2 * Copyright 2012 Advanced Micro Devices, Inc.
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 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the 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 NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
24 * Tom Stellard <thomas.stellard@amd.com>
25 * Michel Dänzer <michel.daenzer@amd.com>
26 * Christian König <christian.koenig@amd.com>
29 #include "gallivm/lp_bld_const.h"
30 #include "gallivm/lp_bld_gather.h"
31 #include "gallivm/lp_bld_intr.h"
32 #include "gallivm/lp_bld_logic.h"
33 #include "gallivm/lp_bld_arit.h"
34 #include "gallivm/lp_bld_flow.h"
35 #include "gallivm/lp_bld_misc.h"
36 #include "util/u_memory.h"
37 #include "util/u_string.h"
38 #include "tgsi/tgsi_build.h"
39 #include "tgsi/tgsi_util.h"
40 #include "tgsi/tgsi_dump.h"
42 #include "ac_binary.h"
43 #include "ac_llvm_util.h"
44 #include "si_shader_internal.h"
49 static const char *scratch_rsrc_dword0_symbol
=
50 "SCRATCH_RSRC_DWORD0";
52 static const char *scratch_rsrc_dword1_symbol
=
53 "SCRATCH_RSRC_DWORD1";
55 struct si_shader_output_values
57 LLVMValueRef values
[4];
58 unsigned semantic_name
;
59 unsigned semantic_index
;
60 ubyte vertex_stream
[4];
63 static void si_init_shader_ctx(struct si_shader_context
*ctx
,
64 struct si_screen
*sscreen
,
65 struct si_shader
*shader
,
66 LLVMTargetMachineRef tm
);
68 static void si_llvm_emit_barrier(const struct lp_build_tgsi_action
*action
,
69 struct lp_build_tgsi_context
*bld_base
,
70 struct lp_build_emit_data
*emit_data
);
72 static void si_dump_shader_key(unsigned shader
, struct si_shader_key
*key
,
75 static void si_build_vs_prolog_function(struct si_shader_context
*ctx
,
76 union si_shader_part_key
*key
);
77 static void si_build_vs_epilog_function(struct si_shader_context
*ctx
,
78 union si_shader_part_key
*key
);
79 static void si_build_tcs_epilog_function(struct si_shader_context
*ctx
,
80 union si_shader_part_key
*key
);
81 static void si_build_ps_prolog_function(struct si_shader_context
*ctx
,
82 union si_shader_part_key
*key
);
83 static void si_build_ps_epilog_function(struct si_shader_context
*ctx
,
84 union si_shader_part_key
*key
);
86 /* Ideally pass the sample mask input to the PS epilog as v13, which
87 * is its usual location, so that the shader doesn't have to add v_mov.
89 #define PS_EPILOG_SAMPLEMASK_MIN_LOC 13
91 /* The VS location of the PrimitiveID input is the same in the epilog,
92 * so that the main shader part doesn't have to move it.
94 #define VS_EPILOG_PRIMID_LOC 2
102 * Returns a unique index for a semantic name and index. The index must be
103 * less than 64, so that a 64-bit bitmask of used inputs or outputs can be
106 unsigned si_shader_io_get_unique_index(unsigned semantic_name
, unsigned index
)
108 switch (semantic_name
) {
109 case TGSI_SEMANTIC_POSITION
:
111 case TGSI_SEMANTIC_PSIZE
:
113 case TGSI_SEMANTIC_CLIPDIST
:
116 case TGSI_SEMANTIC_GENERIC
:
120 assert(!"invalid generic index");
123 /* patch indices are completely separate and thus start from 0 */
124 case TGSI_SEMANTIC_TESSOUTER
:
126 case TGSI_SEMANTIC_TESSINNER
:
128 case TGSI_SEMANTIC_PATCH
:
132 assert(!"invalid semantic name");
137 unsigned si_shader_io_get_unique_index2(unsigned name
, unsigned index
)
140 case TGSI_SEMANTIC_FOG
:
142 case TGSI_SEMANTIC_LAYER
:
144 case TGSI_SEMANTIC_VIEWPORT_INDEX
:
146 case TGSI_SEMANTIC_PRIMID
:
148 case TGSI_SEMANTIC_COLOR
: /* these alias */
149 case TGSI_SEMANTIC_BCOLOR
:
151 case TGSI_SEMANTIC_TEXCOORD
:
154 assert(!"invalid semantic name");
160 * Get the value of a shader input parameter and extract a bitfield.
162 static LLVMValueRef
unpack_param(struct si_shader_context
*ctx
,
163 unsigned param
, unsigned rshift
,
166 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
167 LLVMValueRef value
= LLVMGetParam(ctx
->main_fn
,
170 if (LLVMGetTypeKind(LLVMTypeOf(value
)) == LLVMFloatTypeKind
)
171 value
= bitcast(&ctx
->bld_base
,
172 TGSI_TYPE_UNSIGNED
, value
);
175 value
= LLVMBuildLShr(gallivm
->builder
, value
,
176 lp_build_const_int32(gallivm
, rshift
), "");
178 if (rshift
+ bitwidth
< 32) {
179 unsigned mask
= (1 << bitwidth
) - 1;
180 value
= LLVMBuildAnd(gallivm
->builder
, value
,
181 lp_build_const_int32(gallivm
, mask
), "");
187 static LLVMValueRef
get_rel_patch_id(struct si_shader_context
*ctx
)
190 case PIPE_SHADER_TESS_CTRL
:
191 return unpack_param(ctx
, SI_PARAM_REL_IDS
, 0, 8);
193 case PIPE_SHADER_TESS_EVAL
:
194 return LLVMGetParam(ctx
->main_fn
,
195 ctx
->param_tes_rel_patch_id
);
203 /* Tessellation shaders pass outputs to the next shader using LDS.
205 * LS outputs = TCS inputs
206 * TCS outputs = TES inputs
209 * - TCS inputs for patch 0
210 * - TCS inputs for patch 1
211 * - TCS inputs for patch 2 = get_tcs_in_current_patch_offset (if RelPatchID==2)
213 * - TCS outputs for patch 0 = get_tcs_out_patch0_offset
214 * - Per-patch TCS outputs for patch 0 = get_tcs_out_patch0_patch_data_offset
215 * - TCS outputs for patch 1
216 * - Per-patch TCS outputs for patch 1
217 * - TCS outputs for patch 2 = get_tcs_out_current_patch_offset (if RelPatchID==2)
218 * - Per-patch TCS outputs for patch 2 = get_tcs_out_current_patch_data_offset (if RelPatchID==2)
221 * All three shaders VS(LS), TCS, TES share the same LDS space.
225 get_tcs_in_patch_stride(struct si_shader_context
*ctx
)
227 if (ctx
->type
== PIPE_SHADER_VERTEX
)
228 return unpack_param(ctx
, SI_PARAM_LS_OUT_LAYOUT
, 0, 13);
229 else if (ctx
->type
== PIPE_SHADER_TESS_CTRL
)
230 return unpack_param(ctx
, SI_PARAM_TCS_IN_LAYOUT
, 0, 13);
238 get_tcs_out_patch_stride(struct si_shader_context
*ctx
)
240 return unpack_param(ctx
, SI_PARAM_TCS_OUT_LAYOUT
, 0, 13);
244 get_tcs_out_patch0_offset(struct si_shader_context
*ctx
)
246 return lp_build_mul_imm(&ctx
->bld_base
.uint_bld
,
248 SI_PARAM_TCS_OUT_OFFSETS
,
254 get_tcs_out_patch0_patch_data_offset(struct si_shader_context
*ctx
)
256 return lp_build_mul_imm(&ctx
->bld_base
.uint_bld
,
258 SI_PARAM_TCS_OUT_OFFSETS
,
264 get_tcs_in_current_patch_offset(struct si_shader_context
*ctx
)
266 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
267 LLVMValueRef patch_stride
= get_tcs_in_patch_stride(ctx
);
268 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
270 return LLVMBuildMul(gallivm
->builder
, patch_stride
, rel_patch_id
, "");
274 get_tcs_out_current_patch_offset(struct si_shader_context
*ctx
)
276 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
277 LLVMValueRef patch0_offset
= get_tcs_out_patch0_offset(ctx
);
278 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(ctx
);
279 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
281 return LLVMBuildAdd(gallivm
->builder
, patch0_offset
,
282 LLVMBuildMul(gallivm
->builder
, patch_stride
,
288 get_tcs_out_current_patch_data_offset(struct si_shader_context
*ctx
)
290 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
291 LLVMValueRef patch0_patch_data_offset
=
292 get_tcs_out_patch0_patch_data_offset(ctx
);
293 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(ctx
);
294 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
296 return LLVMBuildAdd(gallivm
->builder
, patch0_patch_data_offset
,
297 LLVMBuildMul(gallivm
->builder
, patch_stride
,
302 static LLVMValueRef
get_instance_index_for_fetch(
303 struct si_shader_context
*radeon_bld
,
304 unsigned param_start_instance
, unsigned divisor
)
306 struct si_shader_context
*ctx
=
307 si_shader_context(&radeon_bld
->bld_base
);
308 struct gallivm_state
*gallivm
= radeon_bld
->bld_base
.base
.gallivm
;
310 LLVMValueRef result
= LLVMGetParam(radeon_bld
->main_fn
,
311 ctx
->param_instance_id
);
313 /* The division must be done before START_INSTANCE is added. */
315 result
= LLVMBuildUDiv(gallivm
->builder
, result
,
316 lp_build_const_int32(gallivm
, divisor
), "");
318 return LLVMBuildAdd(gallivm
->builder
, result
,
319 LLVMGetParam(radeon_bld
->main_fn
, param_start_instance
), "");
322 /* Bitcast <4 x float> to <2 x double>, extract the component, and convert
324 static LLVMValueRef
extract_double_to_float(struct si_shader_context
*ctx
,
326 unsigned double_index
)
328 LLVMBuilderRef builder
= ctx
->gallivm
.builder
;
329 LLVMTypeRef f64
= LLVMDoubleTypeInContext(ctx
->gallivm
.context
);
330 LLVMValueRef dvec2
= LLVMBuildBitCast(builder
, vec4
,
331 LLVMVectorType(f64
, 2), "");
332 LLVMValueRef index
= LLVMConstInt(ctx
->i32
, double_index
, 0);
333 LLVMValueRef value
= LLVMBuildExtractElement(builder
, dvec2
, index
, "");
334 return LLVMBuildFPTrunc(builder
, value
, ctx
->f32
, "");
337 static void declare_input_vs(
338 struct si_shader_context
*ctx
,
339 unsigned input_index
,
340 const struct tgsi_full_declaration
*decl
,
343 struct lp_build_context
*base
= &ctx
->bld_base
.base
;
344 struct gallivm_state
*gallivm
= base
->gallivm
;
348 unsigned num_fetches
;
349 unsigned fetch_stride
;
351 LLVMValueRef t_list_ptr
;
352 LLVMValueRef t_offset
;
354 LLVMValueRef vertex_index
;
355 LLVMValueRef input
[3];
357 /* Load the T list */
358 t_list_ptr
= LLVMGetParam(ctx
->main_fn
, SI_PARAM_VERTEX_BUFFERS
);
360 t_offset
= lp_build_const_int32(gallivm
, input_index
);
362 t_list
= ac_build_indexed_load_const(&ctx
->ac
, t_list_ptr
, t_offset
);
364 vertex_index
= LLVMGetParam(ctx
->main_fn
,
365 ctx
->param_vertex_index0
+
368 fix_fetch
= ctx
->shader
->key
.mono
.vs
.fix_fetch
[input_index
];
370 /* Do multiple loads for special formats. */
372 case SI_FIX_FETCH_RGB_64_FLOAT
:
373 num_fetches
= 3; /* 3 2-dword loads */
376 case SI_FIX_FETCH_RGBA_64_FLOAT
:
377 num_fetches
= 2; /* 2 4-dword loads */
380 case SI_FIX_FETCH_RGB_8
:
381 case SI_FIX_FETCH_RGB_8_INT
:
385 case SI_FIX_FETCH_RGB_16
:
386 case SI_FIX_FETCH_RGB_16_INT
:
395 for (unsigned i
= 0; i
< num_fetches
; i
++) {
396 LLVMValueRef voffset
= LLVMConstInt(ctx
->i32
, fetch_stride
* i
, 0);
398 input
[i
] = ac_build_buffer_load_format(&ctx
->ac
, t_list
,
399 vertex_index
, voffset
,
403 /* Break up the vec4 into individual components */
404 for (chan
= 0; chan
< 4; chan
++) {
405 LLVMValueRef llvm_chan
= lp_build_const_int32(gallivm
, chan
);
406 out
[chan
] = LLVMBuildExtractElement(gallivm
->builder
,
407 input
[0], llvm_chan
, "");
411 case SI_FIX_FETCH_A2_SNORM
:
412 case SI_FIX_FETCH_A2_SSCALED
:
413 case SI_FIX_FETCH_A2_SINT
: {
414 /* The hardware returns an unsigned value; convert it to a
417 LLVMValueRef tmp
= out
[3];
418 LLVMValueRef c30
= LLVMConstInt(ctx
->i32
, 30, 0);
420 /* First, recover the sign-extended signed integer value. */
421 if (fix_fetch
== SI_FIX_FETCH_A2_SSCALED
)
422 tmp
= LLVMBuildFPToUI(gallivm
->builder
, tmp
, ctx
->i32
, "");
424 tmp
= LLVMBuildBitCast(gallivm
->builder
, tmp
, ctx
->i32
, "");
426 /* For the integer-like cases, do a natural sign extension.
428 * For the SNORM case, the values are 0.0, 0.333, 0.666, 1.0
429 * and happen to contain 0, 1, 2, 3 as the two LSBs of the
432 tmp
= LLVMBuildShl(gallivm
->builder
, tmp
,
433 fix_fetch
== SI_FIX_FETCH_A2_SNORM
?
434 LLVMConstInt(ctx
->i32
, 7, 0) : c30
, "");
435 tmp
= LLVMBuildAShr(gallivm
->builder
, tmp
, c30
, "");
437 /* Convert back to the right type. */
438 if (fix_fetch
== SI_FIX_FETCH_A2_SNORM
) {
440 LLVMValueRef neg_one
= LLVMConstReal(ctx
->f32
, -1.0);
441 tmp
= LLVMBuildSIToFP(gallivm
->builder
, tmp
, ctx
->f32
, "");
442 clamp
= LLVMBuildFCmp(gallivm
->builder
, LLVMRealULT
, tmp
, neg_one
, "");
443 tmp
= LLVMBuildSelect(gallivm
->builder
, clamp
, neg_one
, tmp
, "");
444 } else if (fix_fetch
== SI_FIX_FETCH_A2_SSCALED
) {
445 tmp
= LLVMBuildSIToFP(gallivm
->builder
, tmp
, ctx
->f32
, "");
451 case SI_FIX_FETCH_RGBA_32_UNORM
:
452 case SI_FIX_FETCH_RGBX_32_UNORM
:
453 for (chan
= 0; chan
< 4; chan
++) {
454 out
[chan
] = LLVMBuildBitCast(gallivm
->builder
, out
[chan
],
456 out
[chan
] = LLVMBuildUIToFP(gallivm
->builder
,
457 out
[chan
], ctx
->f32
, "");
458 out
[chan
] = LLVMBuildFMul(gallivm
->builder
, out
[chan
],
459 LLVMConstReal(ctx
->f32
, 1.0 / UINT_MAX
), "");
461 /* RGBX UINT returns 1 in alpha, which would be rounded to 0 by normalizing. */
462 if (fix_fetch
== SI_FIX_FETCH_RGBX_32_UNORM
)
463 out
[3] = LLVMConstReal(ctx
->f32
, 1);
465 case SI_FIX_FETCH_RGBA_32_SNORM
:
466 case SI_FIX_FETCH_RGBX_32_SNORM
:
467 case SI_FIX_FETCH_RGBA_32_FIXED
:
468 case SI_FIX_FETCH_RGBX_32_FIXED
: {
470 if (fix_fetch
>= SI_FIX_FETCH_RGBA_32_FIXED
)
471 scale
= 1.0 / 0x10000;
473 scale
= 1.0 / INT_MAX
;
475 for (chan
= 0; chan
< 4; chan
++) {
476 out
[chan
] = LLVMBuildBitCast(gallivm
->builder
, out
[chan
],
478 out
[chan
] = LLVMBuildSIToFP(gallivm
->builder
,
479 out
[chan
], ctx
->f32
, "");
480 out
[chan
] = LLVMBuildFMul(gallivm
->builder
, out
[chan
],
481 LLVMConstReal(ctx
->f32
, scale
), "");
483 /* RGBX SINT returns 1 in alpha, which would be rounded to 0 by normalizing. */
484 if (fix_fetch
== SI_FIX_FETCH_RGBX_32_SNORM
||
485 fix_fetch
== SI_FIX_FETCH_RGBX_32_FIXED
)
486 out
[3] = LLVMConstReal(ctx
->f32
, 1);
489 case SI_FIX_FETCH_RGBA_32_USCALED
:
490 for (chan
= 0; chan
< 4; chan
++) {
491 out
[chan
] = LLVMBuildBitCast(gallivm
->builder
, out
[chan
],
493 out
[chan
] = LLVMBuildUIToFP(gallivm
->builder
,
494 out
[chan
], ctx
->f32
, "");
497 case SI_FIX_FETCH_RGBA_32_SSCALED
:
498 for (chan
= 0; chan
< 4; chan
++) {
499 out
[chan
] = LLVMBuildBitCast(gallivm
->builder
, out
[chan
],
501 out
[chan
] = LLVMBuildSIToFP(gallivm
->builder
,
502 out
[chan
], ctx
->f32
, "");
505 case SI_FIX_FETCH_RG_64_FLOAT
:
506 for (chan
= 0; chan
< 2; chan
++)
507 out
[chan
] = extract_double_to_float(ctx
, input
[0], chan
);
509 out
[2] = LLVMConstReal(ctx
->f32
, 0);
510 out
[3] = LLVMConstReal(ctx
->f32
, 1);
512 case SI_FIX_FETCH_RGB_64_FLOAT
:
513 for (chan
= 0; chan
< 3; chan
++)
514 out
[chan
] = extract_double_to_float(ctx
, input
[chan
], 0);
516 out
[3] = LLVMConstReal(ctx
->f32
, 1);
518 case SI_FIX_FETCH_RGBA_64_FLOAT
:
519 for (chan
= 0; chan
< 4; chan
++) {
520 out
[chan
] = extract_double_to_float(ctx
, input
[chan
/ 2],
524 case SI_FIX_FETCH_RGB_8
:
525 case SI_FIX_FETCH_RGB_8_INT
:
526 case SI_FIX_FETCH_RGB_16
:
527 case SI_FIX_FETCH_RGB_16_INT
:
528 for (chan
= 0; chan
< 3; chan
++) {
529 out
[chan
] = LLVMBuildExtractElement(gallivm
->builder
,
533 if (fix_fetch
== SI_FIX_FETCH_RGB_8
||
534 fix_fetch
== SI_FIX_FETCH_RGB_16
) {
535 out
[3] = LLVMConstReal(ctx
->f32
, 1);
537 out
[3] = LLVMBuildBitCast(gallivm
->builder
, ctx
->i32_1
,
544 static LLVMValueRef
get_primitive_id(struct lp_build_tgsi_context
*bld_base
,
547 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
550 return bld_base
->uint_bld
.zero
;
553 case PIPE_SHADER_VERTEX
:
554 return LLVMGetParam(ctx
->main_fn
,
555 ctx
->param_vs_prim_id
);
556 case PIPE_SHADER_TESS_CTRL
:
557 return LLVMGetParam(ctx
->main_fn
,
559 case PIPE_SHADER_TESS_EVAL
:
560 return LLVMGetParam(ctx
->main_fn
,
561 ctx
->param_tes_patch_id
);
562 case PIPE_SHADER_GEOMETRY
:
563 return LLVMGetParam(ctx
->main_fn
,
564 SI_PARAM_PRIMITIVE_ID
);
567 return bld_base
->uint_bld
.zero
;
572 * Return the value of tgsi_ind_register for indexing.
573 * This is the indirect index with the constant offset added to it.
575 static LLVMValueRef
get_indirect_index(struct si_shader_context
*ctx
,
576 const struct tgsi_ind_register
*ind
,
579 struct gallivm_state
*gallivm
= ctx
->bld_base
.base
.gallivm
;
582 result
= ctx
->addrs
[ind
->Index
][ind
->Swizzle
];
583 result
= LLVMBuildLoad(gallivm
->builder
, result
, "");
584 result
= LLVMBuildAdd(gallivm
->builder
, result
,
585 lp_build_const_int32(gallivm
, rel_index
), "");
590 * Like get_indirect_index, but restricts the return value to a (possibly
591 * undefined) value inside [0..num).
593 static LLVMValueRef
get_bounded_indirect_index(struct si_shader_context
*ctx
,
594 const struct tgsi_ind_register
*ind
,
595 int rel_index
, unsigned num
)
597 LLVMValueRef result
= get_indirect_index(ctx
, ind
, rel_index
);
599 /* LLVM 3.8: If indirect resource indexing is used:
603 if (HAVE_LLVM
== 0x0308)
604 return LLVMGetUndef(ctx
->i32
);
606 return si_llvm_bound_index(ctx
, result
, num
);
611 * Calculate a dword address given an input or output register and a stride.
613 static LLVMValueRef
get_dw_address(struct si_shader_context
*ctx
,
614 const struct tgsi_full_dst_register
*dst
,
615 const struct tgsi_full_src_register
*src
,
616 LLVMValueRef vertex_dw_stride
,
617 LLVMValueRef base_addr
)
619 struct gallivm_state
*gallivm
= ctx
->bld_base
.base
.gallivm
;
620 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
621 ubyte
*name
, *index
, *array_first
;
623 struct tgsi_full_dst_register reg
;
625 /* Set the register description. The address computation is the same
626 * for sources and destinations. */
628 reg
.Register
.File
= src
->Register
.File
;
629 reg
.Register
.Index
= src
->Register
.Index
;
630 reg
.Register
.Indirect
= src
->Register
.Indirect
;
631 reg
.Register
.Dimension
= src
->Register
.Dimension
;
632 reg
.Indirect
= src
->Indirect
;
633 reg
.Dimension
= src
->Dimension
;
634 reg
.DimIndirect
= src
->DimIndirect
;
638 /* If the register is 2-dimensional (e.g. an array of vertices
639 * in a primitive), calculate the base address of the vertex. */
640 if (reg
.Register
.Dimension
) {
643 if (reg
.Dimension
.Indirect
)
644 index
= get_indirect_index(ctx
, ®
.DimIndirect
,
645 reg
.Dimension
.Index
);
647 index
= lp_build_const_int32(gallivm
, reg
.Dimension
.Index
);
649 base_addr
= LLVMBuildAdd(gallivm
->builder
, base_addr
,
650 LLVMBuildMul(gallivm
->builder
, index
,
651 vertex_dw_stride
, ""), "");
654 /* Get information about the register. */
655 if (reg
.Register
.File
== TGSI_FILE_INPUT
) {
656 name
= info
->input_semantic_name
;
657 index
= info
->input_semantic_index
;
658 array_first
= info
->input_array_first
;
659 } else if (reg
.Register
.File
== TGSI_FILE_OUTPUT
) {
660 name
= info
->output_semantic_name
;
661 index
= info
->output_semantic_index
;
662 array_first
= info
->output_array_first
;
668 if (reg
.Register
.Indirect
) {
669 /* Add the relative address of the element. */
670 LLVMValueRef ind_index
;
672 if (reg
.Indirect
.ArrayID
)
673 first
= array_first
[reg
.Indirect
.ArrayID
];
675 first
= reg
.Register
.Index
;
677 ind_index
= get_indirect_index(ctx
, ®
.Indirect
,
678 reg
.Register
.Index
- first
);
680 base_addr
= LLVMBuildAdd(gallivm
->builder
, base_addr
,
681 LLVMBuildMul(gallivm
->builder
, ind_index
,
682 lp_build_const_int32(gallivm
, 4), ""), "");
684 param
= si_shader_io_get_unique_index(name
[first
], index
[first
]);
686 param
= si_shader_io_get_unique_index(name
[reg
.Register
.Index
],
687 index
[reg
.Register
.Index
]);
690 /* Add the base address of the element. */
691 return LLVMBuildAdd(gallivm
->builder
, base_addr
,
692 lp_build_const_int32(gallivm
, param
* 4), "");
695 /* The offchip buffer layout for TCS->TES is
697 * - attribute 0 of patch 0 vertex 0
698 * - attribute 0 of patch 0 vertex 1
699 * - attribute 0 of patch 0 vertex 2
701 * - attribute 0 of patch 1 vertex 0
702 * - attribute 0 of patch 1 vertex 1
704 * - attribute 1 of patch 0 vertex 0
705 * - attribute 1 of patch 0 vertex 1
707 * - per patch attribute 0 of patch 0
708 * - per patch attribute 0 of patch 1
711 * Note that every attribute has 4 components.
713 static LLVMValueRef
get_tcs_tes_buffer_address(struct si_shader_context
*ctx
,
714 LLVMValueRef rel_patch_id
,
715 LLVMValueRef vertex_index
,
716 LLVMValueRef param_index
)
718 struct gallivm_state
*gallivm
= ctx
->bld_base
.base
.gallivm
;
719 LLVMValueRef base_addr
, vertices_per_patch
, num_patches
, total_vertices
;
720 LLVMValueRef param_stride
, constant16
;
722 vertices_per_patch
= unpack_param(ctx
, SI_PARAM_TCS_OFFCHIP_LAYOUT
, 9, 6);
723 num_patches
= unpack_param(ctx
, SI_PARAM_TCS_OFFCHIP_LAYOUT
, 0, 9);
724 total_vertices
= LLVMBuildMul(gallivm
->builder
, vertices_per_patch
,
727 constant16
= lp_build_const_int32(gallivm
, 16);
729 base_addr
= LLVMBuildMul(gallivm
->builder
, rel_patch_id
,
730 vertices_per_patch
, "");
732 base_addr
= LLVMBuildAdd(gallivm
->builder
, base_addr
,
735 param_stride
= total_vertices
;
737 base_addr
= rel_patch_id
;
738 param_stride
= num_patches
;
741 base_addr
= LLVMBuildAdd(gallivm
->builder
, base_addr
,
742 LLVMBuildMul(gallivm
->builder
, param_index
,
743 param_stride
, ""), "");
745 base_addr
= LLVMBuildMul(gallivm
->builder
, base_addr
, constant16
, "");
748 LLVMValueRef patch_data_offset
=
749 unpack_param(ctx
, SI_PARAM_TCS_OFFCHIP_LAYOUT
, 16, 16);
751 base_addr
= LLVMBuildAdd(gallivm
->builder
, base_addr
,
752 patch_data_offset
, "");
757 static LLVMValueRef
get_tcs_tes_buffer_address_from_reg(
758 struct si_shader_context
*ctx
,
759 const struct tgsi_full_dst_register
*dst
,
760 const struct tgsi_full_src_register
*src
)
762 struct gallivm_state
*gallivm
= ctx
->bld_base
.base
.gallivm
;
763 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
764 ubyte
*name
, *index
, *array_first
;
765 struct tgsi_full_src_register reg
;
766 LLVMValueRef vertex_index
= NULL
;
767 LLVMValueRef param_index
= NULL
;
768 unsigned param_index_base
, param_base
;
770 reg
= src
? *src
: tgsi_full_src_register_from_dst(dst
);
772 if (reg
.Register
.Dimension
) {
774 if (reg
.Dimension
.Indirect
)
775 vertex_index
= get_indirect_index(ctx
, ®
.DimIndirect
,
776 reg
.Dimension
.Index
);
778 vertex_index
= lp_build_const_int32(gallivm
,
779 reg
.Dimension
.Index
);
782 /* Get information about the register. */
783 if (reg
.Register
.File
== TGSI_FILE_INPUT
) {
784 name
= info
->input_semantic_name
;
785 index
= info
->input_semantic_index
;
786 array_first
= info
->input_array_first
;
787 } else if (reg
.Register
.File
== TGSI_FILE_OUTPUT
) {
788 name
= info
->output_semantic_name
;
789 index
= info
->output_semantic_index
;
790 array_first
= info
->output_array_first
;
796 if (reg
.Register
.Indirect
) {
797 if (reg
.Indirect
.ArrayID
)
798 param_base
= array_first
[reg
.Indirect
.ArrayID
];
800 param_base
= reg
.Register
.Index
;
802 param_index
= get_indirect_index(ctx
, ®
.Indirect
,
803 reg
.Register
.Index
- param_base
);
806 param_base
= reg
.Register
.Index
;
807 param_index
= lp_build_const_int32(gallivm
, 0);
810 param_index_base
= si_shader_io_get_unique_index(name
[param_base
],
813 param_index
= LLVMBuildAdd(gallivm
->builder
, param_index
,
814 lp_build_const_int32(gallivm
, param_index_base
),
817 return get_tcs_tes_buffer_address(ctx
, get_rel_patch_id(ctx
),
818 vertex_index
, param_index
);
821 static LLVMValueRef
buffer_load(struct lp_build_tgsi_context
*bld_base
,
822 enum tgsi_opcode_type type
, unsigned swizzle
,
823 LLVMValueRef buffer
, LLVMValueRef offset
,
824 LLVMValueRef base
, bool readonly_memory
)
826 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
827 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
828 LLVMValueRef value
, value2
;
829 LLVMTypeRef llvm_type
= tgsi2llvmtype(bld_base
, type
);
830 LLVMTypeRef vec_type
= LLVMVectorType(llvm_type
, 4);
833 value
= ac_build_buffer_load(&ctx
->ac
, buffer
, 4, NULL
, base
, offset
,
834 0, 1, 0, readonly_memory
);
836 return LLVMBuildBitCast(gallivm
->builder
, value
, vec_type
, "");
839 if (!tgsi_type_is_64bit(type
)) {
840 value
= ac_build_buffer_load(&ctx
->ac
, buffer
, 4, NULL
, base
, offset
,
841 0, 1, 0, readonly_memory
);
843 value
= LLVMBuildBitCast(gallivm
->builder
, value
, vec_type
, "");
844 return LLVMBuildExtractElement(gallivm
->builder
, value
,
845 lp_build_const_int32(gallivm
, swizzle
), "");
848 value
= ac_build_buffer_load(&ctx
->ac
, buffer
, 1, NULL
, base
, offset
,
849 swizzle
* 4, 1, 0, readonly_memory
);
851 value2
= ac_build_buffer_load(&ctx
->ac
, buffer
, 1, NULL
, base
, offset
,
852 swizzle
* 4 + 4, 1, 0, readonly_memory
);
854 return si_llvm_emit_fetch_64bit(bld_base
, type
, value
, value2
);
860 * \param type output value type
861 * \param swizzle offset (typically 0..3); it can be ~0, which loads a vec4
862 * \param dw_addr address in dwords
864 static LLVMValueRef
lds_load(struct lp_build_tgsi_context
*bld_base
,
865 enum tgsi_opcode_type type
, unsigned swizzle
,
866 LLVMValueRef dw_addr
)
868 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
869 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
873 LLVMValueRef values
[TGSI_NUM_CHANNELS
];
875 for (unsigned chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++)
876 values
[chan
] = lds_load(bld_base
, type
, chan
, dw_addr
);
878 return lp_build_gather_values(bld_base
->base
.gallivm
, values
,
882 dw_addr
= lp_build_add(&bld_base
->uint_bld
, dw_addr
,
883 lp_build_const_int32(gallivm
, swizzle
));
885 value
= ac_build_indexed_load(&ctx
->ac
, ctx
->lds
, dw_addr
, false);
886 if (tgsi_type_is_64bit(type
)) {
888 dw_addr
= lp_build_add(&bld_base
->uint_bld
, dw_addr
,
889 lp_build_const_int32(gallivm
, 1));
890 value2
= ac_build_indexed_load(&ctx
->ac
, ctx
->lds
, dw_addr
, false);
891 return si_llvm_emit_fetch_64bit(bld_base
, type
, value
, value2
);
894 return LLVMBuildBitCast(gallivm
->builder
, value
,
895 tgsi2llvmtype(bld_base
, type
), "");
901 * \param swizzle offset (typically 0..3)
902 * \param dw_addr address in dwords
903 * \param value value to store
905 static void lds_store(struct lp_build_tgsi_context
*bld_base
,
906 unsigned swizzle
, LLVMValueRef dw_addr
,
909 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
910 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
912 dw_addr
= lp_build_add(&bld_base
->uint_bld
, dw_addr
,
913 lp_build_const_int32(gallivm
, swizzle
));
915 value
= LLVMBuildBitCast(gallivm
->builder
, value
, ctx
->i32
, "");
916 ac_build_indexed_store(&ctx
->ac
, ctx
->lds
,
920 static LLVMValueRef
fetch_input_tcs(
921 struct lp_build_tgsi_context
*bld_base
,
922 const struct tgsi_full_src_register
*reg
,
923 enum tgsi_opcode_type type
, unsigned swizzle
)
925 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
926 LLVMValueRef dw_addr
, stride
;
928 stride
= unpack_param(ctx
, SI_PARAM_TCS_IN_LAYOUT
, 13, 8);
929 dw_addr
= get_tcs_in_current_patch_offset(ctx
);
930 dw_addr
= get_dw_address(ctx
, NULL
, reg
, stride
, dw_addr
);
932 return lds_load(bld_base
, type
, swizzle
, dw_addr
);
935 static LLVMValueRef
fetch_output_tcs(
936 struct lp_build_tgsi_context
*bld_base
,
937 const struct tgsi_full_src_register
*reg
,
938 enum tgsi_opcode_type type
, unsigned swizzle
)
940 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
941 LLVMValueRef dw_addr
, stride
;
943 if (reg
->Register
.Dimension
) {
944 stride
= unpack_param(ctx
, SI_PARAM_TCS_OUT_LAYOUT
, 13, 8);
945 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
946 dw_addr
= get_dw_address(ctx
, NULL
, reg
, stride
, dw_addr
);
948 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
949 dw_addr
= get_dw_address(ctx
, NULL
, reg
, NULL
, dw_addr
);
952 return lds_load(bld_base
, type
, swizzle
, dw_addr
);
955 static LLVMValueRef
fetch_input_tes(
956 struct lp_build_tgsi_context
*bld_base
,
957 const struct tgsi_full_src_register
*reg
,
958 enum tgsi_opcode_type type
, unsigned swizzle
)
960 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
961 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
962 LLVMValueRef rw_buffers
, buffer
, base
, addr
;
964 rw_buffers
= LLVMGetParam(ctx
->main_fn
,
965 SI_PARAM_RW_BUFFERS
);
966 buffer
= ac_build_indexed_load_const(&ctx
->ac
, rw_buffers
,
967 lp_build_const_int32(gallivm
, SI_HS_RING_TESS_OFFCHIP
));
969 base
= LLVMGetParam(ctx
->main_fn
, ctx
->param_oc_lds
);
970 addr
= get_tcs_tes_buffer_address_from_reg(ctx
, NULL
, reg
);
972 return buffer_load(bld_base
, type
, swizzle
, buffer
, base
, addr
, true);
975 static void store_output_tcs(struct lp_build_tgsi_context
*bld_base
,
976 const struct tgsi_full_instruction
*inst
,
977 const struct tgsi_opcode_info
*info
,
980 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
981 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
982 const struct tgsi_full_dst_register
*reg
= &inst
->Dst
[0];
983 const struct tgsi_shader_info
*sh_info
= &ctx
->shader
->selector
->info
;
985 LLVMValueRef dw_addr
, stride
;
986 LLVMValueRef rw_buffers
, buffer
, base
, buf_addr
;
987 LLVMValueRef values
[4];
989 bool is_tess_factor
= false;
991 /* Only handle per-patch and per-vertex outputs here.
992 * Vectors will be lowered to scalars and this function will be called again.
994 if (reg
->Register
.File
!= TGSI_FILE_OUTPUT
||
995 (dst
[0] && LLVMGetTypeKind(LLVMTypeOf(dst
[0])) == LLVMVectorTypeKind
)) {
996 si_llvm_emit_store(bld_base
, inst
, info
, dst
);
1000 if (reg
->Register
.Dimension
) {
1001 stride
= unpack_param(ctx
, SI_PARAM_TCS_OUT_LAYOUT
, 13, 8);
1002 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
1003 dw_addr
= get_dw_address(ctx
, reg
, NULL
, stride
, dw_addr
);
1004 skip_lds_store
= !sh_info
->reads_pervertex_outputs
;
1006 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
1007 dw_addr
= get_dw_address(ctx
, reg
, NULL
, NULL
, dw_addr
);
1008 skip_lds_store
= !sh_info
->reads_perpatch_outputs
;
1010 if (!reg
->Register
.Indirect
) {
1011 int name
= sh_info
->output_semantic_name
[reg
->Register
.Index
];
1013 /* Always write tess factors into LDS for the TCS epilog. */
1014 if (name
== TGSI_SEMANTIC_TESSINNER
||
1015 name
== TGSI_SEMANTIC_TESSOUTER
) {
1016 skip_lds_store
= false;
1017 is_tess_factor
= true;
1022 rw_buffers
= LLVMGetParam(ctx
->main_fn
,
1023 SI_PARAM_RW_BUFFERS
);
1024 buffer
= ac_build_indexed_load_const(&ctx
->ac
, rw_buffers
,
1025 lp_build_const_int32(gallivm
, SI_HS_RING_TESS_OFFCHIP
));
1027 base
= LLVMGetParam(ctx
->main_fn
, ctx
->param_oc_lds
);
1028 buf_addr
= get_tcs_tes_buffer_address_from_reg(ctx
, reg
, NULL
);
1031 TGSI_FOR_EACH_DST0_ENABLED_CHANNEL(inst
, chan_index
) {
1032 LLVMValueRef value
= dst
[chan_index
];
1034 if (inst
->Instruction
.Saturate
)
1035 value
= ac_build_clamp(&ctx
->ac
, value
);
1037 /* Skip LDS stores if there is no LDS read of this output. */
1038 if (!skip_lds_store
)
1039 lds_store(bld_base
, chan_index
, dw_addr
, value
);
1041 value
= LLVMBuildBitCast(gallivm
->builder
, value
, ctx
->i32
, "");
1042 values
[chan_index
] = value
;
1044 if (inst
->Dst
[0].Register
.WriteMask
!= 0xF && !is_tess_factor
) {
1045 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, value
, 1,
1047 4 * chan_index
, 1, 0, true, false);
1051 if (inst
->Dst
[0].Register
.WriteMask
== 0xF && !is_tess_factor
) {
1052 LLVMValueRef value
= lp_build_gather_values(bld_base
->base
.gallivm
,
1054 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, value
, 4, buf_addr
,
1055 base
, 0, 1, 0, true, false);
1059 static LLVMValueRef
fetch_input_gs(
1060 struct lp_build_tgsi_context
*bld_base
,
1061 const struct tgsi_full_src_register
*reg
,
1062 enum tgsi_opcode_type type
,
1065 struct lp_build_context
*base
= &bld_base
->base
;
1066 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1067 struct si_shader
*shader
= ctx
->shader
;
1068 struct lp_build_context
*uint
= &ctx
->bld_base
.uint_bld
;
1069 struct gallivm_state
*gallivm
= base
->gallivm
;
1070 LLVMValueRef vtx_offset
, soffset
;
1071 unsigned vtx_offset_param
;
1072 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
1073 unsigned semantic_name
= info
->input_semantic_name
[reg
->Register
.Index
];
1074 unsigned semantic_index
= info
->input_semantic_index
[reg
->Register
.Index
];
1078 if (swizzle
!= ~0 && semantic_name
== TGSI_SEMANTIC_PRIMID
)
1079 return get_primitive_id(bld_base
, swizzle
);
1081 if (!reg
->Register
.Dimension
)
1084 if (swizzle
== ~0) {
1085 LLVMValueRef values
[TGSI_NUM_CHANNELS
];
1087 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
1088 values
[chan
] = fetch_input_gs(bld_base
, reg
, type
, chan
);
1090 return lp_build_gather_values(bld_base
->base
.gallivm
, values
,
1094 /* Get the vertex offset parameter */
1095 vtx_offset_param
= reg
->Dimension
.Index
;
1096 if (vtx_offset_param
< 2) {
1097 vtx_offset_param
+= SI_PARAM_VTX0_OFFSET
;
1099 assert(vtx_offset_param
< 6);
1100 vtx_offset_param
+= SI_PARAM_VTX2_OFFSET
- 2;
1102 vtx_offset
= lp_build_mul_imm(uint
,
1103 LLVMGetParam(ctx
->main_fn
,
1107 param
= si_shader_io_get_unique_index(semantic_name
, semantic_index
);
1108 soffset
= LLVMConstInt(ctx
->i32
, (param
* 4 + swizzle
) * 256, 0);
1110 value
= ac_build_buffer_load(&ctx
->ac
, ctx
->esgs_ring
, 1, uint
->zero
,
1111 vtx_offset
, soffset
, 0, 1, 0, true);
1112 if (tgsi_type_is_64bit(type
)) {
1113 LLVMValueRef value2
;
1114 soffset
= LLVMConstInt(ctx
->i32
, (param
* 4 + swizzle
+ 1) * 256, 0);
1116 value2
= ac_build_buffer_load(&ctx
->ac
, ctx
->esgs_ring
, 1,
1117 uint
->zero
, vtx_offset
, soffset
,
1119 return si_llvm_emit_fetch_64bit(bld_base
, type
,
1122 return LLVMBuildBitCast(gallivm
->builder
,
1124 tgsi2llvmtype(bld_base
, type
), "");
1127 static int lookup_interp_param_index(unsigned interpolate
, unsigned location
)
1129 switch (interpolate
) {
1130 case TGSI_INTERPOLATE_CONSTANT
:
1133 case TGSI_INTERPOLATE_LINEAR
:
1134 if (location
== TGSI_INTERPOLATE_LOC_SAMPLE
)
1135 return SI_PARAM_LINEAR_SAMPLE
;
1136 else if (location
== TGSI_INTERPOLATE_LOC_CENTROID
)
1137 return SI_PARAM_LINEAR_CENTROID
;
1139 return SI_PARAM_LINEAR_CENTER
;
1141 case TGSI_INTERPOLATE_COLOR
:
1142 case TGSI_INTERPOLATE_PERSPECTIVE
:
1143 if (location
== TGSI_INTERPOLATE_LOC_SAMPLE
)
1144 return SI_PARAM_PERSP_SAMPLE
;
1145 else if (location
== TGSI_INTERPOLATE_LOC_CENTROID
)
1146 return SI_PARAM_PERSP_CENTROID
;
1148 return SI_PARAM_PERSP_CENTER
;
1151 fprintf(stderr
, "Warning: Unhandled interpolation mode.\n");
1157 * Interpolate a fragment shader input.
1159 * @param ctx context
1160 * @param input_index index of the input in hardware
1161 * @param semantic_name TGSI_SEMANTIC_*
1162 * @param semantic_index semantic index
1163 * @param num_interp_inputs number of all interpolated inputs (= BCOLOR offset)
1164 * @param colors_read_mask color components read (4 bits for each color, 8 bits in total)
1165 * @param interp_param interpolation weights (i,j)
1166 * @param prim_mask SI_PARAM_PRIM_MASK
1167 * @param face SI_PARAM_FRONT_FACE
1168 * @param result the return value (4 components)
1170 static void interp_fs_input(struct si_shader_context
*ctx
,
1171 unsigned input_index
,
1172 unsigned semantic_name
,
1173 unsigned semantic_index
,
1174 unsigned num_interp_inputs
,
1175 unsigned colors_read_mask
,
1176 LLVMValueRef interp_param
,
1177 LLVMValueRef prim_mask
,
1179 LLVMValueRef result
[4])
1181 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
1182 struct lp_build_context
*base
= &bld_base
->base
;
1183 struct lp_build_context
*uint
= &bld_base
->uint_bld
;
1184 struct gallivm_state
*gallivm
= base
->gallivm
;
1185 LLVMValueRef attr_number
;
1190 /* fs.constant returns the param from the middle vertex, so it's not
1191 * really useful for flat shading. It's meant to be used for custom
1192 * interpolation (but the intrinsic can't fetch from the other two
1195 * Luckily, it doesn't matter, because we rely on the FLAT_SHADE state
1196 * to do the right thing. The only reason we use fs.constant is that
1197 * fs.interp cannot be used on integers, because they can be equal
1200 * When interp is false we will use fs.constant or for newer llvm,
1201 * amdgcn.interp.mov.
1203 bool interp
= interp_param
!= NULL
;
1205 attr_number
= lp_build_const_int32(gallivm
, input_index
);
1208 interp_param
= LLVMBuildBitCast(gallivm
->builder
, interp_param
,
1209 LLVMVectorType(ctx
->f32
, 2), "");
1211 i
= LLVMBuildExtractElement(gallivm
->builder
, interp_param
,
1213 j
= LLVMBuildExtractElement(gallivm
->builder
, interp_param
,
1217 if (semantic_name
== TGSI_SEMANTIC_COLOR
&&
1218 ctx
->shader
->key
.part
.ps
.prolog
.color_two_side
) {
1219 LLVMValueRef is_face_positive
;
1220 LLVMValueRef back_attr_number
;
1222 /* If BCOLOR0 is used, BCOLOR1 is at offset "num_inputs + 1",
1223 * otherwise it's at offset "num_inputs".
1225 unsigned back_attr_offset
= num_interp_inputs
;
1226 if (semantic_index
== 1 && colors_read_mask
& 0xf)
1227 back_attr_offset
+= 1;
1229 back_attr_number
= lp_build_const_int32(gallivm
, back_attr_offset
);
1231 is_face_positive
= LLVMBuildICmp(gallivm
->builder
, LLVMIntNE
,
1232 face
, uint
->zero
, "");
1234 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
1235 LLVMValueRef llvm_chan
= lp_build_const_int32(gallivm
, chan
);
1236 LLVMValueRef front
, back
;
1239 front
= ac_build_fs_interp(&ctx
->ac
, llvm_chan
,
1240 attr_number
, prim_mask
,
1242 back
= ac_build_fs_interp(&ctx
->ac
, llvm_chan
,
1243 back_attr_number
, prim_mask
,
1246 front
= ac_build_fs_interp_mov(&ctx
->ac
,
1247 lp_build_const_int32(gallivm
, 2), /* P0 */
1248 llvm_chan
, attr_number
, prim_mask
);
1249 back
= ac_build_fs_interp_mov(&ctx
->ac
,
1250 lp_build_const_int32(gallivm
, 2), /* P0 */
1251 llvm_chan
, back_attr_number
, prim_mask
);
1254 result
[chan
] = LLVMBuildSelect(gallivm
->builder
,
1260 } else if (semantic_name
== TGSI_SEMANTIC_FOG
) {
1262 result
[0] = ac_build_fs_interp(&ctx
->ac
, uint
->zero
,
1263 attr_number
, prim_mask
, i
, j
);
1265 result
[0] = ac_build_fs_interp_mov(&ctx
->ac
, uint
->zero
,
1266 lp_build_const_int32(gallivm
, 2), /* P0 */
1267 attr_number
, prim_mask
);
1270 result
[2] = lp_build_const_float(gallivm
, 0.0f
);
1271 result
[3] = lp_build_const_float(gallivm
, 1.0f
);
1273 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
1274 LLVMValueRef llvm_chan
= lp_build_const_int32(gallivm
, chan
);
1277 result
[chan
] = ac_build_fs_interp(&ctx
->ac
,
1278 llvm_chan
, attr_number
, prim_mask
, i
, j
);
1280 result
[chan
] = ac_build_fs_interp_mov(&ctx
->ac
,
1281 lp_build_const_int32(gallivm
, 2), /* P0 */
1282 llvm_chan
, attr_number
, prim_mask
);
1288 static void declare_input_fs(
1289 struct si_shader_context
*radeon_bld
,
1290 unsigned input_index
,
1291 const struct tgsi_full_declaration
*decl
,
1292 LLVMValueRef out
[4])
1294 struct lp_build_context
*base
= &radeon_bld
->bld_base
.base
;
1295 struct si_shader_context
*ctx
=
1296 si_shader_context(&radeon_bld
->bld_base
);
1297 struct si_shader
*shader
= ctx
->shader
;
1298 LLVMValueRef main_fn
= radeon_bld
->main_fn
;
1299 LLVMValueRef interp_param
= NULL
;
1300 int interp_param_idx
;
1302 /* Get colors from input VGPRs (set by the prolog). */
1303 if (decl
->Semantic
.Name
== TGSI_SEMANTIC_COLOR
) {
1304 unsigned i
= decl
->Semantic
.Index
;
1305 unsigned colors_read
= shader
->selector
->info
.colors_read
;
1306 unsigned mask
= colors_read
>> (i
* 4);
1307 unsigned offset
= SI_PARAM_POS_FIXED_PT
+ 1 +
1308 (i
? util_bitcount(colors_read
& 0xf) : 0);
1310 out
[0] = mask
& 0x1 ? LLVMGetParam(main_fn
, offset
++) : base
->undef
;
1311 out
[1] = mask
& 0x2 ? LLVMGetParam(main_fn
, offset
++) : base
->undef
;
1312 out
[2] = mask
& 0x4 ? LLVMGetParam(main_fn
, offset
++) : base
->undef
;
1313 out
[3] = mask
& 0x8 ? LLVMGetParam(main_fn
, offset
++) : base
->undef
;
1317 interp_param_idx
= lookup_interp_param_index(decl
->Interp
.Interpolate
,
1318 decl
->Interp
.Location
);
1319 if (interp_param_idx
== -1)
1321 else if (interp_param_idx
) {
1322 interp_param
= LLVMGetParam(ctx
->main_fn
, interp_param_idx
);
1325 if (decl
->Semantic
.Name
== TGSI_SEMANTIC_COLOR
&&
1326 decl
->Interp
.Interpolate
== TGSI_INTERPOLATE_COLOR
&&
1327 ctx
->shader
->key
.part
.ps
.prolog
.flatshade_colors
)
1328 interp_param
= NULL
; /* load the constant color */
1330 interp_fs_input(ctx
, input_index
, decl
->Semantic
.Name
,
1331 decl
->Semantic
.Index
, shader
->selector
->info
.num_inputs
,
1332 shader
->selector
->info
.colors_read
, interp_param
,
1333 LLVMGetParam(main_fn
, SI_PARAM_PRIM_MASK
),
1334 LLVMGetParam(main_fn
, SI_PARAM_FRONT_FACE
),
1338 static LLVMValueRef
get_sample_id(struct si_shader_context
*radeon_bld
)
1340 return unpack_param(si_shader_context(&radeon_bld
->bld_base
),
1341 SI_PARAM_ANCILLARY
, 8, 4);
1346 * Load a dword from a constant buffer.
1348 static LLVMValueRef
buffer_load_const(struct si_shader_context
*ctx
,
1349 LLVMValueRef resource
,
1350 LLVMValueRef offset
)
1352 LLVMBuilderRef builder
= ctx
->gallivm
.builder
;
1353 LLVMValueRef args
[2] = {resource
, offset
};
1355 return lp_build_intrinsic(builder
, "llvm.SI.load.const", ctx
->f32
, args
, 2,
1356 LP_FUNC_ATTR_READNONE
|
1357 LP_FUNC_ATTR_LEGACY
);
1360 static LLVMValueRef
load_sample_position(struct si_shader_context
*radeon_bld
, LLVMValueRef sample_id
)
1362 struct si_shader_context
*ctx
=
1363 si_shader_context(&radeon_bld
->bld_base
);
1364 struct lp_build_context
*uint_bld
= &radeon_bld
->bld_base
.uint_bld
;
1365 struct gallivm_state
*gallivm
= &radeon_bld
->gallivm
;
1366 LLVMBuilderRef builder
= gallivm
->builder
;
1367 LLVMValueRef desc
= LLVMGetParam(ctx
->main_fn
, SI_PARAM_RW_BUFFERS
);
1368 LLVMValueRef buf_index
= lp_build_const_int32(gallivm
, SI_PS_CONST_SAMPLE_POSITIONS
);
1369 LLVMValueRef resource
= ac_build_indexed_load_const(&ctx
->ac
, desc
, buf_index
);
1371 /* offset = sample_id * 8 (8 = 2 floats containing samplepos.xy) */
1372 LLVMValueRef offset0
= lp_build_mul_imm(uint_bld
, sample_id
, 8);
1373 LLVMValueRef offset1
= LLVMBuildAdd(builder
, offset0
, lp_build_const_int32(gallivm
, 4), "");
1375 LLVMValueRef pos
[4] = {
1376 buffer_load_const(ctx
, resource
, offset0
),
1377 buffer_load_const(ctx
, resource
, offset1
),
1378 lp_build_const_float(gallivm
, 0),
1379 lp_build_const_float(gallivm
, 0)
1382 return lp_build_gather_values(gallivm
, pos
, 4);
1385 static void declare_system_value(
1386 struct si_shader_context
*radeon_bld
,
1388 const struct tgsi_full_declaration
*decl
)
1390 struct si_shader_context
*ctx
=
1391 si_shader_context(&radeon_bld
->bld_base
);
1392 struct lp_build_context
*bld
= &radeon_bld
->bld_base
.base
;
1393 struct gallivm_state
*gallivm
= &radeon_bld
->gallivm
;
1394 LLVMValueRef value
= 0;
1396 switch (decl
->Semantic
.Name
) {
1397 case TGSI_SEMANTIC_INSTANCEID
:
1398 value
= LLVMGetParam(radeon_bld
->main_fn
,
1399 ctx
->param_instance_id
);
1402 case TGSI_SEMANTIC_VERTEXID
:
1403 value
= LLVMBuildAdd(gallivm
->builder
,
1404 LLVMGetParam(radeon_bld
->main_fn
,
1405 ctx
->param_vertex_id
),
1406 LLVMGetParam(radeon_bld
->main_fn
,
1407 SI_PARAM_BASE_VERTEX
), "");
1410 case TGSI_SEMANTIC_VERTEXID_NOBASE
:
1411 value
= LLVMGetParam(radeon_bld
->main_fn
,
1412 ctx
->param_vertex_id
);
1415 case TGSI_SEMANTIC_BASEVERTEX
:
1416 value
= LLVMGetParam(radeon_bld
->main_fn
,
1417 SI_PARAM_BASE_VERTEX
);
1420 case TGSI_SEMANTIC_BASEINSTANCE
:
1421 value
= LLVMGetParam(radeon_bld
->main_fn
,
1422 SI_PARAM_START_INSTANCE
);
1425 case TGSI_SEMANTIC_DRAWID
:
1426 value
= LLVMGetParam(radeon_bld
->main_fn
,
1430 case TGSI_SEMANTIC_INVOCATIONID
:
1431 if (ctx
->type
== PIPE_SHADER_TESS_CTRL
)
1432 value
= unpack_param(ctx
, SI_PARAM_REL_IDS
, 8, 5);
1433 else if (ctx
->type
== PIPE_SHADER_GEOMETRY
)
1434 value
= LLVMGetParam(radeon_bld
->main_fn
,
1435 SI_PARAM_GS_INSTANCE_ID
);
1437 assert(!"INVOCATIONID not implemented");
1440 case TGSI_SEMANTIC_POSITION
:
1442 LLVMValueRef pos
[4] = {
1443 LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_POS_X_FLOAT
),
1444 LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_POS_Y_FLOAT
),
1445 LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_POS_Z_FLOAT
),
1446 lp_build_emit_llvm_unary(&radeon_bld
->bld_base
, TGSI_OPCODE_RCP
,
1447 LLVMGetParam(radeon_bld
->main_fn
,
1448 SI_PARAM_POS_W_FLOAT
)),
1450 value
= lp_build_gather_values(gallivm
, pos
, 4);
1454 case TGSI_SEMANTIC_FACE
:
1455 value
= LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_FRONT_FACE
);
1458 case TGSI_SEMANTIC_SAMPLEID
:
1459 value
= get_sample_id(radeon_bld
);
1462 case TGSI_SEMANTIC_SAMPLEPOS
: {
1463 LLVMValueRef pos
[4] = {
1464 LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_POS_X_FLOAT
),
1465 LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_POS_Y_FLOAT
),
1466 lp_build_const_float(gallivm
, 0),
1467 lp_build_const_float(gallivm
, 0)
1469 pos
[0] = lp_build_emit_llvm_unary(&radeon_bld
->bld_base
,
1470 TGSI_OPCODE_FRC
, pos
[0]);
1471 pos
[1] = lp_build_emit_llvm_unary(&radeon_bld
->bld_base
,
1472 TGSI_OPCODE_FRC
, pos
[1]);
1473 value
= lp_build_gather_values(gallivm
, pos
, 4);
1477 case TGSI_SEMANTIC_SAMPLEMASK
:
1478 /* This can only occur with the OpenGL Core profile, which
1479 * doesn't support smoothing.
1481 value
= LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_SAMPLE_COVERAGE
);
1484 case TGSI_SEMANTIC_TESSCOORD
:
1486 LLVMValueRef coord
[4] = {
1487 LLVMGetParam(radeon_bld
->main_fn
, ctx
->param_tes_u
),
1488 LLVMGetParam(radeon_bld
->main_fn
, ctx
->param_tes_v
),
1493 /* For triangles, the vector should be (u, v, 1-u-v). */
1494 if (ctx
->shader
->selector
->info
.properties
[TGSI_PROPERTY_TES_PRIM_MODE
] ==
1495 PIPE_PRIM_TRIANGLES
)
1496 coord
[2] = lp_build_sub(bld
, bld
->one
,
1497 lp_build_add(bld
, coord
[0], coord
[1]));
1499 value
= lp_build_gather_values(gallivm
, coord
, 4);
1503 case TGSI_SEMANTIC_VERTICESIN
:
1504 if (ctx
->type
== PIPE_SHADER_TESS_CTRL
)
1505 value
= unpack_param(ctx
, SI_PARAM_TCS_OUT_LAYOUT
, 26, 6);
1506 else if (ctx
->type
== PIPE_SHADER_TESS_EVAL
)
1507 value
= unpack_param(ctx
, SI_PARAM_TCS_OFFCHIP_LAYOUT
, 9, 7);
1509 assert(!"invalid shader stage for TGSI_SEMANTIC_VERTICESIN");
1512 case TGSI_SEMANTIC_TESSINNER
:
1513 case TGSI_SEMANTIC_TESSOUTER
:
1515 LLVMValueRef rw_buffers
, buffer
, base
, addr
;
1516 int param
= si_shader_io_get_unique_index(decl
->Semantic
.Name
, 0);
1518 rw_buffers
= LLVMGetParam(ctx
->main_fn
,
1519 SI_PARAM_RW_BUFFERS
);
1520 buffer
= ac_build_indexed_load_const(&ctx
->ac
, rw_buffers
,
1521 lp_build_const_int32(gallivm
, SI_HS_RING_TESS_OFFCHIP
));
1523 base
= LLVMGetParam(ctx
->main_fn
, ctx
->param_oc_lds
);
1524 addr
= get_tcs_tes_buffer_address(ctx
, get_rel_patch_id(ctx
), NULL
,
1525 lp_build_const_int32(gallivm
, param
));
1527 value
= buffer_load(&radeon_bld
->bld_base
, TGSI_TYPE_FLOAT
,
1528 ~0, buffer
, base
, addr
, true);
1533 case TGSI_SEMANTIC_DEFAULT_TESSOUTER_SI
:
1534 case TGSI_SEMANTIC_DEFAULT_TESSINNER_SI
:
1536 LLVMValueRef buf
, slot
, val
[4];
1539 slot
= lp_build_const_int32(gallivm
, SI_HS_CONST_DEFAULT_TESS_LEVELS
);
1540 buf
= LLVMGetParam(ctx
->main_fn
, SI_PARAM_RW_BUFFERS
);
1541 buf
= ac_build_indexed_load_const(&ctx
->ac
, buf
, slot
);
1542 offset
= decl
->Semantic
.Name
== TGSI_SEMANTIC_DEFAULT_TESSINNER_SI
? 4 : 0;
1544 for (i
= 0; i
< 4; i
++)
1545 val
[i
] = buffer_load_const(ctx
, buf
,
1546 lp_build_const_int32(gallivm
, (offset
+ i
) * 4));
1547 value
= lp_build_gather_values(gallivm
, val
, 4);
1551 case TGSI_SEMANTIC_PRIMID
:
1552 value
= get_primitive_id(&radeon_bld
->bld_base
, 0);
1555 case TGSI_SEMANTIC_GRID_SIZE
:
1556 value
= LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_GRID_SIZE
);
1559 case TGSI_SEMANTIC_BLOCK_SIZE
:
1561 LLVMValueRef values
[3];
1563 unsigned *properties
= ctx
->shader
->selector
->info
.properties
;
1565 if (properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
] != 0) {
1566 unsigned sizes
[3] = {
1567 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
],
1568 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT
],
1569 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH
]
1572 for (i
= 0; i
< 3; ++i
)
1573 values
[i
] = lp_build_const_int32(gallivm
, sizes
[i
]);
1575 value
= lp_build_gather_values(gallivm
, values
, 3);
1577 value
= LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_BLOCK_SIZE
);
1582 case TGSI_SEMANTIC_BLOCK_ID
:
1583 value
= LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_BLOCK_ID
);
1586 case TGSI_SEMANTIC_THREAD_ID
:
1587 value
= LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_THREAD_ID
);
1590 case TGSI_SEMANTIC_HELPER_INVOCATION
:
1591 if (HAVE_LLVM
>= 0x0309) {
1592 value
= lp_build_intrinsic(gallivm
->builder
,
1593 "llvm.amdgcn.ps.live",
1595 LP_FUNC_ATTR_READNONE
);
1596 value
= LLVMBuildNot(gallivm
->builder
, value
, "");
1597 value
= LLVMBuildSExt(gallivm
->builder
, value
, ctx
->i32
, "");
1599 assert(!"TGSI_SEMANTIC_HELPER_INVOCATION unsupported");
1605 assert(!"unknown system value");
1609 radeon_bld
->system_values
[index
] = value
;
1612 static void declare_compute_memory(struct si_shader_context
*radeon_bld
,
1613 const struct tgsi_full_declaration
*decl
)
1615 struct si_shader_context
*ctx
=
1616 si_shader_context(&radeon_bld
->bld_base
);
1617 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
1618 struct gallivm_state
*gallivm
= &radeon_bld
->gallivm
;
1620 LLVMTypeRef i8p
= LLVMPointerType(ctx
->i8
, LOCAL_ADDR_SPACE
);
1623 assert(decl
->Declaration
.MemType
== TGSI_MEMORY_TYPE_SHARED
);
1624 assert(decl
->Range
.First
== decl
->Range
.Last
);
1625 assert(!ctx
->shared_memory
);
1627 var
= LLVMAddGlobalInAddressSpace(gallivm
->module
,
1628 LLVMArrayType(ctx
->i8
, sel
->local_size
),
1631 LLVMSetAlignment(var
, 4);
1633 ctx
->shared_memory
= LLVMBuildBitCast(gallivm
->builder
, var
, i8p
, "");
1636 static LLVMValueRef
load_const_buffer_desc(struct si_shader_context
*ctx
, int i
)
1638 LLVMValueRef list_ptr
= LLVMGetParam(ctx
->main_fn
,
1639 SI_PARAM_CONST_BUFFERS
);
1641 return ac_build_indexed_load_const(&ctx
->ac
, list_ptr
,
1642 LLVMConstInt(ctx
->i32
, i
, 0));
1645 static LLVMValueRef
fetch_constant(
1646 struct lp_build_tgsi_context
*bld_base
,
1647 const struct tgsi_full_src_register
*reg
,
1648 enum tgsi_opcode_type type
,
1651 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1652 struct lp_build_context
*base
= &bld_base
->base
;
1653 const struct tgsi_ind_register
*ireg
= ®
->Indirect
;
1656 LLVMValueRef addr
, bufp
;
1657 LLVMValueRef result
;
1659 if (swizzle
== LP_CHAN_ALL
) {
1661 LLVMValueRef values
[4];
1662 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
)
1663 values
[chan
] = fetch_constant(bld_base
, reg
, type
, chan
);
1665 return lp_build_gather_values(bld_base
->base
.gallivm
, values
, 4);
1668 buf
= reg
->Register
.Dimension
? reg
->Dimension
.Index
: 0;
1669 idx
= reg
->Register
.Index
* 4 + swizzle
;
1671 if (reg
->Register
.Dimension
&& reg
->Dimension
.Indirect
) {
1672 LLVMValueRef ptr
= LLVMGetParam(ctx
->main_fn
, SI_PARAM_CONST_BUFFERS
);
1674 index
= get_bounded_indirect_index(ctx
, ®
->DimIndirect
,
1675 reg
->Dimension
.Index
,
1676 SI_NUM_CONST_BUFFERS
);
1677 bufp
= ac_build_indexed_load_const(&ctx
->ac
, ptr
, index
);
1679 bufp
= load_const_buffer_desc(ctx
, buf
);
1681 if (reg
->Register
.Indirect
) {
1682 addr
= ctx
->addrs
[ireg
->Index
][ireg
->Swizzle
];
1683 addr
= LLVMBuildLoad(base
->gallivm
->builder
, addr
, "load addr reg");
1684 addr
= lp_build_mul_imm(&bld_base
->uint_bld
, addr
, 16);
1685 addr
= lp_build_add(&bld_base
->uint_bld
, addr
,
1686 lp_build_const_int32(base
->gallivm
, idx
* 4));
1688 addr
= LLVMConstInt(ctx
->i32
, idx
* 4, 0);
1691 result
= buffer_load_const(ctx
, bufp
, addr
);
1693 if (!tgsi_type_is_64bit(type
))
1694 result
= bitcast(bld_base
, type
, result
);
1696 LLVMValueRef addr2
, result2
;
1698 addr2
= lp_build_add(&bld_base
->uint_bld
, addr
,
1699 LLVMConstInt(ctx
->i32
, 4, 0));
1700 result2
= buffer_load_const(ctx
, bufp
, addr2
);
1702 result
= si_llvm_emit_fetch_64bit(bld_base
, type
,
1708 /* Upper 16 bits must be zero. */
1709 static LLVMValueRef
si_llvm_pack_two_int16(struct gallivm_state
*gallivm
,
1710 LLVMValueRef val
[2])
1712 return LLVMBuildOr(gallivm
->builder
, val
[0],
1713 LLVMBuildShl(gallivm
->builder
, val
[1],
1714 lp_build_const_int32(gallivm
, 16),
1718 /* Upper 16 bits are ignored and will be dropped. */
1719 static LLVMValueRef
si_llvm_pack_two_int32_as_int16(struct gallivm_state
*gallivm
,
1720 LLVMValueRef val
[2])
1722 LLVMValueRef v
[2] = {
1723 LLVMBuildAnd(gallivm
->builder
, val
[0],
1724 lp_build_const_int32(gallivm
, 0xffff), ""),
1727 return si_llvm_pack_two_int16(gallivm
, v
);
1730 /* Initialize arguments for the shader export intrinsic */
1731 static void si_llvm_init_export_args(struct lp_build_tgsi_context
*bld_base
,
1732 LLVMValueRef
*values
,
1734 struct ac_export_args
*args
)
1736 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1737 struct lp_build_context
*base
= &bld_base
->base
;
1738 struct gallivm_state
*gallivm
= base
->gallivm
;
1739 LLVMBuilderRef builder
= base
->gallivm
->builder
;
1740 LLVMValueRef val
[4];
1741 unsigned spi_shader_col_format
= V_028714_SPI_SHADER_32_ABGR
;
1743 bool is_int8
, is_int10
;
1745 /* Default is 0xf. Adjusted below depending on the format. */
1746 args
->enabled_channels
= 0xf; /* writemask */
1748 /* Specify whether the EXEC mask represents the valid mask */
1749 args
->valid_mask
= 0;
1751 /* Specify whether this is the last export */
1754 /* Specify the target we are exporting */
1755 args
->target
= target
;
1757 if (ctx
->type
== PIPE_SHADER_FRAGMENT
) {
1758 const struct si_shader_key
*key
= &ctx
->shader
->key
;
1759 unsigned col_formats
= key
->part
.ps
.epilog
.spi_shader_col_format
;
1760 int cbuf
= target
- V_008DFC_SQ_EXP_MRT
;
1762 assert(cbuf
>= 0 && cbuf
< 8);
1763 spi_shader_col_format
= (col_formats
>> (cbuf
* 4)) & 0xf;
1764 is_int8
= (key
->part
.ps
.epilog
.color_is_int8
>> cbuf
) & 0x1;
1765 is_int10
= (key
->part
.ps
.epilog
.color_is_int10
>> cbuf
) & 0x1;
1768 args
->compr
= false;
1769 args
->out
[0] = base
->undef
;
1770 args
->out
[1] = base
->undef
;
1771 args
->out
[2] = base
->undef
;
1772 args
->out
[3] = base
->undef
;
1774 switch (spi_shader_col_format
) {
1775 case V_028714_SPI_SHADER_ZERO
:
1776 args
->enabled_channels
= 0; /* writemask */
1777 args
->target
= V_008DFC_SQ_EXP_NULL
;
1780 case V_028714_SPI_SHADER_32_R
:
1781 args
->enabled_channels
= 1; /* writemask */
1782 args
->out
[0] = values
[0];
1785 case V_028714_SPI_SHADER_32_GR
:
1786 args
->enabled_channels
= 0x3; /* writemask */
1787 args
->out
[0] = values
[0];
1788 args
->out
[1] = values
[1];
1791 case V_028714_SPI_SHADER_32_AR
:
1792 args
->enabled_channels
= 0x9; /* writemask */
1793 args
->out
[0] = values
[0];
1794 args
->out
[3] = values
[3];
1797 case V_028714_SPI_SHADER_FP16_ABGR
:
1798 args
->compr
= 1; /* COMPR flag */
1800 for (chan
= 0; chan
< 2; chan
++) {
1801 LLVMValueRef pack_args
[2] = {
1803 values
[2 * chan
+ 1]
1805 LLVMValueRef packed
;
1807 packed
= ac_build_cvt_pkrtz_f16(&ctx
->ac
, pack_args
);
1809 LLVMBuildBitCast(base
->gallivm
->builder
,
1810 packed
, ctx
->f32
, "");
1814 case V_028714_SPI_SHADER_UNORM16_ABGR
:
1815 for (chan
= 0; chan
< 4; chan
++) {
1816 val
[chan
] = ac_build_clamp(&ctx
->ac
, values
[chan
]);
1817 val
[chan
] = LLVMBuildFMul(builder
, val
[chan
],
1818 lp_build_const_float(gallivm
, 65535), "");
1819 val
[chan
] = LLVMBuildFAdd(builder
, val
[chan
],
1820 lp_build_const_float(gallivm
, 0.5), "");
1821 val
[chan
] = LLVMBuildFPToUI(builder
, val
[chan
],
1825 args
->compr
= 1; /* COMPR flag */
1826 args
->out
[0] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
1827 si_llvm_pack_two_int16(gallivm
, val
));
1828 args
->out
[1] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
1829 si_llvm_pack_two_int16(gallivm
, val
+2));
1832 case V_028714_SPI_SHADER_SNORM16_ABGR
:
1833 for (chan
= 0; chan
< 4; chan
++) {
1834 /* Clamp between [-1, 1]. */
1835 val
[chan
] = lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_MIN
,
1837 lp_build_const_float(gallivm
, 1));
1838 val
[chan
] = lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_MAX
,
1840 lp_build_const_float(gallivm
, -1));
1841 /* Convert to a signed integer in [-32767, 32767]. */
1842 val
[chan
] = LLVMBuildFMul(builder
, val
[chan
],
1843 lp_build_const_float(gallivm
, 32767), "");
1844 /* If positive, add 0.5, else add -0.5. */
1845 val
[chan
] = LLVMBuildFAdd(builder
, val
[chan
],
1846 LLVMBuildSelect(builder
,
1847 LLVMBuildFCmp(builder
, LLVMRealOGE
,
1848 val
[chan
], base
->zero
, ""),
1849 lp_build_const_float(gallivm
, 0.5),
1850 lp_build_const_float(gallivm
, -0.5), ""), "");
1851 val
[chan
] = LLVMBuildFPToSI(builder
, val
[chan
], ctx
->i32
, "");
1854 args
->compr
= 1; /* COMPR flag */
1855 args
->out
[0] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
1856 si_llvm_pack_two_int32_as_int16(gallivm
, val
));
1857 args
->out
[1] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
1858 si_llvm_pack_two_int32_as_int16(gallivm
, val
+2));
1861 case V_028714_SPI_SHADER_UINT16_ABGR
: {
1862 LLVMValueRef max_rgb
= lp_build_const_int32(gallivm
,
1863 is_int8
? 255 : is_int10
? 1023 : 65535);
1864 LLVMValueRef max_alpha
=
1865 !is_int10
? max_rgb
: lp_build_const_int32(gallivm
, 3);
1868 for (chan
= 0; chan
< 4; chan
++) {
1869 val
[chan
] = bitcast(bld_base
, TGSI_TYPE_UNSIGNED
, values
[chan
]);
1870 val
[chan
] = lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_UMIN
,
1872 chan
== 3 ? max_alpha
: max_rgb
);
1875 args
->compr
= 1; /* COMPR flag */
1876 args
->out
[0] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
1877 si_llvm_pack_two_int16(gallivm
, val
));
1878 args
->out
[1] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
1879 si_llvm_pack_two_int16(gallivm
, val
+2));
1883 case V_028714_SPI_SHADER_SINT16_ABGR
: {
1884 LLVMValueRef max_rgb
= lp_build_const_int32(gallivm
,
1885 is_int8
? 127 : is_int10
? 511 : 32767);
1886 LLVMValueRef min_rgb
= lp_build_const_int32(gallivm
,
1887 is_int8
? -128 : is_int10
? -512 : -32768);
1888 LLVMValueRef max_alpha
=
1889 !is_int10
? max_rgb
: lp_build_const_int32(gallivm
, 1);
1890 LLVMValueRef min_alpha
=
1891 !is_int10
? min_rgb
: lp_build_const_int32(gallivm
, -2);
1894 for (chan
= 0; chan
< 4; chan
++) {
1895 val
[chan
] = bitcast(bld_base
, TGSI_TYPE_UNSIGNED
, values
[chan
]);
1896 val
[chan
] = lp_build_emit_llvm_binary(bld_base
,
1898 val
[chan
], chan
== 3 ? max_alpha
: max_rgb
);
1899 val
[chan
] = lp_build_emit_llvm_binary(bld_base
,
1901 val
[chan
], chan
== 3 ? min_alpha
: min_rgb
);
1904 args
->compr
= 1; /* COMPR flag */
1905 args
->out
[0] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
1906 si_llvm_pack_two_int32_as_int16(gallivm
, val
));
1907 args
->out
[1] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
1908 si_llvm_pack_two_int32_as_int16(gallivm
, val
+2));
1912 case V_028714_SPI_SHADER_32_ABGR
:
1913 memcpy(&args
->out
[0], values
, sizeof(values
[0]) * 4);
1918 static void si_alpha_test(struct lp_build_tgsi_context
*bld_base
,
1921 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1922 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
1924 if (ctx
->shader
->key
.part
.ps
.epilog
.alpha_func
!= PIPE_FUNC_NEVER
) {
1925 LLVMValueRef alpha_ref
= LLVMGetParam(ctx
->main_fn
,
1926 SI_PARAM_ALPHA_REF
);
1928 LLVMValueRef alpha_pass
=
1929 lp_build_cmp(&bld_base
->base
,
1930 ctx
->shader
->key
.part
.ps
.epilog
.alpha_func
,
1933 lp_build_select(&bld_base
->base
,
1935 lp_build_const_float(gallivm
, 1.0f
),
1936 lp_build_const_float(gallivm
, -1.0f
));
1938 ac_build_kill(&ctx
->ac
, arg
);
1940 ac_build_kill(&ctx
->ac
, NULL
);
1944 static LLVMValueRef
si_scale_alpha_by_sample_mask(struct lp_build_tgsi_context
*bld_base
,
1946 unsigned samplemask_param
)
1948 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1949 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
1950 LLVMValueRef coverage
;
1952 /* alpha = alpha * popcount(coverage) / SI_NUM_SMOOTH_AA_SAMPLES */
1953 coverage
= LLVMGetParam(ctx
->main_fn
,
1955 coverage
= bitcast(bld_base
, TGSI_TYPE_SIGNED
, coverage
);
1957 coverage
= lp_build_intrinsic(gallivm
->builder
, "llvm.ctpop.i32",
1959 &coverage
, 1, LP_FUNC_ATTR_READNONE
);
1961 coverage
= LLVMBuildUIToFP(gallivm
->builder
, coverage
,
1964 coverage
= LLVMBuildFMul(gallivm
->builder
, coverage
,
1965 lp_build_const_float(gallivm
,
1966 1.0 / SI_NUM_SMOOTH_AA_SAMPLES
), "");
1968 return LLVMBuildFMul(gallivm
->builder
, alpha
, coverage
, "");
1971 static void si_llvm_emit_clipvertex(struct lp_build_tgsi_context
*bld_base
,
1972 struct ac_export_args
*pos
, LLVMValueRef
*out_elts
)
1974 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1975 struct lp_build_context
*base
= &bld_base
->base
;
1978 unsigned const_chan
;
1979 LLVMValueRef base_elt
;
1980 LLVMValueRef ptr
= LLVMGetParam(ctx
->main_fn
, SI_PARAM_RW_BUFFERS
);
1981 LLVMValueRef constbuf_index
= lp_build_const_int32(base
->gallivm
,
1982 SI_VS_CONST_CLIP_PLANES
);
1983 LLVMValueRef const_resource
= ac_build_indexed_load_const(&ctx
->ac
, ptr
, constbuf_index
);
1985 for (reg_index
= 0; reg_index
< 2; reg_index
++) {
1986 struct ac_export_args
*args
= &pos
[2 + reg_index
];
1991 args
->out
[3] = lp_build_const_float(base
->gallivm
, 0.0f
);
1993 /* Compute dot products of position and user clip plane vectors */
1994 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
1995 for (const_chan
= 0; const_chan
< TGSI_NUM_CHANNELS
; const_chan
++) {
1997 LLVMConstInt(ctx
->i32
, ((reg_index
* 4 + chan
) * 4 +
1998 const_chan
) * 4, 0);
1999 base_elt
= buffer_load_const(ctx
, const_resource
,
2002 lp_build_add(base
, args
->out
[chan
],
2003 lp_build_mul(base
, base_elt
,
2004 out_elts
[const_chan
]));
2008 args
->enabled_channels
= 0xf;
2009 args
->valid_mask
= 0;
2011 args
->target
= V_008DFC_SQ_EXP_POS
+ 2 + reg_index
;
2016 static void si_dump_streamout(struct pipe_stream_output_info
*so
)
2020 if (so
->num_outputs
)
2021 fprintf(stderr
, "STREAMOUT\n");
2023 for (i
= 0; i
< so
->num_outputs
; i
++) {
2024 unsigned mask
= ((1 << so
->output
[i
].num_components
) - 1) <<
2025 so
->output
[i
].start_component
;
2026 fprintf(stderr
, " %i: BUF%i[%i..%i] <- OUT[%i].%s%s%s%s\n",
2027 i
, so
->output
[i
].output_buffer
,
2028 so
->output
[i
].dst_offset
, so
->output
[i
].dst_offset
+ so
->output
[i
].num_components
- 1,
2029 so
->output
[i
].register_index
,
2030 mask
& 1 ? "x" : "",
2031 mask
& 2 ? "y" : "",
2032 mask
& 4 ? "z" : "",
2033 mask
& 8 ? "w" : "");
2037 static void emit_streamout_output(struct si_shader_context
*ctx
,
2038 LLVMValueRef
const *so_buffers
,
2039 LLVMValueRef
const *so_write_offsets
,
2040 struct pipe_stream_output
*stream_out
,
2041 struct si_shader_output_values
*shader_out
)
2043 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
2044 LLVMBuilderRef builder
= gallivm
->builder
;
2045 unsigned buf_idx
= stream_out
->output_buffer
;
2046 unsigned start
= stream_out
->start_component
;
2047 unsigned num_comps
= stream_out
->num_components
;
2048 LLVMValueRef out
[4];
2050 assert(num_comps
&& num_comps
<= 4);
2051 if (!num_comps
|| num_comps
> 4)
2054 /* Load the output as int. */
2055 for (int j
= 0; j
< num_comps
; j
++) {
2056 assert(stream_out
->stream
== shader_out
->vertex_stream
[start
+ j
]);
2058 out
[j
] = LLVMBuildBitCast(builder
,
2059 shader_out
->values
[start
+ j
],
2063 /* Pack the output. */
2064 LLVMValueRef vdata
= NULL
;
2066 switch (num_comps
) {
2067 case 1: /* as i32 */
2070 case 2: /* as v2i32 */
2071 case 3: /* as v4i32 (aligned to 4) */
2072 case 4: /* as v4i32 */
2073 vdata
= LLVMGetUndef(LLVMVectorType(ctx
->i32
, util_next_power_of_two(num_comps
)));
2074 for (int j
= 0; j
< num_comps
; j
++) {
2075 vdata
= LLVMBuildInsertElement(builder
, vdata
, out
[j
],
2076 LLVMConstInt(ctx
->i32
, j
, 0), "");
2081 ac_build_buffer_store_dword(&ctx
->ac
, so_buffers
[buf_idx
],
2083 so_write_offsets
[buf_idx
],
2084 LLVMConstInt(ctx
->i32
, 0, 0),
2085 stream_out
->dst_offset
* 4, 1, 1, true, false);
2089 * Write streamout data to buffers for vertex stream @p stream (different
2090 * vertex streams can occur for GS copy shaders).
2092 static void si_llvm_emit_streamout(struct si_shader_context
*ctx
,
2093 struct si_shader_output_values
*outputs
,
2094 unsigned noutput
, unsigned stream
)
2096 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
2097 struct pipe_stream_output_info
*so
= &sel
->so
;
2098 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
2099 LLVMBuilderRef builder
= gallivm
->builder
;
2101 struct lp_build_if_state if_ctx
;
2103 /* Get bits [22:16], i.e. (so_param >> 16) & 127; */
2104 LLVMValueRef so_vtx_count
=
2105 unpack_param(ctx
, ctx
->param_streamout_config
, 16, 7);
2107 LLVMValueRef tid
= ac_get_thread_id(&ctx
->ac
);
2109 /* can_emit = tid < so_vtx_count; */
2110 LLVMValueRef can_emit
=
2111 LLVMBuildICmp(builder
, LLVMIntULT
, tid
, so_vtx_count
, "");
2113 /* Emit the streamout code conditionally. This actually avoids
2114 * out-of-bounds buffer access. The hw tells us via the SGPR
2115 * (so_vtx_count) which threads are allowed to emit streamout data. */
2116 lp_build_if(&if_ctx
, gallivm
, can_emit
);
2118 /* The buffer offset is computed as follows:
2119 * ByteOffset = streamout_offset[buffer_id]*4 +
2120 * (streamout_write_index + thread_id)*stride[buffer_id] +
2124 LLVMValueRef so_write_index
=
2125 LLVMGetParam(ctx
->main_fn
,
2126 ctx
->param_streamout_write_index
);
2128 /* Compute (streamout_write_index + thread_id). */
2129 so_write_index
= LLVMBuildAdd(builder
, so_write_index
, tid
, "");
2131 /* Load the descriptor and compute the write offset for each
2132 * enabled buffer. */
2133 LLVMValueRef so_write_offset
[4] = {};
2134 LLVMValueRef so_buffers
[4];
2135 LLVMValueRef buf_ptr
= LLVMGetParam(ctx
->main_fn
,
2136 SI_PARAM_RW_BUFFERS
);
2138 for (i
= 0; i
< 4; i
++) {
2142 LLVMValueRef offset
= lp_build_const_int32(gallivm
,
2143 SI_VS_STREAMOUT_BUF0
+ i
);
2145 so_buffers
[i
] = ac_build_indexed_load_const(&ctx
->ac
, buf_ptr
, offset
);
2147 LLVMValueRef so_offset
= LLVMGetParam(ctx
->main_fn
,
2148 ctx
->param_streamout_offset
[i
]);
2149 so_offset
= LLVMBuildMul(builder
, so_offset
, LLVMConstInt(ctx
->i32
, 4, 0), "");
2151 so_write_offset
[i
] = LLVMBuildMul(builder
, so_write_index
,
2152 LLVMConstInt(ctx
->i32
, so
->stride
[i
]*4, 0), "");
2153 so_write_offset
[i
] = LLVMBuildAdd(builder
, so_write_offset
[i
], so_offset
, "");
2156 /* Write streamout data. */
2157 for (i
= 0; i
< so
->num_outputs
; i
++) {
2158 unsigned reg
= so
->output
[i
].register_index
;
2163 if (stream
!= so
->output
[i
].stream
)
2166 emit_streamout_output(ctx
, so_buffers
, so_write_offset
,
2167 &so
->output
[i
], &outputs
[reg
]);
2170 lp_build_endif(&if_ctx
);
2174 /* Generate export instructions for hardware VS shader stage */
2175 static void si_llvm_export_vs(struct lp_build_tgsi_context
*bld_base
,
2176 struct si_shader_output_values
*outputs
,
2179 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2180 struct si_shader
*shader
= ctx
->shader
;
2181 struct lp_build_context
*base
= &bld_base
->base
;
2182 struct ac_export_args args
, pos_args
[4] = {};
2183 LLVMValueRef psize_value
= NULL
, edgeflag_value
= NULL
, layer_value
= NULL
, viewport_index_value
= NULL
;
2184 unsigned semantic_name
, semantic_index
;
2186 unsigned param_count
= 0;
2190 for (i
= 0; i
< noutput
; i
++) {
2191 semantic_name
= outputs
[i
].semantic_name
;
2192 semantic_index
= outputs
[i
].semantic_index
;
2193 bool export_param
= true;
2195 switch (semantic_name
) {
2196 case TGSI_SEMANTIC_POSITION
: /* ignore these */
2197 case TGSI_SEMANTIC_PSIZE
:
2198 case TGSI_SEMANTIC_CLIPVERTEX
:
2199 case TGSI_SEMANTIC_EDGEFLAG
:
2201 case TGSI_SEMANTIC_GENERIC
:
2202 case TGSI_SEMANTIC_CLIPDIST
:
2203 if (shader
->key
.opt
.hw_vs
.kill_outputs
&
2204 (1ull << si_shader_io_get_unique_index(semantic_name
, semantic_index
)))
2205 export_param
= false;
2208 if (shader
->key
.opt
.hw_vs
.kill_outputs2
&
2209 (1u << si_shader_io_get_unique_index2(semantic_name
, semantic_index
)))
2210 export_param
= false;
2214 if (outputs
[i
].vertex_stream
[0] != 0 &&
2215 outputs
[i
].vertex_stream
[1] != 0 &&
2216 outputs
[i
].vertex_stream
[2] != 0 &&
2217 outputs
[i
].vertex_stream
[3] != 0)
2218 export_param
= false;
2221 /* Select the correct target */
2222 switch(semantic_name
) {
2223 case TGSI_SEMANTIC_PSIZE
:
2224 psize_value
= outputs
[i
].values
[0];
2226 case TGSI_SEMANTIC_EDGEFLAG
:
2227 edgeflag_value
= outputs
[i
].values
[0];
2229 case TGSI_SEMANTIC_LAYER
:
2230 layer_value
= outputs
[i
].values
[0];
2231 semantic_name
= TGSI_SEMANTIC_GENERIC
;
2232 goto handle_semantic
;
2233 case TGSI_SEMANTIC_VIEWPORT_INDEX
:
2234 viewport_index_value
= outputs
[i
].values
[0];
2235 semantic_name
= TGSI_SEMANTIC_GENERIC
;
2236 goto handle_semantic
;
2237 case TGSI_SEMANTIC_POSITION
:
2238 target
= V_008DFC_SQ_EXP_POS
;
2240 case TGSI_SEMANTIC_CLIPDIST
:
2241 if (shader
->key
.opt
.hw_vs
.clip_disable
) {
2242 semantic_name
= TGSI_SEMANTIC_GENERIC
;
2243 goto handle_semantic
;
2245 target
= V_008DFC_SQ_EXP_POS
+ 2 + semantic_index
;
2247 case TGSI_SEMANTIC_CLIPVERTEX
:
2248 if (shader
->key
.opt
.hw_vs
.clip_disable
)
2250 si_llvm_emit_clipvertex(bld_base
, pos_args
, outputs
[i
].values
);
2252 case TGSI_SEMANTIC_COLOR
:
2253 case TGSI_SEMANTIC_BCOLOR
:
2254 case TGSI_SEMANTIC_PRIMID
:
2255 case TGSI_SEMANTIC_FOG
:
2256 case TGSI_SEMANTIC_TEXCOORD
:
2257 case TGSI_SEMANTIC_GENERIC
:
2260 target
= V_008DFC_SQ_EXP_PARAM
+ param_count
;
2261 assert(i
< ARRAY_SIZE(shader
->info
.vs_output_param_offset
));
2262 shader
->info
.vs_output_param_offset
[i
] = param_count
;
2268 "Warning: SI unhandled vs output type:%d\n",
2272 si_llvm_init_export_args(bld_base
, outputs
[i
].values
, target
, &args
);
2274 if (target
>= V_008DFC_SQ_EXP_POS
&&
2275 target
<= (V_008DFC_SQ_EXP_POS
+ 3)) {
2276 memcpy(&pos_args
[target
- V_008DFC_SQ_EXP_POS
],
2277 &args
, sizeof(args
));
2279 ac_build_export(&ctx
->ac
, &args
);
2282 if (semantic_name
== TGSI_SEMANTIC_CLIPDIST
) {
2283 semantic_name
= TGSI_SEMANTIC_GENERIC
;
2284 goto handle_semantic
;
2288 shader
->info
.nr_param_exports
= param_count
;
2290 /* We need to add the position output manually if it's missing. */
2291 if (!pos_args
[0].out
[0]) {
2292 pos_args
[0].enabled_channels
= 0xf; /* writemask */
2293 pos_args
[0].valid_mask
= 0; /* EXEC mask */
2294 pos_args
[0].done
= 0; /* last export? */
2295 pos_args
[0].target
= V_008DFC_SQ_EXP_POS
;
2296 pos_args
[0].compr
= 0; /* COMPR flag */
2297 pos_args
[0].out
[0] = base
->zero
; /* X */
2298 pos_args
[0].out
[1] = base
->zero
; /* Y */
2299 pos_args
[0].out
[2] = base
->zero
; /* Z */
2300 pos_args
[0].out
[3] = base
->one
; /* W */
2303 /* Write the misc vector (point size, edgeflag, layer, viewport). */
2304 if (shader
->selector
->info
.writes_psize
||
2305 shader
->selector
->info
.writes_edgeflag
||
2306 shader
->selector
->info
.writes_viewport_index
||
2307 shader
->selector
->info
.writes_layer
) {
2308 pos_args
[1].enabled_channels
= shader
->selector
->info
.writes_psize
|
2309 (shader
->selector
->info
.writes_edgeflag
<< 1) |
2310 (shader
->selector
->info
.writes_layer
<< 2) |
2311 (shader
->selector
->info
.writes_viewport_index
<< 3);
2312 pos_args
[1].valid_mask
= 0; /* EXEC mask */
2313 pos_args
[1].done
= 0; /* last export? */
2314 pos_args
[1].target
= V_008DFC_SQ_EXP_POS
+ 1;
2315 pos_args
[1].compr
= 0; /* COMPR flag */
2316 pos_args
[1].out
[0] = base
->zero
; /* X */
2317 pos_args
[1].out
[1] = base
->zero
; /* Y */
2318 pos_args
[1].out
[2] = base
->zero
; /* Z */
2319 pos_args
[1].out
[3] = base
->zero
; /* W */
2321 if (shader
->selector
->info
.writes_psize
)
2322 pos_args
[1].out
[0] = psize_value
;
2324 if (shader
->selector
->info
.writes_edgeflag
) {
2325 /* The output is a float, but the hw expects an integer
2326 * with the first bit containing the edge flag. */
2327 edgeflag_value
= LLVMBuildFPToUI(base
->gallivm
->builder
,
2330 edgeflag_value
= lp_build_min(&bld_base
->int_bld
,
2332 bld_base
->int_bld
.one
);
2334 /* The LLVM intrinsic expects a float. */
2335 pos_args
[1].out
[1] = LLVMBuildBitCast(base
->gallivm
->builder
,
2340 if (shader
->selector
->info
.writes_layer
)
2341 pos_args
[1].out
[2] = layer_value
;
2343 if (shader
->selector
->info
.writes_viewport_index
)
2344 pos_args
[1].out
[3] = viewport_index_value
;
2347 for (i
= 0; i
< 4; i
++)
2348 if (pos_args
[i
].out
[0])
2349 shader
->info
.nr_pos_exports
++;
2352 for (i
= 0; i
< 4; i
++) {
2353 if (!pos_args
[i
].out
[0])
2356 /* Specify the target we are exporting */
2357 pos_args
[i
].target
= V_008DFC_SQ_EXP_POS
+ pos_idx
++;
2359 if (pos_idx
== shader
->info
.nr_pos_exports
)
2360 /* Specify that this is the last export */
2361 pos_args
[i
].done
= 1;
2363 ac_build_export(&ctx
->ac
, &pos_args
[i
]);
2368 * Forward all outputs from the vertex shader to the TES. This is only used
2369 * for the fixed function TCS.
2371 static void si_copy_tcs_inputs(struct lp_build_tgsi_context
*bld_base
)
2373 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2374 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2375 LLVMValueRef invocation_id
, rw_buffers
, buffer
, buffer_offset
;
2376 LLVMValueRef lds_vertex_stride
, lds_vertex_offset
, lds_base
;
2379 invocation_id
= unpack_param(ctx
, SI_PARAM_REL_IDS
, 8, 5);
2381 rw_buffers
= LLVMGetParam(ctx
->main_fn
, SI_PARAM_RW_BUFFERS
);
2382 buffer
= ac_build_indexed_load_const(&ctx
->ac
, rw_buffers
,
2383 lp_build_const_int32(gallivm
, SI_HS_RING_TESS_OFFCHIP
));
2385 buffer_offset
= LLVMGetParam(ctx
->main_fn
, ctx
->param_oc_lds
);
2387 lds_vertex_stride
= unpack_param(ctx
, SI_PARAM_TCS_IN_LAYOUT
, 13, 8);
2388 lds_vertex_offset
= LLVMBuildMul(gallivm
->builder
, invocation_id
,
2389 lds_vertex_stride
, "");
2390 lds_base
= get_tcs_in_current_patch_offset(ctx
);
2391 lds_base
= LLVMBuildAdd(gallivm
->builder
, lds_base
, lds_vertex_offset
, "");
2393 inputs
= ctx
->shader
->key
.mono
.tcs
.inputs_to_copy
;
2395 unsigned i
= u_bit_scan64(&inputs
);
2397 LLVMValueRef lds_ptr
= LLVMBuildAdd(gallivm
->builder
, lds_base
,
2398 lp_build_const_int32(gallivm
, 4 * i
),
2401 LLVMValueRef buffer_addr
= get_tcs_tes_buffer_address(ctx
,
2402 get_rel_patch_id(ctx
),
2404 lp_build_const_int32(gallivm
, i
));
2406 LLVMValueRef value
= lds_load(bld_base
, TGSI_TYPE_SIGNED
, ~0,
2409 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, value
, 4, buffer_addr
,
2410 buffer_offset
, 0, 1, 0, true, false);
2414 static void si_write_tess_factors(struct lp_build_tgsi_context
*bld_base
,
2415 LLVMValueRef rel_patch_id
,
2416 LLVMValueRef invocation_id
,
2417 LLVMValueRef tcs_out_current_patch_data_offset
)
2419 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2420 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2421 struct si_shader
*shader
= ctx
->shader
;
2422 unsigned tess_inner_index
, tess_outer_index
;
2423 LLVMValueRef lds_base
, lds_inner
, lds_outer
, byteoffset
, buffer
;
2424 LLVMValueRef out
[6], vec0
, vec1
, rw_buffers
, tf_base
, inner
[4], outer
[4];
2425 unsigned stride
, outer_comps
, inner_comps
, i
;
2426 struct lp_build_if_state if_ctx
, inner_if_ctx
;
2428 si_llvm_emit_barrier(NULL
, bld_base
, NULL
);
2430 /* Do this only for invocation 0, because the tess levels are per-patch,
2433 * This can't jump, because invocation 0 executes this. It should
2434 * at least mask out the loads and stores for other invocations.
2436 lp_build_if(&if_ctx
, gallivm
,
2437 LLVMBuildICmp(gallivm
->builder
, LLVMIntEQ
,
2438 invocation_id
, bld_base
->uint_bld
.zero
, ""));
2440 /* Determine the layout of one tess factor element in the buffer. */
2441 switch (shader
->key
.part
.tcs
.epilog
.prim_mode
) {
2442 case PIPE_PRIM_LINES
:
2443 stride
= 2; /* 2 dwords, 1 vec2 store */
2447 case PIPE_PRIM_TRIANGLES
:
2448 stride
= 4; /* 4 dwords, 1 vec4 store */
2452 case PIPE_PRIM_QUADS
:
2453 stride
= 6; /* 6 dwords, 2 stores (vec4 + vec2) */
2462 /* Load tess_inner and tess_outer from LDS.
2463 * Any invocation can write them, so we can't get them from a temporary.
2465 tess_inner_index
= si_shader_io_get_unique_index(TGSI_SEMANTIC_TESSINNER
, 0);
2466 tess_outer_index
= si_shader_io_get_unique_index(TGSI_SEMANTIC_TESSOUTER
, 0);
2468 lds_base
= tcs_out_current_patch_data_offset
;
2469 lds_inner
= LLVMBuildAdd(gallivm
->builder
, lds_base
,
2470 lp_build_const_int32(gallivm
,
2471 tess_inner_index
* 4), "");
2472 lds_outer
= LLVMBuildAdd(gallivm
->builder
, lds_base
,
2473 lp_build_const_int32(gallivm
,
2474 tess_outer_index
* 4), "");
2476 for (i
= 0; i
< 4; i
++) {
2477 inner
[i
] = LLVMGetUndef(ctx
->i32
);
2478 outer
[i
] = LLVMGetUndef(ctx
->i32
);
2481 if (shader
->key
.part
.tcs
.epilog
.prim_mode
== PIPE_PRIM_LINES
) {
2482 /* For isolines, the hardware expects tess factors in the
2483 * reverse order from what GLSL / TGSI specify.
2485 outer
[0] = out
[1] = lds_load(bld_base
, TGSI_TYPE_SIGNED
, 0, lds_outer
);
2486 outer
[1] = out
[0] = lds_load(bld_base
, TGSI_TYPE_SIGNED
, 1, lds_outer
);
2488 for (i
= 0; i
< outer_comps
; i
++) {
2490 lds_load(bld_base
, TGSI_TYPE_SIGNED
, i
, lds_outer
);
2492 for (i
= 0; i
< inner_comps
; i
++) {
2493 inner
[i
] = out
[outer_comps
+i
] =
2494 lds_load(bld_base
, TGSI_TYPE_SIGNED
, i
, lds_inner
);
2498 /* Convert the outputs to vectors for stores. */
2499 vec0
= lp_build_gather_values(gallivm
, out
, MIN2(stride
, 4));
2503 vec1
= lp_build_gather_values(gallivm
, out
+4, stride
- 4);
2505 /* Get the buffer. */
2506 rw_buffers
= LLVMGetParam(ctx
->main_fn
,
2507 SI_PARAM_RW_BUFFERS
);
2508 buffer
= ac_build_indexed_load_const(&ctx
->ac
, rw_buffers
,
2509 lp_build_const_int32(gallivm
, SI_HS_RING_TESS_FACTOR
));
2511 /* Get the offset. */
2512 tf_base
= LLVMGetParam(ctx
->main_fn
,
2513 SI_PARAM_TESS_FACTOR_OFFSET
);
2514 byteoffset
= LLVMBuildMul(gallivm
->builder
, rel_patch_id
,
2515 lp_build_const_int32(gallivm
, 4 * stride
), "");
2517 lp_build_if(&inner_if_ctx
, gallivm
,
2518 LLVMBuildICmp(gallivm
->builder
, LLVMIntEQ
,
2519 rel_patch_id
, bld_base
->uint_bld
.zero
, ""));
2521 /* Store the dynamic HS control word. */
2522 ac_build_buffer_store_dword(&ctx
->ac
, buffer
,
2523 lp_build_const_int32(gallivm
, 0x80000000),
2524 1, lp_build_const_int32(gallivm
, 0), tf_base
,
2525 0, 1, 0, true, false);
2527 lp_build_endif(&inner_if_ctx
);
2529 /* Store the tessellation factors. */
2530 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec0
,
2531 MIN2(stride
, 4), byteoffset
, tf_base
,
2532 4, 1, 0, true, false);
2534 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec1
,
2535 stride
- 4, byteoffset
, tf_base
,
2536 20, 1, 0, true, false);
2538 /* Store the tess factors into the offchip buffer if TES reads them. */
2539 if (shader
->key
.part
.tcs
.epilog
.tes_reads_tess_factors
) {
2540 LLVMValueRef buf
, base
, inner_vec
, outer_vec
, tf_outer_offset
;
2541 LLVMValueRef tf_inner_offset
;
2542 unsigned param_outer
, param_inner
;
2544 buf
= ac_build_indexed_load_const(&ctx
->ac
, rw_buffers
,
2545 LLVMConstInt(ctx
->i32
, SI_HS_RING_TESS_OFFCHIP
, 0));
2546 base
= LLVMGetParam(ctx
->main_fn
, ctx
->param_oc_lds
);
2548 param_outer
= si_shader_io_get_unique_index(
2549 TGSI_SEMANTIC_TESSOUTER
, 0);
2550 tf_outer_offset
= get_tcs_tes_buffer_address(ctx
, rel_patch_id
, NULL
,
2551 LLVMConstInt(ctx
->i32
, param_outer
, 0));
2553 outer_vec
= lp_build_gather_values(gallivm
, outer
,
2554 util_next_power_of_two(outer_comps
));
2556 ac_build_buffer_store_dword(&ctx
->ac
, buf
, outer_vec
,
2557 outer_comps
, tf_outer_offset
,
2558 base
, 0, 1, 0, true, false);
2560 param_inner
= si_shader_io_get_unique_index(
2561 TGSI_SEMANTIC_TESSINNER
, 0);
2562 tf_inner_offset
= get_tcs_tes_buffer_address(ctx
, rel_patch_id
, NULL
,
2563 LLVMConstInt(ctx
->i32
, param_inner
, 0));
2565 inner_vec
= inner_comps
== 1 ? inner
[0] :
2566 lp_build_gather_values(gallivm
, inner
, inner_comps
);
2567 ac_build_buffer_store_dword(&ctx
->ac
, buf
, inner_vec
,
2568 inner_comps
, tf_inner_offset
,
2569 base
, 0, 1, 0, true, false);
2573 lp_build_endif(&if_ctx
);
2576 /* This only writes the tessellation factor levels. */
2577 static void si_llvm_emit_tcs_epilogue(struct lp_build_tgsi_context
*bld_base
)
2579 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2580 LLVMValueRef rel_patch_id
, invocation_id
, tf_lds_offset
;
2581 LLVMValueRef offchip_soffset
, offchip_layout
;
2583 si_copy_tcs_inputs(bld_base
);
2585 rel_patch_id
= get_rel_patch_id(ctx
);
2586 invocation_id
= unpack_param(ctx
, SI_PARAM_REL_IDS
, 8, 5);
2587 tf_lds_offset
= get_tcs_out_current_patch_data_offset(ctx
);
2589 /* Return epilog parameters from this function. */
2590 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
2591 LLVMValueRef ret
= ctx
->return_value
;
2592 LLVMValueRef rw_buffers
, rw0
, rw1
, tf_soffset
;
2595 /* RW_BUFFERS pointer */
2596 rw_buffers
= LLVMGetParam(ctx
->main_fn
,
2597 SI_PARAM_RW_BUFFERS
);
2598 rw_buffers
= LLVMBuildPtrToInt(builder
, rw_buffers
, ctx
->i64
, "");
2599 rw_buffers
= LLVMBuildBitCast(builder
, rw_buffers
, ctx
->v2i32
, "");
2600 rw0
= LLVMBuildExtractElement(builder
, rw_buffers
,
2601 bld_base
->uint_bld
.zero
, "");
2602 rw1
= LLVMBuildExtractElement(builder
, rw_buffers
,
2603 bld_base
->uint_bld
.one
, "");
2604 ret
= LLVMBuildInsertValue(builder
, ret
, rw0
, 0, "");
2605 ret
= LLVMBuildInsertValue(builder
, ret
, rw1
, 1, "");
2607 /* Tess offchip and factor buffer soffset are after user SGPRs. */
2608 offchip_layout
= LLVMGetParam(ctx
->main_fn
,
2609 SI_PARAM_TCS_OFFCHIP_LAYOUT
);
2610 offchip_soffset
= LLVMGetParam(ctx
->main_fn
, ctx
->param_oc_lds
);
2611 tf_soffset
= LLVMGetParam(ctx
->main_fn
,
2612 SI_PARAM_TESS_FACTOR_OFFSET
);
2613 ret
= LLVMBuildInsertValue(builder
, ret
, offchip_layout
,
2614 SI_SGPR_TCS_OFFCHIP_LAYOUT
, "");
2615 ret
= LLVMBuildInsertValue(builder
, ret
, offchip_soffset
,
2616 SI_TCS_NUM_USER_SGPR
, "");
2617 ret
= LLVMBuildInsertValue(builder
, ret
, tf_soffset
,
2618 SI_TCS_NUM_USER_SGPR
+ 1, "");
2621 rel_patch_id
= bitcast(bld_base
, TGSI_TYPE_FLOAT
, rel_patch_id
);
2622 invocation_id
= bitcast(bld_base
, TGSI_TYPE_FLOAT
, invocation_id
);
2623 tf_lds_offset
= bitcast(bld_base
, TGSI_TYPE_FLOAT
, tf_lds_offset
);
2625 vgpr
= SI_TCS_NUM_USER_SGPR
+ 2;
2626 ret
= LLVMBuildInsertValue(builder
, ret
, rel_patch_id
, vgpr
++, "");
2627 ret
= LLVMBuildInsertValue(builder
, ret
, invocation_id
, vgpr
++, "");
2628 ret
= LLVMBuildInsertValue(builder
, ret
, tf_lds_offset
, vgpr
++, "");
2629 ctx
->return_value
= ret
;
2632 static void si_llvm_emit_ls_epilogue(struct lp_build_tgsi_context
*bld_base
)
2634 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2635 struct si_shader
*shader
= ctx
->shader
;
2636 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
2637 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2639 LLVMValueRef vertex_id
= LLVMGetParam(ctx
->main_fn
,
2640 ctx
->param_rel_auto_id
);
2641 LLVMValueRef vertex_dw_stride
=
2642 unpack_param(ctx
, SI_PARAM_LS_OUT_LAYOUT
, 13, 8);
2643 LLVMValueRef base_dw_addr
= LLVMBuildMul(gallivm
->builder
, vertex_id
,
2644 vertex_dw_stride
, "");
2646 /* Write outputs to LDS. The next shader (TCS aka HS) will read
2647 * its inputs from it. */
2648 for (i
= 0; i
< info
->num_outputs
; i
++) {
2649 LLVMValueRef
*out_ptr
= ctx
->outputs
[i
];
2650 unsigned name
= info
->output_semantic_name
[i
];
2651 unsigned index
= info
->output_semantic_index
[i
];
2652 int param
= si_shader_io_get_unique_index(name
, index
);
2653 LLVMValueRef dw_addr
= LLVMBuildAdd(gallivm
->builder
, base_dw_addr
,
2654 lp_build_const_int32(gallivm
, param
* 4), "");
2656 for (chan
= 0; chan
< 4; chan
++) {
2657 lds_store(bld_base
, chan
, dw_addr
,
2658 LLVMBuildLoad(gallivm
->builder
, out_ptr
[chan
], ""));
2663 static void si_llvm_emit_es_epilogue(struct lp_build_tgsi_context
*bld_base
)
2665 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2666 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2667 struct si_shader
*es
= ctx
->shader
;
2668 struct tgsi_shader_info
*info
= &es
->selector
->info
;
2669 LLVMValueRef soffset
= LLVMGetParam(ctx
->main_fn
,
2670 ctx
->param_es2gs_offset
);
2674 for (i
= 0; i
< info
->num_outputs
; i
++) {
2675 LLVMValueRef
*out_ptr
= ctx
->outputs
[i
];
2678 if (info
->output_semantic_name
[i
] == TGSI_SEMANTIC_VIEWPORT_INDEX
||
2679 info
->output_semantic_name
[i
] == TGSI_SEMANTIC_LAYER
)
2682 param_index
= si_shader_io_get_unique_index(info
->output_semantic_name
[i
],
2683 info
->output_semantic_index
[i
]);
2685 for (chan
= 0; chan
< 4; chan
++) {
2686 LLVMValueRef out_val
= LLVMBuildLoad(gallivm
->builder
, out_ptr
[chan
], "");
2687 out_val
= LLVMBuildBitCast(gallivm
->builder
, out_val
, ctx
->i32
, "");
2689 ac_build_buffer_store_dword(&ctx
->ac
,
2691 out_val
, 1, NULL
, soffset
,
2692 (4 * param_index
+ chan
) * 4,
2698 static void si_llvm_emit_gs_epilogue(struct lp_build_tgsi_context
*bld_base
)
2700 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2702 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_NOP
| AC_SENDMSG_GS_DONE
,
2703 LLVMGetParam(ctx
->main_fn
, SI_PARAM_GS_WAVE_ID
));
2706 static void si_llvm_emit_vs_epilogue(struct lp_build_tgsi_context
*bld_base
)
2708 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2709 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2710 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
2711 struct si_shader_output_values
*outputs
= NULL
;
2714 assert(!ctx
->shader
->is_gs_copy_shader
);
2716 outputs
= MALLOC((info
->num_outputs
+ 1) * sizeof(outputs
[0]));
2718 /* Vertex color clamping.
2720 * This uses a state constant loaded in a user data SGPR and
2721 * an IF statement is added that clamps all colors if the constant
2724 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
2725 struct lp_build_if_state if_ctx
;
2726 LLVMValueRef cond
= NULL
;
2727 LLVMValueRef addr
, val
;
2729 for (i
= 0; i
< info
->num_outputs
; i
++) {
2730 if (info
->output_semantic_name
[i
] != TGSI_SEMANTIC_COLOR
&&
2731 info
->output_semantic_name
[i
] != TGSI_SEMANTIC_BCOLOR
)
2734 /* We've found a color. */
2736 /* The state is in the first bit of the user SGPR. */
2737 cond
= LLVMGetParam(ctx
->main_fn
,
2738 SI_PARAM_VS_STATE_BITS
);
2739 cond
= LLVMBuildTrunc(gallivm
->builder
, cond
,
2741 lp_build_if(&if_ctx
, gallivm
, cond
);
2744 for (j
= 0; j
< 4; j
++) {
2745 addr
= ctx
->outputs
[i
][j
];
2746 val
= LLVMBuildLoad(gallivm
->builder
, addr
, "");
2747 val
= ac_build_clamp(&ctx
->ac
, val
);
2748 LLVMBuildStore(gallivm
->builder
, val
, addr
);
2753 lp_build_endif(&if_ctx
);
2756 for (i
= 0; i
< info
->num_outputs
; i
++) {
2757 outputs
[i
].semantic_name
= info
->output_semantic_name
[i
];
2758 outputs
[i
].semantic_index
= info
->output_semantic_index
[i
];
2760 for (j
= 0; j
< 4; j
++) {
2761 outputs
[i
].values
[j
] =
2762 LLVMBuildLoad(gallivm
->builder
,
2765 outputs
[i
].vertex_stream
[j
] =
2766 (info
->output_streams
[i
] >> (2 * j
)) & 3;
2771 /* Return the primitive ID from the LLVM function. */
2773 LLVMBuildInsertValue(gallivm
->builder
,
2775 bitcast(bld_base
, TGSI_TYPE_FLOAT
,
2776 get_primitive_id(bld_base
, 0)),
2777 VS_EPILOG_PRIMID_LOC
, "");
2779 if (ctx
->shader
->selector
->so
.num_outputs
)
2780 si_llvm_emit_streamout(ctx
, outputs
, i
, 0);
2781 si_llvm_export_vs(bld_base
, outputs
, i
);
2785 struct si_ps_exports
{
2787 struct ac_export_args args
[10];
2790 unsigned si_get_spi_shader_z_format(bool writes_z
, bool writes_stencil
,
2791 bool writes_samplemask
)
2794 /* Z needs 32 bits. */
2795 if (writes_samplemask
)
2796 return V_028710_SPI_SHADER_32_ABGR
;
2797 else if (writes_stencil
)
2798 return V_028710_SPI_SHADER_32_GR
;
2800 return V_028710_SPI_SHADER_32_R
;
2801 } else if (writes_stencil
|| writes_samplemask
) {
2802 /* Both stencil and sample mask need only 16 bits. */
2803 return V_028710_SPI_SHADER_UINT16_ABGR
;
2805 return V_028710_SPI_SHADER_ZERO
;
2809 static void si_export_mrt_z(struct lp_build_tgsi_context
*bld_base
,
2810 LLVMValueRef depth
, LLVMValueRef stencil
,
2811 LLVMValueRef samplemask
, struct si_ps_exports
*exp
)
2813 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2814 struct lp_build_context
*base
= &bld_base
->base
;
2815 struct ac_export_args args
;
2817 unsigned format
= si_get_spi_shader_z_format(depth
!= NULL
,
2819 samplemask
!= NULL
);
2821 assert(depth
|| stencil
|| samplemask
);
2823 args
.valid_mask
= 1; /* whether the EXEC mask is valid */
2824 args
.done
= 1; /* DONE bit */
2826 /* Specify the target we are exporting */
2827 args
.target
= V_008DFC_SQ_EXP_MRTZ
;
2829 args
.compr
= 0; /* COMP flag */
2830 args
.out
[0] = base
->undef
; /* R, depth */
2831 args
.out
[1] = base
->undef
; /* G, stencil test value[0:7], stencil op value[8:15] */
2832 args
.out
[2] = base
->undef
; /* B, sample mask */
2833 args
.out
[3] = base
->undef
; /* A, alpha to mask */
2835 if (format
== V_028710_SPI_SHADER_UINT16_ABGR
) {
2837 args
.compr
= 1; /* COMPR flag */
2840 /* Stencil should be in X[23:16]. */
2841 stencil
= bitcast(bld_base
, TGSI_TYPE_UNSIGNED
, stencil
);
2842 stencil
= LLVMBuildShl(base
->gallivm
->builder
, stencil
,
2843 LLVMConstInt(ctx
->i32
, 16, 0), "");
2844 args
.out
[0] = bitcast(bld_base
, TGSI_TYPE_FLOAT
, stencil
);
2848 /* SampleMask should be in Y[15:0]. */
2849 args
.out
[1] = samplemask
;
2854 args
.out
[0] = depth
;
2858 args
.out
[1] = stencil
;
2862 args
.out
[2] = samplemask
;
2867 /* SI (except OLAND and HAINAN) has a bug that it only looks
2868 * at the X writemask component. */
2869 if (ctx
->screen
->b
.chip_class
== SI
&&
2870 ctx
->screen
->b
.family
!= CHIP_OLAND
&&
2871 ctx
->screen
->b
.family
!= CHIP_HAINAN
)
2874 /* Specify which components to enable */
2875 args
.enabled_channels
= mask
;
2877 memcpy(&exp
->args
[exp
->num
++], &args
, sizeof(args
));
2880 static void si_export_mrt_color(struct lp_build_tgsi_context
*bld_base
,
2881 LLVMValueRef
*color
, unsigned index
,
2882 unsigned samplemask_param
,
2883 bool is_last
, struct si_ps_exports
*exp
)
2885 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2886 struct lp_build_context
*base
= &bld_base
->base
;
2890 if (ctx
->shader
->key
.part
.ps
.epilog
.clamp_color
)
2891 for (i
= 0; i
< 4; i
++)
2892 color
[i
] = ac_build_clamp(&ctx
->ac
, color
[i
]);
2895 if (ctx
->shader
->key
.part
.ps
.epilog
.alpha_to_one
)
2896 color
[3] = base
->one
;
2900 ctx
->shader
->key
.part
.ps
.epilog
.alpha_func
!= PIPE_FUNC_ALWAYS
)
2901 si_alpha_test(bld_base
, color
[3]);
2903 /* Line & polygon smoothing */
2904 if (ctx
->shader
->key
.part
.ps
.epilog
.poly_line_smoothing
)
2905 color
[3] = si_scale_alpha_by_sample_mask(bld_base
, color
[3],
2908 /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
2909 if (ctx
->shader
->key
.part
.ps
.epilog
.last_cbuf
> 0) {
2910 struct ac_export_args args
[8];
2913 /* Get the export arguments, also find out what the last one is. */
2914 for (c
= 0; c
<= ctx
->shader
->key
.part
.ps
.epilog
.last_cbuf
; c
++) {
2915 si_llvm_init_export_args(bld_base
, color
,
2916 V_008DFC_SQ_EXP_MRT
+ c
, &args
[c
]);
2917 if (args
[c
].enabled_channels
)
2921 /* Emit all exports. */
2922 for (c
= 0; c
<= ctx
->shader
->key
.part
.ps
.epilog
.last_cbuf
; c
++) {
2923 if (is_last
&& last
== c
) {
2924 args
[c
].valid_mask
= 1; /* whether the EXEC mask is valid */
2925 args
[c
].done
= 1; /* DONE bit */
2926 } else if (!args
[c
].enabled_channels
)
2927 continue; /* unnecessary NULL export */
2929 memcpy(&exp
->args
[exp
->num
++], &args
[c
], sizeof(args
[c
]));
2932 struct ac_export_args args
;
2935 si_llvm_init_export_args(bld_base
, color
, V_008DFC_SQ_EXP_MRT
+ index
,
2938 args
.valid_mask
= 1; /* whether the EXEC mask is valid */
2939 args
.done
= 1; /* DONE bit */
2940 } else if (!args
.enabled_channels
)
2941 return; /* unnecessary NULL export */
2943 memcpy(&exp
->args
[exp
->num
++], &args
, sizeof(args
));
2947 static void si_emit_ps_exports(struct si_shader_context
*ctx
,
2948 struct si_ps_exports
*exp
)
2950 for (unsigned i
= 0; i
< exp
->num
; i
++)
2951 ac_build_export(&ctx
->ac
, &exp
->args
[i
]);
2954 static void si_export_null(struct lp_build_tgsi_context
*bld_base
)
2956 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2957 struct lp_build_context
*base
= &bld_base
->base
;
2958 struct ac_export_args args
;
2960 args
.enabled_channels
= 0x0; /* enabled channels */
2961 args
.valid_mask
= 1; /* whether the EXEC mask is valid */
2962 args
.done
= 1; /* DONE bit */
2963 args
.target
= V_008DFC_SQ_EXP_NULL
;
2964 args
.compr
= 0; /* COMPR flag (0 = 32-bit export) */
2965 args
.out
[0] = base
->undef
; /* R */
2966 args
.out
[1] = base
->undef
; /* G */
2967 args
.out
[2] = base
->undef
; /* B */
2968 args
.out
[3] = base
->undef
; /* A */
2970 ac_build_export(&ctx
->ac
, &args
);
2974 * Return PS outputs in this order:
2976 * v[0:3] = color0.xyzw
2977 * v[4:7] = color1.xyzw
2982 * vN+3 = SampleMaskIn (used for OpenGL smoothing)
2984 * The alpha-ref SGPR is returned via its original location.
2986 static void si_llvm_return_fs_outputs(struct lp_build_tgsi_context
*bld_base
)
2988 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2989 struct si_shader
*shader
= ctx
->shader
;
2990 struct lp_build_context
*base
= &bld_base
->base
;
2991 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
2992 LLVMBuilderRef builder
= base
->gallivm
->builder
;
2993 unsigned i
, j
, first_vgpr
, vgpr
;
2995 LLVMValueRef color
[8][4] = {};
2996 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
2999 /* Read the output values. */
3000 for (i
= 0; i
< info
->num_outputs
; i
++) {
3001 unsigned semantic_name
= info
->output_semantic_name
[i
];
3002 unsigned semantic_index
= info
->output_semantic_index
[i
];
3004 switch (semantic_name
) {
3005 case TGSI_SEMANTIC_COLOR
:
3006 assert(semantic_index
< 8);
3007 for (j
= 0; j
< 4; j
++) {
3008 LLVMValueRef ptr
= ctx
->outputs
[i
][j
];
3009 LLVMValueRef result
= LLVMBuildLoad(builder
, ptr
, "");
3010 color
[semantic_index
][j
] = result
;
3013 case TGSI_SEMANTIC_POSITION
:
3014 depth
= LLVMBuildLoad(builder
,
3015 ctx
->outputs
[i
][2], "");
3017 case TGSI_SEMANTIC_STENCIL
:
3018 stencil
= LLVMBuildLoad(builder
,
3019 ctx
->outputs
[i
][1], "");
3021 case TGSI_SEMANTIC_SAMPLEMASK
:
3022 samplemask
= LLVMBuildLoad(builder
,
3023 ctx
->outputs
[i
][0], "");
3026 fprintf(stderr
, "Warning: SI unhandled fs output type:%d\n",
3031 /* Fill the return structure. */
3032 ret
= ctx
->return_value
;
3035 ret
= LLVMBuildInsertValue(builder
, ret
,
3036 bitcast(bld_base
, TGSI_TYPE_SIGNED
,
3037 LLVMGetParam(ctx
->main_fn
,
3038 SI_PARAM_ALPHA_REF
)),
3039 SI_SGPR_ALPHA_REF
, "");
3042 first_vgpr
= vgpr
= SI_SGPR_ALPHA_REF
+ 1;
3043 for (i
= 0; i
< ARRAY_SIZE(color
); i
++) {
3047 for (j
= 0; j
< 4; j
++)
3048 ret
= LLVMBuildInsertValue(builder
, ret
, color
[i
][j
], vgpr
++, "");
3051 ret
= LLVMBuildInsertValue(builder
, ret
, depth
, vgpr
++, "");
3053 ret
= LLVMBuildInsertValue(builder
, ret
, stencil
, vgpr
++, "");
3055 ret
= LLVMBuildInsertValue(builder
, ret
, samplemask
, vgpr
++, "");
3057 /* Add the input sample mask for smoothing at the end. */
3058 if (vgpr
< first_vgpr
+ PS_EPILOG_SAMPLEMASK_MIN_LOC
)
3059 vgpr
= first_vgpr
+ PS_EPILOG_SAMPLEMASK_MIN_LOC
;
3060 ret
= LLVMBuildInsertValue(builder
, ret
,
3061 LLVMGetParam(ctx
->main_fn
,
3062 SI_PARAM_SAMPLE_COVERAGE
), vgpr
++, "");
3064 ctx
->return_value
= ret
;
3068 * Given a v8i32 resource descriptor for a buffer, extract the size of the
3069 * buffer in number of elements and return it as an i32.
3071 static LLVMValueRef
get_buffer_size(
3072 struct lp_build_tgsi_context
*bld_base
,
3073 LLVMValueRef descriptor
)
3075 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3076 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3077 LLVMBuilderRef builder
= gallivm
->builder
;
3079 LLVMBuildExtractElement(builder
, descriptor
,
3080 lp_build_const_int32(gallivm
, 2), "");
3082 if (ctx
->screen
->b
.chip_class
>= VI
) {
3083 /* On VI, the descriptor contains the size in bytes,
3084 * but TXQ must return the size in elements.
3085 * The stride is always non-zero for resources using TXQ.
3087 LLVMValueRef stride
=
3088 LLVMBuildExtractElement(builder
, descriptor
,
3089 lp_build_const_int32(gallivm
, 1), "");
3090 stride
= LLVMBuildLShr(builder
, stride
,
3091 lp_build_const_int32(gallivm
, 16), "");
3092 stride
= LLVMBuildAnd(builder
, stride
,
3093 lp_build_const_int32(gallivm
, 0x3FFF), "");
3095 size
= LLVMBuildUDiv(builder
, size
, stride
, "");
3101 static void build_tex_intrinsic(const struct lp_build_tgsi_action
*action
,
3102 struct lp_build_tgsi_context
*bld_base
,
3103 struct lp_build_emit_data
*emit_data
);
3105 /* Prevent optimizations (at least of memory accesses) across the current
3106 * point in the program by emitting empty inline assembly that is marked as
3107 * having side effects.
3109 #if 0 /* unused currently */
3110 static void emit_optimization_barrier(struct si_shader_context
*ctx
)
3112 LLVMBuilderRef builder
= ctx
->gallivm
.builder
;
3113 LLVMTypeRef ftype
= LLVMFunctionType(ctx
->voidt
, NULL
, 0, false);
3114 LLVMValueRef inlineasm
= LLVMConstInlineAsm(ftype
, "", "", true, false);
3115 LLVMBuildCall(builder
, inlineasm
, NULL
, 0, "");
3119 /* Combine these with & instead of |. */
3120 #define NOOP_WAITCNT 0xf7f
3121 #define LGKM_CNT 0x07f
3122 #define VM_CNT 0xf70
3124 static void emit_waitcnt(struct si_shader_context
*ctx
, unsigned simm16
)
3126 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
3127 LLVMBuilderRef builder
= gallivm
->builder
;
3128 LLVMValueRef args
[1] = {
3129 lp_build_const_int32(gallivm
, simm16
)
3131 lp_build_intrinsic(builder
, "llvm.amdgcn.s.waitcnt",
3132 ctx
->voidt
, args
, 1, 0);
3135 static void membar_emit(
3136 const struct lp_build_tgsi_action
*action
,
3137 struct lp_build_tgsi_context
*bld_base
,
3138 struct lp_build_emit_data
*emit_data
)
3140 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3141 LLVMValueRef src0
= lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, 0);
3142 unsigned flags
= LLVMConstIntGetZExtValue(src0
);
3143 unsigned waitcnt
= NOOP_WAITCNT
;
3145 if (flags
& TGSI_MEMBAR_THREAD_GROUP
)
3146 waitcnt
&= VM_CNT
& LGKM_CNT
;
3148 if (flags
& (TGSI_MEMBAR_ATOMIC_BUFFER
|
3149 TGSI_MEMBAR_SHADER_BUFFER
|
3150 TGSI_MEMBAR_SHADER_IMAGE
))
3153 if (flags
& TGSI_MEMBAR_SHARED
)
3154 waitcnt
&= LGKM_CNT
;
3156 if (waitcnt
!= NOOP_WAITCNT
)
3157 emit_waitcnt(ctx
, waitcnt
);
3161 shader_buffer_fetch_rsrc(struct si_shader_context
*ctx
,
3162 const struct tgsi_full_src_register
*reg
)
3165 LLVMValueRef rsrc_ptr
= LLVMGetParam(ctx
->main_fn
,
3166 SI_PARAM_SHADER_BUFFERS
);
3168 if (!reg
->Register
.Indirect
)
3169 index
= LLVMConstInt(ctx
->i32
, reg
->Register
.Index
, 0);
3171 index
= get_bounded_indirect_index(ctx
, ®
->Indirect
,
3172 reg
->Register
.Index
,
3173 SI_NUM_SHADER_BUFFERS
);
3175 return ac_build_indexed_load_const(&ctx
->ac
, rsrc_ptr
, index
);
3178 static bool tgsi_is_array_sampler(unsigned target
)
3180 return target
== TGSI_TEXTURE_1D_ARRAY
||
3181 target
== TGSI_TEXTURE_SHADOW1D_ARRAY
||
3182 target
== TGSI_TEXTURE_2D_ARRAY
||
3183 target
== TGSI_TEXTURE_SHADOW2D_ARRAY
||
3184 target
== TGSI_TEXTURE_CUBE_ARRAY
||
3185 target
== TGSI_TEXTURE_SHADOWCUBE_ARRAY
||
3186 target
== TGSI_TEXTURE_2D_ARRAY_MSAA
;
3189 static bool tgsi_is_array_image(unsigned target
)
3191 return target
== TGSI_TEXTURE_3D
||
3192 target
== TGSI_TEXTURE_CUBE
||
3193 target
== TGSI_TEXTURE_1D_ARRAY
||
3194 target
== TGSI_TEXTURE_2D_ARRAY
||
3195 target
== TGSI_TEXTURE_CUBE_ARRAY
||
3196 target
== TGSI_TEXTURE_2D_ARRAY_MSAA
;
3200 * Given a 256-bit resource descriptor, force the DCC enable bit to off.
3202 * At least on Tonga, executing image stores on images with DCC enabled and
3203 * non-trivial can eventually lead to lockups. This can occur when an
3204 * application binds an image as read-only but then uses a shader that writes
3205 * to it. The OpenGL spec allows almost arbitrarily bad behavior (including
3206 * program termination) in this case, but it doesn't cost much to be a bit
3207 * nicer: disabling DCC in the shader still leads to undefined results but
3208 * avoids the lockup.
3210 static LLVMValueRef
force_dcc_off(struct si_shader_context
*ctx
,
3213 if (ctx
->screen
->b
.chip_class
<= CIK
) {
3216 LLVMBuilderRef builder
= ctx
->gallivm
.builder
;
3217 LLVMValueRef i32_6
= LLVMConstInt(ctx
->i32
, 6, 0);
3218 LLVMValueRef i32_C
= LLVMConstInt(ctx
->i32
, C_008F28_COMPRESSION_EN
, 0);
3221 tmp
= LLVMBuildExtractElement(builder
, rsrc
, i32_6
, "");
3222 tmp
= LLVMBuildAnd(builder
, tmp
, i32_C
, "");
3223 return LLVMBuildInsertElement(builder
, rsrc
, tmp
, i32_6
, "");
3227 static LLVMTypeRef
const_array(LLVMTypeRef elem_type
, int num_elements
)
3229 return LLVMPointerType(LLVMArrayType(elem_type
, num_elements
),
3234 * Load the resource descriptor for \p image.
3238 struct lp_build_tgsi_context
*bld_base
,
3239 const struct tgsi_full_src_register
*image
,
3240 bool is_store
, unsigned target
,
3243 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3244 LLVMValueRef rsrc_ptr
= LLVMGetParam(ctx
->main_fn
,
3246 LLVMValueRef index
, tmp
;
3247 bool dcc_off
= target
!= TGSI_TEXTURE_BUFFER
&& is_store
;
3249 assert(image
->Register
.File
== TGSI_FILE_IMAGE
);
3251 if (!image
->Register
.Indirect
) {
3252 const struct tgsi_shader_info
*info
= bld_base
->info
;
3253 unsigned images_writemask
= info
->images_store
|
3254 info
->images_atomic
;
3256 index
= LLVMConstInt(ctx
->i32
, image
->Register
.Index
, 0);
3258 if (images_writemask
& (1 << image
->Register
.Index
) &&
3259 target
!= TGSI_TEXTURE_BUFFER
)
3262 /* From the GL_ARB_shader_image_load_store extension spec:
3264 * If a shader performs an image load, store, or atomic
3265 * operation using an image variable declared as an array,
3266 * and if the index used to select an individual element is
3267 * negative or greater than or equal to the size of the
3268 * array, the results of the operation are undefined but may
3269 * not lead to termination.
3271 index
= get_bounded_indirect_index(ctx
, &image
->Indirect
,
3272 image
->Register
.Index
,
3276 if (target
== TGSI_TEXTURE_BUFFER
) {
3277 LLVMBuilderRef builder
= ctx
->gallivm
.builder
;
3279 rsrc_ptr
= LLVMBuildPointerCast(builder
, rsrc_ptr
,
3280 const_array(ctx
->v4i32
, 0), "");
3281 index
= LLVMBuildMul(builder
, index
,
3282 LLVMConstInt(ctx
->i32
, 2, 0), "");
3283 index
= LLVMBuildAdd(builder
, index
,
3284 LLVMConstInt(ctx
->i32
, 1, 0), "");
3285 *rsrc
= ac_build_indexed_load_const(&ctx
->ac
, rsrc_ptr
, index
);
3289 tmp
= ac_build_indexed_load_const(&ctx
->ac
, rsrc_ptr
, index
);
3291 tmp
= force_dcc_off(ctx
, tmp
);
3295 static LLVMValueRef
image_fetch_coords(
3296 struct lp_build_tgsi_context
*bld_base
,
3297 const struct tgsi_full_instruction
*inst
,
3300 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3301 LLVMBuilderRef builder
= gallivm
->builder
;
3302 unsigned target
= inst
->Memory
.Texture
;
3303 unsigned num_coords
= tgsi_util_get_texture_coord_dim(target
);
3304 LLVMValueRef coords
[4];
3308 for (chan
= 0; chan
< num_coords
; ++chan
) {
3309 tmp
= lp_build_emit_fetch(bld_base
, inst
, src
, chan
);
3310 tmp
= LLVMBuildBitCast(builder
, tmp
, bld_base
->uint_bld
.elem_type
, "");
3314 if (num_coords
== 1)
3317 if (num_coords
== 3) {
3318 /* LLVM has difficulties lowering 3-element vectors. */
3319 coords
[3] = bld_base
->uint_bld
.undef
;
3323 return lp_build_gather_values(gallivm
, coords
, num_coords
);
3327 * Append the extra mode bits that are used by image load and store.
3329 static void image_append_args(
3330 struct si_shader_context
*ctx
,
3331 struct lp_build_emit_data
* emit_data
,
3336 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3337 LLVMValueRef i1false
= LLVMConstInt(ctx
->i1
, 0, 0);
3338 LLVMValueRef i1true
= LLVMConstInt(ctx
->i1
, 1, 0);
3339 LLVMValueRef r128
= i1false
;
3340 LLVMValueRef da
= tgsi_is_array_image(target
) ? i1true
: i1false
;
3343 inst
->Memory
.Qualifier
& (TGSI_MEMORY_COHERENT
| TGSI_MEMORY_VOLATILE
) ?
3345 LLVMValueRef slc
= i1false
;
3346 LLVMValueRef lwe
= i1false
;
3348 if (atomic
|| (HAVE_LLVM
<= 0x0309)) {
3349 emit_data
->args
[emit_data
->arg_count
++] = r128
;
3350 emit_data
->args
[emit_data
->arg_count
++] = da
;
3352 emit_data
->args
[emit_data
->arg_count
++] = glc
;
3354 emit_data
->args
[emit_data
->arg_count
++] = slc
;
3358 /* HAVE_LLVM >= 0x0400 */
3359 emit_data
->args
[emit_data
->arg_count
++] = glc
;
3360 emit_data
->args
[emit_data
->arg_count
++] = slc
;
3361 emit_data
->args
[emit_data
->arg_count
++] = lwe
;
3362 emit_data
->args
[emit_data
->arg_count
++] = da
;
3366 * Append the resource and indexing arguments for buffer intrinsics.
3368 * \param rsrc the v4i32 buffer resource
3369 * \param index index into the buffer (stride-based)
3370 * \param offset byte offset into the buffer
3372 static void buffer_append_args(
3373 struct si_shader_context
*ctx
,
3374 struct lp_build_emit_data
*emit_data
,
3377 LLVMValueRef offset
,
3381 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3382 LLVMValueRef i1false
= LLVMConstInt(ctx
->i1
, 0, 0);
3383 LLVMValueRef i1true
= LLVMConstInt(ctx
->i1
, 1, 0);
3385 emit_data
->args
[emit_data
->arg_count
++] = rsrc
;
3386 emit_data
->args
[emit_data
->arg_count
++] = index
; /* vindex */
3387 emit_data
->args
[emit_data
->arg_count
++] = offset
; /* voffset */
3389 emit_data
->args
[emit_data
->arg_count
++] =
3391 inst
->Memory
.Qualifier
& (TGSI_MEMORY_COHERENT
| TGSI_MEMORY_VOLATILE
) ?
3392 i1true
: i1false
; /* glc */
3394 emit_data
->args
[emit_data
->arg_count
++] = i1false
; /* slc */
3397 static void load_fetch_args(
3398 struct lp_build_tgsi_context
* bld_base
,
3399 struct lp_build_emit_data
* emit_data
)
3401 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3402 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3403 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
3404 unsigned target
= inst
->Memory
.Texture
;
3407 emit_data
->dst_type
= LLVMVectorType(bld_base
->base
.elem_type
, 4);
3409 if (inst
->Src
[0].Register
.File
== TGSI_FILE_BUFFER
) {
3410 LLVMBuilderRef builder
= gallivm
->builder
;
3411 LLVMValueRef offset
;
3414 rsrc
= shader_buffer_fetch_rsrc(ctx
, &inst
->Src
[0]);
3416 tmp
= lp_build_emit_fetch(bld_base
, inst
, 1, 0);
3417 offset
= LLVMBuildBitCast(builder
, tmp
, bld_base
->uint_bld
.elem_type
, "");
3419 buffer_append_args(ctx
, emit_data
, rsrc
, bld_base
->uint_bld
.zero
,
3420 offset
, false, false);
3421 } else if (inst
->Src
[0].Register
.File
== TGSI_FILE_IMAGE
) {
3422 LLVMValueRef coords
;
3424 image_fetch_rsrc(bld_base
, &inst
->Src
[0], false, target
, &rsrc
);
3425 coords
= image_fetch_coords(bld_base
, inst
, 1);
3427 if (target
== TGSI_TEXTURE_BUFFER
) {
3428 buffer_append_args(ctx
, emit_data
, rsrc
, coords
,
3429 bld_base
->uint_bld
.zero
, false, false);
3431 emit_data
->args
[0] = coords
;
3432 emit_data
->args
[1] = rsrc
;
3433 emit_data
->args
[2] = lp_build_const_int32(gallivm
, 15); /* dmask */
3434 emit_data
->arg_count
= 3;
3436 image_append_args(ctx
, emit_data
, target
, false, false);
3441 static unsigned get_load_intr_attribs(bool readonly_memory
)
3443 /* READNONE means writes can't affect it, while READONLY means that
3444 * writes can affect it. */
3445 return readonly_memory
&& HAVE_LLVM
>= 0x0400 ?
3446 LP_FUNC_ATTR_READNONE
:
3447 LP_FUNC_ATTR_READONLY
;
3450 static unsigned get_store_intr_attribs(bool writeonly_memory
)
3452 return writeonly_memory
&& HAVE_LLVM
>= 0x0400 ?
3453 LP_FUNC_ATTR_INACCESSIBLE_MEM_ONLY
:
3454 LP_FUNC_ATTR_WRITEONLY
;
3457 static void load_emit_buffer(struct si_shader_context
*ctx
,
3458 struct lp_build_emit_data
*emit_data
,
3459 bool readonly_memory
)
3461 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3462 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
3463 LLVMBuilderRef builder
= gallivm
->builder
;
3464 uint writemask
= inst
->Dst
[0].Register
.WriteMask
;
3465 uint count
= util_last_bit(writemask
);
3466 const char *intrinsic_name
;
3467 LLVMTypeRef dst_type
;
3471 intrinsic_name
= "llvm.amdgcn.buffer.load.f32";
3472 dst_type
= ctx
->f32
;
3475 intrinsic_name
= "llvm.amdgcn.buffer.load.v2f32";
3476 dst_type
= LLVMVectorType(ctx
->f32
, 2);
3479 intrinsic_name
= "llvm.amdgcn.buffer.load.v4f32";
3480 dst_type
= ctx
->v4f32
;
3484 emit_data
->output
[emit_data
->chan
] = lp_build_intrinsic(
3485 builder
, intrinsic_name
, dst_type
,
3486 emit_data
->args
, emit_data
->arg_count
,
3487 get_load_intr_attribs(readonly_memory
));
3490 static LLVMValueRef
get_memory_ptr(struct si_shader_context
*ctx
,
3491 const struct tgsi_full_instruction
*inst
,
3492 LLVMTypeRef type
, int arg
)
3494 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
3495 LLVMBuilderRef builder
= gallivm
->builder
;
3496 LLVMValueRef offset
, ptr
;
3499 offset
= lp_build_emit_fetch(&ctx
->bld_base
, inst
, arg
, 0);
3500 offset
= LLVMBuildBitCast(builder
, offset
, ctx
->i32
, "");
3502 ptr
= ctx
->shared_memory
;
3503 ptr
= LLVMBuildGEP(builder
, ptr
, &offset
, 1, "");
3504 addr_space
= LLVMGetPointerAddressSpace(LLVMTypeOf(ptr
));
3505 ptr
= LLVMBuildBitCast(builder
, ptr
, LLVMPointerType(type
, addr_space
), "");
3510 static void load_emit_memory(
3511 struct si_shader_context
*ctx
,
3512 struct lp_build_emit_data
*emit_data
)
3514 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3515 struct lp_build_context
*base
= &ctx
->bld_base
.base
;
3516 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
3517 LLVMBuilderRef builder
= gallivm
->builder
;
3518 unsigned writemask
= inst
->Dst
[0].Register
.WriteMask
;
3519 LLVMValueRef channels
[4], ptr
, derived_ptr
, index
;
3522 ptr
= get_memory_ptr(ctx
, inst
, base
->elem_type
, 1);
3524 for (chan
= 0; chan
< 4; ++chan
) {
3525 if (!(writemask
& (1 << chan
))) {
3526 channels
[chan
] = LLVMGetUndef(base
->elem_type
);
3530 index
= lp_build_const_int32(gallivm
, chan
);
3531 derived_ptr
= LLVMBuildGEP(builder
, ptr
, &index
, 1, "");
3532 channels
[chan
] = LLVMBuildLoad(builder
, derived_ptr
, "");
3534 emit_data
->output
[emit_data
->chan
] = lp_build_gather_values(gallivm
, channels
, 4);
3538 * Return true if the memory accessed by a LOAD or STORE instruction is
3539 * read-only or write-only, respectively.
3541 * \param shader_buffers_reverse_access_mask
3542 * For LOAD, set this to (store | atomic) slot usage in the shader.
3543 * For STORE, set this to (load | atomic) slot usage in the shader.
3544 * \param images_reverse_access_mask Same as above, but for images.
3546 static bool is_oneway_access_only(const struct tgsi_full_instruction
*inst
,
3547 const struct tgsi_shader_info
*info
,
3548 unsigned shader_buffers_reverse_access_mask
,
3549 unsigned images_reverse_access_mask
)
3551 /* RESTRICT means NOALIAS.
3552 * If there are no writes, we can assume the accessed memory is read-only.
3553 * If there are no reads, we can assume the accessed memory is write-only.
3555 if (inst
->Memory
.Qualifier
& TGSI_MEMORY_RESTRICT
) {
3556 unsigned reverse_access_mask
;
3558 if (inst
->Src
[0].Register
.File
== TGSI_FILE_BUFFER
) {
3559 reverse_access_mask
= shader_buffers_reverse_access_mask
;
3560 } else if (inst
->Memory
.Texture
== TGSI_TEXTURE_BUFFER
) {
3561 reverse_access_mask
= info
->images_buffers
&
3562 images_reverse_access_mask
;
3564 reverse_access_mask
= ~info
->images_buffers
&
3565 images_reverse_access_mask
;
3568 if (inst
->Src
[0].Register
.Indirect
) {
3569 if (!reverse_access_mask
)
3572 if (!(reverse_access_mask
&
3573 (1u << inst
->Src
[0].Register
.Index
)))
3578 /* If there are no buffer writes (for both shader buffers & image
3579 * buffers), it implies that buffer memory is read-only.
3580 * If there are no buffer reads (for both shader buffers & image
3581 * buffers), it implies that buffer memory is write-only.
3583 * Same for the case when there are no writes/reads for non-buffer
3586 if (inst
->Src
[0].Register
.File
== TGSI_FILE_BUFFER
||
3587 (inst
->Src
[0].Register
.File
== TGSI_FILE_IMAGE
&&
3588 inst
->Memory
.Texture
== TGSI_TEXTURE_BUFFER
)) {
3589 if (!shader_buffers_reverse_access_mask
&&
3590 !(info
->images_buffers
& images_reverse_access_mask
))
3593 if (!(~info
->images_buffers
& images_reverse_access_mask
))
3599 static void load_emit(
3600 const struct lp_build_tgsi_action
*action
,
3601 struct lp_build_tgsi_context
*bld_base
,
3602 struct lp_build_emit_data
*emit_data
)
3604 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3605 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3606 LLVMBuilderRef builder
= gallivm
->builder
;
3607 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
3608 const struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
3609 char intrinsic_name
[64];
3610 bool readonly_memory
= false;
3612 if (inst
->Src
[0].Register
.File
== TGSI_FILE_MEMORY
) {
3613 load_emit_memory(ctx
, emit_data
);
3617 if (inst
->Memory
.Qualifier
& TGSI_MEMORY_VOLATILE
)
3618 emit_waitcnt(ctx
, VM_CNT
);
3620 readonly_memory
= !(inst
->Memory
.Qualifier
& TGSI_MEMORY_VOLATILE
) &&
3621 is_oneway_access_only(inst
, info
,
3622 info
->shader_buffers_store
|
3623 info
->shader_buffers_atomic
,
3624 info
->images_store
|
3625 info
->images_atomic
);
3627 if (inst
->Src
[0].Register
.File
== TGSI_FILE_BUFFER
) {
3628 load_emit_buffer(ctx
, emit_data
, readonly_memory
);
3632 if (inst
->Memory
.Texture
== TGSI_TEXTURE_BUFFER
) {
3633 emit_data
->output
[emit_data
->chan
] =
3635 builder
, "llvm.amdgcn.buffer.load.format.v4f32", emit_data
->dst_type
,
3636 emit_data
->args
, emit_data
->arg_count
,
3637 get_load_intr_attribs(readonly_memory
));
3639 ac_get_image_intr_name("llvm.amdgcn.image.load",
3640 emit_data
->dst_type
, /* vdata */
3641 LLVMTypeOf(emit_data
->args
[0]), /* coords */
3642 LLVMTypeOf(emit_data
->args
[1]), /* rsrc */
3643 intrinsic_name
, sizeof(intrinsic_name
));
3645 emit_data
->output
[emit_data
->chan
] =
3647 builder
, intrinsic_name
, emit_data
->dst_type
,
3648 emit_data
->args
, emit_data
->arg_count
,
3649 get_load_intr_attribs(readonly_memory
));
3653 static void store_fetch_args(
3654 struct lp_build_tgsi_context
* bld_base
,
3655 struct lp_build_emit_data
* emit_data
)
3657 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3658 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3659 LLVMBuilderRef builder
= gallivm
->builder
;
3660 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
3661 struct tgsi_full_src_register memory
;
3662 LLVMValueRef chans
[4];
3667 emit_data
->dst_type
= LLVMVoidTypeInContext(gallivm
->context
);
3669 for (chan
= 0; chan
< 4; ++chan
) {
3670 chans
[chan
] = lp_build_emit_fetch(bld_base
, inst
, 1, chan
);
3672 data
= lp_build_gather_values(gallivm
, chans
, 4);
3674 emit_data
->args
[emit_data
->arg_count
++] = data
;
3676 memory
= tgsi_full_src_register_from_dst(&inst
->Dst
[0]);
3678 if (inst
->Dst
[0].Register
.File
== TGSI_FILE_BUFFER
) {
3679 LLVMValueRef offset
;
3682 rsrc
= shader_buffer_fetch_rsrc(ctx
, &memory
);
3684 tmp
= lp_build_emit_fetch(bld_base
, inst
, 0, 0);
3685 offset
= LLVMBuildBitCast(builder
, tmp
, bld_base
->uint_bld
.elem_type
, "");
3687 buffer_append_args(ctx
, emit_data
, rsrc
, bld_base
->uint_bld
.zero
,
3688 offset
, false, false);
3689 } else if (inst
->Dst
[0].Register
.File
== TGSI_FILE_IMAGE
) {
3690 unsigned target
= inst
->Memory
.Texture
;
3691 LLVMValueRef coords
;
3693 /* 8bit/16bit TC L1 write corruption bug on SI.
3694 * All store opcodes not aligned to a dword are affected.
3696 * The only way to get unaligned stores in radeonsi is through
3699 bool force_glc
= ctx
->screen
->b
.chip_class
== SI
;
3701 coords
= image_fetch_coords(bld_base
, inst
, 0);
3703 if (target
== TGSI_TEXTURE_BUFFER
) {
3704 image_fetch_rsrc(bld_base
, &memory
, true, target
, &rsrc
);
3705 buffer_append_args(ctx
, emit_data
, rsrc
, coords
,
3706 bld_base
->uint_bld
.zero
, false, force_glc
);
3708 emit_data
->args
[1] = coords
;
3709 image_fetch_rsrc(bld_base
, &memory
, true, target
,
3710 &emit_data
->args
[2]);
3711 emit_data
->args
[3] = lp_build_const_int32(gallivm
, 15); /* dmask */
3712 emit_data
->arg_count
= 4;
3714 image_append_args(ctx
, emit_data
, target
, false, force_glc
);
3719 static void store_emit_buffer(
3720 struct si_shader_context
*ctx
,
3721 struct lp_build_emit_data
*emit_data
,
3722 bool writeonly_memory
)
3724 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3725 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
3726 LLVMBuilderRef builder
= gallivm
->builder
;
3727 struct lp_build_context
*uint_bld
= &ctx
->bld_base
.uint_bld
;
3728 LLVMValueRef base_data
= emit_data
->args
[0];
3729 LLVMValueRef base_offset
= emit_data
->args
[3];
3730 unsigned writemask
= inst
->Dst
[0].Register
.WriteMask
;
3734 const char *intrinsic_name
;
3736 LLVMValueRef offset
;
3739 u_bit_scan_consecutive_range(&writemask
, &start
, &count
);
3741 /* Due to an LLVM limitation, split 3-element writes
3742 * into a 2-element and a 1-element write. */
3744 writemask
|= 1 << (start
+ 2);
3750 intrinsic_name
= "llvm.amdgcn.buffer.store.v4f32";
3751 } else if (count
== 2) {
3752 LLVMTypeRef v2f32
= LLVMVectorType(ctx
->f32
, 2);
3754 tmp
= LLVMBuildExtractElement(
3756 lp_build_const_int32(gallivm
, start
), "");
3757 data
= LLVMBuildInsertElement(
3758 builder
, LLVMGetUndef(v2f32
), tmp
,
3759 uint_bld
->zero
, "");
3761 tmp
= LLVMBuildExtractElement(
3763 lp_build_const_int32(gallivm
, start
+ 1), "");
3764 data
= LLVMBuildInsertElement(
3765 builder
, data
, tmp
, uint_bld
->one
, "");
3767 intrinsic_name
= "llvm.amdgcn.buffer.store.v2f32";
3770 data
= LLVMBuildExtractElement(
3772 lp_build_const_int32(gallivm
, start
), "");
3773 intrinsic_name
= "llvm.amdgcn.buffer.store.f32";
3776 offset
= base_offset
;
3778 offset
= LLVMBuildAdd(
3780 lp_build_const_int32(gallivm
, start
* 4), "");
3783 emit_data
->args
[0] = data
;
3784 emit_data
->args
[3] = offset
;
3787 builder
, intrinsic_name
, emit_data
->dst_type
,
3788 emit_data
->args
, emit_data
->arg_count
,
3789 get_store_intr_attribs(writeonly_memory
));
3793 static void store_emit_memory(
3794 struct si_shader_context
*ctx
,
3795 struct lp_build_emit_data
*emit_data
)
3797 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3798 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
3799 struct lp_build_context
*base
= &ctx
->bld_base
.base
;
3800 LLVMBuilderRef builder
= gallivm
->builder
;
3801 unsigned writemask
= inst
->Dst
[0].Register
.WriteMask
;
3802 LLVMValueRef ptr
, derived_ptr
, data
, index
;
3805 ptr
= get_memory_ptr(ctx
, inst
, base
->elem_type
, 0);
3807 for (chan
= 0; chan
< 4; ++chan
) {
3808 if (!(writemask
& (1 << chan
))) {
3811 data
= lp_build_emit_fetch(&ctx
->bld_base
, inst
, 1, chan
);
3812 index
= lp_build_const_int32(gallivm
, chan
);
3813 derived_ptr
= LLVMBuildGEP(builder
, ptr
, &index
, 1, "");
3814 LLVMBuildStore(builder
, data
, derived_ptr
);
3818 static void store_emit(
3819 const struct lp_build_tgsi_action
*action
,
3820 struct lp_build_tgsi_context
*bld_base
,
3821 struct lp_build_emit_data
*emit_data
)
3823 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3824 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3825 LLVMBuilderRef builder
= gallivm
->builder
;
3826 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
3827 const struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
3828 unsigned target
= inst
->Memory
.Texture
;
3829 char intrinsic_name
[64];
3830 bool writeonly_memory
= false;
3832 if (inst
->Dst
[0].Register
.File
== TGSI_FILE_MEMORY
) {
3833 store_emit_memory(ctx
, emit_data
);
3837 if (inst
->Memory
.Qualifier
& TGSI_MEMORY_VOLATILE
)
3838 emit_waitcnt(ctx
, VM_CNT
);
3840 writeonly_memory
= is_oneway_access_only(inst
, info
,
3841 info
->shader_buffers_load
|
3842 info
->shader_buffers_atomic
,
3844 info
->images_atomic
);
3846 if (inst
->Dst
[0].Register
.File
== TGSI_FILE_BUFFER
) {
3847 store_emit_buffer(ctx
, emit_data
, writeonly_memory
);
3851 if (target
== TGSI_TEXTURE_BUFFER
) {
3852 emit_data
->output
[emit_data
->chan
] = lp_build_intrinsic(
3853 builder
, "llvm.amdgcn.buffer.store.format.v4f32",
3854 emit_data
->dst_type
, emit_data
->args
,
3855 emit_data
->arg_count
,
3856 get_store_intr_attribs(writeonly_memory
));
3858 ac_get_image_intr_name("llvm.amdgcn.image.store",
3859 LLVMTypeOf(emit_data
->args
[0]), /* vdata */
3860 LLVMTypeOf(emit_data
->args
[1]), /* coords */
3861 LLVMTypeOf(emit_data
->args
[2]), /* rsrc */
3862 intrinsic_name
, sizeof(intrinsic_name
));
3864 emit_data
->output
[emit_data
->chan
] =
3866 builder
, intrinsic_name
, emit_data
->dst_type
,
3867 emit_data
->args
, emit_data
->arg_count
,
3868 get_store_intr_attribs(writeonly_memory
));
3872 static void atomic_fetch_args(
3873 struct lp_build_tgsi_context
* bld_base
,
3874 struct lp_build_emit_data
* emit_data
)
3876 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3877 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3878 LLVMBuilderRef builder
= gallivm
->builder
;
3879 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
3880 LLVMValueRef data1
, data2
;
3884 emit_data
->dst_type
= bld_base
->base
.elem_type
;
3886 tmp
= lp_build_emit_fetch(bld_base
, inst
, 2, 0);
3887 data1
= LLVMBuildBitCast(builder
, tmp
, bld_base
->uint_bld
.elem_type
, "");
3889 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_ATOMCAS
) {
3890 tmp
= lp_build_emit_fetch(bld_base
, inst
, 3, 0);
3891 data2
= LLVMBuildBitCast(builder
, tmp
, bld_base
->uint_bld
.elem_type
, "");
3894 /* llvm.amdgcn.image/buffer.atomic.cmpswap reflect the hardware order
3895 * of arguments, which is reversed relative to TGSI (and GLSL)
3897 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_ATOMCAS
)
3898 emit_data
->args
[emit_data
->arg_count
++] = data2
;
3899 emit_data
->args
[emit_data
->arg_count
++] = data1
;
3901 if (inst
->Src
[0].Register
.File
== TGSI_FILE_BUFFER
) {
3902 LLVMValueRef offset
;
3904 rsrc
= shader_buffer_fetch_rsrc(ctx
, &inst
->Src
[0]);
3906 tmp
= lp_build_emit_fetch(bld_base
, inst
, 1, 0);
3907 offset
= LLVMBuildBitCast(builder
, tmp
, bld_base
->uint_bld
.elem_type
, "");
3909 buffer_append_args(ctx
, emit_data
, rsrc
, bld_base
->uint_bld
.zero
,
3910 offset
, true, false);
3911 } else if (inst
->Src
[0].Register
.File
== TGSI_FILE_IMAGE
) {
3912 unsigned target
= inst
->Memory
.Texture
;
3913 LLVMValueRef coords
;
3915 image_fetch_rsrc(bld_base
, &inst
->Src
[0], true, target
, &rsrc
);
3916 coords
= image_fetch_coords(bld_base
, inst
, 1);
3918 if (target
== TGSI_TEXTURE_BUFFER
) {
3919 buffer_append_args(ctx
, emit_data
, rsrc
, coords
,
3920 bld_base
->uint_bld
.zero
, true, false);
3922 emit_data
->args
[emit_data
->arg_count
++] = coords
;
3923 emit_data
->args
[emit_data
->arg_count
++] = rsrc
;
3925 image_append_args(ctx
, emit_data
, target
, true, false);
3930 static void atomic_emit_memory(struct si_shader_context
*ctx
,
3931 struct lp_build_emit_data
*emit_data
) {
3932 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
3933 LLVMBuilderRef builder
= gallivm
->builder
;
3934 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
3935 LLVMValueRef ptr
, result
, arg
;
3937 ptr
= get_memory_ptr(ctx
, inst
, ctx
->i32
, 1);
3939 arg
= lp_build_emit_fetch(&ctx
->bld_base
, inst
, 2, 0);
3940 arg
= LLVMBuildBitCast(builder
, arg
, ctx
->i32
, "");
3942 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_ATOMCAS
) {
3943 LLVMValueRef new_data
;
3944 new_data
= lp_build_emit_fetch(&ctx
->bld_base
,
3947 new_data
= LLVMBuildBitCast(builder
, new_data
, ctx
->i32
, "");
3949 #if HAVE_LLVM >= 0x309
3950 result
= LLVMBuildAtomicCmpXchg(builder
, ptr
, arg
, new_data
,
3951 LLVMAtomicOrderingSequentiallyConsistent
,
3952 LLVMAtomicOrderingSequentiallyConsistent
,
3956 result
= LLVMBuildExtractValue(builder
, result
, 0, "");
3958 LLVMAtomicRMWBinOp op
;
3960 switch(inst
->Instruction
.Opcode
) {
3961 case TGSI_OPCODE_ATOMUADD
:
3962 op
= LLVMAtomicRMWBinOpAdd
;
3964 case TGSI_OPCODE_ATOMXCHG
:
3965 op
= LLVMAtomicRMWBinOpXchg
;
3967 case TGSI_OPCODE_ATOMAND
:
3968 op
= LLVMAtomicRMWBinOpAnd
;
3970 case TGSI_OPCODE_ATOMOR
:
3971 op
= LLVMAtomicRMWBinOpOr
;
3973 case TGSI_OPCODE_ATOMXOR
:
3974 op
= LLVMAtomicRMWBinOpXor
;
3976 case TGSI_OPCODE_ATOMUMIN
:
3977 op
= LLVMAtomicRMWBinOpUMin
;
3979 case TGSI_OPCODE_ATOMUMAX
:
3980 op
= LLVMAtomicRMWBinOpUMax
;
3982 case TGSI_OPCODE_ATOMIMIN
:
3983 op
= LLVMAtomicRMWBinOpMin
;
3985 case TGSI_OPCODE_ATOMIMAX
:
3986 op
= LLVMAtomicRMWBinOpMax
;
3989 unreachable("unknown atomic opcode");
3992 result
= LLVMBuildAtomicRMW(builder
, op
, ptr
, arg
,
3993 LLVMAtomicOrderingSequentiallyConsistent
,
3996 emit_data
->output
[emit_data
->chan
] = LLVMBuildBitCast(builder
, result
, emit_data
->dst_type
, "");
3999 static void atomic_emit(
4000 const struct lp_build_tgsi_action
*action
,
4001 struct lp_build_tgsi_context
*bld_base
,
4002 struct lp_build_emit_data
*emit_data
)
4004 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4005 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4006 LLVMBuilderRef builder
= gallivm
->builder
;
4007 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
4008 char intrinsic_name
[40];
4011 if (inst
->Src
[0].Register
.File
== TGSI_FILE_MEMORY
) {
4012 atomic_emit_memory(ctx
, emit_data
);
4016 if (inst
->Src
[0].Register
.File
== TGSI_FILE_BUFFER
||
4017 inst
->Memory
.Texture
== TGSI_TEXTURE_BUFFER
) {
4018 snprintf(intrinsic_name
, sizeof(intrinsic_name
),
4019 "llvm.amdgcn.buffer.atomic.%s", action
->intr_name
);
4021 LLVMValueRef coords
;
4022 char coords_type
[8];
4024 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_ATOMCAS
)
4025 coords
= emit_data
->args
[2];
4027 coords
= emit_data
->args
[1];
4029 ac_build_type_name_for_intr(LLVMTypeOf(coords
), coords_type
, sizeof(coords_type
));
4030 snprintf(intrinsic_name
, sizeof(intrinsic_name
),
4031 "llvm.amdgcn.image.atomic.%s.%s",
4032 action
->intr_name
, coords_type
);
4035 tmp
= lp_build_intrinsic(
4036 builder
, intrinsic_name
, bld_base
->uint_bld
.elem_type
,
4037 emit_data
->args
, emit_data
->arg_count
, 0);
4038 emit_data
->output
[emit_data
->chan
] =
4039 LLVMBuildBitCast(builder
, tmp
, bld_base
->base
.elem_type
, "");
4042 static void set_tex_fetch_args(struct si_shader_context
*ctx
,
4043 struct lp_build_emit_data
*emit_data
,
4045 LLVMValueRef res_ptr
, LLVMValueRef samp_ptr
,
4046 LLVMValueRef
*param
, unsigned count
,
4049 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
4050 struct ac_image_args args
= {};
4052 /* Pad to power of two vector */
4053 while (count
< util_next_power_of_two(count
))
4054 param
[count
++] = LLVMGetUndef(ctx
->i32
);
4057 args
.addr
= lp_build_gather_values(gallivm
, param
, count
);
4059 args
.addr
= param
[0];
4061 args
.resource
= res_ptr
;
4062 args
.sampler
= samp_ptr
;
4064 args
.unorm
= target
== TGSI_TEXTURE_RECT
||
4065 target
== TGSI_TEXTURE_SHADOWRECT
;
4066 args
.da
= tgsi_is_array_sampler(target
);
4068 /* Ugly, but we seem to have no other choice right now. */
4069 STATIC_ASSERT(sizeof(args
) <= sizeof(emit_data
->args
));
4070 memcpy(emit_data
->args
, &args
, sizeof(args
));
4073 static void resq_fetch_args(
4074 struct lp_build_tgsi_context
* bld_base
,
4075 struct lp_build_emit_data
* emit_data
)
4077 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4078 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
4079 const struct tgsi_full_src_register
*reg
= &inst
->Src
[0];
4081 emit_data
->dst_type
= ctx
->v4i32
;
4083 if (reg
->Register
.File
== TGSI_FILE_BUFFER
) {
4084 emit_data
->args
[0] = shader_buffer_fetch_rsrc(ctx
, reg
);
4085 emit_data
->arg_count
= 1;
4086 } else if (inst
->Memory
.Texture
== TGSI_TEXTURE_BUFFER
) {
4087 image_fetch_rsrc(bld_base
, reg
, false, inst
->Memory
.Texture
,
4088 &emit_data
->args
[0]);
4089 emit_data
->arg_count
= 1;
4091 LLVMValueRef res_ptr
;
4092 unsigned image_target
;
4094 if (inst
->Memory
.Texture
== TGSI_TEXTURE_3D
)
4095 image_target
= TGSI_TEXTURE_2D_ARRAY
;
4097 image_target
= inst
->Memory
.Texture
;
4099 image_fetch_rsrc(bld_base
, reg
, false, inst
->Memory
.Texture
,
4101 set_tex_fetch_args(ctx
, emit_data
, image_target
,
4102 res_ptr
, NULL
, &bld_base
->uint_bld
.zero
, 1,
4107 static void resq_emit(
4108 const struct lp_build_tgsi_action
*action
,
4109 struct lp_build_tgsi_context
*bld_base
,
4110 struct lp_build_emit_data
*emit_data
)
4112 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4113 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4114 LLVMBuilderRef builder
= gallivm
->builder
;
4115 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
4118 if (inst
->Src
[0].Register
.File
== TGSI_FILE_BUFFER
) {
4119 out
= LLVMBuildExtractElement(builder
, emit_data
->args
[0],
4120 lp_build_const_int32(gallivm
, 2), "");
4121 } else if (inst
->Memory
.Texture
== TGSI_TEXTURE_BUFFER
) {
4122 out
= get_buffer_size(bld_base
, emit_data
->args
[0]);
4124 struct ac_image_args args
;
4126 memcpy(&args
, emit_data
->args
, sizeof(args
)); /* ugly */
4127 args
.opcode
= ac_image_get_resinfo
;
4128 out
= ac_build_image_opcode(&ctx
->ac
, &args
);
4130 /* Divide the number of layers by 6 to get the number of cubes. */
4131 if (inst
->Memory
.Texture
== TGSI_TEXTURE_CUBE_ARRAY
) {
4132 LLVMValueRef imm2
= lp_build_const_int32(gallivm
, 2);
4133 LLVMValueRef imm6
= lp_build_const_int32(gallivm
, 6);
4135 LLVMValueRef z
= LLVMBuildExtractElement(builder
, out
, imm2
, "");
4136 z
= LLVMBuildSDiv(builder
, z
, imm6
, "");
4137 out
= LLVMBuildInsertElement(builder
, out
, z
, imm2
, "");
4141 emit_data
->output
[emit_data
->chan
] = out
;
4144 static const struct lp_build_tgsi_action tex_action
;
4154 * Load an image view, fmask view. or sampler state descriptor.
4156 static LLVMValueRef
load_sampler_desc_custom(struct si_shader_context
*ctx
,
4157 LLVMValueRef list
, LLVMValueRef index
,
4158 enum desc_type type
)
4160 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
4161 LLVMBuilderRef builder
= gallivm
->builder
;
4165 /* The image is at [0:7]. */
4166 index
= LLVMBuildMul(builder
, index
, LLVMConstInt(ctx
->i32
, 2, 0), "");
4169 /* The buffer is in [4:7]. */
4170 index
= LLVMBuildMul(builder
, index
, LLVMConstInt(ctx
->i32
, 4, 0), "");
4171 index
= LLVMBuildAdd(builder
, index
, LLVMConstInt(ctx
->i32
, 1, 0), "");
4172 list
= LLVMBuildPointerCast(builder
, list
,
4173 const_array(ctx
->v4i32
, 0), "");
4176 /* The FMASK is at [8:15]. */
4177 index
= LLVMBuildMul(builder
, index
, LLVMConstInt(ctx
->i32
, 2, 0), "");
4178 index
= LLVMBuildAdd(builder
, index
, LLVMConstInt(ctx
->i32
, 1, 0), "");
4181 /* The sampler state is at [12:15]. */
4182 index
= LLVMBuildMul(builder
, index
, LLVMConstInt(ctx
->i32
, 4, 0), "");
4183 index
= LLVMBuildAdd(builder
, index
, LLVMConstInt(ctx
->i32
, 3, 0), "");
4184 list
= LLVMBuildPointerCast(builder
, list
,
4185 const_array(ctx
->v4i32
, 0), "");
4189 return ac_build_indexed_load_const(&ctx
->ac
, list
, index
);
4192 static LLVMValueRef
load_sampler_desc(struct si_shader_context
*ctx
,
4193 LLVMValueRef index
, enum desc_type type
)
4195 LLVMValueRef list
= LLVMGetParam(ctx
->main_fn
,
4198 return load_sampler_desc_custom(ctx
, list
, index
, type
);
4201 /* Disable anisotropic filtering if BASE_LEVEL == LAST_LEVEL.
4204 * If BASE_LEVEL == LAST_LEVEL, the shader must disable anisotropic
4205 * filtering manually. The driver sets img7 to a mask clearing
4206 * MAX_ANISO_RATIO if BASE_LEVEL == LAST_LEVEL. The shader must do:
4207 * s_and_b32 samp0, samp0, img7
4210 * The ANISO_OVERRIDE sampler field enables this fix in TA.
4212 static LLVMValueRef
sici_fix_sampler_aniso(struct si_shader_context
*ctx
,
4213 LLVMValueRef res
, LLVMValueRef samp
)
4215 LLVMBuilderRef builder
= ctx
->gallivm
.builder
;
4216 LLVMValueRef img7
, samp0
;
4218 if (ctx
->screen
->b
.chip_class
>= VI
)
4221 img7
= LLVMBuildExtractElement(builder
, res
,
4222 LLVMConstInt(ctx
->i32
, 7, 0), "");
4223 samp0
= LLVMBuildExtractElement(builder
, samp
,
4224 LLVMConstInt(ctx
->i32
, 0, 0), "");
4225 samp0
= LLVMBuildAnd(builder
, samp0
, img7
, "");
4226 return LLVMBuildInsertElement(builder
, samp
, samp0
,
4227 LLVMConstInt(ctx
->i32
, 0, 0), "");
4230 static void tex_fetch_ptrs(
4231 struct lp_build_tgsi_context
*bld_base
,
4232 struct lp_build_emit_data
*emit_data
,
4233 LLVMValueRef
*res_ptr
, LLVMValueRef
*samp_ptr
, LLVMValueRef
*fmask_ptr
)
4235 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4236 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
4237 unsigned target
= inst
->Texture
.Texture
;
4238 unsigned sampler_src
;
4239 unsigned sampler_index
;
4242 sampler_src
= emit_data
->inst
->Instruction
.NumSrcRegs
- 1;
4243 sampler_index
= emit_data
->inst
->Src
[sampler_src
].Register
.Index
;
4245 if (emit_data
->inst
->Src
[sampler_src
].Register
.Indirect
) {
4246 const struct tgsi_full_src_register
*reg
= &emit_data
->inst
->Src
[sampler_src
];
4248 index
= get_bounded_indirect_index(ctx
,
4250 reg
->Register
.Index
,
4253 index
= LLVMConstInt(ctx
->i32
, sampler_index
, 0);
4256 if (target
== TGSI_TEXTURE_BUFFER
)
4257 *res_ptr
= load_sampler_desc(ctx
, index
, DESC_BUFFER
);
4259 *res_ptr
= load_sampler_desc(ctx
, index
, DESC_IMAGE
);
4266 if (target
== TGSI_TEXTURE_2D_MSAA
||
4267 target
== TGSI_TEXTURE_2D_ARRAY_MSAA
) {
4269 *fmask_ptr
= load_sampler_desc(ctx
, index
, DESC_FMASK
);
4270 } else if (target
!= TGSI_TEXTURE_BUFFER
) {
4272 *samp_ptr
= load_sampler_desc(ctx
, index
, DESC_SAMPLER
);
4273 *samp_ptr
= sici_fix_sampler_aniso(ctx
, *res_ptr
, *samp_ptr
);
4278 static void txq_fetch_args(
4279 struct lp_build_tgsi_context
*bld_base
,
4280 struct lp_build_emit_data
*emit_data
)
4282 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4283 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
4284 unsigned target
= inst
->Texture
.Texture
;
4285 LLVMValueRef res_ptr
;
4286 LLVMValueRef address
;
4288 tex_fetch_ptrs(bld_base
, emit_data
, &res_ptr
, NULL
, NULL
);
4290 if (target
== TGSI_TEXTURE_BUFFER
) {
4291 /* Read the size from the buffer descriptor directly. */
4292 emit_data
->args
[0] = get_buffer_size(bld_base
, res_ptr
);
4296 /* Textures - set the mip level. */
4297 address
= lp_build_emit_fetch(bld_base
, inst
, 0, TGSI_CHAN_X
);
4299 set_tex_fetch_args(ctx
, emit_data
, target
, res_ptr
,
4300 NULL
, &address
, 1, 0xf);
4303 static void txq_emit(const struct lp_build_tgsi_action
*action
,
4304 struct lp_build_tgsi_context
*bld_base
,
4305 struct lp_build_emit_data
*emit_data
)
4307 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4308 struct ac_image_args args
;
4309 unsigned target
= emit_data
->inst
->Texture
.Texture
;
4311 if (target
== TGSI_TEXTURE_BUFFER
) {
4312 /* Just return the buffer size. */
4313 emit_data
->output
[emit_data
->chan
] = emit_data
->args
[0];
4317 memcpy(&args
, emit_data
->args
, sizeof(args
)); /* ugly */
4319 args
.opcode
= ac_image_get_resinfo
;
4320 emit_data
->output
[emit_data
->chan
] =
4321 ac_build_image_opcode(&ctx
->ac
, &args
);
4323 /* Divide the number of layers by 6 to get the number of cubes. */
4324 if (target
== TGSI_TEXTURE_CUBE_ARRAY
||
4325 target
== TGSI_TEXTURE_SHADOWCUBE_ARRAY
) {
4326 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
4327 LLVMValueRef two
= lp_build_const_int32(bld_base
->base
.gallivm
, 2);
4328 LLVMValueRef six
= lp_build_const_int32(bld_base
->base
.gallivm
, 6);
4330 LLVMValueRef v4
= emit_data
->output
[emit_data
->chan
];
4331 LLVMValueRef z
= LLVMBuildExtractElement(builder
, v4
, two
, "");
4332 z
= LLVMBuildSDiv(builder
, z
, six
, "");
4334 emit_data
->output
[emit_data
->chan
] =
4335 LLVMBuildInsertElement(builder
, v4
, z
, two
, "");
4339 static void tex_fetch_args(
4340 struct lp_build_tgsi_context
*bld_base
,
4341 struct lp_build_emit_data
*emit_data
)
4343 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4344 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4345 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
4346 unsigned opcode
= inst
->Instruction
.Opcode
;
4347 unsigned target
= inst
->Texture
.Texture
;
4348 LLVMValueRef coords
[5], derivs
[6];
4349 LLVMValueRef address
[16];
4350 unsigned num_coords
= tgsi_util_get_texture_coord_dim(target
);
4351 int ref_pos
= tgsi_util_get_shadow_ref_src_index(target
);
4354 unsigned num_deriv_channels
= 0;
4355 bool has_offset
= inst
->Texture
.NumOffsets
> 0;
4356 LLVMValueRef res_ptr
, samp_ptr
, fmask_ptr
= NULL
;
4357 unsigned dmask
= 0xf;
4359 tex_fetch_ptrs(bld_base
, emit_data
, &res_ptr
, &samp_ptr
, &fmask_ptr
);
4361 if (target
== TGSI_TEXTURE_BUFFER
) {
4362 emit_data
->dst_type
= ctx
->v4f32
;
4363 emit_data
->args
[0] = LLVMBuildBitCast(gallivm
->builder
, res_ptr
,
4365 emit_data
->args
[1] = bld_base
->uint_bld
.zero
;
4366 emit_data
->args
[2] = lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, TGSI_CHAN_X
);
4367 emit_data
->arg_count
= 3;
4371 /* Fetch and project texture coordinates */
4372 coords
[3] = lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, TGSI_CHAN_W
);
4373 for (chan
= 0; chan
< 3; chan
++ ) {
4374 coords
[chan
] = lp_build_emit_fetch(bld_base
,
4377 if (opcode
== TGSI_OPCODE_TXP
)
4378 coords
[chan
] = lp_build_emit_llvm_binary(bld_base
,
4384 if (opcode
== TGSI_OPCODE_TXP
)
4385 coords
[3] = bld_base
->base
.one
;
4388 if (has_offset
&& opcode
!= TGSI_OPCODE_TXF
) {
4389 /* The offsets are six-bit signed integers packed like this:
4390 * X=[5:0], Y=[13:8], and Z=[21:16].
4392 LLVMValueRef offset
[3], pack
;
4394 assert(inst
->Texture
.NumOffsets
== 1);
4396 for (chan
= 0; chan
< 3; chan
++) {
4397 offset
[chan
] = lp_build_emit_fetch_texoffset(bld_base
,
4398 emit_data
->inst
, 0, chan
);
4399 offset
[chan
] = LLVMBuildAnd(gallivm
->builder
, offset
[chan
],
4400 lp_build_const_int32(gallivm
, 0x3f), "");
4402 offset
[chan
] = LLVMBuildShl(gallivm
->builder
, offset
[chan
],
4403 lp_build_const_int32(gallivm
, chan
*8), "");
4406 pack
= LLVMBuildOr(gallivm
->builder
, offset
[0], offset
[1], "");
4407 pack
= LLVMBuildOr(gallivm
->builder
, pack
, offset
[2], "");
4408 address
[count
++] = pack
;
4411 /* Pack LOD bias value */
4412 if (opcode
== TGSI_OPCODE_TXB
)
4413 address
[count
++] = coords
[3];
4414 if (opcode
== TGSI_OPCODE_TXB2
)
4415 address
[count
++] = lp_build_emit_fetch(bld_base
, inst
, 1, TGSI_CHAN_X
);
4417 /* Pack depth comparison value */
4418 if (tgsi_is_shadow_target(target
) && opcode
!= TGSI_OPCODE_LODQ
) {
4421 if (target
== TGSI_TEXTURE_SHADOWCUBE_ARRAY
) {
4422 z
= lp_build_emit_fetch(bld_base
, inst
, 1, TGSI_CHAN_X
);
4424 assert(ref_pos
>= 0);
4425 z
= coords
[ref_pos
];
4428 /* TC-compatible HTILE promotes Z16 and Z24 to Z32_FLOAT,
4429 * so the depth comparison value isn't clamped for Z16 and
4430 * Z24 anymore. Do it manually here.
4432 * It's unnecessary if the original texture format was
4433 * Z32_FLOAT, but we don't know that here.
4435 if (ctx
->screen
->b
.chip_class
== VI
)
4436 z
= ac_build_clamp(&ctx
->ac
, z
);
4438 address
[count
++] = z
;
4441 /* Pack user derivatives */
4442 if (opcode
== TGSI_OPCODE_TXD
) {
4443 int param
, num_src_deriv_channels
;
4446 case TGSI_TEXTURE_3D
:
4447 num_src_deriv_channels
= 3;
4448 num_deriv_channels
= 3;
4450 case TGSI_TEXTURE_2D
:
4451 case TGSI_TEXTURE_SHADOW2D
:
4452 case TGSI_TEXTURE_RECT
:
4453 case TGSI_TEXTURE_SHADOWRECT
:
4454 case TGSI_TEXTURE_2D_ARRAY
:
4455 case TGSI_TEXTURE_SHADOW2D_ARRAY
:
4456 num_src_deriv_channels
= 2;
4457 num_deriv_channels
= 2;
4459 case TGSI_TEXTURE_CUBE
:
4460 case TGSI_TEXTURE_SHADOWCUBE
:
4461 case TGSI_TEXTURE_CUBE_ARRAY
:
4462 case TGSI_TEXTURE_SHADOWCUBE_ARRAY
:
4463 /* Cube derivatives will be converted to 2D. */
4464 num_src_deriv_channels
= 3;
4465 num_deriv_channels
= 2;
4467 case TGSI_TEXTURE_1D
:
4468 case TGSI_TEXTURE_SHADOW1D
:
4469 case TGSI_TEXTURE_1D_ARRAY
:
4470 case TGSI_TEXTURE_SHADOW1D_ARRAY
:
4471 num_src_deriv_channels
= 1;
4472 num_deriv_channels
= 1;
4475 unreachable("invalid target");
4478 for (param
= 0; param
< 2; param
++)
4479 for (chan
= 0; chan
< num_src_deriv_channels
; chan
++)
4480 derivs
[param
* num_src_deriv_channels
+ chan
] =
4481 lp_build_emit_fetch(bld_base
, inst
, param
+1, chan
);
4484 if (target
== TGSI_TEXTURE_CUBE
||
4485 target
== TGSI_TEXTURE_CUBE_ARRAY
||
4486 target
== TGSI_TEXTURE_SHADOWCUBE
||
4487 target
== TGSI_TEXTURE_SHADOWCUBE_ARRAY
)
4488 ac_prepare_cube_coords(&ctx
->ac
,
4489 opcode
== TGSI_OPCODE_TXD
,
4490 target
== TGSI_TEXTURE_CUBE_ARRAY
||
4491 target
== TGSI_TEXTURE_SHADOWCUBE_ARRAY
,
4494 if (opcode
== TGSI_OPCODE_TXD
)
4495 for (int i
= 0; i
< num_deriv_channels
* 2; i
++)
4496 address
[count
++] = derivs
[i
];
4498 /* Pack texture coordinates */
4499 address
[count
++] = coords
[0];
4501 address
[count
++] = coords
[1];
4503 address
[count
++] = coords
[2];
4505 /* Pack LOD or sample index */
4506 if (opcode
== TGSI_OPCODE_TXL
|| opcode
== TGSI_OPCODE_TXF
)
4507 address
[count
++] = coords
[3];
4508 else if (opcode
== TGSI_OPCODE_TXL2
)
4509 address
[count
++] = lp_build_emit_fetch(bld_base
, inst
, 1, TGSI_CHAN_X
);
4512 assert(!"Cannot handle more than 16 texture address parameters");
4516 for (chan
= 0; chan
< count
; chan
++ ) {
4517 address
[chan
] = LLVMBuildBitCast(gallivm
->builder
,
4518 address
[chan
], ctx
->i32
, "");
4521 /* Adjust the sample index according to FMASK.
4523 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
4524 * which is the identity mapping. Each nibble says which physical sample
4525 * should be fetched to get that sample.
4527 * For example, 0x11111100 means there are only 2 samples stored and
4528 * the second sample covers 3/4 of the pixel. When reading samples 0
4529 * and 1, return physical sample 0 (determined by the first two 0s
4530 * in FMASK), otherwise return physical sample 1.
4532 * The sample index should be adjusted as follows:
4533 * sample_index = (fmask >> (sample_index * 4)) & 0xF;
4535 if (target
== TGSI_TEXTURE_2D_MSAA
||
4536 target
== TGSI_TEXTURE_2D_ARRAY_MSAA
) {
4537 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
4538 struct lp_build_emit_data txf_emit_data
= *emit_data
;
4539 LLVMValueRef txf_address
[4];
4540 /* We only need .xy for non-arrays, and .xyz for arrays. */
4541 unsigned txf_count
= target
== TGSI_TEXTURE_2D_MSAA
? 2 : 3;
4542 struct tgsi_full_instruction inst
= {};
4544 memcpy(txf_address
, address
, sizeof(txf_address
));
4546 /* Read FMASK using TXF. */
4547 inst
.Instruction
.Opcode
= TGSI_OPCODE_TXF
;
4548 inst
.Texture
.Texture
= target
;
4549 txf_emit_data
.inst
= &inst
;
4550 txf_emit_data
.chan
= 0;
4551 set_tex_fetch_args(ctx
, &txf_emit_data
,
4552 target
, fmask_ptr
, NULL
,
4553 txf_address
, txf_count
, 0xf);
4554 build_tex_intrinsic(&tex_action
, bld_base
, &txf_emit_data
);
4556 /* Initialize some constants. */
4557 LLVMValueRef four
= LLVMConstInt(ctx
->i32
, 4, 0);
4558 LLVMValueRef F
= LLVMConstInt(ctx
->i32
, 0xF, 0);
4560 /* Apply the formula. */
4561 LLVMValueRef fmask
=
4562 LLVMBuildExtractElement(gallivm
->builder
,
4563 txf_emit_data
.output
[0],
4564 uint_bld
->zero
, "");
4566 unsigned sample_chan
= txf_count
; /* the sample index is last */
4568 LLVMValueRef sample_index4
=
4569 LLVMBuildMul(gallivm
->builder
, address
[sample_chan
], four
, "");
4571 LLVMValueRef shifted_fmask
=
4572 LLVMBuildLShr(gallivm
->builder
, fmask
, sample_index4
, "");
4574 LLVMValueRef final_sample
=
4575 LLVMBuildAnd(gallivm
->builder
, shifted_fmask
, F
, "");
4577 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
4578 * resource descriptor is 0 (invalid),
4580 LLVMValueRef fmask_desc
=
4581 LLVMBuildBitCast(gallivm
->builder
, fmask_ptr
,
4584 LLVMValueRef fmask_word1
=
4585 LLVMBuildExtractElement(gallivm
->builder
, fmask_desc
,
4588 LLVMValueRef word1_is_nonzero
=
4589 LLVMBuildICmp(gallivm
->builder
, LLVMIntNE
,
4590 fmask_word1
, uint_bld
->zero
, "");
4592 /* Replace the MSAA sample index. */
4593 address
[sample_chan
] =
4594 LLVMBuildSelect(gallivm
->builder
, word1_is_nonzero
,
4595 final_sample
, address
[sample_chan
], "");
4598 if (opcode
== TGSI_OPCODE_TXF
) {
4599 /* add tex offsets */
4600 if (inst
->Texture
.NumOffsets
) {
4601 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
4602 const struct tgsi_texture_offset
*off
= inst
->TexOffsets
;
4604 assert(inst
->Texture
.NumOffsets
== 1);
4607 case TGSI_TEXTURE_3D
:
4608 address
[2] = lp_build_add(uint_bld
, address
[2],
4609 ctx
->imms
[off
->Index
* TGSI_NUM_CHANNELS
+ off
->SwizzleZ
]);
4611 case TGSI_TEXTURE_2D
:
4612 case TGSI_TEXTURE_SHADOW2D
:
4613 case TGSI_TEXTURE_RECT
:
4614 case TGSI_TEXTURE_SHADOWRECT
:
4615 case TGSI_TEXTURE_2D_ARRAY
:
4616 case TGSI_TEXTURE_SHADOW2D_ARRAY
:
4618 lp_build_add(uint_bld
, address
[1],
4619 ctx
->imms
[off
->Index
* TGSI_NUM_CHANNELS
+ off
->SwizzleY
]);
4621 case TGSI_TEXTURE_1D
:
4622 case TGSI_TEXTURE_SHADOW1D
:
4623 case TGSI_TEXTURE_1D_ARRAY
:
4624 case TGSI_TEXTURE_SHADOW1D_ARRAY
:
4626 lp_build_add(uint_bld
, address
[0],
4627 ctx
->imms
[off
->Index
* TGSI_NUM_CHANNELS
+ off
->SwizzleX
]);
4629 /* texture offsets do not apply to other texture targets */
4634 if (opcode
== TGSI_OPCODE_TG4
) {
4635 unsigned gather_comp
= 0;
4637 /* DMASK was repurposed for GATHER4. 4 components are always
4638 * returned and DMASK works like a swizzle - it selects
4639 * the component to fetch. The only valid DMASK values are
4640 * 1=red, 2=green, 4=blue, 8=alpha. (e.g. 1 returns
4641 * (red,red,red,red) etc.) The ISA document doesn't mention
4645 /* Get the component index from src1.x for Gather4. */
4646 if (!tgsi_is_shadow_target(target
)) {
4647 LLVMValueRef comp_imm
;
4648 struct tgsi_src_register src1
= inst
->Src
[1].Register
;
4650 assert(src1
.File
== TGSI_FILE_IMMEDIATE
);
4652 comp_imm
= ctx
->imms
[src1
.Index
* TGSI_NUM_CHANNELS
+ src1
.SwizzleX
];
4653 gather_comp
= LLVMConstIntGetZExtValue(comp_imm
);
4654 gather_comp
= CLAMP(gather_comp
, 0, 3);
4657 dmask
= 1 << gather_comp
;
4660 set_tex_fetch_args(ctx
, emit_data
, target
, res_ptr
,
4661 samp_ptr
, address
, count
, dmask
);
4664 /* Gather4 should follow the same rules as bilinear filtering, but the hardware
4665 * incorrectly forces nearest filtering if the texture format is integer.
4666 * The only effect it has on Gather4, which always returns 4 texels for
4667 * bilinear filtering, is that the final coordinates are off by 0.5 of
4670 * The workaround is to subtract 0.5 from the unnormalized coordinates,
4671 * or (0.5 / size) from the normalized coordinates.
4673 static void si_lower_gather4_integer(struct si_shader_context
*ctx
,
4674 struct ac_image_args
*args
,
4677 LLVMBuilderRef builder
= ctx
->gallivm
.builder
;
4678 LLVMValueRef coord
= args
->addr
;
4679 LLVMValueRef half_texel
[2];
4680 /* Texture coordinates start after:
4681 * {offset, bias, z-compare, derivatives}
4682 * Only the offset and z-compare can occur here.
4684 unsigned coord_vgpr_index
= (int)args
->offset
+ (int)args
->compare
;
4687 if (target
== TGSI_TEXTURE_RECT
||
4688 target
== TGSI_TEXTURE_SHADOWRECT
) {
4689 half_texel
[0] = half_texel
[1] = LLVMConstReal(ctx
->f32
, -0.5);
4691 struct tgsi_full_instruction txq_inst
= {};
4692 struct lp_build_emit_data txq_emit_data
= {};
4694 /* Query the texture size. */
4695 txq_inst
.Texture
.Texture
= target
;
4696 txq_emit_data
.inst
= &txq_inst
;
4697 txq_emit_data
.dst_type
= ctx
->v4i32
;
4698 set_tex_fetch_args(ctx
, &txq_emit_data
, target
,
4699 args
->resource
, NULL
,
4700 &ctx
->bld_base
.uint_bld
.zero
,
4702 txq_emit(NULL
, &ctx
->bld_base
, &txq_emit_data
);
4704 /* Compute -0.5 / size. */
4705 for (c
= 0; c
< 2; c
++) {
4707 LLVMBuildExtractElement(builder
, txq_emit_data
.output
[0],
4708 LLVMConstInt(ctx
->i32
, c
, 0), "");
4709 half_texel
[c
] = LLVMBuildUIToFP(builder
, half_texel
[c
], ctx
->f32
, "");
4711 lp_build_emit_llvm_unary(&ctx
->bld_base
,
4712 TGSI_OPCODE_RCP
, half_texel
[c
]);
4713 half_texel
[c
] = LLVMBuildFMul(builder
, half_texel
[c
],
4714 LLVMConstReal(ctx
->f32
, -0.5), "");
4718 for (c
= 0; c
< 2; c
++) {
4720 LLVMValueRef index
= LLVMConstInt(ctx
->i32
, coord_vgpr_index
+ c
, 0);
4722 tmp
= LLVMBuildExtractElement(builder
, coord
, index
, "");
4723 tmp
= LLVMBuildBitCast(builder
, tmp
, ctx
->f32
, "");
4724 tmp
= LLVMBuildFAdd(builder
, tmp
, half_texel
[c
], "");
4725 tmp
= LLVMBuildBitCast(builder
, tmp
, ctx
->i32
, "");
4726 coord
= LLVMBuildInsertElement(builder
, coord
, tmp
, index
, "");
4732 static void build_tex_intrinsic(const struct lp_build_tgsi_action
*action
,
4733 struct lp_build_tgsi_context
*bld_base
,
4734 struct lp_build_emit_data
*emit_data
)
4736 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4737 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
4738 struct ac_image_args args
;
4739 unsigned opcode
= inst
->Instruction
.Opcode
;
4740 unsigned target
= inst
->Texture
.Texture
;
4742 if (target
== TGSI_TEXTURE_BUFFER
) {
4743 emit_data
->output
[emit_data
->chan
] =
4744 ac_build_buffer_load_format(&ctx
->ac
,
4752 memcpy(&args
, emit_data
->args
, sizeof(args
)); /* ugly */
4754 args
.opcode
= ac_image_sample
;
4755 args
.compare
= tgsi_is_shadow_target(target
);
4756 args
.offset
= inst
->Texture
.NumOffsets
> 0;
4759 case TGSI_OPCODE_TXF
:
4760 args
.opcode
= target
== TGSI_TEXTURE_2D_MSAA
||
4761 target
== TGSI_TEXTURE_2D_ARRAY_MSAA
?
4762 ac_image_load
: ac_image_load_mip
;
4763 args
.compare
= false;
4764 args
.offset
= false;
4766 case TGSI_OPCODE_LODQ
:
4767 args
.opcode
= ac_image_get_lod
;
4768 args
.compare
= false;
4769 args
.offset
= false;
4771 case TGSI_OPCODE_TEX
:
4772 case TGSI_OPCODE_TEX2
:
4773 case TGSI_OPCODE_TXP
:
4774 if (ctx
->type
!= PIPE_SHADER_FRAGMENT
)
4775 args
.level_zero
= true;
4777 case TGSI_OPCODE_TXB
:
4778 case TGSI_OPCODE_TXB2
:
4779 assert(ctx
->type
== PIPE_SHADER_FRAGMENT
);
4782 case TGSI_OPCODE_TXL
:
4783 case TGSI_OPCODE_TXL2
:
4786 case TGSI_OPCODE_TXD
:
4789 case TGSI_OPCODE_TG4
:
4790 args
.opcode
= ac_image_gather4
;
4791 args
.level_zero
= true;
4798 /* The hardware needs special lowering for Gather4 with integer formats. */
4799 if (opcode
== TGSI_OPCODE_TG4
) {
4800 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
4801 /* This will also work with non-constant indexing because of how
4802 * glsl_to_tgsi works and we intent to preserve that behavior.
4804 const unsigned src_idx
= 2;
4805 unsigned sampler
= inst
->Src
[src_idx
].Register
.Index
;
4807 assert(inst
->Src
[src_idx
].Register
.File
== TGSI_FILE_SAMPLER
);
4809 if (info
->sampler_type
[sampler
] == TGSI_RETURN_TYPE_SINT
||
4810 info
->sampler_type
[sampler
] == TGSI_RETURN_TYPE_UINT
)
4811 si_lower_gather4_integer(ctx
, &args
, target
);
4814 emit_data
->output
[emit_data
->chan
] =
4815 ac_build_image_opcode(&ctx
->ac
, &args
);
4818 static void si_llvm_emit_txqs(
4819 const struct lp_build_tgsi_action
*action
,
4820 struct lp_build_tgsi_context
*bld_base
,
4821 struct lp_build_emit_data
*emit_data
)
4823 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4824 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4825 LLVMBuilderRef builder
= gallivm
->builder
;
4826 LLVMValueRef res
, samples
;
4827 LLVMValueRef res_ptr
, samp_ptr
, fmask_ptr
= NULL
;
4829 tex_fetch_ptrs(bld_base
, emit_data
, &res_ptr
, &samp_ptr
, &fmask_ptr
);
4832 /* Read the samples from the descriptor directly. */
4833 res
= LLVMBuildBitCast(builder
, res_ptr
, ctx
->v8i32
, "");
4834 samples
= LLVMBuildExtractElement(
4836 lp_build_const_int32(gallivm
, 3), "");
4837 samples
= LLVMBuildLShr(builder
, samples
,
4838 lp_build_const_int32(gallivm
, 16), "");
4839 samples
= LLVMBuildAnd(builder
, samples
,
4840 lp_build_const_int32(gallivm
, 0xf), "");
4841 samples
= LLVMBuildShl(builder
, lp_build_const_int32(gallivm
, 1),
4844 emit_data
->output
[emit_data
->chan
] = samples
;
4847 static void si_llvm_emit_ddxy(
4848 const struct lp_build_tgsi_action
*action
,
4849 struct lp_build_tgsi_context
*bld_base
,
4850 struct lp_build_emit_data
*emit_data
)
4852 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4853 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4854 unsigned opcode
= emit_data
->info
->opcode
;
4859 if (opcode
== TGSI_OPCODE_DDX_FINE
)
4860 mask
= AC_TID_MASK_LEFT
;
4861 else if (opcode
== TGSI_OPCODE_DDY_FINE
)
4862 mask
= AC_TID_MASK_TOP
;
4864 mask
= AC_TID_MASK_TOP_LEFT
;
4866 /* for DDX we want to next X pixel, DDY next Y pixel. */
4867 idx
= (opcode
== TGSI_OPCODE_DDX
|| opcode
== TGSI_OPCODE_DDX_FINE
) ? 1 : 2;
4869 val
= LLVMBuildBitCast(gallivm
->builder
, emit_data
->args
[0], ctx
->i32
, "");
4870 val
= ac_build_ddxy(&ctx
->ac
, ctx
->screen
->has_ds_bpermute
,
4871 mask
, idx
, ctx
->lds
, val
);
4872 emit_data
->output
[emit_data
->chan
] = val
;
4876 * this takes an I,J coordinate pair,
4877 * and works out the X and Y derivatives.
4878 * it returns DDX(I), DDX(J), DDY(I), DDY(J).
4880 static LLVMValueRef
si_llvm_emit_ddxy_interp(
4881 struct lp_build_tgsi_context
*bld_base
,
4882 LLVMValueRef interp_ij
)
4884 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4885 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4886 LLVMValueRef result
[4], a
;
4889 for (i
= 0; i
< 2; i
++) {
4890 a
= LLVMBuildExtractElement(gallivm
->builder
, interp_ij
,
4891 LLVMConstInt(ctx
->i32
, i
, 0), "");
4892 result
[i
] = lp_build_emit_llvm_unary(bld_base
, TGSI_OPCODE_DDX
, a
);
4893 result
[2+i
] = lp_build_emit_llvm_unary(bld_base
, TGSI_OPCODE_DDY
, a
);
4896 return lp_build_gather_values(gallivm
, result
, 4);
4899 static void interp_fetch_args(
4900 struct lp_build_tgsi_context
*bld_base
,
4901 struct lp_build_emit_data
*emit_data
)
4903 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4904 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4905 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
4907 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_OFFSET
) {
4908 /* offset is in second src, first two channels */
4909 emit_data
->args
[0] = lp_build_emit_fetch(bld_base
,
4912 emit_data
->args
[1] = lp_build_emit_fetch(bld_base
,
4915 emit_data
->arg_count
= 2;
4916 } else if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_SAMPLE
) {
4917 LLVMValueRef sample_position
;
4918 LLVMValueRef sample_id
;
4919 LLVMValueRef halfval
= lp_build_const_float(gallivm
, 0.5f
);
4921 /* fetch sample ID, then fetch its sample position,
4922 * and place into first two channels.
4924 sample_id
= lp_build_emit_fetch(bld_base
,
4925 emit_data
->inst
, 1, TGSI_CHAN_X
);
4926 sample_id
= LLVMBuildBitCast(gallivm
->builder
, sample_id
,
4928 sample_position
= load_sample_position(ctx
, sample_id
);
4930 emit_data
->args
[0] = LLVMBuildExtractElement(gallivm
->builder
,
4932 lp_build_const_int32(gallivm
, 0), "");
4934 emit_data
->args
[0] = LLVMBuildFSub(gallivm
->builder
, emit_data
->args
[0], halfval
, "");
4935 emit_data
->args
[1] = LLVMBuildExtractElement(gallivm
->builder
,
4937 lp_build_const_int32(gallivm
, 1), "");
4938 emit_data
->args
[1] = LLVMBuildFSub(gallivm
->builder
, emit_data
->args
[1], halfval
, "");
4939 emit_data
->arg_count
= 2;
4943 static void build_interp_intrinsic(const struct lp_build_tgsi_action
*action
,
4944 struct lp_build_tgsi_context
*bld_base
,
4945 struct lp_build_emit_data
*emit_data
)
4947 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4948 struct si_shader
*shader
= ctx
->shader
;
4949 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4950 struct lp_build_context
*uint
= &bld_base
->uint_bld
;
4951 LLVMValueRef interp_param
;
4952 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
4953 int input_index
= inst
->Src
[0].Register
.Index
;
4956 LLVMValueRef attr_number
;
4957 LLVMValueRef params
= LLVMGetParam(ctx
->main_fn
, SI_PARAM_PRIM_MASK
);
4958 int interp_param_idx
;
4959 unsigned interp
= shader
->selector
->info
.input_interpolate
[input_index
];
4962 assert(inst
->Src
[0].Register
.File
== TGSI_FILE_INPUT
);
4964 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_OFFSET
||
4965 inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_SAMPLE
)
4966 location
= TGSI_INTERPOLATE_LOC_CENTER
;
4968 location
= TGSI_INTERPOLATE_LOC_CENTROID
;
4970 interp_param_idx
= lookup_interp_param_index(interp
, location
);
4971 if (interp_param_idx
== -1)
4973 else if (interp_param_idx
)
4974 interp_param
= LLVMGetParam(ctx
->main_fn
, interp_param_idx
);
4976 interp_param
= NULL
;
4978 attr_number
= lp_build_const_int32(gallivm
, input_index
);
4980 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_OFFSET
||
4981 inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_SAMPLE
) {
4982 LLVMValueRef ij_out
[2];
4983 LLVMValueRef ddxy_out
= si_llvm_emit_ddxy_interp(bld_base
, interp_param
);
4986 * take the I then J parameters, and the DDX/Y for it, and
4987 * calculate the IJ inputs for the interpolator.
4988 * temp1 = ddx * offset/sample.x + I;
4989 * interp_param.I = ddy * offset/sample.y + temp1;
4990 * temp1 = ddx * offset/sample.x + J;
4991 * interp_param.J = ddy * offset/sample.y + temp1;
4993 for (i
= 0; i
< 2; i
++) {
4994 LLVMValueRef ix_ll
= lp_build_const_int32(gallivm
, i
);
4995 LLVMValueRef iy_ll
= lp_build_const_int32(gallivm
, i
+ 2);
4996 LLVMValueRef ddx_el
= LLVMBuildExtractElement(gallivm
->builder
,
4997 ddxy_out
, ix_ll
, "");
4998 LLVMValueRef ddy_el
= LLVMBuildExtractElement(gallivm
->builder
,
4999 ddxy_out
, iy_ll
, "");
5000 LLVMValueRef interp_el
= LLVMBuildExtractElement(gallivm
->builder
,
5001 interp_param
, ix_ll
, "");
5002 LLVMValueRef temp1
, temp2
;
5004 interp_el
= LLVMBuildBitCast(gallivm
->builder
, interp_el
,
5007 temp1
= LLVMBuildFMul(gallivm
->builder
, ddx_el
, emit_data
->args
[0], "");
5009 temp1
= LLVMBuildFAdd(gallivm
->builder
, temp1
, interp_el
, "");
5011 temp2
= LLVMBuildFMul(gallivm
->builder
, ddy_el
, emit_data
->args
[1], "");
5013 ij_out
[i
] = LLVMBuildFAdd(gallivm
->builder
, temp2
, temp1
, "");
5015 interp_param
= lp_build_gather_values(bld_base
->base
.gallivm
, ij_out
, 2);
5018 for (chan
= 0; chan
< 4; chan
++) {
5019 LLVMValueRef llvm_chan
;
5022 schan
= tgsi_util_get_full_src_register_swizzle(&inst
->Src
[0], chan
);
5023 llvm_chan
= lp_build_const_int32(gallivm
, schan
);
5026 interp_param
= LLVMBuildBitCast(gallivm
->builder
,
5027 interp_param
, LLVMVectorType(ctx
->f32
, 2), "");
5028 LLVMValueRef i
= LLVMBuildExtractElement(
5029 gallivm
->builder
, interp_param
, uint
->zero
, "");
5030 LLVMValueRef j
= LLVMBuildExtractElement(
5031 gallivm
->builder
, interp_param
, uint
->one
, "");
5032 emit_data
->output
[chan
] = ac_build_fs_interp(&ctx
->ac
,
5033 llvm_chan
, attr_number
, params
,
5036 emit_data
->output
[chan
] = ac_build_fs_interp_mov(&ctx
->ac
,
5037 lp_build_const_int32(gallivm
, 2), /* P0 */
5038 llvm_chan
, attr_number
, params
);
5043 static unsigned si_llvm_get_stream(struct lp_build_tgsi_context
*bld_base
,
5044 struct lp_build_emit_data
*emit_data
)
5046 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
5047 struct tgsi_src_register src0
= emit_data
->inst
->Src
[0].Register
;
5051 assert(src0
.File
== TGSI_FILE_IMMEDIATE
);
5053 imm
= ctx
->imms
[src0
.Index
* TGSI_NUM_CHANNELS
+ src0
.SwizzleX
];
5054 stream
= LLVMConstIntGetZExtValue(imm
) & 0x3;
5058 /* Emit one vertex from the geometry shader */
5059 static void si_llvm_emit_vertex(
5060 const struct lp_build_tgsi_action
*action
,
5061 struct lp_build_tgsi_context
*bld_base
,
5062 struct lp_build_emit_data
*emit_data
)
5064 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
5065 struct lp_build_context
*uint
= &bld_base
->uint_bld
;
5066 struct si_shader
*shader
= ctx
->shader
;
5067 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
5068 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
5069 struct lp_build_if_state if_state
;
5070 LLVMValueRef soffset
= LLVMGetParam(ctx
->main_fn
,
5071 SI_PARAM_GS2VS_OFFSET
);
5072 LLVMValueRef gs_next_vertex
;
5073 LLVMValueRef can_emit
, kill
;
5074 unsigned chan
, offset
;
5078 stream
= si_llvm_get_stream(bld_base
, emit_data
);
5080 /* Write vertex attribute values to GSVS ring */
5081 gs_next_vertex
= LLVMBuildLoad(gallivm
->builder
,
5082 ctx
->gs_next_vertex
[stream
],
5085 /* If this thread has already emitted the declared maximum number of
5086 * vertices, skip the write: excessive vertex emissions are not
5087 * supposed to have any effect.
5089 * If the shader has no writes to memory, kill it instead. This skips
5090 * further memory loads and may allow LLVM to skip to the end
5093 can_emit
= LLVMBuildICmp(gallivm
->builder
, LLVMIntULT
, gs_next_vertex
,
5094 lp_build_const_int32(gallivm
,
5095 shader
->selector
->gs_max_out_vertices
), "");
5097 bool use_kill
= !info
->writes_memory
;
5099 kill
= lp_build_select(&bld_base
->base
, can_emit
,
5100 lp_build_const_float(gallivm
, 1.0f
),
5101 lp_build_const_float(gallivm
, -1.0f
));
5103 ac_build_kill(&ctx
->ac
, kill
);
5105 lp_build_if(&if_state
, gallivm
, can_emit
);
5109 for (i
= 0; i
< info
->num_outputs
; i
++) {
5110 LLVMValueRef
*out_ptr
= ctx
->outputs
[i
];
5112 for (chan
= 0; chan
< 4; chan
++) {
5113 if (!(info
->output_usagemask
[i
] & (1 << chan
)) ||
5114 ((info
->output_streams
[i
] >> (2 * chan
)) & 3) != stream
)
5117 LLVMValueRef out_val
= LLVMBuildLoad(gallivm
->builder
, out_ptr
[chan
], "");
5118 LLVMValueRef voffset
=
5119 lp_build_const_int32(gallivm
, offset
*
5120 shader
->selector
->gs_max_out_vertices
);
5123 voffset
= lp_build_add(uint
, voffset
, gs_next_vertex
);
5124 voffset
= lp_build_mul_imm(uint
, voffset
, 4);
5126 out_val
= LLVMBuildBitCast(gallivm
->builder
, out_val
, ctx
->i32
, "");
5128 ac_build_buffer_store_dword(&ctx
->ac
,
5129 ctx
->gsvs_ring
[stream
],
5131 voffset
, soffset
, 0,
5136 gs_next_vertex
= lp_build_add(uint
, gs_next_vertex
,
5137 lp_build_const_int32(gallivm
, 1));
5139 LLVMBuildStore(gallivm
->builder
, gs_next_vertex
, ctx
->gs_next_vertex
[stream
]);
5141 /* Signal vertex emission */
5142 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_EMIT
| AC_SENDMSG_GS
| (stream
<< 8),
5143 LLVMGetParam(ctx
->main_fn
, SI_PARAM_GS_WAVE_ID
));
5145 lp_build_endif(&if_state
);
5148 /* Cut one primitive from the geometry shader */
5149 static void si_llvm_emit_primitive(
5150 const struct lp_build_tgsi_action
*action
,
5151 struct lp_build_tgsi_context
*bld_base
,
5152 struct lp_build_emit_data
*emit_data
)
5154 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
5157 /* Signal primitive cut */
5158 stream
= si_llvm_get_stream(bld_base
, emit_data
);
5159 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_CUT
| AC_SENDMSG_GS
| (stream
<< 8),
5160 LLVMGetParam(ctx
->main_fn
, SI_PARAM_GS_WAVE_ID
));
5163 static void si_llvm_emit_barrier(const struct lp_build_tgsi_action
*action
,
5164 struct lp_build_tgsi_context
*bld_base
,
5165 struct lp_build_emit_data
*emit_data
)
5167 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
5168 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
5170 /* SI only (thanks to a hw bug workaround):
5171 * The real barrier instruction isn’t needed, because an entire patch
5172 * always fits into a single wave.
5174 if (HAVE_LLVM
>= 0x0309 &&
5175 ctx
->screen
->b
.chip_class
== SI
&&
5176 ctx
->type
== PIPE_SHADER_TESS_CTRL
) {
5177 emit_waitcnt(ctx
, LGKM_CNT
& VM_CNT
);
5181 lp_build_intrinsic(gallivm
->builder
,
5182 HAVE_LLVM
>= 0x0309 ? "llvm.amdgcn.s.barrier"
5183 : "llvm.AMDGPU.barrier.local",
5184 ctx
->voidt
, NULL
, 0, LP_FUNC_ATTR_CONVERGENT
);
5187 static const struct lp_build_tgsi_action tex_action
= {
5188 .fetch_args
= tex_fetch_args
,
5189 .emit
= build_tex_intrinsic
,
5192 static const struct lp_build_tgsi_action interp_action
= {
5193 .fetch_args
= interp_fetch_args
,
5194 .emit
= build_interp_intrinsic
,
5197 static void si_create_function(struct si_shader_context
*ctx
,
5199 LLVMTypeRef
*returns
, unsigned num_returns
,
5200 LLVMTypeRef
*params
, unsigned num_params
,
5205 si_llvm_create_func(ctx
, name
, returns
, num_returns
,
5206 params
, num_params
);
5207 si_llvm_shader_type(ctx
->main_fn
, ctx
->type
);
5208 ctx
->return_value
= LLVMGetUndef(ctx
->return_type
);
5210 for (i
= 0; i
<= last_sgpr
; ++i
) {
5211 LLVMValueRef P
= LLVMGetParam(ctx
->main_fn
, i
);
5213 /* The combination of:
5217 * allows the optimization passes to move loads and reduces
5218 * SGPR spilling significantly.
5220 if (LLVMGetTypeKind(LLVMTypeOf(P
)) == LLVMPointerTypeKind
) {
5221 lp_add_function_attr(ctx
->main_fn
, i
+ 1, LP_FUNC_ATTR_BYVAL
);
5222 lp_add_function_attr(ctx
->main_fn
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
5223 lp_add_attr_dereferenceable(P
, UINT64_MAX
);
5225 lp_add_function_attr(ctx
->main_fn
, i
+ 1, LP_FUNC_ATTR_INREG
);
5228 LLVMAddTargetDependentFunctionAttr(ctx
->main_fn
,
5229 "no-signed-zeros-fp-math",
5232 if (ctx
->screen
->b
.debug_flags
& DBG_UNSAFE_MATH
) {
5233 /* These were copied from some LLVM test. */
5234 LLVMAddTargetDependentFunctionAttr(ctx
->main_fn
,
5235 "less-precise-fpmad",
5237 LLVMAddTargetDependentFunctionAttr(ctx
->main_fn
,
5240 LLVMAddTargetDependentFunctionAttr(ctx
->main_fn
,
5243 LLVMAddTargetDependentFunctionAttr(ctx
->main_fn
,
5249 static void declare_streamout_params(struct si_shader_context
*ctx
,
5250 struct pipe_stream_output_info
*so
,
5251 LLVMTypeRef
*params
, LLVMTypeRef i32
,
5252 unsigned *num_params
)
5256 /* Streamout SGPRs. */
5257 if (so
->num_outputs
) {
5258 if (ctx
->type
!= PIPE_SHADER_TESS_EVAL
)
5259 params
[ctx
->param_streamout_config
= (*num_params
)++] = i32
;
5261 ctx
->param_streamout_config
= *num_params
- 1;
5263 params
[ctx
->param_streamout_write_index
= (*num_params
)++] = i32
;
5265 /* A streamout buffer offset is loaded if the stride is non-zero. */
5266 for (i
= 0; i
< 4; i
++) {
5270 params
[ctx
->param_streamout_offset
[i
] = (*num_params
)++] = i32
;
5274 static unsigned llvm_get_type_size(LLVMTypeRef type
)
5276 LLVMTypeKind kind
= LLVMGetTypeKind(type
);
5279 case LLVMIntegerTypeKind
:
5280 return LLVMGetIntTypeWidth(type
) / 8;
5281 case LLVMFloatTypeKind
:
5283 case LLVMPointerTypeKind
:
5285 case LLVMVectorTypeKind
:
5286 return LLVMGetVectorSize(type
) *
5287 llvm_get_type_size(LLVMGetElementType(type
));
5288 case LLVMArrayTypeKind
:
5289 return LLVMGetArrayLength(type
) *
5290 llvm_get_type_size(LLVMGetElementType(type
));
5297 static void declare_tess_lds(struct si_shader_context
*ctx
)
5299 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
5300 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
5301 struct lp_build_context
*uint
= &bld_base
->uint_bld
;
5303 unsigned lds_size
= ctx
->screen
->b
.chip_class
>= CIK
? 65536 : 32768;
5304 ctx
->lds
= LLVMBuildIntToPtr(gallivm
->builder
, uint
->zero
,
5305 LLVMPointerType(LLVMArrayType(ctx
->i32
, lds_size
/ 4), LOCAL_ADDR_SPACE
),
5309 static unsigned si_get_max_workgroup_size(struct si_shader
*shader
)
5311 const unsigned *properties
= shader
->selector
->info
.properties
;
5312 unsigned max_work_group_size
=
5313 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
] *
5314 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT
] *
5315 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH
];
5317 if (!max_work_group_size
) {
5318 /* This is a variable group size compute shader,
5319 * compile it for the maximum possible group size.
5321 max_work_group_size
= SI_MAX_VARIABLE_THREADS_PER_BLOCK
;
5323 return max_work_group_size
;
5326 static void create_function(struct si_shader_context
*ctx
)
5328 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
5329 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
5330 struct si_shader
*shader
= ctx
->shader
;
5331 LLVMTypeRef params
[SI_NUM_PARAMS
+ SI_MAX_ATTRIBS
], v3i32
;
5332 LLVMTypeRef returns
[16+32*4];
5333 unsigned i
, last_sgpr
, num_params
, num_return_sgprs
;
5334 unsigned num_returns
= 0;
5335 unsigned num_prolog_vgprs
= 0;
5337 v3i32
= LLVMVectorType(ctx
->i32
, 3);
5339 params
[SI_PARAM_RW_BUFFERS
] = const_array(ctx
->v16i8
, SI_NUM_RW_BUFFERS
);
5340 params
[SI_PARAM_CONST_BUFFERS
] = const_array(ctx
->v16i8
, SI_NUM_CONST_BUFFERS
);
5341 params
[SI_PARAM_SAMPLERS
] = const_array(ctx
->v8i32
, SI_NUM_SAMPLERS
);
5342 params
[SI_PARAM_IMAGES
] = const_array(ctx
->v8i32
, SI_NUM_IMAGES
);
5343 params
[SI_PARAM_SHADER_BUFFERS
] = const_array(ctx
->v4i32
, SI_NUM_SHADER_BUFFERS
);
5345 switch (ctx
->type
) {
5346 case PIPE_SHADER_VERTEX
:
5347 params
[SI_PARAM_VERTEX_BUFFERS
] = const_array(ctx
->v16i8
, SI_MAX_ATTRIBS
);
5348 params
[SI_PARAM_BASE_VERTEX
] = ctx
->i32
;
5349 params
[SI_PARAM_START_INSTANCE
] = ctx
->i32
;
5350 params
[SI_PARAM_DRAWID
] = ctx
->i32
;
5351 num_params
= SI_PARAM_DRAWID
+1;
5353 if (shader
->key
.as_es
) {
5354 params
[ctx
->param_es2gs_offset
= num_params
++] = ctx
->i32
;
5355 } else if (shader
->key
.as_ls
) {
5356 params
[SI_PARAM_LS_OUT_LAYOUT
] = ctx
->i32
;
5357 num_params
= SI_PARAM_LS_OUT_LAYOUT
+1;
5359 if (shader
->is_gs_copy_shader
) {
5360 num_params
= SI_PARAM_RW_BUFFERS
+1;
5362 params
[SI_PARAM_VS_STATE_BITS
] = ctx
->i32
;
5363 num_params
= SI_PARAM_VS_STATE_BITS
+1;
5366 /* The locations of the other parameters are assigned dynamically. */
5367 declare_streamout_params(ctx
, &shader
->selector
->so
,
5368 params
, ctx
->i32
, &num_params
);
5371 last_sgpr
= num_params
-1;
5374 params
[ctx
->param_vertex_id
= num_params
++] = ctx
->i32
;
5375 params
[ctx
->param_rel_auto_id
= num_params
++] = ctx
->i32
;
5376 params
[ctx
->param_vs_prim_id
= num_params
++] = ctx
->i32
;
5377 params
[ctx
->param_instance_id
= num_params
++] = ctx
->i32
;
5379 if (!shader
->is_gs_copy_shader
) {
5380 /* Vertex load indices. */
5381 ctx
->param_vertex_index0
= num_params
;
5383 for (i
= 0; i
< shader
->selector
->info
.num_inputs
; i
++)
5384 params
[num_params
++] = ctx
->i32
;
5386 num_prolog_vgprs
+= shader
->selector
->info
.num_inputs
;
5388 /* PrimitiveID output. */
5389 if (!shader
->key
.as_es
&& !shader
->key
.as_ls
)
5390 for (i
= 0; i
<= VS_EPILOG_PRIMID_LOC
; i
++)
5391 returns
[num_returns
++] = ctx
->f32
;
5395 case PIPE_SHADER_TESS_CTRL
:
5396 params
[SI_PARAM_TCS_OFFCHIP_LAYOUT
] = ctx
->i32
;
5397 params
[SI_PARAM_TCS_OUT_OFFSETS
] = ctx
->i32
;
5398 params
[SI_PARAM_TCS_OUT_LAYOUT
] = ctx
->i32
;
5399 params
[SI_PARAM_TCS_IN_LAYOUT
] = ctx
->i32
;
5400 params
[ctx
->param_oc_lds
= SI_PARAM_TCS_OC_LDS
] = ctx
->i32
;
5401 params
[SI_PARAM_TESS_FACTOR_OFFSET
] = ctx
->i32
;
5402 last_sgpr
= SI_PARAM_TESS_FACTOR_OFFSET
;
5405 params
[SI_PARAM_PATCH_ID
] = ctx
->i32
;
5406 params
[SI_PARAM_REL_IDS
] = ctx
->i32
;
5407 num_params
= SI_PARAM_REL_IDS
+1;
5409 /* SI_PARAM_TCS_OC_LDS and PARAM_TESS_FACTOR_OFFSET are
5410 * placed after the user SGPRs.
5412 for (i
= 0; i
< SI_TCS_NUM_USER_SGPR
+ 2; i
++)
5413 returns
[num_returns
++] = ctx
->i32
; /* SGPRs */
5415 for (i
= 0; i
< 3; i
++)
5416 returns
[num_returns
++] = ctx
->f32
; /* VGPRs */
5419 case PIPE_SHADER_TESS_EVAL
:
5420 params
[SI_PARAM_TCS_OFFCHIP_LAYOUT
] = ctx
->i32
;
5421 num_params
= SI_PARAM_TCS_OFFCHIP_LAYOUT
+1;
5423 if (shader
->key
.as_es
) {
5424 params
[ctx
->param_oc_lds
= num_params
++] = ctx
->i32
;
5425 params
[num_params
++] = ctx
->i32
;
5426 params
[ctx
->param_es2gs_offset
= num_params
++] = ctx
->i32
;
5428 params
[num_params
++] = ctx
->i32
;
5429 declare_streamout_params(ctx
, &shader
->selector
->so
,
5430 params
, ctx
->i32
, &num_params
);
5431 params
[ctx
->param_oc_lds
= num_params
++] = ctx
->i32
;
5433 last_sgpr
= num_params
- 1;
5436 params
[ctx
->param_tes_u
= num_params
++] = ctx
->f32
;
5437 params
[ctx
->param_tes_v
= num_params
++] = ctx
->f32
;
5438 params
[ctx
->param_tes_rel_patch_id
= num_params
++] = ctx
->i32
;
5439 params
[ctx
->param_tes_patch_id
= num_params
++] = ctx
->i32
;
5441 /* PrimitiveID output. */
5442 if (!shader
->key
.as_es
)
5443 for (i
= 0; i
<= VS_EPILOG_PRIMID_LOC
; i
++)
5444 returns
[num_returns
++] = ctx
->f32
;
5447 case PIPE_SHADER_GEOMETRY
:
5448 params
[SI_PARAM_GS2VS_OFFSET
] = ctx
->i32
;
5449 params
[SI_PARAM_GS_WAVE_ID
] = ctx
->i32
;
5450 last_sgpr
= SI_PARAM_GS_WAVE_ID
;
5453 params
[SI_PARAM_VTX0_OFFSET
] = ctx
->i32
;
5454 params
[SI_PARAM_VTX1_OFFSET
] = ctx
->i32
;
5455 params
[SI_PARAM_PRIMITIVE_ID
] = ctx
->i32
;
5456 params
[SI_PARAM_VTX2_OFFSET
] = ctx
->i32
;
5457 params
[SI_PARAM_VTX3_OFFSET
] = ctx
->i32
;
5458 params
[SI_PARAM_VTX4_OFFSET
] = ctx
->i32
;
5459 params
[SI_PARAM_VTX5_OFFSET
] = ctx
->i32
;
5460 params
[SI_PARAM_GS_INSTANCE_ID
] = ctx
->i32
;
5461 num_params
= SI_PARAM_GS_INSTANCE_ID
+1;
5464 case PIPE_SHADER_FRAGMENT
:
5465 params
[SI_PARAM_ALPHA_REF
] = ctx
->f32
;
5466 params
[SI_PARAM_PRIM_MASK
] = ctx
->i32
;
5467 last_sgpr
= SI_PARAM_PRIM_MASK
;
5468 params
[SI_PARAM_PERSP_SAMPLE
] = ctx
->v2i32
;
5469 params
[SI_PARAM_PERSP_CENTER
] = ctx
->v2i32
;
5470 params
[SI_PARAM_PERSP_CENTROID
] = ctx
->v2i32
;
5471 params
[SI_PARAM_PERSP_PULL_MODEL
] = v3i32
;
5472 params
[SI_PARAM_LINEAR_SAMPLE
] = ctx
->v2i32
;
5473 params
[SI_PARAM_LINEAR_CENTER
] = ctx
->v2i32
;
5474 params
[SI_PARAM_LINEAR_CENTROID
] = ctx
->v2i32
;
5475 params
[SI_PARAM_LINE_STIPPLE_TEX
] = ctx
->f32
;
5476 params
[SI_PARAM_POS_X_FLOAT
] = ctx
->f32
;
5477 params
[SI_PARAM_POS_Y_FLOAT
] = ctx
->f32
;
5478 params
[SI_PARAM_POS_Z_FLOAT
] = ctx
->f32
;
5479 params
[SI_PARAM_POS_W_FLOAT
] = ctx
->f32
;
5480 params
[SI_PARAM_FRONT_FACE
] = ctx
->i32
;
5481 shader
->info
.face_vgpr_index
= 20;
5482 params
[SI_PARAM_ANCILLARY
] = ctx
->i32
;
5483 params
[SI_PARAM_SAMPLE_COVERAGE
] = ctx
->f32
;
5484 params
[SI_PARAM_POS_FIXED_PT
] = ctx
->i32
;
5485 num_params
= SI_PARAM_POS_FIXED_PT
+1;
5487 /* Color inputs from the prolog. */
5488 if (shader
->selector
->info
.colors_read
) {
5489 unsigned num_color_elements
=
5490 util_bitcount(shader
->selector
->info
.colors_read
);
5492 assert(num_params
+ num_color_elements
<= ARRAY_SIZE(params
));
5493 for (i
= 0; i
< num_color_elements
; i
++)
5494 params
[num_params
++] = ctx
->f32
;
5496 num_prolog_vgprs
+= num_color_elements
;
5499 /* Outputs for the epilog. */
5500 num_return_sgprs
= SI_SGPR_ALPHA_REF
+ 1;
5503 util_bitcount(shader
->selector
->info
.colors_written
) * 4 +
5504 shader
->selector
->info
.writes_z
+
5505 shader
->selector
->info
.writes_stencil
+
5506 shader
->selector
->info
.writes_samplemask
+
5507 1 /* SampleMaskIn */;
5509 num_returns
= MAX2(num_returns
,
5511 PS_EPILOG_SAMPLEMASK_MIN_LOC
+ 1);
5513 for (i
= 0; i
< num_return_sgprs
; i
++)
5514 returns
[i
] = ctx
->i32
;
5515 for (; i
< num_returns
; i
++)
5516 returns
[i
] = ctx
->f32
;
5519 case PIPE_SHADER_COMPUTE
:
5520 params
[SI_PARAM_GRID_SIZE
] = v3i32
;
5521 params
[SI_PARAM_BLOCK_SIZE
] = v3i32
;
5522 params
[SI_PARAM_BLOCK_ID
] = v3i32
;
5523 last_sgpr
= SI_PARAM_BLOCK_ID
;
5525 params
[SI_PARAM_THREAD_ID
] = v3i32
;
5526 num_params
= SI_PARAM_THREAD_ID
+ 1;
5529 assert(0 && "unimplemented shader");
5533 assert(num_params
<= ARRAY_SIZE(params
));
5535 si_create_function(ctx
, "main", returns
, num_returns
, params
,
5536 num_params
, last_sgpr
);
5538 /* Reserve register locations for VGPR inputs the PS prolog may need. */
5539 if (ctx
->type
== PIPE_SHADER_FRAGMENT
&&
5540 ctx
->separate_prolog
) {
5541 si_llvm_add_attribute(ctx
->main_fn
,
5542 "InitialPSInputAddr",
5543 S_0286D0_PERSP_SAMPLE_ENA(1) |
5544 S_0286D0_PERSP_CENTER_ENA(1) |
5545 S_0286D0_PERSP_CENTROID_ENA(1) |
5546 S_0286D0_LINEAR_SAMPLE_ENA(1) |
5547 S_0286D0_LINEAR_CENTER_ENA(1) |
5548 S_0286D0_LINEAR_CENTROID_ENA(1) |
5549 S_0286D0_FRONT_FACE_ENA(1) |
5550 S_0286D0_POS_FIXED_PT_ENA(1));
5551 } else if (ctx
->type
== PIPE_SHADER_COMPUTE
) {
5552 si_llvm_add_attribute(ctx
->main_fn
,
5553 "amdgpu-max-work-group-size",
5554 si_get_max_workgroup_size(shader
));
5557 shader
->info
.num_input_sgprs
= 0;
5558 shader
->info
.num_input_vgprs
= 0;
5560 for (i
= 0; i
<= last_sgpr
; ++i
)
5561 shader
->info
.num_input_sgprs
+= llvm_get_type_size(params
[i
]) / 4;
5563 for (; i
< num_params
; ++i
)
5564 shader
->info
.num_input_vgprs
+= llvm_get_type_size(params
[i
]) / 4;
5566 assert(shader
->info
.num_input_vgprs
>= num_prolog_vgprs
);
5567 shader
->info
.num_input_vgprs
-= num_prolog_vgprs
;
5569 if (!ctx
->screen
->has_ds_bpermute
&&
5571 (bld_base
->info
->opcode_count
[TGSI_OPCODE_DDX
] > 0 ||
5572 bld_base
->info
->opcode_count
[TGSI_OPCODE_DDY
] > 0 ||
5573 bld_base
->info
->opcode_count
[TGSI_OPCODE_DDX_FINE
] > 0 ||
5574 bld_base
->info
->opcode_count
[TGSI_OPCODE_DDY_FINE
] > 0 ||
5575 bld_base
->info
->opcode_count
[TGSI_OPCODE_INTERP_OFFSET
] > 0 ||
5576 bld_base
->info
->opcode_count
[TGSI_OPCODE_INTERP_SAMPLE
] > 0))
5578 LLVMAddGlobalInAddressSpace(gallivm
->module
,
5579 LLVMArrayType(ctx
->i32
, 64),
5583 if ((ctx
->type
== PIPE_SHADER_VERTEX
&& shader
->key
.as_ls
) ||
5584 ctx
->type
== PIPE_SHADER_TESS_CTRL
)
5585 declare_tess_lds(ctx
);
5589 * Load ESGS and GSVS ring buffer resource descriptors and save the variables
5592 static void preload_ring_buffers(struct si_shader_context
*ctx
)
5594 struct gallivm_state
*gallivm
= ctx
->bld_base
.base
.gallivm
;
5595 LLVMBuilderRef builder
= gallivm
->builder
;
5597 LLVMValueRef buf_ptr
= LLVMGetParam(ctx
->main_fn
,
5598 SI_PARAM_RW_BUFFERS
);
5600 if ((ctx
->type
== PIPE_SHADER_VERTEX
&&
5601 ctx
->shader
->key
.as_es
) ||
5602 (ctx
->type
== PIPE_SHADER_TESS_EVAL
&&
5603 ctx
->shader
->key
.as_es
) ||
5604 ctx
->type
== PIPE_SHADER_GEOMETRY
) {
5606 ctx
->type
== PIPE_SHADER_GEOMETRY
? SI_GS_RING_ESGS
5608 LLVMValueRef offset
= lp_build_const_int32(gallivm
, ring
);
5611 ac_build_indexed_load_const(&ctx
->ac
, buf_ptr
, offset
);
5614 if (ctx
->shader
->is_gs_copy_shader
) {
5615 LLVMValueRef offset
= lp_build_const_int32(gallivm
, SI_RING_GSVS
);
5618 ac_build_indexed_load_const(&ctx
->ac
, buf_ptr
, offset
);
5619 } else if (ctx
->type
== PIPE_SHADER_GEOMETRY
) {
5620 const struct si_shader_selector
*sel
= ctx
->shader
->selector
;
5621 struct lp_build_context
*uint
= &ctx
->bld_base
.uint_bld
;
5622 LLVMValueRef offset
= lp_build_const_int32(gallivm
, SI_RING_GSVS
);
5623 LLVMValueRef base_ring
;
5625 base_ring
= ac_build_indexed_load_const(&ctx
->ac
, buf_ptr
, offset
);
5627 /* The conceptual layout of the GSVS ring is
5628 * v0c0 .. vLv0 v0c1 .. vLc1 ..
5629 * but the real memory layout is swizzled across
5631 * t0v0c0 .. t15v0c0 t0v1c0 .. t15v1c0 ... t15vLcL
5633 * Override the buffer descriptor accordingly.
5635 LLVMTypeRef v2i64
= LLVMVectorType(ctx
->i64
, 2);
5636 uint64_t stream_offset
= 0;
5638 for (unsigned stream
= 0; stream
< 4; ++stream
) {
5639 unsigned num_components
;
5641 unsigned num_records
;
5642 LLVMValueRef ring
, tmp
;
5644 num_components
= sel
->info
.num_stream_output_components
[stream
];
5645 if (!num_components
)
5648 stride
= 4 * num_components
* sel
->gs_max_out_vertices
;
5650 /* Limit on the stride field for <= CIK. */
5651 assert(stride
< (1 << 14));
5655 ring
= LLVMBuildBitCast(builder
, base_ring
, v2i64
, "");
5656 tmp
= LLVMBuildExtractElement(builder
, ring
, uint
->zero
, "");
5657 tmp
= LLVMBuildAdd(builder
, tmp
,
5658 LLVMConstInt(ctx
->i64
,
5659 stream_offset
, 0), "");
5660 stream_offset
+= stride
* 64;
5662 ring
= LLVMBuildInsertElement(builder
, ring
, tmp
, uint
->zero
, "");
5663 ring
= LLVMBuildBitCast(builder
, ring
, ctx
->v4i32
, "");
5664 tmp
= LLVMBuildExtractElement(builder
, ring
, uint
->one
, "");
5665 tmp
= LLVMBuildOr(builder
, tmp
,
5666 LLVMConstInt(ctx
->i32
,
5667 S_008F04_STRIDE(stride
) |
5668 S_008F04_SWIZZLE_ENABLE(1), 0), "");
5669 ring
= LLVMBuildInsertElement(builder
, ring
, tmp
, uint
->one
, "");
5670 ring
= LLVMBuildInsertElement(builder
, ring
,
5671 LLVMConstInt(ctx
->i32
, num_records
, 0),
5672 LLVMConstInt(ctx
->i32
, 2, 0), "");
5673 ring
= LLVMBuildInsertElement(builder
, ring
,
5674 LLVMConstInt(ctx
->i32
,
5675 S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
5676 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
5677 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
5678 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
5679 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
5680 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
5681 S_008F0C_ELEMENT_SIZE(1) | /* element_size = 4 (bytes) */
5682 S_008F0C_INDEX_STRIDE(1) | /* index_stride = 16 (elements) */
5683 S_008F0C_ADD_TID_ENABLE(1),
5685 LLVMConstInt(ctx
->i32
, 3, 0), "");
5686 ring
= LLVMBuildBitCast(builder
, ring
, ctx
->v16i8
, "");
5688 ctx
->gsvs_ring
[stream
] = ring
;
5693 static void si_llvm_emit_polygon_stipple(struct si_shader_context
*ctx
,
5694 LLVMValueRef param_rw_buffers
,
5695 unsigned param_pos_fixed_pt
)
5697 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
5698 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
5699 LLVMBuilderRef builder
= gallivm
->builder
;
5700 LLVMValueRef slot
, desc
, offset
, row
, bit
, address
[2];
5702 /* Use the fixed-point gl_FragCoord input.
5703 * Since the stipple pattern is 32x32 and it repeats, just get 5 bits
5704 * per coordinate to get the repeating effect.
5706 address
[0] = unpack_param(ctx
, param_pos_fixed_pt
, 0, 5);
5707 address
[1] = unpack_param(ctx
, param_pos_fixed_pt
, 16, 5);
5709 /* Load the buffer descriptor. */
5710 slot
= lp_build_const_int32(gallivm
, SI_PS_CONST_POLY_STIPPLE
);
5711 desc
= ac_build_indexed_load_const(&ctx
->ac
, param_rw_buffers
, slot
);
5713 /* The stipple pattern is 32x32, each row has 32 bits. */
5714 offset
= LLVMBuildMul(builder
, address
[1],
5715 LLVMConstInt(ctx
->i32
, 4, 0), "");
5716 row
= buffer_load_const(ctx
, desc
, offset
);
5717 row
= LLVMBuildBitCast(builder
, row
, ctx
->i32
, "");
5718 bit
= LLVMBuildLShr(builder
, row
, address
[0], "");
5719 bit
= LLVMBuildTrunc(builder
, bit
, ctx
->i1
, "");
5721 /* The intrinsic kills the thread if arg < 0. */
5722 bit
= LLVMBuildSelect(builder
, bit
, LLVMConstReal(ctx
->f32
, 0),
5723 LLVMConstReal(ctx
->f32
, -1), "");
5724 ac_build_kill(&ctx
->ac
, bit
);
5727 void si_shader_binary_read_config(struct ac_shader_binary
*binary
,
5728 struct si_shader_config
*conf
,
5729 unsigned symbol_offset
)
5732 const unsigned char *config
=
5733 ac_shader_binary_config_start(binary
, symbol_offset
);
5734 bool really_needs_scratch
= false;
5736 /* LLVM adds SGPR spills to the scratch size.
5737 * Find out if we really need the scratch buffer.
5739 for (i
= 0; i
< binary
->reloc_count
; i
++) {
5740 const struct ac_shader_reloc
*reloc
= &binary
->relocs
[i
];
5742 if (!strcmp(scratch_rsrc_dword0_symbol
, reloc
->name
) ||
5743 !strcmp(scratch_rsrc_dword1_symbol
, reloc
->name
)) {
5744 really_needs_scratch
= true;
5749 /* XXX: We may be able to emit some of these values directly rather than
5750 * extracting fields to be emitted later.
5753 for (i
= 0; i
< binary
->config_size_per_symbol
; i
+= 8) {
5754 unsigned reg
= util_le32_to_cpu(*(uint32_t*)(config
+ i
));
5755 unsigned value
= util_le32_to_cpu(*(uint32_t*)(config
+ i
+ 4));
5757 case R_00B028_SPI_SHADER_PGM_RSRC1_PS
:
5758 case R_00B128_SPI_SHADER_PGM_RSRC1_VS
:
5759 case R_00B228_SPI_SHADER_PGM_RSRC1_GS
:
5760 case R_00B848_COMPUTE_PGM_RSRC1
:
5761 conf
->num_sgprs
= MAX2(conf
->num_sgprs
, (G_00B028_SGPRS(value
) + 1) * 8);
5762 conf
->num_vgprs
= MAX2(conf
->num_vgprs
, (G_00B028_VGPRS(value
) + 1) * 4);
5763 conf
->float_mode
= G_00B028_FLOAT_MODE(value
);
5764 conf
->rsrc1
= value
;
5766 case R_00B02C_SPI_SHADER_PGM_RSRC2_PS
:
5767 conf
->lds_size
= MAX2(conf
->lds_size
, G_00B02C_EXTRA_LDS_SIZE(value
));
5769 case R_00B84C_COMPUTE_PGM_RSRC2
:
5770 conf
->lds_size
= MAX2(conf
->lds_size
, G_00B84C_LDS_SIZE(value
));
5771 conf
->rsrc2
= value
;
5773 case R_0286CC_SPI_PS_INPUT_ENA
:
5774 conf
->spi_ps_input_ena
= value
;
5776 case R_0286D0_SPI_PS_INPUT_ADDR
:
5777 conf
->spi_ps_input_addr
= value
;
5779 case R_0286E8_SPI_TMPRING_SIZE
:
5780 case R_00B860_COMPUTE_TMPRING_SIZE
:
5781 /* WAVESIZE is in units of 256 dwords. */
5782 if (really_needs_scratch
)
5783 conf
->scratch_bytes_per_wave
=
5784 G_00B860_WAVESIZE(value
) * 256 * 4;
5786 case 0x4: /* SPILLED_SGPRS */
5787 conf
->spilled_sgprs
= value
;
5789 case 0x8: /* SPILLED_VGPRS */
5790 conf
->spilled_vgprs
= value
;
5794 static bool printed
;
5797 fprintf(stderr
, "Warning: LLVM emitted unknown "
5798 "config register: 0x%x\n", reg
);
5806 if (!conf
->spi_ps_input_addr
)
5807 conf
->spi_ps_input_addr
= conf
->spi_ps_input_ena
;
5810 void si_shader_apply_scratch_relocs(struct si_context
*sctx
,
5811 struct si_shader
*shader
,
5812 struct si_shader_config
*config
,
5813 uint64_t scratch_va
)
5816 uint32_t scratch_rsrc_dword0
= scratch_va
;
5817 uint32_t scratch_rsrc_dword1
=
5818 S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32);
5820 /* Enable scratch coalescing if LLVM sets ELEMENT_SIZE & INDEX_STRIDE
5823 if (HAVE_LLVM
>= 0x0309)
5824 scratch_rsrc_dword1
|= S_008F04_SWIZZLE_ENABLE(1);
5826 scratch_rsrc_dword1
|=
5827 S_008F04_STRIDE(config
->scratch_bytes_per_wave
/ 64);
5829 for (i
= 0 ; i
< shader
->binary
.reloc_count
; i
++) {
5830 const struct ac_shader_reloc
*reloc
=
5831 &shader
->binary
.relocs
[i
];
5832 if (!strcmp(scratch_rsrc_dword0_symbol
, reloc
->name
)) {
5833 util_memcpy_cpu_to_le32(shader
->binary
.code
+ reloc
->offset
,
5834 &scratch_rsrc_dword0
, 4);
5835 } else if (!strcmp(scratch_rsrc_dword1_symbol
, reloc
->name
)) {
5836 util_memcpy_cpu_to_le32(shader
->binary
.code
+ reloc
->offset
,
5837 &scratch_rsrc_dword1
, 4);
5842 static unsigned si_get_shader_binary_size(struct si_shader
*shader
)
5844 unsigned size
= shader
->binary
.code_size
;
5847 size
+= shader
->prolog
->binary
.code_size
;
5849 size
+= shader
->epilog
->binary
.code_size
;
5853 int si_shader_binary_upload(struct si_screen
*sscreen
, struct si_shader
*shader
)
5855 const struct ac_shader_binary
*prolog
=
5856 shader
->prolog
? &shader
->prolog
->binary
: NULL
;
5857 const struct ac_shader_binary
*epilog
=
5858 shader
->epilog
? &shader
->epilog
->binary
: NULL
;
5859 const struct ac_shader_binary
*mainb
= &shader
->binary
;
5860 unsigned bo_size
= si_get_shader_binary_size(shader
) +
5861 (!epilog
? mainb
->rodata_size
: 0);
5864 assert(!prolog
|| !prolog
->rodata_size
);
5865 assert((!prolog
&& !epilog
) || !mainb
->rodata_size
);
5866 assert(!epilog
|| !epilog
->rodata_size
);
5868 r600_resource_reference(&shader
->bo
, NULL
);
5869 shader
->bo
= (struct r600_resource
*)
5870 pipe_buffer_create(&sscreen
->b
.b
, 0,
5871 PIPE_USAGE_IMMUTABLE
,
5872 align(bo_size
, SI_CPDMA_ALIGNMENT
));
5877 ptr
= sscreen
->b
.ws
->buffer_map(shader
->bo
->buf
, NULL
,
5878 PIPE_TRANSFER_READ_WRITE
);
5881 util_memcpy_cpu_to_le32(ptr
, prolog
->code
, prolog
->code_size
);
5882 ptr
+= prolog
->code_size
;
5885 util_memcpy_cpu_to_le32(ptr
, mainb
->code
, mainb
->code_size
);
5886 ptr
+= mainb
->code_size
;
5889 util_memcpy_cpu_to_le32(ptr
, epilog
->code
, epilog
->code_size
);
5890 else if (mainb
->rodata_size
> 0)
5891 util_memcpy_cpu_to_le32(ptr
, mainb
->rodata
, mainb
->rodata_size
);
5893 sscreen
->b
.ws
->buffer_unmap(shader
->bo
->buf
);
5897 static void si_shader_dump_disassembly(const struct ac_shader_binary
*binary
,
5898 struct pipe_debug_callback
*debug
,
5899 const char *name
, FILE *file
)
5904 if (binary
->disasm_string
) {
5905 fprintf(file
, "Shader %s disassembly:\n", name
);
5906 fprintf(file
, "%s", binary
->disasm_string
);
5908 if (debug
&& debug
->debug_message
) {
5909 /* Very long debug messages are cut off, so send the
5910 * disassembly one line at a time. This causes more
5911 * overhead, but on the plus side it simplifies
5912 * parsing of resulting logs.
5914 pipe_debug_message(debug
, SHADER_INFO
,
5915 "Shader Disassembly Begin");
5917 line
= binary
->disasm_string
;
5919 p
= util_strchrnul(line
, '\n');
5923 pipe_debug_message(debug
, SHADER_INFO
,
5924 "%.*s", count
, line
);
5932 pipe_debug_message(debug
, SHADER_INFO
,
5933 "Shader Disassembly End");
5936 fprintf(file
, "Shader %s binary:\n", name
);
5937 for (i
= 0; i
< binary
->code_size
; i
+= 4) {
5938 fprintf(file
, "@0x%x: %02x%02x%02x%02x\n", i
,
5939 binary
->code
[i
+ 3], binary
->code
[i
+ 2],
5940 binary
->code
[i
+ 1], binary
->code
[i
]);
5945 static void si_shader_dump_stats(struct si_screen
*sscreen
,
5946 struct si_shader
*shader
,
5947 struct pipe_debug_callback
*debug
,
5950 bool check_debug_option
)
5952 struct si_shader_config
*conf
= &shader
->config
;
5953 unsigned num_inputs
= shader
->selector
? shader
->selector
->info
.num_inputs
: 0;
5954 unsigned code_size
= si_get_shader_binary_size(shader
);
5955 unsigned lds_increment
= sscreen
->b
.chip_class
>= CIK
? 512 : 256;
5956 unsigned lds_per_wave
= 0;
5957 unsigned max_simd_waves
= 10;
5959 /* Compute LDS usage for PS. */
5960 switch (processor
) {
5961 case PIPE_SHADER_FRAGMENT
:
5962 /* The minimum usage per wave is (num_inputs * 48). The maximum
5963 * usage is (num_inputs * 48 * 16).
5964 * We can get anything in between and it varies between waves.
5966 * The 48 bytes per input for a single primitive is equal to
5967 * 4 bytes/component * 4 components/input * 3 points.
5969 * Other stages don't know the size at compile time or don't
5970 * allocate LDS per wave, but instead they do it per thread group.
5972 lds_per_wave
= conf
->lds_size
* lds_increment
+
5973 align(num_inputs
* 48, lds_increment
);
5975 case PIPE_SHADER_COMPUTE
:
5976 if (shader
->selector
) {
5977 unsigned max_workgroup_size
=
5978 si_get_max_workgroup_size(shader
);
5979 lds_per_wave
= (conf
->lds_size
* lds_increment
) /
5980 DIV_ROUND_UP(max_workgroup_size
, 64);
5985 /* Compute the per-SIMD wave counts. */
5986 if (conf
->num_sgprs
) {
5987 if (sscreen
->b
.chip_class
>= VI
)
5988 max_simd_waves
= MIN2(max_simd_waves
, 800 / conf
->num_sgprs
);
5990 max_simd_waves
= MIN2(max_simd_waves
, 512 / conf
->num_sgprs
);
5993 if (conf
->num_vgprs
)
5994 max_simd_waves
= MIN2(max_simd_waves
, 256 / conf
->num_vgprs
);
5996 /* LDS is 64KB per CU (4 SIMDs), which is 16KB per SIMD (usage above
5997 * 16KB makes some SIMDs unoccupied). */
5999 max_simd_waves
= MIN2(max_simd_waves
, 16384 / lds_per_wave
);
6001 if (!check_debug_option
||
6002 r600_can_dump_shader(&sscreen
->b
, processor
)) {
6003 if (processor
== PIPE_SHADER_FRAGMENT
) {
6004 fprintf(file
, "*** SHADER CONFIG ***\n"
6005 "SPI_PS_INPUT_ADDR = 0x%04x\n"
6006 "SPI_PS_INPUT_ENA = 0x%04x\n",
6007 conf
->spi_ps_input_addr
, conf
->spi_ps_input_ena
);
6010 fprintf(file
, "*** SHADER STATS ***\n"
6013 "Spilled SGPRs: %d\n"
6014 "Spilled VGPRs: %d\n"
6015 "Private memory VGPRs: %d\n"
6016 "Code Size: %d bytes\n"
6018 "Scratch: %d bytes per wave\n"
6020 "********************\n\n\n",
6021 conf
->num_sgprs
, conf
->num_vgprs
,
6022 conf
->spilled_sgprs
, conf
->spilled_vgprs
,
6023 conf
->private_mem_vgprs
, code_size
,
6024 conf
->lds_size
, conf
->scratch_bytes_per_wave
,
6028 pipe_debug_message(debug
, SHADER_INFO
,
6029 "Shader Stats: SGPRS: %d VGPRS: %d Code Size: %d "
6030 "LDS: %d Scratch: %d Max Waves: %d Spilled SGPRs: %d "
6031 "Spilled VGPRs: %d PrivMem VGPRs: %d",
6032 conf
->num_sgprs
, conf
->num_vgprs
, code_size
,
6033 conf
->lds_size
, conf
->scratch_bytes_per_wave
,
6034 max_simd_waves
, conf
->spilled_sgprs
,
6035 conf
->spilled_vgprs
, conf
->private_mem_vgprs
);
6038 const char *si_get_shader_name(struct si_shader
*shader
, unsigned processor
)
6040 switch (processor
) {
6041 case PIPE_SHADER_VERTEX
:
6042 if (shader
->key
.as_es
)
6043 return "Vertex Shader as ES";
6044 else if (shader
->key
.as_ls
)
6045 return "Vertex Shader as LS";
6047 return "Vertex Shader as VS";
6048 case PIPE_SHADER_TESS_CTRL
:
6049 return "Tessellation Control Shader";
6050 case PIPE_SHADER_TESS_EVAL
:
6051 if (shader
->key
.as_es
)
6052 return "Tessellation Evaluation Shader as ES";
6054 return "Tessellation Evaluation Shader as VS";
6055 case PIPE_SHADER_GEOMETRY
:
6056 if (shader
->is_gs_copy_shader
)
6057 return "GS Copy Shader as VS";
6059 return "Geometry Shader";
6060 case PIPE_SHADER_FRAGMENT
:
6061 return "Pixel Shader";
6062 case PIPE_SHADER_COMPUTE
:
6063 return "Compute Shader";
6065 return "Unknown Shader";
6069 void si_shader_dump(struct si_screen
*sscreen
, struct si_shader
*shader
,
6070 struct pipe_debug_callback
*debug
, unsigned processor
,
6071 FILE *file
, bool check_debug_option
)
6073 if (!check_debug_option
||
6074 r600_can_dump_shader(&sscreen
->b
, processor
))
6075 si_dump_shader_key(processor
, &shader
->key
, file
);
6077 if (!check_debug_option
&& shader
->binary
.llvm_ir_string
) {
6078 fprintf(file
, "\n%s - main shader part - LLVM IR:\n\n",
6079 si_get_shader_name(shader
, processor
));
6080 fprintf(file
, "%s\n", shader
->binary
.llvm_ir_string
);
6083 if (!check_debug_option
||
6084 (r600_can_dump_shader(&sscreen
->b
, processor
) &&
6085 !(sscreen
->b
.debug_flags
& DBG_NO_ASM
))) {
6086 fprintf(file
, "\n%s:\n", si_get_shader_name(shader
, processor
));
6089 si_shader_dump_disassembly(&shader
->prolog
->binary
,
6090 debug
, "prolog", file
);
6092 si_shader_dump_disassembly(&shader
->binary
, debug
, "main", file
);
6095 si_shader_dump_disassembly(&shader
->epilog
->binary
,
6096 debug
, "epilog", file
);
6097 fprintf(file
, "\n");
6100 si_shader_dump_stats(sscreen
, shader
, debug
, processor
, file
,
6101 check_debug_option
);
6104 int si_compile_llvm(struct si_screen
*sscreen
,
6105 struct ac_shader_binary
*binary
,
6106 struct si_shader_config
*conf
,
6107 LLVMTargetMachineRef tm
,
6109 struct pipe_debug_callback
*debug
,
6114 unsigned count
= p_atomic_inc_return(&sscreen
->b
.num_compilations
);
6116 if (r600_can_dump_shader(&sscreen
->b
, processor
)) {
6117 fprintf(stderr
, "radeonsi: Compiling shader %d\n", count
);
6119 if (!(sscreen
->b
.debug_flags
& (DBG_NO_IR
| DBG_PREOPT_IR
))) {
6120 fprintf(stderr
, "%s LLVM IR:\n\n", name
);
6121 ac_dump_module(mod
);
6122 fprintf(stderr
, "\n");
6126 if (sscreen
->record_llvm_ir
) {
6127 char *ir
= LLVMPrintModuleToString(mod
);
6128 binary
->llvm_ir_string
= strdup(ir
);
6129 LLVMDisposeMessage(ir
);
6132 if (!si_replace_shader(count
, binary
)) {
6133 r
= si_llvm_compile(mod
, binary
, tm
, debug
);
6138 si_shader_binary_read_config(binary
, conf
, 0);
6140 /* Enable 64-bit and 16-bit denormals, because there is no performance
6143 * If denormals are enabled, all floating-point output modifiers are
6146 * Don't enable denormals for 32-bit floats, because:
6147 * - Floating-point output modifiers would be ignored by the hw.
6148 * - Some opcodes don't support denormals, such as v_mad_f32. We would
6149 * have to stop using those.
6150 * - SI & CI would be very slow.
6152 conf
->float_mode
|= V_00B028_FP_64_DENORMS
;
6154 FREE(binary
->config
);
6155 FREE(binary
->global_symbol_offsets
);
6156 binary
->config
= NULL
;
6157 binary
->global_symbol_offsets
= NULL
;
6159 /* Some shaders can't have rodata because their binaries can be
6162 if (binary
->rodata_size
&&
6163 (processor
== PIPE_SHADER_VERTEX
||
6164 processor
== PIPE_SHADER_TESS_CTRL
||
6165 processor
== PIPE_SHADER_TESS_EVAL
||
6166 processor
== PIPE_SHADER_FRAGMENT
)) {
6167 fprintf(stderr
, "radeonsi: The shader can't have rodata.");
6174 static void si_llvm_build_ret(struct si_shader_context
*ctx
, LLVMValueRef ret
)
6176 if (LLVMGetTypeKind(LLVMTypeOf(ret
)) == LLVMVoidTypeKind
)
6177 LLVMBuildRetVoid(ctx
->gallivm
.builder
);
6179 LLVMBuildRet(ctx
->gallivm
.builder
, ret
);
6182 /* Generate code for the hardware VS shader stage to go with a geometry shader */
6184 si_generate_gs_copy_shader(struct si_screen
*sscreen
,
6185 LLVMTargetMachineRef tm
,
6186 struct si_shader_selector
*gs_selector
,
6187 struct pipe_debug_callback
*debug
)
6189 struct si_shader_context ctx
;
6190 struct si_shader
*shader
;
6191 struct gallivm_state
*gallivm
= &ctx
.gallivm
;
6192 LLVMBuilderRef builder
;
6193 struct lp_build_tgsi_context
*bld_base
= &ctx
.bld_base
;
6194 struct lp_build_context
*uint
= &bld_base
->uint_bld
;
6195 struct si_shader_output_values
*outputs
;
6196 struct tgsi_shader_info
*gsinfo
= &gs_selector
->info
;
6199 outputs
= MALLOC(gsinfo
->num_outputs
* sizeof(outputs
[0]));
6204 shader
= CALLOC_STRUCT(si_shader
);
6211 shader
->selector
= gs_selector
;
6212 shader
->is_gs_copy_shader
= true;
6214 si_init_shader_ctx(&ctx
, sscreen
, shader
, tm
);
6215 ctx
.type
= PIPE_SHADER_VERTEX
;
6217 builder
= gallivm
->builder
;
6219 create_function(&ctx
);
6220 preload_ring_buffers(&ctx
);
6222 LLVMValueRef voffset
=
6223 lp_build_mul_imm(uint
, LLVMGetParam(ctx
.main_fn
,
6224 ctx
.param_vertex_id
), 4);
6226 /* Fetch the vertex stream ID.*/
6227 LLVMValueRef stream_id
;
6229 if (gs_selector
->so
.num_outputs
)
6230 stream_id
= unpack_param(&ctx
, ctx
.param_streamout_config
, 24, 2);
6232 stream_id
= uint
->zero
;
6234 /* Fill in output information. */
6235 for (i
= 0; i
< gsinfo
->num_outputs
; ++i
) {
6236 outputs
[i
].semantic_name
= gsinfo
->output_semantic_name
[i
];
6237 outputs
[i
].semantic_index
= gsinfo
->output_semantic_index
[i
];
6239 for (int chan
= 0; chan
< 4; chan
++) {
6240 outputs
[i
].vertex_stream
[chan
] =
6241 (gsinfo
->output_streams
[i
] >> (2 * chan
)) & 3;
6245 LLVMBasicBlockRef end_bb
;
6246 LLVMValueRef switch_inst
;
6248 end_bb
= LLVMAppendBasicBlockInContext(gallivm
->context
, ctx
.main_fn
, "end");
6249 switch_inst
= LLVMBuildSwitch(builder
, stream_id
, end_bb
, 4);
6251 for (int stream
= 0; stream
< 4; stream
++) {
6252 LLVMBasicBlockRef bb
;
6255 if (!gsinfo
->num_stream_output_components
[stream
])
6258 if (stream
> 0 && !gs_selector
->so
.num_outputs
)
6261 bb
= LLVMInsertBasicBlockInContext(gallivm
->context
, end_bb
, "out");
6262 LLVMAddCase(switch_inst
, lp_build_const_int32(gallivm
, stream
), bb
);
6263 LLVMPositionBuilderAtEnd(builder
, bb
);
6265 /* Fetch vertex data from GSVS ring */
6267 for (i
= 0; i
< gsinfo
->num_outputs
; ++i
) {
6268 for (unsigned chan
= 0; chan
< 4; chan
++) {
6269 if (!(gsinfo
->output_usagemask
[i
] & (1 << chan
)) ||
6270 outputs
[i
].vertex_stream
[chan
] != stream
) {
6271 outputs
[i
].values
[chan
] = ctx
.bld_base
.base
.undef
;
6275 LLVMValueRef soffset
= LLVMConstInt(ctx
.i32
,
6276 offset
* gs_selector
->gs_max_out_vertices
* 16 * 4, 0);
6279 outputs
[i
].values
[chan
] =
6280 ac_build_buffer_load(&ctx
.ac
,
6281 ctx
.gsvs_ring
[0], 1,
6282 uint
->zero
, voffset
,
6283 soffset
, 0, 1, 1, true);
6287 /* Streamout and exports. */
6288 if (gs_selector
->so
.num_outputs
) {
6289 si_llvm_emit_streamout(&ctx
, outputs
,
6290 gsinfo
->num_outputs
,
6295 si_llvm_export_vs(bld_base
, outputs
, gsinfo
->num_outputs
);
6297 LLVMBuildBr(builder
, end_bb
);
6300 LLVMPositionBuilderAtEnd(builder
, end_bb
);
6302 LLVMBuildRetVoid(gallivm
->builder
);
6304 /* Dump LLVM IR before any optimization passes */
6305 if (sscreen
->b
.debug_flags
& DBG_PREOPT_IR
&&
6306 r600_can_dump_shader(&sscreen
->b
, PIPE_SHADER_GEOMETRY
))
6307 ac_dump_module(bld_base
->base
.gallivm
->module
);
6309 si_llvm_finalize_module(&ctx
,
6310 r600_extra_shader_checks(&sscreen
->b
, PIPE_SHADER_GEOMETRY
));
6312 r
= si_compile_llvm(sscreen
, &ctx
.shader
->binary
,
6313 &ctx
.shader
->config
, ctx
.tm
,
6314 bld_base
->base
.gallivm
->module
,
6315 debug
, PIPE_SHADER_GEOMETRY
,
6318 if (r600_can_dump_shader(&sscreen
->b
, PIPE_SHADER_GEOMETRY
))
6319 fprintf(stderr
, "GS Copy Shader:\n");
6320 si_shader_dump(sscreen
, ctx
.shader
, debug
,
6321 PIPE_SHADER_GEOMETRY
, stderr
, true);
6322 r
= si_shader_binary_upload(sscreen
, ctx
.shader
);
6325 si_llvm_dispose(&ctx
);
6336 static void si_dump_shader_key(unsigned shader
, struct si_shader_key
*key
,
6341 fprintf(f
, "SHADER KEY\n");
6344 case PIPE_SHADER_VERTEX
:
6345 fprintf(f
, " part.vs.prolog.instance_divisors = {");
6346 for (i
= 0; i
< ARRAY_SIZE(key
->part
.vs
.prolog
.instance_divisors
); i
++)
6347 fprintf(f
, !i
? "%u" : ", %u",
6348 key
->part
.vs
.prolog
.instance_divisors
[i
]);
6350 fprintf(f
, " part.vs.epilog.export_prim_id = %u\n", key
->part
.vs
.epilog
.export_prim_id
);
6351 fprintf(f
, " as_es = %u\n", key
->as_es
);
6352 fprintf(f
, " as_ls = %u\n", key
->as_ls
);
6354 fprintf(f
, " mono.vs.fix_fetch = {");
6355 for (i
= 0; i
< SI_MAX_ATTRIBS
; i
++)
6356 fprintf(f
, !i
? "%u" : ", %u", key
->mono
.vs
.fix_fetch
[i
]);
6360 case PIPE_SHADER_TESS_CTRL
:
6361 fprintf(f
, " part.tcs.epilog.prim_mode = %u\n", key
->part
.tcs
.epilog
.prim_mode
);
6362 fprintf(f
, " mono.tcs.inputs_to_copy = 0x%"PRIx64
"\n", key
->mono
.tcs
.inputs_to_copy
);
6365 case PIPE_SHADER_TESS_EVAL
:
6366 fprintf(f
, " part.tes.epilog.export_prim_id = %u\n", key
->part
.tes
.epilog
.export_prim_id
);
6367 fprintf(f
, " as_es = %u\n", key
->as_es
);
6370 case PIPE_SHADER_GEOMETRY
:
6371 fprintf(f
, " part.gs.prolog.tri_strip_adj_fix = %u\n", key
->part
.gs
.prolog
.tri_strip_adj_fix
);
6374 case PIPE_SHADER_COMPUTE
:
6377 case PIPE_SHADER_FRAGMENT
:
6378 fprintf(f
, " part.ps.prolog.color_two_side = %u\n", key
->part
.ps
.prolog
.color_two_side
);
6379 fprintf(f
, " part.ps.prolog.flatshade_colors = %u\n", key
->part
.ps
.prolog
.flatshade_colors
);
6380 fprintf(f
, " part.ps.prolog.poly_stipple = %u\n", key
->part
.ps
.prolog
.poly_stipple
);
6381 fprintf(f
, " part.ps.prolog.force_persp_sample_interp = %u\n", key
->part
.ps
.prolog
.force_persp_sample_interp
);
6382 fprintf(f
, " part.ps.prolog.force_linear_sample_interp = %u\n", key
->part
.ps
.prolog
.force_linear_sample_interp
);
6383 fprintf(f
, " part.ps.prolog.force_persp_center_interp = %u\n", key
->part
.ps
.prolog
.force_persp_center_interp
);
6384 fprintf(f
, " part.ps.prolog.force_linear_center_interp = %u\n", key
->part
.ps
.prolog
.force_linear_center_interp
);
6385 fprintf(f
, " part.ps.prolog.bc_optimize_for_persp = %u\n", key
->part
.ps
.prolog
.bc_optimize_for_persp
);
6386 fprintf(f
, " part.ps.prolog.bc_optimize_for_linear = %u\n", key
->part
.ps
.prolog
.bc_optimize_for_linear
);
6387 fprintf(f
, " part.ps.epilog.spi_shader_col_format = 0x%x\n", key
->part
.ps
.epilog
.spi_shader_col_format
);
6388 fprintf(f
, " part.ps.epilog.color_is_int8 = 0x%X\n", key
->part
.ps
.epilog
.color_is_int8
);
6389 fprintf(f
, " part.ps.epilog.color_is_int10 = 0x%X\n", key
->part
.ps
.epilog
.color_is_int10
);
6390 fprintf(f
, " part.ps.epilog.last_cbuf = %u\n", key
->part
.ps
.epilog
.last_cbuf
);
6391 fprintf(f
, " part.ps.epilog.alpha_func = %u\n", key
->part
.ps
.epilog
.alpha_func
);
6392 fprintf(f
, " part.ps.epilog.alpha_to_one = %u\n", key
->part
.ps
.epilog
.alpha_to_one
);
6393 fprintf(f
, " part.ps.epilog.poly_line_smoothing = %u\n", key
->part
.ps
.epilog
.poly_line_smoothing
);
6394 fprintf(f
, " part.ps.epilog.clamp_color = %u\n", key
->part
.ps
.epilog
.clamp_color
);
6401 if ((shader
== PIPE_SHADER_GEOMETRY
||
6402 shader
== PIPE_SHADER_TESS_EVAL
||
6403 shader
== PIPE_SHADER_VERTEX
) &&
6404 !key
->as_es
&& !key
->as_ls
) {
6405 fprintf(f
, " opt.hw_vs.kill_outputs = 0x%"PRIx64
"\n", key
->opt
.hw_vs
.kill_outputs
);
6406 fprintf(f
, " opt.hw_vs.kill_outputs2 = 0x%x\n", key
->opt
.hw_vs
.kill_outputs2
);
6407 fprintf(f
, " opt.hw_vs.clip_disable = %u\n", key
->opt
.hw_vs
.clip_disable
);
6411 static void si_init_shader_ctx(struct si_shader_context
*ctx
,
6412 struct si_screen
*sscreen
,
6413 struct si_shader
*shader
,
6414 LLVMTargetMachineRef tm
)
6416 struct lp_build_tgsi_context
*bld_base
;
6417 struct lp_build_tgsi_action tmpl
= {};
6419 si_llvm_context_init(ctx
, sscreen
, shader
, tm
,
6420 (shader
&& shader
->selector
) ? &shader
->selector
->info
: NULL
,
6421 (shader
&& shader
->selector
) ? shader
->selector
->tokens
: NULL
);
6423 bld_base
= &ctx
->bld_base
;
6424 bld_base
->emit_fetch_funcs
[TGSI_FILE_CONSTANT
] = fetch_constant
;
6426 bld_base
->op_actions
[TGSI_OPCODE_INTERP_CENTROID
] = interp_action
;
6427 bld_base
->op_actions
[TGSI_OPCODE_INTERP_SAMPLE
] = interp_action
;
6428 bld_base
->op_actions
[TGSI_OPCODE_INTERP_OFFSET
] = interp_action
;
6430 bld_base
->op_actions
[TGSI_OPCODE_TEX
] = tex_action
;
6431 bld_base
->op_actions
[TGSI_OPCODE_TEX2
] = tex_action
;
6432 bld_base
->op_actions
[TGSI_OPCODE_TXB
] = tex_action
;
6433 bld_base
->op_actions
[TGSI_OPCODE_TXB2
] = tex_action
;
6434 bld_base
->op_actions
[TGSI_OPCODE_TXD
] = tex_action
;
6435 bld_base
->op_actions
[TGSI_OPCODE_TXF
] = tex_action
;
6436 bld_base
->op_actions
[TGSI_OPCODE_TXL
] = tex_action
;
6437 bld_base
->op_actions
[TGSI_OPCODE_TXL2
] = tex_action
;
6438 bld_base
->op_actions
[TGSI_OPCODE_TXP
] = tex_action
;
6439 bld_base
->op_actions
[TGSI_OPCODE_TXQ
].fetch_args
= txq_fetch_args
;
6440 bld_base
->op_actions
[TGSI_OPCODE_TXQ
].emit
= txq_emit
;
6441 bld_base
->op_actions
[TGSI_OPCODE_TG4
] = tex_action
;
6442 bld_base
->op_actions
[TGSI_OPCODE_LODQ
] = tex_action
;
6443 bld_base
->op_actions
[TGSI_OPCODE_TXQS
].emit
= si_llvm_emit_txqs
;
6445 bld_base
->op_actions
[TGSI_OPCODE_LOAD
].fetch_args
= load_fetch_args
;
6446 bld_base
->op_actions
[TGSI_OPCODE_LOAD
].emit
= load_emit
;
6447 bld_base
->op_actions
[TGSI_OPCODE_STORE
].fetch_args
= store_fetch_args
;
6448 bld_base
->op_actions
[TGSI_OPCODE_STORE
].emit
= store_emit
;
6449 bld_base
->op_actions
[TGSI_OPCODE_RESQ
].fetch_args
= resq_fetch_args
;
6450 bld_base
->op_actions
[TGSI_OPCODE_RESQ
].emit
= resq_emit
;
6452 tmpl
.fetch_args
= atomic_fetch_args
;
6453 tmpl
.emit
= atomic_emit
;
6454 bld_base
->op_actions
[TGSI_OPCODE_ATOMUADD
] = tmpl
;
6455 bld_base
->op_actions
[TGSI_OPCODE_ATOMUADD
].intr_name
= "add";
6456 bld_base
->op_actions
[TGSI_OPCODE_ATOMXCHG
] = tmpl
;
6457 bld_base
->op_actions
[TGSI_OPCODE_ATOMXCHG
].intr_name
= "swap";
6458 bld_base
->op_actions
[TGSI_OPCODE_ATOMCAS
] = tmpl
;
6459 bld_base
->op_actions
[TGSI_OPCODE_ATOMCAS
].intr_name
= "cmpswap";
6460 bld_base
->op_actions
[TGSI_OPCODE_ATOMAND
] = tmpl
;
6461 bld_base
->op_actions
[TGSI_OPCODE_ATOMAND
].intr_name
= "and";
6462 bld_base
->op_actions
[TGSI_OPCODE_ATOMOR
] = tmpl
;
6463 bld_base
->op_actions
[TGSI_OPCODE_ATOMOR
].intr_name
= "or";
6464 bld_base
->op_actions
[TGSI_OPCODE_ATOMXOR
] = tmpl
;
6465 bld_base
->op_actions
[TGSI_OPCODE_ATOMXOR
].intr_name
= "xor";
6466 bld_base
->op_actions
[TGSI_OPCODE_ATOMUMIN
] = tmpl
;
6467 bld_base
->op_actions
[TGSI_OPCODE_ATOMUMIN
].intr_name
= "umin";
6468 bld_base
->op_actions
[TGSI_OPCODE_ATOMUMAX
] = tmpl
;
6469 bld_base
->op_actions
[TGSI_OPCODE_ATOMUMAX
].intr_name
= "umax";
6470 bld_base
->op_actions
[TGSI_OPCODE_ATOMIMIN
] = tmpl
;
6471 bld_base
->op_actions
[TGSI_OPCODE_ATOMIMIN
].intr_name
= "smin";
6472 bld_base
->op_actions
[TGSI_OPCODE_ATOMIMAX
] = tmpl
;
6473 bld_base
->op_actions
[TGSI_OPCODE_ATOMIMAX
].intr_name
= "smax";
6475 bld_base
->op_actions
[TGSI_OPCODE_MEMBAR
].emit
= membar_emit
;
6477 bld_base
->op_actions
[TGSI_OPCODE_DDX
].emit
= si_llvm_emit_ddxy
;
6478 bld_base
->op_actions
[TGSI_OPCODE_DDY
].emit
= si_llvm_emit_ddxy
;
6479 bld_base
->op_actions
[TGSI_OPCODE_DDX_FINE
].emit
= si_llvm_emit_ddxy
;
6480 bld_base
->op_actions
[TGSI_OPCODE_DDY_FINE
].emit
= si_llvm_emit_ddxy
;
6482 bld_base
->op_actions
[TGSI_OPCODE_EMIT
].emit
= si_llvm_emit_vertex
;
6483 bld_base
->op_actions
[TGSI_OPCODE_ENDPRIM
].emit
= si_llvm_emit_primitive
;
6484 bld_base
->op_actions
[TGSI_OPCODE_BARRIER
].emit
= si_llvm_emit_barrier
;
6487 #define EXP_TARGET (HAVE_LLVM >= 0x0500 ? 0 : 3)
6488 #define EXP_OUT0 (HAVE_LLVM >= 0x0500 ? 2 : 5)
6490 /* Return true if the PARAM export has been eliminated. */
6491 static bool si_eliminate_const_output(struct si_shader_context
*ctx
,
6492 LLVMValueRef inst
, unsigned offset
)
6494 struct si_shader
*shader
= ctx
->shader
;
6495 unsigned num_outputs
= shader
->selector
->info
.num_outputs
;
6496 unsigned i
, default_val
; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
6497 bool is_zero
[4] = {}, is_one
[4] = {};
6499 for (i
= 0; i
< 4; i
++) {
6500 LLVMBool loses_info
;
6501 LLVMValueRef p
= LLVMGetOperand(inst
, EXP_OUT0
+ i
);
6503 /* It's a constant expression. Undef outputs are eliminated too. */
6504 if (LLVMIsUndef(p
)) {
6507 } else if (LLVMIsAConstantFP(p
)) {
6508 double a
= LLVMConstRealGetDouble(p
, &loses_info
);
6515 return false; /* other constant */
6520 /* Only certain combinations of 0 and 1 can be eliminated. */
6521 if (is_zero
[0] && is_zero
[1] && is_zero
[2])
6522 default_val
= is_zero
[3] ? 0 : 1;
6523 else if (is_one
[0] && is_one
[1] && is_one
[2])
6524 default_val
= is_zero
[3] ? 2 : 3;
6528 /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
6529 LLVMInstructionEraseFromParent(inst
);
6531 /* Change OFFSET to DEFAULT_VAL. */
6532 for (i
= 0; i
< num_outputs
; i
++) {
6533 if (shader
->info
.vs_output_param_offset
[i
] == offset
) {
6534 shader
->info
.vs_output_param_offset
[i
] =
6535 EXP_PARAM_DEFAULT_VAL_0000
+ default_val
;
6542 struct si_vs_exports
{
6544 unsigned offset
[SI_MAX_VS_OUTPUTS
];
6545 LLVMValueRef inst
[SI_MAX_VS_OUTPUTS
];
6548 static void si_eliminate_const_vs_outputs(struct si_shader_context
*ctx
)
6550 struct si_shader
*shader
= ctx
->shader
;
6551 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
6552 LLVMBasicBlockRef bb
;
6553 struct si_vs_exports exports
;
6554 bool removed_any
= false;
6558 if (ctx
->type
== PIPE_SHADER_FRAGMENT
||
6559 ctx
->type
== PIPE_SHADER_COMPUTE
||
6560 shader
->key
.as_es
||
6564 /* Process all LLVM instructions. */
6565 bb
= LLVMGetFirstBasicBlock(ctx
->main_fn
);
6567 LLVMValueRef inst
= LLVMGetFirstInstruction(bb
);
6570 LLVMValueRef cur
= inst
;
6571 inst
= LLVMGetNextInstruction(inst
);
6573 if (LLVMGetInstructionOpcode(cur
) != LLVMCall
)
6576 LLVMValueRef callee
= lp_get_called_value(cur
);
6578 if (!lp_is_function(callee
))
6581 const char *name
= LLVMGetValueName(callee
);
6582 unsigned num_args
= LLVMCountParams(callee
);
6584 /* Check if this is an export instruction. */
6585 if ((num_args
!= 9 && num_args
!= 8) ||
6586 (strcmp(name
, "llvm.SI.export") &&
6587 strcmp(name
, "llvm.amdgcn.exp.f32")))
6590 LLVMValueRef arg
= LLVMGetOperand(cur
, EXP_TARGET
);
6591 unsigned target
= LLVMConstIntGetZExtValue(arg
);
6593 if (target
< V_008DFC_SQ_EXP_PARAM
)
6596 target
-= V_008DFC_SQ_EXP_PARAM
;
6598 /* Eliminate constant value PARAM exports. */
6599 if (si_eliminate_const_output(ctx
, cur
, target
)) {
6602 exports
.offset
[exports
.num
] = target
;
6603 exports
.inst
[exports
.num
] = cur
;
6607 bb
= LLVMGetNextBasicBlock(bb
);
6610 /* Remove holes in export memory due to removed PARAM exports.
6611 * This is done by renumbering all PARAM exports.
6614 ubyte current_offset
[SI_MAX_VS_OUTPUTS
];
6615 unsigned new_count
= 0;
6618 /* Make a copy of the offsets. We need the old version while
6619 * we are modifying some of them. */
6620 assert(sizeof(current_offset
) ==
6621 sizeof(shader
->info
.vs_output_param_offset
));
6622 memcpy(current_offset
, shader
->info
.vs_output_param_offset
,
6623 sizeof(current_offset
));
6625 for (i
= 0; i
< exports
.num
; i
++) {
6626 unsigned offset
= exports
.offset
[i
];
6628 for (out
= 0; out
< info
->num_outputs
; out
++) {
6629 if (current_offset
[out
] != offset
)
6632 LLVMSetOperand(exports
.inst
[i
], EXP_TARGET
,
6633 LLVMConstInt(ctx
->i32
,
6634 V_008DFC_SQ_EXP_PARAM
+ new_count
, 0));
6635 shader
->info
.vs_output_param_offset
[out
] = new_count
;
6640 shader
->info
.nr_param_exports
= new_count
;
6644 static void si_count_scratch_private_memory(struct si_shader_context
*ctx
)
6646 ctx
->shader
->config
.private_mem_vgprs
= 0;
6648 /* Process all LLVM instructions. */
6649 LLVMBasicBlockRef bb
= LLVMGetFirstBasicBlock(ctx
->main_fn
);
6651 LLVMValueRef next
= LLVMGetFirstInstruction(bb
);
6654 LLVMValueRef inst
= next
;
6655 next
= LLVMGetNextInstruction(next
);
6657 if (LLVMGetInstructionOpcode(inst
) != LLVMAlloca
)
6660 LLVMTypeRef type
= LLVMGetElementType(LLVMTypeOf(inst
));
6661 /* No idea why LLVM aligns allocas to 4 elements. */
6662 unsigned alignment
= LLVMGetAlignment(inst
);
6663 unsigned dw_size
= align(llvm_get_type_size(type
) / 4, alignment
);
6664 ctx
->shader
->config
.private_mem_vgprs
+= dw_size
;
6666 bb
= LLVMGetNextBasicBlock(bb
);
6670 static bool si_compile_tgsi_main(struct si_shader_context
*ctx
,
6671 struct si_shader
*shader
)
6673 struct si_shader_selector
*sel
= shader
->selector
;
6674 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
6676 switch (ctx
->type
) {
6677 case PIPE_SHADER_VERTEX
:
6678 ctx
->load_input
= declare_input_vs
;
6679 if (shader
->key
.as_ls
)
6680 bld_base
->emit_epilogue
= si_llvm_emit_ls_epilogue
;
6681 else if (shader
->key
.as_es
)
6682 bld_base
->emit_epilogue
= si_llvm_emit_es_epilogue
;
6684 bld_base
->emit_epilogue
= si_llvm_emit_vs_epilogue
;
6686 case PIPE_SHADER_TESS_CTRL
:
6687 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = fetch_input_tcs
;
6688 bld_base
->emit_fetch_funcs
[TGSI_FILE_OUTPUT
] = fetch_output_tcs
;
6689 bld_base
->emit_store
= store_output_tcs
;
6690 bld_base
->emit_epilogue
= si_llvm_emit_tcs_epilogue
;
6692 case PIPE_SHADER_TESS_EVAL
:
6693 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = fetch_input_tes
;
6694 if (shader
->key
.as_es
)
6695 bld_base
->emit_epilogue
= si_llvm_emit_es_epilogue
;
6697 bld_base
->emit_epilogue
= si_llvm_emit_vs_epilogue
;
6699 case PIPE_SHADER_GEOMETRY
:
6700 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = fetch_input_gs
;
6701 bld_base
->emit_epilogue
= si_llvm_emit_gs_epilogue
;
6703 case PIPE_SHADER_FRAGMENT
:
6704 ctx
->load_input
= declare_input_fs
;
6705 bld_base
->emit_epilogue
= si_llvm_return_fs_outputs
;
6707 case PIPE_SHADER_COMPUTE
:
6708 ctx
->declare_memory_region
= declare_compute_memory
;
6711 assert(!"Unsupported shader type");
6715 create_function(ctx
);
6716 preload_ring_buffers(ctx
);
6718 if (ctx
->type
== PIPE_SHADER_GEOMETRY
) {
6720 for (i
= 0; i
< 4; i
++) {
6721 ctx
->gs_next_vertex
[i
] =
6722 lp_build_alloca(bld_base
->base
.gallivm
,
6727 if (!lp_build_tgsi_llvm(bld_base
, sel
->tokens
)) {
6728 fprintf(stderr
, "Failed to translate shader from TGSI to LLVM\n");
6732 si_llvm_build_ret(ctx
, ctx
->return_value
);
6737 * Compute the VS prolog key, which contains all the information needed to
6738 * build the VS prolog function, and set shader->info bits where needed.
6740 static void si_get_vs_prolog_key(struct si_shader
*shader
,
6741 union si_shader_part_key
*key
)
6743 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
6745 memset(key
, 0, sizeof(*key
));
6746 key
->vs_prolog
.states
= shader
->key
.part
.vs
.prolog
;
6747 key
->vs_prolog
.num_input_sgprs
= shader
->info
.num_input_sgprs
;
6748 key
->vs_prolog
.last_input
= MAX2(1, info
->num_inputs
) - 1;
6750 /* Set the instanceID flag. */
6751 for (unsigned i
= 0; i
< info
->num_inputs
; i
++)
6752 if (key
->vs_prolog
.states
.instance_divisors
[i
])
6753 shader
->info
.uses_instanceid
= true;
6757 * Compute the VS epilog key, which contains all the information needed to
6758 * build the VS epilog function, and set the PrimitiveID output offset.
6760 static void si_get_vs_epilog_key(struct si_shader
*shader
,
6761 struct si_vs_epilog_bits
*states
,
6762 union si_shader_part_key
*key
)
6764 memset(key
, 0, sizeof(*key
));
6765 key
->vs_epilog
.states
= *states
;
6767 /* Set up the PrimitiveID output. */
6768 if (shader
->key
.part
.vs
.epilog
.export_prim_id
) {
6769 unsigned index
= shader
->selector
->info
.num_outputs
;
6770 unsigned offset
= shader
->info
.nr_param_exports
++;
6772 key
->vs_epilog
.prim_id_param_offset
= offset
;
6773 assert(index
< ARRAY_SIZE(shader
->info
.vs_output_param_offset
));
6774 shader
->info
.vs_output_param_offset
[index
] = offset
;
6779 * Compute the PS prolog key, which contains all the information needed to
6780 * build the PS prolog function, and set related bits in shader->config.
6782 static void si_get_ps_prolog_key(struct si_shader
*shader
,
6783 union si_shader_part_key
*key
,
6784 bool separate_prolog
)
6786 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
6788 memset(key
, 0, sizeof(*key
));
6789 key
->ps_prolog
.states
= shader
->key
.part
.ps
.prolog
;
6790 key
->ps_prolog
.colors_read
= info
->colors_read
;
6791 key
->ps_prolog
.num_input_sgprs
= shader
->info
.num_input_sgprs
;
6792 key
->ps_prolog
.num_input_vgprs
= shader
->info
.num_input_vgprs
;
6793 key
->ps_prolog
.wqm
= info
->uses_derivatives
&&
6794 (key
->ps_prolog
.colors_read
||
6795 key
->ps_prolog
.states
.force_persp_sample_interp
||
6796 key
->ps_prolog
.states
.force_linear_sample_interp
||
6797 key
->ps_prolog
.states
.force_persp_center_interp
||
6798 key
->ps_prolog
.states
.force_linear_center_interp
||
6799 key
->ps_prolog
.states
.bc_optimize_for_persp
||
6800 key
->ps_prolog
.states
.bc_optimize_for_linear
);
6802 if (info
->colors_read
) {
6803 unsigned *color
= shader
->selector
->color_attr_index
;
6805 if (shader
->key
.part
.ps
.prolog
.color_two_side
) {
6806 /* BCOLORs are stored after the last input. */
6807 key
->ps_prolog
.num_interp_inputs
= info
->num_inputs
;
6808 key
->ps_prolog
.face_vgpr_index
= shader
->info
.face_vgpr_index
;
6809 shader
->config
.spi_ps_input_ena
|= S_0286CC_FRONT_FACE_ENA(1);
6812 for (unsigned i
= 0; i
< 2; i
++) {
6813 unsigned interp
= info
->input_interpolate
[color
[i
]];
6814 unsigned location
= info
->input_interpolate_loc
[color
[i
]];
6816 if (!(info
->colors_read
& (0xf << i
*4)))
6819 key
->ps_prolog
.color_attr_index
[i
] = color
[i
];
6821 if (shader
->key
.part
.ps
.prolog
.flatshade_colors
&&
6822 interp
== TGSI_INTERPOLATE_COLOR
)
6823 interp
= TGSI_INTERPOLATE_CONSTANT
;
6826 case TGSI_INTERPOLATE_CONSTANT
:
6827 key
->ps_prolog
.color_interp_vgpr_index
[i
] = -1;
6829 case TGSI_INTERPOLATE_PERSPECTIVE
:
6830 case TGSI_INTERPOLATE_COLOR
:
6831 /* Force the interpolation location for colors here. */
6832 if (shader
->key
.part
.ps
.prolog
.force_persp_sample_interp
)
6833 location
= TGSI_INTERPOLATE_LOC_SAMPLE
;
6834 if (shader
->key
.part
.ps
.prolog
.force_persp_center_interp
)
6835 location
= TGSI_INTERPOLATE_LOC_CENTER
;
6838 case TGSI_INTERPOLATE_LOC_SAMPLE
:
6839 key
->ps_prolog
.color_interp_vgpr_index
[i
] = 0;
6840 shader
->config
.spi_ps_input_ena
|=
6841 S_0286CC_PERSP_SAMPLE_ENA(1);
6843 case TGSI_INTERPOLATE_LOC_CENTER
:
6844 key
->ps_prolog
.color_interp_vgpr_index
[i
] = 2;
6845 shader
->config
.spi_ps_input_ena
|=
6846 S_0286CC_PERSP_CENTER_ENA(1);
6848 case TGSI_INTERPOLATE_LOC_CENTROID
:
6849 key
->ps_prolog
.color_interp_vgpr_index
[i
] = 4;
6850 shader
->config
.spi_ps_input_ena
|=
6851 S_0286CC_PERSP_CENTROID_ENA(1);
6857 case TGSI_INTERPOLATE_LINEAR
:
6858 /* Force the interpolation location for colors here. */
6859 if (shader
->key
.part
.ps
.prolog
.force_linear_sample_interp
)
6860 location
= TGSI_INTERPOLATE_LOC_SAMPLE
;
6861 if (shader
->key
.part
.ps
.prolog
.force_linear_center_interp
)
6862 location
= TGSI_INTERPOLATE_LOC_CENTER
;
6864 /* The VGPR assignment for non-monolithic shaders
6865 * works because InitialPSInputAddr is set on the
6866 * main shader and PERSP_PULL_MODEL is never used.
6869 case TGSI_INTERPOLATE_LOC_SAMPLE
:
6870 key
->ps_prolog
.color_interp_vgpr_index
[i
] =
6871 separate_prolog
? 6 : 9;
6872 shader
->config
.spi_ps_input_ena
|=
6873 S_0286CC_LINEAR_SAMPLE_ENA(1);
6875 case TGSI_INTERPOLATE_LOC_CENTER
:
6876 key
->ps_prolog
.color_interp_vgpr_index
[i
] =
6877 separate_prolog
? 8 : 11;
6878 shader
->config
.spi_ps_input_ena
|=
6879 S_0286CC_LINEAR_CENTER_ENA(1);
6881 case TGSI_INTERPOLATE_LOC_CENTROID
:
6882 key
->ps_prolog
.color_interp_vgpr_index
[i
] =
6883 separate_prolog
? 10 : 13;
6884 shader
->config
.spi_ps_input_ena
|=
6885 S_0286CC_LINEAR_CENTROID_ENA(1);
6899 * Check whether a PS prolog is required based on the key.
6901 static bool si_need_ps_prolog(const union si_shader_part_key
*key
)
6903 return key
->ps_prolog
.colors_read
||
6904 key
->ps_prolog
.states
.force_persp_sample_interp
||
6905 key
->ps_prolog
.states
.force_linear_sample_interp
||
6906 key
->ps_prolog
.states
.force_persp_center_interp
||
6907 key
->ps_prolog
.states
.force_linear_center_interp
||
6908 key
->ps_prolog
.states
.bc_optimize_for_persp
||
6909 key
->ps_prolog
.states
.bc_optimize_for_linear
||
6910 key
->ps_prolog
.states
.poly_stipple
;
6914 * Compute the PS epilog key, which contains all the information needed to
6915 * build the PS epilog function.
6917 static void si_get_ps_epilog_key(struct si_shader
*shader
,
6918 union si_shader_part_key
*key
)
6920 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
6921 memset(key
, 0, sizeof(*key
));
6922 key
->ps_epilog
.colors_written
= info
->colors_written
;
6923 key
->ps_epilog
.writes_z
= info
->writes_z
;
6924 key
->ps_epilog
.writes_stencil
= info
->writes_stencil
;
6925 key
->ps_epilog
.writes_samplemask
= info
->writes_samplemask
;
6926 key
->ps_epilog
.states
= shader
->key
.part
.ps
.epilog
;
6930 * Build the GS prolog function. Rotate the input vertices for triangle strips
6933 static void si_build_gs_prolog_function(struct si_shader_context
*ctx
,
6934 union si_shader_part_key
*key
)
6936 const unsigned num_sgprs
= SI_GS_NUM_USER_SGPR
+ 2;
6937 const unsigned num_vgprs
= 8;
6938 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
6939 LLVMBuilderRef builder
= gallivm
->builder
;
6940 LLVMTypeRef params
[32];
6941 LLVMTypeRef returns
[32];
6942 LLVMValueRef func
, ret
;
6944 for (unsigned i
= 0; i
< num_sgprs
; ++i
) {
6945 params
[i
] = ctx
->i32
;
6946 returns
[i
] = ctx
->i32
;
6949 for (unsigned i
= 0; i
< num_vgprs
; ++i
) {
6950 params
[num_sgprs
+ i
] = ctx
->i32
;
6951 returns
[num_sgprs
+ i
] = ctx
->f32
;
6954 /* Create the function. */
6955 si_create_function(ctx
, "gs_prolog", returns
, num_sgprs
+ num_vgprs
,
6956 params
, num_sgprs
+ num_vgprs
, num_sgprs
- 1);
6957 func
= ctx
->main_fn
;
6959 /* Copy inputs to outputs. This should be no-op, as the registers match,
6960 * but it will prevent the compiler from overwriting them unintentionally.
6962 ret
= ctx
->return_value
;
6963 for (unsigned i
= 0; i
< num_sgprs
; i
++) {
6964 LLVMValueRef p
= LLVMGetParam(func
, i
);
6965 ret
= LLVMBuildInsertValue(builder
, ret
, p
, i
, "");
6967 for (unsigned i
= 0; i
< num_vgprs
; i
++) {
6968 LLVMValueRef p
= LLVMGetParam(func
, num_sgprs
+ i
);
6969 p
= LLVMBuildBitCast(builder
, p
, ctx
->f32
, "");
6970 ret
= LLVMBuildInsertValue(builder
, ret
, p
, num_sgprs
+ i
, "");
6973 if (key
->gs_prolog
.states
.tri_strip_adj_fix
) {
6974 /* Remap the input vertices for every other primitive. */
6975 const unsigned vtx_params
[6] = {
6983 LLVMValueRef prim_id
, rotate
;
6985 prim_id
= LLVMGetParam(func
, num_sgprs
+ 2);
6986 rotate
= LLVMBuildTrunc(builder
, prim_id
, ctx
->i1
, "");
6988 for (unsigned i
= 0; i
< 6; ++i
) {
6989 LLVMValueRef base
, rotated
, actual
;
6990 base
= LLVMGetParam(func
, vtx_params
[i
]);
6991 rotated
= LLVMGetParam(func
, vtx_params
[(i
+ 4) % 6]);
6992 actual
= LLVMBuildSelect(builder
, rotate
, rotated
, base
, "");
6993 actual
= LLVMBuildBitCast(builder
, actual
, ctx
->f32
, "");
6994 ret
= LLVMBuildInsertValue(builder
, ret
, actual
, vtx_params
[i
], "");
6998 LLVMBuildRet(builder
, ret
);
7002 * Given a list of shader part functions, build a wrapper function that
7003 * runs them in sequence to form a monolithic shader.
7005 static void si_build_wrapper_function(struct si_shader_context
*ctx
,
7006 LLVMValueRef
*parts
,
7010 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
7011 LLVMBuilderRef builder
= ctx
->gallivm
.builder
;
7012 /* PS epilog has one arg per color component */
7013 LLVMTypeRef param_types
[48];
7014 LLVMValueRef out
[48];
7015 LLVMTypeRef function_type
;
7016 unsigned num_params
;
7018 MAYBE_UNUSED
unsigned num_out_sgpr
; /* used in debug checks */
7019 unsigned num_sgprs
, num_vgprs
;
7020 unsigned last_sgpr_param
;
7023 for (unsigned i
= 0; i
< num_parts
; ++i
) {
7024 lp_add_function_attr(parts
[i
], -1, LP_FUNC_ATTR_ALWAYSINLINE
);
7025 LLVMSetLinkage(parts
[i
], LLVMPrivateLinkage
);
7028 /* The parameters of the wrapper function correspond to those of the
7029 * first part in terms of SGPRs and VGPRs, but we use the types of the
7030 * main part to get the right types. This is relevant for the
7031 * dereferenceable attribute on descriptor table pointers.
7036 function_type
= LLVMGetElementType(LLVMTypeOf(parts
[0]));
7037 num_params
= LLVMCountParamTypes(function_type
);
7039 for (unsigned i
= 0; i
< num_params
; ++i
) {
7040 LLVMValueRef param
= LLVMGetParam(parts
[0], i
);
7042 if (ac_is_sgpr_param(param
)) {
7043 assert(num_vgprs
== 0);
7044 num_sgprs
+= llvm_get_type_size(LLVMTypeOf(param
)) / 4;
7046 num_vgprs
+= llvm_get_type_size(LLVMTypeOf(param
)) / 4;
7049 assert(num_vgprs
+ num_sgprs
<= ARRAY_SIZE(param_types
));
7052 last_sgpr_param
= 0;
7054 while (gprs
< num_sgprs
+ num_vgprs
) {
7055 LLVMValueRef param
= LLVMGetParam(parts
[main_part
], num_params
);
7058 param_types
[num_params
] = LLVMTypeOf(param
);
7059 if (gprs
< num_sgprs
)
7060 last_sgpr_param
= num_params
;
7061 size
= llvm_get_type_size(param_types
[num_params
]) / 4;
7064 assert(ac_is_sgpr_param(param
) == (gprs
< num_sgprs
));
7065 assert(gprs
+ size
<= num_sgprs
+ num_vgprs
&&
7066 (gprs
>= num_sgprs
|| gprs
+ size
<= num_sgprs
));
7071 si_create_function(ctx
, "wrapper", NULL
, 0, param_types
, num_params
, last_sgpr_param
);
7073 /* Record the arguments of the function as if they were an output of
7079 for (unsigned i
= 0; i
< num_params
; ++i
) {
7080 LLVMValueRef param
= LLVMGetParam(ctx
->main_fn
, i
);
7081 LLVMTypeRef param_type
= LLVMTypeOf(param
);
7082 LLVMTypeRef out_type
= i
<= last_sgpr_param
? ctx
->i32
: ctx
->f32
;
7083 unsigned size
= llvm_get_type_size(param_type
) / 4;
7086 if (param_type
!= out_type
)
7087 param
= LLVMBuildBitCast(builder
, param
, out_type
, "");
7088 out
[num_out
++] = param
;
7090 LLVMTypeRef vector_type
= LLVMVectorType(out_type
, size
);
7092 if (LLVMGetTypeKind(param_type
) == LLVMPointerTypeKind
) {
7093 param
= LLVMBuildPtrToInt(builder
, param
, ctx
->i64
, "");
7094 param_type
= ctx
->i64
;
7097 if (param_type
!= vector_type
)
7098 param
= LLVMBuildBitCast(builder
, param
, vector_type
, "");
7100 for (unsigned j
= 0; j
< size
; ++j
)
7101 out
[num_out
++] = LLVMBuildExtractElement(
7102 builder
, param
, LLVMConstInt(ctx
->i32
, j
, 0), "");
7105 if (i
<= last_sgpr_param
)
7106 num_out_sgpr
= num_out
;
7109 /* Now chain the parts. */
7110 for (unsigned part
= 0; part
< num_parts
; ++part
) {
7111 LLVMValueRef in
[48];
7113 LLVMTypeRef ret_type
;
7114 unsigned out_idx
= 0;
7116 num_params
= LLVMCountParams(parts
[part
]);
7117 assert(num_params
<= ARRAY_SIZE(param_types
));
7119 /* Derive arguments for the next part from outputs of the
7122 for (unsigned param_idx
= 0; param_idx
< num_params
; ++param_idx
) {
7124 LLVMTypeRef param_type
;
7126 unsigned param_size
;
7127 LLVMValueRef arg
= NULL
;
7129 param
= LLVMGetParam(parts
[part
], param_idx
);
7130 param_type
= LLVMTypeOf(param
);
7131 param_size
= llvm_get_type_size(param_type
) / 4;
7132 is_sgpr
= ac_is_sgpr_param(param
);
7135 #if HAVE_LLVM < 0x0400
7136 LLVMRemoveAttribute(param
, LLVMByValAttribute
);
7138 unsigned kind_id
= LLVMGetEnumAttributeKindForName("byval", 5);
7139 LLVMRemoveEnumAttributeAtIndex(parts
[part
], param_idx
+ 1, kind_id
);
7141 lp_add_function_attr(parts
[part
], param_idx
+ 1, LP_FUNC_ATTR_INREG
);
7144 assert(out_idx
+ param_size
<= (is_sgpr
? num_out_sgpr
: num_out
));
7145 assert(is_sgpr
|| out_idx
>= num_out_sgpr
);
7147 if (param_size
== 1)
7150 arg
= lp_build_gather_values(gallivm
, &out
[out_idx
], param_size
);
7152 if (LLVMTypeOf(arg
) != param_type
) {
7153 if (LLVMGetTypeKind(param_type
) == LLVMPointerTypeKind
) {
7154 arg
= LLVMBuildBitCast(builder
, arg
, ctx
->i64
, "");
7155 arg
= LLVMBuildIntToPtr(builder
, arg
, param_type
, "");
7157 arg
= LLVMBuildBitCast(builder
, arg
, param_type
, "");
7161 in
[param_idx
] = arg
;
7162 out_idx
+= param_size
;
7165 ret
= LLVMBuildCall(builder
, parts
[part
], in
, num_params
, "");
7166 ret_type
= LLVMTypeOf(ret
);
7168 /* Extract the returned GPRs. */
7172 if (LLVMGetTypeKind(ret_type
) != LLVMVoidTypeKind
) {
7173 assert(LLVMGetTypeKind(ret_type
) == LLVMStructTypeKind
);
7175 unsigned ret_size
= LLVMCountStructElementTypes(ret_type
);
7177 for (unsigned i
= 0; i
< ret_size
; ++i
) {
7179 LLVMBuildExtractValue(builder
, ret
, i
, "");
7181 out
[num_out
++] = val
;
7183 if (LLVMTypeOf(val
) == ctx
->i32
) {
7184 assert(num_out_sgpr
+ 1 == num_out
);
7185 num_out_sgpr
= num_out
;
7191 LLVMBuildRetVoid(builder
);
7194 int si_compile_tgsi_shader(struct si_screen
*sscreen
,
7195 LLVMTargetMachineRef tm
,
7196 struct si_shader
*shader
,
7198 struct pipe_debug_callback
*debug
)
7200 struct si_shader_selector
*sel
= shader
->selector
;
7201 struct si_shader_context ctx
;
7202 struct lp_build_tgsi_context
*bld_base
;
7206 /* Dump TGSI code before doing TGSI->LLVM conversion in case the
7207 * conversion fails. */
7208 if (r600_can_dump_shader(&sscreen
->b
, sel
->info
.processor
) &&
7209 !(sscreen
->b
.debug_flags
& DBG_NO_TGSI
)) {
7210 tgsi_dump(sel
->tokens
, 0);
7211 si_dump_streamout(&sel
->so
);
7214 si_init_shader_ctx(&ctx
, sscreen
, shader
, tm
);
7215 ctx
.separate_prolog
= !is_monolithic
;
7217 memset(shader
->info
.vs_output_param_offset
, EXP_PARAM_UNDEFINED
,
7218 sizeof(shader
->info
.vs_output_param_offset
));
7220 shader
->info
.uses_instanceid
= sel
->info
.uses_instanceid
;
7222 bld_base
= &ctx
.bld_base
;
7223 ctx
.load_system_value
= declare_system_value
;
7225 if (!si_compile_tgsi_main(&ctx
, shader
)) {
7226 si_llvm_dispose(&ctx
);
7230 if (is_monolithic
&& ctx
.type
== PIPE_SHADER_VERTEX
) {
7231 LLVMValueRef parts
[3];
7235 need_prolog
= sel
->info
.num_inputs
;
7236 need_epilog
= !shader
->key
.as_es
&& !shader
->key
.as_ls
;
7238 parts
[need_prolog
? 1 : 0] = ctx
.main_fn
;
7241 union si_shader_part_key prolog_key
;
7242 si_get_vs_prolog_key(shader
, &prolog_key
);
7243 si_build_vs_prolog_function(&ctx
, &prolog_key
);
7244 parts
[0] = ctx
.main_fn
;
7248 union si_shader_part_key epilog_key
;
7249 si_get_vs_epilog_key(shader
, &shader
->key
.part
.vs
.epilog
, &epilog_key
);
7250 si_build_vs_epilog_function(&ctx
, &epilog_key
);
7251 parts
[need_prolog
? 2 : 1] = ctx
.main_fn
;
7254 si_build_wrapper_function(&ctx
, parts
, 1 + need_prolog
+ need_epilog
,
7255 need_prolog
? 1 : 0);
7256 } else if (is_monolithic
&& ctx
.type
== PIPE_SHADER_TESS_CTRL
) {
7257 LLVMValueRef parts
[2];
7258 union si_shader_part_key epilog_key
;
7260 parts
[0] = ctx
.main_fn
;
7262 memset(&epilog_key
, 0, sizeof(epilog_key
));
7263 epilog_key
.tcs_epilog
.states
= shader
->key
.part
.tcs
.epilog
;
7264 si_build_tcs_epilog_function(&ctx
, &epilog_key
);
7265 parts
[1] = ctx
.main_fn
;
7267 si_build_wrapper_function(&ctx
, parts
, 2, 0);
7268 } else if (is_monolithic
&& ctx
.type
== PIPE_SHADER_TESS_EVAL
&&
7269 !shader
->key
.as_es
) {
7270 LLVMValueRef parts
[2];
7271 union si_shader_part_key epilog_key
;
7273 parts
[0] = ctx
.main_fn
;
7275 si_get_vs_epilog_key(shader
, &shader
->key
.part
.tes
.epilog
, &epilog_key
);
7276 si_build_vs_epilog_function(&ctx
, &epilog_key
);
7277 parts
[1] = ctx
.main_fn
;
7279 si_build_wrapper_function(&ctx
, parts
, 2, 0);
7280 } else if (is_monolithic
&& ctx
.type
== PIPE_SHADER_GEOMETRY
) {
7281 LLVMValueRef parts
[2];
7282 union si_shader_part_key prolog_key
;
7284 parts
[1] = ctx
.main_fn
;
7286 memset(&prolog_key
, 0, sizeof(prolog_key
));
7287 prolog_key
.gs_prolog
.states
= shader
->key
.part
.gs
.prolog
;
7288 si_build_gs_prolog_function(&ctx
, &prolog_key
);
7289 parts
[0] = ctx
.main_fn
;
7291 si_build_wrapper_function(&ctx
, parts
, 2, 1);
7292 } else if (is_monolithic
&& ctx
.type
== PIPE_SHADER_FRAGMENT
) {
7293 LLVMValueRef parts
[3];
7294 union si_shader_part_key prolog_key
;
7295 union si_shader_part_key epilog_key
;
7298 si_get_ps_prolog_key(shader
, &prolog_key
, false);
7299 need_prolog
= si_need_ps_prolog(&prolog_key
);
7301 parts
[need_prolog
? 1 : 0] = ctx
.main_fn
;
7304 si_build_ps_prolog_function(&ctx
, &prolog_key
);
7305 parts
[0] = ctx
.main_fn
;
7308 si_get_ps_epilog_key(shader
, &epilog_key
);
7309 si_build_ps_epilog_function(&ctx
, &epilog_key
);
7310 parts
[need_prolog
? 2 : 1] = ctx
.main_fn
;
7312 si_build_wrapper_function(&ctx
, parts
, need_prolog
? 3 : 2, need_prolog
? 1 : 0);
7315 mod
= bld_base
->base
.gallivm
->module
;
7317 /* Dump LLVM IR before any optimization passes */
7318 if (sscreen
->b
.debug_flags
& DBG_PREOPT_IR
&&
7319 r600_can_dump_shader(&sscreen
->b
, ctx
.type
))
7320 ac_dump_module(mod
);
7322 si_llvm_finalize_module(&ctx
,
7323 r600_extra_shader_checks(&sscreen
->b
, ctx
.type
));
7325 /* Post-optimization transformations and analysis. */
7326 si_eliminate_const_vs_outputs(&ctx
);
7328 if ((debug
&& debug
->debug_message
) ||
7329 r600_can_dump_shader(&sscreen
->b
, ctx
.type
))
7330 si_count_scratch_private_memory(&ctx
);
7332 /* Compile to bytecode. */
7333 r
= si_compile_llvm(sscreen
, &shader
->binary
, &shader
->config
, tm
,
7334 mod
, debug
, ctx
.type
, "TGSI shader");
7335 si_llvm_dispose(&ctx
);
7337 fprintf(stderr
, "LLVM failed to compile shader\n");
7341 /* Validate SGPR and VGPR usage for compute to detect compiler bugs.
7342 * LLVM 3.9svn has this bug.
7344 if (sel
->type
== PIPE_SHADER_COMPUTE
) {
7345 unsigned wave_size
= 64;
7346 unsigned max_vgprs
= 256;
7347 unsigned max_sgprs
= sscreen
->b
.chip_class
>= VI
? 800 : 512;
7348 unsigned max_sgprs_per_wave
= 128;
7349 unsigned max_block_threads
= si_get_max_workgroup_size(shader
);
7350 unsigned min_waves_per_cu
= DIV_ROUND_UP(max_block_threads
, wave_size
);
7351 unsigned min_waves_per_simd
= DIV_ROUND_UP(min_waves_per_cu
, 4);
7353 max_vgprs
= max_vgprs
/ min_waves_per_simd
;
7354 max_sgprs
= MIN2(max_sgprs
/ min_waves_per_simd
, max_sgprs_per_wave
);
7356 if (shader
->config
.num_sgprs
> max_sgprs
||
7357 shader
->config
.num_vgprs
> max_vgprs
) {
7358 fprintf(stderr
, "LLVM failed to compile a shader correctly: "
7359 "SGPR:VGPR usage is %u:%u, but the hw limit is %u:%u\n",
7360 shader
->config
.num_sgprs
, shader
->config
.num_vgprs
,
7361 max_sgprs
, max_vgprs
);
7363 /* Just terminate the process, because dependent
7364 * shaders can hang due to bad input data, but use
7365 * the env var to allow shader-db to work.
7367 if (!debug_get_bool_option("SI_PASS_BAD_SHADERS", false))
7372 /* Add the scratch offset to input SGPRs. */
7373 if (shader
->config
.scratch_bytes_per_wave
)
7374 shader
->info
.num_input_sgprs
+= 1; /* scratch byte offset */
7376 /* Calculate the number of fragment input VGPRs. */
7377 if (ctx
.type
== PIPE_SHADER_FRAGMENT
) {
7378 shader
->info
.num_input_vgprs
= 0;
7379 shader
->info
.face_vgpr_index
= -1;
7381 if (G_0286CC_PERSP_SAMPLE_ENA(shader
->config
.spi_ps_input_addr
))
7382 shader
->info
.num_input_vgprs
+= 2;
7383 if (G_0286CC_PERSP_CENTER_ENA(shader
->config
.spi_ps_input_addr
))
7384 shader
->info
.num_input_vgprs
+= 2;
7385 if (G_0286CC_PERSP_CENTROID_ENA(shader
->config
.spi_ps_input_addr
))
7386 shader
->info
.num_input_vgprs
+= 2;
7387 if (G_0286CC_PERSP_PULL_MODEL_ENA(shader
->config
.spi_ps_input_addr
))
7388 shader
->info
.num_input_vgprs
+= 3;
7389 if (G_0286CC_LINEAR_SAMPLE_ENA(shader
->config
.spi_ps_input_addr
))
7390 shader
->info
.num_input_vgprs
+= 2;
7391 if (G_0286CC_LINEAR_CENTER_ENA(shader
->config
.spi_ps_input_addr
))
7392 shader
->info
.num_input_vgprs
+= 2;
7393 if (G_0286CC_LINEAR_CENTROID_ENA(shader
->config
.spi_ps_input_addr
))
7394 shader
->info
.num_input_vgprs
+= 2;
7395 if (G_0286CC_LINE_STIPPLE_TEX_ENA(shader
->config
.spi_ps_input_addr
))
7396 shader
->info
.num_input_vgprs
+= 1;
7397 if (G_0286CC_POS_X_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
7398 shader
->info
.num_input_vgprs
+= 1;
7399 if (G_0286CC_POS_Y_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
7400 shader
->info
.num_input_vgprs
+= 1;
7401 if (G_0286CC_POS_Z_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
7402 shader
->info
.num_input_vgprs
+= 1;
7403 if (G_0286CC_POS_W_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
7404 shader
->info
.num_input_vgprs
+= 1;
7405 if (G_0286CC_FRONT_FACE_ENA(shader
->config
.spi_ps_input_addr
)) {
7406 shader
->info
.face_vgpr_index
= shader
->info
.num_input_vgprs
;
7407 shader
->info
.num_input_vgprs
+= 1;
7409 if (G_0286CC_ANCILLARY_ENA(shader
->config
.spi_ps_input_addr
))
7410 shader
->info
.num_input_vgprs
+= 1;
7411 if (G_0286CC_SAMPLE_COVERAGE_ENA(shader
->config
.spi_ps_input_addr
))
7412 shader
->info
.num_input_vgprs
+= 1;
7413 if (G_0286CC_POS_FIXED_PT_ENA(shader
->config
.spi_ps_input_addr
))
7414 shader
->info
.num_input_vgprs
+= 1;
7421 * Create, compile and return a shader part (prolog or epilog).
7423 * \param sscreen screen
7424 * \param list list of shader parts of the same category
7425 * \param type shader type
7426 * \param key shader part key
7427 * \param prolog whether the part being requested is a prolog
7428 * \param tm LLVM target machine
7429 * \param debug debug callback
7430 * \param build the callback responsible for building the main function
7431 * \return non-NULL on success
7433 static struct si_shader_part
*
7434 si_get_shader_part(struct si_screen
*sscreen
,
7435 struct si_shader_part
**list
,
7436 enum pipe_shader_type type
,
7438 union si_shader_part_key
*key
,
7439 LLVMTargetMachineRef tm
,
7440 struct pipe_debug_callback
*debug
,
7441 void (*build
)(struct si_shader_context
*,
7442 union si_shader_part_key
*),
7445 struct si_shader_part
*result
;
7447 mtx_lock(&sscreen
->shader_parts_mutex
);
7449 /* Find existing. */
7450 for (result
= *list
; result
; result
= result
->next
) {
7451 if (memcmp(&result
->key
, key
, sizeof(*key
)) == 0) {
7452 mtx_unlock(&sscreen
->shader_parts_mutex
);
7457 /* Compile a new one. */
7458 result
= CALLOC_STRUCT(si_shader_part
);
7461 struct si_shader shader
= {};
7462 struct si_shader_context ctx
;
7463 struct gallivm_state
*gallivm
= &ctx
.gallivm
;
7465 si_init_shader_ctx(&ctx
, sscreen
, &shader
, tm
);
7469 case PIPE_SHADER_VERTEX
:
7471 case PIPE_SHADER_TESS_CTRL
:
7473 shader
.key
.part
.tcs
.epilog
= key
->tcs_epilog
.states
;
7475 case PIPE_SHADER_GEOMETRY
:
7478 case PIPE_SHADER_FRAGMENT
:
7480 shader
.key
.part
.ps
.prolog
= key
->ps_prolog
.states
;
7482 shader
.key
.part
.ps
.epilog
= key
->ps_epilog
.states
;
7485 unreachable("bad shader part");
7491 si_llvm_finalize_module(&ctx
,
7492 r600_extra_shader_checks(&sscreen
->b
, PIPE_SHADER_FRAGMENT
));
7494 if (si_compile_llvm(sscreen
, &result
->binary
, &result
->config
, tm
,
7495 gallivm
->module
, debug
, ctx
.type
, name
)) {
7501 result
->next
= *list
;
7505 si_llvm_dispose(&ctx
);
7506 mtx_unlock(&sscreen
->shader_parts_mutex
);
7511 * Build the vertex shader prolog function.
7513 * The inputs are the same as VS (a lot of SGPRs and 4 VGPR system values).
7514 * All inputs are returned unmodified. The vertex load indices are
7515 * stored after them, which will be used by the API VS for fetching inputs.
7517 * For example, the expected outputs for instance_divisors[] = {0, 1, 2} are:
7522 * (VertexID + BaseVertex),
7523 * (InstanceID + StartInstance),
7524 * (InstanceID / 2 + StartInstance)
7526 static void si_build_vs_prolog_function(struct si_shader_context
*ctx
,
7527 union si_shader_part_key
*key
)
7529 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
7530 LLVMTypeRef
*params
, *returns
;
7531 LLVMValueRef ret
, func
;
7532 int last_sgpr
, num_params
, num_returns
, i
;
7534 ctx
->param_vertex_id
= key
->vs_prolog
.num_input_sgprs
;
7535 ctx
->param_instance_id
= key
->vs_prolog
.num_input_sgprs
+ 3;
7537 /* 4 preloaded VGPRs + vertex load indices as prolog outputs */
7538 params
= alloca((key
->vs_prolog
.num_input_sgprs
+ 4) *
7539 sizeof(LLVMTypeRef
));
7540 returns
= alloca((key
->vs_prolog
.num_input_sgprs
+ 4 +
7541 key
->vs_prolog
.last_input
+ 1) *
7542 sizeof(LLVMTypeRef
));
7546 /* Declare input and output SGPRs. */
7548 for (i
= 0; i
< key
->vs_prolog
.num_input_sgprs
; i
++) {
7549 params
[num_params
++] = ctx
->i32
;
7550 returns
[num_returns
++] = ctx
->i32
;
7552 last_sgpr
= num_params
- 1;
7554 /* 4 preloaded VGPRs (outputs must be floats) */
7555 for (i
= 0; i
< 4; i
++) {
7556 params
[num_params
++] = ctx
->i32
;
7557 returns
[num_returns
++] = ctx
->f32
;
7560 /* Vertex load indices. */
7561 for (i
= 0; i
<= key
->vs_prolog
.last_input
; i
++)
7562 returns
[num_returns
++] = ctx
->f32
;
7564 /* Create the function. */
7565 si_create_function(ctx
, "vs_prolog", returns
, num_returns
, params
,
7566 num_params
, last_sgpr
);
7567 func
= ctx
->main_fn
;
7569 /* Copy inputs to outputs. This should be no-op, as the registers match,
7570 * but it will prevent the compiler from overwriting them unintentionally.
7572 ret
= ctx
->return_value
;
7573 for (i
= 0; i
< key
->vs_prolog
.num_input_sgprs
; i
++) {
7574 LLVMValueRef p
= LLVMGetParam(func
, i
);
7575 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
, p
, i
, "");
7577 for (i
= num_params
- 4; i
< num_params
; i
++) {
7578 LLVMValueRef p
= LLVMGetParam(func
, i
);
7579 p
= LLVMBuildBitCast(gallivm
->builder
, p
, ctx
->f32
, "");
7580 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
, p
, i
, "");
7583 /* Compute vertex load indices from instance divisors. */
7584 for (i
= 0; i
<= key
->vs_prolog
.last_input
; i
++) {
7585 unsigned divisor
= key
->vs_prolog
.states
.instance_divisors
[i
];
7589 /* InstanceID / Divisor + StartInstance */
7590 index
= get_instance_index_for_fetch(ctx
,
7591 SI_SGPR_START_INSTANCE
,
7594 /* VertexID + BaseVertex */
7595 index
= LLVMBuildAdd(gallivm
->builder
,
7596 LLVMGetParam(func
, ctx
->param_vertex_id
),
7597 LLVMGetParam(func
, SI_SGPR_BASE_VERTEX
), "");
7600 index
= LLVMBuildBitCast(gallivm
->builder
, index
, ctx
->f32
, "");
7601 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
, index
,
7605 si_llvm_build_ret(ctx
, ret
);
7609 * Build the vertex shader epilog function. This is also used by the tessellation
7610 * evaluation shader compiled as VS.
7612 * The input is PrimitiveID.
7614 * If PrimitiveID is required by the pixel shader, export it.
7615 * Otherwise, do nothing.
7617 static void si_build_vs_epilog_function(struct si_shader_context
*ctx
,
7618 union si_shader_part_key
*key
)
7620 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
7621 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
7622 LLVMTypeRef params
[5];
7625 /* Declare input VGPRs. */
7626 num_params
= key
->vs_epilog
.states
.export_prim_id
?
7627 (VS_EPILOG_PRIMID_LOC
+ 1) : 0;
7628 assert(num_params
<= ARRAY_SIZE(params
));
7630 for (i
= 0; i
< num_params
; i
++)
7631 params
[i
] = ctx
->f32
;
7633 /* Create the function. */
7634 si_create_function(ctx
, "vs_epilog", NULL
, 0, params
, num_params
, -1);
7637 if (key
->vs_epilog
.states
.export_prim_id
) {
7638 struct lp_build_context
*base
= &bld_base
->base
;
7639 struct ac_export_args args
;
7641 args
.enabled_channels
= 0x1; /* enabled channels */
7642 args
.valid_mask
= 0; /* whether the EXEC mask is valid */
7643 args
.done
= 0; /* DONE bit */
7644 args
.target
= V_008DFC_SQ_EXP_PARAM
+
7645 key
->vs_epilog
.prim_id_param_offset
;
7646 args
.compr
= 0; /* COMPR flag (0 = 32-bit export) */
7647 args
.out
[0] = LLVMGetParam(ctx
->main_fn
,
7648 VS_EPILOG_PRIMID_LOC
); /* X */
7649 args
.out
[1] = base
->undef
; /* Y */
7650 args
.out
[2] = base
->undef
; /* Z */
7651 args
.out
[3] = base
->undef
; /* W */
7653 ac_build_export(&ctx
->ac
, &args
);
7656 LLVMBuildRetVoid(gallivm
->builder
);
7660 * Create & compile a vertex shader epilog. This a helper used by VS and TES.
7662 static bool si_get_vs_epilog(struct si_screen
*sscreen
,
7663 LLVMTargetMachineRef tm
,
7664 struct si_shader
*shader
,
7665 struct pipe_debug_callback
*debug
,
7666 struct si_vs_epilog_bits
*states
)
7668 union si_shader_part_key epilog_key
;
7670 si_get_vs_epilog_key(shader
, states
, &epilog_key
);
7672 shader
->epilog
= si_get_shader_part(sscreen
, &sscreen
->vs_epilogs
,
7673 PIPE_SHADER_VERTEX
, true,
7674 &epilog_key
, tm
, debug
,
7675 si_build_vs_epilog_function
,
7676 "Vertex Shader Epilog");
7677 return shader
->epilog
!= NULL
;
7681 * Select and compile (or reuse) vertex shader parts (prolog & epilog).
7683 static bool si_shader_select_vs_parts(struct si_screen
*sscreen
,
7684 LLVMTargetMachineRef tm
,
7685 struct si_shader
*shader
,
7686 struct pipe_debug_callback
*debug
)
7688 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
7689 union si_shader_part_key prolog_key
;
7691 /* Get the prolog. */
7692 si_get_vs_prolog_key(shader
, &prolog_key
);
7694 /* The prolog is a no-op if there are no inputs. */
7695 if (info
->num_inputs
) {
7697 si_get_shader_part(sscreen
, &sscreen
->vs_prologs
,
7698 PIPE_SHADER_VERTEX
, true,
7699 &prolog_key
, tm
, debug
,
7700 si_build_vs_prolog_function
,
7701 "Vertex Shader Prolog");
7702 if (!shader
->prolog
)
7706 /* Get the epilog. */
7707 if (!shader
->key
.as_es
&& !shader
->key
.as_ls
&&
7708 !si_get_vs_epilog(sscreen
, tm
, shader
, debug
,
7709 &shader
->key
.part
.vs
.epilog
))
7716 * Select and compile (or reuse) TES parts (epilog).
7718 static bool si_shader_select_tes_parts(struct si_screen
*sscreen
,
7719 LLVMTargetMachineRef tm
,
7720 struct si_shader
*shader
,
7721 struct pipe_debug_callback
*debug
)
7723 if (shader
->key
.as_es
)
7726 /* TES compiled as VS. */
7727 return si_get_vs_epilog(sscreen
, tm
, shader
, debug
,
7728 &shader
->key
.part
.tes
.epilog
);
7732 * Compile the TCS epilog function. This writes tesselation factors to memory
7733 * based on the output primitive type of the tesselator (determined by TES).
7735 static void si_build_tcs_epilog_function(struct si_shader_context
*ctx
,
7736 union si_shader_part_key
*key
)
7738 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
7739 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
7740 LLVMTypeRef params
[16];
7742 int last_sgpr
, num_params
;
7744 /* Declare inputs. Only RW_BUFFERS and TESS_FACTOR_OFFSET are used. */
7745 params
[SI_PARAM_RW_BUFFERS
] = const_array(ctx
->v16i8
, SI_NUM_RW_BUFFERS
);
7746 params
[SI_PARAM_CONST_BUFFERS
] = ctx
->i64
;
7747 params
[SI_PARAM_SAMPLERS
] = ctx
->i64
;
7748 params
[SI_PARAM_IMAGES
] = ctx
->i64
;
7749 params
[SI_PARAM_SHADER_BUFFERS
] = ctx
->i64
;
7750 params
[SI_PARAM_TCS_OFFCHIP_LAYOUT
] = ctx
->i32
;
7751 params
[SI_PARAM_TCS_OUT_OFFSETS
] = ctx
->i32
;
7752 params
[SI_PARAM_TCS_OUT_LAYOUT
] = ctx
->i32
;
7753 params
[SI_PARAM_TCS_IN_LAYOUT
] = ctx
->i32
;
7754 params
[ctx
->param_oc_lds
= SI_PARAM_TCS_OC_LDS
] = ctx
->i32
;
7755 params
[SI_PARAM_TESS_FACTOR_OFFSET
] = ctx
->i32
;
7756 last_sgpr
= SI_PARAM_TESS_FACTOR_OFFSET
;
7757 num_params
= last_sgpr
+ 1;
7759 params
[num_params
++] = ctx
->i32
; /* patch index within the wave (REL_PATCH_ID) */
7760 params
[num_params
++] = ctx
->i32
; /* invocation ID within the patch */
7761 params
[num_params
++] = ctx
->i32
; /* LDS offset where tess factors should be loaded from */
7763 /* Create the function. */
7764 si_create_function(ctx
, "tcs_epilog", NULL
, 0, params
, num_params
, last_sgpr
);
7765 declare_tess_lds(ctx
);
7766 func
= ctx
->main_fn
;
7768 si_write_tess_factors(bld_base
,
7769 LLVMGetParam(func
, last_sgpr
+ 1),
7770 LLVMGetParam(func
, last_sgpr
+ 2),
7771 LLVMGetParam(func
, last_sgpr
+ 3));
7773 LLVMBuildRetVoid(gallivm
->builder
);
7777 * Select and compile (or reuse) TCS parts (epilog).
7779 static bool si_shader_select_tcs_parts(struct si_screen
*sscreen
,
7780 LLVMTargetMachineRef tm
,
7781 struct si_shader
*shader
,
7782 struct pipe_debug_callback
*debug
)
7784 union si_shader_part_key epilog_key
;
7786 /* Get the epilog. */
7787 memset(&epilog_key
, 0, sizeof(epilog_key
));
7788 epilog_key
.tcs_epilog
.states
= shader
->key
.part
.tcs
.epilog
;
7790 shader
->epilog
= si_get_shader_part(sscreen
, &sscreen
->tcs_epilogs
,
7791 PIPE_SHADER_TESS_CTRL
, false,
7792 &epilog_key
, tm
, debug
,
7793 si_build_tcs_epilog_function
,
7794 "Tessellation Control Shader Epilog");
7795 return shader
->epilog
!= NULL
;
7799 * Select and compile (or reuse) GS parts (prolog).
7801 static bool si_shader_select_gs_parts(struct si_screen
*sscreen
,
7802 LLVMTargetMachineRef tm
,
7803 struct si_shader
*shader
,
7804 struct pipe_debug_callback
*debug
)
7806 union si_shader_part_key prolog_key
;
7808 if (!shader
->key
.part
.gs
.prolog
.tri_strip_adj_fix
)
7811 memset(&prolog_key
, 0, sizeof(prolog_key
));
7812 prolog_key
.gs_prolog
.states
= shader
->key
.part
.gs
.prolog
;
7814 shader
->prolog
= si_get_shader_part(sscreen
, &sscreen
->gs_prologs
,
7815 PIPE_SHADER_GEOMETRY
, true,
7816 &prolog_key
, tm
, debug
,
7817 si_build_gs_prolog_function
,
7818 "Geometry Shader Prolog");
7819 return shader
->prolog
!= NULL
;
7823 * Build the pixel shader prolog function. This handles:
7824 * - two-side color selection and interpolation
7825 * - overriding interpolation parameters for the API PS
7826 * - polygon stippling
7828 * All preloaded SGPRs and VGPRs are passed through unmodified unless they are
7829 * overriden by other states. (e.g. per-sample interpolation)
7830 * Interpolated colors are stored after the preloaded VGPRs.
7832 static void si_build_ps_prolog_function(struct si_shader_context
*ctx
,
7833 union si_shader_part_key
*key
)
7835 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
7836 LLVMTypeRef
*params
;
7837 LLVMValueRef ret
, func
;
7838 int last_sgpr
, num_params
, num_returns
, i
, num_color_channels
;
7840 assert(si_need_ps_prolog(key
));
7842 /* Number of inputs + 8 color elements. */
7843 params
= alloca((key
->ps_prolog
.num_input_sgprs
+
7844 key
->ps_prolog
.num_input_vgprs
+ 8) *
7845 sizeof(LLVMTypeRef
));
7847 /* Declare inputs. */
7849 for (i
= 0; i
< key
->ps_prolog
.num_input_sgprs
; i
++)
7850 params
[num_params
++] = ctx
->i32
;
7851 last_sgpr
= num_params
- 1;
7853 for (i
= 0; i
< key
->ps_prolog
.num_input_vgprs
; i
++)
7854 params
[num_params
++] = ctx
->f32
;
7856 /* Declare outputs (same as inputs + add colors if needed) */
7857 num_returns
= num_params
;
7858 num_color_channels
= util_bitcount(key
->ps_prolog
.colors_read
);
7859 for (i
= 0; i
< num_color_channels
; i
++)
7860 params
[num_returns
++] = ctx
->f32
;
7862 /* Create the function. */
7863 si_create_function(ctx
, "ps_prolog", params
, num_returns
, params
,
7864 num_params
, last_sgpr
);
7865 func
= ctx
->main_fn
;
7867 /* Copy inputs to outputs. This should be no-op, as the registers match,
7868 * but it will prevent the compiler from overwriting them unintentionally.
7870 ret
= ctx
->return_value
;
7871 for (i
= 0; i
< num_params
; i
++) {
7872 LLVMValueRef p
= LLVMGetParam(func
, i
);
7873 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
, p
, i
, "");
7876 /* Polygon stippling. */
7877 if (key
->ps_prolog
.states
.poly_stipple
) {
7878 /* POS_FIXED_PT is always last. */
7879 unsigned pos
= key
->ps_prolog
.num_input_sgprs
+
7880 key
->ps_prolog
.num_input_vgprs
- 1;
7881 LLVMValueRef ptr
[2], list
;
7883 /* Get the pointer to rw buffers. */
7884 ptr
[0] = LLVMGetParam(func
, SI_SGPR_RW_BUFFERS
);
7885 ptr
[1] = LLVMGetParam(func
, SI_SGPR_RW_BUFFERS_HI
);
7886 list
= lp_build_gather_values(gallivm
, ptr
, 2);
7887 list
= LLVMBuildBitCast(gallivm
->builder
, list
, ctx
->i64
, "");
7888 list
= LLVMBuildIntToPtr(gallivm
->builder
, list
,
7889 const_array(ctx
->v16i8
, SI_NUM_RW_BUFFERS
), "");
7891 si_llvm_emit_polygon_stipple(ctx
, list
, pos
);
7894 if (key
->ps_prolog
.states
.bc_optimize_for_persp
||
7895 key
->ps_prolog
.states
.bc_optimize_for_linear
) {
7896 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7897 LLVMValueRef center
[2], centroid
[2], tmp
, bc_optimize
;
7899 /* The shader should do: if (PRIM_MASK[31]) CENTROID = CENTER;
7900 * The hw doesn't compute CENTROID if the whole wave only
7901 * contains fully-covered quads.
7903 * PRIM_MASK is after user SGPRs.
7905 bc_optimize
= LLVMGetParam(func
, SI_PS_NUM_USER_SGPR
);
7906 bc_optimize
= LLVMBuildLShr(gallivm
->builder
, bc_optimize
,
7907 LLVMConstInt(ctx
->i32
, 31, 0), "");
7908 bc_optimize
= LLVMBuildTrunc(gallivm
->builder
, bc_optimize
,
7911 if (key
->ps_prolog
.states
.bc_optimize_for_persp
) {
7912 /* Read PERSP_CENTER. */
7913 for (i
= 0; i
< 2; i
++)
7914 center
[i
] = LLVMGetParam(func
, base
+ 2 + i
);
7915 /* Read PERSP_CENTROID. */
7916 for (i
= 0; i
< 2; i
++)
7917 centroid
[i
] = LLVMGetParam(func
, base
+ 4 + i
);
7918 /* Select PERSP_CENTROID. */
7919 for (i
= 0; i
< 2; i
++) {
7920 tmp
= LLVMBuildSelect(gallivm
->builder
, bc_optimize
,
7921 center
[i
], centroid
[i
], "");
7922 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
,
7923 tmp
, base
+ 4 + i
, "");
7926 if (key
->ps_prolog
.states
.bc_optimize_for_linear
) {
7927 /* Read LINEAR_CENTER. */
7928 for (i
= 0; i
< 2; i
++)
7929 center
[i
] = LLVMGetParam(func
, base
+ 8 + i
);
7930 /* Read LINEAR_CENTROID. */
7931 for (i
= 0; i
< 2; i
++)
7932 centroid
[i
] = LLVMGetParam(func
, base
+ 10 + i
);
7933 /* Select LINEAR_CENTROID. */
7934 for (i
= 0; i
< 2; i
++) {
7935 tmp
= LLVMBuildSelect(gallivm
->builder
, bc_optimize
,
7936 center
[i
], centroid
[i
], "");
7937 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
,
7938 tmp
, base
+ 10 + i
, "");
7943 /* Force per-sample interpolation. */
7944 if (key
->ps_prolog
.states
.force_persp_sample_interp
) {
7945 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7946 LLVMValueRef persp_sample
[2];
7948 /* Read PERSP_SAMPLE. */
7949 for (i
= 0; i
< 2; i
++)
7950 persp_sample
[i
] = LLVMGetParam(func
, base
+ i
);
7951 /* Overwrite PERSP_CENTER. */
7952 for (i
= 0; i
< 2; i
++)
7953 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
,
7954 persp_sample
[i
], base
+ 2 + i
, "");
7955 /* Overwrite PERSP_CENTROID. */
7956 for (i
= 0; i
< 2; i
++)
7957 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
,
7958 persp_sample
[i
], base
+ 4 + i
, "");
7960 if (key
->ps_prolog
.states
.force_linear_sample_interp
) {
7961 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7962 LLVMValueRef linear_sample
[2];
7964 /* Read LINEAR_SAMPLE. */
7965 for (i
= 0; i
< 2; i
++)
7966 linear_sample
[i
] = LLVMGetParam(func
, base
+ 6 + i
);
7967 /* Overwrite LINEAR_CENTER. */
7968 for (i
= 0; i
< 2; i
++)
7969 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
,
7970 linear_sample
[i
], base
+ 8 + i
, "");
7971 /* Overwrite LINEAR_CENTROID. */
7972 for (i
= 0; i
< 2; i
++)
7973 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
,
7974 linear_sample
[i
], base
+ 10 + i
, "");
7977 /* Force center interpolation. */
7978 if (key
->ps_prolog
.states
.force_persp_center_interp
) {
7979 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7980 LLVMValueRef persp_center
[2];
7982 /* Read PERSP_CENTER. */
7983 for (i
= 0; i
< 2; i
++)
7984 persp_center
[i
] = LLVMGetParam(func
, base
+ 2 + i
);
7985 /* Overwrite PERSP_SAMPLE. */
7986 for (i
= 0; i
< 2; i
++)
7987 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
,
7988 persp_center
[i
], base
+ i
, "");
7989 /* Overwrite PERSP_CENTROID. */
7990 for (i
= 0; i
< 2; i
++)
7991 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
,
7992 persp_center
[i
], base
+ 4 + i
, "");
7994 if (key
->ps_prolog
.states
.force_linear_center_interp
) {
7995 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7996 LLVMValueRef linear_center
[2];
7998 /* Read LINEAR_CENTER. */
7999 for (i
= 0; i
< 2; i
++)
8000 linear_center
[i
] = LLVMGetParam(func
, base
+ 8 + i
);
8001 /* Overwrite LINEAR_SAMPLE. */
8002 for (i
= 0; i
< 2; i
++)
8003 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
,
8004 linear_center
[i
], base
+ 6 + i
, "");
8005 /* Overwrite LINEAR_CENTROID. */
8006 for (i
= 0; i
< 2; i
++)
8007 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
,
8008 linear_center
[i
], base
+ 10 + i
, "");
8011 /* Interpolate colors. */
8012 for (i
= 0; i
< 2; i
++) {
8013 unsigned writemask
= (key
->ps_prolog
.colors_read
>> (i
* 4)) & 0xf;
8014 unsigned face_vgpr
= key
->ps_prolog
.num_input_sgprs
+
8015 key
->ps_prolog
.face_vgpr_index
;
8016 LLVMValueRef interp
[2], color
[4];
8017 LLVMValueRef interp_ij
= NULL
, prim_mask
= NULL
, face
= NULL
;
8022 /* If the interpolation qualifier is not CONSTANT (-1). */
8023 if (key
->ps_prolog
.color_interp_vgpr_index
[i
] != -1) {
8024 unsigned interp_vgpr
= key
->ps_prolog
.num_input_sgprs
+
8025 key
->ps_prolog
.color_interp_vgpr_index
[i
];
8027 /* Get the (i,j) updated by bc_optimize handling. */
8028 interp
[0] = LLVMBuildExtractValue(gallivm
->builder
, ret
,
8030 interp
[1] = LLVMBuildExtractValue(gallivm
->builder
, ret
,
8031 interp_vgpr
+ 1, "");
8032 interp_ij
= lp_build_gather_values(gallivm
, interp
, 2);
8035 /* Use the absolute location of the input. */
8036 prim_mask
= LLVMGetParam(func
, SI_PS_NUM_USER_SGPR
);
8038 if (key
->ps_prolog
.states
.color_two_side
) {
8039 face
= LLVMGetParam(func
, face_vgpr
);
8040 face
= LLVMBuildBitCast(gallivm
->builder
, face
, ctx
->i32
, "");
8043 interp_fs_input(ctx
,
8044 key
->ps_prolog
.color_attr_index
[i
],
8045 TGSI_SEMANTIC_COLOR
, i
,
8046 key
->ps_prolog
.num_interp_inputs
,
8047 key
->ps_prolog
.colors_read
, interp_ij
,
8048 prim_mask
, face
, color
);
8051 unsigned chan
= u_bit_scan(&writemask
);
8052 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
, color
[chan
],
8057 /* Tell LLVM to insert WQM instruction sequence when needed. */
8058 if (key
->ps_prolog
.wqm
) {
8059 LLVMAddTargetDependentFunctionAttr(func
,
8060 "amdgpu-ps-wqm-outputs", "");
8063 si_llvm_build_ret(ctx
, ret
);
8067 * Build the pixel shader epilog function. This handles everything that must be
8068 * emulated for pixel shader exports. (alpha-test, format conversions, etc)
8070 static void si_build_ps_epilog_function(struct si_shader_context
*ctx
,
8071 union si_shader_part_key
*key
)
8073 struct gallivm_state
*gallivm
= &ctx
->gallivm
;
8074 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
8075 LLVMTypeRef params
[16+8*4+3];
8076 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
8077 int last_sgpr
, num_params
, i
;
8078 struct si_ps_exports exp
= {};
8080 /* Declare input SGPRs. */
8081 params
[SI_PARAM_RW_BUFFERS
] = ctx
->i64
;
8082 params
[SI_PARAM_CONST_BUFFERS
] = ctx
->i64
;
8083 params
[SI_PARAM_SAMPLERS
] = ctx
->i64
;
8084 params
[SI_PARAM_IMAGES
] = ctx
->i64
;
8085 params
[SI_PARAM_SHADER_BUFFERS
] = ctx
->i64
;
8086 params
[SI_PARAM_ALPHA_REF
] = ctx
->f32
;
8087 last_sgpr
= SI_PARAM_ALPHA_REF
;
8089 /* Declare input VGPRs. */
8090 num_params
= (last_sgpr
+ 1) +
8091 util_bitcount(key
->ps_epilog
.colors_written
) * 4 +
8092 key
->ps_epilog
.writes_z
+
8093 key
->ps_epilog
.writes_stencil
+
8094 key
->ps_epilog
.writes_samplemask
;
8096 num_params
= MAX2(num_params
,
8097 last_sgpr
+ 1 + PS_EPILOG_SAMPLEMASK_MIN_LOC
+ 1);
8099 assert(num_params
<= ARRAY_SIZE(params
));
8101 for (i
= last_sgpr
+ 1; i
< num_params
; i
++)
8102 params
[i
] = ctx
->f32
;
8104 /* Create the function. */
8105 si_create_function(ctx
, "ps_epilog", NULL
, 0, params
, num_params
, last_sgpr
);
8106 /* Disable elimination of unused inputs. */
8107 si_llvm_add_attribute(ctx
->main_fn
,
8108 "InitialPSInputAddr", 0xffffff);
8110 /* Process colors. */
8111 unsigned vgpr
= last_sgpr
+ 1;
8112 unsigned colors_written
= key
->ps_epilog
.colors_written
;
8113 int last_color_export
= -1;
8115 /* Find the last color export. */
8116 if (!key
->ps_epilog
.writes_z
&&
8117 !key
->ps_epilog
.writes_stencil
&&
8118 !key
->ps_epilog
.writes_samplemask
) {
8119 unsigned spi_format
= key
->ps_epilog
.states
.spi_shader_col_format
;
8121 /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
8122 if (colors_written
== 0x1 && key
->ps_epilog
.states
.last_cbuf
> 0) {
8123 /* Just set this if any of the colorbuffers are enabled. */
8125 ((1llu << (4 * (key
->ps_epilog
.states
.last_cbuf
+ 1))) - 1))
8126 last_color_export
= 0;
8128 for (i
= 0; i
< 8; i
++)
8129 if (colors_written
& (1 << i
) &&
8130 (spi_format
>> (i
* 4)) & 0xf)
8131 last_color_export
= i
;
8135 while (colors_written
) {
8136 LLVMValueRef color
[4];
8137 int mrt
= u_bit_scan(&colors_written
);
8139 for (i
= 0; i
< 4; i
++)
8140 color
[i
] = LLVMGetParam(ctx
->main_fn
, vgpr
++);
8142 si_export_mrt_color(bld_base
, color
, mrt
,
8144 mrt
== last_color_export
, &exp
);
8147 /* Process depth, stencil, samplemask. */
8148 if (key
->ps_epilog
.writes_z
)
8149 depth
= LLVMGetParam(ctx
->main_fn
, vgpr
++);
8150 if (key
->ps_epilog
.writes_stencil
)
8151 stencil
= LLVMGetParam(ctx
->main_fn
, vgpr
++);
8152 if (key
->ps_epilog
.writes_samplemask
)
8153 samplemask
= LLVMGetParam(ctx
->main_fn
, vgpr
++);
8155 if (depth
|| stencil
|| samplemask
)
8156 si_export_mrt_z(bld_base
, depth
, stencil
, samplemask
, &exp
);
8157 else if (last_color_export
== -1)
8158 si_export_null(bld_base
);
8161 si_emit_ps_exports(ctx
, &exp
);
8164 LLVMBuildRetVoid(gallivm
->builder
);
8168 * Select and compile (or reuse) pixel shader parts (prolog & epilog).
8170 static bool si_shader_select_ps_parts(struct si_screen
*sscreen
,
8171 LLVMTargetMachineRef tm
,
8172 struct si_shader
*shader
,
8173 struct pipe_debug_callback
*debug
)
8175 union si_shader_part_key prolog_key
;
8176 union si_shader_part_key epilog_key
;
8178 /* Get the prolog. */
8179 si_get_ps_prolog_key(shader
, &prolog_key
, true);
8181 /* The prolog is a no-op if these aren't set. */
8182 if (si_need_ps_prolog(&prolog_key
)) {
8184 si_get_shader_part(sscreen
, &sscreen
->ps_prologs
,
8185 PIPE_SHADER_FRAGMENT
, true,
8186 &prolog_key
, tm
, debug
,
8187 si_build_ps_prolog_function
,
8188 "Fragment Shader Prolog");
8189 if (!shader
->prolog
)
8193 /* Get the epilog. */
8194 si_get_ps_epilog_key(shader
, &epilog_key
);
8197 si_get_shader_part(sscreen
, &sscreen
->ps_epilogs
,
8198 PIPE_SHADER_FRAGMENT
, false,
8199 &epilog_key
, tm
, debug
,
8200 si_build_ps_epilog_function
,
8201 "Fragment Shader Epilog");
8202 if (!shader
->epilog
)
8205 /* Enable POS_FIXED_PT if polygon stippling is enabled. */
8206 if (shader
->key
.part
.ps
.prolog
.poly_stipple
) {
8207 shader
->config
.spi_ps_input_ena
|= S_0286CC_POS_FIXED_PT_ENA(1);
8208 assert(G_0286CC_POS_FIXED_PT_ENA(shader
->config
.spi_ps_input_addr
));
8211 /* Set up the enable bits for per-sample shading if needed. */
8212 if (shader
->key
.part
.ps
.prolog
.force_persp_sample_interp
&&
8213 (G_0286CC_PERSP_CENTER_ENA(shader
->config
.spi_ps_input_ena
) ||
8214 G_0286CC_PERSP_CENTROID_ENA(shader
->config
.spi_ps_input_ena
))) {
8215 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_CENTER_ENA
;
8216 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_CENTROID_ENA
;
8217 shader
->config
.spi_ps_input_ena
|= S_0286CC_PERSP_SAMPLE_ENA(1);
8219 if (shader
->key
.part
.ps
.prolog
.force_linear_sample_interp
&&
8220 (G_0286CC_LINEAR_CENTER_ENA(shader
->config
.spi_ps_input_ena
) ||
8221 G_0286CC_LINEAR_CENTROID_ENA(shader
->config
.spi_ps_input_ena
))) {
8222 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_CENTER_ENA
;
8223 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_CENTROID_ENA
;
8224 shader
->config
.spi_ps_input_ena
|= S_0286CC_LINEAR_SAMPLE_ENA(1);
8226 if (shader
->key
.part
.ps
.prolog
.force_persp_center_interp
&&
8227 (G_0286CC_PERSP_SAMPLE_ENA(shader
->config
.spi_ps_input_ena
) ||
8228 G_0286CC_PERSP_CENTROID_ENA(shader
->config
.spi_ps_input_ena
))) {
8229 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_SAMPLE_ENA
;
8230 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_CENTROID_ENA
;
8231 shader
->config
.spi_ps_input_ena
|= S_0286CC_PERSP_CENTER_ENA(1);
8233 if (shader
->key
.part
.ps
.prolog
.force_linear_center_interp
&&
8234 (G_0286CC_LINEAR_SAMPLE_ENA(shader
->config
.spi_ps_input_ena
) ||
8235 G_0286CC_LINEAR_CENTROID_ENA(shader
->config
.spi_ps_input_ena
))) {
8236 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_SAMPLE_ENA
;
8237 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_CENTROID_ENA
;
8238 shader
->config
.spi_ps_input_ena
|= S_0286CC_LINEAR_CENTER_ENA(1);
8241 /* POW_W_FLOAT requires that one of the perspective weights is enabled. */
8242 if (G_0286CC_POS_W_FLOAT_ENA(shader
->config
.spi_ps_input_ena
) &&
8243 !(shader
->config
.spi_ps_input_ena
& 0xf)) {
8244 shader
->config
.spi_ps_input_ena
|= S_0286CC_PERSP_CENTER_ENA(1);
8245 assert(G_0286CC_PERSP_CENTER_ENA(shader
->config
.spi_ps_input_addr
));
8248 /* At least one pair of interpolation weights must be enabled. */
8249 if (!(shader
->config
.spi_ps_input_ena
& 0x7f)) {
8250 shader
->config
.spi_ps_input_ena
|= S_0286CC_LINEAR_CENTER_ENA(1);
8251 assert(G_0286CC_LINEAR_CENTER_ENA(shader
->config
.spi_ps_input_addr
));
8254 /* The sample mask input is always enabled, because the API shader always
8255 * passes it through to the epilog. Disable it here if it's unused.
8257 if (!shader
->key
.part
.ps
.epilog
.poly_line_smoothing
&&
8258 !shader
->selector
->info
.reads_samplemask
)
8259 shader
->config
.spi_ps_input_ena
&= C_0286CC_SAMPLE_COVERAGE_ENA
;
8264 void si_multiwave_lds_size_workaround(struct si_screen
*sscreen
,
8267 /* SPI barrier management bug:
8268 * Make sure we have at least 4k of LDS in use to avoid the bug.
8269 * It applies to workgroup sizes of more than one wavefront.
8271 if (sscreen
->b
.family
== CHIP_BONAIRE
||
8272 sscreen
->b
.family
== CHIP_KABINI
||
8273 sscreen
->b
.family
== CHIP_MULLINS
)
8274 *lds_size
= MAX2(*lds_size
, 8);
8277 static void si_fix_resource_usage(struct si_screen
*sscreen
,
8278 struct si_shader
*shader
)
8280 unsigned min_sgprs
= shader
->info
.num_input_sgprs
+ 2; /* VCC */
8282 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
, min_sgprs
);
8284 if (shader
->selector
->type
== PIPE_SHADER_COMPUTE
&&
8285 si_get_max_workgroup_size(shader
) > 64) {
8286 si_multiwave_lds_size_workaround(sscreen
,
8287 &shader
->config
.lds_size
);
8291 int si_shader_create(struct si_screen
*sscreen
, LLVMTargetMachineRef tm
,
8292 struct si_shader
*shader
,
8293 struct pipe_debug_callback
*debug
)
8295 struct si_shader_selector
*sel
= shader
->selector
;
8296 struct si_shader
*mainp
= *si_get_main_shader_part(sel
, &shader
->key
);
8299 /* LS, ES, VS are compiled on demand if the main part hasn't been
8300 * compiled for that stage.
8302 * Vertex shaders are compiled on demand when a vertex fetch
8303 * workaround must be applied.
8305 if (shader
->is_monolithic
) {
8306 /* Monolithic shader (compiled as a whole, has many variants,
8307 * may take a long time to compile).
8309 r
= si_compile_tgsi_shader(sscreen
, tm
, shader
, true, debug
);
8313 /* The shader consists of 2-3 parts:
8315 * - the middle part is the user shader, it has 1 variant only
8316 * and it was compiled during the creation of the shader
8318 * - the prolog part is inserted at the beginning
8319 * - the epilog part is inserted at the end
8321 * The prolog and epilog have many (but simple) variants.
8324 /* Copy the compiled TGSI shader data over. */
8325 shader
->is_binary_shared
= true;
8326 shader
->binary
= mainp
->binary
;
8327 shader
->config
= mainp
->config
;
8328 shader
->info
.num_input_sgprs
= mainp
->info
.num_input_sgprs
;
8329 shader
->info
.num_input_vgprs
= mainp
->info
.num_input_vgprs
;
8330 shader
->info
.face_vgpr_index
= mainp
->info
.face_vgpr_index
;
8331 memcpy(shader
->info
.vs_output_param_offset
,
8332 mainp
->info
.vs_output_param_offset
,
8333 sizeof(mainp
->info
.vs_output_param_offset
));
8334 shader
->info
.uses_instanceid
= mainp
->info
.uses_instanceid
;
8335 shader
->info
.nr_pos_exports
= mainp
->info
.nr_pos_exports
;
8336 shader
->info
.nr_param_exports
= mainp
->info
.nr_param_exports
;
8338 /* Select prologs and/or epilogs. */
8339 switch (sel
->type
) {
8340 case PIPE_SHADER_VERTEX
:
8341 if (!si_shader_select_vs_parts(sscreen
, tm
, shader
, debug
))
8344 case PIPE_SHADER_TESS_CTRL
:
8345 if (!si_shader_select_tcs_parts(sscreen
, tm
, shader
, debug
))
8348 case PIPE_SHADER_TESS_EVAL
:
8349 if (!si_shader_select_tes_parts(sscreen
, tm
, shader
, debug
))
8352 case PIPE_SHADER_GEOMETRY
:
8353 if (!si_shader_select_gs_parts(sscreen
, tm
, shader
, debug
))
8356 case PIPE_SHADER_FRAGMENT
:
8357 if (!si_shader_select_ps_parts(sscreen
, tm
, shader
, debug
))
8360 /* Make sure we have at least as many VGPRs as there
8361 * are allocated inputs.
8363 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
8364 shader
->info
.num_input_vgprs
);
8368 /* Update SGPR and VGPR counts. */
8369 if (shader
->prolog
) {
8370 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
,
8371 shader
->prolog
->config
.num_sgprs
);
8372 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
8373 shader
->prolog
->config
.num_vgprs
);
8375 if (shader
->epilog
) {
8376 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
,
8377 shader
->epilog
->config
.num_sgprs
);
8378 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
8379 shader
->epilog
->config
.num_vgprs
);
8383 si_fix_resource_usage(sscreen
, shader
);
8384 si_shader_dump(sscreen
, shader
, debug
, sel
->info
.processor
,
8388 r
= si_shader_binary_upload(sscreen
, shader
);
8390 fprintf(stderr
, "LLVM failed to upload shader\n");
8397 void si_shader_destroy(struct si_shader
*shader
)
8399 if (shader
->scratch_bo
)
8400 r600_resource_reference(&shader
->scratch_bo
, NULL
);
8402 r600_resource_reference(&shader
->bo
, NULL
);
8404 if (!shader
->is_binary_shared
)
8405 radeon_shader_binary_clean(&shader
->binary
);
8407 free(shader
->shader_log
);