2 * Copyright 2012 Advanced Micro Devices, Inc.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
24 * Tom Stellard <thomas.stellard@amd.com>
25 * Michel Dänzer <michel.daenzer@amd.com>
26 * Christian König <christian.koenig@amd.com>
29 #include "gallivm/lp_bld_const.h"
30 #include "gallivm/lp_bld_gather.h"
31 #include "gallivm/lp_bld_intr.h"
32 #include "gallivm/lp_bld_logic.h"
33 #include "gallivm/lp_bld_arit.h"
34 #include "gallivm/lp_bld_flow.h"
35 #include "gallivm/lp_bld_misc.h"
36 #include "util/u_memory.h"
37 #include "util/u_string.h"
38 #include "tgsi/tgsi_build.h"
39 #include "tgsi/tgsi_util.h"
40 #include "tgsi/tgsi_dump.h"
42 #include "ac_binary.h"
43 #include "ac_llvm_util.h"
44 #include "ac_exp_param.h"
45 #include "si_shader_internal.h"
49 #include "compiler/nir/nir.h"
51 static const char *scratch_rsrc_dword0_symbol
=
52 "SCRATCH_RSRC_DWORD0";
54 static const char *scratch_rsrc_dword1_symbol
=
55 "SCRATCH_RSRC_DWORD1";
57 struct si_shader_output_values
59 LLVMValueRef values
[4];
60 unsigned semantic_name
;
61 unsigned semantic_index
;
62 ubyte vertex_stream
[4];
66 * Used to collect types and other info about arguments of the LLVM function
67 * before the function is created.
69 struct si_function_info
{
70 LLVMTypeRef types
[100];
71 LLVMValueRef
*assign
[100];
72 unsigned num_sgpr_params
;
81 static void si_init_shader_ctx(struct si_shader_context
*ctx
,
82 struct si_screen
*sscreen
,
83 LLVMTargetMachineRef tm
);
85 static void si_llvm_emit_barrier(const struct lp_build_tgsi_action
*action
,
86 struct lp_build_tgsi_context
*bld_base
,
87 struct lp_build_emit_data
*emit_data
);
89 static void si_dump_shader_key(unsigned processor
, const struct si_shader
*shader
,
92 static void si_build_vs_prolog_function(struct si_shader_context
*ctx
,
93 union si_shader_part_key
*key
);
94 static void si_build_tcs_epilog_function(struct si_shader_context
*ctx
,
95 union si_shader_part_key
*key
);
96 static void si_build_ps_prolog_function(struct si_shader_context
*ctx
,
97 union si_shader_part_key
*key
);
98 static void si_build_ps_epilog_function(struct si_shader_context
*ctx
,
99 union si_shader_part_key
*key
);
101 /* Ideally pass the sample mask input to the PS epilog as v14, which
102 * is its usual location, so that the shader doesn't have to add v_mov.
104 #define PS_EPILOG_SAMPLEMASK_MIN_LOC 14
107 CONST_ADDR_SPACE
= 2,
108 LOCAL_ADDR_SPACE
= 3,
111 static bool is_merged_shader(struct si_shader
*shader
)
113 if (shader
->selector
->screen
->b
.chip_class
<= VI
)
116 return shader
->key
.as_ls
||
118 shader
->selector
->type
== PIPE_SHADER_TESS_CTRL
||
119 shader
->selector
->type
== PIPE_SHADER_GEOMETRY
;
122 static void si_init_function_info(struct si_function_info
*fninfo
)
124 fninfo
->num_params
= 0;
125 fninfo
->num_sgpr_params
= 0;
128 static unsigned add_arg_assign(struct si_function_info
*fninfo
,
129 enum si_arg_regfile regfile
, LLVMTypeRef type
,
130 LLVMValueRef
*assign
)
132 assert(regfile
!= ARG_SGPR
|| fninfo
->num_sgpr_params
== fninfo
->num_params
);
134 unsigned idx
= fninfo
->num_params
++;
135 assert(idx
< ARRAY_SIZE(fninfo
->types
));
137 if (regfile
== ARG_SGPR
)
138 fninfo
->num_sgpr_params
= fninfo
->num_params
;
140 fninfo
->types
[idx
] = type
;
141 fninfo
->assign
[idx
] = assign
;
145 static unsigned add_arg(struct si_function_info
*fninfo
,
146 enum si_arg_regfile regfile
, LLVMTypeRef type
)
148 return add_arg_assign(fninfo
, regfile
, type
, NULL
);
151 static void add_arg_assign_checked(struct si_function_info
*fninfo
,
152 enum si_arg_regfile regfile
, LLVMTypeRef type
,
153 LLVMValueRef
*assign
, unsigned idx
)
155 MAYBE_UNUSED
unsigned actual
= add_arg_assign(fninfo
, regfile
, type
, assign
);
156 assert(actual
== idx
);
159 static void add_arg_checked(struct si_function_info
*fninfo
,
160 enum si_arg_regfile regfile
, LLVMTypeRef type
,
163 add_arg_assign_checked(fninfo
, regfile
, type
, NULL
, idx
);
167 * Returns a unique index for a per-patch semantic name and index. The index
168 * must be less than 32, so that a 32-bit bitmask of used inputs or outputs
171 unsigned si_shader_io_get_unique_index_patch(unsigned semantic_name
, unsigned index
)
173 switch (semantic_name
) {
174 case TGSI_SEMANTIC_TESSOUTER
:
176 case TGSI_SEMANTIC_TESSINNER
:
178 case TGSI_SEMANTIC_PATCH
:
183 assert(!"invalid semantic name");
189 * Returns a unique index for a semantic name and index. The index must be
190 * less than 64, so that a 64-bit bitmask of used inputs or outputs can be
193 unsigned si_shader_io_get_unique_index(unsigned semantic_name
, unsigned index
)
195 switch (semantic_name
) {
196 case TGSI_SEMANTIC_POSITION
:
198 case TGSI_SEMANTIC_GENERIC
:
199 /* Since some shader stages use the the highest used IO index
200 * to determine the size to allocate for inputs/outputs
201 * (in LDS, tess and GS rings). GENERIC should be placed right
202 * after POSITION to make that size as small as possible.
204 if (index
< SI_MAX_IO_GENERIC
)
207 assert(!"invalid generic index");
209 case TGSI_SEMANTIC_PSIZE
:
210 return SI_MAX_IO_GENERIC
+ 1;
211 case TGSI_SEMANTIC_CLIPDIST
:
213 return SI_MAX_IO_GENERIC
+ 2 + index
;
214 case TGSI_SEMANTIC_FOG
:
215 return SI_MAX_IO_GENERIC
+ 4;
216 case TGSI_SEMANTIC_LAYER
:
217 return SI_MAX_IO_GENERIC
+ 5;
218 case TGSI_SEMANTIC_VIEWPORT_INDEX
:
219 return SI_MAX_IO_GENERIC
+ 6;
220 case TGSI_SEMANTIC_PRIMID
:
221 return SI_MAX_IO_GENERIC
+ 7;
222 case TGSI_SEMANTIC_COLOR
: /* these alias */
223 case TGSI_SEMANTIC_BCOLOR
:
225 return SI_MAX_IO_GENERIC
+ 8 + index
;
226 case TGSI_SEMANTIC_TEXCOORD
:
228 assert(SI_MAX_IO_GENERIC
+ 10 + index
< 64);
229 return SI_MAX_IO_GENERIC
+ 10 + index
;
231 assert(!"invalid semantic name");
237 * Get the value of a shader input parameter and extract a bitfield.
239 static LLVMValueRef
unpack_param(struct si_shader_context
*ctx
,
240 unsigned param
, unsigned rshift
,
243 LLVMValueRef value
= LLVMGetParam(ctx
->main_fn
,
246 if (LLVMGetTypeKind(LLVMTypeOf(value
)) == LLVMFloatTypeKind
)
247 value
= ac_to_integer(&ctx
->ac
, value
);
250 value
= LLVMBuildLShr(ctx
->ac
.builder
, value
,
251 LLVMConstInt(ctx
->i32
, rshift
, 0), "");
253 if (rshift
+ bitwidth
< 32) {
254 unsigned mask
= (1 << bitwidth
) - 1;
255 value
= LLVMBuildAnd(ctx
->ac
.builder
, value
,
256 LLVMConstInt(ctx
->i32
, mask
, 0), "");
262 static LLVMValueRef
get_rel_patch_id(struct si_shader_context
*ctx
)
265 case PIPE_SHADER_TESS_CTRL
:
266 return unpack_param(ctx
, ctx
->param_tcs_rel_ids
, 0, 8);
268 case PIPE_SHADER_TESS_EVAL
:
269 return LLVMGetParam(ctx
->main_fn
,
270 ctx
->param_tes_rel_patch_id
);
278 /* Tessellation shaders pass outputs to the next shader using LDS.
280 * LS outputs = TCS inputs
281 * TCS outputs = TES inputs
284 * - TCS inputs for patch 0
285 * - TCS inputs for patch 1
286 * - TCS inputs for patch 2 = get_tcs_in_current_patch_offset (if RelPatchID==2)
288 * - TCS outputs for patch 0 = get_tcs_out_patch0_offset
289 * - Per-patch TCS outputs for patch 0 = get_tcs_out_patch0_patch_data_offset
290 * - TCS outputs for patch 1
291 * - Per-patch TCS outputs for patch 1
292 * - TCS outputs for patch 2 = get_tcs_out_current_patch_offset (if RelPatchID==2)
293 * - Per-patch TCS outputs for patch 2 = get_tcs_out_current_patch_data_offset (if RelPatchID==2)
296 * All three shaders VS(LS), TCS, TES share the same LDS space.
300 get_tcs_in_patch_stride(struct si_shader_context
*ctx
)
302 return unpack_param(ctx
, ctx
->param_vs_state_bits
, 8, 13);
305 static unsigned get_tcs_out_vertex_dw_stride_constant(struct si_shader_context
*ctx
)
307 assert(ctx
->type
== PIPE_SHADER_TESS_CTRL
);
309 if (ctx
->shader
->key
.mono
.u
.ff_tcs_inputs_to_copy
)
310 return util_last_bit64(ctx
->shader
->key
.mono
.u
.ff_tcs_inputs_to_copy
) * 4;
312 return util_last_bit64(ctx
->shader
->selector
->outputs_written
) * 4;
315 static LLVMValueRef
get_tcs_out_vertex_dw_stride(struct si_shader_context
*ctx
)
317 unsigned stride
= get_tcs_out_vertex_dw_stride_constant(ctx
);
319 return LLVMConstInt(ctx
->i32
, stride
, 0);
322 static LLVMValueRef
get_tcs_out_patch_stride(struct si_shader_context
*ctx
)
324 if (ctx
->shader
->key
.mono
.u
.ff_tcs_inputs_to_copy
)
325 return unpack_param(ctx
, ctx
->param_tcs_out_lds_layout
, 0, 13);
327 const struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
328 unsigned tcs_out_vertices
= info
->properties
[TGSI_PROPERTY_TCS_VERTICES_OUT
];
329 unsigned vertex_dw_stride
= get_tcs_out_vertex_dw_stride_constant(ctx
);
330 unsigned num_patch_outputs
= util_last_bit64(ctx
->shader
->selector
->patch_outputs_written
);
331 unsigned patch_dw_stride
= tcs_out_vertices
* vertex_dw_stride
+
332 num_patch_outputs
* 4;
333 return LLVMConstInt(ctx
->i32
, patch_dw_stride
, 0);
337 get_tcs_out_patch0_offset(struct si_shader_context
*ctx
)
339 return lp_build_mul_imm(&ctx
->bld_base
.uint_bld
,
341 ctx
->param_tcs_out_lds_offsets
,
347 get_tcs_out_patch0_patch_data_offset(struct si_shader_context
*ctx
)
349 return lp_build_mul_imm(&ctx
->bld_base
.uint_bld
,
351 ctx
->param_tcs_out_lds_offsets
,
357 get_tcs_in_current_patch_offset(struct si_shader_context
*ctx
)
359 LLVMValueRef patch_stride
= get_tcs_in_patch_stride(ctx
);
360 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
362 return LLVMBuildMul(ctx
->ac
.builder
, patch_stride
, rel_patch_id
, "");
366 get_tcs_out_current_patch_offset(struct si_shader_context
*ctx
)
368 LLVMValueRef patch0_offset
= get_tcs_out_patch0_offset(ctx
);
369 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(ctx
);
370 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
372 return LLVMBuildAdd(ctx
->ac
.builder
, patch0_offset
,
373 LLVMBuildMul(ctx
->ac
.builder
, patch_stride
,
379 get_tcs_out_current_patch_data_offset(struct si_shader_context
*ctx
)
381 LLVMValueRef patch0_patch_data_offset
=
382 get_tcs_out_patch0_patch_data_offset(ctx
);
383 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(ctx
);
384 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
386 return LLVMBuildAdd(ctx
->ac
.builder
, patch0_patch_data_offset
,
387 LLVMBuildMul(ctx
->ac
.builder
, patch_stride
,
392 static LLVMValueRef
get_num_tcs_out_vertices(struct si_shader_context
*ctx
)
394 unsigned tcs_out_vertices
=
395 ctx
->shader
->selector
?
396 ctx
->shader
->selector
->info
.properties
[TGSI_PROPERTY_TCS_VERTICES_OUT
] : 0;
398 /* If !tcs_out_vertices, it's either the fixed-func TCS or the TCS epilog. */
399 if (ctx
->type
== PIPE_SHADER_TESS_CTRL
&& tcs_out_vertices
)
400 return LLVMConstInt(ctx
->i32
, tcs_out_vertices
, 0);
402 return unpack_param(ctx
, ctx
->param_tcs_offchip_layout
, 6, 6);
405 static LLVMValueRef
get_tcs_in_vertex_dw_stride(struct si_shader_context
*ctx
)
410 case PIPE_SHADER_VERTEX
:
411 stride
= util_last_bit64(ctx
->shader
->selector
->outputs_written
);
412 return LLVMConstInt(ctx
->i32
, stride
* 4, 0);
414 case PIPE_SHADER_TESS_CTRL
:
415 if (ctx
->screen
->b
.chip_class
>= GFX9
&&
416 ctx
->shader
->is_monolithic
) {
417 stride
= util_last_bit64(ctx
->shader
->key
.part
.tcs
.ls
->outputs_written
);
418 return LLVMConstInt(ctx
->i32
, stride
* 4, 0);
420 return unpack_param(ctx
, ctx
->param_vs_state_bits
, 24, 8);
428 static LLVMValueRef
get_instance_index_for_fetch(
429 struct si_shader_context
*ctx
,
430 unsigned param_start_instance
, LLVMValueRef divisor
)
432 LLVMValueRef result
= ctx
->abi
.instance_id
;
434 /* The division must be done before START_INSTANCE is added. */
435 if (divisor
!= ctx
->i32_1
)
436 result
= LLVMBuildUDiv(ctx
->ac
.builder
, result
, divisor
, "");
438 return LLVMBuildAdd(ctx
->ac
.builder
, result
,
439 LLVMGetParam(ctx
->main_fn
, param_start_instance
), "");
442 /* Bitcast <4 x float> to <2 x double>, extract the component, and convert
444 static LLVMValueRef
extract_double_to_float(struct si_shader_context
*ctx
,
446 unsigned double_index
)
448 LLVMBuilderRef builder
= ctx
->ac
.builder
;
449 LLVMTypeRef f64
= LLVMDoubleTypeInContext(ctx
->ac
.context
);
450 LLVMValueRef dvec2
= LLVMBuildBitCast(builder
, vec4
,
451 LLVMVectorType(f64
, 2), "");
452 LLVMValueRef index
= LLVMConstInt(ctx
->i32
, double_index
, 0);
453 LLVMValueRef value
= LLVMBuildExtractElement(builder
, dvec2
, index
, "");
454 return LLVMBuildFPTrunc(builder
, value
, ctx
->f32
, "");
457 static LLVMValueRef
unpack_sint16(struct si_shader_context
*ctx
,
458 LLVMValueRef i32
, unsigned index
)
463 return LLVMBuildAShr(ctx
->ac
.builder
, i32
,
464 LLVMConstInt(ctx
->i32
, 16, 0), "");
466 return LLVMBuildSExt(ctx
->ac
.builder
,
467 LLVMBuildTrunc(ctx
->ac
.builder
, i32
,
472 void si_llvm_load_input_vs(
473 struct si_shader_context
*ctx
,
474 unsigned input_index
,
477 unsigned vs_blit_property
=
478 ctx
->shader
->selector
->info
.properties
[TGSI_PROPERTY_VS_BLIT_SGPRS
];
480 if (vs_blit_property
) {
481 LLVMValueRef vertex_id
= ctx
->abi
.vertex_id
;
482 LLVMValueRef sel_x1
= LLVMBuildICmp(ctx
->ac
.builder
,
483 LLVMIntULE
, vertex_id
,
485 /* Use LLVMIntNE, because we have 3 vertices and only
486 * the middle one should use y2.
488 LLVMValueRef sel_y1
= LLVMBuildICmp(ctx
->ac
.builder
,
489 LLVMIntNE
, vertex_id
,
492 if (input_index
== 0) {
494 LLVMValueRef x1y1
= LLVMGetParam(ctx
->main_fn
,
495 ctx
->param_vs_blit_inputs
);
496 LLVMValueRef x2y2
= LLVMGetParam(ctx
->main_fn
,
497 ctx
->param_vs_blit_inputs
+ 1);
499 LLVMValueRef x1
= unpack_sint16(ctx
, x1y1
, 0);
500 LLVMValueRef y1
= unpack_sint16(ctx
, x1y1
, 1);
501 LLVMValueRef x2
= unpack_sint16(ctx
, x2y2
, 0);
502 LLVMValueRef y2
= unpack_sint16(ctx
, x2y2
, 1);
504 LLVMValueRef x
= LLVMBuildSelect(ctx
->ac
.builder
, sel_x1
,
506 LLVMValueRef y
= LLVMBuildSelect(ctx
->ac
.builder
, sel_y1
,
509 out
[0] = LLVMBuildSIToFP(ctx
->ac
.builder
, x
, ctx
->f32
, "");
510 out
[1] = LLVMBuildSIToFP(ctx
->ac
.builder
, y
, ctx
->f32
, "");
511 out
[2] = LLVMGetParam(ctx
->main_fn
,
512 ctx
->param_vs_blit_inputs
+ 2);
513 out
[3] = ctx
->ac
.f32_1
;
517 /* Color or texture coordinates: */
518 assert(input_index
== 1);
520 if (vs_blit_property
== SI_VS_BLIT_SGPRS_POS_COLOR
) {
521 for (int i
= 0; i
< 4; i
++) {
522 out
[i
] = LLVMGetParam(ctx
->main_fn
,
523 ctx
->param_vs_blit_inputs
+ 3 + i
);
526 assert(vs_blit_property
== SI_VS_BLIT_SGPRS_POS_TEXCOORD
);
527 LLVMValueRef x1
= LLVMGetParam(ctx
->main_fn
,
528 ctx
->param_vs_blit_inputs
+ 3);
529 LLVMValueRef y1
= LLVMGetParam(ctx
->main_fn
,
530 ctx
->param_vs_blit_inputs
+ 4);
531 LLVMValueRef x2
= LLVMGetParam(ctx
->main_fn
,
532 ctx
->param_vs_blit_inputs
+ 5);
533 LLVMValueRef y2
= LLVMGetParam(ctx
->main_fn
,
534 ctx
->param_vs_blit_inputs
+ 6);
536 out
[0] = LLVMBuildSelect(ctx
->ac
.builder
, sel_x1
,
538 out
[1] = LLVMBuildSelect(ctx
->ac
.builder
, sel_y1
,
540 out
[2] = LLVMGetParam(ctx
->main_fn
,
541 ctx
->param_vs_blit_inputs
+ 7);
542 out
[3] = LLVMGetParam(ctx
->main_fn
,
543 ctx
->param_vs_blit_inputs
+ 8);
550 unsigned num_fetches
;
551 unsigned fetch_stride
;
553 LLVMValueRef t_list_ptr
;
554 LLVMValueRef t_offset
;
556 LLVMValueRef vertex_index
;
557 LLVMValueRef input
[3];
559 /* Load the T list */
560 t_list_ptr
= LLVMGetParam(ctx
->main_fn
, ctx
->param_vertex_buffers
);
562 t_offset
= LLVMConstInt(ctx
->i32
, input_index
, 0);
564 t_list
= ac_build_indexed_load_const(&ctx
->ac
, t_list_ptr
, t_offset
);
566 vertex_index
= LLVMGetParam(ctx
->main_fn
,
567 ctx
->param_vertex_index0
+
570 fix_fetch
= ctx
->shader
->key
.mono
.vs_fix_fetch
[input_index
];
572 /* Do multiple loads for special formats. */
574 case SI_FIX_FETCH_RGB_64_FLOAT
:
575 num_fetches
= 3; /* 3 2-dword loads */
578 case SI_FIX_FETCH_RGBA_64_FLOAT
:
579 num_fetches
= 2; /* 2 4-dword loads */
582 case SI_FIX_FETCH_RGB_8
:
583 case SI_FIX_FETCH_RGB_8_INT
:
587 case SI_FIX_FETCH_RGB_16
:
588 case SI_FIX_FETCH_RGB_16_INT
:
597 for (unsigned i
= 0; i
< num_fetches
; i
++) {
598 LLVMValueRef voffset
= LLVMConstInt(ctx
->i32
, fetch_stride
* i
, 0);
600 input
[i
] = ac_build_buffer_load_format(&ctx
->ac
, t_list
,
601 vertex_index
, voffset
,
605 /* Break up the vec4 into individual components */
606 for (chan
= 0; chan
< 4; chan
++) {
607 LLVMValueRef llvm_chan
= LLVMConstInt(ctx
->i32
, chan
, 0);
608 out
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
,
609 input
[0], llvm_chan
, "");
613 case SI_FIX_FETCH_A2_SNORM
:
614 case SI_FIX_FETCH_A2_SSCALED
:
615 case SI_FIX_FETCH_A2_SINT
: {
616 /* The hardware returns an unsigned value; convert it to a
619 LLVMValueRef tmp
= out
[3];
620 LLVMValueRef c30
= LLVMConstInt(ctx
->i32
, 30, 0);
622 /* First, recover the sign-extended signed integer value. */
623 if (fix_fetch
== SI_FIX_FETCH_A2_SSCALED
)
624 tmp
= LLVMBuildFPToUI(ctx
->ac
.builder
, tmp
, ctx
->i32
, "");
626 tmp
= ac_to_integer(&ctx
->ac
, tmp
);
628 /* For the integer-like cases, do a natural sign extension.
630 * For the SNORM case, the values are 0.0, 0.333, 0.666, 1.0
631 * and happen to contain 0, 1, 2, 3 as the two LSBs of the
634 tmp
= LLVMBuildShl(ctx
->ac
.builder
, tmp
,
635 fix_fetch
== SI_FIX_FETCH_A2_SNORM
?
636 LLVMConstInt(ctx
->i32
, 7, 0) : c30
, "");
637 tmp
= LLVMBuildAShr(ctx
->ac
.builder
, tmp
, c30
, "");
639 /* Convert back to the right type. */
640 if (fix_fetch
== SI_FIX_FETCH_A2_SNORM
) {
642 LLVMValueRef neg_one
= LLVMConstReal(ctx
->f32
, -1.0);
643 tmp
= LLVMBuildSIToFP(ctx
->ac
.builder
, tmp
, ctx
->f32
, "");
644 clamp
= LLVMBuildFCmp(ctx
->ac
.builder
, LLVMRealULT
, tmp
, neg_one
, "");
645 tmp
= LLVMBuildSelect(ctx
->ac
.builder
, clamp
, neg_one
, tmp
, "");
646 } else if (fix_fetch
== SI_FIX_FETCH_A2_SSCALED
) {
647 tmp
= LLVMBuildSIToFP(ctx
->ac
.builder
, tmp
, ctx
->f32
, "");
653 case SI_FIX_FETCH_RGBA_32_UNORM
:
654 case SI_FIX_FETCH_RGBX_32_UNORM
:
655 for (chan
= 0; chan
< 4; chan
++) {
656 out
[chan
] = ac_to_integer(&ctx
->ac
, out
[chan
]);
657 out
[chan
] = LLVMBuildUIToFP(ctx
->ac
.builder
,
658 out
[chan
], ctx
->f32
, "");
659 out
[chan
] = LLVMBuildFMul(ctx
->ac
.builder
, out
[chan
],
660 LLVMConstReal(ctx
->f32
, 1.0 / UINT_MAX
), "");
662 /* RGBX UINT returns 1 in alpha, which would be rounded to 0 by normalizing. */
663 if (fix_fetch
== SI_FIX_FETCH_RGBX_32_UNORM
)
664 out
[3] = LLVMConstReal(ctx
->f32
, 1);
666 case SI_FIX_FETCH_RGBA_32_SNORM
:
667 case SI_FIX_FETCH_RGBX_32_SNORM
:
668 case SI_FIX_FETCH_RGBA_32_FIXED
:
669 case SI_FIX_FETCH_RGBX_32_FIXED
: {
671 if (fix_fetch
>= SI_FIX_FETCH_RGBA_32_FIXED
)
672 scale
= 1.0 / 0x10000;
674 scale
= 1.0 / INT_MAX
;
676 for (chan
= 0; chan
< 4; chan
++) {
677 out
[chan
] = ac_to_integer(&ctx
->ac
, out
[chan
]);
678 out
[chan
] = LLVMBuildSIToFP(ctx
->ac
.builder
,
679 out
[chan
], ctx
->f32
, "");
680 out
[chan
] = LLVMBuildFMul(ctx
->ac
.builder
, out
[chan
],
681 LLVMConstReal(ctx
->f32
, scale
), "");
683 /* RGBX SINT returns 1 in alpha, which would be rounded to 0 by normalizing. */
684 if (fix_fetch
== SI_FIX_FETCH_RGBX_32_SNORM
||
685 fix_fetch
== SI_FIX_FETCH_RGBX_32_FIXED
)
686 out
[3] = LLVMConstReal(ctx
->f32
, 1);
689 case SI_FIX_FETCH_RGBA_32_USCALED
:
690 for (chan
= 0; chan
< 4; chan
++) {
691 out
[chan
] = ac_to_integer(&ctx
->ac
, out
[chan
]);
692 out
[chan
] = LLVMBuildUIToFP(ctx
->ac
.builder
,
693 out
[chan
], ctx
->f32
, "");
696 case SI_FIX_FETCH_RGBA_32_SSCALED
:
697 for (chan
= 0; chan
< 4; chan
++) {
698 out
[chan
] = ac_to_integer(&ctx
->ac
, out
[chan
]);
699 out
[chan
] = LLVMBuildSIToFP(ctx
->ac
.builder
,
700 out
[chan
], ctx
->f32
, "");
703 case SI_FIX_FETCH_RG_64_FLOAT
:
704 for (chan
= 0; chan
< 2; chan
++)
705 out
[chan
] = extract_double_to_float(ctx
, input
[0], chan
);
707 out
[2] = LLVMConstReal(ctx
->f32
, 0);
708 out
[3] = LLVMConstReal(ctx
->f32
, 1);
710 case SI_FIX_FETCH_RGB_64_FLOAT
:
711 for (chan
= 0; chan
< 3; chan
++)
712 out
[chan
] = extract_double_to_float(ctx
, input
[chan
], 0);
714 out
[3] = LLVMConstReal(ctx
->f32
, 1);
716 case SI_FIX_FETCH_RGBA_64_FLOAT
:
717 for (chan
= 0; chan
< 4; chan
++) {
718 out
[chan
] = extract_double_to_float(ctx
, input
[chan
/ 2],
722 case SI_FIX_FETCH_RGB_8
:
723 case SI_FIX_FETCH_RGB_8_INT
:
724 case SI_FIX_FETCH_RGB_16
:
725 case SI_FIX_FETCH_RGB_16_INT
:
726 for (chan
= 0; chan
< 3; chan
++) {
727 out
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
,
731 if (fix_fetch
== SI_FIX_FETCH_RGB_8
||
732 fix_fetch
== SI_FIX_FETCH_RGB_16
) {
733 out
[3] = LLVMConstReal(ctx
->f32
, 1);
735 out
[3] = ac_to_float(&ctx
->ac
, ctx
->i32_1
);
741 static void declare_input_vs(
742 struct si_shader_context
*ctx
,
743 unsigned input_index
,
744 const struct tgsi_full_declaration
*decl
,
747 si_llvm_load_input_vs(ctx
, input_index
, out
);
750 static LLVMValueRef
get_primitive_id(struct si_shader_context
*ctx
,
757 case PIPE_SHADER_VERTEX
:
758 return LLVMGetParam(ctx
->main_fn
,
759 ctx
->param_vs_prim_id
);
760 case PIPE_SHADER_TESS_CTRL
:
761 return LLVMGetParam(ctx
->main_fn
,
762 ctx
->param_tcs_patch_id
);
763 case PIPE_SHADER_TESS_EVAL
:
764 return LLVMGetParam(ctx
->main_fn
,
765 ctx
->param_tes_patch_id
);
766 case PIPE_SHADER_GEOMETRY
:
767 return LLVMGetParam(ctx
->main_fn
,
768 ctx
->param_gs_prim_id
);
776 * Return the value of tgsi_ind_register for indexing.
777 * This is the indirect index with the constant offset added to it.
779 LLVMValueRef
si_get_indirect_index(struct si_shader_context
*ctx
,
780 const struct tgsi_ind_register
*ind
,
786 if (ind
->File
== TGSI_FILE_ADDRESS
) {
787 result
= ctx
->addrs
[ind
->Index
][ind
->Swizzle
];
788 result
= LLVMBuildLoad(ctx
->ac
.builder
, result
, "");
790 struct tgsi_full_src_register src
= {};
792 src
.Register
.File
= ind
->File
;
793 src
.Register
.Index
= ind
->Index
;
795 /* Set the second index to 0 for constants. */
796 if (ind
->File
== TGSI_FILE_CONSTANT
)
797 src
.Register
.Dimension
= 1;
799 result
= ctx
->bld_base
.emit_fetch_funcs
[ind
->File
](&ctx
->bld_base
, &src
,
802 result
= ac_to_integer(&ctx
->ac
, result
);
806 result
= LLVMBuildMul(ctx
->ac
.builder
, result
,
807 LLVMConstInt(ctx
->i32
, addr_mul
, 0), "");
808 result
= LLVMBuildAdd(ctx
->ac
.builder
, result
,
809 LLVMConstInt(ctx
->i32
, rel_index
, 0), "");
814 * Like si_get_indirect_index, but restricts the return value to a (possibly
815 * undefined) value inside [0..num).
817 LLVMValueRef
si_get_bounded_indirect_index(struct si_shader_context
*ctx
,
818 const struct tgsi_ind_register
*ind
,
819 int rel_index
, unsigned num
)
821 LLVMValueRef result
= si_get_indirect_index(ctx
, ind
, 1, rel_index
);
823 return si_llvm_bound_index(ctx
, result
, num
);
828 * Calculate a dword address given an input or output register and a stride.
830 static LLVMValueRef
get_dw_address(struct si_shader_context
*ctx
,
831 const struct tgsi_full_dst_register
*dst
,
832 const struct tgsi_full_src_register
*src
,
833 LLVMValueRef vertex_dw_stride
,
834 LLVMValueRef base_addr
)
836 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
837 ubyte
*name
, *index
, *array_first
;
839 struct tgsi_full_dst_register reg
;
841 /* Set the register description. The address computation is the same
842 * for sources and destinations. */
844 reg
.Register
.File
= src
->Register
.File
;
845 reg
.Register
.Index
= src
->Register
.Index
;
846 reg
.Register
.Indirect
= src
->Register
.Indirect
;
847 reg
.Register
.Dimension
= src
->Register
.Dimension
;
848 reg
.Indirect
= src
->Indirect
;
849 reg
.Dimension
= src
->Dimension
;
850 reg
.DimIndirect
= src
->DimIndirect
;
854 /* If the register is 2-dimensional (e.g. an array of vertices
855 * in a primitive), calculate the base address of the vertex. */
856 if (reg
.Register
.Dimension
) {
859 if (reg
.Dimension
.Indirect
)
860 index
= si_get_indirect_index(ctx
, ®
.DimIndirect
,
861 1, reg
.Dimension
.Index
);
863 index
= LLVMConstInt(ctx
->i32
, reg
.Dimension
.Index
, 0);
865 base_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_addr
,
866 LLVMBuildMul(ctx
->ac
.builder
, index
,
867 vertex_dw_stride
, ""), "");
870 /* Get information about the register. */
871 if (reg
.Register
.File
== TGSI_FILE_INPUT
) {
872 name
= info
->input_semantic_name
;
873 index
= info
->input_semantic_index
;
874 array_first
= info
->input_array_first
;
875 } else if (reg
.Register
.File
== TGSI_FILE_OUTPUT
) {
876 name
= info
->output_semantic_name
;
877 index
= info
->output_semantic_index
;
878 array_first
= info
->output_array_first
;
884 if (reg
.Register
.Indirect
) {
885 /* Add the relative address of the element. */
886 LLVMValueRef ind_index
;
888 if (reg
.Indirect
.ArrayID
)
889 first
= array_first
[reg
.Indirect
.ArrayID
];
891 first
= reg
.Register
.Index
;
893 ind_index
= si_get_indirect_index(ctx
, ®
.Indirect
,
894 1, reg
.Register
.Index
- first
);
896 base_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_addr
,
897 LLVMBuildMul(ctx
->ac
.builder
, ind_index
,
898 LLVMConstInt(ctx
->i32
, 4, 0), ""), "");
900 param
= reg
.Register
.Dimension
?
901 si_shader_io_get_unique_index(name
[first
], index
[first
]) :
902 si_shader_io_get_unique_index_patch(name
[first
], index
[first
]);
904 param
= reg
.Register
.Dimension
?
905 si_shader_io_get_unique_index(name
[reg
.Register
.Index
],
906 index
[reg
.Register
.Index
]) :
907 si_shader_io_get_unique_index_patch(name
[reg
.Register
.Index
],
908 index
[reg
.Register
.Index
]);
911 /* Add the base address of the element. */
912 return LLVMBuildAdd(ctx
->ac
.builder
, base_addr
,
913 LLVMConstInt(ctx
->i32
, param
* 4, 0), "");
916 /* The offchip buffer layout for TCS->TES is
918 * - attribute 0 of patch 0 vertex 0
919 * - attribute 0 of patch 0 vertex 1
920 * - attribute 0 of patch 0 vertex 2
922 * - attribute 0 of patch 1 vertex 0
923 * - attribute 0 of patch 1 vertex 1
925 * - attribute 1 of patch 0 vertex 0
926 * - attribute 1 of patch 0 vertex 1
928 * - per patch attribute 0 of patch 0
929 * - per patch attribute 0 of patch 1
932 * Note that every attribute has 4 components.
934 static LLVMValueRef
get_tcs_tes_buffer_address(struct si_shader_context
*ctx
,
935 LLVMValueRef rel_patch_id
,
936 LLVMValueRef vertex_index
,
937 LLVMValueRef param_index
)
939 LLVMValueRef base_addr
, vertices_per_patch
, num_patches
, total_vertices
;
940 LLVMValueRef param_stride
, constant16
;
942 vertices_per_patch
= get_num_tcs_out_vertices(ctx
);
943 num_patches
= unpack_param(ctx
, ctx
->param_tcs_offchip_layout
, 0, 6);
944 total_vertices
= LLVMBuildMul(ctx
->ac
.builder
, vertices_per_patch
,
947 constant16
= LLVMConstInt(ctx
->i32
, 16, 0);
949 base_addr
= LLVMBuildMul(ctx
->ac
.builder
, rel_patch_id
,
950 vertices_per_patch
, "");
952 base_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_addr
,
955 param_stride
= total_vertices
;
957 base_addr
= rel_patch_id
;
958 param_stride
= num_patches
;
961 base_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_addr
,
962 LLVMBuildMul(ctx
->ac
.builder
, param_index
,
963 param_stride
, ""), "");
965 base_addr
= LLVMBuildMul(ctx
->ac
.builder
, base_addr
, constant16
, "");
968 LLVMValueRef patch_data_offset
=
969 unpack_param(ctx
, ctx
->param_tcs_offchip_layout
, 12, 20);
971 base_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_addr
,
972 patch_data_offset
, "");
977 static LLVMValueRef
get_tcs_tes_buffer_address_from_reg(
978 struct si_shader_context
*ctx
,
979 const struct tgsi_full_dst_register
*dst
,
980 const struct tgsi_full_src_register
*src
)
982 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
983 ubyte
*name
, *index
, *array_first
;
984 struct tgsi_full_src_register reg
;
985 LLVMValueRef vertex_index
= NULL
;
986 LLVMValueRef param_index
= NULL
;
987 unsigned param_index_base
, param_base
;
989 reg
= src
? *src
: tgsi_full_src_register_from_dst(dst
);
991 if (reg
.Register
.Dimension
) {
993 if (reg
.Dimension
.Indirect
)
994 vertex_index
= si_get_indirect_index(ctx
, ®
.DimIndirect
,
995 1, reg
.Dimension
.Index
);
997 vertex_index
= LLVMConstInt(ctx
->i32
, reg
.Dimension
.Index
, 0);
1000 /* Get information about the register. */
1001 if (reg
.Register
.File
== TGSI_FILE_INPUT
) {
1002 name
= info
->input_semantic_name
;
1003 index
= info
->input_semantic_index
;
1004 array_first
= info
->input_array_first
;
1005 } else if (reg
.Register
.File
== TGSI_FILE_OUTPUT
) {
1006 name
= info
->output_semantic_name
;
1007 index
= info
->output_semantic_index
;
1008 array_first
= info
->output_array_first
;
1014 if (reg
.Register
.Indirect
) {
1015 if (reg
.Indirect
.ArrayID
)
1016 param_base
= array_first
[reg
.Indirect
.ArrayID
];
1018 param_base
= reg
.Register
.Index
;
1020 param_index
= si_get_indirect_index(ctx
, ®
.Indirect
,
1021 1, reg
.Register
.Index
- param_base
);
1024 param_base
= reg
.Register
.Index
;
1025 param_index
= ctx
->i32_0
;
1028 param_index_base
= reg
.Register
.Dimension
?
1029 si_shader_io_get_unique_index(name
[param_base
], index
[param_base
]) :
1030 si_shader_io_get_unique_index_patch(name
[param_base
], index
[param_base
]);
1032 param_index
= LLVMBuildAdd(ctx
->ac
.builder
, param_index
,
1033 LLVMConstInt(ctx
->i32
, param_index_base
, 0),
1036 return get_tcs_tes_buffer_address(ctx
, get_rel_patch_id(ctx
),
1037 vertex_index
, param_index
);
1040 static LLVMValueRef
buffer_load(struct lp_build_tgsi_context
*bld_base
,
1041 enum tgsi_opcode_type type
, unsigned swizzle
,
1042 LLVMValueRef buffer
, LLVMValueRef offset
,
1043 LLVMValueRef base
, bool can_speculate
)
1045 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1046 LLVMValueRef value
, value2
;
1047 LLVMTypeRef llvm_type
= tgsi2llvmtype(bld_base
, type
);
1048 LLVMTypeRef vec_type
= LLVMVectorType(llvm_type
, 4);
1050 if (swizzle
== ~0) {
1051 value
= ac_build_buffer_load(&ctx
->ac
, buffer
, 4, NULL
, base
, offset
,
1052 0, 1, 0, can_speculate
, false);
1054 return LLVMBuildBitCast(ctx
->ac
.builder
, value
, vec_type
, "");
1057 if (!tgsi_type_is_64bit(type
)) {
1058 value
= ac_build_buffer_load(&ctx
->ac
, buffer
, 4, NULL
, base
, offset
,
1059 0, 1, 0, can_speculate
, false);
1061 value
= LLVMBuildBitCast(ctx
->ac
.builder
, value
, vec_type
, "");
1062 return LLVMBuildExtractElement(ctx
->ac
.builder
, value
,
1063 LLVMConstInt(ctx
->i32
, swizzle
, 0), "");
1066 value
= ac_build_buffer_load(&ctx
->ac
, buffer
, 1, NULL
, base
, offset
,
1067 swizzle
* 4, 1, 0, can_speculate
, false);
1069 value2
= ac_build_buffer_load(&ctx
->ac
, buffer
, 1, NULL
, base
, offset
,
1070 swizzle
* 4 + 4, 1, 0, can_speculate
, false);
1072 return si_llvm_emit_fetch_64bit(bld_base
, type
, value
, value2
);
1078 * \param type output value type
1079 * \param swizzle offset (typically 0..3); it can be ~0, which loads a vec4
1080 * \param dw_addr address in dwords
1082 static LLVMValueRef
lds_load(struct lp_build_tgsi_context
*bld_base
,
1083 enum tgsi_opcode_type type
, unsigned swizzle
,
1084 LLVMValueRef dw_addr
)
1086 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1089 if (swizzle
== ~0) {
1090 LLVMValueRef values
[TGSI_NUM_CHANNELS
];
1092 for (unsigned chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++)
1093 values
[chan
] = lds_load(bld_base
, type
, chan
, dw_addr
);
1095 return lp_build_gather_values(&ctx
->gallivm
, values
,
1099 dw_addr
= lp_build_add(&bld_base
->uint_bld
, dw_addr
,
1100 LLVMConstInt(ctx
->i32
, swizzle
, 0));
1102 value
= ac_build_indexed_load(&ctx
->ac
, ctx
->lds
, dw_addr
, false);
1103 if (tgsi_type_is_64bit(type
)) {
1104 LLVMValueRef value2
;
1105 dw_addr
= lp_build_add(&bld_base
->uint_bld
, dw_addr
,
1107 value2
= ac_build_indexed_load(&ctx
->ac
, ctx
->lds
, dw_addr
, false);
1108 return si_llvm_emit_fetch_64bit(bld_base
, type
, value
, value2
);
1111 return bitcast(bld_base
, type
, value
);
1117 * \param swizzle offset (typically 0..3)
1118 * \param dw_addr address in dwords
1119 * \param value value to store
1121 static void lds_store(struct lp_build_tgsi_context
*bld_base
,
1122 unsigned dw_offset_imm
, LLVMValueRef dw_addr
,
1125 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1127 dw_addr
= lp_build_add(&bld_base
->uint_bld
, dw_addr
,
1128 LLVMConstInt(ctx
->i32
, dw_offset_imm
, 0));
1130 value
= ac_to_integer(&ctx
->ac
, value
);
1131 ac_build_indexed_store(&ctx
->ac
, ctx
->lds
,
1135 static LLVMValueRef
desc_from_addr_base64k(struct si_shader_context
*ctx
,
1138 LLVMBuilderRef builder
= ctx
->ac
.builder
;
1140 LLVMValueRef addr
= LLVMGetParam(ctx
->main_fn
, param
);
1141 addr
= LLVMBuildZExt(builder
, addr
, ctx
->i64
, "");
1142 addr
= LLVMBuildShl(builder
, addr
, LLVMConstInt(ctx
->i64
, 16, 0), "");
1144 uint64_t desc2
= 0xffffffff;
1145 uint64_t desc3
= S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1146 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1147 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1148 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1149 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1150 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
);
1151 LLVMValueRef hi
= LLVMConstInt(ctx
->i64
, desc2
| (desc3
<< 32), 0);
1153 LLVMValueRef desc
= LLVMGetUndef(LLVMVectorType(ctx
->i64
, 2));
1154 desc
= LLVMBuildInsertElement(builder
, desc
, addr
, ctx
->i32_0
, "");
1155 desc
= LLVMBuildInsertElement(builder
, desc
, hi
, ctx
->i32_1
, "");
1156 return LLVMBuildBitCast(builder
, desc
, ctx
->v4i32
, "");
1159 static LLVMValueRef
fetch_input_tcs(
1160 struct lp_build_tgsi_context
*bld_base
,
1161 const struct tgsi_full_src_register
*reg
,
1162 enum tgsi_opcode_type type
, unsigned swizzle
)
1164 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1165 LLVMValueRef dw_addr
, stride
;
1167 stride
= get_tcs_in_vertex_dw_stride(ctx
);
1168 dw_addr
= get_tcs_in_current_patch_offset(ctx
);
1169 dw_addr
= get_dw_address(ctx
, NULL
, reg
, stride
, dw_addr
);
1171 return lds_load(bld_base
, type
, swizzle
, dw_addr
);
1174 static LLVMValueRef
fetch_output_tcs(
1175 struct lp_build_tgsi_context
*bld_base
,
1176 const struct tgsi_full_src_register
*reg
,
1177 enum tgsi_opcode_type type
, unsigned swizzle
)
1179 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1180 LLVMValueRef dw_addr
, stride
;
1182 if (reg
->Register
.Dimension
) {
1183 stride
= get_tcs_out_vertex_dw_stride(ctx
);
1184 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
1185 dw_addr
= get_dw_address(ctx
, NULL
, reg
, stride
, dw_addr
);
1187 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
1188 dw_addr
= get_dw_address(ctx
, NULL
, reg
, NULL
, dw_addr
);
1191 return lds_load(bld_base
, type
, swizzle
, dw_addr
);
1194 static LLVMValueRef
fetch_input_tes(
1195 struct lp_build_tgsi_context
*bld_base
,
1196 const struct tgsi_full_src_register
*reg
,
1197 enum tgsi_opcode_type type
, unsigned swizzle
)
1199 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1200 LLVMValueRef buffer
, base
, addr
;
1202 buffer
= desc_from_addr_base64k(ctx
, ctx
->param_tcs_offchip_addr_base64k
);
1204 base
= LLVMGetParam(ctx
->main_fn
, ctx
->param_tcs_offchip_offset
);
1205 addr
= get_tcs_tes_buffer_address_from_reg(ctx
, NULL
, reg
);
1207 return buffer_load(bld_base
, type
, swizzle
, buffer
, base
, addr
, true);
1210 static void store_output_tcs(struct lp_build_tgsi_context
*bld_base
,
1211 const struct tgsi_full_instruction
*inst
,
1212 const struct tgsi_opcode_info
*info
,
1214 LLVMValueRef dst
[4])
1216 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1217 const struct tgsi_full_dst_register
*reg
= &inst
->Dst
[index
];
1218 const struct tgsi_shader_info
*sh_info
= &ctx
->shader
->selector
->info
;
1219 unsigned chan_index
;
1220 LLVMValueRef dw_addr
, stride
;
1221 LLVMValueRef buffer
, base
, buf_addr
;
1222 LLVMValueRef values
[4];
1223 bool skip_lds_store
;
1224 bool is_tess_factor
= false, is_tess_inner
= false;
1226 /* Only handle per-patch and per-vertex outputs here.
1227 * Vectors will be lowered to scalars and this function will be called again.
1229 if (reg
->Register
.File
!= TGSI_FILE_OUTPUT
||
1230 (dst
[0] && LLVMGetTypeKind(LLVMTypeOf(dst
[0])) == LLVMVectorTypeKind
)) {
1231 si_llvm_emit_store(bld_base
, inst
, info
, index
, dst
);
1235 if (reg
->Register
.Dimension
) {
1236 stride
= get_tcs_out_vertex_dw_stride(ctx
);
1237 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
1238 dw_addr
= get_dw_address(ctx
, reg
, NULL
, stride
, dw_addr
);
1239 skip_lds_store
= !sh_info
->reads_pervertex_outputs
;
1241 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
1242 dw_addr
= get_dw_address(ctx
, reg
, NULL
, NULL
, dw_addr
);
1243 skip_lds_store
= !sh_info
->reads_perpatch_outputs
;
1245 if (!reg
->Register
.Indirect
) {
1246 int name
= sh_info
->output_semantic_name
[reg
->Register
.Index
];
1248 /* Always write tess factors into LDS for the TCS epilog. */
1249 if (name
== TGSI_SEMANTIC_TESSINNER
||
1250 name
== TGSI_SEMANTIC_TESSOUTER
) {
1251 /* The epilog doesn't read LDS if invocation 0 defines tess factors. */
1252 skip_lds_store
= !sh_info
->reads_tessfactor_outputs
&&
1253 ctx
->shader
->selector
->tcs_info
.tessfactors_are_def_in_all_invocs
;
1254 is_tess_factor
= true;
1255 is_tess_inner
= name
== TGSI_SEMANTIC_TESSINNER
;
1260 buffer
= desc_from_addr_base64k(ctx
, ctx
->param_tcs_offchip_addr_base64k
);
1262 base
= LLVMGetParam(ctx
->main_fn
, ctx
->param_tcs_offchip_offset
);
1263 buf_addr
= get_tcs_tes_buffer_address_from_reg(ctx
, reg
, NULL
);
1265 uint32_t writemask
= reg
->Register
.WriteMask
;
1267 chan_index
= u_bit_scan(&writemask
);
1268 LLVMValueRef value
= dst
[chan_index
];
1270 if (inst
->Instruction
.Saturate
)
1271 value
= ac_build_clamp(&ctx
->ac
, value
);
1273 /* Skip LDS stores if there is no LDS read of this output. */
1274 if (!skip_lds_store
)
1275 lds_store(bld_base
, chan_index
, dw_addr
, value
);
1277 value
= ac_to_integer(&ctx
->ac
, value
);
1278 values
[chan_index
] = value
;
1280 if (reg
->Register
.WriteMask
!= 0xF && !is_tess_factor
) {
1281 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, value
, 1,
1283 4 * chan_index
, 1, 0, true, false);
1286 /* Write tess factors into VGPRs for the epilog. */
1287 if (is_tess_factor
&&
1288 ctx
->shader
->selector
->tcs_info
.tessfactors_are_def_in_all_invocs
) {
1289 if (!is_tess_inner
) {
1290 LLVMBuildStore(ctx
->ac
.builder
, value
, /* outer */
1291 ctx
->invoc0_tess_factors
[chan_index
]);
1292 } else if (chan_index
< 2) {
1293 LLVMBuildStore(ctx
->ac
.builder
, value
, /* inner */
1294 ctx
->invoc0_tess_factors
[4 + chan_index
]);
1299 if (reg
->Register
.WriteMask
== 0xF && !is_tess_factor
) {
1300 LLVMValueRef value
= lp_build_gather_values(&ctx
->gallivm
,
1302 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, value
, 4, buf_addr
,
1303 base
, 0, 1, 0, true, false);
1307 static LLVMValueRef
fetch_input_gs(
1308 struct lp_build_tgsi_context
*bld_base
,
1309 const struct tgsi_full_src_register
*reg
,
1310 enum tgsi_opcode_type type
,
1313 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1314 struct si_shader
*shader
= ctx
->shader
;
1315 struct lp_build_context
*uint
= &ctx
->bld_base
.uint_bld
;
1316 LLVMValueRef vtx_offset
, soffset
;
1317 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
1318 unsigned semantic_name
= info
->input_semantic_name
[reg
->Register
.Index
];
1319 unsigned semantic_index
= info
->input_semantic_index
[reg
->Register
.Index
];
1323 if (swizzle
!= ~0 && semantic_name
== TGSI_SEMANTIC_PRIMID
)
1324 return get_primitive_id(ctx
, swizzle
);
1326 if (!reg
->Register
.Dimension
)
1329 param
= si_shader_io_get_unique_index(semantic_name
, semantic_index
);
1331 /* GFX9 has the ESGS ring in LDS. */
1332 if (ctx
->screen
->b
.chip_class
>= GFX9
) {
1333 unsigned index
= reg
->Dimension
.Index
;
1335 switch (index
/ 2) {
1337 vtx_offset
= unpack_param(ctx
, ctx
->param_gs_vtx01_offset
,
1338 index
% 2 ? 16 : 0, 16);
1341 vtx_offset
= unpack_param(ctx
, ctx
->param_gs_vtx23_offset
,
1342 index
% 2 ? 16 : 0, 16);
1345 vtx_offset
= unpack_param(ctx
, ctx
->param_gs_vtx45_offset
,
1346 index
% 2 ? 16 : 0, 16);
1353 vtx_offset
= LLVMBuildAdd(ctx
->ac
.builder
, vtx_offset
,
1354 LLVMConstInt(ctx
->i32
, param
* 4, 0), "");
1355 return lds_load(bld_base
, type
, swizzle
, vtx_offset
);
1358 /* GFX6: input load from the ESGS ring in memory. */
1359 if (swizzle
== ~0) {
1360 LLVMValueRef values
[TGSI_NUM_CHANNELS
];
1362 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
1363 values
[chan
] = fetch_input_gs(bld_base
, reg
, type
, chan
);
1365 return lp_build_gather_values(&ctx
->gallivm
, values
,
1369 /* Get the vertex offset parameter on GFX6. */
1370 unsigned vtx_offset_param
= reg
->Dimension
.Index
;
1371 if (vtx_offset_param
< 2) {
1372 vtx_offset_param
+= ctx
->param_gs_vtx0_offset
;
1374 assert(vtx_offset_param
< 6);
1375 vtx_offset_param
+= ctx
->param_gs_vtx2_offset
- 2;
1377 vtx_offset
= lp_build_mul_imm(uint
,
1378 LLVMGetParam(ctx
->main_fn
,
1382 soffset
= LLVMConstInt(ctx
->i32
, (param
* 4 + swizzle
) * 256, 0);
1384 value
= ac_build_buffer_load(&ctx
->ac
, ctx
->esgs_ring
, 1, ctx
->i32_0
,
1385 vtx_offset
, soffset
, 0, 1, 0, true, false);
1386 if (tgsi_type_is_64bit(type
)) {
1387 LLVMValueRef value2
;
1388 soffset
= LLVMConstInt(ctx
->i32
, (param
* 4 + swizzle
+ 1) * 256, 0);
1390 value2
= ac_build_buffer_load(&ctx
->ac
, ctx
->esgs_ring
, 1,
1391 ctx
->i32_0
, vtx_offset
, soffset
,
1392 0, 1, 0, true, false);
1393 return si_llvm_emit_fetch_64bit(bld_base
, type
,
1396 return bitcast(bld_base
, type
, value
);
1399 static int lookup_interp_param_index(unsigned interpolate
, unsigned location
)
1401 switch (interpolate
) {
1402 case TGSI_INTERPOLATE_CONSTANT
:
1405 case TGSI_INTERPOLATE_LINEAR
:
1406 if (location
== TGSI_INTERPOLATE_LOC_SAMPLE
)
1407 return SI_PARAM_LINEAR_SAMPLE
;
1408 else if (location
== TGSI_INTERPOLATE_LOC_CENTROID
)
1409 return SI_PARAM_LINEAR_CENTROID
;
1411 return SI_PARAM_LINEAR_CENTER
;
1413 case TGSI_INTERPOLATE_COLOR
:
1414 case TGSI_INTERPOLATE_PERSPECTIVE
:
1415 if (location
== TGSI_INTERPOLATE_LOC_SAMPLE
)
1416 return SI_PARAM_PERSP_SAMPLE
;
1417 else if (location
== TGSI_INTERPOLATE_LOC_CENTROID
)
1418 return SI_PARAM_PERSP_CENTROID
;
1420 return SI_PARAM_PERSP_CENTER
;
1423 fprintf(stderr
, "Warning: Unhandled interpolation mode.\n");
1428 static LLVMValueRef
si_build_fs_interp(struct si_shader_context
*ctx
,
1429 unsigned attr_index
, unsigned chan
,
1430 LLVMValueRef prim_mask
,
1431 LLVMValueRef i
, LLVMValueRef j
)
1434 return ac_build_fs_interp(&ctx
->ac
,
1435 LLVMConstInt(ctx
->i32
, chan
, 0),
1436 LLVMConstInt(ctx
->i32
, attr_index
, 0),
1439 return ac_build_fs_interp_mov(&ctx
->ac
,
1440 LLVMConstInt(ctx
->i32
, 2, 0), /* P0 */
1441 LLVMConstInt(ctx
->i32
, chan
, 0),
1442 LLVMConstInt(ctx
->i32
, attr_index
, 0),
1447 * Interpolate a fragment shader input.
1449 * @param ctx context
1450 * @param input_index index of the input in hardware
1451 * @param semantic_name TGSI_SEMANTIC_*
1452 * @param semantic_index semantic index
1453 * @param num_interp_inputs number of all interpolated inputs (= BCOLOR offset)
1454 * @param colors_read_mask color components read (4 bits for each color, 8 bits in total)
1455 * @param interp_param interpolation weights (i,j)
1456 * @param prim_mask SI_PARAM_PRIM_MASK
1457 * @param face SI_PARAM_FRONT_FACE
1458 * @param result the return value (4 components)
1460 static void interp_fs_input(struct si_shader_context
*ctx
,
1461 unsigned input_index
,
1462 unsigned semantic_name
,
1463 unsigned semantic_index
,
1464 unsigned num_interp_inputs
,
1465 unsigned colors_read_mask
,
1466 LLVMValueRef interp_param
,
1467 LLVMValueRef prim_mask
,
1469 LLVMValueRef result
[4])
1471 LLVMValueRef i
= NULL
, j
= NULL
;
1474 /* fs.constant returns the param from the middle vertex, so it's not
1475 * really useful for flat shading. It's meant to be used for custom
1476 * interpolation (but the intrinsic can't fetch from the other two
1479 * Luckily, it doesn't matter, because we rely on the FLAT_SHADE state
1480 * to do the right thing. The only reason we use fs.constant is that
1481 * fs.interp cannot be used on integers, because they can be equal
1484 * When interp is false we will use fs.constant or for newer llvm,
1485 * amdgcn.interp.mov.
1487 bool interp
= interp_param
!= NULL
;
1490 interp_param
= LLVMBuildBitCast(ctx
->ac
.builder
, interp_param
,
1491 LLVMVectorType(ctx
->f32
, 2), "");
1493 i
= LLVMBuildExtractElement(ctx
->ac
.builder
, interp_param
,
1495 j
= LLVMBuildExtractElement(ctx
->ac
.builder
, interp_param
,
1499 if (semantic_name
== TGSI_SEMANTIC_COLOR
&&
1500 ctx
->shader
->key
.part
.ps
.prolog
.color_two_side
) {
1501 LLVMValueRef is_face_positive
;
1503 /* If BCOLOR0 is used, BCOLOR1 is at offset "num_inputs + 1",
1504 * otherwise it's at offset "num_inputs".
1506 unsigned back_attr_offset
= num_interp_inputs
;
1507 if (semantic_index
== 1 && colors_read_mask
& 0xf)
1508 back_attr_offset
+= 1;
1510 is_face_positive
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntNE
,
1511 face
, ctx
->i32_0
, "");
1513 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
1514 LLVMValueRef front
, back
;
1516 front
= si_build_fs_interp(ctx
,
1519 back
= si_build_fs_interp(ctx
,
1520 back_attr_offset
, chan
,
1523 result
[chan
] = LLVMBuildSelect(ctx
->ac
.builder
,
1529 } else if (semantic_name
== TGSI_SEMANTIC_FOG
) {
1530 result
[0] = si_build_fs_interp(ctx
, input_index
,
1531 0, prim_mask
, i
, j
);
1533 result
[2] = LLVMConstReal(ctx
->f32
, 0.0f
);
1534 result
[3] = LLVMConstReal(ctx
->f32
, 1.0f
);
1536 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
1537 result
[chan
] = si_build_fs_interp(ctx
,
1544 void si_llvm_load_input_fs(
1545 struct si_shader_context
*ctx
,
1546 unsigned input_index
,
1547 LLVMValueRef out
[4])
1549 struct lp_build_context
*base
= &ctx
->bld_base
.base
;
1550 struct si_shader
*shader
= ctx
->shader
;
1551 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
1552 LLVMValueRef main_fn
= ctx
->main_fn
;
1553 LLVMValueRef interp_param
= NULL
;
1554 int interp_param_idx
;
1555 enum tgsi_semantic semantic_name
= info
->input_semantic_name
[input_index
];
1556 unsigned semantic_index
= info
->input_semantic_index
[input_index
];
1557 enum tgsi_interpolate_mode interp_mode
= info
->input_interpolate
[input_index
];
1558 enum tgsi_interpolate_loc interp_loc
= info
->input_interpolate_loc
[input_index
];
1560 /* Get colors from input VGPRs (set by the prolog). */
1561 if (semantic_name
== TGSI_SEMANTIC_COLOR
) {
1562 unsigned colors_read
= shader
->selector
->info
.colors_read
;
1563 unsigned mask
= colors_read
>> (semantic_index
* 4);
1564 unsigned offset
= SI_PARAM_POS_FIXED_PT
+ 1 +
1565 (semantic_index
? util_bitcount(colors_read
& 0xf) : 0);
1567 out
[0] = mask
& 0x1 ? LLVMGetParam(main_fn
, offset
++) : base
->undef
;
1568 out
[1] = mask
& 0x2 ? LLVMGetParam(main_fn
, offset
++) : base
->undef
;
1569 out
[2] = mask
& 0x4 ? LLVMGetParam(main_fn
, offset
++) : base
->undef
;
1570 out
[3] = mask
& 0x8 ? LLVMGetParam(main_fn
, offset
++) : base
->undef
;
1574 interp_param_idx
= lookup_interp_param_index(interp_mode
, interp_loc
);
1575 if (interp_param_idx
== -1)
1577 else if (interp_param_idx
) {
1578 interp_param
= LLVMGetParam(ctx
->main_fn
, interp_param_idx
);
1581 interp_fs_input(ctx
, input_index
, semantic_name
,
1582 semantic_index
, 0, /* this param is unused */
1583 shader
->selector
->info
.colors_read
, interp_param
,
1584 LLVMGetParam(main_fn
, SI_PARAM_PRIM_MASK
),
1585 LLVMGetParam(main_fn
, SI_PARAM_FRONT_FACE
),
1589 static void declare_input_fs(
1590 struct si_shader_context
*ctx
,
1591 unsigned input_index
,
1592 const struct tgsi_full_declaration
*decl
,
1593 LLVMValueRef out
[4])
1595 si_llvm_load_input_fs(ctx
, input_index
, out
);
1598 static LLVMValueRef
get_sample_id(struct si_shader_context
*ctx
)
1600 return unpack_param(ctx
, SI_PARAM_ANCILLARY
, 8, 4);
1605 * Load a dword from a constant buffer.
1607 static LLVMValueRef
buffer_load_const(struct si_shader_context
*ctx
,
1608 LLVMValueRef resource
,
1609 LLVMValueRef offset
)
1611 return ac_build_buffer_load(&ctx
->ac
, resource
, 1, NULL
, offset
, NULL
,
1612 0, 0, 0, true, true);
1615 static LLVMValueRef
load_sample_position(struct si_shader_context
*ctx
, LLVMValueRef sample_id
)
1617 struct lp_build_context
*uint_bld
= &ctx
->bld_base
.uint_bld
;
1618 LLVMValueRef desc
= LLVMGetParam(ctx
->main_fn
, ctx
->param_rw_buffers
);
1619 LLVMValueRef buf_index
= LLVMConstInt(ctx
->i32
, SI_PS_CONST_SAMPLE_POSITIONS
, 0);
1620 LLVMValueRef resource
= ac_build_indexed_load_const(&ctx
->ac
, desc
, buf_index
);
1622 /* offset = sample_id * 8 (8 = 2 floats containing samplepos.xy) */
1623 LLVMValueRef offset0
= lp_build_mul_imm(uint_bld
, sample_id
, 8);
1624 LLVMValueRef offset1
= LLVMBuildAdd(ctx
->ac
.builder
, offset0
, LLVMConstInt(ctx
->i32
, 4, 0), "");
1626 LLVMValueRef pos
[4] = {
1627 buffer_load_const(ctx
, resource
, offset0
),
1628 buffer_load_const(ctx
, resource
, offset1
),
1629 LLVMConstReal(ctx
->f32
, 0),
1630 LLVMConstReal(ctx
->f32
, 0)
1633 return lp_build_gather_values(&ctx
->gallivm
, pos
, 4);
1636 void si_load_system_value(struct si_shader_context
*ctx
,
1638 const struct tgsi_full_declaration
*decl
)
1640 struct lp_build_context
*bld
= &ctx
->bld_base
.base
;
1641 LLVMValueRef value
= 0;
1643 assert(index
< RADEON_LLVM_MAX_SYSTEM_VALUES
);
1645 switch (decl
->Semantic
.Name
) {
1646 case TGSI_SEMANTIC_INSTANCEID
:
1647 value
= ctx
->abi
.instance_id
;
1650 case TGSI_SEMANTIC_VERTEXID
:
1651 value
= LLVMBuildAdd(ctx
->ac
.builder
,
1653 ctx
->abi
.base_vertex
, "");
1656 case TGSI_SEMANTIC_VERTEXID_NOBASE
:
1657 /* Unused. Clarify the meaning in indexed vs. non-indexed
1658 * draws if this is ever used again. */
1662 case TGSI_SEMANTIC_BASEVERTEX
:
1664 /* For non-indexed draws, the base vertex set by the driver
1665 * (for direct draws) or the CP (for indirect draws) is the
1666 * first vertex ID, but GLSL expects 0 to be returned.
1668 LLVMValueRef vs_state
= LLVMGetParam(ctx
->main_fn
, ctx
->param_vs_state_bits
);
1669 LLVMValueRef indexed
;
1671 indexed
= LLVMBuildLShr(ctx
->ac
.builder
, vs_state
, ctx
->i32_1
, "");
1672 indexed
= LLVMBuildTrunc(ctx
->ac
.builder
, indexed
, ctx
->i1
, "");
1674 value
= LLVMBuildSelect(ctx
->ac
.builder
, indexed
,
1675 ctx
->abi
.base_vertex
, ctx
->i32_0
, "");
1679 case TGSI_SEMANTIC_BASEINSTANCE
:
1680 value
= ctx
->abi
.start_instance
;
1683 case TGSI_SEMANTIC_DRAWID
:
1684 value
= ctx
->abi
.draw_id
;
1687 case TGSI_SEMANTIC_INVOCATIONID
:
1688 if (ctx
->type
== PIPE_SHADER_TESS_CTRL
)
1689 value
= unpack_param(ctx
, ctx
->param_tcs_rel_ids
, 8, 5);
1690 else if (ctx
->type
== PIPE_SHADER_GEOMETRY
)
1691 value
= LLVMGetParam(ctx
->main_fn
,
1692 ctx
->param_gs_instance_id
);
1694 assert(!"INVOCATIONID not implemented");
1697 case TGSI_SEMANTIC_POSITION
:
1699 LLVMValueRef pos
[4] = {
1700 LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_X_FLOAT
),
1701 LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_Y_FLOAT
),
1702 LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_Z_FLOAT
),
1703 lp_build_emit_llvm_unary(&ctx
->bld_base
, TGSI_OPCODE_RCP
,
1704 LLVMGetParam(ctx
->main_fn
,
1705 SI_PARAM_POS_W_FLOAT
)),
1707 value
= lp_build_gather_values(&ctx
->gallivm
, pos
, 4);
1711 case TGSI_SEMANTIC_FACE
:
1712 value
= ctx
->abi
.front_face
;
1715 case TGSI_SEMANTIC_SAMPLEID
:
1716 value
= get_sample_id(ctx
);
1719 case TGSI_SEMANTIC_SAMPLEPOS
: {
1720 LLVMValueRef pos
[4] = {
1721 LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_X_FLOAT
),
1722 LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_Y_FLOAT
),
1723 LLVMConstReal(ctx
->f32
, 0),
1724 LLVMConstReal(ctx
->f32
, 0)
1726 pos
[0] = lp_build_emit_llvm_unary(&ctx
->bld_base
,
1727 TGSI_OPCODE_FRC
, pos
[0]);
1728 pos
[1] = lp_build_emit_llvm_unary(&ctx
->bld_base
,
1729 TGSI_OPCODE_FRC
, pos
[1]);
1730 value
= lp_build_gather_values(&ctx
->gallivm
, pos
, 4);
1734 case TGSI_SEMANTIC_SAMPLEMASK
:
1735 /* This can only occur with the OpenGL Core profile, which
1736 * doesn't support smoothing.
1738 value
= LLVMGetParam(ctx
->main_fn
, SI_PARAM_SAMPLE_COVERAGE
);
1741 case TGSI_SEMANTIC_TESSCOORD
:
1743 LLVMValueRef coord
[4] = {
1744 LLVMGetParam(ctx
->main_fn
, ctx
->param_tes_u
),
1745 LLVMGetParam(ctx
->main_fn
, ctx
->param_tes_v
),
1750 /* For triangles, the vector should be (u, v, 1-u-v). */
1751 if (ctx
->shader
->selector
->info
.properties
[TGSI_PROPERTY_TES_PRIM_MODE
] ==
1752 PIPE_PRIM_TRIANGLES
)
1753 coord
[2] = lp_build_sub(bld
, ctx
->ac
.f32_1
,
1754 lp_build_add(bld
, coord
[0], coord
[1]));
1756 value
= lp_build_gather_values(&ctx
->gallivm
, coord
, 4);
1760 case TGSI_SEMANTIC_VERTICESIN
:
1761 if (ctx
->type
== PIPE_SHADER_TESS_CTRL
)
1762 value
= unpack_param(ctx
, ctx
->param_tcs_out_lds_layout
, 26, 6);
1763 else if (ctx
->type
== PIPE_SHADER_TESS_EVAL
)
1764 value
= get_num_tcs_out_vertices(ctx
);
1766 assert(!"invalid shader stage for TGSI_SEMANTIC_VERTICESIN");
1769 case TGSI_SEMANTIC_TESSINNER
:
1770 case TGSI_SEMANTIC_TESSOUTER
:
1772 LLVMValueRef buffer
, base
, addr
;
1773 int param
= si_shader_io_get_unique_index_patch(decl
->Semantic
.Name
, 0);
1775 buffer
= desc_from_addr_base64k(ctx
, ctx
->param_tcs_offchip_addr_base64k
);
1777 base
= LLVMGetParam(ctx
->main_fn
, ctx
->param_tcs_offchip_offset
);
1778 addr
= get_tcs_tes_buffer_address(ctx
, get_rel_patch_id(ctx
), NULL
,
1779 LLVMConstInt(ctx
->i32
, param
, 0));
1781 value
= buffer_load(&ctx
->bld_base
, TGSI_TYPE_FLOAT
,
1782 ~0, buffer
, base
, addr
, true);
1787 case TGSI_SEMANTIC_DEFAULT_TESSOUTER_SI
:
1788 case TGSI_SEMANTIC_DEFAULT_TESSINNER_SI
:
1790 LLVMValueRef buf
, slot
, val
[4];
1793 slot
= LLVMConstInt(ctx
->i32
, SI_HS_CONST_DEFAULT_TESS_LEVELS
, 0);
1794 buf
= LLVMGetParam(ctx
->main_fn
, ctx
->param_rw_buffers
);
1795 buf
= ac_build_indexed_load_const(&ctx
->ac
, buf
, slot
);
1796 offset
= decl
->Semantic
.Name
== TGSI_SEMANTIC_DEFAULT_TESSINNER_SI
? 4 : 0;
1798 for (i
= 0; i
< 4; i
++)
1799 val
[i
] = buffer_load_const(ctx
, buf
,
1800 LLVMConstInt(ctx
->i32
, (offset
+ i
) * 4, 0));
1801 value
= lp_build_gather_values(&ctx
->gallivm
, val
, 4);
1805 case TGSI_SEMANTIC_PRIMID
:
1806 value
= get_primitive_id(ctx
, 0);
1809 case TGSI_SEMANTIC_GRID_SIZE
:
1810 value
= LLVMGetParam(ctx
->main_fn
, ctx
->param_grid_size
);
1813 case TGSI_SEMANTIC_BLOCK_SIZE
:
1815 LLVMValueRef values
[3];
1817 unsigned *properties
= ctx
->shader
->selector
->info
.properties
;
1819 if (properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
] != 0) {
1820 unsigned sizes
[3] = {
1821 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
],
1822 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT
],
1823 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH
]
1826 for (i
= 0; i
< 3; ++i
)
1827 values
[i
] = LLVMConstInt(ctx
->i32
, sizes
[i
], 0);
1829 value
= lp_build_gather_values(&ctx
->gallivm
, values
, 3);
1831 value
= LLVMGetParam(ctx
->main_fn
, ctx
->param_block_size
);
1836 case TGSI_SEMANTIC_BLOCK_ID
:
1838 LLVMValueRef values
[3];
1840 for (int i
= 0; i
< 3; i
++) {
1841 values
[i
] = ctx
->i32_0
;
1842 if (ctx
->param_block_id
[i
] >= 0) {
1843 values
[i
] = LLVMGetParam(ctx
->main_fn
,
1844 ctx
->param_block_id
[i
]);
1847 value
= lp_build_gather_values(&ctx
->gallivm
, values
, 3);
1851 case TGSI_SEMANTIC_THREAD_ID
:
1852 value
= LLVMGetParam(ctx
->main_fn
, ctx
->param_thread_id
);
1855 case TGSI_SEMANTIC_HELPER_INVOCATION
:
1856 value
= lp_build_intrinsic(ctx
->ac
.builder
,
1857 "llvm.amdgcn.ps.live",
1859 LP_FUNC_ATTR_READNONE
);
1860 value
= LLVMBuildNot(ctx
->ac
.builder
, value
, "");
1861 value
= LLVMBuildSExt(ctx
->ac
.builder
, value
, ctx
->i32
, "");
1864 case TGSI_SEMANTIC_SUBGROUP_SIZE
:
1865 value
= LLVMConstInt(ctx
->i32
, 64, 0);
1868 case TGSI_SEMANTIC_SUBGROUP_INVOCATION
:
1869 value
= ac_get_thread_id(&ctx
->ac
);
1872 case TGSI_SEMANTIC_SUBGROUP_EQ_MASK
:
1874 LLVMValueRef id
= ac_get_thread_id(&ctx
->ac
);
1875 id
= LLVMBuildZExt(ctx
->ac
.builder
, id
, ctx
->i64
, "");
1876 value
= LLVMBuildShl(ctx
->ac
.builder
, LLVMConstInt(ctx
->i64
, 1, 0), id
, "");
1877 value
= LLVMBuildBitCast(ctx
->ac
.builder
, value
, ctx
->v2i32
, "");
1881 case TGSI_SEMANTIC_SUBGROUP_GE_MASK
:
1882 case TGSI_SEMANTIC_SUBGROUP_GT_MASK
:
1883 case TGSI_SEMANTIC_SUBGROUP_LE_MASK
:
1884 case TGSI_SEMANTIC_SUBGROUP_LT_MASK
:
1886 LLVMValueRef id
= ac_get_thread_id(&ctx
->ac
);
1887 if (decl
->Semantic
.Name
== TGSI_SEMANTIC_SUBGROUP_GT_MASK
||
1888 decl
->Semantic
.Name
== TGSI_SEMANTIC_SUBGROUP_LE_MASK
) {
1889 /* All bits set except LSB */
1890 value
= LLVMConstInt(ctx
->i64
, -2, 0);
1893 value
= LLVMConstInt(ctx
->i64
, -1, 0);
1895 id
= LLVMBuildZExt(ctx
->ac
.builder
, id
, ctx
->i64
, "");
1896 value
= LLVMBuildShl(ctx
->ac
.builder
, value
, id
, "");
1897 if (decl
->Semantic
.Name
== TGSI_SEMANTIC_SUBGROUP_LE_MASK
||
1898 decl
->Semantic
.Name
== TGSI_SEMANTIC_SUBGROUP_LT_MASK
)
1899 value
= LLVMBuildNot(ctx
->ac
.builder
, value
, "");
1900 value
= LLVMBuildBitCast(ctx
->ac
.builder
, value
, ctx
->v2i32
, "");
1905 assert(!"unknown system value");
1909 ctx
->system_values
[index
] = value
;
1912 void si_declare_compute_memory(struct si_shader_context
*ctx
,
1913 const struct tgsi_full_declaration
*decl
)
1915 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
1917 LLVMTypeRef i8p
= LLVMPointerType(ctx
->i8
, LOCAL_ADDR_SPACE
);
1920 assert(decl
->Declaration
.MemType
== TGSI_MEMORY_TYPE_SHARED
);
1921 assert(decl
->Range
.First
== decl
->Range
.Last
);
1922 assert(!ctx
->shared_memory
);
1924 var
= LLVMAddGlobalInAddressSpace(ctx
->ac
.module
,
1925 LLVMArrayType(ctx
->i8
, sel
->local_size
),
1928 LLVMSetAlignment(var
, 4);
1930 ctx
->shared_memory
= LLVMBuildBitCast(ctx
->ac
.builder
, var
, i8p
, "");
1933 static LLVMValueRef
load_const_buffer_desc(struct si_shader_context
*ctx
, int i
)
1935 LLVMValueRef list_ptr
= LLVMGetParam(ctx
->main_fn
,
1936 ctx
->param_const_and_shader_buffers
);
1938 return ac_build_indexed_load_const(&ctx
->ac
, list_ptr
,
1939 LLVMConstInt(ctx
->i32
, si_get_constbuf_slot(i
), 0));
1942 static LLVMValueRef
load_ubo(struct ac_shader_abi
*abi
, LLVMValueRef index
)
1944 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1945 LLVMValueRef ptr
= LLVMGetParam(ctx
->main_fn
, ctx
->param_const_and_shader_buffers
);
1947 index
= si_llvm_bound_index(ctx
, index
, ctx
->num_const_buffers
);
1948 index
= LLVMBuildAdd(ctx
->ac
.builder
, index
,
1949 LLVMConstInt(ctx
->i32
, SI_NUM_SHADER_BUFFERS
, 0), "");
1951 return ac_build_indexed_load_const(&ctx
->ac
, ptr
, index
);
1955 load_ssbo(struct ac_shader_abi
*abi
, LLVMValueRef index
, bool write
)
1957 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1958 LLVMValueRef rsrc_ptr
= LLVMGetParam(ctx
->main_fn
,
1959 ctx
->param_const_and_shader_buffers
);
1961 index
= si_llvm_bound_index(ctx
, index
, ctx
->num_shader_buffers
);
1962 index
= LLVMBuildSub(ctx
->ac
.builder
,
1963 LLVMConstInt(ctx
->i32
, SI_NUM_SHADER_BUFFERS
- 1, 0),
1966 return ac_build_indexed_load_const(&ctx
->ac
, rsrc_ptr
, index
);
1969 static LLVMValueRef
fetch_constant(
1970 struct lp_build_tgsi_context
*bld_base
,
1971 const struct tgsi_full_src_register
*reg
,
1972 enum tgsi_opcode_type type
,
1975 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1976 const struct tgsi_ind_register
*ireg
= ®
->Indirect
;
1979 LLVMValueRef addr
, bufp
;
1981 if (swizzle
== LP_CHAN_ALL
) {
1983 LLVMValueRef values
[4];
1984 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
)
1985 values
[chan
] = fetch_constant(bld_base
, reg
, type
, chan
);
1987 return lp_build_gather_values(&ctx
->gallivm
, values
, 4);
1990 /* Split 64-bit loads. */
1991 if (tgsi_type_is_64bit(type
)) {
1992 LLVMValueRef lo
, hi
;
1994 lo
= fetch_constant(bld_base
, reg
, TGSI_TYPE_UNSIGNED
, swizzle
);
1995 hi
= fetch_constant(bld_base
, reg
, TGSI_TYPE_UNSIGNED
, swizzle
+ 1);
1996 return si_llvm_emit_fetch_64bit(bld_base
, type
, lo
, hi
);
1999 assert(reg
->Register
.Dimension
);
2000 buf
= reg
->Dimension
.Index
;
2001 idx
= reg
->Register
.Index
* 4 + swizzle
;
2003 if (reg
->Dimension
.Indirect
) {
2004 LLVMValueRef ptr
= LLVMGetParam(ctx
->main_fn
, ctx
->param_const_and_shader_buffers
);
2006 index
= si_get_bounded_indirect_index(ctx
, ®
->DimIndirect
,
2007 reg
->Dimension
.Index
,
2008 ctx
->num_const_buffers
);
2009 index
= LLVMBuildAdd(ctx
->ac
.builder
, index
,
2010 LLVMConstInt(ctx
->i32
, SI_NUM_SHADER_BUFFERS
, 0), "");
2011 bufp
= ac_build_indexed_load_const(&ctx
->ac
, ptr
, index
);
2013 bufp
= load_const_buffer_desc(ctx
, buf
);
2015 if (reg
->Register
.Indirect
) {
2016 addr
= si_get_indirect_index(ctx
, ireg
, 16, idx
* 4);
2018 addr
= LLVMConstInt(ctx
->i32
, idx
* 4, 0);
2021 return bitcast(bld_base
, type
, buffer_load_const(ctx
, bufp
, addr
));
2024 /* Upper 16 bits must be zero. */
2025 static LLVMValueRef
si_llvm_pack_two_int16(struct si_shader_context
*ctx
,
2026 LLVMValueRef val
[2])
2028 return LLVMBuildOr(ctx
->ac
.builder
, val
[0],
2029 LLVMBuildShl(ctx
->ac
.builder
, val
[1],
2030 LLVMConstInt(ctx
->i32
, 16, 0),
2034 /* Upper 16 bits are ignored and will be dropped. */
2035 static LLVMValueRef
si_llvm_pack_two_int32_as_int16(struct si_shader_context
*ctx
,
2036 LLVMValueRef val
[2])
2038 LLVMValueRef v
[2] = {
2039 LLVMBuildAnd(ctx
->ac
.builder
, val
[0],
2040 LLVMConstInt(ctx
->i32
, 0xffff, 0), ""),
2043 return si_llvm_pack_two_int16(ctx
, v
);
2046 /* Initialize arguments for the shader export intrinsic */
2047 static void si_llvm_init_export_args(struct lp_build_tgsi_context
*bld_base
,
2048 LLVMValueRef
*values
,
2050 struct ac_export_args
*args
)
2052 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2053 struct lp_build_context
*base
= &bld_base
->base
;
2054 LLVMBuilderRef builder
= ctx
->ac
.builder
;
2055 LLVMValueRef val
[4];
2056 unsigned spi_shader_col_format
= V_028714_SPI_SHADER_32_ABGR
;
2058 bool is_int8
, is_int10
;
2060 /* Default is 0xf. Adjusted below depending on the format. */
2061 args
->enabled_channels
= 0xf; /* writemask */
2063 /* Specify whether the EXEC mask represents the valid mask */
2064 args
->valid_mask
= 0;
2066 /* Specify whether this is the last export */
2069 /* Specify the target we are exporting */
2070 args
->target
= target
;
2072 if (ctx
->type
== PIPE_SHADER_FRAGMENT
) {
2073 const struct si_shader_key
*key
= &ctx
->shader
->key
;
2074 unsigned col_formats
= key
->part
.ps
.epilog
.spi_shader_col_format
;
2075 int cbuf
= target
- V_008DFC_SQ_EXP_MRT
;
2077 assert(cbuf
>= 0 && cbuf
< 8);
2078 spi_shader_col_format
= (col_formats
>> (cbuf
* 4)) & 0xf;
2079 is_int8
= (key
->part
.ps
.epilog
.color_is_int8
>> cbuf
) & 0x1;
2080 is_int10
= (key
->part
.ps
.epilog
.color_is_int10
>> cbuf
) & 0x1;
2083 args
->compr
= false;
2084 args
->out
[0] = base
->undef
;
2085 args
->out
[1] = base
->undef
;
2086 args
->out
[2] = base
->undef
;
2087 args
->out
[3] = base
->undef
;
2089 switch (spi_shader_col_format
) {
2090 case V_028714_SPI_SHADER_ZERO
:
2091 args
->enabled_channels
= 0; /* writemask */
2092 args
->target
= V_008DFC_SQ_EXP_NULL
;
2095 case V_028714_SPI_SHADER_32_R
:
2096 args
->enabled_channels
= 1; /* writemask */
2097 args
->out
[0] = values
[0];
2100 case V_028714_SPI_SHADER_32_GR
:
2101 args
->enabled_channels
= 0x3; /* writemask */
2102 args
->out
[0] = values
[0];
2103 args
->out
[1] = values
[1];
2106 case V_028714_SPI_SHADER_32_AR
:
2107 args
->enabled_channels
= 0x9; /* writemask */
2108 args
->out
[0] = values
[0];
2109 args
->out
[3] = values
[3];
2112 case V_028714_SPI_SHADER_FP16_ABGR
:
2113 args
->compr
= 1; /* COMPR flag */
2115 for (chan
= 0; chan
< 2; chan
++) {
2116 LLVMValueRef pack_args
[2] = {
2118 values
[2 * chan
+ 1]
2120 LLVMValueRef packed
;
2122 packed
= ac_build_cvt_pkrtz_f16(&ctx
->ac
, pack_args
);
2123 args
->out
[chan
] = ac_to_float(&ctx
->ac
, packed
);
2127 case V_028714_SPI_SHADER_UNORM16_ABGR
:
2128 for (chan
= 0; chan
< 4; chan
++) {
2129 val
[chan
] = ac_build_clamp(&ctx
->ac
, values
[chan
]);
2130 val
[chan
] = LLVMBuildFMul(builder
, val
[chan
],
2131 LLVMConstReal(ctx
->f32
, 65535), "");
2132 val
[chan
] = LLVMBuildFAdd(builder
, val
[chan
],
2133 LLVMConstReal(ctx
->f32
, 0.5), "");
2134 val
[chan
] = LLVMBuildFPToUI(builder
, val
[chan
],
2138 args
->compr
= 1; /* COMPR flag */
2139 args
->out
[0] = ac_to_float(&ctx
->ac
, si_llvm_pack_two_int16(ctx
, val
));
2140 args
->out
[1] = ac_to_float(&ctx
->ac
, si_llvm_pack_two_int16(ctx
, val
+2));
2143 case V_028714_SPI_SHADER_SNORM16_ABGR
:
2144 for (chan
= 0; chan
< 4; chan
++) {
2145 /* Clamp between [-1, 1]. */
2146 val
[chan
] = lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_MIN
,
2148 LLVMConstReal(ctx
->f32
, 1));
2149 val
[chan
] = lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_MAX
,
2151 LLVMConstReal(ctx
->f32
, -1));
2152 /* Convert to a signed integer in [-32767, 32767]. */
2153 val
[chan
] = LLVMBuildFMul(builder
, val
[chan
],
2154 LLVMConstReal(ctx
->f32
, 32767), "");
2155 /* If positive, add 0.5, else add -0.5. */
2156 val
[chan
] = LLVMBuildFAdd(builder
, val
[chan
],
2157 LLVMBuildSelect(builder
,
2158 LLVMBuildFCmp(builder
, LLVMRealOGE
,
2159 val
[chan
], ctx
->ac
.f32_0
, ""),
2160 LLVMConstReal(ctx
->f32
, 0.5),
2161 LLVMConstReal(ctx
->f32
, -0.5), ""), "");
2162 val
[chan
] = LLVMBuildFPToSI(builder
, val
[chan
], ctx
->i32
, "");
2165 args
->compr
= 1; /* COMPR flag */
2166 args
->out
[0] = ac_to_float(&ctx
->ac
, si_llvm_pack_two_int32_as_int16(ctx
, val
));
2167 args
->out
[1] = ac_to_float(&ctx
->ac
, si_llvm_pack_two_int32_as_int16(ctx
, val
+2));
2170 case V_028714_SPI_SHADER_UINT16_ABGR
: {
2171 LLVMValueRef max_rgb
= LLVMConstInt(ctx
->i32
,
2172 is_int8
? 255 : is_int10
? 1023 : 65535, 0);
2173 LLVMValueRef max_alpha
=
2174 !is_int10
? max_rgb
: LLVMConstInt(ctx
->i32
, 3, 0);
2177 for (chan
= 0; chan
< 4; chan
++) {
2178 val
[chan
] = ac_to_integer(&ctx
->ac
, values
[chan
]);
2179 val
[chan
] = lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_UMIN
,
2181 chan
== 3 ? max_alpha
: max_rgb
);
2184 args
->compr
= 1; /* COMPR flag */
2185 args
->out
[0] = ac_to_float(&ctx
->ac
, si_llvm_pack_two_int16(ctx
, val
));
2186 args
->out
[1] = ac_to_float(&ctx
->ac
, si_llvm_pack_two_int16(ctx
, val
+2));
2190 case V_028714_SPI_SHADER_SINT16_ABGR
: {
2191 LLVMValueRef max_rgb
= LLVMConstInt(ctx
->i32
,
2192 is_int8
? 127 : is_int10
? 511 : 32767, 0);
2193 LLVMValueRef min_rgb
= LLVMConstInt(ctx
->i32
,
2194 is_int8
? -128 : is_int10
? -512 : -32768, 0);
2195 LLVMValueRef max_alpha
=
2196 !is_int10
? max_rgb
: ctx
->i32_1
;
2197 LLVMValueRef min_alpha
=
2198 !is_int10
? min_rgb
: LLVMConstInt(ctx
->i32
, -2, 0);
2201 for (chan
= 0; chan
< 4; chan
++) {
2202 val
[chan
] = ac_to_integer(&ctx
->ac
, values
[chan
]);
2203 val
[chan
] = lp_build_emit_llvm_binary(bld_base
,
2205 val
[chan
], chan
== 3 ? max_alpha
: max_rgb
);
2206 val
[chan
] = lp_build_emit_llvm_binary(bld_base
,
2208 val
[chan
], chan
== 3 ? min_alpha
: min_rgb
);
2211 args
->compr
= 1; /* COMPR flag */
2212 args
->out
[0] = ac_to_float(&ctx
->ac
, si_llvm_pack_two_int32_as_int16(ctx
, val
));
2213 args
->out
[1] = ac_to_float(&ctx
->ac
, si_llvm_pack_two_int32_as_int16(ctx
, val
+2));
2217 case V_028714_SPI_SHADER_32_ABGR
:
2218 memcpy(&args
->out
[0], values
, sizeof(values
[0]) * 4);
2223 static void si_alpha_test(struct lp_build_tgsi_context
*bld_base
,
2226 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2228 if (ctx
->shader
->key
.part
.ps
.epilog
.alpha_func
!= PIPE_FUNC_NEVER
) {
2229 LLVMValueRef alpha_ref
= LLVMGetParam(ctx
->main_fn
,
2230 SI_PARAM_ALPHA_REF
);
2232 LLVMValueRef alpha_pass
=
2233 lp_build_cmp(&bld_base
->base
,
2234 ctx
->shader
->key
.part
.ps
.epilog
.alpha_func
,
2237 lp_build_select(&bld_base
->base
,
2239 LLVMConstReal(ctx
->f32
, 1.0f
),
2240 LLVMConstReal(ctx
->f32
, -1.0f
));
2242 ac_build_kill(&ctx
->ac
, arg
);
2244 ac_build_kill(&ctx
->ac
, NULL
);
2248 static LLVMValueRef
si_scale_alpha_by_sample_mask(struct lp_build_tgsi_context
*bld_base
,
2250 unsigned samplemask_param
)
2252 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2253 LLVMValueRef coverage
;
2255 /* alpha = alpha * popcount(coverage) / SI_NUM_SMOOTH_AA_SAMPLES */
2256 coverage
= LLVMGetParam(ctx
->main_fn
,
2258 coverage
= ac_to_integer(&ctx
->ac
, coverage
);
2260 coverage
= lp_build_intrinsic(ctx
->ac
.builder
, "llvm.ctpop.i32",
2262 &coverage
, 1, LP_FUNC_ATTR_READNONE
);
2264 coverage
= LLVMBuildUIToFP(ctx
->ac
.builder
, coverage
,
2267 coverage
= LLVMBuildFMul(ctx
->ac
.builder
, coverage
,
2268 LLVMConstReal(ctx
->f32
,
2269 1.0 / SI_NUM_SMOOTH_AA_SAMPLES
), "");
2271 return LLVMBuildFMul(ctx
->ac
.builder
, alpha
, coverage
, "");
2274 static void si_llvm_emit_clipvertex(struct lp_build_tgsi_context
*bld_base
,
2275 struct ac_export_args
*pos
, LLVMValueRef
*out_elts
)
2277 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2278 struct lp_build_context
*base
= &bld_base
->base
;
2281 unsigned const_chan
;
2282 LLVMValueRef base_elt
;
2283 LLVMValueRef ptr
= LLVMGetParam(ctx
->main_fn
, ctx
->param_rw_buffers
);
2284 LLVMValueRef constbuf_index
= LLVMConstInt(ctx
->i32
,
2285 SI_VS_CONST_CLIP_PLANES
, 0);
2286 LLVMValueRef const_resource
= ac_build_indexed_load_const(&ctx
->ac
, ptr
, constbuf_index
);
2288 for (reg_index
= 0; reg_index
< 2; reg_index
++) {
2289 struct ac_export_args
*args
= &pos
[2 + reg_index
];
2294 args
->out
[3] = LLVMConstReal(ctx
->f32
, 0.0f
);
2296 /* Compute dot products of position and user clip plane vectors */
2297 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
2298 for (const_chan
= 0; const_chan
< TGSI_NUM_CHANNELS
; const_chan
++) {
2300 LLVMConstInt(ctx
->i32
, ((reg_index
* 4 + chan
) * 4 +
2301 const_chan
) * 4, 0);
2302 base_elt
= buffer_load_const(ctx
, const_resource
,
2305 lp_build_add(base
, args
->out
[chan
],
2306 lp_build_mul(base
, base_elt
,
2307 out_elts
[const_chan
]));
2311 args
->enabled_channels
= 0xf;
2312 args
->valid_mask
= 0;
2314 args
->target
= V_008DFC_SQ_EXP_POS
+ 2 + reg_index
;
2319 static void si_dump_streamout(struct pipe_stream_output_info
*so
)
2323 if (so
->num_outputs
)
2324 fprintf(stderr
, "STREAMOUT\n");
2326 for (i
= 0; i
< so
->num_outputs
; i
++) {
2327 unsigned mask
= ((1 << so
->output
[i
].num_components
) - 1) <<
2328 so
->output
[i
].start_component
;
2329 fprintf(stderr
, " %i: BUF%i[%i..%i] <- OUT[%i].%s%s%s%s\n",
2330 i
, so
->output
[i
].output_buffer
,
2331 so
->output
[i
].dst_offset
, so
->output
[i
].dst_offset
+ so
->output
[i
].num_components
- 1,
2332 so
->output
[i
].register_index
,
2333 mask
& 1 ? "x" : "",
2334 mask
& 2 ? "y" : "",
2335 mask
& 4 ? "z" : "",
2336 mask
& 8 ? "w" : "");
2340 static void emit_streamout_output(struct si_shader_context
*ctx
,
2341 LLVMValueRef
const *so_buffers
,
2342 LLVMValueRef
const *so_write_offsets
,
2343 struct pipe_stream_output
*stream_out
,
2344 struct si_shader_output_values
*shader_out
)
2346 unsigned buf_idx
= stream_out
->output_buffer
;
2347 unsigned start
= stream_out
->start_component
;
2348 unsigned num_comps
= stream_out
->num_components
;
2349 LLVMValueRef out
[4];
2351 assert(num_comps
&& num_comps
<= 4);
2352 if (!num_comps
|| num_comps
> 4)
2355 /* Load the output as int. */
2356 for (int j
= 0; j
< num_comps
; j
++) {
2357 assert(stream_out
->stream
== shader_out
->vertex_stream
[start
+ j
]);
2359 out
[j
] = ac_to_integer(&ctx
->ac
, shader_out
->values
[start
+ j
]);
2362 /* Pack the output. */
2363 LLVMValueRef vdata
= NULL
;
2365 switch (num_comps
) {
2366 case 1: /* as i32 */
2369 case 2: /* as v2i32 */
2370 case 3: /* as v4i32 (aligned to 4) */
2371 case 4: /* as v4i32 */
2372 vdata
= LLVMGetUndef(LLVMVectorType(ctx
->i32
, util_next_power_of_two(num_comps
)));
2373 for (int j
= 0; j
< num_comps
; j
++) {
2374 vdata
= LLVMBuildInsertElement(ctx
->ac
.builder
, vdata
, out
[j
],
2375 LLVMConstInt(ctx
->i32
, j
, 0), "");
2380 ac_build_buffer_store_dword(&ctx
->ac
, so_buffers
[buf_idx
],
2382 so_write_offsets
[buf_idx
],
2384 stream_out
->dst_offset
* 4, 1, 1, true, false);
2388 * Write streamout data to buffers for vertex stream @p stream (different
2389 * vertex streams can occur for GS copy shaders).
2391 static void si_llvm_emit_streamout(struct si_shader_context
*ctx
,
2392 struct si_shader_output_values
*outputs
,
2393 unsigned noutput
, unsigned stream
)
2395 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
2396 struct pipe_stream_output_info
*so
= &sel
->so
;
2397 LLVMBuilderRef builder
= ctx
->ac
.builder
;
2399 struct lp_build_if_state if_ctx
;
2401 /* Get bits [22:16], i.e. (so_param >> 16) & 127; */
2402 LLVMValueRef so_vtx_count
=
2403 unpack_param(ctx
, ctx
->param_streamout_config
, 16, 7);
2405 LLVMValueRef tid
= ac_get_thread_id(&ctx
->ac
);
2407 /* can_emit = tid < so_vtx_count; */
2408 LLVMValueRef can_emit
=
2409 LLVMBuildICmp(builder
, LLVMIntULT
, tid
, so_vtx_count
, "");
2411 /* Emit the streamout code conditionally. This actually avoids
2412 * out-of-bounds buffer access. The hw tells us via the SGPR
2413 * (so_vtx_count) which threads are allowed to emit streamout data. */
2414 lp_build_if(&if_ctx
, &ctx
->gallivm
, can_emit
);
2416 /* The buffer offset is computed as follows:
2417 * ByteOffset = streamout_offset[buffer_id]*4 +
2418 * (streamout_write_index + thread_id)*stride[buffer_id] +
2422 LLVMValueRef so_write_index
=
2423 LLVMGetParam(ctx
->main_fn
,
2424 ctx
->param_streamout_write_index
);
2426 /* Compute (streamout_write_index + thread_id). */
2427 so_write_index
= LLVMBuildAdd(builder
, so_write_index
, tid
, "");
2429 /* Load the descriptor and compute the write offset for each
2430 * enabled buffer. */
2431 LLVMValueRef so_write_offset
[4] = {};
2432 LLVMValueRef so_buffers
[4];
2433 LLVMValueRef buf_ptr
= LLVMGetParam(ctx
->main_fn
,
2434 ctx
->param_rw_buffers
);
2436 for (i
= 0; i
< 4; i
++) {
2440 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
,
2441 SI_VS_STREAMOUT_BUF0
+ i
, 0);
2443 so_buffers
[i
] = ac_build_indexed_load_const(&ctx
->ac
, buf_ptr
, offset
);
2445 LLVMValueRef so_offset
= LLVMGetParam(ctx
->main_fn
,
2446 ctx
->param_streamout_offset
[i
]);
2447 so_offset
= LLVMBuildMul(builder
, so_offset
, LLVMConstInt(ctx
->i32
, 4, 0), "");
2449 so_write_offset
[i
] = LLVMBuildMul(builder
, so_write_index
,
2450 LLVMConstInt(ctx
->i32
, so
->stride
[i
]*4, 0), "");
2451 so_write_offset
[i
] = LLVMBuildAdd(builder
, so_write_offset
[i
], so_offset
, "");
2454 /* Write streamout data. */
2455 for (i
= 0; i
< so
->num_outputs
; i
++) {
2456 unsigned reg
= so
->output
[i
].register_index
;
2461 if (stream
!= so
->output
[i
].stream
)
2464 emit_streamout_output(ctx
, so_buffers
, so_write_offset
,
2465 &so
->output
[i
], &outputs
[reg
]);
2468 lp_build_endif(&if_ctx
);
2471 static void si_export_param(struct si_shader_context
*ctx
, unsigned index
,
2472 LLVMValueRef
*values
)
2474 struct ac_export_args args
;
2476 si_llvm_init_export_args(&ctx
->bld_base
, values
,
2477 V_008DFC_SQ_EXP_PARAM
+ index
, &args
);
2478 ac_build_export(&ctx
->ac
, &args
);
2481 static void si_build_param_exports(struct si_shader_context
*ctx
,
2482 struct si_shader_output_values
*outputs
,
2485 struct si_shader
*shader
= ctx
->shader
;
2486 unsigned param_count
= 0;
2488 for (unsigned i
= 0; i
< noutput
; i
++) {
2489 unsigned semantic_name
= outputs
[i
].semantic_name
;
2490 unsigned semantic_index
= outputs
[i
].semantic_index
;
2492 if (outputs
[i
].vertex_stream
[0] != 0 &&
2493 outputs
[i
].vertex_stream
[1] != 0 &&
2494 outputs
[i
].vertex_stream
[2] != 0 &&
2495 outputs
[i
].vertex_stream
[3] != 0)
2498 switch (semantic_name
) {
2499 case TGSI_SEMANTIC_LAYER
:
2500 case TGSI_SEMANTIC_VIEWPORT_INDEX
:
2501 case TGSI_SEMANTIC_CLIPDIST
:
2502 case TGSI_SEMANTIC_COLOR
:
2503 case TGSI_SEMANTIC_BCOLOR
:
2504 case TGSI_SEMANTIC_PRIMID
:
2505 case TGSI_SEMANTIC_FOG
:
2506 case TGSI_SEMANTIC_TEXCOORD
:
2507 case TGSI_SEMANTIC_GENERIC
:
2513 if ((semantic_name
!= TGSI_SEMANTIC_GENERIC
||
2514 semantic_index
< SI_MAX_IO_GENERIC
) &&
2515 shader
->key
.opt
.kill_outputs
&
2516 (1ull << si_shader_io_get_unique_index(semantic_name
, semantic_index
)))
2519 si_export_param(ctx
, param_count
, outputs
[i
].values
);
2521 assert(i
< ARRAY_SIZE(shader
->info
.vs_output_param_offset
));
2522 shader
->info
.vs_output_param_offset
[i
] = param_count
++;
2525 shader
->info
.nr_param_exports
= param_count
;
2528 /* Generate export instructions for hardware VS shader stage */
2529 static void si_llvm_export_vs(struct lp_build_tgsi_context
*bld_base
,
2530 struct si_shader_output_values
*outputs
,
2533 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2534 struct si_shader
*shader
= ctx
->shader
;
2535 struct ac_export_args pos_args
[4] = {};
2536 LLVMValueRef psize_value
= NULL
, edgeflag_value
= NULL
, layer_value
= NULL
, viewport_index_value
= NULL
;
2540 /* Build position exports. */
2541 for (i
= 0; i
< noutput
; i
++) {
2542 switch (outputs
[i
].semantic_name
) {
2543 case TGSI_SEMANTIC_POSITION
:
2544 si_llvm_init_export_args(bld_base
, outputs
[i
].values
,
2545 V_008DFC_SQ_EXP_POS
, &pos_args
[0]);
2547 case TGSI_SEMANTIC_PSIZE
:
2548 psize_value
= outputs
[i
].values
[0];
2550 case TGSI_SEMANTIC_LAYER
:
2551 layer_value
= outputs
[i
].values
[0];
2553 case TGSI_SEMANTIC_VIEWPORT_INDEX
:
2554 viewport_index_value
= outputs
[i
].values
[0];
2556 case TGSI_SEMANTIC_EDGEFLAG
:
2557 edgeflag_value
= outputs
[i
].values
[0];
2559 case TGSI_SEMANTIC_CLIPDIST
:
2560 if (!shader
->key
.opt
.clip_disable
) {
2561 unsigned index
= 2 + outputs
[i
].semantic_index
;
2562 si_llvm_init_export_args(bld_base
, outputs
[i
].values
,
2563 V_008DFC_SQ_EXP_POS
+ index
,
2567 case TGSI_SEMANTIC_CLIPVERTEX
:
2568 if (!shader
->key
.opt
.clip_disable
) {
2569 si_llvm_emit_clipvertex(bld_base
, pos_args
,
2576 /* We need to add the position output manually if it's missing. */
2577 if (!pos_args
[0].out
[0]) {
2578 pos_args
[0].enabled_channels
= 0xf; /* writemask */
2579 pos_args
[0].valid_mask
= 0; /* EXEC mask */
2580 pos_args
[0].done
= 0; /* last export? */
2581 pos_args
[0].target
= V_008DFC_SQ_EXP_POS
;
2582 pos_args
[0].compr
= 0; /* COMPR flag */
2583 pos_args
[0].out
[0] = ctx
->ac
.f32_0
; /* X */
2584 pos_args
[0].out
[1] = ctx
->ac
.f32_0
; /* Y */
2585 pos_args
[0].out
[2] = ctx
->ac
.f32_0
; /* Z */
2586 pos_args
[0].out
[3] = ctx
->ac
.f32_1
; /* W */
2589 /* Write the misc vector (point size, edgeflag, layer, viewport). */
2590 if (shader
->selector
->info
.writes_psize
||
2591 shader
->selector
->info
.writes_edgeflag
||
2592 shader
->selector
->info
.writes_viewport_index
||
2593 shader
->selector
->info
.writes_layer
) {
2594 pos_args
[1].enabled_channels
= shader
->selector
->info
.writes_psize
|
2595 (shader
->selector
->info
.writes_edgeflag
<< 1) |
2596 (shader
->selector
->info
.writes_layer
<< 2);
2598 pos_args
[1].valid_mask
= 0; /* EXEC mask */
2599 pos_args
[1].done
= 0; /* last export? */
2600 pos_args
[1].target
= V_008DFC_SQ_EXP_POS
+ 1;
2601 pos_args
[1].compr
= 0; /* COMPR flag */
2602 pos_args
[1].out
[0] = ctx
->ac
.f32_0
; /* X */
2603 pos_args
[1].out
[1] = ctx
->ac
.f32_0
; /* Y */
2604 pos_args
[1].out
[2] = ctx
->ac
.f32_0
; /* Z */
2605 pos_args
[1].out
[3] = ctx
->ac
.f32_0
; /* W */
2607 if (shader
->selector
->info
.writes_psize
)
2608 pos_args
[1].out
[0] = psize_value
;
2610 if (shader
->selector
->info
.writes_edgeflag
) {
2611 /* The output is a float, but the hw expects an integer
2612 * with the first bit containing the edge flag. */
2613 edgeflag_value
= LLVMBuildFPToUI(ctx
->ac
.builder
,
2616 edgeflag_value
= ac_build_umin(&ctx
->ac
,
2620 /* The LLVM intrinsic expects a float. */
2621 pos_args
[1].out
[1] = ac_to_float(&ctx
->ac
, edgeflag_value
);
2624 if (ctx
->screen
->b
.chip_class
>= GFX9
) {
2625 /* GFX9 has the layer in out.z[10:0] and the viewport
2626 * index in out.z[19:16].
2628 if (shader
->selector
->info
.writes_layer
)
2629 pos_args
[1].out
[2] = layer_value
;
2631 if (shader
->selector
->info
.writes_viewport_index
) {
2632 LLVMValueRef v
= viewport_index_value
;
2634 v
= ac_to_integer(&ctx
->ac
, v
);
2635 v
= LLVMBuildShl(ctx
->ac
.builder
, v
,
2636 LLVMConstInt(ctx
->i32
, 16, 0), "");
2637 v
= LLVMBuildOr(ctx
->ac
.builder
, v
,
2638 ac_to_integer(&ctx
->ac
, pos_args
[1].out
[2]), "");
2639 pos_args
[1].out
[2] = ac_to_float(&ctx
->ac
, v
);
2640 pos_args
[1].enabled_channels
|= 1 << 2;
2643 if (shader
->selector
->info
.writes_layer
)
2644 pos_args
[1].out
[2] = layer_value
;
2646 if (shader
->selector
->info
.writes_viewport_index
) {
2647 pos_args
[1].out
[3] = viewport_index_value
;
2648 pos_args
[1].enabled_channels
|= 1 << 3;
2653 for (i
= 0; i
< 4; i
++)
2654 if (pos_args
[i
].out
[0])
2655 shader
->info
.nr_pos_exports
++;
2658 for (i
= 0; i
< 4; i
++) {
2659 if (!pos_args
[i
].out
[0])
2662 /* Specify the target we are exporting */
2663 pos_args
[i
].target
= V_008DFC_SQ_EXP_POS
+ pos_idx
++;
2665 if (pos_idx
== shader
->info
.nr_pos_exports
)
2666 /* Specify that this is the last export */
2667 pos_args
[i
].done
= 1;
2669 ac_build_export(&ctx
->ac
, &pos_args
[i
]);
2672 /* Build parameter exports. */
2673 si_build_param_exports(ctx
, outputs
, noutput
);
2677 * Forward all outputs from the vertex shader to the TES. This is only used
2678 * for the fixed function TCS.
2680 static void si_copy_tcs_inputs(struct lp_build_tgsi_context
*bld_base
)
2682 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2683 LLVMValueRef invocation_id
, buffer
, buffer_offset
;
2684 LLVMValueRef lds_vertex_stride
, lds_vertex_offset
, lds_base
;
2687 invocation_id
= unpack_param(ctx
, ctx
->param_tcs_rel_ids
, 8, 5);
2688 buffer
= desc_from_addr_base64k(ctx
, ctx
->param_tcs_offchip_addr_base64k
);
2689 buffer_offset
= LLVMGetParam(ctx
->main_fn
, ctx
->param_tcs_offchip_offset
);
2691 lds_vertex_stride
= get_tcs_in_vertex_dw_stride(ctx
);
2692 lds_vertex_offset
= LLVMBuildMul(ctx
->ac
.builder
, invocation_id
,
2693 lds_vertex_stride
, "");
2694 lds_base
= get_tcs_in_current_patch_offset(ctx
);
2695 lds_base
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
, lds_vertex_offset
, "");
2697 inputs
= ctx
->shader
->key
.mono
.u
.ff_tcs_inputs_to_copy
;
2699 unsigned i
= u_bit_scan64(&inputs
);
2701 LLVMValueRef lds_ptr
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
2702 LLVMConstInt(ctx
->i32
, 4 * i
, 0),
2705 LLVMValueRef buffer_addr
= get_tcs_tes_buffer_address(ctx
,
2706 get_rel_patch_id(ctx
),
2708 LLVMConstInt(ctx
->i32
, i
, 0));
2710 LLVMValueRef value
= lds_load(bld_base
, TGSI_TYPE_SIGNED
, ~0,
2713 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, value
, 4, buffer_addr
,
2714 buffer_offset
, 0, 1, 0, true, false);
2718 static void si_write_tess_factors(struct lp_build_tgsi_context
*bld_base
,
2719 LLVMValueRef rel_patch_id
,
2720 LLVMValueRef invocation_id
,
2721 LLVMValueRef tcs_out_current_patch_data_offset
,
2722 LLVMValueRef invoc0_tf_outer
[4],
2723 LLVMValueRef invoc0_tf_inner
[2])
2725 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2726 struct si_shader
*shader
= ctx
->shader
;
2727 unsigned tess_inner_index
, tess_outer_index
;
2728 LLVMValueRef lds_base
, lds_inner
, lds_outer
, byteoffset
, buffer
;
2729 LLVMValueRef out
[6], vec0
, vec1
, tf_base
, inner
[4], outer
[4];
2730 unsigned stride
, outer_comps
, inner_comps
, i
, offset
;
2731 struct lp_build_if_state if_ctx
, inner_if_ctx
;
2733 /* Add a barrier before loading tess factors from LDS. */
2734 if (!shader
->key
.part
.tcs
.epilog
.invoc0_tess_factors_are_def
)
2735 si_llvm_emit_barrier(NULL
, bld_base
, NULL
);
2737 /* Do this only for invocation 0, because the tess levels are per-patch,
2740 * This can't jump, because invocation 0 executes this. It should
2741 * at least mask out the loads and stores for other invocations.
2743 lp_build_if(&if_ctx
, &ctx
->gallivm
,
2744 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
,
2745 invocation_id
, ctx
->i32_0
, ""));
2747 /* Determine the layout of one tess factor element in the buffer. */
2748 switch (shader
->key
.part
.tcs
.epilog
.prim_mode
) {
2749 case PIPE_PRIM_LINES
:
2750 stride
= 2; /* 2 dwords, 1 vec2 store */
2754 case PIPE_PRIM_TRIANGLES
:
2755 stride
= 4; /* 4 dwords, 1 vec4 store */
2759 case PIPE_PRIM_QUADS
:
2760 stride
= 6; /* 6 dwords, 2 stores (vec4 + vec2) */
2769 for (i
= 0; i
< 4; i
++) {
2770 inner
[i
] = LLVMGetUndef(ctx
->i32
);
2771 outer
[i
] = LLVMGetUndef(ctx
->i32
);
2774 if (shader
->key
.part
.tcs
.epilog
.invoc0_tess_factors_are_def
) {
2775 /* Tess factors are in VGPRs. */
2776 for (i
= 0; i
< outer_comps
; i
++)
2777 outer
[i
] = out
[i
] = invoc0_tf_outer
[i
];
2778 for (i
= 0; i
< inner_comps
; i
++)
2779 inner
[i
] = out
[outer_comps
+i
] = invoc0_tf_inner
[i
];
2781 /* Load tess_inner and tess_outer from LDS.
2782 * Any invocation can write them, so we can't get them from a temporary.
2784 tess_inner_index
= si_shader_io_get_unique_index_patch(TGSI_SEMANTIC_TESSINNER
, 0);
2785 tess_outer_index
= si_shader_io_get_unique_index_patch(TGSI_SEMANTIC_TESSOUTER
, 0);
2787 lds_base
= tcs_out_current_patch_data_offset
;
2788 lds_inner
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
2789 LLVMConstInt(ctx
->i32
,
2790 tess_inner_index
* 4, 0), "");
2791 lds_outer
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
2792 LLVMConstInt(ctx
->i32
,
2793 tess_outer_index
* 4, 0), "");
2795 for (i
= 0; i
< outer_comps
; i
++) {
2797 lds_load(bld_base
, TGSI_TYPE_SIGNED
, i
, lds_outer
);
2799 for (i
= 0; i
< inner_comps
; i
++) {
2800 inner
[i
] = out
[outer_comps
+i
] =
2801 lds_load(bld_base
, TGSI_TYPE_SIGNED
, i
, lds_inner
);
2805 if (shader
->key
.part
.tcs
.epilog
.prim_mode
== PIPE_PRIM_LINES
) {
2806 /* For isolines, the hardware expects tess factors in the
2807 * reverse order from what GLSL / TGSI specify.
2809 LLVMValueRef tmp
= out
[0];
2814 /* Convert the outputs to vectors for stores. */
2815 vec0
= lp_build_gather_values(&ctx
->gallivm
, out
, MIN2(stride
, 4));
2819 vec1
= lp_build_gather_values(&ctx
->gallivm
, out
+4, stride
- 4);
2821 /* Get the buffer. */
2822 buffer
= desc_from_addr_base64k(ctx
, ctx
->param_tcs_factor_addr_base64k
);
2824 /* Get the offset. */
2825 tf_base
= LLVMGetParam(ctx
->main_fn
,
2826 ctx
->param_tcs_factor_offset
);
2827 byteoffset
= LLVMBuildMul(ctx
->ac
.builder
, rel_patch_id
,
2828 LLVMConstInt(ctx
->i32
, 4 * stride
, 0), "");
2830 lp_build_if(&inner_if_ctx
, &ctx
->gallivm
,
2831 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
,
2832 rel_patch_id
, ctx
->i32_0
, ""));
2834 /* Store the dynamic HS control word. */
2836 if (ctx
->screen
->b
.chip_class
<= VI
) {
2837 ac_build_buffer_store_dword(&ctx
->ac
, buffer
,
2838 LLVMConstInt(ctx
->i32
, 0x80000000, 0),
2839 1, ctx
->i32_0
, tf_base
,
2840 offset
, 1, 0, true, false);
2844 lp_build_endif(&inner_if_ctx
);
2846 /* Store the tessellation factors. */
2847 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec0
,
2848 MIN2(stride
, 4), byteoffset
, tf_base
,
2849 offset
, 1, 0, true, false);
2852 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec1
,
2853 stride
- 4, byteoffset
, tf_base
,
2854 offset
, 1, 0, true, false);
2856 /* Store the tess factors into the offchip buffer if TES reads them. */
2857 if (shader
->key
.part
.tcs
.epilog
.tes_reads_tess_factors
) {
2858 LLVMValueRef buf
, base
, inner_vec
, outer_vec
, tf_outer_offset
;
2859 LLVMValueRef tf_inner_offset
;
2860 unsigned param_outer
, param_inner
;
2862 buf
= desc_from_addr_base64k(ctx
, ctx
->param_tcs_offchip_addr_base64k
);
2863 base
= LLVMGetParam(ctx
->main_fn
, ctx
->param_tcs_offchip_offset
);
2865 param_outer
= si_shader_io_get_unique_index_patch(
2866 TGSI_SEMANTIC_TESSOUTER
, 0);
2867 tf_outer_offset
= get_tcs_tes_buffer_address(ctx
, rel_patch_id
, NULL
,
2868 LLVMConstInt(ctx
->i32
, param_outer
, 0));
2870 outer_vec
= lp_build_gather_values(&ctx
->gallivm
, outer
,
2871 util_next_power_of_two(outer_comps
));
2873 ac_build_buffer_store_dword(&ctx
->ac
, buf
, outer_vec
,
2874 outer_comps
, tf_outer_offset
,
2875 base
, 0, 1, 0, true, false);
2877 param_inner
= si_shader_io_get_unique_index_patch(
2878 TGSI_SEMANTIC_TESSINNER
, 0);
2879 tf_inner_offset
= get_tcs_tes_buffer_address(ctx
, rel_patch_id
, NULL
,
2880 LLVMConstInt(ctx
->i32
, param_inner
, 0));
2882 inner_vec
= inner_comps
== 1 ? inner
[0] :
2883 lp_build_gather_values(&ctx
->gallivm
, inner
, inner_comps
);
2884 ac_build_buffer_store_dword(&ctx
->ac
, buf
, inner_vec
,
2885 inner_comps
, tf_inner_offset
,
2886 base
, 0, 1, 0, true, false);
2890 lp_build_endif(&if_ctx
);
2894 si_insert_input_ret(struct si_shader_context
*ctx
, LLVMValueRef ret
,
2895 unsigned param
, unsigned return_index
)
2897 return LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
2898 LLVMGetParam(ctx
->main_fn
, param
),
2903 si_insert_input_ret_float(struct si_shader_context
*ctx
, LLVMValueRef ret
,
2904 unsigned param
, unsigned return_index
)
2906 LLVMBuilderRef builder
= ctx
->ac
.builder
;
2907 LLVMValueRef p
= LLVMGetParam(ctx
->main_fn
, param
);
2909 return LLVMBuildInsertValue(builder
, ret
,
2910 ac_to_float(&ctx
->ac
, p
),
2915 si_insert_input_ptr_as_2xi32(struct si_shader_context
*ctx
, LLVMValueRef ret
,
2916 unsigned param
, unsigned return_index
)
2918 LLVMBuilderRef builder
= ctx
->ac
.builder
;
2919 LLVMValueRef ptr
, lo
, hi
;
2921 ptr
= LLVMGetParam(ctx
->main_fn
, param
);
2922 ptr
= LLVMBuildPtrToInt(builder
, ptr
, ctx
->i64
, "");
2923 ptr
= LLVMBuildBitCast(builder
, ptr
, ctx
->v2i32
, "");
2924 lo
= LLVMBuildExtractElement(builder
, ptr
, ctx
->i32_0
, "");
2925 hi
= LLVMBuildExtractElement(builder
, ptr
, ctx
->i32_1
, "");
2926 ret
= LLVMBuildInsertValue(builder
, ret
, lo
, return_index
, "");
2927 return LLVMBuildInsertValue(builder
, ret
, hi
, return_index
+ 1, "");
2930 /* This only writes the tessellation factor levels. */
2931 static void si_llvm_emit_tcs_epilogue(struct lp_build_tgsi_context
*bld_base
)
2933 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2934 LLVMBuilderRef builder
= ctx
->ac
.builder
;
2935 LLVMValueRef rel_patch_id
, invocation_id
, tf_lds_offset
;
2937 si_copy_tcs_inputs(bld_base
);
2939 rel_patch_id
= get_rel_patch_id(ctx
);
2940 invocation_id
= unpack_param(ctx
, ctx
->param_tcs_rel_ids
, 8, 5);
2941 tf_lds_offset
= get_tcs_out_current_patch_data_offset(ctx
);
2943 if (ctx
->screen
->b
.chip_class
>= GFX9
) {
2944 LLVMBasicBlockRef blocks
[2] = {
2945 LLVMGetInsertBlock(builder
),
2946 ctx
->merged_wrap_if_state
.entry_block
2948 LLVMValueRef values
[2];
2950 lp_build_endif(&ctx
->merged_wrap_if_state
);
2952 values
[0] = rel_patch_id
;
2953 values
[1] = LLVMGetUndef(ctx
->i32
);
2954 rel_patch_id
= ac_build_phi(&ctx
->ac
, ctx
->i32
, 2, values
, blocks
);
2956 values
[0] = tf_lds_offset
;
2957 values
[1] = LLVMGetUndef(ctx
->i32
);
2958 tf_lds_offset
= ac_build_phi(&ctx
->ac
, ctx
->i32
, 2, values
, blocks
);
2960 values
[0] = invocation_id
;
2961 values
[1] = ctx
->i32_1
; /* cause the epilog to skip threads */
2962 invocation_id
= ac_build_phi(&ctx
->ac
, ctx
->i32
, 2, values
, blocks
);
2965 /* Return epilog parameters from this function. */
2966 LLVMValueRef ret
= ctx
->return_value
;
2969 if (ctx
->screen
->b
.chip_class
>= GFX9
) {
2970 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_layout
,
2971 8 + GFX9_SGPR_TCS_OFFCHIP_LAYOUT
);
2972 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_addr_base64k
,
2973 8 + GFX9_SGPR_TCS_OFFCHIP_ADDR_BASE64K
);
2974 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_factor_addr_base64k
,
2975 8 + GFX9_SGPR_TCS_FACTOR_ADDR_BASE64K
);
2976 /* Tess offchip and tess factor offsets are at the beginning. */
2977 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_offset
, 2);
2978 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_factor_offset
, 4);
2979 vgpr
= 8 + GFX9_SGPR_TCS_FACTOR_ADDR_BASE64K
+ 1;
2981 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_layout
,
2982 GFX6_SGPR_TCS_OFFCHIP_LAYOUT
);
2983 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_addr_base64k
,
2984 GFX6_SGPR_TCS_OFFCHIP_ADDR_BASE64K
);
2985 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_factor_addr_base64k
,
2986 GFX6_SGPR_TCS_FACTOR_ADDR_BASE64K
);
2987 /* Tess offchip and tess factor offsets are after user SGPRs. */
2988 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_offset
,
2989 GFX6_TCS_NUM_USER_SGPR
);
2990 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_factor_offset
,
2991 GFX6_TCS_NUM_USER_SGPR
+ 1);
2992 vgpr
= GFX6_TCS_NUM_USER_SGPR
+ 2;
2996 rel_patch_id
= ac_to_float(&ctx
->ac
, rel_patch_id
);
2997 invocation_id
= ac_to_float(&ctx
->ac
, invocation_id
);
2998 tf_lds_offset
= ac_to_float(&ctx
->ac
, tf_lds_offset
);
3000 /* Leave a hole corresponding to the two input VGPRs. This ensures that
3001 * the invocation_id output does not alias the param_tcs_rel_ids input,
3002 * which saves a V_MOV on gfx9.
3006 ret
= LLVMBuildInsertValue(builder
, ret
, rel_patch_id
, vgpr
++, "");
3007 ret
= LLVMBuildInsertValue(builder
, ret
, invocation_id
, vgpr
++, "");
3009 if (ctx
->shader
->selector
->tcs_info
.tessfactors_are_def_in_all_invocs
) {
3010 vgpr
++; /* skip the tess factor LDS offset */
3011 for (unsigned i
= 0; i
< 6; i
++) {
3012 LLVMValueRef value
=
3013 LLVMBuildLoad(builder
, ctx
->invoc0_tess_factors
[i
], "");
3014 value
= ac_to_float(&ctx
->ac
, value
);
3015 ret
= LLVMBuildInsertValue(builder
, ret
, value
, vgpr
++, "");
3018 ret
= LLVMBuildInsertValue(builder
, ret
, tf_lds_offset
, vgpr
++, "");
3020 ctx
->return_value
= ret
;
3023 /* Pass TCS inputs from LS to TCS on GFX9. */
3024 static void si_set_ls_return_value_for_tcs(struct si_shader_context
*ctx
)
3026 LLVMValueRef ret
= ctx
->return_value
;
3028 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_offset
, 2);
3029 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_merged_wave_info
, 3);
3030 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_factor_offset
, 4);
3031 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_merged_scratch_offset
, 5);
3033 ret
= si_insert_input_ptr_as_2xi32(ctx
, ret
, ctx
->param_rw_buffers
,
3034 8 + SI_SGPR_RW_BUFFERS
);
3035 ret
= si_insert_input_ptr_as_2xi32(ctx
, ret
,
3036 ctx
->param_bindless_samplers_and_images
,
3037 8 + SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES
);
3039 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_vs_state_bits
,
3040 8 + SI_SGPR_VS_STATE_BITS
);
3041 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_layout
,
3042 8 + GFX9_SGPR_TCS_OFFCHIP_LAYOUT
);
3043 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_out_lds_offsets
,
3044 8 + GFX9_SGPR_TCS_OUT_OFFSETS
);
3045 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_out_lds_layout
,
3046 8 + GFX9_SGPR_TCS_OUT_LAYOUT
);
3047 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_addr_base64k
,
3048 8 + GFX9_SGPR_TCS_OFFCHIP_ADDR_BASE64K
);
3049 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_factor_addr_base64k
,
3050 8 + GFX9_SGPR_TCS_FACTOR_ADDR_BASE64K
);
3052 unsigned desc_param
= ctx
->param_tcs_factor_addr_base64k
+ 2;
3053 ret
= si_insert_input_ptr_as_2xi32(ctx
, ret
, desc_param
,
3054 8 + GFX9_SGPR_TCS_CONST_AND_SHADER_BUFFERS
);
3055 ret
= si_insert_input_ptr_as_2xi32(ctx
, ret
, desc_param
+ 1,
3056 8 + GFX9_SGPR_TCS_SAMPLERS_AND_IMAGES
);
3058 unsigned vgpr
= 8 + GFX9_TCS_NUM_USER_SGPR
;
3059 ret
= si_insert_input_ret_float(ctx
, ret
,
3060 ctx
->param_tcs_patch_id
, vgpr
++);
3061 ret
= si_insert_input_ret_float(ctx
, ret
,
3062 ctx
->param_tcs_rel_ids
, vgpr
++);
3063 ctx
->return_value
= ret
;
3066 /* Pass GS inputs from ES to GS on GFX9. */
3067 static void si_set_es_return_value_for_gs(struct si_shader_context
*ctx
)
3069 LLVMValueRef ret
= ctx
->return_value
;
3071 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_gs2vs_offset
, 2);
3072 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_merged_wave_info
, 3);
3073 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_merged_scratch_offset
, 5);
3075 ret
= si_insert_input_ptr_as_2xi32(ctx
, ret
, ctx
->param_rw_buffers
,
3076 8 + SI_SGPR_RW_BUFFERS
);
3077 ret
= si_insert_input_ptr_as_2xi32(ctx
, ret
,
3078 ctx
->param_bindless_samplers_and_images
,
3079 8 + SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES
);
3081 unsigned desc_param
= ctx
->param_vs_state_bits
+ 1;
3082 ret
= si_insert_input_ptr_as_2xi32(ctx
, ret
, desc_param
,
3083 8 + GFX9_SGPR_GS_CONST_AND_SHADER_BUFFERS
);
3084 ret
= si_insert_input_ptr_as_2xi32(ctx
, ret
, desc_param
+ 1,
3085 8 + GFX9_SGPR_GS_SAMPLERS_AND_IMAGES
);
3087 unsigned vgpr
= 8 + GFX9_GS_NUM_USER_SGPR
;
3088 for (unsigned i
= 0; i
< 5; i
++) {
3089 unsigned param
= ctx
->param_gs_vtx01_offset
+ i
;
3090 ret
= si_insert_input_ret_float(ctx
, ret
, param
, vgpr
++);
3092 ctx
->return_value
= ret
;
3095 static void si_llvm_emit_ls_epilogue(struct lp_build_tgsi_context
*bld_base
)
3097 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3098 struct si_shader
*shader
= ctx
->shader
;
3099 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
3101 LLVMValueRef vertex_id
= LLVMGetParam(ctx
->main_fn
,
3102 ctx
->param_rel_auto_id
);
3103 LLVMValueRef vertex_dw_stride
= get_tcs_in_vertex_dw_stride(ctx
);
3104 LLVMValueRef base_dw_addr
= LLVMBuildMul(ctx
->ac
.builder
, vertex_id
,
3105 vertex_dw_stride
, "");
3107 /* Write outputs to LDS. The next shader (TCS aka HS) will read
3108 * its inputs from it. */
3109 for (i
= 0; i
< info
->num_outputs
; i
++) {
3110 LLVMValueRef
*out_ptr
= ctx
->outputs
[i
];
3111 unsigned name
= info
->output_semantic_name
[i
];
3112 unsigned index
= info
->output_semantic_index
[i
];
3114 /* The ARB_shader_viewport_layer_array spec contains the
3117 * 2) What happens if gl_ViewportIndex or gl_Layer is
3118 * written in the vertex shader and a geometry shader is
3121 * RESOLVED: The value written by the last vertex processing
3122 * stage is used. If the last vertex processing stage
3123 * (vertex, tessellation evaluation or geometry) does not
3124 * statically assign to gl_ViewportIndex or gl_Layer, index
3125 * or layer zero is assumed.
3127 * So writes to those outputs in VS-as-LS are simply ignored.
3129 if (name
== TGSI_SEMANTIC_LAYER
||
3130 name
== TGSI_SEMANTIC_VIEWPORT_INDEX
)
3133 int param
= si_shader_io_get_unique_index(name
, index
);
3134 LLVMValueRef dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_dw_addr
,
3135 LLVMConstInt(ctx
->i32
, param
* 4, 0), "");
3137 for (chan
= 0; chan
< 4; chan
++) {
3138 lds_store(bld_base
, chan
, dw_addr
,
3139 LLVMBuildLoad(ctx
->ac
.builder
, out_ptr
[chan
], ""));
3143 if (ctx
->screen
->b
.chip_class
>= GFX9
)
3144 si_set_ls_return_value_for_tcs(ctx
);
3147 static void si_llvm_emit_es_epilogue(struct lp_build_tgsi_context
*bld_base
)
3149 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3150 struct si_shader
*es
= ctx
->shader
;
3151 struct tgsi_shader_info
*info
= &es
->selector
->info
;
3152 LLVMValueRef soffset
= LLVMGetParam(ctx
->main_fn
,
3153 ctx
->param_es2gs_offset
);
3154 LLVMValueRef lds_base
= NULL
;
3158 if (ctx
->screen
->b
.chip_class
>= GFX9
&& info
->num_outputs
) {
3159 unsigned itemsize_dw
= es
->selector
->esgs_itemsize
/ 4;
3160 LLVMValueRef vertex_idx
= ac_get_thread_id(&ctx
->ac
);
3161 LLVMValueRef wave_idx
= unpack_param(ctx
, ctx
->param_merged_wave_info
, 24, 4);
3162 vertex_idx
= LLVMBuildOr(ctx
->ac
.builder
, vertex_idx
,
3163 LLVMBuildMul(ctx
->ac
.builder
, wave_idx
,
3164 LLVMConstInt(ctx
->i32
, 64, false), ""), "");
3165 lds_base
= LLVMBuildMul(ctx
->ac
.builder
, vertex_idx
,
3166 LLVMConstInt(ctx
->i32
, itemsize_dw
, 0), "");
3169 for (i
= 0; i
< info
->num_outputs
; i
++) {
3170 LLVMValueRef
*out_ptr
= ctx
->outputs
[i
];
3173 if (info
->output_semantic_name
[i
] == TGSI_SEMANTIC_VIEWPORT_INDEX
||
3174 info
->output_semantic_name
[i
] == TGSI_SEMANTIC_LAYER
)
3177 param
= si_shader_io_get_unique_index(info
->output_semantic_name
[i
],
3178 info
->output_semantic_index
[i
]);
3180 for (chan
= 0; chan
< 4; chan
++) {
3181 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->ac
.builder
, out_ptr
[chan
], "");
3182 out_val
= ac_to_integer(&ctx
->ac
, out_val
);
3184 /* GFX9 has the ESGS ring in LDS. */
3185 if (ctx
->screen
->b
.chip_class
>= GFX9
) {
3186 lds_store(bld_base
, param
* 4 + chan
, lds_base
, out_val
);
3190 ac_build_buffer_store_dword(&ctx
->ac
,
3192 out_val
, 1, NULL
, soffset
,
3193 (4 * param
+ chan
) * 4,
3198 if (ctx
->screen
->b
.chip_class
>= GFX9
)
3199 si_set_es_return_value_for_gs(ctx
);
3202 static LLVMValueRef
si_get_gs_wave_id(struct si_shader_context
*ctx
)
3204 if (ctx
->screen
->b
.chip_class
>= GFX9
)
3205 return unpack_param(ctx
, ctx
->param_merged_wave_info
, 16, 8);
3207 return LLVMGetParam(ctx
->main_fn
, ctx
->param_gs_wave_id
);
3210 static void si_llvm_emit_gs_epilogue(struct lp_build_tgsi_context
*bld_base
)
3212 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3214 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_NOP
| AC_SENDMSG_GS_DONE
,
3215 si_get_gs_wave_id(ctx
));
3217 if (ctx
->screen
->b
.chip_class
>= GFX9
)
3218 lp_build_endif(&ctx
->merged_wrap_if_state
);
3221 static void si_llvm_emit_vs_epilogue(struct ac_shader_abi
*abi
,
3222 unsigned max_outputs
,
3223 LLVMValueRef
*addrs
)
3225 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
3226 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
3227 struct si_shader_output_values
*outputs
= NULL
;
3230 assert(!ctx
->shader
->is_gs_copy_shader
);
3231 assert(info
->num_outputs
<= max_outputs
);
3233 outputs
= MALLOC((info
->num_outputs
+ 1) * sizeof(outputs
[0]));
3235 /* Vertex color clamping.
3237 * This uses a state constant loaded in a user data SGPR and
3238 * an IF statement is added that clamps all colors if the constant
3241 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
3242 struct lp_build_if_state if_ctx
;
3243 LLVMValueRef cond
= NULL
;
3244 LLVMValueRef addr
, val
;
3246 for (i
= 0; i
< info
->num_outputs
; i
++) {
3247 if (info
->output_semantic_name
[i
] != TGSI_SEMANTIC_COLOR
&&
3248 info
->output_semantic_name
[i
] != TGSI_SEMANTIC_BCOLOR
)
3251 /* We've found a color. */
3253 /* The state is in the first bit of the user SGPR. */
3254 cond
= LLVMGetParam(ctx
->main_fn
,
3255 ctx
->param_vs_state_bits
);
3256 cond
= LLVMBuildTrunc(ctx
->ac
.builder
, cond
,
3258 lp_build_if(&if_ctx
, &ctx
->gallivm
, cond
);
3261 for (j
= 0; j
< 4; j
++) {
3262 addr
= addrs
[4 * i
+ j
];
3263 val
= LLVMBuildLoad(ctx
->ac
.builder
, addr
, "");
3264 val
= ac_build_clamp(&ctx
->ac
, val
);
3265 LLVMBuildStore(ctx
->ac
.builder
, val
, addr
);
3270 lp_build_endif(&if_ctx
);
3273 for (i
= 0; i
< info
->num_outputs
; i
++) {
3274 outputs
[i
].semantic_name
= info
->output_semantic_name
[i
];
3275 outputs
[i
].semantic_index
= info
->output_semantic_index
[i
];
3277 for (j
= 0; j
< 4; j
++) {
3278 outputs
[i
].values
[j
] =
3279 LLVMBuildLoad(ctx
->ac
.builder
,
3282 outputs
[i
].vertex_stream
[j
] =
3283 (info
->output_streams
[i
] >> (2 * j
)) & 3;
3287 if (ctx
->shader
->selector
->so
.num_outputs
)
3288 si_llvm_emit_streamout(ctx
, outputs
, i
, 0);
3290 /* Export PrimitiveID. */
3291 if (ctx
->shader
->key
.mono
.u
.vs_export_prim_id
) {
3292 outputs
[i
].semantic_name
= TGSI_SEMANTIC_PRIMID
;
3293 outputs
[i
].semantic_index
= 0;
3294 outputs
[i
].values
[0] = ac_to_float(&ctx
->ac
, get_primitive_id(ctx
, 0));
3295 for (j
= 1; j
< 4; j
++)
3296 outputs
[i
].values
[j
] = LLVMConstReal(ctx
->f32
, 0);
3298 memset(outputs
[i
].vertex_stream
, 0,
3299 sizeof(outputs
[i
].vertex_stream
));
3303 si_llvm_export_vs(&ctx
->bld_base
, outputs
, i
);
3307 static void si_tgsi_emit_epilogue(struct lp_build_tgsi_context
*bld_base
)
3309 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3311 ctx
->abi
.emit_outputs(&ctx
->abi
, RADEON_LLVM_MAX_OUTPUTS
,
3312 &ctx
->outputs
[0][0]);
3315 struct si_ps_exports
{
3317 struct ac_export_args args
[10];
3320 unsigned si_get_spi_shader_z_format(bool writes_z
, bool writes_stencil
,
3321 bool writes_samplemask
)
3324 /* Z needs 32 bits. */
3325 if (writes_samplemask
)
3326 return V_028710_SPI_SHADER_32_ABGR
;
3327 else if (writes_stencil
)
3328 return V_028710_SPI_SHADER_32_GR
;
3330 return V_028710_SPI_SHADER_32_R
;
3331 } else if (writes_stencil
|| writes_samplemask
) {
3332 /* Both stencil and sample mask need only 16 bits. */
3333 return V_028710_SPI_SHADER_UINT16_ABGR
;
3335 return V_028710_SPI_SHADER_ZERO
;
3339 static void si_export_mrt_z(struct lp_build_tgsi_context
*bld_base
,
3340 LLVMValueRef depth
, LLVMValueRef stencil
,
3341 LLVMValueRef samplemask
, struct si_ps_exports
*exp
)
3343 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3344 struct lp_build_context
*base
= &bld_base
->base
;
3345 struct ac_export_args args
;
3347 unsigned format
= si_get_spi_shader_z_format(depth
!= NULL
,
3349 samplemask
!= NULL
);
3351 assert(depth
|| stencil
|| samplemask
);
3353 args
.valid_mask
= 1; /* whether the EXEC mask is valid */
3354 args
.done
= 1; /* DONE bit */
3356 /* Specify the target we are exporting */
3357 args
.target
= V_008DFC_SQ_EXP_MRTZ
;
3359 args
.compr
= 0; /* COMP flag */
3360 args
.out
[0] = base
->undef
; /* R, depth */
3361 args
.out
[1] = base
->undef
; /* G, stencil test value[0:7], stencil op value[8:15] */
3362 args
.out
[2] = base
->undef
; /* B, sample mask */
3363 args
.out
[3] = base
->undef
; /* A, alpha to mask */
3365 if (format
== V_028710_SPI_SHADER_UINT16_ABGR
) {
3367 args
.compr
= 1; /* COMPR flag */
3370 /* Stencil should be in X[23:16]. */
3371 stencil
= ac_to_integer(&ctx
->ac
, stencil
);
3372 stencil
= LLVMBuildShl(ctx
->ac
.builder
, stencil
,
3373 LLVMConstInt(ctx
->i32
, 16, 0), "");
3374 args
.out
[0] = ac_to_float(&ctx
->ac
, stencil
);
3378 /* SampleMask should be in Y[15:0]. */
3379 args
.out
[1] = samplemask
;
3384 args
.out
[0] = depth
;
3388 args
.out
[1] = stencil
;
3392 args
.out
[2] = samplemask
;
3397 /* SI (except OLAND and HAINAN) has a bug that it only looks
3398 * at the X writemask component. */
3399 if (ctx
->screen
->b
.chip_class
== SI
&&
3400 ctx
->screen
->b
.family
!= CHIP_OLAND
&&
3401 ctx
->screen
->b
.family
!= CHIP_HAINAN
)
3404 /* Specify which components to enable */
3405 args
.enabled_channels
= mask
;
3407 memcpy(&exp
->args
[exp
->num
++], &args
, sizeof(args
));
3410 static void si_export_mrt_color(struct lp_build_tgsi_context
*bld_base
,
3411 LLVMValueRef
*color
, unsigned index
,
3412 unsigned samplemask_param
,
3413 bool is_last
, struct si_ps_exports
*exp
)
3415 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3419 if (ctx
->shader
->key
.part
.ps
.epilog
.clamp_color
)
3420 for (i
= 0; i
< 4; i
++)
3421 color
[i
] = ac_build_clamp(&ctx
->ac
, color
[i
]);
3424 if (ctx
->shader
->key
.part
.ps
.epilog
.alpha_to_one
)
3425 color
[3] = ctx
->ac
.f32_1
;
3429 ctx
->shader
->key
.part
.ps
.epilog
.alpha_func
!= PIPE_FUNC_ALWAYS
)
3430 si_alpha_test(bld_base
, color
[3]);
3432 /* Line & polygon smoothing */
3433 if (ctx
->shader
->key
.part
.ps
.epilog
.poly_line_smoothing
)
3434 color
[3] = si_scale_alpha_by_sample_mask(bld_base
, color
[3],
3437 /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
3438 if (ctx
->shader
->key
.part
.ps
.epilog
.last_cbuf
> 0) {
3439 struct ac_export_args args
[8];
3442 /* Get the export arguments, also find out what the last one is. */
3443 for (c
= 0; c
<= ctx
->shader
->key
.part
.ps
.epilog
.last_cbuf
; c
++) {
3444 si_llvm_init_export_args(bld_base
, color
,
3445 V_008DFC_SQ_EXP_MRT
+ c
, &args
[c
]);
3446 if (args
[c
].enabled_channels
)
3450 /* Emit all exports. */
3451 for (c
= 0; c
<= ctx
->shader
->key
.part
.ps
.epilog
.last_cbuf
; c
++) {
3452 if (is_last
&& last
== c
) {
3453 args
[c
].valid_mask
= 1; /* whether the EXEC mask is valid */
3454 args
[c
].done
= 1; /* DONE bit */
3455 } else if (!args
[c
].enabled_channels
)
3456 continue; /* unnecessary NULL export */
3458 memcpy(&exp
->args
[exp
->num
++], &args
[c
], sizeof(args
[c
]));
3461 struct ac_export_args args
;
3464 si_llvm_init_export_args(bld_base
, color
, V_008DFC_SQ_EXP_MRT
+ index
,
3467 args
.valid_mask
= 1; /* whether the EXEC mask is valid */
3468 args
.done
= 1; /* DONE bit */
3469 } else if (!args
.enabled_channels
)
3470 return; /* unnecessary NULL export */
3472 memcpy(&exp
->args
[exp
->num
++], &args
, sizeof(args
));
3476 static void si_emit_ps_exports(struct si_shader_context
*ctx
,
3477 struct si_ps_exports
*exp
)
3479 for (unsigned i
= 0; i
< exp
->num
; i
++)
3480 ac_build_export(&ctx
->ac
, &exp
->args
[i
]);
3483 static void si_export_null(struct lp_build_tgsi_context
*bld_base
)
3485 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3486 struct lp_build_context
*base
= &bld_base
->base
;
3487 struct ac_export_args args
;
3489 args
.enabled_channels
= 0x0; /* enabled channels */
3490 args
.valid_mask
= 1; /* whether the EXEC mask is valid */
3491 args
.done
= 1; /* DONE bit */
3492 args
.target
= V_008DFC_SQ_EXP_NULL
;
3493 args
.compr
= 0; /* COMPR flag (0 = 32-bit export) */
3494 args
.out
[0] = base
->undef
; /* R */
3495 args
.out
[1] = base
->undef
; /* G */
3496 args
.out
[2] = base
->undef
; /* B */
3497 args
.out
[3] = base
->undef
; /* A */
3499 ac_build_export(&ctx
->ac
, &args
);
3503 * Return PS outputs in this order:
3505 * v[0:3] = color0.xyzw
3506 * v[4:7] = color1.xyzw
3511 * vN+3 = SampleMaskIn (used for OpenGL smoothing)
3513 * The alpha-ref SGPR is returned via its original location.
3515 static void si_llvm_return_fs_outputs(struct ac_shader_abi
*abi
,
3516 unsigned max_outputs
,
3517 LLVMValueRef
*addrs
)
3519 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
3520 struct si_shader
*shader
= ctx
->shader
;
3521 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
3522 LLVMBuilderRef builder
= ctx
->ac
.builder
;
3523 unsigned i
, j
, first_vgpr
, vgpr
;
3525 LLVMValueRef color
[8][4] = {};
3526 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
3529 if (ctx
->postponed_kill
)
3530 ac_build_kill(&ctx
->ac
, LLVMBuildLoad(builder
, ctx
->postponed_kill
, ""));
3532 /* Read the output values. */
3533 for (i
= 0; i
< info
->num_outputs
; i
++) {
3534 unsigned semantic_name
= info
->output_semantic_name
[i
];
3535 unsigned semantic_index
= info
->output_semantic_index
[i
];
3537 switch (semantic_name
) {
3538 case TGSI_SEMANTIC_COLOR
:
3539 assert(semantic_index
< 8);
3540 for (j
= 0; j
< 4; j
++) {
3541 LLVMValueRef ptr
= addrs
[4 * i
+ j
];
3542 LLVMValueRef result
= LLVMBuildLoad(builder
, ptr
, "");
3543 color
[semantic_index
][j
] = result
;
3546 case TGSI_SEMANTIC_POSITION
:
3547 depth
= LLVMBuildLoad(builder
,
3548 addrs
[4 * i
+ 2], "");
3550 case TGSI_SEMANTIC_STENCIL
:
3551 stencil
= LLVMBuildLoad(builder
,
3552 addrs
[4 * i
+ 1], "");
3554 case TGSI_SEMANTIC_SAMPLEMASK
:
3555 samplemask
= LLVMBuildLoad(builder
,
3556 addrs
[4 * i
+ 0], "");
3559 fprintf(stderr
, "Warning: SI unhandled fs output type:%d\n",
3564 /* Fill the return structure. */
3565 ret
= ctx
->return_value
;
3568 ret
= LLVMBuildInsertValue(builder
, ret
,
3569 ac_to_integer(&ctx
->ac
,
3570 LLVMGetParam(ctx
->main_fn
,
3571 SI_PARAM_ALPHA_REF
)),
3572 SI_SGPR_ALPHA_REF
, "");
3575 first_vgpr
= vgpr
= SI_SGPR_ALPHA_REF
+ 1;
3576 for (i
= 0; i
< ARRAY_SIZE(color
); i
++) {
3580 for (j
= 0; j
< 4; j
++)
3581 ret
= LLVMBuildInsertValue(builder
, ret
, color
[i
][j
], vgpr
++, "");
3584 ret
= LLVMBuildInsertValue(builder
, ret
, depth
, vgpr
++, "");
3586 ret
= LLVMBuildInsertValue(builder
, ret
, stencil
, vgpr
++, "");
3588 ret
= LLVMBuildInsertValue(builder
, ret
, samplemask
, vgpr
++, "");
3590 /* Add the input sample mask for smoothing at the end. */
3591 if (vgpr
< first_vgpr
+ PS_EPILOG_SAMPLEMASK_MIN_LOC
)
3592 vgpr
= first_vgpr
+ PS_EPILOG_SAMPLEMASK_MIN_LOC
;
3593 ret
= LLVMBuildInsertValue(builder
, ret
,
3594 LLVMGetParam(ctx
->main_fn
,
3595 SI_PARAM_SAMPLE_COVERAGE
), vgpr
++, "");
3597 ctx
->return_value
= ret
;
3600 void si_emit_waitcnt(struct si_shader_context
*ctx
, unsigned simm16
)
3602 LLVMValueRef args
[1] = {
3603 LLVMConstInt(ctx
->i32
, simm16
, 0)
3605 lp_build_intrinsic(ctx
->ac
.builder
, "llvm.amdgcn.s.waitcnt",
3606 ctx
->voidt
, args
, 1, 0);
3609 static void membar_emit(
3610 const struct lp_build_tgsi_action
*action
,
3611 struct lp_build_tgsi_context
*bld_base
,
3612 struct lp_build_emit_data
*emit_data
)
3614 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3615 LLVMValueRef src0
= lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, 0);
3616 unsigned flags
= LLVMConstIntGetZExtValue(src0
);
3617 unsigned waitcnt
= NOOP_WAITCNT
;
3619 if (flags
& TGSI_MEMBAR_THREAD_GROUP
)
3620 waitcnt
&= VM_CNT
& LGKM_CNT
;
3622 if (flags
& (TGSI_MEMBAR_ATOMIC_BUFFER
|
3623 TGSI_MEMBAR_SHADER_BUFFER
|
3624 TGSI_MEMBAR_SHADER_IMAGE
))
3627 if (flags
& TGSI_MEMBAR_SHARED
)
3628 waitcnt
&= LGKM_CNT
;
3630 if (waitcnt
!= NOOP_WAITCNT
)
3631 si_emit_waitcnt(ctx
, waitcnt
);
3634 static void clock_emit(
3635 const struct lp_build_tgsi_action
*action
,
3636 struct lp_build_tgsi_context
*bld_base
,
3637 struct lp_build_emit_data
*emit_data
)
3639 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3642 tmp
= lp_build_intrinsic(ctx
->ac
.builder
, "llvm.readcyclecounter",
3643 ctx
->i64
, NULL
, 0, 0);
3644 tmp
= LLVMBuildBitCast(ctx
->ac
.builder
, tmp
, ctx
->v2i32
, "");
3646 emit_data
->output
[0] =
3647 LLVMBuildExtractElement(ctx
->ac
.builder
, tmp
, ctx
->i32_0
, "");
3648 emit_data
->output
[1] =
3649 LLVMBuildExtractElement(ctx
->ac
.builder
, tmp
, ctx
->i32_1
, "");
3652 LLVMTypeRef
si_const_array(LLVMTypeRef elem_type
, int num_elements
)
3654 return LLVMPointerType(LLVMArrayType(elem_type
, num_elements
),
3658 static void si_llvm_emit_ddxy(
3659 const struct lp_build_tgsi_action
*action
,
3660 struct lp_build_tgsi_context
*bld_base
,
3661 struct lp_build_emit_data
*emit_data
)
3663 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3664 unsigned opcode
= emit_data
->info
->opcode
;
3669 if (opcode
== TGSI_OPCODE_DDX_FINE
)
3670 mask
= AC_TID_MASK_LEFT
;
3671 else if (opcode
== TGSI_OPCODE_DDY_FINE
)
3672 mask
= AC_TID_MASK_TOP
;
3674 mask
= AC_TID_MASK_TOP_LEFT
;
3676 /* for DDX we want to next X pixel, DDY next Y pixel. */
3677 idx
= (opcode
== TGSI_OPCODE_DDX
|| opcode
== TGSI_OPCODE_DDX_FINE
) ? 1 : 2;
3679 val
= ac_to_integer(&ctx
->ac
, emit_data
->args
[0]);
3680 val
= ac_build_ddxy(&ctx
->ac
, mask
, idx
, val
);
3681 emit_data
->output
[emit_data
->chan
] = val
;
3685 * this takes an I,J coordinate pair,
3686 * and works out the X and Y derivatives.
3687 * it returns DDX(I), DDX(J), DDY(I), DDY(J).
3689 static LLVMValueRef
si_llvm_emit_ddxy_interp(
3690 struct lp_build_tgsi_context
*bld_base
,
3691 LLVMValueRef interp_ij
)
3693 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3694 LLVMValueRef result
[4], a
;
3697 for (i
= 0; i
< 2; i
++) {
3698 a
= LLVMBuildExtractElement(ctx
->ac
.builder
, interp_ij
,
3699 LLVMConstInt(ctx
->i32
, i
, 0), "");
3700 result
[i
] = lp_build_emit_llvm_unary(bld_base
, TGSI_OPCODE_DDX
, a
);
3701 result
[2+i
] = lp_build_emit_llvm_unary(bld_base
, TGSI_OPCODE_DDY
, a
);
3704 return lp_build_gather_values(&ctx
->gallivm
, result
, 4);
3707 static void interp_fetch_args(
3708 struct lp_build_tgsi_context
*bld_base
,
3709 struct lp_build_emit_data
*emit_data
)
3711 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3712 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3714 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_OFFSET
) {
3715 /* offset is in second src, first two channels */
3716 emit_data
->args
[0] = lp_build_emit_fetch(bld_base
,
3719 emit_data
->args
[1] = lp_build_emit_fetch(bld_base
,
3722 emit_data
->arg_count
= 2;
3723 } else if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_SAMPLE
) {
3724 LLVMValueRef sample_position
;
3725 LLVMValueRef sample_id
;
3726 LLVMValueRef halfval
= LLVMConstReal(ctx
->f32
, 0.5f
);
3728 /* fetch sample ID, then fetch its sample position,
3729 * and place into first two channels.
3731 sample_id
= lp_build_emit_fetch(bld_base
,
3732 emit_data
->inst
, 1, TGSI_CHAN_X
);
3733 sample_id
= ac_to_integer(&ctx
->ac
, sample_id
);
3735 /* Section 8.13.2 (Interpolation Functions) of the OpenGL Shading
3736 * Language 4.50 spec says about interpolateAtSample:
3738 * "Returns the value of the input interpolant variable at
3739 * the location of sample number sample. If multisample
3740 * buffers are not available, the input variable will be
3741 * evaluated at the center of the pixel. If sample sample
3742 * does not exist, the position used to interpolate the
3743 * input variable is undefined."
3745 * This means that sample_id values outside of the valid are
3746 * in fact valid input, and the usual mechanism for loading the
3747 * sample position doesn't work.
3749 if (ctx
->shader
->key
.mono
.u
.ps
.interpolate_at_sample_force_center
) {
3750 LLVMValueRef center
[4] = {
3751 LLVMConstReal(ctx
->f32
, 0.5),
3752 LLVMConstReal(ctx
->f32
, 0.5),
3757 sample_position
= lp_build_gather_values(&ctx
->gallivm
, center
, 4);
3759 sample_position
= load_sample_position(ctx
, sample_id
);
3762 emit_data
->args
[0] = LLVMBuildExtractElement(ctx
->ac
.builder
,
3766 emit_data
->args
[0] = LLVMBuildFSub(ctx
->ac
.builder
, emit_data
->args
[0], halfval
, "");
3767 emit_data
->args
[1] = LLVMBuildExtractElement(ctx
->ac
.builder
,
3770 emit_data
->args
[1] = LLVMBuildFSub(ctx
->ac
.builder
, emit_data
->args
[1], halfval
, "");
3771 emit_data
->arg_count
= 2;
3775 static void build_interp_intrinsic(const struct lp_build_tgsi_action
*action
,
3776 struct lp_build_tgsi_context
*bld_base
,
3777 struct lp_build_emit_data
*emit_data
)
3779 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3780 struct si_shader
*shader
= ctx
->shader
;
3781 const struct tgsi_shader_info
*info
= &shader
->selector
->info
;
3782 LLVMValueRef interp_param
;
3783 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3784 const struct tgsi_full_src_register
*input
= &inst
->Src
[0];
3785 int input_base
, input_array_size
;
3788 LLVMValueRef prim_mask
= LLVMGetParam(ctx
->main_fn
, SI_PARAM_PRIM_MASK
);
3789 LLVMValueRef array_idx
;
3790 int interp_param_idx
;
3794 assert(input
->Register
.File
== TGSI_FILE_INPUT
);
3796 if (input
->Register
.Indirect
) {
3797 unsigned array_id
= input
->Indirect
.ArrayID
;
3800 input_base
= info
->input_array_first
[array_id
];
3801 input_array_size
= info
->input_array_last
[array_id
] - input_base
+ 1;
3803 input_base
= inst
->Src
[0].Register
.Index
;
3804 input_array_size
= info
->num_inputs
- input_base
;
3807 array_idx
= si_get_indirect_index(ctx
, &input
->Indirect
,
3808 1, input
->Register
.Index
- input_base
);
3810 input_base
= inst
->Src
[0].Register
.Index
;
3811 input_array_size
= 1;
3812 array_idx
= ctx
->i32_0
;
3815 interp
= shader
->selector
->info
.input_interpolate
[input_base
];
3817 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_OFFSET
||
3818 inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_SAMPLE
)
3819 location
= TGSI_INTERPOLATE_LOC_CENTER
;
3821 location
= TGSI_INTERPOLATE_LOC_CENTROID
;
3823 interp_param_idx
= lookup_interp_param_index(interp
, location
);
3824 if (interp_param_idx
== -1)
3826 else if (interp_param_idx
)
3827 interp_param
= LLVMGetParam(ctx
->main_fn
, interp_param_idx
);
3829 interp_param
= NULL
;
3831 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_OFFSET
||
3832 inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_SAMPLE
) {
3833 LLVMValueRef ij_out
[2];
3834 LLVMValueRef ddxy_out
= si_llvm_emit_ddxy_interp(bld_base
, interp_param
);
3837 * take the I then J parameters, and the DDX/Y for it, and
3838 * calculate the IJ inputs for the interpolator.
3839 * temp1 = ddx * offset/sample.x + I;
3840 * interp_param.I = ddy * offset/sample.y + temp1;
3841 * temp1 = ddx * offset/sample.x + J;
3842 * interp_param.J = ddy * offset/sample.y + temp1;
3844 for (i
= 0; i
< 2; i
++) {
3845 LLVMValueRef ix_ll
= LLVMConstInt(ctx
->i32
, i
, 0);
3846 LLVMValueRef iy_ll
= LLVMConstInt(ctx
->i32
, i
+ 2, 0);
3847 LLVMValueRef ddx_el
= LLVMBuildExtractElement(ctx
->ac
.builder
,
3848 ddxy_out
, ix_ll
, "");
3849 LLVMValueRef ddy_el
= LLVMBuildExtractElement(ctx
->ac
.builder
,
3850 ddxy_out
, iy_ll
, "");
3851 LLVMValueRef interp_el
= LLVMBuildExtractElement(ctx
->ac
.builder
,
3852 interp_param
, ix_ll
, "");
3853 LLVMValueRef temp1
, temp2
;
3855 interp_el
= ac_to_float(&ctx
->ac
, interp_el
);
3857 temp1
= LLVMBuildFMul(ctx
->ac
.builder
, ddx_el
, emit_data
->args
[0], "");
3859 temp1
= LLVMBuildFAdd(ctx
->ac
.builder
, temp1
, interp_el
, "");
3861 temp2
= LLVMBuildFMul(ctx
->ac
.builder
, ddy_el
, emit_data
->args
[1], "");
3863 ij_out
[i
] = LLVMBuildFAdd(ctx
->ac
.builder
, temp2
, temp1
, "");
3865 interp_param
= lp_build_gather_values(&ctx
->gallivm
, ij_out
, 2);
3869 interp_param
= ac_to_float(&ctx
->ac
, interp_param
);
3871 for (chan
= 0; chan
< 4; chan
++) {
3872 LLVMValueRef gather
= LLVMGetUndef(LLVMVectorType(ctx
->f32
, input_array_size
));
3873 unsigned schan
= tgsi_util_get_full_src_register_swizzle(&inst
->Src
[0], chan
);
3875 for (unsigned idx
= 0; idx
< input_array_size
; ++idx
) {
3876 LLVMValueRef v
, i
= NULL
, j
= NULL
;
3879 i
= LLVMBuildExtractElement(
3880 ctx
->ac
.builder
, interp_param
, ctx
->i32_0
, "");
3881 j
= LLVMBuildExtractElement(
3882 ctx
->ac
.builder
, interp_param
, ctx
->i32_1
, "");
3884 v
= si_build_fs_interp(ctx
, input_base
+ idx
, schan
,
3887 gather
= LLVMBuildInsertElement(ctx
->ac
.builder
,
3888 gather
, v
, LLVMConstInt(ctx
->i32
, idx
, false), "");
3891 emit_data
->output
[chan
] = LLVMBuildExtractElement(
3892 ctx
->ac
.builder
, gather
, array_idx
, "");
3896 static void vote_all_emit(
3897 const struct lp_build_tgsi_action
*action
,
3898 struct lp_build_tgsi_context
*bld_base
,
3899 struct lp_build_emit_data
*emit_data
)
3901 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3903 LLVMValueRef tmp
= ac_build_vote_all(&ctx
->ac
, emit_data
->args
[0]);
3904 emit_data
->output
[emit_data
->chan
] =
3905 LLVMBuildSExt(ctx
->ac
.builder
, tmp
, ctx
->i32
, "");
3908 static void vote_any_emit(
3909 const struct lp_build_tgsi_action
*action
,
3910 struct lp_build_tgsi_context
*bld_base
,
3911 struct lp_build_emit_data
*emit_data
)
3913 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3915 LLVMValueRef tmp
= ac_build_vote_any(&ctx
->ac
, emit_data
->args
[0]);
3916 emit_data
->output
[emit_data
->chan
] =
3917 LLVMBuildSExt(ctx
->ac
.builder
, tmp
, ctx
->i32
, "");
3920 static void vote_eq_emit(
3921 const struct lp_build_tgsi_action
*action
,
3922 struct lp_build_tgsi_context
*bld_base
,
3923 struct lp_build_emit_data
*emit_data
)
3925 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3927 LLVMValueRef tmp
= ac_build_vote_eq(&ctx
->ac
, emit_data
->args
[0]);
3928 emit_data
->output
[emit_data
->chan
] =
3929 LLVMBuildSExt(ctx
->ac
.builder
, tmp
, ctx
->i32
, "");
3932 static void ballot_emit(
3933 const struct lp_build_tgsi_action
*action
,
3934 struct lp_build_tgsi_context
*bld_base
,
3935 struct lp_build_emit_data
*emit_data
)
3937 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3938 LLVMBuilderRef builder
= ctx
->ac
.builder
;
3941 tmp
= lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, TGSI_CHAN_X
);
3942 tmp
= ac_build_ballot(&ctx
->ac
, tmp
);
3943 tmp
= LLVMBuildBitCast(builder
, tmp
, ctx
->v2i32
, "");
3945 emit_data
->output
[0] = LLVMBuildExtractElement(builder
, tmp
, ctx
->i32_0
, "");
3946 emit_data
->output
[1] = LLVMBuildExtractElement(builder
, tmp
, ctx
->i32_1
, "");
3949 static void read_invoc_fetch_args(
3950 struct lp_build_tgsi_context
*bld_base
,
3951 struct lp_build_emit_data
*emit_data
)
3953 emit_data
->args
[0] = lp_build_emit_fetch(bld_base
, emit_data
->inst
,
3954 0, emit_data
->src_chan
);
3956 /* Always read the source invocation (= lane) from the X channel. */
3957 emit_data
->args
[1] = lp_build_emit_fetch(bld_base
, emit_data
->inst
,
3959 emit_data
->arg_count
= 2;
3962 static void read_lane_emit(
3963 const struct lp_build_tgsi_action
*action
,
3964 struct lp_build_tgsi_context
*bld_base
,
3965 struct lp_build_emit_data
*emit_data
)
3967 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3969 /* We currently have no other way to prevent LLVM from lifting the icmp
3970 * calls to a dominating basic block.
3972 ac_build_optimization_barrier(&ctx
->ac
, &emit_data
->args
[0]);
3974 for (unsigned i
= 0; i
< emit_data
->arg_count
; ++i
)
3975 emit_data
->args
[i
] = ac_to_integer(&ctx
->ac
, emit_data
->args
[i
]);
3977 emit_data
->output
[emit_data
->chan
] =
3978 ac_build_intrinsic(&ctx
->ac
, action
->intr_name
,
3979 ctx
->i32
, emit_data
->args
, emit_data
->arg_count
,
3980 AC_FUNC_ATTR_READNONE
|
3981 AC_FUNC_ATTR_CONVERGENT
);
3984 static unsigned si_llvm_get_stream(struct lp_build_tgsi_context
*bld_base
,
3985 struct lp_build_emit_data
*emit_data
)
3987 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3988 struct tgsi_src_register src0
= emit_data
->inst
->Src
[0].Register
;
3992 assert(src0
.File
== TGSI_FILE_IMMEDIATE
);
3994 imm
= ctx
->imms
[src0
.Index
* TGSI_NUM_CHANNELS
+ src0
.SwizzleX
];
3995 stream
= LLVMConstIntGetZExtValue(imm
) & 0x3;
3999 /* Emit one vertex from the geometry shader */
4000 static void si_llvm_emit_vertex(
4001 const struct lp_build_tgsi_action
*action
,
4002 struct lp_build_tgsi_context
*bld_base
,
4003 struct lp_build_emit_data
*emit_data
)
4005 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4006 struct lp_build_context
*uint
= &bld_base
->uint_bld
;
4007 struct si_shader
*shader
= ctx
->shader
;
4008 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
4009 struct lp_build_if_state if_state
;
4010 LLVMValueRef soffset
= LLVMGetParam(ctx
->main_fn
,
4011 ctx
->param_gs2vs_offset
);
4012 LLVMValueRef gs_next_vertex
;
4013 LLVMValueRef can_emit
, kill
;
4014 unsigned chan
, offset
;
4018 stream
= si_llvm_get_stream(bld_base
, emit_data
);
4020 /* Write vertex attribute values to GSVS ring */
4021 gs_next_vertex
= LLVMBuildLoad(ctx
->ac
.builder
,
4022 ctx
->gs_next_vertex
[stream
],
4025 /* If this thread has already emitted the declared maximum number of
4026 * vertices, skip the write: excessive vertex emissions are not
4027 * supposed to have any effect.
4029 * If the shader has no writes to memory, kill it instead. This skips
4030 * further memory loads and may allow LLVM to skip to the end
4033 can_emit
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntULT
, gs_next_vertex
,
4034 LLVMConstInt(ctx
->i32
,
4035 shader
->selector
->gs_max_out_vertices
, 0), "");
4037 bool use_kill
= !info
->writes_memory
;
4039 kill
= lp_build_select(&bld_base
->base
, can_emit
,
4040 LLVMConstReal(ctx
->f32
, 1.0f
),
4041 LLVMConstReal(ctx
->f32
, -1.0f
));
4043 ac_build_kill(&ctx
->ac
, kill
);
4045 lp_build_if(&if_state
, &ctx
->gallivm
, can_emit
);
4049 for (i
= 0; i
< info
->num_outputs
; i
++) {
4050 LLVMValueRef
*out_ptr
= ctx
->outputs
[i
];
4052 for (chan
= 0; chan
< 4; chan
++) {
4053 if (!(info
->output_usagemask
[i
] & (1 << chan
)) ||
4054 ((info
->output_streams
[i
] >> (2 * chan
)) & 3) != stream
)
4057 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->ac
.builder
, out_ptr
[chan
], "");
4058 LLVMValueRef voffset
=
4059 LLVMConstInt(ctx
->i32
, offset
*
4060 shader
->selector
->gs_max_out_vertices
, 0);
4063 voffset
= lp_build_add(uint
, voffset
, gs_next_vertex
);
4064 voffset
= lp_build_mul_imm(uint
, voffset
, 4);
4066 out_val
= ac_to_integer(&ctx
->ac
, out_val
);
4068 ac_build_buffer_store_dword(&ctx
->ac
,
4069 ctx
->gsvs_ring
[stream
],
4071 voffset
, soffset
, 0,
4076 gs_next_vertex
= lp_build_add(uint
, gs_next_vertex
,
4079 LLVMBuildStore(ctx
->ac
.builder
, gs_next_vertex
, ctx
->gs_next_vertex
[stream
]);
4081 /* Signal vertex emission */
4082 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_EMIT
| AC_SENDMSG_GS
| (stream
<< 8),
4083 si_get_gs_wave_id(ctx
));
4085 lp_build_endif(&if_state
);
4088 /* Cut one primitive from the geometry shader */
4089 static void si_llvm_emit_primitive(
4090 const struct lp_build_tgsi_action
*action
,
4091 struct lp_build_tgsi_context
*bld_base
,
4092 struct lp_build_emit_data
*emit_data
)
4094 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4097 /* Signal primitive cut */
4098 stream
= si_llvm_get_stream(bld_base
, emit_data
);
4099 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_CUT
| AC_SENDMSG_GS
| (stream
<< 8),
4100 si_get_gs_wave_id(ctx
));
4103 static void si_llvm_emit_barrier(const struct lp_build_tgsi_action
*action
,
4104 struct lp_build_tgsi_context
*bld_base
,
4105 struct lp_build_emit_data
*emit_data
)
4107 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4109 /* SI only (thanks to a hw bug workaround):
4110 * The real barrier instruction isn’t needed, because an entire patch
4111 * always fits into a single wave.
4113 if (ctx
->screen
->b
.chip_class
== SI
&&
4114 ctx
->type
== PIPE_SHADER_TESS_CTRL
) {
4115 si_emit_waitcnt(ctx
, LGKM_CNT
& VM_CNT
);
4119 lp_build_intrinsic(ctx
->ac
.builder
,
4120 "llvm.amdgcn.s.barrier",
4121 ctx
->voidt
, NULL
, 0, LP_FUNC_ATTR_CONVERGENT
);
4124 static const struct lp_build_tgsi_action interp_action
= {
4125 .fetch_args
= interp_fetch_args
,
4126 .emit
= build_interp_intrinsic
,
4129 static void si_create_function(struct si_shader_context
*ctx
,
4131 LLVMTypeRef
*returns
, unsigned num_returns
,
4132 struct si_function_info
*fninfo
,
4133 unsigned max_workgroup_size
)
4137 si_llvm_create_func(ctx
, name
, returns
, num_returns
,
4138 fninfo
->types
, fninfo
->num_params
);
4139 ctx
->return_value
= LLVMGetUndef(ctx
->return_type
);
4141 for (i
= 0; i
< fninfo
->num_sgpr_params
; ++i
) {
4142 LLVMValueRef P
= LLVMGetParam(ctx
->main_fn
, i
);
4144 /* The combination of:
4148 * allows the optimization passes to move loads and reduces
4149 * SGPR spilling significantly.
4151 if (LLVMGetTypeKind(LLVMTypeOf(P
)) == LLVMPointerTypeKind
) {
4152 lp_add_function_attr(ctx
->main_fn
, i
+ 1, LP_FUNC_ATTR_BYVAL
);
4153 lp_add_function_attr(ctx
->main_fn
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
4154 ac_add_attr_dereferenceable(P
, UINT64_MAX
);
4156 lp_add_function_attr(ctx
->main_fn
, i
+ 1, LP_FUNC_ATTR_INREG
);
4159 for (i
= 0; i
< fninfo
->num_params
; ++i
) {
4160 if (fninfo
->assign
[i
])
4161 *fninfo
->assign
[i
] = LLVMGetParam(ctx
->main_fn
, i
);
4164 if (max_workgroup_size
) {
4165 si_llvm_add_attribute(ctx
->main_fn
, "amdgpu-max-work-group-size",
4166 max_workgroup_size
);
4168 LLVMAddTargetDependentFunctionAttr(ctx
->main_fn
,
4169 "no-signed-zeros-fp-math",
4172 if (ctx
->screen
->b
.debug_flags
& DBG(UNSAFE_MATH
)) {
4173 /* These were copied from some LLVM test. */
4174 LLVMAddTargetDependentFunctionAttr(ctx
->main_fn
,
4175 "less-precise-fpmad",
4177 LLVMAddTargetDependentFunctionAttr(ctx
->main_fn
,
4180 LLVMAddTargetDependentFunctionAttr(ctx
->main_fn
,
4183 LLVMAddTargetDependentFunctionAttr(ctx
->main_fn
,
4189 static void declare_streamout_params(struct si_shader_context
*ctx
,
4190 struct pipe_stream_output_info
*so
,
4191 struct si_function_info
*fninfo
)
4195 /* Streamout SGPRs. */
4196 if (so
->num_outputs
) {
4197 if (ctx
->type
!= PIPE_SHADER_TESS_EVAL
)
4198 ctx
->param_streamout_config
= add_arg(fninfo
, ARG_SGPR
, ctx
->ac
.i32
);
4200 ctx
->param_streamout_config
= fninfo
->num_params
- 1;
4202 ctx
->param_streamout_write_index
= add_arg(fninfo
, ARG_SGPR
, ctx
->ac
.i32
);
4204 /* A streamout buffer offset is loaded if the stride is non-zero. */
4205 for (i
= 0; i
< 4; i
++) {
4209 ctx
->param_streamout_offset
[i
] = add_arg(fninfo
, ARG_SGPR
, ctx
->ac
.i32
);
4213 static void declare_lds_as_pointer(struct si_shader_context
*ctx
)
4215 unsigned lds_size
= ctx
->screen
->b
.chip_class
>= CIK
? 65536 : 32768;
4216 ctx
->lds
= LLVMBuildIntToPtr(ctx
->ac
.builder
, ctx
->i32_0
,
4217 LLVMPointerType(LLVMArrayType(ctx
->i32
, lds_size
/ 4), LOCAL_ADDR_SPACE
),
4221 static unsigned si_get_max_workgroup_size(const struct si_shader
*shader
)
4223 switch (shader
->selector
->type
) {
4224 case PIPE_SHADER_TESS_CTRL
:
4225 /* Return this so that LLVM doesn't remove s_barrier
4226 * instructions on chips where we use s_barrier. */
4227 return shader
->selector
->screen
->b
.chip_class
>= CIK
? 128 : 64;
4229 case PIPE_SHADER_GEOMETRY
:
4230 return shader
->selector
->screen
->b
.chip_class
>= GFX9
? 128 : 64;
4232 case PIPE_SHADER_COMPUTE
:
4233 break; /* see below */
4239 const unsigned *properties
= shader
->selector
->info
.properties
;
4240 unsigned max_work_group_size
=
4241 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
] *
4242 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT
] *
4243 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH
];
4245 if (!max_work_group_size
) {
4246 /* This is a variable group size compute shader,
4247 * compile it for the maximum possible group size.
4249 max_work_group_size
= SI_MAX_VARIABLE_THREADS_PER_BLOCK
;
4251 return max_work_group_size
;
4254 static void declare_per_stage_desc_pointers(struct si_shader_context
*ctx
,
4255 struct si_function_info
*fninfo
,
4258 unsigned const_and_shader_buffers
=
4259 add_arg(fninfo
, ARG_SGPR
,
4260 si_const_array(ctx
->v4i32
,
4261 SI_NUM_SHADER_BUFFERS
+ SI_NUM_CONST_BUFFERS
));
4262 unsigned samplers_and_images
=
4263 add_arg(fninfo
, ARG_SGPR
,
4264 si_const_array(ctx
->v8i32
,
4265 SI_NUM_IMAGES
+ SI_NUM_SAMPLERS
* 2));
4267 if (assign_params
) {
4268 ctx
->param_const_and_shader_buffers
= const_and_shader_buffers
;
4269 ctx
->param_samplers_and_images
= samplers_and_images
;
4273 static void declare_global_desc_pointers(struct si_shader_context
*ctx
,
4274 struct si_function_info
*fninfo
)
4276 ctx
->param_rw_buffers
= add_arg(fninfo
, ARG_SGPR
,
4277 si_const_array(ctx
->v4i32
, SI_NUM_RW_BUFFERS
));
4278 ctx
->param_bindless_samplers_and_images
= add_arg(fninfo
, ARG_SGPR
,
4279 si_const_array(ctx
->v8i32
, 0));
4282 static void declare_vs_specific_input_sgprs(struct si_shader_context
*ctx
,
4283 struct si_function_info
*fninfo
)
4285 ctx
->param_vertex_buffers
= add_arg(fninfo
, ARG_SGPR
,
4286 si_const_array(ctx
->v4i32
, SI_NUM_VERTEX_BUFFERS
));
4287 add_arg_assign(fninfo
, ARG_SGPR
, ctx
->i32
, &ctx
->abi
.base_vertex
);
4288 add_arg_assign(fninfo
, ARG_SGPR
, ctx
->i32
, &ctx
->abi
.start_instance
);
4289 add_arg_assign(fninfo
, ARG_SGPR
, ctx
->i32
, &ctx
->abi
.draw_id
);
4290 ctx
->param_vs_state_bits
= add_arg(fninfo
, ARG_SGPR
, ctx
->i32
);
4293 static void declare_vs_input_vgprs(struct si_shader_context
*ctx
,
4294 struct si_function_info
*fninfo
,
4295 unsigned *num_prolog_vgprs
)
4297 struct si_shader
*shader
= ctx
->shader
;
4299 add_arg_assign(fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.vertex_id
);
4300 if (shader
->key
.as_ls
) {
4301 ctx
->param_rel_auto_id
= add_arg(fninfo
, ARG_VGPR
, ctx
->i32
);
4302 add_arg_assign(fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.instance_id
);
4304 add_arg_assign(fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.instance_id
);
4305 ctx
->param_vs_prim_id
= add_arg(fninfo
, ARG_VGPR
, ctx
->i32
);
4307 add_arg(fninfo
, ARG_VGPR
, ctx
->i32
); /* unused */
4309 if (!shader
->is_gs_copy_shader
) {
4310 /* Vertex load indices. */
4311 ctx
->param_vertex_index0
= fninfo
->num_params
;
4312 for (unsigned i
= 0; i
< shader
->selector
->info
.num_inputs
; i
++)
4313 add_arg(fninfo
, ARG_VGPR
, ctx
->i32
);
4314 *num_prolog_vgprs
+= shader
->selector
->info
.num_inputs
;
4318 static void declare_tes_input_vgprs(struct si_shader_context
*ctx
,
4319 struct si_function_info
*fninfo
)
4321 ctx
->param_tes_u
= add_arg(fninfo
, ARG_VGPR
, ctx
->f32
);
4322 ctx
->param_tes_v
= add_arg(fninfo
, ARG_VGPR
, ctx
->f32
);
4323 ctx
->param_tes_rel_patch_id
= add_arg(fninfo
, ARG_VGPR
, ctx
->i32
);
4324 ctx
->param_tes_patch_id
= add_arg(fninfo
, ARG_VGPR
, ctx
->i32
);
4328 /* Convenient merged shader definitions. */
4329 SI_SHADER_MERGED_VERTEX_TESSCTRL
= PIPE_SHADER_TYPES
,
4330 SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY
,
4333 static void create_function(struct si_shader_context
*ctx
)
4335 struct si_shader
*shader
= ctx
->shader
;
4336 struct si_function_info fninfo
;
4337 LLVMTypeRef returns
[16+32*4];
4338 unsigned i
, num_return_sgprs
;
4339 unsigned num_returns
= 0;
4340 unsigned num_prolog_vgprs
= 0;
4341 unsigned type
= ctx
->type
;
4342 unsigned vs_blit_property
=
4343 shader
->selector
->info
.properties
[TGSI_PROPERTY_VS_BLIT_SGPRS
];
4345 si_init_function_info(&fninfo
);
4347 /* Set MERGED shaders. */
4348 if (ctx
->screen
->b
.chip_class
>= GFX9
) {
4349 if (shader
->key
.as_ls
|| type
== PIPE_SHADER_TESS_CTRL
)
4350 type
= SI_SHADER_MERGED_VERTEX_TESSCTRL
; /* LS or HS */
4351 else if (shader
->key
.as_es
|| type
== PIPE_SHADER_GEOMETRY
)
4352 type
= SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY
;
4355 LLVMTypeRef v3i32
= LLVMVectorType(ctx
->i32
, 3);
4358 case PIPE_SHADER_VERTEX
:
4359 declare_global_desc_pointers(ctx
, &fninfo
);
4361 if (vs_blit_property
) {
4362 ctx
->param_vs_blit_inputs
= fninfo
.num_params
;
4363 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* i16 x1, y1 */
4364 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* i16 x2, y2 */
4365 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* depth */
4367 if (vs_blit_property
== SI_VS_BLIT_SGPRS_POS_COLOR
) {
4368 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* color0 */
4369 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* color1 */
4370 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* color2 */
4371 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* color3 */
4372 } else if (vs_blit_property
== SI_VS_BLIT_SGPRS_POS_TEXCOORD
) {
4373 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* texcoord.x1 */
4374 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* texcoord.y1 */
4375 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* texcoord.x2 */
4376 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* texcoord.y2 */
4377 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* texcoord.z */
4378 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* texcoord.w */
4382 declare_vs_input_vgprs(ctx
, &fninfo
, &num_prolog_vgprs
);
4386 declare_per_stage_desc_pointers(ctx
, &fninfo
, true);
4387 declare_vs_specific_input_sgprs(ctx
, &fninfo
);
4389 if (shader
->key
.as_es
) {
4390 assert(!shader
->selector
->nir
);
4391 ctx
->param_es2gs_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4392 } else if (shader
->key
.as_ls
) {
4393 assert(!shader
->selector
->nir
);
4394 /* no extra parameters */
4396 if (shader
->is_gs_copy_shader
) {
4397 fninfo
.num_params
= ctx
->param_rw_buffers
+ 1;
4398 fninfo
.num_sgpr_params
= fninfo
.num_params
;
4401 /* The locations of the other parameters are assigned dynamically. */
4402 declare_streamout_params(ctx
, &shader
->selector
->so
,
4407 declare_vs_input_vgprs(ctx
, &fninfo
, &num_prolog_vgprs
);
4410 case PIPE_SHADER_TESS_CTRL
: /* SI-CI-VI */
4411 declare_global_desc_pointers(ctx
, &fninfo
);
4412 declare_per_stage_desc_pointers(ctx
, &fninfo
, true);
4413 ctx
->param_tcs_offchip_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4414 ctx
->param_tcs_out_lds_offsets
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4415 ctx
->param_tcs_out_lds_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4416 ctx
->param_vs_state_bits
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4417 ctx
->param_tcs_offchip_addr_base64k
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4418 ctx
->param_tcs_factor_addr_base64k
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4419 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4420 ctx
->param_tcs_factor_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4423 ctx
->param_tcs_patch_id
= add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
4424 ctx
->param_tcs_rel_ids
= add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
4426 /* param_tcs_offchip_offset and param_tcs_factor_offset are
4427 * placed after the user SGPRs.
4429 for (i
= 0; i
< GFX6_TCS_NUM_USER_SGPR
+ 2; i
++)
4430 returns
[num_returns
++] = ctx
->i32
; /* SGPRs */
4431 for (i
= 0; i
< 11; i
++)
4432 returns
[num_returns
++] = ctx
->f32
; /* VGPRs */
4435 case SI_SHADER_MERGED_VERTEX_TESSCTRL
:
4436 /* Merged stages have 8 system SGPRs at the beginning. */
4437 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* SPI_SHADER_USER_DATA_ADDR_LO_HS */
4438 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* SPI_SHADER_USER_DATA_ADDR_HI_HS */
4439 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4440 ctx
->param_merged_wave_info
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4441 ctx
->param_tcs_factor_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4442 ctx
->param_merged_scratch_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4443 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused */
4444 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused */
4446 declare_global_desc_pointers(ctx
, &fninfo
);
4447 declare_per_stage_desc_pointers(ctx
, &fninfo
,
4448 ctx
->type
== PIPE_SHADER_VERTEX
);
4449 declare_vs_specific_input_sgprs(ctx
, &fninfo
);
4451 ctx
->param_tcs_offchip_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4452 ctx
->param_tcs_out_lds_offsets
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4453 ctx
->param_tcs_out_lds_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4454 ctx
->param_tcs_offchip_addr_base64k
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4455 ctx
->param_tcs_factor_addr_base64k
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4456 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused */
4458 declare_per_stage_desc_pointers(ctx
, &fninfo
,
4459 ctx
->type
== PIPE_SHADER_TESS_CTRL
);
4461 /* VGPRs (first TCS, then VS) */
4462 ctx
->param_tcs_patch_id
= add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
4463 ctx
->param_tcs_rel_ids
= add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
4465 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
4466 declare_vs_input_vgprs(ctx
, &fninfo
,
4469 /* LS return values are inputs to the TCS main shader part. */
4470 for (i
= 0; i
< 8 + GFX9_TCS_NUM_USER_SGPR
; i
++)
4471 returns
[num_returns
++] = ctx
->i32
; /* SGPRs */
4472 for (i
= 0; i
< 2; i
++)
4473 returns
[num_returns
++] = ctx
->f32
; /* VGPRs */
4475 /* TCS return values are inputs to the TCS epilog.
4477 * param_tcs_offchip_offset, param_tcs_factor_offset,
4478 * param_tcs_offchip_layout, and param_rw_buffers
4479 * should be passed to the epilog.
4481 for (i
= 0; i
<= 8 + GFX9_SGPR_TCS_FACTOR_ADDR_BASE64K
; i
++)
4482 returns
[num_returns
++] = ctx
->i32
; /* SGPRs */
4483 for (i
= 0; i
< 11; i
++)
4484 returns
[num_returns
++] = ctx
->f32
; /* VGPRs */
4488 case SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY
:
4489 /* Merged stages have 8 system SGPRs at the beginning. */
4490 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused (SPI_SHADER_USER_DATA_ADDR_LO_GS) */
4491 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused (SPI_SHADER_USER_DATA_ADDR_HI_GS) */
4492 ctx
->param_gs2vs_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4493 ctx
->param_merged_wave_info
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4494 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4495 ctx
->param_merged_scratch_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4496 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused (SPI_SHADER_PGM_LO/HI_GS << 8) */
4497 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused (SPI_SHADER_PGM_LO/HI_GS >> 24) */
4499 declare_global_desc_pointers(ctx
, &fninfo
);
4500 declare_per_stage_desc_pointers(ctx
, &fninfo
,
4501 (ctx
->type
== PIPE_SHADER_VERTEX
||
4502 ctx
->type
== PIPE_SHADER_TESS_EVAL
));
4503 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
4504 declare_vs_specific_input_sgprs(ctx
, &fninfo
);
4506 /* TESS_EVAL (and also GEOMETRY):
4507 * Declare as many input SGPRs as the VS has. */
4508 ctx
->param_tcs_offchip_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4509 ctx
->param_tcs_offchip_addr_base64k
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4510 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused */
4511 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused */
4512 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused */
4513 ctx
->param_vs_state_bits
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused */
4516 declare_per_stage_desc_pointers(ctx
, &fninfo
,
4517 ctx
->type
== PIPE_SHADER_GEOMETRY
);
4519 /* VGPRs (first GS, then VS/TES) */
4520 ctx
->param_gs_vtx01_offset
= add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
4521 ctx
->param_gs_vtx23_offset
= add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
4522 ctx
->param_gs_prim_id
= add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
4523 ctx
->param_gs_instance_id
= add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
4524 ctx
->param_gs_vtx45_offset
= add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
4526 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
4527 declare_vs_input_vgprs(ctx
, &fninfo
,
4529 } else if (ctx
->type
== PIPE_SHADER_TESS_EVAL
) {
4530 declare_tes_input_vgprs(ctx
, &fninfo
);
4533 if (ctx
->type
== PIPE_SHADER_VERTEX
||
4534 ctx
->type
== PIPE_SHADER_TESS_EVAL
) {
4535 /* ES return values are inputs to GS. */
4536 for (i
= 0; i
< 8 + GFX9_GS_NUM_USER_SGPR
; i
++)
4537 returns
[num_returns
++] = ctx
->i32
; /* SGPRs */
4538 for (i
= 0; i
< 5; i
++)
4539 returns
[num_returns
++] = ctx
->f32
; /* VGPRs */
4543 case PIPE_SHADER_TESS_EVAL
:
4544 declare_global_desc_pointers(ctx
, &fninfo
);
4545 declare_per_stage_desc_pointers(ctx
, &fninfo
, true);
4546 ctx
->param_tcs_offchip_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4547 ctx
->param_tcs_offchip_addr_base64k
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4549 if (shader
->key
.as_es
) {
4550 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4551 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4552 ctx
->param_es2gs_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4554 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4555 declare_streamout_params(ctx
, &shader
->selector
->so
,
4557 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4561 declare_tes_input_vgprs(ctx
, &fninfo
);
4564 case PIPE_SHADER_GEOMETRY
:
4565 declare_global_desc_pointers(ctx
, &fninfo
);
4566 declare_per_stage_desc_pointers(ctx
, &fninfo
, true);
4567 ctx
->param_gs2vs_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4568 ctx
->param_gs_wave_id
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4571 ctx
->param_gs_vtx0_offset
= add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
4572 ctx
->param_gs_vtx1_offset
= add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
4573 ctx
->param_gs_prim_id
= add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
4574 ctx
->param_gs_vtx2_offset
= add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
4575 ctx
->param_gs_vtx3_offset
= add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
4576 ctx
->param_gs_vtx4_offset
= add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
4577 ctx
->param_gs_vtx5_offset
= add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
4578 ctx
->param_gs_instance_id
= add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
4581 case PIPE_SHADER_FRAGMENT
:
4582 declare_global_desc_pointers(ctx
, &fninfo
);
4583 declare_per_stage_desc_pointers(ctx
, &fninfo
, true);
4584 add_arg_checked(&fninfo
, ARG_SGPR
, ctx
->f32
, SI_PARAM_ALPHA_REF
);
4585 add_arg_checked(&fninfo
, ARG_SGPR
, ctx
->i32
, SI_PARAM_PRIM_MASK
);
4587 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->v2i32
, SI_PARAM_PERSP_SAMPLE
);
4588 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->v2i32
, SI_PARAM_PERSP_CENTER
);
4589 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->v2i32
, SI_PARAM_PERSP_CENTROID
);
4590 add_arg_checked(&fninfo
, ARG_VGPR
, v3i32
, SI_PARAM_PERSP_PULL_MODEL
);
4591 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->v2i32
, SI_PARAM_LINEAR_SAMPLE
);
4592 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->v2i32
, SI_PARAM_LINEAR_CENTER
);
4593 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->v2i32
, SI_PARAM_LINEAR_CENTROID
);
4594 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->f32
, SI_PARAM_LINE_STIPPLE_TEX
);
4595 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->f32
,
4596 &ctx
->abi
.frag_pos
[0], SI_PARAM_POS_X_FLOAT
);
4597 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->f32
,
4598 &ctx
->abi
.frag_pos
[1], SI_PARAM_POS_Y_FLOAT
);
4599 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->f32
,
4600 &ctx
->abi
.frag_pos
[2], SI_PARAM_POS_Z_FLOAT
);
4601 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->f32
,
4602 &ctx
->abi
.frag_pos
[3], SI_PARAM_POS_W_FLOAT
);
4603 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->i32
,
4604 &ctx
->abi
.front_face
, SI_PARAM_FRONT_FACE
);
4605 shader
->info
.face_vgpr_index
= 20;
4606 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->i32
,
4607 &ctx
->abi
.ancillary
, SI_PARAM_ANCILLARY
);
4608 shader
->info
.ancillary_vgpr_index
= 21;
4609 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->f32
,
4610 &ctx
->abi
.sample_coverage
, SI_PARAM_SAMPLE_COVERAGE
);
4611 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->i32
, SI_PARAM_POS_FIXED_PT
);
4613 /* Color inputs from the prolog. */
4614 if (shader
->selector
->info
.colors_read
) {
4615 unsigned num_color_elements
=
4616 util_bitcount(shader
->selector
->info
.colors_read
);
4618 assert(fninfo
.num_params
+ num_color_elements
<= ARRAY_SIZE(fninfo
.types
));
4619 for (i
= 0; i
< num_color_elements
; i
++)
4620 add_arg(&fninfo
, ARG_VGPR
, ctx
->f32
);
4622 num_prolog_vgprs
+= num_color_elements
;
4625 /* Outputs for the epilog. */
4626 num_return_sgprs
= SI_SGPR_ALPHA_REF
+ 1;
4629 util_bitcount(shader
->selector
->info
.colors_written
) * 4 +
4630 shader
->selector
->info
.writes_z
+
4631 shader
->selector
->info
.writes_stencil
+
4632 shader
->selector
->info
.writes_samplemask
+
4633 1 /* SampleMaskIn */;
4635 num_returns
= MAX2(num_returns
,
4637 PS_EPILOG_SAMPLEMASK_MIN_LOC
+ 1);
4639 for (i
= 0; i
< num_return_sgprs
; i
++)
4640 returns
[i
] = ctx
->i32
;
4641 for (; i
< num_returns
; i
++)
4642 returns
[i
] = ctx
->f32
;
4645 case PIPE_SHADER_COMPUTE
:
4646 declare_global_desc_pointers(ctx
, &fninfo
);
4647 declare_per_stage_desc_pointers(ctx
, &fninfo
, true);
4648 if (shader
->selector
->info
.uses_grid_size
)
4649 ctx
->param_grid_size
= add_arg(&fninfo
, ARG_SGPR
, v3i32
);
4650 if (shader
->selector
->info
.uses_block_size
)
4651 ctx
->param_block_size
= add_arg(&fninfo
, ARG_SGPR
, v3i32
);
4653 for (i
= 0; i
< 3; i
++) {
4654 ctx
->param_block_id
[i
] = -1;
4655 if (shader
->selector
->info
.uses_block_id
[i
])
4656 ctx
->param_block_id
[i
] = add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4659 ctx
->param_thread_id
= add_arg(&fninfo
, ARG_VGPR
, v3i32
);
4662 assert(0 && "unimplemented shader");
4666 si_create_function(ctx
, "main", returns
, num_returns
, &fninfo
,
4667 si_get_max_workgroup_size(shader
));
4669 /* Reserve register locations for VGPR inputs the PS prolog may need. */
4670 if (ctx
->type
== PIPE_SHADER_FRAGMENT
&&
4671 ctx
->separate_prolog
) {
4672 si_llvm_add_attribute(ctx
->main_fn
,
4673 "InitialPSInputAddr",
4674 S_0286D0_PERSP_SAMPLE_ENA(1) |
4675 S_0286D0_PERSP_CENTER_ENA(1) |
4676 S_0286D0_PERSP_CENTROID_ENA(1) |
4677 S_0286D0_LINEAR_SAMPLE_ENA(1) |
4678 S_0286D0_LINEAR_CENTER_ENA(1) |
4679 S_0286D0_LINEAR_CENTROID_ENA(1) |
4680 S_0286D0_FRONT_FACE_ENA(1) |
4681 S_0286D0_ANCILLARY_ENA(1) |
4682 S_0286D0_POS_FIXED_PT_ENA(1));
4685 shader
->info
.num_input_sgprs
= 0;
4686 shader
->info
.num_input_vgprs
= 0;
4688 for (i
= 0; i
< fninfo
.num_sgpr_params
; ++i
)
4689 shader
->info
.num_input_sgprs
+= ac_get_type_size(fninfo
.types
[i
]) / 4;
4691 for (; i
< fninfo
.num_params
; ++i
)
4692 shader
->info
.num_input_vgprs
+= ac_get_type_size(fninfo
.types
[i
]) / 4;
4694 assert(shader
->info
.num_input_vgprs
>= num_prolog_vgprs
);
4695 shader
->info
.num_input_vgprs
-= num_prolog_vgprs
;
4697 if (shader
->key
.as_ls
||
4698 ctx
->type
== PIPE_SHADER_TESS_CTRL
||
4699 /* GFX9 has the ESGS ring buffer in LDS. */
4700 (ctx
->screen
->b
.chip_class
>= GFX9
&&
4701 (shader
->key
.as_es
||
4702 ctx
->type
== PIPE_SHADER_GEOMETRY
)))
4703 declare_lds_as_pointer(ctx
);
4707 * Load ESGS and GSVS ring buffer resource descriptors and save the variables
4710 static void preload_ring_buffers(struct si_shader_context
*ctx
)
4712 LLVMBuilderRef builder
= ctx
->ac
.builder
;
4714 LLVMValueRef buf_ptr
= LLVMGetParam(ctx
->main_fn
,
4715 ctx
->param_rw_buffers
);
4717 if (ctx
->screen
->b
.chip_class
<= VI
&&
4718 (ctx
->shader
->key
.as_es
|| ctx
->type
== PIPE_SHADER_GEOMETRY
)) {
4720 ctx
->type
== PIPE_SHADER_GEOMETRY
? SI_GS_RING_ESGS
4722 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
, ring
, 0);
4725 ac_build_indexed_load_const(&ctx
->ac
, buf_ptr
, offset
);
4728 if (ctx
->shader
->is_gs_copy_shader
) {
4729 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
, SI_RING_GSVS
, 0);
4732 ac_build_indexed_load_const(&ctx
->ac
, buf_ptr
, offset
);
4733 } else if (ctx
->type
== PIPE_SHADER_GEOMETRY
) {
4734 const struct si_shader_selector
*sel
= ctx
->shader
->selector
;
4735 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
, SI_RING_GSVS
, 0);
4736 LLVMValueRef base_ring
;
4738 base_ring
= ac_build_indexed_load_const(&ctx
->ac
, buf_ptr
, offset
);
4740 /* The conceptual layout of the GSVS ring is
4741 * v0c0 .. vLv0 v0c1 .. vLc1 ..
4742 * but the real memory layout is swizzled across
4744 * t0v0c0 .. t15v0c0 t0v1c0 .. t15v1c0 ... t15vLcL
4746 * Override the buffer descriptor accordingly.
4748 LLVMTypeRef v2i64
= LLVMVectorType(ctx
->i64
, 2);
4749 uint64_t stream_offset
= 0;
4751 for (unsigned stream
= 0; stream
< 4; ++stream
) {
4752 unsigned num_components
;
4754 unsigned num_records
;
4755 LLVMValueRef ring
, tmp
;
4757 num_components
= sel
->info
.num_stream_output_components
[stream
];
4758 if (!num_components
)
4761 stride
= 4 * num_components
* sel
->gs_max_out_vertices
;
4763 /* Limit on the stride field for <= CIK. */
4764 assert(stride
< (1 << 14));
4768 ring
= LLVMBuildBitCast(builder
, base_ring
, v2i64
, "");
4769 tmp
= LLVMBuildExtractElement(builder
, ring
, ctx
->i32_0
, "");
4770 tmp
= LLVMBuildAdd(builder
, tmp
,
4771 LLVMConstInt(ctx
->i64
,
4772 stream_offset
, 0), "");
4773 stream_offset
+= stride
* 64;
4775 ring
= LLVMBuildInsertElement(builder
, ring
, tmp
, ctx
->i32_0
, "");
4776 ring
= LLVMBuildBitCast(builder
, ring
, ctx
->v4i32
, "");
4777 tmp
= LLVMBuildExtractElement(builder
, ring
, ctx
->i32_1
, "");
4778 tmp
= LLVMBuildOr(builder
, tmp
,
4779 LLVMConstInt(ctx
->i32
,
4780 S_008F04_STRIDE(stride
) |
4781 S_008F04_SWIZZLE_ENABLE(1), 0), "");
4782 ring
= LLVMBuildInsertElement(builder
, ring
, tmp
, ctx
->i32_1
, "");
4783 ring
= LLVMBuildInsertElement(builder
, ring
,
4784 LLVMConstInt(ctx
->i32
, num_records
, 0),
4785 LLVMConstInt(ctx
->i32
, 2, 0), "");
4786 ring
= LLVMBuildInsertElement(builder
, ring
,
4787 LLVMConstInt(ctx
->i32
,
4788 S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
4789 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
4790 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
4791 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
4792 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
4793 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
4794 S_008F0C_ELEMENT_SIZE(1) | /* element_size = 4 (bytes) */
4795 S_008F0C_INDEX_STRIDE(1) | /* index_stride = 16 (elements) */
4796 S_008F0C_ADD_TID_ENABLE(1),
4798 LLVMConstInt(ctx
->i32
, 3, 0), "");
4800 ctx
->gsvs_ring
[stream
] = ring
;
4805 static void si_llvm_emit_polygon_stipple(struct si_shader_context
*ctx
,
4806 LLVMValueRef param_rw_buffers
,
4807 unsigned param_pos_fixed_pt
)
4809 LLVMBuilderRef builder
= ctx
->ac
.builder
;
4810 LLVMValueRef slot
, desc
, offset
, row
, bit
, address
[2];
4812 /* Use the fixed-point gl_FragCoord input.
4813 * Since the stipple pattern is 32x32 and it repeats, just get 5 bits
4814 * per coordinate to get the repeating effect.
4816 address
[0] = unpack_param(ctx
, param_pos_fixed_pt
, 0, 5);
4817 address
[1] = unpack_param(ctx
, param_pos_fixed_pt
, 16, 5);
4819 /* Load the buffer descriptor. */
4820 slot
= LLVMConstInt(ctx
->i32
, SI_PS_CONST_POLY_STIPPLE
, 0);
4821 desc
= ac_build_indexed_load_const(&ctx
->ac
, param_rw_buffers
, slot
);
4823 /* The stipple pattern is 32x32, each row has 32 bits. */
4824 offset
= LLVMBuildMul(builder
, address
[1],
4825 LLVMConstInt(ctx
->i32
, 4, 0), "");
4826 row
= buffer_load_const(ctx
, desc
, offset
);
4827 row
= ac_to_integer(&ctx
->ac
, row
);
4828 bit
= LLVMBuildLShr(builder
, row
, address
[0], "");
4829 bit
= LLVMBuildTrunc(builder
, bit
, ctx
->i1
, "");
4831 /* The intrinsic kills the thread if arg < 0. */
4832 bit
= LLVMBuildSelect(builder
, bit
, LLVMConstReal(ctx
->f32
, 0),
4833 LLVMConstReal(ctx
->f32
, -1), "");
4834 ac_build_kill(&ctx
->ac
, bit
);
4837 void si_shader_binary_read_config(struct ac_shader_binary
*binary
,
4838 struct si_shader_config
*conf
,
4839 unsigned symbol_offset
)
4842 const unsigned char *config
=
4843 ac_shader_binary_config_start(binary
, symbol_offset
);
4844 bool really_needs_scratch
= false;
4846 /* LLVM adds SGPR spills to the scratch size.
4847 * Find out if we really need the scratch buffer.
4849 for (i
= 0; i
< binary
->reloc_count
; i
++) {
4850 const struct ac_shader_reloc
*reloc
= &binary
->relocs
[i
];
4852 if (!strcmp(scratch_rsrc_dword0_symbol
, reloc
->name
) ||
4853 !strcmp(scratch_rsrc_dword1_symbol
, reloc
->name
)) {
4854 really_needs_scratch
= true;
4859 /* XXX: We may be able to emit some of these values directly rather than
4860 * extracting fields to be emitted later.
4863 for (i
= 0; i
< binary
->config_size_per_symbol
; i
+= 8) {
4864 unsigned reg
= util_le32_to_cpu(*(uint32_t*)(config
+ i
));
4865 unsigned value
= util_le32_to_cpu(*(uint32_t*)(config
+ i
+ 4));
4867 case R_00B028_SPI_SHADER_PGM_RSRC1_PS
:
4868 case R_00B128_SPI_SHADER_PGM_RSRC1_VS
:
4869 case R_00B228_SPI_SHADER_PGM_RSRC1_GS
:
4870 case R_00B428_SPI_SHADER_PGM_RSRC1_HS
:
4871 case R_00B848_COMPUTE_PGM_RSRC1
:
4872 conf
->num_sgprs
= MAX2(conf
->num_sgprs
, (G_00B028_SGPRS(value
) + 1) * 8);
4873 conf
->num_vgprs
= MAX2(conf
->num_vgprs
, (G_00B028_VGPRS(value
) + 1) * 4);
4874 conf
->float_mode
= G_00B028_FLOAT_MODE(value
);
4875 conf
->rsrc1
= value
;
4877 case R_00B02C_SPI_SHADER_PGM_RSRC2_PS
:
4878 conf
->lds_size
= MAX2(conf
->lds_size
, G_00B02C_EXTRA_LDS_SIZE(value
));
4880 case R_00B84C_COMPUTE_PGM_RSRC2
:
4881 conf
->lds_size
= MAX2(conf
->lds_size
, G_00B84C_LDS_SIZE(value
));
4882 conf
->rsrc2
= value
;
4884 case R_0286CC_SPI_PS_INPUT_ENA
:
4885 conf
->spi_ps_input_ena
= value
;
4887 case R_0286D0_SPI_PS_INPUT_ADDR
:
4888 conf
->spi_ps_input_addr
= value
;
4890 case R_0286E8_SPI_TMPRING_SIZE
:
4891 case R_00B860_COMPUTE_TMPRING_SIZE
:
4892 /* WAVESIZE is in units of 256 dwords. */
4893 if (really_needs_scratch
)
4894 conf
->scratch_bytes_per_wave
=
4895 G_00B860_WAVESIZE(value
) * 256 * 4;
4897 case 0x4: /* SPILLED_SGPRS */
4898 conf
->spilled_sgprs
= value
;
4900 case 0x8: /* SPILLED_VGPRS */
4901 conf
->spilled_vgprs
= value
;
4905 static bool printed
;
4908 fprintf(stderr
, "Warning: LLVM emitted unknown "
4909 "config register: 0x%x\n", reg
);
4917 if (!conf
->spi_ps_input_addr
)
4918 conf
->spi_ps_input_addr
= conf
->spi_ps_input_ena
;
4921 void si_shader_apply_scratch_relocs(struct si_shader
*shader
,
4922 uint64_t scratch_va
)
4925 uint32_t scratch_rsrc_dword0
= scratch_va
;
4926 uint32_t scratch_rsrc_dword1
=
4927 S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32);
4929 /* Enable scratch coalescing. */
4930 scratch_rsrc_dword1
|= S_008F04_SWIZZLE_ENABLE(1);
4932 for (i
= 0 ; i
< shader
->binary
.reloc_count
; i
++) {
4933 const struct ac_shader_reloc
*reloc
=
4934 &shader
->binary
.relocs
[i
];
4935 if (!strcmp(scratch_rsrc_dword0_symbol
, reloc
->name
)) {
4936 util_memcpy_cpu_to_le32(shader
->binary
.code
+ reloc
->offset
,
4937 &scratch_rsrc_dword0
, 4);
4938 } else if (!strcmp(scratch_rsrc_dword1_symbol
, reloc
->name
)) {
4939 util_memcpy_cpu_to_le32(shader
->binary
.code
+ reloc
->offset
,
4940 &scratch_rsrc_dword1
, 4);
4945 static unsigned si_get_shader_binary_size(const struct si_shader
*shader
)
4947 unsigned size
= shader
->binary
.code_size
;
4950 size
+= shader
->prolog
->binary
.code_size
;
4951 if (shader
->previous_stage
)
4952 size
+= shader
->previous_stage
->binary
.code_size
;
4953 if (shader
->prolog2
)
4954 size
+= shader
->prolog2
->binary
.code_size
;
4956 size
+= shader
->epilog
->binary
.code_size
;
4960 int si_shader_binary_upload(struct si_screen
*sscreen
, struct si_shader
*shader
)
4962 const struct ac_shader_binary
*prolog
=
4963 shader
->prolog
? &shader
->prolog
->binary
: NULL
;
4964 const struct ac_shader_binary
*previous_stage
=
4965 shader
->previous_stage
? &shader
->previous_stage
->binary
: NULL
;
4966 const struct ac_shader_binary
*prolog2
=
4967 shader
->prolog2
? &shader
->prolog2
->binary
: NULL
;
4968 const struct ac_shader_binary
*epilog
=
4969 shader
->epilog
? &shader
->epilog
->binary
: NULL
;
4970 const struct ac_shader_binary
*mainb
= &shader
->binary
;
4971 unsigned bo_size
= si_get_shader_binary_size(shader
) +
4972 (!epilog
? mainb
->rodata_size
: 0);
4975 assert(!prolog
|| !prolog
->rodata_size
);
4976 assert(!previous_stage
|| !previous_stage
->rodata_size
);
4977 assert(!prolog2
|| !prolog2
->rodata_size
);
4978 assert((!prolog
&& !previous_stage
&& !prolog2
&& !epilog
) ||
4979 !mainb
->rodata_size
);
4980 assert(!epilog
|| !epilog
->rodata_size
);
4982 r600_resource_reference(&shader
->bo
, NULL
);
4983 shader
->bo
= (struct r600_resource
*)
4984 pipe_buffer_create(&sscreen
->b
.b
, 0,
4985 PIPE_USAGE_IMMUTABLE
,
4986 align(bo_size
, SI_CPDMA_ALIGNMENT
));
4991 ptr
= sscreen
->b
.ws
->buffer_map(shader
->bo
->buf
, NULL
,
4992 PIPE_TRANSFER_READ_WRITE
|
4993 PIPE_TRANSFER_UNSYNCHRONIZED
);
4995 /* Don't use util_memcpy_cpu_to_le32. LLVM binaries are
4996 * endian-independent. */
4998 memcpy(ptr
, prolog
->code
, prolog
->code_size
);
4999 ptr
+= prolog
->code_size
;
5001 if (previous_stage
) {
5002 memcpy(ptr
, previous_stage
->code
, previous_stage
->code_size
);
5003 ptr
+= previous_stage
->code_size
;
5006 memcpy(ptr
, prolog2
->code
, prolog2
->code_size
);
5007 ptr
+= prolog2
->code_size
;
5010 memcpy(ptr
, mainb
->code
, mainb
->code_size
);
5011 ptr
+= mainb
->code_size
;
5014 memcpy(ptr
, epilog
->code
, epilog
->code_size
);
5015 else if (mainb
->rodata_size
> 0)
5016 memcpy(ptr
, mainb
->rodata
, mainb
->rodata_size
);
5018 sscreen
->b
.ws
->buffer_unmap(shader
->bo
->buf
);
5022 static void si_shader_dump_disassembly(const struct ac_shader_binary
*binary
,
5023 struct pipe_debug_callback
*debug
,
5024 const char *name
, FILE *file
)
5029 if (binary
->disasm_string
) {
5030 fprintf(file
, "Shader %s disassembly:\n", name
);
5031 fprintf(file
, "%s", binary
->disasm_string
);
5033 if (debug
&& debug
->debug_message
) {
5034 /* Very long debug messages are cut off, so send the
5035 * disassembly one line at a time. This causes more
5036 * overhead, but on the plus side it simplifies
5037 * parsing of resulting logs.
5039 pipe_debug_message(debug
, SHADER_INFO
,
5040 "Shader Disassembly Begin");
5042 line
= binary
->disasm_string
;
5044 p
= util_strchrnul(line
, '\n');
5048 pipe_debug_message(debug
, SHADER_INFO
,
5049 "%.*s", count
, line
);
5057 pipe_debug_message(debug
, SHADER_INFO
,
5058 "Shader Disassembly End");
5061 fprintf(file
, "Shader %s binary:\n", name
);
5062 for (i
= 0; i
< binary
->code_size
; i
+= 4) {
5063 fprintf(file
, "@0x%x: %02x%02x%02x%02x\n", i
,
5064 binary
->code
[i
+ 3], binary
->code
[i
+ 2],
5065 binary
->code
[i
+ 1], binary
->code
[i
]);
5070 static void si_shader_dump_stats(struct si_screen
*sscreen
,
5071 const struct si_shader
*shader
,
5072 struct pipe_debug_callback
*debug
,
5075 bool check_debug_option
)
5077 const struct si_shader_config
*conf
= &shader
->config
;
5078 unsigned num_inputs
= shader
->selector
? shader
->selector
->info
.num_inputs
: 0;
5079 unsigned code_size
= si_get_shader_binary_size(shader
);
5080 unsigned lds_increment
= sscreen
->b
.chip_class
>= CIK
? 512 : 256;
5081 unsigned lds_per_wave
= 0;
5082 unsigned max_simd_waves
;
5084 switch (sscreen
->b
.family
) {
5085 /* These always have 8 waves: */
5086 case CHIP_POLARIS10
:
5087 case CHIP_POLARIS11
:
5088 case CHIP_POLARIS12
:
5092 max_simd_waves
= 10;
5095 /* Compute LDS usage for PS. */
5096 switch (processor
) {
5097 case PIPE_SHADER_FRAGMENT
:
5098 /* The minimum usage per wave is (num_inputs * 48). The maximum
5099 * usage is (num_inputs * 48 * 16).
5100 * We can get anything in between and it varies between waves.
5102 * The 48 bytes per input for a single primitive is equal to
5103 * 4 bytes/component * 4 components/input * 3 points.
5105 * Other stages don't know the size at compile time or don't
5106 * allocate LDS per wave, but instead they do it per thread group.
5108 lds_per_wave
= conf
->lds_size
* lds_increment
+
5109 align(num_inputs
* 48, lds_increment
);
5111 case PIPE_SHADER_COMPUTE
:
5112 if (shader
->selector
) {
5113 unsigned max_workgroup_size
=
5114 si_get_max_workgroup_size(shader
);
5115 lds_per_wave
= (conf
->lds_size
* lds_increment
) /
5116 DIV_ROUND_UP(max_workgroup_size
, 64);
5121 /* Compute the per-SIMD wave counts. */
5122 if (conf
->num_sgprs
) {
5123 if (sscreen
->b
.chip_class
>= VI
)
5124 max_simd_waves
= MIN2(max_simd_waves
, 800 / conf
->num_sgprs
);
5126 max_simd_waves
= MIN2(max_simd_waves
, 512 / conf
->num_sgprs
);
5129 if (conf
->num_vgprs
)
5130 max_simd_waves
= MIN2(max_simd_waves
, 256 / conf
->num_vgprs
);
5132 /* LDS is 64KB per CU (4 SIMDs), which is 16KB per SIMD (usage above
5133 * 16KB makes some SIMDs unoccupied). */
5135 max_simd_waves
= MIN2(max_simd_waves
, 16384 / lds_per_wave
);
5137 if (!check_debug_option
||
5138 si_can_dump_shader(&sscreen
->b
, processor
)) {
5139 if (processor
== PIPE_SHADER_FRAGMENT
) {
5140 fprintf(file
, "*** SHADER CONFIG ***\n"
5141 "SPI_PS_INPUT_ADDR = 0x%04x\n"
5142 "SPI_PS_INPUT_ENA = 0x%04x\n",
5143 conf
->spi_ps_input_addr
, conf
->spi_ps_input_ena
);
5146 fprintf(file
, "*** SHADER STATS ***\n"
5149 "Spilled SGPRs: %d\n"
5150 "Spilled VGPRs: %d\n"
5151 "Private memory VGPRs: %d\n"
5152 "Code Size: %d bytes\n"
5154 "Scratch: %d bytes per wave\n"
5156 "********************\n\n\n",
5157 conf
->num_sgprs
, conf
->num_vgprs
,
5158 conf
->spilled_sgprs
, conf
->spilled_vgprs
,
5159 conf
->private_mem_vgprs
, code_size
,
5160 conf
->lds_size
, conf
->scratch_bytes_per_wave
,
5164 pipe_debug_message(debug
, SHADER_INFO
,
5165 "Shader Stats: SGPRS: %d VGPRS: %d Code Size: %d "
5166 "LDS: %d Scratch: %d Max Waves: %d Spilled SGPRs: %d "
5167 "Spilled VGPRs: %d PrivMem VGPRs: %d",
5168 conf
->num_sgprs
, conf
->num_vgprs
, code_size
,
5169 conf
->lds_size
, conf
->scratch_bytes_per_wave
,
5170 max_simd_waves
, conf
->spilled_sgprs
,
5171 conf
->spilled_vgprs
, conf
->private_mem_vgprs
);
5174 const char *si_get_shader_name(const struct si_shader
*shader
, unsigned processor
)
5176 switch (processor
) {
5177 case PIPE_SHADER_VERTEX
:
5178 if (shader
->key
.as_es
)
5179 return "Vertex Shader as ES";
5180 else if (shader
->key
.as_ls
)
5181 return "Vertex Shader as LS";
5183 return "Vertex Shader as VS";
5184 case PIPE_SHADER_TESS_CTRL
:
5185 return "Tessellation Control Shader";
5186 case PIPE_SHADER_TESS_EVAL
:
5187 if (shader
->key
.as_es
)
5188 return "Tessellation Evaluation Shader as ES";
5190 return "Tessellation Evaluation Shader as VS";
5191 case PIPE_SHADER_GEOMETRY
:
5192 if (shader
->is_gs_copy_shader
)
5193 return "GS Copy Shader as VS";
5195 return "Geometry Shader";
5196 case PIPE_SHADER_FRAGMENT
:
5197 return "Pixel Shader";
5198 case PIPE_SHADER_COMPUTE
:
5199 return "Compute Shader";
5201 return "Unknown Shader";
5205 void si_shader_dump(struct si_screen
*sscreen
, const struct si_shader
*shader
,
5206 struct pipe_debug_callback
*debug
, unsigned processor
,
5207 FILE *file
, bool check_debug_option
)
5209 if (!check_debug_option
||
5210 si_can_dump_shader(&sscreen
->b
, processor
))
5211 si_dump_shader_key(processor
, shader
, file
);
5213 if (!check_debug_option
&& shader
->binary
.llvm_ir_string
) {
5214 if (shader
->previous_stage
&&
5215 shader
->previous_stage
->binary
.llvm_ir_string
) {
5216 fprintf(file
, "\n%s - previous stage - LLVM IR:\n\n",
5217 si_get_shader_name(shader
, processor
));
5218 fprintf(file
, "%s\n", shader
->previous_stage
->binary
.llvm_ir_string
);
5221 fprintf(file
, "\n%s - main shader part - LLVM IR:\n\n",
5222 si_get_shader_name(shader
, processor
));
5223 fprintf(file
, "%s\n", shader
->binary
.llvm_ir_string
);
5226 if (!check_debug_option
||
5227 (si_can_dump_shader(&sscreen
->b
, processor
) &&
5228 !(sscreen
->b
.debug_flags
& DBG(NO_ASM
)))) {
5229 fprintf(file
, "\n%s:\n", si_get_shader_name(shader
, processor
));
5232 si_shader_dump_disassembly(&shader
->prolog
->binary
,
5233 debug
, "prolog", file
);
5234 if (shader
->previous_stage
)
5235 si_shader_dump_disassembly(&shader
->previous_stage
->binary
,
5236 debug
, "previous stage", file
);
5237 if (shader
->prolog2
)
5238 si_shader_dump_disassembly(&shader
->prolog2
->binary
,
5239 debug
, "prolog2", file
);
5241 si_shader_dump_disassembly(&shader
->binary
, debug
, "main", file
);
5244 si_shader_dump_disassembly(&shader
->epilog
->binary
,
5245 debug
, "epilog", file
);
5246 fprintf(file
, "\n");
5249 si_shader_dump_stats(sscreen
, shader
, debug
, processor
, file
,
5250 check_debug_option
);
5253 static int si_compile_llvm(struct si_screen
*sscreen
,
5254 struct ac_shader_binary
*binary
,
5255 struct si_shader_config
*conf
,
5256 LLVMTargetMachineRef tm
,
5258 struct pipe_debug_callback
*debug
,
5263 unsigned count
= p_atomic_inc_return(&sscreen
->b
.num_compilations
);
5265 if (si_can_dump_shader(&sscreen
->b
, processor
)) {
5266 fprintf(stderr
, "radeonsi: Compiling shader %d\n", count
);
5268 if (!(sscreen
->b
.debug_flags
& (DBG(NO_IR
) | DBG(PREOPT_IR
)))) {
5269 fprintf(stderr
, "%s LLVM IR:\n\n", name
);
5270 ac_dump_module(mod
);
5271 fprintf(stderr
, "\n");
5275 if (sscreen
->record_llvm_ir
) {
5276 char *ir
= LLVMPrintModuleToString(mod
);
5277 binary
->llvm_ir_string
= strdup(ir
);
5278 LLVMDisposeMessage(ir
);
5281 if (!si_replace_shader(count
, binary
)) {
5282 r
= si_llvm_compile(mod
, binary
, tm
, debug
);
5287 si_shader_binary_read_config(binary
, conf
, 0);
5289 /* Enable 64-bit and 16-bit denormals, because there is no performance
5292 * If denormals are enabled, all floating-point output modifiers are
5295 * Don't enable denormals for 32-bit floats, because:
5296 * - Floating-point output modifiers would be ignored by the hw.
5297 * - Some opcodes don't support denormals, such as v_mad_f32. We would
5298 * have to stop using those.
5299 * - SI & CI would be very slow.
5301 conf
->float_mode
|= V_00B028_FP_64_DENORMS
;
5303 FREE(binary
->config
);
5304 FREE(binary
->global_symbol_offsets
);
5305 binary
->config
= NULL
;
5306 binary
->global_symbol_offsets
= NULL
;
5308 /* Some shaders can't have rodata because their binaries can be
5311 if (binary
->rodata_size
&&
5312 (processor
== PIPE_SHADER_VERTEX
||
5313 processor
== PIPE_SHADER_TESS_CTRL
||
5314 processor
== PIPE_SHADER_TESS_EVAL
||
5315 processor
== PIPE_SHADER_FRAGMENT
)) {
5316 fprintf(stderr
, "radeonsi: The shader can't have rodata.");
5323 static void si_llvm_build_ret(struct si_shader_context
*ctx
, LLVMValueRef ret
)
5325 if (LLVMGetTypeKind(LLVMTypeOf(ret
)) == LLVMVoidTypeKind
)
5326 LLVMBuildRetVoid(ctx
->ac
.builder
);
5328 LLVMBuildRet(ctx
->ac
.builder
, ret
);
5331 /* Generate code for the hardware VS shader stage to go with a geometry shader */
5333 si_generate_gs_copy_shader(struct si_screen
*sscreen
,
5334 LLVMTargetMachineRef tm
,
5335 struct si_shader_selector
*gs_selector
,
5336 struct pipe_debug_callback
*debug
)
5338 struct si_shader_context ctx
;
5339 struct si_shader
*shader
;
5340 LLVMBuilderRef builder
;
5341 struct lp_build_tgsi_context
*bld_base
= &ctx
.bld_base
;
5342 struct lp_build_context
*uint
= &bld_base
->uint_bld
;
5343 struct si_shader_output_values
*outputs
;
5344 struct tgsi_shader_info
*gsinfo
= &gs_selector
->info
;
5347 outputs
= MALLOC(gsinfo
->num_outputs
* sizeof(outputs
[0]));
5352 shader
= CALLOC_STRUCT(si_shader
);
5359 shader
->selector
= gs_selector
;
5360 shader
->is_gs_copy_shader
= true;
5362 si_init_shader_ctx(&ctx
, sscreen
, tm
);
5363 ctx
.shader
= shader
;
5364 ctx
.type
= PIPE_SHADER_VERTEX
;
5366 builder
= ctx
.ac
.builder
;
5368 create_function(&ctx
);
5369 preload_ring_buffers(&ctx
);
5371 LLVMValueRef voffset
=
5372 lp_build_mul_imm(uint
, ctx
.abi
.vertex_id
, 4);
5374 /* Fetch the vertex stream ID.*/
5375 LLVMValueRef stream_id
;
5377 if (gs_selector
->so
.num_outputs
)
5378 stream_id
= unpack_param(&ctx
, ctx
.param_streamout_config
, 24, 2);
5380 stream_id
= ctx
.i32_0
;
5382 /* Fill in output information. */
5383 for (i
= 0; i
< gsinfo
->num_outputs
; ++i
) {
5384 outputs
[i
].semantic_name
= gsinfo
->output_semantic_name
[i
];
5385 outputs
[i
].semantic_index
= gsinfo
->output_semantic_index
[i
];
5387 for (int chan
= 0; chan
< 4; chan
++) {
5388 outputs
[i
].vertex_stream
[chan
] =
5389 (gsinfo
->output_streams
[i
] >> (2 * chan
)) & 3;
5393 LLVMBasicBlockRef end_bb
;
5394 LLVMValueRef switch_inst
;
5396 end_bb
= LLVMAppendBasicBlockInContext(ctx
.ac
.context
, ctx
.main_fn
, "end");
5397 switch_inst
= LLVMBuildSwitch(builder
, stream_id
, end_bb
, 4);
5399 for (int stream
= 0; stream
< 4; stream
++) {
5400 LLVMBasicBlockRef bb
;
5403 if (!gsinfo
->num_stream_output_components
[stream
])
5406 if (stream
> 0 && !gs_selector
->so
.num_outputs
)
5409 bb
= LLVMInsertBasicBlockInContext(ctx
.ac
.context
, end_bb
, "out");
5410 LLVMAddCase(switch_inst
, LLVMConstInt(ctx
.i32
, stream
, 0), bb
);
5411 LLVMPositionBuilderAtEnd(builder
, bb
);
5413 /* Fetch vertex data from GSVS ring */
5415 for (i
= 0; i
< gsinfo
->num_outputs
; ++i
) {
5416 for (unsigned chan
= 0; chan
< 4; chan
++) {
5417 if (!(gsinfo
->output_usagemask
[i
] & (1 << chan
)) ||
5418 outputs
[i
].vertex_stream
[chan
] != stream
) {
5419 outputs
[i
].values
[chan
] = ctx
.bld_base
.base
.undef
;
5423 LLVMValueRef soffset
= LLVMConstInt(ctx
.i32
,
5424 offset
* gs_selector
->gs_max_out_vertices
* 16 * 4, 0);
5427 outputs
[i
].values
[chan
] =
5428 ac_build_buffer_load(&ctx
.ac
,
5429 ctx
.gsvs_ring
[0], 1,
5436 /* Streamout and exports. */
5437 if (gs_selector
->so
.num_outputs
) {
5438 si_llvm_emit_streamout(&ctx
, outputs
,
5439 gsinfo
->num_outputs
,
5444 si_llvm_export_vs(bld_base
, outputs
, gsinfo
->num_outputs
);
5446 LLVMBuildBr(builder
, end_bb
);
5449 LLVMPositionBuilderAtEnd(builder
, end_bb
);
5451 LLVMBuildRetVoid(ctx
.ac
.builder
);
5453 ctx
.type
= PIPE_SHADER_GEOMETRY
; /* override for shader dumping */
5454 si_llvm_optimize_module(&ctx
);
5456 r
= si_compile_llvm(sscreen
, &ctx
.shader
->binary
,
5457 &ctx
.shader
->config
, ctx
.tm
,
5459 debug
, PIPE_SHADER_GEOMETRY
,
5462 if (si_can_dump_shader(&sscreen
->b
, PIPE_SHADER_GEOMETRY
))
5463 fprintf(stderr
, "GS Copy Shader:\n");
5464 si_shader_dump(sscreen
, ctx
.shader
, debug
,
5465 PIPE_SHADER_GEOMETRY
, stderr
, true);
5466 r
= si_shader_binary_upload(sscreen
, ctx
.shader
);
5469 si_llvm_dispose(&ctx
);
5480 static void si_dump_shader_key_vs(const struct si_shader_key
*key
,
5481 const struct si_vs_prolog_bits
*prolog
,
5482 const char *prefix
, FILE *f
)
5484 fprintf(f
, " %s.instance_divisor_is_one = %u\n",
5485 prefix
, prolog
->instance_divisor_is_one
);
5486 fprintf(f
, " %s.instance_divisor_is_fetched = %u\n",
5487 prefix
, prolog
->instance_divisor_is_fetched
);
5488 fprintf(f
, " %s.ls_vgpr_fix = %u\n",
5489 prefix
, prolog
->ls_vgpr_fix
);
5491 fprintf(f
, " mono.vs.fix_fetch = {");
5492 for (int i
= 0; i
< SI_MAX_ATTRIBS
; i
++)
5493 fprintf(f
, !i
? "%u" : ", %u", key
->mono
.vs_fix_fetch
[i
]);
5497 static void si_dump_shader_key(unsigned processor
, const struct si_shader
*shader
,
5500 const struct si_shader_key
*key
= &shader
->key
;
5502 fprintf(f
, "SHADER KEY\n");
5504 switch (processor
) {
5505 case PIPE_SHADER_VERTEX
:
5506 si_dump_shader_key_vs(key
, &key
->part
.vs
.prolog
,
5507 "part.vs.prolog", f
);
5508 fprintf(f
, " as_es = %u\n", key
->as_es
);
5509 fprintf(f
, " as_ls = %u\n", key
->as_ls
);
5510 fprintf(f
, " mono.u.vs_export_prim_id = %u\n",
5511 key
->mono
.u
.vs_export_prim_id
);
5514 case PIPE_SHADER_TESS_CTRL
:
5515 if (shader
->selector
->screen
->b
.chip_class
>= GFX9
) {
5516 si_dump_shader_key_vs(key
, &key
->part
.tcs
.ls_prolog
,
5517 "part.tcs.ls_prolog", f
);
5519 fprintf(f
, " part.tcs.epilog.prim_mode = %u\n", key
->part
.tcs
.epilog
.prim_mode
);
5520 fprintf(f
, " mono.u.ff_tcs_inputs_to_copy = 0x%"PRIx64
"\n", key
->mono
.u
.ff_tcs_inputs_to_copy
);
5523 case PIPE_SHADER_TESS_EVAL
:
5524 fprintf(f
, " as_es = %u\n", key
->as_es
);
5525 fprintf(f
, " mono.u.vs_export_prim_id = %u\n",
5526 key
->mono
.u
.vs_export_prim_id
);
5529 case PIPE_SHADER_GEOMETRY
:
5530 if (shader
->is_gs_copy_shader
)
5533 if (shader
->selector
->screen
->b
.chip_class
>= GFX9
&&
5534 key
->part
.gs
.es
->type
== PIPE_SHADER_VERTEX
) {
5535 si_dump_shader_key_vs(key
, &key
->part
.gs
.vs_prolog
,
5536 "part.gs.vs_prolog", f
);
5538 fprintf(f
, " part.gs.prolog.tri_strip_adj_fix = %u\n", key
->part
.gs
.prolog
.tri_strip_adj_fix
);
5541 case PIPE_SHADER_COMPUTE
:
5544 case PIPE_SHADER_FRAGMENT
:
5545 fprintf(f
, " part.ps.prolog.color_two_side = %u\n", key
->part
.ps
.prolog
.color_two_side
);
5546 fprintf(f
, " part.ps.prolog.flatshade_colors = %u\n", key
->part
.ps
.prolog
.flatshade_colors
);
5547 fprintf(f
, " part.ps.prolog.poly_stipple = %u\n", key
->part
.ps
.prolog
.poly_stipple
);
5548 fprintf(f
, " part.ps.prolog.force_persp_sample_interp = %u\n", key
->part
.ps
.prolog
.force_persp_sample_interp
);
5549 fprintf(f
, " part.ps.prolog.force_linear_sample_interp = %u\n", key
->part
.ps
.prolog
.force_linear_sample_interp
);
5550 fprintf(f
, " part.ps.prolog.force_persp_center_interp = %u\n", key
->part
.ps
.prolog
.force_persp_center_interp
);
5551 fprintf(f
, " part.ps.prolog.force_linear_center_interp = %u\n", key
->part
.ps
.prolog
.force_linear_center_interp
);
5552 fprintf(f
, " part.ps.prolog.bc_optimize_for_persp = %u\n", key
->part
.ps
.prolog
.bc_optimize_for_persp
);
5553 fprintf(f
, " part.ps.prolog.bc_optimize_for_linear = %u\n", key
->part
.ps
.prolog
.bc_optimize_for_linear
);
5554 fprintf(f
, " part.ps.epilog.spi_shader_col_format = 0x%x\n", key
->part
.ps
.epilog
.spi_shader_col_format
);
5555 fprintf(f
, " part.ps.epilog.color_is_int8 = 0x%X\n", key
->part
.ps
.epilog
.color_is_int8
);
5556 fprintf(f
, " part.ps.epilog.color_is_int10 = 0x%X\n", key
->part
.ps
.epilog
.color_is_int10
);
5557 fprintf(f
, " part.ps.epilog.last_cbuf = %u\n", key
->part
.ps
.epilog
.last_cbuf
);
5558 fprintf(f
, " part.ps.epilog.alpha_func = %u\n", key
->part
.ps
.epilog
.alpha_func
);
5559 fprintf(f
, " part.ps.epilog.alpha_to_one = %u\n", key
->part
.ps
.epilog
.alpha_to_one
);
5560 fprintf(f
, " part.ps.epilog.poly_line_smoothing = %u\n", key
->part
.ps
.epilog
.poly_line_smoothing
);
5561 fprintf(f
, " part.ps.epilog.clamp_color = %u\n", key
->part
.ps
.epilog
.clamp_color
);
5568 if ((processor
== PIPE_SHADER_GEOMETRY
||
5569 processor
== PIPE_SHADER_TESS_EVAL
||
5570 processor
== PIPE_SHADER_VERTEX
) &&
5571 !key
->as_es
&& !key
->as_ls
) {
5572 fprintf(f
, " opt.kill_outputs = 0x%"PRIx64
"\n", key
->opt
.kill_outputs
);
5573 fprintf(f
, " opt.clip_disable = %u\n", key
->opt
.clip_disable
);
5577 static void si_init_shader_ctx(struct si_shader_context
*ctx
,
5578 struct si_screen
*sscreen
,
5579 LLVMTargetMachineRef tm
)
5581 struct lp_build_tgsi_context
*bld_base
;
5583 si_llvm_context_init(ctx
, sscreen
, tm
);
5585 bld_base
= &ctx
->bld_base
;
5586 bld_base
->emit_fetch_funcs
[TGSI_FILE_CONSTANT
] = fetch_constant
;
5588 bld_base
->op_actions
[TGSI_OPCODE_INTERP_CENTROID
] = interp_action
;
5589 bld_base
->op_actions
[TGSI_OPCODE_INTERP_SAMPLE
] = interp_action
;
5590 bld_base
->op_actions
[TGSI_OPCODE_INTERP_OFFSET
] = interp_action
;
5592 bld_base
->op_actions
[TGSI_OPCODE_MEMBAR
].emit
= membar_emit
;
5594 bld_base
->op_actions
[TGSI_OPCODE_CLOCK
].emit
= clock_emit
;
5596 bld_base
->op_actions
[TGSI_OPCODE_DDX
].emit
= si_llvm_emit_ddxy
;
5597 bld_base
->op_actions
[TGSI_OPCODE_DDY
].emit
= si_llvm_emit_ddxy
;
5598 bld_base
->op_actions
[TGSI_OPCODE_DDX_FINE
].emit
= si_llvm_emit_ddxy
;
5599 bld_base
->op_actions
[TGSI_OPCODE_DDY_FINE
].emit
= si_llvm_emit_ddxy
;
5601 bld_base
->op_actions
[TGSI_OPCODE_VOTE_ALL
].emit
= vote_all_emit
;
5602 bld_base
->op_actions
[TGSI_OPCODE_VOTE_ANY
].emit
= vote_any_emit
;
5603 bld_base
->op_actions
[TGSI_OPCODE_VOTE_EQ
].emit
= vote_eq_emit
;
5604 bld_base
->op_actions
[TGSI_OPCODE_BALLOT
].emit
= ballot_emit
;
5605 bld_base
->op_actions
[TGSI_OPCODE_READ_FIRST
].intr_name
= "llvm.amdgcn.readfirstlane";
5606 bld_base
->op_actions
[TGSI_OPCODE_READ_FIRST
].emit
= read_lane_emit
;
5607 bld_base
->op_actions
[TGSI_OPCODE_READ_INVOC
].intr_name
= "llvm.amdgcn.readlane";
5608 bld_base
->op_actions
[TGSI_OPCODE_READ_INVOC
].fetch_args
= read_invoc_fetch_args
;
5609 bld_base
->op_actions
[TGSI_OPCODE_READ_INVOC
].emit
= read_lane_emit
;
5611 bld_base
->op_actions
[TGSI_OPCODE_EMIT
].emit
= si_llvm_emit_vertex
;
5612 bld_base
->op_actions
[TGSI_OPCODE_ENDPRIM
].emit
= si_llvm_emit_primitive
;
5613 bld_base
->op_actions
[TGSI_OPCODE_BARRIER
].emit
= si_llvm_emit_barrier
;
5616 static void si_optimize_vs_outputs(struct si_shader_context
*ctx
)
5618 struct si_shader
*shader
= ctx
->shader
;
5619 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
5621 if ((ctx
->type
!= PIPE_SHADER_VERTEX
&&
5622 ctx
->type
!= PIPE_SHADER_TESS_EVAL
) ||
5623 shader
->key
.as_ls
||
5627 ac_optimize_vs_outputs(&ctx
->ac
,
5629 shader
->info
.vs_output_param_offset
,
5631 &shader
->info
.nr_param_exports
);
5634 static void si_count_scratch_private_memory(struct si_shader_context
*ctx
)
5636 ctx
->shader
->config
.private_mem_vgprs
= 0;
5638 /* Process all LLVM instructions. */
5639 LLVMBasicBlockRef bb
= LLVMGetFirstBasicBlock(ctx
->main_fn
);
5641 LLVMValueRef next
= LLVMGetFirstInstruction(bb
);
5644 LLVMValueRef inst
= next
;
5645 next
= LLVMGetNextInstruction(next
);
5647 if (LLVMGetInstructionOpcode(inst
) != LLVMAlloca
)
5650 LLVMTypeRef type
= LLVMGetElementType(LLVMTypeOf(inst
));
5651 /* No idea why LLVM aligns allocas to 4 elements. */
5652 unsigned alignment
= LLVMGetAlignment(inst
);
5653 unsigned dw_size
= align(ac_get_type_size(type
) / 4, alignment
);
5654 ctx
->shader
->config
.private_mem_vgprs
+= dw_size
;
5656 bb
= LLVMGetNextBasicBlock(bb
);
5660 static void si_init_exec_full_mask(struct si_shader_context
*ctx
)
5662 LLVMValueRef full_mask
= LLVMConstInt(ctx
->i64
, ~0ull, 0);
5663 lp_build_intrinsic(ctx
->ac
.builder
,
5664 "llvm.amdgcn.init.exec", ctx
->voidt
,
5665 &full_mask
, 1, LP_FUNC_ATTR_CONVERGENT
);
5668 static void si_init_exec_from_input(struct si_shader_context
*ctx
,
5669 unsigned param
, unsigned bitoffset
)
5671 LLVMValueRef args
[] = {
5672 LLVMGetParam(ctx
->main_fn
, param
),
5673 LLVMConstInt(ctx
->i32
, bitoffset
, 0),
5675 lp_build_intrinsic(ctx
->ac
.builder
,
5676 "llvm.amdgcn.init.exec.from.input",
5677 ctx
->voidt
, args
, 2, LP_FUNC_ATTR_CONVERGENT
);
5680 static bool si_vs_needs_prolog(const struct si_shader_selector
*sel
,
5681 const struct si_vs_prolog_bits
*key
)
5683 /* VGPR initialization fixup for Vega10 and Raven is always done in the
5685 return sel
->vs_needs_prolog
|| key
->ls_vgpr_fix
;
5688 static bool si_compile_tgsi_main(struct si_shader_context
*ctx
,
5691 struct si_shader
*shader
= ctx
->shader
;
5692 struct si_shader_selector
*sel
= shader
->selector
;
5693 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
5695 // TODO clean all this up!
5696 switch (ctx
->type
) {
5697 case PIPE_SHADER_VERTEX
:
5698 ctx
->load_input
= declare_input_vs
;
5699 if (shader
->key
.as_ls
)
5700 bld_base
->emit_epilogue
= si_llvm_emit_ls_epilogue
;
5701 else if (shader
->key
.as_es
)
5702 bld_base
->emit_epilogue
= si_llvm_emit_es_epilogue
;
5704 ctx
->abi
.emit_outputs
= si_llvm_emit_vs_epilogue
;
5705 bld_base
->emit_epilogue
= si_tgsi_emit_epilogue
;
5708 case PIPE_SHADER_TESS_CTRL
:
5709 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = fetch_input_tcs
;
5710 bld_base
->emit_fetch_funcs
[TGSI_FILE_OUTPUT
] = fetch_output_tcs
;
5711 bld_base
->emit_store
= store_output_tcs
;
5712 bld_base
->emit_epilogue
= si_llvm_emit_tcs_epilogue
;
5714 case PIPE_SHADER_TESS_EVAL
:
5715 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = fetch_input_tes
;
5716 if (shader
->key
.as_es
)
5717 bld_base
->emit_epilogue
= si_llvm_emit_es_epilogue
;
5719 ctx
->abi
.emit_outputs
= si_llvm_emit_vs_epilogue
;
5720 bld_base
->emit_epilogue
= si_tgsi_emit_epilogue
;
5723 case PIPE_SHADER_GEOMETRY
:
5724 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = fetch_input_gs
;
5725 bld_base
->emit_epilogue
= si_llvm_emit_gs_epilogue
;
5727 case PIPE_SHADER_FRAGMENT
:
5728 ctx
->load_input
= declare_input_fs
;
5729 ctx
->abi
.emit_outputs
= si_llvm_return_fs_outputs
;
5730 bld_base
->emit_epilogue
= si_tgsi_emit_epilogue
;
5732 case PIPE_SHADER_COMPUTE
:
5735 assert(!"Unsupported shader type");
5739 ctx
->abi
.load_ubo
= load_ubo
;
5740 ctx
->abi
.load_ssbo
= load_ssbo
;
5742 create_function(ctx
);
5743 preload_ring_buffers(ctx
);
5745 /* For GFX9 merged shaders:
5746 * - Set EXEC for the first shader. If the prolog is present, set
5747 * EXEC there instead.
5748 * - Add a barrier before the second shader.
5749 * - In the second shader, reset EXEC to ~0 and wrap the main part in
5750 * an if-statement. This is required for correctness in geometry
5751 * shaders, to ensure that empty GS waves do not send GS_EMIT and
5754 * For monolithic merged shaders, the first shader is wrapped in an
5755 * if-block together with its prolog in si_build_wrapper_function.
5757 if (ctx
->screen
->b
.chip_class
>= GFX9
) {
5758 if (!is_monolithic
&&
5759 sel
->info
.num_instructions
> 1 && /* not empty shader */
5760 (shader
->key
.as_es
|| shader
->key
.as_ls
) &&
5761 (ctx
->type
== PIPE_SHADER_TESS_EVAL
||
5762 (ctx
->type
== PIPE_SHADER_VERTEX
&&
5763 !si_vs_needs_prolog(sel
, &shader
->key
.part
.vs
.prolog
)))) {
5764 si_init_exec_from_input(ctx
,
5765 ctx
->param_merged_wave_info
, 0);
5766 } else if (ctx
->type
== PIPE_SHADER_TESS_CTRL
||
5767 ctx
->type
== PIPE_SHADER_GEOMETRY
) {
5769 si_init_exec_full_mask(ctx
);
5771 /* The barrier must execute for all shaders in a
5774 si_llvm_emit_barrier(NULL
, bld_base
, NULL
);
5776 LLVMValueRef num_threads
= unpack_param(ctx
, ctx
->param_merged_wave_info
, 8, 8);
5778 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntULT
,
5779 ac_get_thread_id(&ctx
->ac
), num_threads
, "");
5780 lp_build_if(&ctx
->merged_wrap_if_state
, &ctx
->gallivm
, ena
);
5784 if (ctx
->type
== PIPE_SHADER_TESS_CTRL
&&
5785 sel
->tcs_info
.tessfactors_are_def_in_all_invocs
) {
5786 for (unsigned i
= 0; i
< 6; i
++) {
5787 ctx
->invoc0_tess_factors
[i
] =
5788 lp_build_alloca_undef(&ctx
->gallivm
, ctx
->i32
, "");
5792 if (ctx
->type
== PIPE_SHADER_GEOMETRY
) {
5794 for (i
= 0; i
< 4; i
++) {
5795 ctx
->gs_next_vertex
[i
] =
5796 lp_build_alloca(&ctx
->gallivm
,
5801 if (ctx
->type
== PIPE_SHADER_FRAGMENT
&& sel
->info
.uses_kill
&&
5802 ctx
->screen
->b
.debug_flags
& DBG(FS_CORRECT_DERIVS_AFTER_KILL
)) {
5803 /* This is initialized to 0.0 = not kill. */
5804 ctx
->postponed_kill
= lp_build_alloca(&ctx
->gallivm
, ctx
->f32
, "");
5808 if (!lp_build_tgsi_llvm(bld_base
, sel
->tokens
)) {
5809 fprintf(stderr
, "Failed to translate shader from TGSI to LLVM\n");
5813 if (!si_nir_build_llvm(ctx
, sel
->nir
)) {
5814 fprintf(stderr
, "Failed to translate shader from NIR to LLVM\n");
5819 si_llvm_build_ret(ctx
, ctx
->return_value
);
5824 * Compute the VS prolog key, which contains all the information needed to
5825 * build the VS prolog function, and set shader->info bits where needed.
5827 * \param info Shader info of the vertex shader.
5828 * \param num_input_sgprs Number of input SGPRs for the vertex shader.
5829 * \param prolog_key Key of the VS prolog
5830 * \param shader_out The vertex shader, or the next shader if merging LS+HS or ES+GS.
5831 * \param key Output shader part key.
5833 static void si_get_vs_prolog_key(const struct tgsi_shader_info
*info
,
5834 unsigned num_input_sgprs
,
5835 const struct si_vs_prolog_bits
*prolog_key
,
5836 struct si_shader
*shader_out
,
5837 union si_shader_part_key
*key
)
5839 memset(key
, 0, sizeof(*key
));
5840 key
->vs_prolog
.states
= *prolog_key
;
5841 key
->vs_prolog
.num_input_sgprs
= num_input_sgprs
;
5842 key
->vs_prolog
.last_input
= MAX2(1, info
->num_inputs
) - 1;
5843 key
->vs_prolog
.as_ls
= shader_out
->key
.as_ls
;
5845 if (shader_out
->selector
->type
== PIPE_SHADER_TESS_CTRL
) {
5846 key
->vs_prolog
.as_ls
= 1;
5847 key
->vs_prolog
.num_merged_next_stage_vgprs
= 2;
5848 } else if (shader_out
->selector
->type
== PIPE_SHADER_GEOMETRY
) {
5849 key
->vs_prolog
.num_merged_next_stage_vgprs
= 5;
5852 /* Enable loading the InstanceID VGPR. */
5853 uint16_t input_mask
= u_bit_consecutive(0, info
->num_inputs
);
5855 if ((key
->vs_prolog
.states
.instance_divisor_is_one
|
5856 key
->vs_prolog
.states
.instance_divisor_is_fetched
) & input_mask
)
5857 shader_out
->info
.uses_instanceid
= true;
5861 * Compute the PS prolog key, which contains all the information needed to
5862 * build the PS prolog function, and set related bits in shader->config.
5864 static void si_get_ps_prolog_key(struct si_shader
*shader
,
5865 union si_shader_part_key
*key
,
5866 bool separate_prolog
)
5868 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
5870 memset(key
, 0, sizeof(*key
));
5871 key
->ps_prolog
.states
= shader
->key
.part
.ps
.prolog
;
5872 key
->ps_prolog
.colors_read
= info
->colors_read
;
5873 key
->ps_prolog
.num_input_sgprs
= shader
->info
.num_input_sgprs
;
5874 key
->ps_prolog
.num_input_vgprs
= shader
->info
.num_input_vgprs
;
5875 key
->ps_prolog
.wqm
= info
->uses_derivatives
&&
5876 (key
->ps_prolog
.colors_read
||
5877 key
->ps_prolog
.states
.force_persp_sample_interp
||
5878 key
->ps_prolog
.states
.force_linear_sample_interp
||
5879 key
->ps_prolog
.states
.force_persp_center_interp
||
5880 key
->ps_prolog
.states
.force_linear_center_interp
||
5881 key
->ps_prolog
.states
.bc_optimize_for_persp
||
5882 key
->ps_prolog
.states
.bc_optimize_for_linear
);
5883 key
->ps_prolog
.ancillary_vgpr_index
= shader
->info
.ancillary_vgpr_index
;
5885 if (info
->colors_read
) {
5886 unsigned *color
= shader
->selector
->color_attr_index
;
5888 if (shader
->key
.part
.ps
.prolog
.color_two_side
) {
5889 /* BCOLORs are stored after the last input. */
5890 key
->ps_prolog
.num_interp_inputs
= info
->num_inputs
;
5891 key
->ps_prolog
.face_vgpr_index
= shader
->info
.face_vgpr_index
;
5892 shader
->config
.spi_ps_input_ena
|= S_0286CC_FRONT_FACE_ENA(1);
5895 for (unsigned i
= 0; i
< 2; i
++) {
5896 unsigned interp
= info
->input_interpolate
[color
[i
]];
5897 unsigned location
= info
->input_interpolate_loc
[color
[i
]];
5899 if (!(info
->colors_read
& (0xf << i
*4)))
5902 key
->ps_prolog
.color_attr_index
[i
] = color
[i
];
5904 if (shader
->key
.part
.ps
.prolog
.flatshade_colors
&&
5905 interp
== TGSI_INTERPOLATE_COLOR
)
5906 interp
= TGSI_INTERPOLATE_CONSTANT
;
5909 case TGSI_INTERPOLATE_CONSTANT
:
5910 key
->ps_prolog
.color_interp_vgpr_index
[i
] = -1;
5912 case TGSI_INTERPOLATE_PERSPECTIVE
:
5913 case TGSI_INTERPOLATE_COLOR
:
5914 /* Force the interpolation location for colors here. */
5915 if (shader
->key
.part
.ps
.prolog
.force_persp_sample_interp
)
5916 location
= TGSI_INTERPOLATE_LOC_SAMPLE
;
5917 if (shader
->key
.part
.ps
.prolog
.force_persp_center_interp
)
5918 location
= TGSI_INTERPOLATE_LOC_CENTER
;
5921 case TGSI_INTERPOLATE_LOC_SAMPLE
:
5922 key
->ps_prolog
.color_interp_vgpr_index
[i
] = 0;
5923 shader
->config
.spi_ps_input_ena
|=
5924 S_0286CC_PERSP_SAMPLE_ENA(1);
5926 case TGSI_INTERPOLATE_LOC_CENTER
:
5927 key
->ps_prolog
.color_interp_vgpr_index
[i
] = 2;
5928 shader
->config
.spi_ps_input_ena
|=
5929 S_0286CC_PERSP_CENTER_ENA(1);
5931 case TGSI_INTERPOLATE_LOC_CENTROID
:
5932 key
->ps_prolog
.color_interp_vgpr_index
[i
] = 4;
5933 shader
->config
.spi_ps_input_ena
|=
5934 S_0286CC_PERSP_CENTROID_ENA(1);
5940 case TGSI_INTERPOLATE_LINEAR
:
5941 /* Force the interpolation location for colors here. */
5942 if (shader
->key
.part
.ps
.prolog
.force_linear_sample_interp
)
5943 location
= TGSI_INTERPOLATE_LOC_SAMPLE
;
5944 if (shader
->key
.part
.ps
.prolog
.force_linear_center_interp
)
5945 location
= TGSI_INTERPOLATE_LOC_CENTER
;
5947 /* The VGPR assignment for non-monolithic shaders
5948 * works because InitialPSInputAddr is set on the
5949 * main shader and PERSP_PULL_MODEL is never used.
5952 case TGSI_INTERPOLATE_LOC_SAMPLE
:
5953 key
->ps_prolog
.color_interp_vgpr_index
[i
] =
5954 separate_prolog
? 6 : 9;
5955 shader
->config
.spi_ps_input_ena
|=
5956 S_0286CC_LINEAR_SAMPLE_ENA(1);
5958 case TGSI_INTERPOLATE_LOC_CENTER
:
5959 key
->ps_prolog
.color_interp_vgpr_index
[i
] =
5960 separate_prolog
? 8 : 11;
5961 shader
->config
.spi_ps_input_ena
|=
5962 S_0286CC_LINEAR_CENTER_ENA(1);
5964 case TGSI_INTERPOLATE_LOC_CENTROID
:
5965 key
->ps_prolog
.color_interp_vgpr_index
[i
] =
5966 separate_prolog
? 10 : 13;
5967 shader
->config
.spi_ps_input_ena
|=
5968 S_0286CC_LINEAR_CENTROID_ENA(1);
5982 * Check whether a PS prolog is required based on the key.
5984 static bool si_need_ps_prolog(const union si_shader_part_key
*key
)
5986 return key
->ps_prolog
.colors_read
||
5987 key
->ps_prolog
.states
.force_persp_sample_interp
||
5988 key
->ps_prolog
.states
.force_linear_sample_interp
||
5989 key
->ps_prolog
.states
.force_persp_center_interp
||
5990 key
->ps_prolog
.states
.force_linear_center_interp
||
5991 key
->ps_prolog
.states
.bc_optimize_for_persp
||
5992 key
->ps_prolog
.states
.bc_optimize_for_linear
||
5993 key
->ps_prolog
.states
.poly_stipple
||
5994 key
->ps_prolog
.states
.samplemask_log_ps_iter
;
5998 * Compute the PS epilog key, which contains all the information needed to
5999 * build the PS epilog function.
6001 static void si_get_ps_epilog_key(struct si_shader
*shader
,
6002 union si_shader_part_key
*key
)
6004 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
6005 memset(key
, 0, sizeof(*key
));
6006 key
->ps_epilog
.colors_written
= info
->colors_written
;
6007 key
->ps_epilog
.writes_z
= info
->writes_z
;
6008 key
->ps_epilog
.writes_stencil
= info
->writes_stencil
;
6009 key
->ps_epilog
.writes_samplemask
= info
->writes_samplemask
;
6010 key
->ps_epilog
.states
= shader
->key
.part
.ps
.epilog
;
6014 * Build the GS prolog function. Rotate the input vertices for triangle strips
6017 static void si_build_gs_prolog_function(struct si_shader_context
*ctx
,
6018 union si_shader_part_key
*key
)
6020 unsigned num_sgprs
, num_vgprs
;
6021 struct si_function_info fninfo
;
6022 LLVMBuilderRef builder
= ctx
->ac
.builder
;
6023 LLVMTypeRef returns
[48];
6024 LLVMValueRef func
, ret
;
6026 si_init_function_info(&fninfo
);
6028 if (ctx
->screen
->b
.chip_class
>= GFX9
) {
6029 num_sgprs
= 8 + GFX9_GS_NUM_USER_SGPR
;
6030 num_vgprs
= 5; /* ES inputs are not needed by GS */
6032 num_sgprs
= GFX6_GS_NUM_USER_SGPR
+ 2;
6036 for (unsigned i
= 0; i
< num_sgprs
; ++i
) {
6037 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
6038 returns
[i
] = ctx
->i32
;
6041 for (unsigned i
= 0; i
< num_vgprs
; ++i
) {
6042 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
6043 returns
[num_sgprs
+ i
] = ctx
->f32
;
6046 /* Create the function. */
6047 si_create_function(ctx
, "gs_prolog", returns
, num_sgprs
+ num_vgprs
,
6049 func
= ctx
->main_fn
;
6051 /* Set the full EXEC mask for the prolog, because we are only fiddling
6052 * with registers here. The main shader part will set the correct EXEC
6055 if (ctx
->screen
->b
.chip_class
>= GFX9
&& !key
->gs_prolog
.is_monolithic
)
6056 si_init_exec_full_mask(ctx
);
6058 /* Copy inputs to outputs. This should be no-op, as the registers match,
6059 * but it will prevent the compiler from overwriting them unintentionally.
6061 ret
= ctx
->return_value
;
6062 for (unsigned i
= 0; i
< num_sgprs
; i
++) {
6063 LLVMValueRef p
= LLVMGetParam(func
, i
);
6064 ret
= LLVMBuildInsertValue(builder
, ret
, p
, i
, "");
6066 for (unsigned i
= 0; i
< num_vgprs
; i
++) {
6067 LLVMValueRef p
= LLVMGetParam(func
, num_sgprs
+ i
);
6068 p
= ac_to_float(&ctx
->ac
, p
);
6069 ret
= LLVMBuildInsertValue(builder
, ret
, p
, num_sgprs
+ i
, "");
6072 if (key
->gs_prolog
.states
.tri_strip_adj_fix
) {
6073 /* Remap the input vertices for every other primitive. */
6074 const unsigned gfx6_vtx_params
[6] = {
6082 const unsigned gfx9_vtx_params
[3] = {
6087 LLVMValueRef vtx_in
[6], vtx_out
[6];
6088 LLVMValueRef prim_id
, rotate
;
6090 if (ctx
->screen
->b
.chip_class
>= GFX9
) {
6091 for (unsigned i
= 0; i
< 3; i
++) {
6092 vtx_in
[i
*2] = unpack_param(ctx
, gfx9_vtx_params
[i
], 0, 16);
6093 vtx_in
[i
*2+1] = unpack_param(ctx
, gfx9_vtx_params
[i
], 16, 16);
6096 for (unsigned i
= 0; i
< 6; i
++)
6097 vtx_in
[i
] = LLVMGetParam(func
, gfx6_vtx_params
[i
]);
6100 prim_id
= LLVMGetParam(func
, num_sgprs
+ 2);
6101 rotate
= LLVMBuildTrunc(builder
, prim_id
, ctx
->i1
, "");
6103 for (unsigned i
= 0; i
< 6; ++i
) {
6104 LLVMValueRef base
, rotated
;
6106 rotated
= vtx_in
[(i
+ 4) % 6];
6107 vtx_out
[i
] = LLVMBuildSelect(builder
, rotate
, rotated
, base
, "");
6110 if (ctx
->screen
->b
.chip_class
>= GFX9
) {
6111 for (unsigned i
= 0; i
< 3; i
++) {
6112 LLVMValueRef hi
, out
;
6114 hi
= LLVMBuildShl(builder
, vtx_out
[i
*2+1],
6115 LLVMConstInt(ctx
->i32
, 16, 0), "");
6116 out
= LLVMBuildOr(builder
, vtx_out
[i
*2], hi
, "");
6117 out
= ac_to_float(&ctx
->ac
, out
);
6118 ret
= LLVMBuildInsertValue(builder
, ret
, out
,
6119 gfx9_vtx_params
[i
], "");
6122 for (unsigned i
= 0; i
< 6; i
++) {
6125 out
= ac_to_float(&ctx
->ac
, vtx_out
[i
]);
6126 ret
= LLVMBuildInsertValue(builder
, ret
, out
,
6127 gfx6_vtx_params
[i
], "");
6132 LLVMBuildRet(builder
, ret
);
6136 * Given a list of shader part functions, build a wrapper function that
6137 * runs them in sequence to form a monolithic shader.
6139 static void si_build_wrapper_function(struct si_shader_context
*ctx
,
6140 LLVMValueRef
*parts
,
6143 unsigned next_shader_first_part
)
6145 LLVMBuilderRef builder
= ctx
->ac
.builder
;
6146 /* PS epilog has one arg per color component; gfx9 merged shader
6147 * prologs need to forward 32 user SGPRs.
6149 struct si_function_info fninfo
;
6150 LLVMValueRef initial
[64], out
[64];
6151 LLVMTypeRef function_type
;
6152 unsigned num_first_params
;
6153 unsigned num_out
, initial_num_out
;
6154 MAYBE_UNUSED
unsigned num_out_sgpr
; /* used in debug checks */
6155 MAYBE_UNUSED
unsigned initial_num_out_sgpr
; /* used in debug checks */
6156 unsigned num_sgprs
, num_vgprs
;
6158 struct lp_build_if_state if_state
;
6160 si_init_function_info(&fninfo
);
6162 for (unsigned i
= 0; i
< num_parts
; ++i
) {
6163 lp_add_function_attr(parts
[i
], -1, LP_FUNC_ATTR_ALWAYSINLINE
);
6164 LLVMSetLinkage(parts
[i
], LLVMPrivateLinkage
);
6167 /* The parameters of the wrapper function correspond to those of the
6168 * first part in terms of SGPRs and VGPRs, but we use the types of the
6169 * main part to get the right types. This is relevant for the
6170 * dereferenceable attribute on descriptor table pointers.
6175 function_type
= LLVMGetElementType(LLVMTypeOf(parts
[0]));
6176 num_first_params
= LLVMCountParamTypes(function_type
);
6178 for (unsigned i
= 0; i
< num_first_params
; ++i
) {
6179 LLVMValueRef param
= LLVMGetParam(parts
[0], i
);
6181 if (ac_is_sgpr_param(param
)) {
6182 assert(num_vgprs
== 0);
6183 num_sgprs
+= ac_get_type_size(LLVMTypeOf(param
)) / 4;
6185 num_vgprs
+= ac_get_type_size(LLVMTypeOf(param
)) / 4;
6190 while (gprs
< num_sgprs
+ num_vgprs
) {
6191 LLVMValueRef param
= LLVMGetParam(parts
[main_part
], fninfo
.num_params
);
6192 LLVMTypeRef type
= LLVMTypeOf(param
);
6193 unsigned size
= ac_get_type_size(type
) / 4;
6195 add_arg(&fninfo
, gprs
< num_sgprs
? ARG_SGPR
: ARG_VGPR
, type
);
6197 assert(ac_is_sgpr_param(param
) == (gprs
< num_sgprs
));
6198 assert(gprs
+ size
<= num_sgprs
+ num_vgprs
&&
6199 (gprs
>= num_sgprs
|| gprs
+ size
<= num_sgprs
));
6204 si_create_function(ctx
, "wrapper", NULL
, 0, &fninfo
,
6205 si_get_max_workgroup_size(ctx
->shader
));
6207 if (is_merged_shader(ctx
->shader
))
6208 si_init_exec_full_mask(ctx
);
6210 /* Record the arguments of the function as if they were an output of
6216 for (unsigned i
= 0; i
< fninfo
.num_params
; ++i
) {
6217 LLVMValueRef param
= LLVMGetParam(ctx
->main_fn
, i
);
6218 LLVMTypeRef param_type
= LLVMTypeOf(param
);
6219 LLVMTypeRef out_type
= i
< fninfo
.num_sgpr_params
? ctx
->i32
: ctx
->f32
;
6220 unsigned size
= ac_get_type_size(param_type
) / 4;
6223 if (param_type
!= out_type
)
6224 param
= LLVMBuildBitCast(builder
, param
, out_type
, "");
6225 out
[num_out
++] = param
;
6227 LLVMTypeRef vector_type
= LLVMVectorType(out_type
, size
);
6229 if (LLVMGetTypeKind(param_type
) == LLVMPointerTypeKind
) {
6230 param
= LLVMBuildPtrToInt(builder
, param
, ctx
->i64
, "");
6231 param_type
= ctx
->i64
;
6234 if (param_type
!= vector_type
)
6235 param
= LLVMBuildBitCast(builder
, param
, vector_type
, "");
6237 for (unsigned j
= 0; j
< size
; ++j
)
6238 out
[num_out
++] = LLVMBuildExtractElement(
6239 builder
, param
, LLVMConstInt(ctx
->i32
, j
, 0), "");
6242 if (i
< fninfo
.num_sgpr_params
)
6243 num_out_sgpr
= num_out
;
6246 memcpy(initial
, out
, sizeof(out
));
6247 initial_num_out
= num_out
;
6248 initial_num_out_sgpr
= num_out_sgpr
;
6250 /* Now chain the parts. */
6251 for (unsigned part
= 0; part
< num_parts
; ++part
) {
6252 LLVMValueRef in
[48];
6254 LLVMTypeRef ret_type
;
6255 unsigned out_idx
= 0;
6256 unsigned num_params
= LLVMCountParams(parts
[part
]);
6258 /* Merged shaders are executed conditionally depending
6259 * on the number of enabled threads passed in the input SGPRs. */
6260 if (is_merged_shader(ctx
->shader
) && part
== 0) {
6261 LLVMValueRef ena
, count
= initial
[3];
6263 count
= LLVMBuildAnd(builder
, count
,
6264 LLVMConstInt(ctx
->i32
, 0x7f, 0), "");
6265 ena
= LLVMBuildICmp(builder
, LLVMIntULT
,
6266 ac_get_thread_id(&ctx
->ac
), count
, "");
6267 lp_build_if(&if_state
, &ctx
->gallivm
, ena
);
6270 /* Derive arguments for the next part from outputs of the
6273 for (unsigned param_idx
= 0; param_idx
< num_params
; ++param_idx
) {
6275 LLVMTypeRef param_type
;
6277 unsigned param_size
;
6278 LLVMValueRef arg
= NULL
;
6280 param
= LLVMGetParam(parts
[part
], param_idx
);
6281 param_type
= LLVMTypeOf(param
);
6282 param_size
= ac_get_type_size(param_type
) / 4;
6283 is_sgpr
= ac_is_sgpr_param(param
);
6286 #if HAVE_LLVM < 0x0400
6287 LLVMRemoveAttribute(param
, LLVMByValAttribute
);
6289 unsigned kind_id
= LLVMGetEnumAttributeKindForName("byval", 5);
6290 LLVMRemoveEnumAttributeAtIndex(parts
[part
], param_idx
+ 1, kind_id
);
6292 lp_add_function_attr(parts
[part
], param_idx
+ 1, LP_FUNC_ATTR_INREG
);
6295 assert(out_idx
+ param_size
<= (is_sgpr
? num_out_sgpr
: num_out
));
6296 assert(is_sgpr
|| out_idx
>= num_out_sgpr
);
6298 if (param_size
== 1)
6301 arg
= lp_build_gather_values(&ctx
->gallivm
, &out
[out_idx
], param_size
);
6303 if (LLVMTypeOf(arg
) != param_type
) {
6304 if (LLVMGetTypeKind(param_type
) == LLVMPointerTypeKind
) {
6305 arg
= LLVMBuildBitCast(builder
, arg
, ctx
->i64
, "");
6306 arg
= LLVMBuildIntToPtr(builder
, arg
, param_type
, "");
6308 arg
= LLVMBuildBitCast(builder
, arg
, param_type
, "");
6312 in
[param_idx
] = arg
;
6313 out_idx
+= param_size
;
6316 ret
= LLVMBuildCall(builder
, parts
[part
], in
, num_params
, "");
6318 if (is_merged_shader(ctx
->shader
) &&
6319 part
+ 1 == next_shader_first_part
) {
6320 lp_build_endif(&if_state
);
6322 /* The second half of the merged shader should use
6323 * the inputs from the toplevel (wrapper) function,
6324 * not the return value from the last call.
6326 * That's because the last call was executed condi-
6327 * tionally, so we can't consume it in the main
6330 memcpy(out
, initial
, sizeof(initial
));
6331 num_out
= initial_num_out
;
6332 num_out_sgpr
= initial_num_out_sgpr
;
6336 /* Extract the returned GPRs. */
6337 ret_type
= LLVMTypeOf(ret
);
6341 if (LLVMGetTypeKind(ret_type
) != LLVMVoidTypeKind
) {
6342 assert(LLVMGetTypeKind(ret_type
) == LLVMStructTypeKind
);
6344 unsigned ret_size
= LLVMCountStructElementTypes(ret_type
);
6346 for (unsigned i
= 0; i
< ret_size
; ++i
) {
6348 LLVMBuildExtractValue(builder
, ret
, i
, "");
6350 assert(num_out
< ARRAY_SIZE(out
));
6351 out
[num_out
++] = val
;
6353 if (LLVMTypeOf(val
) == ctx
->i32
) {
6354 assert(num_out_sgpr
+ 1 == num_out
);
6355 num_out_sgpr
= num_out
;
6361 LLVMBuildRetVoid(builder
);
6364 int si_compile_tgsi_shader(struct si_screen
*sscreen
,
6365 LLVMTargetMachineRef tm
,
6366 struct si_shader
*shader
,
6368 struct pipe_debug_callback
*debug
)
6370 struct si_shader_selector
*sel
= shader
->selector
;
6371 struct si_shader_context ctx
;
6374 /* Dump TGSI code before doing TGSI->LLVM conversion in case the
6375 * conversion fails. */
6376 if (si_can_dump_shader(&sscreen
->b
, sel
->info
.processor
) &&
6377 !(sscreen
->b
.debug_flags
& DBG(NO_TGSI
))) {
6379 tgsi_dump(sel
->tokens
, 0);
6381 nir_print_shader(sel
->nir
, stderr
);
6382 si_dump_streamout(&sel
->so
);
6385 si_init_shader_ctx(&ctx
, sscreen
, tm
);
6386 si_llvm_context_set_tgsi(&ctx
, shader
);
6387 ctx
.separate_prolog
= !is_monolithic
;
6389 memset(shader
->info
.vs_output_param_offset
, AC_EXP_PARAM_UNDEFINED
,
6390 sizeof(shader
->info
.vs_output_param_offset
));
6392 shader
->info
.uses_instanceid
= sel
->info
.uses_instanceid
;
6394 if (!si_compile_tgsi_main(&ctx
, is_monolithic
)) {
6395 si_llvm_dispose(&ctx
);
6399 if (is_monolithic
&& ctx
.type
== PIPE_SHADER_VERTEX
) {
6400 LLVMValueRef parts
[2];
6401 bool need_prolog
= sel
->vs_needs_prolog
;
6403 parts
[1] = ctx
.main_fn
;
6406 union si_shader_part_key prolog_key
;
6407 si_get_vs_prolog_key(&sel
->info
,
6408 shader
->info
.num_input_sgprs
,
6409 &shader
->key
.part
.vs
.prolog
,
6410 shader
, &prolog_key
);
6411 si_build_vs_prolog_function(&ctx
, &prolog_key
);
6412 parts
[0] = ctx
.main_fn
;
6415 si_build_wrapper_function(&ctx
, parts
+ !need_prolog
,
6416 1 + need_prolog
, need_prolog
, 0);
6417 } else if (is_monolithic
&& ctx
.type
== PIPE_SHADER_TESS_CTRL
) {
6418 if (sscreen
->b
.chip_class
>= GFX9
) {
6419 struct si_shader_selector
*ls
= shader
->key
.part
.tcs
.ls
;
6420 LLVMValueRef parts
[4];
6421 bool vs_needs_prolog
=
6422 si_vs_needs_prolog(ls
, &shader
->key
.part
.tcs
.ls_prolog
);
6425 parts
[2] = ctx
.main_fn
;
6428 union si_shader_part_key tcs_epilog_key
;
6429 memset(&tcs_epilog_key
, 0, sizeof(tcs_epilog_key
));
6430 tcs_epilog_key
.tcs_epilog
.states
= shader
->key
.part
.tcs
.epilog
;
6431 si_build_tcs_epilog_function(&ctx
, &tcs_epilog_key
);
6432 parts
[3] = ctx
.main_fn
;
6435 if (vs_needs_prolog
) {
6436 union si_shader_part_key vs_prolog_key
;
6437 si_get_vs_prolog_key(&ls
->info
,
6438 shader
->info
.num_input_sgprs
,
6439 &shader
->key
.part
.tcs
.ls_prolog
,
6440 shader
, &vs_prolog_key
);
6441 vs_prolog_key
.vs_prolog
.is_monolithic
= true;
6442 si_build_vs_prolog_function(&ctx
, &vs_prolog_key
);
6443 parts
[0] = ctx
.main_fn
;
6446 /* VS as LS main part */
6447 struct si_shader shader_ls
= {};
6448 shader_ls
.selector
= ls
;
6449 shader_ls
.key
.as_ls
= 1;
6450 shader_ls
.key
.mono
= shader
->key
.mono
;
6451 shader_ls
.key
.opt
= shader
->key
.opt
;
6452 si_llvm_context_set_tgsi(&ctx
, &shader_ls
);
6454 if (!si_compile_tgsi_main(&ctx
, true)) {
6455 si_llvm_dispose(&ctx
);
6458 shader
->info
.uses_instanceid
|= ls
->info
.uses_instanceid
;
6459 parts
[1] = ctx
.main_fn
;
6461 /* Reset the shader context. */
6462 ctx
.shader
= shader
;
6463 ctx
.type
= PIPE_SHADER_TESS_CTRL
;
6465 si_build_wrapper_function(&ctx
,
6466 parts
+ !vs_needs_prolog
,
6467 4 - !vs_needs_prolog
, 0,
6468 vs_needs_prolog
? 2 : 1);
6470 LLVMValueRef parts
[2];
6471 union si_shader_part_key epilog_key
;
6473 parts
[0] = ctx
.main_fn
;
6475 memset(&epilog_key
, 0, sizeof(epilog_key
));
6476 epilog_key
.tcs_epilog
.states
= shader
->key
.part
.tcs
.epilog
;
6477 si_build_tcs_epilog_function(&ctx
, &epilog_key
);
6478 parts
[1] = ctx
.main_fn
;
6480 si_build_wrapper_function(&ctx
, parts
, 2, 0, 0);
6482 } else if (is_monolithic
&& ctx
.type
== PIPE_SHADER_GEOMETRY
) {
6483 if (ctx
.screen
->b
.chip_class
>= GFX9
) {
6484 struct si_shader_selector
*es
= shader
->key
.part
.gs
.es
;
6485 LLVMValueRef es_prolog
= NULL
;
6486 LLVMValueRef es_main
= NULL
;
6487 LLVMValueRef gs_prolog
= NULL
;
6488 LLVMValueRef gs_main
= ctx
.main_fn
;
6491 union si_shader_part_key gs_prolog_key
;
6492 memset(&gs_prolog_key
, 0, sizeof(gs_prolog_key
));
6493 gs_prolog_key
.gs_prolog
.states
= shader
->key
.part
.gs
.prolog
;
6494 gs_prolog_key
.gs_prolog
.is_monolithic
= true;
6495 si_build_gs_prolog_function(&ctx
, &gs_prolog_key
);
6496 gs_prolog
= ctx
.main_fn
;
6499 if (es
->vs_needs_prolog
) {
6500 union si_shader_part_key vs_prolog_key
;
6501 si_get_vs_prolog_key(&es
->info
,
6502 shader
->info
.num_input_sgprs
,
6503 &shader
->key
.part
.tcs
.ls_prolog
,
6504 shader
, &vs_prolog_key
);
6505 vs_prolog_key
.vs_prolog
.is_monolithic
= true;
6506 si_build_vs_prolog_function(&ctx
, &vs_prolog_key
);
6507 es_prolog
= ctx
.main_fn
;
6511 struct si_shader shader_es
= {};
6512 shader_es
.selector
= es
;
6513 shader_es
.key
.as_es
= 1;
6514 shader_es
.key
.mono
= shader
->key
.mono
;
6515 shader_es
.key
.opt
= shader
->key
.opt
;
6516 si_llvm_context_set_tgsi(&ctx
, &shader_es
);
6518 if (!si_compile_tgsi_main(&ctx
, true)) {
6519 si_llvm_dispose(&ctx
);
6522 shader
->info
.uses_instanceid
|= es
->info
.uses_instanceid
;
6523 es_main
= ctx
.main_fn
;
6525 /* Reset the shader context. */
6526 ctx
.shader
= shader
;
6527 ctx
.type
= PIPE_SHADER_GEOMETRY
;
6529 /* Prepare the array of shader parts. */
6530 LLVMValueRef parts
[4];
6531 unsigned num_parts
= 0, main_part
, next_first_part
;
6534 parts
[num_parts
++] = es_prolog
;
6536 parts
[main_part
= num_parts
++] = es_main
;
6537 parts
[next_first_part
= num_parts
++] = gs_prolog
;
6538 parts
[num_parts
++] = gs_main
;
6540 si_build_wrapper_function(&ctx
, parts
, num_parts
,
6541 main_part
, next_first_part
);
6543 LLVMValueRef parts
[2];
6544 union si_shader_part_key prolog_key
;
6546 parts
[1] = ctx
.main_fn
;
6548 memset(&prolog_key
, 0, sizeof(prolog_key
));
6549 prolog_key
.gs_prolog
.states
= shader
->key
.part
.gs
.prolog
;
6550 si_build_gs_prolog_function(&ctx
, &prolog_key
);
6551 parts
[0] = ctx
.main_fn
;
6553 si_build_wrapper_function(&ctx
, parts
, 2, 1, 0);
6555 } else if (is_monolithic
&& ctx
.type
== PIPE_SHADER_FRAGMENT
) {
6556 LLVMValueRef parts
[3];
6557 union si_shader_part_key prolog_key
;
6558 union si_shader_part_key epilog_key
;
6561 si_get_ps_prolog_key(shader
, &prolog_key
, false);
6562 need_prolog
= si_need_ps_prolog(&prolog_key
);
6564 parts
[need_prolog
? 1 : 0] = ctx
.main_fn
;
6567 si_build_ps_prolog_function(&ctx
, &prolog_key
);
6568 parts
[0] = ctx
.main_fn
;
6571 si_get_ps_epilog_key(shader
, &epilog_key
);
6572 si_build_ps_epilog_function(&ctx
, &epilog_key
);
6573 parts
[need_prolog
? 2 : 1] = ctx
.main_fn
;
6575 si_build_wrapper_function(&ctx
, parts
, need_prolog
? 3 : 2,
6576 need_prolog
? 1 : 0, 0);
6579 si_llvm_optimize_module(&ctx
);
6581 /* Post-optimization transformations and analysis. */
6582 si_optimize_vs_outputs(&ctx
);
6584 if ((debug
&& debug
->debug_message
) ||
6585 si_can_dump_shader(&sscreen
->b
, ctx
.type
))
6586 si_count_scratch_private_memory(&ctx
);
6588 /* Compile to bytecode. */
6589 r
= si_compile_llvm(sscreen
, &shader
->binary
, &shader
->config
, tm
,
6590 ctx
.gallivm
.module
, debug
, ctx
.type
, "TGSI shader");
6591 si_llvm_dispose(&ctx
);
6593 fprintf(stderr
, "LLVM failed to compile shader\n");
6597 /* Validate SGPR and VGPR usage for compute to detect compiler bugs.
6598 * LLVM 3.9svn has this bug.
6600 if (sel
->type
== PIPE_SHADER_COMPUTE
) {
6601 unsigned wave_size
= 64;
6602 unsigned max_vgprs
= 256;
6603 unsigned max_sgprs
= sscreen
->b
.chip_class
>= VI
? 800 : 512;
6604 unsigned max_sgprs_per_wave
= 128;
6605 unsigned max_block_threads
= si_get_max_workgroup_size(shader
);
6606 unsigned min_waves_per_cu
= DIV_ROUND_UP(max_block_threads
, wave_size
);
6607 unsigned min_waves_per_simd
= DIV_ROUND_UP(min_waves_per_cu
, 4);
6609 max_vgprs
= max_vgprs
/ min_waves_per_simd
;
6610 max_sgprs
= MIN2(max_sgprs
/ min_waves_per_simd
, max_sgprs_per_wave
);
6612 if (shader
->config
.num_sgprs
> max_sgprs
||
6613 shader
->config
.num_vgprs
> max_vgprs
) {
6614 fprintf(stderr
, "LLVM failed to compile a shader correctly: "
6615 "SGPR:VGPR usage is %u:%u, but the hw limit is %u:%u\n",
6616 shader
->config
.num_sgprs
, shader
->config
.num_vgprs
,
6617 max_sgprs
, max_vgprs
);
6619 /* Just terminate the process, because dependent
6620 * shaders can hang due to bad input data, but use
6621 * the env var to allow shader-db to work.
6623 if (!debug_get_bool_option("SI_PASS_BAD_SHADERS", false))
6628 /* Add the scratch offset to input SGPRs. */
6629 if (shader
->config
.scratch_bytes_per_wave
&& !is_merged_shader(shader
))
6630 shader
->info
.num_input_sgprs
+= 1; /* scratch byte offset */
6632 /* Calculate the number of fragment input VGPRs. */
6633 if (ctx
.type
== PIPE_SHADER_FRAGMENT
) {
6634 shader
->info
.num_input_vgprs
= 0;
6635 shader
->info
.face_vgpr_index
= -1;
6636 shader
->info
.ancillary_vgpr_index
= -1;
6638 if (G_0286CC_PERSP_SAMPLE_ENA(shader
->config
.spi_ps_input_addr
))
6639 shader
->info
.num_input_vgprs
+= 2;
6640 if (G_0286CC_PERSP_CENTER_ENA(shader
->config
.spi_ps_input_addr
))
6641 shader
->info
.num_input_vgprs
+= 2;
6642 if (G_0286CC_PERSP_CENTROID_ENA(shader
->config
.spi_ps_input_addr
))
6643 shader
->info
.num_input_vgprs
+= 2;
6644 if (G_0286CC_PERSP_PULL_MODEL_ENA(shader
->config
.spi_ps_input_addr
))
6645 shader
->info
.num_input_vgprs
+= 3;
6646 if (G_0286CC_LINEAR_SAMPLE_ENA(shader
->config
.spi_ps_input_addr
))
6647 shader
->info
.num_input_vgprs
+= 2;
6648 if (G_0286CC_LINEAR_CENTER_ENA(shader
->config
.spi_ps_input_addr
))
6649 shader
->info
.num_input_vgprs
+= 2;
6650 if (G_0286CC_LINEAR_CENTROID_ENA(shader
->config
.spi_ps_input_addr
))
6651 shader
->info
.num_input_vgprs
+= 2;
6652 if (G_0286CC_LINE_STIPPLE_TEX_ENA(shader
->config
.spi_ps_input_addr
))
6653 shader
->info
.num_input_vgprs
+= 1;
6654 if (G_0286CC_POS_X_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
6655 shader
->info
.num_input_vgprs
+= 1;
6656 if (G_0286CC_POS_Y_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
6657 shader
->info
.num_input_vgprs
+= 1;
6658 if (G_0286CC_POS_Z_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
6659 shader
->info
.num_input_vgprs
+= 1;
6660 if (G_0286CC_POS_W_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
6661 shader
->info
.num_input_vgprs
+= 1;
6662 if (G_0286CC_FRONT_FACE_ENA(shader
->config
.spi_ps_input_addr
)) {
6663 shader
->info
.face_vgpr_index
= shader
->info
.num_input_vgprs
;
6664 shader
->info
.num_input_vgprs
+= 1;
6666 if (G_0286CC_ANCILLARY_ENA(shader
->config
.spi_ps_input_addr
)) {
6667 shader
->info
.ancillary_vgpr_index
= shader
->info
.num_input_vgprs
;
6668 shader
->info
.num_input_vgprs
+= 1;
6670 if (G_0286CC_SAMPLE_COVERAGE_ENA(shader
->config
.spi_ps_input_addr
))
6671 shader
->info
.num_input_vgprs
+= 1;
6672 if (G_0286CC_POS_FIXED_PT_ENA(shader
->config
.spi_ps_input_addr
))
6673 shader
->info
.num_input_vgprs
+= 1;
6680 * Create, compile and return a shader part (prolog or epilog).
6682 * \param sscreen screen
6683 * \param list list of shader parts of the same category
6684 * \param type shader type
6685 * \param key shader part key
6686 * \param prolog whether the part being requested is a prolog
6687 * \param tm LLVM target machine
6688 * \param debug debug callback
6689 * \param build the callback responsible for building the main function
6690 * \return non-NULL on success
6692 static struct si_shader_part
*
6693 si_get_shader_part(struct si_screen
*sscreen
,
6694 struct si_shader_part
**list
,
6695 enum pipe_shader_type type
,
6697 union si_shader_part_key
*key
,
6698 LLVMTargetMachineRef tm
,
6699 struct pipe_debug_callback
*debug
,
6700 void (*build
)(struct si_shader_context
*,
6701 union si_shader_part_key
*),
6704 struct si_shader_part
*result
;
6706 mtx_lock(&sscreen
->shader_parts_mutex
);
6708 /* Find existing. */
6709 for (result
= *list
; result
; result
= result
->next
) {
6710 if (memcmp(&result
->key
, key
, sizeof(*key
)) == 0) {
6711 mtx_unlock(&sscreen
->shader_parts_mutex
);
6716 /* Compile a new one. */
6717 result
= CALLOC_STRUCT(si_shader_part
);
6720 struct si_shader shader
= {};
6721 struct si_shader_context ctx
;
6723 si_init_shader_ctx(&ctx
, sscreen
, tm
);
6724 ctx
.shader
= &shader
;
6728 case PIPE_SHADER_VERTEX
:
6730 case PIPE_SHADER_TESS_CTRL
:
6732 shader
.key
.part
.tcs
.epilog
= key
->tcs_epilog
.states
;
6734 case PIPE_SHADER_GEOMETRY
:
6737 case PIPE_SHADER_FRAGMENT
:
6739 shader
.key
.part
.ps
.prolog
= key
->ps_prolog
.states
;
6741 shader
.key
.part
.ps
.epilog
= key
->ps_epilog
.states
;
6744 unreachable("bad shader part");
6750 si_llvm_optimize_module(&ctx
);
6752 if (si_compile_llvm(sscreen
, &result
->binary
, &result
->config
, tm
,
6753 ctx
.ac
.module
, debug
, ctx
.type
, name
)) {
6759 result
->next
= *list
;
6763 si_llvm_dispose(&ctx
);
6764 mtx_unlock(&sscreen
->shader_parts_mutex
);
6768 static LLVMValueRef
si_prolog_get_rw_buffers(struct si_shader_context
*ctx
)
6770 LLVMValueRef ptr
[2], list
;
6772 /* Get the pointer to rw buffers. */
6773 ptr
[0] = LLVMGetParam(ctx
->main_fn
, SI_SGPR_RW_BUFFERS
);
6774 ptr
[1] = LLVMGetParam(ctx
->main_fn
, SI_SGPR_RW_BUFFERS_HI
);
6775 list
= lp_build_gather_values(&ctx
->gallivm
, ptr
, 2);
6776 list
= LLVMBuildBitCast(ctx
->ac
.builder
, list
, ctx
->i64
, "");
6777 list
= LLVMBuildIntToPtr(ctx
->ac
.builder
, list
,
6778 si_const_array(ctx
->v4i32
, SI_NUM_RW_BUFFERS
), "");
6783 * Build the vertex shader prolog function.
6785 * The inputs are the same as VS (a lot of SGPRs and 4 VGPR system values).
6786 * All inputs are returned unmodified. The vertex load indices are
6787 * stored after them, which will be used by the API VS for fetching inputs.
6789 * For example, the expected outputs for instance_divisors[] = {0, 1, 2} are:
6794 * (VertexID + BaseVertex),
6795 * (InstanceID + StartInstance),
6796 * (InstanceID / 2 + StartInstance)
6798 static void si_build_vs_prolog_function(struct si_shader_context
*ctx
,
6799 union si_shader_part_key
*key
)
6801 struct si_function_info fninfo
;
6802 LLVMTypeRef
*returns
;
6803 LLVMValueRef ret
, func
;
6805 unsigned first_vs_vgpr
= key
->vs_prolog
.num_merged_next_stage_vgprs
;
6806 unsigned num_input_vgprs
= key
->vs_prolog
.num_merged_next_stage_vgprs
+ 4;
6807 LLVMValueRef input_vgprs
[9];
6808 unsigned num_all_input_regs
= key
->vs_prolog
.num_input_sgprs
+
6810 unsigned user_sgpr_base
= key
->vs_prolog
.num_merged_next_stage_vgprs
? 8 : 0;
6812 si_init_function_info(&fninfo
);
6814 /* 4 preloaded VGPRs + vertex load indices as prolog outputs */
6815 returns
= alloca((num_all_input_regs
+ key
->vs_prolog
.last_input
+ 1) *
6816 sizeof(LLVMTypeRef
));
6819 /* Declare input and output SGPRs. */
6820 for (i
= 0; i
< key
->vs_prolog
.num_input_sgprs
; i
++) {
6821 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
6822 returns
[num_returns
++] = ctx
->i32
;
6825 /* Preloaded VGPRs (outputs must be floats) */
6826 for (i
= 0; i
< num_input_vgprs
; i
++) {
6827 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &input_vgprs
[i
]);
6828 returns
[num_returns
++] = ctx
->f32
;
6831 /* Vertex load indices. */
6832 for (i
= 0; i
<= key
->vs_prolog
.last_input
; i
++)
6833 returns
[num_returns
++] = ctx
->f32
;
6835 /* Create the function. */
6836 si_create_function(ctx
, "vs_prolog", returns
, num_returns
, &fninfo
, 0);
6837 func
= ctx
->main_fn
;
6839 if (key
->vs_prolog
.num_merged_next_stage_vgprs
) {
6840 if (!key
->vs_prolog
.is_monolithic
)
6841 si_init_exec_from_input(ctx
, 3, 0);
6843 if (key
->vs_prolog
.as_ls
&&
6844 (ctx
->screen
->b
.family
== CHIP_VEGA10
||
6845 ctx
->screen
->b
.family
== CHIP_RAVEN
)) {
6846 /* If there are no HS threads, SPI loads the LS VGPRs
6847 * starting at VGPR 0. Shift them back to where they
6850 LLVMValueRef has_hs_threads
=
6851 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntNE
,
6852 unpack_param(ctx
, 3, 8, 8),
6855 for (i
= 4; i
> 0; --i
) {
6856 input_vgprs
[i
+ 1] =
6857 LLVMBuildSelect(ctx
->ac
.builder
, has_hs_threads
,
6859 input_vgprs
[i
- 1], "");
6864 ctx
->abi
.vertex_id
= input_vgprs
[first_vs_vgpr
];
6865 ctx
->abi
.instance_id
= input_vgprs
[first_vs_vgpr
+ (key
->vs_prolog
.as_ls
? 2 : 1)];
6867 /* Copy inputs to outputs. This should be no-op, as the registers match,
6868 * but it will prevent the compiler from overwriting them unintentionally.
6870 ret
= ctx
->return_value
;
6871 for (i
= 0; i
< key
->vs_prolog
.num_input_sgprs
; i
++) {
6872 LLVMValueRef p
= LLVMGetParam(func
, i
);
6873 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, p
, i
, "");
6875 for (i
= 0; i
< num_input_vgprs
; i
++) {
6876 LLVMValueRef p
= input_vgprs
[i
];
6877 p
= ac_to_float(&ctx
->ac
, p
);
6878 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, p
,
6879 key
->vs_prolog
.num_input_sgprs
+ i
, "");
6882 /* Compute vertex load indices from instance divisors. */
6883 LLVMValueRef instance_divisor_constbuf
= NULL
;
6885 if (key
->vs_prolog
.states
.instance_divisor_is_fetched
) {
6886 LLVMValueRef list
= si_prolog_get_rw_buffers(ctx
);
6887 LLVMValueRef buf_index
=
6888 LLVMConstInt(ctx
->i32
, SI_VS_CONST_INSTANCE_DIVISORS
, 0);
6889 instance_divisor_constbuf
=
6890 ac_build_indexed_load_const(&ctx
->ac
, list
, buf_index
);
6893 for (i
= 0; i
<= key
->vs_prolog
.last_input
; i
++) {
6894 bool divisor_is_one
=
6895 key
->vs_prolog
.states
.instance_divisor_is_one
& (1u << i
);
6896 bool divisor_is_fetched
=
6897 key
->vs_prolog
.states
.instance_divisor_is_fetched
& (1u << i
);
6900 if (divisor_is_one
|| divisor_is_fetched
) {
6901 LLVMValueRef divisor
= ctx
->i32_1
;
6903 if (divisor_is_fetched
) {
6904 divisor
= buffer_load_const(ctx
, instance_divisor_constbuf
,
6905 LLVMConstInt(ctx
->i32
, i
* 4, 0));
6906 divisor
= ac_to_integer(&ctx
->ac
, divisor
);
6909 /* InstanceID / Divisor + StartInstance */
6910 index
= get_instance_index_for_fetch(ctx
,
6912 SI_SGPR_START_INSTANCE
,
6915 /* VertexID + BaseVertex */
6916 index
= LLVMBuildAdd(ctx
->ac
.builder
,
6918 LLVMGetParam(func
, user_sgpr_base
+
6919 SI_SGPR_BASE_VERTEX
), "");
6922 index
= ac_to_float(&ctx
->ac
, index
);
6923 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, index
,
6924 fninfo
.num_params
+ i
, "");
6927 si_llvm_build_ret(ctx
, ret
);
6930 static bool si_get_vs_prolog(struct si_screen
*sscreen
,
6931 LLVMTargetMachineRef tm
,
6932 struct si_shader
*shader
,
6933 struct pipe_debug_callback
*debug
,
6934 struct si_shader
*main_part
,
6935 const struct si_vs_prolog_bits
*key
)
6937 struct si_shader_selector
*vs
= main_part
->selector
;
6939 if (!si_vs_needs_prolog(vs
, key
))
6942 /* Get the prolog. */
6943 union si_shader_part_key prolog_key
;
6944 si_get_vs_prolog_key(&vs
->info
, main_part
->info
.num_input_sgprs
,
6945 key
, shader
, &prolog_key
);
6948 si_get_shader_part(sscreen
, &sscreen
->vs_prologs
,
6949 PIPE_SHADER_VERTEX
, true, &prolog_key
, tm
,
6950 debug
, si_build_vs_prolog_function
,
6951 "Vertex Shader Prolog");
6952 return shader
->prolog
!= NULL
;
6956 * Select and compile (or reuse) vertex shader parts (prolog & epilog).
6958 static bool si_shader_select_vs_parts(struct si_screen
*sscreen
,
6959 LLVMTargetMachineRef tm
,
6960 struct si_shader
*shader
,
6961 struct pipe_debug_callback
*debug
)
6963 return si_get_vs_prolog(sscreen
, tm
, shader
, debug
, shader
,
6964 &shader
->key
.part
.vs
.prolog
);
6968 * Compile the TCS epilog function. This writes tesselation factors to memory
6969 * based on the output primitive type of the tesselator (determined by TES).
6971 static void si_build_tcs_epilog_function(struct si_shader_context
*ctx
,
6972 union si_shader_part_key
*key
)
6974 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
6975 struct si_function_info fninfo
;
6978 si_init_function_info(&fninfo
);
6980 if (ctx
->screen
->b
.chip_class
>= GFX9
) {
6981 add_arg(&fninfo
, ARG_SGPR
, ctx
->i64
);
6982 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
6983 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* wave info */
6984 ctx
->param_tcs_factor_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
6985 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
6986 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
6987 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
6988 add_arg(&fninfo
, ARG_SGPR
, ctx
->i64
);
6989 add_arg(&fninfo
, ARG_SGPR
, ctx
->i64
);
6990 add_arg(&fninfo
, ARG_SGPR
, ctx
->i64
);
6991 add_arg(&fninfo
, ARG_SGPR
, ctx
->i64
);
6992 add_arg(&fninfo
, ARG_SGPR
, ctx
->i64
);
6993 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
6994 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
6995 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
6996 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
6997 ctx
->param_tcs_offchip_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
6998 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
6999 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7000 ctx
->param_tcs_offchip_addr_base64k
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7001 ctx
->param_tcs_factor_addr_base64k
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7003 add_arg(&fninfo
, ARG_SGPR
, ctx
->i64
);
7004 add_arg(&fninfo
, ARG_SGPR
, ctx
->i64
);
7005 add_arg(&fninfo
, ARG_SGPR
, ctx
->i64
);
7006 add_arg(&fninfo
, ARG_SGPR
, ctx
->i64
);
7007 ctx
->param_tcs_offchip_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7008 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7009 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7010 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7011 ctx
->param_tcs_offchip_addr_base64k
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7012 ctx
->param_tcs_factor_addr_base64k
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7013 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7014 ctx
->param_tcs_factor_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7017 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
); /* VGPR gap */
7018 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
); /* VGPR gap */
7019 unsigned tess_factors_idx
=
7020 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
); /* patch index within the wave (REL_PATCH_ID) */
7021 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
); /* invocation ID within the patch */
7022 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
); /* LDS offset where tess factors should be loaded from */
7024 for (unsigned i
= 0; i
< 6; i
++)
7025 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
); /* tess factors */
7027 /* Create the function. */
7028 si_create_function(ctx
, "tcs_epilog", NULL
, 0, &fninfo
,
7029 ctx
->screen
->b
.chip_class
>= CIK
? 128 : 64);
7030 declare_lds_as_pointer(ctx
);
7031 func
= ctx
->main_fn
;
7033 LLVMValueRef invoc0_tess_factors
[6];
7034 for (unsigned i
= 0; i
< 6; i
++)
7035 invoc0_tess_factors
[i
] = LLVMGetParam(func
, tess_factors_idx
+ 3 + i
);
7037 si_write_tess_factors(bld_base
,
7038 LLVMGetParam(func
, tess_factors_idx
),
7039 LLVMGetParam(func
, tess_factors_idx
+ 1),
7040 LLVMGetParam(func
, tess_factors_idx
+ 2),
7041 invoc0_tess_factors
, invoc0_tess_factors
+ 4);
7043 LLVMBuildRetVoid(ctx
->ac
.builder
);
7047 * Select and compile (or reuse) TCS parts (epilog).
7049 static bool si_shader_select_tcs_parts(struct si_screen
*sscreen
,
7050 LLVMTargetMachineRef tm
,
7051 struct si_shader
*shader
,
7052 struct pipe_debug_callback
*debug
)
7054 if (sscreen
->b
.chip_class
>= GFX9
) {
7055 struct si_shader
*ls_main_part
=
7056 shader
->key
.part
.tcs
.ls
->main_shader_part_ls
;
7058 if (!si_get_vs_prolog(sscreen
, tm
, shader
, debug
, ls_main_part
,
7059 &shader
->key
.part
.tcs
.ls_prolog
))
7062 shader
->previous_stage
= ls_main_part
;
7065 /* Get the epilog. */
7066 union si_shader_part_key epilog_key
;
7067 memset(&epilog_key
, 0, sizeof(epilog_key
));
7068 epilog_key
.tcs_epilog
.states
= shader
->key
.part
.tcs
.epilog
;
7070 shader
->epilog
= si_get_shader_part(sscreen
, &sscreen
->tcs_epilogs
,
7071 PIPE_SHADER_TESS_CTRL
, false,
7072 &epilog_key
, tm
, debug
,
7073 si_build_tcs_epilog_function
,
7074 "Tessellation Control Shader Epilog");
7075 return shader
->epilog
!= NULL
;
7079 * Select and compile (or reuse) GS parts (prolog).
7081 static bool si_shader_select_gs_parts(struct si_screen
*sscreen
,
7082 LLVMTargetMachineRef tm
,
7083 struct si_shader
*shader
,
7084 struct pipe_debug_callback
*debug
)
7086 if (sscreen
->b
.chip_class
>= GFX9
) {
7087 struct si_shader
*es_main_part
=
7088 shader
->key
.part
.gs
.es
->main_shader_part_es
;
7090 if (shader
->key
.part
.gs
.es
->type
== PIPE_SHADER_VERTEX
&&
7091 !si_get_vs_prolog(sscreen
, tm
, shader
, debug
, es_main_part
,
7092 &shader
->key
.part
.gs
.vs_prolog
))
7095 shader
->previous_stage
= es_main_part
;
7098 if (!shader
->key
.part
.gs
.prolog
.tri_strip_adj_fix
)
7101 union si_shader_part_key prolog_key
;
7102 memset(&prolog_key
, 0, sizeof(prolog_key
));
7103 prolog_key
.gs_prolog
.states
= shader
->key
.part
.gs
.prolog
;
7105 shader
->prolog2
= si_get_shader_part(sscreen
, &sscreen
->gs_prologs
,
7106 PIPE_SHADER_GEOMETRY
, true,
7107 &prolog_key
, tm
, debug
,
7108 si_build_gs_prolog_function
,
7109 "Geometry Shader Prolog");
7110 return shader
->prolog2
!= NULL
;
7114 * Build the pixel shader prolog function. This handles:
7115 * - two-side color selection and interpolation
7116 * - overriding interpolation parameters for the API PS
7117 * - polygon stippling
7119 * All preloaded SGPRs and VGPRs are passed through unmodified unless they are
7120 * overriden by other states. (e.g. per-sample interpolation)
7121 * Interpolated colors are stored after the preloaded VGPRs.
7123 static void si_build_ps_prolog_function(struct si_shader_context
*ctx
,
7124 union si_shader_part_key
*key
)
7126 struct si_function_info fninfo
;
7127 LLVMValueRef ret
, func
;
7128 int num_returns
, i
, num_color_channels
;
7130 assert(si_need_ps_prolog(key
));
7132 si_init_function_info(&fninfo
);
7134 /* Declare inputs. */
7135 for (i
= 0; i
< key
->ps_prolog
.num_input_sgprs
; i
++)
7136 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7138 for (i
= 0; i
< key
->ps_prolog
.num_input_vgprs
; i
++)
7139 add_arg(&fninfo
, ARG_VGPR
, ctx
->f32
);
7141 /* Declare outputs (same as inputs + add colors if needed) */
7142 num_returns
= fninfo
.num_params
;
7143 num_color_channels
= util_bitcount(key
->ps_prolog
.colors_read
);
7144 for (i
= 0; i
< num_color_channels
; i
++)
7145 fninfo
.types
[num_returns
++] = ctx
->f32
;
7147 /* Create the function. */
7148 si_create_function(ctx
, "ps_prolog", fninfo
.types
, num_returns
,
7150 func
= ctx
->main_fn
;
7152 /* Copy inputs to outputs. This should be no-op, as the registers match,
7153 * but it will prevent the compiler from overwriting them unintentionally.
7155 ret
= ctx
->return_value
;
7156 for (i
= 0; i
< fninfo
.num_params
; i
++) {
7157 LLVMValueRef p
= LLVMGetParam(func
, i
);
7158 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, p
, i
, "");
7161 /* Polygon stippling. */
7162 if (key
->ps_prolog
.states
.poly_stipple
) {
7163 /* POS_FIXED_PT is always last. */
7164 unsigned pos
= key
->ps_prolog
.num_input_sgprs
+
7165 key
->ps_prolog
.num_input_vgprs
- 1;
7166 LLVMValueRef list
= si_prolog_get_rw_buffers(ctx
);
7168 si_llvm_emit_polygon_stipple(ctx
, list
, pos
);
7171 if (key
->ps_prolog
.states
.bc_optimize_for_persp
||
7172 key
->ps_prolog
.states
.bc_optimize_for_linear
) {
7173 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7174 LLVMValueRef center
[2], centroid
[2], tmp
, bc_optimize
;
7176 /* The shader should do: if (PRIM_MASK[31]) CENTROID = CENTER;
7177 * The hw doesn't compute CENTROID if the whole wave only
7178 * contains fully-covered quads.
7180 * PRIM_MASK is after user SGPRs.
7182 bc_optimize
= LLVMGetParam(func
, SI_PS_NUM_USER_SGPR
);
7183 bc_optimize
= LLVMBuildLShr(ctx
->ac
.builder
, bc_optimize
,
7184 LLVMConstInt(ctx
->i32
, 31, 0), "");
7185 bc_optimize
= LLVMBuildTrunc(ctx
->ac
.builder
, bc_optimize
,
7188 if (key
->ps_prolog
.states
.bc_optimize_for_persp
) {
7189 /* Read PERSP_CENTER. */
7190 for (i
= 0; i
< 2; i
++)
7191 center
[i
] = LLVMGetParam(func
, base
+ 2 + i
);
7192 /* Read PERSP_CENTROID. */
7193 for (i
= 0; i
< 2; i
++)
7194 centroid
[i
] = LLVMGetParam(func
, base
+ 4 + i
);
7195 /* Select PERSP_CENTROID. */
7196 for (i
= 0; i
< 2; i
++) {
7197 tmp
= LLVMBuildSelect(ctx
->ac
.builder
, bc_optimize
,
7198 center
[i
], centroid
[i
], "");
7199 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7200 tmp
, base
+ 4 + i
, "");
7203 if (key
->ps_prolog
.states
.bc_optimize_for_linear
) {
7204 /* Read LINEAR_CENTER. */
7205 for (i
= 0; i
< 2; i
++)
7206 center
[i
] = LLVMGetParam(func
, base
+ 8 + i
);
7207 /* Read LINEAR_CENTROID. */
7208 for (i
= 0; i
< 2; i
++)
7209 centroid
[i
] = LLVMGetParam(func
, base
+ 10 + i
);
7210 /* Select LINEAR_CENTROID. */
7211 for (i
= 0; i
< 2; i
++) {
7212 tmp
= LLVMBuildSelect(ctx
->ac
.builder
, bc_optimize
,
7213 center
[i
], centroid
[i
], "");
7214 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7215 tmp
, base
+ 10 + i
, "");
7220 /* Force per-sample interpolation. */
7221 if (key
->ps_prolog
.states
.force_persp_sample_interp
) {
7222 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7223 LLVMValueRef persp_sample
[2];
7225 /* Read PERSP_SAMPLE. */
7226 for (i
= 0; i
< 2; i
++)
7227 persp_sample
[i
] = LLVMGetParam(func
, base
+ i
);
7228 /* Overwrite PERSP_CENTER. */
7229 for (i
= 0; i
< 2; i
++)
7230 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7231 persp_sample
[i
], base
+ 2 + i
, "");
7232 /* Overwrite PERSP_CENTROID. */
7233 for (i
= 0; i
< 2; i
++)
7234 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7235 persp_sample
[i
], base
+ 4 + i
, "");
7237 if (key
->ps_prolog
.states
.force_linear_sample_interp
) {
7238 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7239 LLVMValueRef linear_sample
[2];
7241 /* Read LINEAR_SAMPLE. */
7242 for (i
= 0; i
< 2; i
++)
7243 linear_sample
[i
] = LLVMGetParam(func
, base
+ 6 + i
);
7244 /* Overwrite LINEAR_CENTER. */
7245 for (i
= 0; i
< 2; i
++)
7246 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7247 linear_sample
[i
], base
+ 8 + i
, "");
7248 /* Overwrite LINEAR_CENTROID. */
7249 for (i
= 0; i
< 2; i
++)
7250 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7251 linear_sample
[i
], base
+ 10 + i
, "");
7254 /* Force center interpolation. */
7255 if (key
->ps_prolog
.states
.force_persp_center_interp
) {
7256 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7257 LLVMValueRef persp_center
[2];
7259 /* Read PERSP_CENTER. */
7260 for (i
= 0; i
< 2; i
++)
7261 persp_center
[i
] = LLVMGetParam(func
, base
+ 2 + i
);
7262 /* Overwrite PERSP_SAMPLE. */
7263 for (i
= 0; i
< 2; i
++)
7264 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7265 persp_center
[i
], base
+ i
, "");
7266 /* Overwrite PERSP_CENTROID. */
7267 for (i
= 0; i
< 2; i
++)
7268 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7269 persp_center
[i
], base
+ 4 + i
, "");
7271 if (key
->ps_prolog
.states
.force_linear_center_interp
) {
7272 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7273 LLVMValueRef linear_center
[2];
7275 /* Read LINEAR_CENTER. */
7276 for (i
= 0; i
< 2; i
++)
7277 linear_center
[i
] = LLVMGetParam(func
, base
+ 8 + i
);
7278 /* Overwrite LINEAR_SAMPLE. */
7279 for (i
= 0; i
< 2; i
++)
7280 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7281 linear_center
[i
], base
+ 6 + i
, "");
7282 /* Overwrite LINEAR_CENTROID. */
7283 for (i
= 0; i
< 2; i
++)
7284 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7285 linear_center
[i
], base
+ 10 + i
, "");
7288 /* Interpolate colors. */
7289 unsigned color_out_idx
= 0;
7290 for (i
= 0; i
< 2; i
++) {
7291 unsigned writemask
= (key
->ps_prolog
.colors_read
>> (i
* 4)) & 0xf;
7292 unsigned face_vgpr
= key
->ps_prolog
.num_input_sgprs
+
7293 key
->ps_prolog
.face_vgpr_index
;
7294 LLVMValueRef interp
[2], color
[4];
7295 LLVMValueRef interp_ij
= NULL
, prim_mask
= NULL
, face
= NULL
;
7300 /* If the interpolation qualifier is not CONSTANT (-1). */
7301 if (key
->ps_prolog
.color_interp_vgpr_index
[i
] != -1) {
7302 unsigned interp_vgpr
= key
->ps_prolog
.num_input_sgprs
+
7303 key
->ps_prolog
.color_interp_vgpr_index
[i
];
7305 /* Get the (i,j) updated by bc_optimize handling. */
7306 interp
[0] = LLVMBuildExtractValue(ctx
->ac
.builder
, ret
,
7308 interp
[1] = LLVMBuildExtractValue(ctx
->ac
.builder
, ret
,
7309 interp_vgpr
+ 1, "");
7310 interp_ij
= lp_build_gather_values(&ctx
->gallivm
, interp
, 2);
7313 /* Use the absolute location of the input. */
7314 prim_mask
= LLVMGetParam(func
, SI_PS_NUM_USER_SGPR
);
7316 if (key
->ps_prolog
.states
.color_two_side
) {
7317 face
= LLVMGetParam(func
, face_vgpr
);
7318 face
= ac_to_integer(&ctx
->ac
, face
);
7321 interp_fs_input(ctx
,
7322 key
->ps_prolog
.color_attr_index
[i
],
7323 TGSI_SEMANTIC_COLOR
, i
,
7324 key
->ps_prolog
.num_interp_inputs
,
7325 key
->ps_prolog
.colors_read
, interp_ij
,
7326 prim_mask
, face
, color
);
7329 unsigned chan
= u_bit_scan(&writemask
);
7330 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, color
[chan
],
7331 fninfo
.num_params
+ color_out_idx
++, "");
7335 /* Section 15.2.2 (Shader Inputs) of the OpenGL 4.5 (Core Profile) spec
7338 * "When per-sample shading is active due to the use of a fragment
7339 * input qualified by sample or due to the use of the gl_SampleID
7340 * or gl_SamplePosition variables, only the bit for the current
7341 * sample is set in gl_SampleMaskIn. When state specifies multiple
7342 * fragment shader invocations for a given fragment, the sample
7343 * mask for any single fragment shader invocation may specify a
7344 * subset of the covered samples for the fragment. In this case,
7345 * the bit corresponding to each covered sample will be set in
7346 * exactly one fragment shader invocation."
7348 * The samplemask loaded by hardware is always the coverage of the
7349 * entire pixel/fragment, so mask bits out based on the sample ID.
7351 if (key
->ps_prolog
.states
.samplemask_log_ps_iter
) {
7352 /* The bit pattern matches that used by fixed function fragment
7354 static const uint16_t ps_iter_masks
[] = {
7355 0xffff, /* not used */
7361 assert(key
->ps_prolog
.states
.samplemask_log_ps_iter
< ARRAY_SIZE(ps_iter_masks
));
7363 uint32_t ps_iter_mask
= ps_iter_masks
[key
->ps_prolog
.states
.samplemask_log_ps_iter
];
7364 unsigned ancillary_vgpr
= key
->ps_prolog
.num_input_sgprs
+
7365 key
->ps_prolog
.ancillary_vgpr_index
;
7366 LLVMValueRef sampleid
= unpack_param(ctx
, ancillary_vgpr
, 8, 4);
7367 LLVMValueRef samplemask
= LLVMGetParam(func
, ancillary_vgpr
+ 1);
7369 samplemask
= ac_to_integer(&ctx
->ac
, samplemask
);
7370 samplemask
= LLVMBuildAnd(
7373 LLVMBuildShl(ctx
->ac
.builder
,
7374 LLVMConstInt(ctx
->i32
, ps_iter_mask
, false),
7377 samplemask
= ac_to_float(&ctx
->ac
, samplemask
);
7379 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, samplemask
,
7380 ancillary_vgpr
+ 1, "");
7383 /* Tell LLVM to insert WQM instruction sequence when needed. */
7384 if (key
->ps_prolog
.wqm
) {
7385 LLVMAddTargetDependentFunctionAttr(func
,
7386 "amdgpu-ps-wqm-outputs", "");
7389 si_llvm_build_ret(ctx
, ret
);
7393 * Build the pixel shader epilog function. This handles everything that must be
7394 * emulated for pixel shader exports. (alpha-test, format conversions, etc)
7396 static void si_build_ps_epilog_function(struct si_shader_context
*ctx
,
7397 union si_shader_part_key
*key
)
7399 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
7400 struct si_function_info fninfo
;
7401 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
7403 struct si_ps_exports exp
= {};
7405 si_init_function_info(&fninfo
);
7407 /* Declare input SGPRs. */
7408 ctx
->param_rw_buffers
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i64
);
7409 ctx
->param_bindless_samplers_and_images
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i64
);
7410 ctx
->param_const_and_shader_buffers
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i64
);
7411 ctx
->param_samplers_and_images
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i64
);
7412 add_arg_checked(&fninfo
, ARG_SGPR
, ctx
->f32
, SI_PARAM_ALPHA_REF
);
7414 /* Declare input VGPRs. */
7415 unsigned required_num_params
=
7416 fninfo
.num_sgpr_params
+
7417 util_bitcount(key
->ps_epilog
.colors_written
) * 4 +
7418 key
->ps_epilog
.writes_z
+
7419 key
->ps_epilog
.writes_stencil
+
7420 key
->ps_epilog
.writes_samplemask
;
7422 required_num_params
= MAX2(required_num_params
,
7423 fninfo
.num_sgpr_params
+ PS_EPILOG_SAMPLEMASK_MIN_LOC
+ 1);
7425 while (fninfo
.num_params
< required_num_params
)
7426 add_arg(&fninfo
, ARG_VGPR
, ctx
->f32
);
7428 /* Create the function. */
7429 si_create_function(ctx
, "ps_epilog", NULL
, 0, &fninfo
, 0);
7430 /* Disable elimination of unused inputs. */
7431 si_llvm_add_attribute(ctx
->main_fn
,
7432 "InitialPSInputAddr", 0xffffff);
7434 /* Process colors. */
7435 unsigned vgpr
= fninfo
.num_sgpr_params
;
7436 unsigned colors_written
= key
->ps_epilog
.colors_written
;
7437 int last_color_export
= -1;
7439 /* Find the last color export. */
7440 if (!key
->ps_epilog
.writes_z
&&
7441 !key
->ps_epilog
.writes_stencil
&&
7442 !key
->ps_epilog
.writes_samplemask
) {
7443 unsigned spi_format
= key
->ps_epilog
.states
.spi_shader_col_format
;
7445 /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
7446 if (colors_written
== 0x1 && key
->ps_epilog
.states
.last_cbuf
> 0) {
7447 /* Just set this if any of the colorbuffers are enabled. */
7449 ((1ull << (4 * (key
->ps_epilog
.states
.last_cbuf
+ 1))) - 1))
7450 last_color_export
= 0;
7452 for (i
= 0; i
< 8; i
++)
7453 if (colors_written
& (1 << i
) &&
7454 (spi_format
>> (i
* 4)) & 0xf)
7455 last_color_export
= i
;
7459 while (colors_written
) {
7460 LLVMValueRef color
[4];
7461 int mrt
= u_bit_scan(&colors_written
);
7463 for (i
= 0; i
< 4; i
++)
7464 color
[i
] = LLVMGetParam(ctx
->main_fn
, vgpr
++);
7466 si_export_mrt_color(bld_base
, color
, mrt
,
7467 fninfo
.num_params
- 1,
7468 mrt
== last_color_export
, &exp
);
7471 /* Process depth, stencil, samplemask. */
7472 if (key
->ps_epilog
.writes_z
)
7473 depth
= LLVMGetParam(ctx
->main_fn
, vgpr
++);
7474 if (key
->ps_epilog
.writes_stencil
)
7475 stencil
= LLVMGetParam(ctx
->main_fn
, vgpr
++);
7476 if (key
->ps_epilog
.writes_samplemask
)
7477 samplemask
= LLVMGetParam(ctx
->main_fn
, vgpr
++);
7479 if (depth
|| stencil
|| samplemask
)
7480 si_export_mrt_z(bld_base
, depth
, stencil
, samplemask
, &exp
);
7481 else if (last_color_export
== -1)
7482 si_export_null(bld_base
);
7485 si_emit_ps_exports(ctx
, &exp
);
7488 LLVMBuildRetVoid(ctx
->ac
.builder
);
7492 * Select and compile (or reuse) pixel shader parts (prolog & epilog).
7494 static bool si_shader_select_ps_parts(struct si_screen
*sscreen
,
7495 LLVMTargetMachineRef tm
,
7496 struct si_shader
*shader
,
7497 struct pipe_debug_callback
*debug
)
7499 union si_shader_part_key prolog_key
;
7500 union si_shader_part_key epilog_key
;
7502 /* Get the prolog. */
7503 si_get_ps_prolog_key(shader
, &prolog_key
, true);
7505 /* The prolog is a no-op if these aren't set. */
7506 if (si_need_ps_prolog(&prolog_key
)) {
7508 si_get_shader_part(sscreen
, &sscreen
->ps_prologs
,
7509 PIPE_SHADER_FRAGMENT
, true,
7510 &prolog_key
, tm
, debug
,
7511 si_build_ps_prolog_function
,
7512 "Fragment Shader Prolog");
7513 if (!shader
->prolog
)
7517 /* Get the epilog. */
7518 si_get_ps_epilog_key(shader
, &epilog_key
);
7521 si_get_shader_part(sscreen
, &sscreen
->ps_epilogs
,
7522 PIPE_SHADER_FRAGMENT
, false,
7523 &epilog_key
, tm
, debug
,
7524 si_build_ps_epilog_function
,
7525 "Fragment Shader Epilog");
7526 if (!shader
->epilog
)
7529 /* Enable POS_FIXED_PT if polygon stippling is enabled. */
7530 if (shader
->key
.part
.ps
.prolog
.poly_stipple
) {
7531 shader
->config
.spi_ps_input_ena
|= S_0286CC_POS_FIXED_PT_ENA(1);
7532 assert(G_0286CC_POS_FIXED_PT_ENA(shader
->config
.spi_ps_input_addr
));
7535 /* Set up the enable bits for per-sample shading if needed. */
7536 if (shader
->key
.part
.ps
.prolog
.force_persp_sample_interp
&&
7537 (G_0286CC_PERSP_CENTER_ENA(shader
->config
.spi_ps_input_ena
) ||
7538 G_0286CC_PERSP_CENTROID_ENA(shader
->config
.spi_ps_input_ena
))) {
7539 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_CENTER_ENA
;
7540 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_CENTROID_ENA
;
7541 shader
->config
.spi_ps_input_ena
|= S_0286CC_PERSP_SAMPLE_ENA(1);
7543 if (shader
->key
.part
.ps
.prolog
.force_linear_sample_interp
&&
7544 (G_0286CC_LINEAR_CENTER_ENA(shader
->config
.spi_ps_input_ena
) ||
7545 G_0286CC_LINEAR_CENTROID_ENA(shader
->config
.spi_ps_input_ena
))) {
7546 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_CENTER_ENA
;
7547 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_CENTROID_ENA
;
7548 shader
->config
.spi_ps_input_ena
|= S_0286CC_LINEAR_SAMPLE_ENA(1);
7550 if (shader
->key
.part
.ps
.prolog
.force_persp_center_interp
&&
7551 (G_0286CC_PERSP_SAMPLE_ENA(shader
->config
.spi_ps_input_ena
) ||
7552 G_0286CC_PERSP_CENTROID_ENA(shader
->config
.spi_ps_input_ena
))) {
7553 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_SAMPLE_ENA
;
7554 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_CENTROID_ENA
;
7555 shader
->config
.spi_ps_input_ena
|= S_0286CC_PERSP_CENTER_ENA(1);
7557 if (shader
->key
.part
.ps
.prolog
.force_linear_center_interp
&&
7558 (G_0286CC_LINEAR_SAMPLE_ENA(shader
->config
.spi_ps_input_ena
) ||
7559 G_0286CC_LINEAR_CENTROID_ENA(shader
->config
.spi_ps_input_ena
))) {
7560 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_SAMPLE_ENA
;
7561 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_CENTROID_ENA
;
7562 shader
->config
.spi_ps_input_ena
|= S_0286CC_LINEAR_CENTER_ENA(1);
7565 /* POW_W_FLOAT requires that one of the perspective weights is enabled. */
7566 if (G_0286CC_POS_W_FLOAT_ENA(shader
->config
.spi_ps_input_ena
) &&
7567 !(shader
->config
.spi_ps_input_ena
& 0xf)) {
7568 shader
->config
.spi_ps_input_ena
|= S_0286CC_PERSP_CENTER_ENA(1);
7569 assert(G_0286CC_PERSP_CENTER_ENA(shader
->config
.spi_ps_input_addr
));
7572 /* At least one pair of interpolation weights must be enabled. */
7573 if (!(shader
->config
.spi_ps_input_ena
& 0x7f)) {
7574 shader
->config
.spi_ps_input_ena
|= S_0286CC_LINEAR_CENTER_ENA(1);
7575 assert(G_0286CC_LINEAR_CENTER_ENA(shader
->config
.spi_ps_input_addr
));
7578 /* Samplemask fixup requires the sample ID. */
7579 if (shader
->key
.part
.ps
.prolog
.samplemask_log_ps_iter
) {
7580 shader
->config
.spi_ps_input_ena
|= S_0286CC_ANCILLARY_ENA(1);
7581 assert(G_0286CC_ANCILLARY_ENA(shader
->config
.spi_ps_input_addr
));
7584 /* The sample mask input is always enabled, because the API shader always
7585 * passes it through to the epilog. Disable it here if it's unused.
7587 if (!shader
->key
.part
.ps
.epilog
.poly_line_smoothing
&&
7588 !shader
->selector
->info
.reads_samplemask
)
7589 shader
->config
.spi_ps_input_ena
&= C_0286CC_SAMPLE_COVERAGE_ENA
;
7594 void si_multiwave_lds_size_workaround(struct si_screen
*sscreen
,
7597 /* SPI barrier management bug:
7598 * Make sure we have at least 4k of LDS in use to avoid the bug.
7599 * It applies to workgroup sizes of more than one wavefront.
7601 if (sscreen
->b
.family
== CHIP_BONAIRE
||
7602 sscreen
->b
.family
== CHIP_KABINI
||
7603 sscreen
->b
.family
== CHIP_MULLINS
)
7604 *lds_size
= MAX2(*lds_size
, 8);
7607 static void si_fix_resource_usage(struct si_screen
*sscreen
,
7608 struct si_shader
*shader
)
7610 unsigned min_sgprs
= shader
->info
.num_input_sgprs
+ 2; /* VCC */
7612 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
, min_sgprs
);
7614 if (shader
->selector
->type
== PIPE_SHADER_COMPUTE
&&
7615 si_get_max_workgroup_size(shader
) > 64) {
7616 si_multiwave_lds_size_workaround(sscreen
,
7617 &shader
->config
.lds_size
);
7621 int si_shader_create(struct si_screen
*sscreen
, LLVMTargetMachineRef tm
,
7622 struct si_shader
*shader
,
7623 struct pipe_debug_callback
*debug
)
7625 struct si_shader_selector
*sel
= shader
->selector
;
7626 struct si_shader
*mainp
= *si_get_main_shader_part(sel
, &shader
->key
);
7629 /* LS, ES, VS are compiled on demand if the main part hasn't been
7630 * compiled for that stage.
7632 * Vertex shaders are compiled on demand when a vertex fetch
7633 * workaround must be applied.
7635 if (shader
->is_monolithic
) {
7636 /* Monolithic shader (compiled as a whole, has many variants,
7637 * may take a long time to compile).
7639 r
= si_compile_tgsi_shader(sscreen
, tm
, shader
, true, debug
);
7643 /* The shader consists of several parts:
7645 * - the middle part is the user shader, it has 1 variant only
7646 * and it was compiled during the creation of the shader
7648 * - the prolog part is inserted at the beginning
7649 * - the epilog part is inserted at the end
7651 * The prolog and epilog have many (but simple) variants.
7653 * Starting with gfx9, geometry and tessellation control
7654 * shaders also contain the prolog and user shader parts of
7655 * the previous shader stage.
7661 /* Copy the compiled TGSI shader data over. */
7662 shader
->is_binary_shared
= true;
7663 shader
->binary
= mainp
->binary
;
7664 shader
->config
= mainp
->config
;
7665 shader
->info
.num_input_sgprs
= mainp
->info
.num_input_sgprs
;
7666 shader
->info
.num_input_vgprs
= mainp
->info
.num_input_vgprs
;
7667 shader
->info
.face_vgpr_index
= mainp
->info
.face_vgpr_index
;
7668 shader
->info
.ancillary_vgpr_index
= mainp
->info
.ancillary_vgpr_index
;
7669 memcpy(shader
->info
.vs_output_param_offset
,
7670 mainp
->info
.vs_output_param_offset
,
7671 sizeof(mainp
->info
.vs_output_param_offset
));
7672 shader
->info
.uses_instanceid
= mainp
->info
.uses_instanceid
;
7673 shader
->info
.nr_pos_exports
= mainp
->info
.nr_pos_exports
;
7674 shader
->info
.nr_param_exports
= mainp
->info
.nr_param_exports
;
7676 /* Select prologs and/or epilogs. */
7677 switch (sel
->type
) {
7678 case PIPE_SHADER_VERTEX
:
7679 if (!si_shader_select_vs_parts(sscreen
, tm
, shader
, debug
))
7682 case PIPE_SHADER_TESS_CTRL
:
7683 if (!si_shader_select_tcs_parts(sscreen
, tm
, shader
, debug
))
7686 case PIPE_SHADER_TESS_EVAL
:
7688 case PIPE_SHADER_GEOMETRY
:
7689 if (!si_shader_select_gs_parts(sscreen
, tm
, shader
, debug
))
7692 case PIPE_SHADER_FRAGMENT
:
7693 if (!si_shader_select_ps_parts(sscreen
, tm
, shader
, debug
))
7696 /* Make sure we have at least as many VGPRs as there
7697 * are allocated inputs.
7699 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
7700 shader
->info
.num_input_vgprs
);
7704 /* Update SGPR and VGPR counts. */
7705 if (shader
->prolog
) {
7706 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
,
7707 shader
->prolog
->config
.num_sgprs
);
7708 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
7709 shader
->prolog
->config
.num_vgprs
);
7711 if (shader
->previous_stage
) {
7712 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
,
7713 shader
->previous_stage
->config
.num_sgprs
);
7714 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
7715 shader
->previous_stage
->config
.num_vgprs
);
7716 shader
->config
.spilled_sgprs
=
7717 MAX2(shader
->config
.spilled_sgprs
,
7718 shader
->previous_stage
->config
.spilled_sgprs
);
7719 shader
->config
.spilled_vgprs
=
7720 MAX2(shader
->config
.spilled_vgprs
,
7721 shader
->previous_stage
->config
.spilled_vgprs
);
7722 shader
->config
.private_mem_vgprs
=
7723 MAX2(shader
->config
.private_mem_vgprs
,
7724 shader
->previous_stage
->config
.private_mem_vgprs
);
7725 shader
->config
.scratch_bytes_per_wave
=
7726 MAX2(shader
->config
.scratch_bytes_per_wave
,
7727 shader
->previous_stage
->config
.scratch_bytes_per_wave
);
7728 shader
->info
.uses_instanceid
|=
7729 shader
->previous_stage
->info
.uses_instanceid
;
7731 if (shader
->prolog2
) {
7732 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
,
7733 shader
->prolog2
->config
.num_sgprs
);
7734 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
7735 shader
->prolog2
->config
.num_vgprs
);
7737 if (shader
->epilog
) {
7738 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
,
7739 shader
->epilog
->config
.num_sgprs
);
7740 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
7741 shader
->epilog
->config
.num_vgprs
);
7745 si_fix_resource_usage(sscreen
, shader
);
7746 si_shader_dump(sscreen
, shader
, debug
, sel
->info
.processor
,
7750 r
= si_shader_binary_upload(sscreen
, shader
);
7752 fprintf(stderr
, "LLVM failed to upload shader\n");
7759 void si_shader_destroy(struct si_shader
*shader
)
7761 if (shader
->scratch_bo
)
7762 r600_resource_reference(&shader
->scratch_bo
, NULL
);
7764 r600_resource_reference(&shader
->bo
, NULL
);
7766 if (!shader
->is_binary_shared
)
7767 si_radeon_shader_binary_clean(&shader
->binary
);
7769 free(shader
->shader_log
);