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 LLVMValueRef vertex_index
,
1491 LLVMValueRef param_index
,
1492 unsigned const_index
,
1494 unsigned driver_location
,
1501 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1502 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
1503 LLVMValueRef dw_addr
, stride
;
1504 LLVMValueRef buffer
, base
, addr
;
1505 LLVMValueRef values
[4];
1506 bool skip_lds_store
;
1507 bool is_tess_factor
= false, is_tess_inner
= false;
1509 driver_location
= driver_location
/ 4;
1512 /* Add the constant index to the indirect index */
1513 param_index
= LLVMBuildAdd(ctx
->ac
.builder
, param_index
,
1514 LLVMConstInt(ctx
->i32
, const_index
, 0), "");
1516 if (const_index
!= 0)
1517 param_index
= LLVMConstInt(ctx
->i32
, const_index
, 0);
1521 stride
= get_tcs_out_vertex_dw_stride(ctx
);
1522 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
1523 dw_addr
= get_dw_address_from_generic_indices(ctx
, stride
, dw_addr
,
1524 vertex_index
, param_index
,
1526 info
->output_semantic_name
,
1527 info
->output_semantic_index
,
1530 skip_lds_store
= !info
->reads_pervertex_outputs
;
1532 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
1533 dw_addr
= get_dw_address_from_generic_indices(ctx
, NULL
, dw_addr
,
1534 vertex_index
, param_index
,
1536 info
->output_semantic_name
,
1537 info
->output_semantic_index
,
1540 skip_lds_store
= !info
->reads_perpatch_outputs
;
1543 int name
= info
->output_semantic_name
[driver_location
];
1545 /* Always write tess factors into LDS for the TCS epilog. */
1546 if (name
== TGSI_SEMANTIC_TESSINNER
||
1547 name
== TGSI_SEMANTIC_TESSOUTER
) {
1548 /* The epilog doesn't read LDS if invocation 0 defines tess factors. */
1549 skip_lds_store
= !info
->reads_tessfactor_outputs
&&
1550 ctx
->shader
->selector
->tcs_info
.tessfactors_are_def_in_all_invocs
;
1551 is_tess_factor
= true;
1552 is_tess_inner
= name
== TGSI_SEMANTIC_TESSINNER
;
1557 buffer
= get_tess_ring_descriptor(ctx
, TESS_OFFCHIP_RING_TCS
);
1559 base
= LLVMGetParam(ctx
->main_fn
, ctx
->param_tcs_offchip_offset
);
1561 addr
= get_tcs_tes_buffer_address_from_generic_indices(ctx
, vertex_index
,
1562 param_index
, driver_location
,
1563 info
->output_semantic_name
,
1564 info
->output_semantic_index
,
1567 for (unsigned chan
= 0; chan
< 4; chan
++) {
1568 if (!(writemask
& (1 << chan
)))
1570 LLVMValueRef value
= ac_llvm_extract_elem(&ctx
->ac
, src
, chan
- component
);
1572 /* Skip LDS stores if there is no LDS read of this output. */
1573 if (!skip_lds_store
)
1574 lds_store(ctx
, chan
, dw_addr
, value
);
1576 value
= ac_to_integer(&ctx
->ac
, value
);
1577 values
[chan
] = value
;
1579 if (writemask
!= 0xF && !is_tess_factor
) {
1580 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, value
, 1,
1582 4 * chan
, 1, 0, true, false);
1585 /* Write tess factors into VGPRs for the epilog. */
1586 if (is_tess_factor
&&
1587 ctx
->shader
->selector
->tcs_info
.tessfactors_are_def_in_all_invocs
) {
1588 if (!is_tess_inner
) {
1589 LLVMBuildStore(ctx
->ac
.builder
, value
, /* outer */
1590 ctx
->invoc0_tess_factors
[chan
]);
1591 } else if (chan
< 2) {
1592 LLVMBuildStore(ctx
->ac
.builder
, value
, /* inner */
1593 ctx
->invoc0_tess_factors
[4 + chan
]);
1598 if (writemask
== 0xF && !is_tess_factor
) {
1599 LLVMValueRef value
= lp_build_gather_values(&ctx
->gallivm
,
1601 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, value
, 4, addr
,
1602 base
, 0, 1, 0, true, false);
1606 LLVMValueRef
si_llvm_load_input_gs(struct ac_shader_abi
*abi
,
1607 unsigned input_index
,
1608 unsigned vtx_offset_param
,
1612 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1613 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
1614 struct si_shader
*shader
= ctx
->shader
;
1615 struct lp_build_context
*uint
= &ctx
->bld_base
.uint_bld
;
1616 LLVMValueRef vtx_offset
, soffset
;
1617 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
1618 unsigned semantic_name
= info
->input_semantic_name
[input_index
];
1619 unsigned semantic_index
= info
->input_semantic_index
[input_index
];
1623 param
= si_shader_io_get_unique_index(semantic_name
, semantic_index
);
1625 /* GFX9 has the ESGS ring in LDS. */
1626 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
1627 unsigned index
= vtx_offset_param
;
1629 switch (index
/ 2) {
1631 vtx_offset
= unpack_param(ctx
, ctx
->param_gs_vtx01_offset
,
1632 index
% 2 ? 16 : 0, 16);
1635 vtx_offset
= unpack_param(ctx
, ctx
->param_gs_vtx23_offset
,
1636 index
% 2 ? 16 : 0, 16);
1639 vtx_offset
= unpack_param(ctx
, ctx
->param_gs_vtx45_offset
,
1640 index
% 2 ? 16 : 0, 16);
1647 vtx_offset
= LLVMBuildAdd(ctx
->ac
.builder
, vtx_offset
,
1648 LLVMConstInt(ctx
->i32
, param
* 4, 0), "");
1649 return lds_load(bld_base
, type
, swizzle
, vtx_offset
);
1652 /* GFX6: input load from the ESGS ring in memory. */
1653 if (swizzle
== ~0) {
1654 LLVMValueRef values
[TGSI_NUM_CHANNELS
];
1656 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
1657 values
[chan
] = si_llvm_load_input_gs(abi
, input_index
, vtx_offset_param
,
1660 return lp_build_gather_values(&ctx
->gallivm
, values
,
1664 /* Get the vertex offset parameter on GFX6. */
1665 LLVMValueRef gs_vtx_offset
= ctx
->gs_vtx_offset
[vtx_offset_param
];
1667 vtx_offset
= lp_build_mul_imm(uint
, gs_vtx_offset
, 4);
1669 soffset
= LLVMConstInt(ctx
->i32
, (param
* 4 + swizzle
) * 256, 0);
1671 value
= ac_build_buffer_load(&ctx
->ac
, ctx
->esgs_ring
, 1, ctx
->i32_0
,
1672 vtx_offset
, soffset
, 0, 1, 0, true, false);
1673 if (llvm_type_is_64bit(ctx
, type
)) {
1674 LLVMValueRef value2
;
1675 soffset
= LLVMConstInt(ctx
->i32
, (param
* 4 + swizzle
+ 1) * 256, 0);
1677 value2
= ac_build_buffer_load(&ctx
->ac
, ctx
->esgs_ring
, 1,
1678 ctx
->i32_0
, vtx_offset
, soffset
,
1679 0, 1, 0, true, false);
1680 return si_llvm_emit_fetch_64bit(bld_base
, type
, value
, value2
);
1682 return LLVMBuildBitCast(ctx
->ac
.builder
, value
, type
, "");
1685 static LLVMValueRef
fetch_input_gs(
1686 struct lp_build_tgsi_context
*bld_base
,
1687 const struct tgsi_full_src_register
*reg
,
1688 enum tgsi_opcode_type type
,
1691 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1692 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
1694 unsigned semantic_name
= info
->input_semantic_name
[reg
->Register
.Index
];
1695 if (swizzle
!= ~0 && semantic_name
== TGSI_SEMANTIC_PRIMID
)
1696 return get_primitive_id(ctx
, swizzle
);
1698 if (!reg
->Register
.Dimension
)
1701 return si_llvm_load_input_gs(&ctx
->abi
, reg
->Register
.Index
,
1702 reg
->Dimension
.Index
,
1703 tgsi2llvmtype(bld_base
, type
),
1707 static int lookup_interp_param_index(unsigned interpolate
, unsigned location
)
1709 switch (interpolate
) {
1710 case TGSI_INTERPOLATE_CONSTANT
:
1713 case TGSI_INTERPOLATE_LINEAR
:
1714 if (location
== TGSI_INTERPOLATE_LOC_SAMPLE
)
1715 return SI_PARAM_LINEAR_SAMPLE
;
1716 else if (location
== TGSI_INTERPOLATE_LOC_CENTROID
)
1717 return SI_PARAM_LINEAR_CENTROID
;
1719 return SI_PARAM_LINEAR_CENTER
;
1721 case TGSI_INTERPOLATE_COLOR
:
1722 case TGSI_INTERPOLATE_PERSPECTIVE
:
1723 if (location
== TGSI_INTERPOLATE_LOC_SAMPLE
)
1724 return SI_PARAM_PERSP_SAMPLE
;
1725 else if (location
== TGSI_INTERPOLATE_LOC_CENTROID
)
1726 return SI_PARAM_PERSP_CENTROID
;
1728 return SI_PARAM_PERSP_CENTER
;
1731 fprintf(stderr
, "Warning: Unhandled interpolation mode.\n");
1736 static LLVMValueRef
si_build_fs_interp(struct si_shader_context
*ctx
,
1737 unsigned attr_index
, unsigned chan
,
1738 LLVMValueRef prim_mask
,
1739 LLVMValueRef i
, LLVMValueRef j
)
1742 return ac_build_fs_interp(&ctx
->ac
,
1743 LLVMConstInt(ctx
->i32
, chan
, 0),
1744 LLVMConstInt(ctx
->i32
, attr_index
, 0),
1747 return ac_build_fs_interp_mov(&ctx
->ac
,
1748 LLVMConstInt(ctx
->i32
, 2, 0), /* P0 */
1749 LLVMConstInt(ctx
->i32
, chan
, 0),
1750 LLVMConstInt(ctx
->i32
, attr_index
, 0),
1755 * Interpolate a fragment shader input.
1757 * @param ctx context
1758 * @param input_index index of the input in hardware
1759 * @param semantic_name TGSI_SEMANTIC_*
1760 * @param semantic_index semantic index
1761 * @param num_interp_inputs number of all interpolated inputs (= BCOLOR offset)
1762 * @param colors_read_mask color components read (4 bits for each color, 8 bits in total)
1763 * @param interp_param interpolation weights (i,j)
1764 * @param prim_mask SI_PARAM_PRIM_MASK
1765 * @param face SI_PARAM_FRONT_FACE
1766 * @param result the return value (4 components)
1768 static void interp_fs_input(struct si_shader_context
*ctx
,
1769 unsigned input_index
,
1770 unsigned semantic_name
,
1771 unsigned semantic_index
,
1772 unsigned num_interp_inputs
,
1773 unsigned colors_read_mask
,
1774 LLVMValueRef interp_param
,
1775 LLVMValueRef prim_mask
,
1777 LLVMValueRef result
[4])
1779 LLVMValueRef i
= NULL
, j
= NULL
;
1782 /* fs.constant returns the param from the middle vertex, so it's not
1783 * really useful for flat shading. It's meant to be used for custom
1784 * interpolation (but the intrinsic can't fetch from the other two
1787 * Luckily, it doesn't matter, because we rely on the FLAT_SHADE state
1788 * to do the right thing. The only reason we use fs.constant is that
1789 * fs.interp cannot be used on integers, because they can be equal
1792 * When interp is false we will use fs.constant or for newer llvm,
1793 * amdgcn.interp.mov.
1795 bool interp
= interp_param
!= NULL
;
1798 interp_param
= LLVMBuildBitCast(ctx
->ac
.builder
, interp_param
,
1799 LLVMVectorType(ctx
->f32
, 2), "");
1801 i
= LLVMBuildExtractElement(ctx
->ac
.builder
, interp_param
,
1803 j
= LLVMBuildExtractElement(ctx
->ac
.builder
, interp_param
,
1807 if (semantic_name
== TGSI_SEMANTIC_COLOR
&&
1808 ctx
->shader
->key
.part
.ps
.prolog
.color_two_side
) {
1809 LLVMValueRef is_face_positive
;
1811 /* If BCOLOR0 is used, BCOLOR1 is at offset "num_inputs + 1",
1812 * otherwise it's at offset "num_inputs".
1814 unsigned back_attr_offset
= num_interp_inputs
;
1815 if (semantic_index
== 1 && colors_read_mask
& 0xf)
1816 back_attr_offset
+= 1;
1818 is_face_positive
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntNE
,
1819 face
, ctx
->i32_0
, "");
1821 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
1822 LLVMValueRef front
, back
;
1824 front
= si_build_fs_interp(ctx
,
1827 back
= si_build_fs_interp(ctx
,
1828 back_attr_offset
, chan
,
1831 result
[chan
] = LLVMBuildSelect(ctx
->ac
.builder
,
1837 } else if (semantic_name
== TGSI_SEMANTIC_FOG
) {
1838 result
[0] = si_build_fs_interp(ctx
, input_index
,
1839 0, prim_mask
, i
, j
);
1841 result
[2] = LLVMConstReal(ctx
->f32
, 0.0f
);
1842 result
[3] = LLVMConstReal(ctx
->f32
, 1.0f
);
1844 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
1845 result
[chan
] = si_build_fs_interp(ctx
,
1852 void si_llvm_load_input_fs(
1853 struct si_shader_context
*ctx
,
1854 unsigned input_index
,
1855 LLVMValueRef out
[4])
1857 struct lp_build_context
*base
= &ctx
->bld_base
.base
;
1858 struct si_shader
*shader
= ctx
->shader
;
1859 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
1860 LLVMValueRef main_fn
= ctx
->main_fn
;
1861 LLVMValueRef interp_param
= NULL
;
1862 int interp_param_idx
;
1863 enum tgsi_semantic semantic_name
= info
->input_semantic_name
[input_index
];
1864 unsigned semantic_index
= info
->input_semantic_index
[input_index
];
1865 enum tgsi_interpolate_mode interp_mode
= info
->input_interpolate
[input_index
];
1866 enum tgsi_interpolate_loc interp_loc
= info
->input_interpolate_loc
[input_index
];
1868 /* Get colors from input VGPRs (set by the prolog). */
1869 if (semantic_name
== TGSI_SEMANTIC_COLOR
) {
1870 unsigned colors_read
= shader
->selector
->info
.colors_read
;
1871 unsigned mask
= colors_read
>> (semantic_index
* 4);
1872 unsigned offset
= SI_PARAM_POS_FIXED_PT
+ 1 +
1873 (semantic_index
? util_bitcount(colors_read
& 0xf) : 0);
1875 out
[0] = mask
& 0x1 ? LLVMGetParam(main_fn
, offset
++) : base
->undef
;
1876 out
[1] = mask
& 0x2 ? LLVMGetParam(main_fn
, offset
++) : base
->undef
;
1877 out
[2] = mask
& 0x4 ? LLVMGetParam(main_fn
, offset
++) : base
->undef
;
1878 out
[3] = mask
& 0x8 ? LLVMGetParam(main_fn
, offset
++) : base
->undef
;
1882 interp_param_idx
= lookup_interp_param_index(interp_mode
, interp_loc
);
1883 if (interp_param_idx
== -1)
1885 else if (interp_param_idx
) {
1886 interp_param
= LLVMGetParam(ctx
->main_fn
, interp_param_idx
);
1889 interp_fs_input(ctx
, input_index
, semantic_name
,
1890 semantic_index
, 0, /* this param is unused */
1891 shader
->selector
->info
.colors_read
, interp_param
,
1893 LLVMGetParam(main_fn
, SI_PARAM_FRONT_FACE
),
1897 static void declare_input_fs(
1898 struct si_shader_context
*ctx
,
1899 unsigned input_index
,
1900 const struct tgsi_full_declaration
*decl
,
1901 LLVMValueRef out
[4])
1903 si_llvm_load_input_fs(ctx
, input_index
, out
);
1906 static LLVMValueRef
get_sample_id(struct si_shader_context
*ctx
)
1908 return unpack_param(ctx
, SI_PARAM_ANCILLARY
, 8, 4);
1911 static LLVMValueRef
get_block_size(struct ac_shader_abi
*abi
)
1913 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1915 LLVMValueRef values
[3];
1916 LLVMValueRef result
;
1918 unsigned *properties
= ctx
->shader
->selector
->info
.properties
;
1920 if (properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
] != 0) {
1921 unsigned sizes
[3] = {
1922 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
],
1923 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT
],
1924 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH
]
1927 for (i
= 0; i
< 3; ++i
)
1928 values
[i
] = LLVMConstInt(ctx
->i32
, sizes
[i
], 0);
1930 result
= lp_build_gather_values(&ctx
->gallivm
, values
, 3);
1932 result
= LLVMGetParam(ctx
->main_fn
, ctx
->param_block_size
);
1939 * Load a dword from a constant buffer.
1941 static LLVMValueRef
buffer_load_const(struct si_shader_context
*ctx
,
1942 LLVMValueRef resource
,
1943 LLVMValueRef offset
)
1945 return ac_build_buffer_load(&ctx
->ac
, resource
, 1, NULL
, offset
, NULL
,
1946 0, 0, 0, true, true);
1949 static LLVMValueRef
load_sample_position(struct ac_shader_abi
*abi
, LLVMValueRef sample_id
)
1951 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1952 struct lp_build_context
*uint_bld
= &ctx
->bld_base
.uint_bld
;
1953 LLVMValueRef desc
= LLVMGetParam(ctx
->main_fn
, ctx
->param_rw_buffers
);
1954 LLVMValueRef buf_index
= LLVMConstInt(ctx
->i32
, SI_PS_CONST_SAMPLE_POSITIONS
, 0);
1955 LLVMValueRef resource
= ac_build_load_to_sgpr(&ctx
->ac
, desc
, buf_index
);
1957 /* offset = sample_id * 8 (8 = 2 floats containing samplepos.xy) */
1958 LLVMValueRef offset0
= lp_build_mul_imm(uint_bld
, sample_id
, 8);
1959 LLVMValueRef offset1
= LLVMBuildAdd(ctx
->ac
.builder
, offset0
, LLVMConstInt(ctx
->i32
, 4, 0), "");
1961 LLVMValueRef pos
[4] = {
1962 buffer_load_const(ctx
, resource
, offset0
),
1963 buffer_load_const(ctx
, resource
, offset1
),
1964 LLVMConstReal(ctx
->f32
, 0),
1965 LLVMConstReal(ctx
->f32
, 0)
1968 return lp_build_gather_values(&ctx
->gallivm
, pos
, 4);
1971 static LLVMValueRef
load_sample_mask_in(struct ac_shader_abi
*abi
)
1973 return abi
->sample_coverage
;
1976 static LLVMValueRef
si_load_tess_coord(struct ac_shader_abi
*abi
)
1978 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1979 struct lp_build_context
*bld
= &ctx
->bld_base
.base
;
1981 LLVMValueRef coord
[4] = {
1982 LLVMGetParam(ctx
->main_fn
, ctx
->param_tes_u
),
1983 LLVMGetParam(ctx
->main_fn
, ctx
->param_tes_v
),
1988 /* For triangles, the vector should be (u, v, 1-u-v). */
1989 if (ctx
->shader
->selector
->info
.properties
[TGSI_PROPERTY_TES_PRIM_MODE
] ==
1990 PIPE_PRIM_TRIANGLES
)
1991 coord
[2] = lp_build_sub(bld
, ctx
->ac
.f32_1
,
1992 lp_build_add(bld
, coord
[0], coord
[1]));
1994 return lp_build_gather_values(&ctx
->gallivm
, coord
, 4);
1997 static LLVMValueRef
load_tess_level(struct si_shader_context
*ctx
,
1998 unsigned semantic_name
)
2000 LLVMValueRef base
, addr
;
2002 int param
= si_shader_io_get_unique_index_patch(semantic_name
, 0);
2004 base
= LLVMGetParam(ctx
->main_fn
, ctx
->param_tcs_offchip_offset
);
2005 addr
= get_tcs_tes_buffer_address(ctx
, get_rel_patch_id(ctx
), NULL
,
2006 LLVMConstInt(ctx
->i32
, param
, 0));
2008 return buffer_load(&ctx
->bld_base
, ctx
->f32
,
2009 ~0, ctx
->tess_offchip_ring
, base
, addr
, true);
2013 static LLVMValueRef
si_load_tess_level(struct ac_shader_abi
*abi
,
2014 unsigned varying_id
)
2016 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
2017 unsigned semantic_name
;
2019 switch (varying_id
) {
2020 case VARYING_SLOT_TESS_LEVEL_INNER
:
2021 semantic_name
= TGSI_SEMANTIC_TESSINNER
;
2023 case VARYING_SLOT_TESS_LEVEL_OUTER
:
2024 semantic_name
= TGSI_SEMANTIC_TESSOUTER
;
2027 unreachable("unknown tess level");
2030 return load_tess_level(ctx
, semantic_name
);
2034 static LLVMValueRef
si_load_patch_vertices_in(struct ac_shader_abi
*abi
)
2036 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
2037 if (ctx
->type
== PIPE_SHADER_TESS_CTRL
)
2038 return unpack_param(ctx
, ctx
->param_tcs_out_lds_layout
, 13, 6);
2039 else if (ctx
->type
== PIPE_SHADER_TESS_EVAL
)
2040 return get_num_tcs_out_vertices(ctx
);
2042 unreachable("invalid shader stage for TGSI_SEMANTIC_VERTICESIN");
2045 void si_load_system_value(struct si_shader_context
*ctx
,
2047 const struct tgsi_full_declaration
*decl
)
2049 LLVMValueRef value
= 0;
2051 assert(index
< RADEON_LLVM_MAX_SYSTEM_VALUES
);
2053 switch (decl
->Semantic
.Name
) {
2054 case TGSI_SEMANTIC_INSTANCEID
:
2055 value
= ctx
->abi
.instance_id
;
2058 case TGSI_SEMANTIC_VERTEXID
:
2059 value
= LLVMBuildAdd(ctx
->ac
.builder
,
2061 ctx
->abi
.base_vertex
, "");
2064 case TGSI_SEMANTIC_VERTEXID_NOBASE
:
2065 /* Unused. Clarify the meaning in indexed vs. non-indexed
2066 * draws if this is ever used again. */
2070 case TGSI_SEMANTIC_BASEVERTEX
:
2072 /* For non-indexed draws, the base vertex set by the driver
2073 * (for direct draws) or the CP (for indirect draws) is the
2074 * first vertex ID, but GLSL expects 0 to be returned.
2076 LLVMValueRef vs_state
= LLVMGetParam(ctx
->main_fn
, ctx
->param_vs_state_bits
);
2077 LLVMValueRef indexed
;
2079 indexed
= LLVMBuildLShr(ctx
->ac
.builder
, vs_state
, ctx
->i32_1
, "");
2080 indexed
= LLVMBuildTrunc(ctx
->ac
.builder
, indexed
, ctx
->i1
, "");
2082 value
= LLVMBuildSelect(ctx
->ac
.builder
, indexed
,
2083 ctx
->abi
.base_vertex
, ctx
->i32_0
, "");
2087 case TGSI_SEMANTIC_BASEINSTANCE
:
2088 value
= ctx
->abi
.start_instance
;
2091 case TGSI_SEMANTIC_DRAWID
:
2092 value
= ctx
->abi
.draw_id
;
2095 case TGSI_SEMANTIC_INVOCATIONID
:
2096 if (ctx
->type
== PIPE_SHADER_TESS_CTRL
)
2097 value
= unpack_llvm_param(ctx
, ctx
->abi
.tcs_rel_ids
, 8, 5);
2098 else if (ctx
->type
== PIPE_SHADER_GEOMETRY
)
2099 value
= ctx
->abi
.gs_invocation_id
;
2101 assert(!"INVOCATIONID not implemented");
2104 case TGSI_SEMANTIC_POSITION
:
2106 LLVMValueRef pos
[4] = {
2107 LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_X_FLOAT
),
2108 LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_Y_FLOAT
),
2109 LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_Z_FLOAT
),
2110 lp_build_emit_llvm_unary(&ctx
->bld_base
, TGSI_OPCODE_RCP
,
2111 LLVMGetParam(ctx
->main_fn
,
2112 SI_PARAM_POS_W_FLOAT
)),
2114 value
= lp_build_gather_values(&ctx
->gallivm
, pos
, 4);
2118 case TGSI_SEMANTIC_FACE
:
2119 value
= ctx
->abi
.front_face
;
2122 case TGSI_SEMANTIC_SAMPLEID
:
2123 value
= get_sample_id(ctx
);
2126 case TGSI_SEMANTIC_SAMPLEPOS
: {
2127 LLVMValueRef pos
[4] = {
2128 LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_X_FLOAT
),
2129 LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_Y_FLOAT
),
2130 LLVMConstReal(ctx
->f32
, 0),
2131 LLVMConstReal(ctx
->f32
, 0)
2133 pos
[0] = lp_build_emit_llvm_unary(&ctx
->bld_base
,
2134 TGSI_OPCODE_FRC
, pos
[0]);
2135 pos
[1] = lp_build_emit_llvm_unary(&ctx
->bld_base
,
2136 TGSI_OPCODE_FRC
, pos
[1]);
2137 value
= lp_build_gather_values(&ctx
->gallivm
, pos
, 4);
2141 case TGSI_SEMANTIC_SAMPLEMASK
:
2142 /* This can only occur with the OpenGL Core profile, which
2143 * doesn't support smoothing.
2145 value
= LLVMGetParam(ctx
->main_fn
, SI_PARAM_SAMPLE_COVERAGE
);
2148 case TGSI_SEMANTIC_TESSCOORD
:
2149 value
= si_load_tess_coord(&ctx
->abi
);
2152 case TGSI_SEMANTIC_VERTICESIN
:
2153 value
= si_load_patch_vertices_in(&ctx
->abi
);
2156 case TGSI_SEMANTIC_TESSINNER
:
2157 case TGSI_SEMANTIC_TESSOUTER
:
2158 value
= load_tess_level(ctx
, decl
->Semantic
.Name
);
2161 case TGSI_SEMANTIC_DEFAULT_TESSOUTER_SI
:
2162 case TGSI_SEMANTIC_DEFAULT_TESSINNER_SI
:
2164 LLVMValueRef buf
, slot
, val
[4];
2167 slot
= LLVMConstInt(ctx
->i32
, SI_HS_CONST_DEFAULT_TESS_LEVELS
, 0);
2168 buf
= LLVMGetParam(ctx
->main_fn
, ctx
->param_rw_buffers
);
2169 buf
= ac_build_load_to_sgpr(&ctx
->ac
, buf
, slot
);
2170 offset
= decl
->Semantic
.Name
== TGSI_SEMANTIC_DEFAULT_TESSINNER_SI
? 4 : 0;
2172 for (i
= 0; i
< 4; i
++)
2173 val
[i
] = buffer_load_const(ctx
, buf
,
2174 LLVMConstInt(ctx
->i32
, (offset
+ i
) * 4, 0));
2175 value
= lp_build_gather_values(&ctx
->gallivm
, val
, 4);
2179 case TGSI_SEMANTIC_PRIMID
:
2180 value
= get_primitive_id(ctx
, 0);
2183 case TGSI_SEMANTIC_GRID_SIZE
:
2184 value
= ctx
->abi
.num_work_groups
;
2187 case TGSI_SEMANTIC_BLOCK_SIZE
:
2188 value
= get_block_size(&ctx
->abi
);
2191 case TGSI_SEMANTIC_BLOCK_ID
:
2193 LLVMValueRef values
[3];
2195 for (int i
= 0; i
< 3; i
++) {
2196 values
[i
] = ctx
->i32_0
;
2197 if (ctx
->abi
.workgroup_ids
[i
]) {
2198 values
[i
] = ctx
->abi
.workgroup_ids
[i
];
2201 value
= lp_build_gather_values(&ctx
->gallivm
, values
, 3);
2205 case TGSI_SEMANTIC_THREAD_ID
:
2206 value
= ctx
->abi
.local_invocation_ids
;
2209 case TGSI_SEMANTIC_HELPER_INVOCATION
:
2210 value
= lp_build_intrinsic(ctx
->ac
.builder
,
2211 "llvm.amdgcn.ps.live",
2213 LP_FUNC_ATTR_READNONE
);
2214 value
= LLVMBuildNot(ctx
->ac
.builder
, value
, "");
2215 value
= LLVMBuildSExt(ctx
->ac
.builder
, value
, ctx
->i32
, "");
2218 case TGSI_SEMANTIC_SUBGROUP_SIZE
:
2219 value
= LLVMConstInt(ctx
->i32
, 64, 0);
2222 case TGSI_SEMANTIC_SUBGROUP_INVOCATION
:
2223 value
= ac_get_thread_id(&ctx
->ac
);
2226 case TGSI_SEMANTIC_SUBGROUP_EQ_MASK
:
2228 LLVMValueRef id
= ac_get_thread_id(&ctx
->ac
);
2229 id
= LLVMBuildZExt(ctx
->ac
.builder
, id
, ctx
->i64
, "");
2230 value
= LLVMBuildShl(ctx
->ac
.builder
, LLVMConstInt(ctx
->i64
, 1, 0), id
, "");
2231 value
= LLVMBuildBitCast(ctx
->ac
.builder
, value
, ctx
->v2i32
, "");
2235 case TGSI_SEMANTIC_SUBGROUP_GE_MASK
:
2236 case TGSI_SEMANTIC_SUBGROUP_GT_MASK
:
2237 case TGSI_SEMANTIC_SUBGROUP_LE_MASK
:
2238 case TGSI_SEMANTIC_SUBGROUP_LT_MASK
:
2240 LLVMValueRef id
= ac_get_thread_id(&ctx
->ac
);
2241 if (decl
->Semantic
.Name
== TGSI_SEMANTIC_SUBGROUP_GT_MASK
||
2242 decl
->Semantic
.Name
== TGSI_SEMANTIC_SUBGROUP_LE_MASK
) {
2243 /* All bits set except LSB */
2244 value
= LLVMConstInt(ctx
->i64
, -2, 0);
2247 value
= LLVMConstInt(ctx
->i64
, -1, 0);
2249 id
= LLVMBuildZExt(ctx
->ac
.builder
, id
, ctx
->i64
, "");
2250 value
= LLVMBuildShl(ctx
->ac
.builder
, value
, id
, "");
2251 if (decl
->Semantic
.Name
== TGSI_SEMANTIC_SUBGROUP_LE_MASK
||
2252 decl
->Semantic
.Name
== TGSI_SEMANTIC_SUBGROUP_LT_MASK
)
2253 value
= LLVMBuildNot(ctx
->ac
.builder
, value
, "");
2254 value
= LLVMBuildBitCast(ctx
->ac
.builder
, value
, ctx
->v2i32
, "");
2259 assert(!"unknown system value");
2263 ctx
->system_values
[index
] = value
;
2266 void si_declare_compute_memory(struct si_shader_context
*ctx
)
2268 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
2270 LLVMTypeRef i8p
= LLVMPointerType(ctx
->i8
, AC_LOCAL_ADDR_SPACE
);
2273 assert(!ctx
->ac
.lds
);
2275 var
= LLVMAddGlobalInAddressSpace(ctx
->ac
.module
,
2276 LLVMArrayType(ctx
->i8
, sel
->local_size
),
2278 AC_LOCAL_ADDR_SPACE
);
2279 LLVMSetAlignment(var
, 4);
2281 ctx
->ac
.lds
= LLVMBuildBitCast(ctx
->ac
.builder
, var
, i8p
, "");
2284 void si_tgsi_declare_compute_memory(struct si_shader_context
*ctx
,
2285 const struct tgsi_full_declaration
*decl
)
2287 assert(decl
->Declaration
.MemType
== TGSI_MEMORY_TYPE_SHARED
);
2288 assert(decl
->Range
.First
== decl
->Range
.Last
);
2290 si_declare_compute_memory(ctx
);
2293 static LLVMValueRef
load_const_buffer_desc(struct si_shader_context
*ctx
, int i
)
2295 LLVMValueRef list_ptr
= LLVMGetParam(ctx
->main_fn
,
2296 ctx
->param_const_and_shader_buffers
);
2298 return ac_build_load_to_sgpr(&ctx
->ac
, list_ptr
,
2299 LLVMConstInt(ctx
->i32
, si_get_constbuf_slot(i
), 0));
2302 static LLVMValueRef
load_ubo(struct ac_shader_abi
*abi
, LLVMValueRef index
)
2304 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
2305 LLVMValueRef ptr
= LLVMGetParam(ctx
->main_fn
, ctx
->param_const_and_shader_buffers
);
2307 index
= si_llvm_bound_index(ctx
, index
, ctx
->num_const_buffers
);
2308 index
= LLVMBuildAdd(ctx
->ac
.builder
, index
,
2309 LLVMConstInt(ctx
->i32
, SI_NUM_SHADER_BUFFERS
, 0), "");
2311 return ac_build_load_to_sgpr(&ctx
->ac
, ptr
, index
);
2315 load_ssbo(struct ac_shader_abi
*abi
, LLVMValueRef index
, bool write
)
2317 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
2318 LLVMValueRef rsrc_ptr
= LLVMGetParam(ctx
->main_fn
,
2319 ctx
->param_const_and_shader_buffers
);
2321 index
= si_llvm_bound_index(ctx
, index
, ctx
->num_shader_buffers
);
2322 index
= LLVMBuildSub(ctx
->ac
.builder
,
2323 LLVMConstInt(ctx
->i32
, SI_NUM_SHADER_BUFFERS
- 1, 0),
2326 return ac_build_load_to_sgpr(&ctx
->ac
, rsrc_ptr
, index
);
2329 static LLVMValueRef
fetch_constant(
2330 struct lp_build_tgsi_context
*bld_base
,
2331 const struct tgsi_full_src_register
*reg
,
2332 enum tgsi_opcode_type type
,
2335 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2336 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
2337 const struct tgsi_ind_register
*ireg
= ®
->Indirect
;
2340 LLVMValueRef addr
, bufp
;
2342 if (swizzle
== LP_CHAN_ALL
) {
2344 LLVMValueRef values
[4];
2345 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
)
2346 values
[chan
] = fetch_constant(bld_base
, reg
, type
, chan
);
2348 return lp_build_gather_values(&ctx
->gallivm
, values
, 4);
2351 /* Split 64-bit loads. */
2352 if (tgsi_type_is_64bit(type
)) {
2353 LLVMValueRef lo
, hi
;
2355 lo
= fetch_constant(bld_base
, reg
, TGSI_TYPE_UNSIGNED
, swizzle
);
2356 hi
= fetch_constant(bld_base
, reg
, TGSI_TYPE_UNSIGNED
, swizzle
+ 1);
2357 return si_llvm_emit_fetch_64bit(bld_base
, tgsi2llvmtype(bld_base
, type
),
2361 idx
= reg
->Register
.Index
* 4 + swizzle
;
2362 if (reg
->Register
.Indirect
) {
2363 addr
= si_get_indirect_index(ctx
, ireg
, 16, idx
* 4);
2365 addr
= LLVMConstInt(ctx
->i32
, idx
* 4, 0);
2368 /* Fast path when user data SGPRs point to constant buffer 0 directly. */
2369 if (sel
->info
.const_buffers_declared
== 1 &&
2370 sel
->info
.shader_buffers_declared
== 0) {
2372 LLVMGetParam(ctx
->main_fn
, ctx
->param_const_and_shader_buffers
);
2374 /* This enables use of s_load_dword and flat_load_dword for const buffer 0
2375 * loads, and up to x4 load opcode merging. However, it leads to horrible
2376 * code reducing SIMD wave occupancy from 8 to 2 in many cases.
2378 * Using s_buffer_load_dword (x1) seems to be the best option right now.
2380 * LLVM 5.0 on SI doesn't insert a required s_nop between SALU setting
2381 * a descriptor and s_buffer_load_dword using it, so we can't expand
2382 * the pointer into a full descriptor like below. We have to use
2383 * s_load_dword instead. The only case when LLVM 5.0 would select
2384 * s_buffer_load_dword (that we have to prevent) is when we use use
2385 * a literal offset where we don't need bounds checking.
2387 if (ctx
->screen
->info
.chip_class
== SI
&&
2388 HAVE_LLVM
< 0x0600 &&
2389 !reg
->Register
.Indirect
) {
2390 addr
= LLVMBuildLShr(ctx
->ac
.builder
, addr
, LLVMConstInt(ctx
->i32
, 2, 0), "");
2391 LLVMValueRef result
= ac_build_load_invariant(&ctx
->ac
, ptr
, addr
);
2392 return bitcast(bld_base
, type
, result
);
2395 /* Do the bounds checking with a descriptor, because
2396 * doing computation and manual bounds checking of 64-bit
2397 * addresses generates horrible VALU code with very high
2398 * VGPR usage and very low SIMD occupancy.
2400 ptr
= LLVMBuildPtrToInt(ctx
->ac
.builder
, ptr
, ctx
->i64
, "");
2401 ptr
= LLVMBuildBitCast(ctx
->ac
.builder
, ptr
, ctx
->v2i32
, "");
2403 LLVMValueRef desc_elems
[] = {
2404 LLVMBuildExtractElement(ctx
->ac
.builder
, ptr
, ctx
->i32_0
, ""),
2405 LLVMBuildExtractElement(ctx
->ac
.builder
, ptr
, ctx
->i32_1
, ""),
2406 LLVMConstInt(ctx
->i32
, (sel
->info
.const_file_max
[0] + 1) * 16, 0),
2407 LLVMConstInt(ctx
->i32
,
2408 S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
2409 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
2410 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
2411 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
2412 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
2413 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
), 0)
2415 LLVMValueRef desc
= ac_build_gather_values(&ctx
->ac
, desc_elems
, 4);
2416 LLVMValueRef result
= buffer_load_const(ctx
, desc
, addr
);
2417 return bitcast(bld_base
, type
, result
);
2420 assert(reg
->Register
.Dimension
);
2421 buf
= reg
->Dimension
.Index
;
2423 if (reg
->Dimension
.Indirect
) {
2424 LLVMValueRef ptr
= LLVMGetParam(ctx
->main_fn
, ctx
->param_const_and_shader_buffers
);
2426 index
= si_get_bounded_indirect_index(ctx
, ®
->DimIndirect
,
2427 reg
->Dimension
.Index
,
2428 ctx
->num_const_buffers
);
2429 index
= LLVMBuildAdd(ctx
->ac
.builder
, index
,
2430 LLVMConstInt(ctx
->i32
, SI_NUM_SHADER_BUFFERS
, 0), "");
2431 bufp
= ac_build_load_to_sgpr(&ctx
->ac
, ptr
, index
);
2433 bufp
= load_const_buffer_desc(ctx
, buf
);
2435 return bitcast(bld_base
, type
, buffer_load_const(ctx
, bufp
, addr
));
2438 /* Initialize arguments for the shader export intrinsic */
2439 static void si_llvm_init_export_args(struct si_shader_context
*ctx
,
2440 LLVMValueRef
*values
,
2442 struct ac_export_args
*args
)
2444 LLVMValueRef f32undef
= LLVMGetUndef(ctx
->ac
.f32
);
2445 unsigned spi_shader_col_format
= V_028714_SPI_SHADER_32_ABGR
;
2447 bool is_int8
, is_int10
;
2449 /* Default is 0xf. Adjusted below depending on the format. */
2450 args
->enabled_channels
= 0xf; /* writemask */
2452 /* Specify whether the EXEC mask represents the valid mask */
2453 args
->valid_mask
= 0;
2455 /* Specify whether this is the last export */
2458 /* Specify the target we are exporting */
2459 args
->target
= target
;
2461 if (ctx
->type
== PIPE_SHADER_FRAGMENT
) {
2462 const struct si_shader_key
*key
= &ctx
->shader
->key
;
2463 unsigned col_formats
= key
->part
.ps
.epilog
.spi_shader_col_format
;
2464 int cbuf
= target
- V_008DFC_SQ_EXP_MRT
;
2466 assert(cbuf
>= 0 && cbuf
< 8);
2467 spi_shader_col_format
= (col_formats
>> (cbuf
* 4)) & 0xf;
2468 is_int8
= (key
->part
.ps
.epilog
.color_is_int8
>> cbuf
) & 0x1;
2469 is_int10
= (key
->part
.ps
.epilog
.color_is_int10
>> cbuf
) & 0x1;
2472 args
->compr
= false;
2473 args
->out
[0] = f32undef
;
2474 args
->out
[1] = f32undef
;
2475 args
->out
[2] = f32undef
;
2476 args
->out
[3] = f32undef
;
2478 LLVMValueRef (*packf
)(struct ac_llvm_context
*ctx
, LLVMValueRef args
[2]) = NULL
;
2479 LLVMValueRef (*packi
)(struct ac_llvm_context
*ctx
, LLVMValueRef args
[2],
2480 unsigned bits
, bool hi
) = NULL
;
2482 switch (spi_shader_col_format
) {
2483 case V_028714_SPI_SHADER_ZERO
:
2484 args
->enabled_channels
= 0; /* writemask */
2485 args
->target
= V_008DFC_SQ_EXP_NULL
;
2488 case V_028714_SPI_SHADER_32_R
:
2489 args
->enabled_channels
= 1; /* writemask */
2490 args
->out
[0] = values
[0];
2493 case V_028714_SPI_SHADER_32_GR
:
2494 args
->enabled_channels
= 0x3; /* writemask */
2495 args
->out
[0] = values
[0];
2496 args
->out
[1] = values
[1];
2499 case V_028714_SPI_SHADER_32_AR
:
2500 args
->enabled_channels
= 0x9; /* writemask */
2501 args
->out
[0] = values
[0];
2502 args
->out
[3] = values
[3];
2505 case V_028714_SPI_SHADER_FP16_ABGR
:
2506 packf
= ac_build_cvt_pkrtz_f16
;
2509 case V_028714_SPI_SHADER_UNORM16_ABGR
:
2510 packf
= ac_build_cvt_pknorm_u16
;
2513 case V_028714_SPI_SHADER_SNORM16_ABGR
:
2514 packf
= ac_build_cvt_pknorm_i16
;
2517 case V_028714_SPI_SHADER_UINT16_ABGR
:
2518 packi
= ac_build_cvt_pk_u16
;
2521 case V_028714_SPI_SHADER_SINT16_ABGR
:
2522 packi
= ac_build_cvt_pk_i16
;
2525 case V_028714_SPI_SHADER_32_ABGR
:
2526 memcpy(&args
->out
[0], values
, sizeof(values
[0]) * 4);
2530 /* Pack f16 or norm_i16/u16. */
2532 for (chan
= 0; chan
< 2; chan
++) {
2533 LLVMValueRef pack_args
[2] = {
2535 values
[2 * chan
+ 1]
2537 LLVMValueRef packed
;
2539 packed
= packf(&ctx
->ac
, pack_args
);
2540 args
->out
[chan
] = ac_to_float(&ctx
->ac
, packed
);
2542 args
->compr
= 1; /* COMPR flag */
2546 for (chan
= 0; chan
< 2; chan
++) {
2547 LLVMValueRef pack_args
[2] = {
2548 ac_to_integer(&ctx
->ac
, values
[2 * chan
]),
2549 ac_to_integer(&ctx
->ac
, values
[2 * chan
+ 1])
2551 LLVMValueRef packed
;
2553 packed
= packi(&ctx
->ac
, pack_args
,
2554 is_int8
? 8 : is_int10
? 10 : 16,
2556 args
->out
[chan
] = ac_to_float(&ctx
->ac
, packed
);
2558 args
->compr
= 1; /* COMPR flag */
2562 static void si_alpha_test(struct lp_build_tgsi_context
*bld_base
,
2565 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2567 if (ctx
->shader
->key
.part
.ps
.epilog
.alpha_func
!= PIPE_FUNC_NEVER
) {
2568 static LLVMRealPredicate cond_map
[PIPE_FUNC_ALWAYS
+ 1] = {
2569 [PIPE_FUNC_LESS
] = LLVMRealOLT
,
2570 [PIPE_FUNC_EQUAL
] = LLVMRealOEQ
,
2571 [PIPE_FUNC_LEQUAL
] = LLVMRealOLE
,
2572 [PIPE_FUNC_GREATER
] = LLVMRealOGT
,
2573 [PIPE_FUNC_NOTEQUAL
] = LLVMRealONE
,
2574 [PIPE_FUNC_GEQUAL
] = LLVMRealOGE
,
2576 LLVMRealPredicate cond
= cond_map
[ctx
->shader
->key
.part
.ps
.epilog
.alpha_func
];
2579 LLVMValueRef alpha_ref
= LLVMGetParam(ctx
->main_fn
,
2580 SI_PARAM_ALPHA_REF
);
2581 LLVMValueRef alpha_pass
=
2582 LLVMBuildFCmp(ctx
->ac
.builder
, cond
, alpha
, alpha_ref
, "");
2583 ac_build_kill_if_false(&ctx
->ac
, alpha_pass
);
2585 ac_build_kill_if_false(&ctx
->ac
, LLVMConstInt(ctx
->i1
, 0, 0));
2589 static LLVMValueRef
si_scale_alpha_by_sample_mask(struct lp_build_tgsi_context
*bld_base
,
2591 unsigned samplemask_param
)
2593 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2594 LLVMValueRef coverage
;
2596 /* alpha = alpha * popcount(coverage) / SI_NUM_SMOOTH_AA_SAMPLES */
2597 coverage
= LLVMGetParam(ctx
->main_fn
,
2599 coverage
= ac_to_integer(&ctx
->ac
, coverage
);
2601 coverage
= lp_build_intrinsic(ctx
->ac
.builder
, "llvm.ctpop.i32",
2603 &coverage
, 1, LP_FUNC_ATTR_READNONE
);
2605 coverage
= LLVMBuildUIToFP(ctx
->ac
.builder
, coverage
,
2608 coverage
= LLVMBuildFMul(ctx
->ac
.builder
, coverage
,
2609 LLVMConstReal(ctx
->f32
,
2610 1.0 / SI_NUM_SMOOTH_AA_SAMPLES
), "");
2612 return LLVMBuildFMul(ctx
->ac
.builder
, alpha
, coverage
, "");
2615 static void si_llvm_emit_clipvertex(struct si_shader_context
*ctx
,
2616 struct ac_export_args
*pos
, LLVMValueRef
*out_elts
)
2620 unsigned const_chan
;
2621 LLVMValueRef base_elt
;
2622 LLVMValueRef ptr
= LLVMGetParam(ctx
->main_fn
, ctx
->param_rw_buffers
);
2623 LLVMValueRef constbuf_index
= LLVMConstInt(ctx
->i32
,
2624 SI_VS_CONST_CLIP_PLANES
, 0);
2625 LLVMValueRef const_resource
= ac_build_load_to_sgpr(&ctx
->ac
, ptr
, constbuf_index
);
2627 for (reg_index
= 0; reg_index
< 2; reg_index
++) {
2628 struct ac_export_args
*args
= &pos
[2 + reg_index
];
2633 args
->out
[3] = LLVMConstReal(ctx
->f32
, 0.0f
);
2635 /* Compute dot products of position and user clip plane vectors */
2636 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
2637 for (const_chan
= 0; const_chan
< TGSI_NUM_CHANNELS
; const_chan
++) {
2639 LLVMConstInt(ctx
->i32
, ((reg_index
* 4 + chan
) * 4 +
2640 const_chan
) * 4, 0);
2641 base_elt
= buffer_load_const(ctx
, const_resource
,
2644 lp_build_add(&ctx
->bld_base
.base
, args
->out
[chan
],
2645 lp_build_mul(&ctx
->bld_base
.base
, base_elt
,
2646 out_elts
[const_chan
]));
2650 args
->enabled_channels
= 0xf;
2651 args
->valid_mask
= 0;
2653 args
->target
= V_008DFC_SQ_EXP_POS
+ 2 + reg_index
;
2658 static void si_dump_streamout(struct pipe_stream_output_info
*so
)
2662 if (so
->num_outputs
)
2663 fprintf(stderr
, "STREAMOUT\n");
2665 for (i
= 0; i
< so
->num_outputs
; i
++) {
2666 unsigned mask
= ((1 << so
->output
[i
].num_components
) - 1) <<
2667 so
->output
[i
].start_component
;
2668 fprintf(stderr
, " %i: BUF%i[%i..%i] <- OUT[%i].%s%s%s%s\n",
2669 i
, so
->output
[i
].output_buffer
,
2670 so
->output
[i
].dst_offset
, so
->output
[i
].dst_offset
+ so
->output
[i
].num_components
- 1,
2671 so
->output
[i
].register_index
,
2672 mask
& 1 ? "x" : "",
2673 mask
& 2 ? "y" : "",
2674 mask
& 4 ? "z" : "",
2675 mask
& 8 ? "w" : "");
2679 static void emit_streamout_output(struct si_shader_context
*ctx
,
2680 LLVMValueRef
const *so_buffers
,
2681 LLVMValueRef
const *so_write_offsets
,
2682 struct pipe_stream_output
*stream_out
,
2683 struct si_shader_output_values
*shader_out
)
2685 unsigned buf_idx
= stream_out
->output_buffer
;
2686 unsigned start
= stream_out
->start_component
;
2687 unsigned num_comps
= stream_out
->num_components
;
2688 LLVMValueRef out
[4];
2690 assert(num_comps
&& num_comps
<= 4);
2691 if (!num_comps
|| num_comps
> 4)
2694 /* Load the output as int. */
2695 for (int j
= 0; j
< num_comps
; j
++) {
2696 assert(stream_out
->stream
== shader_out
->vertex_stream
[start
+ j
]);
2698 out
[j
] = ac_to_integer(&ctx
->ac
, shader_out
->values
[start
+ j
]);
2701 /* Pack the output. */
2702 LLVMValueRef vdata
= NULL
;
2704 switch (num_comps
) {
2705 case 1: /* as i32 */
2708 case 2: /* as v2i32 */
2709 case 3: /* as v4i32 (aligned to 4) */
2710 case 4: /* as v4i32 */
2711 vdata
= LLVMGetUndef(LLVMVectorType(ctx
->i32
, util_next_power_of_two(num_comps
)));
2712 for (int j
= 0; j
< num_comps
; j
++) {
2713 vdata
= LLVMBuildInsertElement(ctx
->ac
.builder
, vdata
, out
[j
],
2714 LLVMConstInt(ctx
->i32
, j
, 0), "");
2719 ac_build_buffer_store_dword(&ctx
->ac
, so_buffers
[buf_idx
],
2721 so_write_offsets
[buf_idx
],
2723 stream_out
->dst_offset
* 4, 1, 1, true, false);
2727 * Write streamout data to buffers for vertex stream @p stream (different
2728 * vertex streams can occur for GS copy shaders).
2730 static void si_llvm_emit_streamout(struct si_shader_context
*ctx
,
2731 struct si_shader_output_values
*outputs
,
2732 unsigned noutput
, unsigned stream
)
2734 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
2735 struct pipe_stream_output_info
*so
= &sel
->so
;
2736 LLVMBuilderRef builder
= ctx
->ac
.builder
;
2738 struct lp_build_if_state if_ctx
;
2740 /* Get bits [22:16], i.e. (so_param >> 16) & 127; */
2741 LLVMValueRef so_vtx_count
=
2742 unpack_param(ctx
, ctx
->param_streamout_config
, 16, 7);
2744 LLVMValueRef tid
= ac_get_thread_id(&ctx
->ac
);
2746 /* can_emit = tid < so_vtx_count; */
2747 LLVMValueRef can_emit
=
2748 LLVMBuildICmp(builder
, LLVMIntULT
, tid
, so_vtx_count
, "");
2750 /* Emit the streamout code conditionally. This actually avoids
2751 * out-of-bounds buffer access. The hw tells us via the SGPR
2752 * (so_vtx_count) which threads are allowed to emit streamout data. */
2753 lp_build_if(&if_ctx
, &ctx
->gallivm
, can_emit
);
2755 /* The buffer offset is computed as follows:
2756 * ByteOffset = streamout_offset[buffer_id]*4 +
2757 * (streamout_write_index + thread_id)*stride[buffer_id] +
2761 LLVMValueRef so_write_index
=
2762 LLVMGetParam(ctx
->main_fn
,
2763 ctx
->param_streamout_write_index
);
2765 /* Compute (streamout_write_index + thread_id). */
2766 so_write_index
= LLVMBuildAdd(builder
, so_write_index
, tid
, "");
2768 /* Load the descriptor and compute the write offset for each
2769 * enabled buffer. */
2770 LLVMValueRef so_write_offset
[4] = {};
2771 LLVMValueRef so_buffers
[4];
2772 LLVMValueRef buf_ptr
= LLVMGetParam(ctx
->main_fn
,
2773 ctx
->param_rw_buffers
);
2775 for (i
= 0; i
< 4; i
++) {
2779 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
,
2780 SI_VS_STREAMOUT_BUF0
+ i
, 0);
2782 so_buffers
[i
] = ac_build_load_to_sgpr(&ctx
->ac
, buf_ptr
, offset
);
2784 LLVMValueRef so_offset
= LLVMGetParam(ctx
->main_fn
,
2785 ctx
->param_streamout_offset
[i
]);
2786 so_offset
= LLVMBuildMul(builder
, so_offset
, LLVMConstInt(ctx
->i32
, 4, 0), "");
2788 so_write_offset
[i
] = LLVMBuildMul(builder
, so_write_index
,
2789 LLVMConstInt(ctx
->i32
, so
->stride
[i
]*4, 0), "");
2790 so_write_offset
[i
] = LLVMBuildAdd(builder
, so_write_offset
[i
], so_offset
, "");
2793 /* Write streamout data. */
2794 for (i
= 0; i
< so
->num_outputs
; i
++) {
2795 unsigned reg
= so
->output
[i
].register_index
;
2800 if (stream
!= so
->output
[i
].stream
)
2803 emit_streamout_output(ctx
, so_buffers
, so_write_offset
,
2804 &so
->output
[i
], &outputs
[reg
]);
2807 lp_build_endif(&if_ctx
);
2810 static void si_export_param(struct si_shader_context
*ctx
, unsigned index
,
2811 LLVMValueRef
*values
)
2813 struct ac_export_args args
;
2815 si_llvm_init_export_args(ctx
, values
,
2816 V_008DFC_SQ_EXP_PARAM
+ index
, &args
);
2817 ac_build_export(&ctx
->ac
, &args
);
2820 static void si_build_param_exports(struct si_shader_context
*ctx
,
2821 struct si_shader_output_values
*outputs
,
2824 struct si_shader
*shader
= ctx
->shader
;
2825 unsigned param_count
= 0;
2827 for (unsigned i
= 0; i
< noutput
; i
++) {
2828 unsigned semantic_name
= outputs
[i
].semantic_name
;
2829 unsigned semantic_index
= outputs
[i
].semantic_index
;
2831 if (outputs
[i
].vertex_stream
[0] != 0 &&
2832 outputs
[i
].vertex_stream
[1] != 0 &&
2833 outputs
[i
].vertex_stream
[2] != 0 &&
2834 outputs
[i
].vertex_stream
[3] != 0)
2837 switch (semantic_name
) {
2838 case TGSI_SEMANTIC_LAYER
:
2839 case TGSI_SEMANTIC_VIEWPORT_INDEX
:
2840 case TGSI_SEMANTIC_CLIPDIST
:
2841 case TGSI_SEMANTIC_COLOR
:
2842 case TGSI_SEMANTIC_BCOLOR
:
2843 case TGSI_SEMANTIC_PRIMID
:
2844 case TGSI_SEMANTIC_FOG
:
2845 case TGSI_SEMANTIC_TEXCOORD
:
2846 case TGSI_SEMANTIC_GENERIC
:
2852 if ((semantic_name
!= TGSI_SEMANTIC_GENERIC
||
2853 semantic_index
< SI_MAX_IO_GENERIC
) &&
2854 shader
->key
.opt
.kill_outputs
&
2855 (1ull << si_shader_io_get_unique_index(semantic_name
, semantic_index
)))
2858 si_export_param(ctx
, param_count
, outputs
[i
].values
);
2860 assert(i
< ARRAY_SIZE(shader
->info
.vs_output_param_offset
));
2861 shader
->info
.vs_output_param_offset
[i
] = param_count
++;
2864 shader
->info
.nr_param_exports
= param_count
;
2867 /* Generate export instructions for hardware VS shader stage */
2868 static void si_llvm_export_vs(struct si_shader_context
*ctx
,
2869 struct si_shader_output_values
*outputs
,
2872 struct si_shader
*shader
= ctx
->shader
;
2873 struct ac_export_args pos_args
[4] = {};
2874 LLVMValueRef psize_value
= NULL
, edgeflag_value
= NULL
, layer_value
= NULL
, viewport_index_value
= NULL
;
2878 /* Build position exports. */
2879 for (i
= 0; i
< noutput
; i
++) {
2880 switch (outputs
[i
].semantic_name
) {
2881 case TGSI_SEMANTIC_POSITION
:
2882 si_llvm_init_export_args(ctx
, outputs
[i
].values
,
2883 V_008DFC_SQ_EXP_POS
, &pos_args
[0]);
2885 case TGSI_SEMANTIC_PSIZE
:
2886 psize_value
= outputs
[i
].values
[0];
2888 case TGSI_SEMANTIC_LAYER
:
2889 layer_value
= outputs
[i
].values
[0];
2891 case TGSI_SEMANTIC_VIEWPORT_INDEX
:
2892 viewport_index_value
= outputs
[i
].values
[0];
2894 case TGSI_SEMANTIC_EDGEFLAG
:
2895 edgeflag_value
= outputs
[i
].values
[0];
2897 case TGSI_SEMANTIC_CLIPDIST
:
2898 if (!shader
->key
.opt
.clip_disable
) {
2899 unsigned index
= 2 + outputs
[i
].semantic_index
;
2900 si_llvm_init_export_args(ctx
, outputs
[i
].values
,
2901 V_008DFC_SQ_EXP_POS
+ index
,
2905 case TGSI_SEMANTIC_CLIPVERTEX
:
2906 if (!shader
->key
.opt
.clip_disable
) {
2907 si_llvm_emit_clipvertex(ctx
, pos_args
,
2914 /* We need to add the position output manually if it's missing. */
2915 if (!pos_args
[0].out
[0]) {
2916 pos_args
[0].enabled_channels
= 0xf; /* writemask */
2917 pos_args
[0].valid_mask
= 0; /* EXEC mask */
2918 pos_args
[0].done
= 0; /* last export? */
2919 pos_args
[0].target
= V_008DFC_SQ_EXP_POS
;
2920 pos_args
[0].compr
= 0; /* COMPR flag */
2921 pos_args
[0].out
[0] = ctx
->ac
.f32_0
; /* X */
2922 pos_args
[0].out
[1] = ctx
->ac
.f32_0
; /* Y */
2923 pos_args
[0].out
[2] = ctx
->ac
.f32_0
; /* Z */
2924 pos_args
[0].out
[3] = ctx
->ac
.f32_1
; /* W */
2927 /* Write the misc vector (point size, edgeflag, layer, viewport). */
2928 if (shader
->selector
->info
.writes_psize
||
2929 shader
->selector
->info
.writes_edgeflag
||
2930 shader
->selector
->info
.writes_viewport_index
||
2931 shader
->selector
->info
.writes_layer
) {
2932 pos_args
[1].enabled_channels
= shader
->selector
->info
.writes_psize
|
2933 (shader
->selector
->info
.writes_edgeflag
<< 1) |
2934 (shader
->selector
->info
.writes_layer
<< 2);
2936 pos_args
[1].valid_mask
= 0; /* EXEC mask */
2937 pos_args
[1].done
= 0; /* last export? */
2938 pos_args
[1].target
= V_008DFC_SQ_EXP_POS
+ 1;
2939 pos_args
[1].compr
= 0; /* COMPR flag */
2940 pos_args
[1].out
[0] = ctx
->ac
.f32_0
; /* X */
2941 pos_args
[1].out
[1] = ctx
->ac
.f32_0
; /* Y */
2942 pos_args
[1].out
[2] = ctx
->ac
.f32_0
; /* Z */
2943 pos_args
[1].out
[3] = ctx
->ac
.f32_0
; /* W */
2945 if (shader
->selector
->info
.writes_psize
)
2946 pos_args
[1].out
[0] = psize_value
;
2948 if (shader
->selector
->info
.writes_edgeflag
) {
2949 /* The output is a float, but the hw expects an integer
2950 * with the first bit containing the edge flag. */
2951 edgeflag_value
= LLVMBuildFPToUI(ctx
->ac
.builder
,
2954 edgeflag_value
= ac_build_umin(&ctx
->ac
,
2958 /* The LLVM intrinsic expects a float. */
2959 pos_args
[1].out
[1] = ac_to_float(&ctx
->ac
, edgeflag_value
);
2962 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
2963 /* GFX9 has the layer in out.z[10:0] and the viewport
2964 * index in out.z[19:16].
2966 if (shader
->selector
->info
.writes_layer
)
2967 pos_args
[1].out
[2] = layer_value
;
2969 if (shader
->selector
->info
.writes_viewport_index
) {
2970 LLVMValueRef v
= viewport_index_value
;
2972 v
= ac_to_integer(&ctx
->ac
, v
);
2973 v
= LLVMBuildShl(ctx
->ac
.builder
, v
,
2974 LLVMConstInt(ctx
->i32
, 16, 0), "");
2975 v
= LLVMBuildOr(ctx
->ac
.builder
, v
,
2976 ac_to_integer(&ctx
->ac
, pos_args
[1].out
[2]), "");
2977 pos_args
[1].out
[2] = ac_to_float(&ctx
->ac
, v
);
2978 pos_args
[1].enabled_channels
|= 1 << 2;
2981 if (shader
->selector
->info
.writes_layer
)
2982 pos_args
[1].out
[2] = layer_value
;
2984 if (shader
->selector
->info
.writes_viewport_index
) {
2985 pos_args
[1].out
[3] = viewport_index_value
;
2986 pos_args
[1].enabled_channels
|= 1 << 3;
2991 for (i
= 0; i
< 4; i
++)
2992 if (pos_args
[i
].out
[0])
2993 shader
->info
.nr_pos_exports
++;
2996 for (i
= 0; i
< 4; i
++) {
2997 if (!pos_args
[i
].out
[0])
3000 /* Specify the target we are exporting */
3001 pos_args
[i
].target
= V_008DFC_SQ_EXP_POS
+ pos_idx
++;
3003 if (pos_idx
== shader
->info
.nr_pos_exports
)
3004 /* Specify that this is the last export */
3005 pos_args
[i
].done
= 1;
3007 ac_build_export(&ctx
->ac
, &pos_args
[i
]);
3010 /* Build parameter exports. */
3011 si_build_param_exports(ctx
, outputs
, noutput
);
3015 * Forward all outputs from the vertex shader to the TES. This is only used
3016 * for the fixed function TCS.
3018 static void si_copy_tcs_inputs(struct lp_build_tgsi_context
*bld_base
)
3020 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3021 LLVMValueRef invocation_id
, buffer
, buffer_offset
;
3022 LLVMValueRef lds_vertex_stride
, lds_vertex_offset
, lds_base
;
3025 invocation_id
= unpack_llvm_param(ctx
, ctx
->abi
.tcs_rel_ids
, 8, 5);
3026 buffer
= get_tess_ring_descriptor(ctx
, TESS_OFFCHIP_RING_TCS
);
3027 buffer_offset
= LLVMGetParam(ctx
->main_fn
, ctx
->param_tcs_offchip_offset
);
3029 lds_vertex_stride
= get_tcs_in_vertex_dw_stride(ctx
);
3030 lds_vertex_offset
= LLVMBuildMul(ctx
->ac
.builder
, invocation_id
,
3031 lds_vertex_stride
, "");
3032 lds_base
= get_tcs_in_current_patch_offset(ctx
);
3033 lds_base
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
, lds_vertex_offset
, "");
3035 inputs
= ctx
->shader
->key
.mono
.u
.ff_tcs_inputs_to_copy
;
3037 unsigned i
= u_bit_scan64(&inputs
);
3039 LLVMValueRef lds_ptr
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
3040 LLVMConstInt(ctx
->i32
, 4 * i
, 0),
3043 LLVMValueRef buffer_addr
= get_tcs_tes_buffer_address(ctx
,
3044 get_rel_patch_id(ctx
),
3046 LLVMConstInt(ctx
->i32
, i
, 0));
3048 LLVMValueRef value
= lds_load(bld_base
, ctx
->ac
.i32
, ~0,
3051 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, value
, 4, buffer_addr
,
3052 buffer_offset
, 0, 1, 0, true, false);
3056 static void si_write_tess_factors(struct lp_build_tgsi_context
*bld_base
,
3057 LLVMValueRef rel_patch_id
,
3058 LLVMValueRef invocation_id
,
3059 LLVMValueRef tcs_out_current_patch_data_offset
,
3060 LLVMValueRef invoc0_tf_outer
[4],
3061 LLVMValueRef invoc0_tf_inner
[2])
3063 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3064 struct si_shader
*shader
= ctx
->shader
;
3065 unsigned tess_inner_index
, tess_outer_index
;
3066 LLVMValueRef lds_base
, lds_inner
, lds_outer
, byteoffset
, buffer
;
3067 LLVMValueRef out
[6], vec0
, vec1
, tf_base
, inner
[4], outer
[4];
3068 unsigned stride
, outer_comps
, inner_comps
, i
, offset
;
3069 struct lp_build_if_state if_ctx
, inner_if_ctx
;
3071 /* Add a barrier before loading tess factors from LDS. */
3072 if (!shader
->key
.part
.tcs
.epilog
.invoc0_tess_factors_are_def
)
3073 si_llvm_emit_barrier(NULL
, bld_base
, NULL
);
3075 /* Do this only for invocation 0, because the tess levels are per-patch,
3078 * This can't jump, because invocation 0 executes this. It should
3079 * at least mask out the loads and stores for other invocations.
3081 lp_build_if(&if_ctx
, &ctx
->gallivm
,
3082 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
,
3083 invocation_id
, ctx
->i32_0
, ""));
3085 /* Determine the layout of one tess factor element in the buffer. */
3086 switch (shader
->key
.part
.tcs
.epilog
.prim_mode
) {
3087 case PIPE_PRIM_LINES
:
3088 stride
= 2; /* 2 dwords, 1 vec2 store */
3092 case PIPE_PRIM_TRIANGLES
:
3093 stride
= 4; /* 4 dwords, 1 vec4 store */
3097 case PIPE_PRIM_QUADS
:
3098 stride
= 6; /* 6 dwords, 2 stores (vec4 + vec2) */
3107 for (i
= 0; i
< 4; i
++) {
3108 inner
[i
] = LLVMGetUndef(ctx
->i32
);
3109 outer
[i
] = LLVMGetUndef(ctx
->i32
);
3112 if (shader
->key
.part
.tcs
.epilog
.invoc0_tess_factors_are_def
) {
3113 /* Tess factors are in VGPRs. */
3114 for (i
= 0; i
< outer_comps
; i
++)
3115 outer
[i
] = out
[i
] = invoc0_tf_outer
[i
];
3116 for (i
= 0; i
< inner_comps
; i
++)
3117 inner
[i
] = out
[outer_comps
+i
] = invoc0_tf_inner
[i
];
3119 /* Load tess_inner and tess_outer from LDS.
3120 * Any invocation can write them, so we can't get them from a temporary.
3122 tess_inner_index
= si_shader_io_get_unique_index_patch(TGSI_SEMANTIC_TESSINNER
, 0);
3123 tess_outer_index
= si_shader_io_get_unique_index_patch(TGSI_SEMANTIC_TESSOUTER
, 0);
3125 lds_base
= tcs_out_current_patch_data_offset
;
3126 lds_inner
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
3127 LLVMConstInt(ctx
->i32
,
3128 tess_inner_index
* 4, 0), "");
3129 lds_outer
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
3130 LLVMConstInt(ctx
->i32
,
3131 tess_outer_index
* 4, 0), "");
3133 for (i
= 0; i
< outer_comps
; i
++) {
3135 lds_load(bld_base
, ctx
->ac
.i32
, i
, lds_outer
);
3137 for (i
= 0; i
< inner_comps
; i
++) {
3138 inner
[i
] = out
[outer_comps
+i
] =
3139 lds_load(bld_base
, ctx
->ac
.i32
, i
, lds_inner
);
3143 if (shader
->key
.part
.tcs
.epilog
.prim_mode
== PIPE_PRIM_LINES
) {
3144 /* For isolines, the hardware expects tess factors in the
3145 * reverse order from what GLSL / TGSI specify.
3147 LLVMValueRef tmp
= out
[0];
3152 /* Convert the outputs to vectors for stores. */
3153 vec0
= lp_build_gather_values(&ctx
->gallivm
, out
, MIN2(stride
, 4));
3157 vec1
= lp_build_gather_values(&ctx
->gallivm
, out
+4, stride
- 4);
3159 /* Get the buffer. */
3160 buffer
= get_tess_ring_descriptor(ctx
, TCS_FACTOR_RING
);
3162 /* Get the offset. */
3163 tf_base
= LLVMGetParam(ctx
->main_fn
,
3164 ctx
->param_tcs_factor_offset
);
3165 byteoffset
= LLVMBuildMul(ctx
->ac
.builder
, rel_patch_id
,
3166 LLVMConstInt(ctx
->i32
, 4 * stride
, 0), "");
3168 lp_build_if(&inner_if_ctx
, &ctx
->gallivm
,
3169 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
,
3170 rel_patch_id
, ctx
->i32_0
, ""));
3172 /* Store the dynamic HS control word. */
3174 if (ctx
->screen
->info
.chip_class
<= VI
) {
3175 ac_build_buffer_store_dword(&ctx
->ac
, buffer
,
3176 LLVMConstInt(ctx
->i32
, 0x80000000, 0),
3177 1, ctx
->i32_0
, tf_base
,
3178 offset
, 1, 0, true, false);
3182 lp_build_endif(&inner_if_ctx
);
3184 /* Store the tessellation factors. */
3185 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec0
,
3186 MIN2(stride
, 4), byteoffset
, tf_base
,
3187 offset
, 1, 0, true, false);
3190 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec1
,
3191 stride
- 4, byteoffset
, tf_base
,
3192 offset
, 1, 0, true, false);
3194 /* Store the tess factors into the offchip buffer if TES reads them. */
3195 if (shader
->key
.part
.tcs
.epilog
.tes_reads_tess_factors
) {
3196 LLVMValueRef buf
, base
, inner_vec
, outer_vec
, tf_outer_offset
;
3197 LLVMValueRef tf_inner_offset
;
3198 unsigned param_outer
, param_inner
;
3200 buf
= get_tess_ring_descriptor(ctx
, TESS_OFFCHIP_RING_TCS
);
3201 base
= LLVMGetParam(ctx
->main_fn
, ctx
->param_tcs_offchip_offset
);
3203 param_outer
= si_shader_io_get_unique_index_patch(
3204 TGSI_SEMANTIC_TESSOUTER
, 0);
3205 tf_outer_offset
= get_tcs_tes_buffer_address(ctx
, rel_patch_id
, NULL
,
3206 LLVMConstInt(ctx
->i32
, param_outer
, 0));
3208 outer_vec
= lp_build_gather_values(&ctx
->gallivm
, outer
,
3209 util_next_power_of_two(outer_comps
));
3211 ac_build_buffer_store_dword(&ctx
->ac
, buf
, outer_vec
,
3212 outer_comps
, tf_outer_offset
,
3213 base
, 0, 1, 0, true, false);
3215 param_inner
= si_shader_io_get_unique_index_patch(
3216 TGSI_SEMANTIC_TESSINNER
, 0);
3217 tf_inner_offset
= get_tcs_tes_buffer_address(ctx
, rel_patch_id
, NULL
,
3218 LLVMConstInt(ctx
->i32
, param_inner
, 0));
3220 inner_vec
= inner_comps
== 1 ? inner
[0] :
3221 lp_build_gather_values(&ctx
->gallivm
, inner
, inner_comps
);
3222 ac_build_buffer_store_dword(&ctx
->ac
, buf
, inner_vec
,
3223 inner_comps
, tf_inner_offset
,
3224 base
, 0, 1, 0, true, false);
3228 lp_build_endif(&if_ctx
);
3232 si_insert_input_ret(struct si_shader_context
*ctx
, LLVMValueRef ret
,
3233 unsigned param
, unsigned return_index
)
3235 return LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
3236 LLVMGetParam(ctx
->main_fn
, param
),
3241 si_insert_input_ret_float(struct si_shader_context
*ctx
, LLVMValueRef ret
,
3242 unsigned param
, unsigned return_index
)
3244 LLVMBuilderRef builder
= ctx
->ac
.builder
;
3245 LLVMValueRef p
= LLVMGetParam(ctx
->main_fn
, param
);
3247 return LLVMBuildInsertValue(builder
, ret
,
3248 ac_to_float(&ctx
->ac
, p
),
3253 si_insert_input_ptr(struct si_shader_context
*ctx
, LLVMValueRef ret
,
3254 unsigned param
, unsigned return_index
)
3256 LLVMBuilderRef builder
= ctx
->ac
.builder
;
3257 LLVMValueRef ptr
, lo
, hi
;
3259 if (HAVE_32BIT_POINTERS
) {
3260 ptr
= LLVMGetParam(ctx
->main_fn
, param
);
3261 ptr
= LLVMBuildPtrToInt(builder
, ptr
, ctx
->i32
, "");
3262 return LLVMBuildInsertValue(builder
, ret
, ptr
, return_index
, "");
3265 ptr
= LLVMGetParam(ctx
->main_fn
, param
);
3266 ptr
= LLVMBuildPtrToInt(builder
, ptr
, ctx
->i64
, "");
3267 ptr
= LLVMBuildBitCast(builder
, ptr
, ctx
->v2i32
, "");
3268 lo
= LLVMBuildExtractElement(builder
, ptr
, ctx
->i32_0
, "");
3269 hi
= LLVMBuildExtractElement(builder
, ptr
, ctx
->i32_1
, "");
3270 ret
= LLVMBuildInsertValue(builder
, ret
, lo
, return_index
, "");
3271 return LLVMBuildInsertValue(builder
, ret
, hi
, return_index
+ 1, "");
3274 /* This only writes the tessellation factor levels. */
3275 static void si_llvm_emit_tcs_epilogue(struct ac_shader_abi
*abi
,
3276 unsigned max_outputs
,
3277 LLVMValueRef
*addrs
)
3279 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
3280 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
3281 LLVMBuilderRef builder
= ctx
->ac
.builder
;
3282 LLVMValueRef rel_patch_id
, invocation_id
, tf_lds_offset
;
3284 si_copy_tcs_inputs(bld_base
);
3286 rel_patch_id
= get_rel_patch_id(ctx
);
3287 invocation_id
= unpack_llvm_param(ctx
, ctx
->abi
.tcs_rel_ids
, 8, 5);
3288 tf_lds_offset
= get_tcs_out_current_patch_data_offset(ctx
);
3290 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
3291 LLVMBasicBlockRef blocks
[2] = {
3292 LLVMGetInsertBlock(builder
),
3293 ctx
->merged_wrap_if_state
.entry_block
3295 LLVMValueRef values
[2];
3297 lp_build_endif(&ctx
->merged_wrap_if_state
);
3299 values
[0] = rel_patch_id
;
3300 values
[1] = LLVMGetUndef(ctx
->i32
);
3301 rel_patch_id
= ac_build_phi(&ctx
->ac
, ctx
->i32
, 2, values
, blocks
);
3303 values
[0] = tf_lds_offset
;
3304 values
[1] = LLVMGetUndef(ctx
->i32
);
3305 tf_lds_offset
= ac_build_phi(&ctx
->ac
, ctx
->i32
, 2, values
, blocks
);
3307 values
[0] = invocation_id
;
3308 values
[1] = ctx
->i32_1
; /* cause the epilog to skip threads */
3309 invocation_id
= ac_build_phi(&ctx
->ac
, ctx
->i32
, 2, values
, blocks
);
3312 /* Return epilog parameters from this function. */
3313 LLVMValueRef ret
= ctx
->return_value
;
3316 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
3317 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_layout
,
3318 8 + GFX9_SGPR_TCS_OFFCHIP_LAYOUT
);
3319 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_out_lds_layout
,
3320 8 + GFX9_SGPR_TCS_OUT_LAYOUT
);
3321 /* Tess offchip and tess factor offsets are at the beginning. */
3322 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_offset
, 2);
3323 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_factor_offset
, 4);
3324 vgpr
= 8 + GFX9_SGPR_TCS_OUT_LAYOUT
+ 1;
3326 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_layout
,
3327 GFX6_SGPR_TCS_OFFCHIP_LAYOUT
);
3328 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_out_lds_layout
,
3329 GFX6_SGPR_TCS_OUT_LAYOUT
);
3330 /* Tess offchip and tess factor offsets are after user SGPRs. */
3331 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_offset
,
3332 GFX6_TCS_NUM_USER_SGPR
);
3333 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_factor_offset
,
3334 GFX6_TCS_NUM_USER_SGPR
+ 1);
3335 vgpr
= GFX6_TCS_NUM_USER_SGPR
+ 2;
3339 rel_patch_id
= ac_to_float(&ctx
->ac
, rel_patch_id
);
3340 invocation_id
= ac_to_float(&ctx
->ac
, invocation_id
);
3341 tf_lds_offset
= ac_to_float(&ctx
->ac
, tf_lds_offset
);
3343 /* Leave a hole corresponding to the two input VGPRs. This ensures that
3344 * the invocation_id output does not alias the tcs_rel_ids input,
3345 * which saves a V_MOV on gfx9.
3349 ret
= LLVMBuildInsertValue(builder
, ret
, rel_patch_id
, vgpr
++, "");
3350 ret
= LLVMBuildInsertValue(builder
, ret
, invocation_id
, vgpr
++, "");
3352 if (ctx
->shader
->selector
->tcs_info
.tessfactors_are_def_in_all_invocs
) {
3353 vgpr
++; /* skip the tess factor LDS offset */
3354 for (unsigned i
= 0; i
< 6; i
++) {
3355 LLVMValueRef value
=
3356 LLVMBuildLoad(builder
, ctx
->invoc0_tess_factors
[i
], "");
3357 value
= ac_to_float(&ctx
->ac
, value
);
3358 ret
= LLVMBuildInsertValue(builder
, ret
, value
, vgpr
++, "");
3361 ret
= LLVMBuildInsertValue(builder
, ret
, tf_lds_offset
, vgpr
++, "");
3363 ctx
->return_value
= ret
;
3366 /* Pass TCS inputs from LS to TCS on GFX9. */
3367 static void si_set_ls_return_value_for_tcs(struct si_shader_context
*ctx
)
3369 LLVMValueRef ret
= ctx
->return_value
;
3371 ret
= si_insert_input_ptr(ctx
, ret
, 0, 0);
3372 if (HAVE_32BIT_POINTERS
)
3373 ret
= si_insert_input_ptr(ctx
, ret
, 1, 1);
3374 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_offset
, 2);
3375 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_merged_wave_info
, 3);
3376 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_factor_offset
, 4);
3377 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_merged_scratch_offset
, 5);
3379 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->param_rw_buffers
,
3380 8 + SI_SGPR_RW_BUFFERS
);
3381 ret
= si_insert_input_ptr(ctx
, ret
,
3382 ctx
->param_bindless_samplers_and_images
,
3383 8 + SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES
);
3385 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_vs_state_bits
,
3386 8 + SI_SGPR_VS_STATE_BITS
);
3388 #if !HAVE_32BIT_POINTERS
3389 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->param_vs_state_bits
+ 1,
3390 8 + GFX9_SGPR_2ND_SAMPLERS_AND_IMAGES
);
3393 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_layout
,
3394 8 + GFX9_SGPR_TCS_OFFCHIP_LAYOUT
);
3395 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_out_lds_offsets
,
3396 8 + GFX9_SGPR_TCS_OUT_OFFSETS
);
3397 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_out_lds_layout
,
3398 8 + GFX9_SGPR_TCS_OUT_LAYOUT
);
3400 unsigned vgpr
= 8 + GFX9_TCS_NUM_USER_SGPR
;
3401 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
3402 ac_to_float(&ctx
->ac
, ctx
->abi
.tcs_patch_id
),
3404 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
3405 ac_to_float(&ctx
->ac
, ctx
->abi
.tcs_rel_ids
),
3407 ctx
->return_value
= ret
;
3410 /* Pass GS inputs from ES to GS on GFX9. */
3411 static void si_set_es_return_value_for_gs(struct si_shader_context
*ctx
)
3413 LLVMValueRef ret
= ctx
->return_value
;
3415 ret
= si_insert_input_ptr(ctx
, ret
, 0, 0);
3416 if (HAVE_32BIT_POINTERS
)
3417 ret
= si_insert_input_ptr(ctx
, ret
, 1, 1);
3418 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_gs2vs_offset
, 2);
3419 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_merged_wave_info
, 3);
3420 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_merged_scratch_offset
, 5);
3422 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->param_rw_buffers
,
3423 8 + SI_SGPR_RW_BUFFERS
);
3424 ret
= si_insert_input_ptr(ctx
, ret
,
3425 ctx
->param_bindless_samplers_and_images
,
3426 8 + SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES
);
3428 #if !HAVE_32BIT_POINTERS
3429 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->param_vs_state_bits
+ 1,
3430 8 + GFX9_SGPR_2ND_SAMPLERS_AND_IMAGES
);
3434 if (ctx
->type
== PIPE_SHADER_VERTEX
)
3435 vgpr
= 8 + GFX9_VSGS_NUM_USER_SGPR
;
3437 vgpr
= 8 + GFX9_TESGS_NUM_USER_SGPR
;
3439 for (unsigned i
= 0; i
< 5; i
++) {
3440 unsigned param
= ctx
->param_gs_vtx01_offset
+ i
;
3441 ret
= si_insert_input_ret_float(ctx
, ret
, param
, vgpr
++);
3443 ctx
->return_value
= ret
;
3446 static void si_llvm_emit_ls_epilogue(struct ac_shader_abi
*abi
,
3447 unsigned max_outputs
,
3448 LLVMValueRef
*addrs
)
3450 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
3451 struct si_shader
*shader
= ctx
->shader
;
3452 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
3454 LLVMValueRef vertex_id
= LLVMGetParam(ctx
->main_fn
,
3455 ctx
->param_rel_auto_id
);
3456 LLVMValueRef vertex_dw_stride
= get_tcs_in_vertex_dw_stride(ctx
);
3457 LLVMValueRef base_dw_addr
= LLVMBuildMul(ctx
->ac
.builder
, vertex_id
,
3458 vertex_dw_stride
, "");
3460 /* Write outputs to LDS. The next shader (TCS aka HS) will read
3461 * its inputs from it. */
3462 for (i
= 0; i
< info
->num_outputs
; i
++) {
3463 unsigned name
= info
->output_semantic_name
[i
];
3464 unsigned index
= info
->output_semantic_index
[i
];
3466 /* The ARB_shader_viewport_layer_array spec contains the
3469 * 2) What happens if gl_ViewportIndex or gl_Layer is
3470 * written in the vertex shader and a geometry shader is
3473 * RESOLVED: The value written by the last vertex processing
3474 * stage is used. If the last vertex processing stage
3475 * (vertex, tessellation evaluation or geometry) does not
3476 * statically assign to gl_ViewportIndex or gl_Layer, index
3477 * or layer zero is assumed.
3479 * So writes to those outputs in VS-as-LS are simply ignored.
3481 if (name
== TGSI_SEMANTIC_LAYER
||
3482 name
== TGSI_SEMANTIC_VIEWPORT_INDEX
)
3485 int param
= si_shader_io_get_unique_index(name
, index
);
3486 LLVMValueRef dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_dw_addr
,
3487 LLVMConstInt(ctx
->i32
, param
* 4, 0), "");
3489 for (chan
= 0; chan
< 4; chan
++) {
3490 if (!(info
->output_usagemask
[i
] & (1 << chan
)))
3493 lds_store(ctx
, chan
, dw_addr
,
3494 LLVMBuildLoad(ctx
->ac
.builder
, addrs
[4 * i
+ chan
], ""));
3498 if (ctx
->screen
->info
.chip_class
>= GFX9
)
3499 si_set_ls_return_value_for_tcs(ctx
);
3502 static void si_llvm_emit_es_epilogue(struct ac_shader_abi
*abi
,
3503 unsigned max_outputs
,
3504 LLVMValueRef
*addrs
)
3506 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
3507 struct si_shader
*es
= ctx
->shader
;
3508 struct tgsi_shader_info
*info
= &es
->selector
->info
;
3509 LLVMValueRef soffset
= LLVMGetParam(ctx
->main_fn
,
3510 ctx
->param_es2gs_offset
);
3511 LLVMValueRef lds_base
= NULL
;
3515 if (ctx
->screen
->info
.chip_class
>= GFX9
&& info
->num_outputs
) {
3516 unsigned itemsize_dw
= es
->selector
->esgs_itemsize
/ 4;
3517 LLVMValueRef vertex_idx
= ac_get_thread_id(&ctx
->ac
);
3518 LLVMValueRef wave_idx
= unpack_param(ctx
, ctx
->param_merged_wave_info
, 24, 4);
3519 vertex_idx
= LLVMBuildOr(ctx
->ac
.builder
, vertex_idx
,
3520 LLVMBuildMul(ctx
->ac
.builder
, wave_idx
,
3521 LLVMConstInt(ctx
->i32
, 64, false), ""), "");
3522 lds_base
= LLVMBuildMul(ctx
->ac
.builder
, vertex_idx
,
3523 LLVMConstInt(ctx
->i32
, itemsize_dw
, 0), "");
3526 for (i
= 0; i
< info
->num_outputs
; i
++) {
3529 if (info
->output_semantic_name
[i
] == TGSI_SEMANTIC_VIEWPORT_INDEX
||
3530 info
->output_semantic_name
[i
] == TGSI_SEMANTIC_LAYER
)
3533 param
= si_shader_io_get_unique_index(info
->output_semantic_name
[i
],
3534 info
->output_semantic_index
[i
]);
3536 for (chan
= 0; chan
< 4; chan
++) {
3537 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->ac
.builder
, addrs
[4 * i
+ chan
], "");
3538 out_val
= ac_to_integer(&ctx
->ac
, out_val
);
3540 /* GFX9 has the ESGS ring in LDS. */
3541 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
3542 lds_store(ctx
, param
* 4 + chan
, lds_base
, out_val
);
3546 ac_build_buffer_store_dword(&ctx
->ac
,
3548 out_val
, 1, NULL
, soffset
,
3549 (4 * param
+ chan
) * 4,
3554 if (ctx
->screen
->info
.chip_class
>= GFX9
)
3555 si_set_es_return_value_for_gs(ctx
);
3558 static LLVMValueRef
si_get_gs_wave_id(struct si_shader_context
*ctx
)
3560 if (ctx
->screen
->info
.chip_class
>= GFX9
)
3561 return unpack_param(ctx
, ctx
->param_merged_wave_info
, 16, 8);
3563 return LLVMGetParam(ctx
->main_fn
, ctx
->param_gs_wave_id
);
3566 static void emit_gs_epilogue(struct si_shader_context
*ctx
)
3568 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_NOP
| AC_SENDMSG_GS_DONE
,
3569 si_get_gs_wave_id(ctx
));
3571 if (ctx
->screen
->info
.chip_class
>= GFX9
)
3572 lp_build_endif(&ctx
->merged_wrap_if_state
);
3575 static void si_llvm_emit_gs_epilogue(struct ac_shader_abi
*abi
,
3576 unsigned max_outputs
,
3577 LLVMValueRef
*addrs
)
3579 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
3580 struct tgsi_shader_info UNUSED
*info
= &ctx
->shader
->selector
->info
;
3582 assert(info
->num_outputs
<= max_outputs
);
3584 emit_gs_epilogue(ctx
);
3587 static void si_tgsi_emit_gs_epilogue(struct lp_build_tgsi_context
*bld_base
)
3589 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3590 emit_gs_epilogue(ctx
);
3593 static void si_llvm_emit_vs_epilogue(struct ac_shader_abi
*abi
,
3594 unsigned max_outputs
,
3595 LLVMValueRef
*addrs
)
3597 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
3598 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
3599 struct si_shader_output_values
*outputs
= NULL
;
3602 assert(!ctx
->shader
->is_gs_copy_shader
);
3603 assert(info
->num_outputs
<= max_outputs
);
3605 outputs
= MALLOC((info
->num_outputs
+ 1) * sizeof(outputs
[0]));
3607 /* Vertex color clamping.
3609 * This uses a state constant loaded in a user data SGPR and
3610 * an IF statement is added that clamps all colors if the constant
3613 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
3614 struct lp_build_if_state if_ctx
;
3615 LLVMValueRef cond
= NULL
;
3616 LLVMValueRef addr
, val
;
3618 for (i
= 0; i
< info
->num_outputs
; i
++) {
3619 if (info
->output_semantic_name
[i
] != TGSI_SEMANTIC_COLOR
&&
3620 info
->output_semantic_name
[i
] != TGSI_SEMANTIC_BCOLOR
)
3623 /* We've found a color. */
3625 /* The state is in the first bit of the user SGPR. */
3626 cond
= LLVMGetParam(ctx
->main_fn
,
3627 ctx
->param_vs_state_bits
);
3628 cond
= LLVMBuildTrunc(ctx
->ac
.builder
, cond
,
3630 lp_build_if(&if_ctx
, &ctx
->gallivm
, cond
);
3633 for (j
= 0; j
< 4; j
++) {
3634 addr
= addrs
[4 * i
+ j
];
3635 val
= LLVMBuildLoad(ctx
->ac
.builder
, addr
, "");
3636 val
= ac_build_clamp(&ctx
->ac
, val
);
3637 LLVMBuildStore(ctx
->ac
.builder
, val
, addr
);
3642 lp_build_endif(&if_ctx
);
3645 for (i
= 0; i
< info
->num_outputs
; i
++) {
3646 outputs
[i
].semantic_name
= info
->output_semantic_name
[i
];
3647 outputs
[i
].semantic_index
= info
->output_semantic_index
[i
];
3649 for (j
= 0; j
< 4; j
++) {
3650 outputs
[i
].values
[j
] =
3651 LLVMBuildLoad(ctx
->ac
.builder
,
3654 outputs
[i
].vertex_stream
[j
] =
3655 (info
->output_streams
[i
] >> (2 * j
)) & 3;
3659 if (ctx
->shader
->selector
->so
.num_outputs
)
3660 si_llvm_emit_streamout(ctx
, outputs
, i
, 0);
3662 /* Export PrimitiveID. */
3663 if (ctx
->shader
->key
.mono
.u
.vs_export_prim_id
) {
3664 outputs
[i
].semantic_name
= TGSI_SEMANTIC_PRIMID
;
3665 outputs
[i
].semantic_index
= 0;
3666 outputs
[i
].values
[0] = ac_to_float(&ctx
->ac
, get_primitive_id(ctx
, 0));
3667 for (j
= 1; j
< 4; j
++)
3668 outputs
[i
].values
[j
] = LLVMConstReal(ctx
->f32
, 0);
3670 memset(outputs
[i
].vertex_stream
, 0,
3671 sizeof(outputs
[i
].vertex_stream
));
3675 si_llvm_export_vs(ctx
, outputs
, i
);
3679 static void si_tgsi_emit_epilogue(struct lp_build_tgsi_context
*bld_base
)
3681 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3683 ctx
->abi
.emit_outputs(&ctx
->abi
, RADEON_LLVM_MAX_OUTPUTS
,
3684 &ctx
->outputs
[0][0]);
3687 struct si_ps_exports
{
3689 struct ac_export_args args
[10];
3692 static void si_export_mrt_z(struct lp_build_tgsi_context
*bld_base
,
3693 LLVMValueRef depth
, LLVMValueRef stencil
,
3694 LLVMValueRef samplemask
, struct si_ps_exports
*exp
)
3696 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3697 struct ac_export_args args
;
3699 ac_export_mrt_z(&ctx
->ac
, depth
, stencil
, samplemask
, &args
);
3701 memcpy(&exp
->args
[exp
->num
++], &args
, sizeof(args
));
3704 static void si_export_mrt_color(struct lp_build_tgsi_context
*bld_base
,
3705 LLVMValueRef
*color
, unsigned index
,
3706 unsigned samplemask_param
,
3707 bool is_last
, struct si_ps_exports
*exp
)
3709 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3713 if (ctx
->shader
->key
.part
.ps
.epilog
.clamp_color
)
3714 for (i
= 0; i
< 4; i
++)
3715 color
[i
] = ac_build_clamp(&ctx
->ac
, color
[i
]);
3718 if (ctx
->shader
->key
.part
.ps
.epilog
.alpha_to_one
)
3719 color
[3] = ctx
->ac
.f32_1
;
3723 ctx
->shader
->key
.part
.ps
.epilog
.alpha_func
!= PIPE_FUNC_ALWAYS
)
3724 si_alpha_test(bld_base
, color
[3]);
3726 /* Line & polygon smoothing */
3727 if (ctx
->shader
->key
.part
.ps
.epilog
.poly_line_smoothing
)
3728 color
[3] = si_scale_alpha_by_sample_mask(bld_base
, color
[3],
3731 /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
3732 if (ctx
->shader
->key
.part
.ps
.epilog
.last_cbuf
> 0) {
3733 struct ac_export_args args
[8];
3736 /* Get the export arguments, also find out what the last one is. */
3737 for (c
= 0; c
<= ctx
->shader
->key
.part
.ps
.epilog
.last_cbuf
; c
++) {
3738 si_llvm_init_export_args(ctx
, color
,
3739 V_008DFC_SQ_EXP_MRT
+ c
, &args
[c
]);
3740 if (args
[c
].enabled_channels
)
3744 /* Emit all exports. */
3745 for (c
= 0; c
<= ctx
->shader
->key
.part
.ps
.epilog
.last_cbuf
; c
++) {
3746 if (is_last
&& last
== c
) {
3747 args
[c
].valid_mask
= 1; /* whether the EXEC mask is valid */
3748 args
[c
].done
= 1; /* DONE bit */
3749 } else if (!args
[c
].enabled_channels
)
3750 continue; /* unnecessary NULL export */
3752 memcpy(&exp
->args
[exp
->num
++], &args
[c
], sizeof(args
[c
]));
3755 struct ac_export_args args
;
3758 si_llvm_init_export_args(ctx
, color
, V_008DFC_SQ_EXP_MRT
+ index
,
3761 args
.valid_mask
= 1; /* whether the EXEC mask is valid */
3762 args
.done
= 1; /* DONE bit */
3763 } else if (!args
.enabled_channels
)
3764 return; /* unnecessary NULL export */
3766 memcpy(&exp
->args
[exp
->num
++], &args
, sizeof(args
));
3770 static void si_emit_ps_exports(struct si_shader_context
*ctx
,
3771 struct si_ps_exports
*exp
)
3773 for (unsigned i
= 0; i
< exp
->num
; i
++)
3774 ac_build_export(&ctx
->ac
, &exp
->args
[i
]);
3778 * Return PS outputs in this order:
3780 * v[0:3] = color0.xyzw
3781 * v[4:7] = color1.xyzw
3786 * vN+3 = SampleMaskIn (used for OpenGL smoothing)
3788 * The alpha-ref SGPR is returned via its original location.
3790 static void si_llvm_return_fs_outputs(struct ac_shader_abi
*abi
,
3791 unsigned max_outputs
,
3792 LLVMValueRef
*addrs
)
3794 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
3795 struct si_shader
*shader
= ctx
->shader
;
3796 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
3797 LLVMBuilderRef builder
= ctx
->ac
.builder
;
3798 unsigned i
, j
, first_vgpr
, vgpr
;
3800 LLVMValueRef color
[8][4] = {};
3801 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
3804 if (ctx
->postponed_kill
)
3805 ac_build_kill_if_false(&ctx
->ac
, LLVMBuildLoad(builder
, ctx
->postponed_kill
, ""));
3807 /* Read the output values. */
3808 for (i
= 0; i
< info
->num_outputs
; i
++) {
3809 unsigned semantic_name
= info
->output_semantic_name
[i
];
3810 unsigned semantic_index
= info
->output_semantic_index
[i
];
3812 switch (semantic_name
) {
3813 case TGSI_SEMANTIC_COLOR
:
3814 assert(semantic_index
< 8);
3815 for (j
= 0; j
< 4; j
++) {
3816 LLVMValueRef ptr
= addrs
[4 * i
+ j
];
3817 LLVMValueRef result
= LLVMBuildLoad(builder
, ptr
, "");
3818 color
[semantic_index
][j
] = result
;
3821 case TGSI_SEMANTIC_POSITION
:
3822 depth
= LLVMBuildLoad(builder
,
3823 addrs
[4 * i
+ 2], "");
3825 case TGSI_SEMANTIC_STENCIL
:
3826 stencil
= LLVMBuildLoad(builder
,
3827 addrs
[4 * i
+ 1], "");
3829 case TGSI_SEMANTIC_SAMPLEMASK
:
3830 samplemask
= LLVMBuildLoad(builder
,
3831 addrs
[4 * i
+ 0], "");
3834 fprintf(stderr
, "Warning: SI unhandled fs output type:%d\n",
3839 /* Fill the return structure. */
3840 ret
= ctx
->return_value
;
3843 ret
= LLVMBuildInsertValue(builder
, ret
,
3844 ac_to_integer(&ctx
->ac
,
3845 LLVMGetParam(ctx
->main_fn
,
3846 SI_PARAM_ALPHA_REF
)),
3847 SI_SGPR_ALPHA_REF
, "");
3850 first_vgpr
= vgpr
= SI_SGPR_ALPHA_REF
+ 1;
3851 for (i
= 0; i
< ARRAY_SIZE(color
); i
++) {
3855 for (j
= 0; j
< 4; j
++)
3856 ret
= LLVMBuildInsertValue(builder
, ret
, color
[i
][j
], vgpr
++, "");
3859 ret
= LLVMBuildInsertValue(builder
, ret
, depth
, vgpr
++, "");
3861 ret
= LLVMBuildInsertValue(builder
, ret
, stencil
, vgpr
++, "");
3863 ret
= LLVMBuildInsertValue(builder
, ret
, samplemask
, vgpr
++, "");
3865 /* Add the input sample mask for smoothing at the end. */
3866 if (vgpr
< first_vgpr
+ PS_EPILOG_SAMPLEMASK_MIN_LOC
)
3867 vgpr
= first_vgpr
+ PS_EPILOG_SAMPLEMASK_MIN_LOC
;
3868 ret
= LLVMBuildInsertValue(builder
, ret
,
3869 LLVMGetParam(ctx
->main_fn
,
3870 SI_PARAM_SAMPLE_COVERAGE
), vgpr
++, "");
3872 ctx
->return_value
= ret
;
3875 static void membar_emit(
3876 const struct lp_build_tgsi_action
*action
,
3877 struct lp_build_tgsi_context
*bld_base
,
3878 struct lp_build_emit_data
*emit_data
)
3880 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3881 LLVMValueRef src0
= lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, 0);
3882 unsigned flags
= LLVMConstIntGetZExtValue(src0
);
3883 unsigned waitcnt
= NOOP_WAITCNT
;
3885 if (flags
& TGSI_MEMBAR_THREAD_GROUP
)
3886 waitcnt
&= VM_CNT
& LGKM_CNT
;
3888 if (flags
& (TGSI_MEMBAR_ATOMIC_BUFFER
|
3889 TGSI_MEMBAR_SHADER_BUFFER
|
3890 TGSI_MEMBAR_SHADER_IMAGE
))
3893 if (flags
& TGSI_MEMBAR_SHARED
)
3894 waitcnt
&= LGKM_CNT
;
3896 if (waitcnt
!= NOOP_WAITCNT
)
3897 ac_build_waitcnt(&ctx
->ac
, waitcnt
);
3900 static void clock_emit(
3901 const struct lp_build_tgsi_action
*action
,
3902 struct lp_build_tgsi_context
*bld_base
,
3903 struct lp_build_emit_data
*emit_data
)
3905 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3906 LLVMValueRef tmp
= ac_build_shader_clock(&ctx
->ac
);
3908 emit_data
->output
[0] =
3909 LLVMBuildExtractElement(ctx
->ac
.builder
, tmp
, ctx
->i32_0
, "");
3910 emit_data
->output
[1] =
3911 LLVMBuildExtractElement(ctx
->ac
.builder
, tmp
, ctx
->i32_1
, "");
3914 static void si_llvm_emit_ddxy(
3915 const struct lp_build_tgsi_action
*action
,
3916 struct lp_build_tgsi_context
*bld_base
,
3917 struct lp_build_emit_data
*emit_data
)
3919 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3920 unsigned opcode
= emit_data
->info
->opcode
;
3925 if (opcode
== TGSI_OPCODE_DDX_FINE
)
3926 mask
= AC_TID_MASK_LEFT
;
3927 else if (opcode
== TGSI_OPCODE_DDY_FINE
)
3928 mask
= AC_TID_MASK_TOP
;
3930 mask
= AC_TID_MASK_TOP_LEFT
;
3932 /* for DDX we want to next X pixel, DDY next Y pixel. */
3933 idx
= (opcode
== TGSI_OPCODE_DDX
|| opcode
== TGSI_OPCODE_DDX_FINE
) ? 1 : 2;
3935 val
= ac_to_integer(&ctx
->ac
, emit_data
->args
[0]);
3936 val
= ac_build_ddxy(&ctx
->ac
, mask
, idx
, val
);
3937 emit_data
->output
[emit_data
->chan
] = val
;
3941 * this takes an I,J coordinate pair,
3942 * and works out the X and Y derivatives.
3943 * it returns DDX(I), DDX(J), DDY(I), DDY(J).
3945 static LLVMValueRef
si_llvm_emit_ddxy_interp(
3946 struct lp_build_tgsi_context
*bld_base
,
3947 LLVMValueRef interp_ij
)
3949 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3950 LLVMValueRef result
[4], a
;
3953 for (i
= 0; i
< 2; i
++) {
3954 a
= LLVMBuildExtractElement(ctx
->ac
.builder
, interp_ij
,
3955 LLVMConstInt(ctx
->i32
, i
, 0), "");
3956 result
[i
] = lp_build_emit_llvm_unary(bld_base
, TGSI_OPCODE_DDX
, a
);
3957 result
[2+i
] = lp_build_emit_llvm_unary(bld_base
, TGSI_OPCODE_DDY
, a
);
3960 return lp_build_gather_values(&ctx
->gallivm
, result
, 4);
3963 static void interp_fetch_args(
3964 struct lp_build_tgsi_context
*bld_base
,
3965 struct lp_build_emit_data
*emit_data
)
3967 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3968 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3970 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_OFFSET
) {
3971 /* offset is in second src, first two channels */
3972 emit_data
->args
[0] = lp_build_emit_fetch(bld_base
,
3975 emit_data
->args
[1] = lp_build_emit_fetch(bld_base
,
3978 emit_data
->arg_count
= 2;
3979 } else if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_SAMPLE
) {
3980 LLVMValueRef sample_position
;
3981 LLVMValueRef sample_id
;
3982 LLVMValueRef halfval
= LLVMConstReal(ctx
->f32
, 0.5f
);
3984 /* fetch sample ID, then fetch its sample position,
3985 * and place into first two channels.
3987 sample_id
= lp_build_emit_fetch(bld_base
,
3988 emit_data
->inst
, 1, TGSI_CHAN_X
);
3989 sample_id
= ac_to_integer(&ctx
->ac
, sample_id
);
3991 /* Section 8.13.2 (Interpolation Functions) of the OpenGL Shading
3992 * Language 4.50 spec says about interpolateAtSample:
3994 * "Returns the value of the input interpolant variable at
3995 * the location of sample number sample. If multisample
3996 * buffers are not available, the input variable will be
3997 * evaluated at the center of the pixel. If sample sample
3998 * does not exist, the position used to interpolate the
3999 * input variable is undefined."
4001 * This means that sample_id values outside of the valid are
4002 * in fact valid input, and the usual mechanism for loading the
4003 * sample position doesn't work.
4005 if (ctx
->shader
->key
.mono
.u
.ps
.interpolate_at_sample_force_center
) {
4006 LLVMValueRef center
[4] = {
4007 LLVMConstReal(ctx
->f32
, 0.5),
4008 LLVMConstReal(ctx
->f32
, 0.5),
4013 sample_position
= lp_build_gather_values(&ctx
->gallivm
, center
, 4);
4015 sample_position
= load_sample_position(&ctx
->abi
, sample_id
);
4018 emit_data
->args
[0] = LLVMBuildExtractElement(ctx
->ac
.builder
,
4022 emit_data
->args
[0] = LLVMBuildFSub(ctx
->ac
.builder
, emit_data
->args
[0], halfval
, "");
4023 emit_data
->args
[1] = LLVMBuildExtractElement(ctx
->ac
.builder
,
4026 emit_data
->args
[1] = LLVMBuildFSub(ctx
->ac
.builder
, emit_data
->args
[1], halfval
, "");
4027 emit_data
->arg_count
= 2;
4031 static void build_interp_intrinsic(const struct lp_build_tgsi_action
*action
,
4032 struct lp_build_tgsi_context
*bld_base
,
4033 struct lp_build_emit_data
*emit_data
)
4035 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4036 struct si_shader
*shader
= ctx
->shader
;
4037 const struct tgsi_shader_info
*info
= &shader
->selector
->info
;
4038 LLVMValueRef interp_param
;
4039 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
4040 const struct tgsi_full_src_register
*input
= &inst
->Src
[0];
4041 int input_base
, input_array_size
;
4044 LLVMValueRef prim_mask
= ctx
->abi
.prim_mask
;
4045 LLVMValueRef array_idx
;
4046 int interp_param_idx
;
4050 assert(input
->Register
.File
== TGSI_FILE_INPUT
);
4052 if (input
->Register
.Indirect
) {
4053 unsigned array_id
= input
->Indirect
.ArrayID
;
4056 input_base
= info
->input_array_first
[array_id
];
4057 input_array_size
= info
->input_array_last
[array_id
] - input_base
+ 1;
4059 input_base
= inst
->Src
[0].Register
.Index
;
4060 input_array_size
= info
->num_inputs
- input_base
;
4063 array_idx
= si_get_indirect_index(ctx
, &input
->Indirect
,
4064 1, input
->Register
.Index
- input_base
);
4066 input_base
= inst
->Src
[0].Register
.Index
;
4067 input_array_size
= 1;
4068 array_idx
= ctx
->i32_0
;
4071 interp
= shader
->selector
->info
.input_interpolate
[input_base
];
4073 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_OFFSET
||
4074 inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_SAMPLE
)
4075 location
= TGSI_INTERPOLATE_LOC_CENTER
;
4077 location
= TGSI_INTERPOLATE_LOC_CENTROID
;
4079 interp_param_idx
= lookup_interp_param_index(interp
, location
);
4080 if (interp_param_idx
== -1)
4082 else if (interp_param_idx
)
4083 interp_param
= LLVMGetParam(ctx
->main_fn
, interp_param_idx
);
4085 interp_param
= NULL
;
4087 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_OFFSET
||
4088 inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_SAMPLE
) {
4089 LLVMValueRef ij_out
[2];
4090 LLVMValueRef ddxy_out
= si_llvm_emit_ddxy_interp(bld_base
, interp_param
);
4093 * take the I then J parameters, and the DDX/Y for it, and
4094 * calculate the IJ inputs for the interpolator.
4095 * temp1 = ddx * offset/sample.x + I;
4096 * interp_param.I = ddy * offset/sample.y + temp1;
4097 * temp1 = ddx * offset/sample.x + J;
4098 * interp_param.J = ddy * offset/sample.y + temp1;
4100 for (i
= 0; i
< 2; i
++) {
4101 LLVMValueRef ix_ll
= LLVMConstInt(ctx
->i32
, i
, 0);
4102 LLVMValueRef iy_ll
= LLVMConstInt(ctx
->i32
, i
+ 2, 0);
4103 LLVMValueRef ddx_el
= LLVMBuildExtractElement(ctx
->ac
.builder
,
4104 ddxy_out
, ix_ll
, "");
4105 LLVMValueRef ddy_el
= LLVMBuildExtractElement(ctx
->ac
.builder
,
4106 ddxy_out
, iy_ll
, "");
4107 LLVMValueRef interp_el
= LLVMBuildExtractElement(ctx
->ac
.builder
,
4108 interp_param
, ix_ll
, "");
4109 LLVMValueRef temp1
, temp2
;
4111 interp_el
= ac_to_float(&ctx
->ac
, interp_el
);
4113 temp1
= LLVMBuildFMul(ctx
->ac
.builder
, ddx_el
, emit_data
->args
[0], "");
4115 temp1
= LLVMBuildFAdd(ctx
->ac
.builder
, temp1
, interp_el
, "");
4117 temp2
= LLVMBuildFMul(ctx
->ac
.builder
, ddy_el
, emit_data
->args
[1], "");
4119 ij_out
[i
] = LLVMBuildFAdd(ctx
->ac
.builder
, temp2
, temp1
, "");
4121 interp_param
= lp_build_gather_values(&ctx
->gallivm
, ij_out
, 2);
4125 interp_param
= ac_to_float(&ctx
->ac
, interp_param
);
4127 for (chan
= 0; chan
< 4; chan
++) {
4128 LLVMValueRef gather
= LLVMGetUndef(LLVMVectorType(ctx
->f32
, input_array_size
));
4129 unsigned schan
= tgsi_util_get_full_src_register_swizzle(&inst
->Src
[0], chan
);
4131 for (unsigned idx
= 0; idx
< input_array_size
; ++idx
) {
4132 LLVMValueRef v
, i
= NULL
, j
= NULL
;
4135 i
= LLVMBuildExtractElement(
4136 ctx
->ac
.builder
, interp_param
, ctx
->i32_0
, "");
4137 j
= LLVMBuildExtractElement(
4138 ctx
->ac
.builder
, interp_param
, ctx
->i32_1
, "");
4140 v
= si_build_fs_interp(ctx
, input_base
+ idx
, schan
,
4143 gather
= LLVMBuildInsertElement(ctx
->ac
.builder
,
4144 gather
, v
, LLVMConstInt(ctx
->i32
, idx
, false), "");
4147 emit_data
->output
[chan
] = LLVMBuildExtractElement(
4148 ctx
->ac
.builder
, gather
, array_idx
, "");
4152 static void vote_all_emit(
4153 const struct lp_build_tgsi_action
*action
,
4154 struct lp_build_tgsi_context
*bld_base
,
4155 struct lp_build_emit_data
*emit_data
)
4157 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4159 LLVMValueRef tmp
= ac_build_vote_all(&ctx
->ac
, emit_data
->args
[0]);
4160 emit_data
->output
[emit_data
->chan
] =
4161 LLVMBuildSExt(ctx
->ac
.builder
, tmp
, ctx
->i32
, "");
4164 static void vote_any_emit(
4165 const struct lp_build_tgsi_action
*action
,
4166 struct lp_build_tgsi_context
*bld_base
,
4167 struct lp_build_emit_data
*emit_data
)
4169 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4171 LLVMValueRef tmp
= ac_build_vote_any(&ctx
->ac
, emit_data
->args
[0]);
4172 emit_data
->output
[emit_data
->chan
] =
4173 LLVMBuildSExt(ctx
->ac
.builder
, tmp
, ctx
->i32
, "");
4176 static void vote_eq_emit(
4177 const struct lp_build_tgsi_action
*action
,
4178 struct lp_build_tgsi_context
*bld_base
,
4179 struct lp_build_emit_data
*emit_data
)
4181 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4183 LLVMValueRef tmp
= ac_build_vote_eq(&ctx
->ac
, emit_data
->args
[0]);
4184 emit_data
->output
[emit_data
->chan
] =
4185 LLVMBuildSExt(ctx
->ac
.builder
, tmp
, ctx
->i32
, "");
4188 static void ballot_emit(
4189 const struct lp_build_tgsi_action
*action
,
4190 struct lp_build_tgsi_context
*bld_base
,
4191 struct lp_build_emit_data
*emit_data
)
4193 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4194 LLVMBuilderRef builder
= ctx
->ac
.builder
;
4197 tmp
= lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, TGSI_CHAN_X
);
4198 tmp
= ac_build_ballot(&ctx
->ac
, tmp
);
4199 tmp
= LLVMBuildBitCast(builder
, tmp
, ctx
->v2i32
, "");
4201 emit_data
->output
[0] = LLVMBuildExtractElement(builder
, tmp
, ctx
->i32_0
, "");
4202 emit_data
->output
[1] = LLVMBuildExtractElement(builder
, tmp
, ctx
->i32_1
, "");
4205 static void read_invoc_fetch_args(
4206 struct lp_build_tgsi_context
*bld_base
,
4207 struct lp_build_emit_data
*emit_data
)
4209 emit_data
->args
[0] = lp_build_emit_fetch(bld_base
, emit_data
->inst
,
4210 0, emit_data
->src_chan
);
4212 /* Always read the source invocation (= lane) from the X channel. */
4213 emit_data
->args
[1] = lp_build_emit_fetch(bld_base
, emit_data
->inst
,
4215 emit_data
->arg_count
= 2;
4218 static void read_lane_emit(
4219 const struct lp_build_tgsi_action
*action
,
4220 struct lp_build_tgsi_context
*bld_base
,
4221 struct lp_build_emit_data
*emit_data
)
4223 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4225 /* We currently have no other way to prevent LLVM from lifting the icmp
4226 * calls to a dominating basic block.
4228 ac_build_optimization_barrier(&ctx
->ac
, &emit_data
->args
[0]);
4230 for (unsigned i
= 0; i
< emit_data
->arg_count
; ++i
)
4231 emit_data
->args
[i
] = ac_to_integer(&ctx
->ac
, emit_data
->args
[i
]);
4233 emit_data
->output
[emit_data
->chan
] =
4234 ac_build_intrinsic(&ctx
->ac
, action
->intr_name
,
4235 ctx
->i32
, emit_data
->args
, emit_data
->arg_count
,
4236 AC_FUNC_ATTR_READNONE
|
4237 AC_FUNC_ATTR_CONVERGENT
);
4240 static unsigned si_llvm_get_stream(struct lp_build_tgsi_context
*bld_base
,
4241 struct lp_build_emit_data
*emit_data
)
4243 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4244 struct tgsi_src_register src0
= emit_data
->inst
->Src
[0].Register
;
4248 assert(src0
.File
== TGSI_FILE_IMMEDIATE
);
4250 imm
= ctx
->imms
[src0
.Index
* TGSI_NUM_CHANNELS
+ src0
.SwizzleX
];
4251 stream
= LLVMConstIntGetZExtValue(imm
) & 0x3;
4255 /* Emit one vertex from the geometry shader */
4256 static void si_llvm_emit_vertex(struct ac_shader_abi
*abi
,
4258 LLVMValueRef
*addrs
)
4260 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
4261 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
4262 struct lp_build_context
*uint
= &ctx
->bld_base
.uint_bld
;
4263 struct si_shader
*shader
= ctx
->shader
;
4264 struct lp_build_if_state if_state
;
4265 LLVMValueRef soffset
= LLVMGetParam(ctx
->main_fn
,
4266 ctx
->param_gs2vs_offset
);
4267 LLVMValueRef gs_next_vertex
;
4268 LLVMValueRef can_emit
;
4269 unsigned chan
, offset
;
4272 /* Write vertex attribute values to GSVS ring */
4273 gs_next_vertex
= LLVMBuildLoad(ctx
->ac
.builder
,
4274 ctx
->gs_next_vertex
[stream
],
4277 /* If this thread has already emitted the declared maximum number of
4278 * vertices, skip the write: excessive vertex emissions are not
4279 * supposed to have any effect.
4281 * If the shader has no writes to memory, kill it instead. This skips
4282 * further memory loads and may allow LLVM to skip to the end
4285 can_emit
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntULT
, gs_next_vertex
,
4286 LLVMConstInt(ctx
->i32
,
4287 shader
->selector
->gs_max_out_vertices
, 0), "");
4289 bool use_kill
= !info
->writes_memory
;
4291 ac_build_kill_if_false(&ctx
->ac
, can_emit
);
4293 lp_build_if(&if_state
, &ctx
->gallivm
, can_emit
);
4297 for (i
= 0; i
< info
->num_outputs
; i
++) {
4298 for (chan
= 0; chan
< 4; chan
++) {
4299 if (!(info
->output_usagemask
[i
] & (1 << chan
)) ||
4300 ((info
->output_streams
[i
] >> (2 * chan
)) & 3) != stream
)
4303 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->ac
.builder
, addrs
[4 * i
+ chan
], "");
4304 LLVMValueRef voffset
=
4305 LLVMConstInt(ctx
->i32
, offset
*
4306 shader
->selector
->gs_max_out_vertices
, 0);
4309 voffset
= lp_build_add(uint
, voffset
, gs_next_vertex
);
4310 voffset
= lp_build_mul_imm(uint
, voffset
, 4);
4312 out_val
= ac_to_integer(&ctx
->ac
, out_val
);
4314 ac_build_buffer_store_dword(&ctx
->ac
,
4315 ctx
->gsvs_ring
[stream
],
4317 voffset
, soffset
, 0,
4322 gs_next_vertex
= lp_build_add(uint
, gs_next_vertex
,
4325 LLVMBuildStore(ctx
->ac
.builder
, gs_next_vertex
, ctx
->gs_next_vertex
[stream
]);
4327 /* Signal vertex emission */
4328 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_EMIT
| AC_SENDMSG_GS
| (stream
<< 8),
4329 si_get_gs_wave_id(ctx
));
4331 lp_build_endif(&if_state
);
4334 /* Emit one vertex from the geometry shader */
4335 static void si_tgsi_emit_vertex(
4336 const struct lp_build_tgsi_action
*action
,
4337 struct lp_build_tgsi_context
*bld_base
,
4338 struct lp_build_emit_data
*emit_data
)
4340 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4341 unsigned stream
= si_llvm_get_stream(bld_base
, emit_data
);
4343 si_llvm_emit_vertex(&ctx
->abi
, stream
, ctx
->outputs
[0]);
4346 /* Cut one primitive from the geometry shader */
4347 static void si_llvm_emit_primitive(struct ac_shader_abi
*abi
,
4350 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
4352 /* Signal primitive cut */
4353 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_CUT
| AC_SENDMSG_GS
| (stream
<< 8),
4354 si_get_gs_wave_id(ctx
));
4357 /* Cut one primitive from the geometry shader */
4358 static void si_tgsi_emit_primitive(
4359 const struct lp_build_tgsi_action
*action
,
4360 struct lp_build_tgsi_context
*bld_base
,
4361 struct lp_build_emit_data
*emit_data
)
4363 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4365 si_llvm_emit_primitive(&ctx
->abi
, si_llvm_get_stream(bld_base
, emit_data
));
4368 static void si_llvm_emit_barrier(const struct lp_build_tgsi_action
*action
,
4369 struct lp_build_tgsi_context
*bld_base
,
4370 struct lp_build_emit_data
*emit_data
)
4372 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4374 /* SI only (thanks to a hw bug workaround):
4375 * The real barrier instruction isn’t needed, because an entire patch
4376 * always fits into a single wave.
4378 if (ctx
->screen
->info
.chip_class
== SI
&&
4379 ctx
->type
== PIPE_SHADER_TESS_CTRL
) {
4380 ac_build_waitcnt(&ctx
->ac
, LGKM_CNT
& VM_CNT
);
4384 lp_build_intrinsic(ctx
->ac
.builder
,
4385 "llvm.amdgcn.s.barrier",
4386 ctx
->voidt
, NULL
, 0, LP_FUNC_ATTR_CONVERGENT
);
4389 static const struct lp_build_tgsi_action interp_action
= {
4390 .fetch_args
= interp_fetch_args
,
4391 .emit
= build_interp_intrinsic
,
4394 static void si_create_function(struct si_shader_context
*ctx
,
4396 LLVMTypeRef
*returns
, unsigned num_returns
,
4397 struct si_function_info
*fninfo
,
4398 unsigned max_workgroup_size
)
4402 si_llvm_create_func(ctx
, name
, returns
, num_returns
,
4403 fninfo
->types
, fninfo
->num_params
);
4404 ctx
->return_value
= LLVMGetUndef(ctx
->return_type
);
4406 for (i
= 0; i
< fninfo
->num_sgpr_params
; ++i
) {
4407 LLVMValueRef P
= LLVMGetParam(ctx
->main_fn
, i
);
4409 /* The combination of:
4413 * allows the optimization passes to move loads and reduces
4414 * SGPR spilling significantly.
4416 lp_add_function_attr(ctx
->main_fn
, i
+ 1, LP_FUNC_ATTR_INREG
);
4418 if (LLVMGetTypeKind(LLVMTypeOf(P
)) == LLVMPointerTypeKind
) {
4419 lp_add_function_attr(ctx
->main_fn
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
4420 ac_add_attr_dereferenceable(P
, UINT64_MAX
);
4424 for (i
= 0; i
< fninfo
->num_params
; ++i
) {
4425 if (fninfo
->assign
[i
])
4426 *fninfo
->assign
[i
] = LLVMGetParam(ctx
->main_fn
, i
);
4429 si_llvm_add_attribute(ctx
->main_fn
, "amdgpu-32bit-address-high-bits",
4430 ctx
->screen
->info
.address32_hi
);
4432 if (max_workgroup_size
) {
4433 si_llvm_add_attribute(ctx
->main_fn
, "amdgpu-max-work-group-size",
4434 max_workgroup_size
);
4436 LLVMAddTargetDependentFunctionAttr(ctx
->main_fn
,
4437 "no-signed-zeros-fp-math",
4440 if (ctx
->screen
->debug_flags
& DBG(UNSAFE_MATH
)) {
4441 /* These were copied from some LLVM test. */
4442 LLVMAddTargetDependentFunctionAttr(ctx
->main_fn
,
4443 "less-precise-fpmad",
4445 LLVMAddTargetDependentFunctionAttr(ctx
->main_fn
,
4448 LLVMAddTargetDependentFunctionAttr(ctx
->main_fn
,
4451 LLVMAddTargetDependentFunctionAttr(ctx
->main_fn
,
4457 static void declare_streamout_params(struct si_shader_context
*ctx
,
4458 struct pipe_stream_output_info
*so
,
4459 struct si_function_info
*fninfo
)
4463 /* Streamout SGPRs. */
4464 if (so
->num_outputs
) {
4465 if (ctx
->type
!= PIPE_SHADER_TESS_EVAL
)
4466 ctx
->param_streamout_config
= add_arg(fninfo
, ARG_SGPR
, ctx
->ac
.i32
);
4468 ctx
->param_streamout_config
= fninfo
->num_params
- 1;
4470 ctx
->param_streamout_write_index
= add_arg(fninfo
, ARG_SGPR
, ctx
->ac
.i32
);
4472 /* A streamout buffer offset is loaded if the stride is non-zero. */
4473 for (i
= 0; i
< 4; i
++) {
4477 ctx
->param_streamout_offset
[i
] = add_arg(fninfo
, ARG_SGPR
, ctx
->ac
.i32
);
4481 static unsigned si_get_max_workgroup_size(const struct si_shader
*shader
)
4483 switch (shader
->selector
->type
) {
4484 case PIPE_SHADER_TESS_CTRL
:
4485 /* Return this so that LLVM doesn't remove s_barrier
4486 * instructions on chips where we use s_barrier. */
4487 return shader
->selector
->screen
->info
.chip_class
>= CIK
? 128 : 64;
4489 case PIPE_SHADER_GEOMETRY
:
4490 return shader
->selector
->screen
->info
.chip_class
>= GFX9
? 128 : 64;
4492 case PIPE_SHADER_COMPUTE
:
4493 break; /* see below */
4499 const unsigned *properties
= shader
->selector
->info
.properties
;
4500 unsigned max_work_group_size
=
4501 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
] *
4502 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT
] *
4503 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH
];
4505 if (!max_work_group_size
) {
4506 /* This is a variable group size compute shader,
4507 * compile it for the maximum possible group size.
4509 max_work_group_size
= SI_MAX_VARIABLE_THREADS_PER_BLOCK
;
4511 return max_work_group_size
;
4514 static void declare_const_and_shader_buffers(struct si_shader_context
*ctx
,
4515 struct si_function_info
*fninfo
,
4518 LLVMTypeRef const_shader_buf_type
;
4520 if (ctx
->shader
->selector
->info
.const_buffers_declared
== 1 &&
4521 ctx
->shader
->selector
->info
.shader_buffers_declared
== 0)
4522 const_shader_buf_type
= ctx
->f32
;
4524 const_shader_buf_type
= ctx
->v4i32
;
4526 unsigned const_and_shader_buffers
=
4527 add_arg(fninfo
, ARG_SGPR
,
4528 ac_array_in_const32_addr_space(const_shader_buf_type
));
4531 ctx
->param_const_and_shader_buffers
= const_and_shader_buffers
;
4534 static void declare_samplers_and_images(struct si_shader_context
*ctx
,
4535 struct si_function_info
*fninfo
,
4538 unsigned samplers_and_images
=
4539 add_arg(fninfo
, ARG_SGPR
,
4540 ac_array_in_const32_addr_space(ctx
->v8i32
));
4543 ctx
->param_samplers_and_images
= samplers_and_images
;
4546 static void declare_per_stage_desc_pointers(struct si_shader_context
*ctx
,
4547 struct si_function_info
*fninfo
,
4550 declare_const_and_shader_buffers(ctx
, fninfo
, assign_params
);
4551 declare_samplers_and_images(ctx
, fninfo
, assign_params
);
4554 static void declare_global_desc_pointers(struct si_shader_context
*ctx
,
4555 struct si_function_info
*fninfo
)
4557 ctx
->param_rw_buffers
= add_arg(fninfo
, ARG_SGPR
,
4558 ac_array_in_const32_addr_space(ctx
->v4i32
));
4559 ctx
->param_bindless_samplers_and_images
= add_arg(fninfo
, ARG_SGPR
,
4560 ac_array_in_const32_addr_space(ctx
->v8i32
));
4563 static void declare_vs_specific_input_sgprs(struct si_shader_context
*ctx
,
4564 struct si_function_info
*fninfo
)
4566 add_arg_assign(fninfo
, ARG_SGPR
, ctx
->i32
, &ctx
->abi
.base_vertex
);
4567 add_arg_assign(fninfo
, ARG_SGPR
, ctx
->i32
, &ctx
->abi
.start_instance
);
4568 add_arg_assign(fninfo
, ARG_SGPR
, ctx
->i32
, &ctx
->abi
.draw_id
);
4569 ctx
->param_vs_state_bits
= add_arg(fninfo
, ARG_SGPR
, ctx
->i32
);
4572 static void declare_vs_input_vgprs(struct si_shader_context
*ctx
,
4573 struct si_function_info
*fninfo
,
4574 unsigned *num_prolog_vgprs
)
4576 struct si_shader
*shader
= ctx
->shader
;
4578 add_arg_assign(fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.vertex_id
);
4579 if (shader
->key
.as_ls
) {
4580 ctx
->param_rel_auto_id
= add_arg(fninfo
, ARG_VGPR
, ctx
->i32
);
4581 add_arg_assign(fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.instance_id
);
4583 add_arg_assign(fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.instance_id
);
4584 ctx
->param_vs_prim_id
= add_arg(fninfo
, ARG_VGPR
, ctx
->i32
);
4586 add_arg(fninfo
, ARG_VGPR
, ctx
->i32
); /* unused */
4588 if (!shader
->is_gs_copy_shader
) {
4589 /* Vertex load indices. */
4590 ctx
->param_vertex_index0
= fninfo
->num_params
;
4591 for (unsigned i
= 0; i
< shader
->selector
->info
.num_inputs
; i
++)
4592 add_arg(fninfo
, ARG_VGPR
, ctx
->i32
);
4593 *num_prolog_vgprs
+= shader
->selector
->info
.num_inputs
;
4597 static void declare_tes_input_vgprs(struct si_shader_context
*ctx
,
4598 struct si_function_info
*fninfo
)
4600 ctx
->param_tes_u
= add_arg(fninfo
, ARG_VGPR
, ctx
->f32
);
4601 ctx
->param_tes_v
= add_arg(fninfo
, ARG_VGPR
, ctx
->f32
);
4602 ctx
->param_tes_rel_patch_id
= add_arg(fninfo
, ARG_VGPR
, ctx
->i32
);
4603 add_arg_assign(fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.tes_patch_id
);
4607 /* Convenient merged shader definitions. */
4608 SI_SHADER_MERGED_VERTEX_TESSCTRL
= PIPE_SHADER_TYPES
,
4609 SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY
,
4612 static void create_function(struct si_shader_context
*ctx
)
4614 struct si_shader
*shader
= ctx
->shader
;
4615 struct si_function_info fninfo
;
4616 LLVMTypeRef returns
[16+32*4];
4617 unsigned i
, num_return_sgprs
;
4618 unsigned num_returns
= 0;
4619 unsigned num_prolog_vgprs
= 0;
4620 unsigned type
= ctx
->type
;
4621 unsigned vs_blit_property
=
4622 shader
->selector
->info
.properties
[TGSI_PROPERTY_VS_BLIT_SGPRS
];
4624 si_init_function_info(&fninfo
);
4626 /* Set MERGED shaders. */
4627 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
4628 if (shader
->key
.as_ls
|| type
== PIPE_SHADER_TESS_CTRL
)
4629 type
= SI_SHADER_MERGED_VERTEX_TESSCTRL
; /* LS or HS */
4630 else if (shader
->key
.as_es
|| type
== PIPE_SHADER_GEOMETRY
)
4631 type
= SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY
;
4634 LLVMTypeRef v3i32
= LLVMVectorType(ctx
->i32
, 3);
4637 case PIPE_SHADER_VERTEX
:
4638 declare_global_desc_pointers(ctx
, &fninfo
);
4640 if (vs_blit_property
) {
4641 ctx
->param_vs_blit_inputs
= fninfo
.num_params
;
4642 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* i16 x1, y1 */
4643 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* i16 x2, y2 */
4644 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* depth */
4646 if (vs_blit_property
== SI_VS_BLIT_SGPRS_POS_COLOR
) {
4647 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* color0 */
4648 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* color1 */
4649 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* color2 */
4650 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* color3 */
4651 } else if (vs_blit_property
== SI_VS_BLIT_SGPRS_POS_TEXCOORD
) {
4652 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* texcoord.x1 */
4653 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* texcoord.y1 */
4654 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* texcoord.x2 */
4655 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* texcoord.y2 */
4656 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* texcoord.z */
4657 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* texcoord.w */
4661 declare_vs_input_vgprs(ctx
, &fninfo
, &num_prolog_vgprs
);
4665 declare_per_stage_desc_pointers(ctx
, &fninfo
, true);
4666 declare_vs_specific_input_sgprs(ctx
, &fninfo
);
4667 ctx
->param_vertex_buffers
= add_arg(&fninfo
, ARG_SGPR
,
4668 ac_array_in_const32_addr_space(ctx
->v4i32
));
4670 if (shader
->key
.as_es
) {
4671 ctx
->param_es2gs_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4672 } else if (shader
->key
.as_ls
) {
4673 /* no extra parameters */
4675 if (shader
->is_gs_copy_shader
) {
4676 fninfo
.num_params
= ctx
->param_rw_buffers
+ 1;
4677 fninfo
.num_sgpr_params
= fninfo
.num_params
;
4680 /* The locations of the other parameters are assigned dynamically. */
4681 declare_streamout_params(ctx
, &shader
->selector
->so
,
4686 declare_vs_input_vgprs(ctx
, &fninfo
, &num_prolog_vgprs
);
4689 case PIPE_SHADER_TESS_CTRL
: /* SI-CI-VI */
4690 declare_global_desc_pointers(ctx
, &fninfo
);
4691 declare_per_stage_desc_pointers(ctx
, &fninfo
, true);
4692 ctx
->param_tcs_offchip_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4693 ctx
->param_tcs_out_lds_offsets
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4694 ctx
->param_tcs_out_lds_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4695 ctx
->param_vs_state_bits
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4696 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4697 ctx
->param_tcs_factor_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4700 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.tcs_patch_id
);
4701 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.tcs_rel_ids
);
4703 /* param_tcs_offchip_offset and param_tcs_factor_offset are
4704 * placed after the user SGPRs.
4706 for (i
= 0; i
< GFX6_TCS_NUM_USER_SGPR
+ 2; i
++)
4707 returns
[num_returns
++] = ctx
->i32
; /* SGPRs */
4708 for (i
= 0; i
< 11; i
++)
4709 returns
[num_returns
++] = ctx
->f32
; /* VGPRs */
4712 case SI_SHADER_MERGED_VERTEX_TESSCTRL
:
4713 /* Merged stages have 8 system SGPRs at the beginning. */
4714 /* SPI_SHADER_USER_DATA_ADDR_LO/HI_HS */
4715 if (HAVE_32BIT_POINTERS
) {
4716 declare_per_stage_desc_pointers(ctx
, &fninfo
,
4717 ctx
->type
== PIPE_SHADER_TESS_CTRL
);
4719 declare_const_and_shader_buffers(ctx
, &fninfo
,
4720 ctx
->type
== PIPE_SHADER_TESS_CTRL
);
4722 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4723 ctx
->param_merged_wave_info
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4724 ctx
->param_tcs_factor_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4725 ctx
->param_merged_scratch_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4726 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused */
4727 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused */
4729 declare_global_desc_pointers(ctx
, &fninfo
);
4730 declare_per_stage_desc_pointers(ctx
, &fninfo
,
4731 ctx
->type
== PIPE_SHADER_VERTEX
);
4732 declare_vs_specific_input_sgprs(ctx
, &fninfo
);
4734 if (!HAVE_32BIT_POINTERS
) {
4735 declare_samplers_and_images(ctx
, &fninfo
,
4736 ctx
->type
== PIPE_SHADER_TESS_CTRL
);
4738 ctx
->param_tcs_offchip_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4739 ctx
->param_tcs_out_lds_offsets
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4740 ctx
->param_tcs_out_lds_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4741 if (!HAVE_32BIT_POINTERS
) /* Align to 2 dwords. */
4742 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused */
4743 ctx
->param_vertex_buffers
= add_arg(&fninfo
, ARG_SGPR
,
4744 ac_array_in_const32_addr_space(ctx
->v4i32
));
4746 /* VGPRs (first TCS, then VS) */
4747 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.tcs_patch_id
);
4748 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.tcs_rel_ids
);
4750 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
4751 declare_vs_input_vgprs(ctx
, &fninfo
,
4754 /* LS return values are inputs to the TCS main shader part. */
4755 for (i
= 0; i
< 8 + GFX9_TCS_NUM_USER_SGPR
; i
++)
4756 returns
[num_returns
++] = ctx
->i32
; /* SGPRs */
4757 for (i
= 0; i
< 2; i
++)
4758 returns
[num_returns
++] = ctx
->f32
; /* VGPRs */
4760 /* TCS return values are inputs to the TCS epilog.
4762 * param_tcs_offchip_offset, param_tcs_factor_offset,
4763 * param_tcs_offchip_layout, and param_rw_buffers
4764 * should be passed to the epilog.
4766 for (i
= 0; i
<= 8 + GFX9_SGPR_TCS_OUT_LAYOUT
; i
++)
4767 returns
[num_returns
++] = ctx
->i32
; /* SGPRs */
4768 for (i
= 0; i
< 11; i
++)
4769 returns
[num_returns
++] = ctx
->f32
; /* VGPRs */
4773 case SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY
:
4774 /* Merged stages have 8 system SGPRs at the beginning. */
4775 /* SPI_SHADER_USER_DATA_ADDR_LO/HI_GS */
4776 if (HAVE_32BIT_POINTERS
) {
4777 declare_per_stage_desc_pointers(ctx
, &fninfo
,
4778 ctx
->type
== PIPE_SHADER_GEOMETRY
);
4780 declare_const_and_shader_buffers(ctx
, &fninfo
,
4781 ctx
->type
== PIPE_SHADER_GEOMETRY
);
4783 ctx
->param_gs2vs_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4784 ctx
->param_merged_wave_info
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4785 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4786 ctx
->param_merged_scratch_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4787 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused (SPI_SHADER_PGM_LO/HI_GS << 8) */
4788 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused (SPI_SHADER_PGM_LO/HI_GS >> 24) */
4790 declare_global_desc_pointers(ctx
, &fninfo
);
4791 declare_per_stage_desc_pointers(ctx
, &fninfo
,
4792 (ctx
->type
== PIPE_SHADER_VERTEX
||
4793 ctx
->type
== PIPE_SHADER_TESS_EVAL
));
4794 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
4795 declare_vs_specific_input_sgprs(ctx
, &fninfo
);
4797 ctx
->param_tcs_offchip_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4798 ctx
->param_tes_offchip_addr
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4799 if (!HAVE_32BIT_POINTERS
) {
4800 /* Declare as many input SGPRs as the VS has. */
4801 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused */
4802 ctx
->param_vs_state_bits
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused */
4806 if (!HAVE_32BIT_POINTERS
) {
4807 declare_samplers_and_images(ctx
, &fninfo
,
4808 ctx
->type
== PIPE_SHADER_GEOMETRY
);
4810 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
4811 ctx
->param_vertex_buffers
= add_arg(&fninfo
, ARG_SGPR
,
4812 ac_array_in_const32_addr_space(ctx
->v4i32
));
4815 /* VGPRs (first GS, then VS/TES) */
4816 ctx
->param_gs_vtx01_offset
= add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
4817 ctx
->param_gs_vtx23_offset
= add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
4818 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.gs_prim_id
);
4819 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.gs_invocation_id
);
4820 ctx
->param_gs_vtx45_offset
= add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
4822 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
4823 declare_vs_input_vgprs(ctx
, &fninfo
,
4825 } else if (ctx
->type
== PIPE_SHADER_TESS_EVAL
) {
4826 declare_tes_input_vgprs(ctx
, &fninfo
);
4829 if (ctx
->type
== PIPE_SHADER_VERTEX
||
4830 ctx
->type
== PIPE_SHADER_TESS_EVAL
) {
4831 unsigned num_user_sgprs
;
4833 if (ctx
->type
== PIPE_SHADER_VERTEX
)
4834 num_user_sgprs
= GFX9_VSGS_NUM_USER_SGPR
;
4836 num_user_sgprs
= GFX9_TESGS_NUM_USER_SGPR
;
4838 /* ES return values are inputs to GS. */
4839 for (i
= 0; i
< 8 + num_user_sgprs
; i
++)
4840 returns
[num_returns
++] = ctx
->i32
; /* SGPRs */
4841 for (i
= 0; i
< 5; i
++)
4842 returns
[num_returns
++] = ctx
->f32
; /* VGPRs */
4846 case PIPE_SHADER_TESS_EVAL
:
4847 declare_global_desc_pointers(ctx
, &fninfo
);
4848 declare_per_stage_desc_pointers(ctx
, &fninfo
, true);
4849 ctx
->param_tcs_offchip_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4850 ctx
->param_tes_offchip_addr
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4852 if (shader
->key
.as_es
) {
4853 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4854 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4855 ctx
->param_es2gs_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4857 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4858 declare_streamout_params(ctx
, &shader
->selector
->so
,
4860 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4864 declare_tes_input_vgprs(ctx
, &fninfo
);
4867 case PIPE_SHADER_GEOMETRY
:
4868 declare_global_desc_pointers(ctx
, &fninfo
);
4869 declare_per_stage_desc_pointers(ctx
, &fninfo
, true);
4870 ctx
->param_gs2vs_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4871 ctx
->param_gs_wave_id
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4874 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->gs_vtx_offset
[0]);
4875 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->gs_vtx_offset
[1]);
4876 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.gs_prim_id
);
4877 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->gs_vtx_offset
[2]);
4878 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->gs_vtx_offset
[3]);
4879 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->gs_vtx_offset
[4]);
4880 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->gs_vtx_offset
[5]);
4881 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.gs_invocation_id
);
4884 case PIPE_SHADER_FRAGMENT
:
4885 declare_global_desc_pointers(ctx
, &fninfo
);
4886 declare_per_stage_desc_pointers(ctx
, &fninfo
, true);
4887 add_arg_checked(&fninfo
, ARG_SGPR
, ctx
->f32
, SI_PARAM_ALPHA_REF
);
4888 add_arg_assign_checked(&fninfo
, ARG_SGPR
, ctx
->i32
,
4889 &ctx
->abi
.prim_mask
, SI_PARAM_PRIM_MASK
);
4891 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->v2i32
, SI_PARAM_PERSP_SAMPLE
);
4892 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->v2i32
, SI_PARAM_PERSP_CENTER
);
4893 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->v2i32
, SI_PARAM_PERSP_CENTROID
);
4894 add_arg_checked(&fninfo
, ARG_VGPR
, v3i32
, SI_PARAM_PERSP_PULL_MODEL
);
4895 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->v2i32
, SI_PARAM_LINEAR_SAMPLE
);
4896 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->v2i32
, SI_PARAM_LINEAR_CENTER
);
4897 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->v2i32
, SI_PARAM_LINEAR_CENTROID
);
4898 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->f32
, SI_PARAM_LINE_STIPPLE_TEX
);
4899 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->f32
,
4900 &ctx
->abi
.frag_pos
[0], SI_PARAM_POS_X_FLOAT
);
4901 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->f32
,
4902 &ctx
->abi
.frag_pos
[1], SI_PARAM_POS_Y_FLOAT
);
4903 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->f32
,
4904 &ctx
->abi
.frag_pos
[2], SI_PARAM_POS_Z_FLOAT
);
4905 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->f32
,
4906 &ctx
->abi
.frag_pos
[3], SI_PARAM_POS_W_FLOAT
);
4907 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->i32
,
4908 &ctx
->abi
.front_face
, SI_PARAM_FRONT_FACE
);
4909 shader
->info
.face_vgpr_index
= 20;
4910 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->i32
,
4911 &ctx
->abi
.ancillary
, SI_PARAM_ANCILLARY
);
4912 shader
->info
.ancillary_vgpr_index
= 21;
4913 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->f32
,
4914 &ctx
->abi
.sample_coverage
, SI_PARAM_SAMPLE_COVERAGE
);
4915 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->i32
, SI_PARAM_POS_FIXED_PT
);
4917 /* Color inputs from the prolog. */
4918 if (shader
->selector
->info
.colors_read
) {
4919 unsigned num_color_elements
=
4920 util_bitcount(shader
->selector
->info
.colors_read
);
4922 assert(fninfo
.num_params
+ num_color_elements
<= ARRAY_SIZE(fninfo
.types
));
4923 for (i
= 0; i
< num_color_elements
; i
++)
4924 add_arg(&fninfo
, ARG_VGPR
, ctx
->f32
);
4926 num_prolog_vgprs
+= num_color_elements
;
4929 /* Outputs for the epilog. */
4930 num_return_sgprs
= SI_SGPR_ALPHA_REF
+ 1;
4933 util_bitcount(shader
->selector
->info
.colors_written
) * 4 +
4934 shader
->selector
->info
.writes_z
+
4935 shader
->selector
->info
.writes_stencil
+
4936 shader
->selector
->info
.writes_samplemask
+
4937 1 /* SampleMaskIn */;
4939 num_returns
= MAX2(num_returns
,
4941 PS_EPILOG_SAMPLEMASK_MIN_LOC
+ 1);
4943 for (i
= 0; i
< num_return_sgprs
; i
++)
4944 returns
[i
] = ctx
->i32
;
4945 for (; i
< num_returns
; i
++)
4946 returns
[i
] = ctx
->f32
;
4949 case PIPE_SHADER_COMPUTE
:
4950 declare_global_desc_pointers(ctx
, &fninfo
);
4951 declare_per_stage_desc_pointers(ctx
, &fninfo
, true);
4952 if (shader
->selector
->info
.uses_grid_size
)
4953 add_arg_assign(&fninfo
, ARG_SGPR
, v3i32
, &ctx
->abi
.num_work_groups
);
4954 if (shader
->selector
->info
.uses_block_size
)
4955 ctx
->param_block_size
= add_arg(&fninfo
, ARG_SGPR
, v3i32
);
4957 for (i
= 0; i
< 3; i
++) {
4958 ctx
->abi
.workgroup_ids
[i
] = NULL
;
4959 if (shader
->selector
->info
.uses_block_id
[i
])
4960 add_arg_assign(&fninfo
, ARG_SGPR
, ctx
->i32
, &ctx
->abi
.workgroup_ids
[i
]);
4963 add_arg_assign(&fninfo
, ARG_VGPR
, v3i32
, &ctx
->abi
.local_invocation_ids
);
4966 assert(0 && "unimplemented shader");
4970 si_create_function(ctx
, "main", returns
, num_returns
, &fninfo
,
4971 si_get_max_workgroup_size(shader
));
4973 /* Reserve register locations for VGPR inputs the PS prolog may need. */
4974 if (ctx
->type
== PIPE_SHADER_FRAGMENT
&&
4975 ctx
->separate_prolog
) {
4976 si_llvm_add_attribute(ctx
->main_fn
,
4977 "InitialPSInputAddr",
4978 S_0286D0_PERSP_SAMPLE_ENA(1) |
4979 S_0286D0_PERSP_CENTER_ENA(1) |
4980 S_0286D0_PERSP_CENTROID_ENA(1) |
4981 S_0286D0_LINEAR_SAMPLE_ENA(1) |
4982 S_0286D0_LINEAR_CENTER_ENA(1) |
4983 S_0286D0_LINEAR_CENTROID_ENA(1) |
4984 S_0286D0_FRONT_FACE_ENA(1) |
4985 S_0286D0_ANCILLARY_ENA(1) |
4986 S_0286D0_POS_FIXED_PT_ENA(1));
4989 shader
->info
.num_input_sgprs
= 0;
4990 shader
->info
.num_input_vgprs
= 0;
4992 for (i
= 0; i
< fninfo
.num_sgpr_params
; ++i
)
4993 shader
->info
.num_input_sgprs
+= ac_get_type_size(fninfo
.types
[i
]) / 4;
4995 for (; i
< fninfo
.num_params
; ++i
)
4996 shader
->info
.num_input_vgprs
+= ac_get_type_size(fninfo
.types
[i
]) / 4;
4998 assert(shader
->info
.num_input_vgprs
>= num_prolog_vgprs
);
4999 shader
->info
.num_input_vgprs
-= num_prolog_vgprs
;
5001 if (shader
->key
.as_ls
||
5002 ctx
->type
== PIPE_SHADER_TESS_CTRL
||
5003 /* GFX9 has the ESGS ring buffer in LDS. */
5004 type
== SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY
)
5005 ac_declare_lds_as_pointer(&ctx
->ac
);
5009 * Load ESGS and GSVS ring buffer resource descriptors and save the variables
5012 static void preload_ring_buffers(struct si_shader_context
*ctx
)
5014 LLVMBuilderRef builder
= ctx
->ac
.builder
;
5016 LLVMValueRef buf_ptr
= LLVMGetParam(ctx
->main_fn
,
5017 ctx
->param_rw_buffers
);
5019 if (ctx
->screen
->info
.chip_class
<= VI
&&
5020 (ctx
->shader
->key
.as_es
|| ctx
->type
== PIPE_SHADER_GEOMETRY
)) {
5022 ctx
->type
== PIPE_SHADER_GEOMETRY
? SI_GS_RING_ESGS
5024 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
, ring
, 0);
5027 ac_build_load_to_sgpr(&ctx
->ac
, buf_ptr
, offset
);
5030 if (ctx
->shader
->is_gs_copy_shader
) {
5031 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
, SI_RING_GSVS
, 0);
5034 ac_build_load_to_sgpr(&ctx
->ac
, buf_ptr
, offset
);
5035 } else if (ctx
->type
== PIPE_SHADER_GEOMETRY
) {
5036 const struct si_shader_selector
*sel
= ctx
->shader
->selector
;
5037 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
, SI_RING_GSVS
, 0);
5038 LLVMValueRef base_ring
;
5040 base_ring
= ac_build_load_to_sgpr(&ctx
->ac
, buf_ptr
, offset
);
5042 /* The conceptual layout of the GSVS ring is
5043 * v0c0 .. vLv0 v0c1 .. vLc1 ..
5044 * but the real memory layout is swizzled across
5046 * t0v0c0 .. t15v0c0 t0v1c0 .. t15v1c0 ... t15vLcL
5048 * Override the buffer descriptor accordingly.
5050 LLVMTypeRef v2i64
= LLVMVectorType(ctx
->i64
, 2);
5051 uint64_t stream_offset
= 0;
5053 for (unsigned stream
= 0; stream
< 4; ++stream
) {
5054 unsigned num_components
;
5056 unsigned num_records
;
5057 LLVMValueRef ring
, tmp
;
5059 num_components
= sel
->info
.num_stream_output_components
[stream
];
5060 if (!num_components
)
5063 stride
= 4 * num_components
* sel
->gs_max_out_vertices
;
5065 /* Limit on the stride field for <= CIK. */
5066 assert(stride
< (1 << 14));
5070 ring
= LLVMBuildBitCast(builder
, base_ring
, v2i64
, "");
5071 tmp
= LLVMBuildExtractElement(builder
, ring
, ctx
->i32_0
, "");
5072 tmp
= LLVMBuildAdd(builder
, tmp
,
5073 LLVMConstInt(ctx
->i64
,
5074 stream_offset
, 0), "");
5075 stream_offset
+= stride
* 64;
5077 ring
= LLVMBuildInsertElement(builder
, ring
, tmp
, ctx
->i32_0
, "");
5078 ring
= LLVMBuildBitCast(builder
, ring
, ctx
->v4i32
, "");
5079 tmp
= LLVMBuildExtractElement(builder
, ring
, ctx
->i32_1
, "");
5080 tmp
= LLVMBuildOr(builder
, tmp
,
5081 LLVMConstInt(ctx
->i32
,
5082 S_008F04_STRIDE(stride
) |
5083 S_008F04_SWIZZLE_ENABLE(1), 0), "");
5084 ring
= LLVMBuildInsertElement(builder
, ring
, tmp
, ctx
->i32_1
, "");
5085 ring
= LLVMBuildInsertElement(builder
, ring
,
5086 LLVMConstInt(ctx
->i32
, num_records
, 0),
5087 LLVMConstInt(ctx
->i32
, 2, 0), "");
5088 ring
= LLVMBuildInsertElement(builder
, ring
,
5089 LLVMConstInt(ctx
->i32
,
5090 S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
5091 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
5092 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
5093 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
5094 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
5095 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
5096 S_008F0C_ELEMENT_SIZE(1) | /* element_size = 4 (bytes) */
5097 S_008F0C_INDEX_STRIDE(1) | /* index_stride = 16 (elements) */
5098 S_008F0C_ADD_TID_ENABLE(1),
5100 LLVMConstInt(ctx
->i32
, 3, 0), "");
5102 ctx
->gsvs_ring
[stream
] = ring
;
5104 } else if (ctx
->type
== PIPE_SHADER_TESS_EVAL
) {
5105 ctx
->tess_offchip_ring
= get_tess_ring_descriptor(ctx
, TESS_OFFCHIP_RING_TES
);
5109 static void si_llvm_emit_polygon_stipple(struct si_shader_context
*ctx
,
5110 LLVMValueRef param_rw_buffers
,
5111 unsigned param_pos_fixed_pt
)
5113 LLVMBuilderRef builder
= ctx
->ac
.builder
;
5114 LLVMValueRef slot
, desc
, offset
, row
, bit
, address
[2];
5116 /* Use the fixed-point gl_FragCoord input.
5117 * Since the stipple pattern is 32x32 and it repeats, just get 5 bits
5118 * per coordinate to get the repeating effect.
5120 address
[0] = unpack_param(ctx
, param_pos_fixed_pt
, 0, 5);
5121 address
[1] = unpack_param(ctx
, param_pos_fixed_pt
, 16, 5);
5123 /* Load the buffer descriptor. */
5124 slot
= LLVMConstInt(ctx
->i32
, SI_PS_CONST_POLY_STIPPLE
, 0);
5125 desc
= ac_build_load_to_sgpr(&ctx
->ac
, param_rw_buffers
, slot
);
5127 /* The stipple pattern is 32x32, each row has 32 bits. */
5128 offset
= LLVMBuildMul(builder
, address
[1],
5129 LLVMConstInt(ctx
->i32
, 4, 0), "");
5130 row
= buffer_load_const(ctx
, desc
, offset
);
5131 row
= ac_to_integer(&ctx
->ac
, row
);
5132 bit
= LLVMBuildLShr(builder
, row
, address
[0], "");
5133 bit
= LLVMBuildTrunc(builder
, bit
, ctx
->i1
, "");
5134 ac_build_kill_if_false(&ctx
->ac
, bit
);
5137 void si_shader_binary_read_config(struct ac_shader_binary
*binary
,
5138 struct si_shader_config
*conf
,
5139 unsigned symbol_offset
)
5142 const unsigned char *config
=
5143 ac_shader_binary_config_start(binary
, symbol_offset
);
5144 bool really_needs_scratch
= false;
5146 /* LLVM adds SGPR spills to the scratch size.
5147 * Find out if we really need the scratch buffer.
5149 for (i
= 0; i
< binary
->reloc_count
; i
++) {
5150 const struct ac_shader_reloc
*reloc
= &binary
->relocs
[i
];
5152 if (!strcmp(scratch_rsrc_dword0_symbol
, reloc
->name
) ||
5153 !strcmp(scratch_rsrc_dword1_symbol
, reloc
->name
)) {
5154 really_needs_scratch
= true;
5159 /* XXX: We may be able to emit some of these values directly rather than
5160 * extracting fields to be emitted later.
5163 for (i
= 0; i
< binary
->config_size_per_symbol
; i
+= 8) {
5164 unsigned reg
= util_le32_to_cpu(*(uint32_t*)(config
+ i
));
5165 unsigned value
= util_le32_to_cpu(*(uint32_t*)(config
+ i
+ 4));
5167 case R_00B028_SPI_SHADER_PGM_RSRC1_PS
:
5168 case R_00B128_SPI_SHADER_PGM_RSRC1_VS
:
5169 case R_00B228_SPI_SHADER_PGM_RSRC1_GS
:
5170 case R_00B428_SPI_SHADER_PGM_RSRC1_HS
:
5171 case R_00B848_COMPUTE_PGM_RSRC1
:
5172 conf
->num_sgprs
= MAX2(conf
->num_sgprs
, (G_00B028_SGPRS(value
) + 1) * 8);
5173 conf
->num_vgprs
= MAX2(conf
->num_vgprs
, (G_00B028_VGPRS(value
) + 1) * 4);
5174 conf
->float_mode
= G_00B028_FLOAT_MODE(value
);
5175 conf
->rsrc1
= value
;
5177 case R_00B02C_SPI_SHADER_PGM_RSRC2_PS
:
5178 conf
->lds_size
= MAX2(conf
->lds_size
, G_00B02C_EXTRA_LDS_SIZE(value
));
5180 case R_00B84C_COMPUTE_PGM_RSRC2
:
5181 conf
->lds_size
= MAX2(conf
->lds_size
, G_00B84C_LDS_SIZE(value
));
5182 conf
->rsrc2
= value
;
5184 case R_0286CC_SPI_PS_INPUT_ENA
:
5185 conf
->spi_ps_input_ena
= value
;
5187 case R_0286D0_SPI_PS_INPUT_ADDR
:
5188 conf
->spi_ps_input_addr
= value
;
5190 case R_0286E8_SPI_TMPRING_SIZE
:
5191 case R_00B860_COMPUTE_TMPRING_SIZE
:
5192 /* WAVESIZE is in units of 256 dwords. */
5193 if (really_needs_scratch
)
5194 conf
->scratch_bytes_per_wave
=
5195 G_00B860_WAVESIZE(value
) * 256 * 4;
5197 case 0x4: /* SPILLED_SGPRS */
5198 conf
->spilled_sgprs
= value
;
5200 case 0x8: /* SPILLED_VGPRS */
5201 conf
->spilled_vgprs
= value
;
5205 static bool printed
;
5208 fprintf(stderr
, "Warning: LLVM emitted unknown "
5209 "config register: 0x%x\n", reg
);
5217 if (!conf
->spi_ps_input_addr
)
5218 conf
->spi_ps_input_addr
= conf
->spi_ps_input_ena
;
5221 void si_shader_apply_scratch_relocs(struct si_shader
*shader
,
5222 uint64_t scratch_va
)
5225 uint32_t scratch_rsrc_dword0
= scratch_va
;
5226 uint32_t scratch_rsrc_dword1
=
5227 S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32);
5229 /* Enable scratch coalescing. */
5230 scratch_rsrc_dword1
|= S_008F04_SWIZZLE_ENABLE(1);
5232 for (i
= 0 ; i
< shader
->binary
.reloc_count
; i
++) {
5233 const struct ac_shader_reloc
*reloc
=
5234 &shader
->binary
.relocs
[i
];
5235 if (!strcmp(scratch_rsrc_dword0_symbol
, reloc
->name
)) {
5236 util_memcpy_cpu_to_le32(shader
->binary
.code
+ reloc
->offset
,
5237 &scratch_rsrc_dword0
, 4);
5238 } else if (!strcmp(scratch_rsrc_dword1_symbol
, reloc
->name
)) {
5239 util_memcpy_cpu_to_le32(shader
->binary
.code
+ reloc
->offset
,
5240 &scratch_rsrc_dword1
, 4);
5245 static unsigned si_get_shader_binary_size(const struct si_shader
*shader
)
5247 unsigned size
= shader
->binary
.code_size
;
5250 size
+= shader
->prolog
->binary
.code_size
;
5251 if (shader
->previous_stage
)
5252 size
+= shader
->previous_stage
->binary
.code_size
;
5253 if (shader
->prolog2
)
5254 size
+= shader
->prolog2
->binary
.code_size
;
5256 size
+= shader
->epilog
->binary
.code_size
;
5260 int si_shader_binary_upload(struct si_screen
*sscreen
, struct si_shader
*shader
)
5262 const struct ac_shader_binary
*prolog
=
5263 shader
->prolog
? &shader
->prolog
->binary
: NULL
;
5264 const struct ac_shader_binary
*previous_stage
=
5265 shader
->previous_stage
? &shader
->previous_stage
->binary
: NULL
;
5266 const struct ac_shader_binary
*prolog2
=
5267 shader
->prolog2
? &shader
->prolog2
->binary
: NULL
;
5268 const struct ac_shader_binary
*epilog
=
5269 shader
->epilog
? &shader
->epilog
->binary
: NULL
;
5270 const struct ac_shader_binary
*mainb
= &shader
->binary
;
5271 unsigned bo_size
= si_get_shader_binary_size(shader
) +
5272 (!epilog
? mainb
->rodata_size
: 0);
5275 assert(!prolog
|| !prolog
->rodata_size
);
5276 assert(!previous_stage
|| !previous_stage
->rodata_size
);
5277 assert(!prolog2
|| !prolog2
->rodata_size
);
5278 assert((!prolog
&& !previous_stage
&& !prolog2
&& !epilog
) ||
5279 !mainb
->rodata_size
);
5280 assert(!epilog
|| !epilog
->rodata_size
);
5282 r600_resource_reference(&shader
->bo
, NULL
);
5283 shader
->bo
= (struct r600_resource
*)
5284 si_aligned_buffer_create(&sscreen
->b
,
5285 sscreen
->cpdma_prefetch_writes_memory
?
5286 0 : R600_RESOURCE_FLAG_READ_ONLY
,
5287 PIPE_USAGE_IMMUTABLE
,
5288 align(bo_size
, SI_CPDMA_ALIGNMENT
),
5294 ptr
= sscreen
->ws
->buffer_map(shader
->bo
->buf
, NULL
,
5295 PIPE_TRANSFER_READ_WRITE
|
5296 PIPE_TRANSFER_UNSYNCHRONIZED
);
5298 /* Don't use util_memcpy_cpu_to_le32. LLVM binaries are
5299 * endian-independent. */
5301 memcpy(ptr
, prolog
->code
, prolog
->code_size
);
5302 ptr
+= prolog
->code_size
;
5304 if (previous_stage
) {
5305 memcpy(ptr
, previous_stage
->code
, previous_stage
->code_size
);
5306 ptr
+= previous_stage
->code_size
;
5309 memcpy(ptr
, prolog2
->code
, prolog2
->code_size
);
5310 ptr
+= prolog2
->code_size
;
5313 memcpy(ptr
, mainb
->code
, mainb
->code_size
);
5314 ptr
+= mainb
->code_size
;
5317 memcpy(ptr
, epilog
->code
, epilog
->code_size
);
5318 else if (mainb
->rodata_size
> 0)
5319 memcpy(ptr
, mainb
->rodata
, mainb
->rodata_size
);
5321 sscreen
->ws
->buffer_unmap(shader
->bo
->buf
);
5325 static void si_shader_dump_disassembly(const struct ac_shader_binary
*binary
,
5326 struct pipe_debug_callback
*debug
,
5327 const char *name
, FILE *file
)
5332 if (binary
->disasm_string
) {
5333 fprintf(file
, "Shader %s disassembly:\n", name
);
5334 fprintf(file
, "%s", binary
->disasm_string
);
5336 if (debug
&& debug
->debug_message
) {
5337 /* Very long debug messages are cut off, so send the
5338 * disassembly one line at a time. This causes more
5339 * overhead, but on the plus side it simplifies
5340 * parsing of resulting logs.
5342 pipe_debug_message(debug
, SHADER_INFO
,
5343 "Shader Disassembly Begin");
5345 line
= binary
->disasm_string
;
5347 p
= util_strchrnul(line
, '\n');
5351 pipe_debug_message(debug
, SHADER_INFO
,
5352 "%.*s", count
, line
);
5360 pipe_debug_message(debug
, SHADER_INFO
,
5361 "Shader Disassembly End");
5364 fprintf(file
, "Shader %s binary:\n", name
);
5365 for (i
= 0; i
< binary
->code_size
; i
+= 4) {
5366 fprintf(file
, "@0x%x: %02x%02x%02x%02x\n", i
,
5367 binary
->code
[i
+ 3], binary
->code
[i
+ 2],
5368 binary
->code
[i
+ 1], binary
->code
[i
]);
5373 static void si_calculate_max_simd_waves(struct si_shader
*shader
)
5375 struct si_screen
*sscreen
= shader
->selector
->screen
;
5376 struct si_shader_config
*conf
= &shader
->config
;
5377 unsigned num_inputs
= shader
->selector
->info
.num_inputs
;
5378 unsigned lds_increment
= sscreen
->info
.chip_class
>= CIK
? 512 : 256;
5379 unsigned lds_per_wave
= 0;
5380 unsigned max_simd_waves
;
5382 switch (sscreen
->info
.family
) {
5383 /* These always have 8 waves: */
5384 case CHIP_POLARIS10
:
5385 case CHIP_POLARIS11
:
5386 case CHIP_POLARIS12
:
5390 max_simd_waves
= 10;
5393 /* Compute LDS usage for PS. */
5394 switch (shader
->selector
->type
) {
5395 case PIPE_SHADER_FRAGMENT
:
5396 /* The minimum usage per wave is (num_inputs * 48). The maximum
5397 * usage is (num_inputs * 48 * 16).
5398 * We can get anything in between and it varies between waves.
5400 * The 48 bytes per input for a single primitive is equal to
5401 * 4 bytes/component * 4 components/input * 3 points.
5403 * Other stages don't know the size at compile time or don't
5404 * allocate LDS per wave, but instead they do it per thread group.
5406 lds_per_wave
= conf
->lds_size
* lds_increment
+
5407 align(num_inputs
* 48, lds_increment
);
5409 case PIPE_SHADER_COMPUTE
:
5410 if (shader
->selector
) {
5411 unsigned max_workgroup_size
=
5412 si_get_max_workgroup_size(shader
);
5413 lds_per_wave
= (conf
->lds_size
* lds_increment
) /
5414 DIV_ROUND_UP(max_workgroup_size
, 64);
5419 /* Compute the per-SIMD wave counts. */
5420 if (conf
->num_sgprs
) {
5421 if (sscreen
->info
.chip_class
>= VI
)
5422 max_simd_waves
= MIN2(max_simd_waves
, 800 / conf
->num_sgprs
);
5424 max_simd_waves
= MIN2(max_simd_waves
, 512 / conf
->num_sgprs
);
5427 if (conf
->num_vgprs
)
5428 max_simd_waves
= MIN2(max_simd_waves
, 256 / conf
->num_vgprs
);
5430 /* LDS is 64KB per CU (4 SIMDs), which is 16KB per SIMD (usage above
5431 * 16KB makes some SIMDs unoccupied). */
5433 max_simd_waves
= MIN2(max_simd_waves
, 16384 / lds_per_wave
);
5435 conf
->max_simd_waves
= max_simd_waves
;
5438 void si_shader_dump_stats_for_shader_db(const struct si_shader
*shader
,
5439 struct pipe_debug_callback
*debug
)
5441 const struct si_shader_config
*conf
= &shader
->config
;
5443 pipe_debug_message(debug
, SHADER_INFO
,
5444 "Shader Stats: SGPRS: %d VGPRS: %d Code Size: %d "
5445 "LDS: %d Scratch: %d Max Waves: %d Spilled SGPRs: %d "
5446 "Spilled VGPRs: %d PrivMem VGPRs: %d",
5447 conf
->num_sgprs
, conf
->num_vgprs
,
5448 si_get_shader_binary_size(shader
),
5449 conf
->lds_size
, conf
->scratch_bytes_per_wave
,
5450 conf
->max_simd_waves
, conf
->spilled_sgprs
,
5451 conf
->spilled_vgprs
, conf
->private_mem_vgprs
);
5454 static void si_shader_dump_stats(struct si_screen
*sscreen
,
5455 const struct si_shader
*shader
,
5458 bool check_debug_option
)
5460 const struct si_shader_config
*conf
= &shader
->config
;
5462 if (!check_debug_option
||
5463 si_can_dump_shader(sscreen
, processor
)) {
5464 if (processor
== PIPE_SHADER_FRAGMENT
) {
5465 fprintf(file
, "*** SHADER CONFIG ***\n"
5466 "SPI_PS_INPUT_ADDR = 0x%04x\n"
5467 "SPI_PS_INPUT_ENA = 0x%04x\n",
5468 conf
->spi_ps_input_addr
, conf
->spi_ps_input_ena
);
5471 fprintf(file
, "*** SHADER STATS ***\n"
5474 "Spilled SGPRs: %d\n"
5475 "Spilled VGPRs: %d\n"
5476 "Private memory VGPRs: %d\n"
5477 "Code Size: %d bytes\n"
5479 "Scratch: %d bytes per wave\n"
5481 "********************\n\n\n",
5482 conf
->num_sgprs
, conf
->num_vgprs
,
5483 conf
->spilled_sgprs
, conf
->spilled_vgprs
,
5484 conf
->private_mem_vgprs
,
5485 si_get_shader_binary_size(shader
),
5486 conf
->lds_size
, conf
->scratch_bytes_per_wave
,
5487 conf
->max_simd_waves
);
5491 const char *si_get_shader_name(const struct si_shader
*shader
, unsigned processor
)
5493 switch (processor
) {
5494 case PIPE_SHADER_VERTEX
:
5495 if (shader
->key
.as_es
)
5496 return "Vertex Shader as ES";
5497 else if (shader
->key
.as_ls
)
5498 return "Vertex Shader as LS";
5500 return "Vertex Shader as VS";
5501 case PIPE_SHADER_TESS_CTRL
:
5502 return "Tessellation Control Shader";
5503 case PIPE_SHADER_TESS_EVAL
:
5504 if (shader
->key
.as_es
)
5505 return "Tessellation Evaluation Shader as ES";
5507 return "Tessellation Evaluation Shader as VS";
5508 case PIPE_SHADER_GEOMETRY
:
5509 if (shader
->is_gs_copy_shader
)
5510 return "GS Copy Shader as VS";
5512 return "Geometry Shader";
5513 case PIPE_SHADER_FRAGMENT
:
5514 return "Pixel Shader";
5515 case PIPE_SHADER_COMPUTE
:
5516 return "Compute Shader";
5518 return "Unknown Shader";
5522 void si_shader_dump(struct si_screen
*sscreen
, const struct si_shader
*shader
,
5523 struct pipe_debug_callback
*debug
, unsigned processor
,
5524 FILE *file
, bool check_debug_option
)
5526 if (!check_debug_option
||
5527 si_can_dump_shader(sscreen
, processor
))
5528 si_dump_shader_key(processor
, shader
, file
);
5530 if (!check_debug_option
&& shader
->binary
.llvm_ir_string
) {
5531 if (shader
->previous_stage
&&
5532 shader
->previous_stage
->binary
.llvm_ir_string
) {
5533 fprintf(file
, "\n%s - previous stage - LLVM IR:\n\n",
5534 si_get_shader_name(shader
, processor
));
5535 fprintf(file
, "%s\n", shader
->previous_stage
->binary
.llvm_ir_string
);
5538 fprintf(file
, "\n%s - main shader part - LLVM IR:\n\n",
5539 si_get_shader_name(shader
, processor
));
5540 fprintf(file
, "%s\n", shader
->binary
.llvm_ir_string
);
5543 if (!check_debug_option
||
5544 (si_can_dump_shader(sscreen
, processor
) &&
5545 !(sscreen
->debug_flags
& DBG(NO_ASM
)))) {
5546 fprintf(file
, "\n%s:\n", si_get_shader_name(shader
, processor
));
5549 si_shader_dump_disassembly(&shader
->prolog
->binary
,
5550 debug
, "prolog", file
);
5551 if (shader
->previous_stage
)
5552 si_shader_dump_disassembly(&shader
->previous_stage
->binary
,
5553 debug
, "previous stage", file
);
5554 if (shader
->prolog2
)
5555 si_shader_dump_disassembly(&shader
->prolog2
->binary
,
5556 debug
, "prolog2", file
);
5558 si_shader_dump_disassembly(&shader
->binary
, debug
, "main", file
);
5561 si_shader_dump_disassembly(&shader
->epilog
->binary
,
5562 debug
, "epilog", file
);
5563 fprintf(file
, "\n");
5566 si_shader_dump_stats(sscreen
, shader
, processor
, file
,
5567 check_debug_option
);
5570 static int si_compile_llvm(struct si_screen
*sscreen
,
5571 struct ac_shader_binary
*binary
,
5572 struct si_shader_config
*conf
,
5573 LLVMTargetMachineRef tm
,
5575 struct pipe_debug_callback
*debug
,
5580 unsigned count
= p_atomic_inc_return(&sscreen
->num_compilations
);
5582 if (si_can_dump_shader(sscreen
, processor
)) {
5583 fprintf(stderr
, "radeonsi: Compiling shader %d\n", count
);
5585 if (!(sscreen
->debug_flags
& (DBG(NO_IR
) | DBG(PREOPT_IR
)))) {
5586 fprintf(stderr
, "%s LLVM IR:\n\n", name
);
5587 ac_dump_module(mod
);
5588 fprintf(stderr
, "\n");
5592 if (sscreen
->record_llvm_ir
) {
5593 char *ir
= LLVMPrintModuleToString(mod
);
5594 binary
->llvm_ir_string
= strdup(ir
);
5595 LLVMDisposeMessage(ir
);
5598 if (!si_replace_shader(count
, binary
)) {
5599 r
= si_llvm_compile(mod
, binary
, tm
, debug
);
5604 si_shader_binary_read_config(binary
, conf
, 0);
5606 /* Enable 64-bit and 16-bit denormals, because there is no performance
5609 * If denormals are enabled, all floating-point output modifiers are
5612 * Don't enable denormals for 32-bit floats, because:
5613 * - Floating-point output modifiers would be ignored by the hw.
5614 * - Some opcodes don't support denormals, such as v_mad_f32. We would
5615 * have to stop using those.
5616 * - SI & CI would be very slow.
5618 conf
->float_mode
|= V_00B028_FP_64_DENORMS
;
5620 FREE(binary
->config
);
5621 FREE(binary
->global_symbol_offsets
);
5622 binary
->config
= NULL
;
5623 binary
->global_symbol_offsets
= NULL
;
5625 /* Some shaders can't have rodata because their binaries can be
5628 if (binary
->rodata_size
&&
5629 (processor
== PIPE_SHADER_VERTEX
||
5630 processor
== PIPE_SHADER_TESS_CTRL
||
5631 processor
== PIPE_SHADER_TESS_EVAL
||
5632 processor
== PIPE_SHADER_FRAGMENT
)) {
5633 fprintf(stderr
, "radeonsi: The shader can't have rodata.");
5640 static void si_llvm_build_ret(struct si_shader_context
*ctx
, LLVMValueRef ret
)
5642 if (LLVMGetTypeKind(LLVMTypeOf(ret
)) == LLVMVoidTypeKind
)
5643 LLVMBuildRetVoid(ctx
->ac
.builder
);
5645 LLVMBuildRet(ctx
->ac
.builder
, ret
);
5648 /* Generate code for the hardware VS shader stage to go with a geometry shader */
5650 si_generate_gs_copy_shader(struct si_screen
*sscreen
,
5651 LLVMTargetMachineRef tm
,
5652 struct si_shader_selector
*gs_selector
,
5653 struct pipe_debug_callback
*debug
)
5655 struct si_shader_context ctx
;
5656 struct si_shader
*shader
;
5657 LLVMBuilderRef builder
;
5658 struct lp_build_tgsi_context
*bld_base
= &ctx
.bld_base
;
5659 struct lp_build_context
*uint
= &bld_base
->uint_bld
;
5660 struct si_shader_output_values
*outputs
;
5661 struct tgsi_shader_info
*gsinfo
= &gs_selector
->info
;
5664 outputs
= MALLOC(gsinfo
->num_outputs
* sizeof(outputs
[0]));
5669 shader
= CALLOC_STRUCT(si_shader
);
5675 /* We can leave the fence as permanently signaled because the GS copy
5676 * shader only becomes visible globally after it has been compiled. */
5677 util_queue_fence_init(&shader
->ready
);
5679 shader
->selector
= gs_selector
;
5680 shader
->is_gs_copy_shader
= true;
5682 si_init_shader_ctx(&ctx
, sscreen
, tm
);
5683 ctx
.shader
= shader
;
5684 ctx
.type
= PIPE_SHADER_VERTEX
;
5686 builder
= ctx
.ac
.builder
;
5688 create_function(&ctx
);
5689 preload_ring_buffers(&ctx
);
5691 LLVMValueRef voffset
=
5692 lp_build_mul_imm(uint
, ctx
.abi
.vertex_id
, 4);
5694 /* Fetch the vertex stream ID.*/
5695 LLVMValueRef stream_id
;
5697 if (gs_selector
->so
.num_outputs
)
5698 stream_id
= unpack_param(&ctx
, ctx
.param_streamout_config
, 24, 2);
5700 stream_id
= ctx
.i32_0
;
5702 /* Fill in output information. */
5703 for (i
= 0; i
< gsinfo
->num_outputs
; ++i
) {
5704 outputs
[i
].semantic_name
= gsinfo
->output_semantic_name
[i
];
5705 outputs
[i
].semantic_index
= gsinfo
->output_semantic_index
[i
];
5707 for (int chan
= 0; chan
< 4; chan
++) {
5708 outputs
[i
].vertex_stream
[chan
] =
5709 (gsinfo
->output_streams
[i
] >> (2 * chan
)) & 3;
5713 LLVMBasicBlockRef end_bb
;
5714 LLVMValueRef switch_inst
;
5716 end_bb
= LLVMAppendBasicBlockInContext(ctx
.ac
.context
, ctx
.main_fn
, "end");
5717 switch_inst
= LLVMBuildSwitch(builder
, stream_id
, end_bb
, 4);
5719 for (int stream
= 0; stream
< 4; stream
++) {
5720 LLVMBasicBlockRef bb
;
5723 if (!gsinfo
->num_stream_output_components
[stream
])
5726 if (stream
> 0 && !gs_selector
->so
.num_outputs
)
5729 bb
= LLVMInsertBasicBlockInContext(ctx
.ac
.context
, end_bb
, "out");
5730 LLVMAddCase(switch_inst
, LLVMConstInt(ctx
.i32
, stream
, 0), bb
);
5731 LLVMPositionBuilderAtEnd(builder
, bb
);
5733 /* Fetch vertex data from GSVS ring */
5735 for (i
= 0; i
< gsinfo
->num_outputs
; ++i
) {
5736 for (unsigned chan
= 0; chan
< 4; chan
++) {
5737 if (!(gsinfo
->output_usagemask
[i
] & (1 << chan
)) ||
5738 outputs
[i
].vertex_stream
[chan
] != stream
) {
5739 outputs
[i
].values
[chan
] = ctx
.bld_base
.base
.undef
;
5743 LLVMValueRef soffset
= LLVMConstInt(ctx
.i32
,
5744 offset
* gs_selector
->gs_max_out_vertices
* 16 * 4, 0);
5747 outputs
[i
].values
[chan
] =
5748 ac_build_buffer_load(&ctx
.ac
,
5749 ctx
.gsvs_ring
[0], 1,
5756 /* Streamout and exports. */
5757 if (gs_selector
->so
.num_outputs
) {
5758 si_llvm_emit_streamout(&ctx
, outputs
,
5759 gsinfo
->num_outputs
,
5764 si_llvm_export_vs(&ctx
, outputs
, gsinfo
->num_outputs
);
5766 LLVMBuildBr(builder
, end_bb
);
5769 LLVMPositionBuilderAtEnd(builder
, end_bb
);
5771 LLVMBuildRetVoid(ctx
.ac
.builder
);
5773 ctx
.type
= PIPE_SHADER_GEOMETRY
; /* override for shader dumping */
5774 si_llvm_optimize_module(&ctx
);
5776 r
= si_compile_llvm(sscreen
, &ctx
.shader
->binary
,
5777 &ctx
.shader
->config
, ctx
.tm
,
5779 debug
, PIPE_SHADER_GEOMETRY
,
5782 if (si_can_dump_shader(sscreen
, PIPE_SHADER_GEOMETRY
))
5783 fprintf(stderr
, "GS Copy Shader:\n");
5784 si_shader_dump(sscreen
, ctx
.shader
, debug
,
5785 PIPE_SHADER_GEOMETRY
, stderr
, true);
5786 r
= si_shader_binary_upload(sscreen
, ctx
.shader
);
5789 si_llvm_dispose(&ctx
);
5800 static void si_dump_shader_key_vs(const struct si_shader_key
*key
,
5801 const struct si_vs_prolog_bits
*prolog
,
5802 const char *prefix
, FILE *f
)
5804 fprintf(f
, " %s.instance_divisor_is_one = %u\n",
5805 prefix
, prolog
->instance_divisor_is_one
);
5806 fprintf(f
, " %s.instance_divisor_is_fetched = %u\n",
5807 prefix
, prolog
->instance_divisor_is_fetched
);
5808 fprintf(f
, " %s.ls_vgpr_fix = %u\n",
5809 prefix
, prolog
->ls_vgpr_fix
);
5811 fprintf(f
, " mono.vs.fix_fetch = {");
5812 for (int i
= 0; i
< SI_MAX_ATTRIBS
; i
++)
5813 fprintf(f
, !i
? "%u" : ", %u", key
->mono
.vs_fix_fetch
[i
]);
5817 static void si_dump_shader_key(unsigned processor
, const struct si_shader
*shader
,
5820 const struct si_shader_key
*key
= &shader
->key
;
5822 fprintf(f
, "SHADER KEY\n");
5824 switch (processor
) {
5825 case PIPE_SHADER_VERTEX
:
5826 si_dump_shader_key_vs(key
, &key
->part
.vs
.prolog
,
5827 "part.vs.prolog", f
);
5828 fprintf(f
, " as_es = %u\n", key
->as_es
);
5829 fprintf(f
, " as_ls = %u\n", key
->as_ls
);
5830 fprintf(f
, " mono.u.vs_export_prim_id = %u\n",
5831 key
->mono
.u
.vs_export_prim_id
);
5834 case PIPE_SHADER_TESS_CTRL
:
5835 if (shader
->selector
->screen
->info
.chip_class
>= GFX9
) {
5836 si_dump_shader_key_vs(key
, &key
->part
.tcs
.ls_prolog
,
5837 "part.tcs.ls_prolog", f
);
5839 fprintf(f
, " part.tcs.epilog.prim_mode = %u\n", key
->part
.tcs
.epilog
.prim_mode
);
5840 fprintf(f
, " mono.u.ff_tcs_inputs_to_copy = 0x%"PRIx64
"\n", key
->mono
.u
.ff_tcs_inputs_to_copy
);
5843 case PIPE_SHADER_TESS_EVAL
:
5844 fprintf(f
, " as_es = %u\n", key
->as_es
);
5845 fprintf(f
, " mono.u.vs_export_prim_id = %u\n",
5846 key
->mono
.u
.vs_export_prim_id
);
5849 case PIPE_SHADER_GEOMETRY
:
5850 if (shader
->is_gs_copy_shader
)
5853 if (shader
->selector
->screen
->info
.chip_class
>= GFX9
&&
5854 key
->part
.gs
.es
->type
== PIPE_SHADER_VERTEX
) {
5855 si_dump_shader_key_vs(key
, &key
->part
.gs
.vs_prolog
,
5856 "part.gs.vs_prolog", f
);
5858 fprintf(f
, " part.gs.prolog.tri_strip_adj_fix = %u\n", key
->part
.gs
.prolog
.tri_strip_adj_fix
);
5861 case PIPE_SHADER_COMPUTE
:
5864 case PIPE_SHADER_FRAGMENT
:
5865 fprintf(f
, " part.ps.prolog.color_two_side = %u\n", key
->part
.ps
.prolog
.color_two_side
);
5866 fprintf(f
, " part.ps.prolog.flatshade_colors = %u\n", key
->part
.ps
.prolog
.flatshade_colors
);
5867 fprintf(f
, " part.ps.prolog.poly_stipple = %u\n", key
->part
.ps
.prolog
.poly_stipple
);
5868 fprintf(f
, " part.ps.prolog.force_persp_sample_interp = %u\n", key
->part
.ps
.prolog
.force_persp_sample_interp
);
5869 fprintf(f
, " part.ps.prolog.force_linear_sample_interp = %u\n", key
->part
.ps
.prolog
.force_linear_sample_interp
);
5870 fprintf(f
, " part.ps.prolog.force_persp_center_interp = %u\n", key
->part
.ps
.prolog
.force_persp_center_interp
);
5871 fprintf(f
, " part.ps.prolog.force_linear_center_interp = %u\n", key
->part
.ps
.prolog
.force_linear_center_interp
);
5872 fprintf(f
, " part.ps.prolog.bc_optimize_for_persp = %u\n", key
->part
.ps
.prolog
.bc_optimize_for_persp
);
5873 fprintf(f
, " part.ps.prolog.bc_optimize_for_linear = %u\n", key
->part
.ps
.prolog
.bc_optimize_for_linear
);
5874 fprintf(f
, " part.ps.epilog.spi_shader_col_format = 0x%x\n", key
->part
.ps
.epilog
.spi_shader_col_format
);
5875 fprintf(f
, " part.ps.epilog.color_is_int8 = 0x%X\n", key
->part
.ps
.epilog
.color_is_int8
);
5876 fprintf(f
, " part.ps.epilog.color_is_int10 = 0x%X\n", key
->part
.ps
.epilog
.color_is_int10
);
5877 fprintf(f
, " part.ps.epilog.last_cbuf = %u\n", key
->part
.ps
.epilog
.last_cbuf
);
5878 fprintf(f
, " part.ps.epilog.alpha_func = %u\n", key
->part
.ps
.epilog
.alpha_func
);
5879 fprintf(f
, " part.ps.epilog.alpha_to_one = %u\n", key
->part
.ps
.epilog
.alpha_to_one
);
5880 fprintf(f
, " part.ps.epilog.poly_line_smoothing = %u\n", key
->part
.ps
.epilog
.poly_line_smoothing
);
5881 fprintf(f
, " part.ps.epilog.clamp_color = %u\n", key
->part
.ps
.epilog
.clamp_color
);
5888 if ((processor
== PIPE_SHADER_GEOMETRY
||
5889 processor
== PIPE_SHADER_TESS_EVAL
||
5890 processor
== PIPE_SHADER_VERTEX
) &&
5891 !key
->as_es
&& !key
->as_ls
) {
5892 fprintf(f
, " opt.kill_outputs = 0x%"PRIx64
"\n", key
->opt
.kill_outputs
);
5893 fprintf(f
, " opt.clip_disable = %u\n", key
->opt
.clip_disable
);
5897 static void si_init_shader_ctx(struct si_shader_context
*ctx
,
5898 struct si_screen
*sscreen
,
5899 LLVMTargetMachineRef tm
)
5901 struct lp_build_tgsi_context
*bld_base
;
5903 si_llvm_context_init(ctx
, sscreen
, tm
);
5905 bld_base
= &ctx
->bld_base
;
5906 bld_base
->emit_fetch_funcs
[TGSI_FILE_CONSTANT
] = fetch_constant
;
5908 bld_base
->op_actions
[TGSI_OPCODE_INTERP_CENTROID
] = interp_action
;
5909 bld_base
->op_actions
[TGSI_OPCODE_INTERP_SAMPLE
] = interp_action
;
5910 bld_base
->op_actions
[TGSI_OPCODE_INTERP_OFFSET
] = interp_action
;
5912 bld_base
->op_actions
[TGSI_OPCODE_MEMBAR
].emit
= membar_emit
;
5914 bld_base
->op_actions
[TGSI_OPCODE_CLOCK
].emit
= clock_emit
;
5916 bld_base
->op_actions
[TGSI_OPCODE_DDX
].emit
= si_llvm_emit_ddxy
;
5917 bld_base
->op_actions
[TGSI_OPCODE_DDY
].emit
= si_llvm_emit_ddxy
;
5918 bld_base
->op_actions
[TGSI_OPCODE_DDX_FINE
].emit
= si_llvm_emit_ddxy
;
5919 bld_base
->op_actions
[TGSI_OPCODE_DDY_FINE
].emit
= si_llvm_emit_ddxy
;
5921 bld_base
->op_actions
[TGSI_OPCODE_VOTE_ALL
].emit
= vote_all_emit
;
5922 bld_base
->op_actions
[TGSI_OPCODE_VOTE_ANY
].emit
= vote_any_emit
;
5923 bld_base
->op_actions
[TGSI_OPCODE_VOTE_EQ
].emit
= vote_eq_emit
;
5924 bld_base
->op_actions
[TGSI_OPCODE_BALLOT
].emit
= ballot_emit
;
5925 bld_base
->op_actions
[TGSI_OPCODE_READ_FIRST
].intr_name
= "llvm.amdgcn.readfirstlane";
5926 bld_base
->op_actions
[TGSI_OPCODE_READ_FIRST
].emit
= read_lane_emit
;
5927 bld_base
->op_actions
[TGSI_OPCODE_READ_INVOC
].intr_name
= "llvm.amdgcn.readlane";
5928 bld_base
->op_actions
[TGSI_OPCODE_READ_INVOC
].fetch_args
= read_invoc_fetch_args
;
5929 bld_base
->op_actions
[TGSI_OPCODE_READ_INVOC
].emit
= read_lane_emit
;
5931 bld_base
->op_actions
[TGSI_OPCODE_EMIT
].emit
= si_tgsi_emit_vertex
;
5932 bld_base
->op_actions
[TGSI_OPCODE_ENDPRIM
].emit
= si_tgsi_emit_primitive
;
5933 bld_base
->op_actions
[TGSI_OPCODE_BARRIER
].emit
= si_llvm_emit_barrier
;
5936 static void si_optimize_vs_outputs(struct si_shader_context
*ctx
)
5938 struct si_shader
*shader
= ctx
->shader
;
5939 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
5941 if ((ctx
->type
!= PIPE_SHADER_VERTEX
&&
5942 ctx
->type
!= PIPE_SHADER_TESS_EVAL
) ||
5943 shader
->key
.as_ls
||
5947 ac_optimize_vs_outputs(&ctx
->ac
,
5949 shader
->info
.vs_output_param_offset
,
5951 &shader
->info
.nr_param_exports
);
5954 static void si_count_scratch_private_memory(struct si_shader_context
*ctx
)
5956 ctx
->shader
->config
.private_mem_vgprs
= 0;
5958 /* Process all LLVM instructions. */
5959 LLVMBasicBlockRef bb
= LLVMGetFirstBasicBlock(ctx
->main_fn
);
5961 LLVMValueRef next
= LLVMGetFirstInstruction(bb
);
5964 LLVMValueRef inst
= next
;
5965 next
= LLVMGetNextInstruction(next
);
5967 if (LLVMGetInstructionOpcode(inst
) != LLVMAlloca
)
5970 LLVMTypeRef type
= LLVMGetElementType(LLVMTypeOf(inst
));
5971 /* No idea why LLVM aligns allocas to 4 elements. */
5972 unsigned alignment
= LLVMGetAlignment(inst
);
5973 unsigned dw_size
= align(ac_get_type_size(type
) / 4, alignment
);
5974 ctx
->shader
->config
.private_mem_vgprs
+= dw_size
;
5976 bb
= LLVMGetNextBasicBlock(bb
);
5980 static void si_init_exec_from_input(struct si_shader_context
*ctx
,
5981 unsigned param
, unsigned bitoffset
)
5983 LLVMValueRef args
[] = {
5984 LLVMGetParam(ctx
->main_fn
, param
),
5985 LLVMConstInt(ctx
->i32
, bitoffset
, 0),
5987 lp_build_intrinsic(ctx
->ac
.builder
,
5988 "llvm.amdgcn.init.exec.from.input",
5989 ctx
->voidt
, args
, 2, LP_FUNC_ATTR_CONVERGENT
);
5992 static bool si_vs_needs_prolog(const struct si_shader_selector
*sel
,
5993 const struct si_vs_prolog_bits
*key
)
5995 /* VGPR initialization fixup for Vega10 and Raven is always done in the
5997 return sel
->vs_needs_prolog
|| key
->ls_vgpr_fix
;
6000 static bool si_compile_tgsi_main(struct si_shader_context
*ctx
,
6003 struct si_shader
*shader
= ctx
->shader
;
6004 struct si_shader_selector
*sel
= shader
->selector
;
6005 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
6007 // TODO clean all this up!
6008 switch (ctx
->type
) {
6009 case PIPE_SHADER_VERTEX
:
6010 ctx
->load_input
= declare_input_vs
;
6011 if (shader
->key
.as_ls
)
6012 ctx
->abi
.emit_outputs
= si_llvm_emit_ls_epilogue
;
6013 else if (shader
->key
.as_es
)
6014 ctx
->abi
.emit_outputs
= si_llvm_emit_es_epilogue
;
6016 ctx
->abi
.emit_outputs
= si_llvm_emit_vs_epilogue
;
6017 bld_base
->emit_epilogue
= si_tgsi_emit_epilogue
;
6019 case PIPE_SHADER_TESS_CTRL
:
6020 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = fetch_input_tcs
;
6021 ctx
->abi
.load_tess_varyings
= si_nir_load_tcs_varyings
;
6022 bld_base
->emit_fetch_funcs
[TGSI_FILE_OUTPUT
] = fetch_output_tcs
;
6023 bld_base
->emit_store
= store_output_tcs
;
6024 ctx
->abi
.store_tcs_outputs
= si_nir_store_output_tcs
;
6025 ctx
->abi
.emit_outputs
= si_llvm_emit_tcs_epilogue
;
6026 ctx
->abi
.load_patch_vertices_in
= si_load_patch_vertices_in
;
6027 bld_base
->emit_epilogue
= si_tgsi_emit_epilogue
;
6029 case PIPE_SHADER_TESS_EVAL
:
6030 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = fetch_input_tes
;
6031 ctx
->abi
.load_tess_varyings
= si_nir_load_input_tes
;
6032 ctx
->abi
.load_tess_coord
= si_load_tess_coord
;
6033 ctx
->abi
.load_tess_level
= si_load_tess_level
;
6034 ctx
->abi
.load_patch_vertices_in
= si_load_patch_vertices_in
;
6035 if (shader
->key
.as_es
)
6036 ctx
->abi
.emit_outputs
= si_llvm_emit_es_epilogue
;
6038 ctx
->abi
.emit_outputs
= si_llvm_emit_vs_epilogue
;
6039 bld_base
->emit_epilogue
= si_tgsi_emit_epilogue
;
6041 case PIPE_SHADER_GEOMETRY
:
6042 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = fetch_input_gs
;
6043 ctx
->abi
.load_inputs
= si_nir_load_input_gs
;
6044 ctx
->abi
.emit_vertex
= si_llvm_emit_vertex
;
6045 ctx
->abi
.emit_primitive
= si_llvm_emit_primitive
;
6046 ctx
->abi
.emit_outputs
= si_llvm_emit_gs_epilogue
;
6047 bld_base
->emit_epilogue
= si_tgsi_emit_gs_epilogue
;
6049 case PIPE_SHADER_FRAGMENT
:
6050 ctx
->load_input
= declare_input_fs
;
6051 ctx
->abi
.emit_outputs
= si_llvm_return_fs_outputs
;
6052 bld_base
->emit_epilogue
= si_tgsi_emit_epilogue
;
6053 ctx
->abi
.lookup_interp_param
= si_nir_lookup_interp_param
;
6054 ctx
->abi
.load_sample_position
= load_sample_position
;
6055 ctx
->abi
.load_sample_mask_in
= load_sample_mask_in
;
6057 case PIPE_SHADER_COMPUTE
:
6058 ctx
->abi
.load_local_group_size
= get_block_size
;
6061 assert(!"Unsupported shader type");
6065 ctx
->abi
.load_ubo
= load_ubo
;
6066 ctx
->abi
.load_ssbo
= load_ssbo
;
6068 create_function(ctx
);
6069 preload_ring_buffers(ctx
);
6071 /* For GFX9 merged shaders:
6072 * - Set EXEC for the first shader. If the prolog is present, set
6073 * EXEC there instead.
6074 * - Add a barrier before the second shader.
6075 * - In the second shader, reset EXEC to ~0 and wrap the main part in
6076 * an if-statement. This is required for correctness in geometry
6077 * shaders, to ensure that empty GS waves do not send GS_EMIT and
6080 * For monolithic merged shaders, the first shader is wrapped in an
6081 * if-block together with its prolog in si_build_wrapper_function.
6083 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
6084 if (!is_monolithic
&&
6085 sel
->info
.num_instructions
> 1 && /* not empty shader */
6086 (shader
->key
.as_es
|| shader
->key
.as_ls
) &&
6087 (ctx
->type
== PIPE_SHADER_TESS_EVAL
||
6088 (ctx
->type
== PIPE_SHADER_VERTEX
&&
6089 !si_vs_needs_prolog(sel
, &shader
->key
.part
.vs
.prolog
)))) {
6090 si_init_exec_from_input(ctx
,
6091 ctx
->param_merged_wave_info
, 0);
6092 } else if (ctx
->type
== PIPE_SHADER_TESS_CTRL
||
6093 ctx
->type
== PIPE_SHADER_GEOMETRY
) {
6095 ac_init_exec_full_mask(&ctx
->ac
);
6097 /* The barrier must execute for all shaders in a
6100 si_llvm_emit_barrier(NULL
, bld_base
, NULL
);
6102 LLVMValueRef num_threads
= unpack_param(ctx
, ctx
->param_merged_wave_info
, 8, 8);
6104 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntULT
,
6105 ac_get_thread_id(&ctx
->ac
), num_threads
, "");
6106 lp_build_if(&ctx
->merged_wrap_if_state
, &ctx
->gallivm
, ena
);
6110 if (ctx
->type
== PIPE_SHADER_TESS_CTRL
&&
6111 sel
->tcs_info
.tessfactors_are_def_in_all_invocs
) {
6112 for (unsigned i
= 0; i
< 6; i
++) {
6113 ctx
->invoc0_tess_factors
[i
] =
6114 lp_build_alloca_undef(&ctx
->gallivm
, ctx
->i32
, "");
6118 if (ctx
->type
== PIPE_SHADER_GEOMETRY
) {
6120 for (i
= 0; i
< 4; i
++) {
6121 ctx
->gs_next_vertex
[i
] =
6122 lp_build_alloca(&ctx
->gallivm
,
6127 if (sel
->force_correct_derivs_after_kill
) {
6128 ctx
->postponed_kill
= lp_build_alloca_undef(&ctx
->gallivm
, ctx
->i1
, "");
6129 /* true = don't kill. */
6130 LLVMBuildStore(ctx
->ac
.builder
, LLVMConstInt(ctx
->i1
, 1, 0),
6131 ctx
->postponed_kill
);
6135 if (!lp_build_tgsi_llvm(bld_base
, sel
->tokens
)) {
6136 fprintf(stderr
, "Failed to translate shader from TGSI to LLVM\n");
6140 if (!si_nir_build_llvm(ctx
, sel
->nir
)) {
6141 fprintf(stderr
, "Failed to translate shader from NIR to LLVM\n");
6146 si_llvm_build_ret(ctx
, ctx
->return_value
);
6151 * Compute the VS prolog key, which contains all the information needed to
6152 * build the VS prolog function, and set shader->info bits where needed.
6154 * \param info Shader info of the vertex shader.
6155 * \param num_input_sgprs Number of input SGPRs for the vertex shader.
6156 * \param prolog_key Key of the VS prolog
6157 * \param shader_out The vertex shader, or the next shader if merging LS+HS or ES+GS.
6158 * \param key Output shader part key.
6160 static void si_get_vs_prolog_key(const struct tgsi_shader_info
*info
,
6161 unsigned num_input_sgprs
,
6162 const struct si_vs_prolog_bits
*prolog_key
,
6163 struct si_shader
*shader_out
,
6164 union si_shader_part_key
*key
)
6166 memset(key
, 0, sizeof(*key
));
6167 key
->vs_prolog
.states
= *prolog_key
;
6168 key
->vs_prolog
.num_input_sgprs
= num_input_sgprs
;
6169 key
->vs_prolog
.last_input
= MAX2(1, info
->num_inputs
) - 1;
6170 key
->vs_prolog
.as_ls
= shader_out
->key
.as_ls
;
6171 key
->vs_prolog
.as_es
= shader_out
->key
.as_es
;
6173 if (shader_out
->selector
->type
== PIPE_SHADER_TESS_CTRL
) {
6174 key
->vs_prolog
.as_ls
= 1;
6175 key
->vs_prolog
.num_merged_next_stage_vgprs
= 2;
6176 } else if (shader_out
->selector
->type
== PIPE_SHADER_GEOMETRY
) {
6177 key
->vs_prolog
.as_es
= 1;
6178 key
->vs_prolog
.num_merged_next_stage_vgprs
= 5;
6181 /* Enable loading the InstanceID VGPR. */
6182 uint16_t input_mask
= u_bit_consecutive(0, info
->num_inputs
);
6184 if ((key
->vs_prolog
.states
.instance_divisor_is_one
|
6185 key
->vs_prolog
.states
.instance_divisor_is_fetched
) & input_mask
)
6186 shader_out
->info
.uses_instanceid
= true;
6190 * Compute the PS prolog key, which contains all the information needed to
6191 * build the PS prolog function, and set related bits in shader->config.
6193 static void si_get_ps_prolog_key(struct si_shader
*shader
,
6194 union si_shader_part_key
*key
,
6195 bool separate_prolog
)
6197 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
6199 memset(key
, 0, sizeof(*key
));
6200 key
->ps_prolog
.states
= shader
->key
.part
.ps
.prolog
;
6201 key
->ps_prolog
.colors_read
= info
->colors_read
;
6202 key
->ps_prolog
.num_input_sgprs
= shader
->info
.num_input_sgprs
;
6203 key
->ps_prolog
.num_input_vgprs
= shader
->info
.num_input_vgprs
;
6204 key
->ps_prolog
.wqm
= info
->uses_derivatives
&&
6205 (key
->ps_prolog
.colors_read
||
6206 key
->ps_prolog
.states
.force_persp_sample_interp
||
6207 key
->ps_prolog
.states
.force_linear_sample_interp
||
6208 key
->ps_prolog
.states
.force_persp_center_interp
||
6209 key
->ps_prolog
.states
.force_linear_center_interp
||
6210 key
->ps_prolog
.states
.bc_optimize_for_persp
||
6211 key
->ps_prolog
.states
.bc_optimize_for_linear
);
6212 key
->ps_prolog
.ancillary_vgpr_index
= shader
->info
.ancillary_vgpr_index
;
6214 if (info
->colors_read
) {
6215 unsigned *color
= shader
->selector
->color_attr_index
;
6217 if (shader
->key
.part
.ps
.prolog
.color_two_side
) {
6218 /* BCOLORs are stored after the last input. */
6219 key
->ps_prolog
.num_interp_inputs
= info
->num_inputs
;
6220 key
->ps_prolog
.face_vgpr_index
= shader
->info
.face_vgpr_index
;
6221 shader
->config
.spi_ps_input_ena
|= S_0286CC_FRONT_FACE_ENA(1);
6224 for (unsigned i
= 0; i
< 2; i
++) {
6225 unsigned interp
= info
->input_interpolate
[color
[i
]];
6226 unsigned location
= info
->input_interpolate_loc
[color
[i
]];
6228 if (!(info
->colors_read
& (0xf << i
*4)))
6231 key
->ps_prolog
.color_attr_index
[i
] = color
[i
];
6233 if (shader
->key
.part
.ps
.prolog
.flatshade_colors
&&
6234 interp
== TGSI_INTERPOLATE_COLOR
)
6235 interp
= TGSI_INTERPOLATE_CONSTANT
;
6238 case TGSI_INTERPOLATE_CONSTANT
:
6239 key
->ps_prolog
.color_interp_vgpr_index
[i
] = -1;
6241 case TGSI_INTERPOLATE_PERSPECTIVE
:
6242 case TGSI_INTERPOLATE_COLOR
:
6243 /* Force the interpolation location for colors here. */
6244 if (shader
->key
.part
.ps
.prolog
.force_persp_sample_interp
)
6245 location
= TGSI_INTERPOLATE_LOC_SAMPLE
;
6246 if (shader
->key
.part
.ps
.prolog
.force_persp_center_interp
)
6247 location
= TGSI_INTERPOLATE_LOC_CENTER
;
6250 case TGSI_INTERPOLATE_LOC_SAMPLE
:
6251 key
->ps_prolog
.color_interp_vgpr_index
[i
] = 0;
6252 shader
->config
.spi_ps_input_ena
|=
6253 S_0286CC_PERSP_SAMPLE_ENA(1);
6255 case TGSI_INTERPOLATE_LOC_CENTER
:
6256 key
->ps_prolog
.color_interp_vgpr_index
[i
] = 2;
6257 shader
->config
.spi_ps_input_ena
|=
6258 S_0286CC_PERSP_CENTER_ENA(1);
6260 case TGSI_INTERPOLATE_LOC_CENTROID
:
6261 key
->ps_prolog
.color_interp_vgpr_index
[i
] = 4;
6262 shader
->config
.spi_ps_input_ena
|=
6263 S_0286CC_PERSP_CENTROID_ENA(1);
6269 case TGSI_INTERPOLATE_LINEAR
:
6270 /* Force the interpolation location for colors here. */
6271 if (shader
->key
.part
.ps
.prolog
.force_linear_sample_interp
)
6272 location
= TGSI_INTERPOLATE_LOC_SAMPLE
;
6273 if (shader
->key
.part
.ps
.prolog
.force_linear_center_interp
)
6274 location
= TGSI_INTERPOLATE_LOC_CENTER
;
6276 /* The VGPR assignment for non-monolithic shaders
6277 * works because InitialPSInputAddr is set on the
6278 * main shader and PERSP_PULL_MODEL is never used.
6281 case TGSI_INTERPOLATE_LOC_SAMPLE
:
6282 key
->ps_prolog
.color_interp_vgpr_index
[i
] =
6283 separate_prolog
? 6 : 9;
6284 shader
->config
.spi_ps_input_ena
|=
6285 S_0286CC_LINEAR_SAMPLE_ENA(1);
6287 case TGSI_INTERPOLATE_LOC_CENTER
:
6288 key
->ps_prolog
.color_interp_vgpr_index
[i
] =
6289 separate_prolog
? 8 : 11;
6290 shader
->config
.spi_ps_input_ena
|=
6291 S_0286CC_LINEAR_CENTER_ENA(1);
6293 case TGSI_INTERPOLATE_LOC_CENTROID
:
6294 key
->ps_prolog
.color_interp_vgpr_index
[i
] =
6295 separate_prolog
? 10 : 13;
6296 shader
->config
.spi_ps_input_ena
|=
6297 S_0286CC_LINEAR_CENTROID_ENA(1);
6311 * Check whether a PS prolog is required based on the key.
6313 static bool si_need_ps_prolog(const union si_shader_part_key
*key
)
6315 return key
->ps_prolog
.colors_read
||
6316 key
->ps_prolog
.states
.force_persp_sample_interp
||
6317 key
->ps_prolog
.states
.force_linear_sample_interp
||
6318 key
->ps_prolog
.states
.force_persp_center_interp
||
6319 key
->ps_prolog
.states
.force_linear_center_interp
||
6320 key
->ps_prolog
.states
.bc_optimize_for_persp
||
6321 key
->ps_prolog
.states
.bc_optimize_for_linear
||
6322 key
->ps_prolog
.states
.poly_stipple
||
6323 key
->ps_prolog
.states
.samplemask_log_ps_iter
;
6327 * Compute the PS epilog key, which contains all the information needed to
6328 * build the PS epilog function.
6330 static void si_get_ps_epilog_key(struct si_shader
*shader
,
6331 union si_shader_part_key
*key
)
6333 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
6334 memset(key
, 0, sizeof(*key
));
6335 key
->ps_epilog
.colors_written
= info
->colors_written
;
6336 key
->ps_epilog
.writes_z
= info
->writes_z
;
6337 key
->ps_epilog
.writes_stencil
= info
->writes_stencil
;
6338 key
->ps_epilog
.writes_samplemask
= info
->writes_samplemask
;
6339 key
->ps_epilog
.states
= shader
->key
.part
.ps
.epilog
;
6343 * Build the GS prolog function. Rotate the input vertices for triangle strips
6346 static void si_build_gs_prolog_function(struct si_shader_context
*ctx
,
6347 union si_shader_part_key
*key
)
6349 unsigned num_sgprs
, num_vgprs
;
6350 struct si_function_info fninfo
;
6351 LLVMBuilderRef builder
= ctx
->ac
.builder
;
6352 LLVMTypeRef returns
[48];
6353 LLVMValueRef func
, ret
;
6355 si_init_function_info(&fninfo
);
6357 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
6358 if (key
->gs_prolog
.states
.gfx9_prev_is_vs
)
6359 num_sgprs
= 8 + GFX9_VSGS_NUM_USER_SGPR
;
6361 num_sgprs
= 8 + GFX9_TESGS_NUM_USER_SGPR
;
6362 num_vgprs
= 5; /* ES inputs are not needed by GS */
6364 num_sgprs
= GFX6_GS_NUM_USER_SGPR
+ 2;
6368 for (unsigned i
= 0; i
< num_sgprs
; ++i
) {
6369 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
6370 returns
[i
] = ctx
->i32
;
6373 for (unsigned i
= 0; i
< num_vgprs
; ++i
) {
6374 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
6375 returns
[num_sgprs
+ i
] = ctx
->f32
;
6378 /* Create the function. */
6379 si_create_function(ctx
, "gs_prolog", returns
, num_sgprs
+ num_vgprs
,
6381 func
= ctx
->main_fn
;
6383 /* Set the full EXEC mask for the prolog, because we are only fiddling
6384 * with registers here. The main shader part will set the correct EXEC
6387 if (ctx
->screen
->info
.chip_class
>= GFX9
&& !key
->gs_prolog
.is_monolithic
)
6388 ac_init_exec_full_mask(&ctx
->ac
);
6390 /* Copy inputs to outputs. This should be no-op, as the registers match,
6391 * but it will prevent the compiler from overwriting them unintentionally.
6393 ret
= ctx
->return_value
;
6394 for (unsigned i
= 0; i
< num_sgprs
; i
++) {
6395 LLVMValueRef p
= LLVMGetParam(func
, i
);
6396 ret
= LLVMBuildInsertValue(builder
, ret
, p
, i
, "");
6398 for (unsigned i
= 0; i
< num_vgprs
; i
++) {
6399 LLVMValueRef p
= LLVMGetParam(func
, num_sgprs
+ i
);
6400 p
= ac_to_float(&ctx
->ac
, p
);
6401 ret
= LLVMBuildInsertValue(builder
, ret
, p
, num_sgprs
+ i
, "");
6404 if (key
->gs_prolog
.states
.tri_strip_adj_fix
) {
6405 /* Remap the input vertices for every other primitive. */
6406 const unsigned gfx6_vtx_params
[6] = {
6414 const unsigned gfx9_vtx_params
[3] = {
6419 LLVMValueRef vtx_in
[6], vtx_out
[6];
6420 LLVMValueRef prim_id
, rotate
;
6422 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
6423 for (unsigned i
= 0; i
< 3; i
++) {
6424 vtx_in
[i
*2] = unpack_param(ctx
, gfx9_vtx_params
[i
], 0, 16);
6425 vtx_in
[i
*2+1] = unpack_param(ctx
, gfx9_vtx_params
[i
], 16, 16);
6428 for (unsigned i
= 0; i
< 6; i
++)
6429 vtx_in
[i
] = LLVMGetParam(func
, gfx6_vtx_params
[i
]);
6432 prim_id
= LLVMGetParam(func
, num_sgprs
+ 2);
6433 rotate
= LLVMBuildTrunc(builder
, prim_id
, ctx
->i1
, "");
6435 for (unsigned i
= 0; i
< 6; ++i
) {
6436 LLVMValueRef base
, rotated
;
6438 rotated
= vtx_in
[(i
+ 4) % 6];
6439 vtx_out
[i
] = LLVMBuildSelect(builder
, rotate
, rotated
, base
, "");
6442 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
6443 for (unsigned i
= 0; i
< 3; i
++) {
6444 LLVMValueRef hi
, out
;
6446 hi
= LLVMBuildShl(builder
, vtx_out
[i
*2+1],
6447 LLVMConstInt(ctx
->i32
, 16, 0), "");
6448 out
= LLVMBuildOr(builder
, vtx_out
[i
*2], hi
, "");
6449 out
= ac_to_float(&ctx
->ac
, out
);
6450 ret
= LLVMBuildInsertValue(builder
, ret
, out
,
6451 gfx9_vtx_params
[i
], "");
6454 for (unsigned i
= 0; i
< 6; i
++) {
6457 out
= ac_to_float(&ctx
->ac
, vtx_out
[i
]);
6458 ret
= LLVMBuildInsertValue(builder
, ret
, out
,
6459 gfx6_vtx_params
[i
], "");
6464 LLVMBuildRet(builder
, ret
);
6468 * Given a list of shader part functions, build a wrapper function that
6469 * runs them in sequence to form a monolithic shader.
6471 static void si_build_wrapper_function(struct si_shader_context
*ctx
,
6472 LLVMValueRef
*parts
,
6475 unsigned next_shader_first_part
)
6477 LLVMBuilderRef builder
= ctx
->ac
.builder
;
6478 /* PS epilog has one arg per color component; gfx9 merged shader
6479 * prologs need to forward 32 user SGPRs.
6481 struct si_function_info fninfo
;
6482 LLVMValueRef initial
[64], out
[64];
6483 LLVMTypeRef function_type
;
6484 unsigned num_first_params
;
6485 unsigned num_out
, initial_num_out
;
6486 MAYBE_UNUSED
unsigned num_out_sgpr
; /* used in debug checks */
6487 MAYBE_UNUSED
unsigned initial_num_out_sgpr
; /* used in debug checks */
6488 unsigned num_sgprs
, num_vgprs
;
6490 struct lp_build_if_state if_state
;
6492 si_init_function_info(&fninfo
);
6494 for (unsigned i
= 0; i
< num_parts
; ++i
) {
6495 lp_add_function_attr(parts
[i
], -1, LP_FUNC_ATTR_ALWAYSINLINE
);
6496 LLVMSetLinkage(parts
[i
], LLVMPrivateLinkage
);
6499 /* The parameters of the wrapper function correspond to those of the
6500 * first part in terms of SGPRs and VGPRs, but we use the types of the
6501 * main part to get the right types. This is relevant for the
6502 * dereferenceable attribute on descriptor table pointers.
6507 function_type
= LLVMGetElementType(LLVMTypeOf(parts
[0]));
6508 num_first_params
= LLVMCountParamTypes(function_type
);
6510 for (unsigned i
= 0; i
< num_first_params
; ++i
) {
6511 LLVMValueRef param
= LLVMGetParam(parts
[0], i
);
6513 if (ac_is_sgpr_param(param
)) {
6514 assert(num_vgprs
== 0);
6515 num_sgprs
+= ac_get_type_size(LLVMTypeOf(param
)) / 4;
6517 num_vgprs
+= ac_get_type_size(LLVMTypeOf(param
)) / 4;
6522 while (gprs
< num_sgprs
+ num_vgprs
) {
6523 LLVMValueRef param
= LLVMGetParam(parts
[main_part
], fninfo
.num_params
);
6524 LLVMTypeRef type
= LLVMTypeOf(param
);
6525 unsigned size
= ac_get_type_size(type
) / 4;
6527 add_arg(&fninfo
, gprs
< num_sgprs
? ARG_SGPR
: ARG_VGPR
, type
);
6529 assert(ac_is_sgpr_param(param
) == (gprs
< num_sgprs
));
6530 assert(gprs
+ size
<= num_sgprs
+ num_vgprs
&&
6531 (gprs
>= num_sgprs
|| gprs
+ size
<= num_sgprs
));
6536 si_create_function(ctx
, "wrapper", NULL
, 0, &fninfo
,
6537 si_get_max_workgroup_size(ctx
->shader
));
6539 if (is_merged_shader(ctx
->shader
))
6540 ac_init_exec_full_mask(&ctx
->ac
);
6542 /* Record the arguments of the function as if they were an output of
6548 for (unsigned i
= 0; i
< fninfo
.num_params
; ++i
) {
6549 LLVMValueRef param
= LLVMGetParam(ctx
->main_fn
, i
);
6550 LLVMTypeRef param_type
= LLVMTypeOf(param
);
6551 LLVMTypeRef out_type
= i
< fninfo
.num_sgpr_params
? ctx
->i32
: ctx
->f32
;
6552 unsigned size
= ac_get_type_size(param_type
) / 4;
6555 if (LLVMGetTypeKind(param_type
) == LLVMPointerTypeKind
) {
6556 param
= LLVMBuildPtrToInt(builder
, param
, ctx
->i32
, "");
6557 param_type
= ctx
->i32
;
6560 if (param_type
!= out_type
)
6561 param
= LLVMBuildBitCast(builder
, param
, out_type
, "");
6562 out
[num_out
++] = param
;
6564 LLVMTypeRef vector_type
= LLVMVectorType(out_type
, size
);
6566 if (LLVMGetTypeKind(param_type
) == LLVMPointerTypeKind
) {
6567 param
= LLVMBuildPtrToInt(builder
, param
, ctx
->i64
, "");
6568 param_type
= ctx
->i64
;
6571 if (param_type
!= vector_type
)
6572 param
= LLVMBuildBitCast(builder
, param
, vector_type
, "");
6574 for (unsigned j
= 0; j
< size
; ++j
)
6575 out
[num_out
++] = LLVMBuildExtractElement(
6576 builder
, param
, LLVMConstInt(ctx
->i32
, j
, 0), "");
6579 if (i
< fninfo
.num_sgpr_params
)
6580 num_out_sgpr
= num_out
;
6583 memcpy(initial
, out
, sizeof(out
));
6584 initial_num_out
= num_out
;
6585 initial_num_out_sgpr
= num_out_sgpr
;
6587 /* Now chain the parts. */
6588 for (unsigned part
= 0; part
< num_parts
; ++part
) {
6589 LLVMValueRef in
[48];
6591 LLVMTypeRef ret_type
;
6592 unsigned out_idx
= 0;
6593 unsigned num_params
= LLVMCountParams(parts
[part
]);
6595 /* Merged shaders are executed conditionally depending
6596 * on the number of enabled threads passed in the input SGPRs. */
6597 if (is_merged_shader(ctx
->shader
) && part
== 0) {
6598 LLVMValueRef ena
, count
= initial
[3];
6600 count
= LLVMBuildAnd(builder
, count
,
6601 LLVMConstInt(ctx
->i32
, 0x7f, 0), "");
6602 ena
= LLVMBuildICmp(builder
, LLVMIntULT
,
6603 ac_get_thread_id(&ctx
->ac
), count
, "");
6604 lp_build_if(&if_state
, &ctx
->gallivm
, ena
);
6607 /* Derive arguments for the next part from outputs of the
6610 for (unsigned param_idx
= 0; param_idx
< num_params
; ++param_idx
) {
6612 LLVMTypeRef param_type
;
6614 unsigned param_size
;
6615 LLVMValueRef arg
= NULL
;
6617 param
= LLVMGetParam(parts
[part
], param_idx
);
6618 param_type
= LLVMTypeOf(param
);
6619 param_size
= ac_get_type_size(param_type
) / 4;
6620 is_sgpr
= ac_is_sgpr_param(param
);
6623 lp_add_function_attr(parts
[part
], param_idx
+ 1, LP_FUNC_ATTR_INREG
);
6624 else if (out_idx
< num_out_sgpr
) {
6625 /* Skip returned SGPRs the current part doesn't
6626 * declare on the input. */
6627 out_idx
= num_out_sgpr
;
6630 assert(out_idx
+ param_size
<= (is_sgpr
? num_out_sgpr
: num_out
));
6632 if (param_size
== 1)
6635 arg
= lp_build_gather_values(&ctx
->gallivm
, &out
[out_idx
], param_size
);
6637 if (LLVMTypeOf(arg
) != param_type
) {
6638 if (LLVMGetTypeKind(param_type
) == LLVMPointerTypeKind
) {
6639 if (LLVMGetPointerAddressSpace(param_type
) ==
6640 AC_CONST_32BIT_ADDR_SPACE
) {
6641 arg
= LLVMBuildBitCast(builder
, arg
, ctx
->i32
, "");
6642 arg
= LLVMBuildIntToPtr(builder
, arg
, param_type
, "");
6644 arg
= LLVMBuildBitCast(builder
, arg
, ctx
->i64
, "");
6645 arg
= LLVMBuildIntToPtr(builder
, arg
, param_type
, "");
6648 arg
= LLVMBuildBitCast(builder
, arg
, param_type
, "");
6652 in
[param_idx
] = arg
;
6653 out_idx
+= param_size
;
6656 ret
= LLVMBuildCall(builder
, parts
[part
], in
, num_params
, "");
6658 if (is_merged_shader(ctx
->shader
) &&
6659 part
+ 1 == next_shader_first_part
) {
6660 lp_build_endif(&if_state
);
6662 /* The second half of the merged shader should use
6663 * the inputs from the toplevel (wrapper) function,
6664 * not the return value from the last call.
6666 * That's because the last call was executed condi-
6667 * tionally, so we can't consume it in the main
6670 memcpy(out
, initial
, sizeof(initial
));
6671 num_out
= initial_num_out
;
6672 num_out_sgpr
= initial_num_out_sgpr
;
6676 /* Extract the returned GPRs. */
6677 ret_type
= LLVMTypeOf(ret
);
6681 if (LLVMGetTypeKind(ret_type
) != LLVMVoidTypeKind
) {
6682 assert(LLVMGetTypeKind(ret_type
) == LLVMStructTypeKind
);
6684 unsigned ret_size
= LLVMCountStructElementTypes(ret_type
);
6686 for (unsigned i
= 0; i
< ret_size
; ++i
) {
6688 LLVMBuildExtractValue(builder
, ret
, i
, "");
6690 assert(num_out
< ARRAY_SIZE(out
));
6691 out
[num_out
++] = val
;
6693 if (LLVMTypeOf(val
) == ctx
->i32
) {
6694 assert(num_out_sgpr
+ 1 == num_out
);
6695 num_out_sgpr
= num_out
;
6701 LLVMBuildRetVoid(builder
);
6704 int si_compile_tgsi_shader(struct si_screen
*sscreen
,
6705 LLVMTargetMachineRef tm
,
6706 struct si_shader
*shader
,
6708 struct pipe_debug_callback
*debug
)
6710 struct si_shader_selector
*sel
= shader
->selector
;
6711 struct si_shader_context ctx
;
6714 /* Dump TGSI code before doing TGSI->LLVM conversion in case the
6715 * conversion fails. */
6716 if (si_can_dump_shader(sscreen
, sel
->info
.processor
) &&
6717 !(sscreen
->debug_flags
& DBG(NO_TGSI
))) {
6719 tgsi_dump(sel
->tokens
, 0);
6721 nir_print_shader(sel
->nir
, stderr
);
6722 si_dump_streamout(&sel
->so
);
6725 si_init_shader_ctx(&ctx
, sscreen
, tm
);
6726 si_llvm_context_set_tgsi(&ctx
, shader
);
6727 ctx
.separate_prolog
= !is_monolithic
;
6729 memset(shader
->info
.vs_output_param_offset
, AC_EXP_PARAM_UNDEFINED
,
6730 sizeof(shader
->info
.vs_output_param_offset
));
6732 shader
->info
.uses_instanceid
= sel
->info
.uses_instanceid
;
6734 if (!si_compile_tgsi_main(&ctx
, is_monolithic
)) {
6735 si_llvm_dispose(&ctx
);
6739 if (is_monolithic
&& ctx
.type
== PIPE_SHADER_VERTEX
) {
6740 LLVMValueRef parts
[2];
6741 bool need_prolog
= sel
->vs_needs_prolog
;
6743 parts
[1] = ctx
.main_fn
;
6746 union si_shader_part_key prolog_key
;
6747 si_get_vs_prolog_key(&sel
->info
,
6748 shader
->info
.num_input_sgprs
,
6749 &shader
->key
.part
.vs
.prolog
,
6750 shader
, &prolog_key
);
6751 si_build_vs_prolog_function(&ctx
, &prolog_key
);
6752 parts
[0] = ctx
.main_fn
;
6755 si_build_wrapper_function(&ctx
, parts
+ !need_prolog
,
6756 1 + need_prolog
, need_prolog
, 0);
6757 } else if (is_monolithic
&& ctx
.type
== PIPE_SHADER_TESS_CTRL
) {
6758 if (sscreen
->info
.chip_class
>= GFX9
) {
6759 struct si_shader_selector
*ls
= shader
->key
.part
.tcs
.ls
;
6760 LLVMValueRef parts
[4];
6761 bool vs_needs_prolog
=
6762 si_vs_needs_prolog(ls
, &shader
->key
.part
.tcs
.ls_prolog
);
6765 parts
[2] = ctx
.main_fn
;
6768 union si_shader_part_key tcs_epilog_key
;
6769 memset(&tcs_epilog_key
, 0, sizeof(tcs_epilog_key
));
6770 tcs_epilog_key
.tcs_epilog
.states
= shader
->key
.part
.tcs
.epilog
;
6771 si_build_tcs_epilog_function(&ctx
, &tcs_epilog_key
);
6772 parts
[3] = ctx
.main_fn
;
6774 /* VS as LS main part */
6775 struct si_shader shader_ls
= {};
6776 shader_ls
.selector
= ls
;
6777 shader_ls
.key
.as_ls
= 1;
6778 shader_ls
.key
.mono
= shader
->key
.mono
;
6779 shader_ls
.key
.opt
= shader
->key
.opt
;
6780 si_llvm_context_set_tgsi(&ctx
, &shader_ls
);
6782 if (!si_compile_tgsi_main(&ctx
, true)) {
6783 si_llvm_dispose(&ctx
);
6786 shader
->info
.uses_instanceid
|= ls
->info
.uses_instanceid
;
6787 parts
[1] = ctx
.main_fn
;
6790 if (vs_needs_prolog
) {
6791 union si_shader_part_key vs_prolog_key
;
6792 si_get_vs_prolog_key(&ls
->info
,
6793 shader_ls
.info
.num_input_sgprs
,
6794 &shader
->key
.part
.tcs
.ls_prolog
,
6795 shader
, &vs_prolog_key
);
6796 vs_prolog_key
.vs_prolog
.is_monolithic
= true;
6797 si_build_vs_prolog_function(&ctx
, &vs_prolog_key
);
6798 parts
[0] = ctx
.main_fn
;
6801 /* Reset the shader context. */
6802 ctx
.shader
= shader
;
6803 ctx
.type
= PIPE_SHADER_TESS_CTRL
;
6805 si_build_wrapper_function(&ctx
,
6806 parts
+ !vs_needs_prolog
,
6807 4 - !vs_needs_prolog
, 0,
6808 vs_needs_prolog
? 2 : 1);
6810 LLVMValueRef parts
[2];
6811 union si_shader_part_key epilog_key
;
6813 parts
[0] = ctx
.main_fn
;
6815 memset(&epilog_key
, 0, sizeof(epilog_key
));
6816 epilog_key
.tcs_epilog
.states
= shader
->key
.part
.tcs
.epilog
;
6817 si_build_tcs_epilog_function(&ctx
, &epilog_key
);
6818 parts
[1] = ctx
.main_fn
;
6820 si_build_wrapper_function(&ctx
, parts
, 2, 0, 0);
6822 } else if (is_monolithic
&& ctx
.type
== PIPE_SHADER_GEOMETRY
) {
6823 if (ctx
.screen
->info
.chip_class
>= GFX9
) {
6824 struct si_shader_selector
*es
= shader
->key
.part
.gs
.es
;
6825 LLVMValueRef es_prolog
= NULL
;
6826 LLVMValueRef es_main
= NULL
;
6827 LLVMValueRef gs_prolog
= NULL
;
6828 LLVMValueRef gs_main
= ctx
.main_fn
;
6831 union si_shader_part_key gs_prolog_key
;
6832 memset(&gs_prolog_key
, 0, sizeof(gs_prolog_key
));
6833 gs_prolog_key
.gs_prolog
.states
= shader
->key
.part
.gs
.prolog
;
6834 gs_prolog_key
.gs_prolog
.is_monolithic
= true;
6835 si_build_gs_prolog_function(&ctx
, &gs_prolog_key
);
6836 gs_prolog
= ctx
.main_fn
;
6839 struct si_shader shader_es
= {};
6840 shader_es
.selector
= es
;
6841 shader_es
.key
.as_es
= 1;
6842 shader_es
.key
.mono
= shader
->key
.mono
;
6843 shader_es
.key
.opt
= shader
->key
.opt
;
6844 si_llvm_context_set_tgsi(&ctx
, &shader_es
);
6846 if (!si_compile_tgsi_main(&ctx
, true)) {
6847 si_llvm_dispose(&ctx
);
6850 shader
->info
.uses_instanceid
|= es
->info
.uses_instanceid
;
6851 es_main
= ctx
.main_fn
;
6854 if (es
->vs_needs_prolog
) {
6855 union si_shader_part_key vs_prolog_key
;
6856 si_get_vs_prolog_key(&es
->info
,
6857 shader_es
.info
.num_input_sgprs
,
6858 &shader
->key
.part
.gs
.vs_prolog
,
6859 shader
, &vs_prolog_key
);
6860 vs_prolog_key
.vs_prolog
.is_monolithic
= true;
6861 si_build_vs_prolog_function(&ctx
, &vs_prolog_key
);
6862 es_prolog
= ctx
.main_fn
;
6865 /* Reset the shader context. */
6866 ctx
.shader
= shader
;
6867 ctx
.type
= PIPE_SHADER_GEOMETRY
;
6869 /* Prepare the array of shader parts. */
6870 LLVMValueRef parts
[4];
6871 unsigned num_parts
= 0, main_part
, next_first_part
;
6874 parts
[num_parts
++] = es_prolog
;
6876 parts
[main_part
= num_parts
++] = es_main
;
6877 parts
[next_first_part
= num_parts
++] = gs_prolog
;
6878 parts
[num_parts
++] = gs_main
;
6880 si_build_wrapper_function(&ctx
, parts
, num_parts
,
6881 main_part
, next_first_part
);
6883 LLVMValueRef parts
[2];
6884 union si_shader_part_key prolog_key
;
6886 parts
[1] = ctx
.main_fn
;
6888 memset(&prolog_key
, 0, sizeof(prolog_key
));
6889 prolog_key
.gs_prolog
.states
= shader
->key
.part
.gs
.prolog
;
6890 si_build_gs_prolog_function(&ctx
, &prolog_key
);
6891 parts
[0] = ctx
.main_fn
;
6893 si_build_wrapper_function(&ctx
, parts
, 2, 1, 0);
6895 } else if (is_monolithic
&& ctx
.type
== PIPE_SHADER_FRAGMENT
) {
6896 LLVMValueRef parts
[3];
6897 union si_shader_part_key prolog_key
;
6898 union si_shader_part_key epilog_key
;
6901 si_get_ps_prolog_key(shader
, &prolog_key
, false);
6902 need_prolog
= si_need_ps_prolog(&prolog_key
);
6904 parts
[need_prolog
? 1 : 0] = ctx
.main_fn
;
6907 si_build_ps_prolog_function(&ctx
, &prolog_key
);
6908 parts
[0] = ctx
.main_fn
;
6911 si_get_ps_epilog_key(shader
, &epilog_key
);
6912 si_build_ps_epilog_function(&ctx
, &epilog_key
);
6913 parts
[need_prolog
? 2 : 1] = ctx
.main_fn
;
6915 si_build_wrapper_function(&ctx
, parts
, need_prolog
? 3 : 2,
6916 need_prolog
? 1 : 0, 0);
6919 si_llvm_optimize_module(&ctx
);
6921 /* Post-optimization transformations and analysis. */
6922 si_optimize_vs_outputs(&ctx
);
6924 if ((debug
&& debug
->debug_message
) ||
6925 si_can_dump_shader(sscreen
, ctx
.type
))
6926 si_count_scratch_private_memory(&ctx
);
6928 /* Compile to bytecode. */
6929 r
= si_compile_llvm(sscreen
, &shader
->binary
, &shader
->config
, tm
,
6930 ctx
.gallivm
.module
, debug
, ctx
.type
, "TGSI shader");
6931 si_llvm_dispose(&ctx
);
6933 fprintf(stderr
, "LLVM failed to compile shader\n");
6937 /* Validate SGPR and VGPR usage for compute to detect compiler bugs.
6938 * LLVM 3.9svn has this bug.
6940 if (sel
->type
== PIPE_SHADER_COMPUTE
) {
6941 unsigned wave_size
= 64;
6942 unsigned max_vgprs
= 256;
6943 unsigned max_sgprs
= sscreen
->info
.chip_class
>= VI
? 800 : 512;
6944 unsigned max_sgprs_per_wave
= 128;
6945 unsigned max_block_threads
= si_get_max_workgroup_size(shader
);
6946 unsigned min_waves_per_cu
= DIV_ROUND_UP(max_block_threads
, wave_size
);
6947 unsigned min_waves_per_simd
= DIV_ROUND_UP(min_waves_per_cu
, 4);
6949 max_vgprs
= max_vgprs
/ min_waves_per_simd
;
6950 max_sgprs
= MIN2(max_sgprs
/ min_waves_per_simd
, max_sgprs_per_wave
);
6952 if (shader
->config
.num_sgprs
> max_sgprs
||
6953 shader
->config
.num_vgprs
> max_vgprs
) {
6954 fprintf(stderr
, "LLVM failed to compile a shader correctly: "
6955 "SGPR:VGPR usage is %u:%u, but the hw limit is %u:%u\n",
6956 shader
->config
.num_sgprs
, shader
->config
.num_vgprs
,
6957 max_sgprs
, max_vgprs
);
6959 /* Just terminate the process, because dependent
6960 * shaders can hang due to bad input data, but use
6961 * the env var to allow shader-db to work.
6963 if (!debug_get_bool_option("SI_PASS_BAD_SHADERS", false))
6968 /* Add the scratch offset to input SGPRs. */
6969 if (shader
->config
.scratch_bytes_per_wave
&& !is_merged_shader(shader
))
6970 shader
->info
.num_input_sgprs
+= 1; /* scratch byte offset */
6972 /* Calculate the number of fragment input VGPRs. */
6973 if (ctx
.type
== PIPE_SHADER_FRAGMENT
) {
6974 shader
->info
.num_input_vgprs
= 0;
6975 shader
->info
.face_vgpr_index
= -1;
6976 shader
->info
.ancillary_vgpr_index
= -1;
6978 if (G_0286CC_PERSP_SAMPLE_ENA(shader
->config
.spi_ps_input_addr
))
6979 shader
->info
.num_input_vgprs
+= 2;
6980 if (G_0286CC_PERSP_CENTER_ENA(shader
->config
.spi_ps_input_addr
))
6981 shader
->info
.num_input_vgprs
+= 2;
6982 if (G_0286CC_PERSP_CENTROID_ENA(shader
->config
.spi_ps_input_addr
))
6983 shader
->info
.num_input_vgprs
+= 2;
6984 if (G_0286CC_PERSP_PULL_MODEL_ENA(shader
->config
.spi_ps_input_addr
))
6985 shader
->info
.num_input_vgprs
+= 3;
6986 if (G_0286CC_LINEAR_SAMPLE_ENA(shader
->config
.spi_ps_input_addr
))
6987 shader
->info
.num_input_vgprs
+= 2;
6988 if (G_0286CC_LINEAR_CENTER_ENA(shader
->config
.spi_ps_input_addr
))
6989 shader
->info
.num_input_vgprs
+= 2;
6990 if (G_0286CC_LINEAR_CENTROID_ENA(shader
->config
.spi_ps_input_addr
))
6991 shader
->info
.num_input_vgprs
+= 2;
6992 if (G_0286CC_LINE_STIPPLE_TEX_ENA(shader
->config
.spi_ps_input_addr
))
6993 shader
->info
.num_input_vgprs
+= 1;
6994 if (G_0286CC_POS_X_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
6995 shader
->info
.num_input_vgprs
+= 1;
6996 if (G_0286CC_POS_Y_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
6997 shader
->info
.num_input_vgprs
+= 1;
6998 if (G_0286CC_POS_Z_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
6999 shader
->info
.num_input_vgprs
+= 1;
7000 if (G_0286CC_POS_W_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
7001 shader
->info
.num_input_vgprs
+= 1;
7002 if (G_0286CC_FRONT_FACE_ENA(shader
->config
.spi_ps_input_addr
)) {
7003 shader
->info
.face_vgpr_index
= shader
->info
.num_input_vgprs
;
7004 shader
->info
.num_input_vgprs
+= 1;
7006 if (G_0286CC_ANCILLARY_ENA(shader
->config
.spi_ps_input_addr
)) {
7007 shader
->info
.ancillary_vgpr_index
= shader
->info
.num_input_vgprs
;
7008 shader
->info
.num_input_vgprs
+= 1;
7010 if (G_0286CC_SAMPLE_COVERAGE_ENA(shader
->config
.spi_ps_input_addr
))
7011 shader
->info
.num_input_vgprs
+= 1;
7012 if (G_0286CC_POS_FIXED_PT_ENA(shader
->config
.spi_ps_input_addr
))
7013 shader
->info
.num_input_vgprs
+= 1;
7016 si_calculate_max_simd_waves(shader
);
7017 si_shader_dump_stats_for_shader_db(shader
, debug
);
7022 * Create, compile and return a shader part (prolog or epilog).
7024 * \param sscreen screen
7025 * \param list list of shader parts of the same category
7026 * \param type shader type
7027 * \param key shader part key
7028 * \param prolog whether the part being requested is a prolog
7029 * \param tm LLVM target machine
7030 * \param debug debug callback
7031 * \param build the callback responsible for building the main function
7032 * \return non-NULL on success
7034 static struct si_shader_part
*
7035 si_get_shader_part(struct si_screen
*sscreen
,
7036 struct si_shader_part
**list
,
7037 enum pipe_shader_type type
,
7039 union si_shader_part_key
*key
,
7040 LLVMTargetMachineRef tm
,
7041 struct pipe_debug_callback
*debug
,
7042 void (*build
)(struct si_shader_context
*,
7043 union si_shader_part_key
*),
7046 struct si_shader_part
*result
;
7048 mtx_lock(&sscreen
->shader_parts_mutex
);
7050 /* Find existing. */
7051 for (result
= *list
; result
; result
= result
->next
) {
7052 if (memcmp(&result
->key
, key
, sizeof(*key
)) == 0) {
7053 mtx_unlock(&sscreen
->shader_parts_mutex
);
7058 /* Compile a new one. */
7059 result
= CALLOC_STRUCT(si_shader_part
);
7062 struct si_shader shader
= {};
7063 struct si_shader_context ctx
;
7065 si_init_shader_ctx(&ctx
, sscreen
, tm
);
7066 ctx
.shader
= &shader
;
7070 case PIPE_SHADER_VERTEX
:
7071 shader
.key
.as_ls
= key
->vs_prolog
.as_ls
;
7072 shader
.key
.as_es
= key
->vs_prolog
.as_es
;
7074 case PIPE_SHADER_TESS_CTRL
:
7076 shader
.key
.part
.tcs
.epilog
= key
->tcs_epilog
.states
;
7078 case PIPE_SHADER_GEOMETRY
:
7081 case PIPE_SHADER_FRAGMENT
:
7083 shader
.key
.part
.ps
.prolog
= key
->ps_prolog
.states
;
7085 shader
.key
.part
.ps
.epilog
= key
->ps_epilog
.states
;
7088 unreachable("bad shader part");
7094 si_llvm_optimize_module(&ctx
);
7096 if (si_compile_llvm(sscreen
, &result
->binary
, &result
->config
, tm
,
7097 ctx
.ac
.module
, debug
, ctx
.type
, name
)) {
7103 result
->next
= *list
;
7107 si_llvm_dispose(&ctx
);
7108 mtx_unlock(&sscreen
->shader_parts_mutex
);
7112 static LLVMValueRef
si_prolog_get_rw_buffers(struct si_shader_context
*ctx
)
7114 LLVMValueRef ptr
[2], list
;
7115 bool is_merged_shader
=
7116 ctx
->screen
->info
.chip_class
>= GFX9
&&
7117 (ctx
->type
== PIPE_SHADER_TESS_CTRL
||
7118 ctx
->type
== PIPE_SHADER_GEOMETRY
||
7119 ctx
->shader
->key
.as_ls
|| ctx
->shader
->key
.as_es
);
7121 if (HAVE_32BIT_POINTERS
) {
7122 ptr
[0] = LLVMGetParam(ctx
->main_fn
, (is_merged_shader
? 8 : 0) + SI_SGPR_RW_BUFFERS
);
7123 list
= LLVMBuildIntToPtr(ctx
->ac
.builder
, ptr
[0],
7124 ac_array_in_const32_addr_space(ctx
->v4i32
), "");
7128 /* Get the pointer to rw buffers. */
7129 ptr
[0] = LLVMGetParam(ctx
->main_fn
, (is_merged_shader
? 8 : 0) + SI_SGPR_RW_BUFFERS
);
7130 ptr
[1] = LLVMGetParam(ctx
->main_fn
, (is_merged_shader
? 8 : 0) + SI_SGPR_RW_BUFFERS
+ 1);
7131 list
= lp_build_gather_values(&ctx
->gallivm
, ptr
, 2);
7132 list
= LLVMBuildBitCast(ctx
->ac
.builder
, list
, ctx
->i64
, "");
7133 list
= LLVMBuildIntToPtr(ctx
->ac
.builder
, list
,
7134 ac_array_in_const_addr_space(ctx
->v4i32
), "");
7139 * Build the vertex shader prolog function.
7141 * The inputs are the same as VS (a lot of SGPRs and 4 VGPR system values).
7142 * All inputs are returned unmodified. The vertex load indices are
7143 * stored after them, which will be used by the API VS for fetching inputs.
7145 * For example, the expected outputs for instance_divisors[] = {0, 1, 2} are:
7150 * (VertexID + BaseVertex),
7151 * (InstanceID + StartInstance),
7152 * (InstanceID / 2 + StartInstance)
7154 static void si_build_vs_prolog_function(struct si_shader_context
*ctx
,
7155 union si_shader_part_key
*key
)
7157 struct si_function_info fninfo
;
7158 LLVMTypeRef
*returns
;
7159 LLVMValueRef ret
, func
;
7161 unsigned first_vs_vgpr
= key
->vs_prolog
.num_merged_next_stage_vgprs
;
7162 unsigned num_input_vgprs
= key
->vs_prolog
.num_merged_next_stage_vgprs
+ 4;
7163 LLVMValueRef input_vgprs
[9];
7164 unsigned num_all_input_regs
= key
->vs_prolog
.num_input_sgprs
+
7166 unsigned user_sgpr_base
= key
->vs_prolog
.num_merged_next_stage_vgprs
? 8 : 0;
7168 si_init_function_info(&fninfo
);
7170 /* 4 preloaded VGPRs + vertex load indices as prolog outputs */
7171 returns
= alloca((num_all_input_regs
+ key
->vs_prolog
.last_input
+ 1) *
7172 sizeof(LLVMTypeRef
));
7175 /* Declare input and output SGPRs. */
7176 for (i
= 0; i
< key
->vs_prolog
.num_input_sgprs
; i
++) {
7177 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7178 returns
[num_returns
++] = ctx
->i32
;
7181 /* Preloaded VGPRs (outputs must be floats) */
7182 for (i
= 0; i
< num_input_vgprs
; i
++) {
7183 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &input_vgprs
[i
]);
7184 returns
[num_returns
++] = ctx
->f32
;
7187 /* Vertex load indices. */
7188 for (i
= 0; i
<= key
->vs_prolog
.last_input
; i
++)
7189 returns
[num_returns
++] = ctx
->f32
;
7191 /* Create the function. */
7192 si_create_function(ctx
, "vs_prolog", returns
, num_returns
, &fninfo
, 0);
7193 func
= ctx
->main_fn
;
7195 if (key
->vs_prolog
.num_merged_next_stage_vgprs
) {
7196 if (!key
->vs_prolog
.is_monolithic
)
7197 si_init_exec_from_input(ctx
, 3, 0);
7199 if (key
->vs_prolog
.as_ls
&&
7200 ctx
->screen
->has_ls_vgpr_init_bug
) {
7201 /* If there are no HS threads, SPI loads the LS VGPRs
7202 * starting at VGPR 0. Shift them back to where they
7205 LLVMValueRef has_hs_threads
=
7206 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntNE
,
7207 unpack_param(ctx
, 3, 8, 8),
7210 for (i
= 4; i
> 0; --i
) {
7211 input_vgprs
[i
+ 1] =
7212 LLVMBuildSelect(ctx
->ac
.builder
, has_hs_threads
,
7214 input_vgprs
[i
- 1], "");
7219 ctx
->abi
.vertex_id
= input_vgprs
[first_vs_vgpr
];
7220 ctx
->abi
.instance_id
= input_vgprs
[first_vs_vgpr
+ (key
->vs_prolog
.as_ls
? 2 : 1)];
7222 /* Copy inputs to outputs. This should be no-op, as the registers match,
7223 * but it will prevent the compiler from overwriting them unintentionally.
7225 ret
= ctx
->return_value
;
7226 for (i
= 0; i
< key
->vs_prolog
.num_input_sgprs
; i
++) {
7227 LLVMValueRef p
= LLVMGetParam(func
, i
);
7228 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, p
, i
, "");
7230 for (i
= 0; i
< num_input_vgprs
; i
++) {
7231 LLVMValueRef p
= input_vgprs
[i
];
7232 p
= ac_to_float(&ctx
->ac
, p
);
7233 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, p
,
7234 key
->vs_prolog
.num_input_sgprs
+ i
, "");
7237 /* Compute vertex load indices from instance divisors. */
7238 LLVMValueRef instance_divisor_constbuf
= NULL
;
7240 if (key
->vs_prolog
.states
.instance_divisor_is_fetched
) {
7241 LLVMValueRef list
= si_prolog_get_rw_buffers(ctx
);
7242 LLVMValueRef buf_index
=
7243 LLVMConstInt(ctx
->i32
, SI_VS_CONST_INSTANCE_DIVISORS
, 0);
7244 instance_divisor_constbuf
=
7245 ac_build_load_to_sgpr(&ctx
->ac
, list
, buf_index
);
7248 for (i
= 0; i
<= key
->vs_prolog
.last_input
; i
++) {
7249 bool divisor_is_one
=
7250 key
->vs_prolog
.states
.instance_divisor_is_one
& (1u << i
);
7251 bool divisor_is_fetched
=
7252 key
->vs_prolog
.states
.instance_divisor_is_fetched
& (1u << i
);
7255 if (divisor_is_one
|| divisor_is_fetched
) {
7256 LLVMValueRef divisor
= ctx
->i32_1
;
7258 if (divisor_is_fetched
) {
7259 divisor
= buffer_load_const(ctx
, instance_divisor_constbuf
,
7260 LLVMConstInt(ctx
->i32
, i
* 4, 0));
7261 divisor
= ac_to_integer(&ctx
->ac
, divisor
);
7264 /* InstanceID / Divisor + StartInstance */
7265 index
= get_instance_index_for_fetch(ctx
,
7267 SI_SGPR_START_INSTANCE
,
7270 /* VertexID + BaseVertex */
7271 index
= LLVMBuildAdd(ctx
->ac
.builder
,
7273 LLVMGetParam(func
, user_sgpr_base
+
7274 SI_SGPR_BASE_VERTEX
), "");
7277 index
= ac_to_float(&ctx
->ac
, index
);
7278 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, index
,
7279 fninfo
.num_params
+ i
, "");
7282 si_llvm_build_ret(ctx
, ret
);
7285 static bool si_get_vs_prolog(struct si_screen
*sscreen
,
7286 LLVMTargetMachineRef tm
,
7287 struct si_shader
*shader
,
7288 struct pipe_debug_callback
*debug
,
7289 struct si_shader
*main_part
,
7290 const struct si_vs_prolog_bits
*key
)
7292 struct si_shader_selector
*vs
= main_part
->selector
;
7294 if (!si_vs_needs_prolog(vs
, key
))
7297 /* Get the prolog. */
7298 union si_shader_part_key prolog_key
;
7299 si_get_vs_prolog_key(&vs
->info
, main_part
->info
.num_input_sgprs
,
7300 key
, shader
, &prolog_key
);
7303 si_get_shader_part(sscreen
, &sscreen
->vs_prologs
,
7304 PIPE_SHADER_VERTEX
, true, &prolog_key
, tm
,
7305 debug
, si_build_vs_prolog_function
,
7306 "Vertex Shader Prolog");
7307 return shader
->prolog
!= NULL
;
7311 * Select and compile (or reuse) vertex shader parts (prolog & epilog).
7313 static bool si_shader_select_vs_parts(struct si_screen
*sscreen
,
7314 LLVMTargetMachineRef tm
,
7315 struct si_shader
*shader
,
7316 struct pipe_debug_callback
*debug
)
7318 return si_get_vs_prolog(sscreen
, tm
, shader
, debug
, shader
,
7319 &shader
->key
.part
.vs
.prolog
);
7323 * Compile the TCS epilog function. This writes tesselation factors to memory
7324 * based on the output primitive type of the tesselator (determined by TES).
7326 static void si_build_tcs_epilog_function(struct si_shader_context
*ctx
,
7327 union si_shader_part_key
*key
)
7329 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
7330 struct si_function_info fninfo
;
7333 si_init_function_info(&fninfo
);
7335 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
7336 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7337 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7338 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7339 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* wave info */
7340 ctx
->param_tcs_factor_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7341 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7342 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7343 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7344 add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7345 add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7346 add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7347 add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7348 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7349 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7350 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7351 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7352 if (!HAVE_32BIT_POINTERS
)
7353 add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7354 ctx
->param_tcs_offchip_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7355 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7356 ctx
->param_tcs_out_lds_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7358 add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7359 add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7360 add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7361 add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7362 ctx
->param_tcs_offchip_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7363 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7364 ctx
->param_tcs_out_lds_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7365 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7366 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7367 ctx
->param_tcs_factor_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7370 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
); /* VGPR gap */
7371 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
); /* VGPR gap */
7372 unsigned tess_factors_idx
=
7373 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
); /* patch index within the wave (REL_PATCH_ID) */
7374 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
); /* invocation ID within the patch */
7375 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
); /* LDS offset where tess factors should be loaded from */
7377 for (unsigned i
= 0; i
< 6; i
++)
7378 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
); /* tess factors */
7380 /* Create the function. */
7381 si_create_function(ctx
, "tcs_epilog", NULL
, 0, &fninfo
,
7382 ctx
->screen
->info
.chip_class
>= CIK
? 128 : 64);
7383 ac_declare_lds_as_pointer(&ctx
->ac
);
7384 func
= ctx
->main_fn
;
7386 LLVMValueRef invoc0_tess_factors
[6];
7387 for (unsigned i
= 0; i
< 6; i
++)
7388 invoc0_tess_factors
[i
] = LLVMGetParam(func
, tess_factors_idx
+ 3 + i
);
7390 si_write_tess_factors(bld_base
,
7391 LLVMGetParam(func
, tess_factors_idx
),
7392 LLVMGetParam(func
, tess_factors_idx
+ 1),
7393 LLVMGetParam(func
, tess_factors_idx
+ 2),
7394 invoc0_tess_factors
, invoc0_tess_factors
+ 4);
7396 LLVMBuildRetVoid(ctx
->ac
.builder
);
7400 * Select and compile (or reuse) TCS parts (epilog).
7402 static bool si_shader_select_tcs_parts(struct si_screen
*sscreen
,
7403 LLVMTargetMachineRef tm
,
7404 struct si_shader
*shader
,
7405 struct pipe_debug_callback
*debug
)
7407 if (sscreen
->info
.chip_class
>= GFX9
) {
7408 struct si_shader
*ls_main_part
=
7409 shader
->key
.part
.tcs
.ls
->main_shader_part_ls
;
7411 if (!si_get_vs_prolog(sscreen
, tm
, shader
, debug
, ls_main_part
,
7412 &shader
->key
.part
.tcs
.ls_prolog
))
7415 shader
->previous_stage
= ls_main_part
;
7418 /* Get the epilog. */
7419 union si_shader_part_key epilog_key
;
7420 memset(&epilog_key
, 0, sizeof(epilog_key
));
7421 epilog_key
.tcs_epilog
.states
= shader
->key
.part
.tcs
.epilog
;
7423 shader
->epilog
= si_get_shader_part(sscreen
, &sscreen
->tcs_epilogs
,
7424 PIPE_SHADER_TESS_CTRL
, false,
7425 &epilog_key
, tm
, debug
,
7426 si_build_tcs_epilog_function
,
7427 "Tessellation Control Shader Epilog");
7428 return shader
->epilog
!= NULL
;
7432 * Select and compile (or reuse) GS parts (prolog).
7434 static bool si_shader_select_gs_parts(struct si_screen
*sscreen
,
7435 LLVMTargetMachineRef tm
,
7436 struct si_shader
*shader
,
7437 struct pipe_debug_callback
*debug
)
7439 if (sscreen
->info
.chip_class
>= GFX9
) {
7440 struct si_shader
*es_main_part
=
7441 shader
->key
.part
.gs
.es
->main_shader_part_es
;
7443 if (shader
->key
.part
.gs
.es
->type
== PIPE_SHADER_VERTEX
&&
7444 !si_get_vs_prolog(sscreen
, tm
, shader
, debug
, es_main_part
,
7445 &shader
->key
.part
.gs
.vs_prolog
))
7448 shader
->previous_stage
= es_main_part
;
7451 if (!shader
->key
.part
.gs
.prolog
.tri_strip_adj_fix
)
7454 union si_shader_part_key prolog_key
;
7455 memset(&prolog_key
, 0, sizeof(prolog_key
));
7456 prolog_key
.gs_prolog
.states
= shader
->key
.part
.gs
.prolog
;
7458 shader
->prolog2
= si_get_shader_part(sscreen
, &sscreen
->gs_prologs
,
7459 PIPE_SHADER_GEOMETRY
, true,
7460 &prolog_key
, tm
, debug
,
7461 si_build_gs_prolog_function
,
7462 "Geometry Shader Prolog");
7463 return shader
->prolog2
!= NULL
;
7467 * Build the pixel shader prolog function. This handles:
7468 * - two-side color selection and interpolation
7469 * - overriding interpolation parameters for the API PS
7470 * - polygon stippling
7472 * All preloaded SGPRs and VGPRs are passed through unmodified unless they are
7473 * overriden by other states. (e.g. per-sample interpolation)
7474 * Interpolated colors are stored after the preloaded VGPRs.
7476 static void si_build_ps_prolog_function(struct si_shader_context
*ctx
,
7477 union si_shader_part_key
*key
)
7479 struct si_function_info fninfo
;
7480 LLVMValueRef ret
, func
;
7481 int num_returns
, i
, num_color_channels
;
7483 assert(si_need_ps_prolog(key
));
7485 si_init_function_info(&fninfo
);
7487 /* Declare inputs. */
7488 for (i
= 0; i
< key
->ps_prolog
.num_input_sgprs
; i
++)
7489 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7491 for (i
= 0; i
< key
->ps_prolog
.num_input_vgprs
; i
++)
7492 add_arg(&fninfo
, ARG_VGPR
, ctx
->f32
);
7494 /* Declare outputs (same as inputs + add colors if needed) */
7495 num_returns
= fninfo
.num_params
;
7496 num_color_channels
= util_bitcount(key
->ps_prolog
.colors_read
);
7497 for (i
= 0; i
< num_color_channels
; i
++)
7498 fninfo
.types
[num_returns
++] = ctx
->f32
;
7500 /* Create the function. */
7501 si_create_function(ctx
, "ps_prolog", fninfo
.types
, num_returns
,
7503 func
= ctx
->main_fn
;
7505 /* Copy inputs to outputs. This should be no-op, as the registers match,
7506 * but it will prevent the compiler from overwriting them unintentionally.
7508 ret
= ctx
->return_value
;
7509 for (i
= 0; i
< fninfo
.num_params
; i
++) {
7510 LLVMValueRef p
= LLVMGetParam(func
, i
);
7511 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, p
, i
, "");
7514 /* Polygon stippling. */
7515 if (key
->ps_prolog
.states
.poly_stipple
) {
7516 /* POS_FIXED_PT is always last. */
7517 unsigned pos
= key
->ps_prolog
.num_input_sgprs
+
7518 key
->ps_prolog
.num_input_vgprs
- 1;
7519 LLVMValueRef list
= si_prolog_get_rw_buffers(ctx
);
7521 si_llvm_emit_polygon_stipple(ctx
, list
, pos
);
7524 if (key
->ps_prolog
.states
.bc_optimize_for_persp
||
7525 key
->ps_prolog
.states
.bc_optimize_for_linear
) {
7526 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7527 LLVMValueRef center
[2], centroid
[2], tmp
, bc_optimize
;
7529 /* The shader should do: if (PRIM_MASK[31]) CENTROID = CENTER;
7530 * The hw doesn't compute CENTROID if the whole wave only
7531 * contains fully-covered quads.
7533 * PRIM_MASK is after user SGPRs.
7535 bc_optimize
= LLVMGetParam(func
, SI_PS_NUM_USER_SGPR
);
7536 bc_optimize
= LLVMBuildLShr(ctx
->ac
.builder
, bc_optimize
,
7537 LLVMConstInt(ctx
->i32
, 31, 0), "");
7538 bc_optimize
= LLVMBuildTrunc(ctx
->ac
.builder
, bc_optimize
,
7541 if (key
->ps_prolog
.states
.bc_optimize_for_persp
) {
7542 /* Read PERSP_CENTER. */
7543 for (i
= 0; i
< 2; i
++)
7544 center
[i
] = LLVMGetParam(func
, base
+ 2 + i
);
7545 /* Read PERSP_CENTROID. */
7546 for (i
= 0; i
< 2; i
++)
7547 centroid
[i
] = LLVMGetParam(func
, base
+ 4 + i
);
7548 /* Select PERSP_CENTROID. */
7549 for (i
= 0; i
< 2; i
++) {
7550 tmp
= LLVMBuildSelect(ctx
->ac
.builder
, bc_optimize
,
7551 center
[i
], centroid
[i
], "");
7552 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7553 tmp
, base
+ 4 + i
, "");
7556 if (key
->ps_prolog
.states
.bc_optimize_for_linear
) {
7557 /* Read LINEAR_CENTER. */
7558 for (i
= 0; i
< 2; i
++)
7559 center
[i
] = LLVMGetParam(func
, base
+ 8 + i
);
7560 /* Read LINEAR_CENTROID. */
7561 for (i
= 0; i
< 2; i
++)
7562 centroid
[i
] = LLVMGetParam(func
, base
+ 10 + i
);
7563 /* Select LINEAR_CENTROID. */
7564 for (i
= 0; i
< 2; i
++) {
7565 tmp
= LLVMBuildSelect(ctx
->ac
.builder
, bc_optimize
,
7566 center
[i
], centroid
[i
], "");
7567 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7568 tmp
, base
+ 10 + i
, "");
7573 /* Force per-sample interpolation. */
7574 if (key
->ps_prolog
.states
.force_persp_sample_interp
) {
7575 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7576 LLVMValueRef persp_sample
[2];
7578 /* Read PERSP_SAMPLE. */
7579 for (i
= 0; i
< 2; i
++)
7580 persp_sample
[i
] = LLVMGetParam(func
, base
+ i
);
7581 /* Overwrite PERSP_CENTER. */
7582 for (i
= 0; i
< 2; i
++)
7583 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7584 persp_sample
[i
], base
+ 2 + i
, "");
7585 /* Overwrite PERSP_CENTROID. */
7586 for (i
= 0; i
< 2; i
++)
7587 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7588 persp_sample
[i
], base
+ 4 + i
, "");
7590 if (key
->ps_prolog
.states
.force_linear_sample_interp
) {
7591 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7592 LLVMValueRef linear_sample
[2];
7594 /* Read LINEAR_SAMPLE. */
7595 for (i
= 0; i
< 2; i
++)
7596 linear_sample
[i
] = LLVMGetParam(func
, base
+ 6 + i
);
7597 /* Overwrite LINEAR_CENTER. */
7598 for (i
= 0; i
< 2; i
++)
7599 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7600 linear_sample
[i
], base
+ 8 + i
, "");
7601 /* Overwrite LINEAR_CENTROID. */
7602 for (i
= 0; i
< 2; i
++)
7603 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7604 linear_sample
[i
], base
+ 10 + i
, "");
7607 /* Force center interpolation. */
7608 if (key
->ps_prolog
.states
.force_persp_center_interp
) {
7609 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7610 LLVMValueRef persp_center
[2];
7612 /* Read PERSP_CENTER. */
7613 for (i
= 0; i
< 2; i
++)
7614 persp_center
[i
] = LLVMGetParam(func
, base
+ 2 + i
);
7615 /* Overwrite PERSP_SAMPLE. */
7616 for (i
= 0; i
< 2; i
++)
7617 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7618 persp_center
[i
], base
+ i
, "");
7619 /* Overwrite PERSP_CENTROID. */
7620 for (i
= 0; i
< 2; i
++)
7621 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7622 persp_center
[i
], base
+ 4 + i
, "");
7624 if (key
->ps_prolog
.states
.force_linear_center_interp
) {
7625 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7626 LLVMValueRef linear_center
[2];
7628 /* Read LINEAR_CENTER. */
7629 for (i
= 0; i
< 2; i
++)
7630 linear_center
[i
] = LLVMGetParam(func
, base
+ 8 + i
);
7631 /* Overwrite LINEAR_SAMPLE. */
7632 for (i
= 0; i
< 2; i
++)
7633 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7634 linear_center
[i
], base
+ 6 + i
, "");
7635 /* Overwrite LINEAR_CENTROID. */
7636 for (i
= 0; i
< 2; i
++)
7637 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7638 linear_center
[i
], base
+ 10 + i
, "");
7641 /* Interpolate colors. */
7642 unsigned color_out_idx
= 0;
7643 for (i
= 0; i
< 2; i
++) {
7644 unsigned writemask
= (key
->ps_prolog
.colors_read
>> (i
* 4)) & 0xf;
7645 unsigned face_vgpr
= key
->ps_prolog
.num_input_sgprs
+
7646 key
->ps_prolog
.face_vgpr_index
;
7647 LLVMValueRef interp
[2], color
[4];
7648 LLVMValueRef interp_ij
= NULL
, prim_mask
= NULL
, face
= NULL
;
7653 /* If the interpolation qualifier is not CONSTANT (-1). */
7654 if (key
->ps_prolog
.color_interp_vgpr_index
[i
] != -1) {
7655 unsigned interp_vgpr
= key
->ps_prolog
.num_input_sgprs
+
7656 key
->ps_prolog
.color_interp_vgpr_index
[i
];
7658 /* Get the (i,j) updated by bc_optimize handling. */
7659 interp
[0] = LLVMBuildExtractValue(ctx
->ac
.builder
, ret
,
7661 interp
[1] = LLVMBuildExtractValue(ctx
->ac
.builder
, ret
,
7662 interp_vgpr
+ 1, "");
7663 interp_ij
= lp_build_gather_values(&ctx
->gallivm
, interp
, 2);
7666 /* Use the absolute location of the input. */
7667 prim_mask
= LLVMGetParam(func
, SI_PS_NUM_USER_SGPR
);
7669 if (key
->ps_prolog
.states
.color_two_side
) {
7670 face
= LLVMGetParam(func
, face_vgpr
);
7671 face
= ac_to_integer(&ctx
->ac
, face
);
7674 interp_fs_input(ctx
,
7675 key
->ps_prolog
.color_attr_index
[i
],
7676 TGSI_SEMANTIC_COLOR
, i
,
7677 key
->ps_prolog
.num_interp_inputs
,
7678 key
->ps_prolog
.colors_read
, interp_ij
,
7679 prim_mask
, face
, color
);
7682 unsigned chan
= u_bit_scan(&writemask
);
7683 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, color
[chan
],
7684 fninfo
.num_params
+ color_out_idx
++, "");
7688 /* Section 15.2.2 (Shader Inputs) of the OpenGL 4.5 (Core Profile) spec
7691 * "When per-sample shading is active due to the use of a fragment
7692 * input qualified by sample or due to the use of the gl_SampleID
7693 * or gl_SamplePosition variables, only the bit for the current
7694 * sample is set in gl_SampleMaskIn. When state specifies multiple
7695 * fragment shader invocations for a given fragment, the sample
7696 * mask for any single fragment shader invocation may specify a
7697 * subset of the covered samples for the fragment. In this case,
7698 * the bit corresponding to each covered sample will be set in
7699 * exactly one fragment shader invocation."
7701 * The samplemask loaded by hardware is always the coverage of the
7702 * entire pixel/fragment, so mask bits out based on the sample ID.
7704 if (key
->ps_prolog
.states
.samplemask_log_ps_iter
) {
7705 /* The bit pattern matches that used by fixed function fragment
7707 static const uint16_t ps_iter_masks
[] = {
7708 0xffff, /* not used */
7714 assert(key
->ps_prolog
.states
.samplemask_log_ps_iter
< ARRAY_SIZE(ps_iter_masks
));
7716 uint32_t ps_iter_mask
= ps_iter_masks
[key
->ps_prolog
.states
.samplemask_log_ps_iter
];
7717 unsigned ancillary_vgpr
= key
->ps_prolog
.num_input_sgprs
+
7718 key
->ps_prolog
.ancillary_vgpr_index
;
7719 LLVMValueRef sampleid
= unpack_param(ctx
, ancillary_vgpr
, 8, 4);
7720 LLVMValueRef samplemask
= LLVMGetParam(func
, ancillary_vgpr
+ 1);
7722 samplemask
= ac_to_integer(&ctx
->ac
, samplemask
);
7723 samplemask
= LLVMBuildAnd(
7726 LLVMBuildShl(ctx
->ac
.builder
,
7727 LLVMConstInt(ctx
->i32
, ps_iter_mask
, false),
7730 samplemask
= ac_to_float(&ctx
->ac
, samplemask
);
7732 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, samplemask
,
7733 ancillary_vgpr
+ 1, "");
7736 /* Tell LLVM to insert WQM instruction sequence when needed. */
7737 if (key
->ps_prolog
.wqm
) {
7738 LLVMAddTargetDependentFunctionAttr(func
,
7739 "amdgpu-ps-wqm-outputs", "");
7742 si_llvm_build_ret(ctx
, ret
);
7746 * Build the pixel shader epilog function. This handles everything that must be
7747 * emulated for pixel shader exports. (alpha-test, format conversions, etc)
7749 static void si_build_ps_epilog_function(struct si_shader_context
*ctx
,
7750 union si_shader_part_key
*key
)
7752 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
7753 struct si_function_info fninfo
;
7754 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
7756 struct si_ps_exports exp
= {};
7758 si_init_function_info(&fninfo
);
7760 /* Declare input SGPRs. */
7761 ctx
->param_rw_buffers
= add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7762 ctx
->param_bindless_samplers_and_images
= add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7763 ctx
->param_const_and_shader_buffers
= add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7764 ctx
->param_samplers_and_images
= add_arg(&fninfo
, ARG_SGPR
, ctx
->ac
.intptr
);
7765 add_arg_checked(&fninfo
, ARG_SGPR
, ctx
->f32
, SI_PARAM_ALPHA_REF
);
7767 /* Declare input VGPRs. */
7768 unsigned required_num_params
=
7769 fninfo
.num_sgpr_params
+
7770 util_bitcount(key
->ps_epilog
.colors_written
) * 4 +
7771 key
->ps_epilog
.writes_z
+
7772 key
->ps_epilog
.writes_stencil
+
7773 key
->ps_epilog
.writes_samplemask
;
7775 required_num_params
= MAX2(required_num_params
,
7776 fninfo
.num_sgpr_params
+ PS_EPILOG_SAMPLEMASK_MIN_LOC
+ 1);
7778 while (fninfo
.num_params
< required_num_params
)
7779 add_arg(&fninfo
, ARG_VGPR
, ctx
->f32
);
7781 /* Create the function. */
7782 si_create_function(ctx
, "ps_epilog", NULL
, 0, &fninfo
, 0);
7783 /* Disable elimination of unused inputs. */
7784 si_llvm_add_attribute(ctx
->main_fn
,
7785 "InitialPSInputAddr", 0xffffff);
7787 /* Process colors. */
7788 unsigned vgpr
= fninfo
.num_sgpr_params
;
7789 unsigned colors_written
= key
->ps_epilog
.colors_written
;
7790 int last_color_export
= -1;
7792 /* Find the last color export. */
7793 if (!key
->ps_epilog
.writes_z
&&
7794 !key
->ps_epilog
.writes_stencil
&&
7795 !key
->ps_epilog
.writes_samplemask
) {
7796 unsigned spi_format
= key
->ps_epilog
.states
.spi_shader_col_format
;
7798 /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
7799 if (colors_written
== 0x1 && key
->ps_epilog
.states
.last_cbuf
> 0) {
7800 /* Just set this if any of the colorbuffers are enabled. */
7802 ((1ull << (4 * (key
->ps_epilog
.states
.last_cbuf
+ 1))) - 1))
7803 last_color_export
= 0;
7805 for (i
= 0; i
< 8; i
++)
7806 if (colors_written
& (1 << i
) &&
7807 (spi_format
>> (i
* 4)) & 0xf)
7808 last_color_export
= i
;
7812 while (colors_written
) {
7813 LLVMValueRef color
[4];
7814 int mrt
= u_bit_scan(&colors_written
);
7816 for (i
= 0; i
< 4; i
++)
7817 color
[i
] = LLVMGetParam(ctx
->main_fn
, vgpr
++);
7819 si_export_mrt_color(bld_base
, color
, mrt
,
7820 fninfo
.num_params
- 1,
7821 mrt
== last_color_export
, &exp
);
7824 /* Process depth, stencil, samplemask. */
7825 if (key
->ps_epilog
.writes_z
)
7826 depth
= LLVMGetParam(ctx
->main_fn
, vgpr
++);
7827 if (key
->ps_epilog
.writes_stencil
)
7828 stencil
= LLVMGetParam(ctx
->main_fn
, vgpr
++);
7829 if (key
->ps_epilog
.writes_samplemask
)
7830 samplemask
= LLVMGetParam(ctx
->main_fn
, vgpr
++);
7832 if (depth
|| stencil
|| samplemask
)
7833 si_export_mrt_z(bld_base
, depth
, stencil
, samplemask
, &exp
);
7834 else if (last_color_export
== -1)
7835 ac_build_export_null(&ctx
->ac
);
7838 si_emit_ps_exports(ctx
, &exp
);
7841 LLVMBuildRetVoid(ctx
->ac
.builder
);
7845 * Select and compile (or reuse) pixel shader parts (prolog & epilog).
7847 static bool si_shader_select_ps_parts(struct si_screen
*sscreen
,
7848 LLVMTargetMachineRef tm
,
7849 struct si_shader
*shader
,
7850 struct pipe_debug_callback
*debug
)
7852 union si_shader_part_key prolog_key
;
7853 union si_shader_part_key epilog_key
;
7855 /* Get the prolog. */
7856 si_get_ps_prolog_key(shader
, &prolog_key
, true);
7858 /* The prolog is a no-op if these aren't set. */
7859 if (si_need_ps_prolog(&prolog_key
)) {
7861 si_get_shader_part(sscreen
, &sscreen
->ps_prologs
,
7862 PIPE_SHADER_FRAGMENT
, true,
7863 &prolog_key
, tm
, debug
,
7864 si_build_ps_prolog_function
,
7865 "Fragment Shader Prolog");
7866 if (!shader
->prolog
)
7870 /* Get the epilog. */
7871 si_get_ps_epilog_key(shader
, &epilog_key
);
7874 si_get_shader_part(sscreen
, &sscreen
->ps_epilogs
,
7875 PIPE_SHADER_FRAGMENT
, false,
7876 &epilog_key
, tm
, debug
,
7877 si_build_ps_epilog_function
,
7878 "Fragment Shader Epilog");
7879 if (!shader
->epilog
)
7882 /* Enable POS_FIXED_PT if polygon stippling is enabled. */
7883 if (shader
->key
.part
.ps
.prolog
.poly_stipple
) {
7884 shader
->config
.spi_ps_input_ena
|= S_0286CC_POS_FIXED_PT_ENA(1);
7885 assert(G_0286CC_POS_FIXED_PT_ENA(shader
->config
.spi_ps_input_addr
));
7888 /* Set up the enable bits for per-sample shading if needed. */
7889 if (shader
->key
.part
.ps
.prolog
.force_persp_sample_interp
&&
7890 (G_0286CC_PERSP_CENTER_ENA(shader
->config
.spi_ps_input_ena
) ||
7891 G_0286CC_PERSP_CENTROID_ENA(shader
->config
.spi_ps_input_ena
))) {
7892 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_CENTER_ENA
;
7893 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_CENTROID_ENA
;
7894 shader
->config
.spi_ps_input_ena
|= S_0286CC_PERSP_SAMPLE_ENA(1);
7896 if (shader
->key
.part
.ps
.prolog
.force_linear_sample_interp
&&
7897 (G_0286CC_LINEAR_CENTER_ENA(shader
->config
.spi_ps_input_ena
) ||
7898 G_0286CC_LINEAR_CENTROID_ENA(shader
->config
.spi_ps_input_ena
))) {
7899 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_CENTER_ENA
;
7900 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_CENTROID_ENA
;
7901 shader
->config
.spi_ps_input_ena
|= S_0286CC_LINEAR_SAMPLE_ENA(1);
7903 if (shader
->key
.part
.ps
.prolog
.force_persp_center_interp
&&
7904 (G_0286CC_PERSP_SAMPLE_ENA(shader
->config
.spi_ps_input_ena
) ||
7905 G_0286CC_PERSP_CENTROID_ENA(shader
->config
.spi_ps_input_ena
))) {
7906 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_SAMPLE_ENA
;
7907 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_CENTROID_ENA
;
7908 shader
->config
.spi_ps_input_ena
|= S_0286CC_PERSP_CENTER_ENA(1);
7910 if (shader
->key
.part
.ps
.prolog
.force_linear_center_interp
&&
7911 (G_0286CC_LINEAR_SAMPLE_ENA(shader
->config
.spi_ps_input_ena
) ||
7912 G_0286CC_LINEAR_CENTROID_ENA(shader
->config
.spi_ps_input_ena
))) {
7913 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_SAMPLE_ENA
;
7914 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_CENTROID_ENA
;
7915 shader
->config
.spi_ps_input_ena
|= S_0286CC_LINEAR_CENTER_ENA(1);
7918 /* POW_W_FLOAT requires that one of the perspective weights is enabled. */
7919 if (G_0286CC_POS_W_FLOAT_ENA(shader
->config
.spi_ps_input_ena
) &&
7920 !(shader
->config
.spi_ps_input_ena
& 0xf)) {
7921 shader
->config
.spi_ps_input_ena
|= S_0286CC_PERSP_CENTER_ENA(1);
7922 assert(G_0286CC_PERSP_CENTER_ENA(shader
->config
.spi_ps_input_addr
));
7925 /* At least one pair of interpolation weights must be enabled. */
7926 if (!(shader
->config
.spi_ps_input_ena
& 0x7f)) {
7927 shader
->config
.spi_ps_input_ena
|= S_0286CC_LINEAR_CENTER_ENA(1);
7928 assert(G_0286CC_LINEAR_CENTER_ENA(shader
->config
.spi_ps_input_addr
));
7931 /* Samplemask fixup requires the sample ID. */
7932 if (shader
->key
.part
.ps
.prolog
.samplemask_log_ps_iter
) {
7933 shader
->config
.spi_ps_input_ena
|= S_0286CC_ANCILLARY_ENA(1);
7934 assert(G_0286CC_ANCILLARY_ENA(shader
->config
.spi_ps_input_addr
));
7937 /* The sample mask input is always enabled, because the API shader always
7938 * passes it through to the epilog. Disable it here if it's unused.
7940 if (!shader
->key
.part
.ps
.epilog
.poly_line_smoothing
&&
7941 !shader
->selector
->info
.reads_samplemask
)
7942 shader
->config
.spi_ps_input_ena
&= C_0286CC_SAMPLE_COVERAGE_ENA
;
7947 void si_multiwave_lds_size_workaround(struct si_screen
*sscreen
,
7950 /* SPI barrier management bug:
7951 * Make sure we have at least 4k of LDS in use to avoid the bug.
7952 * It applies to workgroup sizes of more than one wavefront.
7954 if (sscreen
->info
.family
== CHIP_BONAIRE
||
7955 sscreen
->info
.family
== CHIP_KABINI
||
7956 sscreen
->info
.family
== CHIP_MULLINS
)
7957 *lds_size
= MAX2(*lds_size
, 8);
7960 static void si_fix_resource_usage(struct si_screen
*sscreen
,
7961 struct si_shader
*shader
)
7963 unsigned min_sgprs
= shader
->info
.num_input_sgprs
+ 2; /* VCC */
7965 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
, min_sgprs
);
7967 if (shader
->selector
->type
== PIPE_SHADER_COMPUTE
&&
7968 si_get_max_workgroup_size(shader
) > 64) {
7969 si_multiwave_lds_size_workaround(sscreen
,
7970 &shader
->config
.lds_size
);
7974 int si_shader_create(struct si_screen
*sscreen
, LLVMTargetMachineRef tm
,
7975 struct si_shader
*shader
,
7976 struct pipe_debug_callback
*debug
)
7978 struct si_shader_selector
*sel
= shader
->selector
;
7979 struct si_shader
*mainp
= *si_get_main_shader_part(sel
, &shader
->key
);
7982 /* LS, ES, VS are compiled on demand if the main part hasn't been
7983 * compiled for that stage.
7985 * Vertex shaders are compiled on demand when a vertex fetch
7986 * workaround must be applied.
7988 if (shader
->is_monolithic
) {
7989 /* Monolithic shader (compiled as a whole, has many variants,
7990 * may take a long time to compile).
7992 r
= si_compile_tgsi_shader(sscreen
, tm
, shader
, true, debug
);
7996 /* The shader consists of several parts:
7998 * - the middle part is the user shader, it has 1 variant only
7999 * and it was compiled during the creation of the shader
8001 * - the prolog part is inserted at the beginning
8002 * - the epilog part is inserted at the end
8004 * The prolog and epilog have many (but simple) variants.
8006 * Starting with gfx9, geometry and tessellation control
8007 * shaders also contain the prolog and user shader parts of
8008 * the previous shader stage.
8014 /* Copy the compiled TGSI shader data over. */
8015 shader
->is_binary_shared
= true;
8016 shader
->binary
= mainp
->binary
;
8017 shader
->config
= mainp
->config
;
8018 shader
->info
.num_input_sgprs
= mainp
->info
.num_input_sgprs
;
8019 shader
->info
.num_input_vgprs
= mainp
->info
.num_input_vgprs
;
8020 shader
->info
.face_vgpr_index
= mainp
->info
.face_vgpr_index
;
8021 shader
->info
.ancillary_vgpr_index
= mainp
->info
.ancillary_vgpr_index
;
8022 memcpy(shader
->info
.vs_output_param_offset
,
8023 mainp
->info
.vs_output_param_offset
,
8024 sizeof(mainp
->info
.vs_output_param_offset
));
8025 shader
->info
.uses_instanceid
= mainp
->info
.uses_instanceid
;
8026 shader
->info
.nr_pos_exports
= mainp
->info
.nr_pos_exports
;
8027 shader
->info
.nr_param_exports
= mainp
->info
.nr_param_exports
;
8029 /* Select prologs and/or epilogs. */
8030 switch (sel
->type
) {
8031 case PIPE_SHADER_VERTEX
:
8032 if (!si_shader_select_vs_parts(sscreen
, tm
, shader
, debug
))
8035 case PIPE_SHADER_TESS_CTRL
:
8036 if (!si_shader_select_tcs_parts(sscreen
, tm
, shader
, debug
))
8039 case PIPE_SHADER_TESS_EVAL
:
8041 case PIPE_SHADER_GEOMETRY
:
8042 if (!si_shader_select_gs_parts(sscreen
, tm
, shader
, debug
))
8045 case PIPE_SHADER_FRAGMENT
:
8046 if (!si_shader_select_ps_parts(sscreen
, tm
, shader
, debug
))
8049 /* Make sure we have at least as many VGPRs as there
8050 * are allocated inputs.
8052 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
8053 shader
->info
.num_input_vgprs
);
8057 /* Update SGPR and VGPR counts. */
8058 if (shader
->prolog
) {
8059 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
,
8060 shader
->prolog
->config
.num_sgprs
);
8061 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
8062 shader
->prolog
->config
.num_vgprs
);
8064 if (shader
->previous_stage
) {
8065 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
,
8066 shader
->previous_stage
->config
.num_sgprs
);
8067 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
8068 shader
->previous_stage
->config
.num_vgprs
);
8069 shader
->config
.spilled_sgprs
=
8070 MAX2(shader
->config
.spilled_sgprs
,
8071 shader
->previous_stage
->config
.spilled_sgprs
);
8072 shader
->config
.spilled_vgprs
=
8073 MAX2(shader
->config
.spilled_vgprs
,
8074 shader
->previous_stage
->config
.spilled_vgprs
);
8075 shader
->config
.private_mem_vgprs
=
8076 MAX2(shader
->config
.private_mem_vgprs
,
8077 shader
->previous_stage
->config
.private_mem_vgprs
);
8078 shader
->config
.scratch_bytes_per_wave
=
8079 MAX2(shader
->config
.scratch_bytes_per_wave
,
8080 shader
->previous_stage
->config
.scratch_bytes_per_wave
);
8081 shader
->info
.uses_instanceid
|=
8082 shader
->previous_stage
->info
.uses_instanceid
;
8084 if (shader
->prolog2
) {
8085 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
,
8086 shader
->prolog2
->config
.num_sgprs
);
8087 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
8088 shader
->prolog2
->config
.num_vgprs
);
8090 if (shader
->epilog
) {
8091 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
,
8092 shader
->epilog
->config
.num_sgprs
);
8093 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
8094 shader
->epilog
->config
.num_vgprs
);
8096 si_calculate_max_simd_waves(shader
);
8099 si_fix_resource_usage(sscreen
, shader
);
8100 si_shader_dump(sscreen
, shader
, debug
, sel
->info
.processor
,
8104 r
= si_shader_binary_upload(sscreen
, shader
);
8106 fprintf(stderr
, "LLVM failed to upload shader\n");
8113 void si_shader_destroy(struct si_shader
*shader
)
8115 if (shader
->scratch_bo
)
8116 r600_resource_reference(&shader
->scratch_bo
, NULL
);
8118 r600_resource_reference(&shader
->bo
, NULL
);
8120 if (!shader
->is_binary_shared
)
8121 ac_shader_binary_clean(&shader
->binary
);
8123 free(shader
->shader_log
);