = vtn_ssa_value(b, w[i + 5])->def;
431 
432    switch (entrypoint) {
433    case GLSLstd450Radians:
434       val->ssa->def = nir_fmul(nb, src[0], nir_imm_float(nb, 0.01745329251));
435       return;
436    case GLSLstd450Degrees:
437       val->ssa->def = nir_fmul(nb, src[0], nir_imm_float(nb, 57.2957795131));
438       return;
439    case GLSLstd450Tan:
440       val->ssa->def = nir_fdiv(nb, nir_fsin(nb, src[0]),
441                                nir_fcos(nb, src[0]));
442       return;
443 
444    case GLSLstd450Modf: {
445       nir_ssa_def *sign = nir_fsign(nb, src[0]);
446       nir_ssa_def *abs = nir_fabs(nb, src[0]);
447       val->ssa->def = nir_fmul(nb, sign, nir_ffract(nb, abs));
448       nir_store_deref_var(nb, vtn_nir_deref(b, w[6]),
449                           nir_fmul(nb, sign, nir_ffloor(nb, abs)), 0xf);
450       return;
451    }
452 
453    case GLSLstd450ModfStruct: {
454       nir_ssa_def *sign = nir_fsign(nb, src[0]);
455       nir_ssa_def *abs = nir_fabs(nb, src[0]);
456       assert(glsl_type_is_struct(val->ssa->type));
457       val->ssa->elems[0]->def = nir_fmul(nb, sign, nir_ffract(nb, abs));
458       val->ssa->elems[1]->def = nir_fmul(nb, sign, nir_ffloor(nb, abs));
459       return;
460    }
461 
462    case GLSLstd450Step:
463       val->ssa->def = nir_sge(nb, src[1], src[0]);
464       return;
465 
466    case GLSLstd450Length:
467       val->ssa->def = build_length(nb, src[0]);
468       return;
469    case GLSLstd450Distance:
470       val->ssa->def = build_length(nb, nir_fsub(nb, src[0], src[1]));
471       return;
472    case GLSLstd450Normalize:
473       val->ssa->def = nir_fdiv(nb, src[0], build_length(nb, src[0]));
474       return;
475 
476    case GLSLstd450Exp:
477       val->ssa->def = build_exp(nb, src[0]);
478       return;
479 
480    case GLSLstd450Log:
481       val->ssa->def = build_log(nb, src[0]);
482       return;
483 
484    case GLSLstd450FClamp:
485       val->ssa->def = build_fclamp(nb, src[0], src[1], src[2]);
486       return;
487    case GLSLstd450UClamp:
488       val->ssa->def = nir_umin(nb, nir_umax(nb, src[0], src[1]), src[2]);
489       return;
490    case GLSLstd450SClamp:
491       val->ssa->def = nir_imin(nb, nir_imax(nb, src[0], src[1]), src[2]);
492       return;
493 
494    case GLSLstd450Cross: {
495       unsigned yzx[4] = { 1, 2, 0, 0 };
496       unsigned zxy[4] = { 2, 0, 1, 0 };
497       val->ssa->def =
498          nir_fsub(nb, nir_fmul(nb, nir_swizzle(nb, src[0], yzx, 3, true),
499                                    nir_swizzle(nb, src[1], zxy, 3, true)),
500                       nir_fmul(nb, nir_swizzle(nb, src[0], zxy, 3, true),
501                                    nir_swizzle(nb, src[1], yzx, 3, true)));
502       return;
503    }
504 
505    case GLSLstd450SmoothStep: {
506       /* t = clamp((x - edge0) / (edge1 - edge0), 0, 1) */
507       nir_ssa_def *t =
508          build_fclamp(nb, nir_fdiv(nb, nir_fsub(nb, src[2], src[0]),
509                                        nir_fsub(nb, src[1], src[0])),
510                           nir_imm_float(nb, 0.0), nir_imm_float(nb, 1.0));
511       /* result = t * t * (3 - 2 * t) */
512       val->ssa->def =
513          nir_fmul(nb, t, nir_fmul(nb, t,
514             nir_fsub(nb, nir_imm_float(nb, 3.0),
515                          nir_fmul(nb, nir_imm_float(nb, 2.0), t))));
516       return;
517    }
518 
519    case GLSLstd450FaceForward:
520       val->ssa->def =
521          nir_bcsel(nb, nir_flt(nb, nir_fdot(nb, src[2], src[1]),
522                                    nir_imm_float(nb, 0.0)),
523                        src[0], nir_fneg(nb, src[0]));
524       return;
525 
526    case GLSLstd450Reflect:
527       /* I - 2 * dot(N, I) * N */
528       val->ssa->def =
529          nir_fsub(nb, src[0], nir_fmul(nb, nir_imm_float(nb, 2.0),
530                               nir_fmul(nb, nir_fdot(nb, src[0], src[1]),
531                                            src[1])));
532       return;
533 
534    case GLSLstd450Refract: {
535       nir_ssa_def *I = src[0];
536       nir_ssa_def *N = src[1];
537       nir_ssa_def *eta = src[2];
538       nir_ssa_def *n_dot_i = nir_fdot(nb, N, I);
539       nir_ssa_def *one = nir_imm_float(nb, 1.0);
540       nir_ssa_def *zero = nir_imm_float(nb, 0.0);
541       /* k = 1.0 - eta * eta * (1.0 - dot(N, I) * dot(N, I)) */
542       nir_ssa_def *k =
543          nir_fsub(nb, one, nir_fmul(nb, eta, nir_fmul(nb, eta,
544                       nir_fsub(nb, one, nir_fmul(nb, n_dot_i, n_dot_i)))));
545       nir_ssa_def *result =
546          nir_fsub(nb, nir_fmul(nb, eta, I),
547                       nir_fmul(nb, nir_fadd(nb, nir_fmul(nb, eta, n_dot_i),
548                                                 nir_fsqrt(nb, k)), N));
549       /* XXX: bcsel, or if statement? */
550       val->ssa->def = nir_bcsel(nb, nir_flt(nb, k, zero), zero, result);
551       return;
552    }
553 
554    case GLSLstd450Sinh:
555       /* 0.5 * (e^x - e^(-x)) */
556       val->ssa->def =
557          nir_fmul(nb, nir_imm_float(nb, 0.5f),
558                       nir_fsub(nb, build_exp(nb, src[0]),
559                                    build_exp(nb, nir_fneg(nb, src[0]))));
560       return;
561 
562    case GLSLstd450Cosh:
563       /* 0.5 * (e^x + e^(-x)) */
564       val->ssa->def =
565          nir_fmul(nb, nir_imm_float(nb, 0.5f),
566                       nir_fadd(nb, build_exp(nb, src[0]),
567                                    build_exp(nb, nir_fneg(nb, src[0]))));
568       return;
569 
570    case GLSLstd450Tanh: {
571       /* tanh(x) := (0.5 * (e^x - e^(-x))) / (0.5 * (e^x + e^(-x)))
572        *
573        * With a little algebra this reduces to (e^2x - 1) / (e^2x + 1)
574        *
575        * We clamp x to (-inf, +10] to avoid precision problems.  When x > 10,
576        * e^2x is so much larger than 1.0 that 1.0 gets flushed to zero in the
577        * computation e^2x +/- 1 so it can be ignored.
578        */
579       nir_ssa_def *x = nir_fmin(nb, src[0], nir_imm_float(nb, 10));
580       nir_ssa_def *exp2x = build_exp(nb, nir_fmul(nb, x, nir_imm_float(nb, 2)));
581       val->ssa->def = nir_fdiv(nb, nir_fsub(nb, exp2x, nir_imm_float(nb, 1)),
582                                    nir_fadd(nb, exp2x, nir_imm_float(nb, 1)));
583       return;
584    }
585 
586    case GLSLstd450Asinh:
587       val->ssa->def = nir_fmul(nb, nir_fsign(nb, src[0]),
588          build_log(nb, nir_fadd(nb, nir_fabs(nb, src[0]),
589                        nir_fsqrt(nb, nir_fadd(nb, nir_fmul(nb, src[0], src[0]),
590                                                   nir_imm_float(nb, 1.0f))))));
591       return;
592    case GLSLstd450Acosh:
593       val->ssa->def = build_log(nb, nir_fadd(nb, src[0],
594          nir_fsqrt(nb, nir_fsub(nb, nir_fmul(nb, src[0], src[0]),
595                                     nir_imm_float(nb, 1.0f)))));
596       return;
597    case GLSLstd450Atanh: {
598       nir_ssa_def *one = nir_imm_float(nb, 1.0);
599       val->ssa->def = nir_fmul(nb, nir_imm_float(nb, 0.5f),
600          build_log(nb, nir_fdiv(nb, nir_fadd(nb, one, src[0]),
601                                     nir_fsub(nb, one, src[0]))));
602       return;
603    }
604 
605    case GLSLstd450Asin:
606       val->ssa->def = build_asin(nb, src[0], 0.086566724, -0.03102955);
607       return;
608 
609    case GLSLstd450Acos:
610       val->ssa->def = nir_fsub(nb, nir_imm_float(nb, M_PI_2f),
611                                build_asin(nb, src[0], 0.08132463, -0.02363318));
612       return;
613 
614    case GLSLstd450Atan:
615       val->ssa->def = build_atan(nb, src[0]);
616       return;
617 
618    case GLSLstd450Atan2:
619       val->ssa->def = build_atan2(nb, src[0], src[1]);
620       return;
621 
622    case GLSLstd450Frexp: {
623       nir_ssa_def *exponent;
624       val->ssa->def = build_frexp(nb, src[0], &exponent);
625       nir_store_deref_var(nb, vtn_nir_deref(b, w[6]), exponent, 0xf);
626       return;
627    }
628 
629    case GLSLstd450FrexpStruct: {
630       assert(glsl_type_is_struct(val->ssa->type));
631       val->ssa->elems[0]->def = build_frexp(nb, src[0],
632                                             &val->ssa->elems[1]->def);
633       return;
634    }
635 
636    default:
637       val->ssa->def =
638          nir_build_alu(&b->nb, vtn_nir_alu_op_for_spirv_glsl_opcode(entrypoint),
639                        src[0], src[1], src[2], NULL);
640       return;
641    }
642 }
643 
644 static void
handle_glsl450_interpolation(struct vtn_builder * b,enum GLSLstd450 opcode,const uint32_t * w,unsigned count)645 handle_glsl450_interpolation(struct vtn_builder *b, enum GLSLstd450 opcode,
646                              const uint32_t *w, unsigned count)
647 {
648    const struct glsl_type *dest_type =
649       vtn_value(b, w[1], vtn_value_type_type)->type->type;
650 
651    struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
652    val->ssa = vtn_create_ssa_value(b, dest_type);
653 
654    nir_intrinsic_op op;
655    switch (opcode) {
656    case GLSLstd450InterpolateAtCentroid:
657       op = nir_intrinsic_interp_var_at_centroid;
658       break;
659    case GLSLstd450InterpolateAtSample:
660       op = nir_intrinsic_interp_var_at_sample;
661       break;
662    case GLSLstd450InterpolateAtOffset:
663       op = nir_intrinsic_interp_var_at_offset;
664       break;
665    default:
666       unreachable("Invalid opcode");
667    }
668 
669    nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(b->nb.shader, op);
670 
671    nir_deref_var *deref = vtn_nir_deref(b, w[5]);
672    intrin->variables[0] = nir_deref_var_clone(deref, intrin);
673 
674    switch (opcode) {
675    case GLSLstd450InterpolateAtCentroid:
676       break;
677    case GLSLstd450InterpolateAtSample:
678    case GLSLstd450InterpolateAtOffset:
679       intrin->src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[6])->def);
680       break;
681    default:
682       unreachable("Invalid opcode");
683    }
684 
685    intrin->num_components = glsl_get_vector_elements(dest_type);
686    nir_ssa_dest_init(&intrin->instr, &intrin->dest,
687                      glsl_get_vector_elements(dest_type),
688                      glsl_get_bit_size(dest_type), NULL);
689    val->ssa->def = &intrin->dest.ssa;
690 
691    nir_builder_instr_insert(&b->nb, &intrin->instr);
692 }
693 
694 bool
vtn_handle_glsl450_instruction(struct vtn_builder * b,uint32_t ext_opcode,const uint32_t * w,unsigned count)695 vtn_handle_glsl450_instruction(struct vtn_builder *b, uint32_t ext_opcode,
696                                const uint32_t *w, unsigned count)
697 {
698    switch ((enum GLSLstd450)ext_opcode) {
699    case GLSLstd450Determinant: {
700       struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
701       val->ssa = rzalloc(b, struct vtn_ssa_value);
702       val->ssa->type = vtn_value(b, w[1], vtn_value_type_type)->type->type;
703       val->ssa->def = build_mat_det(b, vtn_ssa_value(b, w[5]));
704       break;
705    }
706 
707    case GLSLstd450MatrixInverse: {
708       struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
709       val->ssa = matrix_inverse(b, vtn_ssa_value(b, w[5]));
710       break;
711    }
712 
713    case GLSLstd450InterpolateAtCentroid:
714    case GLSLstd450InterpolateAtSample:
715    case GLSLstd450InterpolateAtOffset:
716       handle_glsl450_interpolation(b, ext_opcode, w, count);
717       break;
718 
719    default:
720       handle_glsl450_alu(b, (enum GLSLstd450)ext_opcode, w, count);
721    }
722 
723    return true;
724 }
725