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 #include "gallivm/lp_bld_const.h"
25 #include "gallivm/lp_bld_gather.h"
26 #include "gallivm/lp_bld_intr.h"
27 #include "gallivm/lp_bld_logic.h"
28 #include "gallivm/lp_bld_arit.h"
29 #include "gallivm/lp_bld_flow.h"
30 #include "gallivm/lp_bld_misc.h"
31 #include "util/u_memory.h"
32 #include "util/u_string.h"
33 #include "tgsi/tgsi_build.h"
34 #include "tgsi/tgsi_util.h"
35 #include "tgsi/tgsi_dump.h"
37 #include "ac_binary.h"
38 #include "ac_llvm_util.h"
39 #include "ac_exp_param.h"
40 #include "ac_shader_util.h"
41 #include "si_shader_internal.h"
45 #include "compiler/nir/nir.h"
47 static const char *scratch_rsrc_dword0_symbol
=
48 "SCRATCH_RSRC_DWORD0";
50 static const char *scratch_rsrc_dword1_symbol
=
51 "SCRATCH_RSRC_DWORD1";
53 struct si_shader_output_values
55 LLVMValueRef values
[4];
56 unsigned semantic_name
;
57 unsigned semantic_index
;
58 ubyte vertex_stream
[4];
62 * Used to collect types and other info about arguments of the LLVM function
63 * before the function is created.
65 struct si_function_info
{
66 LLVMTypeRef types
[100];
67 LLVMValueRef
*assign
[100];
68 unsigned num_sgpr_params
;
77 static void si_init_shader_ctx(struct si_shader_context
*ctx
,
78 struct si_screen
*sscreen
,
79 LLVMTargetMachineRef tm
);
81 static void si_llvm_emit_barrier(const struct lp_build_tgsi_action
*action
,
82 struct lp_build_tgsi_context
*bld_base
,
83 struct lp_build_emit_data
*emit_data
);
85 static void si_dump_shader_key(unsigned processor
, const struct si_shader
*shader
,
88 static void si_build_vs_prolog_function(struct si_shader_context
*ctx
,
89 union si_shader_part_key
*key
);
90 static void si_build_tcs_epilog_function(struct si_shader_context
*ctx
,
91 union si_shader_part_key
*key
);
92 static void si_build_ps_prolog_function(struct si_shader_context
*ctx
,
93 union si_shader_part_key
*key
);
94 static void si_build_ps_epilog_function(struct si_shader_context
*ctx
,
95 union si_shader_part_key
*key
);
97 /* Ideally pass the sample mask input to the PS epilog as v14, which
98 * is its usual location, so that the shader doesn't have to add v_mov.
100 #define PS_EPILOG_SAMPLEMASK_MIN_LOC 14
102 static bool llvm_type_is_64bit(struct si_shader_context
*ctx
,
105 if (type
== ctx
->ac
.i64
|| type
== ctx
->ac
.f64
)
111 static bool is_merged_shader(struct si_shader
*shader
)
113 if (shader
->selector
->screen
->info
.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_llvm_param(struct si_shader_context
*ctx
,
240 LLVMValueRef value
, unsigned rshift
,
243 if (LLVMGetTypeKind(LLVMTypeOf(value
)) == LLVMFloatTypeKind
)
244 value
= ac_to_integer(&ctx
->ac
, value
);
247 value
= LLVMBuildLShr(ctx
->ac
.builder
, value
,
248 LLVMConstInt(ctx
->i32
, rshift
, 0), "");
250 if (rshift
+ bitwidth
< 32) {
251 unsigned mask
= (1 << bitwidth
) - 1;
252 value
= LLVMBuildAnd(ctx
->ac
.builder
, value
,
253 LLVMConstInt(ctx
->i32
, mask
, 0), "");
259 static LLVMValueRef
unpack_param(struct si_shader_context
*ctx
,
260 unsigned param
, unsigned rshift
,
263 LLVMValueRef value
= LLVMGetParam(ctx
->main_fn
, param
);
265 return unpack_llvm_param(ctx
, value
, rshift
, bitwidth
);
268 static LLVMValueRef
get_rel_patch_id(struct si_shader_context
*ctx
)
271 case PIPE_SHADER_TESS_CTRL
:
272 return unpack_llvm_param(ctx
, ctx
->abi
.tcs_rel_ids
, 0, 8);
274 case PIPE_SHADER_TESS_EVAL
:
275 return LLVMGetParam(ctx
->main_fn
,
276 ctx
->param_tes_rel_patch_id
);
284 /* Tessellation shaders pass outputs to the next shader using LDS.
286 * LS outputs = TCS inputs
287 * TCS outputs = TES inputs
290 * - TCS inputs for patch 0
291 * - TCS inputs for patch 1
292 * - TCS inputs for patch 2 = get_tcs_in_current_patch_offset (if RelPatchID==2)
294 * - TCS outputs for patch 0 = get_tcs_out_patch0_offset
295 * - Per-patch TCS outputs for patch 0 = get_tcs_out_patch0_patch_data_offset
296 * - TCS outputs for patch 1
297 * - Per-patch TCS outputs for patch 1
298 * - TCS outputs for patch 2 = get_tcs_out_current_patch_offset (if RelPatchID==2)
299 * - Per-patch TCS outputs for patch 2 = get_tcs_out_current_patch_data_offset (if RelPatchID==2)
302 * All three shaders VS(LS), TCS, TES share the same LDS space.
306 get_tcs_in_patch_stride(struct si_shader_context
*ctx
)
308 return unpack_param(ctx
, ctx
->param_vs_state_bits
, 8, 13);
311 static unsigned get_tcs_out_vertex_dw_stride_constant(struct si_shader_context
*ctx
)
313 assert(ctx
->type
== PIPE_SHADER_TESS_CTRL
);
315 if (ctx
->shader
->key
.mono
.u
.ff_tcs_inputs_to_copy
)
316 return util_last_bit64(ctx
->shader
->key
.mono
.u
.ff_tcs_inputs_to_copy
) * 4;
318 return util_last_bit64(ctx
->shader
->selector
->outputs_written
) * 4;
321 static LLVMValueRef
get_tcs_out_vertex_dw_stride(struct si_shader_context
*ctx
)
323 unsigned stride
= get_tcs_out_vertex_dw_stride_constant(ctx
);
325 return LLVMConstInt(ctx
->i32
, stride
, 0);
328 static LLVMValueRef
get_tcs_out_patch_stride(struct si_shader_context
*ctx
)
330 if (ctx
->shader
->key
.mono
.u
.ff_tcs_inputs_to_copy
)
331 return unpack_param(ctx
, ctx
->param_tcs_out_lds_layout
, 0, 13);
333 const struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
334 unsigned tcs_out_vertices
= info
->properties
[TGSI_PROPERTY_TCS_VERTICES_OUT
];
335 unsigned vertex_dw_stride
= get_tcs_out_vertex_dw_stride_constant(ctx
);
336 unsigned num_patch_outputs
= util_last_bit64(ctx
->shader
->selector
->patch_outputs_written
);
337 unsigned patch_dw_stride
= tcs_out_vertices
* vertex_dw_stride
+
338 num_patch_outputs
* 4;
339 return LLVMConstInt(ctx
->i32
, patch_dw_stride
, 0);
343 get_tcs_out_patch0_offset(struct si_shader_context
*ctx
)
345 return lp_build_mul_imm(&ctx
->bld_base
.uint_bld
,
347 ctx
->param_tcs_out_lds_offsets
,
353 get_tcs_out_patch0_patch_data_offset(struct si_shader_context
*ctx
)
355 return lp_build_mul_imm(&ctx
->bld_base
.uint_bld
,
357 ctx
->param_tcs_out_lds_offsets
,
363 get_tcs_in_current_patch_offset(struct si_shader_context
*ctx
)
365 LLVMValueRef patch_stride
= get_tcs_in_patch_stride(ctx
);
366 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
368 return LLVMBuildMul(ctx
->ac
.builder
, patch_stride
, rel_patch_id
, "");
372 get_tcs_out_current_patch_offset(struct si_shader_context
*ctx
)
374 LLVMValueRef patch0_offset
= get_tcs_out_patch0_offset(ctx
);
375 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(ctx
);
376 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
378 return LLVMBuildAdd(ctx
->ac
.builder
, patch0_offset
,
379 LLVMBuildMul(ctx
->ac
.builder
, patch_stride
,
385 get_tcs_out_current_patch_data_offset(struct si_shader_context
*ctx
)
387 LLVMValueRef patch0_patch_data_offset
=
388 get_tcs_out_patch0_patch_data_offset(ctx
);
389 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(ctx
);
390 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
392 return LLVMBuildAdd(ctx
->ac
.builder
, patch0_patch_data_offset
,
393 LLVMBuildMul(ctx
->ac
.builder
, patch_stride
,
398 static LLVMValueRef
get_num_tcs_out_vertices(struct si_shader_context
*ctx
)
400 unsigned tcs_out_vertices
=
401 ctx
->shader
->selector
?
402 ctx
->shader
->selector
->info
.properties
[TGSI_PROPERTY_TCS_VERTICES_OUT
] : 0;
404 /* If !tcs_out_vertices, it's either the fixed-func TCS or the TCS epilog. */
405 if (ctx
->type
== PIPE_SHADER_TESS_CTRL
&& tcs_out_vertices
)
406 return LLVMConstInt(ctx
->i32
, tcs_out_vertices
, 0);
408 return unpack_param(ctx
, ctx
->param_tcs_offchip_layout
, 6, 6);
411 static LLVMValueRef
get_tcs_in_vertex_dw_stride(struct si_shader_context
*ctx
)
416 case PIPE_SHADER_VERTEX
:
417 stride
= util_last_bit64(ctx
->shader
->selector
->outputs_written
);
418 return LLVMConstInt(ctx
->i32
, stride
* 4, 0);
420 case PIPE_SHADER_TESS_CTRL
:
421 if (ctx
->screen
->info
.chip_class
>= GFX9
&&
422 ctx
->shader
->is_monolithic
) {
423 stride
= util_last_bit64(ctx
->shader
->key
.part
.tcs
.ls
->outputs_written
);
424 return LLVMConstInt(ctx
->i32
, stride
* 4, 0);
426 return unpack_param(ctx
, ctx
->param_vs_state_bits
, 24, 8);
434 static LLVMValueRef
get_instance_index_for_fetch(
435 struct si_shader_context
*ctx
,
436 unsigned param_start_instance
, LLVMValueRef divisor
)
438 LLVMValueRef result
= ctx
->abi
.instance_id
;
440 /* The division must be done before START_INSTANCE is added. */
441 if (divisor
!= ctx
->i32_1
)
442 result
= LLVMBuildUDiv(ctx
->ac
.builder
, result
, divisor
, "");
444 return LLVMBuildAdd(ctx
->ac
.builder
, result
,
445 LLVMGetParam(ctx
->main_fn
, param_start_instance
), "");
448 /* Bitcast <4 x float> to <2 x double>, extract the component, and convert
450 static LLVMValueRef
extract_double_to_float(struct si_shader_context
*ctx
,
452 unsigned double_index
)
454 LLVMBuilderRef builder
= ctx
->ac
.builder
;
455 LLVMTypeRef f64
= LLVMDoubleTypeInContext(ctx
->ac
.context
);
456 LLVMValueRef dvec2
= LLVMBuildBitCast(builder
, vec4
,
457 LLVMVectorType(f64
, 2), "");
458 LLVMValueRef index
= LLVMConstInt(ctx
->i32
, double_index
, 0);
459 LLVMValueRef value
= LLVMBuildExtractElement(builder
, dvec2
, index
, "");
460 return LLVMBuildFPTrunc(builder
, value
, ctx
->f32
, "");
463 static LLVMValueRef
unpack_sint16(struct si_shader_context
*ctx
,
464 LLVMValueRef i32
, unsigned index
)
469 return LLVMBuildAShr(ctx
->ac
.builder
, i32
,
470 LLVMConstInt(ctx
->i32
, 16, 0), "");
472 return LLVMBuildSExt(ctx
->ac
.builder
,
473 LLVMBuildTrunc(ctx
->ac
.builder
, i32
,
478 void si_llvm_load_input_vs(
479 struct si_shader_context
*ctx
,
480 unsigned input_index
,
483 const struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
484 unsigned vs_blit_property
= info
->properties
[TGSI_PROPERTY_VS_BLIT_SGPRS
];
486 if (vs_blit_property
) {
487 LLVMValueRef vertex_id
= ctx
->abi
.vertex_id
;
488 LLVMValueRef sel_x1
= LLVMBuildICmp(ctx
->ac
.builder
,
489 LLVMIntULE
, vertex_id
,
491 /* Use LLVMIntNE, because we have 3 vertices and only
492 * the middle one should use y2.
494 LLVMValueRef sel_y1
= LLVMBuildICmp(ctx
->ac
.builder
,
495 LLVMIntNE
, vertex_id
,
498 if (input_index
== 0) {
500 LLVMValueRef x1y1
= LLVMGetParam(ctx
->main_fn
,
501 ctx
->param_vs_blit_inputs
);
502 LLVMValueRef x2y2
= LLVMGetParam(ctx
->main_fn
,
503 ctx
->param_vs_blit_inputs
+ 1);
505 LLVMValueRef x1
= unpack_sint16(ctx
, x1y1
, 0);
506 LLVMValueRef y1
= unpack_sint16(ctx
, x1y1
, 1);
507 LLVMValueRef x2
= unpack_sint16(ctx
, x2y2
, 0);
508 LLVMValueRef y2
= unpack_sint16(ctx
, x2y2
, 1);
510 LLVMValueRef x
= LLVMBuildSelect(ctx
->ac
.builder
, sel_x1
,
512 LLVMValueRef y
= LLVMBuildSelect(ctx
->ac
.builder
, sel_y1
,
515 out
[0] = LLVMBuildSIToFP(ctx
->ac
.builder
, x
, ctx
->f32
, "");
516 out
[1] = LLVMBuildSIToFP(ctx
->ac
.builder
, y
, ctx
->f32
, "");
517 out
[2] = LLVMGetParam(ctx
->main_fn
,
518 ctx
->param_vs_blit_inputs
+ 2);
519 out
[3] = ctx
->ac
.f32_1
;
523 /* Color or texture coordinates: */
524 assert(input_index
== 1);
526 if (vs_blit_property
== SI_VS_BLIT_SGPRS_POS_COLOR
) {
527 for (int i
= 0; i
< 4; i
++) {
528 out
[i
] = LLVMGetParam(ctx
->main_fn
,
529 ctx
->param_vs_blit_inputs
+ 3 + i
);
532 assert(vs_blit_property
== SI_VS_BLIT_SGPRS_POS_TEXCOORD
);
533 LLVMValueRef x1
= LLVMGetParam(ctx
->main_fn
,
534 ctx
->param_vs_blit_inputs
+ 3);
535 LLVMValueRef y1
= LLVMGetParam(ctx
->main_fn
,
536 ctx
->param_vs_blit_inputs
+ 4);
537 LLVMValueRef x2
= LLVMGetParam(ctx
->main_fn
,
538 ctx
->param_vs_blit_inputs
+ 5);
539 LLVMValueRef y2
= LLVMGetParam(ctx
->main_fn
,
540 ctx
->param_vs_blit_inputs
+ 6);
542 out
[0] = LLVMBuildSelect(ctx
->ac
.builder
, sel_x1
,
544 out
[1] = LLVMBuildSelect(ctx
->ac
.builder
, sel_y1
,
546 out
[2] = LLVMGetParam(ctx
->main_fn
,
547 ctx
->param_vs_blit_inputs
+ 7);
548 out
[3] = LLVMGetParam(ctx
->main_fn
,
549 ctx
->param_vs_blit_inputs
+ 8);
556 unsigned num_fetches
;
557 unsigned fetch_stride
;
558 unsigned num_channels
;
560 LLVMValueRef t_list_ptr
;
561 LLVMValueRef t_offset
;
563 LLVMValueRef vertex_index
;
564 LLVMValueRef input
[3];
566 /* Load the T list */
567 t_list_ptr
= LLVMGetParam(ctx
->main_fn
, ctx
->param_vertex_buffers
);
569 t_offset
= LLVMConstInt(ctx
->i32
, input_index
, 0);
571 t_list
= ac_build_load_to_sgpr(&ctx
->ac
, t_list_ptr
, t_offset
);
573 vertex_index
= LLVMGetParam(ctx
->main_fn
,
574 ctx
->param_vertex_index0
+
577 fix_fetch
= ctx
->shader
->key
.mono
.vs_fix_fetch
[input_index
];
579 /* Do multiple loads for special formats. */
581 case SI_FIX_FETCH_RGB_64_FLOAT
:
582 num_fetches
= 3; /* 3 2-dword loads */
586 case SI_FIX_FETCH_RGBA_64_FLOAT
:
587 num_fetches
= 2; /* 2 4-dword loads */
591 case SI_FIX_FETCH_RGB_8
:
592 case SI_FIX_FETCH_RGB_8_INT
:
597 case SI_FIX_FETCH_RGB_16
:
598 case SI_FIX_FETCH_RGB_16_INT
:
606 num_channels
= util_last_bit(info
->input_usage_mask
[input_index
]);
609 for (unsigned i
= 0; i
< num_fetches
; i
++) {
610 LLVMValueRef voffset
= LLVMConstInt(ctx
->i32
, fetch_stride
* i
, 0);
612 input
[i
] = ac_build_buffer_load_format(&ctx
->ac
, t_list
,
613 vertex_index
, voffset
,
614 num_channels
, false, true);
615 input
[i
] = ac_build_expand_to_vec4(&ctx
->ac
, input
[i
], num_channels
);
618 /* Break up the vec4 into individual components */
619 for (chan
= 0; chan
< 4; chan
++) {
620 LLVMValueRef llvm_chan
= LLVMConstInt(ctx
->i32
, chan
, 0);
621 out
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
,
622 input
[0], llvm_chan
, "");
626 case SI_FIX_FETCH_A2_SNORM
:
627 case SI_FIX_FETCH_A2_SSCALED
:
628 case SI_FIX_FETCH_A2_SINT
: {
629 /* The hardware returns an unsigned value; convert it to a
632 LLVMValueRef tmp
= out
[3];
633 LLVMValueRef c30
= LLVMConstInt(ctx
->i32
, 30, 0);
635 /* First, recover the sign-extended signed integer value. */
636 if (fix_fetch
== SI_FIX_FETCH_A2_SSCALED
)
637 tmp
= LLVMBuildFPToUI(ctx
->ac
.builder
, tmp
, ctx
->i32
, "");
639 tmp
= ac_to_integer(&ctx
->ac
, tmp
);
641 /* For the integer-like cases, do a natural sign extension.
643 * For the SNORM case, the values are 0.0, 0.333, 0.666, 1.0
644 * and happen to contain 0, 1, 2, 3 as the two LSBs of the
647 tmp
= LLVMBuildShl(ctx
->ac
.builder
, tmp
,
648 fix_fetch
== SI_FIX_FETCH_A2_SNORM
?
649 LLVMConstInt(ctx
->i32
, 7, 0) : c30
, "");
650 tmp
= LLVMBuildAShr(ctx
->ac
.builder
, tmp
, c30
, "");
652 /* Convert back to the right type. */
653 if (fix_fetch
== SI_FIX_FETCH_A2_SNORM
) {
655 LLVMValueRef neg_one
= LLVMConstReal(ctx
->f32
, -1.0);
656 tmp
= LLVMBuildSIToFP(ctx
->ac
.builder
, tmp
, ctx
->f32
, "");
657 clamp
= LLVMBuildFCmp(ctx
->ac
.builder
, LLVMRealULT
, tmp
, neg_one
, "");
658 tmp
= LLVMBuildSelect(ctx
->ac
.builder
, clamp
, neg_one
, tmp
, "");
659 } else if (fix_fetch
== SI_FIX_FETCH_A2_SSCALED
) {
660 tmp
= LLVMBuildSIToFP(ctx
->ac
.builder
, tmp
, ctx
->f32
, "");
666 case SI_FIX_FETCH_RGBA_32_UNORM
:
667 case SI_FIX_FETCH_RGBX_32_UNORM
:
668 for (chan
= 0; chan
< 4; chan
++) {
669 out
[chan
] = ac_to_integer(&ctx
->ac
, out
[chan
]);
670 out
[chan
] = LLVMBuildUIToFP(ctx
->ac
.builder
,
671 out
[chan
], ctx
->f32
, "");
672 out
[chan
] = LLVMBuildFMul(ctx
->ac
.builder
, out
[chan
],
673 LLVMConstReal(ctx
->f32
, 1.0 / UINT_MAX
), "");
675 /* RGBX UINT returns 1 in alpha, which would be rounded to 0 by normalizing. */
676 if (fix_fetch
== SI_FIX_FETCH_RGBX_32_UNORM
)
677 out
[3] = LLVMConstReal(ctx
->f32
, 1);
679 case SI_FIX_FETCH_RGBA_32_SNORM
:
680 case SI_FIX_FETCH_RGBX_32_SNORM
:
681 case SI_FIX_FETCH_RGBA_32_FIXED
:
682 case SI_FIX_FETCH_RGBX_32_FIXED
: {
684 if (fix_fetch
>= SI_FIX_FETCH_RGBA_32_FIXED
)
685 scale
= 1.0 / 0x10000;
687 scale
= 1.0 / INT_MAX
;
689 for (chan
= 0; chan
< 4; chan
++) {
690 out
[chan
] = ac_to_integer(&ctx
->ac
, out
[chan
]);
691 out
[chan
] = LLVMBuildSIToFP(ctx
->ac
.builder
,
692 out
[chan
], ctx
->f32
, "");
693 out
[chan
] = LLVMBuildFMul(ctx
->ac
.builder
, out
[chan
],
694 LLVMConstReal(ctx
->f32
, scale
), "");
696 /* RGBX SINT returns 1 in alpha, which would be rounded to 0 by normalizing. */
697 if (fix_fetch
== SI_FIX_FETCH_RGBX_32_SNORM
||
698 fix_fetch
== SI_FIX_FETCH_RGBX_32_FIXED
)
699 out
[3] = LLVMConstReal(ctx
->f32
, 1);
702 case SI_FIX_FETCH_RGBA_32_USCALED
:
703 for (chan
= 0; chan
< 4; chan
++) {
704 out
[chan
] = ac_to_integer(&ctx
->ac
, out
[chan
]);
705 out
[chan
] = LLVMBuildUIToFP(ctx
->ac
.builder
,
706 out
[chan
], ctx
->f32
, "");
709 case SI_FIX_FETCH_RGBA_32_SSCALED
:
710 for (chan
= 0; chan
< 4; chan
++) {
711 out
[chan
] = ac_to_integer(&ctx
->ac
, out
[chan
]);
712 out
[chan
] = LLVMBuildSIToFP(ctx
->ac
.builder
,
713 out
[chan
], ctx
->f32
, "");
716 case SI_FIX_FETCH_RG_64_FLOAT
:
717 for (chan
= 0; chan
< 2; chan
++)
718 out
[chan
] = extract_double_to_float(ctx
, input
[0], chan
);
720 out
[2] = LLVMConstReal(ctx
->f32
, 0);
721 out
[3] = LLVMConstReal(ctx
->f32
, 1);
723 case SI_FIX_FETCH_RGB_64_FLOAT
:
724 for (chan
= 0; chan
< 3; chan
++)
725 out
[chan
] = extract_double_to_float(ctx
, input
[chan
], 0);
727 out
[3] = LLVMConstReal(ctx
->f32
, 1);
729 case SI_FIX_FETCH_RGBA_64_FLOAT
:
730 for (chan
= 0; chan
< 4; chan
++) {
731 out
[chan
] = extract_double_to_float(ctx
, input
[chan
/ 2],
735 case SI_FIX_FETCH_RGB_8
:
736 case SI_FIX_FETCH_RGB_8_INT
:
737 case SI_FIX_FETCH_RGB_16
:
738 case SI_FIX_FETCH_RGB_16_INT
:
739 for (chan
= 0; chan
< 3; chan
++) {
740 out
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
,
744 if (fix_fetch
== SI_FIX_FETCH_RGB_8
||
745 fix_fetch
== SI_FIX_FETCH_RGB_16
) {
746 out
[3] = LLVMConstReal(ctx
->f32
, 1);
748 out
[3] = ac_to_float(&ctx
->ac
, ctx
->i32_1
);
754 static void declare_input_vs(
755 struct si_shader_context
*ctx
,
756 unsigned input_index
,
757 const struct tgsi_full_declaration
*decl
,
760 si_llvm_load_input_vs(ctx
, input_index
, out
);
763 static LLVMValueRef
get_primitive_id(struct si_shader_context
*ctx
,
770 case PIPE_SHADER_VERTEX
:
771 return LLVMGetParam(ctx
->main_fn
,
772 ctx
->param_vs_prim_id
);
773 case PIPE_SHADER_TESS_CTRL
:
774 return ctx
->abi
.tcs_patch_id
;
775 case PIPE_SHADER_TESS_EVAL
:
776 return ctx
->abi
.tes_patch_id
;
777 case PIPE_SHADER_GEOMETRY
:
778 return ctx
->abi
.gs_prim_id
;
786 * Return the value of tgsi_ind_register for indexing.
787 * This is the indirect index with the constant offset added to it.
789 LLVMValueRef
si_get_indirect_index(struct si_shader_context
*ctx
,
790 const struct tgsi_ind_register
*ind
,
796 if (ind
->File
== TGSI_FILE_ADDRESS
) {
797 result
= ctx
->addrs
[ind
->Index
][ind
->Swizzle
];
798 result
= LLVMBuildLoad(ctx
->ac
.builder
, result
, "");
800 struct tgsi_full_src_register src
= {};
802 src
.Register
.File
= ind
->File
;
803 src
.Register
.Index
= ind
->Index
;
805 /* Set the second index to 0 for constants. */
806 if (ind
->File
== TGSI_FILE_CONSTANT
)
807 src
.Register
.Dimension
= 1;
809 result
= ctx
->bld_base
.emit_fetch_funcs
[ind
->File
](&ctx
->bld_base
, &src
,
812 result
= ac_to_integer(&ctx
->ac
, result
);
816 result
= LLVMBuildMul(ctx
->ac
.builder
, result
,
817 LLVMConstInt(ctx
->i32
, addr_mul
, 0), "");
818 result
= LLVMBuildAdd(ctx
->ac
.builder
, result
,
819 LLVMConstInt(ctx
->i32
, rel_index
, 0), "");
824 * Like si_get_indirect_index, but restricts the return value to a (possibly
825 * undefined) value inside [0..num).
827 LLVMValueRef
si_get_bounded_indirect_index(struct si_shader_context
*ctx
,
828 const struct tgsi_ind_register
*ind
,
829 int rel_index
, unsigned num
)
831 LLVMValueRef result
= si_get_indirect_index(ctx
, ind
, 1, rel_index
);
833 return si_llvm_bound_index(ctx
, result
, num
);
836 static LLVMValueRef
get_dw_address_from_generic_indices(struct si_shader_context
*ctx
,
837 LLVMValueRef vertex_dw_stride
,
838 LLVMValueRef base_addr
,
839 LLVMValueRef vertex_index
,
840 LLVMValueRef param_index
,
841 unsigned input_index
,
846 if (vertex_dw_stride
) {
847 base_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_addr
,
848 LLVMBuildMul(ctx
->ac
.builder
, vertex_index
,
849 vertex_dw_stride
, ""), "");
853 base_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_addr
,
854 LLVMBuildMul(ctx
->ac
.builder
, param_index
,
855 LLVMConstInt(ctx
->i32
, 4, 0), ""), "");
858 int param
= is_patch
?
859 si_shader_io_get_unique_index_patch(name
[input_index
],
860 index
[input_index
]) :
861 si_shader_io_get_unique_index(name
[input_index
],
864 /* Add the base address of the element. */
865 return LLVMBuildAdd(ctx
->ac
.builder
, base_addr
,
866 LLVMConstInt(ctx
->i32
, param
* 4, 0), "");
870 * Calculate a dword address given an input or output register and a stride.
872 static LLVMValueRef
get_dw_address(struct si_shader_context
*ctx
,
873 const struct tgsi_full_dst_register
*dst
,
874 const struct tgsi_full_src_register
*src
,
875 LLVMValueRef vertex_dw_stride
,
876 LLVMValueRef base_addr
)
878 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
879 ubyte
*name
, *index
, *array_first
;
881 struct tgsi_full_dst_register reg
;
882 LLVMValueRef vertex_index
= NULL
;
883 LLVMValueRef ind_index
= NULL
;
885 /* Set the register description. The address computation is the same
886 * for sources and destinations. */
888 reg
.Register
.File
= src
->Register
.File
;
889 reg
.Register
.Index
= src
->Register
.Index
;
890 reg
.Register
.Indirect
= src
->Register
.Indirect
;
891 reg
.Register
.Dimension
= src
->Register
.Dimension
;
892 reg
.Indirect
= src
->Indirect
;
893 reg
.Dimension
= src
->Dimension
;
894 reg
.DimIndirect
= src
->DimIndirect
;
898 /* If the register is 2-dimensional (e.g. an array of vertices
899 * in a primitive), calculate the base address of the vertex. */
900 if (reg
.Register
.Dimension
) {
901 if (reg
.Dimension
.Indirect
)
902 vertex_index
= si_get_indirect_index(ctx
, ®
.DimIndirect
,
903 1, reg
.Dimension
.Index
);
905 vertex_index
= LLVMConstInt(ctx
->i32
, reg
.Dimension
.Index
, 0);
908 /* Get information about the register. */
909 if (reg
.Register
.File
== TGSI_FILE_INPUT
) {
910 name
= info
->input_semantic_name
;
911 index
= info
->input_semantic_index
;
912 array_first
= info
->input_array_first
;
913 } else if (reg
.Register
.File
== TGSI_FILE_OUTPUT
) {
914 name
= info
->output_semantic_name
;
915 index
= info
->output_semantic_index
;
916 array_first
= info
->output_array_first
;
922 if (reg
.Register
.Indirect
) {
923 /* Add the relative address of the element. */
924 if (reg
.Indirect
.ArrayID
)
925 input_index
= array_first
[reg
.Indirect
.ArrayID
];
927 input_index
= reg
.Register
.Index
;
929 ind_index
= si_get_indirect_index(ctx
, ®
.Indirect
,
930 1, reg
.Register
.Index
- input_index
);
932 input_index
= reg
.Register
.Index
;
935 return get_dw_address_from_generic_indices(ctx
, vertex_dw_stride
,
936 base_addr
, vertex_index
,
937 ind_index
, input_index
,
939 !reg
.Register
.Dimension
);
942 /* The offchip buffer layout for TCS->TES is
944 * - attribute 0 of patch 0 vertex 0
945 * - attribute 0 of patch 0 vertex 1
946 * - attribute 0 of patch 0 vertex 2
948 * - attribute 0 of patch 1 vertex 0
949 * - attribute 0 of patch 1 vertex 1
951 * - attribute 1 of patch 0 vertex 0
952 * - attribute 1 of patch 0 vertex 1
954 * - per patch attribute 0 of patch 0
955 * - per patch attribute 0 of patch 1
958 * Note that every attribute has 4 components.
960 static LLVMValueRef
get_tcs_tes_buffer_address(struct si_shader_context
*ctx
,
961 LLVMValueRef rel_patch_id
,
962 LLVMValueRef vertex_index
,
963 LLVMValueRef param_index
)
965 LLVMValueRef base_addr
, vertices_per_patch
, num_patches
, total_vertices
;
966 LLVMValueRef param_stride
, constant16
;
968 vertices_per_patch
= get_num_tcs_out_vertices(ctx
);
969 num_patches
= unpack_param(ctx
, ctx
->param_tcs_offchip_layout
, 0, 6);
970 total_vertices
= LLVMBuildMul(ctx
->ac
.builder
, vertices_per_patch
,
973 constant16
= LLVMConstInt(ctx
->i32
, 16, 0);
975 base_addr
= LLVMBuildMul(ctx
->ac
.builder
, rel_patch_id
,
976 vertices_per_patch
, "");
978 base_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_addr
,
981 param_stride
= total_vertices
;
983 base_addr
= rel_patch_id
;
984 param_stride
= num_patches
;
987 base_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_addr
,
988 LLVMBuildMul(ctx
->ac
.builder
, param_index
,
989 param_stride
, ""), "");
991 base_addr
= LLVMBuildMul(ctx
->ac
.builder
, base_addr
, constant16
, "");
994 LLVMValueRef patch_data_offset
=
995 unpack_param(ctx
, ctx
->param_tcs_offchip_layout
, 12, 20);
997 base_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_addr
,
998 patch_data_offset
, "");
1003 /* This is a generic helper that can be shared by the NIR and TGSI backends */
1004 static LLVMValueRef
get_tcs_tes_buffer_address_from_generic_indices(
1005 struct si_shader_context
*ctx
,
1006 LLVMValueRef vertex_index
,
1007 LLVMValueRef param_index
,
1008 unsigned param_base
,
1013 unsigned param_index_base
;
1015 param_index_base
= is_patch
?
1016 si_shader_io_get_unique_index_patch(name
[param_base
], index
[param_base
]) :
1017 si_shader_io_get_unique_index(name
[param_base
], index
[param_base
]);
1020 param_index
= LLVMBuildAdd(ctx
->ac
.builder
, param_index
,
1021 LLVMConstInt(ctx
->i32
, param_index_base
, 0),
1024 param_index
= LLVMConstInt(ctx
->i32
, param_index_base
, 0);
1027 return get_tcs_tes_buffer_address(ctx
, get_rel_patch_id(ctx
),
1028 vertex_index
, param_index
);
1031 static LLVMValueRef
get_tcs_tes_buffer_address_from_reg(
1032 struct si_shader_context
*ctx
,
1033 const struct tgsi_full_dst_register
*dst
,
1034 const struct tgsi_full_src_register
*src
)
1036 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
1037 ubyte
*name
, *index
, *array_first
;
1038 struct tgsi_full_src_register reg
;
1039 LLVMValueRef vertex_index
= NULL
;
1040 LLVMValueRef param_index
= NULL
;
1041 unsigned param_base
;
1043 reg
= src
? *src
: tgsi_full_src_register_from_dst(dst
);
1045 if (reg
.Register
.Dimension
) {
1047 if (reg
.Dimension
.Indirect
)
1048 vertex_index
= si_get_indirect_index(ctx
, ®
.DimIndirect
,
1049 1, reg
.Dimension
.Index
);
1051 vertex_index
= LLVMConstInt(ctx
->i32
, reg
.Dimension
.Index
, 0);
1054 /* Get information about the register. */
1055 if (reg
.Register
.File
== TGSI_FILE_INPUT
) {
1056 name
= info
->input_semantic_name
;
1057 index
= info
->input_semantic_index
;
1058 array_first
= info
->input_array_first
;
1059 } else if (reg
.Register
.File
== TGSI_FILE_OUTPUT
) {
1060 name
= info
->output_semantic_name
;
1061 index
= info
->output_semantic_index
;
1062 array_first
= info
->output_array_first
;
1068 if (reg
.Register
.Indirect
) {
1069 if (reg
.Indirect
.ArrayID
)
1070 param_base
= array_first
[reg
.Indirect
.ArrayID
];
1072 param_base
= reg
.Register
.Index
;
1074 param_index
= si_get_indirect_index(ctx
, ®
.Indirect
,
1075 1, reg
.Register
.Index
- param_base
);
1078 param_base
= reg
.Register
.Index
;
1081 return get_tcs_tes_buffer_address_from_generic_indices(ctx
, vertex_index
,
1082 param_index
, param_base
,
1083 name
, index
, !reg
.Register
.Dimension
);
1086 static LLVMValueRef
buffer_load(struct lp_build_tgsi_context
*bld_base
,
1087 LLVMTypeRef type
, unsigned swizzle
,
1088 LLVMValueRef buffer
, LLVMValueRef offset
,
1089 LLVMValueRef base
, bool can_speculate
)
1091 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1092 LLVMValueRef value
, value2
;
1093 LLVMTypeRef vec_type
= LLVMVectorType(type
, 4);
1095 if (swizzle
== ~0) {
1096 value
= ac_build_buffer_load(&ctx
->ac
, buffer
, 4, NULL
, base
, offset
,
1097 0, 1, 0, can_speculate
, false);
1099 return LLVMBuildBitCast(ctx
->ac
.builder
, value
, vec_type
, "");
1102 if (!llvm_type_is_64bit(ctx
, type
)) {
1103 value
= ac_build_buffer_load(&ctx
->ac
, buffer
, 4, NULL
, base
, offset
,
1104 0, 1, 0, can_speculate
, false);
1106 value
= LLVMBuildBitCast(ctx
->ac
.builder
, value
, vec_type
, "");
1107 return LLVMBuildExtractElement(ctx
->ac
.builder
, value
,
1108 LLVMConstInt(ctx
->i32
, swizzle
, 0), "");
1111 value
= ac_build_buffer_load(&ctx
->ac
, buffer
, 1, NULL
, base
, offset
,
1112 swizzle
* 4, 1, 0, can_speculate
, false);
1114 value2
= ac_build_buffer_load(&ctx
->ac
, buffer
, 1, NULL
, base
, offset
,
1115 swizzle
* 4 + 4, 1, 0, can_speculate
, false);
1117 return si_llvm_emit_fetch_64bit(bld_base
, type
, value
, value2
);
1123 * \param type output value type
1124 * \param swizzle offset (typically 0..3); it can be ~0, which loads a vec4
1125 * \param dw_addr address in dwords
1127 static LLVMValueRef
lds_load(struct lp_build_tgsi_context
*bld_base
,
1128 LLVMTypeRef type
, unsigned swizzle
,
1129 LLVMValueRef dw_addr
)
1131 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1134 if (swizzle
== ~0) {
1135 LLVMValueRef values
[TGSI_NUM_CHANNELS
];
1137 for (unsigned chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++)
1138 values
[chan
] = lds_load(bld_base
, type
, chan
, dw_addr
);
1140 return lp_build_gather_values(&ctx
->gallivm
, values
,
1144 /* Split 64-bit loads. */
1145 if (llvm_type_is_64bit(ctx
, type
)) {
1146 LLVMValueRef lo
, hi
;
1148 lo
= lds_load(bld_base
, ctx
->i32
, swizzle
, dw_addr
);
1149 hi
= lds_load(bld_base
, ctx
->i32
, swizzle
+ 1, dw_addr
);
1150 return si_llvm_emit_fetch_64bit(bld_base
, type
, lo
, hi
);
1153 dw_addr
= lp_build_add(&bld_base
->uint_bld
, dw_addr
,
1154 LLVMConstInt(ctx
->i32
, swizzle
, 0));
1156 value
= ac_lds_load(&ctx
->ac
, dw_addr
);
1158 return LLVMBuildBitCast(ctx
->ac
.builder
, value
, type
, "");
1164 * \param swizzle offset (typically 0..3)
1165 * \param dw_addr address in dwords
1166 * \param value value to store
1168 static void lds_store(struct si_shader_context
*ctx
,
1169 unsigned dw_offset_imm
, LLVMValueRef dw_addr
,
1172 dw_addr
= lp_build_add(&ctx
->bld_base
.uint_bld
, dw_addr
,
1173 LLVMConstInt(ctx
->i32
, dw_offset_imm
, 0));
1175 ac_lds_store(&ctx
->ac
, dw_addr
, value
);
1180 TESS_OFFCHIP_RING_TCS
,
1181 TESS_OFFCHIP_RING_TES
,
1184 static LLVMValueRef
get_tess_ring_descriptor(struct si_shader_context
*ctx
,
1185 enum si_tess_ring ring
)
1187 LLVMBuilderRef builder
= ctx
->ac
.builder
;
1188 unsigned param
= ring
== TESS_OFFCHIP_RING_TES
? ctx
->param_tes_offchip_addr
:
1189 ctx
->param_tcs_out_lds_layout
;
1190 LLVMValueRef addr
= LLVMGetParam(ctx
->main_fn
, param
);
1192 /* TCS only receives high 13 bits of the address. */
1193 if (ring
== TESS_OFFCHIP_RING_TCS
|| ring
== TCS_FACTOR_RING
) {
1194 addr
= LLVMBuildAnd(builder
, addr
,
1195 LLVMConstInt(ctx
->i32
, 0xfff80000, 0), "");
1198 if (ring
== TCS_FACTOR_RING
) {
1199 unsigned tf_offset
= ctx
->screen
->tess_offchip_ring_size
;
1200 addr
= LLVMBuildAdd(builder
, addr
,
1201 LLVMConstInt(ctx
->i32
, tf_offset
, 0), "");
1204 LLVMValueRef desc
[4];
1206 desc
[1] = LLVMConstInt(ctx
->i32
,
1207 S_008F04_BASE_ADDRESS_HI(ctx
->screen
->info
.address32_hi
), 0);
1208 desc
[2] = LLVMConstInt(ctx
->i32
, 0xffffffff, 0);
1209 desc
[3] = LLVMConstInt(ctx
->i32
,
1210 S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1211 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1212 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1213 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1214 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1215 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
), 0);
1217 return ac_build_gather_values(&ctx
->ac
, desc
, 4);
1220 static LLVMValueRef
fetch_input_tcs(
1221 struct lp_build_tgsi_context
*bld_base
,
1222 const struct tgsi_full_src_register
*reg
,
1223 enum tgsi_opcode_type type
, unsigned swizzle
)
1225 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1226 LLVMValueRef dw_addr
, stride
;
1228 stride
= get_tcs_in_vertex_dw_stride(ctx
);
1229 dw_addr
= get_tcs_in_current_patch_offset(ctx
);
1230 dw_addr
= get_dw_address(ctx
, NULL
, reg
, stride
, dw_addr
);
1232 return lds_load(bld_base
, tgsi2llvmtype(bld_base
, type
), swizzle
, dw_addr
);
1235 static LLVMValueRef
si_nir_load_tcs_varyings(struct ac_shader_abi
*abi
,
1237 LLVMValueRef vertex_index
,
1238 LLVMValueRef param_index
,
1239 unsigned const_index
,
1241 unsigned driver_location
,
1243 unsigned num_components
,
1248 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1249 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
1250 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
1251 LLVMValueRef dw_addr
, stride
;
1253 driver_location
= driver_location
/ 4;
1256 stride
= get_tcs_in_vertex_dw_stride(ctx
);
1257 dw_addr
= get_tcs_in_current_patch_offset(ctx
);
1261 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
1263 stride
= get_tcs_out_vertex_dw_stride(ctx
);
1264 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
1269 /* Add the constant index to the indirect index */
1270 param_index
= LLVMBuildAdd(ctx
->ac
.builder
, param_index
,
1271 LLVMConstInt(ctx
->i32
, const_index
, 0), "");
1273 param_index
= LLVMConstInt(ctx
->i32
, const_index
, 0);
1279 names
= info
->input_semantic_name
;
1280 indices
= info
->input_semantic_index
;
1282 names
= info
->output_semantic_name
;
1283 indices
= info
->output_semantic_index
;
1286 dw_addr
= get_dw_address_from_generic_indices(ctx
, stride
, dw_addr
,
1287 vertex_index
, param_index
,
1292 LLVMValueRef value
[4];
1293 for (unsigned i
= 0; i
< num_components
+ component
; i
++) {
1294 unsigned offset
= i
;
1295 if (llvm_type_is_64bit(ctx
, type
))
1298 value
[i
] = lds_load(bld_base
, type
, offset
, dw_addr
);
1301 return ac_build_varying_gather_values(&ctx
->ac
, value
, num_components
, component
);
1304 static LLVMValueRef
fetch_output_tcs(
1305 struct lp_build_tgsi_context
*bld_base
,
1306 const struct tgsi_full_src_register
*reg
,
1307 enum tgsi_opcode_type type
, unsigned swizzle
)
1309 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1310 LLVMValueRef dw_addr
, stride
;
1312 if (reg
->Register
.Dimension
) {
1313 stride
= get_tcs_out_vertex_dw_stride(ctx
);
1314 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
1315 dw_addr
= get_dw_address(ctx
, NULL
, reg
, stride
, dw_addr
);
1317 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
1318 dw_addr
= get_dw_address(ctx
, NULL
, reg
, NULL
, dw_addr
);
1321 return lds_load(bld_base
, tgsi2llvmtype(bld_base
, type
), swizzle
, dw_addr
);
1324 static LLVMValueRef
fetch_input_tes(
1325 struct lp_build_tgsi_context
*bld_base
,
1326 const struct tgsi_full_src_register
*reg
,
1327 enum tgsi_opcode_type type
, unsigned swizzle
)
1329 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1330 LLVMValueRef base
, addr
;
1332 base
= LLVMGetParam(ctx
->main_fn
, ctx
->param_tcs_offchip_offset
);
1333 addr
= get_tcs_tes_buffer_address_from_reg(ctx
, NULL
, reg
);
1335 return buffer_load(bld_base
, tgsi2llvmtype(bld_base
, type
), swizzle
,
1336 ctx
->tess_offchip_ring
, base
, addr
, true);
1339 LLVMValueRef
si_nir_load_input_tes(struct ac_shader_abi
*abi
,
1341 LLVMValueRef vertex_index
,
1342 LLVMValueRef param_index
,
1343 unsigned const_index
,
1345 unsigned driver_location
,
1347 unsigned num_components
,
1352 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1353 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
1354 LLVMValueRef base
, addr
;
1356 driver_location
= driver_location
/ 4;
1358 base
= LLVMGetParam(ctx
->main_fn
, ctx
->param_tcs_offchip_offset
);
1361 /* Add the constant index to the indirect index */
1362 param_index
= LLVMBuildAdd(ctx
->ac
.builder
, param_index
,
1363 LLVMConstInt(ctx
->i32
, const_index
, 0), "");
1365 param_index
= LLVMConstInt(ctx
->i32
, const_index
, 0);
1368 addr
= get_tcs_tes_buffer_address_from_generic_indices(ctx
, vertex_index
,
1369 param_index
, driver_location
,
1370 info
->input_semantic_name
,
1371 info
->input_semantic_index
,
1374 /* TODO: This will generate rather ordinary llvm code, although it
1375 * should be easy for the optimiser to fix up. In future we might want
1376 * to refactor buffer_load(), but for now this maximises code sharing
1377 * between the NIR and TGSI backends.
1379 LLVMValueRef value
[4];
1380 for (unsigned i
= component
; i
< num_components
+ component
; i
++) {
1381 unsigned offset
= i
;
1382 if (llvm_type_is_64bit(ctx
, type
))
1385 value
[i
] = buffer_load(&ctx
->bld_base
, type
, offset
,
1386 ctx
->tess_offchip_ring
, base
, addr
, true);
1389 return ac_build_varying_gather_values(&ctx
->ac
, value
, num_components
, component
);
1392 static void store_output_tcs(struct lp_build_tgsi_context
*bld_base
,
1393 const struct tgsi_full_instruction
*inst
,
1394 const struct tgsi_opcode_info
*info
,
1396 LLVMValueRef dst
[4])
1398 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1399 const struct tgsi_full_dst_register
*reg
= &inst
->Dst
[index
];
1400 const struct tgsi_shader_info
*sh_info
= &ctx
->shader
->selector
->info
;
1401 unsigned chan_index
;
1402 LLVMValueRef dw_addr
, stride
;
1403 LLVMValueRef buffer
, base
, buf_addr
;
1404 LLVMValueRef values
[4];
1405 bool skip_lds_store
;
1406 bool is_tess_factor
= false, is_tess_inner
= false;
1408 /* Only handle per-patch and per-vertex outputs here.
1409 * Vectors will be lowered to scalars and this function will be called again.
1411 if (reg
->Register
.File
!= TGSI_FILE_OUTPUT
||
1412 (dst
[0] && LLVMGetTypeKind(LLVMTypeOf(dst
[0])) == LLVMVectorTypeKind
)) {
1413 si_llvm_emit_store(bld_base
, inst
, info
, index
, dst
);
1417 if (reg
->Register
.Dimension
) {
1418 stride
= get_tcs_out_vertex_dw_stride(ctx
);
1419 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
1420 dw_addr
= get_dw_address(ctx
, reg
, NULL
, stride
, dw_addr
);
1421 skip_lds_store
= !sh_info
->reads_pervertex_outputs
;
1423 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
1424 dw_addr
= get_dw_address(ctx
, reg
, NULL
, NULL
, dw_addr
);
1425 skip_lds_store
= !sh_info
->reads_perpatch_outputs
;
1427 if (!reg
->Register
.Indirect
) {
1428 int name
= sh_info
->output_semantic_name
[reg
->Register
.Index
];
1430 /* Always write tess factors into LDS for the TCS epilog. */
1431 if (name
== TGSI_SEMANTIC_TESSINNER
||
1432 name
== TGSI_SEMANTIC_TESSOUTER
) {
1433 /* The epilog doesn't read LDS if invocation 0 defines tess factors. */
1434 skip_lds_store
= !sh_info
->reads_tessfactor_outputs
&&
1435 ctx
->shader
->selector
->tcs_info
.tessfactors_are_def_in_all_invocs
;
1436 is_tess_factor
= true;
1437 is_tess_inner
= name
== TGSI_SEMANTIC_TESSINNER
;
1442 buffer
= get_tess_ring_descriptor(ctx
, TESS_OFFCHIP_RING_TCS
);
1444 base
= LLVMGetParam(ctx
->main_fn
, ctx
->param_tcs_offchip_offset
);
1445 buf_addr
= get_tcs_tes_buffer_address_from_reg(ctx
, reg
, NULL
);
1447 uint32_t writemask
= reg
->Register
.WriteMask
;
1449 chan_index
= u_bit_scan(&writemask
);
1450 LLVMValueRef value
= dst
[chan_index
];
1452 if (inst
->Instruction
.Saturate
)
1453 value
= ac_build_clamp(&ctx
->ac
, value
);
1455 /* Skip LDS stores if there is no LDS read of this output. */
1456 if (!skip_lds_store
)
1457 lds_store(ctx
, chan_index
, dw_addr
, value
);
1459 value
= ac_to_integer(&ctx
->ac
, value
);
1460 values
[chan_index
] = value
;
1462 if (reg
->Register
.WriteMask
!= 0xF && !is_tess_factor
) {
1463 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, value
, 1,
1465 4 * chan_index
, 1, 0, true, false);
1468 /* Write tess factors into VGPRs for the epilog. */
1469 if (is_tess_factor
&&
1470 ctx
->shader
->selector
->tcs_info
.tessfactors_are_def_in_all_invocs
) {
1471 if (!is_tess_inner
) {
1472 LLVMBuildStore(ctx
->ac
.builder
, value
, /* outer */
1473 ctx
->invoc0_tess_factors
[chan_index
]);
1474 } else if (chan_index
< 2) {
1475 LLVMBuildStore(ctx
->ac
.builder
, value
, /* inner */
1476 ctx
->invoc0_tess_factors
[4 + chan_index
]);
1481 if (reg
->Register
.WriteMask
== 0xF && !is_tess_factor
) {
1482 LLVMValueRef value
= lp_build_gather_values(&ctx
->gallivm
,
1484 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, value
, 4, buf_addr
,
1485 base
, 0, 1, 0, true, false);
1489 static void si_nir_store_output_tcs(struct ac_shader_abi
*abi
,
1490 const struct nir_variable
*var
,
1491 LLVMValueRef vertex_index
,
1492 LLVMValueRef param_index
,
1493 unsigned const_index
,
1497 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1498 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
1499 const unsigned component
= var
->data
.location_frac
;
1500 const bool is_patch
= var
->data
.patch
;
1501 unsigned driver_location
= var
->data
.driver_location
;
1502 LLVMValueRef dw_addr
, stride
;
1503 LLVMValueRef buffer
, base
, addr
;
1504 LLVMValueRef values
[4];
1505 bool skip_lds_store
;
1506 bool is_tess_factor
= false, is_tess_inner
= false;
1508 driver_location
= driver_location
/ 4;
1511 /* Add the constant index to the indirect index */
1512 param_index
= LLVMBuildAdd(ctx
->ac
.builder
, param_index
,
1513 LLVMConstInt(ctx
->i32
, const_index
, 0), "");
1515 if (const_index
!= 0)
1516 param_index
= LLVMConstInt(ctx
->i32
, const_index
, 0);
1520 stride
= get_tcs_out_vertex_dw_stride(ctx
);
1521 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
1522 dw_addr
= get_dw_address_from_generic_indices(ctx
, stride
, dw_addr
,
1523 vertex_index
, param_index
,
1525 info
->output_semantic_name
,
1526 info
->output_semantic_index
,
1529 skip_lds_store
= !info
->reads_pervertex_outputs
;
1531 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
1532 dw_addr
= get_dw_address_from_generic_indices(ctx
, NULL
, dw_addr
,
1533 vertex_index
, param_index
,
1535 info
->output_semantic_name
,
1536 info
->output_semantic_index
,
1539 skip_lds_store
= !info
->reads_perpatch_outputs
;
1542 int name
= info
->output_semantic_name
[driver_location
];
1544 /* Always write tess factors into LDS for the TCS epilog. */
1545 if (name
== TGSI_SEMANTIC_TESSINNER
||
1546 name
== TGSI_SEMANTIC_TESSOUTER
) {
1547 /* The epilog doesn't read LDS if invocation 0 defines tess factors. */
1548 skip_lds_store
= !info
->reads_tessfactor_outputs
&&
1549 ctx
->shader
->selector
->tcs_info
.tessfactors_are_def_in_all_invocs
;
1550 is_tess_factor
= true;
1551 is_tess_inner
= name
== TGSI_SEMANTIC_TESSINNER
;
1556 buffer
= get_tess_ring_descriptor(ctx
, TESS_OFFCHIP_RING_TCS
);
1558 base
= LLVMGetParam(ctx
->main_fn
, ctx
->param_tcs_offchip_offset
);
1560 addr
= get_tcs_tes_buffer_address_from_generic_indices(ctx
, vertex_index
,
1561 param_index
, driver_location
,
1562 info
->output_semantic_name
,
1563 info
->output_semantic_index
,
1566 for (unsigned chan
= 0; chan
< 4; chan
++) {
1567 if (!(writemask
& (1 << chan
)))
1569 LLVMValueRef value
= ac_llvm_extract_elem(&ctx
->ac
, src
, chan
- component
);
1571 /* Skip LDS stores if there is no LDS read of this output. */
1572 if (!skip_lds_store
)
1573 lds_store(ctx
, chan
, dw_addr
, value
);
1575 value
= ac_to_integer(&ctx
->ac
, value
);
1576 values
[chan
] = value
;
1578 if (writemask
!= 0xF && !is_tess_factor
) {
1579 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, value
, 1,
1581 4 * chan
, 1, 0, true, false);
1584 /* Write tess factors into VGPRs for the epilog. */
1585 if (is_tess_factor
&&
1586 ctx
->shader
->selector
->tcs_info
.tessfactors_are_def_in_all_invocs
) {
1587 if (!is_tess_inner
) {
1588 LLVMBuildStore(ctx
->ac
.builder
, value
, /* outer */
1589 ctx
->invoc0_tess_factors
[chan
]);
1590 } else if (chan
< 2) {
1591 LLVMBuildStore(ctx
->ac
.builder
, value
, /* inner */
1592 ctx
->invoc0_tess_factors
[4 + chan
]);
1597 if (writemask
== 0xF && !is_tess_factor
) {
1598 LLVMValueRef value
= lp_build_gather_values(&ctx
->gallivm
,
1600 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, value
, 4, addr
,
1601 base
, 0, 1, 0, true, false);
1605 LLVMValueRef
si_llvm_load_input_gs(struct ac_shader_abi
*abi
,
1606 unsigned input_index
,
1607 unsigned vtx_offset_param
,
1611 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1612 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
1613 struct si_shader
*shader
= ctx
->shader
;
1614 struct lp_build_context
*uint
= &ctx
->bld_base
.uint_bld
;
1615 LLVMValueRef vtx_offset
, soffset
;
1616 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
1617 unsigned semantic_name
= info
->input_semantic_name
[input_index
];
1618 unsigned semantic_index
= info
->input_semantic_index
[input_index
];
1622 param
= si_shader_io_get_unique_index(semantic_name
, semantic_index
);
1624 /* GFX9 has the ESGS ring in LDS. */
1625 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
1626 unsigned index
= vtx_offset_param
;
1628 switch (index
/ 2) {
1630 vtx_offset
= unpack_param(ctx
, ctx
->param_gs_vtx01_offset
,
1631 index
% 2 ? 16 : 0, 16);
1634 vtx_offset
= unpack_param(ctx
, ctx
->param_gs_vtx23_offset
,
1635 index
% 2 ? 16 : 0, 16);
1638 vtx_offset
= unpack_param(ctx
, ctx
->param_gs_vtx45_offset
,
1639 index
% 2 ? 16 : 0, 16);
1646 vtx_offset
= LLVMBuildAdd(ctx
->ac
.builder
, vtx_offset
,
1647 LLVMConstInt(ctx
->i32
, param
* 4, 0), "");
1648 return lds_load(bld_base
, type
, swizzle
, vtx_offset
);
1651 /* GFX6: input load from the ESGS ring in memory. */
1652 if (swizzle
== ~0) {
1653 LLVMValueRef values
[TGSI_NUM_CHANNELS
];
1655 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
1656 values
[chan
] = si_llvm_load_input_gs(abi
, input_index
, vtx_offset_param
,
1659 return lp_build_gather_values(&ctx
->gallivm
, values
,
1663 /* Get the vertex offset parameter on GFX6. */
1664 LLVMValueRef gs_vtx_offset
= ctx
->gs_vtx_offset
[vtx_offset_param
];
1666 vtx_offset
= lp_build_mul_imm(uint
, gs_vtx_offset
, 4);
1668 soffset
= LLVMConstInt(ctx
->i32
, (param
* 4 + swizzle
) * 256, 0);
1670 value
= ac_build_buffer_load(&ctx
->ac
, ctx
->esgs_ring
, 1, ctx
->i32_0
,
1671 vtx_offset
, soffset
, 0, 1, 0, true, false);
1672 if (llvm_type_is_64bit(ctx
, type
)) {
1673 LLVMValueRef value2
;
1674 soffset
= LLVMConstInt(ctx
->i32
, (param
* 4 + swizzle
+ 1) * 256, 0);
1676 value2
= ac_build_buffer_load(&ctx
->ac
, ctx
->esgs_ring
, 1,
1677 ctx
->i32_0
, vtx_offset
, soffset
,
1678 0, 1, 0, true, false);
1679 return si_llvm_emit_fetch_64bit(bld_base
, type
, value
, value2
);
1681 return LLVMBuildBitCast(ctx
->ac
.builder
, value
, type
, "");
1684 static LLVMValueRef
si_nir_load_input_gs(struct ac_shader_abi
*abi
,
1686 unsigned driver_location
,
1688 unsigned num_components
,
1689 unsigned vertex_index
,
1690 unsigned const_index
,
1693 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1695 LLVMValueRef value
[4];
1696 for (unsigned i
= component
; i
< num_components
+ component
; i
++) {
1697 unsigned offset
= i
;
1698 if (llvm_type_is_64bit(ctx
, type
))
1701 value
[i
] = si_llvm_load_input_gs(&ctx
->abi
, driver_location
/ 4,
1702 vertex_index
, type
, offset
);
1705 return ac_build_varying_gather_values(&ctx
->ac
, value
, num_components
, component
);
1708 static LLVMValueRef
fetch_input_gs(
1709 struct lp_build_tgsi_context
*bld_base
,
1710 const struct tgsi_full_src_register
*reg
,
1711 enum tgsi_opcode_type type
,
1714 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1715 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
1717 unsigned semantic_name
= info
->input_semantic_name
[reg
->Register
.Index
];
1718 if (swizzle
!= ~0 && semantic_name
== TGSI_SEMANTIC_PRIMID
)
1719 return get_primitive_id(ctx
, swizzle
);
1721 if (!reg
->Register
.Dimension
)
1724 return si_llvm_load_input_gs(&ctx
->abi
, reg
->Register
.Index
,
1725 reg
->Dimension
.Index
,
1726 tgsi2llvmtype(bld_base
, type
),
1730 static int lookup_interp_param_index(unsigned interpolate
, unsigned location
)
1732 switch (interpolate
) {
1733 case TGSI_INTERPOLATE_CONSTANT
:
1736 case TGSI_INTERPOLATE_LINEAR
:
1737 if (location
== TGSI_INTERPOLATE_LOC_SAMPLE
)
1738 return SI_PARAM_LINEAR_SAMPLE
;
1739 else if (location
== TGSI_INTERPOLATE_LOC_CENTROID
)
1740 return SI_PARAM_LINEAR_CENTROID
;
1742 return SI_PARAM_LINEAR_CENTER
;
1744 case TGSI_INTERPOLATE_COLOR
:
1745 case TGSI_INTERPOLATE_PERSPECTIVE
:
1746 if (location
== TGSI_INTERPOLATE_LOC_SAMPLE
)
1747 return SI_PARAM_PERSP_SAMPLE
;
1748 else if (location
== TGSI_INTERPOLATE_LOC_CENTROID
)
1749 return SI_PARAM_PERSP_CENTROID
;
1751 return SI_PARAM_PERSP_CENTER
;
1754 fprintf(stderr
, "Warning: Unhandled interpolation mode.\n");
1759 static LLVMValueRef
si_build_fs_interp(struct si_shader_context
*ctx
,
1760 unsigned attr_index
, unsigned chan
,
1761 LLVMValueRef prim_mask
,
1762 LLVMValueRef i
, LLVMValueRef j
)
1765 return ac_build_fs_interp(&ctx
->ac
,
1766 LLVMConstInt(ctx
->i32
, chan
, 0),
1767 LLVMConstInt(ctx
->i32
, attr_index
, 0),
1770 return ac_build_fs_interp_mov(&ctx
->ac
,
1771 LLVMConstInt(ctx
->i32
, 2, 0), /* P0 */
1772 LLVMConstInt(ctx
->i32
, chan
, 0),
1773 LLVMConstInt(ctx
->i32
, attr_index
, 0),
1778 * Interpolate a fragment shader input.
1780 * @param ctx context
1781 * @param input_index index of the input in hardware
1782 * @param semantic_name TGSI_SEMANTIC_*
1783 * @param semantic_index semantic index
1784 * @param num_interp_inputs number of all interpolated inputs (= BCOLOR offset)
1785 * @param colors_read_mask color components read (4 bits for each color, 8 bits in total)
1786 * @param interp_param interpolation weights (i,j)
1787 * @param prim_mask SI_PARAM_PRIM_MASK
1788 * @param face SI_PARAM_FRONT_FACE
1789 * @param result the return value (4 components)
1791 static void interp_fs_input(struct si_shader_context
*ctx
,
1792 unsigned input_index
,
1793 unsigned semantic_name
,
1794 unsigned semantic_index
,
1795 unsigned num_interp_inputs
,
1796 unsigned colors_read_mask
,
1797 LLVMValueRef interp_param
,
1798 LLVMValueRef prim_mask
,
1800 LLVMValueRef result
[4])
1802 LLVMValueRef i
= NULL
, j
= NULL
;
1805 /* fs.constant returns the param from the middle vertex, so it's not
1806 * really useful for flat shading. It's meant to be used for custom
1807 * interpolation (but the intrinsic can't fetch from the other two
1810 * Luckily, it doesn't matter, because we rely on the FLAT_SHADE state
1811 * to do the right thing. The only reason we use fs.constant is that
1812 * fs.interp cannot be used on integers, because they can be equal
1815 * When interp is false we will use fs.constant or for newer llvm,
1816 * amdgcn.interp.mov.
1818 bool interp
= interp_param
!= NULL
;
1821 interp_param
= LLVMBuildBitCast(ctx
->ac
.builder
, interp_param
,
1822 LLVMVectorType(ctx
->f32
, 2), "");
1824 i
= LLVMBuildExtractElement(ctx
->ac
.builder
, interp_param
,
1826 j
= LLVMBuildExtractElement(ctx
->ac
.builder
, interp_param
,
1830 if (semantic_name
== TGSI_SEMANTIC_COLOR
&&
1831 ctx
->shader
->key
.part
.ps
.prolog
.color_two_side
) {
1832 LLVMValueRef is_face_positive
;
1834 /* If BCOLOR0 is used, BCOLOR1 is at offset "num_inputs + 1",
1835 * otherwise it's at offset "num_inputs".
1837 unsigned back_attr_offset
= num_interp_inputs
;
1838 if (semantic_index
== 1 && colors_read_mask
& 0xf)
1839 back_attr_offset
+= 1;
1841 is_face_positive
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntNE
,
1842 face
, ctx
->i32_0
, "");
1844 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
1845 LLVMValueRef front
, back
;
1847 front
= si_build_fs_interp(ctx
,
1850 back
= si_build_fs_interp(ctx
,
1851 back_attr_offset
, chan
,
1854 result
[chan
] = LLVMBuildSelect(ctx
->ac
.builder
,
1860 } else if (semantic_name
== TGSI_SEMANTIC_FOG
) {
1861 result
[0] = si_build_fs_interp(ctx
, input_index
,
1862 0, prim_mask
, i
, j
);
1864 result
[2] = LLVMConstReal(ctx
->f32
, 0.0f
);
1865 result
[3] = LLVMConstReal(ctx
->f32
, 1.0f
);
1867 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
1868 result
[chan
] = si_build_fs_interp(ctx
,
1875 void si_llvm_load_input_fs(
1876 struct si_shader_context
*ctx
,
1877 unsigned input_index
,
1878 LLVMValueRef out
[4])
1880 struct lp_build_context
*base
= &ctx
->bld_base
.base
;
1881 struct si_shader
*shader
= ctx
->shader
;
1882 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
1883 LLVMValueRef main_fn
= ctx
->main_fn
;
1884 LLVMValueRef interp_param
= NULL
;
1885 int interp_param_idx
;
1886 enum tgsi_semantic semantic_name
= info
->input_semantic_name
[input_index
];
1887 unsigned semantic_index
= info
->input_semantic_index
[input_index
];
1888 enum tgsi_interpolate_mode interp_mode
= info
->input_interpolate
[input_index
];
1889 enum tgsi_interpolate_loc interp_loc
= info
->input_interpolate_loc
[input_index
];
1891 /* Get colors from input VGPRs (set by the prolog). */
1892 if (semantic_name
== TGSI_SEMANTIC_COLOR
) {
1893 unsigned colors_read
= shader
->selector
->info
.colors_read
;
1894 unsigned mask
= colors_read
>> (semantic_index
* 4);
1895 unsigned offset
= SI_PARAM_POS_FIXED_PT
+ 1 +
1896 (semantic_index
? util_bitcount(colors_read
& 0xf) : 0);
1898 out
[0] = mask
& 0x1 ? LLVMGetParam(main_fn
, offset
++) : base
->undef
;
1899 out
[1] = mask
& 0x2 ? LLVMGetParam(main_fn
, offset
++) : base
->undef
;
1900 out
[2] = mask
& 0x4 ? LLVMGetParam(main_fn
, offset
++) : base
->undef
;
1901 out
[3] = mask
& 0x8 ? LLVMGetParam(main_fn
, offset
++) : base
->undef
;
1905 interp_param_idx
= lookup_interp_param_index(interp_mode
, interp_loc
);
1906 if (interp_param_idx
== -1)
1908 else if (interp_param_idx
) {
1909 interp_param
= LLVMGetParam(ctx
->main_fn
, interp_param_idx
);
1912 interp_fs_input(ctx
, input_index
, semantic_name
,
1913 semantic_index
, 0, /* this param is unused */
1914 shader
->selector
->info
.colors_read
, interp_param
,
1916 LLVMGetParam(main_fn
, SI_PARAM_FRONT_FACE
),
1920 static void declare_input_fs(
1921 struct si_shader_context
*ctx
,
1922 unsigned input_index
,
1923 const struct tgsi_full_declaration
*decl
,
1924 LLVMValueRef out
[4])
1926 si_llvm_load_input_fs(ctx
, input_index
, out
);
1929 static LLVMValueRef
get_sample_id(struct si_shader_context
*ctx
)
1931 return unpack_param(ctx
, SI_PARAM_ANCILLARY
, 8, 4);
1934 static LLVMValueRef
get_base_vertex(struct ac_shader_abi
*abi
)
1936 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1938 /* For non-indexed draws, the base vertex set by the driver
1939 * (for direct draws) or the CP (for indirect draws) is the
1940 * first vertex ID, but GLSL expects 0 to be returned.
1942 LLVMValueRef vs_state
= LLVMGetParam(ctx
->main_fn
,
1943 ctx
->param_vs_state_bits
);
1944 LLVMValueRef indexed
;
1946 indexed
= LLVMBuildLShr(ctx
->ac
.builder
, vs_state
, ctx
->i32_1
, "");
1947 indexed
= LLVMBuildTrunc(ctx
->ac
.builder
, indexed
, ctx
->i1
, "");
1949 return LLVMBuildSelect(ctx
->ac
.builder
, indexed
, ctx
->abi
.base_vertex
,
1953 static LLVMValueRef
get_block_size(struct ac_shader_abi
*abi
)
1955 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1957 LLVMValueRef values
[3];
1958 LLVMValueRef result
;
1960 unsigned *properties
= ctx
->shader
->selector
->info
.properties
;
1962 if (properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
] != 0) {
1963 unsigned sizes
[3] = {
1964 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
],
1965 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT
],
1966 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH
]
1969 for (i
= 0; i
< 3; ++i
)
1970 values
[i
] = LLVMConstInt(ctx
->i32
, sizes
[i
], 0);
1972 result
= lp_build_gather_values(&ctx
->gallivm
, values
, 3);
1974 result
= LLVMGetParam(ctx
->main_fn
, ctx
->param_block_size
);
1981 * Load a dword from a constant buffer.
1983 static LLVMValueRef
buffer_load_const(struct si_shader_context
*ctx
,
1984 LLVMValueRef resource
,
1985 LLVMValueRef offset
)
1987 return ac_build_buffer_load(&ctx
->ac
, resource
, 1, NULL
, offset
, NULL
,
1988 0, 0, 0, true, true);
1991 static LLVMValueRef
load_sample_position(struct ac_shader_abi
*abi
, LLVMValueRef sample_id
)
1993 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1994 struct lp_build_context
*uint_bld
= &ctx
->bld_base
.uint_bld
;
1995 LLVMValueRef desc
= LLVMGetParam(ctx
->main_fn
, ctx
->param_rw_buffers
);
1996 LLVMValueRef buf_index
= LLVMConstInt(ctx
->i32
, SI_PS_CONST_SAMPLE_POSITIONS
, 0);
1997 LLVMValueRef resource
= ac_build_load_to_sgpr(&ctx
->ac
, desc
, buf_index
);
1999 /* offset = sample_id * 8 (8 = 2 floats containing samplepos.xy) */
2000 LLVMValueRef offset0
= lp_build_mul_imm(uint_bld
, sample_id
, 8);
2001 LLVMValueRef offset1
= LLVMBuildAdd(ctx
->ac
.builder
, offset0
, LLVMConstInt(ctx
->i32
, 4, 0), "");
2003 LLVMValueRef pos
[4] = {
2004 buffer_load_const(ctx
, resource
, offset0
),
2005 buffer_load_const(ctx
, resource
, offset1
),
2006 LLVMConstReal(ctx
->f32
, 0),
2007 LLVMConstReal(ctx
->f32
, 0)
2010 return lp_build_gather_values(&ctx
->gallivm
, pos
, 4);
2013 static LLVMValueRef
load_sample_mask_in(struct ac_shader_abi
*abi
)
2015 return abi
->sample_coverage
;
2018 static LLVMValueRef
si_load_tess_coord(struct ac_shader_abi
*abi
)
2020 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
2021 struct lp_build_context
*bld
= &ctx
->bld_base
.base
;
2023 LLVMValueRef coord
[4] = {
2024 LLVMGetParam(ctx
->main_fn
, ctx
->param_tes_u
),
2025 LLVMGetParam(ctx
->main_fn
, ctx
->param_tes_v
),
2030 /* For triangles, the vector should be (u, v, 1-u-v). */
2031 if (ctx
->shader
->selector
->info
.properties
[TGSI_PROPERTY_TES_PRIM_MODE
] ==
2032 PIPE_PRIM_TRIANGLES
)
2033 coord
[2] = lp_build_sub(bld
, ctx
->ac
.f32_1
,
2034 lp_build_add(bld
, coord
[0], coord
[1]));
2036 return lp_build_gather_values(&ctx
->gallivm
, coord
, 4);
2039 static LLVMValueRef
load_tess_level(struct si_shader_context
*ctx
,
2040 unsigned semantic_name
)
2042 LLVMValueRef base
, addr
;
2044 int param
= si_shader_io_get_unique_index_patch(semantic_name
, 0);
2046 base
= LLVMGetParam(ctx
->main_fn
, ctx
->param_tcs_offchip_offset
);
2047 addr
= get_tcs_tes_buffer_address(ctx
, get_rel_patch_id(ctx
), NULL
,
2048 LLVMConstInt(ctx
->i32
, param
, 0));
2050 return buffer_load(&ctx
->bld_base
, ctx
->f32
,
2051 ~0, ctx
->tess_offchip_ring
, base
, addr
, true);
2055 static LLVMValueRef
si_load_tess_level(struct ac_shader_abi
*abi
,
2056 unsigned varying_id
)
2058 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
2059 unsigned semantic_name
;
2061 switch (varying_id
) {
2062 case VARYING_SLOT_TESS_LEVEL_INNER
:
2063 semantic_name
= TGSI_SEMANTIC_TESSINNER
;
2065 case VARYING_SLOT_TESS_LEVEL_OUTER
:
2066 semantic_name
= TGSI_SEMANTIC_TESSOUTER
;
2069 unreachable("unknown tess level");
2072 return load_tess_level(ctx
, semantic_name
);
2076 static LLVMValueRef
si_load_patch_vertices_in(struct ac_shader_abi
*abi
)
2078 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
2079 if (ctx
->type
== PIPE_SHADER_TESS_CTRL
)
2080 return unpack_param(ctx
, ctx
->param_tcs_out_lds_layout
, 13, 6);
2081 else if (ctx
->type
== PIPE_SHADER_TESS_EVAL
)
2082 return get_num_tcs_out_vertices(ctx
);
2084 unreachable("invalid shader stage for TGSI_SEMANTIC_VERTICESIN");
2087 void si_load_system_value(struct si_shader_context
*ctx
,
2089 const struct tgsi_full_declaration
*decl
)
2091 LLVMValueRef value
= 0;
2093 assert(index
< RADEON_LLVM_MAX_SYSTEM_VALUES
);
2095 switch (decl
->Semantic
.Name
) {
2096 case TGSI_SEMANTIC_INSTANCEID
:
2097 value
= ctx
->abi
.instance_id
;
2100 case TGSI_SEMANTIC_VERTEXID
:
2101 value
= LLVMBuildAdd(ctx
->ac
.builder
,
2103 ctx
->abi
.base_vertex
, "");
2106 case TGSI_SEMANTIC_VERTEXID_NOBASE
:
2107 /* Unused. Clarify the meaning in indexed vs. non-indexed
2108 * draws if this is ever used again. */
2112 case TGSI_SEMANTIC_BASEVERTEX
:
2113 value
= get_base_vertex(&ctx
->abi
);
2116 case TGSI_SEMANTIC_BASEINSTANCE
:
2117 value
= ctx
->abi
.start_instance
;
2120 case TGSI_SEMANTIC_DRAWID
:
2121 value
= ctx
->abi
.draw_id
;
2124 case TGSI_SEMANTIC_INVOCATIONID
:
2125 if (ctx
->type
== PIPE_SHADER_TESS_CTRL
)
2126 value
= unpack_llvm_param(ctx
, ctx
->abi
.tcs_rel_ids
, 8, 5);
2127 else if (ctx
->type
== PIPE_SHADER_GEOMETRY
)
2128 value
= ctx
->abi
.gs_invocation_id
;
2130 assert(!"INVOCATIONID not implemented");
2133 case TGSI_SEMANTIC_POSITION
:
2135 LLVMValueRef pos
[4] = {
2136 LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_X_FLOAT
),
2137 LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_Y_FLOAT
),
2138 LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_Z_FLOAT
),
2139 lp_build_emit_llvm_unary(&ctx
->bld_base
, TGSI_OPCODE_RCP
,
2140 LLVMGetParam(ctx
->main_fn
,
2141 SI_PARAM_POS_W_FLOAT
)),
2143 value
= lp_build_gather_values(&ctx
->gallivm
, pos
, 4);
2147 case TGSI_SEMANTIC_FACE
:
2148 value
= ctx
->abi
.front_face
;
2151 case TGSI_SEMANTIC_SAMPLEID
:
2152 value
= get_sample_id(ctx
);
2155 case TGSI_SEMANTIC_SAMPLEPOS
: {
2156 LLVMValueRef pos
[4] = {
2157 LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_X_FLOAT
),
2158 LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_Y_FLOAT
),
2159 LLVMConstReal(ctx
->f32
, 0),
2160 LLVMConstReal(ctx
->f32
, 0)
2162 pos
[0] = lp_build_emit_llvm_unary(&ctx
->bld_base
,
2163 TGSI_OPCODE_FRC
, pos
[0]);
2164 pos
[1] = lp_build_emit_llvm_unary(&ctx
->bld_base
,
2165 TGSI_OPCODE_FRC
, pos
[1]);
2166 value
= lp_build_gather_values(&ctx
->gallivm
, pos
, 4);
2170 case TGSI_SEMANTIC_SAMPLEMASK
:
2171 /* This can only occur with the OpenGL Core profile, which
2172 * doesn't support smoothing.
2174 value
= LLVMGetParam(ctx
->main_fn
, SI_PARAM_SAMPLE_COVERAGE
);
2177 case TGSI_SEMANTIC_TESSCOORD
:
2178 value
= si_load_tess_coord(&ctx
->abi
);
2181 case TGSI_SEMANTIC_VERTICESIN
:
2182 value
= si_load_patch_vertices_in(&ctx
->abi
);
2185 case TGSI_SEMANTIC_TESSINNER
:
2186 case TGSI_SEMANTIC_TESSOUTER
:
2187 value
= load_tess_level(ctx
, decl
->Semantic
.Name
);
2190 case TGSI_SEMANTIC_DEFAULT_TESSOUTER_SI
:
2191 case TGSI_SEMANTIC_DEFAULT_TESSINNER_SI
:
2193 LLVMValueRef buf
, slot
, val
[4];
2196 slot
= LLVMConstInt(ctx
->i32
, SI_HS_CONST_DEFAULT_TESS_LEVELS
, 0);
2197 buf
= LLVMGetParam(ctx
->main_fn
, ctx
->param_rw_buffers
);
2198 buf
= ac_build_load_to_sgpr(&ctx
->ac
, buf
, slot
);
2199 offset
= decl
->Semantic
.Name
== TGSI_SEMANTIC_DEFAULT_TESSINNER_SI
? 4 : 0;
2201 for (i
= 0; i
< 4; i
++)
2202 val
[i
] = buffer_load_const(ctx
, buf
,
2203 LLVMConstInt(ctx
->i32
, (offset
+ i
) * 4, 0));
2204 value
= lp_build_gather_values(&ctx
->gallivm
, val
, 4);
2208 case TGSI_SEMANTIC_PRIMID
:
2209 value
= get_primitive_id(ctx
, 0);
2212 case TGSI_SEMANTIC_GRID_SIZE
:
2213 value
= ctx
->abi
.num_work_groups
;
2216 case TGSI_SEMANTIC_BLOCK_SIZE
:
2217 value
= get_block_size(&ctx
->abi
);
2220 case TGSI_SEMANTIC_BLOCK_ID
:
2222 LLVMValueRef values
[3];
2224 for (int i
= 0; i
< 3; i
++) {
2225 values
[i
] = ctx
->i32_0
;
2226 if (ctx
->abi
.workgroup_ids
[i
]) {
2227 values
[i
] = ctx
->abi
.workgroup_ids
[i
];
2230 value
= lp_build_gather_values(&ctx
->gallivm
, values
, 3);
2234 case TGSI_SEMANTIC_THREAD_ID
:
2235 value
= ctx
->abi
.local_invocation_ids
;
2238 case TGSI_SEMANTIC_HELPER_INVOCATION
:
2239 value
= lp_build_intrinsic(ctx
->ac
.builder
,
2240 "llvm.amdgcn.ps.live",
2242 LP_FUNC_ATTR_READNONE
);
2243 value
= LLVMBuildNot(ctx
->ac
.builder
, value
, "");
2244 value
= LLVMBuildSExt(ctx
->ac
.builder
, value
, ctx
->i32
, "");
2247 case TGSI_SEMANTIC_SUBGROUP_SIZE
:
2248 value
= LLVMConstInt(ctx
->i32
, 64, 0);
2251 case TGSI_SEMANTIC_SUBGROUP_INVOCATION
:
2252 value
= ac_get_thread_id(&ctx
->ac
);
2255 case TGSI_SEMANTIC_SUBGROUP_EQ_MASK
:
2257 LLVMValueRef id
= ac_get_thread_id(&ctx
->ac
);
2258 id
= LLVMBuildZExt(ctx
->ac
.builder
, id
, ctx
->i64
, "");
2259 value
= LLVMBuildShl(ctx
->ac
.builder
, LLVMConstInt(ctx
->i64
, 1, 0), id
, "");
2260 value
= LLVMBuildBitCast(ctx
->ac
.builder
, value
, ctx
->v2i32
, "");
2264 case TGSI_SEMANTIC_SUBGROUP_GE_MASK
:
2265 case TGSI_SEMANTIC_SUBGROUP_GT_MASK
:
2266 case TGSI_SEMANTIC_SUBGROUP_LE_MASK
:
2267 case TGSI_SEMANTIC_SUBGROUP_LT_MASK
:
2269 LLVMValueRef id
= ac_get_thread_id(&ctx
->ac
);
2270 if (decl
->Semantic
.Name
== TGSI_SEMANTIC_SUBGROUP_GT_MASK
||
2271 decl
->Semantic
.Name
== TGSI_SEMANTIC_SUBGROUP_LE_MASK
) {
2272 /* All bits set except LSB */
2273 value
= LLVMConstInt(ctx
->i64
, -2, 0);
2276 value
= LLVMConstInt(ctx
->i64
, -1, 0);
2278 id
= LLVMBuildZExt(ctx
->ac
.builder
, id
, ctx
->i64
, "");
2279 value
= LLVMBuildShl(ctx
->ac
.builder
, value
, id
, "");
2280 if (decl
->Semantic
.Name
== TGSI_SEMANTIC_SUBGROUP_LE_MASK
||
2281 decl
->Semantic
.Name
== TGSI_SEMANTIC_SUBGROUP_LT_MASK
)
2282 value
= LLVMBuildNot(ctx
->ac
.builder
, value
, "");
2283 value
= LLVMBuildBitCast(ctx
->ac
.builder
, value
, ctx
->v2i32
, "");
2288 assert(!"unknown system value");
2292 ctx
->system_values
[index
] = value
;
2295 void si_declare_compute_memory(struct si_shader_context
*ctx
)
2297 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
2299 LLVMTypeRef i8p
= LLVMPointerType(ctx
->i8
, AC_LOCAL_ADDR_SPACE
);
2302 assert(!ctx
->ac
.lds
);
2304 var
= LLVMAddGlobalInAddressSpace(ctx
->ac
.module
,
2305 LLVMArrayType(ctx
->i8
, sel
->local_size
),
2307 AC_LOCAL_ADDR_SPACE
);
2308 LLVMSetAlignment(var
, 4);
2310 ctx
->ac
.lds
= LLVMBuildBitCast(ctx
->ac
.builder
, var
, i8p
, "");
2313 void si_tgsi_declare_compute_memory(struct si_shader_context
*ctx
,
2314 const struct tgsi_full_declaration
*decl
)
2316 assert(decl
->Declaration
.MemType
== TGSI_MEMORY_TYPE_SHARED
);
2317 assert(decl
->Range
.First
== decl
->Range
.Last
);
2319 si_declare_compute_memory(ctx
);
2322 static LLVMValueRef
load_const_buffer_desc_fast_path(struct si_shader_context
*ctx
)
2325 LLVMGetParam(ctx
->main_fn
, ctx
->param_const_and_shader_buffers
);
2326 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
2328 /* Do the bounds checking with a descriptor, because
2329 * doing computation and manual bounds checking of 64-bit
2330 * addresses generates horrible VALU code with very high
2331 * VGPR usage and very low SIMD occupancy.
2333 ptr
= LLVMBuildPtrToInt(ctx
->ac
.builder
, ptr
, ctx
->ac
.intptr
, "");
2335 LLVMValueRef desc0
, desc1
;
2336 if (HAVE_32BIT_POINTERS
) {
2338 desc1
= LLVMConstInt(ctx
->i32
,
2339 S_008F04_BASE_ADDRESS_HI(ctx
->screen
->info
.address32_hi
), 0);
2341 ptr
= LLVMBuildBitCast(ctx
->ac
.builder
, ptr
, ctx
->v2i32
, "");
2342 desc0
= LLVMBuildExtractElement(ctx
->ac
.builder
, ptr
, ctx
->i32_0
, "");
2343 desc1
= LLVMBuildExtractElement(ctx
->ac
.builder
, ptr
, ctx
->i32_1
, "");
2344 /* Mask out all bits except BASE_ADDRESS_HI. */
2345 desc1
= LLVMBuildAnd(ctx
->ac
.builder
, desc1
,
2346 LLVMConstInt(ctx
->i32
, ~C_008F04_BASE_ADDRESS_HI
, 0), "");
2349 LLVMValueRef desc_elems
[] = {
2352 LLVMConstInt(ctx
->i32
, (sel
->info
.const_file_max
[0] + 1) * 16, 0),
2353 LLVMConstInt(ctx
->i32
,
2354 S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
2355 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
2356 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
2357 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
2358 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
2359 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
), 0)
2362 return ac_build_gather_values(&ctx
->ac
, desc_elems
, 4);
2365 static LLVMValueRef
load_const_buffer_desc(struct si_shader_context
*ctx
, int i
)
2367 LLVMValueRef list_ptr
= LLVMGetParam(ctx
->main_fn
,
2368 ctx
->param_const_and_shader_buffers
);
2370 return ac_build_load_to_sgpr(&ctx
->ac
, list_ptr
,
2371 LLVMConstInt(ctx
->i32
, si_get_constbuf_slot(i
), 0));
2374 static LLVMValueRef
load_ubo(struct ac_shader_abi
*abi
, LLVMValueRef index
)
2376 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
2377 LLVMValueRef ptr
= LLVMGetParam(ctx
->main_fn
, ctx
->param_const_and_shader_buffers
);
2379 index
= si_llvm_bound_index(ctx
, index
, ctx
->num_const_buffers
);
2380 index
= LLVMBuildAdd(ctx
->ac
.builder
, index
,
2381 LLVMConstInt(ctx
->i32
, SI_NUM_SHADER_BUFFERS
, 0), "");
2383 return ac_build_load_to_sgpr(&ctx
->ac
, ptr
, index
);
2387 load_ssbo(struct ac_shader_abi
*abi
, LLVMValueRef index
, bool write
)
2389 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
2390 LLVMValueRef rsrc_ptr
= LLVMGetParam(ctx
->main_fn
,
2391 ctx
->param_const_and_shader_buffers
);
2393 index
= si_llvm_bound_index(ctx
, index
, ctx
->num_shader_buffers
);
2394 index
= LLVMBuildSub(ctx
->ac
.builder
,
2395 LLVMConstInt(ctx
->i32
, SI_NUM_SHADER_BUFFERS
- 1, 0),
2398 return ac_build_load_to_sgpr(&ctx
->ac
, rsrc_ptr
, index
);
2401 static LLVMValueRef
fetch_constant(
2402 struct lp_build_tgsi_context
*bld_base
,
2403 const struct tgsi_full_src_register
*reg
,
2404 enum tgsi_opcode_type type
,
2407 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2408 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
2409 const struct tgsi_ind_register
*ireg
= ®
->Indirect
;
2412 LLVMValueRef addr
, bufp
;
2414 if (swizzle
== LP_CHAN_ALL
) {
2416 LLVMValueRef values
[4];
2417 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
)
2418 values
[chan
] = fetch_constant(bld_base
, reg
, type
, chan
);
2420 return lp_build_gather_values(&ctx
->gallivm
, values
, 4);
2423 /* Split 64-bit loads. */
2424 if (tgsi_type_is_64bit(type
)) {
2425 LLVMValueRef lo
, hi
;
2427 lo
= fetch_constant(bld_base
, reg
, TGSI_TYPE_UNSIGNED
, swizzle
);
2428 hi
= fetch_constant(bld_base
, reg
, TGSI_TYPE_UNSIGNED
, swizzle
+ 1);
2429 return si_llvm_emit_fetch_64bit(bld_base
, tgsi2llvmtype(bld_base
, type
),
2433 idx
= reg
->Register
.Index
* 4 + swizzle
;
2434 if (reg
->Register
.Indirect
) {
2435 addr
= si_get_indirect_index(ctx
, ireg
, 16, idx
* 4);
2437 addr
= LLVMConstInt(ctx
->i32
, idx
* 4, 0);
2440 /* Fast path when user data SGPRs point to constant buffer 0 directly. */
2441 if (sel
->info
.const_buffers_declared
== 1 &&
2442 sel
->info
.shader_buffers_declared
== 0) {
2444 /* This enables use of s_load_dword and flat_load_dword for const buffer 0
2445 * loads, and up to x4 load opcode merging. However, it leads to horrible
2446 * code reducing SIMD wave occupancy from 8 to 2 in many cases.
2448 * Using s_buffer_load_dword (x1) seems to be the best option right now.
2450 * LLVM 5.0 on SI doesn't insert a required s_nop between SALU setting
2451 * a descriptor and s_buffer_load_dword using it, so we can't expand
2452 * the pointer into a full descriptor like below. We have to use
2453 * s_load_dword instead. The only case when LLVM 5.0 would select
2454 * s_buffer_load_dword (that we have to prevent) is when we use use
2455 * a literal offset where we don't need bounds checking.
2457 if (ctx
->screen
->info
.chip_class
== SI
&& HAVE_LLVM
< 0x0600 &&
2458 !reg
->Register
.Indirect
) {
2460 LLVMGetParam(ctx
->main_fn
, ctx
->param_const_and_shader_buffers
);
2462 addr
= LLVMBuildLShr(ctx
->ac
.builder
, addr
, LLVMConstInt(ctx
->i32
, 2, 0), "");
2463 LLVMValueRef result
= ac_build_load_invariant(&ctx
->ac
, ptr
, addr
);
2464 return bitcast(bld_base
, type
, result
);
2467 LLVMValueRef desc
= load_const_buffer_desc_fast_path(ctx
);
2468 LLVMValueRef result
= buffer_load_const(ctx
, desc
, addr
);
2469 return bitcast(bld_base
, type
, result
);
2472 assert(reg
->Register
.Dimension
);
2473 buf
= reg
->Dimension
.Index
;
2475 if (reg
->Dimension
.Indirect
) {
2476 LLVMValueRef ptr
= LLVMGetParam(ctx
->main_fn
, ctx
->param_const_and_shader_buffers
);
2478 index
= si_get_bounded_indirect_index(ctx
, ®
->DimIndirect
,
2479 reg
->Dimension
.Index
,
2480 ctx
->num_const_buffers
);
2481 index
= LLVMBuildAdd(ctx
->ac
.builder
, index
,
2482 LLVMConstInt(ctx
->i32
, SI_NUM_SHADER_BUFFERS
, 0), "");
2483 bufp
= ac_build_load_to_sgpr(&ctx
->ac
, ptr
, index
);
2485 bufp
= load_const_buffer_desc(ctx
, buf
);
2487 return bitcast(bld_base
, type
, buffer_load_const(ctx
, bufp
, addr
));
2490 /* Initialize arguments for the shader export intrinsic */
2491 static void si_llvm_init_export_args(struct si_shader_context
*ctx
,
2492 LLVMValueRef
*values
,
2494 struct ac_export_args
*args
)
2496 LLVMValueRef f32undef
= LLVMGetUndef(ctx
->ac
.f32
);
2497 unsigned spi_shader_col_format
= V_028714_SPI_SHADER_32_ABGR
;
2499 bool is_int8
, is_int10
;
2501 /* Default is 0xf. Adjusted below depending on the format. */
2502 args
->enabled_channels
= 0xf; /* writemask */
2504 /* Specify whether the EXEC mask represents the valid mask */
2505 args
->valid_mask
= 0;
2507 /* Specify whether this is the last export */
2510 /* Specify the target we are exporting */
2511 args
->target
= target
;
2513 if (ctx
->type
== PIPE_SHADER_FRAGMENT
) {
2514 const struct si_shader_key
*key
= &ctx
->shader
->key
;
2515 unsigned col_formats
= key
->part
.ps
.epilog
.spi_shader_col_format
;
2516 int cbuf
= target
- V_008DFC_SQ_EXP_MRT
;
2518 assert(cbuf
>= 0 && cbuf
< 8);
2519 spi_shader_col_format
= (col_formats
>> (cbuf
* 4)) & 0xf;
2520 is_int8
= (key
->part
.ps
.epilog
.color_is_int8
>> cbuf
) & 0x1;
2521 is_int10
= (key
->part
.ps
.epilog
.color_is_int10
>> cbuf
) & 0x1;
2524 args
->compr
= false;
2525 args
->out
[0] = f32undef
;
2526 args
->out
[1] = f32undef
;
2527 args
->out
[2] = f32undef
;
2528 args
->out
[3] = f32undef
;
2530 LLVMValueRef (*packf
)(struct ac_llvm_context
*ctx
, LLVMValueRef args
[2]) = NULL
;
2531 LLVMValueRef (*packi
)(struct ac_llvm_context
*ctx
, LLVMValueRef args
[2],
2532 unsigned bits
, bool hi
) = NULL
;
2534 switch (spi_shader_col_format
) {
2535 case V_028714_SPI_SHADER_ZERO
:
2536 args
->enabled_channels
= 0; /* writemask */
2537 args
->target
= V_008DFC_SQ_EXP_NULL
;
2540 case V_028714_SPI_SHADER_32_R
:
2541 args
->enabled_channels
= 1; /* writemask */
2542 args
->out
[0] = values
[0];
2545 case V_028714_SPI_SHADER_32_GR
:
2546 args
->enabled_channels
= 0x3; /* writemask */
2547 args
->out
[0] = values
[0];
2548 args
->out
[1] = values
[1];
2551 case V_028714_SPI_SHADER_32_AR
:
2552 args
->enabled_channels
= 0x9; /* writemask */
2553 args
->out
[0] = values
[0];
2554 args
->out
[3] = values
[3];
2557 case V_028714_SPI_SHADER_FP16_ABGR
:
2558 packf
= ac_build_cvt_pkrtz_f16
;
2561 case V_028714_SPI_SHADER_UNORM16_ABGR
:
2562 packf
= ac_build_cvt_pknorm_u16
;
2565 case V_028714_SPI_SHADER_SNORM16_ABGR
:
2566 packf
= ac_build_cvt_pknorm_i16
;
2569 case V_028714_SPI_SHADER_UINT16_ABGR
:
2570 packi
= ac_build_cvt_pk_u16
;
2573 case V_028714_SPI_SHADER_SINT16_ABGR
:
2574 packi
= ac_build_cvt_pk_i16
;
2577 case V_028714_SPI_SHADER_32_ABGR
:
2578 memcpy(&args
->out
[0], values
, sizeof(values
[0]) * 4);
2582 /* Pack f16 or norm_i16/u16. */
2584 for (chan
= 0; chan
< 2; chan
++) {
2585 LLVMValueRef pack_args
[2] = {
2587 values
[2 * chan
+ 1]
2589 LLVMValueRef packed
;
2591 packed
= packf(&ctx
->ac
, pack_args
);
2592 args
->out
[chan
] = ac_to_float(&ctx
->ac
, packed
);
2594 args
->compr
= 1; /* COMPR flag */
2598 for (chan
= 0; chan
< 2; chan
++) {
2599 LLVMValueRef pack_args
[2] = {
2600 ac_to_integer(&ctx
->ac
, values
[2 * chan
]),
2601 ac_to_integer(&ctx
->ac
, values
[2 * chan
+ 1])
2603 LLVMValueRef packed
;
2605 packed
= packi(&ctx
->ac
, pack_args
,
2606 is_int8
? 8 : is_int10
? 10 : 16,
2608 args
->out
[chan
] = ac_to_float(&ctx
->ac
, packed
);
2610 args
->compr
= 1; /* COMPR flag */
2614 static void si_alpha_test(struct lp_build_tgsi_context
*bld_base
,
2617 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2619 if (ctx
->shader
->key
.part
.ps
.epilog
.alpha_func
!= PIPE_FUNC_NEVER
) {
2620 static LLVMRealPredicate cond_map
[PIPE_FUNC_ALWAYS
+ 1] = {
2621 [PIPE_FUNC_LESS
] = LLVMRealOLT
,
2622 [PIPE_FUNC_EQUAL
] = LLVMRealOEQ
,
2623 [PIPE_FUNC_LEQUAL
] = LLVMRealOLE
,
2624 [PIPE_FUNC_GREATER
] = LLVMRealOGT
,
2625 [PIPE_FUNC_NOTEQUAL
] = LLVMRealONE
,
2626 [PIPE_FUNC_GEQUAL
] = LLVMRealOGE
,
2628 LLVMRealPredicate cond
= cond_map
[ctx
->shader
->key
.part
.ps
.epilog
.alpha_func
];
2631 LLVMValueRef alpha_ref
= LLVMGetParam(ctx
->main_fn
,
2632 SI_PARAM_ALPHA_REF
);
2633 LLVMValueRef alpha_pass
=
2634 LLVMBuildFCmp(ctx
->ac
.builder
, cond
, alpha
, alpha_ref
, "");
2635 ac_build_kill_if_false(&ctx
->ac
, alpha_pass
);
2637 ac_build_kill_if_false(&ctx
->ac
, LLVMConstInt(ctx
->i1
, 0, 0));
2641 static LLVMValueRef
si_scale_alpha_by_sample_mask(struct lp_build_tgsi_context
*bld_base
,
2643 unsigned samplemask_param
)
2645 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2646 LLVMValueRef coverage
;
2648 /* alpha = alpha * popcount(coverage) / SI_NUM_SMOOTH_AA_SAMPLES */
2649 coverage
= LLVMGetParam(ctx
->main_fn
,
2651 coverage
= ac_to_integer(&ctx
->ac
, coverage
);
2653 coverage
= lp_build_intrinsic(ctx
->ac
.builder
, "llvm.ctpop.i32",
2655 &coverage
, 1, LP_FUNC_ATTR_READNONE
);
2657 coverage
= LLVMBuildUIToFP(ctx
->ac
.builder
, coverage
,
2660 coverage
= LLVMBuildFMul(ctx
->ac
.builder
, coverage
,
2661 LLVMConstReal(ctx
->f32
,
2662 1.0 / SI_NUM_SMOOTH_AA_SAMPLES
), "");
2664 return LLVMBuildFMul(ctx
->ac
.builder
, alpha
, coverage
, "");
2667 static void si_llvm_emit_clipvertex(struct si_shader_context
*ctx
,
2668 struct ac_export_args
*pos
, LLVMValueRef
*out_elts
)
2672 unsigned const_chan
;
2673 LLVMValueRef base_elt
;
2674 LLVMValueRef ptr
= LLVMGetParam(ctx
->main_fn
, ctx
->param_rw_buffers
);
2675 LLVMValueRef constbuf_index
= LLVMConstInt(ctx
->i32
,
2676 SI_VS_CONST_CLIP_PLANES
, 0);
2677 LLVMValueRef const_resource
= ac_build_load_to_sgpr(&ctx
->ac
, ptr
, constbuf_index
);
2679 for (reg_index
= 0; reg_index
< 2; reg_index
++) {
2680 struct ac_export_args
*args
= &pos
[2 + reg_index
];
2685 args
->out
[3] = LLVMConstReal(ctx
->f32
, 0.0f
);
2687 /* Compute dot products of position and user clip plane vectors */
2688 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
2689 for (const_chan
= 0; const_chan
< TGSI_NUM_CHANNELS
; const_chan
++) {
2691 LLVMConstInt(ctx
->i32
, ((reg_index
* 4 + chan
) * 4 +
2692 const_chan
) * 4, 0);
2693 base_elt
= buffer_load_const(ctx
, const_resource
,
2696 lp_build_add(&ctx
->bld_base
.base
, args
->out
[chan
],
2697 lp_build_mul(&ctx
->bld_base
.base
, base_elt
,
2698 out_elts
[const_chan
]));
2702 args
->enabled_channels
= 0xf;
2703 args
->valid_mask
= 0;
2705 args
->target
= V_008DFC_SQ_EXP_POS
+ 2 + reg_index
;
2710 static void si_dump_streamout(struct pipe_stream_output_info
*so
)
2714 if (so
->num_outputs
)
2715 fprintf(stderr
, "STREAMOUT\n");
2717 for (i
= 0; i
< so
->num_outputs
; i
++) {
2718 unsigned mask
= ((1 << so
->output
[i
].num_components
) - 1) <<
2719 so
->output
[i
].start_component
;
2720 fprintf(stderr
, " %i: BUF%i[%i..%i] <- OUT[%i].%s%s%s%s\n",
2721 i
, so
->output
[i
].output_buffer
,
2722 so
->output
[i
].dst_offset
, so
->output
[i
].dst_offset
+ so
->output
[i
].num_components
- 1,
2723 so
->output
[i
].register_index
,
2724 mask
& 1 ? "x" : "",
2725 mask
& 2 ? "y" : "",
2726 mask
& 4 ? "z" : "",
2727 mask
& 8 ? "w" : "");
2731 static void emit_streamout_output(struct si_shader_context
*ctx
,
2732 LLVMValueRef
const *so_buffers
,
2733 LLVMValueRef
const *so_write_offsets
,
2734 struct pipe_stream_output
*stream_out
,
2735 struct si_shader_output_values
*shader_out
)
2737 unsigned buf_idx
= stream_out
->output_buffer
;
2738 unsigned start
= stream_out
->start_component
;
2739 unsigned num_comps
= stream_out
->num_components
;
2740 LLVMValueRef out
[4];
2742 assert(num_comps
&& num_comps
<= 4);
2743 if (!num_comps
|| num_comps
> 4)
2746 /* Load the output as int. */
2747 for (int j
= 0; j
< num_comps
; j
++) {
2748 assert(stream_out
->stream
== shader_out
->vertex_stream
[start
+ j
]);
2750 out
[j
] = ac_to_integer(&ctx
->ac
, shader_out
->values
[start
+ j
]);
2753 /* Pack the output. */
2754 LLVMValueRef vdata
= NULL
;
2756 switch (num_comps
) {
2757 case 1: /* as i32 */
2760 case 2: /* as v2i32 */
2761 case 3: /* as v4i32 (aligned to 4) */
2762 case 4: /* as v4i32 */
2763 vdata
= LLVMGetUndef(LLVMVectorType(ctx
->i32
, util_next_power_of_two(num_comps
)));
2764 for (int j
= 0; j
< num_comps
; j
++) {
2765 vdata
= LLVMBuildInsertElement(ctx
->ac
.builder
, vdata
, out
[j
],
2766 LLVMConstInt(ctx
->i32
, j
, 0), "");
2771 ac_build_buffer_store_dword(&ctx
->ac
, so_buffers
[buf_idx
],
2773 so_write_offsets
[buf_idx
],
2775 stream_out
->dst_offset
* 4, 1, 1, true, false);
2779 * Write streamout data to buffers for vertex stream @p stream (different
2780 * vertex streams can occur for GS copy shaders).
2782 static void si_llvm_emit_streamout(struct si_shader_context
*ctx
,
2783 struct si_shader_output_values
*outputs
,
2784 unsigned noutput
, unsigned stream
)
2786 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
2787 struct pipe_stream_output_info
*so
= &sel
->so
;
2788 LLVMBuilderRef builder
= ctx
->ac
.builder
;
2790 struct lp_build_if_state if_ctx
;
2792 /* Get bits [22:16], i.e. (so_param >> 16) & 127; */
2793 LLVMValueRef so_vtx_count
=
2794 unpack_param(ctx
, ctx
->param_streamout_config
, 16, 7);
2796 LLVMValueRef tid
= ac_get_thread_id(&ctx
->ac
);
2798 /* can_emit = tid < so_vtx_count; */
2799 LLVMValueRef can_emit
=
2800 LLVMBuildICmp(builder
, LLVMIntULT
, tid
, so_vtx_count
, "");
2802 /* Emit the streamout code conditionally. This actually avoids
2803 * out-of-bounds buffer access. The hw tells us via the SGPR
2804 * (so_vtx_count) which threads are allowed to emit streamout data. */
2805 lp_build_if(&if_ctx
, &ctx
->gallivm
, can_emit
);
2807 /* The buffer offset is computed as follows:
2808 * ByteOffset = streamout_offset[buffer_id]*4 +
2809 * (streamout_write_index + thread_id)*stride[buffer_id] +
2813 LLVMValueRef so_write_index
=
2814 LLVMGetParam(ctx
->main_fn
,
2815 ctx
->param_streamout_write_index
);
2817 /* Compute (streamout_write_index + thread_id). */
2818 so_write_index
= LLVMBuildAdd(builder
, so_write_index
, tid
, "");
2820 /* Load the descriptor and compute the write offset for each
2821 * enabled buffer. */
2822 LLVMValueRef so_write_offset
[4] = {};
2823 LLVMValueRef so_buffers
[4];
2824 LLVMValueRef buf_ptr
= LLVMGetParam(ctx
->main_fn
,
2825 ctx
->param_rw_buffers
);
2827 for (i
= 0; i
< 4; i
++) {
2831 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
,
2832 SI_VS_STREAMOUT_BUF0
+ i
, 0);
2834 so_buffers
[i
] = ac_build_load_to_sgpr(&ctx
->ac
, buf_ptr
, offset
);
2836 LLVMValueRef so_offset
= LLVMGetParam(ctx
->main_fn
,
2837 ctx
->param_streamout_offset
[i
]);
2838 so_offset
= LLVMBuildMul(builder
, so_offset
, LLVMConstInt(ctx
->i32
, 4, 0), "");
2840 so_write_offset
[i
] = LLVMBuildMul(builder
, so_write_index
,
2841 LLVMConstInt(ctx
->i32
, so
->stride
[i
]*4, 0), "");
2842 so_write_offset
[i
] = LLVMBuildAdd(builder
, so_write_offset
[i
], so_offset
, "");
2845 /* Write streamout data. */
2846 for (i
= 0; i
< so
->num_outputs
; i
++) {
2847 unsigned reg
= so
->output
[i
].register_index
;
2852 if (stream
!= so
->output
[i
].stream
)
2855 emit_streamout_output(ctx
, so_buffers
, so_write_offset
,
2856 &so
->output
[i
], &outputs
[reg
]);
2859 lp_build_endif(&if_ctx
);
2862 static void si_export_param(struct si_shader_context
*ctx
, unsigned index
,
2863 LLVMValueRef
*values
)
2865 struct ac_export_args args
;
2867 si_llvm_init_export_args(ctx
, values
,
2868 V_008DFC_SQ_EXP_PARAM
+ index
, &args
);
2869 ac_build_export(&ctx
->ac
, &args
);
2872 static void si_build_param_exports(struct si_shader_context
*ctx
,
2873 struct si_shader_output_values
*outputs
,
2876 struct si_shader
*shader
= ctx
->shader
;
2877 unsigned param_count
= 0;
2879 for (unsigned i
= 0; i
< noutput
; i
++) {
2880 unsigned semantic_name
= outputs
[i
].semantic_name
;
2881 unsigned semantic_index
= outputs
[i
].semantic_index
;
2883 if (outputs
[i
].vertex_stream
[0] != 0 &&
2884 outputs
[i
].vertex_stream
[1] != 0 &&
2885 outputs
[i
].vertex_stream
[2] != 0 &&
2886 outputs
[i
].vertex_stream
[3] != 0)
2889 switch (semantic_name
) {
2890 case TGSI_SEMANTIC_LAYER
:
2891 case TGSI_SEMANTIC_VIEWPORT_INDEX
:
2892 case TGSI_SEMANTIC_CLIPDIST
:
2893 case TGSI_SEMANTIC_COLOR
:
2894 case TGSI_SEMANTIC_BCOLOR
:
2895 case TGSI_SEMANTIC_PRIMID
:
2896 case TGSI_SEMANTIC_FOG
:
2897 case TGSI_SEMANTIC_TEXCOORD
:
2898 case TGSI_SEMANTIC_GENERIC
:
2904 if ((semantic_name
!= TGSI_SEMANTIC_GENERIC
||
2905 semantic_index
< SI_MAX_IO_GENERIC
) &&
2906 shader
->key
.opt
.kill_outputs
&
2907 (1ull << si_shader_io_get_unique_index(semantic_name
, semantic_index
)))
2910 si_export_param(ctx
, param_count
, outputs
[i
].values
);
2912 assert(i
< ARRAY_SIZE(shader
->info
.vs_output_param_offset
));
2913 shader
->info
.vs_output_param_offset
[i
] = param_count
++;
2916 shader
->info
.nr_param_exports
= param_count
;
2919 /* Generate export instructions for hardware VS shader stage */
2920 static void si_llvm_export_vs(struct si_shader_context
*ctx
,
2921 struct si_shader_output_values
*outputs
,
2924 struct si_shader
*shader
= ctx
->shader
;
2925 struct ac_export_args pos_args
[4] = {};
2926 LLVMValueRef psize_value
= NULL
, edgeflag_value
= NULL
, layer_value
= NULL
, viewport_index_value
= NULL
;
2930 /* Build position exports. */
2931 for (i
= 0; i
< noutput
; i
++) {
2932 switch (outputs
[i
].semantic_name
) {
2933 case TGSI_SEMANTIC_POSITION
:
2934 si_llvm_init_export_args(ctx
, outputs
[i
].values
,
2935 V_008DFC_SQ_EXP_POS
, &pos_args
[0]);
2937 case TGSI_SEMANTIC_PSIZE
:
2938 psize_value
= outputs
[i
].values
[0];
2940 case TGSI_SEMANTIC_LAYER
:
2941 layer_value
= outputs
[i
].values
[0];
2943 case TGSI_SEMANTIC_VIEWPORT_INDEX
:
2944 viewport_index_value
= outputs
[i
].values
[0];
2946 case TGSI_SEMANTIC_EDGEFLAG
:
2947 edgeflag_value
= outputs
[i
].values
[0];
2949 case TGSI_SEMANTIC_CLIPDIST
:
2950 if (!shader
->key
.opt
.clip_disable
) {
2951 unsigned index
= 2 + outputs
[i
].semantic_index
;
2952 si_llvm_init_export_args(ctx
, outputs
[i
].values
,
2953 V_008DFC_SQ_EXP_POS
+ index
,
2957 case TGSI_SEMANTIC_CLIPVERTEX
:
2958 if (!shader
->key
.opt
.clip_disable
) {
2959 si_llvm_emit_clipvertex(ctx
, pos_args
,
2966 /* We need to add the position output manually if it's missing. */
2967 if (!pos_args
[0].out
[0]) {
2968 pos_args
[0].enabled_channels
= 0xf; /* writemask */
2969 pos_args
[0].valid_mask
= 0; /* EXEC mask */
2970 pos_args
[0].done
= 0; /* last export? */
2971 pos_args
[0].target
= V_008DFC_SQ_EXP_POS
;
2972 pos_args
[0].compr
= 0; /* COMPR flag */
2973 pos_args
[0].out
[0] = ctx
->ac
.f32_0
; /* X */
2974 pos_args
[0].out
[1] = ctx
->ac
.f32_0
; /* Y */
2975 pos_args
[0].out
[2] = ctx
->ac
.f32_0
; /* Z */
2976 pos_args
[0].out
[3] = ctx
->ac
.f32_1
; /* W */
2979 /* Write the misc vector (point size, edgeflag, layer, viewport). */
2980 if (shader
->selector
->info
.writes_psize
||
2981 shader
->selector
->info
.writes_edgeflag
||
2982 shader
->selector
->info
.writes_viewport_index
||
2983 shader
->selector
->info
.writes_layer
) {
2984 pos_args
[1].enabled_channels
= shader
->selector
->info
.writes_psize
|
2985 (shader
->selector
->info
.writes_edgeflag
<< 1) |
2986 (shader
->selector
->info
.writes_layer
<< 2);
2988 pos_args
[1].valid_mask
= 0; /* EXEC mask */
2989 pos_args
[1].done
= 0; /* last export? */
2990 pos_args
[1].target
= V_008DFC_SQ_EXP_POS
+ 1;
2991 pos_args
[1].compr
= 0; /* COMPR flag */
2992 pos_args
[1].out
[0] = ctx
->ac
.f32_0
; /* X */
2993 pos_args
[1].out
[1] = ctx
->ac
.f32_0
; /* Y */
2994 pos_args
[1].out
[2] = ctx
->ac
.f32_0
; /* Z */
2995 pos_args
[1].out
[3] = ctx
->ac
.f32_0
; /* W */
2997 if (shader
->selector
->info
.writes_psize
)
2998 pos_args
[1].out
[0] = psize_value
;
3000 if (shader
->selector
->info
.writes_edgeflag
) {
3001 /* The output is a float, but the hw expects an integer
3002 * with the first bit containing the edge flag. */
3003 edgeflag_value
= LLVMBuildFPToUI(ctx
->ac
.builder
,
3006 edgeflag_value
= ac_build_umin(&ctx
->ac
,
3010 /* The LLVM intrinsic expects a float. */
3011 pos_args
[1].out
[1] = ac_to_float(&ctx
->ac
, edgeflag_value
);
3014 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
3015 /* GFX9 has the layer in out.z[10:0] and the viewport
3016 * index in out.z[19:16].
3018 if (shader
->selector
->info
.writes_layer
)
3019 pos_args
[1].out
[2] = layer_value
;
3021 if (shader
->selector
->info
.writes_viewport_index
) {
3022 LLVMValueRef v
= viewport_index_value
;
3024 v
= ac_to_integer(&ctx
->ac
, v
);
3025 v
= LLVMBuildShl(ctx
->ac
.builder
, v
,
3026 LLVMConstInt(ctx
->i32
, 16, 0), "");
3027 v
= LLVMBuildOr(ctx
->ac
.builder
, v
,
3028 ac_to_integer(&ctx
->ac
, pos_args
[1].out
[2]), "");
3029 pos_args
[1].out
[2] = ac_to_float(&ctx
->ac
, v
);
3030 pos_args
[1].enabled_channels
|= 1 << 2;
3033 if (shader
->selector
->info
.writes_layer
)
3034 pos_args
[1].out
[2] = layer_value
;
3036 if (shader
->selector
->info
.writes_viewport_index
) {
3037 pos_args
[1].out
[3] = viewport_index_value
;
3038 pos_args
[1].enabled_channels
|= 1 << 3;
3043 for (i
= 0; i
< 4; i
++)
3044 if (pos_args
[i
].out
[0])
3045 shader
->info
.nr_pos_exports
++;
3048 for (i
= 0; i
< 4; i
++) {
3049 if (!pos_args
[i
].out
[0])
3052 /* Specify the target we are exporting */
3053 pos_args
[i
].target
= V_008DFC_SQ_EXP_POS
+ pos_idx
++;
3055 if (pos_idx
== shader
->info
.nr_pos_exports
)
3056 /* Specify that this is the last export */
3057 pos_args
[i
].done
= 1;
3059 ac_build_export(&ctx
->ac
, &pos_args
[i
]);
3062 /* Build parameter exports. */
3063 si_build_param_exports(ctx
, outputs
, noutput
);
3067 * Forward all outputs from the vertex shader to the TES. This is only used
3068 * for the fixed function TCS.
3070 static void si_copy_tcs_inputs(struct lp_build_tgsi_context
*bld_base
)
3072 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3073 LLVMValueRef invocation_id
, buffer
, buffer_offset
;
3074 LLVMValueRef lds_vertex_stride
, lds_vertex_offset
, lds_base
;
3077 invocation_id
= unpack_llvm_param(ctx
, ctx
->abi
.tcs_rel_ids
, 8, 5);
3078 buffer
= get_tess_ring_descriptor(ctx
, TESS_OFFCHIP_RING_TCS
);
3079 buffer_offset
= LLVMGetParam(ctx
->main_fn
, ctx
->param_tcs_offchip_offset
);
3081 lds_vertex_stride
= get_tcs_in_vertex_dw_stride(ctx
);
3082 lds_vertex_offset
= LLVMBuildMul(ctx
->ac
.builder
, invocation_id
,
3083 lds_vertex_stride
, "");
3084 lds_base
= get_tcs_in_current_patch_offset(ctx
);
3085 lds_base
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
, lds_vertex_offset
, "");
3087 inputs
= ctx
->shader
->key
.mono
.u
.ff_tcs_inputs_to_copy
;
3089 unsigned i
= u_bit_scan64(&inputs
);
3091 LLVMValueRef lds_ptr
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
3092 LLVMConstInt(ctx
->i32
, 4 * i
, 0),
3095 LLVMValueRef buffer_addr
= get_tcs_tes_buffer_address(ctx
,
3096 get_rel_patch_id(ctx
),
3098 LLVMConstInt(ctx
->i32
, i
, 0));
3100 LLVMValueRef value
= lds_load(bld_base
, ctx
->ac
.i32
, ~0,
3103 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, value
, 4, buffer_addr
,
3104 buffer_offset
, 0, 1, 0, true, false);
3108 static void si_write_tess_factors(struct lp_build_tgsi_context
*bld_base
,
3109 LLVMValueRef rel_patch_id
,
3110 LLVMValueRef invocation_id
,
3111 LLVMValueRef tcs_out_current_patch_data_offset
,
3112 LLVMValueRef invoc0_tf_outer
[4],
3113 LLVMValueRef invoc0_tf_inner
[2])
3115 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3116 struct si_shader
*shader
= ctx
->shader
;
3117 unsigned tess_inner_index
, tess_outer_index
;
3118 LLVMValueRef lds_base
, lds_inner
, lds_outer
, byteoffset
, buffer
;
3119 LLVMValueRef out
[6], vec0
, vec1
, tf_base
, inner
[4], outer
[4];
3120 unsigned stride
, outer_comps
, inner_comps
, i
, offset
;
3121 struct lp_build_if_state if_ctx
, inner_if_ctx
;
3123 /* Add a barrier before loading tess factors from LDS. */
3124 if (!shader
->key
.part
.tcs
.epilog
.invoc0_tess_factors_are_def
)
3125 si_llvm_emit_barrier(NULL
, bld_base
, NULL
);
3127 /* Do this only for invocation 0, because the tess levels are per-patch,
3130 * This can't jump, because invocation 0 executes this. It should
3131 * at least mask out the loads and stores for other invocations.
3133 lp_build_if(&if_ctx
, &ctx
->gallivm
,
3134 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
,
3135 invocation_id
, ctx
->i32_0
, ""));
3137 /* Determine the layout of one tess factor element in the buffer. */
3138 switch (shader
->key
.part
.tcs
.epilog
.prim_mode
) {
3139 case PIPE_PRIM_LINES
:
3140 stride
= 2; /* 2 dwords, 1 vec2 store */
3144 case PIPE_PRIM_TRIANGLES
:
3145 stride
= 4; /* 4 dwords, 1 vec4 store */
3149 case PIPE_PRIM_QUADS
:
3150 stride
= 6; /* 6 dwords, 2 stores (vec4 + vec2) */
3159 for (i
= 0; i
< 4; i
++) {
3160 inner
[i
] = LLVMGetUndef(ctx
->i32
);
3161 outer
[i
] = LLVMGetUndef(ctx
->i32
);
3164 if (shader
->key
.part
.tcs
.epilog
.invoc0_tess_factors_are_def
) {
3165 /* Tess factors are in VGPRs. */
3166 for (i
= 0; i
< outer_comps
; i
++)
3167 outer
[i
] = out
[i
] = invoc0_tf_outer
[i
];
3168 for (i
= 0; i
< inner_comps
; i
++)
3169 inner
[i
] = out
[outer_comps
+i
] = invoc0_tf_inner
[i
];
3171 /* Load tess_inner and tess_outer from LDS.
3172 * Any invocation can write them, so we can't get them from a temporary.
3174 tess_inner_index
= si_shader_io_get_unique_index_patch(TGSI_SEMANTIC_TESSINNER
, 0);
3175 tess_outer_index
= si_shader_io_get_unique_index_patch(TGSI_SEMANTIC_TESSOUTER
, 0);
3177 lds_base
= tcs_out_current_patch_data_offset
;
3178 lds_inner
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
3179 LLVMConstInt(ctx
->i32
,
3180 tess_inner_index
* 4, 0), "");
3181 lds_outer
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
3182 LLVMConstInt(ctx
->i32
,
3183 tess_outer_index
* 4, 0), "");
3185 for (i
= 0; i
< outer_comps
; i
++) {
3187 lds_load(bld_base
, ctx
->ac
.i32
, i
, lds_outer
);
3189 for (i
= 0; i
< inner_comps
; i
++) {
3190 inner
[i
] = out
[outer_comps
+i
] =
3191 lds_load(bld_base
, ctx
->ac
.i32
, i
, lds_inner
);
3195 if (shader
->key
.part
.tcs
.epilog
.prim_mode
== PIPE_PRIM_LINES
) {
3196 /* For isolines, the hardware expects tess factors in the
3197 * reverse order from what GLSL / TGSI specify.
3199 LLVMValueRef tmp
= out
[0];
3204 /* Convert the outputs to vectors for stores. */
3205 vec0
= lp_build_gather_values(&ctx
->gallivm
, out
, MIN2(stride
, 4));
3209 vec1
= lp_build_gather_values(&ctx
->gallivm
, out
+4, stride
- 4);
3211 /* Get the buffer. */
3212 buffer
= get_tess_ring_descriptor(ctx
, TCS_FACTOR_RING
);
3214 /* Get the offset. */
3215 tf_base
= LLVMGetParam(ctx
->main_fn
,
3216 ctx
->param_tcs_factor_offset
);
3217 byteoffset
= LLVMBuildMul(ctx
->ac
.builder
, rel_patch_id
,
3218 LLVMConstInt(ctx
->i32
, 4 * stride
, 0), "");
3220 lp_build_if(&inner_if_ctx
, &ctx
->gallivm
,
3221 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
,
3222 rel_patch_id
, ctx
->i32_0
, ""));
3224 /* Store the dynamic HS control word. */
3226 if (ctx
->screen
->info
.chip_class
<= VI
) {
3227 ac_build_buffer_store_dword(&ctx
->ac
, buffer
,
3228 LLVMConstInt(ctx
->i32
, 0x80000000, 0),
3229 1, ctx
->i32_0
, tf_base
,
3230 offset
, 1, 0, true, false);
3234 lp_build_endif(&inner_if_ctx
);
3236 /* Store the tessellation factors. */
3237 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec0
,
3238 MIN2(stride
, 4), byteoffset
, tf_base
,
3239 offset
, 1, 0, true, false);
3242 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec1
,
3243 stride
- 4, byteoffset
, tf_base
,
3244 offset
, 1, 0, true, false);
3246 /* Store the tess factors into the offchip buffer if TES reads them. */
3247 if (shader
->key
.part
.tcs
.epilog
.tes_reads_tess_factors
) {
3248 LLVMValueRef buf
, base
, inner_vec
, outer_vec
, tf_outer_offset
;
3249 LLVMValueRef tf_inner_offset
;
3250 unsigned param_outer
, param_inner
;
3252 buf
= get_tess_ring_descriptor(ctx
, TESS_OFFCHIP_RING_TCS
);
3253 base
= LLVMGetParam(ctx
->main_fn
, ctx
->param_tcs_offchip_offset
);
3255 param_outer
= si_shader_io_get_unique_index_patch(
3256 TGSI_SEMANTIC_TESSOUTER
, 0);
3257 tf_outer_offset
= get_tcs_tes_buffer_address(ctx
, rel_patch_id
, NULL
,
3258 LLVMConstInt(ctx
->i32
, param_outer
, 0));
3260 outer_vec
= lp_build_gather_values(&ctx
->gallivm
, outer
,
3261 util_next_power_of_two(outer_comps
));
3263 ac_build_buffer_store_dword(&ctx
->ac
, buf
, outer_vec
,
3264 outer_comps
, tf_outer_offset
,
3265 base
, 0, 1, 0, true, false);
3267 param_inner
= si_shader_io_get_unique_index_patch(
3268 TGSI_SEMANTIC_TESSINNER
, 0);
3269 tf_inner_offset
= get_tcs_tes_buffer_address(ctx
, rel_patch_id
, NULL
,
3270 LLVMConstInt(ctx
->i32
, param_inner
, 0));
3272 inner_vec
= inner_comps
== 1 ? inner
[0] :
3273 lp_build_gather_values(&ctx
->gallivm
, inner
, inner_comps
);
3274 ac_build_buffer_store_dword(&ctx
->ac
, buf
, inner_vec
,
3275 inner_comps
, tf_inner_offset
,
3276 base
, 0, 1, 0, true, false);
3280 lp_build_endif(&if_ctx
);
3284 si_insert_input_ret(struct si_shader_context
*ctx
, LLVMValueRef ret
,
3285 unsigned param
, unsigned return_index
)
3287 return LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
3288 LLVMGetParam(ctx
->main_fn
, param
),
3293 si_insert_input_ret_float(struct si_shader_context
*ctx
, LLVMValueRef ret
,
3294 unsigned param
, unsigned return_index
)
3296 LLVMBuilderRef builder
= ctx
->ac
.builder
;
3297 LLVMValueRef p
= LLVMGetParam(ctx
->main_fn
, param
);
3299 return LLVMBuildInsertValue(builder
, ret
,
3300 ac_to_float(&ctx
->ac
, p
),
3305 si_insert_input_ptr(struct si_shader_context
*ctx
, LLVMValueRef ret
,
3306 unsigned param
, unsigned return_index
)
3308 LLVMBuilderRef builder
= ctx
->ac
.builder
;
3309 LLVMValueRef ptr
, lo
, hi
;
3311 if (HAVE_32BIT_POINTERS
) {
3312 ptr
= LLVMGetParam(ctx
->main_fn
, param
);
3313 ptr
= LLVMBuildPtrToInt(builder
, ptr
, ctx
->i32
, "");
3314 return LLVMBuildInsertValue(builder
, ret
, ptr
, return_index
, "");
3317 ptr
= LLVMGetParam(ctx
->main_fn
, param
);
3318 ptr
= LLVMBuildPtrToInt(builder
, ptr
, ctx
->i64
, "");
3319 ptr
= LLVMBuildBitCast(builder
, ptr
, ctx
->v2i32
, "");
3320 lo
= LLVMBuildExtractElement(builder
, ptr
, ctx
->i32_0
, "");
3321 hi
= LLVMBuildExtractElement(builder
, ptr
, ctx
->i32_1
, "");
3322 ret
= LLVMBuildInsertValue(builder
, ret
, lo
, return_index
, "");
3323 return LLVMBuildInsertValue(builder
, ret
, hi
, return_index
+ 1, "");
3326 /* This only writes the tessellation factor levels. */
3327 static void si_llvm_emit_tcs_epilogue(struct ac_shader_abi
*abi
,
3328 unsigned max_outputs
,
3329 LLVMValueRef
*addrs
)
3331 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
3332 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
3333 LLVMBuilderRef builder
= ctx
->ac
.builder
;
3334 LLVMValueRef rel_patch_id
, invocation_id
, tf_lds_offset
;
3336 si_copy_tcs_inputs(bld_base
);
3338 rel_patch_id
= get_rel_patch_id(ctx
);
3339 invocation_id
= unpack_llvm_param(ctx
, ctx
->abi
.tcs_rel_ids
, 8, 5);
3340 tf_lds_offset
= get_tcs_out_current_patch_data_offset(ctx
);
3342 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
3343 LLVMBasicBlockRef blocks
[2] = {
3344 LLVMGetInsertBlock(builder
),
3345 ctx
->merged_wrap_if_state
.entry_block
3347 LLVMValueRef values
[2];
3349 lp_build_endif(&ctx
->merged_wrap_if_state
);
3351 values
[0] = rel_patch_id
;
3352 values
[1] = LLVMGetUndef(ctx
->i32
);
3353 rel_patch_id
= ac_build_phi(&ctx
->ac
, ctx
->i32
, 2, values
, blocks
);
3355 values
[0] = tf_lds_offset
;
3356 values
[1] = LLVMGetUndef(ctx
->i32
);
3357 tf_lds_offset
= ac_build_phi(&ctx
->ac
, ctx
->i32
, 2, values
, blocks
);
3359 values
[0] = invocation_id
;
3360 values
[1] = ctx
->i32_1
; /* cause the epilog to skip threads */
3361 invocation_id
= ac_build_phi(&ctx
->ac
, ctx
->i32
, 2, values
, blocks
);
3364 /* Return epilog parameters from this function. */
3365 LLVMValueRef ret
= ctx
->return_value
;
3368 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
3369 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_layout
,
3370 8 + GFX9_SGPR_TCS_OFFCHIP_LAYOUT
);
3371 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_out_lds_layout
,
3372 8 + GFX9_SGPR_TCS_OUT_LAYOUT
);
3373 /* Tess offchip and tess factor offsets are at the beginning. */
3374 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_offset
, 2);
3375 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_factor_offset
, 4);
3376 vgpr
= 8 + GFX9_SGPR_TCS_OUT_LAYOUT
+ 1;
3378 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_layout
,
3379 GFX6_SGPR_TCS_OFFCHIP_LAYOUT
);
3380 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_out_lds_layout
,
3381 GFX6_SGPR_TCS_OUT_LAYOUT
);
3382 /* Tess offchip and tess factor offsets are after user SGPRs. */
3383 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_offset
,
3384 GFX6_TCS_NUM_USER_SGPR
);
3385 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_factor_offset
,
3386 GFX6_TCS_NUM_USER_SGPR
+ 1);
3387 vgpr
= GFX6_TCS_NUM_USER_SGPR
+ 2;
3391 rel_patch_id
= ac_to_float(&ctx
->ac
, rel_patch_id
);
3392 invocation_id
= ac_to_float(&ctx
->ac
, invocation_id
);
3393 tf_lds_offset
= ac_to_float(&ctx
->ac
, tf_lds_offset
);
3395 /* Leave a hole corresponding to the two input VGPRs. This ensures that
3396 * the invocation_id output does not alias the tcs_rel_ids input,
3397 * which saves a V_MOV on gfx9.
3401 ret
= LLVMBuildInsertValue(builder
, ret
, rel_patch_id
, vgpr
++, "");
3402 ret
= LLVMBuildInsertValue(builder
, ret
, invocation_id
, vgpr
++, "");
3404 if (ctx
->shader
->selector
->tcs_info
.tessfactors_are_def_in_all_invocs
) {
3405 vgpr
++; /* skip the tess factor LDS offset */
3406 for (unsigned i
= 0; i
< 6; i
++) {
3407 LLVMValueRef value
=
3408 LLVMBuildLoad(builder
, ctx
->invoc0_tess_factors
[i
], "");
3409 value
= ac_to_float(&ctx
->ac
, value
);
3410 ret
= LLVMBuildInsertValue(builder
, ret
, value
, vgpr
++, "");
3413 ret
= LLVMBuildInsertValue(builder
, ret
, tf_lds_offset
, vgpr
++, "");
3415 ctx
->return_value
= ret
;
3418 /* Pass TCS inputs from LS to TCS on GFX9. */
3419 static void si_set_ls_return_value_for_tcs(struct si_shader_context
*ctx
)
3421 LLVMValueRef ret
= ctx
->return_value
;
3423 ret
= si_insert_input_ptr(ctx
, ret
, 0, 0);
3424 if (HAVE_32BIT_POINTERS
)
3425 ret
= si_insert_input_ptr(ctx
, ret
, 1, 1);
3426 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_offset
, 2);
3427 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_merged_wave_info
, 3);
3428 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_factor_offset
, 4);
3429 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_merged_scratch_offset
, 5);
3431 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->param_rw_buffers
,
3432 8 + SI_SGPR_RW_BUFFERS
);
3433 ret
= si_insert_input_ptr(ctx
, ret
,
3434 ctx
->param_bindless_samplers_and_images
,
3435 8 + SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES
);
3437 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_vs_state_bits
,
3438 8 + SI_SGPR_VS_STATE_BITS
);
3440 #if !HAVE_32BIT_POINTERS
3441 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->param_vs_state_bits
+ 1,
3442 8 + GFX9_SGPR_2ND_SAMPLERS_AND_IMAGES
);
3445 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_layout
,
3446 8 + GFX9_SGPR_TCS_OFFCHIP_LAYOUT
);
3447 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_out_lds_offsets
,
3448 8 + GFX9_SGPR_TCS_OUT_OFFSETS
);
3449 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_out_lds_layout
,
3450 8 + GFX9_SGPR_TCS_OUT_LAYOUT
);
3452 unsigned vgpr
= 8 + GFX9_TCS_NUM_USER_SGPR
;
3453 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
3454 ac_to_float(&ctx
->ac
, ctx
->abi
.tcs_patch_id
),
3456 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
3457 ac_to_float(&ctx
->ac
, ctx
->abi
.tcs_rel_ids
),
3459 ctx
->return_value
= ret
;
3462 /* Pass GS inputs from ES to GS on GFX9. */
3463 static void si_set_es_return_value_for_gs(struct si_shader_context
*ctx
)
3465 LLVMValueRef ret
= ctx
->return_value
;
3467 ret
= si_insert_input_ptr(ctx
, ret
, 0, 0);
3468 if (HAVE_32BIT_POINTERS
)
3469 ret
= si_insert_input_ptr(ctx
, ret
, 1, 1);
3470 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_gs2vs_offset
, 2);
3471 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_merged_wave_info
, 3);
3472 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_merged_scratch_offset
, 5);
3474 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->param_rw_buffers
,
3475 8 + SI_SGPR_RW_BUFFERS
);
3476 ret
= si_insert_input_ptr(ctx
, ret
,
3477 ctx
->param_bindless_samplers_and_images
,
3478 8 + SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES
);
3480 #if !HAVE_32BIT_POINTERS
3481 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->param_vs_state_bits
+ 1,
3482 8 + GFX9_SGPR_2ND_SAMPLERS_AND_IMAGES
);
3486 if (ctx
->type
== PIPE_SHADER_VERTEX
)
3487 vgpr
= 8 + GFX9_VSGS_NUM_USER_SGPR
;
3489 vgpr
= 8 + GFX9_TESGS_NUM_USER_SGPR
;
3491 for (unsigned i
= 0; i
< 5; i
++) {
3492 unsigned param
= ctx
->param_gs_vtx01_offset
+ i
;
3493 ret
= si_insert_input_ret_float(ctx
, ret
, param
, vgpr
++);
3495 ctx
->return_value
= ret
;
3498 static void si_llvm_emit_ls_epilogue(struct ac_shader_abi
*abi
,
3499 unsigned max_outputs
,
3500 LLVMValueRef
*addrs
)
3502 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
3503 struct si_shader
*shader
= ctx
->shader
;
3504 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
3506 LLVMValueRef vertex_id
= LLVMGetParam(ctx
->main_fn
,
3507 ctx
->param_rel_auto_id
);
3508 LLVMValueRef vertex_dw_stride
= get_tcs_in_vertex_dw_stride(ctx
);
3509 LLVMValueRef base_dw_addr
= LLVMBuildMul(ctx
->ac
.builder
, vertex_id
,
3510 vertex_dw_stride
, "");
3512 /* Write outputs to LDS. The next shader (TCS aka HS) will read
3513 * its inputs from it. */
3514 for (i
= 0; i
< info
->num_outputs
; i
++) {
3515 unsigned name
= info
->output_semantic_name
[i
];
3516 unsigned index
= info
->output_semantic_index
[i
];
3518 /* The ARB_shader_viewport_layer_array spec contains the
3521 * 2) What happens if gl_ViewportIndex or gl_Layer is
3522 * written in the vertex shader and a geometry shader is
3525 * RESOLVED: The value written by the last vertex processing
3526 * stage is used. If the last vertex processing stage
3527 * (vertex, tessellation evaluation or geometry) does not
3528 * statically assign to gl_ViewportIndex or gl_Layer, index
3529 * or layer zero is assumed.
3531 * So writes to those outputs in VS-as-LS are simply ignored.
3533 if (name
== TGSI_SEMANTIC_LAYER
||
3534 name
== TGSI_SEMANTIC_VIEWPORT_INDEX
)
3537 int param
= si_shader_io_get_unique_index(name
, index
);
3538 LLVMValueRef dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_dw_addr
,
3539 LLVMConstInt(ctx
->i32
, param
* 4, 0), "");
3541 for (chan
= 0; chan
< 4; chan
++) {
3542 if (!(info
->output_usagemask
[i
] & (1 << chan
)))
3545 lds_store(ctx
, chan
, dw_addr
,
3546 LLVMBuildLoad(ctx
->ac
.builder
, addrs
[4 * i
+ chan
], ""));
3550 if (ctx
->screen
->info
.chip_class
>= GFX9
)
3551 si_set_ls_return_value_for_tcs(ctx
);
3554 static void si_llvm_emit_es_epilogue(struct ac_shader_abi
*abi
,
3555 unsigned max_outputs
,
3556 LLVMValueRef
*addrs
)
3558 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
3559 struct si_shader
*es
= ctx
->shader
;
3560 struct tgsi_shader_info
*info
= &es
->selector
->info
;
3561 LLVMValueRef soffset
= LLVMGetParam(ctx
->main_fn
,
3562 ctx
->param_es2gs_offset
);
3563 LLVMValueRef lds_base
= NULL
;
3567 if (ctx
->screen
->info
.chip_class
>= GFX9
&& info
->num_outputs
) {
3568 unsigned itemsize_dw
= es
->selector
->esgs_itemsize
/ 4;
3569 LLVMValueRef vertex_idx
= ac_get_thread_id(&ctx
->ac
);
3570 LLVMValueRef wave_idx
= unpack_param(ctx
, ctx
->param_merged_wave_info
, 24, 4);
3571 vertex_idx
= LLVMBuildOr(ctx
->ac
.builder
, vertex_idx
,
3572 LLVMBuildMul(ctx
->ac
.builder
, wave_idx
,
3573 LLVMConstInt(ctx
->i32
, 64, false), ""), "");
3574 lds_base
= LLVMBuildMul(ctx
->ac
.builder
, vertex_idx
,
3575 LLVMConstInt(ctx
->i32
, itemsize_dw
, 0), "");
3578 for (i
= 0; i
< info
->num_outputs
; i
++) {
3581 if (info
->output_semantic_name
[i
] == TGSI_SEMANTIC_VIEWPORT_INDEX
||
3582 info
->output_semantic_name
[i
] == TGSI_SEMANTIC_LAYER
)
3585 param
= si_shader_io_get_unique_index(info
->output_semantic_name
[i
],
3586 info
->output_semantic_index
[i
]);
3588 for (chan
= 0; chan
< 4; chan
++) {
3589 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->ac
.builder
, addrs
[4 * i
+ chan
], "");
3590 out_val
= ac_to_integer(&ctx
->ac
, out_val
);
3592 /* GFX9 has the ESGS ring in LDS. */
3593 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
3594 lds_store(ctx
, param
* 4 + chan
, lds_base
, out_val
);
3598 ac_build_buffer_store_dword(&ctx
->ac
,
3600 out_val
, 1, NULL
, soffset
,
3601 (4 * param
+ chan
) * 4,
3606 if (ctx
->screen
->info
.chip_class
>= GFX9
)
3607 si_set_es_return_value_for_gs(ctx
);
3610 static LLVMValueRef
si_get_gs_wave_id(struct si_shader_context
*ctx
)
3612 if (ctx
->screen
->info
.chip_class
>= GFX9
)
3613 return unpack_param(ctx
, ctx
->param_merged_wave_info
, 16, 8);
3615 return LLVMGetParam(ctx
->main_fn
, ctx
->param_gs_wave_id
);
3618 static void emit_gs_epilogue(struct si_shader_context
*ctx
)
3620 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_NOP
| AC_SENDMSG_GS_DONE
,
3621 si_get_gs_wave_id(ctx
));
3623 if (ctx
->screen
->info
.chip_class
>= GFX9
)
3624 lp_build_endif(&ctx
->merged_wrap_if_state
);
3627 static void si_llvm_emit_gs_epilogue(struct ac_shader_abi
*abi
,
3628 unsigned max_outputs
,
3629 LLVMValueRef
*addrs
)
3631 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
3632 struct tgsi_shader_info UNUSED
*info
= &ctx
->shader
->selector
->info
;
3634 assert(info
->num_outputs
<= max_outputs
);
3636 emit_gs_epilogue(ctx
);
3639 static void si_tgsi_emit_gs_epilogue(struct lp_build_tgsi_context
*bld_base
)
3641 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3642 emit_gs_epilogue(ctx
);
3645 static void si_llvm_emit_vs_epilogue(struct ac_shader_abi
*abi
,
3646 unsigned max_outputs
,
3647 LLVMValueRef
*addrs
)
3649 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
3650 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
3651 struct si_shader_output_values
*outputs
= NULL
;
3654 assert(!ctx
->shader
->is_gs_copy_shader
);
3655 assert(info
->num_outputs
<= max_outputs
);
3657 outputs
= MALLOC((info
->num_outputs
+ 1) * sizeof(outputs
[0]));
3659 /* Vertex color clamping.
3661 * This uses a state constant loaded in a user data SGPR and
3662 * an IF statement is added that clamps all colors if the constant
3665 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
3666 struct lp_build_if_state if_ctx
;
3667 LLVMValueRef cond
= NULL
;
3668 LLVMValueRef addr
, val
;
3670 for (i
= 0; i
< info
->num_outputs
; i
++) {
3671 if (info
->output_semantic_name
[i
] != TGSI_SEMANTIC_COLOR
&&
3672 info
->output_semantic_name
[i
] != TGSI_SEMANTIC_BCOLOR
)
3675 /* We've found a color. */
3677 /* The state is in the first bit of the user SGPR. */
3678 cond
= LLVMGetParam(ctx
->main_fn
,
3679 ctx
->param_vs_state_bits
);
3680 cond
= LLVMBuildTrunc(ctx
->ac
.builder
, cond
,
3682 lp_build_if(&if_ctx
, &ctx
->gallivm
, cond
);
3685 for (j
= 0; j
< 4; j
++) {
3686 addr
= addrs
[4 * i
+ j
];
3687 val
= LLVMBuildLoad(ctx
->ac
.builder
, addr
, "");
3688 val
= ac_build_clamp(&ctx
->ac
, val
);
3689 LLVMBuildStore(ctx
->ac
.builder
, val
, addr
);
3694 lp_build_endif(&if_ctx
);
3697 for (i
= 0; i
< info
->num_outputs
; i
++) {
3698 outputs
[i
].semantic_name
= info
->output_semantic_name
[i
];
3699 outputs
[i
].semantic_index
= info
->output_semantic_index
[i
];
3701 for (j
= 0; j
< 4; j
++) {
3702 outputs
[i
].values
[j
] =
3703 LLVMBuildLoad(ctx
->ac
.builder
,
3706 outputs
[i
].vertex_stream
[j
] =
3707 (info
->output_streams
[i
] >> (2 * j
)) & 3;
3711 if (ctx
->shader
->selector
->so
.num_outputs
)
3712 si_llvm_emit_streamout(ctx
, outputs
, i
, 0);
3714 /* Export PrimitiveID. */
3715 if (ctx
->shader
->key
.mono
.u
.vs_export_prim_id
) {
3716 outputs
[i
].semantic_name
= TGSI_SEMANTIC_PRIMID
;
3717 outputs
[i
].semantic_index
= 0;
3718 outputs
[i
].values
[0] = ac_to_float(&ctx
->ac
, get_primitive_id(ctx
, 0));
3719 for (j
= 1; j
< 4; j
++)
3720 outputs
[i
].values
[j
] = LLVMConstReal(ctx
->f32
, 0);
3722 memset(outputs
[i
].vertex_stream
, 0,
3723 sizeof(outputs
[i
].vertex_stream
));
3727 si_llvm_export_vs(ctx
, outputs
, i
);
3731 static void si_tgsi_emit_epilogue(struct lp_build_tgsi_context
*bld_base
)
3733 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3735 ctx
->abi
.emit_outputs(&ctx
->abi
, RADEON_LLVM_MAX_OUTPUTS
,
3736 &ctx
->outputs
[0][0]);
3739 struct si_ps_exports
{
3741 struct ac_export_args args
[10];
3744 static void si_export_mrt_z(struct lp_build_tgsi_context
*bld_base
,
3745 LLVMValueRef depth
, LLVMValueRef stencil
,
3746 LLVMValueRef samplemask
, struct si_ps_exports
*exp
)
3748 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3749 struct ac_export_args args
;
3751 ac_export_mrt_z(&ctx
->ac
, depth
, stencil
, samplemask
, &args
);
3753 memcpy(&exp
->args
[exp
->num
++], &args
, sizeof(args
));
3756 static void si_export_mrt_color(struct lp_build_tgsi_context
*bld_base
,
3757 LLVMValueRef
*color
, unsigned index
,
3758 unsigned samplemask_param
,
3759 bool is_last
, struct si_ps_exports
*exp
)
3761 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3765 if (ctx
->shader
->key
.part
.ps
.epilog
.clamp_color
)
3766 for (i
= 0; i
< 4; i
++)
3767 color
[i
] = ac_build_clamp(&ctx
->ac
, color
[i
]);
3770 if (ctx
->shader
->key
.part
.ps
.epilog
.alpha_to_one
)
3771 color
[3] = ctx
->ac
.f32_1
;
3775 ctx
->shader
->key
.part
.ps
.epilog
.alpha_func
!= PIPE_FUNC_ALWAYS
)
3776 si_alpha_test(bld_base
, color
[3]);
3778 /* Line & polygon smoothing */
3779 if (ctx
->shader
->key
.part
.ps
.epilog
.poly_line_smoothing
)
3780 color
[3] = si_scale_alpha_by_sample_mask(bld_base
, color
[3],
3783 /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
3784 if (ctx
->shader
->key
.part
.ps
.epilog
.last_cbuf
> 0) {
3785 struct ac_export_args args
[8];
3788 /* Get the export arguments, also find out what the last one is. */
3789 for (c
= 0; c
<= ctx
->shader
->key
.part
.ps
.epilog
.last_cbuf
; c
++) {
3790 si_llvm_init_export_args(ctx
, color
,
3791 V_008DFC_SQ_EXP_MRT
+ c
, &args
[c
]);
3792 if (args
[c
].enabled_channels
)
3796 /* Emit all exports. */
3797 for (c
= 0; c
<= ctx
->shader
->key
.part
.ps
.epilog
.last_cbuf
; c
++) {
3798 if (is_last
&& last
== c
) {
3799 args
[c
].valid_mask
= 1; /* whether the EXEC mask is valid */
3800 args
[c
].done
= 1; /* DONE bit */
3801 } else if (!args
[c
].enabled_channels
)
3802 continue; /* unnecessary NULL export */
3804 memcpy(&exp
->args
[exp
->num
++], &args
[c
], sizeof(args
[c
]));
3807 struct ac_export_args args
;
3810 si_llvm_init_export_args(ctx
, color
, V_008DFC_SQ_EXP_MRT
+ index
,
3813 args
.valid_mask
= 1; /* whether the EXEC mask is valid */
3814 args
.done
= 1; /* DONE bit */
3815 } else if (!args
.enabled_channels
)
3816 return; /* unnecessary NULL export */
3818 memcpy(&exp
->args
[exp
->num
++], &args
, sizeof(args
));
3822 static void si_emit_ps_exports(struct si_shader_context
*ctx
,
3823 struct si_ps_exports
*exp
)
3825 for (unsigned i
= 0; i
< exp
->num
; i
++)
3826 ac_build_export(&ctx
->ac
, &exp
->args
[i
]);
3830 * Return PS outputs in this order:
3832 * v[0:3] = color0.xyzw
3833 * v[4:7] = color1.xyzw
3838 * vN+3 = SampleMaskIn (used for OpenGL smoothing)
3840 * The alpha-ref SGPR is returned via its original location.
3842 static void si_llvm_return_fs_outputs(struct ac_shader_abi
*abi
,
3843 unsigned max_outputs
,
3844 LLVMValueRef
*addrs
)
3846 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
3847 struct si_shader
*shader
= ctx
->shader
;
3848 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
3849 LLVMBuilderRef builder
= ctx
->ac
.builder
;
3850 unsigned i
, j
, first_vgpr
, vgpr
;
3852 LLVMValueRef color
[8][4] = {};
3853 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
3856 if (ctx
->postponed_kill
)
3857 ac_build_kill_if_false(&ctx
->ac
, LLVMBuildLoad(builder
, ctx
->postponed_kill
, ""));
3859 /* Read the output values. */
3860 for (i
= 0; i
< info
->num_outputs
; i
++) {
3861 unsigned semantic_name
= info
->output_semantic_name
[i
];
3862 unsigned semantic_index
= info
->output_semantic_index
[i
];
3864 switch (semantic_name
) {
3865 case TGSI_SEMANTIC_COLOR
:
3866 assert(semantic_index
< 8);
3867 for (j
= 0; j
< 4; j
++) {
3868 LLVMValueRef ptr
= addrs
[4 * i
+ j
];
3869 LLVMValueRef result
= LLVMBuildLoad(builder
, ptr
, "");
3870 color
[semantic_index
][j
] = result
;
3873 case TGSI_SEMANTIC_POSITION
:
3874 depth
= LLVMBuildLoad(builder
,
3875 addrs
[4 * i
+ 2], "");
3877 case TGSI_SEMANTIC_STENCIL
:
3878 stencil
= LLVMBuildLoad(builder
,
3879 addrs
[4 * i
+ 1], "");
3881 case TGSI_SEMANTIC_SAMPLEMASK
:
3882 samplemask
= LLVMBuildLoad(builder
,
3883 addrs
[4 * i
+ 0], "");
3886 fprintf(stderr
, "Warning: SI unhandled fs output type:%d\n",
3891 /* Fill the return structure. */
3892 ret
= ctx
->return_value
;
3895 ret
= LLVMBuildInsertValue(builder
, ret
,
3896 ac_to_integer(&ctx
->ac
,
3897 LLVMGetParam(ctx
->main_fn
,
3898 SI_PARAM_ALPHA_REF
)),
3899 SI_SGPR_ALPHA_REF
, "");
3902 first_vgpr
= vgpr
= SI_SGPR_ALPHA_REF
+ 1;
3903 for (i
= 0; i
< ARRAY_SIZE(color
); i
++) {
3907 for (j
= 0; j
< 4; j
++)
3908 ret
= LLVMBuildInsertValue(builder
, ret
, color
[i
][j
], vgpr
++, "");
3911 ret
= LLVMBuildInsertValue(builder
, ret
, depth
, vgpr
++, "");
3913 ret
= LLVMBuildInsertValue(builder
, ret
, stencil
, vgpr
++, "");
3915 ret
= LLVMBuildInsertValue(builder
, ret
, samplemask
, vgpr
++, "");
3917 /* Add the input sample mask for smoothing at the end. */
3918 if (vgpr
< first_vgpr
+ PS_EPILOG_SAMPLEMASK_MIN_LOC
)
3919 vgpr
= first_vgpr
+ PS_EPILOG_SAMPLEMASK_MIN_LOC
;
3920 ret
= LLVMBuildInsertValue(builder
, ret
,
3921 LLVMGetParam(ctx
->main_fn
,
3922 SI_PARAM_SAMPLE_COVERAGE
), vgpr
++, "");
3924 ctx
->return_value
= ret
;
3927 static void membar_emit(
3928 const struct lp_build_tgsi_action
*action
,
3929 struct lp_build_tgsi_context
*bld_base
,
3930 struct lp_build_emit_data
*emit_data
)
3932 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3933 LLVMValueRef src0
= lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, 0);
3934 unsigned flags
= LLVMConstIntGetZExtValue(src0
);
3935 unsigned waitcnt
= NOOP_WAITCNT
;
3937 if (flags
& TGSI_MEMBAR_THREAD_GROUP
)
3938 waitcnt
&= VM_CNT
& LGKM_CNT
;
3940 if (flags
& (TGSI_MEMBAR_ATOMIC_BUFFER
|
3941 TGSI_MEMBAR_SHADER_BUFFER
|
3942 TGSI_MEMBAR_SHADER_IMAGE
))
3945 if (flags
& TGSI_MEMBAR_SHARED
)
3946 waitcnt
&= LGKM_CNT
;
3948 if (waitcnt
!= NOOP_WAITCNT
)
3949 ac_build_waitcnt(&ctx
->ac
, waitcnt
);
3952 static void clock_emit(
3953 const struct lp_build_tgsi_action
*action
,
3954 struct lp_build_tgsi_context
*bld_base
,
3955 struct lp_build_emit_data
*emit_data
)
3957 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3958 LLVMValueRef tmp
= ac_build_shader_clock(&ctx
->ac
);
3960 emit_data
->output
[0] =
3961 LLVMBuildExtractElement(ctx
->ac
.builder
, tmp
, ctx
->i32_0
, "");
3962 emit_data
->output
[1] =
3963 LLVMBuildExtractElement(ctx
->ac
.builder
, tmp
, ctx
->i32_1
, "");
3966 static void si_llvm_emit_ddxy(
3967 const struct lp_build_tgsi_action
*action
,
3968 struct lp_build_tgsi_context
*bld_base
,
3969 struct lp_build_emit_data
*emit_data
)
3971 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3972 unsigned opcode
= emit_data
->info
->opcode
;
3977 if (opcode
== TGSI_OPCODE_DDX_FINE
)
3978 mask
= AC_TID_MASK_LEFT
;
3979 else if (opcode
== TGSI_OPCODE_DDY_FINE
)
3980 mask
= AC_TID_MASK_TOP
;
3982 mask
= AC_TID_MASK_TOP_LEFT
;
3984 /* for DDX we want to next X pixel, DDY next Y pixel. */
3985 idx
= (opcode
== TGSI_OPCODE_DDX
|| opcode
== TGSI_OPCODE_DDX_FINE
) ? 1 : 2;
3987 val
= ac_to_integer(&ctx
->ac
, emit_data
->args
[0]);
3988 val
= ac_build_ddxy(&ctx
->ac
, mask
, idx
, val
);
3989 emit_data
->output
[emit_data
->chan
] = val
;
3993 * this takes an I,J coordinate pair,
3994 * and works out the X and Y derivatives.
3995 * it returns DDX(I), DDX(J), DDY(I), DDY(J).
3997 static LLVMValueRef
si_llvm_emit_ddxy_interp(
3998 struct lp_build_tgsi_context
*bld_base
,
3999 LLVMValueRef interp_ij
)
4001 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4002 LLVMValueRef result
[4], a
;
4005 for (i
= 0; i
< 2; i
++) {
4006 a
= LLVMBuildExtractElement(ctx
->ac
.builder
, interp_ij
,
4007 LLVMConstInt(ctx
->i32
, i
, 0), "");
4008 result
[i
] = lp_build_emit_llvm_unary(bld_base
, TGSI_OPCODE_DDX
, a
);
4009 result
[2+i
] = lp_build_emit_llvm_unary(bld_base
, TGSI_OPCODE_DDY
, a
);
4012 return lp_build_gather_values(&ctx
->gallivm
, result
, 4);
4015 static void interp_fetch_args(
4016 struct lp_build_tgsi_context
*bld_base
,
4017 struct lp_build_emit_data
*emit_data
)
4019 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4020 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
4022 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_OFFSET
) {
4023 /* offset is in second src, first two channels */
4024 emit_data
->args
[0] = lp_build_emit_fetch(bld_base
,
4027 emit_data
->args
[1] = lp_build_emit_fetch(bld_base
,
4030 emit_data
->arg_count
= 2;
4031 } else if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_SAMPLE
) {
4032 LLVMValueRef sample_position
;
4033 LLVMValueRef sample_id
;
4034 LLVMValueRef halfval
= LLVMConstReal(ctx
->f32
, 0.5f
);
4036 /* fetch sample ID, then fetch its sample position,
4037 * and place into first two channels.
4039 sample_id
= lp_build_emit_fetch(bld_base
,
4040 emit_data
->inst
, 1, TGSI_CHAN_X
);
4041 sample_id
= ac_to_integer(&ctx
->ac
, sample_id
);
4043 /* Section 8.13.2 (Interpolation Functions) of the OpenGL Shading
4044 * Language 4.50 spec says about interpolateAtSample:
4046 * "Returns the value of the input interpolant variable at
4047 * the location of sample number sample. If multisample
4048 * buffers are not available, the input variable will be
4049 * evaluated at the center of the pixel. If sample sample
4050 * does not exist, the position used to interpolate the
4051 * input variable is undefined."
4053 * This means that sample_id values outside of the valid are
4054 * in fact valid input, and the usual mechanism for loading the
4055 * sample position doesn't work.
4057 if (ctx
->shader
->key
.mono
.u
.ps
.interpolate_at_sample_force_center
) {
4058 LLVMValueRef center
[4] = {
4059 LLVMConstReal(ctx
->f32
, 0.5),
4060 LLVMConstReal(ctx
->f32
, 0.5),
4065 sample_position
= lp_build_gather_values(&ctx
->gallivm
, center
, 4);
4067 sample_position
= load_sample_position(&ctx
->abi
, sample_id
);
4070 emit_data
->args
[0] = LLVMBuildExtractElement(ctx
->ac
.builder
,
4074 emit_data
->args
[0] = LLVMBuildFSub(ctx
->ac
.builder
, emit_data
->args
[0], halfval
, "");
4075 emit_data
->args
[1] = LLVMBuildExtractElement(ctx
->ac
.builder
,
4078 emit_data
->args
[1] = LLVMBuildFSub(ctx
->ac
.builder
, emit_data
->args
[1], halfval
, "");
4079 emit_data
->arg_count
= 2;
4083 static void build_interp_intrinsic(const struct lp_build_tgsi_action
*action
,
4084 struct lp_build_tgsi_context
*bld_base
,
4085 struct lp_build_emit_data
*emit_data
)
4087 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4088 struct si_shader
*shader
= ctx
->shader
;
4089 const struct tgsi_shader_info
*info
= &shader
->selector
->info
;
4090 LLVMValueRef interp_param
;
4091 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
4092 const struct tgsi_full_src_register
*input
= &inst
->Src
[0];
4093 int input_base
, input_array_size
;
4096 LLVMValueRef prim_mask
= ctx
->abi
.prim_mask
;
4097 LLVMValueRef array_idx
;
4098 int interp_param_idx
;
4102 assert(input
->Register
.File
== TGSI_FILE_INPUT
);
4104 if (input
->Register
.Indirect
) {
4105 unsigned array_id
= input
->Indirect
.ArrayID
;
4108 input_base
= info
->input_array_first
[array_id
];
4109 input_array_size
= info
->input_array_last
[array_id
] - input_base
+ 1;
4111 input_base
= inst
->Src
[0].Register
.Index
;
4112 input_array_size
= info
->num_inputs
- input_base
;
4115 array_idx
= si_get_indirect_index(ctx
, &input
->Indirect
,
4116 1, input
->Register
.Index
- input_base
);
4118 input_base
= inst
->Src
[0].Register
.Index
;
4119 input_array_size
= 1;
4120 array_idx
= ctx
->i32_0
;
4123 interp
= shader
->selector
->info
.input_interpolate
[input_base
];
4125 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_OFFSET
||
4126 inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_SAMPLE
)
4127 location
= TGSI_INTERPOLATE_LOC_CENTER
;
4129 location
= TGSI_INTERPOLATE_LOC_CENTROID
;
4131 interp_param_idx
= lookup_interp_param_index(interp
, location
);
4132 if (interp_param_idx
== -1)
4134 else if (interp_param_idx
)
4135 interp_param
= LLVMGetParam(ctx
->main_fn
, interp_param_idx
);
4137 interp_param
= NULL
;
4139 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_OFFSET
||
4140 inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_SAMPLE
) {
4141 LLVMValueRef ij_out
[2];
4142 LLVMValueRef ddxy_out
= si_llvm_emit_ddxy_interp(bld_base
, interp_param
);
4145 * take the I then J parameters, and the DDX/Y for it, and
4146 * calculate the IJ inputs for the interpolator.
4147 * temp1 = ddx * offset/sample.x + I;
4148 * interp_param.I = ddy * offset/sample.y + temp1;
4149 * temp1 = ddx * offset/sample.x + J;
4150 * interp_param.J = ddy * offset/sample.y + temp1;
4152 for (i
= 0; i
< 2; i
++) {
4153 LLVMValueRef ix_ll
= LLVMConstInt(ctx
->i32
, i
, 0);
4154 LLVMValueRef iy_ll
= LLVMConstInt(ctx
->i32
, i
+ 2, 0);
4155 LLVMValueRef ddx_el
= LLVMBuildExtractElement(ctx
->ac
.builder
,
4156 ddxy_out
, ix_ll
, "");
4157 LLVMValueRef ddy_el
= LLVMBuildExtractElement(ctx
->ac
.builder
,
4158 ddxy_out
, iy_ll
, "");
4159 LLVMValueRef interp_el
= LLVMBuildExtractElement(ctx
->ac
.builder
,
4160 interp_param
, ix_ll
, "");
4161 LLVMValueRef temp1
, temp2
;
4163 interp_el
= ac_to_float(&ctx
->ac
, interp_el
);
4165 temp1
= LLVMBuildFMul(ctx
->ac
.builder
, ddx_el
, emit_data
->args
[0], "");
4167 temp1
= LLVMBuildFAdd(ctx
->ac
.builder
, temp1
, interp_el
, "");
4169 temp2
= LLVMBuildFMul(ctx
->ac
.builder
, ddy_el
, emit_data
->args
[1], "");
4171 ij_out
[i
] = LLVMBuildFAdd(ctx
->ac
.builder
, temp2
, temp1
, "");
4173 interp_param
= lp_build_gather_values(&ctx
->gallivm
, ij_out
, 2);
4177 interp_param
= ac_to_float(&ctx
->ac
, interp_param
);
4179 for (chan
= 0; chan
< 4; chan
++) {
4180 LLVMValueRef gather
= LLVMGetUndef(LLVMVectorType(ctx
->f32
, input_array_size
));
4181 unsigned schan
= tgsi_util_get_full_src_register_swizzle(&inst
->Src
[0], chan
);
4183 for (unsigned idx
= 0; idx
< input_array_size
; ++idx
) {
4184 LLVMValueRef v
, i
= NULL
, j
= NULL
;
4187 i
= LLVMBuildExtractElement(
4188 ctx
->ac
.builder
, interp_param
, ctx
->i32_0
, "");
4189 j
= LLVMBuildExtractElement(
4190 ctx
->ac
.builder
, interp_param
, ctx
->i32_1
, "");
4192 v
= si_build_fs_interp(ctx
, input_base
+ idx
, schan
,
4195 gather
= LLVMBuildInsertElement(ctx
->ac
.builder
,
4196 gather
, v
, LLVMConstInt(ctx
->i32
, idx
, false), "");
4199 emit_data
->output
[chan
] = LLVMBuildExtractElement(
4200 ctx
->ac
.builder
, gather
, array_idx
, "");
4204 static void vote_all_emit(
4205 const struct lp_build_tgsi_action
*action
,
4206 struct lp_build_tgsi_context
*bld_base
,
4207 struct lp_build_emit_data
*emit_data
)
4209 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4211 LLVMValueRef tmp
= ac_build_vote_all(&ctx
->ac
, emit_data
->args
[0]);
4212 emit_data
->output
[emit_data
->chan
] =
4213 LLVMBuildSExt(ctx
->ac
.builder
, tmp
, ctx
->i32
, "");
4216 static void vote_any_emit(
4217 const struct lp_build_tgsi_action
*action
,
4218 struct lp_build_tgsi_context
*bld_base
,
4219 struct lp_build_emit_data
*emit_data
)
4221 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4223 LLVMValueRef tmp
= ac_build_vote_any(&ctx
->ac
, emit_data
->args
[0]);
4224 emit_data
->output
[emit_data
->chan
] =
4225 LLVMBuildSExt(ctx
->ac
.builder
, tmp
, ctx
->i32
, "");
4228 static void vote_eq_emit(
4229 const struct lp_build_tgsi_action
*action
,
4230 struct lp_build_tgsi_context
*bld_base
,
4231 struct lp_build_emit_data
*emit_data
)
4233 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4235 LLVMValueRef tmp
= ac_build_vote_eq(&ctx
->ac
, emit_data
->args
[0]);
4236 emit_data
->output
[emit_data
->chan
] =
4237 LLVMBuildSExt(ctx
->ac
.builder
, tmp
, ctx
->i32
, "");
4240 static void ballot_emit(
4241 const struct lp_build_tgsi_action
*action
,
4242 struct lp_build_tgsi_context
*bld_base
,
4243 struct lp_build_emit_data
*emit_data
)
4245 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4246 LLVMBuilderRef builder
= ctx
->ac
.builder
;
4249 tmp
= lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, TGSI_CHAN_X
);
4250 tmp
= ac_build_ballot(&ctx
->ac
, tmp
);
4251 tmp
= LLVMBuildBitCast(builder
, tmp
, ctx
->v2i32
, "");
4253 emit_data
->output
[0] = LLVMBuildExtractElement(builder
, tmp
, ctx
->i32_0
, "");
4254 emit_data
->output
[1] = LLVMBuildExtractElement(builder
, tmp
, ctx
->i32_1
, "");
4257 static void read_invoc_fetch_args(
4258 struct lp_build_tgsi_context
*bld_base
,
4259 struct lp_build_emit_data
*emit_data
)
4261 emit_data
->args
[0] = lp_build_emit_fetch(bld_base
, emit_data
->inst
,
4262 0, emit_data
->src_chan
);
4264 /* Always read the source invocation (= lane) from the X channel. */
4265 emit_data
->args
[1] = lp_build_emit_fetch(bld_base
, emit_data
->inst
,
4267 emit_data
->arg_count
= 2;
4270 static void read_lane_emit(
4271 const struct lp_build_tgsi_action
*action
,
4272 struct lp_build_tgsi_context
*bld_base
,
4273 struct lp_build_emit_data
*emit_data
)
4275 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4277 /* We currently have no other way to prevent LLVM from lifting the icmp
4278 * calls to a dominating basic block.
4280 ac_build_optimization_barrier(&ctx
->ac
, &emit_data
->args
[0]);
4282 for (unsigned i
= 0; i
< emit_data
->arg_count
; ++i
)
4283 emit_data
->args
[i
] = ac_to_integer(&ctx
->ac
, emit_data
->args
[i
]);
4285 emit_data
->output
[emit_data
->chan
] =
4286 ac_build_intrinsic(&ctx
->ac
, action
->intr_name
,
4287 ctx
->i32
, emit_data
->args
, emit_data
->arg_count
,
4288 AC_FUNC_ATTR_READNONE
|
4289 AC_FUNC_ATTR_CONVERGENT
);
4292 static unsigned si_llvm_get_stream(struct lp_build_tgsi_context
*bld_base
,
4293 struct lp_build_emit_data
*emit_data
)
4295 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4296 struct tgsi_src_register src0
= emit_data
->inst
->Src
[0].Register
;
4300 assert(src0
.File
== TGSI_FILE_IMMEDIATE
);
4302 imm
= ctx
->imms
[src0
.Index
* TGSI_NUM_CHANNELS
+ src0
.SwizzleX
];
4303 stream
= LLVMConstIntGetZExtValue(imm
) & 0x3;
4307 /* Emit one vertex from the geometry shader */
4308 static void si_llvm_emit_vertex(struct ac_shader_abi
*abi
,
4310 LLVMValueRef
*addrs
)
4312 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
4313 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
4314 struct lp_build_context
*uint
= &ctx
->bld_base
.uint_bld
;
4315 struct si_shader
*shader
= ctx
->shader
;
4316 struct lp_build_if_state if_state
;
4317 LLVMValueRef soffset
= LLVMGetParam(ctx
->main_fn
,
4318 ctx
->param_gs2vs_offset
);
4319 LLVMValueRef gs_next_vertex
;
4320 LLVMValueRef can_emit
;
4321 unsigned chan
, offset
;
4324 /* Write vertex attribute values to GSVS ring */
4325 gs_next_vertex
= LLVMBuildLoad(ctx
->ac
.builder
,
4326 ctx
->gs_next_vertex
[stream
],
4329 /* If this thread has already emitted the declared maximum number of
4330 * vertices, skip the write: excessive vertex emissions are not
4331 * supposed to have any effect.
4333 * If the shader has no writes to memory, kill it instead. This skips
4334 * further memory loads and may allow LLVM to skip to the end
4337 can_emit
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntULT
, gs_next_vertex
,
4338 LLVMConstInt(ctx
->i32
,
4339 shader
->selector
->gs_max_out_vertices
, 0), "");
4341 bool use_kill
= !info
->writes_memory
;
4343 ac_build_kill_if_false(&ctx
->ac
, can_emit
);
4345 lp_build_if(&if_state
, &ctx
->gallivm
, can_emit
);
4349 for (i
= 0; i
< info
->num_outputs
; i
++) {
4350 for (chan
= 0; chan
< 4; chan
++) {
4351 if (!(info
->output_usagemask
[i
] & (1 << chan
)) ||
4352 ((info
->output_streams
[i
] >> (2 * chan
)) & 3) != stream
)
4355 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->ac
.builder
, addrs
[4 * i
+ chan
], "");
4356 LLVMValueRef voffset
=
4357 LLVMConstInt(ctx
->i32
, offset
*
4358 shader
->selector
->gs_max_out_vertices
, 0);
4361 voffset
= lp_build_add(uint
, voffset
, gs_next_vertex
);
4362 voffset
= lp_build_mul_imm(uint
, voffset
, 4);
4364 out_val
= ac_to_integer(&ctx
->ac
, out_val
);
4366 ac_build_buffer_store_dword(&ctx
->ac
,
4367 ctx
->gsvs_ring
[stream
],
4369 voffset
, soffset
, 0,
4374 gs_next_vertex
= lp_build_add(uint
, gs_next_vertex
,
4377 LLVMBuildStore(ctx
->ac
.builder
, gs_next_vertex
, ctx
->gs_next_vertex
[stream
]);
4379 /* Signal vertex emission */
4380 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_EMIT
| AC_SENDMSG_GS
| (stream
<< 8),
4381 si_get_gs_wave_id(ctx
));
4383 lp_build_endif(&if_state
);
4386 /* Emit one vertex from the geometry shader */
4387 static void si_tgsi_emit_vertex(
4388 const struct lp_build_tgsi_action
*action
,
4389 struct lp_build_tgsi_context
*bld_base
,
4390 struct lp_build_emit_data
*emit_data
)
4392 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4393 unsigned stream
= si_llvm_get_stream(bld_base
, emit_data
);
4395 si_llvm_emit_vertex(&ctx
->abi
, stream
, ctx
->outputs
[0]);
4398 /* Cut one primitive from the geometry shader */
4399 static void si_llvm_emit_primitive(struct ac_shader_abi
*abi
,
4402 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
4404 /* Signal primitive cut */
4405 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_CUT
| AC_SENDMSG_GS
| (stream
<< 8),
4406 si_get_gs_wave_id(ctx
));
4409 /* Cut one primitive from the geometry shader */
4410 static void si_tgsi_emit_primitive(
4411 const struct lp_build_tgsi_action
*action
,
4412 struct lp_build_tgsi_context
*bld_base
,
4413 struct lp_build_emit_data
*emit_data
)
4415 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4417 si_llvm_emit_primitive(&ctx
->abi
, si_llvm_get_stream(bld_base
, emit_data
));
4420 static void si_llvm_emit_barrier(const struct lp_build_tgsi_action
*action
,
4421 struct lp_build_tgsi_context
*bld_base
,
4422 struct lp_build_emit_data
*emit_data
)
4424 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4426 /* SI only (thanks to a hw bug workaround):
4427 * The real barrier instruction isn’t needed, because an entire patch
4428 * always fits into a single wave.
4430 if (ctx
->screen
->info
.chip_class
== SI
&&
4431 ctx
->type
== PIPE_SHADER_TESS_CTRL
) {
4432 ac_build_waitcnt(&ctx
->ac
, LGKM_CNT
& VM_CNT
);
4436 lp_build_intrinsic(ctx
->ac
.builder
,
4437 "llvm.amdgcn.s.barrier",
4438 ctx
->voidt
, NULL
, 0, LP_FUNC_ATTR_CONVERGENT
);
4441 static const struct lp_build_tgsi_action interp_action
= {
4442 .fetch_args
= interp_fetch_args
,
4443 .emit
= build_interp_intrinsic
,
4446 static void si_create_function(struct si_shader_context
*ctx
,
4448 LLVMTypeRef
*returns
, unsigned num_returns
,
4449 struct si_function_info
*fninfo
,
4450 unsigned max_workgroup_size
)
4454 si_llvm_create_func(ctx
, name
, returns
, num_returns
,
4455 fninfo
->types
, fninfo
->num_params
);
4456 ctx
->return_value
= LLVMGetUndef(ctx
->return_type
);
4458 for (i
= 0; i
< fninfo
->num_sgpr_params
; ++i
) {
4459 LLVMValueRef P
= LLVMGetParam(ctx
->main_fn
, i
);
4461 /* The combination of:
4465 * allows the optimization passes to move loads and reduces
4466 * SGPR spilling significantly.
4468 lp_add_function_attr(ctx
->main_fn
, i
+ 1, LP_FUNC_ATTR_INREG
);
4470 if (LLVMGetTypeKind(LLVMTypeOf(P
)) == LLVMPointerTypeKind
) {
4471 lp_add_function_attr(ctx
->main_fn
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
4472 ac_add_attr_dereferenceable(P
, UINT64_MAX
);
4476 for (i
= 0; i
< fninfo
->num_params
; ++i
) {
4477 if (fninfo
->assign
[i
])
4478 *fninfo
->assign
[i
] = LLVMGetParam(ctx
->main_fn
, i
);
4481 if (ctx
->screen
->info
.address32_hi
) {
4482 ac_llvm_add_target_dep_function_attr(ctx
->main_fn
,
4483 "amdgpu-32bit-address-high-bits",
4484 ctx
->screen
->info
.address32_hi
);
4487 if (max_workgroup_size
) {
4488 ac_llvm_add_target_dep_function_attr(ctx
->main_fn
,
4489 "amdgpu-max-work-group-size",
4490 max_workgroup_size
);
4492 LLVMAddTargetDependentFunctionAttr(ctx
->main_fn
,
4493 "no-signed-zeros-fp-math",
4496 if (ctx
->screen
->debug_flags
& DBG(UNSAFE_MATH
)) {
4497 /* These were copied from some LLVM test. */
4498 LLVMAddTargetDependentFunctionAttr(ctx
->main_fn
,
4499 "less-precise-fpmad",
4501 LLVMAddTargetDependentFunctionAttr(ctx
->main_fn
,
4504 LLVMAddTargetDependentFunctionAttr(ctx
->main_fn
,
4507 LLVMAddTargetDependentFunctionAttr(ctx
->main_fn
,
4513 static void declare_streamout_params(struct si_shader_context
*ctx
,
4514 struct pipe_stream_output_info
*so
,
4515 struct si_function_info
*fninfo
)
4519 /* Streamout SGPRs. */
4520 if (so
->num_outputs
) {
4521 if (ctx
->type
!= PIPE_SHADER_TESS_EVAL
)
4522 ctx
->param_streamout_config
= add_arg(fninfo
, ARG_SGPR
, ctx
->ac
.i32
);
4524 ctx
->param_streamout_config
= fninfo
->num_params
- 1;
4526 ctx
->param_streamout_write_index
= add_arg(fninfo
, ARG_SGPR
, ctx
->ac
.i32
);
4528 /* A streamout buffer offset is loaded if the stride is non-zero. */
4529 for (i
= 0; i
< 4; i
++) {
4533 ctx
->param_streamout_offset
[i
] = add_arg(fninfo
, ARG_SGPR
, ctx
->ac
.i32
);
4537 static unsigned si_get_max_workgroup_size(const struct si_shader
*shader
)
4539 switch (shader
->selector
->type
) {
4540 case PIPE_SHADER_TESS_CTRL
:
4541 /* Return this so that LLVM doesn't remove s_barrier
4542 * instructions on chips where we use s_barrier. */
4543 return shader
->selector
->screen
->info
.chip_class
>= CIK
? 128 : 64;
4545 case PIPE_SHADER_GEOMETRY
:
4546 return shader
->selector
->screen
->info
.chip_class
>= GFX9
? 128 : 64;
4548 case PIPE_SHADER_COMPUTE
:
4549 break; /* see below */
4555 const unsigned *properties
= shader
->selector
->info
.properties
;
4556 unsigned max_work_group_size
=
4557 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
] *
4558 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT
] *
4559 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH
];
4561 if (!max_work_group_size
) {
4562 /* This is a variable group size compute shader,
4563 * compile it for the maximum possible group size.
4565 max_work_group_size
= SI_MAX_VARIABLE_THREADS_PER_BLOCK
;
4567 return max_work_group_size
;
4570 static void declare_const_and_shader_buffers(struct si_shader_context
*ctx
,
4571 struct si_function_info
*fninfo
,
4574 LLVMTypeRef const_shader_buf_type
;
4576 if (ctx
->shader
->selector
->info
.const_buffers_declared
== 1 &&
4577 ctx
->shader
->selector
->info
.shader_buffers_declared
== 0)
4578 const_shader_buf_type
= ctx
->f32
;
4580 const_shader_buf_type
= ctx
->v4i32
;
4582 unsigned const_and_shader_buffers
=
4583 add_arg(fninfo
, ARG_SGPR
,
4584 ac_array_in_const32_addr_space(const_shader_buf_type
));
4587 ctx
->param_const_and_shader_buffers
= const_and_shader_buffers
;
4590 static void declare_samplers_and_images(struct si_shader_context
*ctx
,
4591 struct si_function_info
*fninfo
,
4594 unsigned samplers_and_images
=
4595 add_arg(fninfo
, ARG_SGPR
,
4596 ac_array_in_const32_addr_space(ctx
->v8i32
));
4599 ctx
->param_samplers_and_images
= samplers_and_images
;
4602 static void declare_per_stage_desc_pointers(struct si_shader_context
*ctx
,
4603 struct si_function_info
*fninfo
,
4606 declare_const_and_shader_buffers(ctx
, fninfo
, assign_params
);
4607 declare_samplers_and_images(ctx
, fninfo
, assign_params
);
4610 static void declare_global_desc_pointers(struct si_shader_context
*ctx
,
4611 struct si_function_info
*fninfo
)
4613 ctx
->param_rw_buffers
= add_arg(fninfo
, ARG_SGPR
,
4614 ac_array_in_const32_addr_space(ctx
->v4i32
));
4615 ctx
->param_bindless_samplers_and_images
= add_arg(fninfo
, ARG_SGPR
,
4616 ac_array_in_const32_addr_space(ctx
->v8i32
));
4619 static void declare_vs_specific_input_sgprs(struct si_shader_context
*ctx
,
4620 struct si_function_info
*fninfo
)
4622 add_arg_assign(fninfo
, ARG_SGPR
, ctx
->i32
, &ctx
->abi
.base_vertex
);
4623 add_arg_assign(fninfo
, ARG_SGPR
, ctx
->i32
, &ctx
->abi
.start_instance
);
4624 add_arg_assign(fninfo
, ARG_SGPR
, ctx
->i32
, &ctx
->abi
.draw_id
);
4625 ctx
->param_vs_state_bits
= add_arg(fninfo
, ARG_SGPR
, ctx
->i32
);
4628 static void declare_vs_input_vgprs(struct si_shader_context
*ctx
,
4629 struct si_function_info
*fninfo
,
4630 unsigned *num_prolog_vgprs
)
4632 struct si_shader
*shader
= ctx
->shader
;
4634 add_arg_assign(fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.vertex_id
);
4635 if (shader
->key
.as_ls
) {
4636 ctx
->param_rel_auto_id
= add_arg(fninfo
, ARG_VGPR
, ctx
->i32
);
4637 add_arg_assign(fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.instance_id
);
4639 add_arg_assign(fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.instance_id
);
4640 ctx
->param_vs_prim_id
= add_arg(fninfo
, ARG_VGPR
, ctx
->i32
);
4642 add_arg(fninfo
, ARG_VGPR
, ctx
->i32
); /* unused */
4644 if (!shader
->is_gs_copy_shader
) {
4645 /* Vertex load indices. */
4646 ctx
->param_vertex_index0
= fninfo
->num_params
;
4647 for (unsigned i
= 0; i
< shader
->selector
->info
.num_inputs
; i
++)
4648 add_arg(fninfo
, ARG_VGPR
, ctx
->i32
);
4649 *num_prolog_vgprs
+= shader
->selector
->info
.num_inputs
;
4653 static void declare_tes_input_vgprs(struct si_shader_context
*ctx
,
4654 struct si_function_info
*fninfo
)
4656 ctx
->param_tes_u
= add_arg(fninfo
, ARG_VGPR
, ctx
->f32
);
4657 ctx
->param_tes_v
= add_arg(fninfo
, ARG_VGPR
, ctx
->f32
);
4658 ctx
->param_tes_rel_patch_id
= add_arg(fninfo
, ARG_VGPR
, ctx
->i32
);
4659 add_arg_assign(fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.tes_patch_id
);
4663 /* Convenient merged shader definitions. */
4664 SI_SHADER_MERGED_VERTEX_TESSCTRL
= PIPE_SHADER_TYPES
,
4665 SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY
,
4668 static void create_function(struct si_shader_context
*ctx
)
4670 struct si_shader
*shader
= ctx
->shader
;
4671 struct si_function_info fninfo
;
4672 LLVMTypeRef returns
[16+32*4];
4673 unsigned i
, num_return_sgprs
;
4674 unsigned num_returns
= 0;
4675 unsigned num_prolog_vgprs
= 0;
4676 unsigned type
= ctx
->type
;
4677 unsigned vs_blit_property
=
4678 shader
->selector
->info
.properties
[TGSI_PROPERTY_VS_BLIT_SGPRS
];
4680 si_init_function_info(&fninfo
);
4682 /* Set MERGED shaders. */
4683 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
4684 if (shader
->key
.as_ls
|| type
== PIPE_SHADER_TESS_CTRL
)
4685 type
= SI_SHADER_MERGED_VERTEX_TESSCTRL
; /* LS or HS */
4686 else if (shader
->key
.as_es
|| type
== PIPE_SHADER_GEOMETRY
)
4687 type
= SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY
;
4690 LLVMTypeRef v3i32
= LLVMVectorType(ctx
->i32
, 3);
4693 case PIPE_SHADER_VERTEX
:
4694 declare_global_desc_pointers(ctx
, &fninfo
);
4696 if (vs_blit_property
) {
4697 ctx
->param_vs_blit_inputs
= fninfo
.num_params
;
4698 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* i16 x1, y1 */
4699 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* i16 x2, y2 */
4700 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* depth */
4702 if (vs_blit_property
== SI_VS_BLIT_SGPRS_POS_COLOR
) {
4703 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* color0 */
4704 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* color1 */
4705 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* color2 */
4706 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* color3 */
4707 } else if (vs_blit_property
== SI_VS_BLIT_SGPRS_POS_TEXCOORD
) {
4708 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* texcoord.x1 */
4709 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* texcoord.y1 */
4710 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* texcoord.x2 */
4711 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* texcoord.y2 */
4712 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* texcoord.z */
4713 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* texcoord.w */
4717 declare_vs_input_vgprs(ctx
, &fninfo
, &num_prolog_vgprs
);
4721 declare_per_stage_desc_pointers(ctx
, &fninfo
, true);
4722 declare_vs_specific_input_sgprs(ctx
, &fninfo
);
4723 ctx
->param_vertex_buffers
= add_arg(&fninfo
, ARG_SGPR
,
4724 ac_array_in_const32_addr_space(ctx
->v4i32
));
4726 if (shader
->key
.as_es
) {
4727 ctx
->param_es2gs_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4728 } else if (shader
->key
.as_ls
) {
4729 /* no extra parameters */
4731 if (shader
->is_gs_copy_shader
) {
4732 fninfo
.num_params
= ctx
->param_rw_buffers
+ 1;
4733 fninfo
.num_sgpr_params
= fninfo
.num_params
;
4736 /* The locations of the other parameters are assigned dynamically. */
4737 declare_streamout_params(ctx
, &shader
->selector
->so
,
4742 declare_vs_input_vgprs(ctx
, &fninfo
, &num_prolog_vgprs
);
4745 case PIPE_SHADER_TESS_CTRL
: /* SI-CI-VI */
4746 declare_global_desc_pointers(ctx
, &fninfo
);
4747 declare_per_stage_desc_pointers(ctx
, &fninfo
, true);
4748 ctx
->param_tcs_offchip_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4749 ctx
->param_tcs_out_lds_offsets
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4750 ctx
->param_tcs_out_lds_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4751 ctx
->param_vs_state_bits
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4752 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4753 ctx
->param_tcs_factor_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4756 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.tcs_patch_id
);
4757 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.tcs_rel_ids
);
4759 /* param_tcs_offchip_offset and param_tcs_factor_offset are
4760 * placed after the user SGPRs.
4762 for (i
= 0; i
< GFX6_TCS_NUM_USER_SGPR
+ 2; i
++)
4763 returns
[num_returns
++] = ctx
->i32
; /* SGPRs */
4764 for (i
= 0; i
< 11; i
++)
4765 returns
[num_returns
++] = ctx
->f32
; /* VGPRs */
4768 case SI_SHADER_MERGED_VERTEX_TESSCTRL
:
4769 /* Merged stages have 8 system SGPRs at the beginning. */
4770 /* SPI_SHADER_USER_DATA_ADDR_LO/HI_HS */
4771 if (HAVE_32BIT_POINTERS
) {
4772 declare_per_stage_desc_pointers(ctx
, &fninfo
,
4773 ctx
->type
== PIPE_SHADER_TESS_CTRL
);
4775 declare_const_and_shader_buffers(ctx
, &fninfo
,
4776 ctx
->type
== PIPE_SHADER_TESS_CTRL
);
4778 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4779 ctx
->param_merged_wave_info
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4780 ctx
->param_tcs_factor_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4781 ctx
->param_merged_scratch_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4782 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused */
4783 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused */
4785 declare_global_desc_pointers(ctx
, &fninfo
);
4786 declare_per_stage_desc_pointers(ctx
, &fninfo
,
4787 ctx
->type
== PIPE_SHADER_VERTEX
);
4788 declare_vs_specific_input_sgprs(ctx
, &fninfo
);
4790 if (!HAVE_32BIT_POINTERS
) {
4791 declare_samplers_and_images(ctx
, &fninfo
,
4792 ctx
->type
== PIPE_SHADER_TESS_CTRL
);
4794 ctx
->param_tcs_offchip_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4795 ctx
->param_tcs_out_lds_offsets
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4796 ctx
->param_tcs_out_lds_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4797 if (!HAVE_32BIT_POINTERS
) /* Align to 2 dwords. */
4798 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused */
4799 ctx
->param_vertex_buffers
= add_arg(&fninfo
, ARG_SGPR
,
4800 ac_array_in_const32_addr_space(ctx
->v4i32
));
4802 /* VGPRs (first TCS, then VS) */
4803 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.tcs_patch_id
);
4804 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.tcs_rel_ids
);
4806 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
4807 declare_vs_input_vgprs(ctx
, &fninfo
,
4810 /* LS return values are inputs to the TCS main shader part. */
4811 for (i
= 0; i
< 8 + GFX9_TCS_NUM_USER_SGPR
; i
++)
4812 returns
[num_returns
++] = ctx
->i32
; /* SGPRs */
4813 for (i
= 0; i
< 2; i
++)
4814 returns
[num_returns
++] = ctx
->f32
; /* VGPRs */
4816 /* TCS return values are inputs to the TCS epilog.
4818 * param_tcs_offchip_offset, param_tcs_factor_offset,
4819 * param_tcs_offchip_layout, and param_rw_buffers
4820 * should be passed to the epilog.
4822 for (i
= 0; i
<= 8 + GFX9_SGPR_TCS_OUT_LAYOUT
; i
++)
4823 returns
[num_returns
++] = ctx
->i32
; /* SGPRs */
4824 for (i
= 0; i
< 11; i
++)
4825 returns
[num_returns
++] = ctx
->f32
; /* VGPRs */
4829 case SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY
:
4830 /* Merged stages have 8 system SGPRs at the beginning. */
4831 /* SPI_SHADER_USER_DATA_ADDR_LO/HI_GS */
4832 if (HAVE_32BIT_POINTERS
) {
4833 declare_per_stage_desc_pointers(ctx
, &fninfo
,
4834 ctx
->type
== PIPE_SHADER_GEOMETRY
);
4836 declare_const_and_shader_buffers(ctx
, &fninfo
,
4837 ctx
->type
== PIPE_SHADER_GEOMETRY
);
4839 ctx
->param_gs2vs_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4840 ctx
->param_merged_wave_info
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4841 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4842 ctx
->param_merged_scratch_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4843 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused (SPI_SHADER_PGM_LO/HI_GS << 8) */
4844 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused (SPI_SHADER_PGM_LO/HI_GS >> 24) */
4846 declare_global_desc_pointers(ctx
, &fninfo
);
4847 declare_per_stage_desc_pointers(ctx
, &fninfo
,
4848 (ctx
->type
== PIPE_SHADER_VERTEX
||
4849 ctx
->type
== PIPE_SHADER_TESS_EVAL
));
4850 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
4851 declare_vs_specific_input_sgprs(ctx
, &fninfo
);
4853 ctx
->param_tcs_offchip_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4854 ctx
->param_tes_offchip_addr
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4855 if (!HAVE_32BIT_POINTERS
) {
4856 /* Declare as many input SGPRs as the VS has. */
4857 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused */
4858 ctx
->param_vs_state_bits
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused */
4862 if (!HAVE_32BIT_POINTERS
) {
4863 declare_samplers_and_images(ctx
, &fninfo
,
4864 ctx
->type
== PIPE_SHADER_GEOMETRY
);
4866 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
4867 ctx
->param_vertex_buffers
= add_arg(&fninfo
, ARG_SGPR
,
4868 ac_array_in_const32_addr_space(ctx
->v4i32
));
4871 /* VGPRs (first GS, then VS/TES) */
4872 ctx
->param_gs_vtx01_offset
= add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
4873 ctx
->param_gs_vtx23_offset
= add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
4874 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.gs_prim_id
);
4875 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.gs_invocation_id
);
4876 ctx
->param_gs_vtx45_offset
= add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
4878 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
4879 declare_vs_input_vgprs(ctx
, &fninfo
,
4881 } else if (ctx
->type
== PIPE_SHADER_TESS_EVAL
) {
4882 declare_tes_input_vgprs(ctx
, &fninfo
);
4885 if (ctx
->type
== PIPE_SHADER_VERTEX
||
4886 ctx
->type
== PIPE_SHADER_TESS_EVAL
) {
4887 unsigned num_user_sgprs
;
4889 if (ctx
->type
== PIPE_SHADER_VERTEX
)
4890 num_user_sgprs
= GFX9_VSGS_NUM_USER_SGPR
;
4892 num_user_sgprs
= GFX9_TESGS_NUM_USER_SGPR
;
4894 /* ES return values are inputs to GS. */
4895 for (i
= 0; i
< 8 + num_user_sgprs
; i
++)
4896 returns
[num_returns
++] = ctx
->i32
; /* SGPRs */
4897 for (i
= 0; i
< 5; i
++)
4898 returns
[num_returns
++] = ctx
->f32
; /* VGPRs */
4902 case PIPE_SHADER_TESS_EVAL
:
4903 declare_global_desc_pointers(ctx
, &fninfo
);
4904 declare_per_stage_desc_pointers(ctx
, &fninfo
, true);
4905 ctx
->param_tcs_offchip_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4906 ctx
->param_tes_offchip_addr
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4908 if (shader
->key
.as_es
) {
4909 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4910 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4911 ctx
->param_es2gs_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4913 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4914 declare_streamout_params(ctx
, &shader
->selector
->so
,
4916 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4920 declare_tes_input_vgprs(ctx
, &fninfo
);
4923 case PIPE_SHADER_GEOMETRY
:
4924 declare_global_desc_pointers(ctx
, &fninfo
);
4925 declare_per_stage_desc_pointers(ctx
, &fninfo
, true);
4926 ctx
->param_gs2vs_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4927 ctx
->param_gs_wave_id
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4930 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->gs_vtx_offset
[0]);
4931 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->gs_vtx_offset
[1]);
4932 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.gs_prim_id
);
4933 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->gs_vtx_offset
[2]);
4934 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->gs_vtx_offset
[3]);
4935 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->gs_vtx_offset
[4]);
4936 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->gs_vtx_offset
[5]);
4937 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.gs_invocation_id
);
4940 case PIPE_SHADER_FRAGMENT
:
4941 declare_global_desc_pointers(ctx
, &fninfo
);
4942 declare_per_stage_desc_pointers(ctx
, &fninfo
, true);
4943 add_arg_checked(&fninfo
, ARG_SGPR
, ctx
->f32
, SI_PARAM_ALPHA_REF
);
4944 add_arg_assign_checked(&fninfo
, ARG_SGPR
, ctx
->i32
,
4945 &ctx
->abi
.prim_mask
, SI_PARAM_PRIM_MASK
);
4947 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->v2i32
, SI_PARAM_PERSP_SAMPLE
);
4948 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->v2i32
, SI_PARAM_PERSP_CENTER
);
4949 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->v2i32
, SI_PARAM_PERSP_CENTROID
);
4950 add_arg_checked(&fninfo
, ARG_VGPR
, v3i32
, SI_PARAM_PERSP_PULL_MODEL
);
4951 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->v2i32
, SI_PARAM_LINEAR_SAMPLE
);
4952 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->v2i32
, SI_PARAM_LINEAR_CENTER
);
4953 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->v2i32
, SI_PARAM_LINEAR_CENTROID
);
4954 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->f32
, SI_PARAM_LINE_STIPPLE_TEX
);
4955 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->f32
,
4956 &ctx
->abi
.frag_pos
[0], SI_PARAM_POS_X_FLOAT
);
4957 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->f32
,
4958 &ctx
->abi
.frag_pos
[1], SI_PARAM_POS_Y_FLOAT
);
4959 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->f32
,
4960 &ctx
->abi
.frag_pos
[2], SI_PARAM_POS_Z_FLOAT
);
4961 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->f32
,
4962 &ctx
->abi
.frag_pos
[3], SI_PARAM_POS_W_FLOAT
);
4963 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->i32
,
4964 &ctx
->abi
.front_face
, SI_PARAM_FRONT_FACE
);
4965 shader
->info
.face_vgpr_index
= 20;
4966 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->i32
,
4967 &ctx
->abi
.ancillary
, SI_PARAM_ANCILLARY
);
4968 shader
->info
.ancillary_vgpr_index
= 21;
4969 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->f32
,
4970 &ctx
->abi
.sample_coverage
, SI_PARAM_SAMPLE_COVERAGE
);
4971 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->i32
, SI_PARAM_POS_FIXED_PT
);
4973 /* Color inputs from the prolog. */
4974 if (shader
->selector
->info
.colors_read
) {
4975 unsigned num_color_elements
=
4976 util_bitcount(shader
->selector
->info
.colors_read
);
4978 assert(fninfo
.num_params
+ num_color_elements
<= ARRAY_SIZE(fninfo
.types
));
4979 for (i
= 0; i
< num_color_elements
; i
++)
4980 add_arg(&fninfo
, ARG_VGPR
, ctx
->f32
);
4982 num_prolog_vgprs
+= num_color_elements
;
4985 /* Outputs for the epilog. */
4986 num_return_sgprs
= SI_SGPR_ALPHA_REF
+ 1;
4989 util_bitcount(shader
->selector
->info
.colors_written
) * 4 +
4990 shader
->selector
->info
.writes_z
+
4991 shader
->selector
->info
.writes_stencil
+
4992 shader
->selector
->info
.writes_samplemask
+
4993 1 /* SampleMaskIn */;
4995 num_returns
= MAX2(num_returns
,
4997 PS_EPILOG_SAMPLEMASK_MIN_LOC
+ 1);
4999 for (i
= 0; i
< num_return_sgprs
; i
++)
5000 returns
[i
] = ctx
->i32
;
5001 for (; i
< num_returns
; i
++)
5002 returns
[i
] = ctx
->f32
;
5005 case PIPE_SHADER_COMPUTE
:
5006 declare_global_desc_pointers(ctx
, &fninfo
);
5007 declare_per_stage_desc_pointers(ctx
, &fninfo
, true);
5008 if (shader
->selector
->info
.uses_grid_size
)
5009 add_arg_assign(&fninfo
, ARG_SGPR
, v3i32
, &ctx
->abi
.num_work_groups
);
5010 if (shader
->selector
->info
.uses_block_size
)
5011 ctx
->param_block_size
= add_arg(&fninfo
, ARG_SGPR
, v3i32
);
5013 for (i
= 0; i
< 3; i
++) {
5014 ctx
->abi
.workgroup_ids
[i
] = NULL
;
5015 if (shader
->selector
->info
.uses_block_id
[i
])
5016 add_arg_assign(&fninfo
, ARG_SGPR
, ctx
->i32
, &ctx
->abi
.workgroup_ids
[i
]);
5019 add_arg_assign(&fninfo
, ARG_VGPR
, v3i32
, &ctx
->abi
.local_invocation_ids
);
5022 assert(0 && "unimplemented shader");
5026 si_create_function(ctx
, "main", returns
, num_returns
, &fninfo
,
5027 si_get_max_workgroup_size(shader
));
5029 /* Reserve register locations for VGPR inputs the PS prolog may need. */
5030 if (ctx
->type
== PIPE_SHADER_FRAGMENT
&&
5031 ctx
->separate_prolog
) {
5032 ac_llvm_add_target_dep_function_attr(ctx
->main_fn
,
5033 "InitialPSInputAddr",
5034 S_0286D0_PERSP_SAMPLE_ENA(1) |
5035 S_0286D0_PERSP_CENTER_ENA(1) |
5036 S_0286D0_PERSP_CENTROID_ENA(1) |
5037 S_0286D0_LINEAR_SAMPLE_ENA(1) |
5038 S_0286D0_LINEAR_CENTER_ENA(1) |
5039 S_0286D0_LINEAR_CENTROID_ENA(1) |
5040 S_0286D0_FRONT_FACE_ENA(1) |
5041 S_0286D0_ANCILLARY_ENA(1) |
5042 S_0286D0_POS_FIXED_PT_ENA(1));
5045 shader
->info
.num_input_sgprs
= 0;
5046 shader
->info
.num_input_vgprs
= 0;
5048 for (i
= 0; i
< fninfo
.num_sgpr_params
; ++i
)
5049 shader
->info
.num_input_sgprs
+= ac_get_type_size(fninfo
.types
[i
]) / 4;
5051 for (; i
< fninfo
.num_params
; ++i
)
5052 shader
->info
.num_input_vgprs
+= ac_get_type_size(fninfo
.types
[i
]) / 4;
5054 assert(shader
->info
.num_input_vgprs
>= num_prolog_vgprs
);
5055 shader
->info
.num_input_vgprs
-= num_prolog_vgprs
;
5057 if (shader
->key
.as_ls
||
5058 ctx
->type
== PIPE_SHADER_TESS_CTRL
||
5059 /* GFX9 has the ESGS ring buffer in LDS. */
5060 type
== SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY
)
5061 ac_declare_lds_as_pointer(&ctx
->ac
);
5065 * Load ESGS and GSVS ring buffer resource descriptors and save the variables
5068 static void preload_ring_buffers(struct si_shader_context
*ctx
)
5070 LLVMBuilderRef builder
= ctx
->ac
.builder
;
5072 LLVMValueRef buf_ptr
= LLVMGetParam(ctx
->main_fn
,
5073 ctx
->param_rw_buffers
);
5075 if (ctx
->screen
->info
.chip_class
<= VI
&&
5076 (ctx
->shader
->key
.as_es
|| ctx
->type
== PIPE_SHADER_GEOMETRY
)) {
5078 ctx
->type
== PIPE_SHADER_GEOMETRY
? SI_GS_RING_ESGS
5080 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
, ring
, 0);
5083 ac_build_load_to_sgpr(&ctx
->ac
, buf_ptr
, offset
);
5086 if (ctx
->shader
->is_gs_copy_shader
) {
5087 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
, SI_RING_GSVS
, 0);
5090 ac_build_load_to_sgpr(&ctx
->ac
, buf_ptr
, offset
);
5091 } else if (ctx
->type
== PIPE_SHADER_GEOMETRY
) {
5092 const struct si_shader_selector
*sel
= ctx
->shader
->selector
;
5093 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
, SI_RING_GSVS
, 0);
5094 LLVMValueRef base_ring
;
5096 base_ring
= ac_build_load_to_sgpr(&ctx
->ac
, buf_ptr
, offset
);
5098 /* The conceptual layout of the GSVS ring is
5099 * v0c0 .. vLv0 v0c1 .. vLc1 ..
5100 * but the real memory layout is swizzled across
5102 * t0v0c0 .. t15v0c0 t0v1c0 .. t15v1c0 ... t15vLcL
5104 * Override the buffer descriptor accordingly.
5106 LLVMTypeRef v2i64
= LLVMVectorType(ctx
->i64
, 2);
5107 uint64_t stream_offset
= 0;
5109 for (unsigned stream
= 0; stream
< 4; ++stream
) {
5110 unsigned num_components
;
5112 unsigned num_records
;
5113 LLVMValueRef ring
, tmp
;
5115 num_components
= sel
->info
.num_stream_output_components
[stream
];
5116 if (!num_components
)
5119 stride
= 4 * num_components
* sel
->gs_max_out_vertices
;
5121 /* Limit on the stride field for <= CIK. */
5122 assert(stride
< (1 << 14));
5126 ring
= LLVMBuildBitCast(builder
, base_ring
, v2i64
, "");
5127 tmp
= LLVMBuildExtractElement(builder
, ring
, ctx
->i32_0
, "");
5128 tmp
= LLVMBuildAdd(builder
, tmp
,
5129 LLVMConstInt(ctx
->i64
,
5130 stream_offset
, 0), "");
5131 stream_offset
+= stride
* 64;
5133 ring
= LLVMBuildInsertElement(builder
, ring
, tmp
, ctx
->i32_0
, "");
5134 ring
= LLVMBuildBitCast(builder
, ring
, ctx
->v4i32
, "");
5135 tmp
= LLVMBuildExtractElement(builder
, ring
, ctx
->i32_1
, "");
5136 tmp
= LLVMBuildOr(builder
, tmp
,
5137 LLVMConstInt(ctx
->i32
,
5138 S_008F04_STRIDE(stride
) |
5139 S_008F04_SWIZZLE_ENABLE(1), 0), "");
5140 ring
= LLVMBuildInsertElement(builder
, ring
, tmp
, ctx
->i32_1
, "");
5141 ring
= LLVMBuildInsertElement(builder
, ring
,
5142 LLVMConstInt(ctx
->i32
, num_records
, 0),
5143 LLVMConstInt(ctx
->i32
, 2, 0), "");
5144 ring
= LLVMBuildInsertElement(builder
, ring
,
5145 LLVMConstInt(ctx
->i32
,
5146 S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
5147 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
5148 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
5149 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
5150 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
5151 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
5152 S_008F0C_ELEMENT_SIZE(1) | /* element_size = 4 (bytes) */
5153 S_008F0C_INDEX_STRIDE(1) | /* index_stride = 16 (elements) */
5154 S_008F0C_ADD_TID_ENABLE(1),
5156 LLVMConstInt(ctx
->i32
, 3, 0), "");
5158 ctx
->gsvs_ring
[stream
] = ring
;
5160 } else if (ctx
->type
== PIPE_SHADER_TESS_EVAL
) {
5161 ctx
->tess_offchip_ring
= get_tess_ring_descriptor(ctx
, TESS_OFFCHIP_RING_TES
);
5165 static void si_llvm_emit_polygon_stipple(struct si_shader_context
*ctx
,
5166 LLVMValueRef param_rw_buffers
,
5167 unsigned param_pos_fixed_pt
)
5169 LLVMBuilderRef builder
= ctx
->ac
.builder
;
5170 LLVMValueRef slot
, desc
, offset
, row
, bit
, address
[2];
5172 /* Use the fixed-point gl_FragCoord input.
5173 * Since the stipple pattern is 32x32 and it repeats, just get 5 bits
5174 * per coordinate to get the repeating effect.
5176 address
[0] = unpack_param(ctx
, param_pos_fixed_pt
, 0, 5);
5177 address
[1] = unpack_param(ctx
, param_pos_fixed_pt
, 16, 5);
5179 /* Load the buffer descriptor. */
5180 slot
= LLVMConstInt(ctx
->i32
, SI_PS_CONST_POLY_STIPPLE
, 0);
5181 desc
= ac_build_load_to_sgpr(&ctx
->ac
, param_rw_buffers
, slot
);
5183 /* The stipple pattern is 32x32, each row has 32 bits. */
5184 offset
= LLVMBuildMul(builder
, address
[1],
5185 LLVMConstInt(ctx
->i32
, 4, 0), "");
5186 row
= buffer_load_const(ctx
, desc
, offset
);
5187 row
= ac_to_integer(&ctx
->ac
, row
);
5188 bit
= LLVMBuildLShr(builder
, row
, address
[0], "");
5189 bit
= LLVMBuildTrunc(builder
, bit
, ctx
->i1
, "");
5190 ac_build_kill_if_false(&ctx
->ac
, bit
);
5193 void si_shader_binary_read_config(struct ac_shader_binary
*binary
,
5194 struct si_shader_config
*conf
,
5195 unsigned symbol_offset
)
5198 const unsigned char *config
=
5199 ac_shader_binary_config_start(binary
, symbol_offset
);
5200 bool really_needs_scratch
= false;
5202 /* LLVM adds SGPR spills to the scratch size.
5203 * Find out if we really need the scratch buffer.
5205 for (i
= 0; i
< binary
->reloc_count
; i
++) {
5206 const struct ac_shader_reloc
*reloc
= &binary
->relocs
[i
];
5208 if (!strcmp(scratch_rsrc_dword0_symbol
, reloc
->name
) ||
5209 !strcmp(scratch_rsrc_dword1_symbol
, reloc
->name
)) {
5210 really_needs_scratch
= true;
5215 /* XXX: We may be able to emit some of these values directly rather than
5216 * extracting fields to be emitted later.
5219 for (i
= 0; i
< binary
->config_size_per_symbol
; i
+= 8) {
5220 unsigned reg
= util_le32_to_cpu(*(uint32_t*)(config
+ i
));
5221 unsigned value
= util_le32_to_cpu(*(uint32_t*)(config
+ i
+ 4));
5223 case R_00B028_SPI_SHADER_PGM_RSRC1_PS
:
5224 case R_00B128_SPI_SHADER_PGM_RSRC1_VS
:
5225 case R_00B228_SPI_SHADER_PGM_RSRC1_GS
:
5226 case R_00B428_SPI_SHADER_PGM_RSRC1_HS
:
5227 case R_00B848_COMPUTE_PGM_RSRC1
:
5228 conf
->num_sgprs
= MAX2(conf
->num_sgprs
, (G_00B028_SGPRS(value
) + 1) * 8);
5229 conf
->num_vgprs
= MAX2(conf
->num_vgprs
, (G_00B028_VGPRS(value
) + 1) * 4);
5230 conf
->float_mode
= G_00B028_FLOAT_MODE(value
);
5231 conf
->rsrc1
= value
;
5233 case R_00B02C_SPI_SHADER_PGM_RSRC2_PS
:
5234 conf
->lds_size
= MAX2(conf
->lds_size
, G_00B02C_EXTRA_LDS_SIZE(value
));
5236 case R_00B84C_COMPUTE_PGM_RSRC2
:
5237 conf
->lds_size
= MAX2(conf
->lds_size
, G_00B84C_LDS_SIZE(value
));
5238 conf
->rsrc2
= value
;
5240 case R_0286CC_SPI_PS_INPUT_ENA
:
5241 conf
->spi_ps_input_ena
= value
;
5243 case R_0286D0_SPI_PS_INPUT_ADDR
:
5244 conf
->spi_ps_input_addr
= value
;
5246 case R_0286E8_SPI_TMPRING_SIZE
:
5247 case R_00B860_COMPUTE_TMPRING_SIZE
:
5248 /* WAVESIZE is in units of 256 dwords. */
5249 if (really_needs_scratch
)
5250 conf
->scratch_bytes_per_wave
=
5251 G_00B860_WAVESIZE(value
) * 256 * 4;
5253 case 0x4: /* SPILLED_SGPRS */
5254 conf
->spilled_sgprs
= value
;
5256 case 0x8: /* SPILLED_VGPRS */
5257 conf
->spilled_vgprs
= value
;
5261 static bool printed
;
5264 fprintf(stderr
, "Warning: LLVM emitted unknown "
5265 "config register: 0x%x\n", reg
);
5273 if (!conf
->spi_ps_input_addr
)
5274 conf
->spi_ps_input_addr
= conf
->spi_ps_input_ena
;
5277 void si_shader_apply_scratch_relocs(struct si_shader
*shader
,
5278 uint64_t scratch_va
)
5281 uint32_t scratch_rsrc_dword0
= scratch_va
;
5282 uint32_t scratch_rsrc_dword1
=
5283 S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32);
5285 /* Enable scratch coalescing. */
5286 scratch_rsrc_dword1
|= S_008F04_SWIZZLE_ENABLE(1);
5288 for (i
= 0 ; i
< shader
->binary
.reloc_count
; i
++) {
5289 const struct ac_shader_reloc
*reloc
=
5290 &shader
->binary
.relocs
[i
];
5291 if (!strcmp(scratch_rsrc_dword0_symbol
, reloc
->name
)) {
5292 util_memcpy_cpu_to_le32(shader
->binary
.code
+ reloc
->offset
,
5293 &scratch_rsrc_dword0
, 4);
5294 } else if (!strcmp(scratch_rsrc_dword1_symbol
, reloc
->name
)) {
5295 util_memcpy_cpu_to_le32(shader
->binary
.code
+ reloc
->offset
,
5296 &scratch_rsrc_dword1
, 4);
5301 static unsigned si_get_shader_binary_size(const struct si_shader
*shader
)
5303 unsigned size
= shader
->binary
.code_size
;
5306 size
+= shader
->prolog
->binary
.code_size
;
5307 if (shader
->previous_stage
)
5308 size
+= shader
->previous_stage
->binary
.code_size
;
5309 if (shader
->prolog2
)
5310 size
+= shader
->prolog2
->binary
.code_size
;
5312 size
+= shader
->epilog
->binary
.code_size
;
5316 int si_shader_binary_upload(struct si_screen
*sscreen
, struct si_shader
*shader
)
5318 const struct ac_shader_binary
*prolog
=
5319 shader
->prolog
? &shader
->prolog
->binary
: NULL
;
5320 const struct ac_shader_binary
*previous_stage
=
5321 shader
->previous_stage
? &shader
->previous_stage
->binary
: NULL
;
5322 const struct ac_shader_binary
*prolog2
=
5323 shader
->prolog2
? &shader
->prolog2
->binary
: NULL
;
5324 const struct ac_shader_binary
*epilog
=
5325 shader
->epilog
? &shader
->epilog
->binary
: NULL
;
5326 const struct ac_shader_binary
*mainb
= &shader
->binary
;
5327 unsigned bo_size
= si_get_shader_binary_size(shader
) +
5328 (!epilog
? mainb
->rodata_size
: 0);
5331 assert(!prolog
|| !prolog
->rodata_size
);
5332 assert(!previous_stage
|| !previous_stage
->rodata_size
);
5333 assert(!prolog2
|| !prolog2
->rodata_size
);
5334 assert((!prolog
&& !previous_stage
&& !prolog2
&& !epilog
) ||
5335 !mainb
->rodata_size
);
5336 assert(!epilog
|| !epilog
->rodata_size
);
5338 r600_resource_reference(&shader
->bo
, NULL
);
5339 shader
->bo
= (struct r600_resource
*)
5340 si_aligned_buffer_create(&sscreen
->b
,
5341 sscreen
->cpdma_prefetch_writes_memory
?
5342 0 : R600_RESOURCE_FLAG_READ_ONLY
,
5343 PIPE_USAGE_IMMUTABLE
,
5344 align(bo_size
, SI_CPDMA_ALIGNMENT
),
5350 ptr
= sscreen
->ws
->buffer_map(shader
->bo
->buf
, NULL
,
5351 PIPE_TRANSFER_READ_WRITE
|
5352 PIPE_TRANSFER_UNSYNCHRONIZED
);
5354 /* Don't use util_memcpy_cpu_to_le32. LLVM binaries are
5355 * endian-independent. */
5357 memcpy(ptr
, prolog
->code
, prolog
->code_size
);
5358 ptr
+= prolog
->code_size
;
5360 if (previous_stage
) {
5361 memcpy(ptr
, previous_stage
->code
, previous_stage
->code_size
);
5362 ptr
+= previous_stage
->code_size
;
5365 memcpy(ptr
, prolog2
->code
, prolog2
->code_size
);
5366 ptr
+= prolog2
->code_size
;
5369 memcpy(ptr
, mainb
->code
, mainb
->code_size
);
5370 ptr
+= mainb
->code_size
;
5373 memcpy(ptr
, epilog
->code
, epilog
->code_size
);
5374 else if (mainb
->rodata_size
> 0)
5375 memcpy(ptr
, mainb
->rodata
, mainb
->rodata_size
);
5377 sscreen
->ws
->buffer_unmap(shader
->bo
->buf
);
5381 static void si_shader_dump_disassembly(const struct ac_shader_binary
*binary
,
5382 struct pipe_debug_callback
*debug
,
5383 const char *name
, FILE *file
)
5388 if (binary
->disasm_string
) {
5389 fprintf(file
, "Shader %s disassembly:\n", name
);
5390 fprintf(file
, "%s", binary
->disasm_string
);
5392 if (debug
&& debug
->debug_message
) {
5393 /* Very long debug messages are cut off, so send the
5394 * disassembly one line at a time. This causes more
5395 * overhead, but on the plus side it simplifies
5396 * parsing of resulting logs.
5398 pipe_debug_message(debug
, SHADER_INFO
,
5399 "Shader Disassembly Begin");
5401 line
= binary
->disasm_string
;
5403 p
= util_strchrnul(line
, '\n');
5407 pipe_debug_message(debug
, SHADER_INFO
,
5408 "%.*s", count
, line
);
5416 pipe_debug_message(debug
, SHADER_INFO
,
5417 "Shader Disassembly End");
5420 fprintf(file
, "Shader %s binary:\n", name
);
5421 for (i
= 0; i
< binary
->code_size
; i
+= 4) {
5422 fprintf(file
, "@0x%x: %02x%02x%02x%02x\n", i
,
5423 binary
->code
[i
+ 3], binary
->code
[i
+ 2],
5424 binary
->code
[i
+ 1], binary
->code
[i
]);
5429 static void si_calculate_max_simd_waves(struct si_shader
*shader
)
5431 struct si_screen
*sscreen
= shader
->selector
->screen
;
5432 struct si_shader_config
*conf
= &shader
->config
;
5433 unsigned num_inputs
= shader
->selector
->info
.num_inputs
;
5434 unsigned lds_increment
= sscreen
->info
.chip_class
>= CIK
? 512 : 256;
5435 unsigned lds_per_wave
= 0;
5436 unsigned max_simd_waves
;
5438 switch (sscreen
->info
.family
) {
5439 /* These always have 8 waves: */
5440 case CHIP_POLARIS10
:
5441 case CHIP_POLARIS11
:
5442 case CHIP_POLARIS12
:
5446 max_simd_waves
= 10;
5449 /* Compute LDS usage for PS. */
5450 switch (shader
->selector
->type
) {
5451 case PIPE_SHADER_FRAGMENT
:
5452 /* The minimum usage per wave is (num_inputs * 48). The maximum
5453 * usage is (num_inputs * 48 * 16).
5454 * We can get anything in between and it varies between waves.
5456 * The 48 bytes per input for a single primitive is equal to
5457 * 4 bytes/component * 4 components/input * 3 points.
5459 * Other stages don't know the size at compile time or don't
5460 * allocate LDS per wave, but instead they do it per thread group.
5462 lds_per_wave
= conf
->lds_size
* lds_increment
+
5463 align(num_inputs
* 48, lds_increment
);
5465 case PIPE_SHADER_COMPUTE
:
5466 if (shader
->selector
) {
5467 unsigned max_workgroup_size
=
5468 si_get_max_workgroup_size(shader
);
5469 lds_per_wave
= (conf
->lds_size
* lds_increment
) /
5470 DIV_ROUND_UP(max_workgroup_size
, 64);
5475 /* Compute the per-SIMD wave counts. */
5476 if (conf
->num_sgprs
) {
5477 if (sscreen
->info
.chip_class
>= VI
)
5478 max_simd_waves
= MIN2(max_simd_waves
, 800 / conf
->num_sgprs
);
5480 max_simd_waves
= MIN2(max_simd_waves
, 512 / conf
->num_sgprs
);
5483 if (conf
->num_vgprs
)
5484 max_simd_waves
= MIN2(max_simd_waves
, 256 / conf
->num_vgprs
);
5486 /* LDS is 64KB per CU (4 SIMDs), which is 16KB per SIMD (usage above
5487 * 16KB makes some SIMDs unoccupied). */
5489 max_simd_waves
= MIN2(max_simd_waves
, 16384 / lds_per_wave
);
5491 conf
->max_simd_waves
= max_simd_waves
;
5494 void si_shader_dump_stats_for_shader_db(const struct si_shader
*shader
,
5495 struct pipe_debug_callback
*debug
)
5497 const struct si_shader_config
*conf
= &shader
->config
;
5499 pipe_debug_message(debug
, SHADER_INFO
,
5500 "Shader Stats: SGPRS: %d VGPRS: %d Code Size: %d "
5501 "LDS: %d Scratch: %d Max Waves: %d Spilled SGPRs: %d "
5502 "Spilled VGPRs: %d PrivMem VGPRs: %d",
5503 conf
->num_sgprs
, conf
->num_vgprs
,
5504 si_get_shader_binary_size(shader
),
5505 conf
->lds_size
, conf
->scratch_bytes_per_wave
,
5506 conf
->max_simd_waves
, conf
->spilled_sgprs
,
5507 conf
->spilled_vgprs
, conf
->private_mem_vgprs
);
5510 static void si_shader_dump_stats(struct si_screen
*sscreen
,
5511 const struct si_shader
*shader
,
5514 bool check_debug_option
)
5516 const struct si_shader_config
*conf
= &shader
->config
;
5518 if (!check_debug_option
||
5519 si_can_dump_shader(sscreen
, processor
)) {
5520 if (processor
== PIPE_SHADER_FRAGMENT
) {
5521 fprintf(file
, "*** SHADER CONFIG ***\n"
5522 "SPI_PS_INPUT_ADDR = 0x%04x\n"
5523 "SPI_PS_INPUT_ENA = 0x%04x\n",
5524 conf
->spi_ps_input_addr
, conf
->spi_ps_input_ena
);
5527 fprintf(file
, "*** SHADER STATS ***\n"
5530 "Spilled SGPRs: %d\n"
5531 "Spilled VGPRs: %d\n"
5532 "Private memory VGPRs: %d\n"
5533 "Code Size: %d bytes\n"
5535 "Scratch: %d bytes per wave\n"
5537 "********************\n\n\n",
5538 conf
->num_sgprs
, conf
->num_vgprs
,
5539 conf
->spilled_sgprs
, conf
->spilled_vgprs
,
5540 conf
->private_mem_vgprs
,
5541 si_get_shader_binary_size(shader
),
5542 conf
->lds_size
, conf
->scratch_bytes_per_wave
,
5543 conf
->max_simd_waves
);
5547 const char *si_get_shader_name(const struct si_shader
*shader
, unsigned processor
)
5549 switch (processor
) {
5550 case PIPE_SHADER_VERTEX
:
5551 if (shader
->key
.as_es
)
5552 return "Vertex Shader as ES";
5553 else if (shader
->key
.as_ls
)
5554 return "Vertex Shader as LS";
5556 return "Vertex Shader as VS";
5557 case PIPE_SHADER_TESS_CTRL
:
5558 return "Tessellation Control Shader";
5559 case PIPE_SHADER_TESS_EVAL
:
5560 if (shader
->key
.as_es
)
5561 return "Tessellation Evaluation Shader as ES";
5563 return "Tessellation Evaluation Shader as VS";
5564 case PIPE_SHADER_GEOMETRY
:
5565 if (shader
->is_gs_copy_shader
)
5566 return "GS Copy Shader as VS";
5568 return "Geometry Shader";
5569 case PIPE_SHADER_FRAGMENT
:
5570 return "Pixel Shader";
5571 case PIPE_SHADER_COMPUTE
:
5572 return "Compute Shader";
5574 return "Unknown Shader";
5578 void si_shader_dump(struct si_screen
*sscreen
, const struct si_shader
*shader
,
5579 struct pipe_debug_callback
*debug
, unsigned processor
,
5580 FILE *file
, bool check_debug_option
)
5582 if (!check_debug_option
||
5583 si_can_dump_shader(sscreen
, processor
))
5584 si_dump_shader_key(processor
, shader
, file
);
5586 if (!check_debug_option
&& shader
->binary
.llvm_ir_string
) {
5587 if (shader
->previous_stage
&&
5588 shader
->previous_stage
->binary
.llvm_ir_string
) {
5589 fprintf(file
, "\n%s - previous stage - LLVM IR:\n\n",
5590 si_get_shader_name(shader
, processor
));
5591 fprintf(file
, "%s\n", shader
->previous_stage
->binary
.llvm_ir_string
);
5594 fprintf(file
, "\n%s - main shader part - LLVM IR:\n\n",
5595 si_get_shader_name(shader
, processor
));
5596 fprintf(file
, "%s\n", shader
->binary
.llvm_ir_string
);
5599 if (!check_debug_option
||
5600 (si_can_dump_shader(sscreen
, processor
) &&
5601 !(sscreen
->debug_flags
& DBG(NO_ASM
)))) {
5602 fprintf(file
, "\n%s:\n", si_get_shader_name(shader
, processor
));
5605 si_shader_dump_disassembly(&shader
->prolog
->binary
,
5606 debug
, "prolog", file
);
5607 if (shader
->previous_stage
)
5608 si_shader_dump_disassembly(&shader
->previous_stage
->binary
,
5609 debug
, "previous stage", file
);
5610 if (shader
->prolog2
)
5611 si_shader_dump_disassembly(&shader
->prolog2
->binary
,
5612 debug
, "prolog2", file
);
5614 si_shader_dump_disassembly(&shader
->binary
, debug
, "main", file
);
5617 si_shader_dump_disassembly(&shader
->epilog
->binary
,
5618 debug
, "epilog", file
);
5619 fprintf(file
, "\n");
5622 si_shader_dump_stats(sscreen
, shader
, processor
, file
,
5623 check_debug_option
);
5626 static int si_compile_llvm(struct si_screen
*sscreen
,
5627 struct ac_shader_binary
*binary
,
5628 struct si_shader_config
*conf
,
5629 LLVMTargetMachineRef tm
,
5631 struct pipe_debug_callback
*debug
,
5636 unsigned count
= p_atomic_inc_return(&sscreen
->num_compilations
);
5638 if (si_can_dump_shader(sscreen
, processor
)) {
5639 fprintf(stderr
, "radeonsi: Compiling shader %d\n", count
);
5641 if (!(sscreen
->debug_flags
& (DBG(NO_IR
) | DBG(PREOPT_IR
)))) {
5642 fprintf(stderr
, "%s LLVM IR:\n\n", name
);
5643 ac_dump_module(mod
);
5644 fprintf(stderr
, "\n");
5648 if (sscreen
->record_llvm_ir
) {
5649 char *ir
= LLVMPrintModuleToString(mod
);
5650 binary
->llvm_ir_string
= strdup(ir
);
5651 LLVMDisposeMessage(ir
);
5654 if (!si_replace_shader(count
, binary
)) {
5655 r
= si_llvm_compile(mod
, binary
, tm
, debug
);
5660 si_shader_binary_read_config(binary
, conf
, 0);
5662 /* Enable 64-bit and 16-bit denormals, because there is no performance
5665 * If denormals are enabled, all floating-point output modifiers are
5668 * Don't enable denormals for 32-bit floats, because:
5669 * - Floating-point output modifiers would be ignored by the hw.
5670 * - Some opcodes don't support denormals, such as v_mad_f32. We would
5671 * have to stop using those.
5672 * - SI & CI would be very slow.
5674 conf
->float_mode
|= V_00B028_FP_64_DENORMS
;
5676 FREE(binary
->config
);
5677 FREE(binary
->global_symbol_offsets
);
5678 binary
->config
= NULL
;
5679 binary
->global_symbol_offsets
= NULL
;
5681 /* Some shaders can't have rodata because their binaries can be
5684 if (binary
->rodata_size
&&
5685 (processor
== PIPE_SHADER_VERTEX
||
5686 processor
== PIPE_SHADER_TESS_CTRL
||
5687 processor
== PIPE_SHADER_TESS_EVAL
||
5688 processor
== PIPE_SHADER_FRAGMENT
)) {
5689 fprintf(stderr
, "radeonsi: The shader can't have rodata.");
5696 static void si_llvm_build_ret(struct si_shader_context
*ctx
, LLVMValueRef ret
)
5698 if (LLVMGetTypeKind(LLVMTypeOf(ret
)) == LLVMVoidTypeKind
)
5699 LLVMBuildRetVoid(ctx
->ac
.builder
);
5701 LLVMBuildRet(ctx
->ac
.builder
, ret
);
5704 /* Generate code for the hardware VS shader stage to go with a geometry shader */
5706 si_generate_gs_copy_shader(struct si_screen
*sscreen
,
5707 LLVMTargetMachineRef tm
,
5708 struct si_shader_selector
*gs_selector
,
5709 struct pipe_debug_callback
*debug
)
5711 struct si_shader_context ctx
;
5712 struct si_shader
*shader
;
5713 LLVMBuilderRef builder
;
5714 struct lp_build_tgsi_context
*bld_base
= &ctx
.bld_base
;
5715 struct lp_build_context
*uint
= &bld_base
->uint_bld
;
5716 struct si_shader_output_values
*outputs
;
5717 struct tgsi_shader_info
*gsinfo
= &gs_selector
->info
;
5720 outputs
= MALLOC(gsinfo
->num_outputs
* sizeof(outputs
[0]));
5725 shader
= CALLOC_STRUCT(si_shader
);
5731 /* We can leave the fence as permanently signaled because the GS copy
5732 * shader only becomes visible globally after it has been compiled. */
5733 util_queue_fence_init(&shader
->ready
);
5735 shader
->selector
= gs_selector
;
5736 shader
->is_gs_copy_shader
= true;
5738 si_init_shader_ctx(&ctx
, sscreen
, tm
);
5739 ctx
.shader
= shader
;
5740 ctx
.type
= PIPE_SHADER_VERTEX
;
5742 builder
= ctx
.ac
.builder
;
5744 create_function(&ctx
);
5745 preload_ring_buffers(&ctx
);
5747 LLVMValueRef voffset
=
5748 lp_build_mul_imm(uint
, ctx
.abi
.vertex_id
, 4);
5750 /* Fetch the vertex stream ID.*/
5751 LLVMValueRef stream_id
;
5753 if (gs_selector
->so
.num_outputs
)
5754 stream_id
= unpack_param(&ctx
, ctx
.param_streamout_config
, 24, 2);
5756 stream_id
= ctx
.i32_0
;
5758 /* Fill in output information. */
5759 for (i
= 0; i
< gsinfo
->num_outputs
; ++i
) {
5760 outputs
[i
].semantic_name
= gsinfo
->output_semantic_name
[i
];
5761 outputs
[i
].semantic_index
= gsinfo
->output_semantic_index
[i
];
5763 for (int chan
= 0; chan
< 4; chan
++) {
5764 outputs
[i
].vertex_stream
[chan
] =
5765 (gsinfo
->output_streams
[i
] >> (2 * chan
)) & 3;
5769 LLVMBasicBlockRef end_bb
;
5770 LLVMValueRef switch_inst
;
5772 end_bb
= LLVMAppendBasicBlockInContext(ctx
.ac
.context
, ctx
.main_fn
, "end");
5773 switch_inst
= LLVMBuildSwitch(builder
, stream_id
, end_bb
, 4);
5775 for (int stream
= 0; stream
< 4; stream
++) {
5776 LLVMBasicBlockRef bb
;
5779 if (!gsinfo
->num_stream_output_components
[stream
])
5782 if (stream
> 0 && !gs_selector
->so
.num_outputs
)
5785 bb
= LLVMInsertBasicBlockInContext(ctx
.ac
.context
, end_bb
, "out");
5786 LLVMAddCase(switch_inst
, LLVMConstInt(ctx
.i32
, stream
, 0), bb
);
5787 LLVMPositionBuilderAtEnd(builder
, bb
);
5789 /* Fetch vertex data from GSVS ring */
5791 for (i
= 0; i
< gsinfo
->num_outputs
; ++i
) {
5792 for (unsigned chan
= 0; chan
< 4; chan
++) {
5793 if (!(gsinfo
->output_usagemask
[i
] & (1 << chan
)) ||
5794 outputs
[i
].vertex_stream
[chan
] != stream
) {
5795 outputs
[i
].values
[chan
] = ctx
.bld_base
.base
.undef
;
5799 LLVMValueRef soffset
= LLVMConstInt(ctx
.i32
,
5800 offset
* gs_selector
->gs_max_out_vertices
* 16 * 4, 0);
5803 outputs
[i
].values
[chan
] =
5804 ac_build_buffer_load(&ctx
.ac
,
5805 ctx
.gsvs_ring
[0], 1,
5812 /* Streamout and exports. */
5813 if (gs_selector
->so
.num_outputs
) {
5814 si_llvm_emit_streamout(&ctx
, outputs
,
5815 gsinfo
->num_outputs
,
5820 si_llvm_export_vs(&ctx
, outputs
, gsinfo
->num_outputs
);
5822 LLVMBuildBr(builder
, end_bb
);
5825 LLVMPositionBuilderAtEnd(builder
, end_bb
);
5827 LLVMBuildRetVoid(ctx
.ac
.builder
);
5829 ctx
.type
= PIPE_SHADER_GEOMETRY
; /* override for shader dumping */
5830 si_llvm_optimize_module(&ctx
);
5832 r
= si_compile_llvm(sscreen
, &ctx
.shader
->binary
,
5833 &ctx
.shader
->config
, ctx
.tm
,
5835 debug
, PIPE_SHADER_GEOMETRY
,
5838 if (si_can_dump_shader(sscreen
, PIPE_SHADER_GEOMETRY
))
5839 fprintf(stderr
, "GS Copy Shader:\n");
5840 si_shader_dump(sscreen
, ctx
.shader
, debug
,
5841 PIPE_SHADER_GEOMETRY
, stderr
, true);
5842 r
= si_shader_binary_upload(sscreen
, ctx
.shader
);
5845 si_llvm_dispose(&ctx
);
5856 static void si_dump_shader_key_vs(const struct si_shader_key
*key
,
5857 const struct si_vs_prolog_bits
*prolog
,
5858 const char *prefix
, FILE *f
)
5860 fprintf(f
, " %s.instance_divisor_is_one = %u\n",
5861 prefix
, prolog
->instance_divisor_is_one
);
5862 fprintf(f
, " %s.instance_divisor_is_fetched = %u\n",
5863 prefix
, prolog
->instance_divisor_is_fetched
);
5864 fprintf(f
, " %s.ls_vgpr_fix = %u\n",
5865 prefix
, prolog
->ls_vgpr_fix
);
5867 fprintf(f
, " mono.vs.fix_fetch = {");
5868 for (int i
= 0; i
< SI_MAX_ATTRIBS
; i
++)
5869 fprintf(f
, !i
? "%u" : ", %u", key
->mono
.vs_fix_fetch
[i
]);
5873 static void si_dump_shader_key(unsigned processor
, const struct si_shader
*shader
,
5876 const struct si_shader_key
*key
= &shader
->key
;
5878 fprintf(f
, "SHADER KEY\n");
5880 switch (processor
) {
5881 case PIPE_SHADER_VERTEX
:
5882 si_dump_shader_key_vs(key
, &key
->part
.vs
.prolog
,
5883 "part.vs.prolog", f
);
5884 fprintf(f
, " as_es = %u\n", key
->as_es
);
5885 fprintf(f
, " as_ls = %u\n", key
->as_ls
);
5886 fprintf(f
, " mono.u.vs_export_prim_id = %u\n",
5887 key
->mono
.u
.vs_export_prim_id
);
5890 case PIPE_SHADER_TESS_CTRL
:
5891 if (shader
->selector
->screen
->info
.chip_class
>= GFX9
) {
5892 si_dump_shader_key_vs(key
, &key
->part
.tcs
.ls_prolog
,
5893 "part.tcs.ls_prolog", f
);
5895 fprintf(f
, " part.tcs.epilog.prim_mode = %u\n", key
->part
.tcs
.epilog
.prim_mode
);
5896 fprintf(f
, " mono.u.ff_tcs_inputs_to_copy = 0x%"PRIx64
"\n", key
->mono
.u
.ff_tcs_inputs_to_copy
);
5899 case PIPE_SHADER_TESS_EVAL
:
5900 fprintf(f
, " as_es = %u\n", key
->as_es
);
5901 fprintf(f
, " mono.u.vs_export_prim_id = %u\n",
5902 key
->mono
.u
.vs_export_prim_id
);
5905 case PIPE_SHADER_GEOMETRY
:
5906 if (shader
->is_gs_copy_shader
)
5909 if (shader
->selector
->screen
->info
.chip_class
>= GFX9
&&
5910 key
->part
.gs
.es
->type
== PIPE_SHADER_VERTEX
) {
5911 si_dump_shader_key_vs(key
, &key
->part
.gs
.vs_prolog
,
5912 "part.gs.vs_prolog", f
);
5914 fprintf(f
, " part.gs.prolog.tri_strip_adj_fix = %u\n", key
->part
.gs
.prolog
.tri_strip_adj_fix
);
5917 case PIPE_SHADER_COMPUTE
:
5920 case PIPE_SHADER_FRAGMENT
:
5921 fprintf(f
, " part.ps.prolog.color_two_side = %u\n", key
->part
.ps
.prolog
.color_two_side
);
5922 fprintf(f
, " part.ps.prolog.flatshade_colors = %u\n", key
->part
.ps
.prolog
.flatshade_colors
);
5923 fprintf(f
, " part.ps.prolog.poly_stipple = %u\n", key
->part
.ps
.prolog
.poly_stipple
);
5924 fprintf(f
, " part.ps.prolog.force_persp_sample_interp = %u\n", key
->part
.ps
.prolog
.force_persp_sample_interp
);
5925 fprintf(f
, " part.ps.prolog.force_linear_sample_interp = %u\n", key
->part
.ps
.prolog
.force_linear_sample_interp
);
5926 fprintf(f
, " part.ps.prolog.force_persp_center_interp = %u\n", key
->part
.ps
.prolog
.force_persp_center_interp
);
5927 fprintf(f
, " part.ps.prolog.force_linear_center_interp = %u\n", key
->part
.ps
.prolog
.force_linear_center_interp
);
5928 fprintf(f
, " part.ps.prolog.bc_optimize_for_persp = %u\n", key
->part
.ps
.prolog
.bc_optimize_for_persp
);
5929 fprintf(f
, " part.ps.prolog.bc_optimize_for_linear = %u\n", key
->part
.ps
.prolog
.bc_optimize_for_linear
);
5930 fprintf(f
, " part.ps.epilog.spi_shader_col_format = 0x%x\n", key
->part
.ps
.epilog
.spi_shader_col_format
);
5931 fprintf(f
, " part.ps.epilog.color_is_int8 = 0x%X\n", key
->part
.ps
.epilog
.color_is_int8
);
5932 fprintf(f
, " part.ps.epilog.color_is_int10 = 0x%X\n", key
->part
.ps
.epilog
.color_is_int10
);
5933 fprintf(f
, " part.ps.epilog.last_cbuf = %u\n", key
->part
.ps
.epilog
.last_cbuf
);
5934 fprintf(f
, " part.ps.epilog.alpha_func = %u\n", key
->part
.ps
.epilog
.alpha_func
);
5935 fprintf(f
, " part.ps.epilog.alpha_to_one = %u\n", key
->part
.ps
.epilog
.alpha_to_one
);
5936 fprintf(f
, " part.ps.epilog.poly_line_smoothing = %u\n", key
->part
.ps
.epilog
.poly_line_smoothing
);
5937 fprintf(f
, " part.ps.epilog.clamp_color = %u\n", key
->part
.ps
.epilog
.clamp_color
);
5944 if ((processor
== PIPE_SHADER_GEOMETRY
||
5945 processor
== PIPE_SHADER_TESS_EVAL
||
5946 processor
== PIPE_SHADER_VERTEX
) &&
5947 !key
->as_es
&& !key
->as_ls
) {
5948 fprintf(f
, " opt.kill_outputs = 0x%"PRIx64
"\n", key
->opt
.kill_outputs
);
5949 fprintf(f
, " opt.clip_disable = %u\n", key
->opt
.clip_disable
);
5953 static void si_init_shader_ctx(struct si_shader_context
*ctx
,
5954 struct si_screen
*sscreen
,
5955 LLVMTargetMachineRef tm
)
5957 struct lp_build_tgsi_context
*bld_base
;
5959 si_llvm_context_init(ctx
, sscreen
, tm
);
5961 bld_base
= &ctx
->bld_base
;
5962 bld_base
->emit_fetch_funcs
[TGSI_FILE_CONSTANT
] = fetch_constant
;
5964 bld_base
->op_actions
[TGSI_OPCODE_INTERP_CENTROID
] = interp_action
;
5965 bld_base
->op_actions
[TGSI_OPCODE_INTERP_SAMPLE
] = interp_action
;
5966 bld_base
->op_actions
[TGSI_OPCODE_INTERP_OFFSET
] = interp_action
;
5968 bld_base
->op_actions
[TGSI_OPCODE_MEMBAR
].emit
= membar_emit
;
5970 bld_base
->op_actions
[TGSI_OPCODE_CLOCK
].emit
= clock_emit
;
5972 bld_base
->op_actions
[TGSI_OPCODE_DDX
].emit
= si_llvm_emit_ddxy
;
5973 bld_base
->op_actions
[TGSI_OPCODE_DDY
].emit
= si_llvm_emit_ddxy
;
5974 bld_base
->op_actions
[TGSI_OPCODE_DDX_FINE
].emit
= si_llvm_emit_ddxy
;
5975 bld_base
->op_actions
[TGSI_OPCODE_DDY_FINE
].emit
= si_llvm_emit_ddxy
;
5977 bld_base
->op_actions
[TGSI_OPCODE_VOTE_ALL
].emit
= vote_all_emit
;
5978 bld_base
->op_actions
[TGSI_OPCODE_VOTE_ANY
].emit
= vote_any_emit
;
5979 bld_base
->op_actions
[TGSI_OPCODE_VOTE_EQ
].emit
= vote_eq_emit
;
5980 bld_base
->op_actions
[TGSI_OPCODE_BALLOT
].emit
= ballot_emit
;
5981 bld_base
->op_actions
[TGSI_OPCODE_READ_FIRST
].intr_name
= "llvm.amdgcn.readfirstlane";
5982 bld_base
->op_actions
[TGSI_OPCODE_READ_FIRST
].emit
= read_lane_emit
;
5983 bld_base
->op_actions
[TGSI_OPCODE_READ_INVOC
].intr_name
= "llvm.amdgcn.readlane";
5984 bld_base
->op_actions
[TGSI_OPCODE_READ_INVOC
].fetch_args
= read_invoc_fetch_args
;
5985 bld_base
->op_actions
[TGSI_OPCODE_READ_INVOC
].emit
= read_lane_emit
;
5987 bld_base
->op_actions
[TGSI_OPCODE_EMIT
].emit
= si_tgsi_emit_vertex
;
5988 bld_base
->op_actions
[TGSI_OPCODE_ENDPRIM
].emit
= si_tgsi_emit_primitive
;
5989 bld_base
->op_actions
[TGSI_OPCODE_BARRIER
].emit
= si_llvm_emit_barrier
;
5992 static void si_optimize_vs_outputs(struct si_shader_context
*ctx
)
5994 struct si_shader
*shader
= ctx
->shader
;
5995 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
5997 if ((ctx
->type
!= PIPE_SHADER_VERTEX
&&
5998 ctx
->type
!= PIPE_SHADER_TESS_EVAL
) ||
5999 shader
->key
.as_ls
||
6003 ac_optimize_vs_outputs(&ctx
->ac
,
6005 shader
->info
.vs_output_param_offset
,
6007 &shader
->info
.nr_param_exports
);
6010 static void si_init_exec_from_input(struct si_shader_context
*ctx
,
6011 unsigned param
, unsigned bitoffset
)
6013 LLVMValueRef args
[] = {
6014 LLVMGetParam(ctx
->main_fn
, param
),
6015 LLVMConstInt(ctx
->i32
, bitoffset
, 0),
6017 lp_build_intrinsic(ctx
->ac
.builder
,
6018 "llvm.amdgcn.init.exec.from.input",
6019 ctx
->voidt
, args
, 2, LP_FUNC_ATTR_CONVERGENT
);
6022 static bool si_vs_needs_prolog(const struct si_shader_selector
*sel
,
6023 const struct si_vs_prolog_bits
*key
)
6025 /* VGPR initialization fixup for Vega10 and Raven is always done in the
6027 return sel
->vs_needs_prolog
|| key
->ls_vgpr_fix
;
6030 static bool si_compile_tgsi_main(struct si_shader_context
*ctx
,
6033 struct si_shader
*shader
= ctx
->shader
;
6034 struct si_shader_selector
*sel
= shader
->selector
;
6035 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
6037 // TODO clean all this up!
6038 switch (ctx
->type
) {
6039 case PIPE_SHADER_VERTEX
:
6040 ctx
->load_input
= declare_input_vs
;
6041 if (shader
->key
.as_ls
)
6042 ctx
->abi
.emit_outputs
= si_llvm_emit_ls_epilogue
;
6043 else if (shader
->key
.as_es
)
6044 ctx
->abi
.emit_outputs
= si_llvm_emit_es_epilogue
;
6046 ctx
->abi
.emit_outputs
= si_llvm_emit_vs_epilogue
;
6047 bld_base
->emit_epilogue
= si_tgsi_emit_epilogue
;
6048 ctx
->abi
.load_base_vertex
= get_base_vertex
;
6050 case PIPE_SHADER_TESS_CTRL
:
6051 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = fetch_input_tcs
;
6052 ctx
->abi
.load_tess_varyings
= si_nir_load_tcs_varyings
;
6053 bld_base
->emit_fetch_funcs
[TGSI_FILE_OUTPUT
] = fetch_output_tcs
;
6054 bld_base
->emit_store
= store_output_tcs
;
6055 ctx
->abi
.store_tcs_outputs
= si_nir_store_output_tcs
;
6056 ctx
->abi
.emit_outputs
= si_llvm_emit_tcs_epilogue
;
6057 ctx
->abi
.load_patch_vertices_in
= si_load_patch_vertices_in
;
6058 bld_base
->emit_epilogue
= si_tgsi_emit_epilogue
;
6060 case PIPE_SHADER_TESS_EVAL
:
6061 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = fetch_input_tes
;
6062 ctx
->abi
.load_tess_varyings
= si_nir_load_input_tes
;
6063 ctx
->abi
.load_tess_coord
= si_load_tess_coord
;
6064 ctx
->abi
.load_tess_level
= si_load_tess_level
;
6065 ctx
->abi
.load_patch_vertices_in
= si_load_patch_vertices_in
;
6066 if (shader
->key
.as_es
)
6067 ctx
->abi
.emit_outputs
= si_llvm_emit_es_epilogue
;
6069 ctx
->abi
.emit_outputs
= si_llvm_emit_vs_epilogue
;
6070 bld_base
->emit_epilogue
= si_tgsi_emit_epilogue
;
6072 case PIPE_SHADER_GEOMETRY
:
6073 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = fetch_input_gs
;
6074 ctx
->abi
.load_inputs
= si_nir_load_input_gs
;
6075 ctx
->abi
.emit_vertex
= si_llvm_emit_vertex
;
6076 ctx
->abi
.emit_primitive
= si_llvm_emit_primitive
;
6077 ctx
->abi
.emit_outputs
= si_llvm_emit_gs_epilogue
;
6078 bld_base
->emit_epilogue
= si_tgsi_emit_gs_epilogue
;
6080 case PIPE_SHADER_FRAGMENT
:
6081 ctx
->load_input
= declare_input_fs
;
6082 ctx
->abi
.emit_outputs
= si_llvm_return_fs_outputs
;
6083 bld_base
->emit_epilogue
= si_tgsi_emit_epilogue
;
6084 ctx
->abi
.lookup_interp_param
= si_nir_lookup_interp_param
;
6085 ctx
->abi
.load_sample_position
= load_sample_position
;
6086 ctx
->abi
.load_sample_mask_in
= load_sample_mask_in
;
6087 ctx
->abi
.emit_kill
= si_llvm_emit_kill
;
6089 case PIPE_SHADER_COMPUTE
:
6090 ctx
->abi
.load_local_group_size
= get_block_size
;
6093 assert(!"Unsupported shader type");
6097 ctx
->abi
.load_ubo
= load_ubo
;
6098 ctx
->abi
.load_ssbo
= load_ssbo
;
6100 create_function(ctx
);
6101 preload_ring_buffers(ctx
);
6103 /* For GFX9 merged shaders:
6104 * - Set EXEC for the first shader. If the prolog is present, set
6105 * EXEC there instead.
6106 * - Add a barrier before the second shader.
6107 * - In the second shader, reset EXEC to ~0 and wrap the main part in
6108 * an if-statement. This is required for correctness in geometry
6109 * shaders, to ensure that empty GS waves do not send GS_EMIT and
6112 * For monolithic merged shaders, the first shader is wrapped in an
6113 * if-block together with its prolog in si_build_wrapper_function.
6115 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
6116 if (!is_monolithic
&&
6117 sel
->info
.num_instructions
> 1 && /* not empty shader */
6118 (shader
->key
.as_es
|| shader
->key
.as_ls
) &&
6119 (ctx
->type
== PIPE_SHADER_TESS_EVAL
||
6120 (ctx
->type
== PIPE_SHADER_VERTEX
&&
6121 !si_vs_needs_prolog(sel
, &shader
->key
.part
.vs
.prolog
)))) {
6122 si_init_exec_from_input(ctx
,
6123 ctx
->param_merged_wave_info
, 0);
6124 } else if (ctx
->type
== PIPE_SHADER_TESS_CTRL
||
6125 ctx
->type
== PIPE_SHADER_GEOMETRY
) {
6127 ac_init_exec_full_mask(&ctx
->ac
);
6129 /* The barrier must execute for all shaders in a
6132 si_llvm_emit_barrier(NULL
, bld_base
, NULL
);
6134 LLVMValueRef num_threads
= unpack_param(ctx
, ctx
->param_merged_wave_info
, 8, 8);
6136 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntULT
,
6137 ac_get_thread_id(&ctx
->ac
), num_threads
, "");
6138 lp_build_if(&ctx
->merged_wrap_if_state
, &ctx
->gallivm
, ena
);
6142 if (ctx
->type
== PIPE_SHADER_TESS_CTRL
&&
6143 sel
->tcs_info
.tessfactors_are_def_in_all_invocs
) {
6144 for (unsigned i
= 0; i
< 6; i
++) {
6145 ctx
->invoc0_tess_factors
[i
] =
6146 lp_build_alloca_undef(&ctx
->gallivm
, ctx
->i32
, "");
6150 if (ctx
->type
== PIPE_SHADER_GEOMETRY
) {
6152 for (i
= 0; i
< 4; i
++) {
6153 ctx
->gs_next_vertex
[i
] =
6154 lp_build_alloca(&ctx
->gallivm
,
6159 if (sel
->force_correct_derivs_after_kill
) {
6160 ctx
->postponed_kill
= lp_build_alloca_undef(&ctx
->gallivm
, ctx
->i1
, "");
6161 /* true = don't kill. */
6162 LLVMBuildStore(ctx
->ac
.builder
, LLVMConstInt(ctx
->i1
, 1, 0),
6163 ctx
->postponed_kill
);
6167 if (!lp_build_tgsi_llvm(bld_base
, sel
->tokens
)) {
6168 fprintf(stderr
, "Failed to translate shader from TGSI to LLVM\n");
6172 if (!si_nir_build_llvm(ctx
, sel
->nir
)) {
6173 fprintf(stderr
, "Failed to translate shader from NIR to LLVM\n");
6178 si_llvm_build_ret(ctx
, ctx
->return_value
);
6183 * Compute the VS prolog key, which contains all the information needed to
6184 * build the VS prolog function, and set shader->info bits where needed.
6186 * \param info Shader info of the vertex shader.
6187 * \param num_input_sgprs Number of input SGPRs for the vertex shader.
6188 * \param prolog_key Key of the VS prolog
6189 * \param shader_out The vertex shader, or the next shader if merging LS+HS or ES+GS.
6190 * \param key Output shader part key.
6192 static void si_get_vs_prolog_key(const struct tgsi_shader_info
*info
,
6193 unsigned num_input_sgprs
,
6194 const struct si_vs_prolog_bits
*prolog_key
,
6195 struct si_shader
*shader_out
,
6196 union si_shader_part_key
*key
)
6198 memset(key
, 0, sizeof(*key
));
6199 key
->vs_prolog
.states
= *prolog_key
;
6200 key
->vs_prolog
.num_input_sgprs
= num_input_sgprs
;
6201 key
->vs_prolog
.last_input
= MAX2(1, info
->num_inputs
) - 1;
6202 key
->vs_prolog
.as_ls
= shader_out
->key
.as_ls
;
6203 key
->vs_prolog
.as_es
= shader_out
->key
.as_es
;
6205 if (shader_out
->selector
->type
== PIPE_SHADER_TESS_CTRL
) {
6206 key
->vs_prolog
.as_ls
= 1;
6207 key
->vs_prolog
.num_merged_next_stage_vgprs
= 2;
6208 } else if (shader_out
->selector
->type
== PIPE_SHADER_GEOMETRY
) {
6209 key
->vs_prolog
.as_es
= 1;
6210 key
->vs_prolog
.num_merged_next_stage_vgprs
= 5;
6213 /* Enable loading the InstanceID VGPR. */
6214 uint16_t input_mask
= u_bit_consecutive(0, info
->num_inputs
);
6216 if ((key
->vs_prolog
.states
.instance_divisor_is_one
|
6217 key
->vs_prolog
.states
.instance_divisor_is_fetched
) & input_mask
)
6218 shader_out
->info
.uses_instanceid
= true;
6222 * Compute the PS prolog key, which contains all the information needed to
6223 * build the PS prolog function, and set related bits in shader->config.
6225 static void si_get_ps_prolog_key(struct si_shader
*shader
,
6226 union si_shader_part_key
*key
,
6227 bool separate_prolog
)
6229 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
6231 memset(key
, 0, sizeof(*key
));
6232 key
->ps_prolog
.states
= shader
->key
.part
.ps
.prolog
;
6233 key
->ps_prolog
.colors_read
= info
->colors_read
;
6234 key
->ps_prolog
.num_input_sgprs
= shader
->info
.num_input_sgprs
;
6235 key
->ps_prolog
.num_input_vgprs
= shader
->info
.num_input_vgprs
;
6236 key
->ps_prolog
.wqm
= info
->uses_derivatives
&&
6237 (key
->ps_prolog
.colors_read
||
6238 key
->ps_prolog
.states
.force_persp_sample_interp
||
6239 key
->ps_prolog
.states
.force_linear_sample_interp
||
6240 key
->ps_prolog
.states
.force_persp_center_interp
||
6241 key
->ps_prolog
.states
.force_linear_center_interp
||
6242 key
->ps_prolog
.states
.bc_optimize_for_persp
||
6243 key
->ps_prolog
.states
.bc_optimize_for_linear
);
6244 key
->ps_prolog
.ancillary_vgpr_index
= shader
->info
.ancillary_vgpr_index
;
6246 if (info
->colors_read
) {
6247 unsigned *color
= shader
->selector
->color_attr_index
;
6249 if (shader
->key
.part
.ps
.prolog
.color_two_side
) {
6250 /* BCOLORs are stored after the last input. */
6251 key
->ps_prolog
.num_interp_inputs
= info
->num_inputs
;
6252 key
->ps_prolog
.face_vgpr_index
= shader
->info
.face_vgpr_index
;
6253 shader
->config
.spi_ps_input_ena
|= S_0286CC_FRONT_FACE_ENA(1);
6256 for (unsigned i
= 0; i
< 2; i
++) {
6257 unsigned interp
= info
->input_interpolate
[color
[i
]];
6258 unsigned location
= info
->input_interpolate_loc
[color
[i
]];
6260 if (!(info
->colors_read
& (0xf << i
*4)))
6263 key
->ps_prolog
.color_attr_index
[i
] = color
[i
];
6265 if (shader
->key
.part
.ps
.prolog
.flatshade_colors
&&
6266 interp
== TGSI_INTERPOLATE_COLOR
)
6267 interp
= TGSI_INTERPOLATE_CONSTANT
;
6270 case TGSI_INTERPOLATE_CONSTANT
:
6271 key
->ps_prolog
.color_interp_vgpr_index
[i
] = -1;
6273 case TGSI_INTERPOLATE_PERSPECTIVE
:
6274 case TGSI_INTERPOLATE_COLOR
:
6275 /* Force the interpolation location for colors here. */
6276 if (shader
->key
.part
.ps
.prolog
.force_persp_sample_interp
)
6277 location
= TGSI_INTERPOLATE_LOC_SAMPLE
;
6278 if (shader
->key
.part
.ps
.prolog
.force_persp_center_interp
)
6279 location
= TGSI_INTERPOLATE_LOC_CENTER
;
6282 case TGSI_INTERPOLATE_LOC_SAMPLE
:
6283 key
->ps_prolog
.color_interp_vgpr_index
[i
] = 0;
6284 shader
->config
.spi_ps_input_ena
|=
6285 S_0286CC_PERSP_SAMPLE_ENA(1);
6287 case TGSI_INTERPOLATE_LOC_CENTER
:
6288 key
->ps_prolog
.color_interp_vgpr_index
[i
] = 2;
6289 shader
->config
.spi_ps_input_ena
|=
6290 S_0286CC_PERSP_CENTER_ENA(1);
6292 case TGSI_INTERPOLATE_LOC_CENTROID
:
6293 key
->ps_prolog
.color_interp_vgpr_index
[i
] = 4;
6294 shader
->config
.spi_ps_input_ena
|=
6295 S_0286CC_PERSP_CENTROID_ENA(1);
6301 case TGSI_INTERPOLATE_LINEAR
:
6302 /* Force the interpolation location for colors here. */
6303 if (shader
->key
.part
.ps
.prolog
.force_linear_sample_interp
)
6304 location
= TGSI_INTERPOLATE_LOC_SAMPLE
;
6305 if (shader
->key
.part
.ps
.prolog
.force_linear_center_interp
)
6306 location
= TGSI_INTERPOLATE_LOC_CENTER
;
6308 /* The VGPR assignment for non-monolithic shaders
6309 * works because InitialPSInputAddr is set on the
6310 * main shader and PERSP_PULL_MODEL is never used.
6313 case TGSI_INTERPOLATE_LOC_SAMPLE
:
6314 key
->ps_prolog
.color_interp_vgpr_index
[i
] =
6315 separate_prolog
? 6 : 9;
6316 shader
->config
.spi_ps_input_ena
|=
6317 S_0286CC_LINEAR_SAMPLE_ENA(1);
6319 case TGSI_INTERPOLATE_LOC_CENTER
:
6320 key
->ps_prolog
.color_interp_vgpr_index
[i
] =
6321 separate_prolog
? 8 : 11;
6322 shader
->config
.spi_ps_input_ena
|=
6323 S_0286CC_LINEAR_CENTER_ENA(1);
6325 case TGSI_INTERPOLATE_LOC_CENTROID
:
6326 key
->ps_prolog
.color_interp_vgpr_index
[i
] =
6327 separate_prolog
? 10 : 13;
6328 shader
->config
.spi_ps_input_ena
|=
6329 S_0286CC_LINEAR_CENTROID_ENA(1);
6343 * Check whether a PS prolog is required based on the key.
6345 static bool si_need_ps_prolog(const union si_shader_part_key
*key
)
6347 return key
->ps_prolog
.colors_read
||
6348 key
->ps_prolog
.states
.force_persp_sample_interp
||
6349 key
->ps_prolog
.states
.force_linear_sample_interp
||
6350 key
->ps_prolog
.states
.force_persp_center_interp
||
6351 key
->ps_prolog
.states
.force_linear_center_interp
||
6352 key
->ps_prolog
.states
.bc_optimize_for_persp
||
6353 key
->ps_prolog
.states
.bc_optimize_for_linear
||
6354 key
->ps_prolog
.states
.poly_stipple
||
6355 key
->ps_prolog
.states
.samplemask_log_ps_iter
;
6359 * Compute the PS epilog key, which contains all the information needed to
6360 * build the PS epilog function.
6362 static void si_get_ps_epilog_key(struct si_shader
*shader
,
6363 union si_shader_part_key
*key
)
6365 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
6366 memset(key
, 0, sizeof(*key
));
6367 key
->ps_epilog
.colors_written
= info
->colors_written
;
6368 key
->ps_epilog
.writes_z
= info
->writes_z
;
6369 key
->ps_epilog
.writes_stencil
= info
->writes_stencil
;
6370 key
->ps_epilog
.writes_samplemask
= info
->writes_samplemask
;
6371 key
->ps_epilog
.states
= shader
->key
.part
.ps
.epilog
;
6375 * Build the GS prolog function. Rotate the input vertices for triangle strips
6378 static void si_build_gs_prolog_function(struct si_shader_context
*ctx
,
6379 union si_shader_part_key
*key
)
6381 unsigned num_sgprs
, num_vgprs
;
6382 struct si_function_info fninfo
;
6383 LLVMBuilderRef builder
= ctx
->ac
.builder
;
6384 LLVMTypeRef returns
[48];
6385 LLVMValueRef func
, ret
;
6387 si_init_function_info(&fninfo
);
6389 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
6390 if (key
->gs_prolog
.states
.gfx9_prev_is_vs
)
6391 num_sgprs
= 8 + GFX9_VSGS_NUM_USER_SGPR
;
6393 num_sgprs
= 8 + GFX9_TESGS_NUM_USER_SGPR
;
6394 num_vgprs
= 5; /* ES inputs are not needed by GS */
6396 num_sgprs
= GFX6_GS_NUM_USER_SGPR
+ 2;
6400 for (unsigned i
= 0; i
< num_sgprs
; ++i
) {
6401 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
6402 returns
[i
] = ctx
->i32
;
6405 for (unsigned i
= 0; i
< num_vgprs
; ++i
) {
6406 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
6407 returns
[num_sgprs
+ i
] = ctx
->f32
;
6410 /* Create the function. */
6411 si_create_function(ctx
, "gs_prolog", returns
, num_sgprs
+ num_vgprs
,
6413 func
= ctx
->main_fn
;
6415 /* Set the full EXEC mask for the prolog, because we are only fiddling
6416 * with registers here. The main shader part will set the correct EXEC
6419 if (ctx
->screen
->info
.chip_class
>= GFX9
&& !key
->gs_prolog
.is_monolithic
)
6420 ac_init_exec_full_mask(&ctx
->ac
);
6422 /* Copy inputs to outputs. This should be no-op, as the registers match,
6423 * but it will prevent the compiler from overwriting them unintentionally.
6425 ret
= ctx
->return_value
;
6426 for (unsigned i
= 0; i
< num_sgprs
; i
++) {
6427 LLVMValueRef p
= LLVMGetParam(func
, i
);
6428 ret
= LLVMBuildInsertValue(builder
, ret
, p
, i
, "");
6430 for (unsigned i
= 0; i
< num_vgprs
; i
++) {
6431 LLVMValueRef p
= LLVMGetParam(func
, num_sgprs
+ i
);
6432 p
= ac_to_float(&ctx
->ac
, p
);
6433 ret
= LLVMBuildInsertValue(builder
, ret
, p
, num_sgprs
+ i
, "");
6436 if (key
->gs_prolog
.states
.tri_strip_adj_fix
) {
6437 /* Remap the input vertices for every other primitive. */
6438 const unsigned gfx6_vtx_params
[6] = {
6446 const unsigned gfx9_vtx_params
[3] = {
6451 LLVMValueRef vtx_in
[6], vtx_out
[6];
6452 LLVMValueRef prim_id
, rotate
;
6454 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
6455 for (unsigned i
= 0; i
< 3; i
++) {
6456 vtx_in
[i
*2] = unpack_param(ctx
, gfx9_vtx_params
[i
], 0, 16);
6457 vtx_in
[i
*2+1] = unpack_param(ctx
, gfx9_vtx_params
[i
], 16, 16);
6460 for (unsigned i
= 0; i
< 6; i
++)
6461 vtx_in
[i
] = LLVMGetParam(func
, gfx6_vtx_params
[i
]);
6464 prim_id
= LLVMGetParam(func
, num_sgprs
+ 2);
6465 rotate
= LLVMBuildTrunc(builder
, prim_id
, ctx
->i1
, "");
6467 for (unsigned i
= 0; i
< 6; ++i
) {
6468 LLVMValueRef base
, rotated
;
6470 rotated
= vtx_in
[(i
+ 4) % 6];
6471 vtx_out
[i
] = LLVMBuildSelect(builder
, rotate
, rotated
, base
, "");
6474 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
6475 for (unsigned i
= 0; i
< 3; i
++) {
6476 LLVMValueRef hi
, out
;
6478 hi
= LLVMBuildShl(builder
, vtx_out
[i
*2+1],
6479 LLVMConstInt(ctx
->i32
, 16, 0), "");
6480 out
= LLVMBuildOr(builder
, vtx_out
[i
*2], hi
, "");
6481 out
= ac_to_float(&ctx
->ac
, out
);
6482 ret
= LLVMBuildInsertValue(builder
, ret
, out
,
6483 gfx9_vtx_params
[i
], "");
6486 for (unsigned i
= 0; i
< 6; i
++) {
6489 out
= ac_to_float(&ctx
->ac
, vtx_out
[i
]);
6490 ret
= LLVMBuildInsertValue(builder
, ret
, out
,
6491 gfx6_vtx_params
[i
], "");
6496 LLVMBuildRet(builder
, ret
);
6500 * Given a list of shader part functions, build a wrapper function that
6501 * runs them in sequence to form a monolithic shader.
6503 static void si_build_wrapper_function(struct si_shader_context
*ctx
,
6504 LLVMValueRef
*parts
,
6507 unsigned next_shader_first_part
)
6509 LLVMBuilderRef builder
= ctx
->ac
.builder
;
6510 /* PS epilog has one arg per color component; gfx9 merged shader
6511 * prologs need to forward 32 user SGPRs.
6513 struct si_function_info fninfo
;
6514 LLVMValueRef initial
[64], out
[64];
6515 LLVMTypeRef function_type
;
6516 unsigned num_first_params
;
6517 unsigned num_out
, initial_num_out
;
6518 MAYBE_UNUSED
unsigned num_out_sgpr
; /* used in debug checks */
6519 MAYBE_UNUSED
unsigned initial_num_out_sgpr
; /* used in debug checks */
6520 unsigned num_sgprs
, num_vgprs
;
6522 struct lp_build_if_state if_state
;
6524 si_init_function_info(&fninfo
);
6526 for (unsigned i
= 0; i
< num_parts
; ++i
) {
6527 lp_add_function_attr(parts
[i
], -1, LP_FUNC_ATTR_ALWAYSINLINE
);
6528 LLVMSetLinkage(parts
[i
], LLVMPrivateLinkage
);
6531 /* The parameters of the wrapper function correspond to those of the
6532 * first part in terms of SGPRs and VGPRs, but we use the types of the
6533 * main part to get the right types. This is relevant for the
6534 * dereferenceable attribute on descriptor table pointers.
6539 function_type
= LLVMGetElementType(LLVMTypeOf(parts
[0]));
6540 num_first_params
= LLVMCountParamTypes(function_type
);
6542 for (unsigned i
= 0; i
< num_first_params
; ++i
) {
6543 LLVMValueRef param
= LLVMGetParam(parts
[0], i
);
6545 if (ac_is_sgpr_param(param
)) {
6546 assert(num_vgprs
== 0);
6547 num_sgprs
+= ac_get_type_size(LLVMTypeOf(param
)) / 4;
6549 num_vgprs
+= ac_get_type_size(LLVMTypeOf(param
)) / 4;
6554 while (gprs
< num_sgprs
+ num_vgprs
) {
6555 LLVMValueRef param
= LLVMGetParam(parts
[main_part
], fninfo
.num_params
);
6556 LLVMTypeRef type
= LLVMTypeOf(param
);
6557 unsigned size
= ac_get_type_size(type
) / 4;
6559 add_arg(&fninfo
, gprs
< num_sgprs
? ARG_SGPR
: ARG_VGPR
, type
);
6561 assert(ac_is_sgpr_param(param
) == (gprs
< num_sgprs
));
6562 assert(gprs
+ size
<= num_sgprs
+ num_vgprs
&&
6563 (gprs
>= num_sgprs
|| gprs
+ size
<= num_sgprs
));
6568 si_create_function(ctx
, "wrapper", NULL
, 0, &fninfo
,
6569 si_get_max_workgroup_size(ctx
->shader
));
6571 if (is_merged_shader(ctx
->shader
))
6572 ac_init_exec_full_mask(&ctx
->ac
);
6574 /* Record the arguments of the function as if they were an output of
6580 for (unsigned i
= 0; i
< fninfo
.num_params
; ++i
) {
6581 LLVMValueRef param
= LLVMGetParam(ctx
->main_fn
, i
);
6582 LLVMTypeRef param_type
= LLVMTypeOf(param
);
6583 LLVMTypeRef out_type
= i
< fninfo
.num_sgpr_params
? ctx
->i32
: ctx
->f32
;
6584 unsigned size
= ac_get_type_size(param_type
) / 4;
6587 if (LLVMGetTypeKind(param_type
) == LLVMPointerTypeKind
) {
6588 param
= LLVMBuildPtrToInt(builder
, param
, ctx
->i32
, "");
6589 param_type
= ctx
->i32
;
6592 if (param_type
!= out_type
)
6593 param
= LLVMBuildBitCast(builder
, param
, out_type
, "");
6594 out
[num_out
++] = param
;
6596 LLVMTypeRef vector_type
= LLVMVectorType(out_type
, size
);
6598 if (LLVMGetTypeKind(param_type
) == LLVMPointerTypeKind
) {
6599 param
= LLVMBuildPtrToInt(builder
, param
, ctx
->i64
, "");
6600 param_type
= ctx
->i64
;
6603 if (param_type
!= vector_type
)
6604 param
= LLVMBuildBitCast(builder
, param
, vector_type
, "");
6606 for (unsigned j
= 0; j
< size
; ++j
)
6607 out
[num_out
++] = LLVMBuildExtractElement(
6608 builder
, param
, LLVMConstInt(ctx
->i32
, j
, 0), "");
6611 if (i
< fninfo
.num_sgpr_params
)
6612 num_out_sgpr
= num_out
;
6615 memcpy(initial
, out
, sizeof(out
));
6616 initial_num_out
= num_out
;
6617 initial_num_out_sgpr
= num_out_sgpr
;
6619 /* Now chain the parts. */
6620 for (unsigned part
= 0; part
< num_parts
; ++part
) {
6621 LLVMValueRef in
[48];
6623 LLVMTypeRef ret_type
;
6624 unsigned out_idx
= 0;
6625 unsigned num_params
= LLVMCountParams(parts
[part
]);
6627 /* Merged shaders are executed conditionally depending
6628 * on the number of enabled threads passed in the input SGPRs. */
6629 if (is_merged_shader(ctx
->shader
) && part
== 0) {
6630 LLVMValueRef ena
, count
= initial
[3];
6632 count
= LLVMBuildAnd(builder
, count
,
6633 LLVMConstInt(ctx
->i32
, 0x7f, 0), "");
6634 ena
= LLVMBuildICmp(builder
, LLVMIntULT
,
6635 ac_get_thread_id(&ctx
->ac
), count
, "");
6636 lp_build_if(&if_state
, &ctx
->gallivm
, ena
);
6639 /* Derive arguments for the next part from outputs of the
6642 for (unsigned param_idx
= 0; param_idx
< num_params
; ++param_idx
) {
6644 LLVMTypeRef param_type
;
6646 unsigned param_size
;
6647 LLVMValueRef arg
= NULL
;
6649 param
= LLVMGetParam(parts
[part
], param_idx
);
6650 param_type
= LLVMTypeOf(param
);
6651 param_size
= ac_get_type_size(param_type
) / 4;
6652 is_sgpr
= ac_is_sgpr_param(param
);
6655 lp_add_function_attr(parts
[part
], param_idx
+ 1, LP_FUNC_ATTR_INREG
);
6656 else if (out_idx
< num_out_sgpr
) {
6657 /* Skip returned SGPRs the current part doesn't
6658 * declare on the input. */
6659 out_idx
= num_out_sgpr
;
6662 assert(out_idx
+ param_size
<= (is_sgpr
? num_out_sgpr
: num_out
));
6664 if (param_size
== 1)
6667 arg
= lp_build_gather_values(&ctx
->gallivm
, &out
[out_idx
], param_size
);
6669 if (LLVMTypeOf(arg
) != param_type
) {
6670 if (LLVMGetTypeKind(param_type
) == LLVMPointerTypeKind
) {
6671 if (LLVMGetPointerAddressSpace(param_type
) ==
6672 AC_CONST_32BIT_ADDR_SPACE
) {
6673 arg
= LLVMBuildBitCast(builder
, arg
, ctx
->i32
, "");
6674 arg
= LLVMBuildIntToPtr(builder
, arg
, param_type
, "");
6676 arg
= LLVMBuildBitCast(builder
, arg
, ctx
->i64
, "");
6677 arg
= LLVMBuildIntToPtr(builder
, arg
, param_type
, "");
6680 arg
= LLVMBuildBitCast(builder
, arg
, param_type
, "");
6684 in
[param_idx
] = arg
;
6685 out_idx
+= param_size
;
6688 ret
= LLVMBuildCall(builder
, parts
[part
], in
, num_params
, "");
6690 if (is_merged_shader(ctx
->shader
) &&
6691 part
+ 1 == next_shader_first_part
) {
6692 lp_build_endif(&if_state
);
6694 /* The second half of the merged shader should use
6695 * the inputs from the toplevel (wrapper) function,
6696 * not the return value from the last call.
6698 * That's because the last call was executed condi-
6699 * tionally, so we can't consume it in the main
6702 memcpy(out
, initial
, sizeof(initial
));
6703 num_out
= initial_num_out
;
6704 num_out_sgpr
= initial_num_out_sgpr
;
6708 /* Extract the returned GPRs. */
6709 ret_type
= LLVMTypeOf(ret
);
6713 if (LLVMGetTypeKind(ret_type
) != LLVMVoidTypeKind
) {
6714 assert(LLVMGetTypeKind(ret_type
) == LLVMStructTypeKind
);
6716 unsigned ret_size
= LLVMCountStructElementTypes(ret_type
);
6718 for (unsigned i
= 0; i
< ret_size
; ++i
) {
6720 LLVMBuildExtractValue(builder
, ret
, i
, "");
6722 assert(num_out
< ARRAY_SIZE(out
));
6723 out
[num_out
++] = val
;
6725 if (LLVMTypeOf(val
) == ctx
->i32
) {
6726 assert(num_out_sgpr
+ 1 == num_out
);
6727 num_out_sgpr
= num_out
;
6733 LLVMBuildRetVoid(builder
);
6736 int si_compile_tgsi_shader(struct si_screen
*sscreen
,
6737 LLVMTargetMachineRef tm
,
6738 struct si_shader
*shader
,
6740 struct pipe_debug_callback
*debug
)
6742 struct si_shader_selector
*sel
= shader
->selector
;
6743 struct si_shader_context ctx
;
6746 /* Dump TGSI code before doing TGSI->LLVM conversion in case the
6747 * conversion fails. */
6748 if (si_can_dump_shader(sscreen
, sel
->info
.processor
) &&
6749 !(sscreen
->debug_flags
& DBG(NO_TGSI
))) {
6751 tgsi_dump(sel
->tokens
, 0);
6753 nir_print_shader(sel
->nir
, stderr
);
6754 si_dump_streamout(&sel
->so
);
6757 si_init_shader_ctx(&ctx
, sscreen
, tm
);
6758 si_llvm_context_set_tgsi(&ctx
, shader
);
6759 ctx
.separate_prolog
= !is_monolithic
;
6761 memset(shader
->info
.vs_output_param_offset
, AC_EXP_PARAM_UNDEFINED
,
6762 sizeof(shader
->info
.vs_output_param_offset
));
6764 shader
->info
.uses_instanceid
= sel
->info
.uses_instanceid
;
6766 if (!si_compile_tgsi_main(&ctx
, is_monolithic
)) {
6767 si_llvm_dispose(&ctx
);
6771 if (is_monolithic
&& ctx
.type
== PIPE_SHADER_VERTEX
) {
6772 LLVMValueRef parts
[2];
6773 bool need_prolog
= sel
->vs_needs_prolog
;
6775 parts
[1] = ctx
.main_fn
;
6778 union si_shader_part_key prolog_key
;
6779 si_get_vs_prolog_key(&sel
->info
,
6780 shader
->info
.num_input_sgprs
,
6781 &shader
->key
.part
.vs
.prolog
,
6782 shader
, &prolog_key
);
6783 si_build_vs_prolog_function(&ctx
, &prolog_key
);
6784 parts
[0] = ctx
.main_fn
;
6787 si_build_wrapper_function(&ctx
, parts
+ !need_prolog
,
6788 1 + need_prolog
, need_prolog
, 0);
6789 } else if (is_monolithic
&& ctx
.type
== PIPE_SHADER_TESS_CTRL
) {
6790 if (sscreen
->info
.chip_class
>= GFX9
) {
6791 struct si_shader_selector
*ls
= shader
->key
.part
.tcs
.ls
;
6792 LLVMValueRef parts
[4];
6793 bool vs_needs_prolog
=
6794 si_vs_needs_prolog(ls
, &shader
->key
.part
.tcs
.ls_prolog
);
6797 parts
[2] = ctx
.main_fn
;
6800 union si_shader_part_key tcs_epilog_key
;
6801 memset(&tcs_epilog_key
, 0, sizeof(tcs_epilog_key
));
6802 tcs_epilog_key
.tcs_epilog
.states
= shader
->key
.part
.tcs
.epilog
;
6803 si_build_tcs_epilog_function(&ctx
, &tcs_epilog_key
);
6804 parts
[3] = ctx
.main_fn
;
6806 /* VS as LS main part */
6807 struct si_shader shader_ls
= {};
6808 shader_ls
.selector
= ls
;
6809 shader_ls
.key
.as_ls
= 1;
6810 shader_ls
.key
.mono
= shader
->key
.mono
;
6811 shader_ls
.key
.opt
= shader
->key
.opt
;
6812 si_llvm_context_set_tgsi(&ctx
, &shader_ls
);
6814 if (!si_compile_tgsi_main(&ctx
, true)) {
6815 si_llvm_dispose(&ctx
);
6818 shader
->info
.uses_instanceid
|= ls
->info
.uses_instanceid
;
6819 parts
[1] = ctx
.main_fn
;
6822 if (vs_needs_prolog
) {
6823 union si_shader_part_key vs_prolog_key
;
6824 si_get_vs_prolog_key(&ls
->info
,
6825 shader_ls
.info
.num_input_sgprs
,
6826 &shader
->key
.part
.tcs
.ls_prolog
,
6827 shader
, &vs_prolog_key
);
6828 vs_prolog_key
.vs_prolog
.is_monolithic
= true;
6829 si_build_vs_prolog_function(&ctx
, &vs_prolog_key
);
6830 parts
[0] = ctx
.main_fn
;
6833 /* Reset the shader context. */
6834 ctx
.shader
= shader
;
6835 ctx
.type
= PIPE_SHADER_TESS_CTRL
;
6837 si_build_wrapper_function(&ctx
,
6838 parts
+ !vs_needs_prolog
,
6839 4 - !vs_needs_prolog
, 0,
6840 vs_needs_prolog
? 2 : 1);
6842 LLVMValueRef parts
[2];
6843 union si_shader_part_key epilog_key
;
6845 parts
[0] = ctx
.main_fn
;
6847 memset(&epilog_key
, 0, sizeof(epilog_key
));
6848 epilog_key
.tcs_epilog
.states
= shader
->key
.part
.tcs
.epilog
;
6849 si_build_tcs_epilog_function(&ctx
, &epilog_key
);
6850 parts
[1] = ctx
.main_fn
;
6852 si_build_wrapper_function(&ctx
, parts
, 2, 0, 0);
6854 } else if (is_monolithic
&& ctx
.type
== PIPE_SHADER_GEOMETRY
) {
6855 if (ctx
.screen
->info
.chip_class
>= GFX9
) {
6856 struct si_shader_selector
*es
= shader
->key
.part
.gs
.es
;
6857 LLVMValueRef es_prolog
= NULL
;
6858 LLVMValueRef es_main
= NULL
;
6859 LLVMValueRef gs_prolog
= NULL
;
6860 LLVMValueRef gs_main
= ctx
.main_fn
;
6863 union si_shader_part_key gs_prolog_key
;
6864 memset(&gs_prolog_key
, 0, sizeof(gs_prolog_key
));
6865 gs_prolog_key
.gs_prolog
.states
= shader
->key
.part
.gs
.prolog
;
6866 gs_prolog_key
.gs_prolog
.is_monolithic
= true;
6867 si_build_gs_prolog_function(&ctx
, &gs_prolog_key
);
6868 gs_prolog
= ctx
.main_fn
;
6871 struct si_shader shader_es
= {};
6872 shader_es
.selector
= es
;
6873 shader_es
.key
.as_es
= 1;
6874 shader_es
.key
.mono
= shader
->key
.mono
;
6875 shader_es
.key
.opt
= shader
->key
.opt
;
6876 si_llvm_context_set_tgsi(&ctx
, &shader_es
);
6878 if (!si_compile_tgsi_main(&ctx
, true)) {
6879 si_llvm_dispose(&ctx
);
6882 shader
->info
.uses_instanceid
|= es
->info
.uses_instanceid
;
6883 es_main
= ctx
.main_fn
;
6886 if (es
->vs_needs_prolog
) {
6887 union si_shader_part_key vs_prolog_key
;
6888 si_get_vs_prolog_key(&es
->info
,
6889 shader_es
.info
.num_input_sgprs
,
6890 &shader
->key
.part
.gs
.vs_prolog
,
6891 shader
, &vs_prolog_key
);
6892 vs_prolog_key
.vs_prolog
.is_monolithic
= true;
6893 si_build_vs_prolog_function(&ctx
, &vs_prolog_key
);
6894 es_prolog
= ctx
.main_fn
;
6897 /* Reset the shader context. */
6898 ctx
.shader
= shader
;
6899 ctx
.type
= PIPE_SHADER_GEOMETRY
;
6901 /* Prepare the array of shader parts. */
6902 LLVMValueRef parts
[4];
6903 unsigned num_parts
= 0, main_part
, next_first_part
;
6906 parts
[num_parts
++] = es_prolog
;
6908 parts
[main_part
= num_parts
++] = es_main
;
6909 parts
[next_first_part
= num_parts
++] = gs_prolog
;
6910 parts
[num_parts
++] = gs_main
;
6912 si_build_wrapper_function(&ctx
, parts
, num_parts
,
6913 main_part
, next_first_part
);
6915 LLVMValueRef parts
[2];
6916 union si_shader_part_key prolog_key
;
6918 parts
[1] = ctx
.main_fn
;
6920 memset(&prolog_key
, 0, sizeof(prolog_key
));
6921 prolog_key
.gs_prolog
.states
= shader
->key
.part
.gs
.prolog
;
6922 si_build_gs_prolog_function(&ctx
, &prolog_key
);
6923 parts
[0] = ctx
.main_fn
;
6925 si_build_wrapper_function(&ctx
, parts
, 2, 1, 0);
6927 } else if (is_monolithic
&& ctx
.type
== PIPE_SHADER_FRAGMENT
) {
6928 LLVMValueRef parts
[3];
6929 union si_shader_part_key prolog_key
;
6930 union si_shader_part_key epilog_key
;
6933 si_get_ps_prolog_key(shader
, &prolog_key
, false);
6934 need_prolog
= si_need_ps_prolog(&prolog_key
);
6936 parts
[need_prolog
? 1 : 0] = ctx
.main_fn
;
6939 si_build_ps_prolog_function(&ctx
, &prolog_key
);
6940 parts
[0] = ctx
.main_fn
;
6943 si_get_ps_epilog_key(shader
, &epilog_key
);
6944 si_build_ps_epilog_function(&ctx
, &epilog_key
);
6945 parts
[need_prolog
? 2 : 1] = ctx
.main_fn
;
6947 si_build_wrapper_function(&ctx
, parts
, need_prolog
? 3 : 2,
6948 need_prolog
? 1 : 0, 0);
6951 si_llvm_optimize_module(&ctx
);
6953 /* Post-optimization transformations and analysis. */
6954 si_optimize_vs_outputs(&ctx
);
6956 if ((debug
&& debug
->debug_message
) ||
6957 si_can_dump_shader(sscreen
, ctx
.type
)) {
6958 ctx
.shader
->config
.private_mem_vgprs
=
6959 ac_count_scratch_private_memory(ctx
.main_fn
);
6962 /* Compile to bytecode. */
6963 r
= si_compile_llvm(sscreen
, &shader
->binary
, &shader
->config
, tm
,
6964 ctx
.gallivm
.module
, debug
, ctx
.type
, "TGSI shader");
6965 si_llvm_dispose(&ctx
);
6967 fprintf(stderr
, "LLVM failed to compile shader\n");
6971 /* Validate SGPR and VGPR usage for compute to detect compiler bugs.
6972 * LLVM 3.9svn has this bug.
6974 if (sel
->type
== PIPE_SHADER_COMPUTE
) {
6975 unsigned wave_size
= 64;
6976 unsigned max_vgprs
= 256;
6977 unsigned max_sgprs
= sscreen
->info
.chip_class
>= VI
? 800 : 512;
6978 unsigned max_sgprs_per_wave
= 128;
6979 unsigned max_block_threads
= si_get_max_workgroup_size(shader
);
6980 unsigned min_waves_per_cu
= DIV_ROUND_UP(max_block_threads
, wave_size
);
6981 unsigned min_waves_per_simd
= DIV_ROUND_UP(min_waves_per_cu
, 4);
6983 max_vgprs
= max_vgprs
/ min_waves_per_simd
;
6984 max_sgprs
= MIN2(max_sgprs
/ min_waves_per_simd
, max_sgprs_per_wave
);
6986 if (shader
->config
.num_sgprs
> max_sgprs
||
6987 shader
->config
.num_vgprs
> max_vgprs
) {
6988 fprintf(stderr
, "LLVM failed to compile a shader correctly: "
6989 "SGPR:VGPR usage is %u:%u, but the hw limit is %u:%u\n",
6990 shader
->config
.num_sgprs
, shader
->config
.num_vgprs
,
6991 max_sgprs
, max_vgprs
);
6993 /* Just terminate the process, because dependent
6994 * shaders can hang due to bad input data, but use
6995 * the env var to allow shader-db to work.
6997 if (!debug_get_bool_option("SI_PASS_BAD_SHADERS", false))
7002 /* Add the scratch offset to input SGPRs. */
7003 if (shader
->config
.scratch_bytes_per_wave
&& !is_merged_shader(shader
))
7004 shader
->info
.num_input_sgprs
+= 1; /* scratch byte offset */
7006 /* Calculate the number of fragment input VGPRs. */
7007 if (ctx
.type
== PIPE_SHADER_FRAGMENT
) {
7008 shader
->info
.num_input_vgprs
= 0;
7009 shader
->info
.face_vgpr_index
= -1;
7010 shader
->info
.ancillary_vgpr_index
= -1;
7012 if (G_0286CC_PERSP_SAMPLE_ENA(shader
->config
.spi_ps_input_addr
))
7013 shader
->info
.num_input_vgprs
+= 2;
7014 if (G_0286CC_PERSP_CENTER_ENA(shader
->config
.spi_ps_input_addr
))
7015 shader
->info
.num_input_vgprs
+= 2;
7016 if (G_0286CC_PERSP_CENTROID_ENA(shader
->config
.spi_ps_input_addr
))
7017 shader
->info
.num_input_vgprs
+= 2;
7018 if (G_0286CC_PERSP_PULL_MODEL_ENA(shader
->config
.spi_ps_input_addr
))
7019 shader
->info
.num_input_vgprs
+= 3;
7020 if (G_0286CC_LINEAR_SAMPLE_ENA(shader
->config
.spi_ps_input_addr
))
7021 shader
->info
.num_input_vgprs
+= 2;
7022 if (G_0286CC_LINEAR_CENTER_ENA(shader
->config
.spi_ps_input_addr
))
7023 shader
->info
.num_input_vgprs
+= 2;
7024 if (G_0286CC_LINEAR_CENTROID_ENA(shader
->config
.spi_ps_input_addr
))
7025 shader
->info
.num_input_vgprs
+= 2;
7026 if (G_0286CC_LINE_STIPPLE_TEX_ENA(shader
->config
.spi_ps_input_addr
))
7027 shader
->info
.num_input_vgprs
+= 1;
7028 if (G_0286CC_POS_X_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
7029 shader
->info
.num_input_vgprs
+= 1;
7030 if (G_0286CC_POS_Y_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
7031 shader
->info
.num_input_vgprs
+= 1;
7032 if (G_0286CC_POS_Z_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
7033 shader
->info
.num_input_vgprs
+= 1;
7034 if (G_0286CC_POS_W_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
7035 shader
->info
.num_input_vgprs
+= 1;
7036 if (G_0286CC_FRONT_FACE_ENA(shader
->config
.spi_ps_input_addr
)) {
7037 shader
->info
.face_vgpr_index
= shader
->info
.num_input_vgprs
;
7038 shader
->info
.num_input_vgprs
+= 1;
7040 if (G_0286CC_ANCILLARY_ENA(shader
->config
.spi_ps_input_addr
)) {
7041 shader
->info
.ancillary_vgpr_index
= shader
->info
.num_input_vgprs
;
7042 shader
->info
.num_input_vgprs
+= 1;
7044 if (G_0286CC_SAMPLE_COVERAGE_ENA(shader
->config
.spi_ps_input_addr
))
7045 shader
->info
.num_input_vgprs
+= 1;
7046 if (G_0286CC_POS_FIXED_PT_ENA(shader
->config
.spi_ps_input_addr
))
7047 shader
->info
.num_input_vgprs
+= 1;
7050 si_calculate_max_simd_waves(shader
);
7051 si_shader_dump_stats_for_shader_db(shader
, debug
);
7056 * Create, compile and return a shader part (prolog or epilog).
7058 * \param sscreen screen
7059 * \param list list of shader parts of the same category
7060 * \param type shader type
7061 * \param key shader part key
7062 * \param prolog whether the part being requested is a prolog
7063 * \param tm LLVM target machine
7064 * \param debug debug callback
7065 * \param build the callback responsible for building the main function
7066 * \return non-NULL on success
7068 static struct si_shader_part
*
7069 si_get_shader_part(struct si_screen
*sscreen
,
7070 struct si_shader_part
**list
,
7071 enum pipe_shader_type type
,
7073 union si_shader_part_key
*key
,
7074 LLVMTargetMachineRef tm
,
7075 struct pipe_debug_callback
*debug
,
7076 void (*build
)(struct si_shader_context
*,
7077 union si_shader_part_key
*),
7080 struct si_shader_part
*result
;
7082 mtx_lock(&sscreen
->shader_parts_mutex
);
7084 /* Find existing. */
7085 for (result
= *list
; result
; result
= result
->next
) {
7086 if (memcmp(&result
->key
, key
, sizeof(*key
)) == 0) {
7087 mtx_unlock(&sscreen
->shader_parts_mutex
);
7092 /* Compile a new one. */
7093 result
= CALLOC_STRUCT(si_shader_part
);
7096 struct si_shader shader
= {};
7097 struct si_shader_context ctx
;
7099 si_init_shader_ctx(&ctx
, sscreen
, tm
);
7100 ctx
.shader
= &shader
;
7104 case PIPE_SHADER_VERTEX
:
7105 shader
.key
.as_ls
= key
->vs_prolog
.as_ls
;
7106 shader
.key
.as_es
= key
->vs_prolog
.as_es
;
7108 case PIPE_SHADER_TESS_CTRL
:
7110 shader
.key
.part
.tcs
.epilog
= key
->tcs_epilog
.states
;
7112 case PIPE_SHADER_GEOMETRY
:
7115 case PIPE_SHADER_FRAGMENT
:
7117 shader
.key
.part
.ps
.prolog
= key
->ps_prolog
.states
;
7119 shader
.key
.part
.ps
.epilog
= key
->ps_epilog
.states
;
7122 unreachable("bad shader part");
7128 si_llvm_optimize_module(&ctx
);
7130 if (si_compile_llvm(sscreen
, &result
->binary
, &result
->config
, tm
,
7131 ctx
.ac
.module
, debug
, ctx
.type
, name
)) {
7137 result
->next
= *list
;
7141 si_llvm_dispose(&ctx
);
7142 mtx_unlock(&sscreen
->shader_parts_mutex
);
7146 static LLVMValueRef
si_prolog_get_rw_buffers(struct si_shader_context
*ctx
)
7148 LLVMValueRef ptr
[2], list
;
7149 bool is_merged_shader
=
7150 ctx
->screen
->info
.chip_class
>= GFX9
&&
7151 (ctx
->type
== PIPE_SHADER_TESS_CTRL
||
7152 ctx
->type
== PIPE_SHADER_GEOMETRY
||
7153 ctx
->shader
->key
.as_ls
|| ctx
->shader
->key
.as_es
);
7155 if (HAVE_32BIT_POINTERS
) {
7156 ptr
[0] = LLVMGetParam(ctx
->main_fn
, (is_merged_shader
? 8 : 0) + SI_SGPR_RW_BUFFERS
);
7157 list
= LLVMBuildIntToPtr(ctx
->ac
.builder
, ptr
[0],
7158 ac_array_in_const32_addr_space(ctx
->v4i32
), "");
7162 /* Get the pointer to rw buffers. */
7163 ptr
[0] = LLVMGetParam(ctx
->main_fn
, (is_merged_shader
? 8 : 0) + SI_SGPR_RW_BUFFERS
);
7164 ptr
[1] = LLVMGetParam(ctx
->main_fn
, (is_merged_shader
? 8 : 0) + SI_SGPR_RW_BUFFERS
+ 1);
7165 list
= lp_build_gather_values(&ctx
->gallivm
, ptr
, 2);
7166 list
= LLVMBuildBitCast(ctx
->ac
.builder
, list
, ctx
->i64
, "");
7167 list
= LLVMBuildIntToPtr(ctx
->ac
.builder
, list
,
7168 ac_array_in_const_addr_space(ctx
->v4i32
), "");
7173 * Build the vertex shader prolog function.
7175 * The inputs are the same as VS (a lot of SGPRs and 4 VGPR system values).
7176 * All inputs are returned unmodified. The vertex load indices are
7177 * stored after them, which will be used by the API VS for fetching inputs.
7179 * For example, the expected outputs for instance_divisors[] = {0, 1, 2} are:
7184 * (VertexID + BaseVertex),
7185 * (InstanceID + StartInstance),
7186 * (InstanceID / 2 + StartInstance)
7188 static void si_build_vs_prolog_function(struct si_shader_context
*ctx
,
7189 union si_shader_part_key
*key
)
7191 struct si_function_info fninfo
;
7192 LLVMTypeRef
*returns
;
7193 LLVMValueRef ret
, func
;
7195 unsigned first_vs_vgpr
= key
->vs_prolog
.num_merged_next_stage_vgprs
;
7196 unsigned num_input_vgprs
= key
->vs_prolog
.num_merged_next_stage_vgprs
+ 4;
7197 LLVMValueRef input_vgprs
[9];
7198 unsigned num_all_input_regs
= key
->vs_prolog
.num_input_sgprs
+
7200 unsigned user_sgpr_base
= key
->vs_prolog
.num_merged_next_stage_vgprs
? 8 : 0;
7202 si_init_function_info(&fninfo
);
7204 /* 4 preloaded VGPRs + vertex load indices as prolog outputs */
7205 returns
= alloca((num_all_input_regs
+ key
->vs_prolog
.last_input
+ 1) *
7206 sizeof(LLVMTypeRef
));
7209 /* Declare input and output SGPRs. */
7210 for (i
= 0; i
< key
->vs_prolog
.num_input_sgprs
; i
++) {
7211 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7212 returns
[num_returns
++] = ctx
->i32
;
7215 /* Preloaded VGPRs (outputs must be floats) */
7216 for (i
= 0; i
< num_input_vgprs
; i
++) {
7217 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &input_vgprs
[i
]);
7218 returns
[num_returns
++] = ctx
->f32
;
7221 /* Vertex load indices. */
7222 for (i
= 0; i
<= key
->vs_prolog
.last_input
; i
++)
7223 returns
[num_returns
++] = ctx
->f32
;
7225 /* Create the function. */
7226 si_create_function(ctx
, "vs_prolog", returns
, num_returns
, &fninfo
, 0);
7227 func
= ctx
->main_fn
;
7229 if (key
->vs_prolog
.num_merged_next_stage_vgprs
) {
7230 if (!key
->vs_prolog
.is_monolithic
)
7231 si_init_exec_from_input(ctx
, 3, 0);
7233 if (key
->vs_prolog
.as_ls
&&
7234 ctx
->screen
->has_ls_vgpr_init_bug
) {
7235 /* If there are no HS threads, SPI loads the LS VGPRs
7236 * starting at VGPR 0. Shift them back to where they
7239 LLVMValueRef has_hs_threads
=
7240 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntNE
,
7241 unpack_param(ctx
, 3, 8, 8),
7244 for (i
= 4; i
> 0; --i
) {
7245 input_vgprs
[i
+ 1] =
7246 LLVMBuildSelect(ctx
->ac
.builder
, has_hs_threads
,
7248 input_vgprs
[i
- 1], "");
7253 ctx
->abi
.vertex_id
= input_vgprs
[first_vs_vgpr
];
7254 ctx
->abi
.instance_id
= input_vgprs
[first_vs_vgpr
+ (key
->vs_prolog
.as_ls
? 2 : 1)];
7256 /* Copy inputs to outputs. This should be no-op, as the registers match,
7257 * but it will prevent the compiler from overwriting them unintentionally.
7259 ret
= ctx
->return_value
;
7260 for (i
= 0; i
< key
->vs_prolog
.num_input_sgprs
; i
++) {
7261 LLVMValueRef p
= LLVMGetParam(func
, i
);
7262 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, p
, i
, "");
7264 for (i
= 0; i
< num_input_vgprs
; i
++) {
7265 LLVMValueRef p
= input_vgprs
[i
];
7266 p
= ac_to_float(&ctx
->ac
, p
);
7267 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, p
,
7268 key
->vs_prolog
.num_input_sgprs
+ i
, "");
7271 /* Compute vertex load indices from instance divisors. */
7272 LLVMValueRef instance_divisor_constbuf
= NULL
;
7274 if (key
->vs_prolog
.states
.instance_divisor_is_fetched
) {
7275 LLVMValueRef list
= si_prolog_get_rw_buffers(ctx
);
7276 LLVMValueRef buf_index
=
7277 LLVMConstInt(ctx
->i32
, SI_VS_CONST_INSTANCE_DIVISORS
, 0);
7278 instance_divisor_constbuf
=
7279 ac_build_load_to_sgpr(&ctx
->ac
, list
, buf_index
);
7282 for (i
= 0; i
<= key
->vs_prolog
.last_input
; i
++) {
7283 bool divisor_is_one
=
7284 key
->vs_prolog
.states
.instance_divisor_is_one
& (1u << i
);
7285 bool divisor_is_fetched
=
7286 key
->vs_prolog
.states
.instance_divisor_is_fetched
& (1u << i
);
7289 if (divisor_is_one
|| divisor_is_fetched
) {
7290 LLVMValueRef divisor
= ctx
->i32_1
;
7292 if (divisor_is_fetched
) {
7293 divisor
= buffer_load_const(ctx
, instance_divisor_constbuf
,
7294 LLVMConstInt(ctx
->i32
, i
* 4, 0));
7295 divisor
= ac_to_integer(&ctx
->ac
, divisor
);
7298 /* InstanceID / Divisor + StartInstance */
7299 index
= get_instance_index_for_fetch(ctx
,
7301 SI_SGPR_START_INSTANCE
,
7304 /* VertexID + BaseVertex */
7305 index
= LLVMBuildAdd(ctx
->ac
.builder
,
7307 LLVMGetParam(func
, user_sgpr_base
+
7308 SI_SGPR_BASE_VERTEX
), "");
7311 index
= ac_to_float(&ctx
->ac
, index
);
7312 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, index
,
7313 fninfo
.num_params
+ i
, "");
7316 si_llvm_build_ret(ctx
, ret
);
7319 static bool si_get_vs_prolog(struct si_screen
*sscreen
,
7320 LLVMTargetMachineRef tm
,
7321 struct si_shader
*shader
,
7322 struct pipe_debug_callback
*debug
,
7323 struct si_shader
*main_part
,
7324 const struct si_vs_prolog_bits
*key
)
7326 struct si_shader_selector
*vs
= main_part
->selector
;
7328 if (!si_vs_needs_prolog(vs
, key
))
7331 /* Get the prolog. */
7332 union si_shader_part_key prolog_key
;
7333 si_get_vs_prolog_key(&vs
->info
, main_part
->info
.num_input_sgprs
,
7334 key
, shader
, &prolog_key
);
7337 si_get_shader_part(sscreen
, &sscreen
->vs_prologs
,
7338 PIPE_SHADER_VERTEX
, true, &prolog_key
, tm
,
7339 debug
, si_build_vs_prolog_function
,
7340 "Vertex Shader Prolog");
7341 return shader
->prolog
!= NULL
;
7345 * Select and compile (or reuse) vertex shader parts (prolog & epilog).
7347 static bool si_shader_select_vs_parts(struct si_screen
*sscreen
,
7348 LLVMTargetMachineRef tm
,
7349 struct si_shader
*shader
,
7350 struct pipe_debug_callback
*debug
)
7352 return si_get_vs_prolog(sscreen
, tm
, shader
, debug
, shader
,
7353 &shader
->key
.part
.vs
.prolog
);
7357 * Compile the TCS epilog function. This writes tesselation factors to memory
7358 * based on the output primitive type of the tesselator (determined by TES).
7360 static void si_build_tcs_epilog_function(struct si_shader_context
*ctx
,
7361 union si_shader_part_key
*key
)
7363 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
7364 struct si_function_info fninfo
;
7367 si_init_function_info(&fninfo
);
7369 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
7370 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7371 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7372 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7373 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* wave info */
7374 ctx
->param_tcs_factor_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7375 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7376 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7377 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7378 add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7379 add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7380 add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7381 add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7382 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7383 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7384 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7385 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7386 if (!HAVE_32BIT_POINTERS
)
7387 add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7388 ctx
->param_tcs_offchip_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7389 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7390 ctx
->param_tcs_out_lds_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7392 add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7393 add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7394 add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7395 add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7396 ctx
->param_tcs_offchip_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7397 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7398 ctx
->param_tcs_out_lds_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7399 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7400 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7401 ctx
->param_tcs_factor_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7404 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
); /* VGPR gap */
7405 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
); /* VGPR gap */
7406 unsigned tess_factors_idx
=
7407 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
); /* patch index within the wave (REL_PATCH_ID) */
7408 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
); /* invocation ID within the patch */
7409 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
); /* LDS offset where tess factors should be loaded from */
7411 for (unsigned i
= 0; i
< 6; i
++)
7412 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
); /* tess factors */
7414 /* Create the function. */
7415 si_create_function(ctx
, "tcs_epilog", NULL
, 0, &fninfo
,
7416 ctx
->screen
->info
.chip_class
>= CIK
? 128 : 64);
7417 ac_declare_lds_as_pointer(&ctx
->ac
);
7418 func
= ctx
->main_fn
;
7420 LLVMValueRef invoc0_tess_factors
[6];
7421 for (unsigned i
= 0; i
< 6; i
++)
7422 invoc0_tess_factors
[i
] = LLVMGetParam(func
, tess_factors_idx
+ 3 + i
);
7424 si_write_tess_factors(bld_base
,
7425 LLVMGetParam(func
, tess_factors_idx
),
7426 LLVMGetParam(func
, tess_factors_idx
+ 1),
7427 LLVMGetParam(func
, tess_factors_idx
+ 2),
7428 invoc0_tess_factors
, invoc0_tess_factors
+ 4);
7430 LLVMBuildRetVoid(ctx
->ac
.builder
);
7434 * Select and compile (or reuse) TCS parts (epilog).
7436 static bool si_shader_select_tcs_parts(struct si_screen
*sscreen
,
7437 LLVMTargetMachineRef tm
,
7438 struct si_shader
*shader
,
7439 struct pipe_debug_callback
*debug
)
7441 if (sscreen
->info
.chip_class
>= GFX9
) {
7442 struct si_shader
*ls_main_part
=
7443 shader
->key
.part
.tcs
.ls
->main_shader_part_ls
;
7445 if (!si_get_vs_prolog(sscreen
, tm
, shader
, debug
, ls_main_part
,
7446 &shader
->key
.part
.tcs
.ls_prolog
))
7449 shader
->previous_stage
= ls_main_part
;
7452 /* Get the epilog. */
7453 union si_shader_part_key epilog_key
;
7454 memset(&epilog_key
, 0, sizeof(epilog_key
));
7455 epilog_key
.tcs_epilog
.states
= shader
->key
.part
.tcs
.epilog
;
7457 shader
->epilog
= si_get_shader_part(sscreen
, &sscreen
->tcs_epilogs
,
7458 PIPE_SHADER_TESS_CTRL
, false,
7459 &epilog_key
, tm
, debug
,
7460 si_build_tcs_epilog_function
,
7461 "Tessellation Control Shader Epilog");
7462 return shader
->epilog
!= NULL
;
7466 * Select and compile (or reuse) GS parts (prolog).
7468 static bool si_shader_select_gs_parts(struct si_screen
*sscreen
,
7469 LLVMTargetMachineRef tm
,
7470 struct si_shader
*shader
,
7471 struct pipe_debug_callback
*debug
)
7473 if (sscreen
->info
.chip_class
>= GFX9
) {
7474 struct si_shader
*es_main_part
=
7475 shader
->key
.part
.gs
.es
->main_shader_part_es
;
7477 if (shader
->key
.part
.gs
.es
->type
== PIPE_SHADER_VERTEX
&&
7478 !si_get_vs_prolog(sscreen
, tm
, shader
, debug
, es_main_part
,
7479 &shader
->key
.part
.gs
.vs_prolog
))
7482 shader
->previous_stage
= es_main_part
;
7485 if (!shader
->key
.part
.gs
.prolog
.tri_strip_adj_fix
)
7488 union si_shader_part_key prolog_key
;
7489 memset(&prolog_key
, 0, sizeof(prolog_key
));
7490 prolog_key
.gs_prolog
.states
= shader
->key
.part
.gs
.prolog
;
7492 shader
->prolog2
= si_get_shader_part(sscreen
, &sscreen
->gs_prologs
,
7493 PIPE_SHADER_GEOMETRY
, true,
7494 &prolog_key
, tm
, debug
,
7495 si_build_gs_prolog_function
,
7496 "Geometry Shader Prolog");
7497 return shader
->prolog2
!= NULL
;
7501 * Build the pixel shader prolog function. This handles:
7502 * - two-side color selection and interpolation
7503 * - overriding interpolation parameters for the API PS
7504 * - polygon stippling
7506 * All preloaded SGPRs and VGPRs are passed through unmodified unless they are
7507 * overriden by other states. (e.g. per-sample interpolation)
7508 * Interpolated colors are stored after the preloaded VGPRs.
7510 static void si_build_ps_prolog_function(struct si_shader_context
*ctx
,
7511 union si_shader_part_key
*key
)
7513 struct si_function_info fninfo
;
7514 LLVMValueRef ret
, func
;
7515 int num_returns
, i
, num_color_channels
;
7517 assert(si_need_ps_prolog(key
));
7519 si_init_function_info(&fninfo
);
7521 /* Declare inputs. */
7522 for (i
= 0; i
< key
->ps_prolog
.num_input_sgprs
; i
++)
7523 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7525 for (i
= 0; i
< key
->ps_prolog
.num_input_vgprs
; i
++)
7526 add_arg(&fninfo
, ARG_VGPR
, ctx
->f32
);
7528 /* Declare outputs (same as inputs + add colors if needed) */
7529 num_returns
= fninfo
.num_params
;
7530 num_color_channels
= util_bitcount(key
->ps_prolog
.colors_read
);
7531 for (i
= 0; i
< num_color_channels
; i
++)
7532 fninfo
.types
[num_returns
++] = ctx
->f32
;
7534 /* Create the function. */
7535 si_create_function(ctx
, "ps_prolog", fninfo
.types
, num_returns
,
7537 func
= ctx
->main_fn
;
7539 /* Copy inputs to outputs. This should be no-op, as the registers match,
7540 * but it will prevent the compiler from overwriting them unintentionally.
7542 ret
= ctx
->return_value
;
7543 for (i
= 0; i
< fninfo
.num_params
; i
++) {
7544 LLVMValueRef p
= LLVMGetParam(func
, i
);
7545 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, p
, i
, "");
7548 /* Polygon stippling. */
7549 if (key
->ps_prolog
.states
.poly_stipple
) {
7550 /* POS_FIXED_PT is always last. */
7551 unsigned pos
= key
->ps_prolog
.num_input_sgprs
+
7552 key
->ps_prolog
.num_input_vgprs
- 1;
7553 LLVMValueRef list
= si_prolog_get_rw_buffers(ctx
);
7555 si_llvm_emit_polygon_stipple(ctx
, list
, pos
);
7558 if (key
->ps_prolog
.states
.bc_optimize_for_persp
||
7559 key
->ps_prolog
.states
.bc_optimize_for_linear
) {
7560 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7561 LLVMValueRef center
[2], centroid
[2], tmp
, bc_optimize
;
7563 /* The shader should do: if (PRIM_MASK[31]) CENTROID = CENTER;
7564 * The hw doesn't compute CENTROID if the whole wave only
7565 * contains fully-covered quads.
7567 * PRIM_MASK is after user SGPRs.
7569 bc_optimize
= LLVMGetParam(func
, SI_PS_NUM_USER_SGPR
);
7570 bc_optimize
= LLVMBuildLShr(ctx
->ac
.builder
, bc_optimize
,
7571 LLVMConstInt(ctx
->i32
, 31, 0), "");
7572 bc_optimize
= LLVMBuildTrunc(ctx
->ac
.builder
, bc_optimize
,
7575 if (key
->ps_prolog
.states
.bc_optimize_for_persp
) {
7576 /* Read PERSP_CENTER. */
7577 for (i
= 0; i
< 2; i
++)
7578 center
[i
] = LLVMGetParam(func
, base
+ 2 + i
);
7579 /* Read PERSP_CENTROID. */
7580 for (i
= 0; i
< 2; i
++)
7581 centroid
[i
] = LLVMGetParam(func
, base
+ 4 + i
);
7582 /* Select PERSP_CENTROID. */
7583 for (i
= 0; i
< 2; i
++) {
7584 tmp
= LLVMBuildSelect(ctx
->ac
.builder
, bc_optimize
,
7585 center
[i
], centroid
[i
], "");
7586 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7587 tmp
, base
+ 4 + i
, "");
7590 if (key
->ps_prolog
.states
.bc_optimize_for_linear
) {
7591 /* Read LINEAR_CENTER. */
7592 for (i
= 0; i
< 2; i
++)
7593 center
[i
] = LLVMGetParam(func
, base
+ 8 + i
);
7594 /* Read LINEAR_CENTROID. */
7595 for (i
= 0; i
< 2; i
++)
7596 centroid
[i
] = LLVMGetParam(func
, base
+ 10 + i
);
7597 /* Select LINEAR_CENTROID. */
7598 for (i
= 0; i
< 2; i
++) {
7599 tmp
= LLVMBuildSelect(ctx
->ac
.builder
, bc_optimize
,
7600 center
[i
], centroid
[i
], "");
7601 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7602 tmp
, base
+ 10 + i
, "");
7607 /* Force per-sample interpolation. */
7608 if (key
->ps_prolog
.states
.force_persp_sample_interp
) {
7609 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7610 LLVMValueRef persp_sample
[2];
7612 /* Read PERSP_SAMPLE. */
7613 for (i
= 0; i
< 2; i
++)
7614 persp_sample
[i
] = LLVMGetParam(func
, base
+ i
);
7615 /* Overwrite PERSP_CENTER. */
7616 for (i
= 0; i
< 2; i
++)
7617 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7618 persp_sample
[i
], base
+ 2 + i
, "");
7619 /* Overwrite PERSP_CENTROID. */
7620 for (i
= 0; i
< 2; i
++)
7621 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7622 persp_sample
[i
], base
+ 4 + i
, "");
7624 if (key
->ps_prolog
.states
.force_linear_sample_interp
) {
7625 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7626 LLVMValueRef linear_sample
[2];
7628 /* Read LINEAR_SAMPLE. */
7629 for (i
= 0; i
< 2; i
++)
7630 linear_sample
[i
] = LLVMGetParam(func
, base
+ 6 + i
);
7631 /* Overwrite LINEAR_CENTER. */
7632 for (i
= 0; i
< 2; i
++)
7633 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7634 linear_sample
[i
], base
+ 8 + i
, "");
7635 /* Overwrite LINEAR_CENTROID. */
7636 for (i
= 0; i
< 2; i
++)
7637 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7638 linear_sample
[i
], base
+ 10 + i
, "");
7641 /* Force center interpolation. */
7642 if (key
->ps_prolog
.states
.force_persp_center_interp
) {
7643 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7644 LLVMValueRef persp_center
[2];
7646 /* Read PERSP_CENTER. */
7647 for (i
= 0; i
< 2; i
++)
7648 persp_center
[i
] = LLVMGetParam(func
, base
+ 2 + i
);
7649 /* Overwrite PERSP_SAMPLE. */
7650 for (i
= 0; i
< 2; i
++)
7651 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7652 persp_center
[i
], base
+ i
, "");
7653 /* Overwrite PERSP_CENTROID. */
7654 for (i
= 0; i
< 2; i
++)
7655 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7656 persp_center
[i
], base
+ 4 + i
, "");
7658 if (key
->ps_prolog
.states
.force_linear_center_interp
) {
7659 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7660 LLVMValueRef linear_center
[2];
7662 /* Read LINEAR_CENTER. */
7663 for (i
= 0; i
< 2; i
++)
7664 linear_center
[i
] = LLVMGetParam(func
, base
+ 8 + i
);
7665 /* Overwrite LINEAR_SAMPLE. */
7666 for (i
= 0; i
< 2; i
++)
7667 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7668 linear_center
[i
], base
+ 6 + i
, "");
7669 /* Overwrite LINEAR_CENTROID. */
7670 for (i
= 0; i
< 2; i
++)
7671 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7672 linear_center
[i
], base
+ 10 + i
, "");
7675 /* Interpolate colors. */
7676 unsigned color_out_idx
= 0;
7677 for (i
= 0; i
< 2; i
++) {
7678 unsigned writemask
= (key
->ps_prolog
.colors_read
>> (i
* 4)) & 0xf;
7679 unsigned face_vgpr
= key
->ps_prolog
.num_input_sgprs
+
7680 key
->ps_prolog
.face_vgpr_index
;
7681 LLVMValueRef interp
[2], color
[4];
7682 LLVMValueRef interp_ij
= NULL
, prim_mask
= NULL
, face
= NULL
;
7687 /* If the interpolation qualifier is not CONSTANT (-1). */
7688 if (key
->ps_prolog
.color_interp_vgpr_index
[i
] != -1) {
7689 unsigned interp_vgpr
= key
->ps_prolog
.num_input_sgprs
+
7690 key
->ps_prolog
.color_interp_vgpr_index
[i
];
7692 /* Get the (i,j) updated by bc_optimize handling. */
7693 interp
[0] = LLVMBuildExtractValue(ctx
->ac
.builder
, ret
,
7695 interp
[1] = LLVMBuildExtractValue(ctx
->ac
.builder
, ret
,
7696 interp_vgpr
+ 1, "");
7697 interp_ij
= lp_build_gather_values(&ctx
->gallivm
, interp
, 2);
7700 /* Use the absolute location of the input. */
7701 prim_mask
= LLVMGetParam(func
, SI_PS_NUM_USER_SGPR
);
7703 if (key
->ps_prolog
.states
.color_two_side
) {
7704 face
= LLVMGetParam(func
, face_vgpr
);
7705 face
= ac_to_integer(&ctx
->ac
, face
);
7708 interp_fs_input(ctx
,
7709 key
->ps_prolog
.color_attr_index
[i
],
7710 TGSI_SEMANTIC_COLOR
, i
,
7711 key
->ps_prolog
.num_interp_inputs
,
7712 key
->ps_prolog
.colors_read
, interp_ij
,
7713 prim_mask
, face
, color
);
7716 unsigned chan
= u_bit_scan(&writemask
);
7717 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, color
[chan
],
7718 fninfo
.num_params
+ color_out_idx
++, "");
7722 /* Section 15.2.2 (Shader Inputs) of the OpenGL 4.5 (Core Profile) spec
7725 * "When per-sample shading is active due to the use of a fragment
7726 * input qualified by sample or due to the use of the gl_SampleID
7727 * or gl_SamplePosition variables, only the bit for the current
7728 * sample is set in gl_SampleMaskIn. When state specifies multiple
7729 * fragment shader invocations for a given fragment, the sample
7730 * mask for any single fragment shader invocation may specify a
7731 * subset of the covered samples for the fragment. In this case,
7732 * the bit corresponding to each covered sample will be set in
7733 * exactly one fragment shader invocation."
7735 * The samplemask loaded by hardware is always the coverage of the
7736 * entire pixel/fragment, so mask bits out based on the sample ID.
7738 if (key
->ps_prolog
.states
.samplemask_log_ps_iter
) {
7739 /* The bit pattern matches that used by fixed function fragment
7741 static const uint16_t ps_iter_masks
[] = {
7742 0xffff, /* not used */
7748 assert(key
->ps_prolog
.states
.samplemask_log_ps_iter
< ARRAY_SIZE(ps_iter_masks
));
7750 uint32_t ps_iter_mask
= ps_iter_masks
[key
->ps_prolog
.states
.samplemask_log_ps_iter
];
7751 unsigned ancillary_vgpr
= key
->ps_prolog
.num_input_sgprs
+
7752 key
->ps_prolog
.ancillary_vgpr_index
;
7753 LLVMValueRef sampleid
= unpack_param(ctx
, ancillary_vgpr
, 8, 4);
7754 LLVMValueRef samplemask
= LLVMGetParam(func
, ancillary_vgpr
+ 1);
7756 samplemask
= ac_to_integer(&ctx
->ac
, samplemask
);
7757 samplemask
= LLVMBuildAnd(
7760 LLVMBuildShl(ctx
->ac
.builder
,
7761 LLVMConstInt(ctx
->i32
, ps_iter_mask
, false),
7764 samplemask
= ac_to_float(&ctx
->ac
, samplemask
);
7766 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, samplemask
,
7767 ancillary_vgpr
+ 1, "");
7770 /* Tell LLVM to insert WQM instruction sequence when needed. */
7771 if (key
->ps_prolog
.wqm
) {
7772 LLVMAddTargetDependentFunctionAttr(func
,
7773 "amdgpu-ps-wqm-outputs", "");
7776 si_llvm_build_ret(ctx
, ret
);
7780 * Build the pixel shader epilog function. This handles everything that must be
7781 * emulated for pixel shader exports. (alpha-test, format conversions, etc)
7783 static void si_build_ps_epilog_function(struct si_shader_context
*ctx
,
7784 union si_shader_part_key
*key
)
7786 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
7787 struct si_function_info fninfo
;
7788 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
7790 struct si_ps_exports exp
= {};
7792 si_init_function_info(&fninfo
);
7794 /* Declare input SGPRs. */
7795 ctx
->param_rw_buffers
= add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7796 ctx
->param_bindless_samplers_and_images
= add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7797 ctx
->param_const_and_shader_buffers
= add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7798 ctx
->param_samplers_and_images
= add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7799 add_arg_checked(&fninfo
, ARG_SGPR
, ctx
->f32
, SI_PARAM_ALPHA_REF
);
7801 /* Declare input VGPRs. */
7802 unsigned required_num_params
=
7803 fninfo
.num_sgpr_params
+
7804 util_bitcount(key
->ps_epilog
.colors_written
) * 4 +
7805 key
->ps_epilog
.writes_z
+
7806 key
->ps_epilog
.writes_stencil
+
7807 key
->ps_epilog
.writes_samplemask
;
7809 required_num_params
= MAX2(required_num_params
,
7810 fninfo
.num_sgpr_params
+ PS_EPILOG_SAMPLEMASK_MIN_LOC
+ 1);
7812 while (fninfo
.num_params
< required_num_params
)
7813 add_arg(&fninfo
, ARG_VGPR
, ctx
->f32
);
7815 /* Create the function. */
7816 si_create_function(ctx
, "ps_epilog", NULL
, 0, &fninfo
, 0);
7817 /* Disable elimination of unused inputs. */
7818 ac_llvm_add_target_dep_function_attr(ctx
->main_fn
,
7819 "InitialPSInputAddr", 0xffffff);
7821 /* Process colors. */
7822 unsigned vgpr
= fninfo
.num_sgpr_params
;
7823 unsigned colors_written
= key
->ps_epilog
.colors_written
;
7824 int last_color_export
= -1;
7826 /* Find the last color export. */
7827 if (!key
->ps_epilog
.writes_z
&&
7828 !key
->ps_epilog
.writes_stencil
&&
7829 !key
->ps_epilog
.writes_samplemask
) {
7830 unsigned spi_format
= key
->ps_epilog
.states
.spi_shader_col_format
;
7832 /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
7833 if (colors_written
== 0x1 && key
->ps_epilog
.states
.last_cbuf
> 0) {
7834 /* Just set this if any of the colorbuffers are enabled. */
7836 ((1ull << (4 * (key
->ps_epilog
.states
.last_cbuf
+ 1))) - 1))
7837 last_color_export
= 0;
7839 for (i
= 0; i
< 8; i
++)
7840 if (colors_written
& (1 << i
) &&
7841 (spi_format
>> (i
* 4)) & 0xf)
7842 last_color_export
= i
;
7846 while (colors_written
) {
7847 LLVMValueRef color
[4];
7848 int mrt
= u_bit_scan(&colors_written
);
7850 for (i
= 0; i
< 4; i
++)
7851 color
[i
] = LLVMGetParam(ctx
->main_fn
, vgpr
++);
7853 si_export_mrt_color(bld_base
, color
, mrt
,
7854 fninfo
.num_params
- 1,
7855 mrt
== last_color_export
, &exp
);
7858 /* Process depth, stencil, samplemask. */
7859 if (key
->ps_epilog
.writes_z
)
7860 depth
= LLVMGetParam(ctx
->main_fn
, vgpr
++);
7861 if (key
->ps_epilog
.writes_stencil
)
7862 stencil
= LLVMGetParam(ctx
->main_fn
, vgpr
++);
7863 if (key
->ps_epilog
.writes_samplemask
)
7864 samplemask
= LLVMGetParam(ctx
->main_fn
, vgpr
++);
7866 if (depth
|| stencil
|| samplemask
)
7867 si_export_mrt_z(bld_base
, depth
, stencil
, samplemask
, &exp
);
7868 else if (last_color_export
== -1)
7869 ac_build_export_null(&ctx
->ac
);
7872 si_emit_ps_exports(ctx
, &exp
);
7875 LLVMBuildRetVoid(ctx
->ac
.builder
);
7879 * Select and compile (or reuse) pixel shader parts (prolog & epilog).
7881 static bool si_shader_select_ps_parts(struct si_screen
*sscreen
,
7882 LLVMTargetMachineRef tm
,
7883 struct si_shader
*shader
,
7884 struct pipe_debug_callback
*debug
)
7886 union si_shader_part_key prolog_key
;
7887 union si_shader_part_key epilog_key
;
7889 /* Get the prolog. */
7890 si_get_ps_prolog_key(shader
, &prolog_key
, true);
7892 /* The prolog is a no-op if these aren't set. */
7893 if (si_need_ps_prolog(&prolog_key
)) {
7895 si_get_shader_part(sscreen
, &sscreen
->ps_prologs
,
7896 PIPE_SHADER_FRAGMENT
, true,
7897 &prolog_key
, tm
, debug
,
7898 si_build_ps_prolog_function
,
7899 "Fragment Shader Prolog");
7900 if (!shader
->prolog
)
7904 /* Get the epilog. */
7905 si_get_ps_epilog_key(shader
, &epilog_key
);
7908 si_get_shader_part(sscreen
, &sscreen
->ps_epilogs
,
7909 PIPE_SHADER_FRAGMENT
, false,
7910 &epilog_key
, tm
, debug
,
7911 si_build_ps_epilog_function
,
7912 "Fragment Shader Epilog");
7913 if (!shader
->epilog
)
7916 /* Enable POS_FIXED_PT if polygon stippling is enabled. */
7917 if (shader
->key
.part
.ps
.prolog
.poly_stipple
) {
7918 shader
->config
.spi_ps_input_ena
|= S_0286CC_POS_FIXED_PT_ENA(1);
7919 assert(G_0286CC_POS_FIXED_PT_ENA(shader
->config
.spi_ps_input_addr
));
7922 /* Set up the enable bits for per-sample shading if needed. */
7923 if (shader
->key
.part
.ps
.prolog
.force_persp_sample_interp
&&
7924 (G_0286CC_PERSP_CENTER_ENA(shader
->config
.spi_ps_input_ena
) ||
7925 G_0286CC_PERSP_CENTROID_ENA(shader
->config
.spi_ps_input_ena
))) {
7926 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_CENTER_ENA
;
7927 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_CENTROID_ENA
;
7928 shader
->config
.spi_ps_input_ena
|= S_0286CC_PERSP_SAMPLE_ENA(1);
7930 if (shader
->key
.part
.ps
.prolog
.force_linear_sample_interp
&&
7931 (G_0286CC_LINEAR_CENTER_ENA(shader
->config
.spi_ps_input_ena
) ||
7932 G_0286CC_LINEAR_CENTROID_ENA(shader
->config
.spi_ps_input_ena
))) {
7933 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_CENTER_ENA
;
7934 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_CENTROID_ENA
;
7935 shader
->config
.spi_ps_input_ena
|= S_0286CC_LINEAR_SAMPLE_ENA(1);
7937 if (shader
->key
.part
.ps
.prolog
.force_persp_center_interp
&&
7938 (G_0286CC_PERSP_SAMPLE_ENA(shader
->config
.spi_ps_input_ena
) ||
7939 G_0286CC_PERSP_CENTROID_ENA(shader
->config
.spi_ps_input_ena
))) {
7940 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_SAMPLE_ENA
;
7941 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_CENTROID_ENA
;
7942 shader
->config
.spi_ps_input_ena
|= S_0286CC_PERSP_CENTER_ENA(1);
7944 if (shader
->key
.part
.ps
.prolog
.force_linear_center_interp
&&
7945 (G_0286CC_LINEAR_SAMPLE_ENA(shader
->config
.spi_ps_input_ena
) ||
7946 G_0286CC_LINEAR_CENTROID_ENA(shader
->config
.spi_ps_input_ena
))) {
7947 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_SAMPLE_ENA
;
7948 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_CENTROID_ENA
;
7949 shader
->config
.spi_ps_input_ena
|= S_0286CC_LINEAR_CENTER_ENA(1);
7952 /* POW_W_FLOAT requires that one of the perspective weights is enabled. */
7953 if (G_0286CC_POS_W_FLOAT_ENA(shader
->config
.spi_ps_input_ena
) &&
7954 !(shader
->config
.spi_ps_input_ena
& 0xf)) {
7955 shader
->config
.spi_ps_input_ena
|= S_0286CC_PERSP_CENTER_ENA(1);
7956 assert(G_0286CC_PERSP_CENTER_ENA(shader
->config
.spi_ps_input_addr
));
7959 /* At least one pair of interpolation weights must be enabled. */
7960 if (!(shader
->config
.spi_ps_input_ena
& 0x7f)) {
7961 shader
->config
.spi_ps_input_ena
|= S_0286CC_LINEAR_CENTER_ENA(1);
7962 assert(G_0286CC_LINEAR_CENTER_ENA(shader
->config
.spi_ps_input_addr
));
7965 /* Samplemask fixup requires the sample ID. */
7966 if (shader
->key
.part
.ps
.prolog
.samplemask_log_ps_iter
) {
7967 shader
->config
.spi_ps_input_ena
|= S_0286CC_ANCILLARY_ENA(1);
7968 assert(G_0286CC_ANCILLARY_ENA(shader
->config
.spi_ps_input_addr
));
7971 /* The sample mask input is always enabled, because the API shader always
7972 * passes it through to the epilog. Disable it here if it's unused.
7974 if (!shader
->key
.part
.ps
.epilog
.poly_line_smoothing
&&
7975 !shader
->selector
->info
.reads_samplemask
)
7976 shader
->config
.spi_ps_input_ena
&= C_0286CC_SAMPLE_COVERAGE_ENA
;
7981 void si_multiwave_lds_size_workaround(struct si_screen
*sscreen
,
7984 /* If tessellation is all offchip and on-chip GS isn't used, this
7985 * workaround is not needed.
7989 /* SPI barrier management bug:
7990 * Make sure we have at least 4k of LDS in use to avoid the bug.
7991 * It applies to workgroup sizes of more than one wavefront.
7993 if (sscreen
->info
.family
== CHIP_BONAIRE
||
7994 sscreen
->info
.family
== CHIP_KABINI
||
7995 sscreen
->info
.family
== CHIP_MULLINS
)
7996 *lds_size
= MAX2(*lds_size
, 8);
7999 static void si_fix_resource_usage(struct si_screen
*sscreen
,
8000 struct si_shader
*shader
)
8002 unsigned min_sgprs
= shader
->info
.num_input_sgprs
+ 2; /* VCC */
8004 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
, min_sgprs
);
8006 if (shader
->selector
->type
== PIPE_SHADER_COMPUTE
&&
8007 si_get_max_workgroup_size(shader
) > 64) {
8008 si_multiwave_lds_size_workaround(sscreen
,
8009 &shader
->config
.lds_size
);
8013 int si_shader_create(struct si_screen
*sscreen
, LLVMTargetMachineRef tm
,
8014 struct si_shader
*shader
,
8015 struct pipe_debug_callback
*debug
)
8017 struct si_shader_selector
*sel
= shader
->selector
;
8018 struct si_shader
*mainp
= *si_get_main_shader_part(sel
, &shader
->key
);
8021 /* LS, ES, VS are compiled on demand if the main part hasn't been
8022 * compiled for that stage.
8024 * Vertex shaders are compiled on demand when a vertex fetch
8025 * workaround must be applied.
8027 if (shader
->is_monolithic
) {
8028 /* Monolithic shader (compiled as a whole, has many variants,
8029 * may take a long time to compile).
8031 r
= si_compile_tgsi_shader(sscreen
, tm
, shader
, true, debug
);
8035 /* The shader consists of several parts:
8037 * - the middle part is the user shader, it has 1 variant only
8038 * and it was compiled during the creation of the shader
8040 * - the prolog part is inserted at the beginning
8041 * - the epilog part is inserted at the end
8043 * The prolog and epilog have many (but simple) variants.
8045 * Starting with gfx9, geometry and tessellation control
8046 * shaders also contain the prolog and user shader parts of
8047 * the previous shader stage.
8053 /* Copy the compiled TGSI shader data over. */
8054 shader
->is_binary_shared
= true;
8055 shader
->binary
= mainp
->binary
;
8056 shader
->config
= mainp
->config
;
8057 shader
->info
.num_input_sgprs
= mainp
->info
.num_input_sgprs
;
8058 shader
->info
.num_input_vgprs
= mainp
->info
.num_input_vgprs
;
8059 shader
->info
.face_vgpr_index
= mainp
->info
.face_vgpr_index
;
8060 shader
->info
.ancillary_vgpr_index
= mainp
->info
.ancillary_vgpr_index
;
8061 memcpy(shader
->info
.vs_output_param_offset
,
8062 mainp
->info
.vs_output_param_offset
,
8063 sizeof(mainp
->info
.vs_output_param_offset
));
8064 shader
->info
.uses_instanceid
= mainp
->info
.uses_instanceid
;
8065 shader
->info
.nr_pos_exports
= mainp
->info
.nr_pos_exports
;
8066 shader
->info
.nr_param_exports
= mainp
->info
.nr_param_exports
;
8068 /* Select prologs and/or epilogs. */
8069 switch (sel
->type
) {
8070 case PIPE_SHADER_VERTEX
:
8071 if (!si_shader_select_vs_parts(sscreen
, tm
, shader
, debug
))
8074 case PIPE_SHADER_TESS_CTRL
:
8075 if (!si_shader_select_tcs_parts(sscreen
, tm
, shader
, debug
))
8078 case PIPE_SHADER_TESS_EVAL
:
8080 case PIPE_SHADER_GEOMETRY
:
8081 if (!si_shader_select_gs_parts(sscreen
, tm
, shader
, debug
))
8084 case PIPE_SHADER_FRAGMENT
:
8085 if (!si_shader_select_ps_parts(sscreen
, tm
, shader
, debug
))
8088 /* Make sure we have at least as many VGPRs as there
8089 * are allocated inputs.
8091 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
8092 shader
->info
.num_input_vgprs
);
8096 /* Update SGPR and VGPR counts. */
8097 if (shader
->prolog
) {
8098 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
,
8099 shader
->prolog
->config
.num_sgprs
);
8100 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
8101 shader
->prolog
->config
.num_vgprs
);
8103 if (shader
->previous_stage
) {
8104 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
,
8105 shader
->previous_stage
->config
.num_sgprs
);
8106 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
8107 shader
->previous_stage
->config
.num_vgprs
);
8108 shader
->config
.spilled_sgprs
=
8109 MAX2(shader
->config
.spilled_sgprs
,
8110 shader
->previous_stage
->config
.spilled_sgprs
);
8111 shader
->config
.spilled_vgprs
=
8112 MAX2(shader
->config
.spilled_vgprs
,
8113 shader
->previous_stage
->config
.spilled_vgprs
);
8114 shader
->config
.private_mem_vgprs
=
8115 MAX2(shader
->config
.private_mem_vgprs
,
8116 shader
->previous_stage
->config
.private_mem_vgprs
);
8117 shader
->config
.scratch_bytes_per_wave
=
8118 MAX2(shader
->config
.scratch_bytes_per_wave
,
8119 shader
->previous_stage
->config
.scratch_bytes_per_wave
);
8120 shader
->info
.uses_instanceid
|=
8121 shader
->previous_stage
->info
.uses_instanceid
;
8123 if (shader
->prolog2
) {
8124 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
,
8125 shader
->prolog2
->config
.num_sgprs
);
8126 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
8127 shader
->prolog2
->config
.num_vgprs
);
8129 if (shader
->epilog
) {
8130 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
,
8131 shader
->epilog
->config
.num_sgprs
);
8132 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
8133 shader
->epilog
->config
.num_vgprs
);
8135 si_calculate_max_simd_waves(shader
);
8138 si_fix_resource_usage(sscreen
, shader
);
8139 si_shader_dump(sscreen
, shader
, debug
, sel
->info
.processor
,
8143 r
= si_shader_binary_upload(sscreen
, shader
);
8145 fprintf(stderr
, "LLVM failed to upload shader\n");
8152 void si_shader_destroy(struct si_shader
*shader
)
8154 if (shader
->scratch_bo
)
8155 r600_resource_reference(&shader
->scratch_bo
, NULL
);
8157 r600_resource_reference(&shader
->bo
, NULL
);
8159 if (!shader
->is_binary_shared
)
8160 ac_shader_binary_clean(&shader
->binary
);
8162 free(shader
->shader_log
);