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
,
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
);
1178 static LLVMValueRef
desc_from_addr_base64k(struct si_shader_context
*ctx
,
1181 LLVMBuilderRef builder
= ctx
->ac
.builder
;
1183 LLVMValueRef addr
= LLVMGetParam(ctx
->main_fn
, param
);
1184 addr
= LLVMBuildZExt(builder
, addr
, ctx
->i64
, "");
1185 addr
= LLVMBuildShl(builder
, addr
, LLVMConstInt(ctx
->i64
, 16, 0), "");
1187 uint64_t desc2
= 0xffffffff;
1188 uint64_t desc3
= S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1189 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1190 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1191 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1192 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1193 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
);
1194 LLVMValueRef hi
= LLVMConstInt(ctx
->i64
, desc2
| (desc3
<< 32), 0);
1196 LLVMValueRef desc
= LLVMGetUndef(LLVMVectorType(ctx
->i64
, 2));
1197 desc
= LLVMBuildInsertElement(builder
, desc
, addr
, ctx
->i32_0
, "");
1198 desc
= LLVMBuildInsertElement(builder
, desc
, hi
, ctx
->i32_1
, "");
1199 return LLVMBuildBitCast(builder
, desc
, ctx
->v4i32
, "");
1202 static LLVMValueRef
fetch_input_tcs(
1203 struct lp_build_tgsi_context
*bld_base
,
1204 const struct tgsi_full_src_register
*reg
,
1205 enum tgsi_opcode_type type
, unsigned swizzle
)
1207 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1208 LLVMValueRef dw_addr
, stride
;
1210 stride
= get_tcs_in_vertex_dw_stride(ctx
);
1211 dw_addr
= get_tcs_in_current_patch_offset(ctx
);
1212 dw_addr
= get_dw_address(ctx
, NULL
, reg
, stride
, dw_addr
);
1214 return lds_load(bld_base
, tgsi2llvmtype(bld_base
, type
), swizzle
, dw_addr
);
1217 static LLVMValueRef
si_nir_load_tcs_varyings(struct ac_shader_abi
*abi
,
1218 LLVMValueRef vertex_index
,
1219 LLVMValueRef param_index
,
1220 unsigned const_index
,
1222 unsigned driver_location
,
1224 unsigned num_components
,
1229 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1230 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
1231 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
1232 LLVMValueRef dw_addr
, stride
;
1234 driver_location
= driver_location
/ 4;
1237 stride
= get_tcs_in_vertex_dw_stride(ctx
);
1238 dw_addr
= get_tcs_in_current_patch_offset(ctx
);
1242 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
1244 stride
= get_tcs_out_vertex_dw_stride(ctx
);
1245 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
1250 /* Add the constant index to the indirect index */
1251 param_index
= LLVMBuildAdd(ctx
->ac
.builder
, param_index
,
1252 LLVMConstInt(ctx
->i32
, const_index
, 0), "");
1254 param_index
= LLVMConstInt(ctx
->i32
, const_index
, 0);
1257 dw_addr
= get_dw_address_from_generic_indices(ctx
, stride
, dw_addr
,
1258 vertex_index
, param_index
,
1260 info
->input_semantic_name
,
1261 info
->input_semantic_index
,
1264 LLVMValueRef value
[4];
1265 for (unsigned i
= 0; i
< num_components
+ component
; i
++) {
1266 value
[i
] = lds_load(bld_base
, ctx
->i32
, i
, dw_addr
);
1269 return ac_build_varying_gather_values(&ctx
->ac
, value
, num_components
, component
);
1272 static LLVMValueRef
fetch_output_tcs(
1273 struct lp_build_tgsi_context
*bld_base
,
1274 const struct tgsi_full_src_register
*reg
,
1275 enum tgsi_opcode_type type
, unsigned swizzle
)
1277 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1278 LLVMValueRef dw_addr
, stride
;
1280 if (reg
->Register
.Dimension
) {
1281 stride
= get_tcs_out_vertex_dw_stride(ctx
);
1282 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
1283 dw_addr
= get_dw_address(ctx
, NULL
, reg
, stride
, dw_addr
);
1285 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
1286 dw_addr
= get_dw_address(ctx
, NULL
, reg
, NULL
, dw_addr
);
1289 return lds_load(bld_base
, tgsi2llvmtype(bld_base
, type
), swizzle
, dw_addr
);
1292 static LLVMValueRef
fetch_input_tes(
1293 struct lp_build_tgsi_context
*bld_base
,
1294 const struct tgsi_full_src_register
*reg
,
1295 enum tgsi_opcode_type type
, unsigned swizzle
)
1297 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1298 LLVMValueRef buffer
, base
, addr
;
1300 buffer
= desc_from_addr_base64k(ctx
, ctx
->param_tcs_offchip_addr_base64k
);
1302 base
= LLVMGetParam(ctx
->main_fn
, ctx
->param_tcs_offchip_offset
);
1303 addr
= get_tcs_tes_buffer_address_from_reg(ctx
, NULL
, reg
);
1305 return buffer_load(bld_base
, tgsi2llvmtype(bld_base
, type
), swizzle
,
1306 buffer
, base
, addr
, true);
1309 LLVMValueRef
si_nir_load_input_tes(struct ac_shader_abi
*abi
,
1310 LLVMValueRef vertex_index
,
1311 LLVMValueRef param_index
,
1312 unsigned const_index
,
1314 unsigned driver_location
,
1316 unsigned num_components
,
1321 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1322 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
1323 LLVMValueRef buffer
, base
, addr
;
1325 driver_location
= driver_location
/ 4;
1327 buffer
= desc_from_addr_base64k(ctx
, ctx
->param_tcs_offchip_addr_base64k
);
1329 base
= LLVMGetParam(ctx
->main_fn
, ctx
->param_tcs_offchip_offset
);
1332 /* Add the constant index to the indirect index */
1333 param_index
= LLVMBuildAdd(ctx
->ac
.builder
, param_index
,
1334 LLVMConstInt(ctx
->i32
, const_index
, 0), "");
1336 param_index
= LLVMConstInt(ctx
->i32
, const_index
, 0);
1339 addr
= get_tcs_tes_buffer_address_from_generic_indices(ctx
, vertex_index
,
1340 param_index
, driver_location
,
1341 info
->input_semantic_name
,
1342 info
->input_semantic_index
,
1345 /* TODO: This will generate rather ordinary llvm code, although it
1346 * should be easy for the optimiser to fix up. In future we might want
1347 * to refactor buffer_load(), but for now this maximises code sharing
1348 * between the NIR and TGSI backends.
1350 LLVMValueRef value
[4];
1351 for (unsigned i
= component
; i
< num_components
+ component
; i
++) {
1352 value
[i
] = buffer_load(&ctx
->bld_base
, ctx
->i32
, i
, buffer
, base
, addr
, true);
1355 return ac_build_varying_gather_values(&ctx
->ac
, value
, num_components
, component
);
1358 static void store_output_tcs(struct lp_build_tgsi_context
*bld_base
,
1359 const struct tgsi_full_instruction
*inst
,
1360 const struct tgsi_opcode_info
*info
,
1362 LLVMValueRef dst
[4])
1364 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1365 const struct tgsi_full_dst_register
*reg
= &inst
->Dst
[index
];
1366 const struct tgsi_shader_info
*sh_info
= &ctx
->shader
->selector
->info
;
1367 unsigned chan_index
;
1368 LLVMValueRef dw_addr
, stride
;
1369 LLVMValueRef buffer
, base
, buf_addr
;
1370 LLVMValueRef values
[4];
1371 bool skip_lds_store
;
1372 bool is_tess_factor
= false, is_tess_inner
= false;
1374 /* Only handle per-patch and per-vertex outputs here.
1375 * Vectors will be lowered to scalars and this function will be called again.
1377 if (reg
->Register
.File
!= TGSI_FILE_OUTPUT
||
1378 (dst
[0] && LLVMGetTypeKind(LLVMTypeOf(dst
[0])) == LLVMVectorTypeKind
)) {
1379 si_llvm_emit_store(bld_base
, inst
, info
, index
, dst
);
1383 if (reg
->Register
.Dimension
) {
1384 stride
= get_tcs_out_vertex_dw_stride(ctx
);
1385 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
1386 dw_addr
= get_dw_address(ctx
, reg
, NULL
, stride
, dw_addr
);
1387 skip_lds_store
= !sh_info
->reads_pervertex_outputs
;
1389 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
1390 dw_addr
= get_dw_address(ctx
, reg
, NULL
, NULL
, dw_addr
);
1391 skip_lds_store
= !sh_info
->reads_perpatch_outputs
;
1393 if (!reg
->Register
.Indirect
) {
1394 int name
= sh_info
->output_semantic_name
[reg
->Register
.Index
];
1396 /* Always write tess factors into LDS for the TCS epilog. */
1397 if (name
== TGSI_SEMANTIC_TESSINNER
||
1398 name
== TGSI_SEMANTIC_TESSOUTER
) {
1399 /* The epilog doesn't read LDS if invocation 0 defines tess factors. */
1400 skip_lds_store
= !sh_info
->reads_tessfactor_outputs
&&
1401 ctx
->shader
->selector
->tcs_info
.tessfactors_are_def_in_all_invocs
;
1402 is_tess_factor
= true;
1403 is_tess_inner
= name
== TGSI_SEMANTIC_TESSINNER
;
1408 buffer
= desc_from_addr_base64k(ctx
, ctx
->param_tcs_offchip_addr_base64k
);
1410 base
= LLVMGetParam(ctx
->main_fn
, ctx
->param_tcs_offchip_offset
);
1411 buf_addr
= get_tcs_tes_buffer_address_from_reg(ctx
, reg
, NULL
);
1413 uint32_t writemask
= reg
->Register
.WriteMask
;
1415 chan_index
= u_bit_scan(&writemask
);
1416 LLVMValueRef value
= dst
[chan_index
];
1418 if (inst
->Instruction
.Saturate
)
1419 value
= ac_build_clamp(&ctx
->ac
, value
);
1421 /* Skip LDS stores if there is no LDS read of this output. */
1422 if (!skip_lds_store
)
1423 lds_store(ctx
, chan_index
, dw_addr
, value
);
1425 value
= ac_to_integer(&ctx
->ac
, value
);
1426 values
[chan_index
] = value
;
1428 if (reg
->Register
.WriteMask
!= 0xF && !is_tess_factor
) {
1429 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, value
, 1,
1431 4 * chan_index
, 1, 0, true, false);
1434 /* Write tess factors into VGPRs for the epilog. */
1435 if (is_tess_factor
&&
1436 ctx
->shader
->selector
->tcs_info
.tessfactors_are_def_in_all_invocs
) {
1437 if (!is_tess_inner
) {
1438 LLVMBuildStore(ctx
->ac
.builder
, value
, /* outer */
1439 ctx
->invoc0_tess_factors
[chan_index
]);
1440 } else if (chan_index
< 2) {
1441 LLVMBuildStore(ctx
->ac
.builder
, value
, /* inner */
1442 ctx
->invoc0_tess_factors
[4 + chan_index
]);
1447 if (reg
->Register
.WriteMask
== 0xF && !is_tess_factor
) {
1448 LLVMValueRef value
= lp_build_gather_values(&ctx
->gallivm
,
1450 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, value
, 4, buf_addr
,
1451 base
, 0, 1, 0, true, false);
1455 static void si_nir_store_output_tcs(struct ac_shader_abi
*abi
,
1456 LLVMValueRef vertex_index
,
1457 LLVMValueRef param_index
,
1458 unsigned const_index
,
1460 unsigned driver_location
,
1467 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1468 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
1469 LLVMValueRef dw_addr
, stride
;
1470 LLVMValueRef buffer
, base
, addr
;
1471 LLVMValueRef values
[4];
1472 bool skip_lds_store
;
1473 bool is_tess_factor
= false, is_tess_inner
= false;
1475 driver_location
= driver_location
/ 4;
1478 /* Add the constant index to the indirect index */
1479 param_index
= LLVMBuildAdd(ctx
->ac
.builder
, param_index
,
1480 LLVMConstInt(ctx
->i32
, const_index
, 0), "");
1482 if (const_index
!= 0)
1483 param_index
= LLVMConstInt(ctx
->i32
, const_index
, 0);
1487 stride
= get_tcs_out_vertex_dw_stride(ctx
);
1488 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
1489 dw_addr
= get_dw_address_from_generic_indices(ctx
, stride
, dw_addr
,
1490 vertex_index
, param_index
,
1492 info
->output_semantic_name
,
1493 info
->output_semantic_index
,
1496 skip_lds_store
= !info
->reads_pervertex_outputs
;
1498 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
1499 dw_addr
= get_dw_address_from_generic_indices(ctx
, NULL
, dw_addr
,
1500 vertex_index
, param_index
,
1502 info
->output_semantic_name
,
1503 info
->output_semantic_index
,
1506 skip_lds_store
= !info
->reads_perpatch_outputs
;
1509 int name
= info
->output_semantic_name
[driver_location
];
1511 /* Always write tess factors into LDS for the TCS epilog. */
1512 if (name
== TGSI_SEMANTIC_TESSINNER
||
1513 name
== TGSI_SEMANTIC_TESSOUTER
) {
1514 /* The epilog doesn't read LDS if invocation 0 defines tess factors. */
1515 skip_lds_store
= !info
->reads_tessfactor_outputs
&&
1516 ctx
->shader
->selector
->tcs_info
.tessfactors_are_def_in_all_invocs
;
1517 is_tess_factor
= true;
1518 is_tess_inner
= name
== TGSI_SEMANTIC_TESSINNER
;
1523 buffer
= desc_from_addr_base64k(ctx
, ctx
->param_tcs_offchip_addr_base64k
);
1525 base
= LLVMGetParam(ctx
->main_fn
, ctx
->param_tcs_offchip_offset
);
1527 addr
= get_tcs_tes_buffer_address_from_generic_indices(ctx
, vertex_index
,
1528 param_index
, driver_location
,
1529 info
->output_semantic_name
,
1530 info
->output_semantic_index
,
1533 for (unsigned chan
= 0; chan
< 4; chan
++) {
1534 if (!(writemask
& (1 << chan
)))
1536 LLVMValueRef value
= ac_llvm_extract_elem(&ctx
->ac
, src
, chan
- component
);
1538 /* Skip LDS stores if there is no LDS read of this output. */
1539 if (!skip_lds_store
)
1540 ac_lds_store(&ctx
->ac
, dw_addr
, value
);
1542 value
= ac_to_integer(&ctx
->ac
, value
);
1543 values
[chan
] = value
;
1545 if (writemask
!= 0xF && !is_tess_factor
) {
1546 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, value
, 1,
1548 4 * chan
, 1, 0, true, false);
1551 /* Write tess factors into VGPRs for the epilog. */
1552 if (is_tess_factor
&&
1553 ctx
->shader
->selector
->tcs_info
.tessfactors_are_def_in_all_invocs
) {
1554 if (!is_tess_inner
) {
1555 LLVMBuildStore(ctx
->ac
.builder
, value
, /* outer */
1556 ctx
->invoc0_tess_factors
[chan
]);
1557 } else if (chan
< 2) {
1558 LLVMBuildStore(ctx
->ac
.builder
, value
, /* inner */
1559 ctx
->invoc0_tess_factors
[4 + chan
]);
1564 if (writemask
== 0xF && !is_tess_factor
) {
1565 LLVMValueRef value
= lp_build_gather_values(&ctx
->gallivm
,
1567 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, value
, 4, addr
,
1568 base
, 0, 1, 0, true, false);
1572 LLVMValueRef
si_llvm_load_input_gs(struct ac_shader_abi
*abi
,
1573 unsigned input_index
,
1574 unsigned vtx_offset_param
,
1578 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1579 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
1580 struct si_shader
*shader
= ctx
->shader
;
1581 struct lp_build_context
*uint
= &ctx
->bld_base
.uint_bld
;
1582 LLVMValueRef vtx_offset
, soffset
;
1583 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
1584 unsigned semantic_name
= info
->input_semantic_name
[input_index
];
1585 unsigned semantic_index
= info
->input_semantic_index
[input_index
];
1589 param
= si_shader_io_get_unique_index(semantic_name
, semantic_index
);
1591 /* GFX9 has the ESGS ring in LDS. */
1592 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
1593 unsigned index
= vtx_offset_param
;
1595 switch (index
/ 2) {
1597 vtx_offset
= unpack_param(ctx
, ctx
->param_gs_vtx01_offset
,
1598 index
% 2 ? 16 : 0, 16);
1601 vtx_offset
= unpack_param(ctx
, ctx
->param_gs_vtx23_offset
,
1602 index
% 2 ? 16 : 0, 16);
1605 vtx_offset
= unpack_param(ctx
, ctx
->param_gs_vtx45_offset
,
1606 index
% 2 ? 16 : 0, 16);
1613 vtx_offset
= LLVMBuildAdd(ctx
->ac
.builder
, vtx_offset
,
1614 LLVMConstInt(ctx
->i32
, param
* 4, 0), "");
1615 return lds_load(bld_base
, type
, swizzle
, vtx_offset
);
1618 /* GFX6: input load from the ESGS ring in memory. */
1619 if (swizzle
== ~0) {
1620 LLVMValueRef values
[TGSI_NUM_CHANNELS
];
1622 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
1623 values
[chan
] = si_llvm_load_input_gs(abi
, input_index
, vtx_offset_param
,
1626 return lp_build_gather_values(&ctx
->gallivm
, values
,
1630 /* Get the vertex offset parameter on GFX6. */
1631 LLVMValueRef gs_vtx_offset
= ctx
->gs_vtx_offset
[vtx_offset_param
];
1633 vtx_offset
= lp_build_mul_imm(uint
, gs_vtx_offset
, 4);
1635 soffset
= LLVMConstInt(ctx
->i32
, (param
* 4 + swizzle
) * 256, 0);
1637 value
= ac_build_buffer_load(&ctx
->ac
, ctx
->esgs_ring
, 1, ctx
->i32_0
,
1638 vtx_offset
, soffset
, 0, 1, 0, true, false);
1639 if (llvm_type_is_64bit(ctx
, type
)) {
1640 LLVMValueRef value2
;
1641 soffset
= LLVMConstInt(ctx
->i32
, (param
* 4 + swizzle
+ 1) * 256, 0);
1643 value2
= ac_build_buffer_load(&ctx
->ac
, ctx
->esgs_ring
, 1,
1644 ctx
->i32_0
, vtx_offset
, soffset
,
1645 0, 1, 0, true, false);
1646 return si_llvm_emit_fetch_64bit(bld_base
, type
, value
, value2
);
1648 return LLVMBuildBitCast(ctx
->ac
.builder
, value
, type
, "");
1651 static LLVMValueRef
fetch_input_gs(
1652 struct lp_build_tgsi_context
*bld_base
,
1653 const struct tgsi_full_src_register
*reg
,
1654 enum tgsi_opcode_type type
,
1657 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1658 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
1660 unsigned semantic_name
= info
->input_semantic_name
[reg
->Register
.Index
];
1661 if (swizzle
!= ~0 && semantic_name
== TGSI_SEMANTIC_PRIMID
)
1662 return get_primitive_id(ctx
, swizzle
);
1664 if (!reg
->Register
.Dimension
)
1667 return si_llvm_load_input_gs(&ctx
->abi
, reg
->Register
.Index
,
1668 reg
->Dimension
.Index
,
1669 tgsi2llvmtype(bld_base
, type
),
1673 static int lookup_interp_param_index(unsigned interpolate
, unsigned location
)
1675 switch (interpolate
) {
1676 case TGSI_INTERPOLATE_CONSTANT
:
1679 case TGSI_INTERPOLATE_LINEAR
:
1680 if (location
== TGSI_INTERPOLATE_LOC_SAMPLE
)
1681 return SI_PARAM_LINEAR_SAMPLE
;
1682 else if (location
== TGSI_INTERPOLATE_LOC_CENTROID
)
1683 return SI_PARAM_LINEAR_CENTROID
;
1685 return SI_PARAM_LINEAR_CENTER
;
1687 case TGSI_INTERPOLATE_COLOR
:
1688 case TGSI_INTERPOLATE_PERSPECTIVE
:
1689 if (location
== TGSI_INTERPOLATE_LOC_SAMPLE
)
1690 return SI_PARAM_PERSP_SAMPLE
;
1691 else if (location
== TGSI_INTERPOLATE_LOC_CENTROID
)
1692 return SI_PARAM_PERSP_CENTROID
;
1694 return SI_PARAM_PERSP_CENTER
;
1697 fprintf(stderr
, "Warning: Unhandled interpolation mode.\n");
1702 static LLVMValueRef
si_build_fs_interp(struct si_shader_context
*ctx
,
1703 unsigned attr_index
, unsigned chan
,
1704 LLVMValueRef prim_mask
,
1705 LLVMValueRef i
, LLVMValueRef j
)
1708 return ac_build_fs_interp(&ctx
->ac
,
1709 LLVMConstInt(ctx
->i32
, chan
, 0),
1710 LLVMConstInt(ctx
->i32
, attr_index
, 0),
1713 return ac_build_fs_interp_mov(&ctx
->ac
,
1714 LLVMConstInt(ctx
->i32
, 2, 0), /* P0 */
1715 LLVMConstInt(ctx
->i32
, chan
, 0),
1716 LLVMConstInt(ctx
->i32
, attr_index
, 0),
1721 * Interpolate a fragment shader input.
1723 * @param ctx context
1724 * @param input_index index of the input in hardware
1725 * @param semantic_name TGSI_SEMANTIC_*
1726 * @param semantic_index semantic index
1727 * @param num_interp_inputs number of all interpolated inputs (= BCOLOR offset)
1728 * @param colors_read_mask color components read (4 bits for each color, 8 bits in total)
1729 * @param interp_param interpolation weights (i,j)
1730 * @param prim_mask SI_PARAM_PRIM_MASK
1731 * @param face SI_PARAM_FRONT_FACE
1732 * @param result the return value (4 components)
1734 static void interp_fs_input(struct si_shader_context
*ctx
,
1735 unsigned input_index
,
1736 unsigned semantic_name
,
1737 unsigned semantic_index
,
1738 unsigned num_interp_inputs
,
1739 unsigned colors_read_mask
,
1740 LLVMValueRef interp_param
,
1741 LLVMValueRef prim_mask
,
1743 LLVMValueRef result
[4])
1745 LLVMValueRef i
= NULL
, j
= NULL
;
1748 /* fs.constant returns the param from the middle vertex, so it's not
1749 * really useful for flat shading. It's meant to be used for custom
1750 * interpolation (but the intrinsic can't fetch from the other two
1753 * Luckily, it doesn't matter, because we rely on the FLAT_SHADE state
1754 * to do the right thing. The only reason we use fs.constant is that
1755 * fs.interp cannot be used on integers, because they can be equal
1758 * When interp is false we will use fs.constant or for newer llvm,
1759 * amdgcn.interp.mov.
1761 bool interp
= interp_param
!= NULL
;
1764 interp_param
= LLVMBuildBitCast(ctx
->ac
.builder
, interp_param
,
1765 LLVMVectorType(ctx
->f32
, 2), "");
1767 i
= LLVMBuildExtractElement(ctx
->ac
.builder
, interp_param
,
1769 j
= LLVMBuildExtractElement(ctx
->ac
.builder
, interp_param
,
1773 if (semantic_name
== TGSI_SEMANTIC_COLOR
&&
1774 ctx
->shader
->key
.part
.ps
.prolog
.color_two_side
) {
1775 LLVMValueRef is_face_positive
;
1777 /* If BCOLOR0 is used, BCOLOR1 is at offset "num_inputs + 1",
1778 * otherwise it's at offset "num_inputs".
1780 unsigned back_attr_offset
= num_interp_inputs
;
1781 if (semantic_index
== 1 && colors_read_mask
& 0xf)
1782 back_attr_offset
+= 1;
1784 is_face_positive
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntNE
,
1785 face
, ctx
->i32_0
, "");
1787 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
1788 LLVMValueRef front
, back
;
1790 front
= si_build_fs_interp(ctx
,
1793 back
= si_build_fs_interp(ctx
,
1794 back_attr_offset
, chan
,
1797 result
[chan
] = LLVMBuildSelect(ctx
->ac
.builder
,
1803 } else if (semantic_name
== TGSI_SEMANTIC_FOG
) {
1804 result
[0] = si_build_fs_interp(ctx
, input_index
,
1805 0, prim_mask
, i
, j
);
1807 result
[2] = LLVMConstReal(ctx
->f32
, 0.0f
);
1808 result
[3] = LLVMConstReal(ctx
->f32
, 1.0f
);
1810 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
1811 result
[chan
] = si_build_fs_interp(ctx
,
1818 void si_llvm_load_input_fs(
1819 struct si_shader_context
*ctx
,
1820 unsigned input_index
,
1821 LLVMValueRef out
[4])
1823 struct lp_build_context
*base
= &ctx
->bld_base
.base
;
1824 struct si_shader
*shader
= ctx
->shader
;
1825 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
1826 LLVMValueRef main_fn
= ctx
->main_fn
;
1827 LLVMValueRef interp_param
= NULL
;
1828 int interp_param_idx
;
1829 enum tgsi_semantic semantic_name
= info
->input_semantic_name
[input_index
];
1830 unsigned semantic_index
= info
->input_semantic_index
[input_index
];
1831 enum tgsi_interpolate_mode interp_mode
= info
->input_interpolate
[input_index
];
1832 enum tgsi_interpolate_loc interp_loc
= info
->input_interpolate_loc
[input_index
];
1834 /* Get colors from input VGPRs (set by the prolog). */
1835 if (semantic_name
== TGSI_SEMANTIC_COLOR
) {
1836 unsigned colors_read
= shader
->selector
->info
.colors_read
;
1837 unsigned mask
= colors_read
>> (semantic_index
* 4);
1838 unsigned offset
= SI_PARAM_POS_FIXED_PT
+ 1 +
1839 (semantic_index
? util_bitcount(colors_read
& 0xf) : 0);
1841 out
[0] = mask
& 0x1 ? LLVMGetParam(main_fn
, offset
++) : base
->undef
;
1842 out
[1] = mask
& 0x2 ? LLVMGetParam(main_fn
, offset
++) : base
->undef
;
1843 out
[2] = mask
& 0x4 ? LLVMGetParam(main_fn
, offset
++) : base
->undef
;
1844 out
[3] = mask
& 0x8 ? LLVMGetParam(main_fn
, offset
++) : base
->undef
;
1848 interp_param_idx
= lookup_interp_param_index(interp_mode
, interp_loc
);
1849 if (interp_param_idx
== -1)
1851 else if (interp_param_idx
) {
1852 interp_param
= LLVMGetParam(ctx
->main_fn
, interp_param_idx
);
1855 interp_fs_input(ctx
, input_index
, semantic_name
,
1856 semantic_index
, 0, /* this param is unused */
1857 shader
->selector
->info
.colors_read
, interp_param
,
1859 LLVMGetParam(main_fn
, SI_PARAM_FRONT_FACE
),
1863 static void declare_input_fs(
1864 struct si_shader_context
*ctx
,
1865 unsigned input_index
,
1866 const struct tgsi_full_declaration
*decl
,
1867 LLVMValueRef out
[4])
1869 si_llvm_load_input_fs(ctx
, input_index
, out
);
1872 static LLVMValueRef
get_sample_id(struct si_shader_context
*ctx
)
1874 return unpack_param(ctx
, SI_PARAM_ANCILLARY
, 8, 4);
1879 * Load a dword from a constant buffer.
1881 static LLVMValueRef
buffer_load_const(struct si_shader_context
*ctx
,
1882 LLVMValueRef resource
,
1883 LLVMValueRef offset
)
1885 return ac_build_buffer_load(&ctx
->ac
, resource
, 1, NULL
, offset
, NULL
,
1886 0, 0, 0, true, true);
1889 static LLVMValueRef
load_sample_position(struct ac_shader_abi
*abi
, LLVMValueRef sample_id
)
1891 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1892 struct lp_build_context
*uint_bld
= &ctx
->bld_base
.uint_bld
;
1893 LLVMValueRef desc
= LLVMGetParam(ctx
->main_fn
, ctx
->param_rw_buffers
);
1894 LLVMValueRef buf_index
= LLVMConstInt(ctx
->i32
, SI_PS_CONST_SAMPLE_POSITIONS
, 0);
1895 LLVMValueRef resource
= ac_build_load_to_sgpr(&ctx
->ac
, desc
, buf_index
);
1897 /* offset = sample_id * 8 (8 = 2 floats containing samplepos.xy) */
1898 LLVMValueRef offset0
= lp_build_mul_imm(uint_bld
, sample_id
, 8);
1899 LLVMValueRef offset1
= LLVMBuildAdd(ctx
->ac
.builder
, offset0
, LLVMConstInt(ctx
->i32
, 4, 0), "");
1901 LLVMValueRef pos
[4] = {
1902 buffer_load_const(ctx
, resource
, offset0
),
1903 buffer_load_const(ctx
, resource
, offset1
),
1904 LLVMConstReal(ctx
->f32
, 0),
1905 LLVMConstReal(ctx
->f32
, 0)
1908 return lp_build_gather_values(&ctx
->gallivm
, pos
, 4);
1911 static LLVMValueRef
si_load_tess_coord(struct ac_shader_abi
*abi
,
1913 unsigned num_components
)
1915 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1916 struct lp_build_context
*bld
= &ctx
->bld_base
.base
;
1918 LLVMValueRef coord
[4] = {
1919 LLVMGetParam(ctx
->main_fn
, ctx
->param_tes_u
),
1920 LLVMGetParam(ctx
->main_fn
, ctx
->param_tes_v
),
1925 /* For triangles, the vector should be (u, v, 1-u-v). */
1926 if (ctx
->shader
->selector
->info
.properties
[TGSI_PROPERTY_TES_PRIM_MODE
] ==
1927 PIPE_PRIM_TRIANGLES
)
1928 coord
[2] = lp_build_sub(bld
, ctx
->ac
.f32_1
,
1929 lp_build_add(bld
, coord
[0], coord
[1]));
1931 return lp_build_gather_values(&ctx
->gallivm
, coord
, 4);
1934 static LLVMValueRef
load_tess_level(struct si_shader_context
*ctx
,
1935 unsigned semantic_name
)
1937 LLVMValueRef buffer
, base
, addr
;
1939 int param
= si_shader_io_get_unique_index_patch(semantic_name
, 0);
1941 buffer
= desc_from_addr_base64k(ctx
, ctx
->param_tcs_offchip_addr_base64k
);
1943 base
= LLVMGetParam(ctx
->main_fn
, ctx
->param_tcs_offchip_offset
);
1944 addr
= get_tcs_tes_buffer_address(ctx
, get_rel_patch_id(ctx
), NULL
,
1945 LLVMConstInt(ctx
->i32
, param
, 0));
1947 return buffer_load(&ctx
->bld_base
, ctx
->f32
,
1948 ~0, buffer
, base
, addr
, true);
1952 static LLVMValueRef
si_load_tess_level(struct ac_shader_abi
*abi
,
1953 unsigned varying_id
)
1955 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1956 unsigned semantic_name
;
1958 switch (varying_id
) {
1959 case VARYING_SLOT_TESS_LEVEL_INNER
:
1960 semantic_name
= TGSI_SEMANTIC_TESSINNER
;
1962 case VARYING_SLOT_TESS_LEVEL_OUTER
:
1963 semantic_name
= TGSI_SEMANTIC_TESSOUTER
;
1966 unreachable("unknown tess level");
1969 return load_tess_level(ctx
, semantic_name
);
1973 static LLVMValueRef
si_load_patch_vertices_in(struct ac_shader_abi
*abi
)
1975 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1976 if (ctx
->type
== PIPE_SHADER_TESS_CTRL
)
1977 return unpack_param(ctx
, ctx
->param_tcs_out_lds_layout
, 26, 6);
1978 else if (ctx
->type
== PIPE_SHADER_TESS_EVAL
)
1979 return get_num_tcs_out_vertices(ctx
);
1981 unreachable("invalid shader stage for TGSI_SEMANTIC_VERTICESIN");
1984 void si_load_system_value(struct si_shader_context
*ctx
,
1986 const struct tgsi_full_declaration
*decl
)
1988 LLVMValueRef value
= 0;
1990 assert(index
< RADEON_LLVM_MAX_SYSTEM_VALUES
);
1992 switch (decl
->Semantic
.Name
) {
1993 case TGSI_SEMANTIC_INSTANCEID
:
1994 value
= ctx
->abi
.instance_id
;
1997 case TGSI_SEMANTIC_VERTEXID
:
1998 value
= LLVMBuildAdd(ctx
->ac
.builder
,
2000 ctx
->abi
.base_vertex
, "");
2003 case TGSI_SEMANTIC_VERTEXID_NOBASE
:
2004 /* Unused. Clarify the meaning in indexed vs. non-indexed
2005 * draws if this is ever used again. */
2009 case TGSI_SEMANTIC_BASEVERTEX
:
2011 /* For non-indexed draws, the base vertex set by the driver
2012 * (for direct draws) or the CP (for indirect draws) is the
2013 * first vertex ID, but GLSL expects 0 to be returned.
2015 LLVMValueRef vs_state
= LLVMGetParam(ctx
->main_fn
, ctx
->param_vs_state_bits
);
2016 LLVMValueRef indexed
;
2018 indexed
= LLVMBuildLShr(ctx
->ac
.builder
, vs_state
, ctx
->i32_1
, "");
2019 indexed
= LLVMBuildTrunc(ctx
->ac
.builder
, indexed
, ctx
->i1
, "");
2021 value
= LLVMBuildSelect(ctx
->ac
.builder
, indexed
,
2022 ctx
->abi
.base_vertex
, ctx
->i32_0
, "");
2026 case TGSI_SEMANTIC_BASEINSTANCE
:
2027 value
= ctx
->abi
.start_instance
;
2030 case TGSI_SEMANTIC_DRAWID
:
2031 value
= ctx
->abi
.draw_id
;
2034 case TGSI_SEMANTIC_INVOCATIONID
:
2035 if (ctx
->type
== PIPE_SHADER_TESS_CTRL
)
2036 value
= unpack_llvm_param(ctx
, ctx
->abi
.tcs_rel_ids
, 8, 5);
2037 else if (ctx
->type
== PIPE_SHADER_GEOMETRY
)
2038 value
= ctx
->abi
.gs_invocation_id
;
2040 assert(!"INVOCATIONID not implemented");
2043 case TGSI_SEMANTIC_POSITION
:
2045 LLVMValueRef pos
[4] = {
2046 LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_X_FLOAT
),
2047 LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_Y_FLOAT
),
2048 LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_Z_FLOAT
),
2049 lp_build_emit_llvm_unary(&ctx
->bld_base
, TGSI_OPCODE_RCP
,
2050 LLVMGetParam(ctx
->main_fn
,
2051 SI_PARAM_POS_W_FLOAT
)),
2053 value
= lp_build_gather_values(&ctx
->gallivm
, pos
, 4);
2057 case TGSI_SEMANTIC_FACE
:
2058 value
= ctx
->abi
.front_face
;
2061 case TGSI_SEMANTIC_SAMPLEID
:
2062 value
= get_sample_id(ctx
);
2065 case TGSI_SEMANTIC_SAMPLEPOS
: {
2066 LLVMValueRef pos
[4] = {
2067 LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_X_FLOAT
),
2068 LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_Y_FLOAT
),
2069 LLVMConstReal(ctx
->f32
, 0),
2070 LLVMConstReal(ctx
->f32
, 0)
2072 pos
[0] = lp_build_emit_llvm_unary(&ctx
->bld_base
,
2073 TGSI_OPCODE_FRC
, pos
[0]);
2074 pos
[1] = lp_build_emit_llvm_unary(&ctx
->bld_base
,
2075 TGSI_OPCODE_FRC
, pos
[1]);
2076 value
= lp_build_gather_values(&ctx
->gallivm
, pos
, 4);
2080 case TGSI_SEMANTIC_SAMPLEMASK
:
2081 /* This can only occur with the OpenGL Core profile, which
2082 * doesn't support smoothing.
2084 value
= LLVMGetParam(ctx
->main_fn
, SI_PARAM_SAMPLE_COVERAGE
);
2087 case TGSI_SEMANTIC_TESSCOORD
:
2088 value
= si_load_tess_coord(&ctx
->abi
, NULL
, 4);
2091 case TGSI_SEMANTIC_VERTICESIN
:
2092 value
= si_load_patch_vertices_in(&ctx
->abi
);
2095 case TGSI_SEMANTIC_TESSINNER
:
2096 case TGSI_SEMANTIC_TESSOUTER
:
2097 value
= load_tess_level(ctx
, decl
->Semantic
.Name
);
2100 case TGSI_SEMANTIC_DEFAULT_TESSOUTER_SI
:
2101 case TGSI_SEMANTIC_DEFAULT_TESSINNER_SI
:
2103 LLVMValueRef buf
, slot
, val
[4];
2106 slot
= LLVMConstInt(ctx
->i32
, SI_HS_CONST_DEFAULT_TESS_LEVELS
, 0);
2107 buf
= LLVMGetParam(ctx
->main_fn
, ctx
->param_rw_buffers
);
2108 buf
= ac_build_load_to_sgpr(&ctx
->ac
, buf
, slot
);
2109 offset
= decl
->Semantic
.Name
== TGSI_SEMANTIC_DEFAULT_TESSINNER_SI
? 4 : 0;
2111 for (i
= 0; i
< 4; i
++)
2112 val
[i
] = buffer_load_const(ctx
, buf
,
2113 LLVMConstInt(ctx
->i32
, (offset
+ i
) * 4, 0));
2114 value
= lp_build_gather_values(&ctx
->gallivm
, val
, 4);
2118 case TGSI_SEMANTIC_PRIMID
:
2119 value
= get_primitive_id(ctx
, 0);
2122 case TGSI_SEMANTIC_GRID_SIZE
:
2123 value
= LLVMGetParam(ctx
->main_fn
, ctx
->param_grid_size
);
2126 case TGSI_SEMANTIC_BLOCK_SIZE
:
2128 LLVMValueRef values
[3];
2130 unsigned *properties
= ctx
->shader
->selector
->info
.properties
;
2132 if (properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
] != 0) {
2133 unsigned sizes
[3] = {
2134 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
],
2135 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT
],
2136 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH
]
2139 for (i
= 0; i
< 3; ++i
)
2140 values
[i
] = LLVMConstInt(ctx
->i32
, sizes
[i
], 0);
2142 value
= lp_build_gather_values(&ctx
->gallivm
, values
, 3);
2144 value
= LLVMGetParam(ctx
->main_fn
, ctx
->param_block_size
);
2149 case TGSI_SEMANTIC_BLOCK_ID
:
2151 LLVMValueRef values
[3];
2153 for (int i
= 0; i
< 3; i
++) {
2154 values
[i
] = ctx
->i32_0
;
2155 if (ctx
->param_block_id
[i
] >= 0) {
2156 values
[i
] = LLVMGetParam(ctx
->main_fn
,
2157 ctx
->param_block_id
[i
]);
2160 value
= lp_build_gather_values(&ctx
->gallivm
, values
, 3);
2164 case TGSI_SEMANTIC_THREAD_ID
:
2165 value
= LLVMGetParam(ctx
->main_fn
, ctx
->param_thread_id
);
2168 case TGSI_SEMANTIC_HELPER_INVOCATION
:
2169 value
= lp_build_intrinsic(ctx
->ac
.builder
,
2170 "llvm.amdgcn.ps.live",
2172 LP_FUNC_ATTR_READNONE
);
2173 value
= LLVMBuildNot(ctx
->ac
.builder
, value
, "");
2174 value
= LLVMBuildSExt(ctx
->ac
.builder
, value
, ctx
->i32
, "");
2177 case TGSI_SEMANTIC_SUBGROUP_SIZE
:
2178 value
= LLVMConstInt(ctx
->i32
, 64, 0);
2181 case TGSI_SEMANTIC_SUBGROUP_INVOCATION
:
2182 value
= ac_get_thread_id(&ctx
->ac
);
2185 case TGSI_SEMANTIC_SUBGROUP_EQ_MASK
:
2187 LLVMValueRef id
= ac_get_thread_id(&ctx
->ac
);
2188 id
= LLVMBuildZExt(ctx
->ac
.builder
, id
, ctx
->i64
, "");
2189 value
= LLVMBuildShl(ctx
->ac
.builder
, LLVMConstInt(ctx
->i64
, 1, 0), id
, "");
2190 value
= LLVMBuildBitCast(ctx
->ac
.builder
, value
, ctx
->v2i32
, "");
2194 case TGSI_SEMANTIC_SUBGROUP_GE_MASK
:
2195 case TGSI_SEMANTIC_SUBGROUP_GT_MASK
:
2196 case TGSI_SEMANTIC_SUBGROUP_LE_MASK
:
2197 case TGSI_SEMANTIC_SUBGROUP_LT_MASK
:
2199 LLVMValueRef id
= ac_get_thread_id(&ctx
->ac
);
2200 if (decl
->Semantic
.Name
== TGSI_SEMANTIC_SUBGROUP_GT_MASK
||
2201 decl
->Semantic
.Name
== TGSI_SEMANTIC_SUBGROUP_LE_MASK
) {
2202 /* All bits set except LSB */
2203 value
= LLVMConstInt(ctx
->i64
, -2, 0);
2206 value
= LLVMConstInt(ctx
->i64
, -1, 0);
2208 id
= LLVMBuildZExt(ctx
->ac
.builder
, id
, ctx
->i64
, "");
2209 value
= LLVMBuildShl(ctx
->ac
.builder
, value
, id
, "");
2210 if (decl
->Semantic
.Name
== TGSI_SEMANTIC_SUBGROUP_LE_MASK
||
2211 decl
->Semantic
.Name
== TGSI_SEMANTIC_SUBGROUP_LT_MASK
)
2212 value
= LLVMBuildNot(ctx
->ac
.builder
, value
, "");
2213 value
= LLVMBuildBitCast(ctx
->ac
.builder
, value
, ctx
->v2i32
, "");
2218 assert(!"unknown system value");
2222 ctx
->system_values
[index
] = value
;
2225 void si_declare_compute_memory(struct si_shader_context
*ctx
,
2226 const struct tgsi_full_declaration
*decl
)
2228 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
2230 LLVMTypeRef i8p
= LLVMPointerType(ctx
->i8
, AC_LOCAL_ADDR_SPACE
);
2233 assert(decl
->Declaration
.MemType
== TGSI_MEMORY_TYPE_SHARED
);
2234 assert(decl
->Range
.First
== decl
->Range
.Last
);
2235 assert(!ctx
->ac
.lds
);
2237 var
= LLVMAddGlobalInAddressSpace(ctx
->ac
.module
,
2238 LLVMArrayType(ctx
->i8
, sel
->local_size
),
2240 AC_LOCAL_ADDR_SPACE
);
2241 LLVMSetAlignment(var
, 4);
2243 ctx
->ac
.lds
= LLVMBuildBitCast(ctx
->ac
.builder
, var
, i8p
, "");
2246 static LLVMValueRef
load_const_buffer_desc(struct si_shader_context
*ctx
, int i
)
2248 LLVMValueRef list_ptr
= LLVMGetParam(ctx
->main_fn
,
2249 ctx
->param_const_and_shader_buffers
);
2251 return ac_build_load_to_sgpr(&ctx
->ac
, list_ptr
,
2252 LLVMConstInt(ctx
->i32
, si_get_constbuf_slot(i
), 0));
2255 static LLVMValueRef
load_ubo(struct ac_shader_abi
*abi
, LLVMValueRef index
)
2257 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
2258 LLVMValueRef ptr
= LLVMGetParam(ctx
->main_fn
, ctx
->param_const_and_shader_buffers
);
2260 index
= si_llvm_bound_index(ctx
, index
, ctx
->num_const_buffers
);
2261 index
= LLVMBuildAdd(ctx
->ac
.builder
, index
,
2262 LLVMConstInt(ctx
->i32
, SI_NUM_SHADER_BUFFERS
, 0), "");
2264 return ac_build_load_to_sgpr(&ctx
->ac
, ptr
, index
);
2268 load_ssbo(struct ac_shader_abi
*abi
, LLVMValueRef index
, bool write
)
2270 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
2271 LLVMValueRef rsrc_ptr
= LLVMGetParam(ctx
->main_fn
,
2272 ctx
->param_const_and_shader_buffers
);
2274 index
= si_llvm_bound_index(ctx
, index
, ctx
->num_shader_buffers
);
2275 index
= LLVMBuildSub(ctx
->ac
.builder
,
2276 LLVMConstInt(ctx
->i32
, SI_NUM_SHADER_BUFFERS
- 1, 0),
2279 return ac_build_load_to_sgpr(&ctx
->ac
, rsrc_ptr
, index
);
2282 static LLVMValueRef
fetch_constant(
2283 struct lp_build_tgsi_context
*bld_base
,
2284 const struct tgsi_full_src_register
*reg
,
2285 enum tgsi_opcode_type type
,
2288 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2289 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
2290 const struct tgsi_ind_register
*ireg
= ®
->Indirect
;
2293 LLVMValueRef addr
, bufp
;
2295 if (swizzle
== LP_CHAN_ALL
) {
2297 LLVMValueRef values
[4];
2298 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
)
2299 values
[chan
] = fetch_constant(bld_base
, reg
, type
, chan
);
2301 return lp_build_gather_values(&ctx
->gallivm
, values
, 4);
2304 /* Split 64-bit loads. */
2305 if (tgsi_type_is_64bit(type
)) {
2306 LLVMValueRef lo
, hi
;
2308 lo
= fetch_constant(bld_base
, reg
, TGSI_TYPE_UNSIGNED
, swizzle
);
2309 hi
= fetch_constant(bld_base
, reg
, TGSI_TYPE_UNSIGNED
, swizzle
+ 1);
2310 return si_llvm_emit_fetch_64bit(bld_base
, tgsi2llvmtype(bld_base
, type
),
2314 idx
= reg
->Register
.Index
* 4 + swizzle
;
2315 if (reg
->Register
.Indirect
) {
2316 addr
= si_get_indirect_index(ctx
, ireg
, 16, idx
* 4);
2318 addr
= LLVMConstInt(ctx
->i32
, idx
* 4, 0);
2321 /* Fast path when user data SGPRs point to constant buffer 0 directly. */
2322 if (sel
->info
.const_buffers_declared
== 1 &&
2323 sel
->info
.shader_buffers_declared
== 0) {
2325 LLVMGetParam(ctx
->main_fn
, ctx
->param_const_and_shader_buffers
);
2327 /* This enables use of s_load_dword and flat_load_dword for const buffer 0
2328 * loads, and up to x4 load opcode merging. However, it leads to horrible
2329 * code reducing SIMD wave occupancy from 8 to 2 in many cases.
2331 * Using s_buffer_load_dword (x1) seems to be the best option right now.
2333 * LLVM 5.0 on SI doesn't insert a required s_nop between SALU setting
2334 * a descriptor and s_buffer_load_dword using it, so we can't expand
2335 * the pointer into a full descriptor like below. We have to use
2336 * s_load_dword instead. The only case when LLVM 5.0 would select
2337 * s_buffer_load_dword (that we have to prevent) is when we use use
2338 * a literal offset where we don't need bounds checking.
2340 if (ctx
->screen
->info
.chip_class
== SI
&&
2341 HAVE_LLVM
< 0x0600 &&
2342 !reg
->Register
.Indirect
) {
2343 addr
= LLVMBuildLShr(ctx
->ac
.builder
, addr
, LLVMConstInt(ctx
->i32
, 2, 0), "");
2344 LLVMValueRef result
= ac_build_load_invariant(&ctx
->ac
, ptr
, addr
);
2345 return bitcast(bld_base
, type
, result
);
2348 /* Do the bounds checking with a descriptor, because
2349 * doing computation and manual bounds checking of 64-bit
2350 * addresses generates horrible VALU code with very high
2351 * VGPR usage and very low SIMD occupancy.
2353 ptr
= LLVMBuildPtrToInt(ctx
->ac
.builder
, ptr
, ctx
->i64
, "");
2354 ptr
= LLVMBuildBitCast(ctx
->ac
.builder
, ptr
, ctx
->v2i32
, "");
2356 LLVMValueRef desc_elems
[] = {
2357 LLVMBuildExtractElement(ctx
->ac
.builder
, ptr
, ctx
->i32_0
, ""),
2358 LLVMBuildExtractElement(ctx
->ac
.builder
, ptr
, ctx
->i32_1
, ""),
2359 LLVMConstInt(ctx
->i32
, (sel
->info
.const_file_max
[0] + 1) * 16, 0),
2360 LLVMConstInt(ctx
->i32
,
2361 S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
2362 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
2363 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
2364 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
2365 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
2366 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
), 0)
2368 LLVMValueRef desc
= ac_build_gather_values(&ctx
->ac
, desc_elems
, 4);
2369 LLVMValueRef result
= buffer_load_const(ctx
, desc
, addr
);
2370 return bitcast(bld_base
, type
, result
);
2373 assert(reg
->Register
.Dimension
);
2374 buf
= reg
->Dimension
.Index
;
2376 if (reg
->Dimension
.Indirect
) {
2377 LLVMValueRef ptr
= LLVMGetParam(ctx
->main_fn
, ctx
->param_const_and_shader_buffers
);
2379 index
= si_get_bounded_indirect_index(ctx
, ®
->DimIndirect
,
2380 reg
->Dimension
.Index
,
2381 ctx
->num_const_buffers
);
2382 index
= LLVMBuildAdd(ctx
->ac
.builder
, index
,
2383 LLVMConstInt(ctx
->i32
, SI_NUM_SHADER_BUFFERS
, 0), "");
2384 bufp
= ac_build_load_to_sgpr(&ctx
->ac
, ptr
, index
);
2386 bufp
= load_const_buffer_desc(ctx
, buf
);
2388 return bitcast(bld_base
, type
, buffer_load_const(ctx
, bufp
, addr
));
2391 /* Upper 16 bits must be zero. */
2392 static LLVMValueRef
si_llvm_pack_two_int16(struct si_shader_context
*ctx
,
2393 LLVMValueRef val
[2])
2395 return LLVMBuildOr(ctx
->ac
.builder
, val
[0],
2396 LLVMBuildShl(ctx
->ac
.builder
, val
[1],
2397 LLVMConstInt(ctx
->i32
, 16, 0),
2401 /* Upper 16 bits are ignored and will be dropped. */
2402 static LLVMValueRef
si_llvm_pack_two_int32_as_int16(struct si_shader_context
*ctx
,
2403 LLVMValueRef val
[2])
2405 LLVMValueRef v
[2] = {
2406 LLVMBuildAnd(ctx
->ac
.builder
, val
[0],
2407 LLVMConstInt(ctx
->i32
, 0xffff, 0), ""),
2410 return si_llvm_pack_two_int16(ctx
, v
);
2413 /* Initialize arguments for the shader export intrinsic */
2414 static void si_llvm_init_export_args(struct si_shader_context
*ctx
,
2415 LLVMValueRef
*values
,
2417 struct ac_export_args
*args
)
2419 LLVMValueRef f32undef
= LLVMGetUndef(ctx
->ac
.f32
);
2420 LLVMBuilderRef builder
= ctx
->ac
.builder
;
2421 LLVMValueRef val
[4];
2422 unsigned spi_shader_col_format
= V_028714_SPI_SHADER_32_ABGR
;
2424 bool is_int8
, is_int10
;
2426 /* Default is 0xf. Adjusted below depending on the format. */
2427 args
->enabled_channels
= 0xf; /* writemask */
2429 /* Specify whether the EXEC mask represents the valid mask */
2430 args
->valid_mask
= 0;
2432 /* Specify whether this is the last export */
2435 /* Specify the target we are exporting */
2436 args
->target
= target
;
2438 if (ctx
->type
== PIPE_SHADER_FRAGMENT
) {
2439 const struct si_shader_key
*key
= &ctx
->shader
->key
;
2440 unsigned col_formats
= key
->part
.ps
.epilog
.spi_shader_col_format
;
2441 int cbuf
= target
- V_008DFC_SQ_EXP_MRT
;
2443 assert(cbuf
>= 0 && cbuf
< 8);
2444 spi_shader_col_format
= (col_formats
>> (cbuf
* 4)) & 0xf;
2445 is_int8
= (key
->part
.ps
.epilog
.color_is_int8
>> cbuf
) & 0x1;
2446 is_int10
= (key
->part
.ps
.epilog
.color_is_int10
>> cbuf
) & 0x1;
2449 args
->compr
= false;
2450 args
->out
[0] = f32undef
;
2451 args
->out
[1] = f32undef
;
2452 args
->out
[2] = f32undef
;
2453 args
->out
[3] = f32undef
;
2455 switch (spi_shader_col_format
) {
2456 case V_028714_SPI_SHADER_ZERO
:
2457 args
->enabled_channels
= 0; /* writemask */
2458 args
->target
= V_008DFC_SQ_EXP_NULL
;
2461 case V_028714_SPI_SHADER_32_R
:
2462 args
->enabled_channels
= 1; /* writemask */
2463 args
->out
[0] = values
[0];
2466 case V_028714_SPI_SHADER_32_GR
:
2467 args
->enabled_channels
= 0x3; /* writemask */
2468 args
->out
[0] = values
[0];
2469 args
->out
[1] = values
[1];
2472 case V_028714_SPI_SHADER_32_AR
:
2473 args
->enabled_channels
= 0x9; /* writemask */
2474 args
->out
[0] = values
[0];
2475 args
->out
[3] = values
[3];
2478 case V_028714_SPI_SHADER_FP16_ABGR
:
2479 args
->compr
= 1; /* COMPR flag */
2481 for (chan
= 0; chan
< 2; chan
++) {
2482 LLVMValueRef pack_args
[2] = {
2484 values
[2 * chan
+ 1]
2486 LLVMValueRef packed
;
2488 packed
= ac_build_cvt_pkrtz_f16(&ctx
->ac
, pack_args
);
2489 args
->out
[chan
] = ac_to_float(&ctx
->ac
, packed
);
2493 case V_028714_SPI_SHADER_UNORM16_ABGR
:
2494 for (chan
= 0; chan
< 4; chan
++) {
2495 val
[chan
] = ac_build_clamp(&ctx
->ac
, values
[chan
]);
2496 val
[chan
] = LLVMBuildFMul(builder
, val
[chan
],
2497 LLVMConstReal(ctx
->f32
, 65535), "");
2498 val
[chan
] = LLVMBuildFAdd(builder
, val
[chan
],
2499 LLVMConstReal(ctx
->f32
, 0.5), "");
2500 val
[chan
] = LLVMBuildFPToUI(builder
, val
[chan
],
2504 args
->compr
= 1; /* COMPR flag */
2505 args
->out
[0] = ac_to_float(&ctx
->ac
, si_llvm_pack_two_int16(ctx
, val
));
2506 args
->out
[1] = ac_to_float(&ctx
->ac
, si_llvm_pack_two_int16(ctx
, val
+2));
2509 case V_028714_SPI_SHADER_SNORM16_ABGR
:
2510 for (chan
= 0; chan
< 4; chan
++) {
2511 /* Clamp between [-1, 1]. */
2512 val
[chan
] = lp_build_emit_llvm_binary(&ctx
->bld_base
, TGSI_OPCODE_MIN
,
2514 LLVMConstReal(ctx
->f32
, 1));
2515 val
[chan
] = lp_build_emit_llvm_binary(&ctx
->bld_base
, TGSI_OPCODE_MAX
,
2517 LLVMConstReal(ctx
->f32
, -1));
2518 /* Convert to a signed integer in [-32767, 32767]. */
2519 val
[chan
] = LLVMBuildFMul(builder
, val
[chan
],
2520 LLVMConstReal(ctx
->f32
, 32767), "");
2521 /* If positive, add 0.5, else add -0.5. */
2522 val
[chan
] = LLVMBuildFAdd(builder
, val
[chan
],
2523 LLVMBuildSelect(builder
,
2524 LLVMBuildFCmp(builder
, LLVMRealOGE
,
2525 val
[chan
], ctx
->ac
.f32_0
, ""),
2526 LLVMConstReal(ctx
->f32
, 0.5),
2527 LLVMConstReal(ctx
->f32
, -0.5), ""), "");
2528 val
[chan
] = LLVMBuildFPToSI(builder
, val
[chan
], ctx
->i32
, "");
2531 args
->compr
= 1; /* COMPR flag */
2532 args
->out
[0] = ac_to_float(&ctx
->ac
, si_llvm_pack_two_int32_as_int16(ctx
, val
));
2533 args
->out
[1] = ac_to_float(&ctx
->ac
, si_llvm_pack_two_int32_as_int16(ctx
, val
+2));
2536 case V_028714_SPI_SHADER_UINT16_ABGR
: {
2537 LLVMValueRef max_rgb
= LLVMConstInt(ctx
->i32
,
2538 is_int8
? 255 : is_int10
? 1023 : 65535, 0);
2539 LLVMValueRef max_alpha
=
2540 !is_int10
? max_rgb
: LLVMConstInt(ctx
->i32
, 3, 0);
2543 for (chan
= 0; chan
< 4; chan
++) {
2544 val
[chan
] = ac_to_integer(&ctx
->ac
, values
[chan
]);
2545 val
[chan
] = lp_build_emit_llvm_binary(&ctx
->bld_base
, TGSI_OPCODE_UMIN
,
2547 chan
== 3 ? max_alpha
: max_rgb
);
2550 args
->compr
= 1; /* COMPR flag */
2551 args
->out
[0] = ac_to_float(&ctx
->ac
, si_llvm_pack_two_int16(ctx
, val
));
2552 args
->out
[1] = ac_to_float(&ctx
->ac
, si_llvm_pack_two_int16(ctx
, val
+2));
2556 case V_028714_SPI_SHADER_SINT16_ABGR
: {
2557 LLVMValueRef max_rgb
= LLVMConstInt(ctx
->i32
,
2558 is_int8
? 127 : is_int10
? 511 : 32767, 0);
2559 LLVMValueRef min_rgb
= LLVMConstInt(ctx
->i32
,
2560 is_int8
? -128 : is_int10
? -512 : -32768, 0);
2561 LLVMValueRef max_alpha
=
2562 !is_int10
? max_rgb
: ctx
->i32_1
;
2563 LLVMValueRef min_alpha
=
2564 !is_int10
? min_rgb
: LLVMConstInt(ctx
->i32
, -2, 0);
2567 for (chan
= 0; chan
< 4; chan
++) {
2568 val
[chan
] = ac_to_integer(&ctx
->ac
, values
[chan
]);
2569 val
[chan
] = lp_build_emit_llvm_binary(&ctx
->bld_base
,
2571 val
[chan
], chan
== 3 ? max_alpha
: max_rgb
);
2572 val
[chan
] = lp_build_emit_llvm_binary(&ctx
->bld_base
,
2574 val
[chan
], chan
== 3 ? min_alpha
: min_rgb
);
2577 args
->compr
= 1; /* COMPR flag */
2578 args
->out
[0] = ac_to_float(&ctx
->ac
, si_llvm_pack_two_int32_as_int16(ctx
, val
));
2579 args
->out
[1] = ac_to_float(&ctx
->ac
, si_llvm_pack_two_int32_as_int16(ctx
, val
+2));
2583 case V_028714_SPI_SHADER_32_ABGR
:
2584 memcpy(&args
->out
[0], values
, sizeof(values
[0]) * 4);
2589 static void si_alpha_test(struct lp_build_tgsi_context
*bld_base
,
2592 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2594 if (ctx
->shader
->key
.part
.ps
.epilog
.alpha_func
!= PIPE_FUNC_NEVER
) {
2595 static LLVMRealPredicate cond_map
[PIPE_FUNC_ALWAYS
+ 1] = {
2596 [PIPE_FUNC_LESS
] = LLVMRealOLT
,
2597 [PIPE_FUNC_EQUAL
] = LLVMRealOEQ
,
2598 [PIPE_FUNC_LEQUAL
] = LLVMRealOLE
,
2599 [PIPE_FUNC_GREATER
] = LLVMRealOGT
,
2600 [PIPE_FUNC_NOTEQUAL
] = LLVMRealONE
,
2601 [PIPE_FUNC_GEQUAL
] = LLVMRealOGE
,
2603 LLVMRealPredicate cond
= cond_map
[ctx
->shader
->key
.part
.ps
.epilog
.alpha_func
];
2606 LLVMValueRef alpha_ref
= LLVMGetParam(ctx
->main_fn
,
2607 SI_PARAM_ALPHA_REF
);
2608 LLVMValueRef alpha_pass
=
2609 LLVMBuildFCmp(ctx
->ac
.builder
, cond
, alpha
, alpha_ref
, "");
2610 ac_build_kill_if_false(&ctx
->ac
, alpha_pass
);
2612 ac_build_kill_if_false(&ctx
->ac
, LLVMConstInt(ctx
->i1
, 0, 0));
2616 static LLVMValueRef
si_scale_alpha_by_sample_mask(struct lp_build_tgsi_context
*bld_base
,
2618 unsigned samplemask_param
)
2620 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2621 LLVMValueRef coverage
;
2623 /* alpha = alpha * popcount(coverage) / SI_NUM_SMOOTH_AA_SAMPLES */
2624 coverage
= LLVMGetParam(ctx
->main_fn
,
2626 coverage
= ac_to_integer(&ctx
->ac
, coverage
);
2628 coverage
= lp_build_intrinsic(ctx
->ac
.builder
, "llvm.ctpop.i32",
2630 &coverage
, 1, LP_FUNC_ATTR_READNONE
);
2632 coverage
= LLVMBuildUIToFP(ctx
->ac
.builder
, coverage
,
2635 coverage
= LLVMBuildFMul(ctx
->ac
.builder
, coverage
,
2636 LLVMConstReal(ctx
->f32
,
2637 1.0 / SI_NUM_SMOOTH_AA_SAMPLES
), "");
2639 return LLVMBuildFMul(ctx
->ac
.builder
, alpha
, coverage
, "");
2642 static void si_llvm_emit_clipvertex(struct si_shader_context
*ctx
,
2643 struct ac_export_args
*pos
, LLVMValueRef
*out_elts
)
2647 unsigned const_chan
;
2648 LLVMValueRef base_elt
;
2649 LLVMValueRef ptr
= LLVMGetParam(ctx
->main_fn
, ctx
->param_rw_buffers
);
2650 LLVMValueRef constbuf_index
= LLVMConstInt(ctx
->i32
,
2651 SI_VS_CONST_CLIP_PLANES
, 0);
2652 LLVMValueRef const_resource
= ac_build_load_to_sgpr(&ctx
->ac
, ptr
, constbuf_index
);
2654 for (reg_index
= 0; reg_index
< 2; reg_index
++) {
2655 struct ac_export_args
*args
= &pos
[2 + reg_index
];
2660 args
->out
[3] = LLVMConstReal(ctx
->f32
, 0.0f
);
2662 /* Compute dot products of position and user clip plane vectors */
2663 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
2664 for (const_chan
= 0; const_chan
< TGSI_NUM_CHANNELS
; const_chan
++) {
2666 LLVMConstInt(ctx
->i32
, ((reg_index
* 4 + chan
) * 4 +
2667 const_chan
) * 4, 0);
2668 base_elt
= buffer_load_const(ctx
, const_resource
,
2671 lp_build_add(&ctx
->bld_base
.base
, args
->out
[chan
],
2672 lp_build_mul(&ctx
->bld_base
.base
, base_elt
,
2673 out_elts
[const_chan
]));
2677 args
->enabled_channels
= 0xf;
2678 args
->valid_mask
= 0;
2680 args
->target
= V_008DFC_SQ_EXP_POS
+ 2 + reg_index
;
2685 static void si_dump_streamout(struct pipe_stream_output_info
*so
)
2689 if (so
->num_outputs
)
2690 fprintf(stderr
, "STREAMOUT\n");
2692 for (i
= 0; i
< so
->num_outputs
; i
++) {
2693 unsigned mask
= ((1 << so
->output
[i
].num_components
) - 1) <<
2694 so
->output
[i
].start_component
;
2695 fprintf(stderr
, " %i: BUF%i[%i..%i] <- OUT[%i].%s%s%s%s\n",
2696 i
, so
->output
[i
].output_buffer
,
2697 so
->output
[i
].dst_offset
, so
->output
[i
].dst_offset
+ so
->output
[i
].num_components
- 1,
2698 so
->output
[i
].register_index
,
2699 mask
& 1 ? "x" : "",
2700 mask
& 2 ? "y" : "",
2701 mask
& 4 ? "z" : "",
2702 mask
& 8 ? "w" : "");
2706 static void emit_streamout_output(struct si_shader_context
*ctx
,
2707 LLVMValueRef
const *so_buffers
,
2708 LLVMValueRef
const *so_write_offsets
,
2709 struct pipe_stream_output
*stream_out
,
2710 struct si_shader_output_values
*shader_out
)
2712 unsigned buf_idx
= stream_out
->output_buffer
;
2713 unsigned start
= stream_out
->start_component
;
2714 unsigned num_comps
= stream_out
->num_components
;
2715 LLVMValueRef out
[4];
2717 assert(num_comps
&& num_comps
<= 4);
2718 if (!num_comps
|| num_comps
> 4)
2721 /* Load the output as int. */
2722 for (int j
= 0; j
< num_comps
; j
++) {
2723 assert(stream_out
->stream
== shader_out
->vertex_stream
[start
+ j
]);
2725 out
[j
] = ac_to_integer(&ctx
->ac
, shader_out
->values
[start
+ j
]);
2728 /* Pack the output. */
2729 LLVMValueRef vdata
= NULL
;
2731 switch (num_comps
) {
2732 case 1: /* as i32 */
2735 case 2: /* as v2i32 */
2736 case 3: /* as v4i32 (aligned to 4) */
2737 case 4: /* as v4i32 */
2738 vdata
= LLVMGetUndef(LLVMVectorType(ctx
->i32
, util_next_power_of_two(num_comps
)));
2739 for (int j
= 0; j
< num_comps
; j
++) {
2740 vdata
= LLVMBuildInsertElement(ctx
->ac
.builder
, vdata
, out
[j
],
2741 LLVMConstInt(ctx
->i32
, j
, 0), "");
2746 ac_build_buffer_store_dword(&ctx
->ac
, so_buffers
[buf_idx
],
2748 so_write_offsets
[buf_idx
],
2750 stream_out
->dst_offset
* 4, 1, 1, true, false);
2754 * Write streamout data to buffers for vertex stream @p stream (different
2755 * vertex streams can occur for GS copy shaders).
2757 static void si_llvm_emit_streamout(struct si_shader_context
*ctx
,
2758 struct si_shader_output_values
*outputs
,
2759 unsigned noutput
, unsigned stream
)
2761 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
2762 struct pipe_stream_output_info
*so
= &sel
->so
;
2763 LLVMBuilderRef builder
= ctx
->ac
.builder
;
2765 struct lp_build_if_state if_ctx
;
2767 /* Get bits [22:16], i.e. (so_param >> 16) & 127; */
2768 LLVMValueRef so_vtx_count
=
2769 unpack_param(ctx
, ctx
->param_streamout_config
, 16, 7);
2771 LLVMValueRef tid
= ac_get_thread_id(&ctx
->ac
);
2773 /* can_emit = tid < so_vtx_count; */
2774 LLVMValueRef can_emit
=
2775 LLVMBuildICmp(builder
, LLVMIntULT
, tid
, so_vtx_count
, "");
2777 /* Emit the streamout code conditionally. This actually avoids
2778 * out-of-bounds buffer access. The hw tells us via the SGPR
2779 * (so_vtx_count) which threads are allowed to emit streamout data. */
2780 lp_build_if(&if_ctx
, &ctx
->gallivm
, can_emit
);
2782 /* The buffer offset is computed as follows:
2783 * ByteOffset = streamout_offset[buffer_id]*4 +
2784 * (streamout_write_index + thread_id)*stride[buffer_id] +
2788 LLVMValueRef so_write_index
=
2789 LLVMGetParam(ctx
->main_fn
,
2790 ctx
->param_streamout_write_index
);
2792 /* Compute (streamout_write_index + thread_id). */
2793 so_write_index
= LLVMBuildAdd(builder
, so_write_index
, tid
, "");
2795 /* Load the descriptor and compute the write offset for each
2796 * enabled buffer. */
2797 LLVMValueRef so_write_offset
[4] = {};
2798 LLVMValueRef so_buffers
[4];
2799 LLVMValueRef buf_ptr
= LLVMGetParam(ctx
->main_fn
,
2800 ctx
->param_rw_buffers
);
2802 for (i
= 0; i
< 4; i
++) {
2806 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
,
2807 SI_VS_STREAMOUT_BUF0
+ i
, 0);
2809 so_buffers
[i
] = ac_build_load_to_sgpr(&ctx
->ac
, buf_ptr
, offset
);
2811 LLVMValueRef so_offset
= LLVMGetParam(ctx
->main_fn
,
2812 ctx
->param_streamout_offset
[i
]);
2813 so_offset
= LLVMBuildMul(builder
, so_offset
, LLVMConstInt(ctx
->i32
, 4, 0), "");
2815 so_write_offset
[i
] = LLVMBuildMul(builder
, so_write_index
,
2816 LLVMConstInt(ctx
->i32
, so
->stride
[i
]*4, 0), "");
2817 so_write_offset
[i
] = LLVMBuildAdd(builder
, so_write_offset
[i
], so_offset
, "");
2820 /* Write streamout data. */
2821 for (i
= 0; i
< so
->num_outputs
; i
++) {
2822 unsigned reg
= so
->output
[i
].register_index
;
2827 if (stream
!= so
->output
[i
].stream
)
2830 emit_streamout_output(ctx
, so_buffers
, so_write_offset
,
2831 &so
->output
[i
], &outputs
[reg
]);
2834 lp_build_endif(&if_ctx
);
2837 static void si_export_param(struct si_shader_context
*ctx
, unsigned index
,
2838 LLVMValueRef
*values
)
2840 struct ac_export_args args
;
2842 si_llvm_init_export_args(ctx
, values
,
2843 V_008DFC_SQ_EXP_PARAM
+ index
, &args
);
2844 ac_build_export(&ctx
->ac
, &args
);
2847 static void si_build_param_exports(struct si_shader_context
*ctx
,
2848 struct si_shader_output_values
*outputs
,
2851 struct si_shader
*shader
= ctx
->shader
;
2852 unsigned param_count
= 0;
2854 for (unsigned i
= 0; i
< noutput
; i
++) {
2855 unsigned semantic_name
= outputs
[i
].semantic_name
;
2856 unsigned semantic_index
= outputs
[i
].semantic_index
;
2858 if (outputs
[i
].vertex_stream
[0] != 0 &&
2859 outputs
[i
].vertex_stream
[1] != 0 &&
2860 outputs
[i
].vertex_stream
[2] != 0 &&
2861 outputs
[i
].vertex_stream
[3] != 0)
2864 switch (semantic_name
) {
2865 case TGSI_SEMANTIC_LAYER
:
2866 case TGSI_SEMANTIC_VIEWPORT_INDEX
:
2867 case TGSI_SEMANTIC_CLIPDIST
:
2868 case TGSI_SEMANTIC_COLOR
:
2869 case TGSI_SEMANTIC_BCOLOR
:
2870 case TGSI_SEMANTIC_PRIMID
:
2871 case TGSI_SEMANTIC_FOG
:
2872 case TGSI_SEMANTIC_TEXCOORD
:
2873 case TGSI_SEMANTIC_GENERIC
:
2879 if ((semantic_name
!= TGSI_SEMANTIC_GENERIC
||
2880 semantic_index
< SI_MAX_IO_GENERIC
) &&
2881 shader
->key
.opt
.kill_outputs
&
2882 (1ull << si_shader_io_get_unique_index(semantic_name
, semantic_index
)))
2885 si_export_param(ctx
, param_count
, outputs
[i
].values
);
2887 assert(i
< ARRAY_SIZE(shader
->info
.vs_output_param_offset
));
2888 shader
->info
.vs_output_param_offset
[i
] = param_count
++;
2891 shader
->info
.nr_param_exports
= param_count
;
2894 /* Generate export instructions for hardware VS shader stage */
2895 static void si_llvm_export_vs(struct si_shader_context
*ctx
,
2896 struct si_shader_output_values
*outputs
,
2899 struct si_shader
*shader
= ctx
->shader
;
2900 struct ac_export_args pos_args
[4] = {};
2901 LLVMValueRef psize_value
= NULL
, edgeflag_value
= NULL
, layer_value
= NULL
, viewport_index_value
= NULL
;
2905 /* Build position exports. */
2906 for (i
= 0; i
< noutput
; i
++) {
2907 switch (outputs
[i
].semantic_name
) {
2908 case TGSI_SEMANTIC_POSITION
:
2909 si_llvm_init_export_args(ctx
, outputs
[i
].values
,
2910 V_008DFC_SQ_EXP_POS
, &pos_args
[0]);
2912 case TGSI_SEMANTIC_PSIZE
:
2913 psize_value
= outputs
[i
].values
[0];
2915 case TGSI_SEMANTIC_LAYER
:
2916 layer_value
= outputs
[i
].values
[0];
2918 case TGSI_SEMANTIC_VIEWPORT_INDEX
:
2919 viewport_index_value
= outputs
[i
].values
[0];
2921 case TGSI_SEMANTIC_EDGEFLAG
:
2922 edgeflag_value
= outputs
[i
].values
[0];
2924 case TGSI_SEMANTIC_CLIPDIST
:
2925 if (!shader
->key
.opt
.clip_disable
) {
2926 unsigned index
= 2 + outputs
[i
].semantic_index
;
2927 si_llvm_init_export_args(ctx
, outputs
[i
].values
,
2928 V_008DFC_SQ_EXP_POS
+ index
,
2932 case TGSI_SEMANTIC_CLIPVERTEX
:
2933 if (!shader
->key
.opt
.clip_disable
) {
2934 si_llvm_emit_clipvertex(ctx
, pos_args
,
2941 /* We need to add the position output manually if it's missing. */
2942 if (!pos_args
[0].out
[0]) {
2943 pos_args
[0].enabled_channels
= 0xf; /* writemask */
2944 pos_args
[0].valid_mask
= 0; /* EXEC mask */
2945 pos_args
[0].done
= 0; /* last export? */
2946 pos_args
[0].target
= V_008DFC_SQ_EXP_POS
;
2947 pos_args
[0].compr
= 0; /* COMPR flag */
2948 pos_args
[0].out
[0] = ctx
->ac
.f32_0
; /* X */
2949 pos_args
[0].out
[1] = ctx
->ac
.f32_0
; /* Y */
2950 pos_args
[0].out
[2] = ctx
->ac
.f32_0
; /* Z */
2951 pos_args
[0].out
[3] = ctx
->ac
.f32_1
; /* W */
2954 /* Write the misc vector (point size, edgeflag, layer, viewport). */
2955 if (shader
->selector
->info
.writes_psize
||
2956 shader
->selector
->info
.writes_edgeflag
||
2957 shader
->selector
->info
.writes_viewport_index
||
2958 shader
->selector
->info
.writes_layer
) {
2959 pos_args
[1].enabled_channels
= shader
->selector
->info
.writes_psize
|
2960 (shader
->selector
->info
.writes_edgeflag
<< 1) |
2961 (shader
->selector
->info
.writes_layer
<< 2);
2963 pos_args
[1].valid_mask
= 0; /* EXEC mask */
2964 pos_args
[1].done
= 0; /* last export? */
2965 pos_args
[1].target
= V_008DFC_SQ_EXP_POS
+ 1;
2966 pos_args
[1].compr
= 0; /* COMPR flag */
2967 pos_args
[1].out
[0] = ctx
->ac
.f32_0
; /* X */
2968 pos_args
[1].out
[1] = ctx
->ac
.f32_0
; /* Y */
2969 pos_args
[1].out
[2] = ctx
->ac
.f32_0
; /* Z */
2970 pos_args
[1].out
[3] = ctx
->ac
.f32_0
; /* W */
2972 if (shader
->selector
->info
.writes_psize
)
2973 pos_args
[1].out
[0] = psize_value
;
2975 if (shader
->selector
->info
.writes_edgeflag
) {
2976 /* The output is a float, but the hw expects an integer
2977 * with the first bit containing the edge flag. */
2978 edgeflag_value
= LLVMBuildFPToUI(ctx
->ac
.builder
,
2981 edgeflag_value
= ac_build_umin(&ctx
->ac
,
2985 /* The LLVM intrinsic expects a float. */
2986 pos_args
[1].out
[1] = ac_to_float(&ctx
->ac
, edgeflag_value
);
2989 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
2990 /* GFX9 has the layer in out.z[10:0] and the viewport
2991 * index in out.z[19:16].
2993 if (shader
->selector
->info
.writes_layer
)
2994 pos_args
[1].out
[2] = layer_value
;
2996 if (shader
->selector
->info
.writes_viewport_index
) {
2997 LLVMValueRef v
= viewport_index_value
;
2999 v
= ac_to_integer(&ctx
->ac
, v
);
3000 v
= LLVMBuildShl(ctx
->ac
.builder
, v
,
3001 LLVMConstInt(ctx
->i32
, 16, 0), "");
3002 v
= LLVMBuildOr(ctx
->ac
.builder
, v
,
3003 ac_to_integer(&ctx
->ac
, pos_args
[1].out
[2]), "");
3004 pos_args
[1].out
[2] = ac_to_float(&ctx
->ac
, v
);
3005 pos_args
[1].enabled_channels
|= 1 << 2;
3008 if (shader
->selector
->info
.writes_layer
)
3009 pos_args
[1].out
[2] = layer_value
;
3011 if (shader
->selector
->info
.writes_viewport_index
) {
3012 pos_args
[1].out
[3] = viewport_index_value
;
3013 pos_args
[1].enabled_channels
|= 1 << 3;
3018 for (i
= 0; i
< 4; i
++)
3019 if (pos_args
[i
].out
[0])
3020 shader
->info
.nr_pos_exports
++;
3023 for (i
= 0; i
< 4; i
++) {
3024 if (!pos_args
[i
].out
[0])
3027 /* Specify the target we are exporting */
3028 pos_args
[i
].target
= V_008DFC_SQ_EXP_POS
+ pos_idx
++;
3030 if (pos_idx
== shader
->info
.nr_pos_exports
)
3031 /* Specify that this is the last export */
3032 pos_args
[i
].done
= 1;
3034 ac_build_export(&ctx
->ac
, &pos_args
[i
]);
3037 /* Build parameter exports. */
3038 si_build_param_exports(ctx
, outputs
, noutput
);
3042 * Forward all outputs from the vertex shader to the TES. This is only used
3043 * for the fixed function TCS.
3045 static void si_copy_tcs_inputs(struct lp_build_tgsi_context
*bld_base
)
3047 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3048 LLVMValueRef invocation_id
, buffer
, buffer_offset
;
3049 LLVMValueRef lds_vertex_stride
, lds_vertex_offset
, lds_base
;
3052 invocation_id
= unpack_llvm_param(ctx
, ctx
->abi
.tcs_rel_ids
, 8, 5);
3053 buffer
= desc_from_addr_base64k(ctx
, ctx
->param_tcs_offchip_addr_base64k
);
3054 buffer_offset
= LLVMGetParam(ctx
->main_fn
, ctx
->param_tcs_offchip_offset
);
3056 lds_vertex_stride
= get_tcs_in_vertex_dw_stride(ctx
);
3057 lds_vertex_offset
= LLVMBuildMul(ctx
->ac
.builder
, invocation_id
,
3058 lds_vertex_stride
, "");
3059 lds_base
= get_tcs_in_current_patch_offset(ctx
);
3060 lds_base
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
, lds_vertex_offset
, "");
3062 inputs
= ctx
->shader
->key
.mono
.u
.ff_tcs_inputs_to_copy
;
3064 unsigned i
= u_bit_scan64(&inputs
);
3066 LLVMValueRef lds_ptr
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
3067 LLVMConstInt(ctx
->i32
, 4 * i
, 0),
3070 LLVMValueRef buffer_addr
= get_tcs_tes_buffer_address(ctx
,
3071 get_rel_patch_id(ctx
),
3073 LLVMConstInt(ctx
->i32
, i
, 0));
3075 LLVMValueRef value
= lds_load(bld_base
, ctx
->ac
.i32
, ~0,
3078 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, value
, 4, buffer_addr
,
3079 buffer_offset
, 0, 1, 0, true, false);
3083 static void si_write_tess_factors(struct lp_build_tgsi_context
*bld_base
,
3084 LLVMValueRef rel_patch_id
,
3085 LLVMValueRef invocation_id
,
3086 LLVMValueRef tcs_out_current_patch_data_offset
,
3087 LLVMValueRef invoc0_tf_outer
[4],
3088 LLVMValueRef invoc0_tf_inner
[2])
3090 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3091 struct si_shader
*shader
= ctx
->shader
;
3092 unsigned tess_inner_index
, tess_outer_index
;
3093 LLVMValueRef lds_base
, lds_inner
, lds_outer
, byteoffset
, buffer
;
3094 LLVMValueRef out
[6], vec0
, vec1
, tf_base
, inner
[4], outer
[4];
3095 unsigned stride
, outer_comps
, inner_comps
, i
, offset
;
3096 struct lp_build_if_state if_ctx
, inner_if_ctx
;
3098 /* Add a barrier before loading tess factors from LDS. */
3099 if (!shader
->key
.part
.tcs
.epilog
.invoc0_tess_factors_are_def
)
3100 si_llvm_emit_barrier(NULL
, bld_base
, NULL
);
3102 /* Do this only for invocation 0, because the tess levels are per-patch,
3105 * This can't jump, because invocation 0 executes this. It should
3106 * at least mask out the loads and stores for other invocations.
3108 lp_build_if(&if_ctx
, &ctx
->gallivm
,
3109 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
,
3110 invocation_id
, ctx
->i32_0
, ""));
3112 /* Determine the layout of one tess factor element in the buffer. */
3113 switch (shader
->key
.part
.tcs
.epilog
.prim_mode
) {
3114 case PIPE_PRIM_LINES
:
3115 stride
= 2; /* 2 dwords, 1 vec2 store */
3119 case PIPE_PRIM_TRIANGLES
:
3120 stride
= 4; /* 4 dwords, 1 vec4 store */
3124 case PIPE_PRIM_QUADS
:
3125 stride
= 6; /* 6 dwords, 2 stores (vec4 + vec2) */
3134 for (i
= 0; i
< 4; i
++) {
3135 inner
[i
] = LLVMGetUndef(ctx
->i32
);
3136 outer
[i
] = LLVMGetUndef(ctx
->i32
);
3139 if (shader
->key
.part
.tcs
.epilog
.invoc0_tess_factors_are_def
) {
3140 /* Tess factors are in VGPRs. */
3141 for (i
= 0; i
< outer_comps
; i
++)
3142 outer
[i
] = out
[i
] = invoc0_tf_outer
[i
];
3143 for (i
= 0; i
< inner_comps
; i
++)
3144 inner
[i
] = out
[outer_comps
+i
] = invoc0_tf_inner
[i
];
3146 /* Load tess_inner and tess_outer from LDS.
3147 * Any invocation can write them, so we can't get them from a temporary.
3149 tess_inner_index
= si_shader_io_get_unique_index_patch(TGSI_SEMANTIC_TESSINNER
, 0);
3150 tess_outer_index
= si_shader_io_get_unique_index_patch(TGSI_SEMANTIC_TESSOUTER
, 0);
3152 lds_base
= tcs_out_current_patch_data_offset
;
3153 lds_inner
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
3154 LLVMConstInt(ctx
->i32
,
3155 tess_inner_index
* 4, 0), "");
3156 lds_outer
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
3157 LLVMConstInt(ctx
->i32
,
3158 tess_outer_index
* 4, 0), "");
3160 for (i
= 0; i
< outer_comps
; i
++) {
3162 lds_load(bld_base
, ctx
->ac
.i32
, i
, lds_outer
);
3164 for (i
= 0; i
< inner_comps
; i
++) {
3165 inner
[i
] = out
[outer_comps
+i
] =
3166 lds_load(bld_base
, ctx
->ac
.i32
, i
, lds_inner
);
3170 if (shader
->key
.part
.tcs
.epilog
.prim_mode
== PIPE_PRIM_LINES
) {
3171 /* For isolines, the hardware expects tess factors in the
3172 * reverse order from what GLSL / TGSI specify.
3174 LLVMValueRef tmp
= out
[0];
3179 /* Convert the outputs to vectors for stores. */
3180 vec0
= lp_build_gather_values(&ctx
->gallivm
, out
, MIN2(stride
, 4));
3184 vec1
= lp_build_gather_values(&ctx
->gallivm
, out
+4, stride
- 4);
3186 /* Get the buffer. */
3187 buffer
= desc_from_addr_base64k(ctx
, ctx
->param_tcs_factor_addr_base64k
);
3189 /* Get the offset. */
3190 tf_base
= LLVMGetParam(ctx
->main_fn
,
3191 ctx
->param_tcs_factor_offset
);
3192 byteoffset
= LLVMBuildMul(ctx
->ac
.builder
, rel_patch_id
,
3193 LLVMConstInt(ctx
->i32
, 4 * stride
, 0), "");
3195 lp_build_if(&inner_if_ctx
, &ctx
->gallivm
,
3196 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
,
3197 rel_patch_id
, ctx
->i32_0
, ""));
3199 /* Store the dynamic HS control word. */
3201 if (ctx
->screen
->info
.chip_class
<= VI
) {
3202 ac_build_buffer_store_dword(&ctx
->ac
, buffer
,
3203 LLVMConstInt(ctx
->i32
, 0x80000000, 0),
3204 1, ctx
->i32_0
, tf_base
,
3205 offset
, 1, 0, true, false);
3209 lp_build_endif(&inner_if_ctx
);
3211 /* Store the tessellation factors. */
3212 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec0
,
3213 MIN2(stride
, 4), byteoffset
, tf_base
,
3214 offset
, 1, 0, true, false);
3217 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec1
,
3218 stride
- 4, byteoffset
, tf_base
,
3219 offset
, 1, 0, true, false);
3221 /* Store the tess factors into the offchip buffer if TES reads them. */
3222 if (shader
->key
.part
.tcs
.epilog
.tes_reads_tess_factors
) {
3223 LLVMValueRef buf
, base
, inner_vec
, outer_vec
, tf_outer_offset
;
3224 LLVMValueRef tf_inner_offset
;
3225 unsigned param_outer
, param_inner
;
3227 buf
= desc_from_addr_base64k(ctx
, ctx
->param_tcs_offchip_addr_base64k
);
3228 base
= LLVMGetParam(ctx
->main_fn
, ctx
->param_tcs_offchip_offset
);
3230 param_outer
= si_shader_io_get_unique_index_patch(
3231 TGSI_SEMANTIC_TESSOUTER
, 0);
3232 tf_outer_offset
= get_tcs_tes_buffer_address(ctx
, rel_patch_id
, NULL
,
3233 LLVMConstInt(ctx
->i32
, param_outer
, 0));
3235 outer_vec
= lp_build_gather_values(&ctx
->gallivm
, outer
,
3236 util_next_power_of_two(outer_comps
));
3238 ac_build_buffer_store_dword(&ctx
->ac
, buf
, outer_vec
,
3239 outer_comps
, tf_outer_offset
,
3240 base
, 0, 1, 0, true, false);
3242 param_inner
= si_shader_io_get_unique_index_patch(
3243 TGSI_SEMANTIC_TESSINNER
, 0);
3244 tf_inner_offset
= get_tcs_tes_buffer_address(ctx
, rel_patch_id
, NULL
,
3245 LLVMConstInt(ctx
->i32
, param_inner
, 0));
3247 inner_vec
= inner_comps
== 1 ? inner
[0] :
3248 lp_build_gather_values(&ctx
->gallivm
, inner
, inner_comps
);
3249 ac_build_buffer_store_dword(&ctx
->ac
, buf
, inner_vec
,
3250 inner_comps
, tf_inner_offset
,
3251 base
, 0, 1, 0, true, false);
3255 lp_build_endif(&if_ctx
);
3259 si_insert_input_ret(struct si_shader_context
*ctx
, LLVMValueRef ret
,
3260 unsigned param
, unsigned return_index
)
3262 return LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
3263 LLVMGetParam(ctx
->main_fn
, param
),
3268 si_insert_input_ret_float(struct si_shader_context
*ctx
, LLVMValueRef ret
,
3269 unsigned param
, unsigned return_index
)
3271 LLVMBuilderRef builder
= ctx
->ac
.builder
;
3272 LLVMValueRef p
= LLVMGetParam(ctx
->main_fn
, param
);
3274 return LLVMBuildInsertValue(builder
, ret
,
3275 ac_to_float(&ctx
->ac
, p
),
3280 si_insert_input_ptr_as_2xi32(struct si_shader_context
*ctx
, LLVMValueRef ret
,
3281 unsigned param
, unsigned return_index
)
3283 LLVMBuilderRef builder
= ctx
->ac
.builder
;
3284 LLVMValueRef ptr
, lo
, hi
;
3286 ptr
= LLVMGetParam(ctx
->main_fn
, param
);
3287 ptr
= LLVMBuildPtrToInt(builder
, ptr
, ctx
->i64
, "");
3288 ptr
= LLVMBuildBitCast(builder
, ptr
, ctx
->v2i32
, "");
3289 lo
= LLVMBuildExtractElement(builder
, ptr
, ctx
->i32_0
, "");
3290 hi
= LLVMBuildExtractElement(builder
, ptr
, ctx
->i32_1
, "");
3291 ret
= LLVMBuildInsertValue(builder
, ret
, lo
, return_index
, "");
3292 return LLVMBuildInsertValue(builder
, ret
, hi
, return_index
+ 1, "");
3295 /* This only writes the tessellation factor levels. */
3296 static void si_llvm_emit_tcs_epilogue(struct ac_shader_abi
*abi
,
3297 unsigned max_outputs
,
3298 LLVMValueRef
*addrs
)
3300 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
3301 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
3302 LLVMBuilderRef builder
= ctx
->ac
.builder
;
3303 LLVMValueRef rel_patch_id
, invocation_id
, tf_lds_offset
;
3305 si_copy_tcs_inputs(bld_base
);
3307 rel_patch_id
= get_rel_patch_id(ctx
);
3308 invocation_id
= unpack_llvm_param(ctx
, ctx
->abi
.tcs_rel_ids
, 8, 5);
3309 tf_lds_offset
= get_tcs_out_current_patch_data_offset(ctx
);
3311 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
3312 LLVMBasicBlockRef blocks
[2] = {
3313 LLVMGetInsertBlock(builder
),
3314 ctx
->merged_wrap_if_state
.entry_block
3316 LLVMValueRef values
[2];
3318 lp_build_endif(&ctx
->merged_wrap_if_state
);
3320 values
[0] = rel_patch_id
;
3321 values
[1] = LLVMGetUndef(ctx
->i32
);
3322 rel_patch_id
= ac_build_phi(&ctx
->ac
, ctx
->i32
, 2, values
, blocks
);
3324 values
[0] = tf_lds_offset
;
3325 values
[1] = LLVMGetUndef(ctx
->i32
);
3326 tf_lds_offset
= ac_build_phi(&ctx
->ac
, ctx
->i32
, 2, values
, blocks
);
3328 values
[0] = invocation_id
;
3329 values
[1] = ctx
->i32_1
; /* cause the epilog to skip threads */
3330 invocation_id
= ac_build_phi(&ctx
->ac
, ctx
->i32
, 2, values
, blocks
);
3333 /* Return epilog parameters from this function. */
3334 LLVMValueRef ret
= ctx
->return_value
;
3337 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
3338 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_layout
,
3339 8 + GFX9_SGPR_TCS_OFFCHIP_LAYOUT
);
3340 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_addr_base64k
,
3341 8 + GFX9_SGPR_TCS_OFFCHIP_ADDR_BASE64K
);
3342 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_factor_addr_base64k
,
3343 8 + GFX9_SGPR_TCS_FACTOR_ADDR_BASE64K
);
3344 /* Tess offchip and tess factor offsets are at the beginning. */
3345 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_offset
, 2);
3346 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_factor_offset
, 4);
3347 vgpr
= 8 + GFX9_SGPR_TCS_FACTOR_ADDR_BASE64K
+ 1;
3349 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_layout
,
3350 GFX6_SGPR_TCS_OFFCHIP_LAYOUT
);
3351 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_addr_base64k
,
3352 GFX6_SGPR_TCS_OFFCHIP_ADDR_BASE64K
);
3353 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_factor_addr_base64k
,
3354 GFX6_SGPR_TCS_FACTOR_ADDR_BASE64K
);
3355 /* Tess offchip and tess factor offsets are after user SGPRs. */
3356 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_offset
,
3357 GFX6_TCS_NUM_USER_SGPR
);
3358 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_factor_offset
,
3359 GFX6_TCS_NUM_USER_SGPR
+ 1);
3360 vgpr
= GFX6_TCS_NUM_USER_SGPR
+ 2;
3364 rel_patch_id
= ac_to_float(&ctx
->ac
, rel_patch_id
);
3365 invocation_id
= ac_to_float(&ctx
->ac
, invocation_id
);
3366 tf_lds_offset
= ac_to_float(&ctx
->ac
, tf_lds_offset
);
3368 /* Leave a hole corresponding to the two input VGPRs. This ensures that
3369 * the invocation_id output does not alias the tcs_rel_ids input,
3370 * which saves a V_MOV on gfx9.
3374 ret
= LLVMBuildInsertValue(builder
, ret
, rel_patch_id
, vgpr
++, "");
3375 ret
= LLVMBuildInsertValue(builder
, ret
, invocation_id
, vgpr
++, "");
3377 if (ctx
->shader
->selector
->tcs_info
.tessfactors_are_def_in_all_invocs
) {
3378 vgpr
++; /* skip the tess factor LDS offset */
3379 for (unsigned i
= 0; i
< 6; i
++) {
3380 LLVMValueRef value
=
3381 LLVMBuildLoad(builder
, ctx
->invoc0_tess_factors
[i
], "");
3382 value
= ac_to_float(&ctx
->ac
, value
);
3383 ret
= LLVMBuildInsertValue(builder
, ret
, value
, vgpr
++, "");
3386 ret
= LLVMBuildInsertValue(builder
, ret
, tf_lds_offset
, vgpr
++, "");
3388 ctx
->return_value
= ret
;
3391 /* Pass TCS inputs from LS to TCS on GFX9. */
3392 static void si_set_ls_return_value_for_tcs(struct si_shader_context
*ctx
)
3394 LLVMValueRef ret
= ctx
->return_value
;
3396 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_offset
, 2);
3397 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_merged_wave_info
, 3);
3398 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_factor_offset
, 4);
3399 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_merged_scratch_offset
, 5);
3401 ret
= si_insert_input_ptr_as_2xi32(ctx
, ret
, ctx
->param_rw_buffers
,
3402 8 + SI_SGPR_RW_BUFFERS
);
3403 ret
= si_insert_input_ptr_as_2xi32(ctx
, ret
,
3404 ctx
->param_bindless_samplers_and_images
,
3405 8 + SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES
);
3407 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_vs_state_bits
,
3408 8 + SI_SGPR_VS_STATE_BITS
);
3409 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_layout
,
3410 8 + GFX9_SGPR_TCS_OFFCHIP_LAYOUT
);
3411 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_out_lds_offsets
,
3412 8 + GFX9_SGPR_TCS_OUT_OFFSETS
);
3413 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_out_lds_layout
,
3414 8 + GFX9_SGPR_TCS_OUT_LAYOUT
);
3415 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_offchip_addr_base64k
,
3416 8 + GFX9_SGPR_TCS_OFFCHIP_ADDR_BASE64K
);
3417 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_tcs_factor_addr_base64k
,
3418 8 + GFX9_SGPR_TCS_FACTOR_ADDR_BASE64K
);
3420 unsigned desc_param
= ctx
->param_tcs_factor_addr_base64k
+ 2;
3421 ret
= si_insert_input_ptr_as_2xi32(ctx
, ret
, desc_param
,
3422 8 + GFX9_SGPR_TCS_CONST_AND_SHADER_BUFFERS
);
3423 ret
= si_insert_input_ptr_as_2xi32(ctx
, ret
, desc_param
+ 1,
3424 8 + GFX9_SGPR_TCS_SAMPLERS_AND_IMAGES
);
3426 unsigned vgpr
= 8 + GFX9_TCS_NUM_USER_SGPR
;
3427 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
3428 ac_to_float(&ctx
->ac
, ctx
->abi
.tcs_patch_id
),
3430 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
3431 ac_to_float(&ctx
->ac
, ctx
->abi
.tcs_rel_ids
),
3433 ctx
->return_value
= ret
;
3436 /* Pass GS inputs from ES to GS on GFX9. */
3437 static void si_set_es_return_value_for_gs(struct si_shader_context
*ctx
)
3439 LLVMValueRef ret
= ctx
->return_value
;
3441 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_gs2vs_offset
, 2);
3442 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_merged_wave_info
, 3);
3443 ret
= si_insert_input_ret(ctx
, ret
, ctx
->param_merged_scratch_offset
, 5);
3445 ret
= si_insert_input_ptr_as_2xi32(ctx
, ret
, ctx
->param_rw_buffers
,
3446 8 + SI_SGPR_RW_BUFFERS
);
3447 ret
= si_insert_input_ptr_as_2xi32(ctx
, ret
,
3448 ctx
->param_bindless_samplers_and_images
,
3449 8 + SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES
);
3451 unsigned desc_param
= ctx
->param_vs_state_bits
+ 1;
3452 ret
= si_insert_input_ptr_as_2xi32(ctx
, ret
, desc_param
,
3453 8 + GFX9_SGPR_GS_CONST_AND_SHADER_BUFFERS
);
3454 ret
= si_insert_input_ptr_as_2xi32(ctx
, ret
, desc_param
+ 1,
3455 8 + GFX9_SGPR_GS_SAMPLERS_AND_IMAGES
);
3457 unsigned vgpr
= 8 + GFX9_GS_NUM_USER_SGPR
;
3458 for (unsigned i
= 0; i
< 5; i
++) {
3459 unsigned param
= ctx
->param_gs_vtx01_offset
+ i
;
3460 ret
= si_insert_input_ret_float(ctx
, ret
, param
, vgpr
++);
3462 ctx
->return_value
= ret
;
3465 static void si_llvm_emit_ls_epilogue(struct ac_shader_abi
*abi
,
3466 unsigned max_outputs
,
3467 LLVMValueRef
*addrs
)
3469 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
3470 struct si_shader
*shader
= ctx
->shader
;
3471 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
3473 LLVMValueRef vertex_id
= LLVMGetParam(ctx
->main_fn
,
3474 ctx
->param_rel_auto_id
);
3475 LLVMValueRef vertex_dw_stride
= get_tcs_in_vertex_dw_stride(ctx
);
3476 LLVMValueRef base_dw_addr
= LLVMBuildMul(ctx
->ac
.builder
, vertex_id
,
3477 vertex_dw_stride
, "");
3479 /* Write outputs to LDS. The next shader (TCS aka HS) will read
3480 * its inputs from it. */
3481 for (i
= 0; i
< info
->num_outputs
; i
++) {
3482 unsigned name
= info
->output_semantic_name
[i
];
3483 unsigned index
= info
->output_semantic_index
[i
];
3485 /* The ARB_shader_viewport_layer_array spec contains the
3488 * 2) What happens if gl_ViewportIndex or gl_Layer is
3489 * written in the vertex shader and a geometry shader is
3492 * RESOLVED: The value written by the last vertex processing
3493 * stage is used. If the last vertex processing stage
3494 * (vertex, tessellation evaluation or geometry) does not
3495 * statically assign to gl_ViewportIndex or gl_Layer, index
3496 * or layer zero is assumed.
3498 * So writes to those outputs in VS-as-LS are simply ignored.
3500 if (name
== TGSI_SEMANTIC_LAYER
||
3501 name
== TGSI_SEMANTIC_VIEWPORT_INDEX
)
3504 int param
= si_shader_io_get_unique_index(name
, index
);
3505 LLVMValueRef dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_dw_addr
,
3506 LLVMConstInt(ctx
->i32
, param
* 4, 0), "");
3508 for (chan
= 0; chan
< 4; chan
++) {
3509 if (!(info
->output_usagemask
[i
] & (1 << chan
)))
3512 lds_store(ctx
, chan
, dw_addr
,
3513 LLVMBuildLoad(ctx
->ac
.builder
, addrs
[4 * i
+ chan
], ""));
3517 if (ctx
->screen
->info
.chip_class
>= GFX9
)
3518 si_set_ls_return_value_for_tcs(ctx
);
3521 static void si_llvm_emit_es_epilogue(struct ac_shader_abi
*abi
,
3522 unsigned max_outputs
,
3523 LLVMValueRef
*addrs
)
3525 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
3526 struct si_shader
*es
= ctx
->shader
;
3527 struct tgsi_shader_info
*info
= &es
->selector
->info
;
3528 LLVMValueRef soffset
= LLVMGetParam(ctx
->main_fn
,
3529 ctx
->param_es2gs_offset
);
3530 LLVMValueRef lds_base
= NULL
;
3534 if (ctx
->screen
->info
.chip_class
>= GFX9
&& info
->num_outputs
) {
3535 unsigned itemsize_dw
= es
->selector
->esgs_itemsize
/ 4;
3536 LLVMValueRef vertex_idx
= ac_get_thread_id(&ctx
->ac
);
3537 LLVMValueRef wave_idx
= unpack_param(ctx
, ctx
->param_merged_wave_info
, 24, 4);
3538 vertex_idx
= LLVMBuildOr(ctx
->ac
.builder
, vertex_idx
,
3539 LLVMBuildMul(ctx
->ac
.builder
, wave_idx
,
3540 LLVMConstInt(ctx
->i32
, 64, false), ""), "");
3541 lds_base
= LLVMBuildMul(ctx
->ac
.builder
, vertex_idx
,
3542 LLVMConstInt(ctx
->i32
, itemsize_dw
, 0), "");
3545 for (i
= 0; i
< info
->num_outputs
; i
++) {
3548 if (info
->output_semantic_name
[i
] == TGSI_SEMANTIC_VIEWPORT_INDEX
||
3549 info
->output_semantic_name
[i
] == TGSI_SEMANTIC_LAYER
)
3552 param
= si_shader_io_get_unique_index(info
->output_semantic_name
[i
],
3553 info
->output_semantic_index
[i
]);
3555 for (chan
= 0; chan
< 4; chan
++) {
3556 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->ac
.builder
, addrs
[4 * i
+ chan
], "");
3557 out_val
= ac_to_integer(&ctx
->ac
, out_val
);
3559 /* GFX9 has the ESGS ring in LDS. */
3560 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
3561 lds_store(ctx
, param
* 4 + chan
, lds_base
, out_val
);
3565 ac_build_buffer_store_dword(&ctx
->ac
,
3567 out_val
, 1, NULL
, soffset
,
3568 (4 * param
+ chan
) * 4,
3573 if (ctx
->screen
->info
.chip_class
>= GFX9
)
3574 si_set_es_return_value_for_gs(ctx
);
3577 static LLVMValueRef
si_get_gs_wave_id(struct si_shader_context
*ctx
)
3579 if (ctx
->screen
->info
.chip_class
>= GFX9
)
3580 return unpack_param(ctx
, ctx
->param_merged_wave_info
, 16, 8);
3582 return LLVMGetParam(ctx
->main_fn
, ctx
->param_gs_wave_id
);
3585 static void emit_gs_epilogue(struct si_shader_context
*ctx
)
3587 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_NOP
| AC_SENDMSG_GS_DONE
,
3588 si_get_gs_wave_id(ctx
));
3590 if (ctx
->screen
->info
.chip_class
>= GFX9
)
3591 lp_build_endif(&ctx
->merged_wrap_if_state
);
3594 static void si_llvm_emit_gs_epilogue(struct ac_shader_abi
*abi
,
3595 unsigned max_outputs
,
3596 LLVMValueRef
*addrs
)
3598 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
3599 struct tgsi_shader_info UNUSED
*info
= &ctx
->shader
->selector
->info
;
3601 assert(info
->num_outputs
<= max_outputs
);
3603 emit_gs_epilogue(ctx
);
3606 static void si_tgsi_emit_gs_epilogue(struct lp_build_tgsi_context
*bld_base
)
3608 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3609 emit_gs_epilogue(ctx
);
3612 static void si_llvm_emit_vs_epilogue(struct ac_shader_abi
*abi
,
3613 unsigned max_outputs
,
3614 LLVMValueRef
*addrs
)
3616 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
3617 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
3618 struct si_shader_output_values
*outputs
= NULL
;
3621 assert(!ctx
->shader
->is_gs_copy_shader
);
3622 assert(info
->num_outputs
<= max_outputs
);
3624 outputs
= MALLOC((info
->num_outputs
+ 1) * sizeof(outputs
[0]));
3626 /* Vertex color clamping.
3628 * This uses a state constant loaded in a user data SGPR and
3629 * an IF statement is added that clamps all colors if the constant
3632 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
3633 struct lp_build_if_state if_ctx
;
3634 LLVMValueRef cond
= NULL
;
3635 LLVMValueRef addr
, val
;
3637 for (i
= 0; i
< info
->num_outputs
; i
++) {
3638 if (info
->output_semantic_name
[i
] != TGSI_SEMANTIC_COLOR
&&
3639 info
->output_semantic_name
[i
] != TGSI_SEMANTIC_BCOLOR
)
3642 /* We've found a color. */
3644 /* The state is in the first bit of the user SGPR. */
3645 cond
= LLVMGetParam(ctx
->main_fn
,
3646 ctx
->param_vs_state_bits
);
3647 cond
= LLVMBuildTrunc(ctx
->ac
.builder
, cond
,
3649 lp_build_if(&if_ctx
, &ctx
->gallivm
, cond
);
3652 for (j
= 0; j
< 4; j
++) {
3653 addr
= addrs
[4 * i
+ j
];
3654 val
= LLVMBuildLoad(ctx
->ac
.builder
, addr
, "");
3655 val
= ac_build_clamp(&ctx
->ac
, val
);
3656 LLVMBuildStore(ctx
->ac
.builder
, val
, addr
);
3661 lp_build_endif(&if_ctx
);
3664 for (i
= 0; i
< info
->num_outputs
; i
++) {
3665 outputs
[i
].semantic_name
= info
->output_semantic_name
[i
];
3666 outputs
[i
].semantic_index
= info
->output_semantic_index
[i
];
3668 for (j
= 0; j
< 4; j
++) {
3669 outputs
[i
].values
[j
] =
3670 LLVMBuildLoad(ctx
->ac
.builder
,
3673 outputs
[i
].vertex_stream
[j
] =
3674 (info
->output_streams
[i
] >> (2 * j
)) & 3;
3678 if (ctx
->shader
->selector
->so
.num_outputs
)
3679 si_llvm_emit_streamout(ctx
, outputs
, i
, 0);
3681 /* Export PrimitiveID. */
3682 if (ctx
->shader
->key
.mono
.u
.vs_export_prim_id
) {
3683 outputs
[i
].semantic_name
= TGSI_SEMANTIC_PRIMID
;
3684 outputs
[i
].semantic_index
= 0;
3685 outputs
[i
].values
[0] = ac_to_float(&ctx
->ac
, get_primitive_id(ctx
, 0));
3686 for (j
= 1; j
< 4; j
++)
3687 outputs
[i
].values
[j
] = LLVMConstReal(ctx
->f32
, 0);
3689 memset(outputs
[i
].vertex_stream
, 0,
3690 sizeof(outputs
[i
].vertex_stream
));
3694 si_llvm_export_vs(ctx
, outputs
, i
);
3698 static void si_tgsi_emit_epilogue(struct lp_build_tgsi_context
*bld_base
)
3700 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3702 ctx
->abi
.emit_outputs(&ctx
->abi
, RADEON_LLVM_MAX_OUTPUTS
,
3703 &ctx
->outputs
[0][0]);
3706 struct si_ps_exports
{
3708 struct ac_export_args args
[10];
3711 static void si_export_mrt_z(struct lp_build_tgsi_context
*bld_base
,
3712 LLVMValueRef depth
, LLVMValueRef stencil
,
3713 LLVMValueRef samplemask
, struct si_ps_exports
*exp
)
3715 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3716 struct ac_export_args args
;
3718 ac_export_mrt_z(&ctx
->ac
, depth
, stencil
, samplemask
, &args
);
3720 memcpy(&exp
->args
[exp
->num
++], &args
, sizeof(args
));
3723 static void si_export_mrt_color(struct lp_build_tgsi_context
*bld_base
,
3724 LLVMValueRef
*color
, unsigned index
,
3725 unsigned samplemask_param
,
3726 bool is_last
, struct si_ps_exports
*exp
)
3728 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3732 if (ctx
->shader
->key
.part
.ps
.epilog
.clamp_color
)
3733 for (i
= 0; i
< 4; i
++)
3734 color
[i
] = ac_build_clamp(&ctx
->ac
, color
[i
]);
3737 if (ctx
->shader
->key
.part
.ps
.epilog
.alpha_to_one
)
3738 color
[3] = ctx
->ac
.f32_1
;
3742 ctx
->shader
->key
.part
.ps
.epilog
.alpha_func
!= PIPE_FUNC_ALWAYS
)
3743 si_alpha_test(bld_base
, color
[3]);
3745 /* Line & polygon smoothing */
3746 if (ctx
->shader
->key
.part
.ps
.epilog
.poly_line_smoothing
)
3747 color
[3] = si_scale_alpha_by_sample_mask(bld_base
, color
[3],
3750 /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
3751 if (ctx
->shader
->key
.part
.ps
.epilog
.last_cbuf
> 0) {
3752 struct ac_export_args args
[8];
3755 /* Get the export arguments, also find out what the last one is. */
3756 for (c
= 0; c
<= ctx
->shader
->key
.part
.ps
.epilog
.last_cbuf
; c
++) {
3757 si_llvm_init_export_args(ctx
, color
,
3758 V_008DFC_SQ_EXP_MRT
+ c
, &args
[c
]);
3759 if (args
[c
].enabled_channels
)
3763 /* Emit all exports. */
3764 for (c
= 0; c
<= ctx
->shader
->key
.part
.ps
.epilog
.last_cbuf
; c
++) {
3765 if (is_last
&& last
== c
) {
3766 args
[c
].valid_mask
= 1; /* whether the EXEC mask is valid */
3767 args
[c
].done
= 1; /* DONE bit */
3768 } else if (!args
[c
].enabled_channels
)
3769 continue; /* unnecessary NULL export */
3771 memcpy(&exp
->args
[exp
->num
++], &args
[c
], sizeof(args
[c
]));
3774 struct ac_export_args args
;
3777 si_llvm_init_export_args(ctx
, color
, V_008DFC_SQ_EXP_MRT
+ index
,
3780 args
.valid_mask
= 1; /* whether the EXEC mask is valid */
3781 args
.done
= 1; /* DONE bit */
3782 } else if (!args
.enabled_channels
)
3783 return; /* unnecessary NULL export */
3785 memcpy(&exp
->args
[exp
->num
++], &args
, sizeof(args
));
3789 static void si_emit_ps_exports(struct si_shader_context
*ctx
,
3790 struct si_ps_exports
*exp
)
3792 for (unsigned i
= 0; i
< exp
->num
; i
++)
3793 ac_build_export(&ctx
->ac
, &exp
->args
[i
]);
3796 static void si_export_null(struct lp_build_tgsi_context
*bld_base
)
3798 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3799 struct lp_build_context
*base
= &bld_base
->base
;
3800 struct ac_export_args args
;
3802 args
.enabled_channels
= 0x0; /* enabled channels */
3803 args
.valid_mask
= 1; /* whether the EXEC mask is valid */
3804 args
.done
= 1; /* DONE bit */
3805 args
.target
= V_008DFC_SQ_EXP_NULL
;
3806 args
.compr
= 0; /* COMPR flag (0 = 32-bit export) */
3807 args
.out
[0] = base
->undef
; /* R */
3808 args
.out
[1] = base
->undef
; /* G */
3809 args
.out
[2] = base
->undef
; /* B */
3810 args
.out
[3] = base
->undef
; /* A */
3812 ac_build_export(&ctx
->ac
, &args
);
3816 * Return PS outputs in this order:
3818 * v[0:3] = color0.xyzw
3819 * v[4:7] = color1.xyzw
3824 * vN+3 = SampleMaskIn (used for OpenGL smoothing)
3826 * The alpha-ref SGPR is returned via its original location.
3828 static void si_llvm_return_fs_outputs(struct ac_shader_abi
*abi
,
3829 unsigned max_outputs
,
3830 LLVMValueRef
*addrs
)
3832 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
3833 struct si_shader
*shader
= ctx
->shader
;
3834 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
3835 LLVMBuilderRef builder
= ctx
->ac
.builder
;
3836 unsigned i
, j
, first_vgpr
, vgpr
;
3838 LLVMValueRef color
[8][4] = {};
3839 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
3842 if (ctx
->postponed_kill
)
3843 ac_build_kill_if_false(&ctx
->ac
, LLVMBuildLoad(builder
, ctx
->postponed_kill
, ""));
3845 /* Read the output values. */
3846 for (i
= 0; i
< info
->num_outputs
; i
++) {
3847 unsigned semantic_name
= info
->output_semantic_name
[i
];
3848 unsigned semantic_index
= info
->output_semantic_index
[i
];
3850 switch (semantic_name
) {
3851 case TGSI_SEMANTIC_COLOR
:
3852 assert(semantic_index
< 8);
3853 for (j
= 0; j
< 4; j
++) {
3854 LLVMValueRef ptr
= addrs
[4 * i
+ j
];
3855 LLVMValueRef result
= LLVMBuildLoad(builder
, ptr
, "");
3856 color
[semantic_index
][j
] = result
;
3859 case TGSI_SEMANTIC_POSITION
:
3860 depth
= LLVMBuildLoad(builder
,
3861 addrs
[4 * i
+ 2], "");
3863 case TGSI_SEMANTIC_STENCIL
:
3864 stencil
= LLVMBuildLoad(builder
,
3865 addrs
[4 * i
+ 1], "");
3867 case TGSI_SEMANTIC_SAMPLEMASK
:
3868 samplemask
= LLVMBuildLoad(builder
,
3869 addrs
[4 * i
+ 0], "");
3872 fprintf(stderr
, "Warning: SI unhandled fs output type:%d\n",
3877 /* Fill the return structure. */
3878 ret
= ctx
->return_value
;
3881 ret
= LLVMBuildInsertValue(builder
, ret
,
3882 ac_to_integer(&ctx
->ac
,
3883 LLVMGetParam(ctx
->main_fn
,
3884 SI_PARAM_ALPHA_REF
)),
3885 SI_SGPR_ALPHA_REF
, "");
3888 first_vgpr
= vgpr
= SI_SGPR_ALPHA_REF
+ 1;
3889 for (i
= 0; i
< ARRAY_SIZE(color
); i
++) {
3893 for (j
= 0; j
< 4; j
++)
3894 ret
= LLVMBuildInsertValue(builder
, ret
, color
[i
][j
], vgpr
++, "");
3897 ret
= LLVMBuildInsertValue(builder
, ret
, depth
, vgpr
++, "");
3899 ret
= LLVMBuildInsertValue(builder
, ret
, stencil
, vgpr
++, "");
3901 ret
= LLVMBuildInsertValue(builder
, ret
, samplemask
, vgpr
++, "");
3903 /* Add the input sample mask for smoothing at the end. */
3904 if (vgpr
< first_vgpr
+ PS_EPILOG_SAMPLEMASK_MIN_LOC
)
3905 vgpr
= first_vgpr
+ PS_EPILOG_SAMPLEMASK_MIN_LOC
;
3906 ret
= LLVMBuildInsertValue(builder
, ret
,
3907 LLVMGetParam(ctx
->main_fn
,
3908 SI_PARAM_SAMPLE_COVERAGE
), vgpr
++, "");
3910 ctx
->return_value
= ret
;
3913 static void membar_emit(
3914 const struct lp_build_tgsi_action
*action
,
3915 struct lp_build_tgsi_context
*bld_base
,
3916 struct lp_build_emit_data
*emit_data
)
3918 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3919 LLVMValueRef src0
= lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, 0);
3920 unsigned flags
= LLVMConstIntGetZExtValue(src0
);
3921 unsigned waitcnt
= NOOP_WAITCNT
;
3923 if (flags
& TGSI_MEMBAR_THREAD_GROUP
)
3924 waitcnt
&= VM_CNT
& LGKM_CNT
;
3926 if (flags
& (TGSI_MEMBAR_ATOMIC_BUFFER
|
3927 TGSI_MEMBAR_SHADER_BUFFER
|
3928 TGSI_MEMBAR_SHADER_IMAGE
))
3931 if (flags
& TGSI_MEMBAR_SHARED
)
3932 waitcnt
&= LGKM_CNT
;
3934 if (waitcnt
!= NOOP_WAITCNT
)
3935 ac_build_waitcnt(&ctx
->ac
, waitcnt
);
3938 static void clock_emit(
3939 const struct lp_build_tgsi_action
*action
,
3940 struct lp_build_tgsi_context
*bld_base
,
3941 struct lp_build_emit_data
*emit_data
)
3943 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3946 tmp
= lp_build_intrinsic(ctx
->ac
.builder
, "llvm.readcyclecounter",
3947 ctx
->i64
, NULL
, 0, 0);
3948 tmp
= LLVMBuildBitCast(ctx
->ac
.builder
, tmp
, ctx
->v2i32
, "");
3950 emit_data
->output
[0] =
3951 LLVMBuildExtractElement(ctx
->ac
.builder
, tmp
, ctx
->i32_0
, "");
3952 emit_data
->output
[1] =
3953 LLVMBuildExtractElement(ctx
->ac
.builder
, tmp
, ctx
->i32_1
, "");
3956 static void si_llvm_emit_ddxy(
3957 const struct lp_build_tgsi_action
*action
,
3958 struct lp_build_tgsi_context
*bld_base
,
3959 struct lp_build_emit_data
*emit_data
)
3961 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3962 unsigned opcode
= emit_data
->info
->opcode
;
3967 if (opcode
== TGSI_OPCODE_DDX_FINE
)
3968 mask
= AC_TID_MASK_LEFT
;
3969 else if (opcode
== TGSI_OPCODE_DDY_FINE
)
3970 mask
= AC_TID_MASK_TOP
;
3972 mask
= AC_TID_MASK_TOP_LEFT
;
3974 /* for DDX we want to next X pixel, DDY next Y pixel. */
3975 idx
= (opcode
== TGSI_OPCODE_DDX
|| opcode
== TGSI_OPCODE_DDX_FINE
) ? 1 : 2;
3977 val
= ac_to_integer(&ctx
->ac
, emit_data
->args
[0]);
3978 val
= ac_build_ddxy(&ctx
->ac
, mask
, idx
, val
);
3979 emit_data
->output
[emit_data
->chan
] = val
;
3983 * this takes an I,J coordinate pair,
3984 * and works out the X and Y derivatives.
3985 * it returns DDX(I), DDX(J), DDY(I), DDY(J).
3987 static LLVMValueRef
si_llvm_emit_ddxy_interp(
3988 struct lp_build_tgsi_context
*bld_base
,
3989 LLVMValueRef interp_ij
)
3991 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3992 LLVMValueRef result
[4], a
;
3995 for (i
= 0; i
< 2; i
++) {
3996 a
= LLVMBuildExtractElement(ctx
->ac
.builder
, interp_ij
,
3997 LLVMConstInt(ctx
->i32
, i
, 0), "");
3998 result
[i
] = lp_build_emit_llvm_unary(bld_base
, TGSI_OPCODE_DDX
, a
);
3999 result
[2+i
] = lp_build_emit_llvm_unary(bld_base
, TGSI_OPCODE_DDY
, a
);
4002 return lp_build_gather_values(&ctx
->gallivm
, result
, 4);
4005 static void interp_fetch_args(
4006 struct lp_build_tgsi_context
*bld_base
,
4007 struct lp_build_emit_data
*emit_data
)
4009 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4010 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
4012 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_OFFSET
) {
4013 /* offset is in second src, first two channels */
4014 emit_data
->args
[0] = lp_build_emit_fetch(bld_base
,
4017 emit_data
->args
[1] = lp_build_emit_fetch(bld_base
,
4020 emit_data
->arg_count
= 2;
4021 } else if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_SAMPLE
) {
4022 LLVMValueRef sample_position
;
4023 LLVMValueRef sample_id
;
4024 LLVMValueRef halfval
= LLVMConstReal(ctx
->f32
, 0.5f
);
4026 /* fetch sample ID, then fetch its sample position,
4027 * and place into first two channels.
4029 sample_id
= lp_build_emit_fetch(bld_base
,
4030 emit_data
->inst
, 1, TGSI_CHAN_X
);
4031 sample_id
= ac_to_integer(&ctx
->ac
, sample_id
);
4033 /* Section 8.13.2 (Interpolation Functions) of the OpenGL Shading
4034 * Language 4.50 spec says about interpolateAtSample:
4036 * "Returns the value of the input interpolant variable at
4037 * the location of sample number sample. If multisample
4038 * buffers are not available, the input variable will be
4039 * evaluated at the center of the pixel. If sample sample
4040 * does not exist, the position used to interpolate the
4041 * input variable is undefined."
4043 * This means that sample_id values outside of the valid are
4044 * in fact valid input, and the usual mechanism for loading the
4045 * sample position doesn't work.
4047 if (ctx
->shader
->key
.mono
.u
.ps
.interpolate_at_sample_force_center
) {
4048 LLVMValueRef center
[4] = {
4049 LLVMConstReal(ctx
->f32
, 0.5),
4050 LLVMConstReal(ctx
->f32
, 0.5),
4055 sample_position
= lp_build_gather_values(&ctx
->gallivm
, center
, 4);
4057 sample_position
= load_sample_position(&ctx
->abi
, sample_id
);
4060 emit_data
->args
[0] = LLVMBuildExtractElement(ctx
->ac
.builder
,
4064 emit_data
->args
[0] = LLVMBuildFSub(ctx
->ac
.builder
, emit_data
->args
[0], halfval
, "");
4065 emit_data
->args
[1] = LLVMBuildExtractElement(ctx
->ac
.builder
,
4068 emit_data
->args
[1] = LLVMBuildFSub(ctx
->ac
.builder
, emit_data
->args
[1], halfval
, "");
4069 emit_data
->arg_count
= 2;
4073 static void build_interp_intrinsic(const struct lp_build_tgsi_action
*action
,
4074 struct lp_build_tgsi_context
*bld_base
,
4075 struct lp_build_emit_data
*emit_data
)
4077 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4078 struct si_shader
*shader
= ctx
->shader
;
4079 const struct tgsi_shader_info
*info
= &shader
->selector
->info
;
4080 LLVMValueRef interp_param
;
4081 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
4082 const struct tgsi_full_src_register
*input
= &inst
->Src
[0];
4083 int input_base
, input_array_size
;
4086 LLVMValueRef prim_mask
= ctx
->abi
.prim_mask
;
4087 LLVMValueRef array_idx
;
4088 int interp_param_idx
;
4092 assert(input
->Register
.File
== TGSI_FILE_INPUT
);
4094 if (input
->Register
.Indirect
) {
4095 unsigned array_id
= input
->Indirect
.ArrayID
;
4098 input_base
= info
->input_array_first
[array_id
];
4099 input_array_size
= info
->input_array_last
[array_id
] - input_base
+ 1;
4101 input_base
= inst
->Src
[0].Register
.Index
;
4102 input_array_size
= info
->num_inputs
- input_base
;
4105 array_idx
= si_get_indirect_index(ctx
, &input
->Indirect
,
4106 1, input
->Register
.Index
- input_base
);
4108 input_base
= inst
->Src
[0].Register
.Index
;
4109 input_array_size
= 1;
4110 array_idx
= ctx
->i32_0
;
4113 interp
= shader
->selector
->info
.input_interpolate
[input_base
];
4115 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_OFFSET
||
4116 inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_SAMPLE
)
4117 location
= TGSI_INTERPOLATE_LOC_CENTER
;
4119 location
= TGSI_INTERPOLATE_LOC_CENTROID
;
4121 interp_param_idx
= lookup_interp_param_index(interp
, location
);
4122 if (interp_param_idx
== -1)
4124 else if (interp_param_idx
)
4125 interp_param
= LLVMGetParam(ctx
->main_fn
, interp_param_idx
);
4127 interp_param
= NULL
;
4129 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_OFFSET
||
4130 inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_SAMPLE
) {
4131 LLVMValueRef ij_out
[2];
4132 LLVMValueRef ddxy_out
= si_llvm_emit_ddxy_interp(bld_base
, interp_param
);
4135 * take the I then J parameters, and the DDX/Y for it, and
4136 * calculate the IJ inputs for the interpolator.
4137 * temp1 = ddx * offset/sample.x + I;
4138 * interp_param.I = ddy * offset/sample.y + temp1;
4139 * temp1 = ddx * offset/sample.x + J;
4140 * interp_param.J = ddy * offset/sample.y + temp1;
4142 for (i
= 0; i
< 2; i
++) {
4143 LLVMValueRef ix_ll
= LLVMConstInt(ctx
->i32
, i
, 0);
4144 LLVMValueRef iy_ll
= LLVMConstInt(ctx
->i32
, i
+ 2, 0);
4145 LLVMValueRef ddx_el
= LLVMBuildExtractElement(ctx
->ac
.builder
,
4146 ddxy_out
, ix_ll
, "");
4147 LLVMValueRef ddy_el
= LLVMBuildExtractElement(ctx
->ac
.builder
,
4148 ddxy_out
, iy_ll
, "");
4149 LLVMValueRef interp_el
= LLVMBuildExtractElement(ctx
->ac
.builder
,
4150 interp_param
, ix_ll
, "");
4151 LLVMValueRef temp1
, temp2
;
4153 interp_el
= ac_to_float(&ctx
->ac
, interp_el
);
4155 temp1
= LLVMBuildFMul(ctx
->ac
.builder
, ddx_el
, emit_data
->args
[0], "");
4157 temp1
= LLVMBuildFAdd(ctx
->ac
.builder
, temp1
, interp_el
, "");
4159 temp2
= LLVMBuildFMul(ctx
->ac
.builder
, ddy_el
, emit_data
->args
[1], "");
4161 ij_out
[i
] = LLVMBuildFAdd(ctx
->ac
.builder
, temp2
, temp1
, "");
4163 interp_param
= lp_build_gather_values(&ctx
->gallivm
, ij_out
, 2);
4167 interp_param
= ac_to_float(&ctx
->ac
, interp_param
);
4169 for (chan
= 0; chan
< 4; chan
++) {
4170 LLVMValueRef gather
= LLVMGetUndef(LLVMVectorType(ctx
->f32
, input_array_size
));
4171 unsigned schan
= tgsi_util_get_full_src_register_swizzle(&inst
->Src
[0], chan
);
4173 for (unsigned idx
= 0; idx
< input_array_size
; ++idx
) {
4174 LLVMValueRef v
, i
= NULL
, j
= NULL
;
4177 i
= LLVMBuildExtractElement(
4178 ctx
->ac
.builder
, interp_param
, ctx
->i32_0
, "");
4179 j
= LLVMBuildExtractElement(
4180 ctx
->ac
.builder
, interp_param
, ctx
->i32_1
, "");
4182 v
= si_build_fs_interp(ctx
, input_base
+ idx
, schan
,
4185 gather
= LLVMBuildInsertElement(ctx
->ac
.builder
,
4186 gather
, v
, LLVMConstInt(ctx
->i32
, idx
, false), "");
4189 emit_data
->output
[chan
] = LLVMBuildExtractElement(
4190 ctx
->ac
.builder
, gather
, array_idx
, "");
4194 static void vote_all_emit(
4195 const struct lp_build_tgsi_action
*action
,
4196 struct lp_build_tgsi_context
*bld_base
,
4197 struct lp_build_emit_data
*emit_data
)
4199 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4201 LLVMValueRef tmp
= ac_build_vote_all(&ctx
->ac
, emit_data
->args
[0]);
4202 emit_data
->output
[emit_data
->chan
] =
4203 LLVMBuildSExt(ctx
->ac
.builder
, tmp
, ctx
->i32
, "");
4206 static void vote_any_emit(
4207 const struct lp_build_tgsi_action
*action
,
4208 struct lp_build_tgsi_context
*bld_base
,
4209 struct lp_build_emit_data
*emit_data
)
4211 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4213 LLVMValueRef tmp
= ac_build_vote_any(&ctx
->ac
, emit_data
->args
[0]);
4214 emit_data
->output
[emit_data
->chan
] =
4215 LLVMBuildSExt(ctx
->ac
.builder
, tmp
, ctx
->i32
, "");
4218 static void vote_eq_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 LLVMValueRef tmp
= ac_build_vote_eq(&ctx
->ac
, emit_data
->args
[0]);
4226 emit_data
->output
[emit_data
->chan
] =
4227 LLVMBuildSExt(ctx
->ac
.builder
, tmp
, ctx
->i32
, "");
4230 static void ballot_emit(
4231 const struct lp_build_tgsi_action
*action
,
4232 struct lp_build_tgsi_context
*bld_base
,
4233 struct lp_build_emit_data
*emit_data
)
4235 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4236 LLVMBuilderRef builder
= ctx
->ac
.builder
;
4239 tmp
= lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, TGSI_CHAN_X
);
4240 tmp
= ac_build_ballot(&ctx
->ac
, tmp
);
4241 tmp
= LLVMBuildBitCast(builder
, tmp
, ctx
->v2i32
, "");
4243 emit_data
->output
[0] = LLVMBuildExtractElement(builder
, tmp
, ctx
->i32_0
, "");
4244 emit_data
->output
[1] = LLVMBuildExtractElement(builder
, tmp
, ctx
->i32_1
, "");
4247 static void read_invoc_fetch_args(
4248 struct lp_build_tgsi_context
*bld_base
,
4249 struct lp_build_emit_data
*emit_data
)
4251 emit_data
->args
[0] = lp_build_emit_fetch(bld_base
, emit_data
->inst
,
4252 0, emit_data
->src_chan
);
4254 /* Always read the source invocation (= lane) from the X channel. */
4255 emit_data
->args
[1] = lp_build_emit_fetch(bld_base
, emit_data
->inst
,
4257 emit_data
->arg_count
= 2;
4260 static void read_lane_emit(
4261 const struct lp_build_tgsi_action
*action
,
4262 struct lp_build_tgsi_context
*bld_base
,
4263 struct lp_build_emit_data
*emit_data
)
4265 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4267 /* We currently have no other way to prevent LLVM from lifting the icmp
4268 * calls to a dominating basic block.
4270 ac_build_optimization_barrier(&ctx
->ac
, &emit_data
->args
[0]);
4272 for (unsigned i
= 0; i
< emit_data
->arg_count
; ++i
)
4273 emit_data
->args
[i
] = ac_to_integer(&ctx
->ac
, emit_data
->args
[i
]);
4275 emit_data
->output
[emit_data
->chan
] =
4276 ac_build_intrinsic(&ctx
->ac
, action
->intr_name
,
4277 ctx
->i32
, emit_data
->args
, emit_data
->arg_count
,
4278 AC_FUNC_ATTR_READNONE
|
4279 AC_FUNC_ATTR_CONVERGENT
);
4282 static unsigned si_llvm_get_stream(struct lp_build_tgsi_context
*bld_base
,
4283 struct lp_build_emit_data
*emit_data
)
4285 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4286 struct tgsi_src_register src0
= emit_data
->inst
->Src
[0].Register
;
4290 assert(src0
.File
== TGSI_FILE_IMMEDIATE
);
4292 imm
= ctx
->imms
[src0
.Index
* TGSI_NUM_CHANNELS
+ src0
.SwizzleX
];
4293 stream
= LLVMConstIntGetZExtValue(imm
) & 0x3;
4297 /* Emit one vertex from the geometry shader */
4298 static void si_llvm_emit_vertex(struct ac_shader_abi
*abi
,
4300 LLVMValueRef
*addrs
)
4302 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
4303 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
4304 struct lp_build_context
*uint
= &ctx
->bld_base
.uint_bld
;
4305 struct si_shader
*shader
= ctx
->shader
;
4306 struct lp_build_if_state if_state
;
4307 LLVMValueRef soffset
= LLVMGetParam(ctx
->main_fn
,
4308 ctx
->param_gs2vs_offset
);
4309 LLVMValueRef gs_next_vertex
;
4310 LLVMValueRef can_emit
;
4311 unsigned chan
, offset
;
4314 /* Write vertex attribute values to GSVS ring */
4315 gs_next_vertex
= LLVMBuildLoad(ctx
->ac
.builder
,
4316 ctx
->gs_next_vertex
[stream
],
4319 /* If this thread has already emitted the declared maximum number of
4320 * vertices, skip the write: excessive vertex emissions are not
4321 * supposed to have any effect.
4323 * If the shader has no writes to memory, kill it instead. This skips
4324 * further memory loads and may allow LLVM to skip to the end
4327 can_emit
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntULT
, gs_next_vertex
,
4328 LLVMConstInt(ctx
->i32
,
4329 shader
->selector
->gs_max_out_vertices
, 0), "");
4331 bool use_kill
= !info
->writes_memory
;
4333 ac_build_kill_if_false(&ctx
->ac
, can_emit
);
4335 lp_build_if(&if_state
, &ctx
->gallivm
, can_emit
);
4339 for (i
= 0; i
< info
->num_outputs
; i
++) {
4340 for (chan
= 0; chan
< 4; chan
++) {
4341 if (!(info
->output_usagemask
[i
] & (1 << chan
)) ||
4342 ((info
->output_streams
[i
] >> (2 * chan
)) & 3) != stream
)
4345 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->ac
.builder
, addrs
[4 * i
+ chan
], "");
4346 LLVMValueRef voffset
=
4347 LLVMConstInt(ctx
->i32
, offset
*
4348 shader
->selector
->gs_max_out_vertices
, 0);
4351 voffset
= lp_build_add(uint
, voffset
, gs_next_vertex
);
4352 voffset
= lp_build_mul_imm(uint
, voffset
, 4);
4354 out_val
= ac_to_integer(&ctx
->ac
, out_val
);
4356 ac_build_buffer_store_dword(&ctx
->ac
,
4357 ctx
->gsvs_ring
[stream
],
4359 voffset
, soffset
, 0,
4364 gs_next_vertex
= lp_build_add(uint
, gs_next_vertex
,
4367 LLVMBuildStore(ctx
->ac
.builder
, gs_next_vertex
, ctx
->gs_next_vertex
[stream
]);
4369 /* Signal vertex emission */
4370 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_EMIT
| AC_SENDMSG_GS
| (stream
<< 8),
4371 si_get_gs_wave_id(ctx
));
4373 lp_build_endif(&if_state
);
4376 /* Emit one vertex from the geometry shader */
4377 static void si_tgsi_emit_vertex(
4378 const struct lp_build_tgsi_action
*action
,
4379 struct lp_build_tgsi_context
*bld_base
,
4380 struct lp_build_emit_data
*emit_data
)
4382 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4383 unsigned stream
= si_llvm_get_stream(bld_base
, emit_data
);
4385 si_llvm_emit_vertex(&ctx
->abi
, stream
, ctx
->outputs
[0]);
4388 /* Cut one primitive from the geometry shader */
4389 static void si_llvm_emit_primitive(struct ac_shader_abi
*abi
,
4392 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
4394 /* Signal primitive cut */
4395 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_CUT
| AC_SENDMSG_GS
| (stream
<< 8),
4396 si_get_gs_wave_id(ctx
));
4399 /* Cut one primitive from the geometry shader */
4400 static void si_tgsi_emit_primitive(
4401 const struct lp_build_tgsi_action
*action
,
4402 struct lp_build_tgsi_context
*bld_base
,
4403 struct lp_build_emit_data
*emit_data
)
4405 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4407 si_llvm_emit_primitive(&ctx
->abi
, si_llvm_get_stream(bld_base
, emit_data
));
4410 static void si_llvm_emit_barrier(const struct lp_build_tgsi_action
*action
,
4411 struct lp_build_tgsi_context
*bld_base
,
4412 struct lp_build_emit_data
*emit_data
)
4414 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4416 /* SI only (thanks to a hw bug workaround):
4417 * The real barrier instruction isn’t needed, because an entire patch
4418 * always fits into a single wave.
4420 if (ctx
->screen
->info
.chip_class
== SI
&&
4421 ctx
->type
== PIPE_SHADER_TESS_CTRL
) {
4422 ac_build_waitcnt(&ctx
->ac
, LGKM_CNT
& VM_CNT
);
4426 lp_build_intrinsic(ctx
->ac
.builder
,
4427 "llvm.amdgcn.s.barrier",
4428 ctx
->voidt
, NULL
, 0, LP_FUNC_ATTR_CONVERGENT
);
4431 static const struct lp_build_tgsi_action interp_action
= {
4432 .fetch_args
= interp_fetch_args
,
4433 .emit
= build_interp_intrinsic
,
4436 static void si_create_function(struct si_shader_context
*ctx
,
4438 LLVMTypeRef
*returns
, unsigned num_returns
,
4439 struct si_function_info
*fninfo
,
4440 unsigned max_workgroup_size
)
4444 si_llvm_create_func(ctx
, name
, returns
, num_returns
,
4445 fninfo
->types
, fninfo
->num_params
);
4446 ctx
->return_value
= LLVMGetUndef(ctx
->return_type
);
4448 for (i
= 0; i
< fninfo
->num_sgpr_params
; ++i
) {
4449 LLVMValueRef P
= LLVMGetParam(ctx
->main_fn
, i
);
4451 /* The combination of:
4455 * allows the optimization passes to move loads and reduces
4456 * SGPR spilling significantly.
4458 lp_add_function_attr(ctx
->main_fn
, i
+ 1, LP_FUNC_ATTR_INREG
);
4460 if (LLVMGetTypeKind(LLVMTypeOf(P
)) == LLVMPointerTypeKind
) {
4461 lp_add_function_attr(ctx
->main_fn
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
4462 ac_add_attr_dereferenceable(P
, UINT64_MAX
);
4466 for (i
= 0; i
< fninfo
->num_params
; ++i
) {
4467 if (fninfo
->assign
[i
])
4468 *fninfo
->assign
[i
] = LLVMGetParam(ctx
->main_fn
, i
);
4471 if (max_workgroup_size
) {
4472 si_llvm_add_attribute(ctx
->main_fn
, "amdgpu-max-work-group-size",
4473 max_workgroup_size
);
4475 LLVMAddTargetDependentFunctionAttr(ctx
->main_fn
,
4476 "no-signed-zeros-fp-math",
4479 if (ctx
->screen
->debug_flags
& DBG(UNSAFE_MATH
)) {
4480 /* These were copied from some LLVM test. */
4481 LLVMAddTargetDependentFunctionAttr(ctx
->main_fn
,
4482 "less-precise-fpmad",
4484 LLVMAddTargetDependentFunctionAttr(ctx
->main_fn
,
4487 LLVMAddTargetDependentFunctionAttr(ctx
->main_fn
,
4490 LLVMAddTargetDependentFunctionAttr(ctx
->main_fn
,
4496 static void declare_streamout_params(struct si_shader_context
*ctx
,
4497 struct pipe_stream_output_info
*so
,
4498 struct si_function_info
*fninfo
)
4502 /* Streamout SGPRs. */
4503 if (so
->num_outputs
) {
4504 if (ctx
->type
!= PIPE_SHADER_TESS_EVAL
)
4505 ctx
->param_streamout_config
= add_arg(fninfo
, ARG_SGPR
, ctx
->ac
.i32
);
4507 ctx
->param_streamout_config
= fninfo
->num_params
- 1;
4509 ctx
->param_streamout_write_index
= add_arg(fninfo
, ARG_SGPR
, ctx
->ac
.i32
);
4511 /* A streamout buffer offset is loaded if the stride is non-zero. */
4512 for (i
= 0; i
< 4; i
++) {
4516 ctx
->param_streamout_offset
[i
] = add_arg(fninfo
, ARG_SGPR
, ctx
->ac
.i32
);
4520 static unsigned si_get_max_workgroup_size(const struct si_shader
*shader
)
4522 switch (shader
->selector
->type
) {
4523 case PIPE_SHADER_TESS_CTRL
:
4524 /* Return this so that LLVM doesn't remove s_barrier
4525 * instructions on chips where we use s_barrier. */
4526 return shader
->selector
->screen
->info
.chip_class
>= CIK
? 128 : 64;
4528 case PIPE_SHADER_GEOMETRY
:
4529 return shader
->selector
->screen
->info
.chip_class
>= GFX9
? 128 : 64;
4531 case PIPE_SHADER_COMPUTE
:
4532 break; /* see below */
4538 const unsigned *properties
= shader
->selector
->info
.properties
;
4539 unsigned max_work_group_size
=
4540 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
] *
4541 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT
] *
4542 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH
];
4544 if (!max_work_group_size
) {
4545 /* This is a variable group size compute shader,
4546 * compile it for the maximum possible group size.
4548 max_work_group_size
= SI_MAX_VARIABLE_THREADS_PER_BLOCK
;
4550 return max_work_group_size
;
4553 static void declare_per_stage_desc_pointers(struct si_shader_context
*ctx
,
4554 struct si_function_info
*fninfo
,
4557 LLVMTypeRef const_shader_buf_type
;
4559 if (ctx
->shader
->selector
->info
.const_buffers_declared
== 1 &&
4560 ctx
->shader
->selector
->info
.shader_buffers_declared
== 0)
4561 const_shader_buf_type
= ctx
->f32
;
4563 const_shader_buf_type
= ctx
->v4i32
;
4565 unsigned const_and_shader_buffers
=
4566 add_arg(fninfo
, ARG_SGPR
,
4567 ac_array_in_const_addr_space(const_shader_buf_type
));
4569 unsigned samplers_and_images
=
4570 add_arg(fninfo
, ARG_SGPR
,
4571 ac_array_in_const_addr_space(ctx
->v8i32
));
4573 if (assign_params
) {
4574 ctx
->param_const_and_shader_buffers
= const_and_shader_buffers
;
4575 ctx
->param_samplers_and_images
= samplers_and_images
;
4579 static void declare_global_desc_pointers(struct si_shader_context
*ctx
,
4580 struct si_function_info
*fninfo
)
4582 ctx
->param_rw_buffers
= add_arg(fninfo
, ARG_SGPR
,
4583 ac_array_in_const_addr_space(ctx
->v4i32
));
4584 ctx
->param_bindless_samplers_and_images
= add_arg(fninfo
, ARG_SGPR
,
4585 ac_array_in_const_addr_space(ctx
->v8i32
));
4588 static void declare_vs_specific_input_sgprs(struct si_shader_context
*ctx
,
4589 struct si_function_info
*fninfo
)
4591 ctx
->param_vertex_buffers
= add_arg(fninfo
, ARG_SGPR
,
4592 ac_array_in_const_addr_space(ctx
->v4i32
));
4593 add_arg_assign(fninfo
, ARG_SGPR
, ctx
->i32
, &ctx
->abi
.base_vertex
);
4594 add_arg_assign(fninfo
, ARG_SGPR
, ctx
->i32
, &ctx
->abi
.start_instance
);
4595 add_arg_assign(fninfo
, ARG_SGPR
, ctx
->i32
, &ctx
->abi
.draw_id
);
4596 ctx
->param_vs_state_bits
= add_arg(fninfo
, ARG_SGPR
, ctx
->i32
);
4599 static void declare_vs_input_vgprs(struct si_shader_context
*ctx
,
4600 struct si_function_info
*fninfo
,
4601 unsigned *num_prolog_vgprs
)
4603 struct si_shader
*shader
= ctx
->shader
;
4605 add_arg_assign(fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.vertex_id
);
4606 if (shader
->key
.as_ls
) {
4607 ctx
->param_rel_auto_id
= add_arg(fninfo
, ARG_VGPR
, ctx
->i32
);
4608 add_arg_assign(fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.instance_id
);
4610 add_arg_assign(fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.instance_id
);
4611 ctx
->param_vs_prim_id
= add_arg(fninfo
, ARG_VGPR
, ctx
->i32
);
4613 add_arg(fninfo
, ARG_VGPR
, ctx
->i32
); /* unused */
4615 if (!shader
->is_gs_copy_shader
) {
4616 /* Vertex load indices. */
4617 ctx
->param_vertex_index0
= fninfo
->num_params
;
4618 for (unsigned i
= 0; i
< shader
->selector
->info
.num_inputs
; i
++)
4619 add_arg(fninfo
, ARG_VGPR
, ctx
->i32
);
4620 *num_prolog_vgprs
+= shader
->selector
->info
.num_inputs
;
4624 static void declare_tes_input_vgprs(struct si_shader_context
*ctx
,
4625 struct si_function_info
*fninfo
)
4627 ctx
->param_tes_u
= add_arg(fninfo
, ARG_VGPR
, ctx
->f32
);
4628 ctx
->param_tes_v
= add_arg(fninfo
, ARG_VGPR
, ctx
->f32
);
4629 ctx
->param_tes_rel_patch_id
= add_arg(fninfo
, ARG_VGPR
, ctx
->i32
);
4630 add_arg_assign(fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.tes_patch_id
);
4634 /* Convenient merged shader definitions. */
4635 SI_SHADER_MERGED_VERTEX_TESSCTRL
= PIPE_SHADER_TYPES
,
4636 SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY
,
4639 static void create_function(struct si_shader_context
*ctx
)
4641 struct si_shader
*shader
= ctx
->shader
;
4642 struct si_function_info fninfo
;
4643 LLVMTypeRef returns
[16+32*4];
4644 unsigned i
, num_return_sgprs
;
4645 unsigned num_returns
= 0;
4646 unsigned num_prolog_vgprs
= 0;
4647 unsigned type
= ctx
->type
;
4648 unsigned vs_blit_property
=
4649 shader
->selector
->info
.properties
[TGSI_PROPERTY_VS_BLIT_SGPRS
];
4651 si_init_function_info(&fninfo
);
4653 /* Set MERGED shaders. */
4654 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
4655 if (shader
->key
.as_ls
|| type
== PIPE_SHADER_TESS_CTRL
)
4656 type
= SI_SHADER_MERGED_VERTEX_TESSCTRL
; /* LS or HS */
4657 else if (shader
->key
.as_es
|| type
== PIPE_SHADER_GEOMETRY
)
4658 type
= SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY
;
4661 LLVMTypeRef v3i32
= LLVMVectorType(ctx
->i32
, 3);
4664 case PIPE_SHADER_VERTEX
:
4665 declare_global_desc_pointers(ctx
, &fninfo
);
4667 if (vs_blit_property
) {
4668 ctx
->param_vs_blit_inputs
= fninfo
.num_params
;
4669 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* i16 x1, y1 */
4670 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* i16 x2, y2 */
4671 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* depth */
4673 if (vs_blit_property
== SI_VS_BLIT_SGPRS_POS_COLOR
) {
4674 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* color0 */
4675 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* color1 */
4676 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* color2 */
4677 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* color3 */
4678 } else if (vs_blit_property
== SI_VS_BLIT_SGPRS_POS_TEXCOORD
) {
4679 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* texcoord.x1 */
4680 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* texcoord.y1 */
4681 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* texcoord.x2 */
4682 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* texcoord.y2 */
4683 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* texcoord.z */
4684 add_arg(&fninfo
, ARG_SGPR
, ctx
->f32
); /* texcoord.w */
4688 declare_vs_input_vgprs(ctx
, &fninfo
, &num_prolog_vgprs
);
4692 declare_per_stage_desc_pointers(ctx
, &fninfo
, true);
4693 declare_vs_specific_input_sgprs(ctx
, &fninfo
);
4695 if (shader
->key
.as_es
) {
4696 ctx
->param_es2gs_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4697 } else if (shader
->key
.as_ls
) {
4698 /* no extra parameters */
4700 if (shader
->is_gs_copy_shader
) {
4701 fninfo
.num_params
= ctx
->param_rw_buffers
+ 1;
4702 fninfo
.num_sgpr_params
= fninfo
.num_params
;
4705 /* The locations of the other parameters are assigned dynamically. */
4706 declare_streamout_params(ctx
, &shader
->selector
->so
,
4711 declare_vs_input_vgprs(ctx
, &fninfo
, &num_prolog_vgprs
);
4714 case PIPE_SHADER_TESS_CTRL
: /* SI-CI-VI */
4715 declare_global_desc_pointers(ctx
, &fninfo
);
4716 declare_per_stage_desc_pointers(ctx
, &fninfo
, true);
4717 ctx
->param_tcs_offchip_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4718 ctx
->param_tcs_out_lds_offsets
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4719 ctx
->param_tcs_out_lds_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4720 ctx
->param_vs_state_bits
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4721 ctx
->param_tcs_offchip_addr_base64k
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4722 ctx
->param_tcs_factor_addr_base64k
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4723 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4724 ctx
->param_tcs_factor_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4727 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.tcs_patch_id
);
4728 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.tcs_rel_ids
);
4730 /* param_tcs_offchip_offset and param_tcs_factor_offset are
4731 * placed after the user SGPRs.
4733 for (i
= 0; i
< GFX6_TCS_NUM_USER_SGPR
+ 2; i
++)
4734 returns
[num_returns
++] = ctx
->i32
; /* SGPRs */
4735 for (i
= 0; i
< 11; i
++)
4736 returns
[num_returns
++] = ctx
->f32
; /* VGPRs */
4739 case SI_SHADER_MERGED_VERTEX_TESSCTRL
:
4740 /* Merged stages have 8 system SGPRs at the beginning. */
4741 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* SPI_SHADER_USER_DATA_ADDR_LO_HS */
4742 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* SPI_SHADER_USER_DATA_ADDR_HI_HS */
4743 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4744 ctx
->param_merged_wave_info
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4745 ctx
->param_tcs_factor_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4746 ctx
->param_merged_scratch_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4747 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused */
4748 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused */
4750 declare_global_desc_pointers(ctx
, &fninfo
);
4751 declare_per_stage_desc_pointers(ctx
, &fninfo
,
4752 ctx
->type
== PIPE_SHADER_VERTEX
);
4753 declare_vs_specific_input_sgprs(ctx
, &fninfo
);
4755 ctx
->param_tcs_offchip_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4756 ctx
->param_tcs_out_lds_offsets
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4757 ctx
->param_tcs_out_lds_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4758 ctx
->param_tcs_offchip_addr_base64k
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4759 ctx
->param_tcs_factor_addr_base64k
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4760 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused */
4762 declare_per_stage_desc_pointers(ctx
, &fninfo
,
4763 ctx
->type
== PIPE_SHADER_TESS_CTRL
);
4765 /* VGPRs (first TCS, then VS) */
4766 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.tcs_patch_id
);
4767 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.tcs_rel_ids
);
4769 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
4770 declare_vs_input_vgprs(ctx
, &fninfo
,
4773 /* LS return values are inputs to the TCS main shader part. */
4774 for (i
= 0; i
< 8 + GFX9_TCS_NUM_USER_SGPR
; i
++)
4775 returns
[num_returns
++] = ctx
->i32
; /* SGPRs */
4776 for (i
= 0; i
< 2; i
++)
4777 returns
[num_returns
++] = ctx
->f32
; /* VGPRs */
4779 /* TCS return values are inputs to the TCS epilog.
4781 * param_tcs_offchip_offset, param_tcs_factor_offset,
4782 * param_tcs_offchip_layout, and param_rw_buffers
4783 * should be passed to the epilog.
4785 for (i
= 0; i
<= 8 + GFX9_SGPR_TCS_FACTOR_ADDR_BASE64K
; i
++)
4786 returns
[num_returns
++] = ctx
->i32
; /* SGPRs */
4787 for (i
= 0; i
< 11; i
++)
4788 returns
[num_returns
++] = ctx
->f32
; /* VGPRs */
4792 case SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY
:
4793 /* Merged stages have 8 system SGPRs at the beginning. */
4794 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused (SPI_SHADER_USER_DATA_ADDR_LO_GS) */
4795 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused (SPI_SHADER_USER_DATA_ADDR_HI_GS) */
4796 ctx
->param_gs2vs_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4797 ctx
->param_merged_wave_info
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4798 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4799 ctx
->param_merged_scratch_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4800 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused (SPI_SHADER_PGM_LO/HI_GS << 8) */
4801 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused (SPI_SHADER_PGM_LO/HI_GS >> 24) */
4803 declare_global_desc_pointers(ctx
, &fninfo
);
4804 declare_per_stage_desc_pointers(ctx
, &fninfo
,
4805 (ctx
->type
== PIPE_SHADER_VERTEX
||
4806 ctx
->type
== PIPE_SHADER_TESS_EVAL
));
4807 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
4808 declare_vs_specific_input_sgprs(ctx
, &fninfo
);
4810 /* TESS_EVAL (and also GEOMETRY):
4811 * Declare as many input SGPRs as the VS has. */
4812 ctx
->param_tcs_offchip_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4813 ctx
->param_tcs_offchip_addr_base64k
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4814 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused */
4815 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused */
4816 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused */
4817 ctx
->param_vs_state_bits
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* unused */
4820 declare_per_stage_desc_pointers(ctx
, &fninfo
,
4821 ctx
->type
== PIPE_SHADER_GEOMETRY
);
4823 /* VGPRs (first GS, then VS/TES) */
4824 ctx
->param_gs_vtx01_offset
= add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
4825 ctx
->param_gs_vtx23_offset
= add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
4826 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.gs_prim_id
);
4827 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.gs_invocation_id
);
4828 ctx
->param_gs_vtx45_offset
= add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
4830 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
4831 declare_vs_input_vgprs(ctx
, &fninfo
,
4833 } else if (ctx
->type
== PIPE_SHADER_TESS_EVAL
) {
4834 declare_tes_input_vgprs(ctx
, &fninfo
);
4837 if (ctx
->type
== PIPE_SHADER_VERTEX
||
4838 ctx
->type
== PIPE_SHADER_TESS_EVAL
) {
4839 /* ES return values are inputs to GS. */
4840 for (i
= 0; i
< 8 + GFX9_GS_NUM_USER_SGPR
; i
++)
4841 returns
[num_returns
++] = ctx
->i32
; /* SGPRs */
4842 for (i
= 0; i
< 5; i
++)
4843 returns
[num_returns
++] = ctx
->f32
; /* VGPRs */
4847 case PIPE_SHADER_TESS_EVAL
:
4848 declare_global_desc_pointers(ctx
, &fninfo
);
4849 declare_per_stage_desc_pointers(ctx
, &fninfo
, true);
4850 ctx
->param_tcs_offchip_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4851 ctx
->param_tcs_offchip_addr_base64k
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4853 if (shader
->key
.as_es
) {
4854 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4855 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4856 ctx
->param_es2gs_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4858 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4859 declare_streamout_params(ctx
, &shader
->selector
->so
,
4861 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4865 declare_tes_input_vgprs(ctx
, &fninfo
);
4868 case PIPE_SHADER_GEOMETRY
:
4869 declare_global_desc_pointers(ctx
, &fninfo
);
4870 declare_per_stage_desc_pointers(ctx
, &fninfo
, true);
4871 ctx
->param_gs2vs_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4872 ctx
->param_gs_wave_id
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4875 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->gs_vtx_offset
[0]);
4876 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->gs_vtx_offset
[1]);
4877 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.gs_prim_id
);
4878 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->gs_vtx_offset
[2]);
4879 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->gs_vtx_offset
[3]);
4880 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->gs_vtx_offset
[4]);
4881 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->gs_vtx_offset
[5]);
4882 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &ctx
->abi
.gs_invocation_id
);
4885 case PIPE_SHADER_FRAGMENT
:
4886 declare_global_desc_pointers(ctx
, &fninfo
);
4887 declare_per_stage_desc_pointers(ctx
, &fninfo
, true);
4888 add_arg_checked(&fninfo
, ARG_SGPR
, ctx
->f32
, SI_PARAM_ALPHA_REF
);
4889 add_arg_assign_checked(&fninfo
, ARG_SGPR
, ctx
->i32
,
4890 &ctx
->abi
.prim_mask
, SI_PARAM_PRIM_MASK
);
4892 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->v2i32
, SI_PARAM_PERSP_SAMPLE
);
4893 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->v2i32
, SI_PARAM_PERSP_CENTER
);
4894 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->v2i32
, SI_PARAM_PERSP_CENTROID
);
4895 add_arg_checked(&fninfo
, ARG_VGPR
, v3i32
, SI_PARAM_PERSP_PULL_MODEL
);
4896 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->v2i32
, SI_PARAM_LINEAR_SAMPLE
);
4897 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->v2i32
, SI_PARAM_LINEAR_CENTER
);
4898 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->v2i32
, SI_PARAM_LINEAR_CENTROID
);
4899 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->f32
, SI_PARAM_LINE_STIPPLE_TEX
);
4900 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->f32
,
4901 &ctx
->abi
.frag_pos
[0], SI_PARAM_POS_X_FLOAT
);
4902 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->f32
,
4903 &ctx
->abi
.frag_pos
[1], SI_PARAM_POS_Y_FLOAT
);
4904 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->f32
,
4905 &ctx
->abi
.frag_pos
[2], SI_PARAM_POS_Z_FLOAT
);
4906 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->f32
,
4907 &ctx
->abi
.frag_pos
[3], SI_PARAM_POS_W_FLOAT
);
4908 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->i32
,
4909 &ctx
->abi
.front_face
, SI_PARAM_FRONT_FACE
);
4910 shader
->info
.face_vgpr_index
= 20;
4911 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->i32
,
4912 &ctx
->abi
.ancillary
, SI_PARAM_ANCILLARY
);
4913 shader
->info
.ancillary_vgpr_index
= 21;
4914 add_arg_assign_checked(&fninfo
, ARG_VGPR
, ctx
->f32
,
4915 &ctx
->abi
.sample_coverage
, SI_PARAM_SAMPLE_COVERAGE
);
4916 add_arg_checked(&fninfo
, ARG_VGPR
, ctx
->i32
, SI_PARAM_POS_FIXED_PT
);
4918 /* Color inputs from the prolog. */
4919 if (shader
->selector
->info
.colors_read
) {
4920 unsigned num_color_elements
=
4921 util_bitcount(shader
->selector
->info
.colors_read
);
4923 assert(fninfo
.num_params
+ num_color_elements
<= ARRAY_SIZE(fninfo
.types
));
4924 for (i
= 0; i
< num_color_elements
; i
++)
4925 add_arg(&fninfo
, ARG_VGPR
, ctx
->f32
);
4927 num_prolog_vgprs
+= num_color_elements
;
4930 /* Outputs for the epilog. */
4931 num_return_sgprs
= SI_SGPR_ALPHA_REF
+ 1;
4934 util_bitcount(shader
->selector
->info
.colors_written
) * 4 +
4935 shader
->selector
->info
.writes_z
+
4936 shader
->selector
->info
.writes_stencil
+
4937 shader
->selector
->info
.writes_samplemask
+
4938 1 /* SampleMaskIn */;
4940 num_returns
= MAX2(num_returns
,
4942 PS_EPILOG_SAMPLEMASK_MIN_LOC
+ 1);
4944 for (i
= 0; i
< num_return_sgprs
; i
++)
4945 returns
[i
] = ctx
->i32
;
4946 for (; i
< num_returns
; i
++)
4947 returns
[i
] = ctx
->f32
;
4950 case PIPE_SHADER_COMPUTE
:
4951 declare_global_desc_pointers(ctx
, &fninfo
);
4952 declare_per_stage_desc_pointers(ctx
, &fninfo
, true);
4953 if (shader
->selector
->info
.uses_grid_size
)
4954 ctx
->param_grid_size
= add_arg(&fninfo
, ARG_SGPR
, v3i32
);
4955 if (shader
->selector
->info
.uses_block_size
)
4956 ctx
->param_block_size
= add_arg(&fninfo
, ARG_SGPR
, v3i32
);
4958 for (i
= 0; i
< 3; i
++) {
4959 ctx
->param_block_id
[i
] = -1;
4960 if (shader
->selector
->info
.uses_block_id
[i
])
4961 ctx
->param_block_id
[i
] = add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
4964 ctx
->param_thread_id
= add_arg(&fninfo
, ARG_VGPR
, v3i32
);
4967 assert(0 && "unimplemented shader");
4971 si_create_function(ctx
, "main", returns
, num_returns
, &fninfo
,
4972 si_get_max_workgroup_size(shader
));
4974 /* Reserve register locations for VGPR inputs the PS prolog may need. */
4975 if (ctx
->type
== PIPE_SHADER_FRAGMENT
&&
4976 ctx
->separate_prolog
) {
4977 si_llvm_add_attribute(ctx
->main_fn
,
4978 "InitialPSInputAddr",
4979 S_0286D0_PERSP_SAMPLE_ENA(1) |
4980 S_0286D0_PERSP_CENTER_ENA(1) |
4981 S_0286D0_PERSP_CENTROID_ENA(1) |
4982 S_0286D0_LINEAR_SAMPLE_ENA(1) |
4983 S_0286D0_LINEAR_CENTER_ENA(1) |
4984 S_0286D0_LINEAR_CENTROID_ENA(1) |
4985 S_0286D0_FRONT_FACE_ENA(1) |
4986 S_0286D0_ANCILLARY_ENA(1) |
4987 S_0286D0_POS_FIXED_PT_ENA(1));
4990 shader
->info
.num_input_sgprs
= 0;
4991 shader
->info
.num_input_vgprs
= 0;
4993 for (i
= 0; i
< fninfo
.num_sgpr_params
; ++i
)
4994 shader
->info
.num_input_sgprs
+= ac_get_type_size(fninfo
.types
[i
]) / 4;
4996 for (; i
< fninfo
.num_params
; ++i
)
4997 shader
->info
.num_input_vgprs
+= ac_get_type_size(fninfo
.types
[i
]) / 4;
4999 assert(shader
->info
.num_input_vgprs
>= num_prolog_vgprs
);
5000 shader
->info
.num_input_vgprs
-= num_prolog_vgprs
;
5002 if (shader
->key
.as_ls
||
5003 ctx
->type
== PIPE_SHADER_TESS_CTRL
||
5004 /* GFX9 has the ESGS ring buffer in LDS. */
5005 type
== SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY
)
5006 ac_declare_lds_as_pointer(&ctx
->ac
);
5010 * Load ESGS and GSVS ring buffer resource descriptors and save the variables
5013 static void preload_ring_buffers(struct si_shader_context
*ctx
)
5015 LLVMBuilderRef builder
= ctx
->ac
.builder
;
5017 LLVMValueRef buf_ptr
= LLVMGetParam(ctx
->main_fn
,
5018 ctx
->param_rw_buffers
);
5020 if (ctx
->screen
->info
.chip_class
<= VI
&&
5021 (ctx
->shader
->key
.as_es
|| ctx
->type
== PIPE_SHADER_GEOMETRY
)) {
5023 ctx
->type
== PIPE_SHADER_GEOMETRY
? SI_GS_RING_ESGS
5025 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
, ring
, 0);
5028 ac_build_load_to_sgpr(&ctx
->ac
, buf_ptr
, offset
);
5031 if (ctx
->shader
->is_gs_copy_shader
) {
5032 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
, SI_RING_GSVS
, 0);
5035 ac_build_load_to_sgpr(&ctx
->ac
, buf_ptr
, offset
);
5036 } else if (ctx
->type
== PIPE_SHADER_GEOMETRY
) {
5037 const struct si_shader_selector
*sel
= ctx
->shader
->selector
;
5038 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
, SI_RING_GSVS
, 0);
5039 LLVMValueRef base_ring
;
5041 base_ring
= ac_build_load_to_sgpr(&ctx
->ac
, buf_ptr
, offset
);
5043 /* The conceptual layout of the GSVS ring is
5044 * v0c0 .. vLv0 v0c1 .. vLc1 ..
5045 * but the real memory layout is swizzled across
5047 * t0v0c0 .. t15v0c0 t0v1c0 .. t15v1c0 ... t15vLcL
5049 * Override the buffer descriptor accordingly.
5051 LLVMTypeRef v2i64
= LLVMVectorType(ctx
->i64
, 2);
5052 uint64_t stream_offset
= 0;
5054 for (unsigned stream
= 0; stream
< 4; ++stream
) {
5055 unsigned num_components
;
5057 unsigned num_records
;
5058 LLVMValueRef ring
, tmp
;
5060 num_components
= sel
->info
.num_stream_output_components
[stream
];
5061 if (!num_components
)
5064 stride
= 4 * num_components
* sel
->gs_max_out_vertices
;
5066 /* Limit on the stride field for <= CIK. */
5067 assert(stride
< (1 << 14));
5071 ring
= LLVMBuildBitCast(builder
, base_ring
, v2i64
, "");
5072 tmp
= LLVMBuildExtractElement(builder
, ring
, ctx
->i32_0
, "");
5073 tmp
= LLVMBuildAdd(builder
, tmp
,
5074 LLVMConstInt(ctx
->i64
,
5075 stream_offset
, 0), "");
5076 stream_offset
+= stride
* 64;
5078 ring
= LLVMBuildInsertElement(builder
, ring
, tmp
, ctx
->i32_0
, "");
5079 ring
= LLVMBuildBitCast(builder
, ring
, ctx
->v4i32
, "");
5080 tmp
= LLVMBuildExtractElement(builder
, ring
, ctx
->i32_1
, "");
5081 tmp
= LLVMBuildOr(builder
, tmp
,
5082 LLVMConstInt(ctx
->i32
,
5083 S_008F04_STRIDE(stride
) |
5084 S_008F04_SWIZZLE_ENABLE(1), 0), "");
5085 ring
= LLVMBuildInsertElement(builder
, ring
, tmp
, ctx
->i32_1
, "");
5086 ring
= LLVMBuildInsertElement(builder
, ring
,
5087 LLVMConstInt(ctx
->i32
, num_records
, 0),
5088 LLVMConstInt(ctx
->i32
, 2, 0), "");
5089 ring
= LLVMBuildInsertElement(builder
, ring
,
5090 LLVMConstInt(ctx
->i32
,
5091 S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
5092 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
5093 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
5094 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
5095 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
5096 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
5097 S_008F0C_ELEMENT_SIZE(1) | /* element_size = 4 (bytes) */
5098 S_008F0C_INDEX_STRIDE(1) | /* index_stride = 16 (elements) */
5099 S_008F0C_ADD_TID_ENABLE(1),
5101 LLVMConstInt(ctx
->i32
, 3, 0), "");
5103 ctx
->gsvs_ring
[stream
] = ring
;
5108 static void si_llvm_emit_polygon_stipple(struct si_shader_context
*ctx
,
5109 LLVMValueRef param_rw_buffers
,
5110 unsigned param_pos_fixed_pt
)
5112 LLVMBuilderRef builder
= ctx
->ac
.builder
;
5113 LLVMValueRef slot
, desc
, offset
, row
, bit
, address
[2];
5115 /* Use the fixed-point gl_FragCoord input.
5116 * Since the stipple pattern is 32x32 and it repeats, just get 5 bits
5117 * per coordinate to get the repeating effect.
5119 address
[0] = unpack_param(ctx
, param_pos_fixed_pt
, 0, 5);
5120 address
[1] = unpack_param(ctx
, param_pos_fixed_pt
, 16, 5);
5122 /* Load the buffer descriptor. */
5123 slot
= LLVMConstInt(ctx
->i32
, SI_PS_CONST_POLY_STIPPLE
, 0);
5124 desc
= ac_build_load_to_sgpr(&ctx
->ac
, param_rw_buffers
, slot
);
5126 /* The stipple pattern is 32x32, each row has 32 bits. */
5127 offset
= LLVMBuildMul(builder
, address
[1],
5128 LLVMConstInt(ctx
->i32
, 4, 0), "");
5129 row
= buffer_load_const(ctx
, desc
, offset
);
5130 row
= ac_to_integer(&ctx
->ac
, row
);
5131 bit
= LLVMBuildLShr(builder
, row
, address
[0], "");
5132 bit
= LLVMBuildTrunc(builder
, bit
, ctx
->i1
, "");
5133 ac_build_kill_if_false(&ctx
->ac
, bit
);
5136 void si_shader_binary_read_config(struct ac_shader_binary
*binary
,
5137 struct si_shader_config
*conf
,
5138 unsigned symbol_offset
)
5141 const unsigned char *config
=
5142 ac_shader_binary_config_start(binary
, symbol_offset
);
5143 bool really_needs_scratch
= false;
5145 /* LLVM adds SGPR spills to the scratch size.
5146 * Find out if we really need the scratch buffer.
5148 for (i
= 0; i
< binary
->reloc_count
; i
++) {
5149 const struct ac_shader_reloc
*reloc
= &binary
->relocs
[i
];
5151 if (!strcmp(scratch_rsrc_dword0_symbol
, reloc
->name
) ||
5152 !strcmp(scratch_rsrc_dword1_symbol
, reloc
->name
)) {
5153 really_needs_scratch
= true;
5158 /* XXX: We may be able to emit some of these values directly rather than
5159 * extracting fields to be emitted later.
5162 for (i
= 0; i
< binary
->config_size_per_symbol
; i
+= 8) {
5163 unsigned reg
= util_le32_to_cpu(*(uint32_t*)(config
+ i
));
5164 unsigned value
= util_le32_to_cpu(*(uint32_t*)(config
+ i
+ 4));
5166 case R_00B028_SPI_SHADER_PGM_RSRC1_PS
:
5167 case R_00B128_SPI_SHADER_PGM_RSRC1_VS
:
5168 case R_00B228_SPI_SHADER_PGM_RSRC1_GS
:
5169 case R_00B428_SPI_SHADER_PGM_RSRC1_HS
:
5170 case R_00B848_COMPUTE_PGM_RSRC1
:
5171 conf
->num_sgprs
= MAX2(conf
->num_sgprs
, (G_00B028_SGPRS(value
) + 1) * 8);
5172 conf
->num_vgprs
= MAX2(conf
->num_vgprs
, (G_00B028_VGPRS(value
) + 1) * 4);
5173 conf
->float_mode
= G_00B028_FLOAT_MODE(value
);
5174 conf
->rsrc1
= value
;
5176 case R_00B02C_SPI_SHADER_PGM_RSRC2_PS
:
5177 conf
->lds_size
= MAX2(conf
->lds_size
, G_00B02C_EXTRA_LDS_SIZE(value
));
5179 case R_00B84C_COMPUTE_PGM_RSRC2
:
5180 conf
->lds_size
= MAX2(conf
->lds_size
, G_00B84C_LDS_SIZE(value
));
5181 conf
->rsrc2
= value
;
5183 case R_0286CC_SPI_PS_INPUT_ENA
:
5184 conf
->spi_ps_input_ena
= value
;
5186 case R_0286D0_SPI_PS_INPUT_ADDR
:
5187 conf
->spi_ps_input_addr
= value
;
5189 case R_0286E8_SPI_TMPRING_SIZE
:
5190 case R_00B860_COMPUTE_TMPRING_SIZE
:
5191 /* WAVESIZE is in units of 256 dwords. */
5192 if (really_needs_scratch
)
5193 conf
->scratch_bytes_per_wave
=
5194 G_00B860_WAVESIZE(value
) * 256 * 4;
5196 case 0x4: /* SPILLED_SGPRS */
5197 conf
->spilled_sgprs
= value
;
5199 case 0x8: /* SPILLED_VGPRS */
5200 conf
->spilled_vgprs
= value
;
5204 static bool printed
;
5207 fprintf(stderr
, "Warning: LLVM emitted unknown "
5208 "config register: 0x%x\n", reg
);
5216 if (!conf
->spi_ps_input_addr
)
5217 conf
->spi_ps_input_addr
= conf
->spi_ps_input_ena
;
5220 void si_shader_apply_scratch_relocs(struct si_shader
*shader
,
5221 uint64_t scratch_va
)
5224 uint32_t scratch_rsrc_dword0
= scratch_va
;
5225 uint32_t scratch_rsrc_dword1
=
5226 S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32);
5228 /* Enable scratch coalescing. */
5229 scratch_rsrc_dword1
|= S_008F04_SWIZZLE_ENABLE(1);
5231 for (i
= 0 ; i
< shader
->binary
.reloc_count
; i
++) {
5232 const struct ac_shader_reloc
*reloc
=
5233 &shader
->binary
.relocs
[i
];
5234 if (!strcmp(scratch_rsrc_dword0_symbol
, reloc
->name
)) {
5235 util_memcpy_cpu_to_le32(shader
->binary
.code
+ reloc
->offset
,
5236 &scratch_rsrc_dword0
, 4);
5237 } else if (!strcmp(scratch_rsrc_dword1_symbol
, reloc
->name
)) {
5238 util_memcpy_cpu_to_le32(shader
->binary
.code
+ reloc
->offset
,
5239 &scratch_rsrc_dword1
, 4);
5244 static unsigned si_get_shader_binary_size(const struct si_shader
*shader
)
5246 unsigned size
= shader
->binary
.code_size
;
5249 size
+= shader
->prolog
->binary
.code_size
;
5250 if (shader
->previous_stage
)
5251 size
+= shader
->previous_stage
->binary
.code_size
;
5252 if (shader
->prolog2
)
5253 size
+= shader
->prolog2
->binary
.code_size
;
5255 size
+= shader
->epilog
->binary
.code_size
;
5259 int si_shader_binary_upload(struct si_screen
*sscreen
, struct si_shader
*shader
)
5261 const struct ac_shader_binary
*prolog
=
5262 shader
->prolog
? &shader
->prolog
->binary
: NULL
;
5263 const struct ac_shader_binary
*previous_stage
=
5264 shader
->previous_stage
? &shader
->previous_stage
->binary
: NULL
;
5265 const struct ac_shader_binary
*prolog2
=
5266 shader
->prolog2
? &shader
->prolog2
->binary
: NULL
;
5267 const struct ac_shader_binary
*epilog
=
5268 shader
->epilog
? &shader
->epilog
->binary
: NULL
;
5269 const struct ac_shader_binary
*mainb
= &shader
->binary
;
5270 unsigned bo_size
= si_get_shader_binary_size(shader
) +
5271 (!epilog
? mainb
->rodata_size
: 0);
5274 assert(!prolog
|| !prolog
->rodata_size
);
5275 assert(!previous_stage
|| !previous_stage
->rodata_size
);
5276 assert(!prolog2
|| !prolog2
->rodata_size
);
5277 assert((!prolog
&& !previous_stage
&& !prolog2
&& !epilog
) ||
5278 !mainb
->rodata_size
);
5279 assert(!epilog
|| !epilog
->rodata_size
);
5281 r600_resource_reference(&shader
->bo
, NULL
);
5282 shader
->bo
= (struct r600_resource
*)
5283 si_aligned_buffer_create(&sscreen
->b
,
5284 sscreen
->cpdma_prefetch_writes_memory
?
5285 0 : R600_RESOURCE_FLAG_READ_ONLY
,
5286 PIPE_USAGE_IMMUTABLE
,
5287 align(bo_size
, SI_CPDMA_ALIGNMENT
),
5293 ptr
= sscreen
->ws
->buffer_map(shader
->bo
->buf
, NULL
,
5294 PIPE_TRANSFER_READ_WRITE
|
5295 PIPE_TRANSFER_UNSYNCHRONIZED
);
5297 /* Don't use util_memcpy_cpu_to_le32. LLVM binaries are
5298 * endian-independent. */
5300 memcpy(ptr
, prolog
->code
, prolog
->code_size
);
5301 ptr
+= prolog
->code_size
;
5303 if (previous_stage
) {
5304 memcpy(ptr
, previous_stage
->code
, previous_stage
->code_size
);
5305 ptr
+= previous_stage
->code_size
;
5308 memcpy(ptr
, prolog2
->code
, prolog2
->code_size
);
5309 ptr
+= prolog2
->code_size
;
5312 memcpy(ptr
, mainb
->code
, mainb
->code_size
);
5313 ptr
+= mainb
->code_size
;
5316 memcpy(ptr
, epilog
->code
, epilog
->code_size
);
5317 else if (mainb
->rodata_size
> 0)
5318 memcpy(ptr
, mainb
->rodata
, mainb
->rodata_size
);
5320 sscreen
->ws
->buffer_unmap(shader
->bo
->buf
);
5324 static void si_shader_dump_disassembly(const struct ac_shader_binary
*binary
,
5325 struct pipe_debug_callback
*debug
,
5326 const char *name
, FILE *file
)
5331 if (binary
->disasm_string
) {
5332 fprintf(file
, "Shader %s disassembly:\n", name
);
5333 fprintf(file
, "%s", binary
->disasm_string
);
5335 if (debug
&& debug
->debug_message
) {
5336 /* Very long debug messages are cut off, so send the
5337 * disassembly one line at a time. This causes more
5338 * overhead, but on the plus side it simplifies
5339 * parsing of resulting logs.
5341 pipe_debug_message(debug
, SHADER_INFO
,
5342 "Shader Disassembly Begin");
5344 line
= binary
->disasm_string
;
5346 p
= util_strchrnul(line
, '\n');
5350 pipe_debug_message(debug
, SHADER_INFO
,
5351 "%.*s", count
, line
);
5359 pipe_debug_message(debug
, SHADER_INFO
,
5360 "Shader Disassembly End");
5363 fprintf(file
, "Shader %s binary:\n", name
);
5364 for (i
= 0; i
< binary
->code_size
; i
+= 4) {
5365 fprintf(file
, "@0x%x: %02x%02x%02x%02x\n", i
,
5366 binary
->code
[i
+ 3], binary
->code
[i
+ 2],
5367 binary
->code
[i
+ 1], binary
->code
[i
]);
5372 static void si_calculate_max_simd_waves(struct si_shader
*shader
)
5374 struct si_screen
*sscreen
= shader
->selector
->screen
;
5375 struct si_shader_config
*conf
= &shader
->config
;
5376 unsigned num_inputs
= shader
->selector
->info
.num_inputs
;
5377 unsigned lds_increment
= sscreen
->info
.chip_class
>= CIK
? 512 : 256;
5378 unsigned lds_per_wave
= 0;
5379 unsigned max_simd_waves
;
5381 switch (sscreen
->info
.family
) {
5382 /* These always have 8 waves: */
5383 case CHIP_POLARIS10
:
5384 case CHIP_POLARIS11
:
5385 case CHIP_POLARIS12
:
5389 max_simd_waves
= 10;
5392 /* Compute LDS usage for PS. */
5393 switch (shader
->selector
->type
) {
5394 case PIPE_SHADER_FRAGMENT
:
5395 /* The minimum usage per wave is (num_inputs * 48). The maximum
5396 * usage is (num_inputs * 48 * 16).
5397 * We can get anything in between and it varies between waves.
5399 * The 48 bytes per input for a single primitive is equal to
5400 * 4 bytes/component * 4 components/input * 3 points.
5402 * Other stages don't know the size at compile time or don't
5403 * allocate LDS per wave, but instead they do it per thread group.
5405 lds_per_wave
= conf
->lds_size
* lds_increment
+
5406 align(num_inputs
* 48, lds_increment
);
5408 case PIPE_SHADER_COMPUTE
:
5409 if (shader
->selector
) {
5410 unsigned max_workgroup_size
=
5411 si_get_max_workgroup_size(shader
);
5412 lds_per_wave
= (conf
->lds_size
* lds_increment
) /
5413 DIV_ROUND_UP(max_workgroup_size
, 64);
5418 /* Compute the per-SIMD wave counts. */
5419 if (conf
->num_sgprs
) {
5420 if (sscreen
->info
.chip_class
>= VI
)
5421 max_simd_waves
= MIN2(max_simd_waves
, 800 / conf
->num_sgprs
);
5423 max_simd_waves
= MIN2(max_simd_waves
, 512 / conf
->num_sgprs
);
5426 if (conf
->num_vgprs
)
5427 max_simd_waves
= MIN2(max_simd_waves
, 256 / conf
->num_vgprs
);
5429 /* LDS is 64KB per CU (4 SIMDs), which is 16KB per SIMD (usage above
5430 * 16KB makes some SIMDs unoccupied). */
5432 max_simd_waves
= MIN2(max_simd_waves
, 16384 / lds_per_wave
);
5434 conf
->max_simd_waves
= max_simd_waves
;
5437 void si_shader_dump_stats_for_shader_db(const struct si_shader
*shader
,
5438 struct pipe_debug_callback
*debug
)
5440 const struct si_shader_config
*conf
= &shader
->config
;
5442 pipe_debug_message(debug
, SHADER_INFO
,
5443 "Shader Stats: SGPRS: %d VGPRS: %d Code Size: %d "
5444 "LDS: %d Scratch: %d Max Waves: %d Spilled SGPRs: %d "
5445 "Spilled VGPRs: %d PrivMem VGPRs: %d",
5446 conf
->num_sgprs
, conf
->num_vgprs
,
5447 si_get_shader_binary_size(shader
),
5448 conf
->lds_size
, conf
->scratch_bytes_per_wave
,
5449 conf
->max_simd_waves
, conf
->spilled_sgprs
,
5450 conf
->spilled_vgprs
, conf
->private_mem_vgprs
);
5453 static void si_shader_dump_stats(struct si_screen
*sscreen
,
5454 const struct si_shader
*shader
,
5457 bool check_debug_option
)
5459 const struct si_shader_config
*conf
= &shader
->config
;
5461 if (!check_debug_option
||
5462 si_can_dump_shader(sscreen
, processor
)) {
5463 if (processor
== PIPE_SHADER_FRAGMENT
) {
5464 fprintf(file
, "*** SHADER CONFIG ***\n"
5465 "SPI_PS_INPUT_ADDR = 0x%04x\n"
5466 "SPI_PS_INPUT_ENA = 0x%04x\n",
5467 conf
->spi_ps_input_addr
, conf
->spi_ps_input_ena
);
5470 fprintf(file
, "*** SHADER STATS ***\n"
5473 "Spilled SGPRs: %d\n"
5474 "Spilled VGPRs: %d\n"
5475 "Private memory VGPRs: %d\n"
5476 "Code Size: %d bytes\n"
5478 "Scratch: %d bytes per wave\n"
5480 "********************\n\n\n",
5481 conf
->num_sgprs
, conf
->num_vgprs
,
5482 conf
->spilled_sgprs
, conf
->spilled_vgprs
,
5483 conf
->private_mem_vgprs
,
5484 si_get_shader_binary_size(shader
),
5485 conf
->lds_size
, conf
->scratch_bytes_per_wave
,
5486 conf
->max_simd_waves
);
5490 const char *si_get_shader_name(const struct si_shader
*shader
, unsigned processor
)
5492 switch (processor
) {
5493 case PIPE_SHADER_VERTEX
:
5494 if (shader
->key
.as_es
)
5495 return "Vertex Shader as ES";
5496 else if (shader
->key
.as_ls
)
5497 return "Vertex Shader as LS";
5499 return "Vertex Shader as VS";
5500 case PIPE_SHADER_TESS_CTRL
:
5501 return "Tessellation Control Shader";
5502 case PIPE_SHADER_TESS_EVAL
:
5503 if (shader
->key
.as_es
)
5504 return "Tessellation Evaluation Shader as ES";
5506 return "Tessellation Evaluation Shader as VS";
5507 case PIPE_SHADER_GEOMETRY
:
5508 if (shader
->is_gs_copy_shader
)
5509 return "GS Copy Shader as VS";
5511 return "Geometry Shader";
5512 case PIPE_SHADER_FRAGMENT
:
5513 return "Pixel Shader";
5514 case PIPE_SHADER_COMPUTE
:
5515 return "Compute Shader";
5517 return "Unknown Shader";
5521 void si_shader_dump(struct si_screen
*sscreen
, const struct si_shader
*shader
,
5522 struct pipe_debug_callback
*debug
, unsigned processor
,
5523 FILE *file
, bool check_debug_option
)
5525 if (!check_debug_option
||
5526 si_can_dump_shader(sscreen
, processor
))
5527 si_dump_shader_key(processor
, shader
, file
);
5529 if (!check_debug_option
&& shader
->binary
.llvm_ir_string
) {
5530 if (shader
->previous_stage
&&
5531 shader
->previous_stage
->binary
.llvm_ir_string
) {
5532 fprintf(file
, "\n%s - previous stage - LLVM IR:\n\n",
5533 si_get_shader_name(shader
, processor
));
5534 fprintf(file
, "%s\n", shader
->previous_stage
->binary
.llvm_ir_string
);
5537 fprintf(file
, "\n%s - main shader part - LLVM IR:\n\n",
5538 si_get_shader_name(shader
, processor
));
5539 fprintf(file
, "%s\n", shader
->binary
.llvm_ir_string
);
5542 if (!check_debug_option
||
5543 (si_can_dump_shader(sscreen
, processor
) &&
5544 !(sscreen
->debug_flags
& DBG(NO_ASM
)))) {
5545 fprintf(file
, "\n%s:\n", si_get_shader_name(shader
, processor
));
5548 si_shader_dump_disassembly(&shader
->prolog
->binary
,
5549 debug
, "prolog", file
);
5550 if (shader
->previous_stage
)
5551 si_shader_dump_disassembly(&shader
->previous_stage
->binary
,
5552 debug
, "previous stage", file
);
5553 if (shader
->prolog2
)
5554 si_shader_dump_disassembly(&shader
->prolog2
->binary
,
5555 debug
, "prolog2", file
);
5557 si_shader_dump_disassembly(&shader
->binary
, debug
, "main", file
);
5560 si_shader_dump_disassembly(&shader
->epilog
->binary
,
5561 debug
, "epilog", file
);
5562 fprintf(file
, "\n");
5565 si_shader_dump_stats(sscreen
, shader
, processor
, file
,
5566 check_debug_option
);
5569 static int si_compile_llvm(struct si_screen
*sscreen
,
5570 struct ac_shader_binary
*binary
,
5571 struct si_shader_config
*conf
,
5572 LLVMTargetMachineRef tm
,
5574 struct pipe_debug_callback
*debug
,
5579 unsigned count
= p_atomic_inc_return(&sscreen
->num_compilations
);
5581 if (si_can_dump_shader(sscreen
, processor
)) {
5582 fprintf(stderr
, "radeonsi: Compiling shader %d\n", count
);
5584 if (!(sscreen
->debug_flags
& (DBG(NO_IR
) | DBG(PREOPT_IR
)))) {
5585 fprintf(stderr
, "%s LLVM IR:\n\n", name
);
5586 ac_dump_module(mod
);
5587 fprintf(stderr
, "\n");
5591 if (sscreen
->record_llvm_ir
) {
5592 char *ir
= LLVMPrintModuleToString(mod
);
5593 binary
->llvm_ir_string
= strdup(ir
);
5594 LLVMDisposeMessage(ir
);
5597 if (!si_replace_shader(count
, binary
)) {
5598 r
= si_llvm_compile(mod
, binary
, tm
, debug
);
5603 si_shader_binary_read_config(binary
, conf
, 0);
5605 /* Enable 64-bit and 16-bit denormals, because there is no performance
5608 * If denormals are enabled, all floating-point output modifiers are
5611 * Don't enable denormals for 32-bit floats, because:
5612 * - Floating-point output modifiers would be ignored by the hw.
5613 * - Some opcodes don't support denormals, such as v_mad_f32. We would
5614 * have to stop using those.
5615 * - SI & CI would be very slow.
5617 conf
->float_mode
|= V_00B028_FP_64_DENORMS
;
5619 FREE(binary
->config
);
5620 FREE(binary
->global_symbol_offsets
);
5621 binary
->config
= NULL
;
5622 binary
->global_symbol_offsets
= NULL
;
5624 /* Some shaders can't have rodata because their binaries can be
5627 if (binary
->rodata_size
&&
5628 (processor
== PIPE_SHADER_VERTEX
||
5629 processor
== PIPE_SHADER_TESS_CTRL
||
5630 processor
== PIPE_SHADER_TESS_EVAL
||
5631 processor
== PIPE_SHADER_FRAGMENT
)) {
5632 fprintf(stderr
, "radeonsi: The shader can't have rodata.");
5639 static void si_llvm_build_ret(struct si_shader_context
*ctx
, LLVMValueRef ret
)
5641 if (LLVMGetTypeKind(LLVMTypeOf(ret
)) == LLVMVoidTypeKind
)
5642 LLVMBuildRetVoid(ctx
->ac
.builder
);
5644 LLVMBuildRet(ctx
->ac
.builder
, ret
);
5647 /* Generate code for the hardware VS shader stage to go with a geometry shader */
5649 si_generate_gs_copy_shader(struct si_screen
*sscreen
,
5650 LLVMTargetMachineRef tm
,
5651 struct si_shader_selector
*gs_selector
,
5652 struct pipe_debug_callback
*debug
)
5654 struct si_shader_context ctx
;
5655 struct si_shader
*shader
;
5656 LLVMBuilderRef builder
;
5657 struct lp_build_tgsi_context
*bld_base
= &ctx
.bld_base
;
5658 struct lp_build_context
*uint
= &bld_base
->uint_bld
;
5659 struct si_shader_output_values
*outputs
;
5660 struct tgsi_shader_info
*gsinfo
= &gs_selector
->info
;
5663 outputs
= MALLOC(gsinfo
->num_outputs
* sizeof(outputs
[0]));
5668 shader
= CALLOC_STRUCT(si_shader
);
5674 /* We can leave the fence as permanently signaled because the GS copy
5675 * shader only becomes visible globally after it has been compiled. */
5676 util_queue_fence_init(&shader
->ready
);
5678 shader
->selector
= gs_selector
;
5679 shader
->is_gs_copy_shader
= true;
5681 si_init_shader_ctx(&ctx
, sscreen
, tm
);
5682 ctx
.shader
= shader
;
5683 ctx
.type
= PIPE_SHADER_VERTEX
;
5685 builder
= ctx
.ac
.builder
;
5687 create_function(&ctx
);
5688 preload_ring_buffers(&ctx
);
5690 LLVMValueRef voffset
=
5691 lp_build_mul_imm(uint
, ctx
.abi
.vertex_id
, 4);
5693 /* Fetch the vertex stream ID.*/
5694 LLVMValueRef stream_id
;
5696 if (gs_selector
->so
.num_outputs
)
5697 stream_id
= unpack_param(&ctx
, ctx
.param_streamout_config
, 24, 2);
5699 stream_id
= ctx
.i32_0
;
5701 /* Fill in output information. */
5702 for (i
= 0; i
< gsinfo
->num_outputs
; ++i
) {
5703 outputs
[i
].semantic_name
= gsinfo
->output_semantic_name
[i
];
5704 outputs
[i
].semantic_index
= gsinfo
->output_semantic_index
[i
];
5706 for (int chan
= 0; chan
< 4; chan
++) {
5707 outputs
[i
].vertex_stream
[chan
] =
5708 (gsinfo
->output_streams
[i
] >> (2 * chan
)) & 3;
5712 LLVMBasicBlockRef end_bb
;
5713 LLVMValueRef switch_inst
;
5715 end_bb
= LLVMAppendBasicBlockInContext(ctx
.ac
.context
, ctx
.main_fn
, "end");
5716 switch_inst
= LLVMBuildSwitch(builder
, stream_id
, end_bb
, 4);
5718 for (int stream
= 0; stream
< 4; stream
++) {
5719 LLVMBasicBlockRef bb
;
5722 if (!gsinfo
->num_stream_output_components
[stream
])
5725 if (stream
> 0 && !gs_selector
->so
.num_outputs
)
5728 bb
= LLVMInsertBasicBlockInContext(ctx
.ac
.context
, end_bb
, "out");
5729 LLVMAddCase(switch_inst
, LLVMConstInt(ctx
.i32
, stream
, 0), bb
);
5730 LLVMPositionBuilderAtEnd(builder
, bb
);
5732 /* Fetch vertex data from GSVS ring */
5734 for (i
= 0; i
< gsinfo
->num_outputs
; ++i
) {
5735 for (unsigned chan
= 0; chan
< 4; chan
++) {
5736 if (!(gsinfo
->output_usagemask
[i
] & (1 << chan
)) ||
5737 outputs
[i
].vertex_stream
[chan
] != stream
) {
5738 outputs
[i
].values
[chan
] = ctx
.bld_base
.base
.undef
;
5742 LLVMValueRef soffset
= LLVMConstInt(ctx
.i32
,
5743 offset
* gs_selector
->gs_max_out_vertices
* 16 * 4, 0);
5746 outputs
[i
].values
[chan
] =
5747 ac_build_buffer_load(&ctx
.ac
,
5748 ctx
.gsvs_ring
[0], 1,
5755 /* Streamout and exports. */
5756 if (gs_selector
->so
.num_outputs
) {
5757 si_llvm_emit_streamout(&ctx
, outputs
,
5758 gsinfo
->num_outputs
,
5763 si_llvm_export_vs(&ctx
, outputs
, gsinfo
->num_outputs
);
5765 LLVMBuildBr(builder
, end_bb
);
5768 LLVMPositionBuilderAtEnd(builder
, end_bb
);
5770 LLVMBuildRetVoid(ctx
.ac
.builder
);
5772 ctx
.type
= PIPE_SHADER_GEOMETRY
; /* override for shader dumping */
5773 si_llvm_optimize_module(&ctx
);
5775 r
= si_compile_llvm(sscreen
, &ctx
.shader
->binary
,
5776 &ctx
.shader
->config
, ctx
.tm
,
5778 debug
, PIPE_SHADER_GEOMETRY
,
5781 if (si_can_dump_shader(sscreen
, PIPE_SHADER_GEOMETRY
))
5782 fprintf(stderr
, "GS Copy Shader:\n");
5783 si_shader_dump(sscreen
, ctx
.shader
, debug
,
5784 PIPE_SHADER_GEOMETRY
, stderr
, true);
5785 r
= si_shader_binary_upload(sscreen
, ctx
.shader
);
5788 si_llvm_dispose(&ctx
);
5799 static void si_dump_shader_key_vs(const struct si_shader_key
*key
,
5800 const struct si_vs_prolog_bits
*prolog
,
5801 const char *prefix
, FILE *f
)
5803 fprintf(f
, " %s.instance_divisor_is_one = %u\n",
5804 prefix
, prolog
->instance_divisor_is_one
);
5805 fprintf(f
, " %s.instance_divisor_is_fetched = %u\n",
5806 prefix
, prolog
->instance_divisor_is_fetched
);
5807 fprintf(f
, " %s.ls_vgpr_fix = %u\n",
5808 prefix
, prolog
->ls_vgpr_fix
);
5810 fprintf(f
, " mono.vs.fix_fetch = {");
5811 for (int i
= 0; i
< SI_MAX_ATTRIBS
; i
++)
5812 fprintf(f
, !i
? "%u" : ", %u", key
->mono
.vs_fix_fetch
[i
]);
5816 static void si_dump_shader_key(unsigned processor
, const struct si_shader
*shader
,
5819 const struct si_shader_key
*key
= &shader
->key
;
5821 fprintf(f
, "SHADER KEY\n");
5823 switch (processor
) {
5824 case PIPE_SHADER_VERTEX
:
5825 si_dump_shader_key_vs(key
, &key
->part
.vs
.prolog
,
5826 "part.vs.prolog", f
);
5827 fprintf(f
, " as_es = %u\n", key
->as_es
);
5828 fprintf(f
, " as_ls = %u\n", key
->as_ls
);
5829 fprintf(f
, " mono.u.vs_export_prim_id = %u\n",
5830 key
->mono
.u
.vs_export_prim_id
);
5833 case PIPE_SHADER_TESS_CTRL
:
5834 if (shader
->selector
->screen
->info
.chip_class
>= GFX9
) {
5835 si_dump_shader_key_vs(key
, &key
->part
.tcs
.ls_prolog
,
5836 "part.tcs.ls_prolog", f
);
5838 fprintf(f
, " part.tcs.epilog.prim_mode = %u\n", key
->part
.tcs
.epilog
.prim_mode
);
5839 fprintf(f
, " mono.u.ff_tcs_inputs_to_copy = 0x%"PRIx64
"\n", key
->mono
.u
.ff_tcs_inputs_to_copy
);
5842 case PIPE_SHADER_TESS_EVAL
:
5843 fprintf(f
, " as_es = %u\n", key
->as_es
);
5844 fprintf(f
, " mono.u.vs_export_prim_id = %u\n",
5845 key
->mono
.u
.vs_export_prim_id
);
5848 case PIPE_SHADER_GEOMETRY
:
5849 if (shader
->is_gs_copy_shader
)
5852 if (shader
->selector
->screen
->info
.chip_class
>= GFX9
&&
5853 key
->part
.gs
.es
->type
== PIPE_SHADER_VERTEX
) {
5854 si_dump_shader_key_vs(key
, &key
->part
.gs
.vs_prolog
,
5855 "part.gs.vs_prolog", f
);
5857 fprintf(f
, " part.gs.prolog.tri_strip_adj_fix = %u\n", key
->part
.gs
.prolog
.tri_strip_adj_fix
);
5860 case PIPE_SHADER_COMPUTE
:
5863 case PIPE_SHADER_FRAGMENT
:
5864 fprintf(f
, " part.ps.prolog.color_two_side = %u\n", key
->part
.ps
.prolog
.color_two_side
);
5865 fprintf(f
, " part.ps.prolog.flatshade_colors = %u\n", key
->part
.ps
.prolog
.flatshade_colors
);
5866 fprintf(f
, " part.ps.prolog.poly_stipple = %u\n", key
->part
.ps
.prolog
.poly_stipple
);
5867 fprintf(f
, " part.ps.prolog.force_persp_sample_interp = %u\n", key
->part
.ps
.prolog
.force_persp_sample_interp
);
5868 fprintf(f
, " part.ps.prolog.force_linear_sample_interp = %u\n", key
->part
.ps
.prolog
.force_linear_sample_interp
);
5869 fprintf(f
, " part.ps.prolog.force_persp_center_interp = %u\n", key
->part
.ps
.prolog
.force_persp_center_interp
);
5870 fprintf(f
, " part.ps.prolog.force_linear_center_interp = %u\n", key
->part
.ps
.prolog
.force_linear_center_interp
);
5871 fprintf(f
, " part.ps.prolog.bc_optimize_for_persp = %u\n", key
->part
.ps
.prolog
.bc_optimize_for_persp
);
5872 fprintf(f
, " part.ps.prolog.bc_optimize_for_linear = %u\n", key
->part
.ps
.prolog
.bc_optimize_for_linear
);
5873 fprintf(f
, " part.ps.epilog.spi_shader_col_format = 0x%x\n", key
->part
.ps
.epilog
.spi_shader_col_format
);
5874 fprintf(f
, " part.ps.epilog.color_is_int8 = 0x%X\n", key
->part
.ps
.epilog
.color_is_int8
);
5875 fprintf(f
, " part.ps.epilog.color_is_int10 = 0x%X\n", key
->part
.ps
.epilog
.color_is_int10
);
5876 fprintf(f
, " part.ps.epilog.last_cbuf = %u\n", key
->part
.ps
.epilog
.last_cbuf
);
5877 fprintf(f
, " part.ps.epilog.alpha_func = %u\n", key
->part
.ps
.epilog
.alpha_func
);
5878 fprintf(f
, " part.ps.epilog.alpha_to_one = %u\n", key
->part
.ps
.epilog
.alpha_to_one
);
5879 fprintf(f
, " part.ps.epilog.poly_line_smoothing = %u\n", key
->part
.ps
.epilog
.poly_line_smoothing
);
5880 fprintf(f
, " part.ps.epilog.clamp_color = %u\n", key
->part
.ps
.epilog
.clamp_color
);
5887 if ((processor
== PIPE_SHADER_GEOMETRY
||
5888 processor
== PIPE_SHADER_TESS_EVAL
||
5889 processor
== PIPE_SHADER_VERTEX
) &&
5890 !key
->as_es
&& !key
->as_ls
) {
5891 fprintf(f
, " opt.kill_outputs = 0x%"PRIx64
"\n", key
->opt
.kill_outputs
);
5892 fprintf(f
, " opt.clip_disable = %u\n", key
->opt
.clip_disable
);
5896 static void si_init_shader_ctx(struct si_shader_context
*ctx
,
5897 struct si_screen
*sscreen
,
5898 LLVMTargetMachineRef tm
)
5900 struct lp_build_tgsi_context
*bld_base
;
5902 si_llvm_context_init(ctx
, sscreen
, tm
);
5904 bld_base
= &ctx
->bld_base
;
5905 bld_base
->emit_fetch_funcs
[TGSI_FILE_CONSTANT
] = fetch_constant
;
5907 bld_base
->op_actions
[TGSI_OPCODE_INTERP_CENTROID
] = interp_action
;
5908 bld_base
->op_actions
[TGSI_OPCODE_INTERP_SAMPLE
] = interp_action
;
5909 bld_base
->op_actions
[TGSI_OPCODE_INTERP_OFFSET
] = interp_action
;
5911 bld_base
->op_actions
[TGSI_OPCODE_MEMBAR
].emit
= membar_emit
;
5913 bld_base
->op_actions
[TGSI_OPCODE_CLOCK
].emit
= clock_emit
;
5915 bld_base
->op_actions
[TGSI_OPCODE_DDX
].emit
= si_llvm_emit_ddxy
;
5916 bld_base
->op_actions
[TGSI_OPCODE_DDY
].emit
= si_llvm_emit_ddxy
;
5917 bld_base
->op_actions
[TGSI_OPCODE_DDX_FINE
].emit
= si_llvm_emit_ddxy
;
5918 bld_base
->op_actions
[TGSI_OPCODE_DDY_FINE
].emit
= si_llvm_emit_ddxy
;
5920 bld_base
->op_actions
[TGSI_OPCODE_VOTE_ALL
].emit
= vote_all_emit
;
5921 bld_base
->op_actions
[TGSI_OPCODE_VOTE_ANY
].emit
= vote_any_emit
;
5922 bld_base
->op_actions
[TGSI_OPCODE_VOTE_EQ
].emit
= vote_eq_emit
;
5923 bld_base
->op_actions
[TGSI_OPCODE_BALLOT
].emit
= ballot_emit
;
5924 bld_base
->op_actions
[TGSI_OPCODE_READ_FIRST
].intr_name
= "llvm.amdgcn.readfirstlane";
5925 bld_base
->op_actions
[TGSI_OPCODE_READ_FIRST
].emit
= read_lane_emit
;
5926 bld_base
->op_actions
[TGSI_OPCODE_READ_INVOC
].intr_name
= "llvm.amdgcn.readlane";
5927 bld_base
->op_actions
[TGSI_OPCODE_READ_INVOC
].fetch_args
= read_invoc_fetch_args
;
5928 bld_base
->op_actions
[TGSI_OPCODE_READ_INVOC
].emit
= read_lane_emit
;
5930 bld_base
->op_actions
[TGSI_OPCODE_EMIT
].emit
= si_tgsi_emit_vertex
;
5931 bld_base
->op_actions
[TGSI_OPCODE_ENDPRIM
].emit
= si_tgsi_emit_primitive
;
5932 bld_base
->op_actions
[TGSI_OPCODE_BARRIER
].emit
= si_llvm_emit_barrier
;
5935 static void si_optimize_vs_outputs(struct si_shader_context
*ctx
)
5937 struct si_shader
*shader
= ctx
->shader
;
5938 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
5940 if ((ctx
->type
!= PIPE_SHADER_VERTEX
&&
5941 ctx
->type
!= PIPE_SHADER_TESS_EVAL
) ||
5942 shader
->key
.as_ls
||
5946 ac_optimize_vs_outputs(&ctx
->ac
,
5948 shader
->info
.vs_output_param_offset
,
5950 &shader
->info
.nr_param_exports
);
5953 static void si_count_scratch_private_memory(struct si_shader_context
*ctx
)
5955 ctx
->shader
->config
.private_mem_vgprs
= 0;
5957 /* Process all LLVM instructions. */
5958 LLVMBasicBlockRef bb
= LLVMGetFirstBasicBlock(ctx
->main_fn
);
5960 LLVMValueRef next
= LLVMGetFirstInstruction(bb
);
5963 LLVMValueRef inst
= next
;
5964 next
= LLVMGetNextInstruction(next
);
5966 if (LLVMGetInstructionOpcode(inst
) != LLVMAlloca
)
5969 LLVMTypeRef type
= LLVMGetElementType(LLVMTypeOf(inst
));
5970 /* No idea why LLVM aligns allocas to 4 elements. */
5971 unsigned alignment
= LLVMGetAlignment(inst
);
5972 unsigned dw_size
= align(ac_get_type_size(type
) / 4, alignment
);
5973 ctx
->shader
->config
.private_mem_vgprs
+= dw_size
;
5975 bb
= LLVMGetNextBasicBlock(bb
);
5979 static void si_init_exec_from_input(struct si_shader_context
*ctx
,
5980 unsigned param
, unsigned bitoffset
)
5982 LLVMValueRef args
[] = {
5983 LLVMGetParam(ctx
->main_fn
, param
),
5984 LLVMConstInt(ctx
->i32
, bitoffset
, 0),
5986 lp_build_intrinsic(ctx
->ac
.builder
,
5987 "llvm.amdgcn.init.exec.from.input",
5988 ctx
->voidt
, args
, 2, LP_FUNC_ATTR_CONVERGENT
);
5991 static bool si_vs_needs_prolog(const struct si_shader_selector
*sel
,
5992 const struct si_vs_prolog_bits
*key
)
5994 /* VGPR initialization fixup for Vega10 and Raven is always done in the
5996 return sel
->vs_needs_prolog
|| key
->ls_vgpr_fix
;
5999 static bool si_compile_tgsi_main(struct si_shader_context
*ctx
,
6002 struct si_shader
*shader
= ctx
->shader
;
6003 struct si_shader_selector
*sel
= shader
->selector
;
6004 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
6006 // TODO clean all this up!
6007 switch (ctx
->type
) {
6008 case PIPE_SHADER_VERTEX
:
6009 ctx
->load_input
= declare_input_vs
;
6010 if (shader
->key
.as_ls
)
6011 ctx
->abi
.emit_outputs
= si_llvm_emit_ls_epilogue
;
6012 else if (shader
->key
.as_es
)
6013 ctx
->abi
.emit_outputs
= si_llvm_emit_es_epilogue
;
6015 ctx
->abi
.emit_outputs
= si_llvm_emit_vs_epilogue
;
6016 bld_base
->emit_epilogue
= si_tgsi_emit_epilogue
;
6018 case PIPE_SHADER_TESS_CTRL
:
6019 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = fetch_input_tcs
;
6020 ctx
->abi
.load_tess_varyings
= si_nir_load_tcs_varyings
;
6021 bld_base
->emit_fetch_funcs
[TGSI_FILE_OUTPUT
] = fetch_output_tcs
;
6022 bld_base
->emit_store
= store_output_tcs
;
6023 ctx
->abi
.store_tcs_outputs
= si_nir_store_output_tcs
;
6024 ctx
->abi
.emit_outputs
= si_llvm_emit_tcs_epilogue
;
6025 ctx
->abi
.load_patch_vertices_in
= si_load_patch_vertices_in
;
6026 bld_base
->emit_epilogue
= si_tgsi_emit_epilogue
;
6028 case PIPE_SHADER_TESS_EVAL
:
6029 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = fetch_input_tes
;
6030 ctx
->abi
.load_tess_varyings
= si_nir_load_input_tes
;
6031 ctx
->abi
.load_tess_coord
= si_load_tess_coord
;
6032 ctx
->abi
.load_tess_level
= si_load_tess_level
;
6033 ctx
->abi
.load_patch_vertices_in
= si_load_patch_vertices_in
;
6034 if (shader
->key
.as_es
)
6035 ctx
->abi
.emit_outputs
= si_llvm_emit_es_epilogue
;
6037 ctx
->abi
.emit_outputs
= si_llvm_emit_vs_epilogue
;
6038 bld_base
->emit_epilogue
= si_tgsi_emit_epilogue
;
6040 case PIPE_SHADER_GEOMETRY
:
6041 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = fetch_input_gs
;
6042 ctx
->abi
.load_inputs
= si_nir_load_input_gs
;
6043 ctx
->abi
.emit_vertex
= si_llvm_emit_vertex
;
6044 ctx
->abi
.emit_primitive
= si_llvm_emit_primitive
;
6045 ctx
->abi
.emit_outputs
= si_llvm_emit_gs_epilogue
;
6046 bld_base
->emit_epilogue
= si_tgsi_emit_gs_epilogue
;
6048 case PIPE_SHADER_FRAGMENT
:
6049 ctx
->load_input
= declare_input_fs
;
6050 ctx
->abi
.emit_outputs
= si_llvm_return_fs_outputs
;
6051 bld_base
->emit_epilogue
= si_tgsi_emit_epilogue
;
6052 ctx
->abi
.lookup_interp_param
= si_nir_lookup_interp_param
;
6053 ctx
->abi
.load_sample_position
= load_sample_position
;
6055 case PIPE_SHADER_COMPUTE
:
6058 assert(!"Unsupported shader type");
6062 ctx
->abi
.load_ubo
= load_ubo
;
6063 ctx
->abi
.load_ssbo
= load_ssbo
;
6065 create_function(ctx
);
6066 preload_ring_buffers(ctx
);
6068 /* For GFX9 merged shaders:
6069 * - Set EXEC for the first shader. If the prolog is present, set
6070 * EXEC there instead.
6071 * - Add a barrier before the second shader.
6072 * - In the second shader, reset EXEC to ~0 and wrap the main part in
6073 * an if-statement. This is required for correctness in geometry
6074 * shaders, to ensure that empty GS waves do not send GS_EMIT and
6077 * For monolithic merged shaders, the first shader is wrapped in an
6078 * if-block together with its prolog in si_build_wrapper_function.
6080 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
6081 if (!is_monolithic
&&
6082 sel
->info
.num_instructions
> 1 && /* not empty shader */
6083 (shader
->key
.as_es
|| shader
->key
.as_ls
) &&
6084 (ctx
->type
== PIPE_SHADER_TESS_EVAL
||
6085 (ctx
->type
== PIPE_SHADER_VERTEX
&&
6086 !si_vs_needs_prolog(sel
, &shader
->key
.part
.vs
.prolog
)))) {
6087 si_init_exec_from_input(ctx
,
6088 ctx
->param_merged_wave_info
, 0);
6089 } else if (ctx
->type
== PIPE_SHADER_TESS_CTRL
||
6090 ctx
->type
== PIPE_SHADER_GEOMETRY
) {
6092 ac_init_exec_full_mask(&ctx
->ac
);
6094 /* The barrier must execute for all shaders in a
6097 si_llvm_emit_barrier(NULL
, bld_base
, NULL
);
6099 LLVMValueRef num_threads
= unpack_param(ctx
, ctx
->param_merged_wave_info
, 8, 8);
6101 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntULT
,
6102 ac_get_thread_id(&ctx
->ac
), num_threads
, "");
6103 lp_build_if(&ctx
->merged_wrap_if_state
, &ctx
->gallivm
, ena
);
6107 if (ctx
->type
== PIPE_SHADER_TESS_CTRL
&&
6108 sel
->tcs_info
.tessfactors_are_def_in_all_invocs
) {
6109 for (unsigned i
= 0; i
< 6; i
++) {
6110 ctx
->invoc0_tess_factors
[i
] =
6111 lp_build_alloca_undef(&ctx
->gallivm
, ctx
->i32
, "");
6115 if (ctx
->type
== PIPE_SHADER_GEOMETRY
) {
6117 for (i
= 0; i
< 4; i
++) {
6118 ctx
->gs_next_vertex
[i
] =
6119 lp_build_alloca(&ctx
->gallivm
,
6124 if (sel
->force_correct_derivs_after_kill
) {
6125 ctx
->postponed_kill
= lp_build_alloca_undef(&ctx
->gallivm
, ctx
->i1
, "");
6126 /* true = don't kill. */
6127 LLVMBuildStore(ctx
->ac
.builder
, LLVMConstInt(ctx
->i1
, 1, 0),
6128 ctx
->postponed_kill
);
6132 if (!lp_build_tgsi_llvm(bld_base
, sel
->tokens
)) {
6133 fprintf(stderr
, "Failed to translate shader from TGSI to LLVM\n");
6137 if (!si_nir_build_llvm(ctx
, sel
->nir
)) {
6138 fprintf(stderr
, "Failed to translate shader from NIR to LLVM\n");
6143 si_llvm_build_ret(ctx
, ctx
->return_value
);
6148 * Compute the VS prolog key, which contains all the information needed to
6149 * build the VS prolog function, and set shader->info bits where needed.
6151 * \param info Shader info of the vertex shader.
6152 * \param num_input_sgprs Number of input SGPRs for the vertex shader.
6153 * \param prolog_key Key of the VS prolog
6154 * \param shader_out The vertex shader, or the next shader if merging LS+HS or ES+GS.
6155 * \param key Output shader part key.
6157 static void si_get_vs_prolog_key(const struct tgsi_shader_info
*info
,
6158 unsigned num_input_sgprs
,
6159 const struct si_vs_prolog_bits
*prolog_key
,
6160 struct si_shader
*shader_out
,
6161 union si_shader_part_key
*key
)
6163 memset(key
, 0, sizeof(*key
));
6164 key
->vs_prolog
.states
= *prolog_key
;
6165 key
->vs_prolog
.num_input_sgprs
= num_input_sgprs
;
6166 key
->vs_prolog
.last_input
= MAX2(1, info
->num_inputs
) - 1;
6167 key
->vs_prolog
.as_ls
= shader_out
->key
.as_ls
;
6168 key
->vs_prolog
.as_es
= shader_out
->key
.as_es
;
6170 if (shader_out
->selector
->type
== PIPE_SHADER_TESS_CTRL
) {
6171 key
->vs_prolog
.as_ls
= 1;
6172 key
->vs_prolog
.num_merged_next_stage_vgprs
= 2;
6173 } else if (shader_out
->selector
->type
== PIPE_SHADER_GEOMETRY
) {
6174 key
->vs_prolog
.as_es
= 1;
6175 key
->vs_prolog
.num_merged_next_stage_vgprs
= 5;
6178 /* Enable loading the InstanceID VGPR. */
6179 uint16_t input_mask
= u_bit_consecutive(0, info
->num_inputs
);
6181 if ((key
->vs_prolog
.states
.instance_divisor_is_one
|
6182 key
->vs_prolog
.states
.instance_divisor_is_fetched
) & input_mask
)
6183 shader_out
->info
.uses_instanceid
= true;
6187 * Compute the PS prolog key, which contains all the information needed to
6188 * build the PS prolog function, and set related bits in shader->config.
6190 static void si_get_ps_prolog_key(struct si_shader
*shader
,
6191 union si_shader_part_key
*key
,
6192 bool separate_prolog
)
6194 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
6196 memset(key
, 0, sizeof(*key
));
6197 key
->ps_prolog
.states
= shader
->key
.part
.ps
.prolog
;
6198 key
->ps_prolog
.colors_read
= info
->colors_read
;
6199 key
->ps_prolog
.num_input_sgprs
= shader
->info
.num_input_sgprs
;
6200 key
->ps_prolog
.num_input_vgprs
= shader
->info
.num_input_vgprs
;
6201 key
->ps_prolog
.wqm
= info
->uses_derivatives
&&
6202 (key
->ps_prolog
.colors_read
||
6203 key
->ps_prolog
.states
.force_persp_sample_interp
||
6204 key
->ps_prolog
.states
.force_linear_sample_interp
||
6205 key
->ps_prolog
.states
.force_persp_center_interp
||
6206 key
->ps_prolog
.states
.force_linear_center_interp
||
6207 key
->ps_prolog
.states
.bc_optimize_for_persp
||
6208 key
->ps_prolog
.states
.bc_optimize_for_linear
);
6209 key
->ps_prolog
.ancillary_vgpr_index
= shader
->info
.ancillary_vgpr_index
;
6211 if (info
->colors_read
) {
6212 unsigned *color
= shader
->selector
->color_attr_index
;
6214 if (shader
->key
.part
.ps
.prolog
.color_two_side
) {
6215 /* BCOLORs are stored after the last input. */
6216 key
->ps_prolog
.num_interp_inputs
= info
->num_inputs
;
6217 key
->ps_prolog
.face_vgpr_index
= shader
->info
.face_vgpr_index
;
6218 shader
->config
.spi_ps_input_ena
|= S_0286CC_FRONT_FACE_ENA(1);
6221 for (unsigned i
= 0; i
< 2; i
++) {
6222 unsigned interp
= info
->input_interpolate
[color
[i
]];
6223 unsigned location
= info
->input_interpolate_loc
[color
[i
]];
6225 if (!(info
->colors_read
& (0xf << i
*4)))
6228 key
->ps_prolog
.color_attr_index
[i
] = color
[i
];
6230 if (shader
->key
.part
.ps
.prolog
.flatshade_colors
&&
6231 interp
== TGSI_INTERPOLATE_COLOR
)
6232 interp
= TGSI_INTERPOLATE_CONSTANT
;
6235 case TGSI_INTERPOLATE_CONSTANT
:
6236 key
->ps_prolog
.color_interp_vgpr_index
[i
] = -1;
6238 case TGSI_INTERPOLATE_PERSPECTIVE
:
6239 case TGSI_INTERPOLATE_COLOR
:
6240 /* Force the interpolation location for colors here. */
6241 if (shader
->key
.part
.ps
.prolog
.force_persp_sample_interp
)
6242 location
= TGSI_INTERPOLATE_LOC_SAMPLE
;
6243 if (shader
->key
.part
.ps
.prolog
.force_persp_center_interp
)
6244 location
= TGSI_INTERPOLATE_LOC_CENTER
;
6247 case TGSI_INTERPOLATE_LOC_SAMPLE
:
6248 key
->ps_prolog
.color_interp_vgpr_index
[i
] = 0;
6249 shader
->config
.spi_ps_input_ena
|=
6250 S_0286CC_PERSP_SAMPLE_ENA(1);
6252 case TGSI_INTERPOLATE_LOC_CENTER
:
6253 key
->ps_prolog
.color_interp_vgpr_index
[i
] = 2;
6254 shader
->config
.spi_ps_input_ena
|=
6255 S_0286CC_PERSP_CENTER_ENA(1);
6257 case TGSI_INTERPOLATE_LOC_CENTROID
:
6258 key
->ps_prolog
.color_interp_vgpr_index
[i
] = 4;
6259 shader
->config
.spi_ps_input_ena
|=
6260 S_0286CC_PERSP_CENTROID_ENA(1);
6266 case TGSI_INTERPOLATE_LINEAR
:
6267 /* Force the interpolation location for colors here. */
6268 if (shader
->key
.part
.ps
.prolog
.force_linear_sample_interp
)
6269 location
= TGSI_INTERPOLATE_LOC_SAMPLE
;
6270 if (shader
->key
.part
.ps
.prolog
.force_linear_center_interp
)
6271 location
= TGSI_INTERPOLATE_LOC_CENTER
;
6273 /* The VGPR assignment for non-monolithic shaders
6274 * works because InitialPSInputAddr is set on the
6275 * main shader and PERSP_PULL_MODEL is never used.
6278 case TGSI_INTERPOLATE_LOC_SAMPLE
:
6279 key
->ps_prolog
.color_interp_vgpr_index
[i
] =
6280 separate_prolog
? 6 : 9;
6281 shader
->config
.spi_ps_input_ena
|=
6282 S_0286CC_LINEAR_SAMPLE_ENA(1);
6284 case TGSI_INTERPOLATE_LOC_CENTER
:
6285 key
->ps_prolog
.color_interp_vgpr_index
[i
] =
6286 separate_prolog
? 8 : 11;
6287 shader
->config
.spi_ps_input_ena
|=
6288 S_0286CC_LINEAR_CENTER_ENA(1);
6290 case TGSI_INTERPOLATE_LOC_CENTROID
:
6291 key
->ps_prolog
.color_interp_vgpr_index
[i
] =
6292 separate_prolog
? 10 : 13;
6293 shader
->config
.spi_ps_input_ena
|=
6294 S_0286CC_LINEAR_CENTROID_ENA(1);
6308 * Check whether a PS prolog is required based on the key.
6310 static bool si_need_ps_prolog(const union si_shader_part_key
*key
)
6312 return key
->ps_prolog
.colors_read
||
6313 key
->ps_prolog
.states
.force_persp_sample_interp
||
6314 key
->ps_prolog
.states
.force_linear_sample_interp
||
6315 key
->ps_prolog
.states
.force_persp_center_interp
||
6316 key
->ps_prolog
.states
.force_linear_center_interp
||
6317 key
->ps_prolog
.states
.bc_optimize_for_persp
||
6318 key
->ps_prolog
.states
.bc_optimize_for_linear
||
6319 key
->ps_prolog
.states
.poly_stipple
||
6320 key
->ps_prolog
.states
.samplemask_log_ps_iter
;
6324 * Compute the PS epilog key, which contains all the information needed to
6325 * build the PS epilog function.
6327 static void si_get_ps_epilog_key(struct si_shader
*shader
,
6328 union si_shader_part_key
*key
)
6330 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
6331 memset(key
, 0, sizeof(*key
));
6332 key
->ps_epilog
.colors_written
= info
->colors_written
;
6333 key
->ps_epilog
.writes_z
= info
->writes_z
;
6334 key
->ps_epilog
.writes_stencil
= info
->writes_stencil
;
6335 key
->ps_epilog
.writes_samplemask
= info
->writes_samplemask
;
6336 key
->ps_epilog
.states
= shader
->key
.part
.ps
.epilog
;
6340 * Build the GS prolog function. Rotate the input vertices for triangle strips
6343 static void si_build_gs_prolog_function(struct si_shader_context
*ctx
,
6344 union si_shader_part_key
*key
)
6346 unsigned num_sgprs
, num_vgprs
;
6347 struct si_function_info fninfo
;
6348 LLVMBuilderRef builder
= ctx
->ac
.builder
;
6349 LLVMTypeRef returns
[48];
6350 LLVMValueRef func
, ret
;
6352 si_init_function_info(&fninfo
);
6354 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
6355 num_sgprs
= 8 + GFX9_GS_NUM_USER_SGPR
;
6356 num_vgprs
= 5; /* ES inputs are not needed by GS */
6358 num_sgprs
= GFX6_GS_NUM_USER_SGPR
+ 2;
6362 for (unsigned i
= 0; i
< num_sgprs
; ++i
) {
6363 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
6364 returns
[i
] = ctx
->i32
;
6367 for (unsigned i
= 0; i
< num_vgprs
; ++i
) {
6368 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
);
6369 returns
[num_sgprs
+ i
] = ctx
->f32
;
6372 /* Create the function. */
6373 si_create_function(ctx
, "gs_prolog", returns
, num_sgprs
+ num_vgprs
,
6375 func
= ctx
->main_fn
;
6377 /* Set the full EXEC mask for the prolog, because we are only fiddling
6378 * with registers here. The main shader part will set the correct EXEC
6381 if (ctx
->screen
->info
.chip_class
>= GFX9
&& !key
->gs_prolog
.is_monolithic
)
6382 ac_init_exec_full_mask(&ctx
->ac
);
6384 /* Copy inputs to outputs. This should be no-op, as the registers match,
6385 * but it will prevent the compiler from overwriting them unintentionally.
6387 ret
= ctx
->return_value
;
6388 for (unsigned i
= 0; i
< num_sgprs
; i
++) {
6389 LLVMValueRef p
= LLVMGetParam(func
, i
);
6390 ret
= LLVMBuildInsertValue(builder
, ret
, p
, i
, "");
6392 for (unsigned i
= 0; i
< num_vgprs
; i
++) {
6393 LLVMValueRef p
= LLVMGetParam(func
, num_sgprs
+ i
);
6394 p
= ac_to_float(&ctx
->ac
, p
);
6395 ret
= LLVMBuildInsertValue(builder
, ret
, p
, num_sgprs
+ i
, "");
6398 if (key
->gs_prolog
.states
.tri_strip_adj_fix
) {
6399 /* Remap the input vertices for every other primitive. */
6400 const unsigned gfx6_vtx_params
[6] = {
6408 const unsigned gfx9_vtx_params
[3] = {
6413 LLVMValueRef vtx_in
[6], vtx_out
[6];
6414 LLVMValueRef prim_id
, rotate
;
6416 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
6417 for (unsigned i
= 0; i
< 3; i
++) {
6418 vtx_in
[i
*2] = unpack_param(ctx
, gfx9_vtx_params
[i
], 0, 16);
6419 vtx_in
[i
*2+1] = unpack_param(ctx
, gfx9_vtx_params
[i
], 16, 16);
6422 for (unsigned i
= 0; i
< 6; i
++)
6423 vtx_in
[i
] = LLVMGetParam(func
, gfx6_vtx_params
[i
]);
6426 prim_id
= LLVMGetParam(func
, num_sgprs
+ 2);
6427 rotate
= LLVMBuildTrunc(builder
, prim_id
, ctx
->i1
, "");
6429 for (unsigned i
= 0; i
< 6; ++i
) {
6430 LLVMValueRef base
, rotated
;
6432 rotated
= vtx_in
[(i
+ 4) % 6];
6433 vtx_out
[i
] = LLVMBuildSelect(builder
, rotate
, rotated
, base
, "");
6436 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
6437 for (unsigned i
= 0; i
< 3; i
++) {
6438 LLVMValueRef hi
, out
;
6440 hi
= LLVMBuildShl(builder
, vtx_out
[i
*2+1],
6441 LLVMConstInt(ctx
->i32
, 16, 0), "");
6442 out
= LLVMBuildOr(builder
, vtx_out
[i
*2], hi
, "");
6443 out
= ac_to_float(&ctx
->ac
, out
);
6444 ret
= LLVMBuildInsertValue(builder
, ret
, out
,
6445 gfx9_vtx_params
[i
], "");
6448 for (unsigned i
= 0; i
< 6; i
++) {
6451 out
= ac_to_float(&ctx
->ac
, vtx_out
[i
]);
6452 ret
= LLVMBuildInsertValue(builder
, ret
, out
,
6453 gfx6_vtx_params
[i
], "");
6458 LLVMBuildRet(builder
, ret
);
6462 * Given a list of shader part functions, build a wrapper function that
6463 * runs them in sequence to form a monolithic shader.
6465 static void si_build_wrapper_function(struct si_shader_context
*ctx
,
6466 LLVMValueRef
*parts
,
6469 unsigned next_shader_first_part
)
6471 LLVMBuilderRef builder
= ctx
->ac
.builder
;
6472 /* PS epilog has one arg per color component; gfx9 merged shader
6473 * prologs need to forward 32 user SGPRs.
6475 struct si_function_info fninfo
;
6476 LLVMValueRef initial
[64], out
[64];
6477 LLVMTypeRef function_type
;
6478 unsigned num_first_params
;
6479 unsigned num_out
, initial_num_out
;
6480 MAYBE_UNUSED
unsigned num_out_sgpr
; /* used in debug checks */
6481 MAYBE_UNUSED
unsigned initial_num_out_sgpr
; /* used in debug checks */
6482 unsigned num_sgprs
, num_vgprs
;
6484 struct lp_build_if_state if_state
;
6486 si_init_function_info(&fninfo
);
6488 for (unsigned i
= 0; i
< num_parts
; ++i
) {
6489 lp_add_function_attr(parts
[i
], -1, LP_FUNC_ATTR_ALWAYSINLINE
);
6490 LLVMSetLinkage(parts
[i
], LLVMPrivateLinkage
);
6493 /* The parameters of the wrapper function correspond to those of the
6494 * first part in terms of SGPRs and VGPRs, but we use the types of the
6495 * main part to get the right types. This is relevant for the
6496 * dereferenceable attribute on descriptor table pointers.
6501 function_type
= LLVMGetElementType(LLVMTypeOf(parts
[0]));
6502 num_first_params
= LLVMCountParamTypes(function_type
);
6504 for (unsigned i
= 0; i
< num_first_params
; ++i
) {
6505 LLVMValueRef param
= LLVMGetParam(parts
[0], i
);
6507 if (ac_is_sgpr_param(param
)) {
6508 assert(num_vgprs
== 0);
6509 num_sgprs
+= ac_get_type_size(LLVMTypeOf(param
)) / 4;
6511 num_vgprs
+= ac_get_type_size(LLVMTypeOf(param
)) / 4;
6516 while (gprs
< num_sgprs
+ num_vgprs
) {
6517 LLVMValueRef param
= LLVMGetParam(parts
[main_part
], fninfo
.num_params
);
6518 LLVMTypeRef type
= LLVMTypeOf(param
);
6519 unsigned size
= ac_get_type_size(type
) / 4;
6521 add_arg(&fninfo
, gprs
< num_sgprs
? ARG_SGPR
: ARG_VGPR
, type
);
6523 assert(ac_is_sgpr_param(param
) == (gprs
< num_sgprs
));
6524 assert(gprs
+ size
<= num_sgprs
+ num_vgprs
&&
6525 (gprs
>= num_sgprs
|| gprs
+ size
<= num_sgprs
));
6530 si_create_function(ctx
, "wrapper", NULL
, 0, &fninfo
,
6531 si_get_max_workgroup_size(ctx
->shader
));
6533 if (is_merged_shader(ctx
->shader
))
6534 ac_init_exec_full_mask(&ctx
->ac
);
6536 /* Record the arguments of the function as if they were an output of
6542 for (unsigned i
= 0; i
< fninfo
.num_params
; ++i
) {
6543 LLVMValueRef param
= LLVMGetParam(ctx
->main_fn
, i
);
6544 LLVMTypeRef param_type
= LLVMTypeOf(param
);
6545 LLVMTypeRef out_type
= i
< fninfo
.num_sgpr_params
? ctx
->i32
: ctx
->f32
;
6546 unsigned size
= ac_get_type_size(param_type
) / 4;
6549 if (param_type
!= out_type
)
6550 param
= LLVMBuildBitCast(builder
, param
, out_type
, "");
6551 out
[num_out
++] = param
;
6553 LLVMTypeRef vector_type
= LLVMVectorType(out_type
, size
);
6555 if (LLVMGetTypeKind(param_type
) == LLVMPointerTypeKind
) {
6556 param
= LLVMBuildPtrToInt(builder
, param
, ctx
->i64
, "");
6557 param_type
= ctx
->i64
;
6560 if (param_type
!= vector_type
)
6561 param
= LLVMBuildBitCast(builder
, param
, vector_type
, "");
6563 for (unsigned j
= 0; j
< size
; ++j
)
6564 out
[num_out
++] = LLVMBuildExtractElement(
6565 builder
, param
, LLVMConstInt(ctx
->i32
, j
, 0), "");
6568 if (i
< fninfo
.num_sgpr_params
)
6569 num_out_sgpr
= num_out
;
6572 memcpy(initial
, out
, sizeof(out
));
6573 initial_num_out
= num_out
;
6574 initial_num_out_sgpr
= num_out_sgpr
;
6576 /* Now chain the parts. */
6577 for (unsigned part
= 0; part
< num_parts
; ++part
) {
6578 LLVMValueRef in
[48];
6580 LLVMTypeRef ret_type
;
6581 unsigned out_idx
= 0;
6582 unsigned num_params
= LLVMCountParams(parts
[part
]);
6584 /* Merged shaders are executed conditionally depending
6585 * on the number of enabled threads passed in the input SGPRs. */
6586 if (is_merged_shader(ctx
->shader
) && part
== 0) {
6587 LLVMValueRef ena
, count
= initial
[3];
6589 count
= LLVMBuildAnd(builder
, count
,
6590 LLVMConstInt(ctx
->i32
, 0x7f, 0), "");
6591 ena
= LLVMBuildICmp(builder
, LLVMIntULT
,
6592 ac_get_thread_id(&ctx
->ac
), count
, "");
6593 lp_build_if(&if_state
, &ctx
->gallivm
, ena
);
6596 /* Derive arguments for the next part from outputs of the
6599 for (unsigned param_idx
= 0; param_idx
< num_params
; ++param_idx
) {
6601 LLVMTypeRef param_type
;
6603 unsigned param_size
;
6604 LLVMValueRef arg
= NULL
;
6606 param
= LLVMGetParam(parts
[part
], param_idx
);
6607 param_type
= LLVMTypeOf(param
);
6608 param_size
= ac_get_type_size(param_type
) / 4;
6609 is_sgpr
= ac_is_sgpr_param(param
);
6612 lp_add_function_attr(parts
[part
], param_idx
+ 1, LP_FUNC_ATTR_INREG
);
6614 assert(out_idx
+ param_size
<= (is_sgpr
? num_out_sgpr
: num_out
));
6615 assert(is_sgpr
|| out_idx
>= num_out_sgpr
);
6617 if (param_size
== 1)
6620 arg
= lp_build_gather_values(&ctx
->gallivm
, &out
[out_idx
], param_size
);
6622 if (LLVMTypeOf(arg
) != param_type
) {
6623 if (LLVMGetTypeKind(param_type
) == LLVMPointerTypeKind
) {
6624 arg
= LLVMBuildBitCast(builder
, arg
, ctx
->i64
, "");
6625 arg
= LLVMBuildIntToPtr(builder
, arg
, param_type
, "");
6627 arg
= LLVMBuildBitCast(builder
, arg
, param_type
, "");
6631 in
[param_idx
] = arg
;
6632 out_idx
+= param_size
;
6635 ret
= LLVMBuildCall(builder
, parts
[part
], in
, num_params
, "");
6637 if (is_merged_shader(ctx
->shader
) &&
6638 part
+ 1 == next_shader_first_part
) {
6639 lp_build_endif(&if_state
);
6641 /* The second half of the merged shader should use
6642 * the inputs from the toplevel (wrapper) function,
6643 * not the return value from the last call.
6645 * That's because the last call was executed condi-
6646 * tionally, so we can't consume it in the main
6649 memcpy(out
, initial
, sizeof(initial
));
6650 num_out
= initial_num_out
;
6651 num_out_sgpr
= initial_num_out_sgpr
;
6655 /* Extract the returned GPRs. */
6656 ret_type
= LLVMTypeOf(ret
);
6660 if (LLVMGetTypeKind(ret_type
) != LLVMVoidTypeKind
) {
6661 assert(LLVMGetTypeKind(ret_type
) == LLVMStructTypeKind
);
6663 unsigned ret_size
= LLVMCountStructElementTypes(ret_type
);
6665 for (unsigned i
= 0; i
< ret_size
; ++i
) {
6667 LLVMBuildExtractValue(builder
, ret
, i
, "");
6669 assert(num_out
< ARRAY_SIZE(out
));
6670 out
[num_out
++] = val
;
6672 if (LLVMTypeOf(val
) == ctx
->i32
) {
6673 assert(num_out_sgpr
+ 1 == num_out
);
6674 num_out_sgpr
= num_out
;
6680 LLVMBuildRetVoid(builder
);
6683 int si_compile_tgsi_shader(struct si_screen
*sscreen
,
6684 LLVMTargetMachineRef tm
,
6685 struct si_shader
*shader
,
6687 struct pipe_debug_callback
*debug
)
6689 struct si_shader_selector
*sel
= shader
->selector
;
6690 struct si_shader_context ctx
;
6693 /* Dump TGSI code before doing TGSI->LLVM conversion in case the
6694 * conversion fails. */
6695 if (si_can_dump_shader(sscreen
, sel
->info
.processor
) &&
6696 !(sscreen
->debug_flags
& DBG(NO_TGSI
))) {
6698 tgsi_dump(sel
->tokens
, 0);
6700 nir_print_shader(sel
->nir
, stderr
);
6701 si_dump_streamout(&sel
->so
);
6704 si_init_shader_ctx(&ctx
, sscreen
, tm
);
6705 si_llvm_context_set_tgsi(&ctx
, shader
);
6706 ctx
.separate_prolog
= !is_monolithic
;
6708 memset(shader
->info
.vs_output_param_offset
, AC_EXP_PARAM_UNDEFINED
,
6709 sizeof(shader
->info
.vs_output_param_offset
));
6711 shader
->info
.uses_instanceid
= sel
->info
.uses_instanceid
;
6713 if (!si_compile_tgsi_main(&ctx
, is_monolithic
)) {
6714 si_llvm_dispose(&ctx
);
6718 if (is_monolithic
&& ctx
.type
== PIPE_SHADER_VERTEX
) {
6719 LLVMValueRef parts
[2];
6720 bool need_prolog
= sel
->vs_needs_prolog
;
6722 parts
[1] = ctx
.main_fn
;
6725 union si_shader_part_key prolog_key
;
6726 si_get_vs_prolog_key(&sel
->info
,
6727 shader
->info
.num_input_sgprs
,
6728 &shader
->key
.part
.vs
.prolog
,
6729 shader
, &prolog_key
);
6730 si_build_vs_prolog_function(&ctx
, &prolog_key
);
6731 parts
[0] = ctx
.main_fn
;
6734 si_build_wrapper_function(&ctx
, parts
+ !need_prolog
,
6735 1 + need_prolog
, need_prolog
, 0);
6736 } else if (is_monolithic
&& ctx
.type
== PIPE_SHADER_TESS_CTRL
) {
6737 if (sscreen
->info
.chip_class
>= GFX9
) {
6738 struct si_shader_selector
*ls
= shader
->key
.part
.tcs
.ls
;
6739 LLVMValueRef parts
[4];
6740 bool vs_needs_prolog
=
6741 si_vs_needs_prolog(ls
, &shader
->key
.part
.tcs
.ls_prolog
);
6744 parts
[2] = ctx
.main_fn
;
6747 union si_shader_part_key tcs_epilog_key
;
6748 memset(&tcs_epilog_key
, 0, sizeof(tcs_epilog_key
));
6749 tcs_epilog_key
.tcs_epilog
.states
= shader
->key
.part
.tcs
.epilog
;
6750 si_build_tcs_epilog_function(&ctx
, &tcs_epilog_key
);
6751 parts
[3] = ctx
.main_fn
;
6754 if (vs_needs_prolog
) {
6755 union si_shader_part_key vs_prolog_key
;
6756 si_get_vs_prolog_key(&ls
->info
,
6757 shader
->info
.num_input_sgprs
,
6758 &shader
->key
.part
.tcs
.ls_prolog
,
6759 shader
, &vs_prolog_key
);
6760 vs_prolog_key
.vs_prolog
.is_monolithic
= true;
6761 si_build_vs_prolog_function(&ctx
, &vs_prolog_key
);
6762 parts
[0] = ctx
.main_fn
;
6765 /* VS as LS main part */
6766 struct si_shader shader_ls
= {};
6767 shader_ls
.selector
= ls
;
6768 shader_ls
.key
.as_ls
= 1;
6769 shader_ls
.key
.mono
= shader
->key
.mono
;
6770 shader_ls
.key
.opt
= shader
->key
.opt
;
6771 si_llvm_context_set_tgsi(&ctx
, &shader_ls
);
6773 if (!si_compile_tgsi_main(&ctx
, true)) {
6774 si_llvm_dispose(&ctx
);
6777 shader
->info
.uses_instanceid
|= ls
->info
.uses_instanceid
;
6778 parts
[1] = ctx
.main_fn
;
6780 /* Reset the shader context. */
6781 ctx
.shader
= shader
;
6782 ctx
.type
= PIPE_SHADER_TESS_CTRL
;
6784 si_build_wrapper_function(&ctx
,
6785 parts
+ !vs_needs_prolog
,
6786 4 - !vs_needs_prolog
, 0,
6787 vs_needs_prolog
? 2 : 1);
6789 LLVMValueRef parts
[2];
6790 union si_shader_part_key epilog_key
;
6792 parts
[0] = ctx
.main_fn
;
6794 memset(&epilog_key
, 0, sizeof(epilog_key
));
6795 epilog_key
.tcs_epilog
.states
= shader
->key
.part
.tcs
.epilog
;
6796 si_build_tcs_epilog_function(&ctx
, &epilog_key
);
6797 parts
[1] = ctx
.main_fn
;
6799 si_build_wrapper_function(&ctx
, parts
, 2, 0, 0);
6801 } else if (is_monolithic
&& ctx
.type
== PIPE_SHADER_GEOMETRY
) {
6802 if (ctx
.screen
->info
.chip_class
>= GFX9
) {
6803 struct si_shader_selector
*es
= shader
->key
.part
.gs
.es
;
6804 LLVMValueRef es_prolog
= NULL
;
6805 LLVMValueRef es_main
= NULL
;
6806 LLVMValueRef gs_prolog
= NULL
;
6807 LLVMValueRef gs_main
= ctx
.main_fn
;
6810 union si_shader_part_key gs_prolog_key
;
6811 memset(&gs_prolog_key
, 0, sizeof(gs_prolog_key
));
6812 gs_prolog_key
.gs_prolog
.states
= shader
->key
.part
.gs
.prolog
;
6813 gs_prolog_key
.gs_prolog
.is_monolithic
= true;
6814 si_build_gs_prolog_function(&ctx
, &gs_prolog_key
);
6815 gs_prolog
= ctx
.main_fn
;
6818 if (es
->vs_needs_prolog
) {
6819 union si_shader_part_key vs_prolog_key
;
6820 si_get_vs_prolog_key(&es
->info
,
6821 shader
->info
.num_input_sgprs
,
6822 &shader
->key
.part
.gs
.vs_prolog
,
6823 shader
, &vs_prolog_key
);
6824 vs_prolog_key
.vs_prolog
.is_monolithic
= true;
6825 si_build_vs_prolog_function(&ctx
, &vs_prolog_key
);
6826 es_prolog
= ctx
.main_fn
;
6830 struct si_shader shader_es
= {};
6831 shader_es
.selector
= es
;
6832 shader_es
.key
.as_es
= 1;
6833 shader_es
.key
.mono
= shader
->key
.mono
;
6834 shader_es
.key
.opt
= shader
->key
.opt
;
6835 si_llvm_context_set_tgsi(&ctx
, &shader_es
);
6837 if (!si_compile_tgsi_main(&ctx
, true)) {
6838 si_llvm_dispose(&ctx
);
6841 shader
->info
.uses_instanceid
|= es
->info
.uses_instanceid
;
6842 es_main
= ctx
.main_fn
;
6844 /* Reset the shader context. */
6845 ctx
.shader
= shader
;
6846 ctx
.type
= PIPE_SHADER_GEOMETRY
;
6848 /* Prepare the array of shader parts. */
6849 LLVMValueRef parts
[4];
6850 unsigned num_parts
= 0, main_part
, next_first_part
;
6853 parts
[num_parts
++] = es_prolog
;
6855 parts
[main_part
= num_parts
++] = es_main
;
6856 parts
[next_first_part
= num_parts
++] = gs_prolog
;
6857 parts
[num_parts
++] = gs_main
;
6859 si_build_wrapper_function(&ctx
, parts
, num_parts
,
6860 main_part
, next_first_part
);
6862 LLVMValueRef parts
[2];
6863 union si_shader_part_key prolog_key
;
6865 parts
[1] = ctx
.main_fn
;
6867 memset(&prolog_key
, 0, sizeof(prolog_key
));
6868 prolog_key
.gs_prolog
.states
= shader
->key
.part
.gs
.prolog
;
6869 si_build_gs_prolog_function(&ctx
, &prolog_key
);
6870 parts
[0] = ctx
.main_fn
;
6872 si_build_wrapper_function(&ctx
, parts
, 2, 1, 0);
6874 } else if (is_monolithic
&& ctx
.type
== PIPE_SHADER_FRAGMENT
) {
6875 LLVMValueRef parts
[3];
6876 union si_shader_part_key prolog_key
;
6877 union si_shader_part_key epilog_key
;
6880 si_get_ps_prolog_key(shader
, &prolog_key
, false);
6881 need_prolog
= si_need_ps_prolog(&prolog_key
);
6883 parts
[need_prolog
? 1 : 0] = ctx
.main_fn
;
6886 si_build_ps_prolog_function(&ctx
, &prolog_key
);
6887 parts
[0] = ctx
.main_fn
;
6890 si_get_ps_epilog_key(shader
, &epilog_key
);
6891 si_build_ps_epilog_function(&ctx
, &epilog_key
);
6892 parts
[need_prolog
? 2 : 1] = ctx
.main_fn
;
6894 si_build_wrapper_function(&ctx
, parts
, need_prolog
? 3 : 2,
6895 need_prolog
? 1 : 0, 0);
6898 si_llvm_optimize_module(&ctx
);
6900 /* Post-optimization transformations and analysis. */
6901 si_optimize_vs_outputs(&ctx
);
6903 if ((debug
&& debug
->debug_message
) ||
6904 si_can_dump_shader(sscreen
, ctx
.type
))
6905 si_count_scratch_private_memory(&ctx
);
6907 /* Compile to bytecode. */
6908 r
= si_compile_llvm(sscreen
, &shader
->binary
, &shader
->config
, tm
,
6909 ctx
.gallivm
.module
, debug
, ctx
.type
, "TGSI shader");
6910 si_llvm_dispose(&ctx
);
6912 fprintf(stderr
, "LLVM failed to compile shader\n");
6916 /* Validate SGPR and VGPR usage for compute to detect compiler bugs.
6917 * LLVM 3.9svn has this bug.
6919 if (sel
->type
== PIPE_SHADER_COMPUTE
) {
6920 unsigned wave_size
= 64;
6921 unsigned max_vgprs
= 256;
6922 unsigned max_sgprs
= sscreen
->info
.chip_class
>= VI
? 800 : 512;
6923 unsigned max_sgprs_per_wave
= 128;
6924 unsigned max_block_threads
= si_get_max_workgroup_size(shader
);
6925 unsigned min_waves_per_cu
= DIV_ROUND_UP(max_block_threads
, wave_size
);
6926 unsigned min_waves_per_simd
= DIV_ROUND_UP(min_waves_per_cu
, 4);
6928 max_vgprs
= max_vgprs
/ min_waves_per_simd
;
6929 max_sgprs
= MIN2(max_sgprs
/ min_waves_per_simd
, max_sgprs_per_wave
);
6931 if (shader
->config
.num_sgprs
> max_sgprs
||
6932 shader
->config
.num_vgprs
> max_vgprs
) {
6933 fprintf(stderr
, "LLVM failed to compile a shader correctly: "
6934 "SGPR:VGPR usage is %u:%u, but the hw limit is %u:%u\n",
6935 shader
->config
.num_sgprs
, shader
->config
.num_vgprs
,
6936 max_sgprs
, max_vgprs
);
6938 /* Just terminate the process, because dependent
6939 * shaders can hang due to bad input data, but use
6940 * the env var to allow shader-db to work.
6942 if (!debug_get_bool_option("SI_PASS_BAD_SHADERS", false))
6947 /* Add the scratch offset to input SGPRs. */
6948 if (shader
->config
.scratch_bytes_per_wave
&& !is_merged_shader(shader
))
6949 shader
->info
.num_input_sgprs
+= 1; /* scratch byte offset */
6951 /* Calculate the number of fragment input VGPRs. */
6952 if (ctx
.type
== PIPE_SHADER_FRAGMENT
) {
6953 shader
->info
.num_input_vgprs
= 0;
6954 shader
->info
.face_vgpr_index
= -1;
6955 shader
->info
.ancillary_vgpr_index
= -1;
6957 if (G_0286CC_PERSP_SAMPLE_ENA(shader
->config
.spi_ps_input_addr
))
6958 shader
->info
.num_input_vgprs
+= 2;
6959 if (G_0286CC_PERSP_CENTER_ENA(shader
->config
.spi_ps_input_addr
))
6960 shader
->info
.num_input_vgprs
+= 2;
6961 if (G_0286CC_PERSP_CENTROID_ENA(shader
->config
.spi_ps_input_addr
))
6962 shader
->info
.num_input_vgprs
+= 2;
6963 if (G_0286CC_PERSP_PULL_MODEL_ENA(shader
->config
.spi_ps_input_addr
))
6964 shader
->info
.num_input_vgprs
+= 3;
6965 if (G_0286CC_LINEAR_SAMPLE_ENA(shader
->config
.spi_ps_input_addr
))
6966 shader
->info
.num_input_vgprs
+= 2;
6967 if (G_0286CC_LINEAR_CENTER_ENA(shader
->config
.spi_ps_input_addr
))
6968 shader
->info
.num_input_vgprs
+= 2;
6969 if (G_0286CC_LINEAR_CENTROID_ENA(shader
->config
.spi_ps_input_addr
))
6970 shader
->info
.num_input_vgprs
+= 2;
6971 if (G_0286CC_LINE_STIPPLE_TEX_ENA(shader
->config
.spi_ps_input_addr
))
6972 shader
->info
.num_input_vgprs
+= 1;
6973 if (G_0286CC_POS_X_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
6974 shader
->info
.num_input_vgprs
+= 1;
6975 if (G_0286CC_POS_Y_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
6976 shader
->info
.num_input_vgprs
+= 1;
6977 if (G_0286CC_POS_Z_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
6978 shader
->info
.num_input_vgprs
+= 1;
6979 if (G_0286CC_POS_W_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
6980 shader
->info
.num_input_vgprs
+= 1;
6981 if (G_0286CC_FRONT_FACE_ENA(shader
->config
.spi_ps_input_addr
)) {
6982 shader
->info
.face_vgpr_index
= shader
->info
.num_input_vgprs
;
6983 shader
->info
.num_input_vgprs
+= 1;
6985 if (G_0286CC_ANCILLARY_ENA(shader
->config
.spi_ps_input_addr
)) {
6986 shader
->info
.ancillary_vgpr_index
= shader
->info
.num_input_vgprs
;
6987 shader
->info
.num_input_vgprs
+= 1;
6989 if (G_0286CC_SAMPLE_COVERAGE_ENA(shader
->config
.spi_ps_input_addr
))
6990 shader
->info
.num_input_vgprs
+= 1;
6991 if (G_0286CC_POS_FIXED_PT_ENA(shader
->config
.spi_ps_input_addr
))
6992 shader
->info
.num_input_vgprs
+= 1;
6995 si_calculate_max_simd_waves(shader
);
6996 si_shader_dump_stats_for_shader_db(shader
, debug
);
7001 * Create, compile and return a shader part (prolog or epilog).
7003 * \param sscreen screen
7004 * \param list list of shader parts of the same category
7005 * \param type shader type
7006 * \param key shader part key
7007 * \param prolog whether the part being requested is a prolog
7008 * \param tm LLVM target machine
7009 * \param debug debug callback
7010 * \param build the callback responsible for building the main function
7011 * \return non-NULL on success
7013 static struct si_shader_part
*
7014 si_get_shader_part(struct si_screen
*sscreen
,
7015 struct si_shader_part
**list
,
7016 enum pipe_shader_type type
,
7018 union si_shader_part_key
*key
,
7019 LLVMTargetMachineRef tm
,
7020 struct pipe_debug_callback
*debug
,
7021 void (*build
)(struct si_shader_context
*,
7022 union si_shader_part_key
*),
7025 struct si_shader_part
*result
;
7027 mtx_lock(&sscreen
->shader_parts_mutex
);
7029 /* Find existing. */
7030 for (result
= *list
; result
; result
= result
->next
) {
7031 if (memcmp(&result
->key
, key
, sizeof(*key
)) == 0) {
7032 mtx_unlock(&sscreen
->shader_parts_mutex
);
7037 /* Compile a new one. */
7038 result
= CALLOC_STRUCT(si_shader_part
);
7041 struct si_shader shader
= {};
7042 struct si_shader_context ctx
;
7044 si_init_shader_ctx(&ctx
, sscreen
, tm
);
7045 ctx
.shader
= &shader
;
7049 case PIPE_SHADER_VERTEX
:
7050 shader
.key
.as_ls
= key
->vs_prolog
.as_ls
;
7051 shader
.key
.as_es
= key
->vs_prolog
.as_es
;
7053 case PIPE_SHADER_TESS_CTRL
:
7055 shader
.key
.part
.tcs
.epilog
= key
->tcs_epilog
.states
;
7057 case PIPE_SHADER_GEOMETRY
:
7060 case PIPE_SHADER_FRAGMENT
:
7062 shader
.key
.part
.ps
.prolog
= key
->ps_prolog
.states
;
7064 shader
.key
.part
.ps
.epilog
= key
->ps_epilog
.states
;
7067 unreachable("bad shader part");
7073 si_llvm_optimize_module(&ctx
);
7075 if (si_compile_llvm(sscreen
, &result
->binary
, &result
->config
, tm
,
7076 ctx
.ac
.module
, debug
, ctx
.type
, name
)) {
7082 result
->next
= *list
;
7086 si_llvm_dispose(&ctx
);
7087 mtx_unlock(&sscreen
->shader_parts_mutex
);
7091 static LLVMValueRef
si_prolog_get_rw_buffers(struct si_shader_context
*ctx
)
7093 LLVMValueRef ptr
[2], list
;
7094 bool is_merged_shader
=
7095 ctx
->screen
->info
.chip_class
>= GFX9
&&
7096 (ctx
->type
== PIPE_SHADER_TESS_CTRL
||
7097 ctx
->type
== PIPE_SHADER_GEOMETRY
||
7098 ctx
->shader
->key
.as_ls
|| ctx
->shader
->key
.as_es
);
7100 /* Get the pointer to rw buffers. */
7101 ptr
[0] = LLVMGetParam(ctx
->main_fn
, (is_merged_shader
? 8 : 0) + SI_SGPR_RW_BUFFERS
);
7102 ptr
[1] = LLVMGetParam(ctx
->main_fn
, (is_merged_shader
? 8 : 0) + SI_SGPR_RW_BUFFERS_HI
);
7103 list
= lp_build_gather_values(&ctx
->gallivm
, ptr
, 2);
7104 list
= LLVMBuildBitCast(ctx
->ac
.builder
, list
, ctx
->i64
, "");
7105 list
= LLVMBuildIntToPtr(ctx
->ac
.builder
, list
,
7106 ac_array_in_const_addr_space(ctx
->v4i32
), "");
7111 * Build the vertex shader prolog function.
7113 * The inputs are the same as VS (a lot of SGPRs and 4 VGPR system values).
7114 * All inputs are returned unmodified. The vertex load indices are
7115 * stored after them, which will be used by the API VS for fetching inputs.
7117 * For example, the expected outputs for instance_divisors[] = {0, 1, 2} are:
7122 * (VertexID + BaseVertex),
7123 * (InstanceID + StartInstance),
7124 * (InstanceID / 2 + StartInstance)
7126 static void si_build_vs_prolog_function(struct si_shader_context
*ctx
,
7127 union si_shader_part_key
*key
)
7129 struct si_function_info fninfo
;
7130 LLVMTypeRef
*returns
;
7131 LLVMValueRef ret
, func
;
7133 unsigned first_vs_vgpr
= key
->vs_prolog
.num_merged_next_stage_vgprs
;
7134 unsigned num_input_vgprs
= key
->vs_prolog
.num_merged_next_stage_vgprs
+ 4;
7135 LLVMValueRef input_vgprs
[9];
7136 unsigned num_all_input_regs
= key
->vs_prolog
.num_input_sgprs
+
7138 unsigned user_sgpr_base
= key
->vs_prolog
.num_merged_next_stage_vgprs
? 8 : 0;
7140 si_init_function_info(&fninfo
);
7142 /* 4 preloaded VGPRs + vertex load indices as prolog outputs */
7143 returns
= alloca((num_all_input_regs
+ key
->vs_prolog
.last_input
+ 1) *
7144 sizeof(LLVMTypeRef
));
7147 /* Declare input and output SGPRs. */
7148 for (i
= 0; i
< key
->vs_prolog
.num_input_sgprs
; i
++) {
7149 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7150 returns
[num_returns
++] = ctx
->i32
;
7153 /* Preloaded VGPRs (outputs must be floats) */
7154 for (i
= 0; i
< num_input_vgprs
; i
++) {
7155 add_arg_assign(&fninfo
, ARG_VGPR
, ctx
->i32
, &input_vgprs
[i
]);
7156 returns
[num_returns
++] = ctx
->f32
;
7159 /* Vertex load indices. */
7160 for (i
= 0; i
<= key
->vs_prolog
.last_input
; i
++)
7161 returns
[num_returns
++] = ctx
->f32
;
7163 /* Create the function. */
7164 si_create_function(ctx
, "vs_prolog", returns
, num_returns
, &fninfo
, 0);
7165 func
= ctx
->main_fn
;
7167 if (key
->vs_prolog
.num_merged_next_stage_vgprs
) {
7168 if (!key
->vs_prolog
.is_monolithic
)
7169 si_init_exec_from_input(ctx
, 3, 0);
7171 if (key
->vs_prolog
.as_ls
&&
7172 ctx
->screen
->has_ls_vgpr_init_bug
) {
7173 /* If there are no HS threads, SPI loads the LS VGPRs
7174 * starting at VGPR 0. Shift them back to where they
7177 LLVMValueRef has_hs_threads
=
7178 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntNE
,
7179 unpack_param(ctx
, 3, 8, 8),
7182 for (i
= 4; i
> 0; --i
) {
7183 input_vgprs
[i
+ 1] =
7184 LLVMBuildSelect(ctx
->ac
.builder
, has_hs_threads
,
7186 input_vgprs
[i
- 1], "");
7191 ctx
->abi
.vertex_id
= input_vgprs
[first_vs_vgpr
];
7192 ctx
->abi
.instance_id
= input_vgprs
[first_vs_vgpr
+ (key
->vs_prolog
.as_ls
? 2 : 1)];
7194 /* Copy inputs to outputs. This should be no-op, as the registers match,
7195 * but it will prevent the compiler from overwriting them unintentionally.
7197 ret
= ctx
->return_value
;
7198 for (i
= 0; i
< key
->vs_prolog
.num_input_sgprs
; i
++) {
7199 LLVMValueRef p
= LLVMGetParam(func
, i
);
7200 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, p
, i
, "");
7202 for (i
= 0; i
< num_input_vgprs
; i
++) {
7203 LLVMValueRef p
= input_vgprs
[i
];
7204 p
= ac_to_float(&ctx
->ac
, p
);
7205 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, p
,
7206 key
->vs_prolog
.num_input_sgprs
+ i
, "");
7209 /* Compute vertex load indices from instance divisors. */
7210 LLVMValueRef instance_divisor_constbuf
= NULL
;
7212 if (key
->vs_prolog
.states
.instance_divisor_is_fetched
) {
7213 LLVMValueRef list
= si_prolog_get_rw_buffers(ctx
);
7214 LLVMValueRef buf_index
=
7215 LLVMConstInt(ctx
->i32
, SI_VS_CONST_INSTANCE_DIVISORS
, 0);
7216 instance_divisor_constbuf
=
7217 ac_build_load_to_sgpr(&ctx
->ac
, list
, buf_index
);
7220 for (i
= 0; i
<= key
->vs_prolog
.last_input
; i
++) {
7221 bool divisor_is_one
=
7222 key
->vs_prolog
.states
.instance_divisor_is_one
& (1u << i
);
7223 bool divisor_is_fetched
=
7224 key
->vs_prolog
.states
.instance_divisor_is_fetched
& (1u << i
);
7227 if (divisor_is_one
|| divisor_is_fetched
) {
7228 LLVMValueRef divisor
= ctx
->i32_1
;
7230 if (divisor_is_fetched
) {
7231 divisor
= buffer_load_const(ctx
, instance_divisor_constbuf
,
7232 LLVMConstInt(ctx
->i32
, i
* 4, 0));
7233 divisor
= ac_to_integer(&ctx
->ac
, divisor
);
7236 /* InstanceID / Divisor + StartInstance */
7237 index
= get_instance_index_for_fetch(ctx
,
7239 SI_SGPR_START_INSTANCE
,
7242 /* VertexID + BaseVertex */
7243 index
= LLVMBuildAdd(ctx
->ac
.builder
,
7245 LLVMGetParam(func
, user_sgpr_base
+
7246 SI_SGPR_BASE_VERTEX
), "");
7249 index
= ac_to_float(&ctx
->ac
, index
);
7250 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, index
,
7251 fninfo
.num_params
+ i
, "");
7254 si_llvm_build_ret(ctx
, ret
);
7257 static bool si_get_vs_prolog(struct si_screen
*sscreen
,
7258 LLVMTargetMachineRef tm
,
7259 struct si_shader
*shader
,
7260 struct pipe_debug_callback
*debug
,
7261 struct si_shader
*main_part
,
7262 const struct si_vs_prolog_bits
*key
)
7264 struct si_shader_selector
*vs
= main_part
->selector
;
7266 if (!si_vs_needs_prolog(vs
, key
))
7269 /* Get the prolog. */
7270 union si_shader_part_key prolog_key
;
7271 si_get_vs_prolog_key(&vs
->info
, main_part
->info
.num_input_sgprs
,
7272 key
, shader
, &prolog_key
);
7275 si_get_shader_part(sscreen
, &sscreen
->vs_prologs
,
7276 PIPE_SHADER_VERTEX
, true, &prolog_key
, tm
,
7277 debug
, si_build_vs_prolog_function
,
7278 "Vertex Shader Prolog");
7279 return shader
->prolog
!= NULL
;
7283 * Select and compile (or reuse) vertex shader parts (prolog & epilog).
7285 static bool si_shader_select_vs_parts(struct si_screen
*sscreen
,
7286 LLVMTargetMachineRef tm
,
7287 struct si_shader
*shader
,
7288 struct pipe_debug_callback
*debug
)
7290 return si_get_vs_prolog(sscreen
, tm
, shader
, debug
, shader
,
7291 &shader
->key
.part
.vs
.prolog
);
7295 * Compile the TCS epilog function. This writes tesselation factors to memory
7296 * based on the output primitive type of the tesselator (determined by TES).
7298 static void si_build_tcs_epilog_function(struct si_shader_context
*ctx
,
7299 union si_shader_part_key
*key
)
7301 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
7302 struct si_function_info fninfo
;
7305 si_init_function_info(&fninfo
);
7307 if (ctx
->screen
->info
.chip_class
>= GFX9
) {
7308 add_arg(&fninfo
, ARG_SGPR
, ctx
->i64
);
7309 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7310 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
); /* wave info */
7311 ctx
->param_tcs_factor_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7312 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7313 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7314 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7315 add_arg(&fninfo
, ARG_SGPR
, ctx
->i64
);
7316 add_arg(&fninfo
, ARG_SGPR
, ctx
->i64
);
7317 add_arg(&fninfo
, ARG_SGPR
, ctx
->i64
);
7318 add_arg(&fninfo
, ARG_SGPR
, ctx
->i64
);
7319 add_arg(&fninfo
, ARG_SGPR
, ctx
->i64
);
7320 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7321 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7322 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7323 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7324 ctx
->param_tcs_offchip_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7325 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7326 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7327 ctx
->param_tcs_offchip_addr_base64k
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7328 ctx
->param_tcs_factor_addr_base64k
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7330 add_arg(&fninfo
, ARG_SGPR
, ctx
->i64
);
7331 add_arg(&fninfo
, ARG_SGPR
, ctx
->i64
);
7332 add_arg(&fninfo
, ARG_SGPR
, ctx
->i64
);
7333 add_arg(&fninfo
, ARG_SGPR
, ctx
->i64
);
7334 ctx
->param_tcs_offchip_layout
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7335 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7336 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7337 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7338 ctx
->param_tcs_offchip_addr_base64k
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7339 ctx
->param_tcs_factor_addr_base64k
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7340 ctx
->param_tcs_offchip_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7341 ctx
->param_tcs_factor_offset
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7344 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
); /* VGPR gap */
7345 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
); /* VGPR gap */
7346 unsigned tess_factors_idx
=
7347 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
); /* patch index within the wave (REL_PATCH_ID) */
7348 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
); /* invocation ID within the patch */
7349 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
); /* LDS offset where tess factors should be loaded from */
7351 for (unsigned i
= 0; i
< 6; i
++)
7352 add_arg(&fninfo
, ARG_VGPR
, ctx
->i32
); /* tess factors */
7354 /* Create the function. */
7355 si_create_function(ctx
, "tcs_epilog", NULL
, 0, &fninfo
,
7356 ctx
->screen
->info
.chip_class
>= CIK
? 128 : 64);
7357 ac_declare_lds_as_pointer(&ctx
->ac
);
7358 func
= ctx
->main_fn
;
7360 LLVMValueRef invoc0_tess_factors
[6];
7361 for (unsigned i
= 0; i
< 6; i
++)
7362 invoc0_tess_factors
[i
] = LLVMGetParam(func
, tess_factors_idx
+ 3 + i
);
7364 si_write_tess_factors(bld_base
,
7365 LLVMGetParam(func
, tess_factors_idx
),
7366 LLVMGetParam(func
, tess_factors_idx
+ 1),
7367 LLVMGetParam(func
, tess_factors_idx
+ 2),
7368 invoc0_tess_factors
, invoc0_tess_factors
+ 4);
7370 LLVMBuildRetVoid(ctx
->ac
.builder
);
7374 * Select and compile (or reuse) TCS parts (epilog).
7376 static bool si_shader_select_tcs_parts(struct si_screen
*sscreen
,
7377 LLVMTargetMachineRef tm
,
7378 struct si_shader
*shader
,
7379 struct pipe_debug_callback
*debug
)
7381 if (sscreen
->info
.chip_class
>= GFX9
) {
7382 struct si_shader
*ls_main_part
=
7383 shader
->key
.part
.tcs
.ls
->main_shader_part_ls
;
7385 if (!si_get_vs_prolog(sscreen
, tm
, shader
, debug
, ls_main_part
,
7386 &shader
->key
.part
.tcs
.ls_prolog
))
7389 shader
->previous_stage
= ls_main_part
;
7392 /* Get the epilog. */
7393 union si_shader_part_key epilog_key
;
7394 memset(&epilog_key
, 0, sizeof(epilog_key
));
7395 epilog_key
.tcs_epilog
.states
= shader
->key
.part
.tcs
.epilog
;
7397 shader
->epilog
= si_get_shader_part(sscreen
, &sscreen
->tcs_epilogs
,
7398 PIPE_SHADER_TESS_CTRL
, false,
7399 &epilog_key
, tm
, debug
,
7400 si_build_tcs_epilog_function
,
7401 "Tessellation Control Shader Epilog");
7402 return shader
->epilog
!= NULL
;
7406 * Select and compile (or reuse) GS parts (prolog).
7408 static bool si_shader_select_gs_parts(struct si_screen
*sscreen
,
7409 LLVMTargetMachineRef tm
,
7410 struct si_shader
*shader
,
7411 struct pipe_debug_callback
*debug
)
7413 if (sscreen
->info
.chip_class
>= GFX9
) {
7414 struct si_shader
*es_main_part
=
7415 shader
->key
.part
.gs
.es
->main_shader_part_es
;
7417 if (shader
->key
.part
.gs
.es
->type
== PIPE_SHADER_VERTEX
&&
7418 !si_get_vs_prolog(sscreen
, tm
, shader
, debug
, es_main_part
,
7419 &shader
->key
.part
.gs
.vs_prolog
))
7422 shader
->previous_stage
= es_main_part
;
7425 if (!shader
->key
.part
.gs
.prolog
.tri_strip_adj_fix
)
7428 union si_shader_part_key prolog_key
;
7429 memset(&prolog_key
, 0, sizeof(prolog_key
));
7430 prolog_key
.gs_prolog
.states
= shader
->key
.part
.gs
.prolog
;
7432 shader
->prolog2
= si_get_shader_part(sscreen
, &sscreen
->gs_prologs
,
7433 PIPE_SHADER_GEOMETRY
, true,
7434 &prolog_key
, tm
, debug
,
7435 si_build_gs_prolog_function
,
7436 "Geometry Shader Prolog");
7437 return shader
->prolog2
!= NULL
;
7441 * Build the pixel shader prolog function. This handles:
7442 * - two-side color selection and interpolation
7443 * - overriding interpolation parameters for the API PS
7444 * - polygon stippling
7446 * All preloaded SGPRs and VGPRs are passed through unmodified unless they are
7447 * overriden by other states. (e.g. per-sample interpolation)
7448 * Interpolated colors are stored after the preloaded VGPRs.
7450 static void si_build_ps_prolog_function(struct si_shader_context
*ctx
,
7451 union si_shader_part_key
*key
)
7453 struct si_function_info fninfo
;
7454 LLVMValueRef ret
, func
;
7455 int num_returns
, i
, num_color_channels
;
7457 assert(si_need_ps_prolog(key
));
7459 si_init_function_info(&fninfo
);
7461 /* Declare inputs. */
7462 for (i
= 0; i
< key
->ps_prolog
.num_input_sgprs
; i
++)
7463 add_arg(&fninfo
, ARG_SGPR
, ctx
->i32
);
7465 for (i
= 0; i
< key
->ps_prolog
.num_input_vgprs
; i
++)
7466 add_arg(&fninfo
, ARG_VGPR
, ctx
->f32
);
7468 /* Declare outputs (same as inputs + add colors if needed) */
7469 num_returns
= fninfo
.num_params
;
7470 num_color_channels
= util_bitcount(key
->ps_prolog
.colors_read
);
7471 for (i
= 0; i
< num_color_channels
; i
++)
7472 fninfo
.types
[num_returns
++] = ctx
->f32
;
7474 /* Create the function. */
7475 si_create_function(ctx
, "ps_prolog", fninfo
.types
, num_returns
,
7477 func
= ctx
->main_fn
;
7479 /* Copy inputs to outputs. This should be no-op, as the registers match,
7480 * but it will prevent the compiler from overwriting them unintentionally.
7482 ret
= ctx
->return_value
;
7483 for (i
= 0; i
< fninfo
.num_params
; i
++) {
7484 LLVMValueRef p
= LLVMGetParam(func
, i
);
7485 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, p
, i
, "");
7488 /* Polygon stippling. */
7489 if (key
->ps_prolog
.states
.poly_stipple
) {
7490 /* POS_FIXED_PT is always last. */
7491 unsigned pos
= key
->ps_prolog
.num_input_sgprs
+
7492 key
->ps_prolog
.num_input_vgprs
- 1;
7493 LLVMValueRef list
= si_prolog_get_rw_buffers(ctx
);
7495 si_llvm_emit_polygon_stipple(ctx
, list
, pos
);
7498 if (key
->ps_prolog
.states
.bc_optimize_for_persp
||
7499 key
->ps_prolog
.states
.bc_optimize_for_linear
) {
7500 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7501 LLVMValueRef center
[2], centroid
[2], tmp
, bc_optimize
;
7503 /* The shader should do: if (PRIM_MASK[31]) CENTROID = CENTER;
7504 * The hw doesn't compute CENTROID if the whole wave only
7505 * contains fully-covered quads.
7507 * PRIM_MASK is after user SGPRs.
7509 bc_optimize
= LLVMGetParam(func
, SI_PS_NUM_USER_SGPR
);
7510 bc_optimize
= LLVMBuildLShr(ctx
->ac
.builder
, bc_optimize
,
7511 LLVMConstInt(ctx
->i32
, 31, 0), "");
7512 bc_optimize
= LLVMBuildTrunc(ctx
->ac
.builder
, bc_optimize
,
7515 if (key
->ps_prolog
.states
.bc_optimize_for_persp
) {
7516 /* Read PERSP_CENTER. */
7517 for (i
= 0; i
< 2; i
++)
7518 center
[i
] = LLVMGetParam(func
, base
+ 2 + i
);
7519 /* Read PERSP_CENTROID. */
7520 for (i
= 0; i
< 2; i
++)
7521 centroid
[i
] = LLVMGetParam(func
, base
+ 4 + i
);
7522 /* Select PERSP_CENTROID. */
7523 for (i
= 0; i
< 2; i
++) {
7524 tmp
= LLVMBuildSelect(ctx
->ac
.builder
, bc_optimize
,
7525 center
[i
], centroid
[i
], "");
7526 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7527 tmp
, base
+ 4 + i
, "");
7530 if (key
->ps_prolog
.states
.bc_optimize_for_linear
) {
7531 /* Read LINEAR_CENTER. */
7532 for (i
= 0; i
< 2; i
++)
7533 center
[i
] = LLVMGetParam(func
, base
+ 8 + i
);
7534 /* Read LINEAR_CENTROID. */
7535 for (i
= 0; i
< 2; i
++)
7536 centroid
[i
] = LLVMGetParam(func
, base
+ 10 + i
);
7537 /* Select LINEAR_CENTROID. */
7538 for (i
= 0; i
< 2; i
++) {
7539 tmp
= LLVMBuildSelect(ctx
->ac
.builder
, bc_optimize
,
7540 center
[i
], centroid
[i
], "");
7541 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7542 tmp
, base
+ 10 + i
, "");
7547 /* Force per-sample interpolation. */
7548 if (key
->ps_prolog
.states
.force_persp_sample_interp
) {
7549 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7550 LLVMValueRef persp_sample
[2];
7552 /* Read PERSP_SAMPLE. */
7553 for (i
= 0; i
< 2; i
++)
7554 persp_sample
[i
] = LLVMGetParam(func
, base
+ i
);
7555 /* Overwrite PERSP_CENTER. */
7556 for (i
= 0; i
< 2; i
++)
7557 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7558 persp_sample
[i
], base
+ 2 + i
, "");
7559 /* Overwrite PERSP_CENTROID. */
7560 for (i
= 0; i
< 2; i
++)
7561 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7562 persp_sample
[i
], base
+ 4 + i
, "");
7564 if (key
->ps_prolog
.states
.force_linear_sample_interp
) {
7565 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7566 LLVMValueRef linear_sample
[2];
7568 /* Read LINEAR_SAMPLE. */
7569 for (i
= 0; i
< 2; i
++)
7570 linear_sample
[i
] = LLVMGetParam(func
, base
+ 6 + i
);
7571 /* Overwrite LINEAR_CENTER. */
7572 for (i
= 0; i
< 2; i
++)
7573 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7574 linear_sample
[i
], base
+ 8 + i
, "");
7575 /* Overwrite LINEAR_CENTROID. */
7576 for (i
= 0; i
< 2; i
++)
7577 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7578 linear_sample
[i
], base
+ 10 + i
, "");
7581 /* Force center interpolation. */
7582 if (key
->ps_prolog
.states
.force_persp_center_interp
) {
7583 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7584 LLVMValueRef persp_center
[2];
7586 /* Read PERSP_CENTER. */
7587 for (i
= 0; i
< 2; i
++)
7588 persp_center
[i
] = LLVMGetParam(func
, base
+ 2 + i
);
7589 /* Overwrite PERSP_SAMPLE. */
7590 for (i
= 0; i
< 2; i
++)
7591 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7592 persp_center
[i
], base
+ i
, "");
7593 /* Overwrite PERSP_CENTROID. */
7594 for (i
= 0; i
< 2; i
++)
7595 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7596 persp_center
[i
], base
+ 4 + i
, "");
7598 if (key
->ps_prolog
.states
.force_linear_center_interp
) {
7599 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7600 LLVMValueRef linear_center
[2];
7602 /* Read LINEAR_CENTER. */
7603 for (i
= 0; i
< 2; i
++)
7604 linear_center
[i
] = LLVMGetParam(func
, base
+ 8 + i
);
7605 /* Overwrite LINEAR_SAMPLE. */
7606 for (i
= 0; i
< 2; i
++)
7607 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7608 linear_center
[i
], base
+ 6 + i
, "");
7609 /* Overwrite LINEAR_CENTROID. */
7610 for (i
= 0; i
< 2; i
++)
7611 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
7612 linear_center
[i
], base
+ 10 + i
, "");
7615 /* Interpolate colors. */
7616 unsigned color_out_idx
= 0;
7617 for (i
= 0; i
< 2; i
++) {
7618 unsigned writemask
= (key
->ps_prolog
.colors_read
>> (i
* 4)) & 0xf;
7619 unsigned face_vgpr
= key
->ps_prolog
.num_input_sgprs
+
7620 key
->ps_prolog
.face_vgpr_index
;
7621 LLVMValueRef interp
[2], color
[4];
7622 LLVMValueRef interp_ij
= NULL
, prim_mask
= NULL
, face
= NULL
;
7627 /* If the interpolation qualifier is not CONSTANT (-1). */
7628 if (key
->ps_prolog
.color_interp_vgpr_index
[i
] != -1) {
7629 unsigned interp_vgpr
= key
->ps_prolog
.num_input_sgprs
+
7630 key
->ps_prolog
.color_interp_vgpr_index
[i
];
7632 /* Get the (i,j) updated by bc_optimize handling. */
7633 interp
[0] = LLVMBuildExtractValue(ctx
->ac
.builder
, ret
,
7635 interp
[1] = LLVMBuildExtractValue(ctx
->ac
.builder
, ret
,
7636 interp_vgpr
+ 1, "");
7637 interp_ij
= lp_build_gather_values(&ctx
->gallivm
, interp
, 2);
7640 /* Use the absolute location of the input. */
7641 prim_mask
= LLVMGetParam(func
, SI_PS_NUM_USER_SGPR
);
7643 if (key
->ps_prolog
.states
.color_two_side
) {
7644 face
= LLVMGetParam(func
, face_vgpr
);
7645 face
= ac_to_integer(&ctx
->ac
, face
);
7648 interp_fs_input(ctx
,
7649 key
->ps_prolog
.color_attr_index
[i
],
7650 TGSI_SEMANTIC_COLOR
, i
,
7651 key
->ps_prolog
.num_interp_inputs
,
7652 key
->ps_prolog
.colors_read
, interp_ij
,
7653 prim_mask
, face
, color
);
7656 unsigned chan
= u_bit_scan(&writemask
);
7657 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, color
[chan
],
7658 fninfo
.num_params
+ color_out_idx
++, "");
7662 /* Section 15.2.2 (Shader Inputs) of the OpenGL 4.5 (Core Profile) spec
7665 * "When per-sample shading is active due to the use of a fragment
7666 * input qualified by sample or due to the use of the gl_SampleID
7667 * or gl_SamplePosition variables, only the bit for the current
7668 * sample is set in gl_SampleMaskIn. When state specifies multiple
7669 * fragment shader invocations for a given fragment, the sample
7670 * mask for any single fragment shader invocation may specify a
7671 * subset of the covered samples for the fragment. In this case,
7672 * the bit corresponding to each covered sample will be set in
7673 * exactly one fragment shader invocation."
7675 * The samplemask loaded by hardware is always the coverage of the
7676 * entire pixel/fragment, so mask bits out based on the sample ID.
7678 if (key
->ps_prolog
.states
.samplemask_log_ps_iter
) {
7679 /* The bit pattern matches that used by fixed function fragment
7681 static const uint16_t ps_iter_masks
[] = {
7682 0xffff, /* not used */
7688 assert(key
->ps_prolog
.states
.samplemask_log_ps_iter
< ARRAY_SIZE(ps_iter_masks
));
7690 uint32_t ps_iter_mask
= ps_iter_masks
[key
->ps_prolog
.states
.samplemask_log_ps_iter
];
7691 unsigned ancillary_vgpr
= key
->ps_prolog
.num_input_sgprs
+
7692 key
->ps_prolog
.ancillary_vgpr_index
;
7693 LLVMValueRef sampleid
= unpack_param(ctx
, ancillary_vgpr
, 8, 4);
7694 LLVMValueRef samplemask
= LLVMGetParam(func
, ancillary_vgpr
+ 1);
7696 samplemask
= ac_to_integer(&ctx
->ac
, samplemask
);
7697 samplemask
= LLVMBuildAnd(
7700 LLVMBuildShl(ctx
->ac
.builder
,
7701 LLVMConstInt(ctx
->i32
, ps_iter_mask
, false),
7704 samplemask
= ac_to_float(&ctx
->ac
, samplemask
);
7706 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, samplemask
,
7707 ancillary_vgpr
+ 1, "");
7710 /* Tell LLVM to insert WQM instruction sequence when needed. */
7711 if (key
->ps_prolog
.wqm
) {
7712 LLVMAddTargetDependentFunctionAttr(func
,
7713 "amdgpu-ps-wqm-outputs", "");
7716 si_llvm_build_ret(ctx
, ret
);
7720 * Build the pixel shader epilog function. This handles everything that must be
7721 * emulated for pixel shader exports. (alpha-test, format conversions, etc)
7723 static void si_build_ps_epilog_function(struct si_shader_context
*ctx
,
7724 union si_shader_part_key
*key
)
7726 struct lp_build_tgsi_context
*bld_base
= &ctx
->bld_base
;
7727 struct si_function_info fninfo
;
7728 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
7730 struct si_ps_exports exp
= {};
7732 si_init_function_info(&fninfo
);
7734 /* Declare input SGPRs. */
7735 ctx
->param_rw_buffers
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i64
);
7736 ctx
->param_bindless_samplers_and_images
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i64
);
7737 ctx
->param_const_and_shader_buffers
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i64
);
7738 ctx
->param_samplers_and_images
= add_arg(&fninfo
, ARG_SGPR
, ctx
->i64
);
7739 add_arg_checked(&fninfo
, ARG_SGPR
, ctx
->f32
, SI_PARAM_ALPHA_REF
);
7741 /* Declare input VGPRs. */
7742 unsigned required_num_params
=
7743 fninfo
.num_sgpr_params
+
7744 util_bitcount(key
->ps_epilog
.colors_written
) * 4 +
7745 key
->ps_epilog
.writes_z
+
7746 key
->ps_epilog
.writes_stencil
+
7747 key
->ps_epilog
.writes_samplemask
;
7749 required_num_params
= MAX2(required_num_params
,
7750 fninfo
.num_sgpr_params
+ PS_EPILOG_SAMPLEMASK_MIN_LOC
+ 1);
7752 while (fninfo
.num_params
< required_num_params
)
7753 add_arg(&fninfo
, ARG_VGPR
, ctx
->f32
);
7755 /* Create the function. */
7756 si_create_function(ctx
, "ps_epilog", NULL
, 0, &fninfo
, 0);
7757 /* Disable elimination of unused inputs. */
7758 si_llvm_add_attribute(ctx
->main_fn
,
7759 "InitialPSInputAddr", 0xffffff);
7761 /* Process colors. */
7762 unsigned vgpr
= fninfo
.num_sgpr_params
;
7763 unsigned colors_written
= key
->ps_epilog
.colors_written
;
7764 int last_color_export
= -1;
7766 /* Find the last color export. */
7767 if (!key
->ps_epilog
.writes_z
&&
7768 !key
->ps_epilog
.writes_stencil
&&
7769 !key
->ps_epilog
.writes_samplemask
) {
7770 unsigned spi_format
= key
->ps_epilog
.states
.spi_shader_col_format
;
7772 /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
7773 if (colors_written
== 0x1 && key
->ps_epilog
.states
.last_cbuf
> 0) {
7774 /* Just set this if any of the colorbuffers are enabled. */
7776 ((1ull << (4 * (key
->ps_epilog
.states
.last_cbuf
+ 1))) - 1))
7777 last_color_export
= 0;
7779 for (i
= 0; i
< 8; i
++)
7780 if (colors_written
& (1 << i
) &&
7781 (spi_format
>> (i
* 4)) & 0xf)
7782 last_color_export
= i
;
7786 while (colors_written
) {
7787 LLVMValueRef color
[4];
7788 int mrt
= u_bit_scan(&colors_written
);
7790 for (i
= 0; i
< 4; i
++)
7791 color
[i
] = LLVMGetParam(ctx
->main_fn
, vgpr
++);
7793 si_export_mrt_color(bld_base
, color
, mrt
,
7794 fninfo
.num_params
- 1,
7795 mrt
== last_color_export
, &exp
);
7798 /* Process depth, stencil, samplemask. */
7799 if (key
->ps_epilog
.writes_z
)
7800 depth
= LLVMGetParam(ctx
->main_fn
, vgpr
++);
7801 if (key
->ps_epilog
.writes_stencil
)
7802 stencil
= LLVMGetParam(ctx
->main_fn
, vgpr
++);
7803 if (key
->ps_epilog
.writes_samplemask
)
7804 samplemask
= LLVMGetParam(ctx
->main_fn
, vgpr
++);
7806 if (depth
|| stencil
|| samplemask
)
7807 si_export_mrt_z(bld_base
, depth
, stencil
, samplemask
, &exp
);
7808 else if (last_color_export
== -1)
7809 si_export_null(bld_base
);
7812 si_emit_ps_exports(ctx
, &exp
);
7815 LLVMBuildRetVoid(ctx
->ac
.builder
);
7819 * Select and compile (or reuse) pixel shader parts (prolog & epilog).
7821 static bool si_shader_select_ps_parts(struct si_screen
*sscreen
,
7822 LLVMTargetMachineRef tm
,
7823 struct si_shader
*shader
,
7824 struct pipe_debug_callback
*debug
)
7826 union si_shader_part_key prolog_key
;
7827 union si_shader_part_key epilog_key
;
7829 /* Get the prolog. */
7830 si_get_ps_prolog_key(shader
, &prolog_key
, true);
7832 /* The prolog is a no-op if these aren't set. */
7833 if (si_need_ps_prolog(&prolog_key
)) {
7835 si_get_shader_part(sscreen
, &sscreen
->ps_prologs
,
7836 PIPE_SHADER_FRAGMENT
, true,
7837 &prolog_key
, tm
, debug
,
7838 si_build_ps_prolog_function
,
7839 "Fragment Shader Prolog");
7840 if (!shader
->prolog
)
7844 /* Get the epilog. */
7845 si_get_ps_epilog_key(shader
, &epilog_key
);
7848 si_get_shader_part(sscreen
, &sscreen
->ps_epilogs
,
7849 PIPE_SHADER_FRAGMENT
, false,
7850 &epilog_key
, tm
, debug
,
7851 si_build_ps_epilog_function
,
7852 "Fragment Shader Epilog");
7853 if (!shader
->epilog
)
7856 /* Enable POS_FIXED_PT if polygon stippling is enabled. */
7857 if (shader
->key
.part
.ps
.prolog
.poly_stipple
) {
7858 shader
->config
.spi_ps_input_ena
|= S_0286CC_POS_FIXED_PT_ENA(1);
7859 assert(G_0286CC_POS_FIXED_PT_ENA(shader
->config
.spi_ps_input_addr
));
7862 /* Set up the enable bits for per-sample shading if needed. */
7863 if (shader
->key
.part
.ps
.prolog
.force_persp_sample_interp
&&
7864 (G_0286CC_PERSP_CENTER_ENA(shader
->config
.spi_ps_input_ena
) ||
7865 G_0286CC_PERSP_CENTROID_ENA(shader
->config
.spi_ps_input_ena
))) {
7866 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_CENTER_ENA
;
7867 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_CENTROID_ENA
;
7868 shader
->config
.spi_ps_input_ena
|= S_0286CC_PERSP_SAMPLE_ENA(1);
7870 if (shader
->key
.part
.ps
.prolog
.force_linear_sample_interp
&&
7871 (G_0286CC_LINEAR_CENTER_ENA(shader
->config
.spi_ps_input_ena
) ||
7872 G_0286CC_LINEAR_CENTROID_ENA(shader
->config
.spi_ps_input_ena
))) {
7873 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_CENTER_ENA
;
7874 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_CENTROID_ENA
;
7875 shader
->config
.spi_ps_input_ena
|= S_0286CC_LINEAR_SAMPLE_ENA(1);
7877 if (shader
->key
.part
.ps
.prolog
.force_persp_center_interp
&&
7878 (G_0286CC_PERSP_SAMPLE_ENA(shader
->config
.spi_ps_input_ena
) ||
7879 G_0286CC_PERSP_CENTROID_ENA(shader
->config
.spi_ps_input_ena
))) {
7880 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_SAMPLE_ENA
;
7881 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_CENTROID_ENA
;
7882 shader
->config
.spi_ps_input_ena
|= S_0286CC_PERSP_CENTER_ENA(1);
7884 if (shader
->key
.part
.ps
.prolog
.force_linear_center_interp
&&
7885 (G_0286CC_LINEAR_SAMPLE_ENA(shader
->config
.spi_ps_input_ena
) ||
7886 G_0286CC_LINEAR_CENTROID_ENA(shader
->config
.spi_ps_input_ena
))) {
7887 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_SAMPLE_ENA
;
7888 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_CENTROID_ENA
;
7889 shader
->config
.spi_ps_input_ena
|= S_0286CC_LINEAR_CENTER_ENA(1);
7892 /* POW_W_FLOAT requires that one of the perspective weights is enabled. */
7893 if (G_0286CC_POS_W_FLOAT_ENA(shader
->config
.spi_ps_input_ena
) &&
7894 !(shader
->config
.spi_ps_input_ena
& 0xf)) {
7895 shader
->config
.spi_ps_input_ena
|= S_0286CC_PERSP_CENTER_ENA(1);
7896 assert(G_0286CC_PERSP_CENTER_ENA(shader
->config
.spi_ps_input_addr
));
7899 /* At least one pair of interpolation weights must be enabled. */
7900 if (!(shader
->config
.spi_ps_input_ena
& 0x7f)) {
7901 shader
->config
.spi_ps_input_ena
|= S_0286CC_LINEAR_CENTER_ENA(1);
7902 assert(G_0286CC_LINEAR_CENTER_ENA(shader
->config
.spi_ps_input_addr
));
7905 /* Samplemask fixup requires the sample ID. */
7906 if (shader
->key
.part
.ps
.prolog
.samplemask_log_ps_iter
) {
7907 shader
->config
.spi_ps_input_ena
|= S_0286CC_ANCILLARY_ENA(1);
7908 assert(G_0286CC_ANCILLARY_ENA(shader
->config
.spi_ps_input_addr
));
7911 /* The sample mask input is always enabled, because the API shader always
7912 * passes it through to the epilog. Disable it here if it's unused.
7914 if (!shader
->key
.part
.ps
.epilog
.poly_line_smoothing
&&
7915 !shader
->selector
->info
.reads_samplemask
)
7916 shader
->config
.spi_ps_input_ena
&= C_0286CC_SAMPLE_COVERAGE_ENA
;
7921 void si_multiwave_lds_size_workaround(struct si_screen
*sscreen
,
7924 /* SPI barrier management bug:
7925 * Make sure we have at least 4k of LDS in use to avoid the bug.
7926 * It applies to workgroup sizes of more than one wavefront.
7928 if (sscreen
->info
.family
== CHIP_BONAIRE
||
7929 sscreen
->info
.family
== CHIP_KABINI
||
7930 sscreen
->info
.family
== CHIP_MULLINS
)
7931 *lds_size
= MAX2(*lds_size
, 8);
7934 static void si_fix_resource_usage(struct si_screen
*sscreen
,
7935 struct si_shader
*shader
)
7937 unsigned min_sgprs
= shader
->info
.num_input_sgprs
+ 2; /* VCC */
7939 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
, min_sgprs
);
7941 if (shader
->selector
->type
== PIPE_SHADER_COMPUTE
&&
7942 si_get_max_workgroup_size(shader
) > 64) {
7943 si_multiwave_lds_size_workaround(sscreen
,
7944 &shader
->config
.lds_size
);
7948 int si_shader_create(struct si_screen
*sscreen
, LLVMTargetMachineRef tm
,
7949 struct si_shader
*shader
,
7950 struct pipe_debug_callback
*debug
)
7952 struct si_shader_selector
*sel
= shader
->selector
;
7953 struct si_shader
*mainp
= *si_get_main_shader_part(sel
, &shader
->key
);
7956 /* LS, ES, VS are compiled on demand if the main part hasn't been
7957 * compiled for that stage.
7959 * Vertex shaders are compiled on demand when a vertex fetch
7960 * workaround must be applied.
7962 if (shader
->is_monolithic
) {
7963 /* Monolithic shader (compiled as a whole, has many variants,
7964 * may take a long time to compile).
7966 r
= si_compile_tgsi_shader(sscreen
, tm
, shader
, true, debug
);
7970 /* The shader consists of several parts:
7972 * - the middle part is the user shader, it has 1 variant only
7973 * and it was compiled during the creation of the shader
7975 * - the prolog part is inserted at the beginning
7976 * - the epilog part is inserted at the end
7978 * The prolog and epilog have many (but simple) variants.
7980 * Starting with gfx9, geometry and tessellation control
7981 * shaders also contain the prolog and user shader parts of
7982 * the previous shader stage.
7988 /* Copy the compiled TGSI shader data over. */
7989 shader
->is_binary_shared
= true;
7990 shader
->binary
= mainp
->binary
;
7991 shader
->config
= mainp
->config
;
7992 shader
->info
.num_input_sgprs
= mainp
->info
.num_input_sgprs
;
7993 shader
->info
.num_input_vgprs
= mainp
->info
.num_input_vgprs
;
7994 shader
->info
.face_vgpr_index
= mainp
->info
.face_vgpr_index
;
7995 shader
->info
.ancillary_vgpr_index
= mainp
->info
.ancillary_vgpr_index
;
7996 memcpy(shader
->info
.vs_output_param_offset
,
7997 mainp
->info
.vs_output_param_offset
,
7998 sizeof(mainp
->info
.vs_output_param_offset
));
7999 shader
->info
.uses_instanceid
= mainp
->info
.uses_instanceid
;
8000 shader
->info
.nr_pos_exports
= mainp
->info
.nr_pos_exports
;
8001 shader
->info
.nr_param_exports
= mainp
->info
.nr_param_exports
;
8003 /* Select prologs and/or epilogs. */
8004 switch (sel
->type
) {
8005 case PIPE_SHADER_VERTEX
:
8006 if (!si_shader_select_vs_parts(sscreen
, tm
, shader
, debug
))
8009 case PIPE_SHADER_TESS_CTRL
:
8010 if (!si_shader_select_tcs_parts(sscreen
, tm
, shader
, debug
))
8013 case PIPE_SHADER_TESS_EVAL
:
8015 case PIPE_SHADER_GEOMETRY
:
8016 if (!si_shader_select_gs_parts(sscreen
, tm
, shader
, debug
))
8019 case PIPE_SHADER_FRAGMENT
:
8020 if (!si_shader_select_ps_parts(sscreen
, tm
, shader
, debug
))
8023 /* Make sure we have at least as many VGPRs as there
8024 * are allocated inputs.
8026 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
8027 shader
->info
.num_input_vgprs
);
8031 /* Update SGPR and VGPR counts. */
8032 if (shader
->prolog
) {
8033 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
,
8034 shader
->prolog
->config
.num_sgprs
);
8035 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
8036 shader
->prolog
->config
.num_vgprs
);
8038 if (shader
->previous_stage
) {
8039 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
,
8040 shader
->previous_stage
->config
.num_sgprs
);
8041 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
8042 shader
->previous_stage
->config
.num_vgprs
);
8043 shader
->config
.spilled_sgprs
=
8044 MAX2(shader
->config
.spilled_sgprs
,
8045 shader
->previous_stage
->config
.spilled_sgprs
);
8046 shader
->config
.spilled_vgprs
=
8047 MAX2(shader
->config
.spilled_vgprs
,
8048 shader
->previous_stage
->config
.spilled_vgprs
);
8049 shader
->config
.private_mem_vgprs
=
8050 MAX2(shader
->config
.private_mem_vgprs
,
8051 shader
->previous_stage
->config
.private_mem_vgprs
);
8052 shader
->config
.scratch_bytes_per_wave
=
8053 MAX2(shader
->config
.scratch_bytes_per_wave
,
8054 shader
->previous_stage
->config
.scratch_bytes_per_wave
);
8055 shader
->info
.uses_instanceid
|=
8056 shader
->previous_stage
->info
.uses_instanceid
;
8058 if (shader
->prolog2
) {
8059 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
,
8060 shader
->prolog2
->config
.num_sgprs
);
8061 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
8062 shader
->prolog2
->config
.num_vgprs
);
8064 if (shader
->epilog
) {
8065 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
,
8066 shader
->epilog
->config
.num_sgprs
);
8067 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
8068 shader
->epilog
->config
.num_vgprs
);
8070 si_calculate_max_simd_waves(shader
);
8073 si_fix_resource_usage(sscreen
, shader
);
8074 si_shader_dump(sscreen
, shader
, debug
, sel
->info
.processor
,
8078 r
= si_shader_binary_upload(sscreen
, shader
);
8080 fprintf(stderr
, "LLVM failed to upload shader\n");
8087 void si_shader_destroy(struct si_shader
*shader
)
8089 if (shader
->scratch_bo
)
8090 r600_resource_reference(&shader
->scratch_bo
, NULL
);
8092 r600_resource_reference(&shader
->bo
, NULL
);
8094 if (!shader
->is_binary_shared
)
8095 ac_shader_binary_clean(&shader
->binary
);
8097 free(shader
->shader_log
);