1 /**************************************************************************
3 * Copyright 2009 VMware, Inc.
4 * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sub license, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
20 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
21 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
22 * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
23 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
24 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
25 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
27 **************************************************************************/
31 * Code generate the whole fragment pipeline.
33 * The fragment pipeline consists of the following stages:
37 * - depth/stencil test
40 * This file has only the glue to assemble the fragment pipeline. The actual
41 * plumbing of converting Gallium state into LLVM IR is done elsewhere, in the
42 * lp_bld_*.[ch] files, and in a complete generic and reusable way. Here we
43 * muster the LLVM JIT execution engine to create a function that follows an
44 * established binary interface and that can be called from C directly.
46 * A big source of complexity here is that we often want to run different
47 * stages with different precisions and data types and precisions. For example,
48 * the fragment shader needs typically to be done in floats, but the
49 * depth/stencil test and blending is better done in the type that most closely
50 * matches the depth/stencil and color buffer respectively.
52 * Since the width of a SIMD vector register stays the same regardless of the
53 * element type, different types imply different number of elements, so we must
54 * code generate more instances of the stages with larger types to be able to
55 * feed/consume the stages with smaller types.
57 * @author Jose Fonseca <jfonseca@vmware.com>
61 #include "pipe/p_defines.h"
62 #include "util/u_inlines.h"
63 #include "util/u_memory.h"
64 #include "util/u_pointer.h"
65 #include "util/u_format.h"
66 #include "util/u_dump.h"
67 #include "util/u_string.h"
68 #include "util/u_simple_list.h"
69 #include "util/u_dual_blend.h"
70 #include "os/os_time.h"
71 #include "pipe/p_shader_tokens.h"
72 #include "draw/draw_context.h"
73 #include "tgsi/tgsi_dump.h"
74 #include "tgsi/tgsi_scan.h"
75 #include "tgsi/tgsi_parse.h"
76 #include "gallivm/lp_bld_type.h"
77 #include "gallivm/lp_bld_const.h"
78 #include "gallivm/lp_bld_conv.h"
79 #include "gallivm/lp_bld_init.h"
80 #include "gallivm/lp_bld_intr.h"
81 #include "gallivm/lp_bld_logic.h"
82 #include "gallivm/lp_bld_tgsi.h"
83 #include "gallivm/lp_bld_swizzle.h"
84 #include "gallivm/lp_bld_flow.h"
85 #include "gallivm/lp_bld_debug.h"
86 #include "gallivm/lp_bld_arit.h"
87 #include "gallivm/lp_bld_pack.h"
88 #include "gallivm/lp_bld_format.h"
89 #include "gallivm/lp_bld_quad.h"
91 #include "lp_bld_alpha.h"
92 #include "lp_bld_blend.h"
93 #include "lp_bld_depth.h"
94 #include "lp_bld_interp.h"
95 #include "lp_context.h"
100 #include "lp_tex_sample.h"
101 #include "lp_flush.h"
102 #include "lp_state_fs.h"
106 /** Fragment shader number (for debugging) */
107 static unsigned fs_no
= 0;
111 * Expand the relevant bits of mask_input to a n*4-dword mask for the
112 * n*four pixels in n 2x2 quads. This will set the n*four elements of the
113 * quad mask vector to 0 or ~0.
114 * Grouping is 01, 23 for 2 quad mode hence only 0 and 2 are valid
115 * quad arguments with fs length 8.
117 * \param first_quad which quad(s) of the quad group to test, in [0,3]
118 * \param mask_input bitwise mask for the whole 4x4 stamp
121 generate_quad_mask(struct gallivm_state
*gallivm
,
122 struct lp_type fs_type
,
124 LLVMValueRef mask_input
) /* int32 */
126 LLVMBuilderRef builder
= gallivm
->builder
;
127 struct lp_type mask_type
;
128 LLVMTypeRef i32t
= LLVMInt32TypeInContext(gallivm
->context
);
129 LLVMValueRef bits
[16];
134 * XXX: We'll need a different path for 16 x u8
136 assert(fs_type
.width
== 32);
137 assert(fs_type
.length
<= Elements(bits
));
138 mask_type
= lp_int_type(fs_type
);
141 * mask_input >>= (quad * 4)
143 switch (first_quad
) {
148 assert(fs_type
.length
== 4);
155 assert(fs_type
.length
== 4);
163 mask_input
= LLVMBuildLShr(builder
,
165 LLVMConstInt(i32t
, shift
, 0),
169 * mask = { mask_input & (1 << i), for i in [0,3] }
171 mask
= lp_build_broadcast(gallivm
,
172 lp_build_vec_type(gallivm
, mask_type
),
175 for (i
= 0; i
< fs_type
.length
/ 4; i
++) {
176 unsigned j
= 2 * (i
% 2) + (i
/ 2) * 8;
177 bits
[4*i
+ 0] = LLVMConstInt(i32t
, 1 << (j
+ 0), 0);
178 bits
[4*i
+ 1] = LLVMConstInt(i32t
, 1 << (j
+ 1), 0);
179 bits
[4*i
+ 2] = LLVMConstInt(i32t
, 1 << (j
+ 4), 0);
180 bits
[4*i
+ 3] = LLVMConstInt(i32t
, 1 << (j
+ 5), 0);
182 mask
= LLVMBuildAnd(builder
, mask
, LLVMConstVector(bits
, fs_type
.length
), "");
185 * mask = mask != 0 ? ~0 : 0
187 mask
= lp_build_compare(gallivm
,
188 mask_type
, PIPE_FUNC_NOTEQUAL
,
190 lp_build_const_int_vec(gallivm
, mask_type
, 0));
196 #define EARLY_DEPTH_TEST 0x1
197 #define LATE_DEPTH_TEST 0x2
198 #define EARLY_DEPTH_WRITE 0x4
199 #define LATE_DEPTH_WRITE 0x8
202 find_output_by_semantic( const struct tgsi_shader_info
*info
,
208 for (i
= 0; i
< info
->num_outputs
; i
++)
209 if (info
->output_semantic_name
[i
] == semantic
&&
210 info
->output_semantic_index
[i
] == index
)
218 * Fetch the specified lp_jit_viewport structure for a given viewport_index.
221 lp_llvm_viewport(LLVMValueRef context_ptr
,
222 struct gallivm_state
*gallivm
,
223 LLVMValueRef viewport_index
)
225 LLVMBuilderRef builder
= gallivm
->builder
;
228 struct lp_type viewport_type
=
229 lp_type_float_vec(32, 32 * LP_JIT_VIEWPORT_NUM_FIELDS
);
231 ptr
= lp_jit_context_viewports(gallivm
, context_ptr
);
232 ptr
= LLVMBuildPointerCast(builder
, ptr
,
233 LLVMPointerType(lp_build_vec_type(gallivm
, viewport_type
), 0), "");
235 res
= lp_build_pointer_get(builder
, ptr
, viewport_index
);
242 * Generate the fragment shader, depth/stencil test, and alpha tests.
245 generate_fs_loop(struct gallivm_state
*gallivm
,
246 struct lp_fragment_shader
*shader
,
247 const struct lp_fragment_shader_variant_key
*key
,
248 LLVMBuilderRef builder
,
250 LLVMValueRef context_ptr
,
251 LLVMValueRef num_loop
,
252 struct lp_build_interp_soa_context
*interp
,
253 struct lp_build_sampler_soa
*sampler
,
254 LLVMValueRef mask_store
,
255 LLVMValueRef (*out_color
)[4],
256 LLVMValueRef depth_ptr
,
257 LLVMValueRef depth_stride
,
259 LLVMValueRef thread_data_ptr
)
261 const struct util_format_description
*zs_format_desc
= NULL
;
262 const struct tgsi_token
*tokens
= shader
->base
.tokens
;
263 LLVMTypeRef vec_type
;
264 LLVMValueRef mask_ptr
, mask_val
;
265 LLVMValueRef consts_ptr
;
267 LLVMValueRef z_value
, s_value
;
268 LLVMValueRef z_fb
, s_fb
;
269 LLVMValueRef stencil_refs
[2];
270 LLVMValueRef outputs
[PIPE_MAX_SHADER_OUTPUTS
][TGSI_NUM_CHANNELS
];
271 struct lp_build_for_loop_state loop_state
;
272 struct lp_build_mask_context mask
;
274 * TODO: figure out if simple_shader optimization is really worthwile to
275 * keep. Disabled because it may hide some real bugs in the (depth/stencil)
276 * code since tests tend to take another codepath than real shaders.
278 boolean simple_shader
= (shader
->info
.base
.file_count
[TGSI_FILE_SAMPLER
] == 0 &&
279 shader
->info
.base
.num_inputs
< 3 &&
280 shader
->info
.base
.num_instructions
< 8) && 0;
281 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
282 util_blend_state_is_dual(&key
->blend
, 0);
288 struct lp_bld_tgsi_system_values system_values
;
290 memset(&system_values
, 0, sizeof(system_values
));
292 if (key
->depth
.enabled
||
293 key
->stencil
[0].enabled
) {
295 zs_format_desc
= util_format_description(key
->zsbuf_format
);
296 assert(zs_format_desc
);
298 if (!shader
->info
.base
.writes_z
) {
299 if (key
->alpha
.enabled
|| shader
->info
.base
.uses_kill
) {
300 /* With alpha test and kill, can do the depth test early
301 * and hopefully eliminate some quads. But need to do a
302 * special deferred depth write once the final mask value
303 * is known. This only works though if there's either no
304 * stencil test or the stencil value isn't written.
306 if (key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
307 (key
->stencil
[1].enabled
&&
308 key
->stencil
[1].writemask
)))
309 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
311 depth_mode
= EARLY_DEPTH_TEST
| LATE_DEPTH_WRITE
;
314 depth_mode
= EARLY_DEPTH_TEST
| EARLY_DEPTH_WRITE
;
317 depth_mode
= LATE_DEPTH_TEST
| LATE_DEPTH_WRITE
;
320 if (!(key
->depth
.enabled
&& key
->depth
.writemask
) &&
321 !(key
->stencil
[0].enabled
&& (key
->stencil
[0].writemask
||
322 (key
->stencil
[1].enabled
&&
323 key
->stencil
[1].writemask
))))
324 depth_mode
&= ~(LATE_DEPTH_WRITE
| EARLY_DEPTH_WRITE
);
331 stencil_refs
[0] = lp_jit_context_stencil_ref_front_value(gallivm
, context_ptr
);
332 stencil_refs
[1] = lp_jit_context_stencil_ref_back_value(gallivm
, context_ptr
);
334 vec_type
= lp_build_vec_type(gallivm
, type
);
336 consts_ptr
= lp_jit_context_constants(gallivm
, context_ptr
);
338 lp_build_for_loop_begin(&loop_state
, gallivm
,
339 lp_build_const_int32(gallivm
, 0),
342 lp_build_const_int32(gallivm
, 1));
344 mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
345 &loop_state
.counter
, 1, "mask_ptr");
346 mask_val
= LLVMBuildLoad(builder
, mask_ptr
, "");
348 memset(outputs
, 0, sizeof outputs
);
350 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
351 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
352 out_color
[cbuf
][chan
] = lp_build_array_alloca(gallivm
,
353 lp_build_vec_type(gallivm
,
358 if (dual_source_blend
) {
359 assert(key
->nr_cbufs
<= 1);
360 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
361 out_color
[1][chan
] = lp_build_array_alloca(gallivm
,
362 lp_build_vec_type(gallivm
,
369 /* 'mask' will control execution based on quad's pixel alive/killed state */
370 lp_build_mask_begin(&mask
, gallivm
, type
, mask_val
);
372 if (!(depth_mode
& EARLY_DEPTH_TEST
) && !simple_shader
)
373 lp_build_mask_check(&mask
);
375 lp_build_interp_soa_update_pos_dyn(interp
, gallivm
, loop_state
.counter
);
378 if (depth_mode
& EARLY_DEPTH_TEST
) {
379 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
380 zs_format_desc
, key
->resource_1d
,
381 depth_ptr
, depth_stride
,
382 &z_fb
, &s_fb
, loop_state
.counter
);
383 lp_build_depth_stencil_test(gallivm
,
395 if (depth_mode
& EARLY_DEPTH_WRITE
) {
396 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
397 zs_format_desc
, key
->resource_1d
,
398 NULL
, NULL
, NULL
, loop_state
.counter
,
399 depth_ptr
, depth_stride
,
403 * Note mask check if stencil is enabled must be after ds write not after
404 * stencil test otherwise new stencil values may not get written if all
405 * fragments got killed by depth/stencil test.
407 if (!simple_shader
&& key
->stencil
[0].enabled
)
408 lp_build_mask_check(&mask
);
411 lp_build_interp_soa_update_inputs_dyn(interp
, gallivm
, loop_state
.counter
);
413 /* Build the actual shader */
414 lp_build_tgsi_soa(gallivm
, tokens
, type
, &mask
,
415 consts_ptr
, &system_values
,
417 outputs
, sampler
, &shader
->info
.base
, NULL
);
420 if (key
->alpha
.enabled
) {
421 int color0
= find_output_by_semantic(&shader
->info
.base
,
425 if (color0
!= -1 && outputs
[color0
][3]) {
426 const struct util_format_description
*cbuf_format_desc
;
427 LLVMValueRef alpha
= LLVMBuildLoad(builder
, outputs
[color0
][3], "alpha");
428 LLVMValueRef alpha_ref_value
;
430 alpha_ref_value
= lp_jit_context_alpha_ref_value(gallivm
, context_ptr
);
431 alpha_ref_value
= lp_build_broadcast(gallivm
, vec_type
, alpha_ref_value
);
433 cbuf_format_desc
= util_format_description(key
->cbuf_format
[0]);
435 lp_build_alpha_test(gallivm
, key
->alpha
.func
, type
, cbuf_format_desc
,
436 &mask
, alpha
, alpha_ref_value
,
437 (depth_mode
& LATE_DEPTH_TEST
) != 0);
442 if (depth_mode
& LATE_DEPTH_TEST
) {
443 int pos0
= find_output_by_semantic(&shader
->info
.base
,
444 TGSI_SEMANTIC_POSITION
,
447 if (pos0
!= -1 && outputs
[pos0
][2]) {
448 z
= LLVMBuildLoad(builder
, outputs
[pos0
][2], "output.z");
451 * Clamp according to ARB_depth_clamp semantics.
453 if (key
->depth_clamp
) {
454 LLVMValueRef viewport
, min_depth
, max_depth
;
455 LLVMValueRef viewport_index
;
456 struct lp_build_context f32_bld
;
458 assert(type
.floating
);
459 lp_build_context_init(&f32_bld
, gallivm
, type
);
462 * Assumes clamping of the viewport index will occur in setup/gs. Value
463 * is passed through the rasterization stage via lp_rast_shader_inputs.
465 * See: draw_clamp_viewport_idx and lp_clamp_viewport_idx for clamping
468 viewport_index
= lp_jit_thread_data_raster_state_viewport_index(gallivm
,
472 * Load the min and max depth from the lp_jit_context.viewports
473 * array of lp_jit_viewport structures.
475 viewport
= lp_llvm_viewport(context_ptr
, gallivm
, viewport_index
);
477 /* viewports[viewport_index].min_depth */
478 min_depth
= LLVMBuildExtractElement(builder
, viewport
,
479 lp_build_const_int32(gallivm
, LP_JIT_VIEWPORT_MIN_DEPTH
),
481 min_depth
= lp_build_broadcast_scalar(&f32_bld
, min_depth
);
483 /* viewports[viewport_index].max_depth */
484 max_depth
= LLVMBuildExtractElement(builder
, viewport
,
485 lp_build_const_int32(gallivm
, LP_JIT_VIEWPORT_MAX_DEPTH
),
487 max_depth
= lp_build_broadcast_scalar(&f32_bld
, max_depth
);
490 * Clamp to the min and max depth values for the given viewport.
492 z
= lp_build_clamp(&f32_bld
, z
, min_depth
, max_depth
);
496 lp_build_depth_stencil_load_swizzled(gallivm
, type
,
497 zs_format_desc
, key
->resource_1d
,
498 depth_ptr
, depth_stride
,
499 &z_fb
, &s_fb
, loop_state
.counter
);
501 lp_build_depth_stencil_test(gallivm
,
513 if (depth_mode
& LATE_DEPTH_WRITE
) {
514 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
515 zs_format_desc
, key
->resource_1d
,
516 NULL
, NULL
, NULL
, loop_state
.counter
,
517 depth_ptr
, depth_stride
,
521 else if ((depth_mode
& EARLY_DEPTH_TEST
) &&
522 (depth_mode
& LATE_DEPTH_WRITE
))
524 /* Need to apply a reduced mask to the depth write. Reload the
525 * depth value, update from zs_value with the new mask value and
528 lp_build_depth_stencil_write_swizzled(gallivm
, type
,
529 zs_format_desc
, key
->resource_1d
,
530 &mask
, z_fb
, s_fb
, loop_state
.counter
,
531 depth_ptr
, depth_stride
,
537 for (attrib
= 0; attrib
< shader
->info
.base
.num_outputs
; ++attrib
)
539 unsigned cbuf
= shader
->info
.base
.output_semantic_index
[attrib
];
540 if ((shader
->info
.base
.output_semantic_name
[attrib
] == TGSI_SEMANTIC_COLOR
) &&
541 ((cbuf
< key
->nr_cbufs
) || (cbuf
== 1 && dual_source_blend
)))
543 for(chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
544 if(outputs
[attrib
][chan
]) {
545 /* XXX: just initialize outputs to point at colors[] and
548 LLVMValueRef out
= LLVMBuildLoad(builder
, outputs
[attrib
][chan
], "");
549 LLVMValueRef color_ptr
;
550 color_ptr
= LLVMBuildGEP(builder
, out_color
[cbuf
][chan
],
551 &loop_state
.counter
, 1, "");
552 lp_build_name(out
, "color%u.%c", attrib
, "rgba"[chan
]);
553 LLVMBuildStore(builder
, out
, color_ptr
);
559 if (key
->occlusion_count
) {
560 LLVMValueRef counter
= lp_jit_thread_data_counter(gallivm
, thread_data_ptr
);
561 lp_build_name(counter
, "counter");
562 lp_build_occlusion_count(gallivm
, type
,
563 lp_build_mask_value(&mask
), counter
);
566 mask_val
= lp_build_mask_end(&mask
);
567 LLVMBuildStore(builder
, mask_val
, mask_ptr
);
568 lp_build_for_loop_end(&loop_state
);
573 * This function will reorder pixels from the fragment shader SoA to memory layout AoS
575 * Fragment Shader outputs pixels in small 2x2 blocks
576 * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
578 * However in memory pixels are stored in rows
579 * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
581 * @param type fragment shader type (4x or 8x float)
582 * @param num_fs number of fs_src
583 * @param is_1d whether we're outputting to a 1d resource
584 * @param dst_channels number of output channels
585 * @param fs_src output from fragment shader
586 * @param dst pointer to store result
587 * @param pad_inline is channel padding inline or at end of row
588 * @return the number of dsts
591 generate_fs_twiddle(struct gallivm_state
*gallivm
,
594 unsigned dst_channels
,
595 LLVMValueRef fs_src
[][4],
599 LLVMValueRef src
[16];
605 unsigned pixels
= type
.length
/ 4;
606 unsigned reorder_group
;
607 unsigned src_channels
;
611 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
612 src_count
= num_fs
* src_channels
;
614 assert(pixels
== 2 || pixels
== 1);
615 assert(num_fs
* src_channels
<= Elements(src
));
618 * Transpose from SoA -> AoS
620 for (i
= 0; i
< num_fs
; ++i
) {
621 lp_build_transpose_aos_n(gallivm
, type
, &fs_src
[i
][0], src_channels
, &src
[i
* src_channels
]);
625 * Pick transformation options
632 if (dst_channels
== 1) {
638 } else if (dst_channels
== 2) {
642 } else if (dst_channels
> 2) {
649 if (!pad_inline
&& dst_channels
== 3 && pixels
> 1) {
655 * Split the src in half
658 for (i
= num_fs
; i
> 0; --i
) {
659 src
[(i
- 1)*2 + 1] = lp_build_extract_range(gallivm
, src
[i
- 1], 4, 4);
660 src
[(i
- 1)*2 + 0] = lp_build_extract_range(gallivm
, src
[i
- 1], 0, 4);
668 * Ensure pixels are in memory order
671 /* Twiddle pixels by reordering the array, e.g.:
673 * src_count = 8 -> 0 2 1 3 4 6 5 7
674 * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
676 const unsigned reorder_sw
[] = { 0, 2, 1, 3 };
678 for (i
= 0; i
< src_count
; ++i
) {
679 unsigned group
= i
/ reorder_group
;
680 unsigned block
= (group
/ 4) * 4 * reorder_group
;
681 unsigned j
= block
+ (reorder_sw
[group
% 4] * reorder_group
) + (i
% reorder_group
);
684 } else if (twiddle
) {
685 /* Twiddle pixels across elements of array */
686 lp_bld_quad_twiddle(gallivm
, type
, src
, src_count
, dst
);
689 memcpy(dst
, src
, sizeof(LLVMValueRef
) * src_count
);
693 * Moves any padding between pixels to the end
694 * e.g. RGBXRGBX -> RGBRGBXX
697 unsigned char swizzles
[16];
698 unsigned elems
= pixels
* dst_channels
;
700 for (i
= 0; i
< type
.length
; ++i
) {
702 swizzles
[i
] = i
% dst_channels
+ (i
/ dst_channels
) * 4;
704 swizzles
[i
] = LP_BLD_SWIZZLE_DONTCARE
;
707 for (i
= 0; i
< src_count
; ++i
) {
708 dst
[i
] = lp_build_swizzle_aos_n(gallivm
, dst
[i
], swizzles
, type
.length
, type
.length
);
717 * Load an unswizzled block of pixels from memory
720 load_unswizzled_block(struct gallivm_state
*gallivm
,
721 LLVMValueRef base_ptr
,
723 unsigned block_width
,
724 unsigned block_height
,
726 struct lp_type dst_type
,
728 unsigned dst_alignment
)
730 LLVMBuilderRef builder
= gallivm
->builder
;
731 unsigned row_size
= dst_count
/ block_height
;
734 /* Ensure block exactly fits into dst */
735 assert((block_width
* block_height
) % dst_count
== 0);
737 for (i
= 0; i
< dst_count
; ++i
) {
738 unsigned x
= i
% row_size
;
739 unsigned y
= i
/ row_size
;
741 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (dst_type
.width
/ 8) * dst_type
.length
);
742 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
745 LLVMValueRef dst_ptr
;
747 gep
[0] = lp_build_const_int32(gallivm
, 0);
748 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
750 dst_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
751 dst_ptr
= LLVMBuildBitCast(builder
, dst_ptr
, LLVMPointerType(lp_build_vec_type(gallivm
, dst_type
), 0), "");
753 dst
[i
] = LLVMBuildLoad(builder
, dst_ptr
, "");
755 lp_set_load_alignment(dst
[i
], dst_alignment
);
761 * Store an unswizzled block of pixels to memory
764 store_unswizzled_block(struct gallivm_state
*gallivm
,
765 LLVMValueRef base_ptr
,
767 unsigned block_width
,
768 unsigned block_height
,
770 struct lp_type src_type
,
772 unsigned src_alignment
)
774 LLVMBuilderRef builder
= gallivm
->builder
;
775 unsigned row_size
= src_count
/ block_height
;
778 /* Ensure src exactly fits into block */
779 assert((block_width
* block_height
) % src_count
== 0);
781 for (i
= 0; i
< src_count
; ++i
) {
782 unsigned x
= i
% row_size
;
783 unsigned y
= i
/ row_size
;
785 LLVMValueRef bx
= lp_build_const_int32(gallivm
, x
* (src_type
.width
/ 8) * src_type
.length
);
786 LLVMValueRef by
= LLVMBuildMul(builder
, lp_build_const_int32(gallivm
, y
), stride
, "");
789 LLVMValueRef src_ptr
;
791 gep
[0] = lp_build_const_int32(gallivm
, 0);
792 gep
[1] = LLVMBuildAdd(builder
, bx
, by
, "");
794 src_ptr
= LLVMBuildGEP(builder
, base_ptr
, gep
, 2, "");
795 src_ptr
= LLVMBuildBitCast(builder
, src_ptr
, LLVMPointerType(lp_build_vec_type(gallivm
, src_type
), 0), "");
797 src_ptr
= LLVMBuildStore(builder
, src
[i
], src_ptr
);
799 lp_set_store_alignment(src_ptr
, src_alignment
);
805 * Checks if a format description is an arithmetic format
807 * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
809 static INLINE boolean
810 is_arithmetic_format(const struct util_format_description
*format_desc
)
812 boolean arith
= false;
815 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
816 arith
|= format_desc
->channel
[i
].size
!= format_desc
->channel
[0].size
;
817 arith
|= (format_desc
->channel
[i
].size
% 8) != 0;
825 * Checks if this format requires special handling due to required expansion
826 * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
829 static INLINE boolean
830 format_expands_to_float_soa(const struct util_format_description
*format_desc
)
832 if (format_desc
->format
== PIPE_FORMAT_R11G11B10_FLOAT
||
833 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
841 * Retrieves the type representing the memory layout for a format
843 * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
846 lp_mem_type_from_format_desc(const struct util_format_description
*format_desc
,
847 struct lp_type
* type
)
852 if (format_expands_to_float_soa(format_desc
)) {
853 /* just make this a 32bit uint */
854 type
->floating
= false;
863 for (i
= 0; i
< 4; i
++)
864 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
868 memset(type
, 0, sizeof(struct lp_type
));
869 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
870 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
871 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
872 type
->norm
= format_desc
->channel
[chan
].normalized
;
874 if (is_arithmetic_format(format_desc
)) {
878 for (i
= 0; i
< format_desc
->nr_channels
; ++i
) {
879 type
->width
+= format_desc
->channel
[i
].size
;
882 type
->width
= format_desc
->channel
[chan
].size
;
883 type
->length
= format_desc
->nr_channels
;
889 * Retrieves the type for a format which is usable in the blending code.
891 * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
894 lp_blend_type_from_format_desc(const struct util_format_description
*format_desc
,
895 struct lp_type
* type
)
900 if (format_expands_to_float_soa(format_desc
)) {
901 /* always use ordinary floats for blending */
902 type
->floating
= true;
911 for (i
= 0; i
< 4; i
++)
912 if (format_desc
->channel
[i
].type
!= UTIL_FORMAT_TYPE_VOID
)
916 memset(type
, 0, sizeof(struct lp_type
));
917 type
->floating
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FLOAT
;
918 type
->fixed
= format_desc
->channel
[chan
].type
== UTIL_FORMAT_TYPE_FIXED
;
919 type
->sign
= format_desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_UNSIGNED
;
920 type
->norm
= format_desc
->channel
[chan
].normalized
;
921 type
->width
= format_desc
->channel
[chan
].size
;
922 type
->length
= format_desc
->nr_channels
;
924 for (i
= 1; i
< format_desc
->nr_channels
; ++i
) {
925 if (format_desc
->channel
[i
].size
> type
->width
)
926 type
->width
= format_desc
->channel
[i
].size
;
929 if (type
->floating
) {
932 if (type
->width
<= 8) {
934 } else if (type
->width
<= 16) {
941 if (is_arithmetic_format(format_desc
) && type
->length
== 3) {
948 * Scale a normalized value from src_bits to dst_bits.
950 * The exact calculation is
952 * dst = iround(src * dst_mask / src_mask)
954 * or with integer rounding
956 * dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
960 * src_mask = (1 << src_bits) - 1
961 * dst_mask = (1 << dst_bits) - 1
963 * but we try to avoid division and multiplication through shifts.
965 static INLINE LLVMValueRef
966 scale_bits(struct gallivm_state
*gallivm
,
970 struct lp_type src_type
)
972 LLVMBuilderRef builder
= gallivm
->builder
;
973 LLVMValueRef result
= src
;
975 if (dst_bits
< src_bits
) {
976 int delta_bits
= src_bits
- dst_bits
;
978 if (delta_bits
<= dst_bits
) {
980 * Approximate the rescaling with a single shift.
982 * This gives the wrong rounding.
985 result
= LLVMBuildLShr(builder
,
987 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
992 * Try more accurate rescaling.
996 * Drop the least significant bits to make space for the multiplication.
998 * XXX: A better approach would be to use a wider integer type as intermediate. But
999 * this is enough to convert alpha from 16bits -> 2 when rendering to
1000 * PIPE_FORMAT_R10G10B10A2_UNORM.
1002 result
= LLVMBuildLShr(builder
,
1004 lp_build_const_int_vec(gallivm
, src_type
, dst_bits
),
1008 result
= LLVMBuildMul(builder
,
1010 lp_build_const_int_vec(gallivm
, src_type
, (1LL << dst_bits
) - 1),
1014 * Add a rounding term before the division.
1016 * TODO: Handle signed integers too.
1018 if (!src_type
.sign
) {
1019 result
= LLVMBuildAdd(builder
,
1021 lp_build_const_int_vec(gallivm
, src_type
, (1LL << (delta_bits
- 1))),
1026 * Approximate the division by src_mask with a src_bits shift.
1028 * Given the src has already been shifted by dst_bits, all we need
1029 * to do is to shift by the difference.
1032 result
= LLVMBuildLShr(builder
,
1034 lp_build_const_int_vec(gallivm
, src_type
, delta_bits
),
1038 } else if (dst_bits
> src_bits
) {
1040 int db
= dst_bits
- src_bits
;
1042 /* Shift left by difference in bits */
1043 result
= LLVMBuildShl(builder
,
1045 lp_build_const_int_vec(gallivm
, src_type
, db
),
1048 if (db
< src_bits
) {
1049 /* Enough bits in src to fill the remainder */
1050 LLVMValueRef lower
= LLVMBuildLShr(builder
,
1052 lp_build_const_int_vec(gallivm
, src_type
, src_bits
- db
),
1055 result
= LLVMBuildOr(builder
, result
, lower
, "");
1056 } else if (db
> src_bits
) {
1057 /* Need to repeatedly copy src bits to fill remainder in dst */
1060 for (n
= src_bits
; n
< dst_bits
; n
*= 2) {
1061 LLVMValueRef shuv
= lp_build_const_int_vec(gallivm
, src_type
, n
);
1063 result
= LLVMBuildOr(builder
,
1065 LLVMBuildLShr(builder
, result
, shuv
, ""),
1076 * Convert from memory format to blending format
1078 * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
1081 convert_to_blend_type(struct gallivm_state
*gallivm
,
1082 unsigned block_size
,
1083 const struct util_format_description
*src_fmt
,
1084 struct lp_type src_type
,
1085 struct lp_type dst_type
,
1086 LLVMValueRef
* src
, // and dst
1089 LLVMValueRef
*dst
= src
;
1090 LLVMBuilderRef builder
= gallivm
->builder
;
1091 struct lp_type blend_type
;
1092 struct lp_type mem_type
;
1094 unsigned pixels
= block_size
/ num_srcs
;
1098 * full custom path for packed floats and srgb formats - none of the later
1099 * functions would do anything useful, and given the lp_type representation they
1100 * can't be fixed. Should really have some SoA blend path for these kind of
1101 * formats rather than hacking them in here.
1103 if (format_expands_to_float_soa(src_fmt
)) {
1104 LLVMValueRef tmpsrc
[4];
1106 * This is pretty suboptimal for this case blending in SoA would be much
1107 * better, since conversion gets us SoA values so need to convert back.
1109 assert(src_type
.width
== 32);
1110 assert(dst_type
.floating
);
1111 assert(dst_type
.width
== 32);
1112 assert(dst_type
.length
% 4 == 0);
1113 assert(num_srcs
% 4 == 0);
1115 for (i
= 0; i
< 4; i
++) {
1118 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1119 LLVMValueRef tmpsoa
[4];
1120 LLVMValueRef tmps
= tmpsrc
[i
];
1121 if (dst_type
.length
== 8) {
1122 LLVMValueRef shuffles
[8];
1124 /* fetch was 4 values but need 8-wide output values */
1125 tmps
= lp_build_concat(gallivm
, &tmpsrc
[i
* 2], src_type
, 2);
1127 * for 8-wide aos transpose would give us wrong order not matching
1128 * incoming converted fs values and mask. ARGH.
1130 for (j
= 0; j
< 4; j
++) {
1131 shuffles
[j
] = lp_build_const_int32(gallivm
, j
* 2);
1132 shuffles
[j
+ 4] = lp_build_const_int32(gallivm
, j
* 2 + 1);
1134 tmps
= LLVMBuildShuffleVector(builder
, tmps
, tmps
,
1135 LLVMConstVector(shuffles
, 8), "");
1137 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1138 lp_build_r11g11b10_to_float(gallivm
, tmps
, tmpsoa
);
1141 lp_build_unpack_rgba_soa(gallivm
, src_fmt
, dst_type
, tmps
, tmpsoa
);
1143 lp_build_transpose_aos(gallivm
, dst_type
, tmpsoa
, &src
[i
* 4]);
1148 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1149 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1151 /* Is the format arithmetic */
1152 is_arith
= blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
;
1153 is_arith
&= !(mem_type
.width
== 16 && mem_type
.floating
);
1155 /* Pad if necessary */
1156 if (!is_arith
&& src_type
.length
< dst_type
.length
) {
1157 for (i
= 0; i
< num_srcs
; ++i
) {
1158 dst
[i
] = lp_build_pad_vector(gallivm
, src
[i
], dst_type
.length
);
1161 src_type
.length
= dst_type
.length
;
1164 /* Special case for half-floats */
1165 if (mem_type
.width
== 16 && mem_type
.floating
) {
1166 assert(blend_type
.width
== 32 && blend_type
.floating
);
1167 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1175 src_type
.width
= blend_type
.width
* blend_type
.length
;
1176 blend_type
.length
*= pixels
;
1177 src_type
.length
*= pixels
/ (src_type
.length
/ mem_type
.length
);
1179 for (i
= 0; i
< num_srcs
; ++i
) {
1180 LLVMValueRef chans
[4];
1181 LLVMValueRef res
= NULL
;
1183 dst
[i
] = LLVMBuildZExt(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1185 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1187 unsigned sa
= src_fmt
->channel
[j
].shift
;
1188 #ifdef PIPE_ARCH_LITTLE_ENDIAN
1189 unsigned from_lsb
= j
;
1191 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1194 for (k
= 0; k
< src_fmt
->channel
[j
].size
; ++k
) {
1198 /* Extract bits from source */
1199 chans
[j
] = LLVMBuildLShr(builder
,
1201 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1204 chans
[j
] = LLVMBuildAnd(builder
,
1206 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1210 if (src_type
.norm
) {
1211 chans
[j
] = scale_bits(gallivm
, src_fmt
->channel
[j
].size
,
1212 blend_type
.width
, chans
[j
], src_type
);
1215 /* Insert bits into correct position */
1216 chans
[j
] = LLVMBuildShl(builder
,
1218 lp_build_const_int_vec(gallivm
, src_type
, from_lsb
* blend_type
.width
),
1224 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1228 dst
[i
] = LLVMBuildBitCast(builder
, res
, lp_build_vec_type(gallivm
, blend_type
), "");
1234 * Convert from blending format to memory format
1236 * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
1239 convert_from_blend_type(struct gallivm_state
*gallivm
,
1240 unsigned block_size
,
1241 const struct util_format_description
*src_fmt
,
1242 struct lp_type src_type
,
1243 struct lp_type dst_type
,
1244 LLVMValueRef
* src
, // and dst
1247 LLVMValueRef
* dst
= src
;
1249 struct lp_type mem_type
;
1250 struct lp_type blend_type
;
1251 LLVMBuilderRef builder
= gallivm
->builder
;
1252 unsigned pixels
= block_size
/ num_srcs
;
1256 * full custom path for packed floats and srgb formats - none of the later
1257 * functions would do anything useful, and given the lp_type representation they
1258 * can't be fixed. Should really have some SoA blend path for these kind of
1259 * formats rather than hacking them in here.
1261 if (format_expands_to_float_soa(src_fmt
)) {
1263 * This is pretty suboptimal for this case blending in SoA would be much
1264 * better - we need to transpose the AoS values back to SoA values for
1265 * conversion/packing.
1267 assert(src_type
.floating
);
1268 assert(src_type
.width
== 32);
1269 assert(src_type
.length
% 4 == 0);
1270 assert(dst_type
.width
== 32);
1272 for (i
= 0; i
< num_srcs
/ 4; i
++) {
1273 LLVMValueRef tmpsoa
[4], tmpdst
;
1274 lp_build_transpose_aos(gallivm
, src_type
, &src
[i
* 4], tmpsoa
);
1275 /* really really need SoA here */
1277 if (src_fmt
->format
== PIPE_FORMAT_R11G11B10_FLOAT
) {
1278 tmpdst
= lp_build_float_to_r11g11b10(gallivm
, tmpsoa
);
1281 tmpdst
= lp_build_float_to_srgb_packed(gallivm
, src_fmt
,
1285 if (src_type
.length
== 8) {
1286 LLVMValueRef tmpaos
, shuffles
[8];
1289 * for 8-wide aos transpose has given us wrong order not matching
1290 * output order. HMPF. Also need to split the output values manually.
1292 for (j
= 0; j
< 4; j
++) {
1293 shuffles
[j
* 2] = lp_build_const_int32(gallivm
, j
);
1294 shuffles
[j
* 2 + 1] = lp_build_const_int32(gallivm
, j
+ 4);
1296 tmpaos
= LLVMBuildShuffleVector(builder
, tmpdst
, tmpdst
,
1297 LLVMConstVector(shuffles
, 8), "");
1298 src
[i
* 2] = lp_build_extract_range(gallivm
, tmpaos
, 0, 4);
1299 src
[i
* 2 + 1] = lp_build_extract_range(gallivm
, tmpaos
, 4, 4);
1308 lp_mem_type_from_format_desc(src_fmt
, &mem_type
);
1309 lp_blend_type_from_format_desc(src_fmt
, &blend_type
);
1311 is_arith
= (blend_type
.length
* blend_type
.width
!= mem_type
.width
* mem_type
.length
);
1313 /* Special case for half-floats */
1314 if (mem_type
.width
== 16 && mem_type
.floating
) {
1315 int length
= dst_type
.length
;
1316 assert(blend_type
.width
== 32 && blend_type
.floating
);
1318 dst_type
.length
= src_type
.length
;
1320 lp_build_conv_auto(gallivm
, src_type
, &dst_type
, dst
, num_srcs
, dst
);
1322 dst_type
.length
= length
;
1326 /* Remove any padding */
1327 if (!is_arith
&& (src_type
.length
% mem_type
.length
)) {
1328 src_type
.length
-= (src_type
.length
% mem_type
.length
);
1330 for (i
= 0; i
< num_srcs
; ++i
) {
1331 dst
[i
] = lp_build_extract_range(gallivm
, dst
[i
], 0, src_type
.length
);
1335 /* No bit arithmetic to do */
1340 src_type
.length
= pixels
;
1341 src_type
.width
= blend_type
.length
* blend_type
.width
;
1342 dst_type
.length
= pixels
;
1344 for (i
= 0; i
< num_srcs
; ++i
) {
1345 LLVMValueRef chans
[4];
1346 LLVMValueRef res
= NULL
;
1348 dst
[i
] = LLVMBuildBitCast(builder
, src
[i
], lp_build_vec_type(gallivm
, src_type
), "");
1350 for (j
= 0; j
< src_fmt
->nr_channels
; ++j
) {
1352 unsigned sa
= src_fmt
->channel
[j
].shift
;
1353 #ifdef PIPE_ARCH_LITTLE_ENDIAN
1354 unsigned from_lsb
= j
;
1356 unsigned from_lsb
= src_fmt
->nr_channels
- j
- 1;
1359 assert(blend_type
.width
> src_fmt
->channel
[j
].size
);
1361 for (k
= 0; k
< blend_type
.width
; ++k
) {
1366 chans
[j
] = LLVMBuildLShr(builder
,
1368 lp_build_const_int_vec(gallivm
, src_type
, from_lsb
* blend_type
.width
),
1371 chans
[j
] = LLVMBuildAnd(builder
,
1373 lp_build_const_int_vec(gallivm
, src_type
, mask
),
1376 /* Scale down bits */
1377 if (src_type
.norm
) {
1378 chans
[j
] = scale_bits(gallivm
, blend_type
.width
,
1379 src_fmt
->channel
[j
].size
, chans
[j
], src_type
);
1383 chans
[j
] = LLVMBuildShl(builder
,
1385 lp_build_const_int_vec(gallivm
, src_type
, sa
),
1388 sa
+= src_fmt
->channel
[j
].size
;
1393 res
= LLVMBuildOr(builder
, res
, chans
[j
], "");
1397 assert (dst_type
.width
!= 24);
1399 dst
[i
] = LLVMBuildTrunc(builder
, res
, lp_build_vec_type(gallivm
, dst_type
), "");
1405 * Convert alpha to same blend type as src
1408 convert_alpha(struct gallivm_state
*gallivm
,
1409 struct lp_type row_type
,
1410 struct lp_type alpha_type
,
1411 const unsigned block_size
,
1412 const unsigned block_height
,
1413 const unsigned src_count
,
1414 const unsigned dst_channels
,
1415 const bool pad_inline
,
1416 LLVMValueRef
* src_alpha
)
1418 LLVMBuilderRef builder
= gallivm
->builder
;
1420 unsigned length
= row_type
.length
;
1421 row_type
.length
= alpha_type
.length
;
1423 /* Twiddle the alpha to match pixels */
1424 lp_bld_quad_twiddle(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
);
1427 * TODO this should use single lp_build_conv call for
1428 * src_count == 1 && dst_channels == 1 case (dropping the concat below)
1430 for (i
= 0; i
< block_height
; ++i
) {
1431 lp_build_conv(gallivm
, alpha_type
, row_type
, &src_alpha
[i
], 1, &src_alpha
[i
], 1);
1434 alpha_type
= row_type
;
1435 row_type
.length
= length
;
1437 /* If only one channel we can only need the single alpha value per pixel */
1438 if (src_count
== 1 && dst_channels
== 1) {
1440 lp_build_concat_n(gallivm
, alpha_type
, src_alpha
, block_height
, src_alpha
, src_count
);
1442 /* If there are more srcs than rows then we need to split alpha up */
1443 if (src_count
> block_height
) {
1444 for (i
= src_count
; i
> 0; --i
) {
1445 unsigned pixels
= block_size
/ src_count
;
1446 unsigned idx
= i
- 1;
1448 src_alpha
[idx
] = lp_build_extract_range(gallivm
, src_alpha
[(idx
* pixels
) / 4],
1449 (idx
* pixels
) % 4, pixels
);
1453 /* If there is a src for each pixel broadcast the alpha across whole row */
1454 if (src_count
== block_size
) {
1455 for (i
= 0; i
< src_count
; ++i
) {
1456 src_alpha
[i
] = lp_build_broadcast(gallivm
, lp_build_vec_type(gallivm
, row_type
), src_alpha
[i
]);
1459 unsigned pixels
= block_size
/ src_count
;
1460 unsigned channels
= pad_inline
? TGSI_NUM_CHANNELS
: dst_channels
;
1461 unsigned alpha_span
= 1;
1462 LLVMValueRef shuffles
[LP_MAX_VECTOR_LENGTH
];
1464 /* Check if we need 2 src_alphas for our shuffles */
1465 if (pixels
> alpha_type
.length
) {
1469 /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
1470 for (j
= 0; j
< row_type
.length
; ++j
) {
1471 if (j
< pixels
* channels
) {
1472 shuffles
[j
] = lp_build_const_int32(gallivm
, j
/ channels
);
1474 shuffles
[j
] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
1478 for (i
= 0; i
< src_count
; ++i
) {
1479 unsigned idx1
= i
, idx2
= i
;
1481 if (alpha_span
> 1){
1486 src_alpha
[i
] = LLVMBuildShuffleVector(builder
,
1489 LLVMConstVector(shuffles
, row_type
.length
),
1498 * Generates the blend function for unswizzled colour buffers
1499 * Also generates the read & write from colour buffer
1502 generate_unswizzled_blend(struct gallivm_state
*gallivm
,
1504 struct lp_fragment_shader_variant
*variant
,
1505 enum pipe_format out_format
,
1506 unsigned int num_fs
,
1507 struct lp_type fs_type
,
1508 LLVMValueRef
* fs_mask
,
1509 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][4],
1510 LLVMValueRef context_ptr
,
1511 LLVMValueRef color_ptr
,
1512 LLVMValueRef stride
,
1513 unsigned partial_mask
,
1516 const unsigned alpha_channel
= 3;
1517 const unsigned block_width
= LP_RASTER_BLOCK_SIZE
;
1518 const unsigned block_height
= LP_RASTER_BLOCK_SIZE
;
1519 const unsigned block_size
= block_width
* block_height
;
1520 const unsigned lp_integer_vector_width
= 128;
1522 LLVMBuilderRef builder
= gallivm
->builder
;
1523 LLVMValueRef fs_src
[4][TGSI_NUM_CHANNELS
];
1524 LLVMValueRef fs_src1
[4][TGSI_NUM_CHANNELS
];
1525 LLVMValueRef src_alpha
[4 * 4];
1526 LLVMValueRef src1_alpha
[4 * 4];
1527 LLVMValueRef src_mask
[4 * 4];
1528 LLVMValueRef src
[4 * 4];
1529 LLVMValueRef src1
[4 * 4];
1530 LLVMValueRef dst
[4 * 4];
1531 LLVMValueRef blend_color
;
1532 LLVMValueRef blend_alpha
;
1533 LLVMValueRef i32_zero
;
1534 LLVMValueRef check_mask
;
1535 LLVMValueRef undef_src_val
;
1537 struct lp_build_mask_context mask_ctx
;
1538 struct lp_type mask_type
;
1539 struct lp_type blend_type
;
1540 struct lp_type row_type
;
1541 struct lp_type dst_type
;
1543 unsigned char swizzle
[TGSI_NUM_CHANNELS
];
1544 unsigned vector_width
;
1545 unsigned src_channels
= TGSI_NUM_CHANNELS
;
1546 unsigned dst_channels
;
1551 const struct util_format_description
* out_format_desc
= util_format_description(out_format
);
1553 unsigned dst_alignment
;
1555 bool pad_inline
= is_arithmetic_format(out_format_desc
);
1556 bool has_alpha
= false;
1557 const boolean dual_source_blend
= variant
->key
.blend
.rt
[0].blend_enable
&&
1558 util_blend_state_is_dual(&variant
->key
.blend
, 0);
1560 const boolean is_1d
= variant
->key
.resource_1d
;
1561 unsigned num_fullblock_fs
= is_1d
? 2 * num_fs
: num_fs
;
1562 LLVMValueRef fpstate
= 0;
1564 /* Get type from output format */
1565 lp_blend_type_from_format_desc(out_format_desc
, &row_type
);
1566 lp_mem_type_from_format_desc(out_format_desc
, &dst_type
);
1569 * Technically this code should go into lp_build_smallfloat_to_float
1570 * and lp_build_float_to_smallfloat but due to the
1571 * http://llvm.org/bugs/show_bug.cgi?id=6393
1572 * llvm reorders the mxcsr intrinsics in a way that breaks the code.
1573 * So the ordering is important here and there shouldn't be any
1574 * llvm ir instrunctions in this function before
1575 * this, otherwise half-float format conversions won't work
1576 * (again due to llvm bug #6393).
1578 if (dst_type
.floating
&& dst_type
.width
!= 32) {
1579 /* We need to make sure that denorms are ok for half float
1581 fpstate
= lp_build_fpstate_get(gallivm
);
1582 lp_build_fpstate_set_denorms_zero(gallivm
, FALSE
);
1585 mask_type
= lp_int32_vec4_type();
1586 mask_type
.length
= fs_type
.length
;
1588 for (i
= num_fs
; i
< num_fullblock_fs
; i
++) {
1589 fs_mask
[i
] = lp_build_zero(gallivm
, mask_type
);
1592 /* Do not bother executing code when mask is empty.. */
1594 check_mask
= LLVMConstNull(lp_build_int_vec_type(gallivm
, mask_type
));
1596 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1597 check_mask
= LLVMBuildOr(builder
, check_mask
, fs_mask
[i
], "");
1600 lp_build_mask_begin(&mask_ctx
, gallivm
, mask_type
, check_mask
);
1601 lp_build_mask_check(&mask_ctx
);
1604 partial_mask
|= !variant
->opaque
;
1605 i32_zero
= lp_build_const_int32(gallivm
, 0);
1607 #if HAVE_LLVM < 0x0302
1609 * undef triggers a crash in LLVMBuildTrunc in convert_from_blend_type in some
1610 * cases (seen with r10g10b10a2, 128bit wide vectors) (only used for 1d case).
1612 undef_src_val
= lp_build_zero(gallivm
, fs_type
);
1614 undef_src_val
= lp_build_undef(gallivm
, fs_type
);
1617 row_type
.length
= fs_type
.length
;
1618 vector_width
= dst_type
.floating
? lp_native_vector_width
: lp_integer_vector_width
;
1620 /* Compute correct swizzle and count channels */
1621 memset(swizzle
, LP_BLD_SWIZZLE_DONTCARE
, TGSI_NUM_CHANNELS
);
1624 for (i
= 0; i
< TGSI_NUM_CHANNELS
; ++i
) {
1625 /* Ensure channel is used */
1626 if (out_format_desc
->swizzle
[i
] >= TGSI_NUM_CHANNELS
) {
1630 /* Ensure not already written to (happens in case with GL_ALPHA) */
1631 if (swizzle
[out_format_desc
->swizzle
[i
]] < TGSI_NUM_CHANNELS
) {
1635 /* Ensure we havn't already found all channels */
1636 if (dst_channels
>= out_format_desc
->nr_channels
) {
1640 swizzle
[out_format_desc
->swizzle
[i
]] = i
;
1643 if (i
== alpha_channel
) {
1648 if (format_expands_to_float_soa(out_format_desc
)) {
1650 * the code above can't work for layout_other
1651 * for srgb it would sort of work but we short-circuit swizzles, etc.
1652 * as that is done as part of unpack / pack.
1654 dst_channels
= 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
1660 pad_inline
= true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
1663 /* If 3 channels then pad to include alpha for 4 element transpose */
1664 if (dst_channels
== 3 && !has_alpha
) {
1665 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
1666 if (swizzle
[i
] > TGSI_NUM_CHANNELS
)
1669 if (out_format_desc
->nr_channels
== 4) {
1675 * Load shader output
1677 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1678 /* Always load alpha for use in blending */
1681 alpha
= LLVMBuildLoad(builder
, fs_out_color
[rt
][alpha_channel
][i
], "");
1684 alpha
= undef_src_val
;
1687 /* Load each channel */
1688 for (j
= 0; j
< dst_channels
; ++j
) {
1689 assert(swizzle
[j
] < 4);
1691 fs_src
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[rt
][swizzle
[j
]][i
], "");
1694 fs_src
[i
][j
] = undef_src_val
;
1698 /* If 3 channels then pad to include alpha for 4 element transpose */
1700 * XXX If we include that here maybe could actually use it instead of
1701 * separate alpha for blending?
1703 if (dst_channels
== 3 && !has_alpha
) {
1704 fs_src
[i
][3] = alpha
;
1707 /* We split the row_mask and row_alpha as we want 128bit interleave */
1708 if (fs_type
.length
== 8) {
1709 src_mask
[i
*2 + 0] = lp_build_extract_range(gallivm
, fs_mask
[i
], 0, src_channels
);
1710 src_mask
[i
*2 + 1] = lp_build_extract_range(gallivm
, fs_mask
[i
], src_channels
, src_channels
);
1712 src_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
1713 src_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
, src_channels
, src_channels
);
1715 src_mask
[i
] = fs_mask
[i
];
1716 src_alpha
[i
] = alpha
;
1719 if (dual_source_blend
) {
1720 /* same as above except different src/dst, skip masks and comments... */
1721 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1724 alpha
= LLVMBuildLoad(builder
, fs_out_color
[1][alpha_channel
][i
], "");
1727 alpha
= undef_src_val
;
1730 for (j
= 0; j
< dst_channels
; ++j
) {
1731 assert(swizzle
[j
] < 4);
1733 fs_src1
[i
][j
] = LLVMBuildLoad(builder
, fs_out_color
[1][swizzle
[j
]][i
], "");
1736 fs_src1
[i
][j
] = undef_src_val
;
1739 if (dst_channels
== 3 && !has_alpha
) {
1740 fs_src1
[i
][3] = alpha
;
1742 if (fs_type
.length
== 8) {
1743 src1_alpha
[i
*2 + 0] = lp_build_extract_range(gallivm
, alpha
, 0, src_channels
);
1744 src1_alpha
[i
*2 + 1] = lp_build_extract_range(gallivm
, alpha
, src_channels
, src_channels
);
1746 src1_alpha
[i
] = alpha
;
1751 if (util_format_is_pure_integer(out_format
)) {
1753 * In this case fs_type was really ints or uints disguised as floats,
1756 fs_type
.floating
= 0;
1757 fs_type
.sign
= dst_type
.sign
;
1758 for (i
= 0; i
< num_fullblock_fs
; ++i
) {
1759 for (j
= 0; j
< dst_channels
; ++j
) {
1760 fs_src
[i
][j
] = LLVMBuildBitCast(builder
, fs_src
[i
][j
],
1761 lp_build_vec_type(gallivm
, fs_type
), "");
1763 if (dst_channels
== 3 && !has_alpha
) {
1764 fs_src
[i
][3] = LLVMBuildBitCast(builder
, fs_src
[i
][3],
1765 lp_build_vec_type(gallivm
, fs_type
), "");
1771 * Pixel twiddle from fragment shader order to memory order
1773 src_count
= generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
,
1774 dst_channels
, fs_src
, src
, pad_inline
);
1775 if (dual_source_blend
) {
1776 generate_fs_twiddle(gallivm
, fs_type
, num_fullblock_fs
, dst_channels
,
1777 fs_src1
, src1
, pad_inline
);
1780 src_channels
= dst_channels
< 3 ? dst_channels
: 4;
1781 if (src_count
!= num_fullblock_fs
* src_channels
) {
1782 unsigned ds
= src_count
/ (num_fullblock_fs
* src_channels
);
1783 row_type
.length
/= ds
;
1784 fs_type
.length
= row_type
.length
;
1787 blend_type
= row_type
;
1788 mask_type
.length
= 4;
1790 /* Convert src to row_type */
1791 if (dual_source_blend
) {
1792 struct lp_type old_row_type
= row_type
;
1793 lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
1794 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &old_row_type
, src1
, src_count
, src1
);
1797 src_count
= lp_build_conv_auto(gallivm
, fs_type
, &row_type
, src
, src_count
, src
);
1800 /* If the rows are not an SSE vector, combine them to become SSE size! */
1801 if ((row_type
.width
* row_type
.length
) % 128) {
1802 unsigned bits
= row_type
.width
* row_type
.length
;
1805 assert(src_count
>= (vector_width
/ bits
));
1807 dst_count
= src_count
/ (vector_width
/ bits
);
1809 combined
= lp_build_concat_n(gallivm
, row_type
, src
, src_count
, src
, dst_count
);
1810 if (dual_source_blend
) {
1811 lp_build_concat_n(gallivm
, row_type
, src1
, src_count
, src1
, dst_count
);
1814 row_type
.length
*= combined
;
1815 src_count
/= combined
;
1817 bits
= row_type
.width
* row_type
.length
;
1818 assert(bits
== 128 || bits
== 256);
1823 * Blend Colour conversion
1825 blend_color
= lp_jit_context_f_blend_color(gallivm
, context_ptr
);
1826 blend_color
= LLVMBuildPointerCast(builder
, blend_color
, LLVMPointerType(lp_build_vec_type(gallivm
, fs_type
), 0), "");
1827 blend_color
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, blend_color
, &i32_zero
, 1, ""), "");
1830 lp_build_conv(gallivm
, fs_type
, blend_type
, &blend_color
, 1, &blend_color
, 1);
1832 if (out_format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
) {
1834 * since blending is done with floats, there was no conversion.
1835 * However, the rules according to fixed point renderbuffers still
1836 * apply, that is we must clamp inputs to 0.0/1.0.
1837 * (This would apply to separate alpha conversion too but we currently
1838 * force has_alpha to be true.)
1839 * TODO: should skip this with "fake" blend, since post-blend conversion
1840 * will clamp anyway.
1841 * TODO: could also skip this if fragment color clamping is enabled. We
1842 * don't support it natively so it gets baked into the shader however, so
1843 * can't really tell here.
1845 struct lp_build_context f32_bld
;
1846 assert(row_type
.floating
);
1847 lp_build_context_init(&f32_bld
, gallivm
, row_type
);
1848 for (i
= 0; i
< src_count
; i
++) {
1849 src
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src
[i
]);
1851 if (dual_source_blend
) {
1852 for (i
= 0; i
< src_count
; i
++) {
1853 src1
[i
] = lp_build_clamp_zero_one_nanzero(&f32_bld
, src1
[i
]);
1856 /* probably can't be different than row_type but better safe than sorry... */
1857 lp_build_context_init(&f32_bld
, gallivm
, blend_type
);
1858 blend_color
= lp_build_clamp(&f32_bld
, blend_color
, f32_bld
.zero
, f32_bld
.one
);
1862 blend_alpha
= lp_build_extract_broadcast(gallivm
, blend_type
, row_type
, blend_color
, lp_build_const_int32(gallivm
, 3));
1864 /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
1865 pad_inline
&= (dst_channels
* (block_size
/ src_count
) * row_type
.width
) != vector_width
;
1867 /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
1868 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, TGSI_NUM_CHANNELS
, row_type
.length
);
1870 /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
1871 blend_color
= lp_build_swizzle_aos_n(gallivm
, blend_color
, swizzle
, dst_channels
, row_type
.length
);
1877 lp_bld_quad_twiddle(gallivm
, mask_type
, &src_mask
[0], block_height
, &src_mask
[0]);
1879 if (src_count
< block_height
) {
1880 lp_build_concat_n(gallivm
, mask_type
, src_mask
, 4, src_mask
, src_count
);
1881 } else if (src_count
> block_height
) {
1882 for (i
= src_count
; i
> 0; --i
) {
1883 unsigned pixels
= block_size
/ src_count
;
1884 unsigned idx
= i
- 1;
1886 src_mask
[idx
] = lp_build_extract_range(gallivm
, src_mask
[(idx
* pixels
) / 4],
1887 (idx
* pixels
) % 4, pixels
);
1891 assert(mask_type
.width
== 32);
1893 for (i
= 0; i
< src_count
; ++i
) {
1894 unsigned pixels
= block_size
/ src_count
;
1895 unsigned pixel_width
= row_type
.width
* dst_channels
;
1897 if (pixel_width
== 24) {
1898 mask_type
.width
= 8;
1899 mask_type
.length
= vector_width
/ mask_type
.width
;
1901 mask_type
.length
= pixels
;
1902 mask_type
.width
= row_type
.width
* dst_channels
;
1904 src_mask
[i
] = LLVMBuildIntCast(builder
, src_mask
[i
], lp_build_int_vec_type(gallivm
, mask_type
), "");
1906 mask_type
.length
*= dst_channels
;
1907 mask_type
.width
/= dst_channels
;
1910 src_mask
[i
] = LLVMBuildBitCast(builder
, src_mask
[i
], lp_build_int_vec_type(gallivm
, mask_type
), "");
1911 src_mask
[i
] = lp_build_pad_vector(gallivm
, src_mask
[i
], row_type
.length
);
1918 struct lp_type alpha_type
= fs_type
;
1919 alpha_type
.length
= 4;
1920 convert_alpha(gallivm
, row_type
, alpha_type
,
1921 block_size
, block_height
,
1922 src_count
, dst_channels
,
1923 pad_inline
, src_alpha
);
1924 if (dual_source_blend
) {
1925 convert_alpha(gallivm
, row_type
, alpha_type
,
1926 block_size
, block_height
,
1927 src_count
, dst_channels
,
1928 pad_inline
, src1_alpha
);
1934 * Load dst from memory
1936 if (src_count
< block_height
) {
1937 dst_count
= block_height
;
1939 dst_count
= src_count
;
1942 dst_type
.length
*= block_size
/ dst_count
;
1944 if (format_expands_to_float_soa(out_format_desc
)) {
1946 * we need multiple values at once for the conversion, so can as well
1947 * load them vectorized here too instead of concatenating later.
1948 * (Still need concatenation later for 8-wide vectors).
1950 dst_count
= block_height
;
1951 dst_type
.length
= block_width
;
1955 * Compute the alignment of the destination pointer in bytes
1956 * We fetch 1-4 pixels, if the format has pot alignment then those fetches
1957 * are always aligned by MIN2(16, fetch_width) except for buffers (not
1958 * 1d tex but can't distinguish here) so need to stick with per-pixel
1959 * alignment in this case.
1962 dst_alignment
= (out_format_desc
->block
.bits
+ 7)/(out_format_desc
->block
.width
* 8);
1965 dst_alignment
= dst_type
.length
* dst_type
.width
/ 8;
1967 /* Force power-of-two alignment by extracting only the least-significant-bit */
1968 dst_alignment
= 1 << (ffs(dst_alignment
) - 1);
1970 * Resource base and stride pointers are aligned to 16 bytes, so that's
1971 * the maximum alignment we can guarantee
1973 dst_alignment
= MIN2(16, dst_alignment
);
1976 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
1977 dst
, dst_type
, dst_count
/ 4, dst_alignment
);
1978 for (i
= dst_count
/ 4; i
< dst_count
; i
++) {
1979 dst
[i
] = lp_build_undef(gallivm
, dst_type
);
1984 load_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
1985 dst
, dst_type
, dst_count
, dst_alignment
);
1990 * Convert from dst/output format to src/blending format.
1992 * This is necessary as we can only read 1 row from memory at a time,
1993 * so the minimum dst_count will ever be at this point is 4.
1995 * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
1996 * this will take the 4 dsts and combine them into 1 src so we can perform blending
1997 * on all 16 pixels in that single vector at once.
1999 if (dst_count
> src_count
) {
2000 lp_build_concat_n(gallivm
, dst_type
, dst
, 4, dst
, src_count
);
2006 /* XXX this is broken for RGB8 formats -
2007 * they get expanded from 12 to 16 elements (to include alpha)
2008 * by convert_to_blend_type then reduced to 15 instead of 12
2009 * by convert_from_blend_type (a simple fix though breaks A8...).
2010 * R16G16B16 also crashes differently however something going wrong
2011 * inside llvm handling npot vector sizes seemingly.
2012 * It seems some cleanup could be done here (like skipping conversion/blend
2015 convert_to_blend_type(gallivm
, block_size
, out_format_desc
, dst_type
, row_type
, dst
, src_count
);
2018 * FIXME: Really should get logic ops / masks out of generic blend / row
2019 * format. Logic ops will definitely not work on the blend float format
2020 * used for SRGB here and I think OpenGL expects this to work as expected
2021 * (that is incoming values converted to srgb then logic op applied).
2023 for (i
= 0; i
< src_count
; ++i
) {
2024 dst
[i
] = lp_build_blend_aos(gallivm
,
2025 &variant
->key
.blend
,
2030 has_alpha
? NULL
: src_alpha
[i
],
2032 has_alpha
? NULL
: src1_alpha
[i
],
2034 partial_mask
? src_mask
[i
] : NULL
,
2036 has_alpha
? NULL
: blend_alpha
,
2038 pad_inline
? 4 : dst_channels
);
2041 convert_from_blend_type(gallivm
, block_size
, out_format_desc
, row_type
, dst_type
, dst
, src_count
);
2043 /* Split the blend rows back to memory rows */
2044 if (dst_count
> src_count
) {
2045 row_type
.length
= dst_type
.length
* (dst_count
/ src_count
);
2047 if (src_count
== 1) {
2048 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2049 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2051 row_type
.length
/= 2;
2055 dst
[3] = lp_build_extract_range(gallivm
, dst
[1], row_type
.length
/ 2, row_type
.length
/ 2);
2056 dst
[2] = lp_build_extract_range(gallivm
, dst
[1], 0, row_type
.length
/ 2);
2057 dst
[1] = lp_build_extract_range(gallivm
, dst
[0], row_type
.length
/ 2, row_type
.length
/ 2);
2058 dst
[0] = lp_build_extract_range(gallivm
, dst
[0], 0, row_type
.length
/ 2);
2060 row_type
.length
/= 2;
2065 * Store blend result to memory
2068 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, 1,
2069 dst
, dst_type
, dst_count
/ 4, dst_alignment
);
2072 store_unswizzled_block(gallivm
, color_ptr
, stride
, block_width
, block_height
,
2073 dst
, dst_type
, dst_count
, dst_alignment
);
2076 if (dst_type
.floating
&& dst_type
.width
!= 32) {
2077 lp_build_fpstate_set(gallivm
, fpstate
);
2081 lp_build_mask_end(&mask_ctx
);
2087 * Generate the runtime callable function for the whole fragment pipeline.
2088 * Note that the function which we generate operates on a block of 16
2089 * pixels at at time. The block contains 2x2 quads. Each quad contains
2093 generate_fragment(struct llvmpipe_context
*lp
,
2094 struct lp_fragment_shader
*shader
,
2095 struct lp_fragment_shader_variant
*variant
,
2096 unsigned partial_mask
)
2098 struct gallivm_state
*gallivm
= variant
->gallivm
;
2099 const struct lp_fragment_shader_variant_key
*key
= &variant
->key
;
2100 struct lp_shader_input inputs
[PIPE_MAX_SHADER_INPUTS
];
2101 char func_name
[256];
2102 struct lp_type fs_type
;
2103 struct lp_type blend_type
;
2104 LLVMTypeRef fs_elem_type
;
2105 LLVMTypeRef blend_vec_type
;
2106 LLVMTypeRef arg_types
[13];
2107 LLVMTypeRef func_type
;
2108 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
2109 LLVMTypeRef int8_type
= LLVMInt8TypeInContext(gallivm
->context
);
2110 LLVMValueRef context_ptr
;
2113 LLVMValueRef a0_ptr
;
2114 LLVMValueRef dadx_ptr
;
2115 LLVMValueRef dady_ptr
;
2116 LLVMValueRef color_ptr_ptr
;
2117 LLVMValueRef stride_ptr
;
2118 LLVMValueRef depth_ptr
;
2119 LLVMValueRef depth_stride
;
2120 LLVMValueRef mask_input
;
2121 LLVMValueRef thread_data_ptr
;
2122 LLVMBasicBlockRef block
;
2123 LLVMBuilderRef builder
;
2124 struct lp_build_sampler_soa
*sampler
;
2125 struct lp_build_interp_soa_context interp
;
2126 LLVMValueRef fs_mask
[16 / 4];
2127 LLVMValueRef fs_out_color
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
][16 / 4];
2128 LLVMValueRef function
;
2129 LLVMValueRef facing
;
2134 boolean cbuf0_write_all
;
2135 const boolean dual_source_blend
= key
->blend
.rt
[0].blend_enable
&&
2136 util_blend_state_is_dual(&key
->blend
, 0);
2138 assert(lp_native_vector_width
/ 32 >= 4);
2140 /* Adjust color input interpolation according to flatshade state:
2142 memcpy(inputs
, shader
->inputs
, shader
->info
.base
.num_inputs
* sizeof inputs
[0]);
2143 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
2144 if (inputs
[i
].interp
== LP_INTERP_COLOR
) {
2146 inputs
[i
].interp
= LP_INTERP_CONSTANT
;
2148 inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
2152 /* check if writes to cbuf[0] are to be copied to all cbufs */
2153 cbuf0_write_all
= FALSE
;
2154 for (i
= 0;i
< shader
->info
.base
.num_properties
; i
++) {
2155 if (shader
->info
.base
.properties
[i
].name
==
2156 TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
) {
2157 cbuf0_write_all
= TRUE
;
2162 /* TODO: actually pick these based on the fs and color buffer
2163 * characteristics. */
2165 memset(&fs_type
, 0, sizeof fs_type
);
2166 fs_type
.floating
= TRUE
; /* floating point values */
2167 fs_type
.sign
= TRUE
; /* values are signed */
2168 fs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
2169 fs_type
.width
= 32; /* 32-bit float */
2170 fs_type
.length
= MIN2(lp_native_vector_width
/ 32, 16); /* n*4 elements per vector */
2172 memset(&blend_type
, 0, sizeof blend_type
);
2173 blend_type
.floating
= FALSE
; /* values are integers */
2174 blend_type
.sign
= FALSE
; /* values are unsigned */
2175 blend_type
.norm
= TRUE
; /* values are in [0,1] or [-1,1] */
2176 blend_type
.width
= 8; /* 8-bit ubyte values */
2177 blend_type
.length
= 16; /* 16 elements per vector */
2180 * Generate the function prototype. Any change here must be reflected in
2181 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
2184 fs_elem_type
= lp_build_elem_type(gallivm
, fs_type
);
2186 blend_vec_type
= lp_build_vec_type(gallivm
, blend_type
);
2188 util_snprintf(func_name
, sizeof(func_name
), "fs%u_variant%u_%s",
2189 shader
->no
, variant
->no
, partial_mask
? "partial" : "whole");
2191 arg_types
[0] = variant
->jit_context_ptr_type
; /* context */
2192 arg_types
[1] = int32_type
; /* x */
2193 arg_types
[2] = int32_type
; /* y */
2194 arg_types
[3] = int32_type
; /* facing */
2195 arg_types
[4] = LLVMPointerType(fs_elem_type
, 0); /* a0 */
2196 arg_types
[5] = LLVMPointerType(fs_elem_type
, 0); /* dadx */
2197 arg_types
[6] = LLVMPointerType(fs_elem_type
, 0); /* dady */
2198 arg_types
[7] = LLVMPointerType(LLVMPointerType(blend_vec_type
, 0), 0); /* color */
2199 arg_types
[8] = LLVMPointerType(int8_type
, 0); /* depth */
2200 arg_types
[9] = int32_type
; /* mask_input */
2201 arg_types
[10] = variant
->jit_thread_data_ptr_type
; /* per thread data */
2202 arg_types
[11] = LLVMPointerType(int32_type
, 0); /* stride */
2203 arg_types
[12] = int32_type
; /* depth_stride */
2205 func_type
= LLVMFunctionType(LLVMVoidTypeInContext(gallivm
->context
),
2206 arg_types
, Elements(arg_types
), 0);
2208 function
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
2209 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
2211 variant
->function
[partial_mask
] = function
;
2213 /* XXX: need to propagate noalias down into color param now we are
2214 * passing a pointer-to-pointer?
2216 for(i
= 0; i
< Elements(arg_types
); ++i
)
2217 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
2218 LLVMAddAttribute(LLVMGetParam(function
, i
), LLVMNoAliasAttribute
);
2220 context_ptr
= LLVMGetParam(function
, 0);
2221 x
= LLVMGetParam(function
, 1);
2222 y
= LLVMGetParam(function
, 2);
2223 facing
= LLVMGetParam(function
, 3);
2224 a0_ptr
= LLVMGetParam(function
, 4);
2225 dadx_ptr
= LLVMGetParam(function
, 5);
2226 dady_ptr
= LLVMGetParam(function
, 6);
2227 color_ptr_ptr
= LLVMGetParam(function
, 7);
2228 depth_ptr
= LLVMGetParam(function
, 8);
2229 mask_input
= LLVMGetParam(function
, 9);
2230 thread_data_ptr
= LLVMGetParam(function
, 10);
2231 stride_ptr
= LLVMGetParam(function
, 11);
2232 depth_stride
= LLVMGetParam(function
, 12);
2234 lp_build_name(context_ptr
, "context");
2235 lp_build_name(x
, "x");
2236 lp_build_name(y
, "y");
2237 lp_build_name(a0_ptr
, "a0");
2238 lp_build_name(dadx_ptr
, "dadx");
2239 lp_build_name(dady_ptr
, "dady");
2240 lp_build_name(color_ptr_ptr
, "color_ptr_ptr");
2241 lp_build_name(depth_ptr
, "depth");
2242 lp_build_name(thread_data_ptr
, "thread_data");
2243 lp_build_name(mask_input
, "mask_input");
2244 lp_build_name(stride_ptr
, "stride_ptr");
2245 lp_build_name(depth_stride
, "depth_stride");
2251 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, function
, "entry");
2252 builder
= gallivm
->builder
;
2254 LLVMPositionBuilderAtEnd(builder
, block
);
2256 /* code generated texture sampling */
2257 sampler
= lp_llvm_sampler_soa_create(key
->state
, context_ptr
);
2259 num_fs
= 16 / fs_type
.length
; /* number of loops per 4x4 stamp */
2260 /* for 1d resources only run "upper half" of stamp */
2261 if (key
->resource_1d
)
2265 LLVMValueRef num_loop
= lp_build_const_int32(gallivm
, num_fs
);
2266 LLVMTypeRef mask_type
= lp_build_int_vec_type(gallivm
, fs_type
);
2267 LLVMValueRef mask_store
= lp_build_array_alloca(gallivm
, mask_type
,
2268 num_loop
, "mask_store");
2269 LLVMValueRef color_store
[PIPE_MAX_COLOR_BUFS
][TGSI_NUM_CHANNELS
];
2272 * The shader input interpolation info is not explicitely baked in the
2273 * shader key, but everything it derives from (TGSI, and flatshade) is
2274 * already included in the shader key.
2276 lp_build_interp_soa_init(&interp
,
2278 shader
->info
.base
.num_inputs
,
2280 shader
->info
.base
.pixel_center_integer
,
2282 a0_ptr
, dadx_ptr
, dady_ptr
,
2285 for (i
= 0; i
< num_fs
; i
++) {
2287 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
2288 LLVMValueRef mask_ptr
= LLVMBuildGEP(builder
, mask_store
,
2289 &indexi
, 1, "mask_ptr");
2292 mask
= generate_quad_mask(gallivm
, fs_type
,
2293 i
*fs_type
.length
/4, mask_input
);
2296 mask
= lp_build_const_int_vec(gallivm
, fs_type
, ~0);
2298 LLVMBuildStore(builder
, mask
, mask_ptr
);
2301 generate_fs_loop(gallivm
,
2309 mask_store
, /* output */
2316 for (i
= 0; i
< num_fs
; i
++) {
2317 LLVMValueRef indexi
= lp_build_const_int32(gallivm
, i
);
2318 LLVMValueRef ptr
= LLVMBuildGEP(builder
, mask_store
,
2320 fs_mask
[i
] = LLVMBuildLoad(builder
, ptr
, "mask");
2321 /* This is fucked up need to reorganize things */
2322 for (cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
2323 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
2324 ptr
= LLVMBuildGEP(builder
,
2325 color_store
[cbuf
* !cbuf0_write_all
][chan
],
2327 fs_out_color
[cbuf
][chan
][i
] = ptr
;
2330 if (dual_source_blend
) {
2331 /* only support one dual source blend target hence always use output 1 */
2332 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
2333 ptr
= LLVMBuildGEP(builder
,
2334 color_store
[1][chan
],
2336 fs_out_color
[1][chan
][i
] = ptr
;
2342 sampler
->destroy(sampler
);
2344 /* Loop over color outputs / color buffers to do blending.
2346 for(cbuf
= 0; cbuf
< key
->nr_cbufs
; cbuf
++) {
2347 LLVMValueRef color_ptr
;
2348 LLVMValueRef stride
;
2349 LLVMValueRef index
= lp_build_const_int32(gallivm
, cbuf
);
2351 boolean do_branch
= ((key
->depth
.enabled
2352 || key
->stencil
[0].enabled
2353 || key
->alpha
.enabled
)
2354 && !shader
->info
.base
.uses_kill
);
2356 color_ptr
= LLVMBuildLoad(builder
,
2357 LLVMBuildGEP(builder
, color_ptr_ptr
, &index
, 1, ""),
2360 lp_build_name(color_ptr
, "color_ptr%d", cbuf
);
2362 stride
= LLVMBuildLoad(builder
,
2363 LLVMBuildGEP(builder
, stride_ptr
, &index
, 1, ""),
2366 generate_unswizzled_blend(gallivm
, cbuf
, variant
, key
->cbuf_format
[cbuf
],
2367 num_fs
, fs_type
, fs_mask
, fs_out_color
,
2368 context_ptr
, color_ptr
, stride
, partial_mask
, do_branch
);
2371 LLVMBuildRetVoid(builder
);
2373 gallivm_verify_function(gallivm
, function
);
2375 variant
->nr_instrs
+= lp_build_count_instructions(function
);
2380 dump_fs_variant_key(const struct lp_fragment_shader_variant_key
*key
)
2384 debug_printf("fs variant %p:\n", (void *) key
);
2386 if (key
->flatshade
) {
2387 debug_printf("flatshade = 1\n");
2389 for (i
= 0; i
< key
->nr_cbufs
; ++i
) {
2390 debug_printf("cbuf_format[%u] = %s\n", i
, util_format_name(key
->cbuf_format
[i
]));
2392 if (key
->depth
.enabled
) {
2393 debug_printf("depth.format = %s\n", util_format_name(key
->zsbuf_format
));
2394 debug_printf("depth.func = %s\n", util_dump_func(key
->depth
.func
, TRUE
));
2395 debug_printf("depth.writemask = %u\n", key
->depth
.writemask
);
2398 for (i
= 0; i
< 2; ++i
) {
2399 if (key
->stencil
[i
].enabled
) {
2400 debug_printf("stencil[%u].func = %s\n", i
, util_dump_func(key
->stencil
[i
].func
, TRUE
));
2401 debug_printf("stencil[%u].fail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].fail_op
, TRUE
));
2402 debug_printf("stencil[%u].zpass_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zpass_op
, TRUE
));
2403 debug_printf("stencil[%u].zfail_op = %s\n", i
, util_dump_stencil_op(key
->stencil
[i
].zfail_op
, TRUE
));
2404 debug_printf("stencil[%u].valuemask = 0x%x\n", i
, key
->stencil
[i
].valuemask
);
2405 debug_printf("stencil[%u].writemask = 0x%x\n", i
, key
->stencil
[i
].writemask
);
2409 if (key
->alpha
.enabled
) {
2410 debug_printf("alpha.func = %s\n", util_dump_func(key
->alpha
.func
, TRUE
));
2413 if (key
->occlusion_count
) {
2414 debug_printf("occlusion_count = 1\n");
2417 if (key
->blend
.logicop_enable
) {
2418 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key
->blend
.logicop_func
, TRUE
));
2420 else if (key
->blend
.rt
[0].blend_enable
) {
2421 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].rgb_func
, TRUE
));
2422 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_src_factor
, TRUE
));
2423 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].rgb_dst_factor
, TRUE
));
2424 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key
->blend
.rt
[0].alpha_func
, TRUE
));
2425 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_src_factor
, TRUE
));
2426 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key
->blend
.rt
[0].alpha_dst_factor
, TRUE
));
2428 debug_printf("blend.colormask = 0x%x\n", key
->blend
.rt
[0].colormask
);
2429 for (i
= 0; i
< key
->nr_samplers
; ++i
) {
2430 const struct lp_static_sampler_state
*sampler
= &key
->state
[i
].sampler_state
;
2431 debug_printf("sampler[%u] = \n", i
);
2432 debug_printf(" .wrap = %s %s %s\n",
2433 util_dump_tex_wrap(sampler
->wrap_s
, TRUE
),
2434 util_dump_tex_wrap(sampler
->wrap_t
, TRUE
),
2435 util_dump_tex_wrap(sampler
->wrap_r
, TRUE
));
2436 debug_printf(" .min_img_filter = %s\n",
2437 util_dump_tex_filter(sampler
->min_img_filter
, TRUE
));
2438 debug_printf(" .min_mip_filter = %s\n",
2439 util_dump_tex_mipfilter(sampler
->min_mip_filter
, TRUE
));
2440 debug_printf(" .mag_img_filter = %s\n",
2441 util_dump_tex_filter(sampler
->mag_img_filter
, TRUE
));
2442 if (sampler
->compare_mode
!= PIPE_TEX_COMPARE_NONE
)
2443 debug_printf(" .compare_func = %s\n", util_dump_func(sampler
->compare_func
, TRUE
));
2444 debug_printf(" .normalized_coords = %u\n", sampler
->normalized_coords
);
2445 debug_printf(" .min_max_lod_equal = %u\n", sampler
->min_max_lod_equal
);
2446 debug_printf(" .lod_bias_non_zero = %u\n", sampler
->lod_bias_non_zero
);
2447 debug_printf(" .apply_min_lod = %u\n", sampler
->apply_min_lod
);
2448 debug_printf(" .apply_max_lod = %u\n", sampler
->apply_max_lod
);
2450 for (i
= 0; i
< key
->nr_sampler_views
; ++i
) {
2451 const struct lp_static_texture_state
*texture
= &key
->state
[i
].texture_state
;
2452 debug_printf("texture[%u] = \n", i
);
2453 debug_printf(" .format = %s\n",
2454 util_format_name(texture
->format
));
2455 debug_printf(" .target = %s\n",
2456 util_dump_tex_target(texture
->target
, TRUE
));
2457 debug_printf(" .level_zero_only = %u\n",
2458 texture
->level_zero_only
);
2459 debug_printf(" .pot = %u %u %u\n",
2461 texture
->pot_height
,
2462 texture
->pot_depth
);
2468 lp_debug_fs_variant(const struct lp_fragment_shader_variant
*variant
)
2470 debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
2471 variant
->shader
->no
, variant
->no
);
2472 tgsi_dump(variant
->shader
->base
.tokens
, 0);
2473 dump_fs_variant_key(&variant
->key
);
2474 debug_printf("variant->opaque = %u\n", variant
->opaque
);
2480 * Generate a new fragment shader variant from the shader code and
2481 * other state indicated by the key.
2483 static struct lp_fragment_shader_variant
*
2484 generate_variant(struct llvmpipe_context
*lp
,
2485 struct lp_fragment_shader
*shader
,
2486 const struct lp_fragment_shader_variant_key
*key
)
2488 struct lp_fragment_shader_variant
*variant
;
2489 const struct util_format_description
*cbuf0_format_desc
;
2490 boolean fullcolormask
;
2492 variant
= CALLOC_STRUCT(lp_fragment_shader_variant
);
2496 variant
->gallivm
= gallivm_create();
2497 if (!variant
->gallivm
) {
2502 variant
->shader
= shader
;
2503 variant
->list_item_global
.base
= variant
;
2504 variant
->list_item_local
.base
= variant
;
2505 variant
->no
= shader
->variants_created
++;
2507 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
2510 * Determine whether we are touching all channels in the color buffer.
2512 fullcolormask
= FALSE
;
2513 if (key
->nr_cbufs
== 1) {
2514 cbuf0_format_desc
= util_format_description(key
->cbuf_format
[0]);
2515 fullcolormask
= util_format_colormask_full(cbuf0_format_desc
, key
->blend
.rt
[0].colormask
);
2519 !key
->blend
.logicop_enable
&&
2520 !key
->blend
.rt
[0].blend_enable
&&
2522 !key
->stencil
[0].enabled
&&
2523 !key
->alpha
.enabled
&&
2524 !key
->depth
.enabled
&&
2525 !shader
->info
.base
.uses_kill
2528 if ((shader
->info
.base
.num_tokens
<= 1) &&
2529 !key
->depth
.enabled
&& !key
->stencil
[0].enabled
) {
2530 variant
->ps_inv_multiplier
= 0;
2532 variant
->ps_inv_multiplier
= 1;
2535 if ((LP_DEBUG
& DEBUG_FS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
2536 lp_debug_fs_variant(variant
);
2539 lp_jit_init_types(variant
);
2541 if (variant
->jit_function
[RAST_EDGE_TEST
] == NULL
)
2542 generate_fragment(lp
, shader
, variant
, RAST_EDGE_TEST
);
2544 if (variant
->jit_function
[RAST_WHOLE
] == NULL
) {
2545 if (variant
->opaque
) {
2546 /* Specialized shader, which doesn't need to read the color buffer. */
2547 generate_fragment(lp
, shader
, variant
, RAST_WHOLE
);
2552 * Compile everything
2555 gallivm_compile_module(variant
->gallivm
);
2557 if (variant
->function
[RAST_EDGE_TEST
]) {
2558 variant
->jit_function
[RAST_EDGE_TEST
] = (lp_jit_frag_func
)
2559 gallivm_jit_function(variant
->gallivm
,
2560 variant
->function
[RAST_EDGE_TEST
]);
2563 if (variant
->function
[RAST_WHOLE
]) {
2564 variant
->jit_function
[RAST_WHOLE
] = (lp_jit_frag_func
)
2565 gallivm_jit_function(variant
->gallivm
,
2566 variant
->function
[RAST_WHOLE
]);
2567 } else if (!variant
->jit_function
[RAST_WHOLE
]) {
2568 variant
->jit_function
[RAST_WHOLE
] = variant
->jit_function
[RAST_EDGE_TEST
];
2576 llvmpipe_create_fs_state(struct pipe_context
*pipe
,
2577 const struct pipe_shader_state
*templ
)
2579 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2580 struct lp_fragment_shader
*shader
;
2582 int nr_sampler_views
;
2585 shader
= CALLOC_STRUCT(lp_fragment_shader
);
2589 shader
->no
= fs_no
++;
2590 make_empty_list(&shader
->variants
);
2592 /* get/save the summary info for this shader */
2593 lp_build_tgsi_info(templ
->tokens
, &shader
->info
);
2595 /* we need to keep a local copy of the tokens */
2596 shader
->base
.tokens
= tgsi_dup_tokens(templ
->tokens
);
2598 shader
->draw_data
= draw_create_fragment_shader(llvmpipe
->draw
, templ
);
2599 if (shader
->draw_data
== NULL
) {
2600 FREE((void *) shader
->base
.tokens
);
2605 nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
2606 nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
2608 shader
->variant_key_size
= Offset(struct lp_fragment_shader_variant_key
,
2609 state
[MAX2(nr_samplers
, nr_sampler_views
)]);
2611 for (i
= 0; i
< shader
->info
.base
.num_inputs
; i
++) {
2612 shader
->inputs
[i
].usage_mask
= shader
->info
.base
.input_usage_mask
[i
];
2613 shader
->inputs
[i
].cyl_wrap
= shader
->info
.base
.input_cylindrical_wrap
[i
];
2615 switch (shader
->info
.base
.input_interpolate
[i
]) {
2616 case TGSI_INTERPOLATE_CONSTANT
:
2617 shader
->inputs
[i
].interp
= LP_INTERP_CONSTANT
;
2619 case TGSI_INTERPOLATE_LINEAR
:
2620 shader
->inputs
[i
].interp
= LP_INTERP_LINEAR
;
2622 case TGSI_INTERPOLATE_PERSPECTIVE
:
2623 shader
->inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
2625 case TGSI_INTERPOLATE_COLOR
:
2626 shader
->inputs
[i
].interp
= LP_INTERP_COLOR
;
2633 switch (shader
->info
.base
.input_semantic_name
[i
]) {
2634 case TGSI_SEMANTIC_FACE
:
2635 shader
->inputs
[i
].interp
= LP_INTERP_FACING
;
2637 case TGSI_SEMANTIC_POSITION
:
2638 /* Position was already emitted above
2640 shader
->inputs
[i
].interp
= LP_INTERP_POSITION
;
2641 shader
->inputs
[i
].src_index
= 0;
2645 shader
->inputs
[i
].src_index
= i
+1;
2648 if (LP_DEBUG
& DEBUG_TGSI
) {
2650 debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
2651 shader
->no
, (void *) shader
);
2652 tgsi_dump(templ
->tokens
, 0);
2653 debug_printf("usage masks:\n");
2654 for (attrib
= 0; attrib
< shader
->info
.base
.num_inputs
; ++attrib
) {
2655 unsigned usage_mask
= shader
->info
.base
.input_usage_mask
[attrib
];
2656 debug_printf(" IN[%u].%s%s%s%s\n",
2658 usage_mask
& TGSI_WRITEMASK_X
? "x" : "",
2659 usage_mask
& TGSI_WRITEMASK_Y
? "y" : "",
2660 usage_mask
& TGSI_WRITEMASK_Z
? "z" : "",
2661 usage_mask
& TGSI_WRITEMASK_W
? "w" : "");
2671 llvmpipe_bind_fs_state(struct pipe_context
*pipe
, void *fs
)
2673 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2675 if (llvmpipe
->fs
== fs
)
2678 llvmpipe
->fs
= (struct lp_fragment_shader
*) fs
;
2680 draw_bind_fragment_shader(llvmpipe
->draw
,
2681 (llvmpipe
->fs
? llvmpipe
->fs
->draw_data
: NULL
));
2683 llvmpipe
->dirty
|= LP_NEW_FS
;
2688 * Remove shader variant from two lists: the shader's variant list
2689 * and the context's variant list.
2692 llvmpipe_remove_shader_variant(struct llvmpipe_context
*lp
,
2693 struct lp_fragment_shader_variant
*variant
)
2697 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
2698 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
2699 " #%u v total cached #%u\n",
2700 variant
->shader
->no
,
2702 variant
->shader
->variants_created
,
2703 variant
->shader
->variants_cached
,
2704 lp
->nr_fs_variants
);
2707 /* free all the variant's JIT'd functions */
2708 for (i
= 0; i
< Elements(variant
->function
); i
++) {
2709 if (variant
->function
[i
]) {
2710 gallivm_free_function(variant
->gallivm
,
2711 variant
->function
[i
],
2712 variant
->jit_function
[i
]);
2716 gallivm_destroy(variant
->gallivm
);
2718 /* remove from shader's list */
2719 remove_from_list(&variant
->list_item_local
);
2720 variant
->shader
->variants_cached
--;
2722 /* remove from context's list */
2723 remove_from_list(&variant
->list_item_global
);
2724 lp
->nr_fs_variants
--;
2725 lp
->nr_fs_instrs
-= variant
->nr_instrs
;
2732 llvmpipe_delete_fs_state(struct pipe_context
*pipe
, void *fs
)
2734 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2735 struct lp_fragment_shader
*shader
= fs
;
2736 struct lp_fs_variant_list_item
*li
;
2738 assert(fs
!= llvmpipe
->fs
);
2741 * XXX: we need to flush the context until we have some sort of reference
2742 * counting in fragment shaders as they may still be binned
2743 * Flushing alone might not sufficient we need to wait on it too.
2745 llvmpipe_finish(pipe
, __FUNCTION__
);
2747 /* Delete all the variants */
2748 li
= first_elem(&shader
->variants
);
2749 while(!at_end(&shader
->variants
, li
)) {
2750 struct lp_fs_variant_list_item
*next
= next_elem(li
);
2751 llvmpipe_remove_shader_variant(llvmpipe
, li
->base
);
2755 /* Delete draw module's data */
2756 draw_delete_fragment_shader(llvmpipe
->draw
, shader
->draw_data
);
2758 assert(shader
->variants_cached
== 0);
2759 FREE((void *) shader
->base
.tokens
);
2766 llvmpipe_set_constant_buffer(struct pipe_context
*pipe
,
2767 uint shader
, uint index
,
2768 struct pipe_constant_buffer
*cb
)
2770 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
2771 struct pipe_resource
*constants
= cb
? cb
->buffer
: NULL
;
2773 assert(shader
< PIPE_SHADER_TYPES
);
2774 assert(index
< Elements(llvmpipe
->constants
[shader
]));
2776 /* note: reference counting */
2777 util_copy_constant_buffer(&llvmpipe
->constants
[shader
][index
], cb
);
2779 if (shader
== PIPE_SHADER_VERTEX
||
2780 shader
== PIPE_SHADER_GEOMETRY
) {
2781 /* Pass the constants to the 'draw' module */
2782 const unsigned size
= cb
? cb
->buffer_size
: 0;
2786 data
= (ubyte
*) llvmpipe_resource_data(constants
);
2788 else if (cb
&& cb
->user_buffer
) {
2789 data
= (ubyte
*) cb
->user_buffer
;
2796 data
+= cb
->buffer_offset
;
2798 draw_set_mapped_constant_buffer(llvmpipe
->draw
, shader
,
2802 llvmpipe
->dirty
|= LP_NEW_CONSTANTS
;
2804 if (cb
&& cb
->user_buffer
) {
2805 pipe_resource_reference(&constants
, NULL
);
2811 * Return the blend factor equivalent to a destination alpha of one.
2813 static INLINE
unsigned
2814 force_dst_alpha_one(unsigned factor
, boolean clamped_zero
)
2817 case PIPE_BLENDFACTOR_DST_ALPHA
:
2818 return PIPE_BLENDFACTOR_ONE
;
2819 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
2820 return PIPE_BLENDFACTOR_ZERO
;
2821 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
2823 return PIPE_BLENDFACTOR_ZERO
;
2825 return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
;
2833 * We need to generate several variants of the fragment pipeline to match
2834 * all the combinations of the contributing state atoms.
2836 * TODO: there is actually no reason to tie this to context state -- the
2837 * generated code could be cached globally in the screen.
2840 make_variant_key(struct llvmpipe_context
*lp
,
2841 struct lp_fragment_shader
*shader
,
2842 struct lp_fragment_shader_variant_key
*key
)
2846 memset(key
, 0, shader
->variant_key_size
);
2848 if (lp
->framebuffer
.zsbuf
) {
2849 enum pipe_format zsbuf_format
= lp
->framebuffer
.zsbuf
->format
;
2850 const struct util_format_description
*zsbuf_desc
=
2851 util_format_description(zsbuf_format
);
2853 if (lp
->depth_stencil
->depth
.enabled
&&
2854 util_format_has_depth(zsbuf_desc
)) {
2855 key
->zsbuf_format
= zsbuf_format
;
2856 memcpy(&key
->depth
, &lp
->depth_stencil
->depth
, sizeof key
->depth
);
2858 if (lp
->depth_stencil
->stencil
[0].enabled
&&
2859 util_format_has_stencil(zsbuf_desc
)) {
2860 key
->zsbuf_format
= zsbuf_format
;
2861 memcpy(&key
->stencil
, &lp
->depth_stencil
->stencil
, sizeof key
->stencil
);
2863 if (llvmpipe_resource_is_1d(lp
->framebuffer
.zsbuf
->texture
)) {
2864 key
->resource_1d
= TRUE
;
2869 * Propagate the depth clamp setting from the rasterizer state.
2870 * depth_clip == 0 implies depth clamping is enabled.
2872 * When clip_halfz is enabled, then always clamp the depth values.
2874 if (lp
->rasterizer
->clip_halfz
) {
2875 key
->depth_clamp
= 1;
2877 key
->depth_clamp
= (lp
->rasterizer
->depth_clip
== 0) ? 1 : 0;
2880 /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
2881 if (!lp
->framebuffer
.nr_cbufs
||
2882 !util_format_is_pure_integer(lp
->framebuffer
.cbufs
[0]->format
)) {
2883 key
->alpha
.enabled
= lp
->depth_stencil
->alpha
.enabled
;
2885 if(key
->alpha
.enabled
)
2886 key
->alpha
.func
= lp
->depth_stencil
->alpha
.func
;
2887 /* alpha.ref_value is passed in jit_context */
2889 key
->flatshade
= lp
->rasterizer
->flatshade
;
2890 if (lp
->active_occlusion_queries
) {
2891 key
->occlusion_count
= TRUE
;
2894 if (lp
->framebuffer
.nr_cbufs
) {
2895 memcpy(&key
->blend
, lp
->blend
, sizeof key
->blend
);
2898 key
->nr_cbufs
= lp
->framebuffer
.nr_cbufs
;
2900 if (!key
->blend
.independent_blend_enable
) {
2901 /* we always need independent blend otherwise the fixups below won't work */
2902 for (i
= 1; i
< key
->nr_cbufs
; i
++) {
2903 memcpy(&key
->blend
.rt
[i
], &key
->blend
.rt
[0], sizeof(key
->blend
.rt
[0]));
2905 key
->blend
.independent_blend_enable
= 1;
2908 for (i
= 0; i
< lp
->framebuffer
.nr_cbufs
; i
++) {
2909 enum pipe_format format
= lp
->framebuffer
.cbufs
[i
]->format
;
2910 struct pipe_rt_blend_state
*blend_rt
= &key
->blend
.rt
[i
];
2911 const struct util_format_description
*format_desc
;
2913 key
->cbuf_format
[i
] = format
;
2916 * Figure out if this is a 1d resource. Note that OpenGL allows crazy
2917 * mixing of 2d textures with height 1 and 1d textures, so make sure
2918 * we pick 1d if any cbuf or zsbuf is 1d.
2920 if (llvmpipe_resource_is_1d(lp
->framebuffer
.cbufs
[0]->texture
)) {
2921 key
->resource_1d
= TRUE
;
2924 format_desc
= util_format_description(format
);
2925 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
||
2926 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_SRGB
);
2929 * Mask out color channels not present in the color buffer.
2931 blend_rt
->colormask
&= util_format_colormask(format_desc
);
2934 * Disable blend for integer formats.
2936 if (util_format_is_pure_integer(format
)) {
2937 blend_rt
->blend_enable
= 0;
2941 * Our swizzled render tiles always have an alpha channel, but the linear
2942 * render target format often does not, so force here the dst alpha to be
2945 * This is not a mere optimization. Wrong results will be produced if the
2946 * dst alpha is used, the dst format does not have alpha, and the previous
2947 * rendering was not flushed from the swizzled to linear buffer. For
2948 * example, NonPowTwo DCT.
2950 * TODO: This should be generalized to all channels for better
2951 * performance, but only alpha causes correctness issues.
2953 * Also, force rgb/alpha func/factors match, to make AoS blending easier.
2955 if (format_desc
->swizzle
[3] > UTIL_FORMAT_SWIZZLE_W
||
2956 format_desc
->swizzle
[3] == format_desc
->swizzle
[0]) {
2957 /* Doesn't cover mixed snorm/unorm but can't render to them anyway */
2958 boolean clamped_zero
= !util_format_is_float(format
) &&
2959 !util_format_is_snorm(format
);
2960 blend_rt
->rgb_src_factor
= force_dst_alpha_one(blend_rt
->rgb_src_factor
,
2962 blend_rt
->rgb_dst_factor
= force_dst_alpha_one(blend_rt
->rgb_dst_factor
,
2964 blend_rt
->alpha_func
= blend_rt
->rgb_func
;
2965 blend_rt
->alpha_src_factor
= blend_rt
->rgb_src_factor
;
2966 blend_rt
->alpha_dst_factor
= blend_rt
->rgb_dst_factor
;
2970 /* This value will be the same for all the variants of a given shader:
2972 key
->nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
2974 for(i
= 0; i
< key
->nr_samplers
; ++i
) {
2975 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
2976 lp_sampler_static_sampler_state(&key
->state
[i
].sampler_state
,
2977 lp
->samplers
[PIPE_SHADER_FRAGMENT
][i
]);
2982 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
2983 * are dx10-style? Can't really have mixed opcodes, at least not
2984 * if we want to skip the holes here (without rescanning tgsi).
2986 if (shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
2987 key
->nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
2988 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
2989 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER_VIEW
] & (1 << i
)) {
2990 lp_sampler_static_texture_state(&key
->state
[i
].texture_state
,
2991 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
2996 key
->nr_sampler_views
= key
->nr_samplers
;
2997 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
2998 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
2999 lp_sampler_static_texture_state(&key
->state
[i
].texture_state
,
3000 lp
->sampler_views
[PIPE_SHADER_FRAGMENT
][i
]);
3009 * Update fragment shader state. This is called just prior to drawing
3010 * something when some fragment-related state has changed.
3013 llvmpipe_update_fs(struct llvmpipe_context
*lp
)
3015 struct lp_fragment_shader
*shader
= lp
->fs
;
3016 struct lp_fragment_shader_variant_key key
;
3017 struct lp_fragment_shader_variant
*variant
= NULL
;
3018 struct lp_fs_variant_list_item
*li
;
3020 make_variant_key(lp
, shader
, &key
);
3022 /* Search the variants for one which matches the key */
3023 li
= first_elem(&shader
->variants
);
3024 while(!at_end(&shader
->variants
, li
)) {
3025 if(memcmp(&li
->base
->key
, &key
, shader
->variant_key_size
) == 0) {
3033 /* Move this variant to the head of the list to implement LRU
3034 * deletion of shader's when we have too many.
3036 move_to_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
3039 /* variant not found, create it now */
3042 unsigned variants_to_cull
;
3045 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
3048 lp
->nr_fs_variants
? lp
->nr_fs_instrs
/ lp
->nr_fs_variants
: 0);
3051 /* First, check if we've exceeded the max number of shader variants.
3052 * If so, free 25% of them (the least recently used ones).
3054 variants_to_cull
= lp
->nr_fs_variants
>= LP_MAX_SHADER_VARIANTS
? LP_MAX_SHADER_VARIANTS
/ 4 : 0;
3056 if (variants_to_cull
||
3057 lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
) {
3058 struct pipe_context
*pipe
= &lp
->pipe
;
3061 * XXX: we need to flush the context until we have some sort of
3062 * reference counting in fragment shaders as they may still be binned
3063 * Flushing alone might not be sufficient we need to wait on it too.
3065 llvmpipe_finish(pipe
, __FUNCTION__
);
3068 * We need to re-check lp->nr_fs_variants because an arbitrarliy large
3069 * number of shader variants (potentially all of them) could be
3070 * pending for destruction on flush.
3073 for (i
= 0; i
< variants_to_cull
|| lp
->nr_fs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
; i
++) {
3074 struct lp_fs_variant_list_item
*item
;
3075 if (is_empty_list(&lp
->fs_variants_list
)) {
3078 item
= last_elem(&lp
->fs_variants_list
);
3081 llvmpipe_remove_shader_variant(lp
, item
->base
);
3086 * Generate the new variant.
3089 variant
= generate_variant(lp
, shader
, &key
);
3092 LP_COUNT_ADD(llvm_compile_time
, dt
);
3093 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
3095 llvmpipe_variant_count
++;
3097 /* Put the new variant into the list */
3099 insert_at_head(&shader
->variants
, &variant
->list_item_local
);
3100 insert_at_head(&lp
->fs_variants_list
, &variant
->list_item_global
);
3101 lp
->nr_fs_variants
++;
3102 lp
->nr_fs_instrs
+= variant
->nr_instrs
;
3103 shader
->variants_cached
++;
3107 /* Bind this variant */
3108 lp_setup_set_fs_variant(lp
->setup
, variant
);
3116 llvmpipe_init_fs_funcs(struct llvmpipe_context
*llvmpipe
)
3118 llvmpipe
->pipe
.create_fs_state
= llvmpipe_create_fs_state
;
3119 llvmpipe
->pipe
.bind_fs_state
= llvmpipe_bind_fs_state
;
3120 llvmpipe
->pipe
.delete_fs_state
= llvmpipe_delete_fs_state
;
3122 llvmpipe
->pipe
.set_constant_buffer
= llvmpipe_set_constant_buffer
;
3126 * Rasterization is disabled if there is no pixel shader and
3127 * both depth and stencil testing are disabled:
3128 * http://msdn.microsoft.com/en-us/library/windows/desktop/bb205125
3131 llvmpipe_rasterization_disabled(struct llvmpipe_context
*lp
)
3133 boolean null_fs
= !lp
->fs
|| lp
->fs
->info
.base
.num_tokens
<= 1;
3136 !lp
->depth_stencil
->depth
.enabled
&&
3137 !lp
->depth_stencil
->stencil
[0].enabled
);