2 * Copyright © 2015 Broadcom
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:
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
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
25 * This lowering pass supports (as configured via nir_lower_tex_options)
26 * various texture related conversions:
27 * + texture projector lowering: converts the coordinate division for
28 * texture projection to be done in ALU instructions instead of
29 * asking the texture operation to do so.
30 * + lowering RECT: converts the un-normalized RECT texture coordinates
31 * to normalized coordinates with txs plus ALU instructions
32 * + saturate s/t/r coords: to emulate certain texture clamp/wrap modes,
33 * inserts instructions to clamp specified coordinates to [0.0, 1.0].
34 * Note that this automatically triggers texture projector lowering if
35 * needed, since clamping must happen after projector lowering.
39 #include "nir_builder.h"
42 project_src(nir_builder
*b
, nir_tex_instr
*tex
)
44 /* Find the projector in the srcs list, if present. */
45 int proj_index
= nir_tex_instr_src_index(tex
, nir_tex_src_projector
);
49 b
->cursor
= nir_before_instr(&tex
->instr
);
51 nir_ssa_def
*inv_proj
=
52 nir_frcp(b
, nir_ssa_for_src(b
, tex
->src
[proj_index
].src
, 1));
54 /* Walk through the sources projecting the arguments. */
55 for (unsigned i
= 0; i
< tex
->num_srcs
; i
++) {
56 switch (tex
->src
[i
].src_type
) {
57 case nir_tex_src_coord
:
58 case nir_tex_src_comparator
:
63 nir_ssa_def
*unprojected
=
64 nir_ssa_for_src(b
, tex
->src
[i
].src
, nir_tex_instr_src_size(tex
, i
));
65 nir_ssa_def
*projected
= nir_fmul(b
, unprojected
, inv_proj
);
67 /* Array indices don't get projected, so make an new vector with the
68 * coordinate's array index untouched.
70 if (tex
->is_array
&& tex
->src
[i
].src_type
== nir_tex_src_coord
) {
71 switch (tex
->coord_components
) {
73 projected
= nir_vec4(b
,
74 nir_channel(b
, projected
, 0),
75 nir_channel(b
, projected
, 1),
76 nir_channel(b
, projected
, 2),
77 nir_channel(b
, unprojected
, 3));
80 projected
= nir_vec3(b
,
81 nir_channel(b
, projected
, 0),
82 nir_channel(b
, projected
, 1),
83 nir_channel(b
, unprojected
, 2));
86 projected
= nir_vec2(b
,
87 nir_channel(b
, projected
, 0),
88 nir_channel(b
, unprojected
, 1));
91 unreachable("bad texture coord count for array");
96 nir_instr_rewrite_src(&tex
->instr
,
98 nir_src_for_ssa(projected
));
101 nir_tex_instr_remove_src(tex
, proj_index
);
105 get_texture_size(nir_builder
*b
, nir_tex_instr
*tex
)
107 b
->cursor
= nir_before_instr(&tex
->instr
);
111 unsigned num_srcs
= 1; /* One for the LOD */
112 for (unsigned i
= 0; i
< tex
->num_srcs
; i
++) {
113 if (tex
->src
[i
].src_type
== nir_tex_src_texture_deref
||
114 tex
->src
[i
].src_type
== nir_tex_src_sampler_deref
||
115 tex
->src
[i
].src_type
== nir_tex_src_texture_offset
||
116 tex
->src
[i
].src_type
== nir_tex_src_sampler_offset
)
120 txs
= nir_tex_instr_create(b
->shader
, num_srcs
);
121 txs
->op
= nir_texop_txs
;
122 txs
->sampler_dim
= tex
->sampler_dim
;
123 txs
->is_array
= tex
->is_array
;
124 txs
->is_shadow
= tex
->is_shadow
;
125 txs
->is_new_style_shadow
= tex
->is_new_style_shadow
;
126 txs
->texture_index
= tex
->texture_index
;
127 txs
->texture
= nir_deref_var_clone(tex
->texture
, txs
);
128 txs
->sampler_index
= tex
->sampler_index
;
129 txs
->sampler
= nir_deref_var_clone(tex
->sampler
, txs
);
130 txs
->dest_type
= nir_type_int
;
133 for (unsigned i
= 0; i
< tex
->num_srcs
; i
++) {
134 if (tex
->src
[i
].src_type
== nir_tex_src_texture_deref
||
135 tex
->src
[i
].src_type
== nir_tex_src_sampler_deref
||
136 tex
->src
[i
].src_type
== nir_tex_src_texture_offset
||
137 tex
->src
[i
].src_type
== nir_tex_src_sampler_offset
) {
138 nir_src_copy(&txs
->src
[idx
].src
, &tex
->src
[i
].src
, txs
);
139 txs
->src
[idx
].src_type
= tex
->src
[i
].src_type
;
143 /* Add in an LOD because some back-ends require it */
144 txs
->src
[idx
].src
= nir_src_for_ssa(nir_imm_int(b
, 0));
145 txs
->src
[idx
].src_type
= nir_tex_src_lod
;
147 nir_ssa_dest_init(&txs
->instr
, &txs
->dest
,
148 nir_tex_instr_dest_size(txs
), 32, NULL
);
149 nir_builder_instr_insert(b
, &txs
->instr
);
151 return nir_i2f32(b
, &txs
->dest
.ssa
);
155 lower_offset(nir_builder
*b
, nir_tex_instr
*tex
)
157 int offset_index
= nir_tex_instr_src_index(tex
, nir_tex_src_offset
);
158 if (offset_index
< 0)
161 int coord_index
= nir_tex_instr_src_index(tex
, nir_tex_src_coord
);
162 assert(coord_index
>= 0);
164 assert(tex
->src
[offset_index
].src
.is_ssa
);
165 assert(tex
->src
[coord_index
].src
.is_ssa
);
166 nir_ssa_def
*offset
= tex
->src
[offset_index
].src
.ssa
;
167 nir_ssa_def
*coord
= tex
->src
[coord_index
].src
.ssa
;
169 b
->cursor
= nir_before_instr(&tex
->instr
);
171 nir_ssa_def
*offset_coord
;
172 if (nir_tex_instr_src_type(tex
, coord_index
) == nir_type_float
) {
173 if (tex
->sampler_dim
== GLSL_SAMPLER_DIM_RECT
) {
174 offset_coord
= nir_fadd(b
, coord
, nir_i2f32(b
, offset
));
176 nir_ssa_def
*txs
= get_texture_size(b
, tex
);
177 nir_ssa_def
*scale
= nir_frcp(b
, txs
);
179 offset_coord
= nir_fadd(b
, coord
,
181 nir_i2f32(b
, offset
),
185 offset_coord
= nir_iadd(b
, coord
, offset
);
189 /* The offset is not applied to the array index */
190 if (tex
->coord_components
== 2) {
191 offset_coord
= nir_vec2(b
, nir_channel(b
, offset_coord
, 0),
192 nir_channel(b
, coord
, 1));
193 } else if (tex
->coord_components
== 3) {
194 offset_coord
= nir_vec3(b
, nir_channel(b
, offset_coord
, 0),
195 nir_channel(b
, offset_coord
, 1),
196 nir_channel(b
, coord
, 2));
198 unreachable("Invalid number of components");
202 nir_instr_rewrite_src(&tex
->instr
, &tex
->src
[coord_index
].src
,
203 nir_src_for_ssa(offset_coord
));
205 nir_tex_instr_remove_src(tex
, offset_index
);
211 lower_rect(nir_builder
*b
, nir_tex_instr
*tex
)
213 nir_ssa_def
*txs
= get_texture_size(b
, tex
);
214 nir_ssa_def
*scale
= nir_frcp(b
, txs
);
216 /* Walk through the sources normalizing the requested arguments. */
217 for (unsigned i
= 0; i
< tex
->num_srcs
; i
++) {
218 if (tex
->src
[i
].src_type
!= nir_tex_src_coord
)
221 nir_ssa_def
*coords
=
222 nir_ssa_for_src(b
, tex
->src
[i
].src
, tex
->coord_components
);
223 nir_instr_rewrite_src(&tex
->instr
,
225 nir_src_for_ssa(nir_fmul(b
, coords
, scale
)));
228 tex
->sampler_dim
= GLSL_SAMPLER_DIM_2D
;
232 sample_plane(nir_builder
*b
, nir_tex_instr
*tex
, int plane
)
234 assert(tex
->dest
.is_ssa
);
235 assert(nir_tex_instr_dest_size(tex
) == 4);
236 assert(nir_alu_type_get_base_type(tex
->dest_type
) == nir_type_float
);
237 assert(tex
->op
== nir_texop_tex
);
238 assert(tex
->coord_components
== 2);
240 nir_tex_instr
*plane_tex
=
241 nir_tex_instr_create(b
->shader
, tex
->num_srcs
+ 1);
242 for (unsigned i
= 0; i
< tex
->num_srcs
; i
++) {
243 nir_src_copy(&plane_tex
->src
[i
].src
, &tex
->src
[i
].src
, plane_tex
);
244 plane_tex
->src
[i
].src_type
= tex
->src
[i
].src_type
;
246 plane_tex
->src
[tex
->num_srcs
].src
= nir_src_for_ssa(nir_imm_int(b
, plane
));
247 plane_tex
->src
[tex
->num_srcs
].src_type
= nir_tex_src_plane
;
248 plane_tex
->op
= nir_texop_tex
;
249 plane_tex
->sampler_dim
= GLSL_SAMPLER_DIM_2D
;
250 plane_tex
->dest_type
= nir_type_float
;
251 plane_tex
->coord_components
= 2;
253 plane_tex
->texture_index
= tex
->texture_index
;
254 plane_tex
->texture
= nir_deref_var_clone(tex
->texture
, plane_tex
);
255 plane_tex
->sampler_index
= tex
->sampler_index
;
256 plane_tex
->sampler
= nir_deref_var_clone(tex
->sampler
, plane_tex
);
258 nir_ssa_dest_init(&plane_tex
->instr
, &plane_tex
->dest
, 4, 32, NULL
);
260 nir_builder_instr_insert(b
, &plane_tex
->instr
);
262 return &plane_tex
->dest
.ssa
;
266 convert_yuv_to_rgb(nir_builder
*b
, nir_tex_instr
*tex
,
267 nir_ssa_def
*y
, nir_ssa_def
*u
, nir_ssa_def
*v
)
269 nir_const_value m
[3] = {
270 { .f32
= { 1.0f
, 0.0f
, 1.59602678f
, 0.0f
} },
271 { .f32
= { 1.0f
, -0.39176229f
, -0.81296764f
, 0.0f
} },
272 { .f32
= { 1.0f
, 2.01723214f
, 0.0f
, 0.0f
} }
277 nir_fmul(b
, nir_imm_float(b
, 1.16438356f
),
278 nir_fadd(b
, y
, nir_imm_float(b
, -16.0f
/ 255.0f
))),
279 nir_channel(b
, nir_fadd(b
, u
, nir_imm_float(b
, -128.0f
/ 255.0f
)), 0),
280 nir_channel(b
, nir_fadd(b
, v
, nir_imm_float(b
, -128.0f
/ 255.0f
)), 0),
281 nir_imm_float(b
, 0.0));
283 nir_ssa_def
*red
= nir_fdot4(b
, yuv
, nir_build_imm(b
, 4, 32, m
[0]));
284 nir_ssa_def
*green
= nir_fdot4(b
, yuv
, nir_build_imm(b
, 4, 32, m
[1]));
285 nir_ssa_def
*blue
= nir_fdot4(b
, yuv
, nir_build_imm(b
, 4, 32, m
[2]));
287 nir_ssa_def
*result
= nir_vec4(b
, red
, green
, blue
, nir_imm_float(b
, 1.0f
));
289 nir_ssa_def_rewrite_uses(&tex
->dest
.ssa
, nir_src_for_ssa(result
));
293 lower_y_uv_external(nir_builder
*b
, nir_tex_instr
*tex
)
295 b
->cursor
= nir_after_instr(&tex
->instr
);
297 nir_ssa_def
*y
= sample_plane(b
, tex
, 0);
298 nir_ssa_def
*uv
= sample_plane(b
, tex
, 1);
300 convert_yuv_to_rgb(b
, tex
,
301 nir_channel(b
, y
, 0),
302 nir_channel(b
, uv
, 0),
303 nir_channel(b
, uv
, 1));
307 lower_y_u_v_external(nir_builder
*b
, nir_tex_instr
*tex
)
309 b
->cursor
= nir_after_instr(&tex
->instr
);
311 nir_ssa_def
*y
= sample_plane(b
, tex
, 0);
312 nir_ssa_def
*u
= sample_plane(b
, tex
, 1);
313 nir_ssa_def
*v
= sample_plane(b
, tex
, 2);
315 convert_yuv_to_rgb(b
, tex
,
316 nir_channel(b
, y
, 0),
317 nir_channel(b
, u
, 0),
318 nir_channel(b
, v
, 0));
322 lower_yx_xuxv_external(nir_builder
*b
, nir_tex_instr
*tex
)
324 b
->cursor
= nir_after_instr(&tex
->instr
);
326 nir_ssa_def
*y
= sample_plane(b
, tex
, 0);
327 nir_ssa_def
*xuxv
= sample_plane(b
, tex
, 1);
329 convert_yuv_to_rgb(b
, tex
,
330 nir_channel(b
, y
, 0),
331 nir_channel(b
, xuxv
, 1),
332 nir_channel(b
, xuxv
, 3));
336 lower_xy_uxvx_external(nir_builder
*b
, nir_tex_instr
*tex
)
338 b
->cursor
= nir_after_instr(&tex
->instr
);
340 nir_ssa_def
*y
= sample_plane(b
, tex
, 0);
341 nir_ssa_def
*uxvx
= sample_plane(b
, tex
, 1);
343 convert_yuv_to_rgb(b
, tex
,
344 nir_channel(b
, y
, 1),
345 nir_channel(b
, uxvx
, 0),
346 nir_channel(b
, uxvx
, 2));
350 * Emits a textureLod operation used to replace an existing
351 * textureGrad instruction.
354 replace_gradient_with_lod(nir_builder
*b
, nir_ssa_def
*lod
, nir_tex_instr
*tex
)
356 /* We are going to emit a textureLod() with the same parameters except that
357 * we replace ddx/ddy with lod.
359 int num_srcs
= tex
->num_srcs
- 1;
360 nir_tex_instr
*txl
= nir_tex_instr_create(b
->shader
, num_srcs
);
362 txl
->op
= nir_texop_txl
;
363 txl
->sampler_dim
= tex
->sampler_dim
;
364 txl
->texture_index
= tex
->texture_index
;
365 txl
->dest_type
= tex
->dest_type
;
366 txl
->is_array
= tex
->is_array
;
367 txl
->is_shadow
= tex
->is_shadow
;
368 txl
->is_new_style_shadow
= tex
->is_new_style_shadow
;
369 txl
->sampler_index
= tex
->sampler_index
;
370 txl
->texture
= nir_deref_var_clone(tex
->texture
, txl
);
371 txl
->sampler
= nir_deref_var_clone(tex
->sampler
, txl
);
372 txl
->coord_components
= tex
->coord_components
;
374 nir_ssa_dest_init(&txl
->instr
, &txl
->dest
, 4, 32, NULL
);
377 for (int i
= 0; i
< tex
->num_srcs
; i
++) {
378 if (tex
->src
[i
].src_type
== nir_tex_src_ddx
||
379 tex
->src
[i
].src_type
== nir_tex_src_ddy
)
381 nir_src_copy(&txl
->src
[src_num
].src
, &tex
->src
[i
].src
, txl
);
382 txl
->src
[src_num
].src_type
= tex
->src
[i
].src_type
;
386 txl
->src
[src_num
].src
= nir_src_for_ssa(lod
);
387 txl
->src
[src_num
].src_type
= nir_tex_src_lod
;
390 assert(src_num
== num_srcs
);
392 nir_ssa_dest_init(&txl
->instr
, &txl
->dest
,
393 tex
->dest
.ssa
.num_components
, 32, NULL
);
394 nir_builder_instr_insert(b
, &txl
->instr
);
396 nir_ssa_def_rewrite_uses(&tex
->dest
.ssa
, nir_src_for_ssa(&txl
->dest
.ssa
));
398 nir_instr_remove(&tex
->instr
);
402 lower_gradient_cube_map(nir_builder
*b
, nir_tex_instr
*tex
)
404 assert(tex
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
);
405 assert(tex
->op
== nir_texop_txd
);
406 assert(tex
->dest
.is_ssa
);
408 /* Use textureSize() to get the width and height of LOD 0 */
409 nir_ssa_def
*size
= get_texture_size(b
, tex
);
411 /* Cubemap texture lookups first generate a texture coordinate normalized
412 * to [-1, 1] on the appropiate face. The appropiate face is determined
413 * by which component has largest magnitude and its sign. The texture
414 * coordinate is the quotient of the remaining texture coordinates against
415 * that absolute value of the component of largest magnitude. This
416 * division requires that the computing of the derivative of the texel
417 * coordinate must use the quotient rule. The high level GLSL code is as
422 * vec3 abs_p, Q, dQdx, dQdy;
423 * abs_p = abs(ir->coordinate);
424 * if (abs_p.x >= max(abs_p.y, abs_p.z)) {
425 * Q = ir->coordinate.yzx;
426 * dQdx = ir->lod_info.grad.dPdx.yzx;
427 * dQdy = ir->lod_info.grad.dPdy.yzx;
429 * if (abs_p.y >= max(abs_p.x, abs_p.z)) {
430 * Q = ir->coordinate.xzy;
431 * dQdx = ir->lod_info.grad.dPdx.xzy;
432 * dQdy = ir->lod_info.grad.dPdy.xzy;
434 * if (abs_p.z >= max(abs_p.x, abs_p.y)) {
435 * Q = ir->coordinate;
436 * dQdx = ir->lod_info.grad.dPdx;
437 * dQdy = ir->lod_info.grad.dPdy;
440 * Step 2: use quotient rule to compute derivative. The normalized to
441 * [-1, 1] texel coordinate is given by Q.xy / (sign(Q.z) * Q.z). We are
442 * only concerned with the magnitudes of the derivatives whose values are
443 * not affected by the sign. We drop the sign from the computation.
449 * dx = recip * ( dQdx.xy - Q.xy * (dQdx.z * recip) );
450 * dy = recip * ( dQdy.xy - Q.xy * (dQdy.z * recip) );
452 * Step 3: compute LOD. At this point we have the derivatives of the
453 * texture coordinates normalized to [-1,1]. We take the LOD to be
454 * result = log2(max(sqrt(dot(dx, dx)), sqrt(dy, dy)) * 0.5 * L)
455 * = -1.0 + log2(max(sqrt(dot(dx, dx)), sqrt(dy, dy)) * L)
456 * = -1.0 + log2(sqrt(max(dot(dx, dx), dot(dy,dy))) * L)
457 * = -1.0 + log2(sqrt(L * L * max(dot(dx, dx), dot(dy,dy))))
458 * = -1.0 + 0.5 * log2(L * L * max(dot(dx, dx), dot(dy,dy)))
459 * where L is the dimension of the cubemap. The code is:
462 * M = max(dot(dx, dx), dot(dy, dy));
463 * L = textureSize(sampler, 0).x;
464 * result = -1.0 + 0.5 * log2(L * L * M);
469 tex
->src
[nir_tex_instr_src_index(tex
, nir_tex_src_coord
)].src
.ssa
;
471 /* unmodified dPdx, dPdy values */
473 tex
->src
[nir_tex_instr_src_index(tex
, nir_tex_src_ddx
)].src
.ssa
;
475 tex
->src
[nir_tex_instr_src_index(tex
, nir_tex_src_ddy
)].src
.ssa
;
477 nir_ssa_def
*abs_p
= nir_fabs(b
, p
);
478 nir_ssa_def
*abs_p_x
= nir_channel(b
, abs_p
, 0);
479 nir_ssa_def
*abs_p_y
= nir_channel(b
, abs_p
, 1);
480 nir_ssa_def
*abs_p_z
= nir_channel(b
, abs_p
, 2);
482 /* 1. compute selector */
483 nir_ssa_def
*Q
, *dQdx
, *dQdy
;
485 nir_ssa_def
*cond_z
= nir_fge(b
, abs_p_z
, nir_fmax(b
, abs_p_x
, abs_p_y
));
486 nir_ssa_def
*cond_y
= nir_fge(b
, abs_p_y
, nir_fmax(b
, abs_p_x
, abs_p_z
));
488 unsigned yzx
[4] = { 1, 2, 0, 0 };
489 unsigned xzy
[4] = { 0, 2, 1, 0 };
491 Q
= nir_bcsel(b
, cond_z
,
494 nir_swizzle(b
, p
, xzy
, 3, false),
495 nir_swizzle(b
, p
, yzx
, 3, false)));
497 dQdx
= nir_bcsel(b
, cond_z
,
500 nir_swizzle(b
, dPdx
, xzy
, 3, false),
501 nir_swizzle(b
, dPdx
, yzx
, 3, false)));
503 dQdy
= nir_bcsel(b
, cond_z
,
506 nir_swizzle(b
, dPdy
, xzy
, 3, false),
507 nir_swizzle(b
, dPdy
, yzx
, 3, false)));
509 /* 2. quotient rule */
511 /* tmp = Q.xy * recip;
512 * dx = recip * ( dQdx.xy - (tmp * dQdx.z) );
513 * dy = recip * ( dQdy.xy - (tmp * dQdy.z) );
515 nir_ssa_def
*rcp_Q_z
= nir_frcp(b
, nir_channel(b
, Q
, 2));
517 unsigned xy
[4] = { 0, 1, 0, 0 };
518 nir_ssa_def
*Q_xy
= nir_swizzle(b
, Q
, xy
, 2, false);
519 nir_ssa_def
*tmp
= nir_fmul(b
, Q_xy
, rcp_Q_z
);
521 nir_ssa_def
*dQdx_xy
= nir_swizzle(b
, dQdx
, xy
, 2, false);
522 nir_ssa_def
*dQdx_z
= nir_channel(b
, dQdx
, 2);
524 nir_fmul(b
, rcp_Q_z
, nir_fsub(b
, dQdx_xy
, nir_fmul(b
, tmp
, dQdx_z
)));
526 nir_ssa_def
*dQdy_xy
= nir_swizzle(b
, dQdy
, xy
, 2, false);
527 nir_ssa_def
*dQdy_z
= nir_channel(b
, dQdy
, 2);
529 nir_fmul(b
, rcp_Q_z
, nir_fsub(b
, dQdy_xy
, nir_fmul(b
, tmp
, dQdy_z
)));
531 /* M = max(dot(dx, dx), dot(dy, dy)); */
532 nir_ssa_def
*M
= nir_fmax(b
, nir_fdot(b
, dx
, dx
), nir_fdot(b
, dy
, dy
));
534 /* size has textureSize() of LOD 0 */
535 nir_ssa_def
*L
= nir_channel(b
, size
, 0);
537 /* lod = -1.0 + 0.5 * log2(L * L * M); */
540 nir_imm_float(b
, -1.0f
),
542 nir_imm_float(b
, 0.5f
),
543 nir_flog2(b
, nir_fmul(b
, L
, nir_fmul(b
, L
, M
)))));
545 /* 3. Replace the gradient instruction with an equivalent lod instruction */
546 replace_gradient_with_lod(b
, lod
, tex
);
550 lower_gradient(nir_builder
*b
, nir_tex_instr
*tex
)
552 assert(tex
->sampler_dim
!= GLSL_SAMPLER_DIM_CUBE
);
553 assert(tex
->op
== nir_texop_txd
);
554 assert(tex
->dest
.is_ssa
);
556 /* Use textureSize() to get the width and height of LOD 0 */
557 unsigned component_mask
;
558 switch (tex
->sampler_dim
) {
559 case GLSL_SAMPLER_DIM_3D
:
562 case GLSL_SAMPLER_DIM_1D
:
571 nir_channels(b
, get_texture_size(b
, tex
), component_mask
);
573 /* Scale the gradients by width and height. Effectively, the incoming
574 * gradients are s'(x,y), t'(x,y), and r'(x,y) from equation 3.19 in the
575 * GL 3.0 spec; we want u'(x,y), which is w_t * s'(x,y).
578 tex
->src
[nir_tex_instr_src_index(tex
, nir_tex_src_ddx
)].src
.ssa
;
580 tex
->src
[nir_tex_instr_src_index(tex
, nir_tex_src_ddy
)].src
.ssa
;
582 nir_ssa_def
*dPdx
= nir_fmul(b
, ddx
, size
);
583 nir_ssa_def
*dPdy
= nir_fmul(b
, ddy
, size
);
586 if (dPdx
->num_components
== 1) {
587 rho
= nir_fmax(b
, nir_fabs(b
, dPdx
), nir_fabs(b
, dPdy
));
590 nir_fsqrt(b
, nir_fdot(b
, dPdx
, dPdx
)),
591 nir_fsqrt(b
, nir_fdot(b
, dPdy
, dPdy
)));
594 /* lod = log2(rho). We're ignoring GL state biases for now. */
595 nir_ssa_def
*lod
= nir_flog2(b
, rho
);
597 /* Replace the gradient instruction with an equivalent lod instruction */
598 replace_gradient_with_lod(b
, lod
, tex
);
602 saturate_src(nir_builder
*b
, nir_tex_instr
*tex
, unsigned sat_mask
)
604 b
->cursor
= nir_before_instr(&tex
->instr
);
606 /* Walk through the sources saturating the requested arguments. */
607 for (unsigned i
= 0; i
< tex
->num_srcs
; i
++) {
608 if (tex
->src
[i
].src_type
!= nir_tex_src_coord
)
612 nir_ssa_for_src(b
, tex
->src
[i
].src
, tex
->coord_components
);
614 /* split src into components: */
615 nir_ssa_def
*comp
[4];
617 assume(tex
->coord_components
>= 1);
619 for (unsigned j
= 0; j
< tex
->coord_components
; j
++)
620 comp
[j
] = nir_channel(b
, src
, j
);
622 /* clamp requested components, array index does not get clamped: */
623 unsigned ncomp
= tex
->coord_components
;
627 for (unsigned j
= 0; j
< ncomp
; j
++) {
628 if ((1 << j
) & sat_mask
) {
629 if (tex
->sampler_dim
== GLSL_SAMPLER_DIM_RECT
) {
630 /* non-normalized texture coords, so clamp to texture
631 * size rather than [0.0, 1.0]
633 nir_ssa_def
*txs
= get_texture_size(b
, tex
);
634 comp
[j
] = nir_fmax(b
, comp
[j
], nir_imm_float(b
, 0.0));
635 comp
[j
] = nir_fmin(b
, comp
[j
], nir_channel(b
, txs
, j
));
637 comp
[j
] = nir_fsat(b
, comp
[j
]);
642 /* and move the result back into a single vecN: */
643 src
= nir_vec(b
, comp
, tex
->coord_components
);
645 nir_instr_rewrite_src(&tex
->instr
,
647 nir_src_for_ssa(src
));
652 get_zero_or_one(nir_builder
*b
, nir_alu_type type
, uint8_t swizzle_val
)
656 memset(&v
, 0, sizeof(v
));
658 if (swizzle_val
== 4) {
659 v
.u32
[0] = v
.u32
[1] = v
.u32
[2] = v
.u32
[3] = 0;
661 assert(swizzle_val
== 5);
662 if (type
== nir_type_float
)
663 v
.f32
[0] = v
.f32
[1] = v
.f32
[2] = v
.f32
[3] = 1.0;
665 v
.u32
[0] = v
.u32
[1] = v
.u32
[2] = v
.u32
[3] = 1;
668 return nir_build_imm(b
, 4, 32, v
);
672 swizzle_result(nir_builder
*b
, nir_tex_instr
*tex
, const uint8_t swizzle
[4])
674 assert(tex
->dest
.is_ssa
);
676 b
->cursor
= nir_after_instr(&tex
->instr
);
678 nir_ssa_def
*swizzled
;
679 if (tex
->op
== nir_texop_tg4
) {
680 if (swizzle
[tex
->component
] < 4) {
681 /* This one's easy */
682 tex
->component
= swizzle
[tex
->component
];
685 swizzled
= get_zero_or_one(b
, tex
->dest_type
, swizzle
[tex
->component
]);
688 assert(nir_tex_instr_dest_size(tex
) == 4);
689 if (swizzle
[0] < 4 && swizzle
[1] < 4 &&
690 swizzle
[2] < 4 && swizzle
[3] < 4) {
691 unsigned swiz
[4] = { swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3] };
692 /* We have no 0s or 1s, just emit a swizzling MOV */
693 swizzled
= nir_swizzle(b
, &tex
->dest
.ssa
, swiz
, 4, false);
695 nir_ssa_def
*srcs
[4];
696 for (unsigned i
= 0; i
< 4; i
++) {
697 if (swizzle
[i
] < 4) {
698 srcs
[i
] = nir_channel(b
, &tex
->dest
.ssa
, swizzle
[i
]);
700 srcs
[i
] = get_zero_or_one(b
, tex
->dest_type
, swizzle
[i
]);
703 swizzled
= nir_vec(b
, srcs
, 4);
707 nir_ssa_def_rewrite_uses_after(&tex
->dest
.ssa
, nir_src_for_ssa(swizzled
),
708 swizzled
->parent_instr
);
712 linearize_srgb_result(nir_builder
*b
, nir_tex_instr
*tex
)
714 assert(tex
->dest
.is_ssa
);
715 assert(nir_tex_instr_dest_size(tex
) == 4);
716 assert(nir_alu_type_get_base_type(tex
->dest_type
) == nir_type_float
);
718 b
->cursor
= nir_after_instr(&tex
->instr
);
720 static const unsigned swiz
[4] = {0, 1, 2, 0};
721 nir_ssa_def
*comp
= nir_swizzle(b
, &tex
->dest
.ssa
, swiz
, 3, true);
724 * (comp <= 0.04045) ?
726 * pow((comp + 0.055) / 1.055, 2.4)
728 nir_ssa_def
*low
= nir_fmul(b
, comp
, nir_imm_float(b
, 1.0 / 12.92));
729 nir_ssa_def
*high
= nir_fpow(b
,
733 nir_imm_float(b
, 0.055)),
734 nir_imm_float(b
, 1.0 / 1.055)),
735 nir_imm_float(b
, 2.4));
736 nir_ssa_def
*cond
= nir_fge(b
, nir_imm_float(b
, 0.04045), comp
);
737 nir_ssa_def
*rgb
= nir_bcsel(b
, cond
, low
, high
);
739 /* alpha is untouched: */
740 nir_ssa_def
*result
= nir_vec4(b
,
741 nir_channel(b
, rgb
, 0),
742 nir_channel(b
, rgb
, 1),
743 nir_channel(b
, rgb
, 2),
744 nir_channel(b
, &tex
->dest
.ssa
, 3));
746 nir_ssa_def_rewrite_uses_after(&tex
->dest
.ssa
, nir_src_for_ssa(result
),
747 result
->parent_instr
);
751 nir_lower_tex_block(nir_block
*block
, nir_builder
*b
,
752 const nir_lower_tex_options
*options
)
754 bool progress
= false;
756 nir_foreach_instr_safe(instr
, block
) {
757 if (instr
->type
!= nir_instr_type_tex
)
760 nir_tex_instr
*tex
= nir_instr_as_tex(instr
);
761 bool lower_txp
= !!(options
->lower_txp
& (1 << tex
->sampler_dim
));
763 /* mask of src coords to saturate (clamp): */
764 unsigned sat_mask
= 0;
766 if ((1 << tex
->sampler_index
) & options
->saturate_r
)
767 sat_mask
|= (1 << 2); /* .z */
768 if ((1 << tex
->sampler_index
) & options
->saturate_t
)
769 sat_mask
|= (1 << 1); /* .y */
770 if ((1 << tex
->sampler_index
) & options
->saturate_s
)
771 sat_mask
|= (1 << 0); /* .x */
773 /* If we are clamping any coords, we must lower projector first
774 * as clamping happens *after* projection:
776 if (lower_txp
|| sat_mask
) {
781 if ((tex
->op
== nir_texop_txf
&& options
->lower_txf_offset
) ||
782 (sat_mask
&& nir_tex_instr_src_index(tex
, nir_tex_src_coord
) >= 0) ||
783 (tex
->sampler_dim
== GLSL_SAMPLER_DIM_RECT
&&
784 options
->lower_rect_offset
)) {
785 progress
= lower_offset(b
, tex
) || progress
;
788 if ((tex
->sampler_dim
== GLSL_SAMPLER_DIM_RECT
) && options
->lower_rect
) {
793 if ((1 << tex
->texture_index
) & options
->lower_y_uv_external
) {
794 lower_y_uv_external(b
, tex
);
798 if ((1 << tex
->texture_index
) & options
->lower_y_u_v_external
) {
799 lower_y_u_v_external(b
, tex
);
803 if ((1 << tex
->texture_index
) & options
->lower_yx_xuxv_external
) {
804 lower_yx_xuxv_external(b
, tex
);
808 if ((1 << tex
->texture_index
) & options
->lower_xy_uxvx_external
) {
809 lower_xy_uxvx_external(b
, tex
);
814 saturate_src(b
, tex
, sat_mask
);
818 if (((1 << tex
->texture_index
) & options
->swizzle_result
) &&
819 !nir_tex_instr_is_query(tex
) &&
820 !(tex
->is_shadow
&& tex
->is_new_style_shadow
)) {
821 swizzle_result(b
, tex
, options
->swizzles
[tex
->texture_index
]);
825 /* should be after swizzle so we know which channels are rgb: */
826 if (((1 << tex
->texture_index
) & options
->lower_srgb
) &&
827 !nir_tex_instr_is_query(tex
) && !tex
->is_shadow
) {
828 linearize_srgb_result(b
, tex
);
832 if (tex
->op
== nir_texop_txd
&&
833 tex
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
&&
834 (options
->lower_txd
||
835 options
->lower_txd_cube_map
||
836 (tex
->is_shadow
&& options
->lower_txd_shadow
))) {
837 lower_gradient_cube_map(b
, tex
);
842 if (tex
->op
== nir_texop_txd
&&
843 (options
->lower_txd
||
844 (options
->lower_txd_shadow
&&
845 tex
->is_shadow
&& tex
->sampler_dim
!= GLSL_SAMPLER_DIM_CUBE
))) {
846 lower_gradient(b
, tex
);
851 /* TXF, TXS and TXL require a LOD but not everything we implement using those
852 * three opcodes provides one. Provide a default LOD of 0.
854 if ((nir_tex_instr_src_index(tex
, nir_tex_src_lod
) == -1) &&
855 (tex
->op
== nir_texop_txf
|| tex
->op
== nir_texop_txs
||
856 tex
->op
== nir_texop_txl
|| tex
->op
== nir_texop_query_levels
||
857 (tex
->op
== nir_texop_tex
&&
858 b
->shader
->info
.stage
!= MESA_SHADER_FRAGMENT
))) {
859 b
->cursor
= nir_before_instr(&tex
->instr
);
860 nir_tex_instr_add_src(tex
, nir_tex_src_lod
, nir_src_for_ssa(nir_imm_int(b
, 0)));
870 nir_lower_tex_impl(nir_function_impl
*impl
,
871 const nir_lower_tex_options
*options
)
873 bool progress
= false;
875 nir_builder_init(&builder
, impl
);
877 nir_foreach_block(block
, impl
) {
878 progress
|= nir_lower_tex_block(block
, &builder
, options
);
881 nir_metadata_preserve(impl
, nir_metadata_block_index
|
882 nir_metadata_dominance
);
887 nir_lower_tex(nir_shader
*shader
, const nir_lower_tex_options
*options
)
889 bool progress
= false;
891 nir_foreach_function(function
, shader
) {
893 progress
|= nir_lower_tex_impl(function
->impl
, options
);