2 * Copyright 2012 Advanced Micro Devices, Inc.
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * on the rights to use, copy, modify, merge, publish, distribute, sub
9 * license, and/or sell copies of the Software, and to permit persons to whom
10 * the Software is furnished to do so, subject to the following conditions:
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22 * USE OR OTHER DEALINGS IN THE SOFTWARE.
25 #include "util/u_memory.h"
26 #include "util/u_string.h"
27 #include "tgsi/tgsi_build.h"
28 #include "tgsi/tgsi_strings.h"
29 #include "tgsi/tgsi_util.h"
30 #include "tgsi/tgsi_dump.h"
31 #include "tgsi/tgsi_from_mesa.h"
33 #include "ac_binary.h"
34 #include "ac_exp_param.h"
35 #include "ac_shader_util.h"
37 #include "ac_llvm_util.h"
38 #include "si_shader_internal.h"
42 #include "compiler/nir/nir.h"
44 static const char scratch_rsrc_dword0_symbol
[] =
45 "SCRATCH_RSRC_DWORD0";
47 static const char scratch_rsrc_dword1_symbol
[] =
48 "SCRATCH_RSRC_DWORD1";
50 static void si_init_shader_ctx(struct si_shader_context
*ctx
,
51 struct si_screen
*sscreen
,
52 struct ac_llvm_compiler
*compiler
);
54 static void si_llvm_emit_barrier(const struct lp_build_tgsi_action
*action
,
55 struct lp_build_tgsi_context
*bld_base
,
56 struct lp_build_emit_data
*emit_data
);
58 static void si_dump_shader_key(unsigned processor
, const struct si_shader
*shader
,
61 static void si_build_vs_prolog_function(struct si_shader_context
*ctx
,
62 union si_shader_part_key
*key
);
63 static void si_build_tcs_epilog_function(struct si_shader_context
*ctx
,
64 union si_shader_part_key
*key
);
65 static void si_build_ps_prolog_function(struct si_shader_context
*ctx
,
66 union si_shader_part_key
*key
);
67 static void si_build_ps_epilog_function(struct si_shader_context
*ctx
,
68 union si_shader_part_key
*key
);
69 static void si_fix_resource_usage(struct si_screen
*sscreen
,
70 struct si_shader
*shader
);
72 /* Ideally pass the sample mask input to the PS epilog as v14, which
73 * is its usual location, so that the shader doesn't have to add v_mov.
75 #define PS_EPILOG_SAMPLEMASK_MIN_LOC 14
77 static bool llvm_type_is_64bit(struct si_shader_context
*ctx
,
80 if (type
== ctx
->ac
.i64
|| type
== ctx
->ac
.f64
)
86 /** Whether the shader runs as a combination of multiple API shaders */
87 static bool is_multi_part_shader(struct si_shader_context
*ctx
)
89 if (ctx
->screen
->info
.chip_class
<= GFX8
)
92 return ctx
->shader
->key
.as_ls
||
93 ctx
->shader
->key
.as_es
||
94 ctx
->type
== PIPE_SHADER_TESS_CTRL
||
95 ctx
->type
== PIPE_SHADER_GEOMETRY
;
98 /** Whether the shader runs on a merged HW stage (LSHS or ESGS) */
99 static bool is_merged_shader(struct si_shader_context
*ctx
)
101 return ctx
->shader
->key
.as_ngg
|| is_multi_part_shader(ctx
);
104 void si_init_function_info(struct si_function_info
*fninfo
)
106 fninfo
->num_params
= 0;
107 fninfo
->num_sgpr_params
= 0;
110 unsigned add_arg_assign(struct si_function_info
*fninfo
,
111 enum si_arg_regfile regfile
, LLVMTypeRef type
,
112 LLVMValueRef
*assign
)
114 assert(regfile
!= ARG_SGPR
|| fninfo
->num_sgpr_params
== fninfo
->num_params
);
116 unsigned idx
= fninfo
->num_params
++;
117 assert(idx
< ARRAY_SIZE(fninfo
->types
));
119 if (regfile
== ARG_SGPR
)
120 fninfo
->num_sgpr_params
= fninfo
->num_params
;
122 fninfo
->types
[idx
] = type
;
123 fninfo
->assign
[idx
] = assign
;
127 static unsigned add_arg(struct si_function_info
*fninfo
,
128 enum si_arg_regfile regfile
, LLVMTypeRef type
)
130 return add_arg_assign(fninfo
, regfile
, type
, NULL
);
133 static void add_arg_assign_checked(struct si_function_info
*fninfo
,
134 enum si_arg_regfile regfile
, LLVMTypeRef type
,
135 LLVMValueRef
*assign
, unsigned idx
)
137 MAYBE_UNUSED
unsigned actual
= add_arg_assign(fninfo
, regfile
, type
, assign
);
138 assert(actual
== idx
);
141 static void add_arg_checked(struct si_function_info
*fninfo
,
142 enum si_arg_regfile regfile
, LLVMTypeRef type
,
145 add_arg_assign_checked(fninfo
, regfile
, type
, NULL
, idx
);
149 * Returns a unique index for a per-patch semantic name and index. The index
150 * must be less than 32, so that a 32-bit bitmask of used inputs or outputs
153 unsigned si_shader_io_get_unique_index_patch(unsigned semantic_name
, unsigned index
)
155 switch (semantic_name
) {
156 case TGSI_SEMANTIC_TESSOUTER
:
158 case TGSI_SEMANTIC_TESSINNER
:
160 case TGSI_SEMANTIC_PATCH
:
165 assert(!"invalid semantic name");
171 * Returns a unique index for a semantic name and index. The index must be
172 * less than 64, so that a 64-bit bitmask of used inputs or outputs can be
175 unsigned si_shader_io_get_unique_index(unsigned semantic_name
, unsigned index
,
178 switch (semantic_name
) {
179 case TGSI_SEMANTIC_POSITION
:
181 case TGSI_SEMANTIC_GENERIC
:
182 /* Since some shader stages use the the highest used IO index
183 * to determine the size to allocate for inputs/outputs
184 * (in LDS, tess and GS rings). GENERIC should be placed right
185 * after POSITION to make that size as small as possible.
187 if (index
< SI_MAX_IO_GENERIC
)
190 assert(!"invalid generic index");
192 case TGSI_SEMANTIC_PSIZE
:
193 return SI_MAX_IO_GENERIC
+ 1;
194 case TGSI_SEMANTIC_CLIPDIST
:
196 return SI_MAX_IO_GENERIC
+ 2 + index
;
197 case TGSI_SEMANTIC_FOG
:
198 return SI_MAX_IO_GENERIC
+ 4;
199 case TGSI_SEMANTIC_LAYER
:
200 return SI_MAX_IO_GENERIC
+ 5;
201 case TGSI_SEMANTIC_VIEWPORT_INDEX
:
202 return SI_MAX_IO_GENERIC
+ 6;
203 case TGSI_SEMANTIC_PRIMID
:
204 return SI_MAX_IO_GENERIC
+ 7;
205 case TGSI_SEMANTIC_COLOR
:
207 return SI_MAX_IO_GENERIC
+ 8 + index
;
208 case TGSI_SEMANTIC_BCOLOR
:
210 /* If it's a varying, COLOR and BCOLOR alias. */
212 return SI_MAX_IO_GENERIC
+ 8 + index
;
214 return SI_MAX_IO_GENERIC
+ 10 + index
;
215 case TGSI_SEMANTIC_TEXCOORD
:
217 STATIC_ASSERT(SI_MAX_IO_GENERIC
+ 12 + 8 <= 63);
218 return SI_MAX_IO_GENERIC
+ 12 + index
;
219 case TGSI_SEMANTIC_CLIPVERTEX
:
222 fprintf(stderr
, "invalid semantic name = %u\n", semantic_name
);
223 assert(!"invalid semantic name");
229 * Get the value of a shader input parameter and extract a bitfield.
231 static LLVMValueRef
unpack_llvm_param(struct si_shader_context
*ctx
,
232 LLVMValueRef value
, unsigned rshift
,
235 if (LLVMGetTypeKind(LLVMTypeOf(value
)) == LLVMFloatTypeKind
)
236 value
= ac_to_integer(&ctx
->ac
, value
);
239 value
= LLVMBuildLShr(ctx
->ac
.builder
, value
,
240 LLVMConstInt(ctx
->i32
, rshift
, 0), "");
242 if (rshift
+ bitwidth
< 32) {
243 unsigned mask
= (1 << bitwidth
) - 1;
244 value
= LLVMBuildAnd(ctx
->ac
.builder
, value
,
245 LLVMConstInt(ctx
->i32
, mask
, 0), "");
251 LLVMValueRef
si_unpack_param(struct si_shader_context
*ctx
,
252 unsigned param
, unsigned rshift
,
255 LLVMValueRef value
= LLVMGetParam(ctx
->main_fn
, param
);
257 return unpack_llvm_param(ctx
, value
, rshift
, bitwidth
);
260 static LLVMValueRef
get_rel_patch_id(struct si_shader_context
*ctx
)
263 case PIPE_SHADER_TESS_CTRL
:
264 return unpack_llvm_param(ctx
, ctx
->abi
.tcs_rel_ids
, 0, 8);
266 case PIPE_SHADER_TESS_EVAL
:
267 return LLVMGetParam(ctx
->main_fn
,
268 ctx
->param_tes_rel_patch_id
);
276 /* Tessellation shaders pass outputs to the next shader using LDS.
278 * LS outputs = TCS inputs
279 * TCS outputs = TES inputs
282 * - TCS inputs for patch 0
283 * - TCS inputs for patch 1
284 * - TCS inputs for patch 2 = get_tcs_in_current_patch_offset (if RelPatchID==2)
286 * - TCS outputs for patch 0 = get_tcs_out_patch0_offset
287 * - Per-patch TCS outputs for patch 0 = get_tcs_out_patch0_patch_data_offset
288 * - TCS outputs for patch 1
289 * - Per-patch TCS outputs for patch 1
290 * - TCS outputs for patch 2 = get_tcs_out_current_patch_offset (if RelPatchID==2)
291 * - Per-patch TCS outputs for patch 2 = get_tcs_out_current_patch_data_offset (if RelPatchID==2)
294 * All three shaders VS(LS), TCS, TES share the same LDS space.
298 get_tcs_in_patch_stride(struct si_shader_context
*ctx
)
300 return si_unpack_param(ctx
, ctx
->param_vs_state_bits
, 8, 13);
303 static unsigned get_tcs_out_vertex_dw_stride_constant(struct si_shader_context
*ctx
)
305 assert(ctx
->type
== PIPE_SHADER_TESS_CTRL
);
307 if (ctx
->shader
->key
.mono
.u
.ff_tcs_inputs_to_copy
)
308 return util_last_bit64(ctx
->shader
->key
.mono
.u
.ff_tcs_inputs_to_copy
) * 4;
310 return util_last_bit64(ctx
->shader
->selector
->outputs_written
) * 4;
313 static LLVMValueRef
get_tcs_out_vertex_dw_stride(struct si_shader_context
*ctx
)
315 unsigned stride
= get_tcs_out_vertex_dw_stride_constant(ctx
);
317 return LLVMConstInt(ctx
->i32
, stride
, 0);
320 static LLVMValueRef
get_tcs_out_patch_stride(struct si_shader_context
*ctx
)
322 if (ctx
->shader
->key
.mono
.u
.ff_tcs_inputs_to_copy
)
323 return si_unpack_param(ctx
, ctx
->param_tcs_out_lds_layout
, 0, 13);
325 const struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
326 unsigned tcs_out_vertices
= info
->properties
[TGSI_PROPERTY_TCS_VERTICES_OUT
];
327 unsigned vertex_dw_stride
= get_tcs_out_vertex_dw_stride_constant(ctx
);
328 unsigned num_patch_outputs
= util_last_bit64(ctx
->shader
->selector
->patch_outputs_written
);
329 unsigned patch_dw_stride
= tcs_out_vertices
* vertex_dw_stride
+
330 num_patch_outputs
* 4;
331 return LLVMConstInt(ctx
->i32
, patch_dw_stride
, 0);
335 get_tcs_out_patch0_offset(struct si_shader_context
*ctx
)
337 return LLVMBuildMul(ctx
->ac
.builder
,
339 ctx
->param_tcs_out_lds_offsets
,
341 LLVMConstInt(ctx
->i32
, 4, 0), "");
345 get_tcs_out_patch0_patch_data_offset(struct si_shader_context
*ctx
)
347 return LLVMBuildMul(ctx
->ac
.builder
,
349 ctx
->param_tcs_out_lds_offsets
,
351 LLVMConstInt(ctx
->i32
, 4, 0), "");
355 get_tcs_in_current_patch_offset(struct si_shader_context
*ctx
)
357 LLVMValueRef patch_stride
= get_tcs_in_patch_stride(ctx
);
358 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
360 return LLVMBuildMul(ctx
->ac
.builder
, patch_stride
, rel_patch_id
, "");
364 get_tcs_out_current_patch_offset(struct si_shader_context
*ctx
)
366 LLVMValueRef patch0_offset
= get_tcs_out_patch0_offset(ctx
);
367 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(ctx
);
368 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
370 return ac_build_imad(&ctx
->ac
, patch_stride
, rel_patch_id
, patch0_offset
);
374 get_tcs_out_current_patch_data_offset(struct si_shader_context
*ctx
)
376 LLVMValueRef patch0_patch_data_offset
=
377 get_tcs_out_patch0_patch_data_offset(ctx
);
378 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(ctx
);
379 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
381 return ac_build_imad(&ctx
->ac
, patch_stride
, rel_patch_id
, patch0_patch_data_offset
);
384 static LLVMValueRef
get_num_tcs_out_vertices(struct si_shader_context
*ctx
)
386 unsigned tcs_out_vertices
=
387 ctx
->shader
->selector
?
388 ctx
->shader
->selector
->info
.properties
[TGSI_PROPERTY_TCS_VERTICES_OUT
] : 0;
390 /* If !tcs_out_vertices, it's either the fixed-func TCS or the TCS epilog. */
391 if (ctx
->type
== PIPE_SHADER_TESS_CTRL
&& tcs_out_vertices
)
392 return LLVMConstInt(ctx
->i32
, tcs_out_vertices
, 0);
394 return si_unpack_param(ctx
, ctx
->param_tcs_offchip_layout
, 6, 6);
397 static LLVMValueRef
get_tcs_in_vertex_dw_stride(struct si_shader_context
*ctx
)
402 case PIPE_SHADER_VERTEX
:
403 stride
= ctx
->shader
->selector
->lshs_vertex_stride
/ 4;
404 return LLVMConstInt(ctx
->i32
, stride
, 0);
406 case PIPE_SHADER_TESS_CTRL
:
407 if (ctx
->screen
->info
.chip_class
>= GFX9
&&
408 ctx
->shader
->is_monolithic
) {
409 stride
= ctx
->shader
->key
.part
.tcs
.ls
->lshs_vertex_stride
/ 4;
410 return LLVMConstInt(ctx
->i32
, stride
, 0);
412 return si_unpack_param(ctx
, ctx
->param_vs_state_bits
, 24, 8);
420 static LLVMValueRef
unpack_sint16(struct si_shader_context
*ctx
,
421 LLVMValueRef i32
, unsigned index
)
426 return LLVMBuildAShr(ctx
->ac
.builder
, i32
,
427 LLVMConstInt(ctx
->i32
, 16, 0), "");
429 return LLVMBuildSExt(ctx
->ac
.builder
,
430 LLVMBuildTrunc(ctx
->ac
.builder
, i32
,
435 void si_llvm_load_input_vs(
436 struct si_shader_context
*ctx
,
437 unsigned input_index
,
440 const struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
441 unsigned vs_blit_property
= info
->properties
[TGSI_PROPERTY_VS_BLIT_SGPRS
];
443 if (vs_blit_property
) {
444 LLVMValueRef vertex_id
= ctx
->abi
.vertex_id
;
445 LLVMValueRef sel_x1
= LLVMBuildICmp(ctx
->ac
.builder
,
446 LLVMIntULE
, vertex_id
,
448 /* Use LLVMIntNE, because we have 3 vertices and only
449 * the middle one should use y2.
451 LLVMValueRef sel_y1
= LLVMBuildICmp(ctx
->ac
.builder
,
452 LLVMIntNE
, vertex_id
,
455 if (input_index
== 0) {
457 LLVMValueRef x1y1
= LLVMGetParam(ctx
->main_fn
,
458 ctx
->param_vs_blit_inputs
);
459 LLVMValueRef x2y2
= LLVMGetParam(ctx
->main_fn
,
460 ctx
->param_vs_blit_inputs
+ 1);
462 LLVMValueRef x1
= unpack_sint16(ctx
, x1y1
, 0);
463 LLVMValueRef y1
= unpack_sint16(ctx
, x1y1
, 1);
464 LLVMValueRef x2
= unpack_sint16(ctx
, x2y2
, 0);
465 LLVMValueRef y2
= unpack_sint16(ctx
, x2y2
, 1);
467 LLVMValueRef x
= LLVMBuildSelect(ctx
->ac
.builder
, sel_x1
,
469 LLVMValueRef y
= LLVMBuildSelect(ctx
->ac
.builder
, sel_y1
,
472 out
[0] = LLVMBuildSIToFP(ctx
->ac
.builder
, x
, ctx
->f32
, "");
473 out
[1] = LLVMBuildSIToFP(ctx
->ac
.builder
, y
, ctx
->f32
, "");
474 out
[2] = LLVMGetParam(ctx
->main_fn
,
475 ctx
->param_vs_blit_inputs
+ 2);
476 out
[3] = ctx
->ac
.f32_1
;
480 /* Color or texture coordinates: */
481 assert(input_index
== 1);
483 if (vs_blit_property
== SI_VS_BLIT_SGPRS_POS_COLOR
) {
484 for (int i
= 0; i
< 4; i
++) {
485 out
[i
] = LLVMGetParam(ctx
->main_fn
,
486 ctx
->param_vs_blit_inputs
+ 3 + i
);
489 assert(vs_blit_property
== SI_VS_BLIT_SGPRS_POS_TEXCOORD
);
490 LLVMValueRef x1
= LLVMGetParam(ctx
->main_fn
,
491 ctx
->param_vs_blit_inputs
+ 3);
492 LLVMValueRef y1
= LLVMGetParam(ctx
->main_fn
,
493 ctx
->param_vs_blit_inputs
+ 4);
494 LLVMValueRef x2
= LLVMGetParam(ctx
->main_fn
,
495 ctx
->param_vs_blit_inputs
+ 5);
496 LLVMValueRef y2
= LLVMGetParam(ctx
->main_fn
,
497 ctx
->param_vs_blit_inputs
+ 6);
499 out
[0] = LLVMBuildSelect(ctx
->ac
.builder
, sel_x1
,
501 out
[1] = LLVMBuildSelect(ctx
->ac
.builder
, sel_y1
,
503 out
[2] = LLVMGetParam(ctx
->main_fn
,
504 ctx
->param_vs_blit_inputs
+ 7);
505 out
[3] = LLVMGetParam(ctx
->main_fn
,
506 ctx
->param_vs_blit_inputs
+ 8);
511 union si_vs_fix_fetch fix_fetch
;
512 LLVMValueRef t_list_ptr
;
513 LLVMValueRef t_offset
;
515 LLVMValueRef vertex_index
;
518 /* Load the T list */
519 t_list_ptr
= LLVMGetParam(ctx
->main_fn
, ctx
->param_vertex_buffers
);
521 t_offset
= LLVMConstInt(ctx
->i32
, input_index
, 0);
523 t_list
= ac_build_load_to_sgpr(&ctx
->ac
, t_list_ptr
, t_offset
);
525 vertex_index
= LLVMGetParam(ctx
->main_fn
,
526 ctx
->param_vertex_index0
+
529 /* Use the open-coded implementation for all loads of doubles and
530 * of dword-sized data that needs fixups. We need to insert conversion
531 * code anyway, and the amd/common code does it for us.
533 * Note: On LLVM <= 8, we can only open-code formats with
534 * channel size >= 4 bytes.
536 bool opencode
= ctx
->shader
->key
.mono
.vs_fetch_opencode
& (1 << input_index
);
537 fix_fetch
.bits
= ctx
->shader
->key
.mono
.vs_fix_fetch
[input_index
].bits
;
539 (fix_fetch
.u
.log_size
== 3 && fix_fetch
.u
.format
== AC_FETCH_FORMAT_FLOAT
) ||
540 (fix_fetch
.u
.log_size
== 2)) {
541 tmp
= ac_build_opencoded_load_format(
542 &ctx
->ac
, fix_fetch
.u
.log_size
, fix_fetch
.u
.num_channels_m1
+ 1,
543 fix_fetch
.u
.format
, fix_fetch
.u
.reverse
, !opencode
,
544 t_list
, vertex_index
, ctx
->ac
.i32_0
, ctx
->ac
.i32_0
,
546 for (unsigned i
= 0; i
< 4; ++i
)
547 out
[i
] = LLVMBuildExtractElement(ctx
->ac
.builder
, tmp
, LLVMConstInt(ctx
->i32
, i
, false), "");
551 /* Do multiple loads for special formats. */
552 unsigned required_channels
= util_last_bit(info
->input_usage_mask
[input_index
]);
553 LLVMValueRef fetches
[4];
554 unsigned num_fetches
;
555 unsigned fetch_stride
;
556 unsigned channels_per_fetch
;
558 if (fix_fetch
.u
.log_size
<= 1 && fix_fetch
.u
.num_channels_m1
== 2) {
559 num_fetches
= MIN2(required_channels
, 3);
560 fetch_stride
= 1 << fix_fetch
.u
.log_size
;
561 channels_per_fetch
= 1;
565 channels_per_fetch
= required_channels
;
568 for (unsigned i
= 0; i
< num_fetches
; ++i
) {
569 LLVMValueRef voffset
= LLVMConstInt(ctx
->i32
, fetch_stride
* i
, 0);
570 fetches
[i
] = ac_build_buffer_load_format(&ctx
->ac
, t_list
, vertex_index
, voffset
,
571 channels_per_fetch
, false, true);
574 if (num_fetches
== 1 && channels_per_fetch
> 1) {
575 LLVMValueRef fetch
= fetches
[0];
576 for (unsigned i
= 0; i
< channels_per_fetch
; ++i
) {
577 tmp
= LLVMConstInt(ctx
->i32
, i
, false);
578 fetches
[i
] = LLVMBuildExtractElement(
579 ctx
->ac
.builder
, fetch
, tmp
, "");
581 num_fetches
= channels_per_fetch
;
582 channels_per_fetch
= 1;
585 for (unsigned i
= num_fetches
; i
< 4; ++i
)
586 fetches
[i
] = LLVMGetUndef(ctx
->f32
);
588 if (fix_fetch
.u
.log_size
<= 1 && fix_fetch
.u
.num_channels_m1
== 2 &&
589 required_channels
== 4) {
590 if (fix_fetch
.u
.format
== AC_FETCH_FORMAT_UINT
|| fix_fetch
.u
.format
== AC_FETCH_FORMAT_SINT
)
591 fetches
[3] = ctx
->ac
.i32_1
;
593 fetches
[3] = ctx
->ac
.f32_1
;
594 } else if (fix_fetch
.u
.log_size
== 3 &&
595 (fix_fetch
.u
.format
== AC_FETCH_FORMAT_SNORM
||
596 fix_fetch
.u
.format
== AC_FETCH_FORMAT_SSCALED
||
597 fix_fetch
.u
.format
== AC_FETCH_FORMAT_SINT
) &&
598 required_channels
== 4) {
599 /* For 2_10_10_10, the hardware returns an unsigned value;
600 * convert it to a signed one.
602 LLVMValueRef tmp
= fetches
[3];
603 LLVMValueRef c30
= LLVMConstInt(ctx
->i32
, 30, 0);
605 /* First, recover the sign-extended signed integer value. */
606 if (fix_fetch
.u
.format
== AC_FETCH_FORMAT_SSCALED
)
607 tmp
= LLVMBuildFPToUI(ctx
->ac
.builder
, tmp
, ctx
->i32
, "");
609 tmp
= ac_to_integer(&ctx
->ac
, tmp
);
611 /* For the integer-like cases, do a natural sign extension.
613 * For the SNORM case, the values are 0.0, 0.333, 0.666, 1.0
614 * and happen to contain 0, 1, 2, 3 as the two LSBs of the
617 tmp
= LLVMBuildShl(ctx
->ac
.builder
, tmp
,
618 fix_fetch
.u
.format
== AC_FETCH_FORMAT_SNORM
?
619 LLVMConstInt(ctx
->i32
, 7, 0) : c30
, "");
620 tmp
= LLVMBuildAShr(ctx
->ac
.builder
, tmp
, c30
, "");
622 /* Convert back to the right type. */
623 if (fix_fetch
.u
.format
== AC_FETCH_FORMAT_SNORM
) {
625 LLVMValueRef neg_one
= LLVMConstReal(ctx
->f32
, -1.0);
626 tmp
= LLVMBuildSIToFP(ctx
->ac
.builder
, tmp
, ctx
->f32
, "");
627 clamp
= LLVMBuildFCmp(ctx
->ac
.builder
, LLVMRealULT
, tmp
, neg_one
, "");
628 tmp
= LLVMBuildSelect(ctx
->ac
.builder
, clamp
, neg_one
, tmp
, "");
629 } else if (fix_fetch
.u
.format
== AC_FETCH_FORMAT_SSCALED
) {
630 tmp
= LLVMBuildSIToFP(ctx
->ac
.builder
, tmp
, ctx
->f32
, "");
636 for (unsigned i
= 0; i
< 4; ++i
)
637 out
[i
] = ac_to_float(&ctx
->ac
, fetches
[i
]);
640 static void declare_input_vs(
641 struct si_shader_context
*ctx
,
642 unsigned input_index
,
643 const struct tgsi_full_declaration
*decl
,
646 si_llvm_load_input_vs(ctx
, input_index
, out
);
649 LLVMValueRef
si_get_primitive_id(struct si_shader_context
*ctx
,
656 case PIPE_SHADER_VERTEX
:
657 return LLVMGetParam(ctx
->main_fn
,
658 ctx
->param_vs_prim_id
);
659 case PIPE_SHADER_TESS_CTRL
:
660 return ctx
->abi
.tcs_patch_id
;
661 case PIPE_SHADER_TESS_EVAL
:
662 return ctx
->abi
.tes_patch_id
;
663 case PIPE_SHADER_GEOMETRY
:
664 return ctx
->abi
.gs_prim_id
;
672 * Return the value of tgsi_ind_register for indexing.
673 * This is the indirect index with the constant offset added to it.
675 LLVMValueRef
si_get_indirect_index(struct si_shader_context
*ctx
,
676 const struct tgsi_ind_register
*ind
,
682 if (ind
->File
== TGSI_FILE_ADDRESS
) {
683 result
= ctx
->addrs
[ind
->Index
][ind
->Swizzle
];
684 result
= LLVMBuildLoad(ctx
->ac
.builder
, result
, "");
686 struct tgsi_full_src_register src
= {};
688 src
.Register
.File
= ind
->File
;
689 src
.Register
.Index
= ind
->Index
;
691 /* Set the second index to 0 for constants. */
692 if (ind
->File
== TGSI_FILE_CONSTANT
)
693 src
.Register
.Dimension
= 1;
695 result
= ctx
->bld_base
.emit_fetch_funcs
[ind
->File
](&ctx
->bld_base
, &src
,
698 result
= ac_to_integer(&ctx
->ac
, result
);
701 return ac_build_imad(&ctx
->ac
, result
, LLVMConstInt(ctx
->i32
, addr_mul
, 0),
702 LLVMConstInt(ctx
->i32
, rel_index
, 0));
706 * Like si_get_indirect_index, but restricts the return value to a (possibly
707 * undefined) value inside [0..num).
709 LLVMValueRef
si_get_bounded_indirect_index(struct si_shader_context
*ctx
,
710 const struct tgsi_ind_register
*ind
,
711 int rel_index
, unsigned num
)
713 LLVMValueRef result
= si_get_indirect_index(ctx
, ind
, 1, rel_index
);
715 return si_llvm_bound_index(ctx
, result
, num
);
718 static LLVMValueRef
get_dw_address_from_generic_indices(struct si_shader_context
*ctx
,
719 LLVMValueRef vertex_dw_stride
,
720 LLVMValueRef base_addr
,
721 LLVMValueRef vertex_index
,
722 LLVMValueRef param_index
,
723 unsigned input_index
,
728 if (vertex_dw_stride
) {
729 base_addr
= ac_build_imad(&ctx
->ac
, vertex_index
,
730 vertex_dw_stride
, base_addr
);
734 base_addr
= ac_build_imad(&ctx
->ac
, param_index
,
735 LLVMConstInt(ctx
->i32
, 4, 0), base_addr
);
738 int param
= is_patch
?
739 si_shader_io_get_unique_index_patch(name
[input_index
],
740 index
[input_index
]) :
741 si_shader_io_get_unique_index(name
[input_index
],
742 index
[input_index
], false);
744 /* Add the base address of the element. */
745 return LLVMBuildAdd(ctx
->ac
.builder
, base_addr
,
746 LLVMConstInt(ctx
->i32
, param
* 4, 0), "");
750 * Calculate a dword address given an input or output register and a stride.
752 static LLVMValueRef
get_dw_address(struct si_shader_context
*ctx
,
753 const struct tgsi_full_dst_register
*dst
,
754 const struct tgsi_full_src_register
*src
,
755 LLVMValueRef vertex_dw_stride
,
756 LLVMValueRef base_addr
)
758 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
759 ubyte
*name
, *index
, *array_first
;
761 struct tgsi_full_dst_register reg
;
762 LLVMValueRef vertex_index
= NULL
;
763 LLVMValueRef ind_index
= NULL
;
765 /* Set the register description. The address computation is the same
766 * for sources and destinations. */
768 reg
.Register
.File
= src
->Register
.File
;
769 reg
.Register
.Index
= src
->Register
.Index
;
770 reg
.Register
.Indirect
= src
->Register
.Indirect
;
771 reg
.Register
.Dimension
= src
->Register
.Dimension
;
772 reg
.Indirect
= src
->Indirect
;
773 reg
.Dimension
= src
->Dimension
;
774 reg
.DimIndirect
= src
->DimIndirect
;
778 /* If the register is 2-dimensional (e.g. an array of vertices
779 * in a primitive), calculate the base address of the vertex. */
780 if (reg
.Register
.Dimension
) {
781 if (reg
.Dimension
.Indirect
)
782 vertex_index
= si_get_indirect_index(ctx
, ®
.DimIndirect
,
783 1, reg
.Dimension
.Index
);
785 vertex_index
= LLVMConstInt(ctx
->i32
, reg
.Dimension
.Index
, 0);
788 /* Get information about the register. */
789 if (reg
.Register
.File
== TGSI_FILE_INPUT
) {
790 name
= info
->input_semantic_name
;
791 index
= info
->input_semantic_index
;
792 array_first
= info
->input_array_first
;
793 } else if (reg
.Register
.File
== TGSI_FILE_OUTPUT
) {
794 name
= info
->output_semantic_name
;
795 index
= info
->output_semantic_index
;
796 array_first
= info
->output_array_first
;
802 if (reg
.Register
.Indirect
) {
803 /* Add the relative address of the element. */
804 if (reg
.Indirect
.ArrayID
)
805 input_index
= array_first
[reg
.Indirect
.ArrayID
];
807 input_index
= reg
.Register
.Index
;
809 ind_index
= si_get_indirect_index(ctx
, ®
.Indirect
,
810 1, reg
.Register
.Index
- input_index
);
812 input_index
= reg
.Register
.Index
;
815 return get_dw_address_from_generic_indices(ctx
, vertex_dw_stride
,
816 base_addr
, vertex_index
,
817 ind_index
, input_index
,
819 !reg
.Register
.Dimension
);
822 /* The offchip buffer layout for TCS->TES is
824 * - attribute 0 of patch 0 vertex 0
825 * - attribute 0 of patch 0 vertex 1
826 * - attribute 0 of patch 0 vertex 2
828 * - attribute 0 of patch 1 vertex 0
829 * - attribute 0 of patch 1 vertex 1
831 * - attribute 1 of patch 0 vertex 0
832 * - attribute 1 of patch 0 vertex 1
834 * - per patch attribute 0 of patch 0
835 * - per patch attribute 0 of patch 1
838 * Note that every attribute has 4 components.
840 static LLVMValueRef
get_tcs_tes_buffer_address(struct si_shader_context
*ctx
,
841 LLVMValueRef rel_patch_id
,
842 LLVMValueRef vertex_index
,
843 LLVMValueRef param_index
)
845 LLVMValueRef base_addr
, vertices_per_patch
, num_patches
, total_vertices
;
846 LLVMValueRef param_stride
, constant16
;
848 vertices_per_patch
= get_num_tcs_out_vertices(ctx
);
849 num_patches
= si_unpack_param(ctx
, ctx
->param_tcs_offchip_layout
, 0, 6);
850 total_vertices
= LLVMBuildMul(ctx
->ac
.builder
, vertices_per_patch
,
853 constant16
= LLVMConstInt(ctx
->i32
, 16, 0);
855 base_addr
= ac_build_imad(&ctx
->ac
, rel_patch_id
,
856 vertices_per_patch
, vertex_index
);
857 param_stride
= total_vertices
;
859 base_addr
= rel_patch_id
;
860 param_stride
= num_patches
;
863 base_addr
= ac_build_imad(&ctx
->ac
, param_index
, param_stride
, base_addr
);
864 base_addr
= LLVMBuildMul(ctx
->ac
.builder
, base_addr
, constant16
, "");
867 LLVMValueRef patch_data_offset
=
868 si_unpack_param(ctx
, ctx
->param_tcs_offchip_layout
, 12, 20);
870 base_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_addr
,
871 patch_data_offset
, "");
876 /* This is a generic helper that can be shared by the NIR and TGSI backends */
877 static LLVMValueRef
get_tcs_tes_buffer_address_from_generic_indices(
878 struct si_shader_context
*ctx
,
879 LLVMValueRef vertex_index
,
880 LLVMValueRef param_index
,
886 unsigned param_index_base
;
888 param_index_base
= is_patch
?
889 si_shader_io_get_unique_index_patch(name
[param_base
], index
[param_base
]) :
890 si_shader_io_get_unique_index(name
[param_base
], index
[param_base
], false);
893 param_index
= LLVMBuildAdd(ctx
->ac
.builder
, param_index
,
894 LLVMConstInt(ctx
->i32
, param_index_base
, 0),
897 param_index
= LLVMConstInt(ctx
->i32
, param_index_base
, 0);
900 return get_tcs_tes_buffer_address(ctx
, get_rel_patch_id(ctx
),
901 vertex_index
, param_index
);
904 static LLVMValueRef
get_tcs_tes_buffer_address_from_reg(
905 struct si_shader_context
*ctx
,
906 const struct tgsi_full_dst_register
*dst
,
907 const struct tgsi_full_src_register
*src
)
909 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
910 ubyte
*name
, *index
, *array_first
;
911 struct tgsi_full_src_register reg
;
912 LLVMValueRef vertex_index
= NULL
;
913 LLVMValueRef param_index
= NULL
;
916 reg
= src
? *src
: tgsi_full_src_register_from_dst(dst
);
918 if (reg
.Register
.Dimension
) {
920 if (reg
.Dimension
.Indirect
)
921 vertex_index
= si_get_indirect_index(ctx
, ®
.DimIndirect
,
922 1, reg
.Dimension
.Index
);
924 vertex_index
= LLVMConstInt(ctx
->i32
, reg
.Dimension
.Index
, 0);
927 /* Get information about the register. */
928 if (reg
.Register
.File
== TGSI_FILE_INPUT
) {
929 name
= info
->input_semantic_name
;
930 index
= info
->input_semantic_index
;
931 array_first
= info
->input_array_first
;
932 } else if (reg
.Register
.File
== TGSI_FILE_OUTPUT
) {
933 name
= info
->output_semantic_name
;
934 index
= info
->output_semantic_index
;
935 array_first
= info
->output_array_first
;
941 if (reg
.Register
.Indirect
) {
942 if (reg
.Indirect
.ArrayID
)
943 param_base
= array_first
[reg
.Indirect
.ArrayID
];
945 param_base
= reg
.Register
.Index
;
947 param_index
= si_get_indirect_index(ctx
, ®
.Indirect
,
948 1, reg
.Register
.Index
- param_base
);
951 param_base
= reg
.Register
.Index
;
954 return get_tcs_tes_buffer_address_from_generic_indices(ctx
, vertex_index
,
955 param_index
, param_base
,
956 name
, index
, !reg
.Register
.Dimension
);
959 static LLVMValueRef
buffer_load(struct lp_build_tgsi_context
*bld_base
,
960 LLVMTypeRef type
, unsigned swizzle
,
961 LLVMValueRef buffer
, LLVMValueRef offset
,
962 LLVMValueRef base
, bool can_speculate
)
964 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
965 LLVMValueRef value
, value2
;
966 LLVMTypeRef vec_type
= LLVMVectorType(type
, 4);
969 value
= ac_build_buffer_load(&ctx
->ac
, buffer
, 4, NULL
, base
, offset
,
970 0, 1, 0, can_speculate
, false);
972 return LLVMBuildBitCast(ctx
->ac
.builder
, value
, vec_type
, "");
975 if (!llvm_type_is_64bit(ctx
, type
)) {
976 value
= ac_build_buffer_load(&ctx
->ac
, buffer
, 4, NULL
, base
, offset
,
977 0, 1, 0, can_speculate
, false);
979 value
= LLVMBuildBitCast(ctx
->ac
.builder
, value
, vec_type
, "");
980 return LLVMBuildExtractElement(ctx
->ac
.builder
, value
,
981 LLVMConstInt(ctx
->i32
, swizzle
, 0), "");
984 value
= ac_build_buffer_load(&ctx
->ac
, buffer
, 1, NULL
, base
, offset
,
985 swizzle
* 4, 1, 0, can_speculate
, false);
987 value2
= ac_build_buffer_load(&ctx
->ac
, buffer
, 1, NULL
, base
, offset
,
988 swizzle
* 4 + 4, 1, 0, can_speculate
, false);
990 return si_llvm_emit_fetch_64bit(bld_base
, type
, value
, value2
);
994 * Load from LSHS LDS storage.
996 * \param type output value type
997 * \param swizzle offset (typically 0..3); it can be ~0, which loads a vec4
998 * \param dw_addr address in dwords
1000 static LLVMValueRef
lshs_lds_load(struct lp_build_tgsi_context
*bld_base
,
1001 LLVMTypeRef type
, unsigned swizzle
,
1002 LLVMValueRef dw_addr
)
1004 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1007 if (swizzle
== ~0) {
1008 LLVMValueRef values
[TGSI_NUM_CHANNELS
];
1010 for (unsigned chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++)
1011 values
[chan
] = lshs_lds_load(bld_base
, type
, chan
, dw_addr
);
1013 return ac_build_gather_values(&ctx
->ac
, values
,
1017 /* Split 64-bit loads. */
1018 if (llvm_type_is_64bit(ctx
, type
)) {
1019 LLVMValueRef lo
, hi
;
1021 lo
= lshs_lds_load(bld_base
, ctx
->i32
, swizzle
, dw_addr
);
1022 hi
= lshs_lds_load(bld_base
, ctx
->i32
, swizzle
+ 1, dw_addr
);
1023 return si_llvm_emit_fetch_64bit(bld_base
, type
, lo
, hi
);
1026 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1027 LLVMConstInt(ctx
->i32
, swizzle
, 0), "");
1029 value
= ac_lds_load(&ctx
->ac
, dw_addr
);
1031 return LLVMBuildBitCast(ctx
->ac
.builder
, value
, type
, "");
1035 * Store to LSHS LDS storage.
1037 * \param swizzle offset (typically 0..3)
1038 * \param dw_addr address in dwords
1039 * \param value value to store
1041 static void lshs_lds_store(struct si_shader_context
*ctx
,
1042 unsigned dw_offset_imm
, LLVMValueRef dw_addr
,
1045 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
1046 LLVMConstInt(ctx
->i32
, dw_offset_imm
, 0), "");
1048 ac_lds_store(&ctx
->ac
, dw_addr
, value
);
1053 TESS_OFFCHIP_RING_TCS
,
1054 TESS_OFFCHIP_RING_TES
,
1057 static LLVMValueRef
get_tess_ring_descriptor(struct si_shader_context
*ctx
,
1058 enum si_tess_ring ring
)
1060 LLVMBuilderRef builder
= ctx
->ac
.builder
;
1061 unsigned param
= ring
== TESS_OFFCHIP_RING_TES
? ctx
->param_tes_offchip_addr
:
1062 ctx
->param_tcs_out_lds_layout
;
1063 LLVMValueRef addr
= LLVMGetParam(ctx
->main_fn
, param
);
1065 /* TCS only receives high 13 bits of the address. */
1066 if (ring
== TESS_OFFCHIP_RING_TCS
|| ring
== TCS_FACTOR_RING
) {
1067 addr
= LLVMBuildAnd(builder
, addr
,
1068 LLVMConstInt(ctx
->i32
, 0xfff80000, 0), "");
1071 if (ring
== TCS_FACTOR_RING
) {
1072 unsigned tf_offset
= ctx
->screen
->tess_offchip_ring_size
;
1073 addr
= LLVMBuildAdd(builder
, addr
,
1074 LLVMConstInt(ctx
->i32
, tf_offset
, 0), "");
1077 LLVMValueRef desc
[4];
1079 desc
[1] = LLVMConstInt(ctx
->i32
,
1080 S_008F04_BASE_ADDRESS_HI(ctx
->screen
->info
.address32_hi
), 0);
1081 desc
[2] = LLVMConstInt(ctx
->i32
, 0xffffffff, 0);
1082 desc
[3] = LLVMConstInt(ctx
->i32
,
1083 S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1084 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1085 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1086 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1087 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1088 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
), 0);
1090 return ac_build_gather_values(&ctx
->ac
, desc
, 4);
1093 static LLVMValueRef
fetch_input_tcs(
1094 struct lp_build_tgsi_context
*bld_base
,
1095 const struct tgsi_full_src_register
*reg
,
1096 enum tgsi_opcode_type type
, unsigned swizzle_in
)
1098 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1099 LLVMValueRef dw_addr
, stride
;
1100 unsigned swizzle
= swizzle_in
& 0xffff;
1101 stride
= get_tcs_in_vertex_dw_stride(ctx
);
1102 dw_addr
= get_tcs_in_current_patch_offset(ctx
);
1103 dw_addr
= get_dw_address(ctx
, NULL
, reg
, stride
, dw_addr
);
1105 return lshs_lds_load(bld_base
, tgsi2llvmtype(bld_base
, type
), swizzle
, dw_addr
);
1108 static LLVMValueRef
si_nir_load_tcs_varyings(struct ac_shader_abi
*abi
,
1110 LLVMValueRef vertex_index
,
1111 LLVMValueRef param_index
,
1112 unsigned const_index
,
1114 unsigned driver_location
,
1116 unsigned num_components
,
1121 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1122 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
1123 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
1124 LLVMValueRef dw_addr
, stride
;
1126 driver_location
= driver_location
/ 4;
1129 stride
= get_tcs_in_vertex_dw_stride(ctx
);
1130 dw_addr
= get_tcs_in_current_patch_offset(ctx
);
1134 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
1136 stride
= get_tcs_out_vertex_dw_stride(ctx
);
1137 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
1142 /* Add the constant index to the indirect index */
1143 param_index
= LLVMBuildAdd(ctx
->ac
.builder
, param_index
,
1144 LLVMConstInt(ctx
->i32
, const_index
, 0), "");
1146 param_index
= LLVMConstInt(ctx
->i32
, const_index
, 0);
1152 names
= info
->input_semantic_name
;
1153 indices
= info
->input_semantic_index
;
1155 names
= info
->output_semantic_name
;
1156 indices
= info
->output_semantic_index
;
1159 dw_addr
= get_dw_address_from_generic_indices(ctx
, stride
, dw_addr
,
1160 vertex_index
, param_index
,
1165 LLVMValueRef value
[4];
1166 for (unsigned i
= 0; i
< num_components
; i
++) {
1167 unsigned offset
= i
;
1168 if (llvm_type_is_64bit(ctx
, type
))
1171 offset
+= component
;
1172 value
[i
+ component
] = lshs_lds_load(bld_base
, type
, offset
, dw_addr
);
1175 return ac_build_varying_gather_values(&ctx
->ac
, value
, num_components
, component
);
1178 static LLVMValueRef
fetch_output_tcs(
1179 struct lp_build_tgsi_context
*bld_base
,
1180 const struct tgsi_full_src_register
*reg
,
1181 enum tgsi_opcode_type type
, unsigned swizzle_in
)
1183 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1184 LLVMValueRef dw_addr
, stride
;
1185 unsigned swizzle
= (swizzle_in
& 0xffff);
1187 if (reg
->Register
.Dimension
) {
1188 stride
= get_tcs_out_vertex_dw_stride(ctx
);
1189 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
1190 dw_addr
= get_dw_address(ctx
, NULL
, reg
, stride
, dw_addr
);
1192 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
1193 dw_addr
= get_dw_address(ctx
, NULL
, reg
, NULL
, dw_addr
);
1196 return lshs_lds_load(bld_base
, tgsi2llvmtype(bld_base
, type
), swizzle
, dw_addr
);
1199 static LLVMValueRef
fetch_input_tes(
1200 struct lp_build_tgsi_context
*bld_base
,
1201 const struct tgsi_full_src_register
*reg
,
1202 enum tgsi_opcode_type type
, unsigned swizzle_in
)
1204 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1205 LLVMValueRef base
, addr
;
1206 unsigned swizzle
= (swizzle_in
& 0xffff);
1208 base
= LLVMGetParam(ctx
->main_fn
, ctx
->param_tcs_offchip_offset
);
1209 addr
= get_tcs_tes_buffer_address_from_reg(ctx
, NULL
, reg
);
1211 return buffer_load(bld_base
, tgsi2llvmtype(bld_base
, type
), swizzle
,
1212 ctx
->tess_offchip_ring
, base
, addr
, true);
1215 LLVMValueRef
si_nir_load_input_tes(struct ac_shader_abi
*abi
,
1217 LLVMValueRef vertex_index
,
1218 LLVMValueRef param_index
,
1219 unsigned const_index
,
1221 unsigned driver_location
,
1223 unsigned num_components
,
1228 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1229 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
1230 LLVMValueRef base
, addr
;
1232 driver_location
= driver_location
/ 4;
1234 base
= LLVMGetParam(ctx
->main_fn
, ctx
->param_tcs_offchip_offset
);
1237 /* Add the constant index to the indirect index */
1238 param_index
= LLVMBuildAdd(ctx
->ac
.builder
, param_index
,
1239 LLVMConstInt(ctx
->i32
, const_index
, 0), "");
1241 param_index
= LLVMConstInt(ctx
->i32
, const_index
, 0);
1244 addr
= get_tcs_tes_buffer_address_from_generic_indices(ctx
, vertex_index
,
1245 param_index
, driver_location
,
1246 info
->input_semantic_name
,
1247 info
->input_semantic_index
,
1250 /* TODO: This will generate rather ordinary llvm code, although it
1251 * should be easy for the optimiser to fix up. In future we might want
1252 * to refactor buffer_load(), but for now this maximises code sharing
1253 * between the NIR and TGSI backends.
1255 LLVMValueRef value
[4];
1256 for (unsigned i
= 0; i
< num_components
; i
++) {
1257 unsigned offset
= i
;
1258 if (llvm_type_is_64bit(ctx
, type
)) {
1261 addr
= get_tcs_tes_buffer_address_from_generic_indices(ctx
,
1264 driver_location
+ 1,
1265 info
->input_semantic_name
,
1266 info
->input_semantic_index
,
1270 offset
= offset
% 4;
1273 offset
+= component
;
1274 value
[i
+ component
] = buffer_load(&ctx
->bld_base
, type
, offset
,
1275 ctx
->tess_offchip_ring
, base
, addr
, true);
1278 return ac_build_varying_gather_values(&ctx
->ac
, value
, num_components
, component
);
1281 static void store_output_tcs(struct lp_build_tgsi_context
*bld_base
,
1282 const struct tgsi_full_instruction
*inst
,
1283 const struct tgsi_opcode_info
*info
,
1285 LLVMValueRef dst
[4])
1287 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1288 const struct tgsi_full_dst_register
*reg
= &inst
->Dst
[index
];
1289 const struct tgsi_shader_info
*sh_info
= &ctx
->shader
->selector
->info
;
1290 unsigned chan_index
;
1291 LLVMValueRef dw_addr
, stride
;
1292 LLVMValueRef buffer
, base
, buf_addr
;
1293 LLVMValueRef values
[4];
1294 bool skip_lds_store
;
1295 bool is_tess_factor
= false, is_tess_inner
= false;
1297 /* Only handle per-patch and per-vertex outputs here.
1298 * Vectors will be lowered to scalars and this function will be called again.
1300 if (reg
->Register
.File
!= TGSI_FILE_OUTPUT
||
1301 (dst
[0] && LLVMGetTypeKind(LLVMTypeOf(dst
[0])) == LLVMVectorTypeKind
)) {
1302 si_llvm_emit_store(bld_base
, inst
, info
, index
, dst
);
1306 if (reg
->Register
.Dimension
) {
1307 stride
= get_tcs_out_vertex_dw_stride(ctx
);
1308 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
1309 dw_addr
= get_dw_address(ctx
, reg
, NULL
, stride
, dw_addr
);
1310 skip_lds_store
= !sh_info
->reads_pervertex_outputs
;
1312 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
1313 dw_addr
= get_dw_address(ctx
, reg
, NULL
, NULL
, dw_addr
);
1314 skip_lds_store
= !sh_info
->reads_perpatch_outputs
;
1316 if (!reg
->Register
.Indirect
) {
1317 int name
= sh_info
->output_semantic_name
[reg
->Register
.Index
];
1319 /* Always write tess factors into LDS for the TCS epilog. */
1320 if (name
== TGSI_SEMANTIC_TESSINNER
||
1321 name
== TGSI_SEMANTIC_TESSOUTER
) {
1322 /* The epilog doesn't read LDS if invocation 0 defines tess factors. */
1323 skip_lds_store
= !sh_info
->reads_tessfactor_outputs
&&
1324 ctx
->shader
->selector
->tcs_info
.tessfactors_are_def_in_all_invocs
;
1325 is_tess_factor
= true;
1326 is_tess_inner
= name
== TGSI_SEMANTIC_TESSINNER
;
1331 buffer
= get_tess_ring_descriptor(ctx
, TESS_OFFCHIP_RING_TCS
);
1333 base
= LLVMGetParam(ctx
->main_fn
, ctx
->param_tcs_offchip_offset
);
1334 buf_addr
= get_tcs_tes_buffer_address_from_reg(ctx
, reg
, NULL
);
1336 uint32_t writemask
= reg
->Register
.WriteMask
;
1338 chan_index
= u_bit_scan(&writemask
);
1339 LLVMValueRef value
= dst
[chan_index
];
1341 if (inst
->Instruction
.Saturate
)
1342 value
= ac_build_clamp(&ctx
->ac
, value
);
1344 /* Skip LDS stores if there is no LDS read of this output. */
1345 if (!skip_lds_store
)
1346 lshs_lds_store(ctx
, chan_index
, dw_addr
, value
);
1348 value
= ac_to_integer(&ctx
->ac
, value
);
1349 values
[chan_index
] = value
;
1351 if (reg
->Register
.WriteMask
!= 0xF && !is_tess_factor
) {
1352 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, value
, 1,
1354 4 * chan_index
, 1, 0, false);
1357 /* Write tess factors into VGPRs for the epilog. */
1358 if (is_tess_factor
&&
1359 ctx
->shader
->selector
->tcs_info
.tessfactors_are_def_in_all_invocs
) {
1360 if (!is_tess_inner
) {
1361 LLVMBuildStore(ctx
->ac
.builder
, value
, /* outer */
1362 ctx
->invoc0_tess_factors
[chan_index
]);
1363 } else if (chan_index
< 2) {
1364 LLVMBuildStore(ctx
->ac
.builder
, value
, /* inner */
1365 ctx
->invoc0_tess_factors
[4 + chan_index
]);
1370 if (reg
->Register
.WriteMask
== 0xF && !is_tess_factor
) {
1371 LLVMValueRef value
= ac_build_gather_values(&ctx
->ac
,
1373 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, value
, 4, buf_addr
,
1374 base
, 0, 1, 0, false);
1378 static void si_nir_store_output_tcs(struct ac_shader_abi
*abi
,
1379 const struct nir_variable
*var
,
1380 LLVMValueRef vertex_index
,
1381 LLVMValueRef param_index
,
1382 unsigned const_index
,
1386 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1387 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
1388 const unsigned component
= var
->data
.location_frac
;
1389 const bool is_patch
= var
->data
.patch
;
1390 unsigned driver_location
= var
->data
.driver_location
;
1391 LLVMValueRef dw_addr
, stride
;
1392 LLVMValueRef buffer
, base
, addr
;
1393 LLVMValueRef values
[8];
1394 bool skip_lds_store
;
1395 bool is_tess_factor
= false, is_tess_inner
= false;
1397 driver_location
= driver_location
/ 4;
1400 /* Add the constant index to the indirect index */
1401 param_index
= LLVMBuildAdd(ctx
->ac
.builder
, param_index
,
1402 LLVMConstInt(ctx
->i32
, const_index
, 0), "");
1404 if (const_index
!= 0)
1405 param_index
= LLVMConstInt(ctx
->i32
, const_index
, 0);
1409 stride
= get_tcs_out_vertex_dw_stride(ctx
);
1410 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
1411 dw_addr
= get_dw_address_from_generic_indices(ctx
, stride
, dw_addr
,
1412 vertex_index
, param_index
,
1414 info
->output_semantic_name
,
1415 info
->output_semantic_index
,
1418 skip_lds_store
= !info
->reads_pervertex_outputs
;
1420 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
1421 dw_addr
= get_dw_address_from_generic_indices(ctx
, NULL
, dw_addr
,
1422 vertex_index
, param_index
,
1424 info
->output_semantic_name
,
1425 info
->output_semantic_index
,
1428 skip_lds_store
= !info
->reads_perpatch_outputs
;
1431 int name
= info
->output_semantic_name
[driver_location
];
1433 /* Always write tess factors into LDS for the TCS epilog. */
1434 if (name
== TGSI_SEMANTIC_TESSINNER
||
1435 name
== TGSI_SEMANTIC_TESSOUTER
) {
1436 /* The epilog doesn't read LDS if invocation 0 defines tess factors. */
1437 skip_lds_store
= !info
->reads_tessfactor_outputs
&&
1438 ctx
->shader
->selector
->tcs_info
.tessfactors_are_def_in_all_invocs
;
1439 is_tess_factor
= true;
1440 is_tess_inner
= name
== TGSI_SEMANTIC_TESSINNER
;
1445 buffer
= get_tess_ring_descriptor(ctx
, TESS_OFFCHIP_RING_TCS
);
1447 base
= LLVMGetParam(ctx
->main_fn
, ctx
->param_tcs_offchip_offset
);
1449 addr
= get_tcs_tes_buffer_address_from_generic_indices(ctx
, vertex_index
,
1450 param_index
, driver_location
,
1451 info
->output_semantic_name
,
1452 info
->output_semantic_index
,
1455 for (unsigned chan
= 0; chan
< 8; chan
++) {
1456 if (!(writemask
& (1 << chan
)))
1458 LLVMValueRef value
= ac_llvm_extract_elem(&ctx
->ac
, src
, chan
- component
);
1460 unsigned buffer_store_offset
= chan
% 4;
1462 addr
= get_tcs_tes_buffer_address_from_generic_indices(ctx
,
1465 driver_location
+ 1,
1466 info
->output_semantic_name
,
1467 info
->output_semantic_index
,
1471 /* Skip LDS stores if there is no LDS read of this output. */
1472 if (!skip_lds_store
)
1473 lshs_lds_store(ctx
, chan
, dw_addr
, value
);
1475 value
= ac_to_integer(&ctx
->ac
, value
);
1476 values
[chan
] = value
;
1478 if (writemask
!= 0xF && !is_tess_factor
) {
1479 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, value
, 1,
1481 4 * buffer_store_offset
,
1485 /* Write tess factors into VGPRs for the epilog. */
1486 if (is_tess_factor
&&
1487 ctx
->shader
->selector
->tcs_info
.tessfactors_are_def_in_all_invocs
) {
1488 if (!is_tess_inner
) {
1489 LLVMBuildStore(ctx
->ac
.builder
, value
, /* outer */
1490 ctx
->invoc0_tess_factors
[chan
]);
1491 } else if (chan
< 2) {
1492 LLVMBuildStore(ctx
->ac
.builder
, value
, /* inner */
1493 ctx
->invoc0_tess_factors
[4 + chan
]);
1498 if (writemask
== 0xF && !is_tess_factor
) {
1499 LLVMValueRef value
= ac_build_gather_values(&ctx
->ac
,
1501 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, value
, 4, addr
,
1502 base
, 0, 1, 0, false);
1506 LLVMValueRef
si_llvm_load_input_gs(struct ac_shader_abi
*abi
,
1507 unsigned input_index
,
1508 unsigned vtx_offset_param
,
1512 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1513 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
1514 struct si_shader
*shader
= ctx
->shader
;
1515 LLVMValueRef vtx_offset
, soffset
;
1516 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
1517 unsigned semantic_name
= info
->input_semantic_name
[input_index
];
1518 unsigned semantic_index
= info
->input_semantic_index
[input_index
];
1522 param
= si_shader_io_get_unique_index(semantic_name
, semantic_index
, false);
1524 /* GFX9 has the ESGS ring in LDS. */
1525 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
1526 unsigned index
= vtx_offset_param
;
1528 switch (index
/ 2) {
1530 vtx_offset
= si_unpack_param(ctx
, ctx
->param_gs_vtx01_offset
,
1531 index
% 2 ? 16 : 0, 16);
1534 vtx_offset
= si_unpack_param(ctx
, ctx
->param_gs_vtx23_offset
,
1535 index
% 2 ? 16 : 0, 16);
1538 vtx_offset
= si_unpack_param(ctx
, ctx
->param_gs_vtx45_offset
,
1539 index
% 2 ? 16 : 0, 16);
1546 unsigned offset
= param
* 4 + swizzle
;
1547 vtx_offset
= LLVMBuildAdd(ctx
->ac
.builder
, vtx_offset
,
1548 LLVMConstInt(ctx
->i32
, offset
, false), "");
1550 LLVMValueRef ptr
= ac_build_gep0(&ctx
->ac
, ctx
->esgs_ring
, vtx_offset
);
1551 LLVMValueRef value
= LLVMBuildLoad(ctx
->ac
.builder
, ptr
, "");
1552 if (llvm_type_is_64bit(ctx
, type
)) {
1553 ptr
= LLVMBuildGEP(ctx
->ac
.builder
, ptr
,
1554 &ctx
->ac
.i32_1
, 1, "");
1555 LLVMValueRef values
[2] = {
1557 LLVMBuildLoad(ctx
->ac
.builder
, ptr
, "")
1559 value
= ac_build_gather_values(&ctx
->ac
, values
, 2);
1561 return LLVMBuildBitCast(ctx
->ac
.builder
, value
, type
, "");
1564 /* GFX6: input load from the ESGS ring in memory. */
1565 if (swizzle
== ~0) {
1566 LLVMValueRef values
[TGSI_NUM_CHANNELS
];
1568 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
1569 values
[chan
] = si_llvm_load_input_gs(abi
, input_index
, vtx_offset_param
,
1572 return ac_build_gather_values(&ctx
->ac
, values
,
1576 /* Get the vertex offset parameter on GFX6. */
1577 LLVMValueRef gs_vtx_offset
= ctx
->gs_vtx_offset
[vtx_offset_param
];
1579 vtx_offset
= LLVMBuildMul(ctx
->ac
.builder
, gs_vtx_offset
,
1580 LLVMConstInt(ctx
->i32
, 4, 0), "");
1582 soffset
= LLVMConstInt(ctx
->i32
, (param
* 4 + swizzle
) * 256, 0);
1584 value
= ac_build_buffer_load(&ctx
->ac
, ctx
->esgs_ring
, 1, ctx
->i32_0
,
1585 vtx_offset
, soffset
, 0, 1, 0, true, false);
1586 if (llvm_type_is_64bit(ctx
, type
)) {
1587 LLVMValueRef value2
;
1588 soffset
= LLVMConstInt(ctx
->i32
, (param
* 4 + swizzle
+ 1) * 256, 0);
1590 value2
= ac_build_buffer_load(&ctx
->ac
, ctx
->esgs_ring
, 1,
1591 ctx
->i32_0
, vtx_offset
, soffset
,
1592 0, 1, 0, true, false);
1593 return si_llvm_emit_fetch_64bit(bld_base
, type
, value
, value2
);
1595 return LLVMBuildBitCast(ctx
->ac
.builder
, value
, type
, "");
1598 static LLVMValueRef
si_nir_load_input_gs(struct ac_shader_abi
*abi
,
1600 unsigned driver_location
,
1602 unsigned num_components
,
1603 unsigned vertex_index
,
1604 unsigned const_index
,
1607 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1609 LLVMValueRef value
[4];
1610 for (unsigned i
= 0; i
< num_components
; i
++) {
1611 unsigned offset
= i
;
1612 if (llvm_type_is_64bit(ctx
, type
))
1615 offset
+= component
;
1616 value
[i
+ component
] = si_llvm_load_input_gs(&ctx
->abi
, driver_location
/ 4,
1617 vertex_index
, type
, offset
);
1620 return ac_build_varying_gather_values(&ctx
->ac
, value
, num_components
, component
);
1623 static LLVMValueRef
fetch_input_gs(
1624 struct lp_build_tgsi_context
*bld_base
,
1625 const struct tgsi_full_src_register
*reg
,
1626 enum tgsi_opcode_type type
,
1627 unsigned swizzle_in
)
1629 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1630 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
1631 unsigned swizzle
= swizzle_in
& 0xffff;
1633 unsigned semantic_name
= info
->input_semantic_name
[reg
->Register
.Index
];
1634 if (swizzle
!= ~0 && semantic_name
== TGSI_SEMANTIC_PRIMID
)
1635 return si_get_primitive_id(ctx
, swizzle
);
1637 if (!reg
->Register
.Dimension
)
1640 return si_llvm_load_input_gs(&ctx
->abi
, reg
->Register
.Index
,
1641 reg
->Dimension
.Index
,
1642 tgsi2llvmtype(bld_base
, type
),
1646 static int lookup_interp_param_index(unsigned interpolate
, unsigned location
)
1648 switch (interpolate
) {
1649 case TGSI_INTERPOLATE_CONSTANT
:
1652 case TGSI_INTERPOLATE_LINEAR
:
1653 if (location
== TGSI_INTERPOLATE_LOC_SAMPLE
)
1654 return SI_PARAM_LINEAR_SAMPLE
;
1655 else if (location
== TGSI_INTERPOLATE_LOC_CENTROID
)
1656 return SI_PARAM_LINEAR_CENTROID
;
1658 return SI_PARAM_LINEAR_CENTER
;
1660 case TGSI_INTERPOLATE_COLOR
:
1661 case TGSI_INTERPOLATE_PERSPECTIVE
:
1662 if (location
== TGSI_INTERPOLATE_LOC_SAMPLE
)
1663 return SI_PARAM_PERSP_SAMPLE
;
1664 else if (location
== TGSI_INTERPOLATE_LOC_CENTROID
)
1665 return SI_PARAM_PERSP_CENTROID
;
1667 return SI_PARAM_PERSP_CENTER
;
1670 fprintf(stderr
, "Warning: Unhandled interpolation mode.\n");
1675 static LLVMValueRef
si_build_fs_interp(struct si_shader_context
*ctx
,
1676 unsigned attr_index
, unsigned chan
,
1677 LLVMValueRef prim_mask
,
1678 LLVMValueRef i
, LLVMValueRef j
)
1681 return ac_build_fs_interp(&ctx
->ac
,
1682 LLVMConstInt(ctx
->i32
, chan
, 0),
1683 LLVMConstInt(ctx
->i32
, attr_index
, 0),
1686 return ac_build_fs_interp_mov(&ctx
->ac
,
1687 LLVMConstInt(ctx
->i32
, 2, 0), /* P0 */
1688 LLVMConstInt(ctx
->i32
, chan
, 0),
1689 LLVMConstInt(ctx
->i32
, attr_index
, 0),
1694 * Interpolate a fragment shader input.
1696 * @param ctx context
1697 * @param input_index index of the input in hardware
1698 * @param semantic_name TGSI_SEMANTIC_*
1699 * @param semantic_index semantic index
1700 * @param num_interp_inputs number of all interpolated inputs (= BCOLOR offset)
1701 * @param colors_read_mask color components read (4 bits for each color, 8 bits in total)
1702 * @param interp_param interpolation weights (i,j)
1703 * @param prim_mask SI_PARAM_PRIM_MASK
1704 * @param face SI_PARAM_FRONT_FACE
1705 * @param result the return value (4 components)
1707 static void interp_fs_input(struct si_shader_context
*ctx
,
1708 unsigned input_index
,
1709 unsigned semantic_name
,
1710 unsigned semantic_index
,
1711 unsigned num_interp_inputs
,
1712 unsigned colors_read_mask
,
1713 LLVMValueRef interp_param
,
1714 LLVMValueRef prim_mask
,
1716 LLVMValueRef result
[4])
1718 LLVMValueRef i
= NULL
, j
= NULL
;
1721 /* fs.constant returns the param from the middle vertex, so it's not
1722 * really useful for flat shading. It's meant to be used for custom
1723 * interpolation (but the intrinsic can't fetch from the other two
1726 * Luckily, it doesn't matter, because we rely on the FLAT_SHADE state
1727 * to do the right thing. The only reason we use fs.constant is that
1728 * fs.interp cannot be used on integers, because they can be equal
1731 * When interp is false we will use fs.constant or for newer llvm,
1732 * amdgcn.interp.mov.
1734 bool interp
= interp_param
!= NULL
;
1737 interp_param
= LLVMBuildBitCast(ctx
->ac
.builder
, interp_param
,
1738 LLVMVectorType(ctx
->f32
, 2), "");
1740 i
= LLVMBuildExtractElement(ctx
->ac
.builder
, interp_param
,
1742 j
= LLVMBuildExtractElement(ctx
->ac
.builder
, interp_param
,
1746 if (semantic_name
== TGSI_SEMANTIC_COLOR
&&
1747 ctx
->shader
->key
.part
.ps
.prolog
.color_two_side
) {
1748 LLVMValueRef is_face_positive
;
1750 /* If BCOLOR0 is used, BCOLOR1 is at offset "num_inputs + 1",
1751 * otherwise it's at offset "num_inputs".
1753 unsigned back_attr_offset
= num_interp_inputs
;
1754 if (semantic_index
== 1 && colors_read_mask
& 0xf)
1755 back_attr_offset
+= 1;
1757 is_face_positive
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntNE
,
1758 face
, ctx
->i32_0
, "");
1760 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
1761 LLVMValueRef front
, back
;
1763 front
= si_build_fs_interp(ctx
,
1766 back
= si_build_fs_interp(ctx
,
1767 back_attr_offset
, chan
,
1770 result
[chan
] = LLVMBuildSelect(ctx
->ac
.builder
,
1776 } else if (semantic_name
== TGSI_SEMANTIC_FOG
) {
1777 result
[0] = si_build_fs_interp(ctx
, input_index
,
1778 0, prim_mask
, i
, j
);
1780 result
[2] = LLVMConstReal(ctx
->f32
, 0.0f
);
1781 result
[3] = LLVMConstReal(ctx
->f32
, 1.0f
);
1783 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
1784 result
[chan
] = si_build_fs_interp(ctx
,
1791 void si_llvm_load_input_fs(
1792 struct si_shader_context
*ctx
,
1793 unsigned input_index
,
1794 LLVMValueRef out
[4])
1796 struct si_shader
*shader
= ctx
->shader
;
1797 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
1798 LLVMValueRef main_fn
= ctx
->main_fn
;
1799 LLVMValueRef interp_param
= NULL
;
1800 int interp_param_idx
;
1801 enum tgsi_semantic semantic_name
= info
->input_semantic_name
[input_index
];
1802 unsigned semantic_index
= info
->input_semantic_index
[input_index
];
1803 enum tgsi_interpolate_mode interp_mode
= info
->input_interpolate
[input_index
];
1804 enum tgsi_interpolate_loc interp_loc
= info
->input_interpolate_loc
[input_index
];
1806 /* Get colors from input VGPRs (set by the prolog). */
1807 if (semantic_name
== TGSI_SEMANTIC_COLOR
) {
1808 unsigned colors_read
= shader
->selector
->info
.colors_read
;
1809 unsigned mask
= colors_read
>> (semantic_index
* 4);
1810 unsigned offset
= SI_PARAM_POS_FIXED_PT
+ 1 +
1811 (semantic_index
? util_bitcount(colors_read
& 0xf) : 0);
1812 LLVMValueRef undef
= LLVMGetUndef(ctx
->f32
);
1814 out
[0] = mask
& 0x1 ? LLVMGetParam(main_fn
, offset
++) : undef
;
1815 out
[1] = mask
& 0x2 ? LLVMGetParam(main_fn
, offset
++) : undef
;
1816 out
[2] = mask
& 0x4 ? LLVMGetParam(main_fn
, offset
++) : undef
;
1817 out
[3] = mask
& 0x8 ? LLVMGetParam(main_fn
, offset
++) : undef
;
1821 interp_param_idx
= lookup_interp_param_index(interp_mode
, interp_loc
);
1822 if (interp_param_idx
== -1)
1824 else if (interp_param_idx
) {
1825 interp_param
= LLVMGetParam(ctx
->main_fn
, interp_param_idx
);
1828 interp_fs_input(ctx
, input_index
, semantic_name
,
1829 semantic_index
, 0, /* this param is unused */
1830 shader
->selector
->info
.colors_read
, interp_param
,
1832 LLVMGetParam(main_fn
, SI_PARAM_FRONT_FACE
),
1836 static void declare_input_fs(
1837 struct si_shader_context
*ctx
,
1838 unsigned input_index
,
1839 const struct tgsi_full_declaration
*decl
,
1840 LLVMValueRef out
[4])
1842 si_llvm_load_input_fs(ctx
, input_index
, out
);
1845 LLVMValueRef
si_get_sample_id(struct si_shader_context
*ctx
)
1847 return si_unpack_param(ctx
, SI_PARAM_ANCILLARY
, 8, 4);
1850 static LLVMValueRef
get_base_vertex(struct ac_shader_abi
*abi
)
1852 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1854 /* For non-indexed draws, the base vertex set by the driver
1855 * (for direct draws) or the CP (for indirect draws) is the
1856 * first vertex ID, but GLSL expects 0 to be returned.
1858 LLVMValueRef vs_state
= LLVMGetParam(ctx
->main_fn
,
1859 ctx
->param_vs_state_bits
);
1860 LLVMValueRef indexed
;
1862 indexed
= LLVMBuildLShr(ctx
->ac
.builder
, vs_state
, ctx
->i32_1
, "");
1863 indexed
= LLVMBuildTrunc(ctx
->ac
.builder
, indexed
, ctx
->i1
, "");
1865 return LLVMBuildSelect(ctx
->ac
.builder
, indexed
, ctx
->abi
.base_vertex
,
1869 static LLVMValueRef
get_block_size(struct ac_shader_abi
*abi
)
1871 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1873 LLVMValueRef values
[3];
1874 LLVMValueRef result
;
1876 unsigned *properties
= ctx
->shader
->selector
->info
.properties
;
1878 if (properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
] != 0) {
1879 unsigned sizes
[3] = {
1880 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
],
1881 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT
],
1882 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH
]
1885 for (i
= 0; i
< 3; ++i
)
1886 values
[i
] = LLVMConstInt(ctx
->i32
, sizes
[i
], 0);
1888 result
= ac_build_gather_values(&ctx
->ac
, values
, 3);
1890 result
= LLVMGetParam(ctx
->main_fn
, ctx
->param_block_size
);
1897 * Load a dword from a constant buffer.
1899 static LLVMValueRef
buffer_load_const(struct si_shader_context
*ctx
,
1900 LLVMValueRef resource
,
1901 LLVMValueRef offset
)
1903 return ac_build_buffer_load(&ctx
->ac
, resource
, 1, NULL
, offset
, NULL
,
1904 0, 0, 0, true, true);
1907 static LLVMValueRef
load_sample_position(struct ac_shader_abi
*abi
, LLVMValueRef sample_id
)
1909 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1910 LLVMValueRef desc
= LLVMGetParam(ctx
->main_fn
, ctx
->param_rw_buffers
);
1911 LLVMValueRef buf_index
= LLVMConstInt(ctx
->i32
, SI_PS_CONST_SAMPLE_POSITIONS
, 0);
1912 LLVMValueRef resource
= ac_build_load_to_sgpr(&ctx
->ac
, desc
, buf_index
);
1914 /* offset = sample_id * 8 (8 = 2 floats containing samplepos.xy) */
1915 LLVMValueRef offset0
= LLVMBuildMul(ctx
->ac
.builder
, sample_id
, LLVMConstInt(ctx
->i32
, 8, 0), "");
1916 LLVMValueRef offset1
= LLVMBuildAdd(ctx
->ac
.builder
, offset0
, LLVMConstInt(ctx
->i32
, 4, 0), "");
1918 LLVMValueRef pos
[4] = {
1919 buffer_load_const(ctx
, resource
, offset0
),
1920 buffer_load_const(ctx
, resource
, offset1
),
1921 LLVMConstReal(ctx
->f32
, 0),
1922 LLVMConstReal(ctx
->f32
, 0)
1925 return ac_build_gather_values(&ctx
->ac
, pos
, 4);
1928 static LLVMValueRef
load_sample_mask_in(struct ac_shader_abi
*abi
)
1930 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1931 return ac_to_integer(&ctx
->ac
, abi
->sample_coverage
);
1934 static LLVMValueRef
si_load_tess_coord(struct ac_shader_abi
*abi
)
1936 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1937 LLVMValueRef coord
[4] = {
1938 LLVMGetParam(ctx
->main_fn
, ctx
->param_tes_u
),
1939 LLVMGetParam(ctx
->main_fn
, ctx
->param_tes_v
),
1944 /* For triangles, the vector should be (u, v, 1-u-v). */
1945 if (ctx
->shader
->selector
->info
.properties
[TGSI_PROPERTY_TES_PRIM_MODE
] ==
1946 PIPE_PRIM_TRIANGLES
) {
1947 coord
[2] = LLVMBuildFSub(ctx
->ac
.builder
, ctx
->ac
.f32_1
,
1948 LLVMBuildFAdd(ctx
->ac
.builder
,
1949 coord
[0], coord
[1], ""), "");
1951 return ac_build_gather_values(&ctx
->ac
, coord
, 4);
1954 static LLVMValueRef
load_tess_level(struct si_shader_context
*ctx
,
1955 unsigned semantic_name
)
1957 LLVMValueRef base
, addr
;
1959 int param
= si_shader_io_get_unique_index_patch(semantic_name
, 0);
1961 base
= LLVMGetParam(ctx
->main_fn
, ctx
->param_tcs_offchip_offset
);
1962 addr
= get_tcs_tes_buffer_address(ctx
, get_rel_patch_id(ctx
), NULL
,
1963 LLVMConstInt(ctx
->i32
, param
, 0));
1965 return buffer_load(&ctx
->bld_base
, ctx
->f32
,
1966 ~0, ctx
->tess_offchip_ring
, base
, addr
, true);
1970 static LLVMValueRef
si_load_tess_level(struct ac_shader_abi
*abi
,
1971 unsigned varying_id
)
1973 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1974 unsigned semantic_name
;
1976 switch (varying_id
) {
1977 case VARYING_SLOT_TESS_LEVEL_INNER
:
1978 semantic_name
= TGSI_SEMANTIC_TESSINNER
;
1980 case VARYING_SLOT_TESS_LEVEL_OUTER
:
1981 semantic_name
= TGSI_SEMANTIC_TESSOUTER
;
1984 unreachable("unknown tess level");
1987 return load_tess_level(ctx
, semantic_name
);
1991 static LLVMValueRef
si_load_patch_vertices_in(struct ac_shader_abi
*abi
)
1993 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1994 if (ctx
->type
== PIPE_SHADER_TESS_CTRL
)
1995 return si_unpack_param(ctx
, ctx
->param_tcs_out_lds_layout
, 13, 6);
1996 else if (ctx
->type
== PIPE_SHADER_TESS_EVAL
)
1997 return get_num_tcs_out_vertices(ctx
);
1999 unreachable("invalid shader stage for TGSI_SEMANTIC_VERTICESIN");
2002 void si_load_system_value(struct si_shader_context
*ctx
,
2004 const struct tgsi_full_declaration
*decl
)
2006 LLVMValueRef value
= 0;
2008 assert(index
< RADEON_LLVM_MAX_SYSTEM_VALUES
);
2010 switch (decl
->Semantic
.Name
) {
2011 case TGSI_SEMANTIC_INSTANCEID
:
2012 value
= ctx
->abi
.instance_id
;
2015 case TGSI_SEMANTIC_VERTEXID
:
2016 value
= LLVMBuildAdd(ctx
->ac
.builder
,
2018 ctx
->abi
.base_vertex
, "");
2021 case TGSI_SEMANTIC_VERTEXID_NOBASE
:
2022 /* Unused. Clarify the meaning in indexed vs. non-indexed
2023 * draws if this is ever used again. */
2027 case TGSI_SEMANTIC_BASEVERTEX
:
2028 value
= get_base_vertex(&ctx
->abi
);
2031 case TGSI_SEMANTIC_BASEINSTANCE
:
2032 value
= ctx
->abi
.start_instance
;
2035 case TGSI_SEMANTIC_DRAWID
:
2036 value
= ctx
->abi
.draw_id
;
2039 case TGSI_SEMANTIC_INVOCATIONID
:
2040 if (ctx
->type
== PIPE_SHADER_TESS_CTRL
)
2041 value
= unpack_llvm_param(ctx
, ctx
->abi
.tcs_rel_ids
, 8, 5);
2042 else if (ctx
->type
== PIPE_SHADER_GEOMETRY
)
2043 value
= ctx
->abi
.gs_invocation_id
;
2045 assert(!"INVOCATIONID not implemented");
2048 case TGSI_SEMANTIC_POSITION
:
2050 LLVMValueRef pos
[4] = {
2051 LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_X_FLOAT
),
2052 LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_Y_FLOAT
),
2053 LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_Z_FLOAT
),
2054 ac_build_fdiv(&ctx
->ac
, ctx
->ac
.f32_1
,
2055 LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_W_FLOAT
)),
2057 value
= ac_build_gather_values(&ctx
->ac
, pos
, 4);
2061 case TGSI_SEMANTIC_FACE
:
2062 value
= ctx
->abi
.front_face
;
2065 case TGSI_SEMANTIC_SAMPLEID
:
2066 value
= si_get_sample_id(ctx
);
2069 case TGSI_SEMANTIC_SAMPLEPOS
: {
2070 LLVMValueRef pos
[4] = {
2071 LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_X_FLOAT
),
2072 LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_Y_FLOAT
),
2073 LLVMConstReal(ctx
->f32
, 0),
2074 LLVMConstReal(ctx
->f32
, 0)
2076 pos
[0] = ac_build_fract(&ctx
->ac
, pos
[0], 32);
2077 pos
[1] = ac_build_fract(&ctx
->ac
, pos
[1], 32);
2078 value
= ac_build_gather_values(&ctx
->ac
, pos
, 4);
2082 case TGSI_SEMANTIC_SAMPLEMASK
:
2083 /* This can only occur with the OpenGL Core profile, which
2084 * doesn't support smoothing.
2086 value
= LLVMGetParam(ctx
->main_fn
, SI_PARAM_SAMPLE_COVERAGE
);
2089 case TGSI_SEMANTIC_TESSCOORD
:
2090 value
= si_load_tess_coord(&ctx
->abi
);
2093 case TGSI_SEMANTIC_VERTICESIN
:
2094 value
= si_load_patch_vertices_in(&ctx
->abi
);
2097 case TGSI_SEMANTIC_TESSINNER
:
2098 case TGSI_SEMANTIC_TESSOUTER
:
2099 value
= load_tess_level(ctx
, decl
->Semantic
.Name
);
2102 case TGSI_SEMANTIC_DEFAULT_TESSOUTER_SI
:
2103 case TGSI_SEMANTIC_DEFAULT_TESSINNER_SI
:
2105 LLVMValueRef buf
, slot
, val
[4];
2108 slot
= LLVMConstInt(ctx
->i32
, SI_HS_CONST_DEFAULT_TESS_LEVELS
, 0);
2109 buf
= LLVMGetParam(ctx
->main_fn
, ctx
->param_rw_buffers
);
2110 buf
= ac_build_load_to_sgpr(&ctx
->ac
, buf
, slot
);
2111 offset
= decl
->Semantic
.Name
== TGSI_SEMANTIC_DEFAULT_TESSINNER_SI
? 4 : 0;
2113 for (i
= 0; i
< 4; i
++)
2114 val
[i
] = buffer_load_const(ctx
, buf
,
2115 LLVMConstInt(ctx
->i32
, (offset
+ i
) * 4, 0));
2116 value
= ac_build_gather_values(&ctx
->ac
, val
, 4);
2120 case TGSI_SEMANTIC_PRIMID
:
2121 value
= si_get_primitive_id(ctx
, 0);
2124 case TGSI_SEMANTIC_GRID_SIZE
:
2125 value
= ctx
->abi
.num_work_groups
;
2128 case TGSI_SEMANTIC_BLOCK_SIZE
:
2129 value
= get_block_size(&ctx
->abi
);
2132 case TGSI_SEMANTIC_BLOCK_ID
:
2134 LLVMValueRef values
[3];
2136 for (int i
= 0; i
< 3; i
++) {
2137 values
[i
] = ctx
->i32_0
;
2138 if (ctx
->abi
.workgroup_ids
[i
]) {
2139 values
[i
] = ctx
->abi
.workgroup_ids
[i
];
2142 value
= ac_build_gather_values(&ctx
->ac
, values
, 3);
2146 case TGSI_SEMANTIC_THREAD_ID
:
2147 value
= ctx
->abi
.local_invocation_ids
;
2150 case TGSI_SEMANTIC_HELPER_INVOCATION
:
2151 value
= ac_build_load_helper_invocation(&ctx
->ac
);
2154 case TGSI_SEMANTIC_SUBGROUP_SIZE
:
2155 value
= LLVMConstInt(ctx
->i32
, 64, 0);
2158 case TGSI_SEMANTIC_SUBGROUP_INVOCATION
:
2159 value
= ac_get_thread_id(&ctx
->ac
);
2162 case TGSI_SEMANTIC_SUBGROUP_EQ_MASK
:
2164 LLVMValueRef id
= ac_get_thread_id(&ctx
->ac
);
2165 id
= LLVMBuildZExt(ctx
->ac
.builder
, id
, ctx
->i64
, "");
2166 value
= LLVMBuildShl(ctx
->ac
.builder
, LLVMConstInt(ctx
->i64
, 1, 0), id
, "");
2167 value
= LLVMBuildBitCast(ctx
->ac
.builder
, value
, ctx
->v2i32
, "");
2171 case TGSI_SEMANTIC_SUBGROUP_GE_MASK
:
2172 case TGSI_SEMANTIC_SUBGROUP_GT_MASK
:
2173 case TGSI_SEMANTIC_SUBGROUP_LE_MASK
:
2174 case TGSI_SEMANTIC_SUBGROUP_LT_MASK
:
2176 LLVMValueRef id
= ac_get_thread_id(&ctx
->ac
);
2177 if (decl
->Semantic
.Name
== TGSI_SEMANTIC_SUBGROUP_GT_MASK
||
2178 decl
->Semantic
.Name
== TGSI_SEMANTIC_SUBGROUP_LE_MASK
) {
2179 /* All bits set except LSB */
2180 value
= LLVMConstInt(ctx
->i64
, -2, 0);
2183 value
= LLVMConstInt(ctx
->i64
, -1, 0);
2185 id
= LLVMBuildZExt(ctx
->ac
.builder
, id
, ctx
->i64
, "");
2186 value
= LLVMBuildShl(ctx
->ac
.builder
, value
, id
, "");
2187 if (decl
->Semantic
.Name
== TGSI_SEMANTIC_SUBGROUP_LE_MASK
||
2188 decl
->Semantic
.Name
== TGSI_SEMANTIC_SUBGROUP_LT_MASK
)
2189 value
= LLVMBuildNot(ctx
->ac
.builder
, value
, "");
2190 value
= LLVMBuildBitCast(ctx
->ac
.builder
, value
, ctx
->v2i32
, "");
2194 case TGSI_SEMANTIC_CS_USER_DATA
:
2195 value
= LLVMGetParam(ctx
->main_fn
, ctx
->param_cs_user_data
);
2199 assert(!"unknown system value");
2203 ctx
->system_values
[index
] = value
;
2206 void si_declare_compute_memory(struct si_shader_context
*ctx
)
2208 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
2209 unsigned lds_size
= sel
->info
.properties
[TGSI_PROPERTY_CS_LOCAL_SIZE
];
2211 LLVMTypeRef i8p
= LLVMPointerType(ctx
->i8
, AC_ADDR_SPACE_LDS
);
2214 assert(!ctx
->ac
.lds
);
2216 var
= LLVMAddGlobalInAddressSpace(ctx
->ac
.module
,
2217 LLVMArrayType(ctx
->i8
, lds_size
),
2220 LLVMSetAlignment(var
, 64 * 1024);
2222 ctx
->ac
.lds
= LLVMBuildBitCast(ctx
->ac
.builder
, var
, i8p
, "");
2225 void si_tgsi_declare_compute_memory(struct si_shader_context
*ctx
,
2226 const struct tgsi_full_declaration
*decl
)
2228 assert(decl
->Declaration
.MemType
== TGSI_MEMORY_TYPE_SHARED
);
2229 assert(decl
->Range
.First
== decl
->Range
.Last
);
2231 si_declare_compute_memory(ctx
);
2234 static LLVMValueRef
load_const_buffer_desc_fast_path(struct si_shader_context
*ctx
)
2237 LLVMGetParam(ctx
->main_fn
, ctx
->param_const_and_shader_buffers
);
2238 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
2240 /* Do the bounds checking with a descriptor, because
2241 * doing computation and manual bounds checking of 64-bit
2242 * addresses generates horrible VALU code with very high
2243 * VGPR usage and very low SIMD occupancy.
2245 ptr
= LLVMBuildPtrToInt(ctx
->ac
.builder
, ptr
, ctx
->ac
.intptr
, "");
2247 LLVMValueRef desc0
, desc1
;
2249 desc1
= LLVMConstInt(ctx
->i32
,
2250 S_008F04_BASE_ADDRESS_HI(ctx
->screen
->info
.address32_hi
), 0);
2252 LLVMValueRef desc_elems
[] = {
2255 LLVMConstInt(ctx
->i32
, (sel
->info
.const_file_max
[0] + 1) * 16, 0),
2256 LLVMConstInt(ctx
->i32
,
2257 S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
2258 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
2259 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
2260 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
2261 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
2262 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
), 0)
2265 return ac_build_gather_values(&ctx
->ac
, desc_elems
, 4);
2268 static LLVMValueRef
load_const_buffer_desc(struct si_shader_context
*ctx
, int i
)
2270 LLVMValueRef list_ptr
= LLVMGetParam(ctx
->main_fn
,
2271 ctx
->param_const_and_shader_buffers
);
2273 return ac_build_load_to_sgpr(&ctx
->ac
, list_ptr
,
2274 LLVMConstInt(ctx
->i32
, si_get_constbuf_slot(i
), 0));
2277 static LLVMValueRef
load_ubo(struct ac_shader_abi
*abi
, LLVMValueRef index
)
2279 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
2280 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
2282 LLVMValueRef ptr
= LLVMGetParam(ctx
->main_fn
, ctx
->param_const_and_shader_buffers
);
2284 if (sel
->info
.const_buffers_declared
== 1 &&
2285 sel
->info
.shader_buffers_declared
== 0) {
2286 return load_const_buffer_desc_fast_path(ctx
);
2289 index
= si_llvm_bound_index(ctx
, index
, ctx
->num_const_buffers
);
2290 index
= LLVMBuildAdd(ctx
->ac
.builder
, index
,
2291 LLVMConstInt(ctx
->i32
, SI_NUM_SHADER_BUFFERS
, 0), "");
2293 return ac_build_load_to_sgpr(&ctx
->ac
, ptr
, index
);
2297 load_ssbo(struct ac_shader_abi
*abi
, LLVMValueRef index
, bool write
)
2299 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
2300 LLVMValueRef rsrc_ptr
= LLVMGetParam(ctx
->main_fn
,
2301 ctx
->param_const_and_shader_buffers
);
2303 index
= si_llvm_bound_index(ctx
, index
, ctx
->num_shader_buffers
);
2304 index
= LLVMBuildSub(ctx
->ac
.builder
,
2305 LLVMConstInt(ctx
->i32
, SI_NUM_SHADER_BUFFERS
- 1, 0),
2308 return ac_build_load_to_sgpr(&ctx
->ac
, rsrc_ptr
, index
);
2311 static LLVMValueRef
fetch_constant(
2312 struct lp_build_tgsi_context
*bld_base
,
2313 const struct tgsi_full_src_register
*reg
,
2314 enum tgsi_opcode_type type
,
2315 unsigned swizzle_in
)
2317 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2318 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
2319 const struct tgsi_ind_register
*ireg
= ®
->Indirect
;
2321 unsigned swizzle
= swizzle_in
& 0xffff;
2323 LLVMValueRef addr
, bufp
;
2325 if (swizzle_in
== LP_CHAN_ALL
) {
2327 LLVMValueRef values
[4];
2328 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
)
2329 values
[chan
] = fetch_constant(bld_base
, reg
, type
, chan
);
2331 return ac_build_gather_values(&ctx
->ac
, values
, 4);
2334 /* Split 64-bit loads. */
2335 if (tgsi_type_is_64bit(type
)) {
2336 LLVMValueRef lo
, hi
;
2338 lo
= fetch_constant(bld_base
, reg
, TGSI_TYPE_UNSIGNED
, swizzle
);
2339 hi
= fetch_constant(bld_base
, reg
, TGSI_TYPE_UNSIGNED
, (swizzle_in
>> 16));
2340 return si_llvm_emit_fetch_64bit(bld_base
, tgsi2llvmtype(bld_base
, type
),
2344 idx
= reg
->Register
.Index
* 4 + swizzle
;
2345 if (reg
->Register
.Indirect
) {
2346 addr
= si_get_indirect_index(ctx
, ireg
, 16, idx
* 4);
2348 addr
= LLVMConstInt(ctx
->i32
, idx
* 4, 0);
2351 /* Fast path when user data SGPRs point to constant buffer 0 directly. */
2352 if (sel
->info
.const_buffers_declared
== 1 &&
2353 sel
->info
.shader_buffers_declared
== 0) {
2354 LLVMValueRef desc
= load_const_buffer_desc_fast_path(ctx
);
2355 LLVMValueRef result
= buffer_load_const(ctx
, desc
, addr
);
2356 return bitcast(bld_base
, type
, result
);
2359 assert(reg
->Register
.Dimension
);
2360 buf
= reg
->Dimension
.Index
;
2362 if (reg
->Dimension
.Indirect
) {
2363 LLVMValueRef ptr
= LLVMGetParam(ctx
->main_fn
, ctx
->param_const_and_shader_buffers
);
2365 index
= si_get_bounded_indirect_index(ctx
, ®
->DimIndirect
,
2366 reg
->Dimension
.Index
,
2367 ctx
->num_const_buffers
);
2368 index
= LLVMBuildAdd(ctx
->ac
.builder
, index
,
2369 LLVMConstInt(ctx
->i32
, SI_NUM_SHADER_BUFFERS
, 0), "");
2370 bufp
= ac_build_load_to_sgpr(&ctx
->ac
, ptr
, index
);
2372 bufp
= load_const_buffer_desc(ctx
, buf
);
2374 return bitcast(bld_base
, type
, buffer_load_const(ctx
, bufp
, addr
));
2377 /* Initialize arguments for the shader export intrinsic */
2378 static void si_llvm_init_export_args(struct si_shader_context
*ctx
,
2379 LLVMValueRef
*values
,
2381 struct ac_export_args
*args
)
2383 LLVMValueRef f32undef
= LLVMGetUndef(ctx
->ac
.f32
);
2384 unsigned spi_shader_col_format
= V_028714_SPI_SHADER_32_ABGR
;
2386 bool is_int8
, is_int10
;
2388 /* Default is 0xf. Adjusted below depending on the format. */
2389 args
->enabled_channels
= 0xf; /* writemask */
2391 /* Specify whether the EXEC mask represents the valid mask */
2392 args
->valid_mask
= 0;
2394 /* Specify whether this is the last export */
2397 /* Specify the target we are exporting */
2398 args
->target
= target
;
2400 if (ctx
->type
== PIPE_SHADER_FRAGMENT
) {
2401 const struct si_shader_key
*key
= &ctx
->shader
->key
;
2402 unsigned col_formats
= key
->part
.ps
.epilog
.spi_shader_col_format
;
2403 int cbuf
= target
- V_008DFC_SQ_EXP_MRT
;
2405 assert(cbuf
>= 0 && cbuf
< 8);
2406 spi_shader_col_format
= (col_formats
>> (cbuf
* 4)) & 0xf;
2407 is_int8
= (key
->part
.ps
.epilog
.color_is_int8
>> cbuf
) & 0x1;
2408 is_int10
= (key
->part
.ps
.epilog
.color_is_int10
>> cbuf
) & 0x1;
2411 args
->compr
= false;
2412 args
->out
[0] = f32undef
;
2413 args
->out
[1] = f32undef
;
2414 args
->out
[2] = f32undef
;
2415 args
->out
[3] = f32undef
;
2417 LLVMValueRef (*packf
)(struct ac_llvm_context
*ctx
, LLVMValueRef args
[2]) = NULL
;
2418 LLVMValueRef (*packi
)(struct ac_llvm_context
*ctx
, LLVMValueRef args
[2],
2419 unsigned bits
, bool hi
) = NULL
;
2421 switch (spi_shader_col_format
) {
2422 case V_028714_SPI_SHADER_ZERO
:
2423 args
->enabled_channels
= 0; /* writemask */
2424 args
->target
= V_008DFC_SQ_EXP_NULL
;
2427 case V_028714_SPI_SHADER_32_R
:
2428 args
->enabled_channels
= 1; /* writemask */
2429 args
->out
[0] = values
[0];
2432 case V_028714_SPI_SHADER_32_GR
:
2433 args
->enabled_channels
= 0x3; /* writemask */
2434 args
->out
[0] = values
[0];
2435 args
->out
[1] = values
[1];
2438 case V_028714_SPI_SHADER_32_AR
:
2439 args
->enabled_channels
= 0x9; /* writemask */
2440 args
->out
[0] = values
[0];
2441 args
->out
[3] = values
[3];
2444 case V_028714_SPI_SHADER_FP16_ABGR
:
2445 packf
= ac_build_cvt_pkrtz_f16
;
2448 case V_028714_SPI_SHADER_UNORM16_ABGR
:
2449 packf
= ac_build_cvt_pknorm_u16
;
2452 case V_028714_SPI_SHADER_SNORM16_ABGR
:
2453 packf
= ac_build_cvt_pknorm_i16
;
2456 case V_028714_SPI_SHADER_UINT16_ABGR
:
2457 packi
= ac_build_cvt_pk_u16
;
2460 case V_028714_SPI_SHADER_SINT16_ABGR
:
2461 packi
= ac_build_cvt_pk_i16
;
2464 case V_028714_SPI_SHADER_32_ABGR
:
2465 memcpy(&args
->out
[0], values
, sizeof(values
[0]) * 4);
2469 /* Pack f16 or norm_i16/u16. */
2471 for (chan
= 0; chan
< 2; chan
++) {
2472 LLVMValueRef pack_args
[2] = {
2474 values
[2 * chan
+ 1]
2476 LLVMValueRef packed
;
2478 packed
= packf(&ctx
->ac
, pack_args
);
2479 args
->out
[chan
] = ac_to_float(&ctx
->ac
, packed
);
2481 args
->compr
= 1; /* COMPR flag */
2485 for (chan
= 0; chan
< 2; chan
++) {
2486 LLVMValueRef pack_args
[2] = {
2487 ac_to_integer(&ctx
->ac
, values
[2 * chan
]),
2488 ac_to_integer(&ctx
->ac
, values
[2 * chan
+ 1])
2490 LLVMValueRef packed
;
2492 packed
= packi(&ctx
->ac
, pack_args
,
2493 is_int8
? 8 : is_int10
? 10 : 16,
2495 args
->out
[chan
] = ac_to_float(&ctx
->ac
, packed
);
2497 args
->compr
= 1; /* COMPR flag */
2501 static void si_alpha_test(struct lp_build_tgsi_context
*bld_base
,
2504 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2506 if (ctx
->shader
->key
.part
.ps
.epilog
.alpha_func
!= PIPE_FUNC_NEVER
) {
2507 static LLVMRealPredicate cond_map
[PIPE_FUNC_ALWAYS
+ 1] = {
2508 [PIPE_FUNC_LESS
] = LLVMRealOLT
,
2509 [PIPE_FUNC_EQUAL
] = LLVMRealOEQ
,
2510 [PIPE_FUNC_LEQUAL
] = LLVMRealOLE
,
2511 [PIPE_FUNC_GREATER
] = LLVMRealOGT
,
2512 [PIPE_FUNC_NOTEQUAL
] = LLVMRealONE
,
2513 [PIPE_FUNC_GEQUAL
] = LLVMRealOGE
,
2515 LLVMRealPredicate cond
= cond_map
[ctx
->shader
->key
.part
.ps
.epilog
.alpha_func
];
2518 LLVMValueRef alpha_ref
= LLVMGetParam(ctx
->main_fn
,
2519 SI_PARAM_ALPHA_REF
);
2520 LLVMValueRef alpha_pass
=
2521 LLVMBuildFCmp(ctx
->ac
.builder
, cond
, alpha
, alpha_ref
, "");
2522 ac_build_kill_if_false(&ctx
->ac
, alpha_pass
);
2524 ac_build_kill_if_false(&ctx
->ac
, ctx
->i1false
);
2528 static LLVMValueRef
si_scale_alpha_by_sample_mask(struct lp_build_tgsi_context
*bld_base
,
2530 unsigned samplemask_param
)
2532 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2533 LLVMValueRef coverage
;
2535 /* alpha = alpha * popcount(coverage) / SI_NUM_SMOOTH_AA_SAMPLES */
2536 coverage
= LLVMGetParam(ctx
->main_fn
,
2538 coverage
= ac_to_integer(&ctx
->ac
, coverage
);
2540 coverage
= ac_build_intrinsic(&ctx
->ac
, "llvm.ctpop.i32",
2542 &coverage
, 1, AC_FUNC_ATTR_READNONE
);
2544 coverage
= LLVMBuildUIToFP(ctx
->ac
.builder
, coverage
,
2547 coverage
= LLVMBuildFMul(ctx
->ac
.builder
, coverage
,
2548 LLVMConstReal(ctx
->f32
,
2549 1.0 / SI_NUM_SMOOTH_AA_SAMPLES
), "");
2551 return LLVMBuildFMul(ctx
->ac
.builder
, alpha
, coverage
, "");
2554 static void si_llvm_emit_clipvertex(struct si_shader_context
*ctx
,
2555 struct ac_export_args
*pos
, LLVMValueRef
*out_elts
)
2559 unsigned const_chan
;
2560 LLVMValueRef base_elt
;
2561 LLVMValueRef ptr
= LLVMGetParam(ctx
->main_fn
, ctx
->param_rw_buffers
);
2562 LLVMValueRef constbuf_index
= LLVMConstInt(ctx
->i32
,
2563 SI_VS_CONST_CLIP_PLANES
, 0);
2564 LLVMValueRef const_resource
= ac_build_load_to_sgpr(&ctx
->ac
, ptr
, constbuf_index
);
2566 for (reg_index
= 0; reg_index
< 2; reg_index
++) {
2567 struct ac_export_args
*args
= &pos
[2 + reg_index
];
2572 args
->out
[3] = LLVMConstReal(ctx
->f32
, 0.0f
);
2574 /* Compute dot products of position and user clip plane vectors */
2575 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
2576 for (const_chan
= 0; const_chan
< TGSI_NUM_CHANNELS
; const_chan
++) {
2578 LLVMConstInt(ctx
->i32
, ((reg_index
* 4 + chan
) * 4 +
2579 const_chan
) * 4, 0);
2580 base_elt
= buffer_load_const(ctx
, const_resource
,
2582 args
->out
[chan
] = ac_build_fmad(&ctx
->ac
, base_elt
,
2583 out_elts
[const_chan
], args
->out
[chan
]);
2587 args
->enabled_channels
= 0xf;
2588 args
->valid_mask
= 0;
2590 args
->target
= V_008DFC_SQ_EXP_POS
+ 2 + reg_index
;
2595 static void si_dump_streamout(struct pipe_stream_output_info
*so
)
2599 if (so
->num_outputs
)
2600 fprintf(stderr
, "STREAMOUT\n");
2602 for (i
= 0; i
< so
->num_outputs
; i
++) {
2603 unsigned mask
= ((1 << so
->output
[i
].num_components
) - 1) <<
2604 so
->output
[i
].start_component
;
2605 fprintf(stderr
, " %i: BUF%i[%i..%i] <- OUT[%i].%s%s%s%s\n",
2606 i
, so
->output
[i
].output_buffer
,
2607 so
->output
[i
].dst_offset
, so
->output
[i
].dst_offset
+ so
->output
[i
].num_components
- 1,
2608 so
->output
[i
].register_index
,
2609 mask
& 1 ? "x" : "",
2610 mask
& 2 ? "y" : "",
2611 mask
& 4 ? "z" : "",
2612 mask
& 8 ? "w" : "");
2616 void si_emit_streamout_output(struct si_shader_context
*ctx
,
2617 LLVMValueRef
const *so_buffers
,
2618 LLVMValueRef
const *so_write_offsets
,
2619 struct pipe_stream_output
*stream_out
,
2620 struct si_shader_output_values
*shader_out
)
2622 unsigned buf_idx
= stream_out
->output_buffer
;
2623 unsigned start
= stream_out
->start_component
;
2624 unsigned num_comps
= stream_out
->num_components
;
2625 LLVMValueRef out
[4];
2627 assert(num_comps
&& num_comps
<= 4);
2628 if (!num_comps
|| num_comps
> 4)
2631 /* Load the output as int. */
2632 for (int j
= 0; j
< num_comps
; j
++) {
2633 assert(stream_out
->stream
== shader_out
->vertex_stream
[start
+ j
]);
2635 out
[j
] = ac_to_integer(&ctx
->ac
, shader_out
->values
[start
+ j
]);
2638 /* Pack the output. */
2639 LLVMValueRef vdata
= NULL
;
2641 switch (num_comps
) {
2642 case 1: /* as i32 */
2645 case 2: /* as v2i32 */
2646 case 3: /* as v3i32 */
2647 if (ac_has_vec3_support(ctx
->screen
->info
.chip_class
, false)) {
2648 vdata
= ac_build_gather_values(&ctx
->ac
, out
, num_comps
);
2651 /* as v4i32 (aligned to 4) */
2652 out
[3] = LLVMGetUndef(ctx
->i32
);
2654 case 4: /* as v4i32 */
2655 vdata
= ac_build_gather_values(&ctx
->ac
, out
, util_next_power_of_two(num_comps
));
2659 ac_build_buffer_store_dword(&ctx
->ac
, so_buffers
[buf_idx
],
2661 so_write_offsets
[buf_idx
],
2663 stream_out
->dst_offset
* 4, 1, 1, false);
2667 * Write streamout data to buffers for vertex stream @p stream (different
2668 * vertex streams can occur for GS copy shaders).
2670 static void si_llvm_emit_streamout(struct si_shader_context
*ctx
,
2671 struct si_shader_output_values
*outputs
,
2672 unsigned noutput
, unsigned stream
)
2674 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
2675 struct pipe_stream_output_info
*so
= &sel
->so
;
2676 LLVMBuilderRef builder
= ctx
->ac
.builder
;
2678 struct lp_build_if_state if_ctx
;
2680 /* Get bits [22:16], i.e. (so_param >> 16) & 127; */
2681 LLVMValueRef so_vtx_count
=
2682 si_unpack_param(ctx
, ctx
->param_streamout_config
, 16, 7);
2684 LLVMValueRef tid
= ac_get_thread_id(&ctx
->ac
);
2686 /* can_emit = tid < so_vtx_count; */
2687 LLVMValueRef can_emit
=
2688 LLVMBuildICmp(builder
, LLVMIntULT
, tid
, so_vtx_count
, "");
2690 /* Emit the streamout code conditionally. This actually avoids
2691 * out-of-bounds buffer access. The hw tells us via the SGPR
2692 * (so_vtx_count) which threads are allowed to emit streamout data. */
2693 lp_build_if(&if_ctx
, &ctx
->gallivm
, can_emit
);
2695 /* The buffer offset is computed as follows:
2696 * ByteOffset = streamout_offset[buffer_id]*4 +
2697 * (streamout_write_index + thread_id)*stride[buffer_id] +
2701 LLVMValueRef so_write_index
=
2702 LLVMGetParam(ctx
->main_fn
,
2703 ctx
->param_streamout_write_index
);
2705 /* Compute (streamout_write_index + thread_id). */
2706 so_write_index
= LLVMBuildAdd(builder
, so_write_index
, tid
, "");
2708 /* Load the descriptor and compute the write offset for each
2709 * enabled buffer. */
2710 LLVMValueRef so_write_offset
[4] = {};
2711 LLVMValueRef so_buffers
[4];
2712 LLVMValueRef buf_ptr
= LLVMGetParam(ctx
->main_fn
,
2713 ctx
->param_rw_buffers
);
2715 for (i
= 0; i
< 4; i
++) {
2719 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
,
2720 SI_VS_STREAMOUT_BUF0
+ i
, 0);
2722 so_buffers
[i
] = ac_build_load_to_sgpr(&ctx
->ac
, buf_ptr
, offset
);
2724 LLVMValueRef so_offset
= LLVMGetParam(ctx
->main_fn
,
2725 ctx
->param_streamout_offset
[i
]);
2726 so_offset
= LLVMBuildMul(builder
, so_offset
, LLVMConstInt(ctx
->i32
, 4, 0), "");
2728 so_write_offset
[i
] = ac_build_imad(&ctx
->ac
, so_write_index
,
2729 LLVMConstInt(ctx
->i32
, so
->stride
[i
]*4, 0),
2733 /* Write streamout data. */
2734 for (i
= 0; i
< so
->num_outputs
; i
++) {
2735 unsigned reg
= so
->output
[i
].register_index
;
2740 if (stream
!= so
->output
[i
].stream
)
2743 si_emit_streamout_output(ctx
, so_buffers
, so_write_offset
,
2744 &so
->output
[i
], &outputs
[reg
]);
2747 lp_build_endif(&if_ctx
);
2750 static void si_export_param(struct si_shader_context
*ctx
, unsigned index
,
2751 LLVMValueRef
*values
)
2753 struct ac_export_args args
;
2755 si_llvm_init_export_args(ctx
, values
,
2756 V_008DFC_SQ_EXP_PARAM
+ index
, &args
);
2757 ac_build_export(&ctx
->ac
, &args
);
2760 static void si_build_param_exports(struct si_shader_context
*ctx
,
2761 struct si_shader_output_values
*outputs
,
2764 struct si_shader
*shader
= ctx
->shader
;
2765 unsigned param_count
= 0;
2767 for (unsigned i
= 0; i
< noutput
; i
++) {
2768 unsigned semantic_name
= outputs
[i
].semantic_name
;
2769 unsigned semantic_index
= outputs
[i
].semantic_index
;
2771 if (outputs
[i
].vertex_stream
[0] != 0 &&
2772 outputs
[i
].vertex_stream
[1] != 0 &&
2773 outputs
[i
].vertex_stream
[2] != 0 &&
2774 outputs
[i
].vertex_stream
[3] != 0)
2777 switch (semantic_name
) {
2778 case TGSI_SEMANTIC_LAYER
:
2779 case TGSI_SEMANTIC_VIEWPORT_INDEX
:
2780 case TGSI_SEMANTIC_CLIPDIST
:
2781 case TGSI_SEMANTIC_COLOR
:
2782 case TGSI_SEMANTIC_BCOLOR
:
2783 case TGSI_SEMANTIC_PRIMID
:
2784 case TGSI_SEMANTIC_FOG
:
2785 case TGSI_SEMANTIC_TEXCOORD
:
2786 case TGSI_SEMANTIC_GENERIC
:
2792 if ((semantic_name
!= TGSI_SEMANTIC_GENERIC
||
2793 semantic_index
< SI_MAX_IO_GENERIC
) &&
2794 shader
->key
.opt
.kill_outputs
&
2795 (1ull << si_shader_io_get_unique_index(semantic_name
,
2796 semantic_index
, true)))
2799 si_export_param(ctx
, param_count
, outputs
[i
].values
);
2801 assert(i
< ARRAY_SIZE(shader
->info
.vs_output_param_offset
));
2802 shader
->info
.vs_output_param_offset
[i
] = param_count
++;
2805 shader
->info
.nr_param_exports
= param_count
;
2809 * Vertex color clamping.
2811 * This uses a state constant loaded in a user data SGPR and
2812 * an IF statement is added that clamps all colors if the constant
2815 static void si_vertex_color_clamping(struct si_shader_context
*ctx
,
2816 struct si_shader_output_values
*outputs
,
2819 LLVMValueRef addr
[SI_MAX_VS_OUTPUTS
][4];
2820 bool has_colors
= false;
2822 /* Store original colors to alloca variables. */
2823 for (unsigned i
= 0; i
< noutput
; i
++) {
2824 if (outputs
[i
].semantic_name
!= TGSI_SEMANTIC_COLOR
&&
2825 outputs
[i
].semantic_name
!= TGSI_SEMANTIC_BCOLOR
)
2828 for (unsigned j
= 0; j
< 4; j
++) {
2829 addr
[i
][j
] = ac_build_alloca_undef(&ctx
->ac
, ctx
->f32
, "");
2830 LLVMBuildStore(ctx
->ac
.builder
, outputs
[i
].values
[j
], addr
[i
][j
]);
2838 /* The state is in the first bit of the user SGPR. */
2839 LLVMValueRef cond
= LLVMGetParam(ctx
->main_fn
, ctx
->param_vs_state_bits
);
2840 cond
= LLVMBuildTrunc(ctx
->ac
.builder
, cond
, ctx
->i1
, "");
2842 struct lp_build_if_state if_ctx
;
2843 lp_build_if(&if_ctx
, &ctx
->gallivm
, cond
);
2845 /* Store clamped colors to alloca variables within the conditional block. */
2846 for (unsigned i
= 0; i
< noutput
; i
++) {
2847 if (outputs
[i
].semantic_name
!= TGSI_SEMANTIC_COLOR
&&
2848 outputs
[i
].semantic_name
!= TGSI_SEMANTIC_BCOLOR
)
2851 for (unsigned j
= 0; j
< 4; j
++) {
2852 LLVMBuildStore(ctx
->ac
.builder
,
2853 ac_build_clamp(&ctx
->ac
, outputs
[i
].values
[j
]),
2857 lp_build_endif(&if_ctx
);
2859 /* Load clamped colors */
2860 for (unsigned i
= 0; i
< noutput
; i
++) {
2861 if (outputs
[i
].semantic_name
!= TGSI_SEMANTIC_COLOR
&&
2862 outputs
[i
].semantic_name
!= TGSI_SEMANTIC_BCOLOR
)
2865 for (unsigned j
= 0; j
< 4; j
++) {
2866 outputs
[i
].values
[j
] =
2867 LLVMBuildLoad(ctx
->ac
.builder
, addr
[i
][j
], "");
2872 /* Generate export instructions for hardware VS shader stage or NGG GS stage
2873 * (position and parameter data only).
2875 void si_llvm_export_vs(struct si_shader_context
*ctx
,
2876 struct si_shader_output_values
*outputs
,
2879 struct si_shader
*shader
= ctx
->shader
;
2880 struct ac_export_args pos_args
[4] = {};
2881 LLVMValueRef psize_value
= NULL
, edgeflag_value
= NULL
, layer_value
= NULL
, viewport_index_value
= NULL
;
2885 si_vertex_color_clamping(ctx
, outputs
, noutput
);
2887 /* Build position exports. */
2888 for (i
= 0; i
< noutput
; i
++) {
2889 switch (outputs
[i
].semantic_name
) {
2890 case TGSI_SEMANTIC_POSITION
:
2891 si_llvm_init_export_args(ctx
, outputs
[i
].values
,
2892 V_008DFC_SQ_EXP_POS
, &pos_args
[0]);
2894 case TGSI_SEMANTIC_PSIZE
:
2895 psize_value
= outputs
[i
].values
[0];
2897 case TGSI_SEMANTIC_LAYER
:
2898 layer_value
= outputs
[i
].values
[0];
2900 case TGSI_SEMANTIC_VIEWPORT_INDEX
:
2901 viewport_index_value
= outputs
[i
].values
[0];
2903 case TGSI_SEMANTIC_EDGEFLAG
:
2904 edgeflag_value
= outputs
[i
].values
[0];
2906 case TGSI_SEMANTIC_CLIPDIST
:
2907 if (!shader
->key
.opt
.clip_disable
) {
2908 unsigned index
= 2 + outputs
[i
].semantic_index
;
2909 si_llvm_init_export_args(ctx
, outputs
[i
].values
,
2910 V_008DFC_SQ_EXP_POS
+ index
,
2914 case TGSI_SEMANTIC_CLIPVERTEX
:
2915 if (!shader
->key
.opt
.clip_disable
) {
2916 si_llvm_emit_clipvertex(ctx
, pos_args
,
2923 /* We need to add the position output manually if it's missing. */
2924 if (!pos_args
[0].out
[0]) {
2925 pos_args
[0].enabled_channels
= 0xf; /* writemask */
2926 pos_args
[0].valid_mask
= 0; /* EXEC mask */
2927 pos_args
[0].done
= 0; /* last export? */
2928 pos_args
[0].target
= V_008DFC_SQ_EXP_POS
;
2929 pos_args
[0].compr
= 0; /* COMPR flag */
2930 pos_args
[0].out
[0] = ctx
->ac
.f32_0
; /* X */
2931 pos_args
[0].out
[1] = ctx
->ac
.f32_0
; /* Y */
2932 pos_args
[0].out
[2] = ctx
->ac
.f32_0
; /* Z */
2933 pos_args
[0].out
[3] = ctx
->ac
.f32_1
; /* W */
2936 /* Write the misc vector (point size, edgeflag, layer, viewport). */
2937 if (shader
->selector
->info
.writes_psize
||
2938 shader
->selector
->info
.writes_edgeflag
||
2939 shader
->selector
->info
.writes_viewport_index
||
2940 shader
->selector
->info
.writes_layer
) {
2941 pos_args
[1].enabled_channels
= shader
->selector
->info
.writes_psize
|
2942 (shader
->selector
->info
.writes_edgeflag
<< 1) |
2943 (shader
->selector
->info
.writes_layer
<< 2);
2945 pos_args
[1].valid_mask
= 0; /* EXEC mask */
2946 pos_args
[1].done
= 0; /* last export? */
2947 pos_args
[1].target
= V_008DFC_SQ_EXP_POS
+ 1;
2948 pos_args
[1].compr
= 0; /* COMPR flag */
2949 pos_args
[1].out
[0] = ctx
->ac
.f32_0
; /* X */
2950 pos_args
[1].out
[1] = ctx
->ac
.f32_0
; /* Y */
2951 pos_args
[1].out
[2] = ctx
->ac
.f32_0
; /* Z */
2952 pos_args
[1].out
[3] = ctx
->ac
.f32_0
; /* W */
2954 if (shader
->selector
->info
.writes_psize
)
2955 pos_args
[1].out
[0] = psize_value
;
2957 if (shader
->selector
->info
.writes_edgeflag
) {
2958 /* The output is a float, but the hw expects an integer
2959 * with the first bit containing the edge flag. */
2960 edgeflag_value
= LLVMBuildFPToUI(ctx
->ac
.builder
,
2963 edgeflag_value
= ac_build_umin(&ctx
->ac
,
2967 /* The LLVM intrinsic expects a float. */
2968 pos_args
[1].out
[1] = ac_to_float(&ctx
->ac
, edgeflag_value
);
2971 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
2972 /* GFX9 has the layer in out.z[10:0] and the viewport
2973 * index in out.z[19:16].
2975 if (shader
->selector
->info
.writes_layer
)
2976 pos_args
[1].out
[2] = layer_value
;
2978 if (shader
->selector
->info
.writes_viewport_index
) {
2979 LLVMValueRef v
= viewport_index_value
;
2981 v
= ac_to_integer(&ctx
->ac
, v
);
2982 v
= LLVMBuildShl(ctx
->ac
.builder
, v
,
2983 LLVMConstInt(ctx
->i32
, 16, 0), "");
2984 v
= LLVMBuildOr(ctx
->ac
.builder
, v
,
2985 ac_to_integer(&ctx
->ac
, pos_args
[1].out
[2]), "");
2986 pos_args
[1].out
[2] = ac_to_float(&ctx
->ac
, v
);
2987 pos_args
[1].enabled_channels
|= 1 << 2;
2990 if (shader
->selector
->info
.writes_layer
)
2991 pos_args
[1].out
[2] = layer_value
;
2993 if (shader
->selector
->info
.writes_viewport_index
) {
2994 pos_args
[1].out
[3] = viewport_index_value
;
2995 pos_args
[1].enabled_channels
|= 1 << 3;
3000 for (i
= 0; i
< 4; i
++)
3001 if (pos_args
[i
].out
[0])
3002 shader
->info
.nr_pos_exports
++;
3005 for (i
= 0; i
< 4; i
++) {
3006 if (!pos_args
[i
].out
[0])
3009 /* Specify the target we are exporting */
3010 pos_args
[i
].target
= V_008DFC_SQ_EXP_POS
+ pos_idx
++;
3012 if (pos_idx
== shader
->info
.nr_pos_exports
)
3013 /* Specify that this is the last export */
3014 pos_args
[i
].done
= 1;
3016 ac_build_export(&ctx
->ac
, &pos_args
[i
]);
3019 /* Build parameter exports. */
3020 si_build_param_exports(ctx
, outputs
, noutput
);
3024 * Forward all outputs from the vertex shader to the TES. This is only used
3025 * for the fixed function TCS.
3027 static void si_copy_tcs_inputs(struct lp_build_tgsi_context
*bld_base
)
3029 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3030 LLVMValueRef invocation_id
, buffer
, buffer_offset
;
3031 LLVMValueRef lds_vertex_stride
, lds_base
;
3034 invocation_id
= unpack_llvm_param(ctx
, ctx
->abi
.tcs_rel_ids
, 8, 5);
3035 buffer
= get_tess_ring_descriptor(ctx
, TESS_OFFCHIP_RING_TCS
);
3036 buffer_offset
= LLVMGetParam(ctx
->main_fn
, ctx
->param_tcs_offchip_offset
);
3038 lds_vertex_stride
= get_tcs_in_vertex_dw_stride(ctx
);
3039 lds_base
= get_tcs_in_current_patch_offset(ctx
);
3040 lds_base
= ac_build_imad(&ctx
->ac
, invocation_id
, lds_vertex_stride
,
3043 inputs
= ctx
->shader
->key
.mono
.u
.ff_tcs_inputs_to_copy
;
3045 unsigned i
= u_bit_scan64(&inputs
);
3047 LLVMValueRef lds_ptr
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
3048 LLVMConstInt(ctx
->i32
, 4 * i
, 0),
3051 LLVMValueRef buffer_addr
= get_tcs_tes_buffer_address(ctx
,
3052 get_rel_patch_id(ctx
),
3054 LLVMConstInt(ctx
->i32
, i
, 0));
3056 LLVMValueRef value
= lshs_lds_load(bld_base
, ctx
->ac
.i32
, ~0, lds_ptr
);
3058 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, value
, 4, buffer_addr
,
3059 buffer_offset
, 0, 1, 0, false);
3063 static void si_write_tess_factors(struct lp_build_tgsi_context
*bld_base
,
3064 LLVMValueRef rel_patch_id
,
3065 LLVMValueRef invocation_id
,
3066 LLVMValueRef tcs_out_current_patch_data_offset
,
3067 LLVMValueRef invoc0_tf_outer
[4],
3068 LLVMValueRef invoc0_tf_inner
[2])
3070 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3071 struct si_shader
*shader
= ctx
->shader
;
3072 unsigned tess_inner_index
, tess_outer_index
;
3073 LLVMValueRef lds_base
, lds_inner
, lds_outer
, byteoffset
, buffer
;
3074 LLVMValueRef out
[6], vec0
, vec1
, tf_base
, inner
[4], outer
[4];
3075 unsigned stride
, outer_comps
, inner_comps
, i
, offset
;
3076 struct lp_build_if_state if_ctx
, inner_if_ctx
;
3078 /* Add a barrier before loading tess factors from LDS. */
3079 if (!shader
->key
.part
.tcs
.epilog
.invoc0_tess_factors_are_def
)
3080 si_llvm_emit_barrier(NULL
, bld_base
, NULL
);
3082 /* Do this only for invocation 0, because the tess levels are per-patch,
3085 * This can't jump, because invocation 0 executes this. It should
3086 * at least mask out the loads and stores for other invocations.
3088 lp_build_if(&if_ctx
, &ctx
->gallivm
,
3089 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
,
3090 invocation_id
, ctx
->i32_0
, ""));
3092 /* Determine the layout of one tess factor element in the buffer. */
3093 switch (shader
->key
.part
.tcs
.epilog
.prim_mode
) {
3094 case PIPE_PRIM_LINES
:
3095 stride
= 2; /* 2 dwords, 1 vec2 store */
3099 case PIPE_PRIM_TRIANGLES
:
3100 stride
= 4; /* 4 dwords, 1 vec4 store */
3104 case PIPE_PRIM_QUADS
:
3105 stride
= 6; /* 6 dwords, 2 stores (vec4 + vec2) */
3114 for (i
= 0; i
< 4; i
++) {
3115 inner
[i
] = LLVMGetUndef(ctx
->i32
);
3116 outer
[i
] = LLVMGetUndef(ctx
->i32
);
3119 if (shader
->key
.part
.tcs
.epilog
.invoc0_tess_factors_are_def
) {
3120 /* Tess factors are in VGPRs. */
3121 for (i
= 0; i
< outer_comps
; i
++)
3122 outer
[i
] = out
[i
] = invoc0_tf_outer
[i
];
3123 for (i
= 0; i
< inner_comps
; i
++)
3124 inner
[i
] = out
[outer_comps
+i
] = invoc0_tf_inner
[i
];
3126 /* Load tess_inner and tess_outer from LDS.
3127 * Any invocation can write them, so we can't get them from a temporary.
3129 tess_inner_index
= si_shader_io_get_unique_index_patch(TGSI_SEMANTIC_TESSINNER
, 0);
3130 tess_outer_index
= si_shader_io_get_unique_index_patch(TGSI_SEMANTIC_TESSOUTER
, 0);
3132 lds_base
= tcs_out_current_patch_data_offset
;
3133 lds_inner
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
3134 LLVMConstInt(ctx
->i32
,
3135 tess_inner_index
* 4, 0), "");
3136 lds_outer
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
3137 LLVMConstInt(ctx
->i32
,
3138 tess_outer_index
* 4, 0), "");
3140 for (i
= 0; i
< outer_comps
; i
++) {
3142 lshs_lds_load(bld_base
, ctx
->ac
.i32
, i
, lds_outer
);
3144 for (i
= 0; i
< inner_comps
; i
++) {
3145 inner
[i
] = out
[outer_comps
+i
] =
3146 lshs_lds_load(bld_base
, ctx
->ac
.i32
, i
, lds_inner
);
3150 if (shader
->key
.part
.tcs
.epilog
.prim_mode
== PIPE_PRIM_LINES
) {
3151 /* For isolines, the hardware expects tess factors in the
3152 * reverse order from what GLSL / TGSI specify.
3154 LLVMValueRef tmp
= out
[0];
3159 /* Convert the outputs to vectors for stores. */
3160 vec0
= ac_build_gather_values(&ctx
->ac
, out
, MIN2(stride
, 4));
3164 vec1
= ac_build_gather_values(&ctx
->ac
, out
+4, stride
- 4);
3166 /* Get the buffer. */
3167 buffer
= get_tess_ring_descriptor(ctx
, TCS_FACTOR_RING
);
3169 /* Get the offset. */
3170 tf_base
= LLVMGetParam(ctx
->main_fn
,
3171 ctx
->param_tcs_factor_offset
);
3172 byteoffset
= LLVMBuildMul(ctx
->ac
.builder
, rel_patch_id
,
3173 LLVMConstInt(ctx
->i32
, 4 * stride
, 0), "");
3175 lp_build_if(&inner_if_ctx
, &ctx
->gallivm
,
3176 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
,
3177 rel_patch_id
, ctx
->i32_0
, ""));
3179 /* Store the dynamic HS control word. */
3181 if (ctx
->screen
->info
.chip_class
<= GFX8
) {
3182 ac_build_buffer_store_dword(&ctx
->ac
, buffer
,
3183 LLVMConstInt(ctx
->i32
, 0x80000000, 0),
3184 1, ctx
->i32_0
, tf_base
,
3185 offset
, 1, 0, false);
3189 lp_build_endif(&inner_if_ctx
);
3191 /* Store the tessellation factors. */
3192 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec0
,
3193 MIN2(stride
, 4), byteoffset
, tf_base
,
3194 offset
, 1, 0, false);
3197 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec1
,
3198 stride
- 4, byteoffset
, tf_base
,
3199 offset
, 1, 0, false);
3201 /* Store the tess factors into the offchip buffer if TES reads them. */
3202 if (shader
->key
.part
.tcs
.epilog
.tes_reads_tess_factors
) {
3203 LLVMValueRef buf
, base
, inner_vec
, outer_vec
, tf_outer_offset
;
3204 LLVMValueRef tf_inner_offset
;
3205 unsigned param_outer
, param_inner
;
3207 buf
= get_tess_ring_descriptor(ctx
, TESS_OFFCHIP_RING_TCS
);
3208 base
= LLVMGetParam(ctx
->main_fn
, ctx
->param_tcs_offchip_offset
);
3210 param_outer
= si_shader_io_get_unique_index_patch(
3211 TGSI_SEMANTIC_TESSOUTER
, 0);
3212 tf_outer_offset
= get_tcs_tes_buffer_address(ctx
, rel_patch_id
, NULL
,
3213 LLVMConstInt(ctx
->i32
, param_outer
, 0));
3215 unsigned outer_vec_size
=
3216 ac_has_vec3_support(ctx
->screen
->info
.chip_class
, false) ?
3217 outer_comps
: util_next_power_of_two(outer_comps
);
3218 outer_vec
= ac_build_gather_values(&ctx
->ac
, outer
, outer_vec_size
);
3220 ac_build_buffer_store_dword(&ctx
->ac
, buf
, outer_vec
,
3221 outer_comps
, tf_outer_offset
,
3222 base
, 0, 1, 0, false);
3224 param_inner
= si_shader_io_get_unique_index_patch(
3225 TGSI_SEMANTIC_TESSINNER
, 0);
3226 tf_inner_offset
= get_tcs_tes_buffer_address(ctx
, rel_patch_id
, NULL
,
3227 LLVMConstInt(ctx
->i32
, param_inner
, 0));
3229 inner_vec
= inner_comps
== 1 ? inner
[0] :
3230 ac_build_gather_values(&ctx
->ac
, inner
, inner_comps
);
3231 ac_build_buffer_store_dword(&ctx
->ac
, buf
, inner_vec
,
3232 inner_comps
, tf_inner_offset
,
3233 base
, 0, 1, 0, false);
3237 lp_build_endif(&if_ctx
);
3241 si_insert_input_ret(struct si_shader_context
*ctx
, LLVMValueRef ret
,
3242 unsigned param
, unsigned return_index
)
3244 return LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
3245 LLVMGetParam(ctx
->main_fn
, param
),
3250 si_insert_input_ret_float(struct si_shader_context
*ctx
, LLVMValueRef ret
,
3251 unsigned param
, unsigned return_index
)
3253 LLVMBuilderRef builder
= ctx
->ac
.builder
;
3254 LLVMValueRef p
= LLVMGetParam(ctx
->main_fn
, param
);
3256 return LLVMBuildInsertValue(builder
, ret
,
3257 ac_to_float(&ctx
->ac
, p
),
3262 si_insert_input_ptr(struct si_shader_context
*ctx
, LLVMValueRef ret
,
3263 unsigned param
, unsigned return_index
)
3265 LLVMBuilderRef builder
= ctx
->ac
.builder
;
3266 LLVMValueRef ptr
= LLVMGetParam(ctx
->main_fn
, param
);
3267 ptr
= LLVMBuildPtrToInt(builder
, ptr
, ctx
->i32
, "");
3268 return LLVMBuildInsertValue(builder
, ret
, ptr
, return_index
, "");
3271 /* This only writes the tessellation factor levels. */
3272 static void si_llvm_emit_tcs_epilogue(struct ac_shader_abi
*abi
,
3273 unsigned max_outputs
,
3274 LLVMValueRef
*addrs
)
3276 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
3277 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
3278 LLVMBuilderRef builder
= ctx
->ac
.builder
;
3279 LLVMValueRef rel_patch_id
, invocation_id
, tf_lds_offset
;
3281 si_copy_tcs_inputs(bld_base
);
3283 rel_patch_id
= get_rel_patch_id(ctx
);
3284 invocation_id
= unpack_llvm_param(ctx
, ctx
->abi
.tcs_rel_ids
, 8, 5);
3285 tf_lds_offset
= get_tcs_out_current_patch_data_offset(ctx
);
3287 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
3288 LLVMBasicBlockRef blocks
[2] = {
3289 LLVMGetInsertBlock(builder
),
3290 ctx
->merged_wrap_if_state
.entry_block
3292 LLVMValueRef values
[2];
3294 lp_build_endif(&ctx
->merged_wrap_if_state
);
3296 values
[0] = rel_patch_id
;
3297 values
[1] = LLVMGetUndef(ctx
->i32
);
3298 rel_patch_id
= ac_build_phi(&ctx
->ac
, ctx
->i32
, 2, values
, blocks
);
3300 values
[0] = tf_lds_offset
;
3301 values
[1] = LLVMGetUndef(ctx
->i32
);
3302 tf_lds_offset
= ac_build_phi(&ctx
->ac
, ctx
->i32
, 2, values
, blocks
);
3304 values
[0] = invocation_id
;
3305 values
[1] = ctx
->i32_1
; /* cause the epilog to skip threads */
3306 invocation_id
= ac_build_phi(&ctx
->ac
, ctx
->i32
, 2, values
, blocks
);
3309 /* Return epilog parameters from this function. */
3310 LLVMValueRef ret
= ctx
->return_value
;
3313 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
3314 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_layout
,
3315 8 + GFX9_SGPR_TCS_OFFCHIP_LAYOUT
);
3316 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_out_lds_layout
,
3317 8 + GFX9_SGPR_TCS_OUT_LAYOUT
);
3318 /* Tess offchip and tess factor offsets are at the beginning. */
3319 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_offset
, 2);
3320 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_factor_offset
, 4);
3321 vgpr
= 8 + GFX9_SGPR_TCS_OUT_LAYOUT
+ 1;
3323 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_layout
,
3324 GFX6_SGPR_TCS_OFFCHIP_LAYOUT
);
3325 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_out_lds_layout
,
3326 GFX6_SGPR_TCS_OUT_LAYOUT
);
3327 /* Tess offchip and tess factor offsets are after user SGPRs. */
3328 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_offset
,
3329 GFX6_TCS_NUM_USER_SGPR
);
3330 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_factor_offset
,
3331 GFX6_TCS_NUM_USER_SGPR
+ 1);
3332 vgpr
= GFX6_TCS_NUM_USER_SGPR
+ 2;
3336 rel_patch_id
= ac_to_float(&ctx
->ac
, rel_patch_id
);
3337 invocation_id
= ac_to_float(&ctx
->ac
, invocation_id
);
3338 tf_lds_offset
= ac_to_float(&ctx
->ac
, tf_lds_offset
);
3340 /* Leave a hole corresponding to the two input VGPRs. This ensures that
3341 * the invocation_id output does not alias the tcs_rel_ids input,
3342 * which saves a V_MOV on gfx9.
3346 ret
= LLVMBuildInsertValue(builder
, ret
, rel_patch_id
, vgpr
++, "");
3347 ret
= LLVMBuildInsertValue(builder
, ret
, invocation_id
, vgpr
++, "");
3349 if (ctx
->shader
->selector
->tcs_info
.tessfactors_are_def_in_all_invocs
) {
3350 vgpr
++; /* skip the tess factor LDS offset */
3351 for (unsigned i
= 0; i
< 6; i
++) {
3352 LLVMValueRef value
=
3353 LLVMBuildLoad(builder
, ctx
->invoc0_tess_factors
[i
], "");
3354 value
= ac_to_float(&ctx
->ac
, value
);
3355 ret
= LLVMBuildInsertValue(builder
, ret
, value
, vgpr
++, "");
3358 ret
= LLVMBuildInsertValue(builder
, ret
, tf_lds_offset
, vgpr
++, "");
3360 ctx
->return_value
= ret
;
3363 /* Pass TCS inputs from LS to TCS on GFX9. */
3364 static void si_set_ls_return_value_for_tcs(struct si_shader_context
*ctx
)
3366 LLVMValueRef ret
= ctx
->return_value
;
3368 ret
= si_insert_input_ptr(ctx
, ret
, 0, 0);
3369 ret
= si_insert_input_ptr(ctx
, ret
, 1, 1);
3370 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_offset
, 2);
3371 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_merged_wave_info
, 3);
3372 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_factor_offset
, 4);
3373 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_merged_scratch_offset
, 5);
3375 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->param_rw_buffers
,
3376 8 + SI_SGPR_RW_BUFFERS
);
3377 ret
= si_insert_input_ptr(ctx
, ret
,
3378 ctx
->param_bindless_samplers_and_images
,
3379 8 + SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES
);
3381 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_vs_state_bits
,
3382 8 + SI_SGPR_VS_STATE_BITS
);
3384 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_layout
,
3385 8 + GFX9_SGPR_TCS_OFFCHIP_LAYOUT
);
3386 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_out_lds_offsets
,
3387 8 + GFX9_SGPR_TCS_OUT_OFFSETS
);
3388 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_out_lds_layout
,
3389 8 + GFX9_SGPR_TCS_OUT_LAYOUT
);
3391 unsigned vgpr
= 8 + GFX9_TCS_NUM_USER_SGPR
;
3392 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
3393 ac_to_float(&ctx
->ac
, ctx
->abi
.tcs_patch_id
),
3395 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
3396 ac_to_float(&ctx
->ac
, ctx
->abi
.tcs_rel_ids
),
3398 ctx
->return_value
= ret
;
3401 /* Pass GS inputs from ES to GS on GFX9. */
3402 static void si_set_es_return_value_for_gs(struct si_shader_context
*ctx
)
3404 LLVMBuilderRef builder
= ctx
->ac
.builder
;
3405 LLVMValueRef ret
= ctx
->return_value
;
3407 ret
= si_insert_input_ptr(ctx
, ret
, 0, 0);
3408 ret
= si_insert_input_ptr(ctx
, ret
, 1, 1);
3409 if (ctx
->shader
->key
.as_ngg
)
3410 ret
= LLVMBuildInsertValue(builder
, ret
, ctx
->gs_tg_info
, 2, "");
3412 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_gs2vs_offset
, 2);
3413 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_merged_wave_info
, 3);
3414 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_merged_scratch_offset
, 5);
3416 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->param_rw_buffers
,
3417 8 + SI_SGPR_RW_BUFFERS
);
3418 ret
= si_insert_input_ptr(ctx
, ret
,
3419 ctx
->param_bindless_samplers_and_images
,
3420 8 + SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES
);
3423 if (ctx
->type
== PIPE_SHADER_VERTEX
)
3424 vgpr
= 8 + GFX9_VSGS_NUM_USER_SGPR
;
3426 vgpr
= 8 + GFX9_TESGS_NUM_USER_SGPR
;
3428 for (unsigned i
= 0; i
< 5; i
++) {
3429 unsigned param
= ctx
->param_gs_vtx01_offset
+ i
;
3430 ret
= si_insert_input_ret_float(ctx
, ret
, param
, vgpr
++);
3432 ctx
->return_value
= ret
;
3435 static void si_llvm_emit_ls_epilogue(struct ac_shader_abi
*abi
,
3436 unsigned max_outputs
,
3437 LLVMValueRef
*addrs
)
3439 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
3440 struct si_shader
*shader
= ctx
->shader
;
3441 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
3443 LLVMValueRef vertex_id
= LLVMGetParam(ctx
->main_fn
,
3444 ctx
->param_rel_auto_id
);
3445 LLVMValueRef vertex_dw_stride
= get_tcs_in_vertex_dw_stride(ctx
);
3446 LLVMValueRef base_dw_addr
= LLVMBuildMul(ctx
->ac
.builder
, vertex_id
,
3447 vertex_dw_stride
, "");
3449 /* Write outputs to LDS. The next shader (TCS aka HS) will read
3450 * its inputs from it. */
3451 for (i
= 0; i
< info
->num_outputs
; i
++) {
3452 unsigned name
= info
->output_semantic_name
[i
];
3453 unsigned index
= info
->output_semantic_index
[i
];
3455 /* The ARB_shader_viewport_layer_array spec contains the
3458 * 2) What happens if gl_ViewportIndex or gl_Layer is
3459 * written in the vertex shader and a geometry shader is
3462 * RESOLVED: The value written by the last vertex processing
3463 * stage is used. If the last vertex processing stage
3464 * (vertex, tessellation evaluation or geometry) does not
3465 * statically assign to gl_ViewportIndex or gl_Layer, index
3466 * or layer zero is assumed.
3468 * So writes to those outputs in VS-as-LS are simply ignored.
3470 if (name
== TGSI_SEMANTIC_LAYER
||
3471 name
== TGSI_SEMANTIC_VIEWPORT_INDEX
)
3474 int param
= si_shader_io_get_unique_index(name
, index
, false);
3475 LLVMValueRef dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_dw_addr
,
3476 LLVMConstInt(ctx
->i32
, param
* 4, 0), "");
3478 for (chan
= 0; chan
< 4; chan
++) {
3479 if (!(info
->output_usagemask
[i
] & (1 << chan
)))
3482 lshs_lds_store(ctx
, chan
, dw_addr
,
3483 LLVMBuildLoad(ctx
->ac
.builder
, addrs
[4 * i
+ chan
], ""));
3487 if (ctx
->screen
->info
.chip_class
>= GFX9
)
3488 si_set_ls_return_value_for_tcs(ctx
);
3491 static void si_llvm_emit_es_epilogue(struct ac_shader_abi
*abi
,
3492 unsigned max_outputs
,
3493 LLVMValueRef
*addrs
)
3495 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
3496 struct si_shader
*es
= ctx
->shader
;
3497 struct tgsi_shader_info
*info
= &es
->selector
->info
;
3498 LLVMValueRef soffset
= LLVMGetParam(ctx
->main_fn
,
3499 ctx
->param_es2gs_offset
);
3500 LLVMValueRef lds_base
= NULL
;
3504 if (ctx
->screen
->info
.chip_class
>= GFX9
&& info
->num_outputs
) {
3505 unsigned itemsize_dw
= es
->selector
->esgs_itemsize
/ 4;
3506 LLVMValueRef vertex_idx
= ac_get_thread_id(&ctx
->ac
);
3507 LLVMValueRef wave_idx
= si_unpack_param(ctx
, ctx
->param_merged_wave_info
, 24, 4);
3508 vertex_idx
= LLVMBuildOr(ctx
->ac
.builder
, vertex_idx
,
3509 LLVMBuildMul(ctx
->ac
.builder
, wave_idx
,
3510 LLVMConstInt(ctx
->i32
, 64, false), ""), "");
3511 lds_base
= LLVMBuildMul(ctx
->ac
.builder
, vertex_idx
,
3512 LLVMConstInt(ctx
->i32
, itemsize_dw
, 0), "");
3515 for (i
= 0; i
< info
->num_outputs
; i
++) {
3518 if (info
->output_semantic_name
[i
] == TGSI_SEMANTIC_VIEWPORT_INDEX
||
3519 info
->output_semantic_name
[i
] == TGSI_SEMANTIC_LAYER
)
3522 param
= si_shader_io_get_unique_index(info
->output_semantic_name
[i
],
3523 info
->output_semantic_index
[i
], false);
3525 for (chan
= 0; chan
< 4; chan
++) {
3526 if (!(info
->output_usagemask
[i
] & (1 << chan
)))
3529 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->ac
.builder
, addrs
[4 * i
+ chan
], "");
3530 out_val
= ac_to_integer(&ctx
->ac
, out_val
);
3532 /* GFX9 has the ESGS ring in LDS. */
3533 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
3534 LLVMValueRef idx
= LLVMConstInt(ctx
->i32
, param
* 4 + chan
, false);
3535 idx
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
, idx
, "");
3536 ac_build_indexed_store(&ctx
->ac
, ctx
->esgs_ring
, idx
, out_val
);
3540 ac_build_buffer_store_dword(&ctx
->ac
,
3542 out_val
, 1, NULL
, soffset
,
3543 (4 * param
+ chan
) * 4,
3548 if (ctx
->screen
->info
.chip_class
>= GFX9
)
3549 si_set_es_return_value_for_gs(ctx
);
3552 static LLVMValueRef
si_get_gs_wave_id(struct si_shader_context
*ctx
)
3554 if (ctx
->screen
->info
.chip_class
>= GFX9
)
3555 return si_unpack_param(ctx
, ctx
->param_merged_wave_info
, 16, 8);
3557 return LLVMGetParam(ctx
->main_fn
, ctx
->param_gs_wave_id
);
3560 static void emit_gs_epilogue(struct si_shader_context
*ctx
)
3562 if (ctx
->shader
->key
.as_ngg
) {
3563 gfx10_ngg_gs_emit_epilogue(ctx
);
3567 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_NOP
| AC_SENDMSG_GS_DONE
,
3568 si_get_gs_wave_id(ctx
));
3570 if (ctx
->screen
->info
.chip_class
>= GFX9
)
3571 lp_build_endif(&ctx
->merged_wrap_if_state
);
3574 static void si_llvm_emit_gs_epilogue(struct ac_shader_abi
*abi
,
3575 unsigned max_outputs
,
3576 LLVMValueRef
*addrs
)
3578 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
3579 struct tgsi_shader_info UNUSED
*info
= &ctx
->shader
->selector
->info
;
3581 assert(info
->num_outputs
<= max_outputs
);
3583 emit_gs_epilogue(ctx
);
3586 static void si_tgsi_emit_gs_epilogue(struct lp_build_tgsi_context
*bld_base
)
3588 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3589 emit_gs_epilogue(ctx
);
3592 static void si_llvm_emit_vs_epilogue(struct ac_shader_abi
*abi
,
3593 unsigned max_outputs
,
3594 LLVMValueRef
*addrs
)
3596 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
3597 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
3598 struct si_shader_output_values
*outputs
= NULL
;
3601 assert(!ctx
->shader
->is_gs_copy_shader
);
3602 assert(info
->num_outputs
<= max_outputs
);
3604 outputs
= MALLOC((info
->num_outputs
+ 1) * sizeof(outputs
[0]));
3606 for (i
= 0; i
< info
->num_outputs
; i
++) {
3607 outputs
[i
].semantic_name
= info
->output_semantic_name
[i
];
3608 outputs
[i
].semantic_index
= info
->output_semantic_index
[i
];
3610 for (j
= 0; j
< 4; j
++) {
3611 outputs
[i
].values
[j
] =
3612 LLVMBuildLoad(ctx
->ac
.builder
,
3615 outputs
[i
].vertex_stream
[j
] =
3616 (info
->output_streams
[i
] >> (2 * j
)) & 3;
3620 if (ctx
->shader
->selector
->so
.num_outputs
)
3621 si_llvm_emit_streamout(ctx
, outputs
, i
, 0);
3623 /* Export PrimitiveID. */
3624 if (ctx
->shader
->key
.mono
.u
.vs_export_prim_id
) {
3625 outputs
[i
].semantic_name
= TGSI_SEMANTIC_PRIMID
;
3626 outputs
[i
].semantic_index
= 0;
3627 outputs
[i
].values
[0] = ac_to_float(&ctx
->ac
, si_get_primitive_id(ctx
, 0));
3628 for (j
= 1; j
< 4; j
++)
3629 outputs
[i
].values
[j
] = LLVMConstReal(ctx
->f32
, 0);
3631 memset(outputs
[i
].vertex_stream
, 0,
3632 sizeof(outputs
[i
].vertex_stream
));
3636 si_llvm_export_vs(ctx
, outputs
, i
);
3640 static void si_llvm_emit_prim_discard_cs_epilogue(struct ac_shader_abi
*abi
,
3641 unsigned max_outputs
,
3642 LLVMValueRef
*addrs
)
3644 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
3645 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
3646 LLVMValueRef pos
[4] = {};
3648 assert(info
->num_outputs
<= max_outputs
);
3650 for (unsigned i
= 0; i
< info
->num_outputs
; i
++) {
3651 if (info
->output_semantic_name
[i
] != TGSI_SEMANTIC_POSITION
)
3654 for (unsigned chan
= 0; chan
< 4; chan
++)
3655 pos
[chan
] = LLVMBuildLoad(ctx
->ac
.builder
, addrs
[4 * i
+ chan
], "");
3658 assert(pos
[0] != NULL
);
3660 /* Return the position output. */
3661 LLVMValueRef ret
= ctx
->return_value
;
3662 for (unsigned chan
= 0; chan
< 4; chan
++)
3663 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, pos
[chan
], chan
, "");
3664 ctx
->return_value
= ret
;
3667 static void si_tgsi_emit_epilogue(struct lp_build_tgsi_context
*bld_base
)
3669 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3671 ctx
->abi
.emit_outputs(&ctx
->abi
, RADEON_LLVM_MAX_OUTPUTS
,
3672 &ctx
->outputs
[0][0]);
3675 struct si_ps_exports
{
3677 struct ac_export_args args
[10];
3680 static void si_export_mrt_z(struct lp_build_tgsi_context
*bld_base
,
3681 LLVMValueRef depth
, LLVMValueRef stencil
,
3682 LLVMValueRef samplemask
, struct si_ps_exports
*exp
)
3684 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3685 struct ac_export_args args
;
3687 ac_export_mrt_z(&ctx
->ac
, depth
, stencil
, samplemask
, &args
);
3689 memcpy(&exp
->args
[exp
->num
++], &args
, sizeof(args
));
3692 static void si_export_mrt_color(struct lp_build_tgsi_context
*bld_base
,
3693 LLVMValueRef
*color
, unsigned index
,
3694 unsigned samplemask_param
,
3695 bool is_last
, struct si_ps_exports
*exp
)
3697 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3701 if (ctx
->shader
->key
.part
.ps
.epilog
.clamp_color
)
3702 for (i
= 0; i
< 4; i
++)
3703 color
[i
] = ac_build_clamp(&ctx
->ac
, color
[i
]);
3706 if (ctx
->shader
->key
.part
.ps
.epilog
.alpha_to_one
)
3707 color
[3] = ctx
->ac
.f32_1
;
3711 ctx
->shader
->key
.part
.ps
.epilog
.alpha_func
!= PIPE_FUNC_ALWAYS
)
3712 si_alpha_test(bld_base
, color
[3]);
3714 /* Line & polygon smoothing */
3715 if (ctx
->shader
->key
.part
.ps
.epilog
.poly_line_smoothing
)
3716 color
[3] = si_scale_alpha_by_sample_mask(bld_base
, color
[3],
3719 /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
3720 if (ctx
->shader
->key
.part
.ps
.epilog
.last_cbuf
> 0) {
3721 struct ac_export_args args
[8];
3724 /* Get the export arguments, also find out what the last one is. */
3725 for (c
= 0; c
<= ctx
->shader
->key
.part
.ps
.epilog
.last_cbuf
; c
++) {
3726 si_llvm_init_export_args(ctx
, color
,
3727 V_008DFC_SQ_EXP_MRT
+ c
, &args
[c
]);
3728 if (args
[c
].enabled_channels
)
3732 /* Emit all exports. */
3733 for (c
= 0; c
<= ctx
->shader
->key
.part
.ps
.epilog
.last_cbuf
; c
++) {
3734 if (is_last
&& last
== c
) {
3735 args
[c
].valid_mask
= 1; /* whether the EXEC mask is valid */
3736 args
[c
].done
= 1; /* DONE bit */
3737 } else if (!args
[c
].enabled_channels
)
3738 continue; /* unnecessary NULL export */
3740 memcpy(&exp
->args
[exp
->num
++], &args
[c
], sizeof(args
[c
]));
3743 struct ac_export_args args
;
3746 si_llvm_init_export_args(ctx
, color
, V_008DFC_SQ_EXP_MRT
+ index
,
3749 args
.valid_mask
= 1; /* whether the EXEC mask is valid */
3750 args
.done
= 1; /* DONE bit */
3751 } else if (!args
.enabled_channels
)
3752 return; /* unnecessary NULL export */
3754 memcpy(&exp
->args
[exp
->num
++], &args
, sizeof(args
));
3758 static void si_emit_ps_exports(struct si_shader_context
*ctx
,
3759 struct si_ps_exports
*exp
)
3761 for (unsigned i
= 0; i
< exp
->num
; i
++)
3762 ac_build_export(&ctx
->ac
, &exp
->args
[i
]);
3766 * Return PS outputs in this order:
3768 * v[0:3] = color0.xyzw
3769 * v[4:7] = color1.xyzw
3774 * vN+3 = SampleMaskIn (used for OpenGL smoothing)
3776 * The alpha-ref SGPR is returned via its original location.
3778 static void si_llvm_return_fs_outputs(struct ac_shader_abi
*abi
,
3779 unsigned max_outputs
,
3780 LLVMValueRef
*addrs
)
3782 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
3783 struct si_shader
*shader
= ctx
->shader
;
3784 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
3785 LLVMBuilderRef builder
= ctx
->ac
.builder
;
3786 unsigned i
, j
, first_vgpr
, vgpr
;
3788 LLVMValueRef color
[8][4] = {};
3789 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
3792 if (ctx
->postponed_kill
)
3793 ac_build_kill_if_false(&ctx
->ac
, LLVMBuildLoad(builder
, ctx
->postponed_kill
, ""));
3795 /* Read the output values. */
3796 for (i
= 0; i
< info
->num_outputs
; i
++) {
3797 unsigned semantic_name
= info
->output_semantic_name
[i
];
3798 unsigned semantic_index
= info
->output_semantic_index
[i
];
3800 switch (semantic_name
) {
3801 case TGSI_SEMANTIC_COLOR
:
3802 assert(semantic_index
< 8);
3803 for (j
= 0; j
< 4; j
++) {
3804 LLVMValueRef ptr
= addrs
[4 * i
+ j
];
3805 LLVMValueRef result
= LLVMBuildLoad(builder
, ptr
, "");
3806 color
[semantic_index
][j
] = result
;
3809 case TGSI_SEMANTIC_POSITION
:
3810 depth
= LLVMBuildLoad(builder
,
3811 addrs
[4 * i
+ 2], "");
3813 case TGSI_SEMANTIC_STENCIL
:
3814 stencil
= LLVMBuildLoad(builder
,
3815 addrs
[4 * i
+ 1], "");
3817 case TGSI_SEMANTIC_SAMPLEMASK
:
3818 samplemask
= LLVMBuildLoad(builder
,
3819 addrs
[4 * i
+ 0], "");
3822 fprintf(stderr
, "Warning: GFX6 unhandled fs output type:%d\n",
3827 /* Fill the return structure. */
3828 ret
= ctx
->return_value
;
3831 ret
= LLVMBuildInsertValue(builder
, ret
,
3832 ac_to_integer(&ctx
->ac
,
3833 LLVMGetParam(ctx
->main_fn
,
3834 SI_PARAM_ALPHA_REF
)),
3835 SI_SGPR_ALPHA_REF
, "");
3838 first_vgpr
= vgpr
= SI_SGPR_ALPHA_REF
+ 1;
3839 for (i
= 0; i
< ARRAY_SIZE(color
); i
++) {
3843 for (j
= 0; j
< 4; j
++)
3844 ret
= LLVMBuildInsertValue(builder
, ret
, color
[i
][j
], vgpr
++, "");
3847 ret
= LLVMBuildInsertValue(builder
, ret
, depth
, vgpr
++, "");
3849 ret
= LLVMBuildInsertValue(builder
, ret
, stencil
, vgpr
++, "");
3851 ret
= LLVMBuildInsertValue(builder
, ret
, samplemask
, vgpr
++, "");
3853 /* Add the input sample mask for smoothing at the end. */
3854 if (vgpr
< first_vgpr
+ PS_EPILOG_SAMPLEMASK_MIN_LOC
)
3855 vgpr
= first_vgpr
+ PS_EPILOG_SAMPLEMASK_MIN_LOC
;
3856 ret
= LLVMBuildInsertValue(builder
, ret
,
3857 LLVMGetParam(ctx
->main_fn
,
3858 SI_PARAM_SAMPLE_COVERAGE
), vgpr
++, "");
3860 ctx
->return_value
= ret
;
3863 static void membar_emit(
3864 const struct lp_build_tgsi_action
*action
,
3865 struct lp_build_tgsi_context
*bld_base
,
3866 struct lp_build_emit_data
*emit_data
)
3868 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3869 LLVMValueRef src0
= lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, 0);
3870 unsigned flags
= LLVMConstIntGetZExtValue(src0
);
3871 unsigned waitcnt
= NOOP_WAITCNT
;
3873 if (flags
& TGSI_MEMBAR_THREAD_GROUP
)
3874 waitcnt
&= VM_CNT
& LGKM_CNT
;
3876 if (flags
& (TGSI_MEMBAR_ATOMIC_BUFFER
|
3877 TGSI_MEMBAR_SHADER_BUFFER
|
3878 TGSI_MEMBAR_SHADER_IMAGE
))
3881 if (flags
& TGSI_MEMBAR_SHARED
)
3882 waitcnt
&= LGKM_CNT
;
3884 if (waitcnt
!= NOOP_WAITCNT
)
3885 ac_build_waitcnt(&ctx
->ac
, waitcnt
);
3888 static void clock_emit(
3889 const struct lp_build_tgsi_action
*action
,
3890 struct lp_build_tgsi_context
*bld_base
,
3891 struct lp_build_emit_data
*emit_data
)
3893 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3894 LLVMValueRef tmp
= ac_build_shader_clock(&ctx
->ac
);
3896 emit_data
->output
[0] =
3897 LLVMBuildExtractElement(ctx
->ac
.builder
, tmp
, ctx
->i32_0
, "");
3898 emit_data
->output
[1] =
3899 LLVMBuildExtractElement(ctx
->ac
.builder
, tmp
, ctx
->i32_1
, "");
3902 static void si_llvm_emit_ddxy(
3903 const struct lp_build_tgsi_action
*action
,
3904 struct lp_build_tgsi_context
*bld_base
,
3905 struct lp_build_emit_data
*emit_data
)
3907 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3908 unsigned opcode
= emit_data
->info
->opcode
;
3913 if (opcode
== TGSI_OPCODE_DDX_FINE
)
3914 mask
= AC_TID_MASK_LEFT
;
3915 else if (opcode
== TGSI_OPCODE_DDY_FINE
)
3916 mask
= AC_TID_MASK_TOP
;
3918 mask
= AC_TID_MASK_TOP_LEFT
;
3920 /* for DDX we want to next X pixel, DDY next Y pixel. */
3921 idx
= (opcode
== TGSI_OPCODE_DDX
|| opcode
== TGSI_OPCODE_DDX_FINE
) ? 1 : 2;
3923 val
= ac_to_integer(&ctx
->ac
, emit_data
->args
[0]);
3924 val
= ac_build_ddxy(&ctx
->ac
, mask
, idx
, val
);
3925 emit_data
->output
[emit_data
->chan
] = val
;
3928 static void build_interp_intrinsic(const struct lp_build_tgsi_action
*action
,
3929 struct lp_build_tgsi_context
*bld_base
,
3930 struct lp_build_emit_data
*emit_data
)
3932 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3933 struct si_shader
*shader
= ctx
->shader
;
3934 const struct tgsi_shader_info
*info
= &shader
->selector
->info
;
3935 LLVMValueRef interp_param
;
3936 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3937 const struct tgsi_full_src_register
*input
= &inst
->Src
[0];
3938 int input_base
, input_array_size
;
3941 LLVMValueRef prim_mask
= ctx
->abi
.prim_mask
;
3942 LLVMValueRef array_idx
, offset_x
= NULL
, offset_y
= NULL
;
3943 int interp_param_idx
;
3947 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_OFFSET
) {
3948 /* offset is in second src, first two channels */
3949 offset_x
= lp_build_emit_fetch(bld_base
, emit_data
->inst
, 1,
3951 offset_y
= lp_build_emit_fetch(bld_base
, emit_data
->inst
, 1,
3953 } else if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_SAMPLE
) {
3954 LLVMValueRef sample_position
;
3955 LLVMValueRef sample_id
;
3956 LLVMValueRef halfval
= LLVMConstReal(ctx
->f32
, 0.5f
);
3958 /* fetch sample ID, then fetch its sample position,
3959 * and place into first two channels.
3961 sample_id
= lp_build_emit_fetch(bld_base
,
3962 emit_data
->inst
, 1, TGSI_CHAN_X
);
3963 sample_id
= ac_to_integer(&ctx
->ac
, sample_id
);
3965 /* Section 8.13.2 (Interpolation Functions) of the OpenGL Shading
3966 * Language 4.50 spec says about interpolateAtSample:
3968 * "Returns the value of the input interpolant variable at
3969 * the location of sample number sample. If multisample
3970 * buffers are not available, the input variable will be
3971 * evaluated at the center of the pixel. If sample sample
3972 * does not exist, the position used to interpolate the
3973 * input variable is undefined."
3975 * This means that sample_id values outside of the valid are
3976 * in fact valid input, and the usual mechanism for loading the
3977 * sample position doesn't work.
3979 if (ctx
->shader
->key
.mono
.u
.ps
.interpolate_at_sample_force_center
) {
3980 LLVMValueRef center
[4] = {
3981 LLVMConstReal(ctx
->f32
, 0.5),
3982 LLVMConstReal(ctx
->f32
, 0.5),
3987 sample_position
= ac_build_gather_values(&ctx
->ac
, center
, 4);
3989 sample_position
= load_sample_position(&ctx
->abi
, sample_id
);
3992 offset_x
= LLVMBuildExtractElement(ctx
->ac
.builder
, sample_position
,
3995 offset_x
= LLVMBuildFSub(ctx
->ac
.builder
, offset_x
, halfval
, "");
3996 offset_y
= LLVMBuildExtractElement(ctx
->ac
.builder
, sample_position
,
3998 offset_y
= LLVMBuildFSub(ctx
->ac
.builder
, offset_y
, halfval
, "");
4001 assert(input
->Register
.File
== TGSI_FILE_INPUT
);
4003 if (input
->Register
.Indirect
) {
4004 unsigned array_id
= input
->Indirect
.ArrayID
;
4007 input_base
= info
->input_array_first
[array_id
];
4008 input_array_size
= info
->input_array_last
[array_id
] - input_base
+ 1;
4010 input_base
= inst
->Src
[0].Register
.Index
;
4011 input_array_size
= info
->num_inputs
- input_base
;
4014 array_idx
= si_get_indirect_index(ctx
, &input
->Indirect
,
4015 1, input
->Register
.Index
- input_base
);
4017 input_base
= inst
->Src
[0].Register
.Index
;
4018 input_array_size
= 1;
4019 array_idx
= ctx
->i32_0
;
4022 interp
= shader
->selector
->info
.input_interpolate
[input_base
];
4024 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_OFFSET
||
4025 inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_SAMPLE
)
4026 location
= TGSI_INTERPOLATE_LOC_CENTER
;
4028 location
= TGSI_INTERPOLATE_LOC_CENTROID
;
4030 interp_param_idx
= lookup_interp_param_index(interp
, location
);
4031 if (interp_param_idx
== -1)
4033 else if (interp_param_idx
)
4034 interp_param
= LLVMGetParam(ctx
->main_fn
, interp_param_idx
);
4036 interp_param
= NULL
;
4038 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_OFFSET
||
4039 inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_SAMPLE
) {
4040 LLVMValueRef ij_out
[2];
4041 LLVMValueRef ddxy_out
= ac_build_ddxy_interp(&ctx
->ac
, interp_param
);
4044 * take the I then J parameters, and the DDX/Y for it, and
4045 * calculate the IJ inputs for the interpolator.
4046 * temp1 = ddx * offset/sample.x + I;
4047 * interp_param.I = ddy * offset/sample.y + temp1;
4048 * temp1 = ddx * offset/sample.x + J;
4049 * interp_param.J = ddy * offset/sample.y + temp1;
4051 for (i
= 0; i
< 2; i
++) {
4052 LLVMValueRef ix_ll
= LLVMConstInt(ctx
->i32
, i
, 0);
4053 LLVMValueRef iy_ll
= LLVMConstInt(ctx
->i32
, i
+ 2, 0);
4054 LLVMValueRef ddx_el
= LLVMBuildExtractElement(ctx
->ac
.builder
,
4055 ddxy_out
, ix_ll
, "");
4056 LLVMValueRef ddy_el
= LLVMBuildExtractElement(ctx
->ac
.builder
,
4057 ddxy_out
, iy_ll
, "");
4058 LLVMValueRef interp_el
= LLVMBuildExtractElement(ctx
->ac
.builder
,
4059 interp_param
, ix_ll
, "");
4062 interp_el
= ac_to_float(&ctx
->ac
, interp_el
);
4064 temp
= ac_build_fmad(&ctx
->ac
, ddx_el
, offset_x
, interp_el
);
4065 ij_out
[i
] = ac_build_fmad(&ctx
->ac
, ddy_el
, offset_y
, temp
);
4067 interp_param
= ac_build_gather_values(&ctx
->ac
, ij_out
, 2);
4071 interp_param
= ac_to_float(&ctx
->ac
, interp_param
);
4073 for (chan
= 0; chan
< 4; chan
++) {
4074 LLVMValueRef gather
= LLVMGetUndef(LLVMVectorType(ctx
->f32
, input_array_size
));
4075 unsigned schan
= tgsi_util_get_full_src_register_swizzle(&inst
->Src
[0], chan
);
4077 for (unsigned idx
= 0; idx
< input_array_size
; ++idx
) {
4078 LLVMValueRef v
, i
= NULL
, j
= NULL
;
4081 i
= LLVMBuildExtractElement(
4082 ctx
->ac
.builder
, interp_param
, ctx
->i32_0
, "");
4083 j
= LLVMBuildExtractElement(
4084 ctx
->ac
.builder
, interp_param
, ctx
->i32_1
, "");
4086 v
= si_build_fs_interp(ctx
, input_base
+ idx
, schan
,
4089 gather
= LLVMBuildInsertElement(ctx
->ac
.builder
,
4090 gather
, v
, LLVMConstInt(ctx
->i32
, idx
, false), "");
4093 emit_data
->output
[chan
] = LLVMBuildExtractElement(
4094 ctx
->ac
.builder
, gather
, array_idx
, "");
4098 static void vote_all_emit(
4099 const struct lp_build_tgsi_action
*action
,
4100 struct lp_build_tgsi_context
*bld_base
,
4101 struct lp_build_emit_data
*emit_data
)
4103 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4105 LLVMValueRef tmp
= ac_build_vote_all(&ctx
->ac
, emit_data
->args
[0]);
4106 emit_data
->output
[emit_data
->chan
] =
4107 LLVMBuildSExt(ctx
->ac
.builder
, tmp
, ctx
->i32
, "");
4110 static void vote_any_emit(
4111 const struct lp_build_tgsi_action
*action
,
4112 struct lp_build_tgsi_context
*bld_base
,
4113 struct lp_build_emit_data
*emit_data
)
4115 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4117 LLVMValueRef tmp
= ac_build_vote_any(&ctx
->ac
, emit_data
->args
[0]);
4118 emit_data
->output
[emit_data
->chan
] =
4119 LLVMBuildSExt(ctx
->ac
.builder
, tmp
, ctx
->i32
, "");
4122 static void vote_eq_emit(
4123 const struct lp_build_tgsi_action
*action
,
4124 struct lp_build_tgsi_context
*bld_base
,
4125 struct lp_build_emit_data
*emit_data
)
4127 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4129 LLVMValueRef tmp
= ac_build_vote_eq(&ctx
->ac
, emit_data
->args
[0]);
4130 emit_data
->output
[emit_data
->chan
] =
4131 LLVMBuildSExt(ctx
->ac
.builder
, tmp
, ctx
->i32
, "");
4134 static void ballot_emit(
4135 const struct lp_build_tgsi_action
*action
,
4136 struct lp_build_tgsi_context
*bld_base
,
4137 struct lp_build_emit_data
*emit_data
)
4139 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4140 LLVMBuilderRef builder
= ctx
->ac
.builder
;
4143 tmp
= lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, TGSI_CHAN_X
);
4144 tmp
= ac_build_ballot(&ctx
->ac
, tmp
);
4145 tmp
= LLVMBuildBitCast(builder
, tmp
, ctx
->v2i32
, "");
4147 emit_data
->output
[0] = LLVMBuildExtractElement(builder
, tmp
, ctx
->i32_0
, "");
4148 emit_data
->output
[1] = LLVMBuildExtractElement(builder
, tmp
, ctx
->i32_1
, "");
4151 static void read_lane_emit(
4152 const struct lp_build_tgsi_action
*action
,
4153 struct lp_build_tgsi_context
*bld_base
,
4154 struct lp_build_emit_data
*emit_data
)
4156 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4158 if (emit_data
->inst
->Instruction
.Opcode
== TGSI_OPCODE_READ_INVOC
) {
4159 emit_data
->args
[0] = lp_build_emit_fetch(bld_base
, emit_data
->inst
,
4160 0, emit_data
->src_chan
);
4162 /* Always read the source invocation (= lane) from the X channel. */
4163 emit_data
->args
[1] = lp_build_emit_fetch(bld_base
, emit_data
->inst
,
4165 emit_data
->arg_count
= 2;
4168 /* We currently have no other way to prevent LLVM from lifting the icmp
4169 * calls to a dominating basic block.
4171 ac_build_optimization_barrier(&ctx
->ac
, &emit_data
->args
[0]);
4173 for (unsigned i
= 0; i
< emit_data
->arg_count
; ++i
)
4174 emit_data
->args
[i
] = ac_to_integer(&ctx
->ac
, emit_data
->args
[i
]);
4176 emit_data
->output
[emit_data
->chan
] =
4177 ac_build_intrinsic(&ctx
->ac
, action
->intr_name
,
4178 ctx
->i32
, emit_data
->args
, emit_data
->arg_count
,
4179 AC_FUNC_ATTR_READNONE
|
4180 AC_FUNC_ATTR_CONVERGENT
);
4183 static unsigned si_llvm_get_stream(struct lp_build_tgsi_context
*bld_base
,
4184 struct lp_build_emit_data
*emit_data
)
4186 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4187 struct tgsi_src_register src0
= emit_data
->inst
->Src
[0].Register
;
4191 assert(src0
.File
== TGSI_FILE_IMMEDIATE
);
4193 imm
= ctx
->imms
[src0
.Index
* TGSI_NUM_CHANNELS
+ src0
.SwizzleX
];
4194 stream
= LLVMConstIntGetZExtValue(imm
) & 0x3;
4198 /* Emit one vertex from the geometry shader */
4199 static void si_llvm_emit_vertex(struct ac_shader_abi
*abi
,
4201 LLVMValueRef
*addrs
)
4203 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
4205 if (ctx
->shader
->key
.as_ngg
) {
4206 gfx10_ngg_gs_emit_vertex(ctx
, stream
, addrs
);
4210 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
4211 struct si_shader
*shader
= ctx
->shader
;
4212 struct lp_build_if_state if_state
;
4213 LLVMValueRef soffset
= LLVMGetParam(ctx
->main_fn
,
4214 ctx
->param_gs2vs_offset
);
4215 LLVMValueRef gs_next_vertex
;
4216 LLVMValueRef can_emit
;
4217 unsigned chan
, offset
;
4220 /* Write vertex attribute values to GSVS ring */
4221 gs_next_vertex
= LLVMBuildLoad(ctx
->ac
.builder
,
4222 ctx
->gs_next_vertex
[stream
],
4225 /* If this thread has already emitted the declared maximum number of
4226 * vertices, skip the write: excessive vertex emissions are not
4227 * supposed to have any effect.
4229 * If the shader has no writes to memory, kill it instead. This skips
4230 * further memory loads and may allow LLVM to skip to the end
4233 can_emit
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntULT
, gs_next_vertex
,
4234 LLVMConstInt(ctx
->i32
,
4235 shader
->selector
->gs_max_out_vertices
, 0), "");
4237 bool use_kill
= !info
->writes_memory
;
4239 ac_build_kill_if_false(&ctx
->ac
, can_emit
);
4241 lp_build_if(&if_state
, &ctx
->gallivm
, can_emit
);
4245 for (i
= 0; i
< info
->num_outputs
; i
++) {
4246 for (chan
= 0; chan
< 4; chan
++) {
4247 if (!(info
->output_usagemask
[i
] & (1 << chan
)) ||
4248 ((info
->output_streams
[i
] >> (2 * chan
)) & 3) != stream
)
4251 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->ac
.builder
, addrs
[4 * i
+ chan
], "");
4252 LLVMValueRef voffset
=
4253 LLVMConstInt(ctx
->i32
, offset
*
4254 shader
->selector
->gs_max_out_vertices
, 0);
4257 voffset
= LLVMBuildAdd(ctx
->ac
.builder
, voffset
, gs_next_vertex
, "");
4258 voffset
= LLVMBuildMul(ctx
->ac
.builder
, voffset
,
4259 LLVMConstInt(ctx
->i32
, 4, 0), "");
4261 out_val
= ac_to_integer(&ctx
->ac
, out_val
);
4263 ac_build_buffer_store_dword(&ctx
->ac
,
4264 ctx
->gsvs_ring
[stream
],
4266 voffset
, soffset
, 0,
4271 gs_next_vertex
= LLVMBuildAdd(ctx
->ac
.builder
, gs_next_vertex
, ctx
->i32_1
, "");
4272 LLVMBuildStore(ctx
->ac
.builder
, gs_next_vertex
, ctx
->gs_next_vertex
[stream
]);
4274 /* Signal vertex emission if vertex data was written. */
4276 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_EMIT
| AC_SENDMSG_GS
| (stream
<< 8),
4277 si_get_gs_wave_id(ctx
));
4281 lp_build_endif(&if_state
);
4284 /* Emit one vertex from the geometry shader */
4285 static void si_tgsi_emit_vertex(
4286 const struct lp_build_tgsi_action
*action
,
4287 struct lp_build_tgsi_context
*bld_base
,
4288 struct lp_build_emit_data
*emit_data
)
4290 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4291 unsigned stream
= si_llvm_get_stream(bld_base
, emit_data
);
4293 si_llvm_emit_vertex(&ctx
->abi
, stream
, ctx
->outputs
[0]);
4296 /* Cut one primitive from the geometry shader */
4297 static void si_llvm_emit_primitive(struct ac_shader_abi
*abi
,
4300 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
4302 if (ctx
->shader
->key
.as_ngg
) {
4303 LLVMBuildStore(ctx
->ac
.builder
, ctx
->ac
.i32_0
, ctx
->gs_curprim_verts
[stream
]);
4307 /* Signal primitive cut */
4308 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_CUT
| AC_SENDMSG_GS
| (stream
<< 8),
4309 si_get_gs_wave_id(ctx
));
4312 /* Cut one primitive from the geometry shader */
4313 static void si_tgsi_emit_primitive(
4314 const struct lp_build_tgsi_action
*action
,
4315 struct lp_build_tgsi_context
*bld_base
,
4316 struct lp_build_emit_data
*emit_data
)
4318 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4320 si_llvm_emit_primitive(&ctx
->abi
, si_llvm_get_stream(bld_base
, emit_data
));
4323 static void si_llvm_emit_barrier(const struct lp_build_tgsi_action
*action
,
4324 struct lp_build_tgsi_context
*bld_base
,
4325 struct lp_build_emit_data
*emit_data
)
4327 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4329 /* GFX6 only (thanks to a hw bug workaround):
4330 * The real barrier instruction isn’t needed, because an entire patch
4331 * always fits into a single wave.
4333 if (ctx
->screen
->info
.chip_class
== GFX6
&&
4334 ctx
->type
== PIPE_SHADER_TESS_CTRL
) {
4335 ac_build_waitcnt(&ctx
->ac
, LGKM_CNT
& VM_CNT
);
4339 ac_build_s_barrier(&ctx
->ac
);
4342 void si_create_function(struct si_shader_context
*ctx
,
4344 LLVMTypeRef
*returns
, unsigned num_returns
,
4345 struct si_function_info
*fninfo
,
4346 unsigned max_workgroup_size
)
4350 si_llvm_create_func(ctx
, name
, returns
, num_returns
,
4351 fninfo
->types
, fninfo
->num_params
);
4352 ctx
->return_value
= LLVMGetUndef(ctx
->return_type
);
4354 for (i
= 0; i
< fninfo
->num_sgpr_params
; ++i
) {
4355 LLVMValueRef P
= LLVMGetParam(ctx
->main_fn
, i
);
4357 /* The combination of:
4361 * allows the optimization passes to move loads and reduces
4362 * SGPR spilling significantly.
4364 ac_add_function_attr(ctx
->ac
.context
, ctx
->main_fn
, i
+ 1,
4365 AC_FUNC_ATTR_INREG
);
4367 if (LLVMGetTypeKind(LLVMTypeOf(P
)) == LLVMPointerTypeKind
) {
4368 ac_add_function_attr(ctx
->ac
.context
, ctx
->main_fn
, i
+ 1,
4369 AC_FUNC_ATTR_NOALIAS
);
4370 ac_add_attr_dereferenceable(P
, UINT64_MAX
);
4374 for (i
= 0; i
< fninfo
->num_params
; ++i
) {
4375 if (fninfo
->assign
[i
])
4376 *fninfo
->assign
[i
] = LLVMGetParam(ctx
->main_fn
, i
);
4379 if (ctx
->screen
->info
.address32_hi
) {
4380 ac_llvm_add_target_dep_function_attr(ctx
->main_fn
,
4381 "amdgpu-32bit-address-high-bits",
4382 ctx
->screen
->info
.address32_hi
);
4385 ac_llvm_set_workgroup_size(ctx
->main_fn
, max_workgroup_size
);
4387 LLVMAddTargetDependentFunctionAttr(ctx
->main_fn
,
4388 "no-signed-zeros-fp-math",
4391 if (ctx
->screen
->debug_flags
& DBG(UNSAFE_MATH
)) {
4392 /* These were copied from some LLVM test. */
4393 LLVMAddTargetDependentFunctionAttr(ctx
->main_fn
,
4394 "less-precise-fpmad",
4396 LLVMAddTargetDependentFunctionAttr(ctx
->main_fn
,
4399 LLVMAddTargetDependentFunctionAttr(ctx
->main_fn
,
4402 LLVMAddTargetDependentFunctionAttr(ctx
->main_fn
,
4408 static void declare_streamout_params(struct si_shader_context
*ctx
,
4409 struct pipe_stream_output_info
*so
,
4410 struct si_function_info
*fninfo
)
4414 /* Streamout SGPRs. */
4415 if (so
->num_outputs
) {
4416 if (ctx
->type
!= PIPE_SHADER_TESS_EVAL
)
4417 ctx
->param_streamout_config
= add_arg(fninfo
, ARG_SGPR
, ctx
->ac
.i32
);
4419 ctx
->param_streamout_config
= fninfo
->num_params
- 1;
4421 ctx
->param_streamout_write_index
= add_arg(fninfo
, ARG_SGPR
, ctx
->ac
.i32
);
4423 /* A streamout buffer offset is loaded if the stride is non-zero. */
4424 for (i
= 0; i
< 4; i
++) {
4428 ctx
->param_streamout_offset
[i
] = add_arg(fninfo
, ARG_SGPR
, ctx
->ac
.i32
);
4432 static unsigned si_get_max_workgroup_size(const struct si_shader
*shader
)
4434 switch (shader
->selector
->type
) {
4435 case PIPE_SHADER_VERTEX
:
4436 case PIPE_SHADER_TESS_EVAL
:
4437 return shader
->key
.as_ngg
? 128 : 0;
4439 case PIPE_SHADER_TESS_CTRL
:
4440 /* Return this so that LLVM doesn't remove s_barrier
4441 * instructions on chips where we use s_barrier. */
4442 return shader
->selector
->screen
->info
.chip_class
>= GFX7
? 128 : 64;
4444 case PIPE_SHADER_GEOMETRY
:
4445 return shader
->selector
->screen
->info
.chip_class
>= GFX9
? 128 : 64;
4447 case PIPE_SHADER_COMPUTE
:
4448 break; /* see below */
4454 const unsigned *properties
= shader
->selector
->info
.properties
;
4455 unsigned max_work_group_size
=
4456 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
] *
4457 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT
] *
4458 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH
];
4460 if (!max_work_group_size
) {
4461 /* This is a variable group size compute shader,
4462 * compile it for the maximum possible group size.
4464 max_work_group_size
= SI_MAX_VARIABLE_THREADS_PER_BLOCK
;
4466 return max_work_group_size
;
4469 static void declare_const_and_shader_buffers(struct si_shader_context
*ctx
,
4470 struct si_function_info
*fninfo
,
4473 LLVMTypeRef const_shader_buf_type
;
4475 if (ctx
->shader
->selector
->info
.const_buffers_declared
== 1 &&
4476 ctx
->shader
->selector
->info
.shader_buffers_declared
== 0)
4477 const_shader_buf_type
= ctx
->f32
;
4479 const_shader_buf_type
= ctx
->v4i32
;
4481 unsigned const_and_shader_buffers
=
4482 add_arg(fninfo
, ARG_SGPR
,
4483 ac_array_in_const32_addr_space(const_shader_buf_type
));
4486 ctx
->param_const_and_shader_buffers
= const_and_shader_buffers
;
4489 static void declare_samplers_and_images(struct si_shader_context
*ctx
,
4490 struct si_function_info
*fninfo
,
4493 unsigned samplers_and_images
=
4494 add_arg(fninfo
, ARG_SGPR
,
4495 ac_array_in_const32_addr_space(ctx
->v8i32
));
4498 ctx
->param_samplers_and_images
= samplers_and_images
;
4501 static void declare_per_stage_desc_pointers(struct si_shader_context
*ctx
,
4502 struct si_function_info
*fninfo
,
4505 declare_const_and_shader_buffers(ctx
, fninfo
, assign_params
);
4506 declare_samplers_and_images(ctx
, fninfo
, assign_params
);
4509 static void declare_global_desc_pointers(struct si_shader_context
*ctx
,
4510 struct si_function_info
*fninfo
)
4512 ctx
->param_rw_buffers
= add_arg(fninfo
, ARG_SGPR
,
4513 ac_array_in_const32_addr_space(ctx
->v4i32
));
4514 ctx
->param_bindless_samplers_and_images
= add_arg(fninfo
, ARG_SGPR
,
4515 ac_array_in_const32_addr_space(ctx
->v8i32
));
4518 static void declare_vs_specific_input_sgprs(struct si_shader_context
*ctx
,
4519 struct si_function_info
*fninfo
)
4521 ctx
->param_vs_state_bits
= add_arg(fninfo
, ARG_SGPR
, ctx
->i32
);
4522 add_arg_assign(fninfo
, ARG_SGPR
, ctx
->i32
, &ctx
->abi
.base_vertex
);
4523 add_arg_assign(fninfo
, ARG_SGPR
, ctx
->i32
, &ctx
->abi
.start_instance
);
4524 add_arg_assign(fninfo
, ARG_SGPR
, ctx
->i32
, &ctx
->abi
.draw_id
);
4527 static void declare_vs_input_vgprs(struct si_shader_context
*ctx
,
4528 struct si_function_info
*fninfo
,
4529 unsigned *num_prolog_vgprs
)
4531 struct si_shader
*shader
= ctx
->shader
;
4533 add_arg_assign(fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.vertex_id
);
4534 if (shader
->key
.as_ls
) {
4535 ctx
->param_rel_auto_id
= add_arg(fninfo
, ARG_VGPR
, ctx
->i32
);
4536 add_arg_assign(fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.instance_id
);
4538 add_arg_assign(fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.instance_id
);
4539 ctx
->param_vs_prim_id
= add_arg(fninfo
, ARG_VGPR
, ctx
->i32
);
4541 add_arg(fninfo
, ARG_VGPR
, ctx
->i32
); /* unused */
4543 if (!shader
->is_gs_copy_shader
) {
4544 /* Vertex load indices. */
4545 ctx
->param_vertex_index0
= fninfo
->num_params
;
4546 for (unsigned i
= 0; i
< shader
->selector
->info
.num_inputs
; i
++)
4547 add_arg(fninfo
, ARG_VGPR
, ctx
->i32
);
4548 *num_prolog_vgprs
+= shader
->selector
->info
.num_inputs
;
4552 static void declare_vs_blit_inputs(struct si_shader_context
*ctx
,
4553 struct si_function_info
*fninfo
,
4554 unsigned vs_blit_property
)
4556 ctx
->param_vs_blit_inputs
= fninfo
->num_params
;
4557 add_arg(fninfo
, ARG_SGPR
, ctx
->i32
); /* i16 x1, y1 */
4558 add_arg(fninfo
, ARG_SGPR
, ctx
->i32
); /* i16 x2, y2 */
4559 add_arg(fninfo
, ARG_SGPR
, ctx
->f32
); /* depth */
4561 if (vs_blit_property
== SI_VS_BLIT_SGPRS_POS_COLOR
) {
4562 add_arg(fninfo
, ARG_SGPR
, ctx
->f32
); /* color0 */
4563 add_arg(fninfo
, ARG_SGPR
, ctx
->f32
); /* color1 */
4564 add_arg(fninfo
, ARG_SGPR
, ctx
->f32
); /* color2 */
4565 add_arg(fninfo
, ARG_SGPR
, ctx
->f32
); /* color3 */
4566 } else if (vs_blit_property
== SI_VS_BLIT_SGPRS_POS_TEXCOORD
) {
4567 add_arg(fninfo
, ARG_SGPR
, ctx
->f32
); /* texcoord.x1 */
4568 add_arg(fninfo
, ARG_SGPR
, ctx
->f32
); /* texcoord.y1 */
4569 add_arg(fninfo
, ARG_SGPR
, ctx
->f32
); /* texcoord.x2 */
4570 add_arg(fninfo
, ARG_SGPR
, ctx
->f32
); /* texcoord.y2 */
4571 add_arg(fninfo
, ARG_SGPR
, ctx
->f32
); /* texcoord.z */
4572 add_arg(fninfo
, ARG_SGPR
, ctx
->f32
); /* texcoord.w */
4576 static void declare_tes_input_vgprs(struct si_shader_context
*ctx
,
4577 struct si_function_info
*fninfo
)
4579 ctx
->param_tes_u
= add_arg(fninfo
, ARG_VGPR
, ctx
->f32
);
4580 ctx
->param_tes_v
= add_arg(fninfo
, ARG_VGPR
, ctx
->f32
);
4581 ctx
->param_tes_rel_patch_id
= add_arg(fninfo
, ARG_VGPR
, ctx
->i32
);
4582 add_arg_assign(fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.tes_patch_id
);
4586 /* Convenient merged shader definitions. */
4587 SI_SHADER_MERGED_VERTEX_TESSCTRL
= PIPE_SHADER_TYPES
,
4588 SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY
,
4591 static void create_function(struct si_shader_context
*ctx
)
4593 struct si_shader
*shader
= ctx
->shader
;
4594 struct si_function_info fninfo
;
4595 LLVMTypeRef returns
[16+32*4];
4596 unsigned i
, num_return_sgprs
;
4597 unsigned num_returns
= 0;
4598 unsigned num_prolog_vgprs
= 0;
4599 unsigned type
= ctx
->type
;
4600 unsigned vs_blit_property
=
4601 shader
->selector
->info
.properties
[TGSI_PROPERTY_VS_BLIT_SGPRS
];
4603 si_init_function_info(&fninfo
);
4605 /* Set MERGED shaders. */
4606 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
4607 if (shader
->key
.as_ls
|| type
== PIPE_SHADER_TESS_CTRL
)
4608 type
= SI_SHADER_MERGED_VERTEX_TESSCTRL
; /* LS or HS */
4609 else if (shader
->key
.as_es
|| shader
->key
.as_ngg
|| type
== PIPE_SHADER_GEOMETRY
)
4610 type
= SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY
;
4613 LLVMTypeRef v3i32
= LLVMVectorType(ctx
->i32
, 3);
4616 case PIPE_SHADER_VERTEX
:
4617 declare_global_desc_pointers(ctx
, &fninfo
);
4619 if (vs_blit_property
) {
4620 declare_vs_blit_inputs(ctx
, &fninfo
, vs_blit_property
);
4623 declare_vs_input_vgprs(ctx
, &fninfo
, &num_prolog_vgprs
);
4627 declare_per_stage_desc_pointers(ctx
, &fninfo
, true);
4628 declare_vs_specific_input_sgprs(ctx
, &fninfo
);
4629 ctx
->param_vertex_buffers
= add_arg(&fninfo
, ARG_SGPR
,
4630 ac_array_in_const32_addr_space(ctx
->v4i32
));
4632 if (shader
->key
.as_es
) {
4633 ctx
->param_es2gs_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4634 } else if (shader
->key
.as_ls
) {
4635 /* no extra parameters */
4637 if (shader
->is_gs_copy_shader
) {
4638 fninfo
.num_params
= ctx
->param_vs_state_bits
+ 1;
4639 fninfo
.num_sgpr_params
= fninfo
.num_params
;
4642 /* The locations of the other parameters are assigned dynamically. */
4643 declare_streamout_params(ctx
, &shader
->selector
->so
,
4648 declare_vs_input_vgprs(ctx
, &fninfo
, &num_prolog_vgprs
);
4651 if (shader
->key
.opt
.vs_as_prim_discard_cs
) {
4652 for (i
= 0; i
< 4; i
++)
4653 returns
[num_returns
++] = ctx
->f32
; /* VGPRs */
4657 case PIPE_SHADER_TESS_CTRL
: /* GFX6-GFX8 */
4658 declare_global_desc_pointers(ctx
, &fninfo
);
4659 declare_per_stage_desc_pointers(ctx
, &fninfo
, true);
4660 ctx
->param_tcs_offchip_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4661 ctx
->param_tcs_out_lds_offsets
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4662 ctx
->param_tcs_out_lds_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4663 ctx
->param_vs_state_bits
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4664 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4665 ctx
->param_tcs_factor_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4668 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.tcs_patch_id
);
4669 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.tcs_rel_ids
);
4671 /* param_tcs_offchip_offset and param_tcs_factor_offset are
4672 * placed after the user SGPRs.
4674 for (i
= 0; i
< GFX6_TCS_NUM_USER_SGPR
+ 2; i
++)
4675 returns
[num_returns
++] = ctx
->i32
; /* SGPRs */
4676 for (i
= 0; i
< 11; i
++)
4677 returns
[num_returns
++] = ctx
->f32
; /* VGPRs */
4680 case SI_SHADER_MERGED_VERTEX_TESSCTRL
:
4681 /* Merged stages have 8 system SGPRs at the beginning. */
4682 /* SPI_SHADER_USER_DATA_ADDR_LO/HI_HS */
4683 declare_per_stage_desc_pointers(ctx
, &fninfo
,
4684 ctx
->type
== PIPE_SHADER_TESS_CTRL
);
4685 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4686 ctx
->param_merged_wave_info
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4687 ctx
->param_tcs_factor_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4688 ctx
->param_merged_scratch_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4689 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused */
4690 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused */
4692 declare_global_desc_pointers(ctx
, &fninfo
);
4693 declare_per_stage_desc_pointers(ctx
, &fninfo
,
4694 ctx
->type
== PIPE_SHADER_VERTEX
);
4695 declare_vs_specific_input_sgprs(ctx
, &fninfo
);
4697 ctx
->param_tcs_offchip_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4698 ctx
->param_tcs_out_lds_offsets
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4699 ctx
->param_tcs_out_lds_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4700 ctx
->param_vertex_buffers
= add_arg(&fninfo
, ARG_SGPR
,
4701 ac_array_in_const32_addr_space(ctx
->v4i32
));
4703 /* VGPRs (first TCS, then VS) */
4704 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.tcs_patch_id
);
4705 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.tcs_rel_ids
);
4707 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
4708 declare_vs_input_vgprs(ctx
, &fninfo
,
4711 /* LS return values are inputs to the TCS main shader part. */
4712 for (i
= 0; i
< 8 + GFX9_TCS_NUM_USER_SGPR
; i
++)
4713 returns
[num_returns
++] = ctx
->i32
; /* SGPRs */
4714 for (i
= 0; i
< 2; i
++)
4715 returns
[num_returns
++] = ctx
->f32
; /* VGPRs */
4717 /* TCS return values are inputs to the TCS epilog.
4719 * param_tcs_offchip_offset, param_tcs_factor_offset,
4720 * param_tcs_offchip_layout, and param_rw_buffers
4721 * should be passed to the epilog.
4723 for (i
= 0; i
<= 8 + GFX9_SGPR_TCS_OUT_LAYOUT
; i
++)
4724 returns
[num_returns
++] = ctx
->i32
; /* SGPRs */
4725 for (i
= 0; i
< 11; i
++)
4726 returns
[num_returns
++] = ctx
->f32
; /* VGPRs */
4730 case SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY
:
4731 /* Merged stages have 8 system SGPRs at the beginning. */
4732 /* SPI_SHADER_USER_DATA_ADDR_LO/HI_GS */
4733 declare_per_stage_desc_pointers(ctx
, &fninfo
,
4734 ctx
->type
== PIPE_SHADER_GEOMETRY
);
4736 if (ctx
->shader
->key
.as_ngg
)
4737 add_arg_assign(&fninfo
, ARG_SGPR
, ctx
->i32
, &ctx
->gs_tg_info
);
4739 ctx
->param_gs2vs_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4741 ctx
->param_merged_wave_info
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4742 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4743 ctx
->param_merged_scratch_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4744 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused (SPI_SHADER_PGM_LO/HI_GS << 8) */
4745 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused (SPI_SHADER_PGM_LO/HI_GS >> 24) */
4747 declare_global_desc_pointers(ctx
, &fninfo
);
4748 if (ctx
->type
!= PIPE_SHADER_VERTEX
|| !vs_blit_property
) {
4749 declare_per_stage_desc_pointers(ctx
, &fninfo
,
4750 (ctx
->type
== PIPE_SHADER_VERTEX
||
4751 ctx
->type
== PIPE_SHADER_TESS_EVAL
));
4754 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
4755 if (vs_blit_property
)
4756 declare_vs_blit_inputs(ctx
, &fninfo
, vs_blit_property
);
4758 declare_vs_specific_input_sgprs(ctx
, &fninfo
);
4760 ctx
->param_vs_state_bits
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4761 ctx
->param_tcs_offchip_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4762 ctx
->param_tes_offchip_addr
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4763 /* Declare as many input SGPRs as the VS has. */
4766 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
4767 ctx
->param_vertex_buffers
= add_arg(&fninfo
, ARG_SGPR
,
4768 ac_array_in_const32_addr_space(ctx
->v4i32
));
4771 /* VGPRs (first GS, then VS/TES) */
4772 ctx
->param_gs_vtx01_offset
= add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
4773 ctx
->param_gs_vtx23_offset
= add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
4774 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.gs_prim_id
);
4775 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.gs_invocation_id
);
4776 ctx
->param_gs_vtx45_offset
= add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
4778 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
4779 declare_vs_input_vgprs(ctx
, &fninfo
,
4781 } else if (ctx
->type
== PIPE_SHADER_TESS_EVAL
) {
4782 declare_tes_input_vgprs(ctx
, &fninfo
);
4785 if (ctx
->shader
->key
.as_es
&&
4786 (ctx
->type
== PIPE_SHADER_VERTEX
||
4787 ctx
->type
== PIPE_SHADER_TESS_EVAL
)) {
4788 unsigned num_user_sgprs
;
4790 if (ctx
->type
== PIPE_SHADER_VERTEX
)
4791 num_user_sgprs
= GFX9_VSGS_NUM_USER_SGPR
;
4793 num_user_sgprs
= GFX9_TESGS_NUM_USER_SGPR
;
4795 /* ES return values are inputs to GS. */
4796 for (i
= 0; i
< 8 + num_user_sgprs
; i
++)
4797 returns
[num_returns
++] = ctx
->i32
; /* SGPRs */
4798 for (i
= 0; i
< 5; i
++)
4799 returns
[num_returns
++] = ctx
->f32
; /* VGPRs */
4803 case PIPE_SHADER_TESS_EVAL
:
4804 declare_global_desc_pointers(ctx
, &fninfo
);
4805 declare_per_stage_desc_pointers(ctx
, &fninfo
, true);
4806 ctx
->param_vs_state_bits
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4807 ctx
->param_tcs_offchip_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4808 ctx
->param_tes_offchip_addr
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4810 if (shader
->key
.as_es
) {
4811 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4812 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4813 ctx
->param_es2gs_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4815 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4816 declare_streamout_params(ctx
, &shader
->selector
->so
,
4818 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4822 declare_tes_input_vgprs(ctx
, &fninfo
);
4825 case PIPE_SHADER_GEOMETRY
:
4826 declare_global_desc_pointers(ctx
, &fninfo
);
4827 declare_per_stage_desc_pointers(ctx
, &fninfo
, true);
4828 ctx
->param_gs2vs_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4829 ctx
->param_gs_wave_id
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4832 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->gs_vtx_offset
[0]);
4833 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->gs_vtx_offset
[1]);
4834 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.gs_prim_id
);
4835 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->gs_vtx_offset
[2]);
4836 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->gs_vtx_offset
[3]);
4837 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->gs_vtx_offset
[4]);
4838 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->gs_vtx_offset
[5]);
4839 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.gs_invocation_id
);
4842 case PIPE_SHADER_FRAGMENT
:
4843 declare_global_desc_pointers(ctx
, &fninfo
);
4844 declare_per_stage_desc_pointers(ctx
, &fninfo
, true);
4845 add_arg_checked(&fninfo
, ARG_SGPR
, ctx
->f32
, SI_PARAM_ALPHA_REF
);
4846 add_arg_assign_checked(&fninfo
, ARG_SGPR
, ctx
->i32
,
4847 &ctx
->abi
.prim_mask
, SI_PARAM_PRIM_MASK
);
4849 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->v2i32
, SI_PARAM_PERSP_SAMPLE
);
4850 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->v2i32
, SI_PARAM_PERSP_CENTER
);
4851 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->v2i32
, SI_PARAM_PERSP_CENTROID
);
4852 add_arg_checked(&fninfo
, ARG_VGPR
, v3i32
, SI_PARAM_PERSP_PULL_MODEL
);
4853 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->v2i32
, SI_PARAM_LINEAR_SAMPLE
);
4854 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->v2i32
, SI_PARAM_LINEAR_CENTER
);
4855 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->v2i32
, SI_PARAM_LINEAR_CENTROID
);
4856 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->f32
, SI_PARAM_LINE_STIPPLE_TEX
);
4857 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->f32
,
4858 &ctx
->abi
.frag_pos
[0], SI_PARAM_POS_X_FLOAT
);
4859 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->f32
,
4860 &ctx
->abi
.frag_pos
[1], SI_PARAM_POS_Y_FLOAT
);
4861 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->f32
,
4862 &ctx
->abi
.frag_pos
[2], SI_PARAM_POS_Z_FLOAT
);
4863 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->f32
,
4864 &ctx
->abi
.frag_pos
[3], SI_PARAM_POS_W_FLOAT
);
4865 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->i32
,
4866 &ctx
->abi
.front_face
, SI_PARAM_FRONT_FACE
);
4867 shader
->info
.face_vgpr_index
= 20;
4868 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->i32
,
4869 &ctx
->abi
.ancillary
, SI_PARAM_ANCILLARY
);
4870 shader
->info
.ancillary_vgpr_index
= 21;
4871 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->f32
,
4872 &ctx
->abi
.sample_coverage
, SI_PARAM_SAMPLE_COVERAGE
);
4873 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->i32
, SI_PARAM_POS_FIXED_PT
);
4875 /* Color inputs from the prolog. */
4876 if (shader
->selector
->info
.colors_read
) {
4877 unsigned num_color_elements
=
4878 util_bitcount(shader
->selector
->info
.colors_read
);
4880 assert(fninfo
.num_params
+ num_color_elements
<= ARRAY_SIZE(fninfo
.types
));
4881 for (i
= 0; i
< num_color_elements
; i
++)
4882 add_arg(&fninfo
, ARG_VGPR
, ctx
->f32
);
4884 num_prolog_vgprs
+= num_color_elements
;
4887 /* Outputs for the epilog. */
4888 num_return_sgprs
= SI_SGPR_ALPHA_REF
+ 1;
4891 util_bitcount(shader
->selector
->info
.colors_written
) * 4 +
4892 shader
->selector
->info
.writes_z
+
4893 shader
->selector
->info
.writes_stencil
+
4894 shader
->selector
->info
.writes_samplemask
+
4895 1 /* SampleMaskIn */;
4897 num_returns
= MAX2(num_returns
,
4899 PS_EPILOG_SAMPLEMASK_MIN_LOC
+ 1);
4901 for (i
= 0; i
< num_return_sgprs
; i
++)
4902 returns
[i
] = ctx
->i32
;
4903 for (; i
< num_returns
; i
++)
4904 returns
[i
] = ctx
->f32
;
4907 case PIPE_SHADER_COMPUTE
:
4908 declare_global_desc_pointers(ctx
, &fninfo
);
4909 declare_per_stage_desc_pointers(ctx
, &fninfo
, true);
4910 if (shader
->selector
->info
.uses_grid_size
)
4911 add_arg_assign(&fninfo
, ARG_SGPR
, v3i32
, &ctx
->abi
.num_work_groups
);
4912 if (shader
->selector
->info
.uses_block_size
&&
4913 shader
->selector
->info
.properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
] == 0)
4914 ctx
->param_block_size
= add_arg(&fninfo
, ARG_SGPR
, v3i32
);
4916 unsigned cs_user_data_dwords
=
4917 shader
->selector
->info
.properties
[TGSI_PROPERTY_CS_USER_DATA_DWORDS
];
4918 if (cs_user_data_dwords
) {
4919 ctx
->param_cs_user_data
= add_arg(&fninfo
, ARG_SGPR
,
4920 LLVMVectorType(ctx
->i32
, cs_user_data_dwords
));
4923 for (i
= 0; i
< 3; i
++) {
4924 ctx
->abi
.workgroup_ids
[i
] = NULL
;
4925 if (shader
->selector
->info
.uses_block_id
[i
])
4926 add_arg_assign(&fninfo
, ARG_SGPR
, ctx
->i32
, &ctx
->abi
.workgroup_ids
[i
]);
4929 add_arg_assign(&fninfo
, ARG_VGPR
, v3i32
, &ctx
->abi
.local_invocation_ids
);
4932 assert(0 && "unimplemented shader");
4936 si_create_function(ctx
, "main", returns
, num_returns
, &fninfo
,
4937 si_get_max_workgroup_size(shader
));
4939 /* Reserve register locations for VGPR inputs the PS prolog may need. */
4940 if (ctx
->type
== PIPE_SHADER_FRAGMENT
&& !ctx
->shader
->is_monolithic
) {
4941 ac_llvm_add_target_dep_function_attr(ctx
->main_fn
,
4942 "InitialPSInputAddr",
4943 S_0286D0_PERSP_SAMPLE_ENA(1) |
4944 S_0286D0_PERSP_CENTER_ENA(1) |
4945 S_0286D0_PERSP_CENTROID_ENA(1) |
4946 S_0286D0_LINEAR_SAMPLE_ENA(1) |
4947 S_0286D0_LINEAR_CENTER_ENA(1) |
4948 S_0286D0_LINEAR_CENTROID_ENA(1) |
4949 S_0286D0_FRONT_FACE_ENA(1) |
4950 S_0286D0_ANCILLARY_ENA(1) |
4951 S_0286D0_POS_FIXED_PT_ENA(1));
4954 shader
->info
.num_input_sgprs
= 0;
4955 shader
->info
.num_input_vgprs
= 0;
4957 for (i
= 0; i
< fninfo
.num_sgpr_params
; ++i
)
4958 shader
->info
.num_input_sgprs
+= ac_get_type_size(fninfo
.types
[i
]) / 4;
4960 for (; i
< fninfo
.num_params
; ++i
)
4961 shader
->info
.num_input_vgprs
+= ac_get_type_size(fninfo
.types
[i
]) / 4;
4963 assert(shader
->info
.num_input_vgprs
>= num_prolog_vgprs
);
4964 shader
->info
.num_input_vgprs
-= num_prolog_vgprs
;
4966 if (shader
->key
.as_ls
|| ctx
->type
== PIPE_SHADER_TESS_CTRL
) {
4967 if (USE_LDS_SYMBOLS
&& HAVE_LLVM
>= 0x0900) {
4968 /* The LSHS size is not known until draw time, so we append it
4969 * at the end of whatever LDS use there may be in the rest of
4970 * the shader (currently none, unless LLVM decides to do its
4971 * own LDS-based lowering).
4973 ctx
->ac
.lds
= LLVMAddGlobalInAddressSpace(
4974 ctx
->ac
.module
, LLVMArrayType(ctx
->i32
, 0),
4975 "__lds_end", AC_ADDR_SPACE_LDS
);
4976 LLVMSetAlignment(ctx
->ac
.lds
, 256);
4978 ac_declare_lds_as_pointer(&ctx
->ac
);
4984 * Load ESGS and GSVS ring buffer resource descriptors and save the variables
4987 static void preload_ring_buffers(struct si_shader_context
*ctx
)
4989 LLVMBuilderRef builder
= ctx
->ac
.builder
;
4991 LLVMValueRef buf_ptr
= LLVMGetParam(ctx
->main_fn
,
4992 ctx
->param_rw_buffers
);
4994 if (ctx
->shader
->key
.as_es
|| ctx
->type
== PIPE_SHADER_GEOMETRY
) {
4995 if (ctx
->screen
->info
.chip_class
<= GFX8
) {
4997 ctx
->type
== PIPE_SHADER_GEOMETRY
? SI_GS_RING_ESGS
4999 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
, ring
, 0);
5002 ac_build_load_to_sgpr(&ctx
->ac
, buf_ptr
, offset
);
5004 if (USE_LDS_SYMBOLS
&& HAVE_LLVM
>= 0x0900) {
5005 /* Declare the ESGS ring as an explicit LDS symbol.
5006 * For monolithic shaders, we declare the ring only once.
5008 * We declare it with 64KB alignment as a hint that the
5009 * pointer value will always be 0.
5011 ctx
->esgs_ring
= LLVMAddGlobalInAddressSpace(
5012 ctx
->ac
.module
, LLVMArrayType(ctx
->i32
, 0),
5015 LLVMSetAlignment(ctx
->esgs_ring
, 64 * 1024);
5017 ac_declare_lds_as_pointer(&ctx
->ac
);
5018 ctx
->esgs_ring
= ctx
->ac
.lds
;
5023 if (ctx
->shader
->is_gs_copy_shader
) {
5024 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
, SI_RING_GSVS
, 0);
5027 ac_build_load_to_sgpr(&ctx
->ac
, buf_ptr
, offset
);
5028 } else if (ctx
->type
== PIPE_SHADER_GEOMETRY
) {
5029 const struct si_shader_selector
*sel
= ctx
->shader
->selector
;
5030 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
, SI_RING_GSVS
, 0);
5031 LLVMValueRef base_ring
;
5033 base_ring
= ac_build_load_to_sgpr(&ctx
->ac
, buf_ptr
, offset
);
5035 /* The conceptual layout of the GSVS ring is
5036 * v0c0 .. vLv0 v0c1 .. vLc1 ..
5037 * but the real memory layout is swizzled across
5039 * t0v0c0 .. t15v0c0 t0v1c0 .. t15v1c0 ... t15vLcL
5041 * Override the buffer descriptor accordingly.
5043 LLVMTypeRef v2i64
= LLVMVectorType(ctx
->i64
, 2);
5044 uint64_t stream_offset
= 0;
5046 for (unsigned stream
= 0; stream
< 4; ++stream
) {
5047 unsigned num_components
;
5049 unsigned num_records
;
5050 LLVMValueRef ring
, tmp
;
5052 num_components
= sel
->info
.num_stream_output_components
[stream
];
5053 if (!num_components
)
5056 stride
= 4 * num_components
* sel
->gs_max_out_vertices
;
5058 /* Limit on the stride field for <= GFX7. */
5059 assert(stride
< (1 << 14));
5063 ring
= LLVMBuildBitCast(builder
, base_ring
, v2i64
, "");
5064 tmp
= LLVMBuildExtractElement(builder
, ring
, ctx
->i32_0
, "");
5065 tmp
= LLVMBuildAdd(builder
, tmp
,
5066 LLVMConstInt(ctx
->i64
,
5067 stream_offset
, 0), "");
5068 stream_offset
+= stride
* 64;
5070 ring
= LLVMBuildInsertElement(builder
, ring
, tmp
, ctx
->i32_0
, "");
5071 ring
= LLVMBuildBitCast(builder
, ring
, ctx
->v4i32
, "");
5072 tmp
= LLVMBuildExtractElement(builder
, ring
, ctx
->i32_1
, "");
5073 tmp
= LLVMBuildOr(builder
, tmp
,
5074 LLVMConstInt(ctx
->i32
,
5075 S_008F04_STRIDE(stride
) |
5076 S_008F04_SWIZZLE_ENABLE(1), 0), "");
5077 ring
= LLVMBuildInsertElement(builder
, ring
, tmp
, ctx
->i32_1
, "");
5078 ring
= LLVMBuildInsertElement(builder
, ring
,
5079 LLVMConstInt(ctx
->i32
, num_records
, 0),
5080 LLVMConstInt(ctx
->i32
, 2, 0), "");
5083 S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
5084 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
5085 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
5086 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
5087 S_008F0C_INDEX_STRIDE(1) | /* index_stride = 16 (elements) */
5088 S_008F0C_ADD_TID_ENABLE(1);
5090 if (ctx
->ac
.chip_class
>= GFX10
) {
5091 rsrc3
|= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT
) |
5092 S_008F0C_OOB_SELECT(2) |
5093 S_008F0C_RESOURCE_LEVEL(1);
5095 rsrc3
|= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
5096 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
5097 S_008F0C_ELEMENT_SIZE(1); /* element_size = 4 (bytes) */
5100 ring
= LLVMBuildInsertElement(builder
, ring
,
5101 LLVMConstInt(ctx
->i32
, rsrc3
, false),
5102 LLVMConstInt(ctx
->i32
, 3, 0), "");
5104 ctx
->gsvs_ring
[stream
] = ring
;
5106 } else if (ctx
->type
== PIPE_SHADER_TESS_EVAL
) {
5107 ctx
->tess_offchip_ring
= get_tess_ring_descriptor(ctx
, TESS_OFFCHIP_RING_TES
);
5111 static void si_llvm_emit_polygon_stipple(struct si_shader_context
*ctx
,
5112 LLVMValueRef param_rw_buffers
,
5113 unsigned param_pos_fixed_pt
)
5115 LLVMBuilderRef builder
= ctx
->ac
.builder
;
5116 LLVMValueRef slot
, desc
, offset
, row
, bit
, address
[2];
5118 /* Use the fixed-point gl_FragCoord input.
5119 * Since the stipple pattern is 32x32 and it repeats, just get 5 bits
5120 * per coordinate to get the repeating effect.
5122 address
[0] = si_unpack_param(ctx
, param_pos_fixed_pt
, 0, 5);
5123 address
[1] = si_unpack_param(ctx
, param_pos_fixed_pt
, 16, 5);
5125 /* Load the buffer descriptor. */
5126 slot
= LLVMConstInt(ctx
->i32
, SI_PS_CONST_POLY_STIPPLE
, 0);
5127 desc
= ac_build_load_to_sgpr(&ctx
->ac
, param_rw_buffers
, slot
);
5129 /* The stipple pattern is 32x32, each row has 32 bits. */
5130 offset
= LLVMBuildMul(builder
, address
[1],
5131 LLVMConstInt(ctx
->i32
, 4, 0), "");
5132 row
= buffer_load_const(ctx
, desc
, offset
);
5133 row
= ac_to_integer(&ctx
->ac
, row
);
5134 bit
= LLVMBuildLShr(builder
, row
, address
[0], "");
5135 bit
= LLVMBuildTrunc(builder
, bit
, ctx
->i1
, "");
5136 ac_build_kill_if_false(&ctx
->ac
, bit
);
5139 /* For the UMR disassembler. */
5140 #define DEBUGGER_END_OF_CODE_MARKER 0xbf9f0000 /* invalid instruction */
5141 #define DEBUGGER_NUM_MARKERS 5
5143 static bool si_shader_binary_open(struct si_screen
*screen
,
5144 struct si_shader
*shader
,
5145 struct ac_rtld_binary
*rtld
)
5147 const struct si_shader_selector
*sel
= shader
->selector
;
5148 enum pipe_shader_type shader_type
= sel
? sel
->type
: PIPE_SHADER_COMPUTE
;
5149 const char *part_elfs
[5];
5150 size_t part_sizes
[5];
5151 unsigned num_parts
= 0;
5153 #define add_part(shader_or_part) \
5154 if (shader_or_part) { \
5155 part_elfs[num_parts] = (shader_or_part)->binary.elf_buffer; \
5156 part_sizes[num_parts] = (shader_or_part)->binary.elf_size; \
5160 add_part(shader
->prolog
);
5161 add_part(shader
->previous_stage
);
5162 add_part(shader
->prolog2
);
5164 add_part(shader
->epilog
);
5168 struct ac_rtld_symbol lds_symbols
[2];
5169 unsigned num_lds_symbols
= 0;
5171 if (sel
&& screen
->info
.chip_class
>= GFX9
&&
5172 sel
->type
== PIPE_SHADER_GEOMETRY
&& !shader
->is_gs_copy_shader
) {
5173 /* We add this symbol even on LLVM <= 8 to ensure that
5174 * shader->config.lds_size is set correctly below.
5176 struct ac_rtld_symbol
*sym
= &lds_symbols
[num_lds_symbols
++];
5177 sym
->name
= "esgs_ring";
5178 sym
->size
= shader
->gs_info
.esgs_ring_size
;
5179 sym
->align
= 64 * 1024;
5182 if (shader
->key
.as_ngg
&& sel
->type
== PIPE_SHADER_GEOMETRY
) {
5183 struct ac_rtld_symbol
*sym
= &lds_symbols
[num_lds_symbols
++];
5184 sym
->name
= "ngg_emit";
5185 sym
->size
= shader
->ngg
.ngg_emit_size
* 4;
5189 bool ok
= ac_rtld_open(rtld
, (struct ac_rtld_open_info
){
5190 .info
= &screen
->info
,
5192 .halt_at_entry
= screen
->options
.halt_shaders
,
5194 .shader_type
= tgsi_processor_to_shader_stage(shader_type
),
5195 .num_parts
= num_parts
,
5196 .elf_ptrs
= part_elfs
,
5197 .elf_sizes
= part_sizes
,
5198 .num_shared_lds_symbols
= num_lds_symbols
,
5199 .shared_lds_symbols
= lds_symbols
});
5201 if (rtld
->lds_size
> 0) {
5202 unsigned alloc_granularity
= screen
->info
.chip_class
>= GFX7
? 512 : 256;
5203 shader
->config
.lds_size
=
5204 align(rtld
->lds_size
, alloc_granularity
) / alloc_granularity
;
5210 static unsigned si_get_shader_binary_size(struct si_screen
*screen
, struct si_shader
*shader
)
5212 struct ac_rtld_binary rtld
;
5213 si_shader_binary_open(screen
, shader
, &rtld
);
5214 return rtld
.rx_size
;
5217 static bool si_get_external_symbol(void *data
, const char *name
, uint64_t *value
)
5219 uint64_t *scratch_va
= data
;
5221 if (!strcmp(scratch_rsrc_dword0_symbol
, name
)) {
5222 *value
= (uint32_t)*scratch_va
;
5225 if (!strcmp(scratch_rsrc_dword1_symbol
, name
)) {
5226 /* Enable scratch coalescing. */
5227 *value
= S_008F04_BASE_ADDRESS_HI(*scratch_va
>> 32) |
5228 S_008F04_SWIZZLE_ENABLE(1);
5229 if (HAVE_LLVM
< 0x0800) {
5230 /* Old LLVM created an R_ABS32_HI relocation for
5240 bool si_shader_binary_upload(struct si_screen
*sscreen
, struct si_shader
*shader
,
5241 uint64_t scratch_va
)
5243 struct ac_rtld_binary binary
;
5244 if (!si_shader_binary_open(sscreen
, shader
, &binary
))
5247 si_resource_reference(&shader
->bo
, NULL
);
5248 shader
->bo
= si_aligned_buffer_create(&sscreen
->b
,
5249 sscreen
->cpdma_prefetch_writes_memory
?
5250 0 : SI_RESOURCE_FLAG_READ_ONLY
,
5251 PIPE_USAGE_IMMUTABLE
,
5252 align(binary
.rx_size
, SI_CPDMA_ALIGNMENT
),
5258 struct ac_rtld_upload_info u
= {};
5260 u
.get_external_symbol
= si_get_external_symbol
;
5261 u
.cb_data
= &scratch_va
;
5262 u
.rx_va
= shader
->bo
->gpu_address
;
5263 u
.rx_ptr
= sscreen
->ws
->buffer_map(shader
->bo
->buf
, NULL
,
5264 PIPE_TRANSFER_READ_WRITE
|
5265 PIPE_TRANSFER_UNSYNCHRONIZED
|
5266 RADEON_TRANSFER_TEMPORARY
);
5270 bool ok
= ac_rtld_upload(&u
);
5272 sscreen
->ws
->buffer_unmap(shader
->bo
->buf
);
5273 ac_rtld_close(&binary
);
5278 static void si_shader_dump_disassembly(struct si_screen
*screen
,
5279 const struct si_shader_binary
*binary
,
5280 struct pipe_debug_callback
*debug
,
5281 const char *name
, FILE *file
)
5283 struct ac_rtld_binary rtld_binary
;
5285 if (!ac_rtld_open(&rtld_binary
, (struct ac_rtld_open_info
){
5286 .info
= &screen
->info
,
5288 .elf_ptrs
= &binary
->elf_buffer
,
5289 .elf_sizes
= &binary
->elf_size
}))
5295 if (!ac_rtld_get_section_by_name(&rtld_binary
, ".AMDGPU.disasm", &disasm
, &nbytes
))
5298 if (nbytes
> INT_MAX
)
5301 if (debug
&& debug
->debug_message
) {
5302 /* Very long debug messages are cut off, so send the
5303 * disassembly one line at a time. This causes more
5304 * overhead, but on the plus side it simplifies
5305 * parsing of resulting logs.
5307 pipe_debug_message(debug
, SHADER_INFO
,
5308 "Shader Disassembly Begin");
5311 while (line
< nbytes
) {
5312 int count
= nbytes
- line
;
5313 const char *nl
= memchr(disasm
+ line
, '\n', nbytes
- line
);
5315 count
= nl
- (disasm
+ line
);
5318 pipe_debug_message(debug
, SHADER_INFO
,
5319 "%.*s", count
, disasm
+ line
);
5325 pipe_debug_message(debug
, SHADER_INFO
,
5326 "Shader Disassembly End");
5330 fprintf(file
, "Shader %s disassembly:\n", name
);
5331 fprintf(file
, "%*s", (int)nbytes
, disasm
);
5335 ac_rtld_close(&rtld_binary
);
5338 static void si_calculate_max_simd_waves(struct si_shader
*shader
)
5340 struct si_screen
*sscreen
= shader
->selector
->screen
;
5341 struct ac_shader_config
*conf
= &shader
->config
;
5342 unsigned num_inputs
= shader
->selector
->info
.num_inputs
;
5343 unsigned lds_increment
= sscreen
->info
.chip_class
>= GFX7
? 512 : 256;
5344 unsigned lds_per_wave
= 0;
5345 unsigned max_simd_waves
;
5347 max_simd_waves
= ac_get_max_simd_waves(sscreen
->info
.family
);
5349 /* Compute LDS usage for PS. */
5350 switch (shader
->selector
->type
) {
5351 case PIPE_SHADER_FRAGMENT
:
5352 /* The minimum usage per wave is (num_inputs * 48). The maximum
5353 * usage is (num_inputs * 48 * 16).
5354 * We can get anything in between and it varies between waves.
5356 * The 48 bytes per input for a single primitive is equal to
5357 * 4 bytes/component * 4 components/input * 3 points.
5359 * Other stages don't know the size at compile time or don't
5360 * allocate LDS per wave, but instead they do it per thread group.
5362 lds_per_wave
= conf
->lds_size
* lds_increment
+
5363 align(num_inputs
* 48, lds_increment
);
5365 case PIPE_SHADER_COMPUTE
:
5366 if (shader
->selector
) {
5367 unsigned max_workgroup_size
=
5368 si_get_max_workgroup_size(shader
);
5369 lds_per_wave
= (conf
->lds_size
* lds_increment
) /
5370 DIV_ROUND_UP(max_workgroup_size
, 64);
5375 /* Compute the per-SIMD wave counts. */
5376 if (conf
->num_sgprs
) {
5378 MIN2(max_simd_waves
,
5379 ac_get_num_physical_sgprs(sscreen
->info
.chip_class
) / conf
->num_sgprs
);
5382 if (conf
->num_vgprs
)
5383 max_simd_waves
= MIN2(max_simd_waves
, 256 / conf
->num_vgprs
);
5385 /* LDS is 64KB per CU (4 SIMDs), which is 16KB per SIMD (usage above
5386 * 16KB makes some SIMDs unoccupied). */
5388 max_simd_waves
= MIN2(max_simd_waves
, 16384 / lds_per_wave
);
5390 shader
->info
.max_simd_waves
= max_simd_waves
;
5393 void si_shader_dump_stats_for_shader_db(struct si_screen
*screen
,
5394 struct si_shader
*shader
,
5395 struct pipe_debug_callback
*debug
)
5397 const struct ac_shader_config
*conf
= &shader
->config
;
5399 if (screen
->options
.debug_disassembly
)
5400 si_shader_dump_disassembly(screen
, &shader
->binary
, debug
, "main", NULL
);
5402 pipe_debug_message(debug
, SHADER_INFO
,
5403 "Shader Stats: SGPRS: %d VGPRS: %d Code Size: %d "
5404 "LDS: %d Scratch: %d Max Waves: %d Spilled SGPRs: %d "
5405 "Spilled VGPRs: %d PrivMem VGPRs: %d",
5406 conf
->num_sgprs
, conf
->num_vgprs
,
5407 si_get_shader_binary_size(screen
, shader
),
5408 conf
->lds_size
, conf
->scratch_bytes_per_wave
,
5409 shader
->info
.max_simd_waves
, conf
->spilled_sgprs
,
5410 conf
->spilled_vgprs
, shader
->info
.private_mem_vgprs
);
5413 static void si_shader_dump_stats(struct si_screen
*sscreen
,
5414 struct si_shader
*shader
,
5417 bool check_debug_option
)
5419 const struct ac_shader_config
*conf
= &shader
->config
;
5421 if (!check_debug_option
||
5422 si_can_dump_shader(sscreen
, processor
)) {
5423 if (processor
== PIPE_SHADER_FRAGMENT
) {
5424 fprintf(file
, "*** SHADER CONFIG ***\n"
5425 "SPI_PS_INPUT_ADDR = 0x%04x\n"
5426 "SPI_PS_INPUT_ENA = 0x%04x\n",
5427 conf
->spi_ps_input_addr
, conf
->spi_ps_input_ena
);
5430 fprintf(file
, "*** SHADER STATS ***\n"
5433 "Spilled SGPRs: %d\n"
5434 "Spilled VGPRs: %d\n"
5435 "Private memory VGPRs: %d\n"
5436 "Code Size: %d bytes\n"
5438 "Scratch: %d bytes per wave\n"
5440 "********************\n\n\n",
5441 conf
->num_sgprs
, conf
->num_vgprs
,
5442 conf
->spilled_sgprs
, conf
->spilled_vgprs
,
5443 shader
->info
.private_mem_vgprs
,
5444 si_get_shader_binary_size(sscreen
, shader
),
5445 conf
->lds_size
, conf
->scratch_bytes_per_wave
,
5446 shader
->info
.max_simd_waves
);
5450 const char *si_get_shader_name(const struct si_shader
*shader
, unsigned processor
)
5452 switch (processor
) {
5453 case PIPE_SHADER_VERTEX
:
5454 if (shader
->key
.as_es
)
5455 return "Vertex Shader as ES";
5456 else if (shader
->key
.as_ls
)
5457 return "Vertex Shader as LS";
5458 else if (shader
->key
.opt
.vs_as_prim_discard_cs
)
5459 return "Vertex Shader as Primitive Discard CS";
5460 else if (shader
->key
.as_ngg
)
5461 return "Vertex Shader as ESGS";
5463 return "Vertex Shader as VS";
5464 case PIPE_SHADER_TESS_CTRL
:
5465 return "Tessellation Control Shader";
5466 case PIPE_SHADER_TESS_EVAL
:
5467 if (shader
->key
.as_es
)
5468 return "Tessellation Evaluation Shader as ES";
5469 else if (shader
->key
.as_ngg
)
5470 return "Tessellation Evaluation Shader as ESGS";
5472 return "Tessellation Evaluation Shader as VS";
5473 case PIPE_SHADER_GEOMETRY
:
5474 if (shader
->is_gs_copy_shader
)
5475 return "GS Copy Shader as VS";
5477 return "Geometry Shader";
5478 case PIPE_SHADER_FRAGMENT
:
5479 return "Pixel Shader";
5480 case PIPE_SHADER_COMPUTE
:
5481 return "Compute Shader";
5483 return "Unknown Shader";
5487 void si_shader_dump(struct si_screen
*sscreen
, struct si_shader
*shader
,
5488 struct pipe_debug_callback
*debug
, unsigned processor
,
5489 FILE *file
, bool check_debug_option
)
5491 if (!check_debug_option
||
5492 si_can_dump_shader(sscreen
, processor
))
5493 si_dump_shader_key(processor
, shader
, file
);
5495 if (!check_debug_option
&& shader
->binary
.llvm_ir_string
) {
5496 if (shader
->previous_stage
&&
5497 shader
->previous_stage
->binary
.llvm_ir_string
) {
5498 fprintf(file
, "\n%s - previous stage - LLVM IR:\n\n",
5499 si_get_shader_name(shader
, processor
));
5500 fprintf(file
, "%s\n", shader
->previous_stage
->binary
.llvm_ir_string
);
5503 fprintf(file
, "\n%s - main shader part - LLVM IR:\n\n",
5504 si_get_shader_name(shader
, processor
));
5505 fprintf(file
, "%s\n", shader
->binary
.llvm_ir_string
);
5508 if (!check_debug_option
||
5509 (si_can_dump_shader(sscreen
, processor
) &&
5510 !(sscreen
->debug_flags
& DBG(NO_ASM
)))) {
5511 fprintf(file
, "\n%s:\n", si_get_shader_name(shader
, processor
));
5514 si_shader_dump_disassembly(sscreen
, &shader
->prolog
->binary
,
5515 debug
, "prolog", file
);
5516 if (shader
->previous_stage
)
5517 si_shader_dump_disassembly(sscreen
, &shader
->previous_stage
->binary
,
5518 debug
, "previous stage", file
);
5519 if (shader
->prolog2
)
5520 si_shader_dump_disassembly(sscreen
, &shader
->prolog2
->binary
,
5521 debug
, "prolog2", file
);
5523 si_shader_dump_disassembly(sscreen
, &shader
->binary
, debug
, "main", file
);
5526 si_shader_dump_disassembly(sscreen
, &shader
->epilog
->binary
,
5527 debug
, "epilog", file
);
5528 fprintf(file
, "\n");
5531 si_shader_dump_stats(sscreen
, shader
, processor
, file
,
5532 check_debug_option
);
5535 static int si_compile_llvm(struct si_screen
*sscreen
,
5536 struct si_shader_binary
*binary
,
5537 struct ac_shader_config
*conf
,
5538 struct ac_llvm_compiler
*compiler
,
5540 struct pipe_debug_callback
*debug
,
5543 bool less_optimized
)
5545 unsigned count
= p_atomic_inc_return(&sscreen
->num_compilations
);
5547 if (si_can_dump_shader(sscreen
, processor
)) {
5548 fprintf(stderr
, "radeonsi: Compiling shader %d\n", count
);
5550 if (!(sscreen
->debug_flags
& (DBG(NO_IR
) | DBG(PREOPT_IR
)))) {
5551 fprintf(stderr
, "%s LLVM IR:\n\n", name
);
5552 ac_dump_module(mod
);
5553 fprintf(stderr
, "\n");
5557 if (sscreen
->record_llvm_ir
) {
5558 char *ir
= LLVMPrintModuleToString(mod
);
5559 binary
->llvm_ir_string
= strdup(ir
);
5560 LLVMDisposeMessage(ir
);
5563 if (!si_replace_shader(count
, binary
)) {
5564 unsigned r
= si_llvm_compile(mod
, binary
, compiler
, debug
,
5570 struct ac_rtld_binary rtld
;
5571 if (!ac_rtld_open(&rtld
, (struct ac_rtld_open_info
){
5572 .info
= &sscreen
->info
,
5574 .elf_ptrs
= &binary
->elf_buffer
,
5575 .elf_sizes
= &binary
->elf_size
}))
5578 bool ok
= ac_rtld_read_config(&rtld
, conf
);
5579 ac_rtld_close(&rtld
);
5583 /* Enable 64-bit and 16-bit denormals, because there is no performance
5586 * If denormals are enabled, all floating-point output modifiers are
5589 * Don't enable denormals for 32-bit floats, because:
5590 * - Floating-point output modifiers would be ignored by the hw.
5591 * - Some opcodes don't support denormals, such as v_mad_f32. We would
5592 * have to stop using those.
5593 * - GFX6 & GFX7 would be very slow.
5595 conf
->float_mode
|= V_00B028_FP_64_DENORMS
;
5600 static void si_llvm_build_ret(struct si_shader_context
*ctx
, LLVMValueRef ret
)
5602 if (LLVMGetTypeKind(LLVMTypeOf(ret
)) == LLVMVoidTypeKind
)
5603 LLVMBuildRetVoid(ctx
->ac
.builder
);
5605 LLVMBuildRet(ctx
->ac
.builder
, ret
);
5608 /* Generate code for the hardware VS shader stage to go with a geometry shader */
5610 si_generate_gs_copy_shader(struct si_screen
*sscreen
,
5611 struct ac_llvm_compiler
*compiler
,
5612 struct si_shader_selector
*gs_selector
,
5613 struct pipe_debug_callback
*debug
)
5615 struct si_shader_context ctx
;
5616 struct si_shader
*shader
;
5617 LLVMBuilderRef builder
;
5618 struct si_shader_output_values outputs
[SI_MAX_VS_OUTPUTS
];
5619 struct tgsi_shader_info
*gsinfo
= &gs_selector
->info
;
5623 shader
= CALLOC_STRUCT(si_shader
);
5627 /* We can leave the fence as permanently signaled because the GS copy
5628 * shader only becomes visible globally after it has been compiled. */
5629 util_queue_fence_init(&shader
->ready
);
5631 shader
->selector
= gs_selector
;
5632 shader
->is_gs_copy_shader
= true;
5634 si_init_shader_ctx(&ctx
, sscreen
, compiler
);
5635 ctx
.shader
= shader
;
5636 ctx
.type
= PIPE_SHADER_VERTEX
;
5638 builder
= ctx
.ac
.builder
;
5640 create_function(&ctx
);
5641 preload_ring_buffers(&ctx
);
5643 LLVMValueRef voffset
=
5644 LLVMBuildMul(ctx
.ac
.builder
, ctx
.abi
.vertex_id
,
5645 LLVMConstInt(ctx
.i32
, 4, 0), "");
5647 /* Fetch the vertex stream ID.*/
5648 LLVMValueRef stream_id
;
5650 if (gs_selector
->so
.num_outputs
)
5651 stream_id
= si_unpack_param(&ctx
, ctx
.param_streamout_config
, 24, 2);
5653 stream_id
= ctx
.i32_0
;
5655 /* Fill in output information. */
5656 for (i
= 0; i
< gsinfo
->num_outputs
; ++i
) {
5657 outputs
[i
].semantic_name
= gsinfo
->output_semantic_name
[i
];
5658 outputs
[i
].semantic_index
= gsinfo
->output_semantic_index
[i
];
5660 for (int chan
= 0; chan
< 4; chan
++) {
5661 outputs
[i
].vertex_stream
[chan
] =
5662 (gsinfo
->output_streams
[i
] >> (2 * chan
)) & 3;
5666 LLVMBasicBlockRef end_bb
;
5667 LLVMValueRef switch_inst
;
5669 end_bb
= LLVMAppendBasicBlockInContext(ctx
.ac
.context
, ctx
.main_fn
, "end");
5670 switch_inst
= LLVMBuildSwitch(builder
, stream_id
, end_bb
, 4);
5672 for (int stream
= 0; stream
< 4; stream
++) {
5673 LLVMBasicBlockRef bb
;
5676 if (!gsinfo
->num_stream_output_components
[stream
])
5679 if (stream
> 0 && !gs_selector
->so
.num_outputs
)
5682 bb
= LLVMInsertBasicBlockInContext(ctx
.ac
.context
, end_bb
, "out");
5683 LLVMAddCase(switch_inst
, LLVMConstInt(ctx
.i32
, stream
, 0), bb
);
5684 LLVMPositionBuilderAtEnd(builder
, bb
);
5686 /* Fetch vertex data from GSVS ring */
5688 for (i
= 0; i
< gsinfo
->num_outputs
; ++i
) {
5689 for (unsigned chan
= 0; chan
< 4; chan
++) {
5690 if (!(gsinfo
->output_usagemask
[i
] & (1 << chan
)) ||
5691 outputs
[i
].vertex_stream
[chan
] != stream
) {
5692 outputs
[i
].values
[chan
] = LLVMGetUndef(ctx
.f32
);
5696 LLVMValueRef soffset
= LLVMConstInt(ctx
.i32
,
5697 offset
* gs_selector
->gs_max_out_vertices
* 16 * 4, 0);
5700 outputs
[i
].values
[chan
] =
5701 ac_build_buffer_load(&ctx
.ac
,
5702 ctx
.gsvs_ring
[0], 1,
5709 /* Streamout and exports. */
5710 if (gs_selector
->so
.num_outputs
) {
5711 si_llvm_emit_streamout(&ctx
, outputs
,
5712 gsinfo
->num_outputs
,
5717 si_llvm_export_vs(&ctx
, outputs
, gsinfo
->num_outputs
);
5719 LLVMBuildBr(builder
, end_bb
);
5722 LLVMPositionBuilderAtEnd(builder
, end_bb
);
5724 LLVMBuildRetVoid(ctx
.ac
.builder
);
5726 ctx
.type
= PIPE_SHADER_GEOMETRY
; /* override for shader dumping */
5727 si_llvm_optimize_module(&ctx
);
5730 if (si_compile_llvm(sscreen
, &ctx
.shader
->binary
,
5731 &ctx
.shader
->config
, ctx
.compiler
,
5733 debug
, PIPE_SHADER_GEOMETRY
,
5734 "GS Copy Shader", false) == 0) {
5735 if (si_can_dump_shader(sscreen
, PIPE_SHADER_GEOMETRY
))
5736 fprintf(stderr
, "GS Copy Shader:\n");
5737 si_shader_dump(sscreen
, ctx
.shader
, debug
,
5738 PIPE_SHADER_GEOMETRY
, stderr
, true);
5740 if (!ctx
.shader
->config
.scratch_bytes_per_wave
)
5741 ok
= si_shader_binary_upload(sscreen
, ctx
.shader
, 0);
5746 si_llvm_dispose(&ctx
);
5752 si_fix_resource_usage(sscreen
, shader
);
5757 static void si_dump_shader_key_vs(const struct si_shader_key
*key
,
5758 const struct si_vs_prolog_bits
*prolog
,
5759 const char *prefix
, FILE *f
)
5761 fprintf(f
, " %s.instance_divisor_is_one = %u\n",
5762 prefix
, prolog
->instance_divisor_is_one
);
5763 fprintf(f
, " %s.instance_divisor_is_fetched = %u\n",
5764 prefix
, prolog
->instance_divisor_is_fetched
);
5765 fprintf(f
, " %s.unpack_instance_id_from_vertex_id = %u\n",
5766 prefix
, prolog
->unpack_instance_id_from_vertex_id
);
5767 fprintf(f
, " %s.ls_vgpr_fix = %u\n",
5768 prefix
, prolog
->ls_vgpr_fix
);
5770 fprintf(f
, " mono.vs.fetch_opencode = %x\n", key
->mono
.vs_fetch_opencode
);
5771 fprintf(f
, " mono.vs.fix_fetch = {");
5772 for (int i
= 0; i
< SI_MAX_ATTRIBS
; i
++) {
5773 union si_vs_fix_fetch fix
= key
->mono
.vs_fix_fetch
[i
];
5779 fprintf(f
, "%u.%u.%u.%u", fix
.u
.reverse
, fix
.u
.log_size
,
5780 fix
.u
.num_channels_m1
, fix
.u
.format
);
5785 static void si_dump_shader_key(unsigned processor
, const struct si_shader
*shader
,
5788 const struct si_shader_key
*key
= &shader
->key
;
5790 fprintf(f
, "SHADER KEY\n");
5792 switch (processor
) {
5793 case PIPE_SHADER_VERTEX
:
5794 si_dump_shader_key_vs(key
, &key
->part
.vs
.prolog
,
5795 "part.vs.prolog", f
);
5796 fprintf(f
, " as_es = %u\n", key
->as_es
);
5797 fprintf(f
, " as_ls = %u\n", key
->as_ls
);
5798 fprintf(f
, " mono.u.vs_export_prim_id = %u\n",
5799 key
->mono
.u
.vs_export_prim_id
);
5800 fprintf(f
, " opt.vs_as_prim_discard_cs = %u\n",
5801 key
->opt
.vs_as_prim_discard_cs
);
5802 fprintf(f
, " opt.cs_prim_type = %s\n",
5803 tgsi_primitive_names
[key
->opt
.cs_prim_type
]);
5804 fprintf(f
, " opt.cs_indexed = %u\n",
5805 key
->opt
.cs_indexed
);
5806 fprintf(f
, " opt.cs_instancing = %u\n",
5807 key
->opt
.cs_instancing
);
5808 fprintf(f
, " opt.cs_primitive_restart = %u\n",
5809 key
->opt
.cs_primitive_restart
);
5810 fprintf(f
, " opt.cs_provoking_vertex_first = %u\n",
5811 key
->opt
.cs_provoking_vertex_first
);
5812 fprintf(f
, " opt.cs_need_correct_orientation = %u\n",
5813 key
->opt
.cs_need_correct_orientation
);
5814 fprintf(f
, " opt.cs_cull_front = %u\n",
5815 key
->opt
.cs_cull_front
);
5816 fprintf(f
, " opt.cs_cull_back = %u\n",
5817 key
->opt
.cs_cull_back
);
5818 fprintf(f
, " opt.cs_cull_z = %u\n",
5819 key
->opt
.cs_cull_z
);
5820 fprintf(f
, " opt.cs_halfz_clip_space = %u\n",
5821 key
->opt
.cs_halfz_clip_space
);
5824 case PIPE_SHADER_TESS_CTRL
:
5825 if (shader
->selector
->screen
->info
.chip_class
>= GFX9
) {
5826 si_dump_shader_key_vs(key
, &key
->part
.tcs
.ls_prolog
,
5827 "part.tcs.ls_prolog", f
);
5829 fprintf(f
, " part.tcs.epilog.prim_mode = %u\n", key
->part
.tcs
.epilog
.prim_mode
);
5830 fprintf(f
, " mono.u.ff_tcs_inputs_to_copy = 0x%"PRIx64
"\n", key
->mono
.u
.ff_tcs_inputs_to_copy
);
5833 case PIPE_SHADER_TESS_EVAL
:
5834 fprintf(f
, " as_es = %u\n", key
->as_es
);
5835 fprintf(f
, " mono.u.vs_export_prim_id = %u\n",
5836 key
->mono
.u
.vs_export_prim_id
);
5839 case PIPE_SHADER_GEOMETRY
:
5840 if (shader
->is_gs_copy_shader
)
5843 if (shader
->selector
->screen
->info
.chip_class
>= GFX9
&&
5844 key
->part
.gs
.es
->type
== PIPE_SHADER_VERTEX
) {
5845 si_dump_shader_key_vs(key
, &key
->part
.gs
.vs_prolog
,
5846 "part.gs.vs_prolog", f
);
5848 fprintf(f
, " part.gs.prolog.tri_strip_adj_fix = %u\n", key
->part
.gs
.prolog
.tri_strip_adj_fix
);
5851 case PIPE_SHADER_COMPUTE
:
5854 case PIPE_SHADER_FRAGMENT
:
5855 fprintf(f
, " part.ps.prolog.color_two_side = %u\n", key
->part
.ps
.prolog
.color_two_side
);
5856 fprintf(f
, " part.ps.prolog.flatshade_colors = %u\n", key
->part
.ps
.prolog
.flatshade_colors
);
5857 fprintf(f
, " part.ps.prolog.poly_stipple = %u\n", key
->part
.ps
.prolog
.poly_stipple
);
5858 fprintf(f
, " part.ps.prolog.force_persp_sample_interp = %u\n", key
->part
.ps
.prolog
.force_persp_sample_interp
);
5859 fprintf(f
, " part.ps.prolog.force_linear_sample_interp = %u\n", key
->part
.ps
.prolog
.force_linear_sample_interp
);
5860 fprintf(f
, " part.ps.prolog.force_persp_center_interp = %u\n", key
->part
.ps
.prolog
.force_persp_center_interp
);
5861 fprintf(f
, " part.ps.prolog.force_linear_center_interp = %u\n", key
->part
.ps
.prolog
.force_linear_center_interp
);
5862 fprintf(f
, " part.ps.prolog.bc_optimize_for_persp = %u\n", key
->part
.ps
.prolog
.bc_optimize_for_persp
);
5863 fprintf(f
, " part.ps.prolog.bc_optimize_for_linear = %u\n", key
->part
.ps
.prolog
.bc_optimize_for_linear
);
5864 fprintf(f
, " part.ps.epilog.spi_shader_col_format = 0x%x\n", key
->part
.ps
.epilog
.spi_shader_col_format
);
5865 fprintf(f
, " part.ps.epilog.color_is_int8 = 0x%X\n", key
->part
.ps
.epilog
.color_is_int8
);
5866 fprintf(f
, " part.ps.epilog.color_is_int10 = 0x%X\n", key
->part
.ps
.epilog
.color_is_int10
);
5867 fprintf(f
, " part.ps.epilog.last_cbuf = %u\n", key
->part
.ps
.epilog
.last_cbuf
);
5868 fprintf(f
, " part.ps.epilog.alpha_func = %u\n", key
->part
.ps
.epilog
.alpha_func
);
5869 fprintf(f
, " part.ps.epilog.alpha_to_one = %u\n", key
->part
.ps
.epilog
.alpha_to_one
);
5870 fprintf(f
, " part.ps.epilog.poly_line_smoothing = %u\n", key
->part
.ps
.epilog
.poly_line_smoothing
);
5871 fprintf(f
, " part.ps.epilog.clamp_color = %u\n", key
->part
.ps
.epilog
.clamp_color
);
5878 if ((processor
== PIPE_SHADER_GEOMETRY
||
5879 processor
== PIPE_SHADER_TESS_EVAL
||
5880 processor
== PIPE_SHADER_VERTEX
) &&
5881 !key
->as_es
&& !key
->as_ls
) {
5882 fprintf(f
, " opt.kill_outputs = 0x%"PRIx64
"\n", key
->opt
.kill_outputs
);
5883 fprintf(f
, " opt.clip_disable = %u\n", key
->opt
.clip_disable
);
5887 static void si_init_shader_ctx(struct si_shader_context
*ctx
,
5888 struct si_screen
*sscreen
,
5889 struct ac_llvm_compiler
*compiler
)
5891 struct lp_build_tgsi_context
*bld_base
;
5893 si_llvm_context_init(ctx
, sscreen
, compiler
);
5895 bld_base
= &ctx
->bld_base
;
5896 bld_base
->emit_fetch_funcs
[TGSI_FILE_CONSTANT
] = fetch_constant
;
5898 bld_base
->op_actions
[TGSI_OPCODE_INTERP_CENTROID
].emit
= build_interp_intrinsic
;
5899 bld_base
->op_actions
[TGSI_OPCODE_INTERP_SAMPLE
].emit
= build_interp_intrinsic
;
5900 bld_base
->op_actions
[TGSI_OPCODE_INTERP_OFFSET
].emit
= build_interp_intrinsic
;
5902 bld_base
->op_actions
[TGSI_OPCODE_MEMBAR
].emit
= membar_emit
;
5904 bld_base
->op_actions
[TGSI_OPCODE_CLOCK
].emit
= clock_emit
;
5906 bld_base
->op_actions
[TGSI_OPCODE_DDX
].emit
= si_llvm_emit_ddxy
;
5907 bld_base
->op_actions
[TGSI_OPCODE_DDY
].emit
= si_llvm_emit_ddxy
;
5908 bld_base
->op_actions
[TGSI_OPCODE_DDX_FINE
].emit
= si_llvm_emit_ddxy
;
5909 bld_base
->op_actions
[TGSI_OPCODE_DDY_FINE
].emit
= si_llvm_emit_ddxy
;
5911 bld_base
->op_actions
[TGSI_OPCODE_VOTE_ALL
].emit
= vote_all_emit
;
5912 bld_base
->op_actions
[TGSI_OPCODE_VOTE_ANY
].emit
= vote_any_emit
;
5913 bld_base
->op_actions
[TGSI_OPCODE_VOTE_EQ
].emit
= vote_eq_emit
;
5914 bld_base
->op_actions
[TGSI_OPCODE_BALLOT
].emit
= ballot_emit
;
5915 bld_base
->op_actions
[TGSI_OPCODE_READ_FIRST
].intr_name
= "llvm.amdgcn.readfirstlane";
5916 bld_base
->op_actions
[TGSI_OPCODE_READ_FIRST
].emit
= read_lane_emit
;
5917 bld_base
->op_actions
[TGSI_OPCODE_READ_INVOC
].intr_name
= "llvm.amdgcn.readlane";
5918 bld_base
->op_actions
[TGSI_OPCODE_READ_INVOC
].emit
= read_lane_emit
;
5920 bld_base
->op_actions
[TGSI_OPCODE_EMIT
].emit
= si_tgsi_emit_vertex
;
5921 bld_base
->op_actions
[TGSI_OPCODE_ENDPRIM
].emit
= si_tgsi_emit_primitive
;
5922 bld_base
->op_actions
[TGSI_OPCODE_BARRIER
].emit
= si_llvm_emit_barrier
;
5925 static void si_optimize_vs_outputs(struct si_shader_context
*ctx
)
5927 struct si_shader
*shader
= ctx
->shader
;
5928 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
5930 if ((ctx
->type
!= PIPE_SHADER_VERTEX
&&
5931 ctx
->type
!= PIPE_SHADER_TESS_EVAL
) ||
5932 shader
->key
.as_ls
||
5936 ac_optimize_vs_outputs(&ctx
->ac
,
5938 shader
->info
.vs_output_param_offset
,
5940 &shader
->info
.nr_param_exports
);
5943 static void si_init_exec_from_input(struct si_shader_context
*ctx
,
5944 unsigned param
, unsigned bitoffset
)
5946 LLVMValueRef args
[] = {
5947 LLVMGetParam(ctx
->main_fn
, param
),
5948 LLVMConstInt(ctx
->i32
, bitoffset
, 0),
5950 ac_build_intrinsic(&ctx
->ac
,
5951 "llvm.amdgcn.init.exec.from.input",
5952 ctx
->voidt
, args
, 2, AC_FUNC_ATTR_CONVERGENT
);
5955 static bool si_vs_needs_prolog(const struct si_shader_selector
*sel
,
5956 const struct si_vs_prolog_bits
*key
)
5958 /* VGPR initialization fixup for Vega10 and Raven is always done in the
5960 return sel
->vs_needs_prolog
|| key
->ls_vgpr_fix
;
5963 static bool si_compile_tgsi_main(struct si_shader_context
*ctx
)
5965 struct si_shader
*shader
= ctx
->shader
;
5966 struct si_shader_selector
*sel
= shader
->selector
;
5967 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
5969 // TODO clean all this up!
5970 switch (ctx
->type
) {
5971 case PIPE_SHADER_VERTEX
:
5972 ctx
->load_input
= declare_input_vs
;
5973 if (shader
->key
.as_ls
)
5974 ctx
->abi
.emit_outputs
= si_llvm_emit_ls_epilogue
;
5975 else if (shader
->key
.as_es
)
5976 ctx
->abi
.emit_outputs
= si_llvm_emit_es_epilogue
;
5977 else if (shader
->key
.opt
.vs_as_prim_discard_cs
)
5978 ctx
->abi
.emit_outputs
= si_llvm_emit_prim_discard_cs_epilogue
;
5979 else if (shader
->key
.as_ngg
)
5980 ctx
->abi
.emit_outputs
= gfx10_emit_ngg_epilogue
;
5982 ctx
->abi
.emit_outputs
= si_llvm_emit_vs_epilogue
;
5983 bld_base
->emit_epilogue
= si_tgsi_emit_epilogue
;
5984 ctx
->abi
.load_base_vertex
= get_base_vertex
;
5986 case PIPE_SHADER_TESS_CTRL
:
5987 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = fetch_input_tcs
;
5988 ctx
->abi
.load_tess_varyings
= si_nir_load_tcs_varyings
;
5989 bld_base
->emit_fetch_funcs
[TGSI_FILE_OUTPUT
] = fetch_output_tcs
;
5990 bld_base
->emit_store
= store_output_tcs
;
5991 ctx
->abi
.store_tcs_outputs
= si_nir_store_output_tcs
;
5992 ctx
->abi
.emit_outputs
= si_llvm_emit_tcs_epilogue
;
5993 ctx
->abi
.load_patch_vertices_in
= si_load_patch_vertices_in
;
5994 bld_base
->emit_epilogue
= si_tgsi_emit_epilogue
;
5996 case PIPE_SHADER_TESS_EVAL
:
5997 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = fetch_input_tes
;
5998 ctx
->abi
.load_tess_varyings
= si_nir_load_input_tes
;
5999 ctx
->abi
.load_tess_coord
= si_load_tess_coord
;
6000 ctx
->abi
.load_tess_level
= si_load_tess_level
;
6001 ctx
->abi
.load_patch_vertices_in
= si_load_patch_vertices_in
;
6002 if (shader
->key
.as_es
)
6003 ctx
->abi
.emit_outputs
= si_llvm_emit_es_epilogue
;
6005 if (shader
->key
.as_ngg
)
6006 ctx
->abi
.emit_outputs
= gfx10_emit_ngg_epilogue
;
6008 ctx
->abi
.emit_outputs
= si_llvm_emit_vs_epilogue
;
6010 bld_base
->emit_epilogue
= si_tgsi_emit_epilogue
;
6012 case PIPE_SHADER_GEOMETRY
:
6013 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = fetch_input_gs
;
6014 ctx
->abi
.load_inputs
= si_nir_load_input_gs
;
6015 ctx
->abi
.emit_vertex
= si_llvm_emit_vertex
;
6016 ctx
->abi
.emit_primitive
= si_llvm_emit_primitive
;
6017 ctx
->abi
.emit_outputs
= si_llvm_emit_gs_epilogue
;
6018 bld_base
->emit_epilogue
= si_tgsi_emit_gs_epilogue
;
6020 case PIPE_SHADER_FRAGMENT
:
6021 ctx
->load_input
= declare_input_fs
;
6022 ctx
->abi
.emit_outputs
= si_llvm_return_fs_outputs
;
6023 bld_base
->emit_epilogue
= si_tgsi_emit_epilogue
;
6024 ctx
->abi
.lookup_interp_param
= si_nir_lookup_interp_param
;
6025 ctx
->abi
.load_sample_position
= load_sample_position
;
6026 ctx
->abi
.load_sample_mask_in
= load_sample_mask_in
;
6027 ctx
->abi
.emit_kill
= si_llvm_emit_kill
;
6029 case PIPE_SHADER_COMPUTE
:
6030 ctx
->abi
.load_local_group_size
= get_block_size
;
6033 assert(!"Unsupported shader type");
6037 ctx
->abi
.load_ubo
= load_ubo
;
6038 ctx
->abi
.load_ssbo
= load_ssbo
;
6040 create_function(ctx
);
6041 preload_ring_buffers(ctx
);
6043 /* For GFX9 merged shaders:
6044 * - Set EXEC for the first shader. If the prolog is present, set
6045 * EXEC there instead.
6046 * - Add a barrier before the second shader.
6047 * - In the second shader, reset EXEC to ~0 and wrap the main part in
6048 * an if-statement. This is required for correctness in geometry
6049 * shaders, to ensure that empty GS waves do not send GS_EMIT and
6052 * For monolithic merged shaders, the first shader is wrapped in an
6053 * if-block together with its prolog in si_build_wrapper_function.
6055 * NGG vertex and tess eval shaders running as the last
6056 * vertex/geometry stage handle execution explicitly using
6059 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
6060 if (!shader
->is_monolithic
&&
6061 sel
->info
.num_instructions
> 1 && /* not empty shader */
6062 (shader
->key
.as_es
|| shader
->key
.as_ls
) &&
6063 (ctx
->type
== PIPE_SHADER_TESS_EVAL
||
6064 (ctx
->type
== PIPE_SHADER_VERTEX
&&
6065 !si_vs_needs_prolog(sel
, &shader
->key
.part
.vs
.prolog
)))) {
6066 si_init_exec_from_input(ctx
,
6067 ctx
->param_merged_wave_info
, 0);
6068 } else if (ctx
->type
== PIPE_SHADER_TESS_CTRL
||
6069 ctx
->type
== PIPE_SHADER_GEOMETRY
||
6070 shader
->key
.as_ngg
) {
6071 LLVMValueRef num_threads
;
6072 bool nested_barrier
;
6074 if (!shader
->is_monolithic
)
6075 ac_init_exec_full_mask(&ctx
->ac
);
6077 if (ctx
->type
== PIPE_SHADER_TESS_CTRL
||
6078 ctx
->type
== PIPE_SHADER_GEOMETRY
) {
6079 /* Number of patches / primitives */
6080 num_threads
= si_unpack_param(ctx
, ctx
->param_merged_wave_info
, 8, 8);
6081 nested_barrier
= true;
6083 /* Number of vertices */
6084 num_threads
= si_unpack_param(ctx
, ctx
->param_merged_wave_info
, 0, 8);
6085 nested_barrier
= false;
6089 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntULT
,
6090 ac_get_thread_id(&ctx
->ac
), num_threads
, "");
6091 lp_build_if(&ctx
->merged_wrap_if_state
, &ctx
->gallivm
, ena
);
6093 if (nested_barrier
) {
6094 /* Execute a barrier before the second shader in
6097 * Execute the barrier inside the conditional block,
6098 * so that empty waves can jump directly to s_endpgm,
6099 * which will also signal the barrier.
6101 * This is possible in gfx9, because an empty wave
6102 * for the second shader does not participate in
6103 * the epilogue. With NGG, empty waves may still
6104 * be required to export data (e.g. GS output vertices),
6105 * so we cannot let them exit early.
6107 * If the shader is TCS and the TCS epilog is present
6108 * and contains a barrier, it will wait there and then
6111 si_llvm_emit_barrier(NULL
, bld_base
, NULL
);
6116 if (ctx
->type
== PIPE_SHADER_TESS_CTRL
&&
6117 sel
->tcs_info
.tessfactors_are_def_in_all_invocs
) {
6118 for (unsigned i
= 0; i
< 6; i
++) {
6119 ctx
->invoc0_tess_factors
[i
] =
6120 ac_build_alloca_undef(&ctx
->ac
, ctx
->i32
, "");
6124 if (ctx
->type
== PIPE_SHADER_GEOMETRY
) {
6125 for (unsigned i
= 0; i
< 4; i
++) {
6126 ctx
->gs_next_vertex
[i
] =
6127 ac_build_alloca(&ctx
->ac
, ctx
->i32
, "");
6129 if (shader
->key
.as_ngg
) {
6130 for (unsigned i
= 0; i
< 4; ++i
) {
6131 ctx
->gs_curprim_verts
[i
] =
6132 lp_build_alloca(&ctx
->gallivm
, ctx
->ac
.i32
, "");
6135 LLVMTypeRef a8i32
= LLVMArrayType(ctx
->i32
, 8);
6136 ctx
->gs_ngg_scratch
= LLVMAddGlobalInAddressSpace(ctx
->ac
.module
,
6137 a8i32
, "ngg_scratch", AC_ADDR_SPACE_LDS
);
6138 LLVMSetInitializer(ctx
->gs_ngg_scratch
, LLVMGetUndef(a8i32
));
6139 LLVMSetAlignment(ctx
->gs_ngg_scratch
, 4);
6141 ctx
->gs_ngg_emit
= LLVMAddGlobalInAddressSpace(ctx
->ac
.module
,
6142 LLVMArrayType(ctx
->i32
, 0), "ngg_emit", AC_ADDR_SPACE_LDS
);
6143 LLVMSetLinkage(ctx
->gs_ngg_emit
, LLVMExternalLinkage
);
6144 LLVMSetAlignment(ctx
->gs_ngg_emit
, 4);
6148 if (sel
->force_correct_derivs_after_kill
) {
6149 ctx
->postponed_kill
= ac_build_alloca_undef(&ctx
->ac
, ctx
->i1
, "");
6150 /* true = don't kill. */
6151 LLVMBuildStore(ctx
->ac
.builder
, ctx
->i1true
,
6152 ctx
->postponed_kill
);
6156 if (!lp_build_tgsi_llvm(bld_base
, sel
->tokens
)) {
6157 fprintf(stderr
, "Failed to translate shader from TGSI to LLVM\n");
6161 if (!si_nir_build_llvm(ctx
, sel
->nir
)) {
6162 fprintf(stderr
, "Failed to translate shader from NIR to LLVM\n");
6167 si_llvm_build_ret(ctx
, ctx
->return_value
);
6172 * Compute the VS prolog key, which contains all the information needed to
6173 * build the VS prolog function, and set shader->info bits where needed.
6175 * \param info Shader info of the vertex shader.
6176 * \param num_input_sgprs Number of input SGPRs for the vertex shader.
6177 * \param prolog_key Key of the VS prolog
6178 * \param shader_out The vertex shader, or the next shader if merging LS+HS or ES+GS.
6179 * \param key Output shader part key.
6181 static void si_get_vs_prolog_key(const struct tgsi_shader_info
*info
,
6182 unsigned num_input_sgprs
,
6183 const struct si_vs_prolog_bits
*prolog_key
,
6184 struct si_shader
*shader_out
,
6185 union si_shader_part_key
*key
)
6187 memset(key
, 0, sizeof(*key
));
6188 key
->vs_prolog
.states
= *prolog_key
;
6189 key
->vs_prolog
.num_input_sgprs
= num_input_sgprs
;
6190 key
->vs_prolog
.last_input
= MAX2(1, info
->num_inputs
) - 1;
6191 key
->vs_prolog
.as_ls
= shader_out
->key
.as_ls
;
6192 key
->vs_prolog
.as_es
= shader_out
->key
.as_es
;
6193 key
->vs_prolog
.as_ngg
= shader_out
->key
.as_ngg
;
6195 if (shader_out
->selector
->type
== PIPE_SHADER_TESS_CTRL
) {
6196 key
->vs_prolog
.as_ls
= 1;
6197 key
->vs_prolog
.num_merged_next_stage_vgprs
= 2;
6198 } else if (shader_out
->selector
->type
== PIPE_SHADER_GEOMETRY
) {
6199 key
->vs_prolog
.as_es
= 1;
6200 key
->vs_prolog
.num_merged_next_stage_vgprs
= 5;
6201 } else if (shader_out
->key
.as_ngg
) {
6202 key
->vs_prolog
.num_merged_next_stage_vgprs
= 5;
6205 /* Enable loading the InstanceID VGPR. */
6206 uint16_t input_mask
= u_bit_consecutive(0, info
->num_inputs
);
6208 if ((key
->vs_prolog
.states
.instance_divisor_is_one
|
6209 key
->vs_prolog
.states
.instance_divisor_is_fetched
) & input_mask
)
6210 shader_out
->info
.uses_instanceid
= true;
6214 * Compute the PS prolog key, which contains all the information needed to
6215 * build the PS prolog function, and set related bits in shader->config.
6217 static void si_get_ps_prolog_key(struct si_shader
*shader
,
6218 union si_shader_part_key
*key
,
6219 bool separate_prolog
)
6221 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
6223 memset(key
, 0, sizeof(*key
));
6224 key
->ps_prolog
.states
= shader
->key
.part
.ps
.prolog
;
6225 key
->ps_prolog
.colors_read
= info
->colors_read
;
6226 key
->ps_prolog
.num_input_sgprs
= shader
->info
.num_input_sgprs
;
6227 key
->ps_prolog
.num_input_vgprs
= shader
->info
.num_input_vgprs
;
6228 key
->ps_prolog
.wqm
= info
->uses_derivatives
&&
6229 (key
->ps_prolog
.colors_read
||
6230 key
->ps_prolog
.states
.force_persp_sample_interp
||
6231 key
->ps_prolog
.states
.force_linear_sample_interp
||
6232 key
->ps_prolog
.states
.force_persp_center_interp
||
6233 key
->ps_prolog
.states
.force_linear_center_interp
||
6234 key
->ps_prolog
.states
.bc_optimize_for_persp
||
6235 key
->ps_prolog
.states
.bc_optimize_for_linear
);
6236 key
->ps_prolog
.ancillary_vgpr_index
= shader
->info
.ancillary_vgpr_index
;
6238 if (info
->colors_read
) {
6239 unsigned *color
= shader
->selector
->color_attr_index
;
6241 if (shader
->key
.part
.ps
.prolog
.color_two_side
) {
6242 /* BCOLORs are stored after the last input. */
6243 key
->ps_prolog
.num_interp_inputs
= info
->num_inputs
;
6244 key
->ps_prolog
.face_vgpr_index
= shader
->info
.face_vgpr_index
;
6245 if (separate_prolog
)
6246 shader
->config
.spi_ps_input_ena
|= S_0286CC_FRONT_FACE_ENA(1);
6249 for (unsigned i
= 0; i
< 2; i
++) {
6250 unsigned interp
= info
->input_interpolate
[color
[i
]];
6251 unsigned location
= info
->input_interpolate_loc
[color
[i
]];
6253 if (!(info
->colors_read
& (0xf << i
*4)))
6256 key
->ps_prolog
.color_attr_index
[i
] = color
[i
];
6258 if (shader
->key
.part
.ps
.prolog
.flatshade_colors
&&
6259 interp
== TGSI_INTERPOLATE_COLOR
)
6260 interp
= TGSI_INTERPOLATE_CONSTANT
;
6263 case TGSI_INTERPOLATE_CONSTANT
:
6264 key
->ps_prolog
.color_interp_vgpr_index
[i
] = -1;
6266 case TGSI_INTERPOLATE_PERSPECTIVE
:
6267 case TGSI_INTERPOLATE_COLOR
:
6268 /* Force the interpolation location for colors here. */
6269 if (shader
->key
.part
.ps
.prolog
.force_persp_sample_interp
)
6270 location
= TGSI_INTERPOLATE_LOC_SAMPLE
;
6271 if (shader
->key
.part
.ps
.prolog
.force_persp_center_interp
)
6272 location
= TGSI_INTERPOLATE_LOC_CENTER
;
6275 case TGSI_INTERPOLATE_LOC_SAMPLE
:
6276 key
->ps_prolog
.color_interp_vgpr_index
[i
] = 0;
6277 if (separate_prolog
) {
6278 shader
->config
.spi_ps_input_ena
|=
6279 S_0286CC_PERSP_SAMPLE_ENA(1);
6282 case TGSI_INTERPOLATE_LOC_CENTER
:
6283 key
->ps_prolog
.color_interp_vgpr_index
[i
] = 2;
6284 if (separate_prolog
) {
6285 shader
->config
.spi_ps_input_ena
|=
6286 S_0286CC_PERSP_CENTER_ENA(1);
6289 case TGSI_INTERPOLATE_LOC_CENTROID
:
6290 key
->ps_prolog
.color_interp_vgpr_index
[i
] = 4;
6291 if (separate_prolog
) {
6292 shader
->config
.spi_ps_input_ena
|=
6293 S_0286CC_PERSP_CENTROID_ENA(1);
6300 case TGSI_INTERPOLATE_LINEAR
:
6301 /* Force the interpolation location for colors here. */
6302 if (shader
->key
.part
.ps
.prolog
.force_linear_sample_interp
)
6303 location
= TGSI_INTERPOLATE_LOC_SAMPLE
;
6304 if (shader
->key
.part
.ps
.prolog
.force_linear_center_interp
)
6305 location
= TGSI_INTERPOLATE_LOC_CENTER
;
6307 /* The VGPR assignment for non-monolithic shaders
6308 * works because InitialPSInputAddr is set on the
6309 * main shader and PERSP_PULL_MODEL is never used.
6312 case TGSI_INTERPOLATE_LOC_SAMPLE
:
6313 key
->ps_prolog
.color_interp_vgpr_index
[i
] =
6314 separate_prolog
? 6 : 9;
6315 if (separate_prolog
) {
6316 shader
->config
.spi_ps_input_ena
|=
6317 S_0286CC_LINEAR_SAMPLE_ENA(1);
6320 case TGSI_INTERPOLATE_LOC_CENTER
:
6321 key
->ps_prolog
.color_interp_vgpr_index
[i
] =
6322 separate_prolog
? 8 : 11;
6323 if (separate_prolog
) {
6324 shader
->config
.spi_ps_input_ena
|=
6325 S_0286CC_LINEAR_CENTER_ENA(1);
6328 case TGSI_INTERPOLATE_LOC_CENTROID
:
6329 key
->ps_prolog
.color_interp_vgpr_index
[i
] =
6330 separate_prolog
? 10 : 13;
6331 if (separate_prolog
) {
6332 shader
->config
.spi_ps_input_ena
|=
6333 S_0286CC_LINEAR_CENTROID_ENA(1);
6348 * Check whether a PS prolog is required based on the key.
6350 static bool si_need_ps_prolog(const union si_shader_part_key
*key
)
6352 return key
->ps_prolog
.colors_read
||
6353 key
->ps_prolog
.states
.force_persp_sample_interp
||
6354 key
->ps_prolog
.states
.force_linear_sample_interp
||
6355 key
->ps_prolog
.states
.force_persp_center_interp
||
6356 key
->ps_prolog
.states
.force_linear_center_interp
||
6357 key
->ps_prolog
.states
.bc_optimize_for_persp
||
6358 key
->ps_prolog
.states
.bc_optimize_for_linear
||
6359 key
->ps_prolog
.states
.poly_stipple
||
6360 key
->ps_prolog
.states
.samplemask_log_ps_iter
;
6364 * Compute the PS epilog key, which contains all the information needed to
6365 * build the PS epilog function.
6367 static void si_get_ps_epilog_key(struct si_shader
*shader
,
6368 union si_shader_part_key
*key
)
6370 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
6371 memset(key
, 0, sizeof(*key
));
6372 key
->ps_epilog
.colors_written
= info
->colors_written
;
6373 key
->ps_epilog
.writes_z
= info
->writes_z
;
6374 key
->ps_epilog
.writes_stencil
= info
->writes_stencil
;
6375 key
->ps_epilog
.writes_samplemask
= info
->writes_samplemask
;
6376 key
->ps_epilog
.states
= shader
->key
.part
.ps
.epilog
;
6380 * Build the GS prolog function. Rotate the input vertices for triangle strips
6383 static void si_build_gs_prolog_function(struct si_shader_context
*ctx
,
6384 union si_shader_part_key
*key
)
6386 unsigned num_sgprs
, num_vgprs
;
6387 struct si_function_info fninfo
;
6388 LLVMBuilderRef builder
= ctx
->ac
.builder
;
6389 LLVMTypeRef returns
[48];
6390 LLVMValueRef func
, ret
;
6392 si_init_function_info(&fninfo
);
6394 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
6395 if (key
->gs_prolog
.states
.gfx9_prev_is_vs
)
6396 num_sgprs
= 8 + GFX9_VSGS_NUM_USER_SGPR
;
6398 num_sgprs
= 8 + GFX9_TESGS_NUM_USER_SGPR
;
6399 num_vgprs
= 5; /* ES inputs are not needed by GS */
6401 num_sgprs
= GFX6_GS_NUM_USER_SGPR
+ 2;
6405 for (unsigned i
= 0; i
< num_sgprs
; ++i
) {
6406 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
6407 returns
[i
] = ctx
->i32
;
6410 for (unsigned i
= 0; i
< num_vgprs
; ++i
) {
6411 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
6412 returns
[num_sgprs
+ i
] = ctx
->f32
;
6415 /* Create the function. */
6416 si_create_function(ctx
, "gs_prolog", returns
, num_sgprs
+ num_vgprs
,
6418 func
= ctx
->main_fn
;
6420 /* Set the full EXEC mask for the prolog, because we are only fiddling
6421 * with registers here. The main shader part will set the correct EXEC
6424 if (ctx
->screen
->info
.chip_class
>= GFX9
&& !key
->gs_prolog
.is_monolithic
)
6425 ac_init_exec_full_mask(&ctx
->ac
);
6427 /* Copy inputs to outputs. This should be no-op, as the registers match,
6428 * but it will prevent the compiler from overwriting them unintentionally.
6430 ret
= ctx
->return_value
;
6431 for (unsigned i
= 0; i
< num_sgprs
; i
++) {
6432 LLVMValueRef p
= LLVMGetParam(func
, i
);
6433 ret
= LLVMBuildInsertValue(builder
, ret
, p
, i
, "");
6435 for (unsigned i
= 0; i
< num_vgprs
; i
++) {
6436 LLVMValueRef p
= LLVMGetParam(func
, num_sgprs
+ i
);
6437 p
= ac_to_float(&ctx
->ac
, p
);
6438 ret
= LLVMBuildInsertValue(builder
, ret
, p
, num_sgprs
+ i
, "");
6441 if (key
->gs_prolog
.states
.tri_strip_adj_fix
) {
6442 /* Remap the input vertices for every other primitive. */
6443 const unsigned gfx6_vtx_params
[6] = {
6451 const unsigned gfx9_vtx_params
[3] = {
6456 LLVMValueRef vtx_in
[6], vtx_out
[6];
6457 LLVMValueRef prim_id
, rotate
;
6459 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
6460 for (unsigned i
= 0; i
< 3; i
++) {
6461 vtx_in
[i
*2] = si_unpack_param(ctx
, gfx9_vtx_params
[i
], 0, 16);
6462 vtx_in
[i
*2+1] = si_unpack_param(ctx
, gfx9_vtx_params
[i
], 16, 16);
6465 for (unsigned i
= 0; i
< 6; i
++)
6466 vtx_in
[i
] = LLVMGetParam(func
, gfx6_vtx_params
[i
]);
6469 prim_id
= LLVMGetParam(func
, num_sgprs
+ 2);
6470 rotate
= LLVMBuildTrunc(builder
, prim_id
, ctx
->i1
, "");
6472 for (unsigned i
= 0; i
< 6; ++i
) {
6473 LLVMValueRef base
, rotated
;
6475 rotated
= vtx_in
[(i
+ 4) % 6];
6476 vtx_out
[i
] = LLVMBuildSelect(builder
, rotate
, rotated
, base
, "");
6479 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
6480 for (unsigned i
= 0; i
< 3; i
++) {
6481 LLVMValueRef hi
, out
;
6483 hi
= LLVMBuildShl(builder
, vtx_out
[i
*2+1],
6484 LLVMConstInt(ctx
->i32
, 16, 0), "");
6485 out
= LLVMBuildOr(builder
, vtx_out
[i
*2], hi
, "");
6486 out
= ac_to_float(&ctx
->ac
, out
);
6487 ret
= LLVMBuildInsertValue(builder
, ret
, out
,
6488 gfx9_vtx_params
[i
], "");
6491 for (unsigned i
= 0; i
< 6; i
++) {
6494 out
= ac_to_float(&ctx
->ac
, vtx_out
[i
]);
6495 ret
= LLVMBuildInsertValue(builder
, ret
, out
,
6496 gfx6_vtx_params
[i
], "");
6501 LLVMBuildRet(builder
, ret
);
6505 * Given a list of shader part functions, build a wrapper function that
6506 * runs them in sequence to form a monolithic shader.
6508 static void si_build_wrapper_function(struct si_shader_context
*ctx
,
6509 LLVMValueRef
*parts
,
6512 unsigned next_shader_first_part
)
6514 LLVMBuilderRef builder
= ctx
->ac
.builder
;
6515 /* PS epilog has one arg per color component; gfx9 merged shader
6516 * prologs need to forward 32 user SGPRs.
6518 struct si_function_info fninfo
;
6519 LLVMValueRef initial
[64], out
[64];
6520 LLVMTypeRef function_type
;
6521 unsigned num_first_params
;
6522 unsigned num_out
, initial_num_out
;
6523 MAYBE_UNUSED
unsigned num_out_sgpr
; /* used in debug checks */
6524 MAYBE_UNUSED
unsigned initial_num_out_sgpr
; /* used in debug checks */
6525 unsigned num_sgprs
, num_vgprs
;
6527 struct lp_build_if_state if_state
;
6529 si_init_function_info(&fninfo
);
6531 for (unsigned i
= 0; i
< num_parts
; ++i
) {
6532 ac_add_function_attr(ctx
->ac
.context
, parts
[i
], -1,
6533 AC_FUNC_ATTR_ALWAYSINLINE
);
6534 LLVMSetLinkage(parts
[i
], LLVMPrivateLinkage
);
6537 /* The parameters of the wrapper function correspond to those of the
6538 * first part in terms of SGPRs and VGPRs, but we use the types of the
6539 * main part to get the right types. This is relevant for the
6540 * dereferenceable attribute on descriptor table pointers.
6545 function_type
= LLVMGetElementType(LLVMTypeOf(parts
[0]));
6546 num_first_params
= LLVMCountParamTypes(function_type
);
6548 for (unsigned i
= 0; i
< num_first_params
; ++i
) {
6549 LLVMValueRef param
= LLVMGetParam(parts
[0], i
);
6551 if (ac_is_sgpr_param(param
)) {
6552 assert(num_vgprs
== 0);
6553 num_sgprs
+= ac_get_type_size(LLVMTypeOf(param
)) / 4;
6555 num_vgprs
+= ac_get_type_size(LLVMTypeOf(param
)) / 4;
6560 while (gprs
< num_sgprs
+ num_vgprs
) {
6561 LLVMValueRef param
= LLVMGetParam(parts
[main_part
], fninfo
.num_params
);
6562 LLVMTypeRef type
= LLVMTypeOf(param
);
6563 unsigned size
= ac_get_type_size(type
) / 4;
6565 add_arg(&fninfo
, gprs
< num_sgprs
? ARG_SGPR
: ARG_VGPR
, type
);
6567 assert(ac_is_sgpr_param(param
) == (gprs
< num_sgprs
));
6568 assert(gprs
+ size
<= num_sgprs
+ num_vgprs
&&
6569 (gprs
>= num_sgprs
|| gprs
+ size
<= num_sgprs
));
6574 /* Prepare the return type. */
6575 unsigned num_returns
= 0;
6576 LLVMTypeRef returns
[32], last_func_type
, return_type
;
6578 last_func_type
= LLVMGetElementType(LLVMTypeOf(parts
[num_parts
- 1]));
6579 return_type
= LLVMGetReturnType(last_func_type
);
6581 switch (LLVMGetTypeKind(return_type
)) {
6582 case LLVMStructTypeKind
:
6583 num_returns
= LLVMCountStructElementTypes(return_type
);
6584 assert(num_returns
<= ARRAY_SIZE(returns
));
6585 LLVMGetStructElementTypes(return_type
, returns
);
6587 case LLVMVoidTypeKind
:
6590 unreachable("unexpected type");
6593 si_create_function(ctx
, "wrapper", returns
, num_returns
, &fninfo
,
6594 si_get_max_workgroup_size(ctx
->shader
));
6596 if (is_merged_shader(ctx
))
6597 ac_init_exec_full_mask(&ctx
->ac
);
6599 /* Record the arguments of the function as if they were an output of
6605 for (unsigned i
= 0; i
< fninfo
.num_params
; ++i
) {
6606 LLVMValueRef param
= LLVMGetParam(ctx
->main_fn
, i
);
6607 LLVMTypeRef param_type
= LLVMTypeOf(param
);
6608 LLVMTypeRef out_type
= i
< fninfo
.num_sgpr_params
? ctx
->i32
: ctx
->f32
;
6609 unsigned size
= ac_get_type_size(param_type
) / 4;
6612 if (LLVMGetTypeKind(param_type
) == LLVMPointerTypeKind
) {
6613 param
= LLVMBuildPtrToInt(builder
, param
, ctx
->i32
, "");
6614 param_type
= ctx
->i32
;
6617 if (param_type
!= out_type
)
6618 param
= LLVMBuildBitCast(builder
, param
, out_type
, "");
6619 out
[num_out
++] = param
;
6621 LLVMTypeRef vector_type
= LLVMVectorType(out_type
, size
);
6623 if (LLVMGetTypeKind(param_type
) == LLVMPointerTypeKind
) {
6624 param
= LLVMBuildPtrToInt(builder
, param
, ctx
->i64
, "");
6625 param_type
= ctx
->i64
;
6628 if (param_type
!= vector_type
)
6629 param
= LLVMBuildBitCast(builder
, param
, vector_type
, "");
6631 for (unsigned j
= 0; j
< size
; ++j
)
6632 out
[num_out
++] = LLVMBuildExtractElement(
6633 builder
, param
, LLVMConstInt(ctx
->i32
, j
, 0), "");
6636 if (i
< fninfo
.num_sgpr_params
)
6637 num_out_sgpr
= num_out
;
6640 memcpy(initial
, out
, sizeof(out
));
6641 initial_num_out
= num_out
;
6642 initial_num_out_sgpr
= num_out_sgpr
;
6644 /* Now chain the parts. */
6646 for (unsigned part
= 0; part
< num_parts
; ++part
) {
6647 LLVMValueRef in
[48];
6648 LLVMTypeRef ret_type
;
6649 unsigned out_idx
= 0;
6650 unsigned num_params
= LLVMCountParams(parts
[part
]);
6652 /* Merged shaders are executed conditionally depending
6653 * on the number of enabled threads passed in the input SGPRs. */
6654 if (is_multi_part_shader(ctx
) && part
== 0) {
6655 LLVMValueRef ena
, count
= initial
[3];
6657 count
= LLVMBuildAnd(builder
, count
,
6658 LLVMConstInt(ctx
->i32
, 0x7f, 0), "");
6659 ena
= LLVMBuildICmp(builder
, LLVMIntULT
,
6660 ac_get_thread_id(&ctx
->ac
), count
, "");
6661 lp_build_if(&if_state
, &ctx
->gallivm
, ena
);
6664 /* Derive arguments for the next part from outputs of the
6667 for (unsigned param_idx
= 0; param_idx
< num_params
; ++param_idx
) {
6669 LLVMTypeRef param_type
;
6671 unsigned param_size
;
6672 LLVMValueRef arg
= NULL
;
6674 param
= LLVMGetParam(parts
[part
], param_idx
);
6675 param_type
= LLVMTypeOf(param
);
6676 param_size
= ac_get_type_size(param_type
) / 4;
6677 is_sgpr
= ac_is_sgpr_param(param
);
6680 ac_add_function_attr(ctx
->ac
.context
, parts
[part
],
6681 param_idx
+ 1, AC_FUNC_ATTR_INREG
);
6682 } else if (out_idx
< num_out_sgpr
) {
6683 /* Skip returned SGPRs the current part doesn't
6684 * declare on the input. */
6685 out_idx
= num_out_sgpr
;
6688 assert(out_idx
+ param_size
<= (is_sgpr
? num_out_sgpr
: num_out
));
6690 if (param_size
== 1)
6693 arg
= ac_build_gather_values(&ctx
->ac
, &out
[out_idx
], param_size
);
6695 if (LLVMTypeOf(arg
) != param_type
) {
6696 if (LLVMGetTypeKind(param_type
) == LLVMPointerTypeKind
) {
6697 if (LLVMGetPointerAddressSpace(param_type
) ==
6698 AC_ADDR_SPACE_CONST_32BIT
) {
6699 arg
= LLVMBuildBitCast(builder
, arg
, ctx
->i32
, "");
6700 arg
= LLVMBuildIntToPtr(builder
, arg
, param_type
, "");
6702 arg
= LLVMBuildBitCast(builder
, arg
, ctx
->i64
, "");
6703 arg
= LLVMBuildIntToPtr(builder
, arg
, param_type
, "");
6706 arg
= LLVMBuildBitCast(builder
, arg
, param_type
, "");
6710 in
[param_idx
] = arg
;
6711 out_idx
+= param_size
;
6714 ret
= ac_build_call(&ctx
->ac
, parts
[part
], in
, num_params
);
6716 if (is_multi_part_shader(ctx
) &&
6717 part
+ 1 == next_shader_first_part
) {
6718 lp_build_endif(&if_state
);
6720 /* The second half of the merged shader should use
6721 * the inputs from the toplevel (wrapper) function,
6722 * not the return value from the last call.
6724 * That's because the last call was executed condi-
6725 * tionally, so we can't consume it in the main
6728 memcpy(out
, initial
, sizeof(initial
));
6729 num_out
= initial_num_out
;
6730 num_out_sgpr
= initial_num_out_sgpr
;
6734 /* Extract the returned GPRs. */
6735 ret_type
= LLVMTypeOf(ret
);
6739 if (LLVMGetTypeKind(ret_type
) != LLVMVoidTypeKind
) {
6740 assert(LLVMGetTypeKind(ret_type
) == LLVMStructTypeKind
);
6742 unsigned ret_size
= LLVMCountStructElementTypes(ret_type
);
6744 for (unsigned i
= 0; i
< ret_size
; ++i
) {
6746 LLVMBuildExtractValue(builder
, ret
, i
, "");
6748 assert(num_out
< ARRAY_SIZE(out
));
6749 out
[num_out
++] = val
;
6751 if (LLVMTypeOf(val
) == ctx
->i32
) {
6752 assert(num_out_sgpr
+ 1 == num_out
);
6753 num_out_sgpr
= num_out
;
6759 /* Return the value from the last part. */
6760 if (LLVMGetTypeKind(LLVMTypeOf(ret
)) == LLVMVoidTypeKind
)
6761 LLVMBuildRetVoid(builder
);
6763 LLVMBuildRet(builder
, ret
);
6766 static bool si_should_optimize_less(struct ac_llvm_compiler
*compiler
,
6767 struct si_shader_selector
*sel
)
6769 if (!compiler
->low_opt_passes
)
6772 /* Assume a slow CPU. */
6773 assert(!sel
->screen
->info
.has_dedicated_vram
&&
6774 sel
->screen
->info
.chip_class
<= GFX8
);
6776 /* For a crazy dEQP test containing 2597 memory opcodes, mostly
6778 return sel
->type
== PIPE_SHADER_COMPUTE
&&
6779 sel
->info
.num_memory_instructions
> 1000;
6782 int si_compile_tgsi_shader(struct si_screen
*sscreen
,
6783 struct ac_llvm_compiler
*compiler
,
6784 struct si_shader
*shader
,
6785 struct pipe_debug_callback
*debug
)
6787 struct si_shader_selector
*sel
= shader
->selector
;
6788 struct si_shader_context ctx
;
6791 /* Dump TGSI code before doing TGSI->LLVM conversion in case the
6792 * conversion fails. */
6793 if (si_can_dump_shader(sscreen
, sel
->info
.processor
) &&
6794 !(sscreen
->debug_flags
& DBG(NO_TGSI
))) {
6796 tgsi_dump(sel
->tokens
, 0);
6798 nir_print_shader(sel
->nir
, stderr
);
6799 si_dump_streamout(&sel
->so
);
6802 si_init_shader_ctx(&ctx
, sscreen
, compiler
);
6803 si_llvm_context_set_tgsi(&ctx
, shader
);
6805 memset(shader
->info
.vs_output_param_offset
, AC_EXP_PARAM_UNDEFINED
,
6806 sizeof(shader
->info
.vs_output_param_offset
));
6808 shader
->info
.uses_instanceid
= sel
->info
.uses_instanceid
;
6810 if (!si_compile_tgsi_main(&ctx
)) {
6811 si_llvm_dispose(&ctx
);
6815 if (shader
->is_monolithic
&& ctx
.type
== PIPE_SHADER_VERTEX
) {
6816 LLVMValueRef parts
[2];
6817 bool need_prolog
= sel
->vs_needs_prolog
;
6819 parts
[1] = ctx
.main_fn
;
6822 union si_shader_part_key prolog_key
;
6823 si_get_vs_prolog_key(&sel
->info
,
6824 shader
->info
.num_input_sgprs
,
6825 &shader
->key
.part
.vs
.prolog
,
6826 shader
, &prolog_key
);
6827 si_build_vs_prolog_function(&ctx
, &prolog_key
);
6828 parts
[0] = ctx
.main_fn
;
6831 si_build_wrapper_function(&ctx
, parts
+ !need_prolog
,
6832 1 + need_prolog
, need_prolog
, 0);
6834 if (ctx
.shader
->key
.opt
.vs_as_prim_discard_cs
)
6835 si_build_prim_discard_compute_shader(&ctx
);
6836 } else if (shader
->is_monolithic
&& ctx
.type
== PIPE_SHADER_TESS_CTRL
) {
6837 if (sscreen
->info
.chip_class
>= GFX9
) {
6838 struct si_shader_selector
*ls
= shader
->key
.part
.tcs
.ls
;
6839 LLVMValueRef parts
[4];
6840 bool vs_needs_prolog
=
6841 si_vs_needs_prolog(ls
, &shader
->key
.part
.tcs
.ls_prolog
);
6844 parts
[2] = ctx
.main_fn
;
6847 union si_shader_part_key tcs_epilog_key
;
6848 memset(&tcs_epilog_key
, 0, sizeof(tcs_epilog_key
));
6849 tcs_epilog_key
.tcs_epilog
.states
= shader
->key
.part
.tcs
.epilog
;
6850 si_build_tcs_epilog_function(&ctx
, &tcs_epilog_key
);
6851 parts
[3] = ctx
.main_fn
;
6853 /* VS as LS main part */
6854 struct si_shader shader_ls
= {};
6855 shader_ls
.selector
= ls
;
6856 shader_ls
.key
.as_ls
= 1;
6857 shader_ls
.key
.mono
= shader
->key
.mono
;
6858 shader_ls
.key
.opt
= shader
->key
.opt
;
6859 shader_ls
.is_monolithic
= true;
6860 si_llvm_context_set_tgsi(&ctx
, &shader_ls
);
6862 if (!si_compile_tgsi_main(&ctx
)) {
6863 si_llvm_dispose(&ctx
);
6866 shader
->info
.uses_instanceid
|= ls
->info
.uses_instanceid
;
6867 parts
[1] = ctx
.main_fn
;
6870 if (vs_needs_prolog
) {
6871 union si_shader_part_key vs_prolog_key
;
6872 si_get_vs_prolog_key(&ls
->info
,
6873 shader_ls
.info
.num_input_sgprs
,
6874 &shader
->key
.part
.tcs
.ls_prolog
,
6875 shader
, &vs_prolog_key
);
6876 vs_prolog_key
.vs_prolog
.is_monolithic
= true;
6877 si_build_vs_prolog_function(&ctx
, &vs_prolog_key
);
6878 parts
[0] = ctx
.main_fn
;
6881 /* Reset the shader context. */
6882 ctx
.shader
= shader
;
6883 ctx
.type
= PIPE_SHADER_TESS_CTRL
;
6885 si_build_wrapper_function(&ctx
,
6886 parts
+ !vs_needs_prolog
,
6887 4 - !vs_needs_prolog
, vs_needs_prolog
,
6888 vs_needs_prolog
? 2 : 1);
6890 LLVMValueRef parts
[2];
6891 union si_shader_part_key epilog_key
;
6893 parts
[0] = ctx
.main_fn
;
6895 memset(&epilog_key
, 0, sizeof(epilog_key
));
6896 epilog_key
.tcs_epilog
.states
= shader
->key
.part
.tcs
.epilog
;
6897 si_build_tcs_epilog_function(&ctx
, &epilog_key
);
6898 parts
[1] = ctx
.main_fn
;
6900 si_build_wrapper_function(&ctx
, parts
, 2, 0, 0);
6902 } else if (shader
->is_monolithic
&& ctx
.type
== PIPE_SHADER_GEOMETRY
) {
6903 if (ctx
.screen
->info
.chip_class
>= GFX9
) {
6904 struct si_shader_selector
*es
= shader
->key
.part
.gs
.es
;
6905 LLVMValueRef es_prolog
= NULL
;
6906 LLVMValueRef es_main
= NULL
;
6907 LLVMValueRef gs_prolog
= NULL
;
6908 LLVMValueRef gs_main
= ctx
.main_fn
;
6911 union si_shader_part_key gs_prolog_key
;
6912 memset(&gs_prolog_key
, 0, sizeof(gs_prolog_key
));
6913 gs_prolog_key
.gs_prolog
.states
= shader
->key
.part
.gs
.prolog
;
6914 gs_prolog_key
.gs_prolog
.is_monolithic
= true;
6915 si_build_gs_prolog_function(&ctx
, &gs_prolog_key
);
6916 gs_prolog
= ctx
.main_fn
;
6919 struct si_shader shader_es
= {};
6920 shader_es
.selector
= es
;
6921 shader_es
.key
.as_es
= 1;
6922 shader_es
.key
.mono
= shader
->key
.mono
;
6923 shader_es
.key
.opt
= shader
->key
.opt
;
6924 shader_es
.is_monolithic
= true;
6925 si_llvm_context_set_tgsi(&ctx
, &shader_es
);
6927 if (!si_compile_tgsi_main(&ctx
)) {
6928 si_llvm_dispose(&ctx
);
6931 shader
->info
.uses_instanceid
|= es
->info
.uses_instanceid
;
6932 es_main
= ctx
.main_fn
;
6935 if (es
->vs_needs_prolog
) {
6936 union si_shader_part_key vs_prolog_key
;
6937 si_get_vs_prolog_key(&es
->info
,
6938 shader_es
.info
.num_input_sgprs
,
6939 &shader
->key
.part
.gs
.vs_prolog
,
6940 shader
, &vs_prolog_key
);
6941 vs_prolog_key
.vs_prolog
.is_monolithic
= true;
6942 si_build_vs_prolog_function(&ctx
, &vs_prolog_key
);
6943 es_prolog
= ctx
.main_fn
;
6946 /* Reset the shader context. */
6947 ctx
.shader
= shader
;
6948 ctx
.type
= PIPE_SHADER_GEOMETRY
;
6950 /* Prepare the array of shader parts. */
6951 LLVMValueRef parts
[4];
6952 unsigned num_parts
= 0, main_part
, next_first_part
;
6955 parts
[num_parts
++] = es_prolog
;
6957 parts
[main_part
= num_parts
++] = es_main
;
6958 parts
[next_first_part
= num_parts
++] = gs_prolog
;
6959 parts
[num_parts
++] = gs_main
;
6961 si_build_wrapper_function(&ctx
, parts
, num_parts
,
6962 main_part
, next_first_part
);
6964 LLVMValueRef parts
[2];
6965 union si_shader_part_key prolog_key
;
6967 parts
[1] = ctx
.main_fn
;
6969 memset(&prolog_key
, 0, sizeof(prolog_key
));
6970 prolog_key
.gs_prolog
.states
= shader
->key
.part
.gs
.prolog
;
6971 si_build_gs_prolog_function(&ctx
, &prolog_key
);
6972 parts
[0] = ctx
.main_fn
;
6974 si_build_wrapper_function(&ctx
, parts
, 2, 1, 0);
6976 } else if (shader
->is_monolithic
&& ctx
.type
== PIPE_SHADER_FRAGMENT
) {
6977 LLVMValueRef parts
[3];
6978 union si_shader_part_key prolog_key
;
6979 union si_shader_part_key epilog_key
;
6982 si_get_ps_prolog_key(shader
, &prolog_key
, false);
6983 need_prolog
= si_need_ps_prolog(&prolog_key
);
6985 parts
[need_prolog
? 1 : 0] = ctx
.main_fn
;
6988 si_build_ps_prolog_function(&ctx
, &prolog_key
);
6989 parts
[0] = ctx
.main_fn
;
6992 si_get_ps_epilog_key(shader
, &epilog_key
);
6993 si_build_ps_epilog_function(&ctx
, &epilog_key
);
6994 parts
[need_prolog
? 2 : 1] = ctx
.main_fn
;
6996 si_build_wrapper_function(&ctx
, parts
, need_prolog
? 3 : 2,
6997 need_prolog
? 1 : 0, 0);
7000 si_llvm_optimize_module(&ctx
);
7002 /* Post-optimization transformations and analysis. */
7003 si_optimize_vs_outputs(&ctx
);
7005 if ((debug
&& debug
->debug_message
) ||
7006 si_can_dump_shader(sscreen
, ctx
.type
)) {
7007 ctx
.shader
->info
.private_mem_vgprs
=
7008 ac_count_scratch_private_memory(ctx
.main_fn
);
7011 /* Make sure the input is a pointer and not integer followed by inttoptr. */
7012 assert(LLVMGetTypeKind(LLVMTypeOf(LLVMGetParam(ctx
.main_fn
, 0))) ==
7013 LLVMPointerTypeKind
);
7015 /* Compile to bytecode. */
7016 r
= si_compile_llvm(sscreen
, &shader
->binary
, &shader
->config
, compiler
,
7017 ctx
.ac
.module
, debug
, ctx
.type
,
7018 si_get_shader_name(shader
, ctx
.type
),
7019 si_should_optimize_less(compiler
, shader
->selector
));
7020 si_llvm_dispose(&ctx
);
7022 fprintf(stderr
, "LLVM failed to compile shader\n");
7026 /* Validate SGPR and VGPR usage for compute to detect compiler bugs.
7027 * LLVM 3.9svn has this bug.
7029 if (sel
->type
== PIPE_SHADER_COMPUTE
) {
7030 unsigned wave_size
= 64;
7031 unsigned max_vgprs
= 256;
7032 unsigned max_sgprs
= sscreen
->info
.chip_class
>= GFX8
? 800 : 512;
7033 unsigned max_sgprs_per_wave
= 128;
7034 unsigned max_block_threads
= si_get_max_workgroup_size(shader
);
7035 unsigned min_waves_per_cu
= DIV_ROUND_UP(max_block_threads
, wave_size
);
7036 unsigned min_waves_per_simd
= DIV_ROUND_UP(min_waves_per_cu
, 4);
7038 max_vgprs
= max_vgprs
/ min_waves_per_simd
;
7039 max_sgprs
= MIN2(max_sgprs
/ min_waves_per_simd
, max_sgprs_per_wave
);
7041 if (shader
->config
.num_sgprs
> max_sgprs
||
7042 shader
->config
.num_vgprs
> max_vgprs
) {
7043 fprintf(stderr
, "LLVM failed to compile a shader correctly: "
7044 "SGPR:VGPR usage is %u:%u, but the hw limit is %u:%u\n",
7045 shader
->config
.num_sgprs
, shader
->config
.num_vgprs
,
7046 max_sgprs
, max_vgprs
);
7048 /* Just terminate the process, because dependent
7049 * shaders can hang due to bad input data, but use
7050 * the env var to allow shader-db to work.
7052 if (!debug_get_bool_option("SI_PASS_BAD_SHADERS", false))
7057 /* Add the scratch offset to input SGPRs. */
7058 if (shader
->config
.scratch_bytes_per_wave
&& !is_merged_shader(&ctx
))
7059 shader
->info
.num_input_sgprs
+= 1; /* scratch byte offset */
7061 /* Calculate the number of fragment input VGPRs. */
7062 if (ctx
.type
== PIPE_SHADER_FRAGMENT
) {
7063 shader
->info
.num_input_vgprs
= 0;
7064 shader
->info
.face_vgpr_index
= -1;
7065 shader
->info
.ancillary_vgpr_index
= -1;
7067 if (G_0286CC_PERSP_SAMPLE_ENA(shader
->config
.spi_ps_input_addr
))
7068 shader
->info
.num_input_vgprs
+= 2;
7069 if (G_0286CC_PERSP_CENTER_ENA(shader
->config
.spi_ps_input_addr
))
7070 shader
->info
.num_input_vgprs
+= 2;
7071 if (G_0286CC_PERSP_CENTROID_ENA(shader
->config
.spi_ps_input_addr
))
7072 shader
->info
.num_input_vgprs
+= 2;
7073 if (G_0286CC_PERSP_PULL_MODEL_ENA(shader
->config
.spi_ps_input_addr
))
7074 shader
->info
.num_input_vgprs
+= 3;
7075 if (G_0286CC_LINEAR_SAMPLE_ENA(shader
->config
.spi_ps_input_addr
))
7076 shader
->info
.num_input_vgprs
+= 2;
7077 if (G_0286CC_LINEAR_CENTER_ENA(shader
->config
.spi_ps_input_addr
))
7078 shader
->info
.num_input_vgprs
+= 2;
7079 if (G_0286CC_LINEAR_CENTROID_ENA(shader
->config
.spi_ps_input_addr
))
7080 shader
->info
.num_input_vgprs
+= 2;
7081 if (G_0286CC_LINE_STIPPLE_TEX_ENA(shader
->config
.spi_ps_input_addr
))
7082 shader
->info
.num_input_vgprs
+= 1;
7083 if (G_0286CC_POS_X_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
7084 shader
->info
.num_input_vgprs
+= 1;
7085 if (G_0286CC_POS_Y_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
7086 shader
->info
.num_input_vgprs
+= 1;
7087 if (G_0286CC_POS_Z_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
7088 shader
->info
.num_input_vgprs
+= 1;
7089 if (G_0286CC_POS_W_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
7090 shader
->info
.num_input_vgprs
+= 1;
7091 if (G_0286CC_FRONT_FACE_ENA(shader
->config
.spi_ps_input_addr
)) {
7092 shader
->info
.face_vgpr_index
= shader
->info
.num_input_vgprs
;
7093 shader
->info
.num_input_vgprs
+= 1;
7095 if (G_0286CC_ANCILLARY_ENA(shader
->config
.spi_ps_input_addr
)) {
7096 shader
->info
.ancillary_vgpr_index
= shader
->info
.num_input_vgprs
;
7097 shader
->info
.num_input_vgprs
+= 1;
7099 if (G_0286CC_SAMPLE_COVERAGE_ENA(shader
->config
.spi_ps_input_addr
))
7100 shader
->info
.num_input_vgprs
+= 1;
7101 if (G_0286CC_POS_FIXED_PT_ENA(shader
->config
.spi_ps_input_addr
))
7102 shader
->info
.num_input_vgprs
+= 1;
7105 si_calculate_max_simd_waves(shader
);
7106 si_shader_dump_stats_for_shader_db(sscreen
, shader
, debug
);
7111 * Create, compile and return a shader part (prolog or epilog).
7113 * \param sscreen screen
7114 * \param list list of shader parts of the same category
7115 * \param type shader type
7116 * \param key shader part key
7117 * \param prolog whether the part being requested is a prolog
7118 * \param tm LLVM target machine
7119 * \param debug debug callback
7120 * \param build the callback responsible for building the main function
7121 * \return non-NULL on success
7123 static struct si_shader_part
*
7124 si_get_shader_part(struct si_screen
*sscreen
,
7125 struct si_shader_part
**list
,
7126 enum pipe_shader_type type
,
7128 union si_shader_part_key
*key
,
7129 struct ac_llvm_compiler
*compiler
,
7130 struct pipe_debug_callback
*debug
,
7131 void (*build
)(struct si_shader_context
*,
7132 union si_shader_part_key
*),
7135 struct si_shader_part
*result
;
7137 mtx_lock(&sscreen
->shader_parts_mutex
);
7139 /* Find existing. */
7140 for (result
= *list
; result
; result
= result
->next
) {
7141 if (memcmp(&result
->key
, key
, sizeof(*key
)) == 0) {
7142 mtx_unlock(&sscreen
->shader_parts_mutex
);
7147 /* Compile a new one. */
7148 result
= CALLOC_STRUCT(si_shader_part
);
7151 struct si_shader shader
= {};
7152 struct si_shader_context ctx
;
7154 si_init_shader_ctx(&ctx
, sscreen
, compiler
);
7155 ctx
.shader
= &shader
;
7159 case PIPE_SHADER_VERTEX
:
7160 shader
.key
.as_ls
= key
->vs_prolog
.as_ls
;
7161 shader
.key
.as_es
= key
->vs_prolog
.as_es
;
7162 shader
.key
.as_ngg
= key
->vs_prolog
.as_ngg
;
7164 case PIPE_SHADER_TESS_CTRL
:
7166 shader
.key
.part
.tcs
.epilog
= key
->tcs_epilog
.states
;
7168 case PIPE_SHADER_GEOMETRY
:
7171 case PIPE_SHADER_FRAGMENT
:
7173 shader
.key
.part
.ps
.prolog
= key
->ps_prolog
.states
;
7175 shader
.key
.part
.ps
.epilog
= key
->ps_epilog
.states
;
7178 unreachable("bad shader part");
7184 si_llvm_optimize_module(&ctx
);
7186 if (si_compile_llvm(sscreen
, &result
->binary
, &result
->config
, compiler
,
7187 ctx
.ac
.module
, debug
, ctx
.type
, name
, false)) {
7193 result
->next
= *list
;
7197 si_llvm_dispose(&ctx
);
7198 mtx_unlock(&sscreen
->shader_parts_mutex
);
7202 static LLVMValueRef
si_prolog_get_rw_buffers(struct si_shader_context
*ctx
)
7204 LLVMValueRef ptr
[2], list
;
7205 bool merged_shader
= is_merged_shader(ctx
);
7207 ptr
[0] = LLVMGetParam(ctx
->main_fn
, (merged_shader
? 8 : 0) + SI_SGPR_RW_BUFFERS
);
7208 list
= LLVMBuildIntToPtr(ctx
->ac
.builder
, ptr
[0],
7209 ac_array_in_const32_addr_space(ctx
->v4i32
), "");
7214 * Build the vertex shader prolog function.
7216 * The inputs are the same as VS (a lot of SGPRs and 4 VGPR system values).
7217 * All inputs are returned unmodified. The vertex load indices are
7218 * stored after them, which will be used by the API VS for fetching inputs.
7220 * For example, the expected outputs for instance_divisors[] = {0, 1, 2} are:
7225 * (VertexID + BaseVertex),
7226 * (InstanceID + StartInstance),
7227 * (InstanceID / 2 + StartInstance)
7229 static void si_build_vs_prolog_function(struct si_shader_context
*ctx
,
7230 union si_shader_part_key
*key
)
7232 struct si_function_info fninfo
;
7233 LLVMTypeRef
*returns
;
7234 LLVMValueRef ret
, func
;
7236 unsigned first_vs_vgpr
= key
->vs_prolog
.num_merged_next_stage_vgprs
;
7237 unsigned num_input_vgprs
= key
->vs_prolog
.num_merged_next_stage_vgprs
+ 4;
7238 LLVMValueRef input_vgprs
[9];
7239 unsigned num_all_input_regs
= key
->vs_prolog
.num_input_sgprs
+
7241 unsigned user_sgpr_base
= key
->vs_prolog
.num_merged_next_stage_vgprs
? 8 : 0;
7243 si_init_function_info(&fninfo
);
7245 /* 4 preloaded VGPRs + vertex load indices as prolog outputs */
7246 returns
= alloca((num_all_input_regs
+ key
->vs_prolog
.last_input
+ 1) *
7247 sizeof(LLVMTypeRef
));
7250 /* Declare input and output SGPRs. */
7251 for (i
= 0; i
< key
->vs_prolog
.num_input_sgprs
; i
++) {
7252 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7253 returns
[num_returns
++] = ctx
->i32
;
7256 /* Preloaded VGPRs (outputs must be floats) */
7257 for (i
= 0; i
< num_input_vgprs
; i
++) {
7258 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &input_vgprs
[i
]);
7259 returns
[num_returns
++] = ctx
->f32
;
7262 /* Vertex load indices. */
7263 for (i
= 0; i
<= key
->vs_prolog
.last_input
; i
++)
7264 returns
[num_returns
++] = ctx
->f32
;
7266 /* Create the function. */
7267 si_create_function(ctx
, "vs_prolog", returns
, num_returns
, &fninfo
, 0);
7268 func
= ctx
->main_fn
;
7270 if (key
->vs_prolog
.num_merged_next_stage_vgprs
) {
7271 if (!key
->vs_prolog
.is_monolithic
)
7272 si_init_exec_from_input(ctx
, 3, 0);
7274 if (key
->vs_prolog
.as_ls
&&
7275 ctx
->screen
->has_ls_vgpr_init_bug
) {
7276 /* If there are no HS threads, SPI loads the LS VGPRs
7277 * starting at VGPR 0. Shift them back to where they
7280 LLVMValueRef has_hs_threads
=
7281 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntNE
,
7282 si_unpack_param(ctx
, 3, 8, 8),
7285 for (i
= 4; i
> 0; --i
) {
7286 input_vgprs
[i
+ 1] =
7287 LLVMBuildSelect(ctx
->ac
.builder
, has_hs_threads
,
7289 input_vgprs
[i
- 1], "");
7294 unsigned vertex_id_vgpr
= first_vs_vgpr
;
7295 unsigned instance_id_vgpr
= first_vs_vgpr
+ (key
->vs_prolog
.as_ls
? 2 : 1);
7297 ctx
->abi
.vertex_id
= input_vgprs
[vertex_id_vgpr
];
7298 ctx
->abi
.instance_id
= input_vgprs
[instance_id_vgpr
];
7300 /* InstanceID = VertexID >> 16;
7301 * VertexID = VertexID & 0xffff;
7303 if (key
->vs_prolog
.states
.unpack_instance_id_from_vertex_id
) {
7304 ctx
->abi
.instance_id
= LLVMBuildLShr(ctx
->ac
.builder
, ctx
->abi
.vertex_id
,
7305 LLVMConstInt(ctx
->i32
, 16, 0), "");
7306 ctx
->abi
.vertex_id
= LLVMBuildAnd(ctx
->ac
.builder
, ctx
->abi
.vertex_id
,
7307 LLVMConstInt(ctx
->i32
, 0xffff, 0), "");
7310 /* Copy inputs to outputs. This should be no-op, as the registers match,
7311 * but it will prevent the compiler from overwriting them unintentionally.
7313 ret
= ctx
->return_value
;
7314 for (i
= 0; i
< key
->vs_prolog
.num_input_sgprs
; i
++) {
7315 LLVMValueRef p
= LLVMGetParam(func
, i
);
7316 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, p
, i
, "");
7318 for (i
= 0; i
< num_input_vgprs
; i
++) {
7319 LLVMValueRef p
= input_vgprs
[i
];
7321 if (i
== vertex_id_vgpr
)
7322 p
= ctx
->abi
.vertex_id
;
7323 else if (i
== instance_id_vgpr
)
7324 p
= ctx
->abi
.instance_id
;
7326 p
= ac_to_float(&ctx
->ac
, p
);
7327 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, p
,
7328 key
->vs_prolog
.num_input_sgprs
+ i
, "");
7331 struct lp_build_if_state wrap_if_state
;
7332 LLVMValueRef original_ret
= ret
;
7333 bool wrapped
= false;
7335 if (key
->vs_prolog
.is_monolithic
&& key
->vs_prolog
.as_ngg
) {
7336 LLVMValueRef num_threads
;
7339 num_threads
= si_unpack_param(ctx
, 3, 0, 8);
7340 ena
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntULT
,
7341 ac_get_thread_id(&ctx
->ac
), num_threads
, "");
7342 lp_build_if(&wrap_if_state
, &ctx
->gallivm
, ena
);
7346 /* Compute vertex load indices from instance divisors. */
7347 LLVMValueRef instance_divisor_constbuf
= NULL
;
7349 if (key
->vs_prolog
.states
.instance_divisor_is_fetched
) {
7350 LLVMValueRef list
= si_prolog_get_rw_buffers(ctx
);
7351 LLVMValueRef buf_index
=
7352 LLVMConstInt(ctx
->i32
, SI_VS_CONST_INSTANCE_DIVISORS
, 0);
7353 instance_divisor_constbuf
=
7354 ac_build_load_to_sgpr(&ctx
->ac
, list
, buf_index
);
7357 for (i
= 0; i
<= key
->vs_prolog
.last_input
; i
++) {
7358 bool divisor_is_one
=
7359 key
->vs_prolog
.states
.instance_divisor_is_one
& (1u << i
);
7360 bool divisor_is_fetched
=
7361 key
->vs_prolog
.states
.instance_divisor_is_fetched
& (1u << i
);
7362 LLVMValueRef index
= NULL
;
7364 if (divisor_is_one
) {
7365 index
= ctx
->abi
.instance_id
;
7366 } else if (divisor_is_fetched
) {
7367 LLVMValueRef udiv_factors
[4];
7369 for (unsigned j
= 0; j
< 4; j
++) {
7371 buffer_load_const(ctx
, instance_divisor_constbuf
,
7372 LLVMConstInt(ctx
->i32
, i
*16 + j
*4, 0));
7373 udiv_factors
[j
] = ac_to_integer(&ctx
->ac
, udiv_factors
[j
]);
7375 /* The faster NUW version doesn't work when InstanceID == UINT_MAX.
7376 * Such InstanceID might not be achievable in a reasonable time though.
7378 index
= ac_build_fast_udiv_nuw(&ctx
->ac
, ctx
->abi
.instance_id
,
7379 udiv_factors
[0], udiv_factors
[1],
7380 udiv_factors
[2], udiv_factors
[3]);
7383 if (divisor_is_one
|| divisor_is_fetched
) {
7384 /* Add StartInstance. */
7385 index
= LLVMBuildAdd(ctx
->ac
.builder
, index
,
7386 LLVMGetParam(ctx
->main_fn
, user_sgpr_base
+
7387 SI_SGPR_START_INSTANCE
), "");
7389 /* VertexID + BaseVertex */
7390 index
= LLVMBuildAdd(ctx
->ac
.builder
,
7392 LLVMGetParam(func
, user_sgpr_base
+
7393 SI_SGPR_BASE_VERTEX
), "");
7396 index
= ac_to_float(&ctx
->ac
, index
);
7397 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, index
,
7398 fninfo
.num_params
+ i
, "");
7402 lp_build_endif(&wrap_if_state
);
7404 LLVMValueRef values
[2] = {
7408 LLVMBasicBlockRef bbs
[2] = {
7409 wrap_if_state
.true_block
,
7410 wrap_if_state
.entry_block
7412 ret
= ac_build_phi(&ctx
->ac
, LLVMTypeOf(ret
), 2, values
, bbs
);
7415 si_llvm_build_ret(ctx
, ret
);
7418 static bool si_get_vs_prolog(struct si_screen
*sscreen
,
7419 struct ac_llvm_compiler
*compiler
,
7420 struct si_shader
*shader
,
7421 struct pipe_debug_callback
*debug
,
7422 struct si_shader
*main_part
,
7423 const struct si_vs_prolog_bits
*key
)
7425 struct si_shader_selector
*vs
= main_part
->selector
;
7427 if (!si_vs_needs_prolog(vs
, key
))
7430 /* Get the prolog. */
7431 union si_shader_part_key prolog_key
;
7432 si_get_vs_prolog_key(&vs
->info
, main_part
->info
.num_input_sgprs
,
7433 key
, shader
, &prolog_key
);
7436 si_get_shader_part(sscreen
, &sscreen
->vs_prologs
,
7437 PIPE_SHADER_VERTEX
, true, &prolog_key
, compiler
,
7438 debug
, si_build_vs_prolog_function
,
7439 "Vertex Shader Prolog");
7440 return shader
->prolog
!= NULL
;
7444 * Select and compile (or reuse) vertex shader parts (prolog & epilog).
7446 static bool si_shader_select_vs_parts(struct si_screen
*sscreen
,
7447 struct ac_llvm_compiler
*compiler
,
7448 struct si_shader
*shader
,
7449 struct pipe_debug_callback
*debug
)
7451 return si_get_vs_prolog(sscreen
, compiler
, shader
, debug
, shader
,
7452 &shader
->key
.part
.vs
.prolog
);
7456 * Compile the TCS epilog function. This writes tesselation factors to memory
7457 * based on the output primitive type of the tesselator (determined by TES).
7459 static void si_build_tcs_epilog_function(struct si_shader_context
*ctx
,
7460 union si_shader_part_key
*key
)
7462 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
7463 struct si_function_info fninfo
;
7466 si_init_function_info(&fninfo
);
7468 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
7469 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7470 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7471 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7472 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* wave info */
7473 ctx
->param_tcs_factor_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7474 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7475 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7476 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7477 add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7478 add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7479 add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7480 add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7481 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7482 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7483 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7484 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7485 ctx
->param_tcs_offchip_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7486 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7487 ctx
->param_tcs_out_lds_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7489 add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7490 add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7491 add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7492 add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7493 ctx
->param_tcs_offchip_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7494 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7495 ctx
->param_tcs_out_lds_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7496 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7497 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7498 ctx
->param_tcs_factor_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7501 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
); /* VGPR gap */
7502 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
); /* VGPR gap */
7503 unsigned tess_factors_idx
=
7504 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
); /* patch index within the wave (REL_PATCH_ID) */
7505 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
); /* invocation ID within the patch */
7506 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
); /* LDS offset where tess factors should be loaded from */
7508 for (unsigned i
= 0; i
< 6; i
++)
7509 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
); /* tess factors */
7511 /* Create the function. */
7512 si_create_function(ctx
, "tcs_epilog", NULL
, 0, &fninfo
,
7513 ctx
->screen
->info
.chip_class
>= GFX7
? 128 : 64);
7514 ac_declare_lds_as_pointer(&ctx
->ac
);
7515 func
= ctx
->main_fn
;
7517 LLVMValueRef invoc0_tess_factors
[6];
7518 for (unsigned i
= 0; i
< 6; i
++)
7519 invoc0_tess_factors
[i
] = LLVMGetParam(func
, tess_factors_idx
+ 3 + i
);
7521 si_write_tess_factors(bld_base
,
7522 LLVMGetParam(func
, tess_factors_idx
),
7523 LLVMGetParam(func
, tess_factors_idx
+ 1),
7524 LLVMGetParam(func
, tess_factors_idx
+ 2),
7525 invoc0_tess_factors
, invoc0_tess_factors
+ 4);
7527 LLVMBuildRetVoid(ctx
->ac
.builder
);
7531 * Select and compile (or reuse) TCS parts (epilog).
7533 static bool si_shader_select_tcs_parts(struct si_screen
*sscreen
,
7534 struct ac_llvm_compiler
*compiler
,
7535 struct si_shader
*shader
,
7536 struct pipe_debug_callback
*debug
)
7538 if (sscreen
->info
.chip_class
>= GFX9
) {
7539 struct si_shader
*ls_main_part
=
7540 shader
->key
.part
.tcs
.ls
->main_shader_part_ls
;
7542 if (!si_get_vs_prolog(sscreen
, compiler
, shader
, debug
, ls_main_part
,
7543 &shader
->key
.part
.tcs
.ls_prolog
))
7546 shader
->previous_stage
= ls_main_part
;
7549 /* Get the epilog. */
7550 union si_shader_part_key epilog_key
;
7551 memset(&epilog_key
, 0, sizeof(epilog_key
));
7552 epilog_key
.tcs_epilog
.states
= shader
->key
.part
.tcs
.epilog
;
7554 shader
->epilog
= si_get_shader_part(sscreen
, &sscreen
->tcs_epilogs
,
7555 PIPE_SHADER_TESS_CTRL
, false,
7556 &epilog_key
, compiler
, debug
,
7557 si_build_tcs_epilog_function
,
7558 "Tessellation Control Shader Epilog");
7559 return shader
->epilog
!= NULL
;
7563 * Select and compile (or reuse) GS parts (prolog).
7565 static bool si_shader_select_gs_parts(struct si_screen
*sscreen
,
7566 struct ac_llvm_compiler
*compiler
,
7567 struct si_shader
*shader
,
7568 struct pipe_debug_callback
*debug
)
7570 if (sscreen
->info
.chip_class
>= GFX9
) {
7571 struct si_shader
*es_main_part
=
7572 shader
->key
.part
.gs
.es
->main_shader_part_es
;
7574 if (shader
->key
.part
.gs
.es
->type
== PIPE_SHADER_VERTEX
&&
7575 !si_get_vs_prolog(sscreen
, compiler
, shader
, debug
, es_main_part
,
7576 &shader
->key
.part
.gs
.vs_prolog
))
7579 shader
->previous_stage
= es_main_part
;
7582 if (!shader
->key
.part
.gs
.prolog
.tri_strip_adj_fix
)
7585 union si_shader_part_key prolog_key
;
7586 memset(&prolog_key
, 0, sizeof(prolog_key
));
7587 prolog_key
.gs_prolog
.states
= shader
->key
.part
.gs
.prolog
;
7589 shader
->prolog2
= si_get_shader_part(sscreen
, &sscreen
->gs_prologs
,
7590 PIPE_SHADER_GEOMETRY
, true,
7591 &prolog_key
, compiler
, debug
,
7592 si_build_gs_prolog_function
,
7593 "Geometry Shader Prolog");
7594 return shader
->prolog2
!= NULL
;
7598 * Build the pixel shader prolog function. This handles:
7599 * - two-side color selection and interpolation
7600 * - overriding interpolation parameters for the API PS
7601 * - polygon stippling
7603 * All preloaded SGPRs and VGPRs are passed through unmodified unless they are
7604 * overriden by other states. (e.g. per-sample interpolation)
7605 * Interpolated colors are stored after the preloaded VGPRs.
7607 static void si_build_ps_prolog_function(struct si_shader_context
*ctx
,
7608 union si_shader_part_key
*key
)
7610 struct si_function_info fninfo
;
7611 LLVMValueRef ret
, func
;
7612 int num_returns
, i
, num_color_channels
;
7614 assert(si_need_ps_prolog(key
));
7616 si_init_function_info(&fninfo
);
7618 /* Declare inputs. */
7619 for (i
= 0; i
< key
->ps_prolog
.num_input_sgprs
; i
++)
7620 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7622 for (i
= 0; i
< key
->ps_prolog
.num_input_vgprs
; i
++)
7623 add_arg(&fninfo
, ARG_VGPR
, ctx
->f32
);
7625 /* Declare outputs (same as inputs + add colors if needed) */
7626 num_returns
= fninfo
.num_params
;
7627 num_color_channels
= util_bitcount(key
->ps_prolog
.colors_read
);
7628 for (i
= 0; i
< num_color_channels
; i
++)
7629 fninfo
.types
[num_returns
++] = ctx
->f32
;
7631 /* Create the function. */
7632 si_create_function(ctx
, "ps_prolog", fninfo
.types
, num_returns
,
7634 func
= ctx
->main_fn
;
7636 /* Copy inputs to outputs. This should be no-op, as the registers match,
7637 * but it will prevent the compiler from overwriting them unintentionally.
7639 ret
= ctx
->return_value
;
7640 for (i
= 0; i
< fninfo
.num_params
; i
++) {
7641 LLVMValueRef p
= LLVMGetParam(func
, i
);
7642 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, p
, i
, "");
7645 /* Polygon stippling. */
7646 if (key
->ps_prolog
.states
.poly_stipple
) {
7647 /* POS_FIXED_PT is always last. */
7648 unsigned pos
= key
->ps_prolog
.num_input_sgprs
+
7649 key
->ps_prolog
.num_input_vgprs
- 1;
7650 LLVMValueRef list
= si_prolog_get_rw_buffers(ctx
);
7652 si_llvm_emit_polygon_stipple(ctx
, list
, pos
);
7655 if (key
->ps_prolog
.states
.bc_optimize_for_persp
||
7656 key
->ps_prolog
.states
.bc_optimize_for_linear
) {
7657 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7658 LLVMValueRef center
[2], centroid
[2], tmp
, bc_optimize
;
7660 /* The shader should do: if (PRIM_MASK[31]) CENTROID = CENTER;
7661 * The hw doesn't compute CENTROID if the whole wave only
7662 * contains fully-covered quads.
7664 * PRIM_MASK is after user SGPRs.
7666 bc_optimize
= LLVMGetParam(func
, SI_PS_NUM_USER_SGPR
);
7667 bc_optimize
= LLVMBuildLShr(ctx
->ac
.builder
, bc_optimize
,
7668 LLVMConstInt(ctx
->i32
, 31, 0), "");
7669 bc_optimize
= LLVMBuildTrunc(ctx
->ac
.builder
, bc_optimize
,
7672 if (key
->ps_prolog
.states
.bc_optimize_for_persp
) {
7673 /* Read PERSP_CENTER. */
7674 for (i
= 0; i
< 2; i
++)
7675 center
[i
] = LLVMGetParam(func
, base
+ 2 + i
);
7676 /* Read PERSP_CENTROID. */
7677 for (i
= 0; i
< 2; i
++)
7678 centroid
[i
] = LLVMGetParam(func
, base
+ 4 + i
);
7679 /* Select PERSP_CENTROID. */
7680 for (i
= 0; i
< 2; i
++) {
7681 tmp
= LLVMBuildSelect(ctx
->ac
.builder
, bc_optimize
,
7682 center
[i
], centroid
[i
], "");
7683 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7684 tmp
, base
+ 4 + i
, "");
7687 if (key
->ps_prolog
.states
.bc_optimize_for_linear
) {
7688 /* Read LINEAR_CENTER. */
7689 for (i
= 0; i
< 2; i
++)
7690 center
[i
] = LLVMGetParam(func
, base
+ 8 + i
);
7691 /* Read LINEAR_CENTROID. */
7692 for (i
= 0; i
< 2; i
++)
7693 centroid
[i
] = LLVMGetParam(func
, base
+ 10 + i
);
7694 /* Select LINEAR_CENTROID. */
7695 for (i
= 0; i
< 2; i
++) {
7696 tmp
= LLVMBuildSelect(ctx
->ac
.builder
, bc_optimize
,
7697 center
[i
], centroid
[i
], "");
7698 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7699 tmp
, base
+ 10 + i
, "");
7704 /* Force per-sample interpolation. */
7705 if (key
->ps_prolog
.states
.force_persp_sample_interp
) {
7706 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7707 LLVMValueRef persp_sample
[2];
7709 /* Read PERSP_SAMPLE. */
7710 for (i
= 0; i
< 2; i
++)
7711 persp_sample
[i
] = LLVMGetParam(func
, base
+ i
);
7712 /* Overwrite PERSP_CENTER. */
7713 for (i
= 0; i
< 2; i
++)
7714 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7715 persp_sample
[i
], base
+ 2 + i
, "");
7716 /* Overwrite PERSP_CENTROID. */
7717 for (i
= 0; i
< 2; i
++)
7718 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7719 persp_sample
[i
], base
+ 4 + i
, "");
7721 if (key
->ps_prolog
.states
.force_linear_sample_interp
) {
7722 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7723 LLVMValueRef linear_sample
[2];
7725 /* Read LINEAR_SAMPLE. */
7726 for (i
= 0; i
< 2; i
++)
7727 linear_sample
[i
] = LLVMGetParam(func
, base
+ 6 + i
);
7728 /* Overwrite LINEAR_CENTER. */
7729 for (i
= 0; i
< 2; i
++)
7730 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7731 linear_sample
[i
], base
+ 8 + i
, "");
7732 /* Overwrite LINEAR_CENTROID. */
7733 for (i
= 0; i
< 2; i
++)
7734 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7735 linear_sample
[i
], base
+ 10 + i
, "");
7738 /* Force center interpolation. */
7739 if (key
->ps_prolog
.states
.force_persp_center_interp
) {
7740 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7741 LLVMValueRef persp_center
[2];
7743 /* Read PERSP_CENTER. */
7744 for (i
= 0; i
< 2; i
++)
7745 persp_center
[i
] = LLVMGetParam(func
, base
+ 2 + i
);
7746 /* Overwrite PERSP_SAMPLE. */
7747 for (i
= 0; i
< 2; i
++)
7748 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7749 persp_center
[i
], base
+ i
, "");
7750 /* Overwrite PERSP_CENTROID. */
7751 for (i
= 0; i
< 2; i
++)
7752 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7753 persp_center
[i
], base
+ 4 + i
, "");
7755 if (key
->ps_prolog
.states
.force_linear_center_interp
) {
7756 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7757 LLVMValueRef linear_center
[2];
7759 /* Read LINEAR_CENTER. */
7760 for (i
= 0; i
< 2; i
++)
7761 linear_center
[i
] = LLVMGetParam(func
, base
+ 8 + i
);
7762 /* Overwrite LINEAR_SAMPLE. */
7763 for (i
= 0; i
< 2; i
++)
7764 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7765 linear_center
[i
], base
+ 6 + i
, "");
7766 /* Overwrite LINEAR_CENTROID. */
7767 for (i
= 0; i
< 2; i
++)
7768 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7769 linear_center
[i
], base
+ 10 + i
, "");
7772 /* Interpolate colors. */
7773 unsigned color_out_idx
= 0;
7774 for (i
= 0; i
< 2; i
++) {
7775 unsigned writemask
= (key
->ps_prolog
.colors_read
>> (i
* 4)) & 0xf;
7776 unsigned face_vgpr
= key
->ps_prolog
.num_input_sgprs
+
7777 key
->ps_prolog
.face_vgpr_index
;
7778 LLVMValueRef interp
[2], color
[4];
7779 LLVMValueRef interp_ij
= NULL
, prim_mask
= NULL
, face
= NULL
;
7784 /* If the interpolation qualifier is not CONSTANT (-1). */
7785 if (key
->ps_prolog
.color_interp_vgpr_index
[i
] != -1) {
7786 unsigned interp_vgpr
= key
->ps_prolog
.num_input_sgprs
+
7787 key
->ps_prolog
.color_interp_vgpr_index
[i
];
7789 /* Get the (i,j) updated by bc_optimize handling. */
7790 interp
[0] = LLVMBuildExtractValue(ctx
->ac
.builder
, ret
,
7792 interp
[1] = LLVMBuildExtractValue(ctx
->ac
.builder
, ret
,
7793 interp_vgpr
+ 1, "");
7794 interp_ij
= ac_build_gather_values(&ctx
->ac
, interp
, 2);
7797 /* Use the absolute location of the input. */
7798 prim_mask
= LLVMGetParam(func
, SI_PS_NUM_USER_SGPR
);
7800 if (key
->ps_prolog
.states
.color_two_side
) {
7801 face
= LLVMGetParam(func
, face_vgpr
);
7802 face
= ac_to_integer(&ctx
->ac
, face
);
7805 interp_fs_input(ctx
,
7806 key
->ps_prolog
.color_attr_index
[i
],
7807 TGSI_SEMANTIC_COLOR
, i
,
7808 key
->ps_prolog
.num_interp_inputs
,
7809 key
->ps_prolog
.colors_read
, interp_ij
,
7810 prim_mask
, face
, color
);
7813 unsigned chan
= u_bit_scan(&writemask
);
7814 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, color
[chan
],
7815 fninfo
.num_params
+ color_out_idx
++, "");
7819 /* Section 15.2.2 (Shader Inputs) of the OpenGL 4.5 (Core Profile) spec
7822 * "When per-sample shading is active due to the use of a fragment
7823 * input qualified by sample or due to the use of the gl_SampleID
7824 * or gl_SamplePosition variables, only the bit for the current
7825 * sample is set in gl_SampleMaskIn. When state specifies multiple
7826 * fragment shader invocations for a given fragment, the sample
7827 * mask for any single fragment shader invocation may specify a
7828 * subset of the covered samples for the fragment. In this case,
7829 * the bit corresponding to each covered sample will be set in
7830 * exactly one fragment shader invocation."
7832 * The samplemask loaded by hardware is always the coverage of the
7833 * entire pixel/fragment, so mask bits out based on the sample ID.
7835 if (key
->ps_prolog
.states
.samplemask_log_ps_iter
) {
7836 /* The bit pattern matches that used by fixed function fragment
7838 static const uint16_t ps_iter_masks
[] = {
7839 0xffff, /* not used */
7845 assert(key
->ps_prolog
.states
.samplemask_log_ps_iter
< ARRAY_SIZE(ps_iter_masks
));
7847 uint32_t ps_iter_mask
= ps_iter_masks
[key
->ps_prolog
.states
.samplemask_log_ps_iter
];
7848 unsigned ancillary_vgpr
= key
->ps_prolog
.num_input_sgprs
+
7849 key
->ps_prolog
.ancillary_vgpr_index
;
7850 LLVMValueRef sampleid
= si_unpack_param(ctx
, ancillary_vgpr
, 8, 4);
7851 LLVMValueRef samplemask
= LLVMGetParam(func
, ancillary_vgpr
+ 1);
7853 samplemask
= ac_to_integer(&ctx
->ac
, samplemask
);
7854 samplemask
= LLVMBuildAnd(
7857 LLVMBuildShl(ctx
->ac
.builder
,
7858 LLVMConstInt(ctx
->i32
, ps_iter_mask
, false),
7861 samplemask
= ac_to_float(&ctx
->ac
, samplemask
);
7863 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, samplemask
,
7864 ancillary_vgpr
+ 1, "");
7867 /* Tell LLVM to insert WQM instruction sequence when needed. */
7868 if (key
->ps_prolog
.wqm
) {
7869 LLVMAddTargetDependentFunctionAttr(func
,
7870 "amdgpu-ps-wqm-outputs", "");
7873 si_llvm_build_ret(ctx
, ret
);
7877 * Build the pixel shader epilog function. This handles everything that must be
7878 * emulated for pixel shader exports. (alpha-test, format conversions, etc)
7880 static void si_build_ps_epilog_function(struct si_shader_context
*ctx
,
7881 union si_shader_part_key
*key
)
7883 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
7884 struct si_function_info fninfo
;
7885 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
7887 struct si_ps_exports exp
= {};
7889 si_init_function_info(&fninfo
);
7891 /* Declare input SGPRs. */
7892 ctx
->param_rw_buffers
= add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7893 ctx
->param_bindless_samplers_and_images
= add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7894 ctx
->param_const_and_shader_buffers
= add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7895 ctx
->param_samplers_and_images
= add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7896 add_arg_checked(&fninfo
, ARG_SGPR
, ctx
->f32
, SI_PARAM_ALPHA_REF
);
7898 /* Declare input VGPRs. */
7899 unsigned required_num_params
=
7900 fninfo
.num_sgpr_params
+
7901 util_bitcount(key
->ps_epilog
.colors_written
) * 4 +
7902 key
->ps_epilog
.writes_z
+
7903 key
->ps_epilog
.writes_stencil
+
7904 key
->ps_epilog
.writes_samplemask
;
7906 required_num_params
= MAX2(required_num_params
,
7907 fninfo
.num_sgpr_params
+ PS_EPILOG_SAMPLEMASK_MIN_LOC
+ 1);
7909 while (fninfo
.num_params
< required_num_params
)
7910 add_arg(&fninfo
, ARG_VGPR
, ctx
->f32
);
7912 /* Create the function. */
7913 si_create_function(ctx
, "ps_epilog", NULL
, 0, &fninfo
, 0);
7914 /* Disable elimination of unused inputs. */
7915 ac_llvm_add_target_dep_function_attr(ctx
->main_fn
,
7916 "InitialPSInputAddr", 0xffffff);
7918 /* Process colors. */
7919 unsigned vgpr
= fninfo
.num_sgpr_params
;
7920 unsigned colors_written
= key
->ps_epilog
.colors_written
;
7921 int last_color_export
= -1;
7923 /* Find the last color export. */
7924 if (!key
->ps_epilog
.writes_z
&&
7925 !key
->ps_epilog
.writes_stencil
&&
7926 !key
->ps_epilog
.writes_samplemask
) {
7927 unsigned spi_format
= key
->ps_epilog
.states
.spi_shader_col_format
;
7929 /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
7930 if (colors_written
== 0x1 && key
->ps_epilog
.states
.last_cbuf
> 0) {
7931 /* Just set this if any of the colorbuffers are enabled. */
7933 ((1ull << (4 * (key
->ps_epilog
.states
.last_cbuf
+ 1))) - 1))
7934 last_color_export
= 0;
7936 for (i
= 0; i
< 8; i
++)
7937 if (colors_written
& (1 << i
) &&
7938 (spi_format
>> (i
* 4)) & 0xf)
7939 last_color_export
= i
;
7943 while (colors_written
) {
7944 LLVMValueRef color
[4];
7945 int mrt
= u_bit_scan(&colors_written
);
7947 for (i
= 0; i
< 4; i
++)
7948 color
[i
] = LLVMGetParam(ctx
->main_fn
, vgpr
++);
7950 si_export_mrt_color(bld_base
, color
, mrt
,
7951 fninfo
.num_params
- 1,
7952 mrt
== last_color_export
, &exp
);
7955 /* Process depth, stencil, samplemask. */
7956 if (key
->ps_epilog
.writes_z
)
7957 depth
= LLVMGetParam(ctx
->main_fn
, vgpr
++);
7958 if (key
->ps_epilog
.writes_stencil
)
7959 stencil
= LLVMGetParam(ctx
->main_fn
, vgpr
++);
7960 if (key
->ps_epilog
.writes_samplemask
)
7961 samplemask
= LLVMGetParam(ctx
->main_fn
, vgpr
++);
7963 if (depth
|| stencil
|| samplemask
)
7964 si_export_mrt_z(bld_base
, depth
, stencil
, samplemask
, &exp
);
7965 else if (last_color_export
== -1)
7966 ac_build_export_null(&ctx
->ac
);
7969 si_emit_ps_exports(ctx
, &exp
);
7972 LLVMBuildRetVoid(ctx
->ac
.builder
);
7976 * Select and compile (or reuse) pixel shader parts (prolog & epilog).
7978 static bool si_shader_select_ps_parts(struct si_screen
*sscreen
,
7979 struct ac_llvm_compiler
*compiler
,
7980 struct si_shader
*shader
,
7981 struct pipe_debug_callback
*debug
)
7983 union si_shader_part_key prolog_key
;
7984 union si_shader_part_key epilog_key
;
7986 /* Get the prolog. */
7987 si_get_ps_prolog_key(shader
, &prolog_key
, true);
7989 /* The prolog is a no-op if these aren't set. */
7990 if (si_need_ps_prolog(&prolog_key
)) {
7992 si_get_shader_part(sscreen
, &sscreen
->ps_prologs
,
7993 PIPE_SHADER_FRAGMENT
, true,
7994 &prolog_key
, compiler
, debug
,
7995 si_build_ps_prolog_function
,
7996 "Fragment Shader Prolog");
7997 if (!shader
->prolog
)
8001 /* Get the epilog. */
8002 si_get_ps_epilog_key(shader
, &epilog_key
);
8005 si_get_shader_part(sscreen
, &sscreen
->ps_epilogs
,
8006 PIPE_SHADER_FRAGMENT
, false,
8007 &epilog_key
, compiler
, debug
,
8008 si_build_ps_epilog_function
,
8009 "Fragment Shader Epilog");
8010 if (!shader
->epilog
)
8013 /* Enable POS_FIXED_PT if polygon stippling is enabled. */
8014 if (shader
->key
.part
.ps
.prolog
.poly_stipple
) {
8015 shader
->config
.spi_ps_input_ena
|= S_0286CC_POS_FIXED_PT_ENA(1);
8016 assert(G_0286CC_POS_FIXED_PT_ENA(shader
->config
.spi_ps_input_addr
));
8019 /* Set up the enable bits for per-sample shading if needed. */
8020 if (shader
->key
.part
.ps
.prolog
.force_persp_sample_interp
&&
8021 (G_0286CC_PERSP_CENTER_ENA(shader
->config
.spi_ps_input_ena
) ||
8022 G_0286CC_PERSP_CENTROID_ENA(shader
->config
.spi_ps_input_ena
))) {
8023 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_CENTER_ENA
;
8024 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_CENTROID_ENA
;
8025 shader
->config
.spi_ps_input_ena
|= S_0286CC_PERSP_SAMPLE_ENA(1);
8027 if (shader
->key
.part
.ps
.prolog
.force_linear_sample_interp
&&
8028 (G_0286CC_LINEAR_CENTER_ENA(shader
->config
.spi_ps_input_ena
) ||
8029 G_0286CC_LINEAR_CENTROID_ENA(shader
->config
.spi_ps_input_ena
))) {
8030 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_CENTER_ENA
;
8031 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_CENTROID_ENA
;
8032 shader
->config
.spi_ps_input_ena
|= S_0286CC_LINEAR_SAMPLE_ENA(1);
8034 if (shader
->key
.part
.ps
.prolog
.force_persp_center_interp
&&
8035 (G_0286CC_PERSP_SAMPLE_ENA(shader
->config
.spi_ps_input_ena
) ||
8036 G_0286CC_PERSP_CENTROID_ENA(shader
->config
.spi_ps_input_ena
))) {
8037 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_SAMPLE_ENA
;
8038 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_CENTROID_ENA
;
8039 shader
->config
.spi_ps_input_ena
|= S_0286CC_PERSP_CENTER_ENA(1);
8041 if (shader
->key
.part
.ps
.prolog
.force_linear_center_interp
&&
8042 (G_0286CC_LINEAR_SAMPLE_ENA(shader
->config
.spi_ps_input_ena
) ||
8043 G_0286CC_LINEAR_CENTROID_ENA(shader
->config
.spi_ps_input_ena
))) {
8044 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_SAMPLE_ENA
;
8045 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_CENTROID_ENA
;
8046 shader
->config
.spi_ps_input_ena
|= S_0286CC_LINEAR_CENTER_ENA(1);
8049 /* POW_W_FLOAT requires that one of the perspective weights is enabled. */
8050 if (G_0286CC_POS_W_FLOAT_ENA(shader
->config
.spi_ps_input_ena
) &&
8051 !(shader
->config
.spi_ps_input_ena
& 0xf)) {
8052 shader
->config
.spi_ps_input_ena
|= S_0286CC_PERSP_CENTER_ENA(1);
8053 assert(G_0286CC_PERSP_CENTER_ENA(shader
->config
.spi_ps_input_addr
));
8056 /* At least one pair of interpolation weights must be enabled. */
8057 if (!(shader
->config
.spi_ps_input_ena
& 0x7f)) {
8058 shader
->config
.spi_ps_input_ena
|= S_0286CC_LINEAR_CENTER_ENA(1);
8059 assert(G_0286CC_LINEAR_CENTER_ENA(shader
->config
.spi_ps_input_addr
));
8062 /* Samplemask fixup requires the sample ID. */
8063 if (shader
->key
.part
.ps
.prolog
.samplemask_log_ps_iter
) {
8064 shader
->config
.spi_ps_input_ena
|= S_0286CC_ANCILLARY_ENA(1);
8065 assert(G_0286CC_ANCILLARY_ENA(shader
->config
.spi_ps_input_addr
));
8068 /* The sample mask input is always enabled, because the API shader always
8069 * passes it through to the epilog. Disable it here if it's unused.
8071 if (!shader
->key
.part
.ps
.epilog
.poly_line_smoothing
&&
8072 !shader
->selector
->info
.reads_samplemask
)
8073 shader
->config
.spi_ps_input_ena
&= C_0286CC_SAMPLE_COVERAGE_ENA
;
8078 void si_multiwave_lds_size_workaround(struct si_screen
*sscreen
,
8081 /* If tessellation is all offchip and on-chip GS isn't used, this
8082 * workaround is not needed.
8086 /* SPI barrier management bug:
8087 * Make sure we have at least 4k of LDS in use to avoid the bug.
8088 * It applies to workgroup sizes of more than one wavefront.
8090 if (sscreen
->info
.family
== CHIP_BONAIRE
||
8091 sscreen
->info
.family
== CHIP_KABINI
)
8092 *lds_size
= MAX2(*lds_size
, 8);
8095 static void si_fix_resource_usage(struct si_screen
*sscreen
,
8096 struct si_shader
*shader
)
8098 unsigned min_sgprs
= shader
->info
.num_input_sgprs
+ 2; /* VCC */
8100 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
, min_sgprs
);
8102 if (shader
->selector
->type
== PIPE_SHADER_COMPUTE
&&
8103 si_get_max_workgroup_size(shader
) > 64) {
8104 si_multiwave_lds_size_workaround(sscreen
,
8105 &shader
->config
.lds_size
);
8109 bool si_shader_create(struct si_screen
*sscreen
, struct ac_llvm_compiler
*compiler
,
8110 struct si_shader
*shader
,
8111 struct pipe_debug_callback
*debug
)
8113 struct si_shader_selector
*sel
= shader
->selector
;
8114 struct si_shader
*mainp
= *si_get_main_shader_part(sel
, &shader
->key
);
8117 /* LS, ES, VS are compiled on demand if the main part hasn't been
8118 * compiled for that stage.
8120 * GS are compiled on demand if the main part hasn't been compiled
8121 * for the chosen NGG-ness.
8123 * Vertex shaders are compiled on demand when a vertex fetch
8124 * workaround must be applied.
8126 if (shader
->is_monolithic
) {
8127 /* Monolithic shader (compiled as a whole, has many variants,
8128 * may take a long time to compile).
8130 r
= si_compile_tgsi_shader(sscreen
, compiler
, shader
, debug
);
8134 /* The shader consists of several parts:
8136 * - the middle part is the user shader, it has 1 variant only
8137 * and it was compiled during the creation of the shader
8139 * - the prolog part is inserted at the beginning
8140 * - the epilog part is inserted at the end
8142 * The prolog and epilog have many (but simple) variants.
8144 * Starting with gfx9, geometry and tessellation control
8145 * shaders also contain the prolog and user shader parts of
8146 * the previous shader stage.
8152 /* Copy the compiled TGSI shader data over. */
8153 shader
->is_binary_shared
= true;
8154 shader
->binary
= mainp
->binary
;
8155 shader
->config
= mainp
->config
;
8156 shader
->info
.num_input_sgprs
= mainp
->info
.num_input_sgprs
;
8157 shader
->info
.num_input_vgprs
= mainp
->info
.num_input_vgprs
;
8158 shader
->info
.face_vgpr_index
= mainp
->info
.face_vgpr_index
;
8159 shader
->info
.ancillary_vgpr_index
= mainp
->info
.ancillary_vgpr_index
;
8160 memcpy(shader
->info
.vs_output_param_offset
,
8161 mainp
->info
.vs_output_param_offset
,
8162 sizeof(mainp
->info
.vs_output_param_offset
));
8163 shader
->info
.uses_instanceid
= mainp
->info
.uses_instanceid
;
8164 shader
->info
.nr_pos_exports
= mainp
->info
.nr_pos_exports
;
8165 shader
->info
.nr_param_exports
= mainp
->info
.nr_param_exports
;
8167 /* Select prologs and/or epilogs. */
8168 switch (sel
->type
) {
8169 case PIPE_SHADER_VERTEX
:
8170 if (!si_shader_select_vs_parts(sscreen
, compiler
, shader
, debug
))
8173 case PIPE_SHADER_TESS_CTRL
:
8174 if (!si_shader_select_tcs_parts(sscreen
, compiler
, shader
, debug
))
8177 case PIPE_SHADER_TESS_EVAL
:
8179 case PIPE_SHADER_GEOMETRY
:
8180 if (!si_shader_select_gs_parts(sscreen
, compiler
, shader
, debug
))
8183 case PIPE_SHADER_FRAGMENT
:
8184 if (!si_shader_select_ps_parts(sscreen
, compiler
, shader
, debug
))
8187 /* Make sure we have at least as many VGPRs as there
8188 * are allocated inputs.
8190 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
8191 shader
->info
.num_input_vgprs
);
8195 /* Update SGPR and VGPR counts. */
8196 if (shader
->prolog
) {
8197 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
,
8198 shader
->prolog
->config
.num_sgprs
);
8199 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
8200 shader
->prolog
->config
.num_vgprs
);
8202 if (shader
->previous_stage
) {
8203 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
,
8204 shader
->previous_stage
->config
.num_sgprs
);
8205 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
8206 shader
->previous_stage
->config
.num_vgprs
);
8207 shader
->config
.spilled_sgprs
=
8208 MAX2(shader
->config
.spilled_sgprs
,
8209 shader
->previous_stage
->config
.spilled_sgprs
);
8210 shader
->config
.spilled_vgprs
=
8211 MAX2(shader
->config
.spilled_vgprs
,
8212 shader
->previous_stage
->config
.spilled_vgprs
);
8213 shader
->info
.private_mem_vgprs
=
8214 MAX2(shader
->info
.private_mem_vgprs
,
8215 shader
->previous_stage
->info
.private_mem_vgprs
);
8216 shader
->config
.scratch_bytes_per_wave
=
8217 MAX2(shader
->config
.scratch_bytes_per_wave
,
8218 shader
->previous_stage
->config
.scratch_bytes_per_wave
);
8219 shader
->info
.uses_instanceid
|=
8220 shader
->previous_stage
->info
.uses_instanceid
;
8222 if (shader
->prolog2
) {
8223 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
,
8224 shader
->prolog2
->config
.num_sgprs
);
8225 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
8226 shader
->prolog2
->config
.num_vgprs
);
8228 if (shader
->epilog
) {
8229 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
,
8230 shader
->epilog
->config
.num_sgprs
);
8231 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
8232 shader
->epilog
->config
.num_vgprs
);
8234 si_calculate_max_simd_waves(shader
);
8237 if (shader
->key
.as_ngg
) {
8238 assert(!shader
->key
.as_es
&& !shader
->key
.as_ls
);
8239 gfx10_ngg_calculate_subgroup_info(shader
);
8240 } else if (sscreen
->info
.chip_class
>= GFX9
&& sel
->type
== PIPE_SHADER_GEOMETRY
) {
8241 gfx9_get_gs_info(shader
->previous_stage_sel
, sel
, &shader
->gs_info
);
8244 si_fix_resource_usage(sscreen
, shader
);
8245 si_shader_dump(sscreen
, shader
, debug
, sel
->info
.processor
,
8249 if (!si_shader_binary_upload(sscreen
, shader
, 0)) {
8250 fprintf(stderr
, "LLVM failed to upload shader\n");
8257 void si_shader_destroy(struct si_shader
*shader
)
8259 if (shader
->scratch_bo
)
8260 si_resource_reference(&shader
->scratch_bo
, NULL
);
8262 si_resource_reference(&shader
->bo
, NULL
);
8264 if (!shader
->is_binary_shared
)
8265 si_shader_binary_clean(&shader
->binary
);
8267 free(shader
->shader_log
);