2 * Copyright 2012 Advanced Micro Devices, Inc.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
24 * Tom Stellard <thomas.stellard@amd.com>
25 * Michel Dänzer <michel.daenzer@amd.com>
26 * Christian König <christian.koenig@amd.com>
29 #include "gallivm/lp_bld_const.h"
30 #include "gallivm/lp_bld_gather.h"
31 #include "gallivm/lp_bld_intr.h"
32 #include "gallivm/lp_bld_logic.h"
33 #include "gallivm/lp_bld_arit.h"
34 #include "gallivm/lp_bld_flow.h"
35 #include "gallivm/lp_bld_misc.h"
36 #include "radeon/radeon_llvm.h"
37 #include "radeon/radeon_elf_util.h"
38 #include "radeon/radeon_llvm_emit.h"
39 #include "util/u_memory.h"
40 #include "util/u_string.h"
41 #include "tgsi/tgsi_build.h"
42 #include "tgsi/tgsi_util.h"
43 #include "tgsi/tgsi_dump.h"
49 static const char *scratch_rsrc_dword0_symbol
=
50 "SCRATCH_RSRC_DWORD0";
52 static const char *scratch_rsrc_dword1_symbol
=
53 "SCRATCH_RSRC_DWORD1";
55 struct si_shader_output_values
57 LLVMValueRef values
[4];
62 struct si_shader_context
64 struct radeon_llvm_context radeon_bld
;
65 struct si_shader
*shader
;
66 struct si_screen
*screen
;
68 unsigned type
; /* PIPE_SHADER_* specifies the type of shader. */
69 bool is_gs_copy_shader
;
71 /* Whether to generate the optimized shader variant compiled as a whole
72 * (without a prolog and epilog)
76 int param_streamout_config
;
77 int param_streamout_write_index
;
78 int param_streamout_offset
[4];
80 int param_rel_auto_id
;
82 int param_instance_id
;
83 int param_vertex_index0
;
86 int param_tes_rel_patch_id
;
87 int param_tes_patch_id
;
88 int param_es2gs_offset
;
91 /* Sets a bit if the dynamic HS control word was 0x80000000. The bit is
92 * 0x800000 for VS, 0x1 for ES.
94 int param_tess_offchip
;
96 LLVMTargetMachineRef tm
;
98 unsigned invariant_load_md_kind
;
99 unsigned range_md_kind
;
100 unsigned uniform_md_kind
;
101 LLVMValueRef empty_md
;
103 /* Preloaded descriptors. */
104 LLVMValueRef esgs_ring
;
105 LLVMValueRef gsvs_ring
[4];
108 LLVMValueRef gs_next_vertex
[4];
109 LLVMValueRef return_value
;
124 LLVMValueRef shared_memory
;
127 static struct si_shader_context
*si_shader_context(
128 struct lp_build_tgsi_context
*bld_base
)
130 return (struct si_shader_context
*)bld_base
;
133 static void si_init_shader_ctx(struct si_shader_context
*ctx
,
134 struct si_screen
*sscreen
,
135 struct si_shader
*shader
,
136 LLVMTargetMachineRef tm
);
138 static void si_llvm_emit_barrier(const struct lp_build_tgsi_action
*action
,
139 struct lp_build_tgsi_context
*bld_base
,
140 struct lp_build_emit_data
*emit_data
);
142 static void si_dump_shader_key(unsigned shader
, union si_shader_key
*key
,
145 /* Ideally pass the sample mask input to the PS epilog as v13, which
146 * is its usual location, so that the shader doesn't have to add v_mov.
148 #define PS_EPILOG_SAMPLEMASK_MIN_LOC 13
150 /* The VS location of the PrimitiveID input is the same in the epilog,
151 * so that the main shader part doesn't have to move it.
153 #define VS_EPILOG_PRIMID_LOC 2
156 CONST_ADDR_SPACE
= 2,
157 LOCAL_ADDR_SPACE
= 3,
161 #define SENDMSG_GS_DONE 3
163 #define SENDMSG_GS_OP_NOP (0 << 4)
164 #define SENDMSG_GS_OP_CUT (1 << 4)
165 #define SENDMSG_GS_OP_EMIT (2 << 4)
166 #define SENDMSG_GS_OP_EMIT_CUT (3 << 4)
169 * Returns a unique index for a semantic name and index. The index must be
170 * less than 64, so that a 64-bit bitmask of used inputs or outputs can be
173 unsigned si_shader_io_get_unique_index(unsigned semantic_name
, unsigned index
)
175 switch (semantic_name
) {
176 case TGSI_SEMANTIC_POSITION
:
178 case TGSI_SEMANTIC_PSIZE
:
180 case TGSI_SEMANTIC_CLIPDIST
:
183 case TGSI_SEMANTIC_GENERIC
:
187 /* same explanation as in the default statement,
188 * the only user hitting this is st/nine.
192 /* patch indices are completely separate and thus start from 0 */
193 case TGSI_SEMANTIC_TESSOUTER
:
195 case TGSI_SEMANTIC_TESSINNER
:
197 case TGSI_SEMANTIC_PATCH
:
201 /* Don't fail here. The result of this function is only used
202 * for LS, TCS, TES, and GS, where legacy GL semantics can't
203 * occur, but this function is called for all vertex shaders
204 * before it's known whether LS will be compiled or not.
211 * Get the value of a shader input parameter and extract a bitfield.
213 static LLVMValueRef
unpack_param(struct si_shader_context
*ctx
,
214 unsigned param
, unsigned rshift
,
217 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
218 LLVMValueRef value
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
221 if (LLVMGetTypeKind(LLVMTypeOf(value
)) == LLVMFloatTypeKind
)
222 value
= bitcast(&ctx
->radeon_bld
.soa
.bld_base
,
223 TGSI_TYPE_UNSIGNED
, value
);
226 value
= LLVMBuildLShr(gallivm
->builder
, value
,
227 lp_build_const_int32(gallivm
, rshift
), "");
229 if (rshift
+ bitwidth
< 32) {
230 unsigned mask
= (1 << bitwidth
) - 1;
231 value
= LLVMBuildAnd(gallivm
->builder
, value
,
232 lp_build_const_int32(gallivm
, mask
), "");
238 static LLVMValueRef
get_rel_patch_id(struct si_shader_context
*ctx
)
241 case PIPE_SHADER_TESS_CTRL
:
242 return unpack_param(ctx
, SI_PARAM_REL_IDS
, 0, 8);
244 case PIPE_SHADER_TESS_EVAL
:
245 return LLVMGetParam(ctx
->radeon_bld
.main_fn
,
246 ctx
->param_tes_rel_patch_id
);
254 /* Tessellation shaders pass outputs to the next shader using LDS.
256 * LS outputs = TCS inputs
257 * TCS outputs = TES inputs
260 * - TCS inputs for patch 0
261 * - TCS inputs for patch 1
262 * - TCS inputs for patch 2 = get_tcs_in_current_patch_offset (if RelPatchID==2)
264 * - TCS outputs for patch 0 = get_tcs_out_patch0_offset
265 * - Per-patch TCS outputs for patch 0 = get_tcs_out_patch0_patch_data_offset
266 * - TCS outputs for patch 1
267 * - Per-patch TCS outputs for patch 1
268 * - TCS outputs for patch 2 = get_tcs_out_current_patch_offset (if RelPatchID==2)
269 * - Per-patch TCS outputs for patch 2 = get_tcs_out_current_patch_data_offset (if RelPatchID==2)
272 * All three shaders VS(LS), TCS, TES share the same LDS space.
276 get_tcs_in_patch_stride(struct si_shader_context
*ctx
)
278 if (ctx
->type
== PIPE_SHADER_VERTEX
)
279 return unpack_param(ctx
, SI_PARAM_LS_OUT_LAYOUT
, 0, 13);
280 else if (ctx
->type
== PIPE_SHADER_TESS_CTRL
)
281 return unpack_param(ctx
, SI_PARAM_TCS_IN_LAYOUT
, 0, 13);
289 get_tcs_out_patch_stride(struct si_shader_context
*ctx
)
291 return unpack_param(ctx
, SI_PARAM_TCS_OUT_LAYOUT
, 0, 13);
295 get_tcs_out_patch0_offset(struct si_shader_context
*ctx
)
297 return lp_build_mul_imm(&ctx
->radeon_bld
.soa
.bld_base
.uint_bld
,
299 SI_PARAM_TCS_OUT_OFFSETS
,
305 get_tcs_out_patch0_patch_data_offset(struct si_shader_context
*ctx
)
307 return lp_build_mul_imm(&ctx
->radeon_bld
.soa
.bld_base
.uint_bld
,
309 SI_PARAM_TCS_OUT_OFFSETS
,
315 get_tcs_in_current_patch_offset(struct si_shader_context
*ctx
)
317 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
318 LLVMValueRef patch_stride
= get_tcs_in_patch_stride(ctx
);
319 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
321 return LLVMBuildMul(gallivm
->builder
, patch_stride
, rel_patch_id
, "");
325 get_tcs_out_current_patch_offset(struct si_shader_context
*ctx
)
327 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
328 LLVMValueRef patch0_offset
= get_tcs_out_patch0_offset(ctx
);
329 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(ctx
);
330 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
332 return LLVMBuildAdd(gallivm
->builder
, patch0_offset
,
333 LLVMBuildMul(gallivm
->builder
, patch_stride
,
339 get_tcs_out_current_patch_data_offset(struct si_shader_context
*ctx
)
341 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
342 LLVMValueRef patch0_patch_data_offset
=
343 get_tcs_out_patch0_patch_data_offset(ctx
);
344 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(ctx
);
345 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
347 return LLVMBuildAdd(gallivm
->builder
, patch0_patch_data_offset
,
348 LLVMBuildMul(gallivm
->builder
, patch_stride
,
353 static LLVMValueRef
build_gep0(struct si_shader_context
*ctx
,
354 LLVMValueRef base_ptr
, LLVMValueRef index
)
356 LLVMValueRef indices
[2] = {
357 LLVMConstInt(ctx
->i32
, 0, 0),
360 return LLVMBuildGEP(ctx
->radeon_bld
.gallivm
.builder
, base_ptr
,
364 static void build_indexed_store(struct si_shader_context
*ctx
,
365 LLVMValueRef base_ptr
, LLVMValueRef index
,
368 struct lp_build_tgsi_context
*bld_base
= &ctx
->radeon_bld
.soa
.bld_base
;
369 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
371 LLVMBuildStore(gallivm
->builder
, value
,
372 build_gep0(ctx
, base_ptr
, index
));
376 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
377 * It's equivalent to doing a load from &base_ptr[index].
379 * \param base_ptr Where the array starts.
380 * \param index The element index into the array.
381 * \param uniform Whether the base_ptr and index can be assumed to be
382 * dynamically uniform
384 static LLVMValueRef
build_indexed_load(struct si_shader_context
*ctx
,
385 LLVMValueRef base_ptr
, LLVMValueRef index
,
388 struct lp_build_tgsi_context
*bld_base
= &ctx
->radeon_bld
.soa
.bld_base
;
389 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
390 LLVMValueRef pointer
;
392 pointer
= build_gep0(ctx
, base_ptr
, index
);
394 LLVMSetMetadata(pointer
, ctx
->uniform_md_kind
, ctx
->empty_md
);
395 return LLVMBuildLoad(gallivm
->builder
, pointer
, "");
399 * Do a load from &base_ptr[index], but also add a flag that it's loading
400 * a constant from a dynamically uniform index.
402 static LLVMValueRef
build_indexed_load_const(
403 struct si_shader_context
*ctx
,
404 LLVMValueRef base_ptr
, LLVMValueRef index
)
406 LLVMValueRef result
= build_indexed_load(ctx
, base_ptr
, index
, true);
407 LLVMSetMetadata(result
, ctx
->invariant_load_md_kind
, ctx
->empty_md
);
411 static LLVMValueRef
get_instance_index_for_fetch(
412 struct radeon_llvm_context
*radeon_bld
,
413 unsigned param_start_instance
, unsigned divisor
)
415 struct si_shader_context
*ctx
=
416 si_shader_context(&radeon_bld
->soa
.bld_base
);
417 struct gallivm_state
*gallivm
= radeon_bld
->soa
.bld_base
.base
.gallivm
;
419 LLVMValueRef result
= LLVMGetParam(radeon_bld
->main_fn
,
420 ctx
->param_instance_id
);
422 /* The division must be done before START_INSTANCE is added. */
424 result
= LLVMBuildUDiv(gallivm
->builder
, result
,
425 lp_build_const_int32(gallivm
, divisor
), "");
427 return LLVMBuildAdd(gallivm
->builder
, result
,
428 LLVMGetParam(radeon_bld
->main_fn
, param_start_instance
), "");
431 static void declare_input_vs(
432 struct radeon_llvm_context
*radeon_bld
,
433 unsigned input_index
,
434 const struct tgsi_full_declaration
*decl
,
437 struct lp_build_context
*base
= &radeon_bld
->soa
.bld_base
.base
;
438 struct gallivm_state
*gallivm
= base
->gallivm
;
439 struct si_shader_context
*ctx
=
440 si_shader_context(&radeon_bld
->soa
.bld_base
);
442 ctx
->shader
->key
.vs
.prolog
.instance_divisors
[input_index
];
446 LLVMValueRef t_list_ptr
;
447 LLVMValueRef t_offset
;
449 LLVMValueRef attribute_offset
;
450 LLVMValueRef buffer_index
;
451 LLVMValueRef args
[3];
454 /* Load the T list */
455 t_list_ptr
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_VERTEX_BUFFERS
);
457 t_offset
= lp_build_const_int32(gallivm
, input_index
);
459 t_list
= build_indexed_load_const(ctx
, t_list_ptr
, t_offset
);
461 /* Build the attribute offset */
462 attribute_offset
= lp_build_const_int32(gallivm
, 0);
464 if (!ctx
->is_monolithic
) {
465 buffer_index
= LLVMGetParam(radeon_bld
->main_fn
,
466 ctx
->param_vertex_index0
+
468 } else if (divisor
) {
469 /* Build index from instance ID, start instance and divisor */
470 ctx
->shader
->info
.uses_instanceid
= true;
471 buffer_index
= get_instance_index_for_fetch(&ctx
->radeon_bld
,
472 SI_PARAM_START_INSTANCE
,
475 /* Load the buffer index for vertices. */
476 LLVMValueRef vertex_id
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
477 ctx
->param_vertex_id
);
478 LLVMValueRef base_vertex
= LLVMGetParam(radeon_bld
->main_fn
,
479 SI_PARAM_BASE_VERTEX
);
480 buffer_index
= LLVMBuildAdd(gallivm
->builder
, base_vertex
, vertex_id
, "");
484 args
[1] = attribute_offset
;
485 args
[2] = buffer_index
;
486 input
= lp_build_intrinsic(gallivm
->builder
,
487 "llvm.SI.vs.load.input", ctx
->v4f32
, args
, 3,
488 LLVMReadNoneAttribute
);
490 /* Break up the vec4 into individual components */
491 for (chan
= 0; chan
< 4; chan
++) {
492 LLVMValueRef llvm_chan
= lp_build_const_int32(gallivm
, chan
);
493 out
[chan
] = LLVMBuildExtractElement(gallivm
->builder
,
494 input
, llvm_chan
, "");
498 static LLVMValueRef
get_primitive_id(struct lp_build_tgsi_context
*bld_base
,
501 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
504 return bld_base
->uint_bld
.zero
;
507 case PIPE_SHADER_VERTEX
:
508 return LLVMGetParam(ctx
->radeon_bld
.main_fn
,
509 ctx
->param_vs_prim_id
);
510 case PIPE_SHADER_TESS_CTRL
:
511 return LLVMGetParam(ctx
->radeon_bld
.main_fn
,
513 case PIPE_SHADER_TESS_EVAL
:
514 return LLVMGetParam(ctx
->radeon_bld
.main_fn
,
515 ctx
->param_tes_patch_id
);
516 case PIPE_SHADER_GEOMETRY
:
517 return LLVMGetParam(ctx
->radeon_bld
.main_fn
,
518 SI_PARAM_PRIMITIVE_ID
);
521 return bld_base
->uint_bld
.zero
;
526 * Return the value of tgsi_ind_register for indexing.
527 * This is the indirect index with the constant offset added to it.
529 static LLVMValueRef
get_indirect_index(struct si_shader_context
*ctx
,
530 const struct tgsi_ind_register
*ind
,
533 struct gallivm_state
*gallivm
= ctx
->radeon_bld
.soa
.bld_base
.base
.gallivm
;
536 result
= ctx
->radeon_bld
.soa
.addr
[ind
->Index
][ind
->Swizzle
];
537 result
= LLVMBuildLoad(gallivm
->builder
, result
, "");
538 result
= LLVMBuildAdd(gallivm
->builder
, result
,
539 lp_build_const_int32(gallivm
, rel_index
), "");
544 * Like get_indirect_index, but restricts the return value to a (possibly
545 * undefined) value inside [0..num).
547 static LLVMValueRef
get_bounded_indirect_index(struct si_shader_context
*ctx
,
548 const struct tgsi_ind_register
*ind
,
549 int rel_index
, unsigned num
)
551 LLVMValueRef result
= get_indirect_index(ctx
, ind
, rel_index
);
553 /* LLVM 3.8: If indirect resource indexing is used:
557 if (HAVE_LLVM
<= 0x0308)
558 return LLVMGetUndef(ctx
->i32
);
560 return radeon_llvm_bound_index(&ctx
->radeon_bld
, result
, num
);
565 * Calculate a dword address given an input or output register and a stride.
567 static LLVMValueRef
get_dw_address(struct si_shader_context
*ctx
,
568 const struct tgsi_full_dst_register
*dst
,
569 const struct tgsi_full_src_register
*src
,
570 LLVMValueRef vertex_dw_stride
,
571 LLVMValueRef base_addr
)
573 struct gallivm_state
*gallivm
= ctx
->radeon_bld
.soa
.bld_base
.base
.gallivm
;
574 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
575 ubyte
*name
, *index
, *array_first
;
577 struct tgsi_full_dst_register reg
;
579 /* Set the register description. The address computation is the same
580 * for sources and destinations. */
582 reg
.Register
.File
= src
->Register
.File
;
583 reg
.Register
.Index
= src
->Register
.Index
;
584 reg
.Register
.Indirect
= src
->Register
.Indirect
;
585 reg
.Register
.Dimension
= src
->Register
.Dimension
;
586 reg
.Indirect
= src
->Indirect
;
587 reg
.Dimension
= src
->Dimension
;
588 reg
.DimIndirect
= src
->DimIndirect
;
592 /* If the register is 2-dimensional (e.g. an array of vertices
593 * in a primitive), calculate the base address of the vertex. */
594 if (reg
.Register
.Dimension
) {
597 if (reg
.Dimension
.Indirect
)
598 index
= get_indirect_index(ctx
, ®
.DimIndirect
,
599 reg
.Dimension
.Index
);
601 index
= lp_build_const_int32(gallivm
, reg
.Dimension
.Index
);
603 base_addr
= LLVMBuildAdd(gallivm
->builder
, base_addr
,
604 LLVMBuildMul(gallivm
->builder
, index
,
605 vertex_dw_stride
, ""), "");
608 /* Get information about the register. */
609 if (reg
.Register
.File
== TGSI_FILE_INPUT
) {
610 name
= info
->input_semantic_name
;
611 index
= info
->input_semantic_index
;
612 array_first
= info
->input_array_first
;
613 } else if (reg
.Register
.File
== TGSI_FILE_OUTPUT
) {
614 name
= info
->output_semantic_name
;
615 index
= info
->output_semantic_index
;
616 array_first
= info
->output_array_first
;
622 if (reg
.Register
.Indirect
) {
623 /* Add the relative address of the element. */
624 LLVMValueRef ind_index
;
626 if (reg
.Indirect
.ArrayID
)
627 first
= array_first
[reg
.Indirect
.ArrayID
];
629 first
= reg
.Register
.Index
;
631 ind_index
= get_indirect_index(ctx
, ®
.Indirect
,
632 reg
.Register
.Index
- first
);
634 base_addr
= LLVMBuildAdd(gallivm
->builder
, base_addr
,
635 LLVMBuildMul(gallivm
->builder
, ind_index
,
636 lp_build_const_int32(gallivm
, 4), ""), "");
638 param
= si_shader_io_get_unique_index(name
[first
], index
[first
]);
640 param
= si_shader_io_get_unique_index(name
[reg
.Register
.Index
],
641 index
[reg
.Register
.Index
]);
644 /* Add the base address of the element. */
645 return LLVMBuildAdd(gallivm
->builder
, base_addr
,
646 lp_build_const_int32(gallivm
, param
* 4), "");
649 /* The offchip buffer layout for TCS->TES is
651 * - attribute 0 of patch 0 vertex 0
652 * - attribute 0 of patch 0 vertex 1
653 * - attribute 0 of patch 0 vertex 2
655 * - attribute 0 of patch 1 vertex 0
656 * - attribute 0 of patch 1 vertex 1
658 * - attribute 1 of patch 0 vertex 0
659 * - attribute 1 of patch 0 vertex 1
661 * - per patch attribute 0 of patch 0
662 * - per patch attribute 0 of patch 1
665 * Note that every attribute has 4 components.
667 static LLVMValueRef
get_tcs_tes_buffer_address(struct si_shader_context
*ctx
,
668 LLVMValueRef vertex_index
,
669 LLVMValueRef param_index
)
671 struct gallivm_state
*gallivm
= ctx
->radeon_bld
.soa
.bld_base
.base
.gallivm
;
672 LLVMValueRef base_addr
, vertices_per_patch
, num_patches
, total_vertices
;
673 LLVMValueRef param_stride
, constant16
;
675 vertices_per_patch
= unpack_param(ctx
, SI_PARAM_TCS_OFFCHIP_LAYOUT
, 9, 6);
676 num_patches
= unpack_param(ctx
, SI_PARAM_TCS_OFFCHIP_LAYOUT
, 0, 9);
677 total_vertices
= LLVMBuildMul(gallivm
->builder
, vertices_per_patch
,
680 constant16
= lp_build_const_int32(gallivm
, 16);
682 base_addr
= LLVMBuildMul(gallivm
->builder
, get_rel_patch_id(ctx
),
683 vertices_per_patch
, "");
685 base_addr
= LLVMBuildAdd(gallivm
->builder
, base_addr
,
688 param_stride
= total_vertices
;
690 base_addr
= get_rel_patch_id(ctx
);
691 param_stride
= num_patches
;
694 base_addr
= LLVMBuildAdd(gallivm
->builder
, base_addr
,
695 LLVMBuildMul(gallivm
->builder
, param_index
,
696 param_stride
, ""), "");
698 base_addr
= LLVMBuildMul(gallivm
->builder
, base_addr
, constant16
, "");
701 LLVMValueRef patch_data_offset
=
702 unpack_param(ctx
, SI_PARAM_TCS_OFFCHIP_LAYOUT
, 16, 16);
704 base_addr
= LLVMBuildAdd(gallivm
->builder
, base_addr
,
705 patch_data_offset
, "");
710 static LLVMValueRef
get_tcs_tes_buffer_address_from_reg(
711 struct si_shader_context
*ctx
,
712 const struct tgsi_full_dst_register
*dst
,
713 const struct tgsi_full_src_register
*src
)
715 struct gallivm_state
*gallivm
= ctx
->radeon_bld
.soa
.bld_base
.base
.gallivm
;
716 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
717 ubyte
*name
, *index
, *array_first
;
718 struct tgsi_full_src_register reg
;
719 LLVMValueRef vertex_index
= NULL
;
720 LLVMValueRef param_index
= NULL
;
721 unsigned param_index_base
, param_base
;
723 reg
= src
? *src
: tgsi_full_src_register_from_dst(dst
);
725 if (reg
.Register
.Dimension
) {
727 if (reg
.Dimension
.Indirect
)
728 vertex_index
= get_indirect_index(ctx
, ®
.DimIndirect
,
729 reg
.Dimension
.Index
);
731 vertex_index
= lp_build_const_int32(gallivm
,
732 reg
.Dimension
.Index
);
735 /* Get information about the register. */
736 if (reg
.Register
.File
== TGSI_FILE_INPUT
) {
737 name
= info
->input_semantic_name
;
738 index
= info
->input_semantic_index
;
739 array_first
= info
->input_array_first
;
740 } else if (reg
.Register
.File
== TGSI_FILE_OUTPUT
) {
741 name
= info
->output_semantic_name
;
742 index
= info
->output_semantic_index
;
743 array_first
= info
->output_array_first
;
749 if (reg
.Register
.Indirect
) {
750 if (reg
.Indirect
.ArrayID
)
751 param_base
= array_first
[reg
.Indirect
.ArrayID
];
753 param_base
= reg
.Register
.Index
;
755 param_index
= get_indirect_index(ctx
, ®
.Indirect
,
756 reg
.Register
.Index
- param_base
);
759 param_base
= reg
.Register
.Index
;
760 param_index
= lp_build_const_int32(gallivm
, 0);
763 param_index_base
= si_shader_io_get_unique_index(name
[param_base
],
766 param_index
= LLVMBuildAdd(gallivm
->builder
, param_index
,
767 lp_build_const_int32(gallivm
, param_index_base
),
770 return get_tcs_tes_buffer_address(ctx
, vertex_index
, param_index
);
773 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
774 * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
775 * or v4i32 (num_channels=3,4). */
776 static void build_tbuffer_store(struct si_shader_context
*ctx
,
779 unsigned num_channels
,
781 LLVMValueRef soffset
,
782 unsigned inst_offset
,
791 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
792 LLVMValueRef args
[] = {
795 LLVMConstInt(ctx
->i32
, num_channels
, 0),
798 LLVMConstInt(ctx
->i32
, inst_offset
, 0),
799 LLVMConstInt(ctx
->i32
, dfmt
, 0),
800 LLVMConstInt(ctx
->i32
, nfmt
, 0),
801 LLVMConstInt(ctx
->i32
, offen
, 0),
802 LLVMConstInt(ctx
->i32
, idxen
, 0),
803 LLVMConstInt(ctx
->i32
, glc
, 0),
804 LLVMConstInt(ctx
->i32
, slc
, 0),
805 LLVMConstInt(ctx
->i32
, tfe
, 0)
808 /* The instruction offset field has 12 bits */
809 assert(offen
|| inst_offset
< (1 << 12));
811 /* The intrinsic is overloaded, we need to add a type suffix for overloading to work. */
812 unsigned func
= CLAMP(num_channels
, 1, 3) - 1;
813 const char *types
[] = {"i32", "v2i32", "v4i32"};
815 snprintf(name
, sizeof(name
), "llvm.SI.tbuffer.store.%s", types
[func
]);
817 lp_build_intrinsic(gallivm
->builder
, name
, ctx
->voidt
,
818 args
, ARRAY_SIZE(args
), 0);
821 static void build_tbuffer_store_dwords(struct si_shader_context
*ctx
,
824 unsigned num_channels
,
826 LLVMValueRef soffset
,
827 unsigned inst_offset
)
829 static unsigned dfmt
[] = {
830 V_008F0C_BUF_DATA_FORMAT_32
,
831 V_008F0C_BUF_DATA_FORMAT_32_32
,
832 V_008F0C_BUF_DATA_FORMAT_32_32_32
,
833 V_008F0C_BUF_DATA_FORMAT_32_32_32_32
835 assert(num_channels
>= 1 && num_channels
<= 4);
837 build_tbuffer_store(ctx
, rsrc
, vdata
, num_channels
, vaddr
, soffset
,
838 inst_offset
, dfmt
[num_channels
-1],
839 V_008F0C_BUF_NUM_FORMAT_UINT
, 1, 0, 1, 1, 0);
842 static LLVMValueRef
build_buffer_load(struct si_shader_context
*ctx
,
846 LLVMValueRef voffset
,
847 LLVMValueRef soffset
,
848 unsigned inst_offset
,
852 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
853 unsigned func
= CLAMP(num_channels
, 1, 3) - 1;
855 if (HAVE_LLVM
>= 0x309) {
856 LLVMValueRef args
[] = {
857 LLVMBuildBitCast(gallivm
->builder
, rsrc
, ctx
->v4i32
, ""),
858 vindex
? vindex
: LLVMConstInt(ctx
->i32
, 0, 0),
859 LLVMConstInt(ctx
->i32
, inst_offset
, 0),
860 LLVMConstInt(ctx
->i1
, glc
, 0),
861 LLVMConstInt(ctx
->i1
, slc
, 0)
864 LLVMTypeRef types
[] = {ctx
->f32
, LLVMVectorType(ctx
->f32
, 2),
866 const char *type_names
[] = {"f32", "v2f32", "v4f32"};
870 args
[2] = LLVMBuildAdd(gallivm
->builder
, args
[2], voffset
,
875 args
[2] = LLVMBuildAdd(gallivm
->builder
, args
[2], soffset
,
879 snprintf(name
, sizeof(name
), "llvm.amdgcn.buffer.load.%s",
882 return lp_build_intrinsic(gallivm
->builder
, name
, types
[func
], args
,
883 ARRAY_SIZE(args
), LLVMReadOnlyAttribute
);
885 LLVMValueRef args
[] = {
886 LLVMBuildBitCast(gallivm
->builder
, rsrc
, ctx
->v16i8
, ""),
887 voffset
? voffset
: vindex
,
889 LLVMConstInt(ctx
->i32
, inst_offset
, 0),
890 LLVMConstInt(ctx
->i32
, voffset
? 1 : 0, 0), // offen
891 LLVMConstInt(ctx
->i32
, vindex
? 1 : 0, 0), //idxen
892 LLVMConstInt(ctx
->i32
, glc
, 0),
893 LLVMConstInt(ctx
->i32
, slc
, 0),
894 LLVMConstInt(ctx
->i32
, 0, 0), // TFE
897 LLVMTypeRef types
[] = {ctx
->i32
, LLVMVectorType(ctx
->i32
, 2),
899 const char *type_names
[] = {"i32", "v2i32", "v4i32"};
900 const char *arg_type
= "i32";
903 if (voffset
&& vindex
) {
904 LLVMValueRef vaddr
[] = {vindex
, voffset
};
907 args
[1] = lp_build_gather_values(gallivm
, vaddr
, 2);
910 snprintf(name
, sizeof(name
), "llvm.SI.buffer.load.dword.%s.%s",
911 type_names
[func
], arg_type
);
913 return lp_build_intrinsic(gallivm
->builder
, name
, types
[func
], args
,
914 ARRAY_SIZE(args
), LLVMReadOnlyAttribute
);
918 static LLVMValueRef
buffer_load(struct lp_build_tgsi_context
*bld_base
,
919 enum tgsi_opcode_type type
, unsigned swizzle
,
920 LLVMValueRef buffer
, LLVMValueRef offset
,
923 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
924 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
925 LLVMValueRef value
, value2
;
926 LLVMTypeRef llvm_type
= tgsi2llvmtype(bld_base
, type
);
927 LLVMTypeRef vec_type
= LLVMVectorType(llvm_type
, 4);
930 value
= build_buffer_load(ctx
, buffer
, 4, NULL
, base
, offset
,
933 return LLVMBuildBitCast(gallivm
->builder
, value
, vec_type
, "");
936 if (!tgsi_type_is_64bit(type
)) {
937 value
= build_buffer_load(ctx
, buffer
, 4, NULL
, base
, offset
,
940 value
= LLVMBuildBitCast(gallivm
->builder
, value
, vec_type
, "");
941 return LLVMBuildExtractElement(gallivm
->builder
, value
,
942 lp_build_const_int32(gallivm
, swizzle
), "");
945 value
= build_buffer_load(ctx
, buffer
, 1, NULL
, base
, offset
,
948 value2
= build_buffer_load(ctx
, buffer
, 1, NULL
, base
, offset
,
949 swizzle
* 4 + 4, 1, 0);
951 return radeon_llvm_emit_fetch_64bit(bld_base
, type
, value
, value2
);
957 * \param type output value type
958 * \param swizzle offset (typically 0..3); it can be ~0, which loads a vec4
959 * \param dw_addr address in dwords
961 static LLVMValueRef
lds_load(struct lp_build_tgsi_context
*bld_base
,
962 enum tgsi_opcode_type type
, unsigned swizzle
,
963 LLVMValueRef dw_addr
)
965 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
966 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
970 LLVMValueRef values
[TGSI_NUM_CHANNELS
];
972 for (unsigned chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++)
973 values
[chan
] = lds_load(bld_base
, type
, chan
, dw_addr
);
975 return lp_build_gather_values(bld_base
->base
.gallivm
, values
,
979 dw_addr
= lp_build_add(&bld_base
->uint_bld
, dw_addr
,
980 lp_build_const_int32(gallivm
, swizzle
));
982 value
= build_indexed_load(ctx
, ctx
->lds
, dw_addr
, false);
983 if (tgsi_type_is_64bit(type
)) {
985 dw_addr
= lp_build_add(&bld_base
->uint_bld
, dw_addr
,
986 lp_build_const_int32(gallivm
, swizzle
+ 1));
987 value2
= build_indexed_load(ctx
, ctx
->lds
, dw_addr
, false);
988 return radeon_llvm_emit_fetch_64bit(bld_base
, type
, value
, value2
);
991 return LLVMBuildBitCast(gallivm
->builder
, value
,
992 tgsi2llvmtype(bld_base
, type
), "");
998 * \param swizzle offset (typically 0..3)
999 * \param dw_addr address in dwords
1000 * \param value value to store
1002 static void lds_store(struct lp_build_tgsi_context
*bld_base
,
1003 unsigned swizzle
, LLVMValueRef dw_addr
,
1006 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1007 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
1009 dw_addr
= lp_build_add(&bld_base
->uint_bld
, dw_addr
,
1010 lp_build_const_int32(gallivm
, swizzle
));
1012 value
= LLVMBuildBitCast(gallivm
->builder
, value
, ctx
->i32
, "");
1013 build_indexed_store(ctx
, ctx
->lds
,
1017 static LLVMValueRef
fetch_input_tcs(
1018 struct lp_build_tgsi_context
*bld_base
,
1019 const struct tgsi_full_src_register
*reg
,
1020 enum tgsi_opcode_type type
, unsigned swizzle
)
1022 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1023 LLVMValueRef dw_addr
, stride
;
1025 stride
= unpack_param(ctx
, SI_PARAM_TCS_IN_LAYOUT
, 13, 8);
1026 dw_addr
= get_tcs_in_current_patch_offset(ctx
);
1027 dw_addr
= get_dw_address(ctx
, NULL
, reg
, stride
, dw_addr
);
1029 return lds_load(bld_base
, type
, swizzle
, dw_addr
);
1032 static LLVMValueRef
fetch_output_tcs(
1033 struct lp_build_tgsi_context
*bld_base
,
1034 const struct tgsi_full_src_register
*reg
,
1035 enum tgsi_opcode_type type
, unsigned swizzle
)
1037 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1038 LLVMValueRef dw_addr
, stride
;
1040 if (reg
->Register
.Dimension
) {
1041 stride
= unpack_param(ctx
, SI_PARAM_TCS_OUT_LAYOUT
, 13, 8);
1042 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
1043 dw_addr
= get_dw_address(ctx
, NULL
, reg
, stride
, dw_addr
);
1045 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
1046 dw_addr
= get_dw_address(ctx
, NULL
, reg
, NULL
, dw_addr
);
1049 return lds_load(bld_base
, type
, swizzle
, dw_addr
);
1052 static LLVMValueRef
fetch_input_tes(
1053 struct lp_build_tgsi_context
*bld_base
,
1054 const struct tgsi_full_src_register
*reg
,
1055 enum tgsi_opcode_type type
, unsigned swizzle
)
1057 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1058 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
1059 LLVMValueRef rw_buffers
, buffer
, base
, addr
;
1061 rw_buffers
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
1062 SI_PARAM_RW_BUFFERS
);
1063 buffer
= build_indexed_load_const(ctx
, rw_buffers
,
1064 lp_build_const_int32(gallivm
, SI_HS_RING_TESS_OFFCHIP
));
1066 base
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, ctx
->param_oc_lds
);
1067 addr
= get_tcs_tes_buffer_address_from_reg(ctx
, NULL
, reg
);
1069 return buffer_load(bld_base
, type
, swizzle
, buffer
, base
, addr
);
1072 static void store_output_tcs(struct lp_build_tgsi_context
*bld_base
,
1073 const struct tgsi_full_instruction
*inst
,
1074 const struct tgsi_opcode_info
*info
,
1075 LLVMValueRef dst
[4])
1077 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1078 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
1079 const struct tgsi_full_dst_register
*reg
= &inst
->Dst
[0];
1080 unsigned chan_index
;
1081 LLVMValueRef dw_addr
, stride
;
1082 LLVMValueRef rw_buffers
, buffer
, base
, buf_addr
;
1083 LLVMValueRef values
[4];
1085 /* Only handle per-patch and per-vertex outputs here.
1086 * Vectors will be lowered to scalars and this function will be called again.
1088 if (reg
->Register
.File
!= TGSI_FILE_OUTPUT
||
1089 (dst
[0] && LLVMGetTypeKind(LLVMTypeOf(dst
[0])) == LLVMVectorTypeKind
)) {
1090 radeon_llvm_emit_store(bld_base
, inst
, info
, dst
);
1094 if (reg
->Register
.Dimension
) {
1095 stride
= unpack_param(ctx
, SI_PARAM_TCS_OUT_LAYOUT
, 13, 8);
1096 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
1097 dw_addr
= get_dw_address(ctx
, reg
, NULL
, stride
, dw_addr
);
1099 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
1100 dw_addr
= get_dw_address(ctx
, reg
, NULL
, NULL
, dw_addr
);
1103 rw_buffers
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
1104 SI_PARAM_RW_BUFFERS
);
1105 buffer
= build_indexed_load_const(ctx
, rw_buffers
,
1106 lp_build_const_int32(gallivm
, SI_HS_RING_TESS_OFFCHIP
));
1108 base
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, ctx
->param_oc_lds
);
1109 buf_addr
= get_tcs_tes_buffer_address_from_reg(ctx
, reg
, NULL
);
1112 TGSI_FOR_EACH_DST0_ENABLED_CHANNEL(inst
, chan_index
) {
1113 LLVMValueRef value
= dst
[chan_index
];
1115 if (inst
->Instruction
.Saturate
)
1116 value
= radeon_llvm_saturate(bld_base
, value
);
1118 lds_store(bld_base
, chan_index
, dw_addr
, value
);
1120 value
= LLVMBuildBitCast(gallivm
->builder
, value
, ctx
->i32
, "");
1121 values
[chan_index
] = value
;
1123 if (inst
->Dst
[0].Register
.WriteMask
!= 0xF) {
1124 build_tbuffer_store_dwords(ctx
, buffer
, value
, 1,
1130 if (inst
->Dst
[0].Register
.WriteMask
== 0xF) {
1131 LLVMValueRef value
= lp_build_gather_values(bld_base
->base
.gallivm
,
1133 build_tbuffer_store_dwords(ctx
, buffer
, value
, 4, buf_addr
,
1138 static LLVMValueRef
fetch_input_gs(
1139 struct lp_build_tgsi_context
*bld_base
,
1140 const struct tgsi_full_src_register
*reg
,
1141 enum tgsi_opcode_type type
,
1144 struct lp_build_context
*base
= &bld_base
->base
;
1145 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1146 struct si_shader
*shader
= ctx
->shader
;
1147 struct lp_build_context
*uint
= &ctx
->radeon_bld
.soa
.bld_base
.uint_bld
;
1148 struct gallivm_state
*gallivm
= base
->gallivm
;
1149 LLVMValueRef vtx_offset
;
1150 LLVMValueRef args
[9];
1151 unsigned vtx_offset_param
;
1152 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
1153 unsigned semantic_name
= info
->input_semantic_name
[reg
->Register
.Index
];
1154 unsigned semantic_index
= info
->input_semantic_index
[reg
->Register
.Index
];
1158 if (swizzle
!= ~0 && semantic_name
== TGSI_SEMANTIC_PRIMID
)
1159 return get_primitive_id(bld_base
, swizzle
);
1161 if (!reg
->Register
.Dimension
)
1164 if (swizzle
== ~0) {
1165 LLVMValueRef values
[TGSI_NUM_CHANNELS
];
1167 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
1168 values
[chan
] = fetch_input_gs(bld_base
, reg
, type
, chan
);
1170 return lp_build_gather_values(bld_base
->base
.gallivm
, values
,
1174 /* Get the vertex offset parameter */
1175 vtx_offset_param
= reg
->Dimension
.Index
;
1176 if (vtx_offset_param
< 2) {
1177 vtx_offset_param
+= SI_PARAM_VTX0_OFFSET
;
1179 assert(vtx_offset_param
< 6);
1180 vtx_offset_param
+= SI_PARAM_VTX2_OFFSET
- 2;
1182 vtx_offset
= lp_build_mul_imm(uint
,
1183 LLVMGetParam(ctx
->radeon_bld
.main_fn
,
1187 param
= si_shader_io_get_unique_index(semantic_name
, semantic_index
);
1188 args
[0] = ctx
->esgs_ring
;
1189 args
[1] = vtx_offset
;
1190 args
[2] = lp_build_const_int32(gallivm
, (param
* 4 + swizzle
) * 256);
1191 args
[3] = uint
->zero
;
1192 args
[4] = uint
->one
; /* OFFEN */
1193 args
[5] = uint
->zero
; /* IDXEN */
1194 args
[6] = uint
->one
; /* GLC */
1195 args
[7] = uint
->zero
; /* SLC */
1196 args
[8] = uint
->zero
; /* TFE */
1198 value
= lp_build_intrinsic(gallivm
->builder
,
1199 "llvm.SI.buffer.load.dword.i32.i32",
1201 LLVMReadOnlyAttribute
);
1202 if (tgsi_type_is_64bit(type
)) {
1203 LLVMValueRef value2
;
1204 args
[2] = lp_build_const_int32(gallivm
, (param
* 4 + swizzle
+ 1) * 256);
1205 value2
= lp_build_intrinsic(gallivm
->builder
,
1206 "llvm.SI.buffer.load.dword.i32.i32",
1208 LLVMReadOnlyAttribute
);
1209 return radeon_llvm_emit_fetch_64bit(bld_base
, type
,
1212 return LLVMBuildBitCast(gallivm
->builder
,
1214 tgsi2llvmtype(bld_base
, type
), "");
1217 static int lookup_interp_param_index(unsigned interpolate
, unsigned location
)
1219 switch (interpolate
) {
1220 case TGSI_INTERPOLATE_CONSTANT
:
1223 case TGSI_INTERPOLATE_LINEAR
:
1224 if (location
== TGSI_INTERPOLATE_LOC_SAMPLE
)
1225 return SI_PARAM_LINEAR_SAMPLE
;
1226 else if (location
== TGSI_INTERPOLATE_LOC_CENTROID
)
1227 return SI_PARAM_LINEAR_CENTROID
;
1229 return SI_PARAM_LINEAR_CENTER
;
1231 case TGSI_INTERPOLATE_COLOR
:
1232 case TGSI_INTERPOLATE_PERSPECTIVE
:
1233 if (location
== TGSI_INTERPOLATE_LOC_SAMPLE
)
1234 return SI_PARAM_PERSP_SAMPLE
;
1235 else if (location
== TGSI_INTERPOLATE_LOC_CENTROID
)
1236 return SI_PARAM_PERSP_CENTROID
;
1238 return SI_PARAM_PERSP_CENTER
;
1241 fprintf(stderr
, "Warning: Unhandled interpolation mode.\n");
1246 /* This shouldn't be used by explicit INTERP opcodes. */
1247 static unsigned select_interp_param(struct si_shader_context
*ctx
,
1250 if (!ctx
->is_monolithic
)
1253 if (ctx
->shader
->key
.ps
.prolog
.force_persp_sample_interp
) {
1255 case SI_PARAM_PERSP_CENTROID
:
1256 case SI_PARAM_PERSP_CENTER
:
1257 return SI_PARAM_PERSP_SAMPLE
;
1260 if (ctx
->shader
->key
.ps
.prolog
.force_linear_sample_interp
) {
1262 case SI_PARAM_LINEAR_CENTROID
:
1263 case SI_PARAM_LINEAR_CENTER
:
1264 return SI_PARAM_LINEAR_SAMPLE
;
1267 if (ctx
->shader
->key
.ps
.prolog
.force_persp_center_interp
) {
1269 case SI_PARAM_PERSP_CENTROID
:
1270 case SI_PARAM_PERSP_SAMPLE
:
1271 return SI_PARAM_PERSP_CENTER
;
1274 if (ctx
->shader
->key
.ps
.prolog
.force_linear_center_interp
) {
1276 case SI_PARAM_LINEAR_CENTROID
:
1277 case SI_PARAM_LINEAR_SAMPLE
:
1278 return SI_PARAM_LINEAR_CENTER
;
1286 * Interpolate a fragment shader input.
1288 * @param ctx context
1289 * @param input_index index of the input in hardware
1290 * @param semantic_name TGSI_SEMANTIC_*
1291 * @param semantic_index semantic index
1292 * @param num_interp_inputs number of all interpolated inputs (= BCOLOR offset)
1293 * @param colors_read_mask color components read (4 bits for each color, 8 bits in total)
1294 * @param interp_param interpolation weights (i,j)
1295 * @param prim_mask SI_PARAM_PRIM_MASK
1296 * @param face SI_PARAM_FRONT_FACE
1297 * @param result the return value (4 components)
1299 static void interp_fs_input(struct si_shader_context
*ctx
,
1300 unsigned input_index
,
1301 unsigned semantic_name
,
1302 unsigned semantic_index
,
1303 unsigned num_interp_inputs
,
1304 unsigned colors_read_mask
,
1305 LLVMValueRef interp_param
,
1306 LLVMValueRef prim_mask
,
1308 LLVMValueRef result
[4])
1310 struct lp_build_context
*base
= &ctx
->radeon_bld
.soa
.bld_base
.base
;
1311 struct lp_build_context
*uint
= &ctx
->radeon_bld
.soa
.bld_base
.uint_bld
;
1312 struct gallivm_state
*gallivm
= base
->gallivm
;
1313 const char *intr_name
;
1314 LLVMValueRef attr_number
;
1318 attr_number
= lp_build_const_int32(gallivm
, input_index
);
1320 /* fs.constant returns the param from the middle vertex, so it's not
1321 * really useful for flat shading. It's meant to be used for custom
1322 * interpolation (but the intrinsic can't fetch from the other two
1325 * Luckily, it doesn't matter, because we rely on the FLAT_SHADE state
1326 * to do the right thing. The only reason we use fs.constant is that
1327 * fs.interp cannot be used on integers, because they can be equal
1330 intr_name
= interp_param
? "llvm.SI.fs.interp" : "llvm.SI.fs.constant";
1332 if (semantic_name
== TGSI_SEMANTIC_COLOR
&&
1333 ctx
->shader
->key
.ps
.prolog
.color_two_side
) {
1334 LLVMValueRef args
[4];
1335 LLVMValueRef is_face_positive
;
1336 LLVMValueRef back_attr_number
;
1338 /* If BCOLOR0 is used, BCOLOR1 is at offset "num_inputs + 1",
1339 * otherwise it's at offset "num_inputs".
1341 unsigned back_attr_offset
= num_interp_inputs
;
1342 if (semantic_index
== 1 && colors_read_mask
& 0xf)
1343 back_attr_offset
+= 1;
1345 back_attr_number
= lp_build_const_int32(gallivm
, back_attr_offset
);
1347 is_face_positive
= LLVMBuildICmp(gallivm
->builder
, LLVMIntNE
,
1348 face
, uint
->zero
, "");
1350 args
[2] = prim_mask
;
1351 args
[3] = interp_param
;
1352 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
1353 LLVMValueRef llvm_chan
= lp_build_const_int32(gallivm
, chan
);
1354 LLVMValueRef front
, back
;
1356 args
[0] = llvm_chan
;
1357 args
[1] = attr_number
;
1358 front
= lp_build_intrinsic(gallivm
->builder
, intr_name
,
1359 ctx
->f32
, args
, args
[3] ? 4 : 3,
1360 LLVMReadNoneAttribute
);
1362 args
[1] = back_attr_number
;
1363 back
= lp_build_intrinsic(gallivm
->builder
, intr_name
,
1364 ctx
->f32
, args
, args
[3] ? 4 : 3,
1365 LLVMReadNoneAttribute
);
1367 result
[chan
] = LLVMBuildSelect(gallivm
->builder
,
1373 } else if (semantic_name
== TGSI_SEMANTIC_FOG
) {
1374 LLVMValueRef args
[4];
1376 args
[0] = uint
->zero
;
1377 args
[1] = attr_number
;
1378 args
[2] = prim_mask
;
1379 args
[3] = interp_param
;
1380 result
[0] = lp_build_intrinsic(gallivm
->builder
, intr_name
,
1381 ctx
->f32
, args
, args
[3] ? 4 : 3,
1382 LLVMReadNoneAttribute
);
1384 result
[2] = lp_build_const_float(gallivm
, 0.0f
);
1385 result
[3] = lp_build_const_float(gallivm
, 1.0f
);
1387 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
1388 LLVMValueRef args
[4];
1389 LLVMValueRef llvm_chan
= lp_build_const_int32(gallivm
, chan
);
1391 args
[0] = llvm_chan
;
1392 args
[1] = attr_number
;
1393 args
[2] = prim_mask
;
1394 args
[3] = interp_param
;
1395 result
[chan
] = lp_build_intrinsic(gallivm
->builder
, intr_name
,
1396 ctx
->f32
, args
, args
[3] ? 4 : 3,
1397 LLVMReadNoneAttribute
);
1402 /* LLVMGetParam with bc_optimize resolved. */
1403 static LLVMValueRef
get_interp_param(struct si_shader_context
*ctx
,
1404 int interp_param_idx
)
1406 LLVMBuilderRef builder
= ctx
->radeon_bld
.gallivm
.builder
;
1407 LLVMValueRef main_fn
= ctx
->radeon_bld
.main_fn
;
1408 LLVMValueRef param
= NULL
;
1410 /* Handle PRIM_MASK[31] (bc_optimize). */
1411 if (ctx
->is_monolithic
&&
1412 ((ctx
->shader
->key
.ps
.prolog
.bc_optimize_for_persp
&&
1413 interp_param_idx
== SI_PARAM_PERSP_CENTROID
) ||
1414 (ctx
->shader
->key
.ps
.prolog
.bc_optimize_for_linear
&&
1415 interp_param_idx
== SI_PARAM_LINEAR_CENTROID
))) {
1416 /* The shader should do: if (PRIM_MASK[31]) CENTROID = CENTER;
1417 * The hw doesn't compute CENTROID if the whole wave only
1418 * contains fully-covered quads.
1420 LLVMValueRef bc_optimize
=
1421 LLVMGetParam(main_fn
, SI_PARAM_PRIM_MASK
);
1422 bc_optimize
= LLVMBuildLShr(builder
,
1424 LLVMConstInt(ctx
->i32
, 31, 0), "");
1425 bc_optimize
= LLVMBuildTrunc(builder
, bc_optimize
, ctx
->i1
, "");
1427 if (ctx
->shader
->key
.ps
.prolog
.bc_optimize_for_persp
&&
1428 interp_param_idx
== SI_PARAM_PERSP_CENTROID
) {
1429 param
= LLVMBuildSelect(builder
, bc_optimize
,
1430 LLVMGetParam(main_fn
,
1431 SI_PARAM_PERSP_CENTER
),
1432 LLVMGetParam(main_fn
,
1433 SI_PARAM_PERSP_CENTROID
),
1436 if (ctx
->shader
->key
.ps
.prolog
.bc_optimize_for_linear
&&
1437 interp_param_idx
== SI_PARAM_LINEAR_CENTROID
) {
1438 param
= LLVMBuildSelect(builder
, bc_optimize
,
1439 LLVMGetParam(main_fn
,
1440 SI_PARAM_LINEAR_CENTER
),
1441 LLVMGetParam(main_fn
,
1442 SI_PARAM_LINEAR_CENTROID
),
1448 param
= LLVMGetParam(main_fn
, interp_param_idx
);
1452 static void declare_input_fs(
1453 struct radeon_llvm_context
*radeon_bld
,
1454 unsigned input_index
,
1455 const struct tgsi_full_declaration
*decl
,
1456 LLVMValueRef out
[4])
1458 struct lp_build_context
*base
= &radeon_bld
->soa
.bld_base
.base
;
1459 struct si_shader_context
*ctx
=
1460 si_shader_context(&radeon_bld
->soa
.bld_base
);
1461 struct si_shader
*shader
= ctx
->shader
;
1462 LLVMValueRef main_fn
= radeon_bld
->main_fn
;
1463 LLVMValueRef interp_param
= NULL
;
1464 int interp_param_idx
;
1466 /* Get colors from input VGPRs (set by the prolog). */
1467 if (!ctx
->is_monolithic
&&
1468 decl
->Semantic
.Name
== TGSI_SEMANTIC_COLOR
) {
1469 unsigned i
= decl
->Semantic
.Index
;
1470 unsigned colors_read
= shader
->selector
->info
.colors_read
;
1471 unsigned mask
= colors_read
>> (i
* 4);
1472 unsigned offset
= SI_PARAM_POS_FIXED_PT
+ 1 +
1473 (i
? util_bitcount(colors_read
& 0xf) : 0);
1475 out
[0] = mask
& 0x1 ? LLVMGetParam(main_fn
, offset
++) : base
->undef
;
1476 out
[1] = mask
& 0x2 ? LLVMGetParam(main_fn
, offset
++) : base
->undef
;
1477 out
[2] = mask
& 0x4 ? LLVMGetParam(main_fn
, offset
++) : base
->undef
;
1478 out
[3] = mask
& 0x8 ? LLVMGetParam(main_fn
, offset
++) : base
->undef
;
1482 interp_param_idx
= lookup_interp_param_index(decl
->Interp
.Interpolate
,
1483 decl
->Interp
.Location
);
1484 if (interp_param_idx
== -1)
1486 else if (interp_param_idx
) {
1487 interp_param_idx
= select_interp_param(ctx
,
1489 interp_param
= get_interp_param(ctx
, interp_param_idx
);
1492 if (decl
->Semantic
.Name
== TGSI_SEMANTIC_COLOR
&&
1493 decl
->Interp
.Interpolate
== TGSI_INTERPOLATE_COLOR
&&
1494 ctx
->shader
->key
.ps
.prolog
.flatshade_colors
)
1495 interp_param
= NULL
; /* load the constant color */
1497 interp_fs_input(ctx
, input_index
, decl
->Semantic
.Name
,
1498 decl
->Semantic
.Index
, shader
->selector
->info
.num_inputs
,
1499 shader
->selector
->info
.colors_read
, interp_param
,
1500 LLVMGetParam(main_fn
, SI_PARAM_PRIM_MASK
),
1501 LLVMGetParam(main_fn
, SI_PARAM_FRONT_FACE
),
1505 static LLVMValueRef
get_sample_id(struct radeon_llvm_context
*radeon_bld
)
1507 return unpack_param(si_shader_context(&radeon_bld
->soa
.bld_base
),
1508 SI_PARAM_ANCILLARY
, 8, 4);
1512 * Set range metadata on an instruction. This can only be used on load and
1513 * call instructions. If you know an instruction can only produce the values
1514 * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
1515 * \p lo is the minimum value inclusive.
1516 * \p hi is the maximum value exclusive.
1518 static void set_range_metadata(struct si_shader_context
*ctx
,
1519 LLVMValueRef value
, unsigned lo
, unsigned hi
)
1521 LLVMValueRef range_md
, md_args
[2];
1522 LLVMTypeRef type
= LLVMTypeOf(value
);
1523 LLVMContextRef context
= LLVMGetTypeContext(type
);
1525 md_args
[0] = LLVMConstInt(type
, lo
, false);
1526 md_args
[1] = LLVMConstInt(type
, hi
, false);
1527 range_md
= LLVMMDNodeInContext(context
, md_args
, 2);
1528 LLVMSetMetadata(value
, ctx
->range_md_kind
, range_md
);
1531 static LLVMValueRef
get_thread_id(struct si_shader_context
*ctx
)
1533 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
1536 if (HAVE_LLVM
< 0x0308) {
1537 tid
= lp_build_intrinsic(gallivm
->builder
, "llvm.SI.tid",
1538 ctx
->i32
, NULL
, 0, LLVMReadNoneAttribute
);
1540 LLVMValueRef tid_args
[2];
1541 tid_args
[0] = lp_build_const_int32(gallivm
, 0xffffffff);
1542 tid_args
[1] = lp_build_const_int32(gallivm
, 0);
1543 tid_args
[1] = lp_build_intrinsic(gallivm
->builder
,
1544 "llvm.amdgcn.mbcnt.lo", ctx
->i32
,
1545 tid_args
, 2, LLVMReadNoneAttribute
);
1547 tid
= lp_build_intrinsic(gallivm
->builder
,
1548 "llvm.amdgcn.mbcnt.hi", ctx
->i32
,
1549 tid_args
, 2, LLVMReadNoneAttribute
);
1551 set_range_metadata(ctx
, tid
, 0, 64);
1556 * Load a dword from a constant buffer.
1558 static LLVMValueRef
buffer_load_const(struct si_shader_context
*ctx
,
1559 LLVMValueRef resource
,
1560 LLVMValueRef offset
)
1562 LLVMBuilderRef builder
= ctx
->radeon_bld
.gallivm
.builder
;
1563 LLVMValueRef args
[2] = {resource
, offset
};
1565 return lp_build_intrinsic(builder
, "llvm.SI.load.const", ctx
->f32
, args
, 2,
1566 LLVMReadNoneAttribute
);
1569 static LLVMValueRef
load_sample_position(struct radeon_llvm_context
*radeon_bld
, LLVMValueRef sample_id
)
1571 struct si_shader_context
*ctx
=
1572 si_shader_context(&radeon_bld
->soa
.bld_base
);
1573 struct lp_build_context
*uint_bld
= &radeon_bld
->soa
.bld_base
.uint_bld
;
1574 struct gallivm_state
*gallivm
= &radeon_bld
->gallivm
;
1575 LLVMBuilderRef builder
= gallivm
->builder
;
1576 LLVMValueRef desc
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_RW_BUFFERS
);
1577 LLVMValueRef buf_index
= lp_build_const_int32(gallivm
, SI_PS_CONST_SAMPLE_POSITIONS
);
1578 LLVMValueRef resource
= build_indexed_load_const(ctx
, desc
, buf_index
);
1580 /* offset = sample_id * 8 (8 = 2 floats containing samplepos.xy) */
1581 LLVMValueRef offset0
= lp_build_mul_imm(uint_bld
, sample_id
, 8);
1582 LLVMValueRef offset1
= LLVMBuildAdd(builder
, offset0
, lp_build_const_int32(gallivm
, 4), "");
1584 LLVMValueRef pos
[4] = {
1585 buffer_load_const(ctx
, resource
, offset0
),
1586 buffer_load_const(ctx
, resource
, offset1
),
1587 lp_build_const_float(gallivm
, 0),
1588 lp_build_const_float(gallivm
, 0)
1591 return lp_build_gather_values(gallivm
, pos
, 4);
1594 static void declare_system_value(
1595 struct radeon_llvm_context
*radeon_bld
,
1597 const struct tgsi_full_declaration
*decl
)
1599 struct si_shader_context
*ctx
=
1600 si_shader_context(&radeon_bld
->soa
.bld_base
);
1601 struct lp_build_context
*bld
= &radeon_bld
->soa
.bld_base
.base
;
1602 struct gallivm_state
*gallivm
= &radeon_bld
->gallivm
;
1603 LLVMValueRef value
= 0;
1605 switch (decl
->Semantic
.Name
) {
1606 case TGSI_SEMANTIC_INSTANCEID
:
1607 value
= LLVMGetParam(radeon_bld
->main_fn
,
1608 ctx
->param_instance_id
);
1611 case TGSI_SEMANTIC_VERTEXID
:
1612 value
= LLVMBuildAdd(gallivm
->builder
,
1613 LLVMGetParam(radeon_bld
->main_fn
,
1614 ctx
->param_vertex_id
),
1615 LLVMGetParam(radeon_bld
->main_fn
,
1616 SI_PARAM_BASE_VERTEX
), "");
1619 case TGSI_SEMANTIC_VERTEXID_NOBASE
:
1620 value
= LLVMGetParam(radeon_bld
->main_fn
,
1621 ctx
->param_vertex_id
);
1624 case TGSI_SEMANTIC_BASEVERTEX
:
1625 value
= LLVMGetParam(radeon_bld
->main_fn
,
1626 SI_PARAM_BASE_VERTEX
);
1629 case TGSI_SEMANTIC_BASEINSTANCE
:
1630 value
= LLVMGetParam(radeon_bld
->main_fn
,
1631 SI_PARAM_START_INSTANCE
);
1634 case TGSI_SEMANTIC_DRAWID
:
1635 value
= LLVMGetParam(radeon_bld
->main_fn
,
1639 case TGSI_SEMANTIC_INVOCATIONID
:
1640 if (ctx
->type
== PIPE_SHADER_TESS_CTRL
)
1641 value
= unpack_param(ctx
, SI_PARAM_REL_IDS
, 8, 5);
1642 else if (ctx
->type
== PIPE_SHADER_GEOMETRY
)
1643 value
= LLVMGetParam(radeon_bld
->main_fn
,
1644 SI_PARAM_GS_INSTANCE_ID
);
1646 assert(!"INVOCATIONID not implemented");
1649 case TGSI_SEMANTIC_POSITION
:
1651 LLVMValueRef pos
[4] = {
1652 LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_POS_X_FLOAT
),
1653 LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_POS_Y_FLOAT
),
1654 LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_POS_Z_FLOAT
),
1655 lp_build_emit_llvm_unary(&radeon_bld
->soa
.bld_base
, TGSI_OPCODE_RCP
,
1656 LLVMGetParam(radeon_bld
->main_fn
,
1657 SI_PARAM_POS_W_FLOAT
)),
1659 value
= lp_build_gather_values(gallivm
, pos
, 4);
1663 case TGSI_SEMANTIC_FACE
:
1664 value
= LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_FRONT_FACE
);
1667 case TGSI_SEMANTIC_SAMPLEID
:
1668 value
= get_sample_id(radeon_bld
);
1671 case TGSI_SEMANTIC_SAMPLEPOS
: {
1672 LLVMValueRef pos
[4] = {
1673 LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_POS_X_FLOAT
),
1674 LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_POS_Y_FLOAT
),
1675 lp_build_const_float(gallivm
, 0),
1676 lp_build_const_float(gallivm
, 0)
1678 pos
[0] = lp_build_emit_llvm_unary(&radeon_bld
->soa
.bld_base
,
1679 TGSI_OPCODE_FRC
, pos
[0]);
1680 pos
[1] = lp_build_emit_llvm_unary(&radeon_bld
->soa
.bld_base
,
1681 TGSI_OPCODE_FRC
, pos
[1]);
1682 value
= lp_build_gather_values(gallivm
, pos
, 4);
1686 case TGSI_SEMANTIC_SAMPLEMASK
:
1687 /* This can only occur with the OpenGL Core profile, which
1688 * doesn't support smoothing.
1690 value
= LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_SAMPLE_COVERAGE
);
1693 case TGSI_SEMANTIC_TESSCOORD
:
1695 LLVMValueRef coord
[4] = {
1696 LLVMGetParam(radeon_bld
->main_fn
, ctx
->param_tes_u
),
1697 LLVMGetParam(radeon_bld
->main_fn
, ctx
->param_tes_v
),
1702 /* For triangles, the vector should be (u, v, 1-u-v). */
1703 if (ctx
->shader
->selector
->info
.properties
[TGSI_PROPERTY_TES_PRIM_MODE
] ==
1704 PIPE_PRIM_TRIANGLES
)
1705 coord
[2] = lp_build_sub(bld
, bld
->one
,
1706 lp_build_add(bld
, coord
[0], coord
[1]));
1708 value
= lp_build_gather_values(gallivm
, coord
, 4);
1712 case TGSI_SEMANTIC_VERTICESIN
:
1713 if (ctx
->type
== PIPE_SHADER_TESS_CTRL
)
1714 value
= unpack_param(ctx
, SI_PARAM_TCS_OUT_LAYOUT
, 26, 6);
1715 else if (ctx
->type
== PIPE_SHADER_TESS_EVAL
)
1716 value
= unpack_param(ctx
, SI_PARAM_TCS_OFFCHIP_LAYOUT
, 9, 7);
1718 assert(!"invalid shader stage for TGSI_SEMANTIC_VERTICESIN");
1721 case TGSI_SEMANTIC_TESSINNER
:
1722 case TGSI_SEMANTIC_TESSOUTER
:
1724 LLVMValueRef rw_buffers
, buffer
, base
, addr
;
1725 int param
= si_shader_io_get_unique_index(decl
->Semantic
.Name
, 0);
1727 rw_buffers
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
1728 SI_PARAM_RW_BUFFERS
);
1729 buffer
= build_indexed_load_const(ctx
, rw_buffers
,
1730 lp_build_const_int32(gallivm
, SI_HS_RING_TESS_OFFCHIP
));
1732 base
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, ctx
->param_oc_lds
);
1733 addr
= get_tcs_tes_buffer_address(ctx
, NULL
,
1734 lp_build_const_int32(gallivm
, param
));
1736 value
= buffer_load(&radeon_bld
->soa
.bld_base
, TGSI_TYPE_FLOAT
,
1737 ~0, buffer
, base
, addr
);
1742 case TGSI_SEMANTIC_DEFAULT_TESSOUTER_SI
:
1743 case TGSI_SEMANTIC_DEFAULT_TESSINNER_SI
:
1745 LLVMValueRef buf
, slot
, val
[4];
1748 slot
= lp_build_const_int32(gallivm
, SI_HS_CONST_DEFAULT_TESS_LEVELS
);
1749 buf
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_RW_BUFFERS
);
1750 buf
= build_indexed_load_const(ctx
, buf
, slot
);
1751 offset
= decl
->Semantic
.Name
== TGSI_SEMANTIC_DEFAULT_TESSINNER_SI
? 4 : 0;
1753 for (i
= 0; i
< 4; i
++)
1754 val
[i
] = buffer_load_const(ctx
, buf
,
1755 lp_build_const_int32(gallivm
, (offset
+ i
) * 4));
1756 value
= lp_build_gather_values(gallivm
, val
, 4);
1760 case TGSI_SEMANTIC_PRIMID
:
1761 value
= get_primitive_id(&radeon_bld
->soa
.bld_base
, 0);
1764 case TGSI_SEMANTIC_GRID_SIZE
:
1765 value
= LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_GRID_SIZE
);
1768 case TGSI_SEMANTIC_BLOCK_SIZE
:
1770 LLVMValueRef values
[3];
1772 unsigned *properties
= ctx
->shader
->selector
->info
.properties
;
1773 unsigned sizes
[3] = {
1774 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
],
1775 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT
],
1776 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH
]
1779 for (i
= 0; i
< 3; ++i
)
1780 values
[i
] = lp_build_const_int32(gallivm
, sizes
[i
]);
1782 value
= lp_build_gather_values(gallivm
, values
, 3);
1786 case TGSI_SEMANTIC_BLOCK_ID
:
1787 value
= LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_BLOCK_ID
);
1790 case TGSI_SEMANTIC_THREAD_ID
:
1791 value
= LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_THREAD_ID
);
1794 #if HAVE_LLVM >= 0x0309
1795 case TGSI_SEMANTIC_HELPER_INVOCATION
:
1796 value
= lp_build_intrinsic(gallivm
->builder
,
1797 "llvm.amdgcn.ps.live",
1799 LLVMReadNoneAttribute
);
1800 value
= LLVMBuildNot(gallivm
->builder
, value
, "");
1801 value
= LLVMBuildSExt(gallivm
->builder
, value
, ctx
->i32
, "");
1806 assert(!"unknown system value");
1810 radeon_bld
->system_values
[index
] = value
;
1813 static void declare_compute_memory(struct radeon_llvm_context
*radeon_bld
,
1814 const struct tgsi_full_declaration
*decl
)
1816 struct si_shader_context
*ctx
=
1817 si_shader_context(&radeon_bld
->soa
.bld_base
);
1818 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
1819 struct gallivm_state
*gallivm
= &radeon_bld
->gallivm
;
1821 LLVMTypeRef i8p
= LLVMPointerType(ctx
->i8
, LOCAL_ADDR_SPACE
);
1824 assert(decl
->Declaration
.MemType
== TGSI_MEMORY_TYPE_SHARED
);
1825 assert(decl
->Range
.First
== decl
->Range
.Last
);
1826 assert(!ctx
->shared_memory
);
1828 var
= LLVMAddGlobalInAddressSpace(gallivm
->module
,
1829 LLVMArrayType(ctx
->i8
, sel
->local_size
),
1832 LLVMSetAlignment(var
, 4);
1834 ctx
->shared_memory
= LLVMBuildBitCast(gallivm
->builder
, var
, i8p
, "");
1837 static LLVMValueRef
load_const_buffer_desc(struct si_shader_context
*ctx
, int i
)
1839 LLVMValueRef list_ptr
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
1840 SI_PARAM_CONST_BUFFERS
);
1842 return build_indexed_load_const(ctx
, list_ptr
,
1843 LLVMConstInt(ctx
->i32
, i
, 0));
1846 static LLVMValueRef
fetch_constant(
1847 struct lp_build_tgsi_context
*bld_base
,
1848 const struct tgsi_full_src_register
*reg
,
1849 enum tgsi_opcode_type type
,
1852 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1853 struct lp_build_context
*base
= &bld_base
->base
;
1854 const struct tgsi_ind_register
*ireg
= ®
->Indirect
;
1857 LLVMValueRef addr
, bufp
;
1858 LLVMValueRef result
;
1860 if (swizzle
== LP_CHAN_ALL
) {
1862 LLVMValueRef values
[4];
1863 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
)
1864 values
[chan
] = fetch_constant(bld_base
, reg
, type
, chan
);
1866 return lp_build_gather_values(bld_base
->base
.gallivm
, values
, 4);
1869 buf
= reg
->Register
.Dimension
? reg
->Dimension
.Index
: 0;
1870 idx
= reg
->Register
.Index
* 4 + swizzle
;
1872 if (!reg
->Register
.Indirect
&& !reg
->Dimension
.Indirect
) {
1873 LLVMValueRef c0
, c1
, desc
;
1875 desc
= load_const_buffer_desc(ctx
, buf
);
1876 c0
= buffer_load_const(ctx
, desc
,
1877 LLVMConstInt(ctx
->i32
, idx
* 4, 0));
1879 if (!tgsi_type_is_64bit(type
))
1880 return bitcast(bld_base
, type
, c0
);
1882 c1
= buffer_load_const(ctx
, desc
,
1883 LLVMConstInt(ctx
->i32
,
1885 return radeon_llvm_emit_fetch_64bit(bld_base
, type
,
1890 if (reg
->Register
.Dimension
&& reg
->Dimension
.Indirect
) {
1891 LLVMValueRef ptr
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_CONST_BUFFERS
);
1893 index
= get_bounded_indirect_index(ctx
, ®
->DimIndirect
,
1894 reg
->Dimension
.Index
,
1895 SI_NUM_CONST_BUFFERS
);
1896 bufp
= build_indexed_load_const(ctx
, ptr
, index
);
1898 bufp
= load_const_buffer_desc(ctx
, buf
);
1900 addr
= ctx
->radeon_bld
.soa
.addr
[ireg
->Index
][ireg
->Swizzle
];
1901 addr
= LLVMBuildLoad(base
->gallivm
->builder
, addr
, "load addr reg");
1902 addr
= lp_build_mul_imm(&bld_base
->uint_bld
, addr
, 16);
1903 addr
= lp_build_add(&bld_base
->uint_bld
, addr
,
1904 lp_build_const_int32(base
->gallivm
, idx
* 4));
1906 result
= buffer_load_const(ctx
, bufp
, addr
);
1908 if (!tgsi_type_is_64bit(type
))
1909 result
= bitcast(bld_base
, type
, result
);
1911 LLVMValueRef addr2
, result2
;
1912 addr2
= ctx
->radeon_bld
.soa
.addr
[ireg
->Index
][ireg
->Swizzle
+ 1];
1913 addr2
= LLVMBuildLoad(base
->gallivm
->builder
, addr2
, "load addr reg2");
1914 addr2
= lp_build_mul_imm(&bld_base
->uint_bld
, addr2
, 16);
1915 addr2
= lp_build_add(&bld_base
->uint_bld
, addr2
,
1916 lp_build_const_int32(base
->gallivm
, idx
* 4));
1918 result2
= buffer_load_const(ctx
, bufp
, addr2
);
1920 result
= radeon_llvm_emit_fetch_64bit(bld_base
, type
,
1926 /* Upper 16 bits must be zero. */
1927 static LLVMValueRef
si_llvm_pack_two_int16(struct gallivm_state
*gallivm
,
1928 LLVMValueRef val
[2])
1930 return LLVMBuildOr(gallivm
->builder
, val
[0],
1931 LLVMBuildShl(gallivm
->builder
, val
[1],
1932 lp_build_const_int32(gallivm
, 16),
1936 /* Upper 16 bits are ignored and will be dropped. */
1937 static LLVMValueRef
si_llvm_pack_two_int32_as_int16(struct gallivm_state
*gallivm
,
1938 LLVMValueRef val
[2])
1940 LLVMValueRef v
[2] = {
1941 LLVMBuildAnd(gallivm
->builder
, val
[0],
1942 lp_build_const_int32(gallivm
, 0xffff), ""),
1945 return si_llvm_pack_two_int16(gallivm
, v
);
1948 /* Initialize arguments for the shader export intrinsic */
1949 static void si_llvm_init_export_args(struct lp_build_tgsi_context
*bld_base
,
1950 LLVMValueRef
*values
,
1954 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1955 struct lp_build_context
*uint
=
1956 &ctx
->radeon_bld
.soa
.bld_base
.uint_bld
;
1957 struct lp_build_context
*base
= &bld_base
->base
;
1958 struct gallivm_state
*gallivm
= base
->gallivm
;
1959 LLVMBuilderRef builder
= base
->gallivm
->builder
;
1960 LLVMValueRef val
[4];
1961 unsigned spi_shader_col_format
= V_028714_SPI_SHADER_32_ABGR
;
1965 /* Default is 0xf. Adjusted below depending on the format. */
1966 args
[0] = lp_build_const_int32(base
->gallivm
, 0xf); /* writemask */
1968 /* Specify whether the EXEC mask represents the valid mask */
1969 args
[1] = uint
->zero
;
1971 /* Specify whether this is the last export */
1972 args
[2] = uint
->zero
;
1974 /* Specify the target we are exporting */
1975 args
[3] = lp_build_const_int32(base
->gallivm
, target
);
1977 if (ctx
->type
== PIPE_SHADER_FRAGMENT
) {
1978 const union si_shader_key
*key
= &ctx
->shader
->key
;
1979 unsigned col_formats
= key
->ps
.epilog
.spi_shader_col_format
;
1980 int cbuf
= target
- V_008DFC_SQ_EXP_MRT
;
1982 assert(cbuf
>= 0 && cbuf
< 8);
1983 spi_shader_col_format
= (col_formats
>> (cbuf
* 4)) & 0xf;
1984 is_int8
= (key
->ps
.epilog
.color_is_int8
>> cbuf
) & 0x1;
1987 args
[4] = uint
->zero
; /* COMPR flag */
1988 args
[5] = base
->undef
;
1989 args
[6] = base
->undef
;
1990 args
[7] = base
->undef
;
1991 args
[8] = base
->undef
;
1993 switch (spi_shader_col_format
) {
1994 case V_028714_SPI_SHADER_ZERO
:
1995 args
[0] = uint
->zero
; /* writemask */
1996 args
[3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_NULL
);
1999 case V_028714_SPI_SHADER_32_R
:
2000 args
[0] = uint
->one
; /* writemask */
2001 args
[5] = values
[0];
2004 case V_028714_SPI_SHADER_32_GR
:
2005 args
[0] = lp_build_const_int32(base
->gallivm
, 0x3); /* writemask */
2006 args
[5] = values
[0];
2007 args
[6] = values
[1];
2010 case V_028714_SPI_SHADER_32_AR
:
2011 args
[0] = lp_build_const_int32(base
->gallivm
, 0x9); /* writemask */
2012 args
[5] = values
[0];
2013 args
[8] = values
[3];
2016 case V_028714_SPI_SHADER_FP16_ABGR
:
2017 args
[4] = uint
->one
; /* COMPR flag */
2019 for (chan
= 0; chan
< 2; chan
++) {
2020 LLVMValueRef pack_args
[2] = {
2022 values
[2 * chan
+ 1]
2024 LLVMValueRef packed
;
2026 packed
= lp_build_intrinsic(base
->gallivm
->builder
,
2028 ctx
->i32
, pack_args
, 2,
2029 LLVMReadNoneAttribute
);
2031 LLVMBuildBitCast(base
->gallivm
->builder
,
2032 packed
, ctx
->f32
, "");
2036 case V_028714_SPI_SHADER_UNORM16_ABGR
:
2037 for (chan
= 0; chan
< 4; chan
++) {
2038 val
[chan
] = radeon_llvm_saturate(bld_base
, values
[chan
]);
2039 val
[chan
] = LLVMBuildFMul(builder
, val
[chan
],
2040 lp_build_const_float(gallivm
, 65535), "");
2041 val
[chan
] = LLVMBuildFAdd(builder
, val
[chan
],
2042 lp_build_const_float(gallivm
, 0.5), "");
2043 val
[chan
] = LLVMBuildFPToUI(builder
, val
[chan
],
2047 args
[4] = uint
->one
; /* COMPR flag */
2048 args
[5] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
2049 si_llvm_pack_two_int16(gallivm
, val
));
2050 args
[6] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
2051 si_llvm_pack_two_int16(gallivm
, val
+2));
2054 case V_028714_SPI_SHADER_SNORM16_ABGR
:
2055 for (chan
= 0; chan
< 4; chan
++) {
2056 /* Clamp between [-1, 1]. */
2057 val
[chan
] = lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_MIN
,
2059 lp_build_const_float(gallivm
, 1));
2060 val
[chan
] = lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_MAX
,
2062 lp_build_const_float(gallivm
, -1));
2063 /* Convert to a signed integer in [-32767, 32767]. */
2064 val
[chan
] = LLVMBuildFMul(builder
, val
[chan
],
2065 lp_build_const_float(gallivm
, 32767), "");
2066 /* If positive, add 0.5, else add -0.5. */
2067 val
[chan
] = LLVMBuildFAdd(builder
, val
[chan
],
2068 LLVMBuildSelect(builder
,
2069 LLVMBuildFCmp(builder
, LLVMRealOGE
,
2070 val
[chan
], base
->zero
, ""),
2071 lp_build_const_float(gallivm
, 0.5),
2072 lp_build_const_float(gallivm
, -0.5), ""), "");
2073 val
[chan
] = LLVMBuildFPToSI(builder
, val
[chan
], ctx
->i32
, "");
2076 args
[4] = uint
->one
; /* COMPR flag */
2077 args
[5] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
2078 si_llvm_pack_two_int32_as_int16(gallivm
, val
));
2079 args
[6] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
2080 si_llvm_pack_two_int32_as_int16(gallivm
, val
+2));
2083 case V_028714_SPI_SHADER_UINT16_ABGR
: {
2084 LLVMValueRef max
= lp_build_const_int32(gallivm
, is_int8
?
2087 for (chan
= 0; chan
< 4; chan
++) {
2088 val
[chan
] = bitcast(bld_base
, TGSI_TYPE_UNSIGNED
, values
[chan
]);
2089 val
[chan
] = lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_UMIN
,
2093 args
[4] = uint
->one
; /* COMPR flag */
2094 args
[5] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
2095 si_llvm_pack_two_int16(gallivm
, val
));
2096 args
[6] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
2097 si_llvm_pack_two_int16(gallivm
, val
+2));
2101 case V_028714_SPI_SHADER_SINT16_ABGR
: {
2102 LLVMValueRef max
= lp_build_const_int32(gallivm
, is_int8
?
2104 LLVMValueRef min
= lp_build_const_int32(gallivm
, is_int8
?
2107 for (chan
= 0; chan
< 4; chan
++) {
2108 val
[chan
] = bitcast(bld_base
, TGSI_TYPE_UNSIGNED
, values
[chan
]);
2109 val
[chan
] = lp_build_emit_llvm_binary(bld_base
,
2112 val
[chan
] = lp_build_emit_llvm_binary(bld_base
,
2117 args
[4] = uint
->one
; /* COMPR flag */
2118 args
[5] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
2119 si_llvm_pack_two_int32_as_int16(gallivm
, val
));
2120 args
[6] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
2121 si_llvm_pack_two_int32_as_int16(gallivm
, val
+2));
2125 case V_028714_SPI_SHADER_32_ABGR
:
2126 memcpy(&args
[5], values
, sizeof(values
[0]) * 4);
2131 static void si_alpha_test(struct lp_build_tgsi_context
*bld_base
,
2134 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2135 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2137 if (ctx
->shader
->key
.ps
.epilog
.alpha_func
!= PIPE_FUNC_NEVER
) {
2138 LLVMValueRef alpha_ref
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
2139 SI_PARAM_ALPHA_REF
);
2141 LLVMValueRef alpha_pass
=
2142 lp_build_cmp(&bld_base
->base
,
2143 ctx
->shader
->key
.ps
.epilog
.alpha_func
,
2146 lp_build_select(&bld_base
->base
,
2148 lp_build_const_float(gallivm
, 1.0f
),
2149 lp_build_const_float(gallivm
, -1.0f
));
2151 lp_build_intrinsic(gallivm
->builder
, "llvm.AMDGPU.kill",
2152 ctx
->voidt
, &arg
, 1, 0);
2154 lp_build_intrinsic(gallivm
->builder
, "llvm.AMDGPU.kilp",
2155 ctx
->voidt
, NULL
, 0, 0);
2159 static LLVMValueRef
si_scale_alpha_by_sample_mask(struct lp_build_tgsi_context
*bld_base
,
2161 unsigned samplemask_param
)
2163 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2164 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2165 LLVMValueRef coverage
;
2167 /* alpha = alpha * popcount(coverage) / SI_NUM_SMOOTH_AA_SAMPLES */
2168 coverage
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
2170 coverage
= bitcast(bld_base
, TGSI_TYPE_SIGNED
, coverage
);
2172 coverage
= lp_build_intrinsic(gallivm
->builder
, "llvm.ctpop.i32",
2174 &coverage
, 1, LLVMReadNoneAttribute
);
2176 coverage
= LLVMBuildUIToFP(gallivm
->builder
, coverage
,
2179 coverage
= LLVMBuildFMul(gallivm
->builder
, coverage
,
2180 lp_build_const_float(gallivm
,
2181 1.0 / SI_NUM_SMOOTH_AA_SAMPLES
), "");
2183 return LLVMBuildFMul(gallivm
->builder
, alpha
, coverage
, "");
2186 static void si_llvm_emit_clipvertex(struct lp_build_tgsi_context
*bld_base
,
2187 LLVMValueRef (*pos
)[9], LLVMValueRef
*out_elts
)
2189 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2190 struct lp_build_context
*base
= &bld_base
->base
;
2191 struct lp_build_context
*uint
= &ctx
->radeon_bld
.soa
.bld_base
.uint_bld
;
2194 unsigned const_chan
;
2195 LLVMValueRef base_elt
;
2196 LLVMValueRef ptr
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_RW_BUFFERS
);
2197 LLVMValueRef constbuf_index
= lp_build_const_int32(base
->gallivm
,
2198 SI_VS_CONST_CLIP_PLANES
);
2199 LLVMValueRef const_resource
= build_indexed_load_const(ctx
, ptr
, constbuf_index
);
2201 for (reg_index
= 0; reg_index
< 2; reg_index
++) {
2202 LLVMValueRef
*args
= pos
[2 + reg_index
];
2207 args
[8] = lp_build_const_float(base
->gallivm
, 0.0f
);
2209 /* Compute dot products of position and user clip plane vectors */
2210 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
2211 for (const_chan
= 0; const_chan
< TGSI_NUM_CHANNELS
; const_chan
++) {
2212 args
[1] = lp_build_const_int32(base
->gallivm
,
2213 ((reg_index
* 4 + chan
) * 4 +
2215 base_elt
= buffer_load_const(ctx
, const_resource
,
2218 lp_build_add(base
, args
[5 + chan
],
2219 lp_build_mul(base
, base_elt
,
2220 out_elts
[const_chan
]));
2224 args
[0] = lp_build_const_int32(base
->gallivm
, 0xf);
2225 args
[1] = uint
->zero
;
2226 args
[2] = uint
->zero
;
2227 args
[3] = lp_build_const_int32(base
->gallivm
,
2228 V_008DFC_SQ_EXP_POS
+ 2 + reg_index
);
2229 args
[4] = uint
->zero
;
2233 static void si_dump_streamout(struct pipe_stream_output_info
*so
)
2237 if (so
->num_outputs
)
2238 fprintf(stderr
, "STREAMOUT\n");
2240 for (i
= 0; i
< so
->num_outputs
; i
++) {
2241 unsigned mask
= ((1 << so
->output
[i
].num_components
) - 1) <<
2242 so
->output
[i
].start_component
;
2243 fprintf(stderr
, " %i: BUF%i[%i..%i] <- OUT[%i].%s%s%s%s\n",
2244 i
, so
->output
[i
].output_buffer
,
2245 so
->output
[i
].dst_offset
, so
->output
[i
].dst_offset
+ so
->output
[i
].num_components
- 1,
2246 so
->output
[i
].register_index
,
2247 mask
& 1 ? "x" : "",
2248 mask
& 2 ? "y" : "",
2249 mask
& 4 ? "z" : "",
2250 mask
& 8 ? "w" : "");
2254 /* On SI, the vertex shader is responsible for writing streamout data
2256 static void si_llvm_emit_streamout(struct si_shader_context
*ctx
,
2257 struct si_shader_output_values
*outputs
,
2260 struct pipe_stream_output_info
*so
= &ctx
->shader
->selector
->so
;
2261 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
2262 LLVMBuilderRef builder
= gallivm
->builder
;
2264 struct lp_build_if_state if_ctx
;
2265 LLVMValueRef so_buffers
[4];
2266 LLVMValueRef buf_ptr
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
2267 SI_PARAM_RW_BUFFERS
);
2269 /* Load the descriptors. */
2270 for (i
= 0; i
< 4; ++i
) {
2271 if (ctx
->shader
->selector
->so
.stride
[i
]) {
2272 LLVMValueRef offset
= lp_build_const_int32(gallivm
,
2273 SI_VS_STREAMOUT_BUF0
+ i
);
2275 so_buffers
[i
] = build_indexed_load_const(ctx
, buf_ptr
, offset
);
2279 /* Get bits [22:16], i.e. (so_param >> 16) & 127; */
2280 LLVMValueRef so_vtx_count
=
2281 unpack_param(ctx
, ctx
->param_streamout_config
, 16, 7);
2283 LLVMValueRef tid
= get_thread_id(ctx
);
2285 /* can_emit = tid < so_vtx_count; */
2286 LLVMValueRef can_emit
=
2287 LLVMBuildICmp(builder
, LLVMIntULT
, tid
, so_vtx_count
, "");
2289 LLVMValueRef stream_id
=
2290 unpack_param(ctx
, ctx
->param_streamout_config
, 24, 2);
2292 /* Emit the streamout code conditionally. This actually avoids
2293 * out-of-bounds buffer access. The hw tells us via the SGPR
2294 * (so_vtx_count) which threads are allowed to emit streamout data. */
2295 lp_build_if(&if_ctx
, gallivm
, can_emit
);
2297 /* The buffer offset is computed as follows:
2298 * ByteOffset = streamout_offset[buffer_id]*4 +
2299 * (streamout_write_index + thread_id)*stride[buffer_id] +
2303 LLVMValueRef so_write_index
=
2304 LLVMGetParam(ctx
->radeon_bld
.main_fn
,
2305 ctx
->param_streamout_write_index
);
2307 /* Compute (streamout_write_index + thread_id). */
2308 so_write_index
= LLVMBuildAdd(builder
, so_write_index
, tid
, "");
2310 /* Compute the write offset for each enabled buffer. */
2311 LLVMValueRef so_write_offset
[4] = {};
2312 for (i
= 0; i
< 4; i
++) {
2316 LLVMValueRef so_offset
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
2317 ctx
->param_streamout_offset
[i
]);
2318 so_offset
= LLVMBuildMul(builder
, so_offset
, LLVMConstInt(ctx
->i32
, 4, 0), "");
2320 so_write_offset
[i
] = LLVMBuildMul(builder
, so_write_index
,
2321 LLVMConstInt(ctx
->i32
, so
->stride
[i
]*4, 0), "");
2322 so_write_offset
[i
] = LLVMBuildAdd(builder
, so_write_offset
[i
], so_offset
, "");
2325 /* Write streamout data. */
2326 for (i
= 0; i
< so
->num_outputs
; i
++) {
2327 unsigned buf_idx
= so
->output
[i
].output_buffer
;
2328 unsigned reg
= so
->output
[i
].register_index
;
2329 unsigned start
= so
->output
[i
].start_component
;
2330 unsigned num_comps
= so
->output
[i
].num_components
;
2331 unsigned stream
= so
->output
[i
].stream
;
2332 LLVMValueRef out
[4];
2333 struct lp_build_if_state if_ctx_stream
;
2335 assert(num_comps
&& num_comps
<= 4);
2336 if (!num_comps
|| num_comps
> 4)
2342 /* Load the output as int. */
2343 for (j
= 0; j
< num_comps
; j
++) {
2344 out
[j
] = LLVMBuildBitCast(builder
,
2345 outputs
[reg
].values
[start
+j
],
2349 /* Pack the output. */
2350 LLVMValueRef vdata
= NULL
;
2352 switch (num_comps
) {
2353 case 1: /* as i32 */
2356 case 2: /* as v2i32 */
2357 case 3: /* as v4i32 (aligned to 4) */
2358 case 4: /* as v4i32 */
2359 vdata
= LLVMGetUndef(LLVMVectorType(ctx
->i32
, util_next_power_of_two(num_comps
)));
2360 for (j
= 0; j
< num_comps
; j
++) {
2361 vdata
= LLVMBuildInsertElement(builder
, vdata
, out
[j
],
2362 LLVMConstInt(ctx
->i32
, j
, 0), "");
2367 LLVMValueRef can_emit_stream
=
2368 LLVMBuildICmp(builder
, LLVMIntEQ
,
2370 lp_build_const_int32(gallivm
, stream
), "");
2372 lp_build_if(&if_ctx_stream
, gallivm
, can_emit_stream
);
2373 build_tbuffer_store_dwords(ctx
, so_buffers
[buf_idx
],
2375 so_write_offset
[buf_idx
],
2376 LLVMConstInt(ctx
->i32
, 0, 0),
2377 so
->output
[i
].dst_offset
*4);
2378 lp_build_endif(&if_ctx_stream
);
2381 lp_build_endif(&if_ctx
);
2385 /* Generate export instructions for hardware VS shader stage */
2386 static void si_llvm_export_vs(struct lp_build_tgsi_context
*bld_base
,
2387 struct si_shader_output_values
*outputs
,
2390 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2391 struct si_shader
*shader
= ctx
->shader
;
2392 struct lp_build_context
*base
= &bld_base
->base
;
2393 struct lp_build_context
*uint
=
2394 &ctx
->radeon_bld
.soa
.bld_base
.uint_bld
;
2395 LLVMValueRef args
[9];
2396 LLVMValueRef pos_args
[4][9] = { { 0 } };
2397 LLVMValueRef psize_value
= NULL
, edgeflag_value
= NULL
, layer_value
= NULL
, viewport_index_value
= NULL
;
2398 unsigned semantic_name
, semantic_index
;
2400 unsigned param_count
= 0;
2404 if (outputs
&& ctx
->shader
->selector
->so
.num_outputs
) {
2405 si_llvm_emit_streamout(ctx
, outputs
, noutput
);
2408 for (i
= 0; i
< noutput
; i
++) {
2409 semantic_name
= outputs
[i
].name
;
2410 semantic_index
= outputs
[i
].sid
;
2413 /* Select the correct target */
2414 switch(semantic_name
) {
2415 case TGSI_SEMANTIC_PSIZE
:
2416 psize_value
= outputs
[i
].values
[0];
2418 case TGSI_SEMANTIC_EDGEFLAG
:
2419 edgeflag_value
= outputs
[i
].values
[0];
2421 case TGSI_SEMANTIC_LAYER
:
2422 layer_value
= outputs
[i
].values
[0];
2423 semantic_name
= TGSI_SEMANTIC_GENERIC
;
2424 goto handle_semantic
;
2425 case TGSI_SEMANTIC_VIEWPORT_INDEX
:
2426 viewport_index_value
= outputs
[i
].values
[0];
2427 semantic_name
= TGSI_SEMANTIC_GENERIC
;
2428 goto handle_semantic
;
2429 case TGSI_SEMANTIC_POSITION
:
2430 target
= V_008DFC_SQ_EXP_POS
;
2432 case TGSI_SEMANTIC_COLOR
:
2433 case TGSI_SEMANTIC_BCOLOR
:
2434 target
= V_008DFC_SQ_EXP_PARAM
+ param_count
;
2435 assert(i
< ARRAY_SIZE(shader
->info
.vs_output_param_offset
));
2436 shader
->info
.vs_output_param_offset
[i
] = param_count
;
2439 case TGSI_SEMANTIC_CLIPDIST
:
2440 target
= V_008DFC_SQ_EXP_POS
+ 2 + semantic_index
;
2442 case TGSI_SEMANTIC_CLIPVERTEX
:
2443 si_llvm_emit_clipvertex(bld_base
, pos_args
, outputs
[i
].values
);
2445 case TGSI_SEMANTIC_PRIMID
:
2446 case TGSI_SEMANTIC_FOG
:
2447 case TGSI_SEMANTIC_TEXCOORD
:
2448 case TGSI_SEMANTIC_GENERIC
:
2449 target
= V_008DFC_SQ_EXP_PARAM
+ param_count
;
2450 assert(i
< ARRAY_SIZE(shader
->info
.vs_output_param_offset
));
2451 shader
->info
.vs_output_param_offset
[i
] = param_count
;
2457 "Warning: SI unhandled vs output type:%d\n",
2461 si_llvm_init_export_args(bld_base
, outputs
[i
].values
, target
, args
);
2463 if (target
>= V_008DFC_SQ_EXP_POS
&&
2464 target
<= (V_008DFC_SQ_EXP_POS
+ 3)) {
2465 memcpy(pos_args
[target
- V_008DFC_SQ_EXP_POS
],
2466 args
, sizeof(args
));
2468 lp_build_intrinsic(base
->gallivm
->builder
,
2469 "llvm.SI.export", ctx
->voidt
,
2473 if (semantic_name
== TGSI_SEMANTIC_CLIPDIST
) {
2474 semantic_name
= TGSI_SEMANTIC_GENERIC
;
2475 goto handle_semantic
;
2479 shader
->info
.nr_param_exports
= param_count
;
2481 /* We need to add the position output manually if it's missing. */
2482 if (!pos_args
[0][0]) {
2483 pos_args
[0][0] = lp_build_const_int32(base
->gallivm
, 0xf); /* writemask */
2484 pos_args
[0][1] = uint
->zero
; /* EXEC mask */
2485 pos_args
[0][2] = uint
->zero
; /* last export? */
2486 pos_args
[0][3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_POS
);
2487 pos_args
[0][4] = uint
->zero
; /* COMPR flag */
2488 pos_args
[0][5] = base
->zero
; /* X */
2489 pos_args
[0][6] = base
->zero
; /* Y */
2490 pos_args
[0][7] = base
->zero
; /* Z */
2491 pos_args
[0][8] = base
->one
; /* W */
2494 /* Write the misc vector (point size, edgeflag, layer, viewport). */
2495 if (shader
->selector
->info
.writes_psize
||
2496 shader
->selector
->info
.writes_edgeflag
||
2497 shader
->selector
->info
.writes_viewport_index
||
2498 shader
->selector
->info
.writes_layer
) {
2499 pos_args
[1][0] = lp_build_const_int32(base
->gallivm
, /* writemask */
2500 shader
->selector
->info
.writes_psize
|
2501 (shader
->selector
->info
.writes_edgeflag
<< 1) |
2502 (shader
->selector
->info
.writes_layer
<< 2) |
2503 (shader
->selector
->info
.writes_viewport_index
<< 3));
2504 pos_args
[1][1] = uint
->zero
; /* EXEC mask */
2505 pos_args
[1][2] = uint
->zero
; /* last export? */
2506 pos_args
[1][3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_POS
+ 1);
2507 pos_args
[1][4] = uint
->zero
; /* COMPR flag */
2508 pos_args
[1][5] = base
->zero
; /* X */
2509 pos_args
[1][6] = base
->zero
; /* Y */
2510 pos_args
[1][7] = base
->zero
; /* Z */
2511 pos_args
[1][8] = base
->zero
; /* W */
2513 if (shader
->selector
->info
.writes_psize
)
2514 pos_args
[1][5] = psize_value
;
2516 if (shader
->selector
->info
.writes_edgeflag
) {
2517 /* The output is a float, but the hw expects an integer
2518 * with the first bit containing the edge flag. */
2519 edgeflag_value
= LLVMBuildFPToUI(base
->gallivm
->builder
,
2522 edgeflag_value
= lp_build_min(&bld_base
->int_bld
,
2524 bld_base
->int_bld
.one
);
2526 /* The LLVM intrinsic expects a float. */
2527 pos_args
[1][6] = LLVMBuildBitCast(base
->gallivm
->builder
,
2532 if (shader
->selector
->info
.writes_layer
)
2533 pos_args
[1][7] = layer_value
;
2535 if (shader
->selector
->info
.writes_viewport_index
)
2536 pos_args
[1][8] = viewport_index_value
;
2539 for (i
= 0; i
< 4; i
++)
2541 shader
->info
.nr_pos_exports
++;
2544 for (i
= 0; i
< 4; i
++) {
2545 if (!pos_args
[i
][0])
2548 /* Specify the target we are exporting */
2549 pos_args
[i
][3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_POS
+ pos_idx
++);
2551 if (pos_idx
== shader
->info
.nr_pos_exports
)
2552 /* Specify that this is the last export */
2553 pos_args
[i
][2] = uint
->one
;
2555 lp_build_intrinsic(base
->gallivm
->builder
, "llvm.SI.export",
2556 ctx
->voidt
, pos_args
[i
], 9, 0);
2560 static void si_copy_tcs_inputs(struct lp_build_tgsi_context
*bld_base
)
2562 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2563 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2564 LLVMValueRef invocation_id
, rw_buffers
, buffer
, buffer_offset
;
2565 LLVMValueRef lds_vertex_stride
, lds_vertex_offset
, lds_base
;
2568 invocation_id
= unpack_param(ctx
, SI_PARAM_REL_IDS
, 8, 5);
2570 rw_buffers
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_RW_BUFFERS
);
2571 buffer
= build_indexed_load_const(ctx
, rw_buffers
,
2572 lp_build_const_int32(gallivm
, SI_HS_RING_TESS_OFFCHIP
));
2574 buffer_offset
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, ctx
->param_oc_lds
);
2576 lds_vertex_stride
= unpack_param(ctx
, SI_PARAM_TCS_IN_LAYOUT
, 13, 8);
2577 lds_vertex_offset
= LLVMBuildMul(gallivm
->builder
, invocation_id
,
2578 lds_vertex_stride
, "");
2579 lds_base
= get_tcs_in_current_patch_offset(ctx
);
2580 lds_base
= LLVMBuildAdd(gallivm
->builder
, lds_base
, lds_vertex_offset
, "");
2582 inputs
= ctx
->shader
->key
.tcs
.epilog
.inputs_to_copy
;
2584 unsigned i
= u_bit_scan64(&inputs
);
2586 LLVMValueRef lds_ptr
= LLVMBuildAdd(gallivm
->builder
, lds_base
,
2587 lp_build_const_int32(gallivm
, 4 * i
),
2590 LLVMValueRef buffer_addr
= get_tcs_tes_buffer_address(ctx
,
2592 lp_build_const_int32(gallivm
, i
));
2594 LLVMValueRef value
= lds_load(bld_base
, TGSI_TYPE_SIGNED
, ~0,
2597 build_tbuffer_store_dwords(ctx
, buffer
, value
, 4, buffer_addr
,
2602 static void si_write_tess_factors(struct lp_build_tgsi_context
*bld_base
,
2603 LLVMValueRef rel_patch_id
,
2604 LLVMValueRef invocation_id
,
2605 LLVMValueRef tcs_out_current_patch_data_offset
)
2607 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2608 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2609 struct si_shader
*shader
= ctx
->shader
;
2610 unsigned tess_inner_index
, tess_outer_index
;
2611 LLVMValueRef lds_base
, lds_inner
, lds_outer
, byteoffset
, buffer
;
2612 LLVMValueRef out
[6], vec0
, vec1
, rw_buffers
, tf_base
;
2613 unsigned stride
, outer_comps
, inner_comps
, i
;
2614 struct lp_build_if_state if_ctx
, inner_if_ctx
;
2616 si_llvm_emit_barrier(NULL
, bld_base
, NULL
);
2618 /* Do this only for invocation 0, because the tess levels are per-patch,
2621 * This can't jump, because invocation 0 executes this. It should
2622 * at least mask out the loads and stores for other invocations.
2624 lp_build_if(&if_ctx
, gallivm
,
2625 LLVMBuildICmp(gallivm
->builder
, LLVMIntEQ
,
2626 invocation_id
, bld_base
->uint_bld
.zero
, ""));
2628 /* Determine the layout of one tess factor element in the buffer. */
2629 switch (shader
->key
.tcs
.epilog
.prim_mode
) {
2630 case PIPE_PRIM_LINES
:
2631 stride
= 2; /* 2 dwords, 1 vec2 store */
2635 case PIPE_PRIM_TRIANGLES
:
2636 stride
= 4; /* 4 dwords, 1 vec4 store */
2640 case PIPE_PRIM_QUADS
:
2641 stride
= 6; /* 6 dwords, 2 stores (vec4 + vec2) */
2650 /* Load tess_inner and tess_outer from LDS.
2651 * Any invocation can write them, so we can't get them from a temporary.
2653 tess_inner_index
= si_shader_io_get_unique_index(TGSI_SEMANTIC_TESSINNER
, 0);
2654 tess_outer_index
= si_shader_io_get_unique_index(TGSI_SEMANTIC_TESSOUTER
, 0);
2656 lds_base
= tcs_out_current_patch_data_offset
;
2657 lds_inner
= LLVMBuildAdd(gallivm
->builder
, lds_base
,
2658 lp_build_const_int32(gallivm
,
2659 tess_inner_index
* 4), "");
2660 lds_outer
= LLVMBuildAdd(gallivm
->builder
, lds_base
,
2661 lp_build_const_int32(gallivm
,
2662 tess_outer_index
* 4), "");
2664 for (i
= 0; i
< outer_comps
; i
++)
2665 out
[i
] = lds_load(bld_base
, TGSI_TYPE_SIGNED
, i
, lds_outer
);
2666 for (i
= 0; i
< inner_comps
; i
++)
2667 out
[outer_comps
+i
] = lds_load(bld_base
, TGSI_TYPE_SIGNED
, i
, lds_inner
);
2669 /* Convert the outputs to vectors for stores. */
2670 vec0
= lp_build_gather_values(gallivm
, out
, MIN2(stride
, 4));
2674 vec1
= lp_build_gather_values(gallivm
, out
+4, stride
- 4);
2676 /* Get the buffer. */
2677 rw_buffers
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
2678 SI_PARAM_RW_BUFFERS
);
2679 buffer
= build_indexed_load_const(ctx
, rw_buffers
,
2680 lp_build_const_int32(gallivm
, SI_HS_RING_TESS_FACTOR
));
2682 /* Get the offset. */
2683 tf_base
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
2684 SI_PARAM_TESS_FACTOR_OFFSET
);
2685 byteoffset
= LLVMBuildMul(gallivm
->builder
, rel_patch_id
,
2686 lp_build_const_int32(gallivm
, 4 * stride
), "");
2688 lp_build_if(&inner_if_ctx
, gallivm
,
2689 LLVMBuildICmp(gallivm
->builder
, LLVMIntEQ
,
2690 rel_patch_id
, bld_base
->uint_bld
.zero
, ""));
2692 /* Store the dynamic HS control word. */
2693 build_tbuffer_store_dwords(ctx
, buffer
,
2694 lp_build_const_int32(gallivm
, 0x80000000),
2695 1, lp_build_const_int32(gallivm
, 0), tf_base
, 0);
2697 lp_build_endif(&inner_if_ctx
);
2699 /* Store the tessellation factors. */
2700 build_tbuffer_store_dwords(ctx
, buffer
, vec0
,
2701 MIN2(stride
, 4), byteoffset
, tf_base
, 4);
2703 build_tbuffer_store_dwords(ctx
, buffer
, vec1
,
2704 stride
- 4, byteoffset
, tf_base
, 20);
2705 lp_build_endif(&if_ctx
);
2708 /* This only writes the tessellation factor levels. */
2709 static void si_llvm_emit_tcs_epilogue(struct lp_build_tgsi_context
*bld_base
)
2711 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2712 LLVMValueRef rel_patch_id
, invocation_id
, tf_lds_offset
;
2714 rel_patch_id
= get_rel_patch_id(ctx
);
2715 invocation_id
= unpack_param(ctx
, SI_PARAM_REL_IDS
, 8, 5);
2716 tf_lds_offset
= get_tcs_out_current_patch_data_offset(ctx
);
2718 if (!ctx
->is_monolithic
) {
2719 /* Return epilog parameters from this function. */
2720 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
2721 LLVMValueRef ret
= ctx
->return_value
;
2722 LLVMValueRef rw_buffers
, rw0
, rw1
, tf_soffset
;
2725 /* RW_BUFFERS pointer */
2726 rw_buffers
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
2727 SI_PARAM_RW_BUFFERS
);
2728 rw_buffers
= LLVMBuildPtrToInt(builder
, rw_buffers
, ctx
->i64
, "");
2729 rw_buffers
= LLVMBuildBitCast(builder
, rw_buffers
, ctx
->v2i32
, "");
2730 rw0
= LLVMBuildExtractElement(builder
, rw_buffers
,
2731 bld_base
->uint_bld
.zero
, "");
2732 rw1
= LLVMBuildExtractElement(builder
, rw_buffers
,
2733 bld_base
->uint_bld
.one
, "");
2734 ret
= LLVMBuildInsertValue(builder
, ret
, rw0
, 0, "");
2735 ret
= LLVMBuildInsertValue(builder
, ret
, rw1
, 1, "");
2737 /* Tess factor buffer soffset is after user SGPRs. */
2738 tf_soffset
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
2739 SI_PARAM_TESS_FACTOR_OFFSET
);
2740 ret
= LLVMBuildInsertValue(builder
, ret
, tf_soffset
,
2741 SI_TCS_NUM_USER_SGPR
+ 1, "");
2744 rel_patch_id
= bitcast(bld_base
, TGSI_TYPE_FLOAT
, rel_patch_id
);
2745 invocation_id
= bitcast(bld_base
, TGSI_TYPE_FLOAT
, invocation_id
);
2746 tf_lds_offset
= bitcast(bld_base
, TGSI_TYPE_FLOAT
, tf_lds_offset
);
2748 vgpr
= SI_TCS_NUM_USER_SGPR
+ 2;
2749 ret
= LLVMBuildInsertValue(builder
, ret
, rel_patch_id
, vgpr
++, "");
2750 ret
= LLVMBuildInsertValue(builder
, ret
, invocation_id
, vgpr
++, "");
2751 ret
= LLVMBuildInsertValue(builder
, ret
, tf_lds_offset
, vgpr
++, "");
2752 ctx
->return_value
= ret
;
2756 si_copy_tcs_inputs(bld_base
);
2757 si_write_tess_factors(bld_base
, rel_patch_id
, invocation_id
, tf_lds_offset
);
2760 static void si_llvm_emit_ls_epilogue(struct lp_build_tgsi_context
*bld_base
)
2762 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2763 struct si_shader
*shader
= ctx
->shader
;
2764 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
2765 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2767 LLVMValueRef vertex_id
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
2768 ctx
->param_rel_auto_id
);
2769 LLVMValueRef vertex_dw_stride
=
2770 unpack_param(ctx
, SI_PARAM_LS_OUT_LAYOUT
, 13, 8);
2771 LLVMValueRef base_dw_addr
= LLVMBuildMul(gallivm
->builder
, vertex_id
,
2772 vertex_dw_stride
, "");
2774 /* Write outputs to LDS. The next shader (TCS aka HS) will read
2775 * its inputs from it. */
2776 for (i
= 0; i
< info
->num_outputs
; i
++) {
2777 LLVMValueRef
*out_ptr
= ctx
->radeon_bld
.soa
.outputs
[i
];
2778 unsigned name
= info
->output_semantic_name
[i
];
2779 unsigned index
= info
->output_semantic_index
[i
];
2780 int param
= si_shader_io_get_unique_index(name
, index
);
2781 LLVMValueRef dw_addr
= LLVMBuildAdd(gallivm
->builder
, base_dw_addr
,
2782 lp_build_const_int32(gallivm
, param
* 4), "");
2784 for (chan
= 0; chan
< 4; chan
++) {
2785 lds_store(bld_base
, chan
, dw_addr
,
2786 LLVMBuildLoad(gallivm
->builder
, out_ptr
[chan
], ""));
2791 static void si_llvm_emit_es_epilogue(struct lp_build_tgsi_context
*bld_base
)
2793 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2794 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2795 struct si_shader
*es
= ctx
->shader
;
2796 struct tgsi_shader_info
*info
= &es
->selector
->info
;
2797 LLVMValueRef soffset
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
2798 ctx
->param_es2gs_offset
);
2802 for (i
= 0; i
< info
->num_outputs
; i
++) {
2803 LLVMValueRef
*out_ptr
=
2804 ctx
->radeon_bld
.soa
.outputs
[i
];
2807 if (info
->output_semantic_name
[i
] == TGSI_SEMANTIC_VIEWPORT_INDEX
||
2808 info
->output_semantic_name
[i
] == TGSI_SEMANTIC_LAYER
)
2811 param_index
= si_shader_io_get_unique_index(info
->output_semantic_name
[i
],
2812 info
->output_semantic_index
[i
]);
2814 for (chan
= 0; chan
< 4; chan
++) {
2815 LLVMValueRef out_val
= LLVMBuildLoad(gallivm
->builder
, out_ptr
[chan
], "");
2816 out_val
= LLVMBuildBitCast(gallivm
->builder
, out_val
, ctx
->i32
, "");
2818 build_tbuffer_store(ctx
,
2821 LLVMGetUndef(ctx
->i32
), soffset
,
2822 (4 * param_index
+ chan
) * 4,
2823 V_008F0C_BUF_DATA_FORMAT_32
,
2824 V_008F0C_BUF_NUM_FORMAT_UINT
,
2830 static void si_llvm_emit_gs_epilogue(struct lp_build_tgsi_context
*bld_base
)
2832 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2833 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2834 LLVMValueRef args
[2];
2836 args
[0] = lp_build_const_int32(gallivm
, SENDMSG_GS_OP_NOP
| SENDMSG_GS_DONE
);
2837 args
[1] = LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_GS_WAVE_ID
);
2838 lp_build_intrinsic(gallivm
->builder
, "llvm.SI.sendmsg",
2839 ctx
->voidt
, args
, 2, 0);
2842 static void si_llvm_emit_vs_epilogue(struct lp_build_tgsi_context
*bld_base
)
2844 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2845 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2846 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
2847 struct si_shader_output_values
*outputs
= NULL
;
2850 assert(!ctx
->is_gs_copy_shader
);
2852 outputs
= MALLOC((info
->num_outputs
+ 1) * sizeof(outputs
[0]));
2854 /* Vertex color clamping.
2856 * This uses a state constant loaded in a user data SGPR and
2857 * an IF statement is added that clamps all colors if the constant
2860 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
2861 struct lp_build_if_state if_ctx
;
2862 LLVMValueRef cond
= NULL
;
2863 LLVMValueRef addr
, val
;
2865 for (i
= 0; i
< info
->num_outputs
; i
++) {
2866 if (info
->output_semantic_name
[i
] != TGSI_SEMANTIC_COLOR
&&
2867 info
->output_semantic_name
[i
] != TGSI_SEMANTIC_BCOLOR
)
2870 /* We've found a color. */
2872 /* The state is in the first bit of the user SGPR. */
2873 cond
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
2874 SI_PARAM_VS_STATE_BITS
);
2875 cond
= LLVMBuildTrunc(gallivm
->builder
, cond
,
2877 lp_build_if(&if_ctx
, gallivm
, cond
);
2880 for (j
= 0; j
< 4; j
++) {
2881 addr
= ctx
->radeon_bld
.soa
.outputs
[i
][j
];
2882 val
= LLVMBuildLoad(gallivm
->builder
, addr
, "");
2883 val
= radeon_llvm_saturate(bld_base
, val
);
2884 LLVMBuildStore(gallivm
->builder
, val
, addr
);
2889 lp_build_endif(&if_ctx
);
2892 for (i
= 0; i
< info
->num_outputs
; i
++) {
2893 outputs
[i
].name
= info
->output_semantic_name
[i
];
2894 outputs
[i
].sid
= info
->output_semantic_index
[i
];
2896 for (j
= 0; j
< 4; j
++)
2897 outputs
[i
].values
[j
] =
2898 LLVMBuildLoad(gallivm
->builder
,
2899 ctx
->radeon_bld
.soa
.outputs
[i
][j
],
2903 if (ctx
->is_monolithic
) {
2904 /* Export PrimitiveID when PS needs it. */
2905 if (si_vs_exports_prim_id(ctx
->shader
)) {
2906 outputs
[i
].name
= TGSI_SEMANTIC_PRIMID
;
2908 outputs
[i
].values
[0] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
2909 get_primitive_id(bld_base
, 0));
2910 outputs
[i
].values
[1] = bld_base
->base
.undef
;
2911 outputs
[i
].values
[2] = bld_base
->base
.undef
;
2912 outputs
[i
].values
[3] = bld_base
->base
.undef
;
2916 /* Return the primitive ID from the LLVM function. */
2918 LLVMBuildInsertValue(gallivm
->builder
,
2920 bitcast(bld_base
, TGSI_TYPE_FLOAT
,
2921 get_primitive_id(bld_base
, 0)),
2922 VS_EPILOG_PRIMID_LOC
, "");
2925 si_llvm_export_vs(bld_base
, outputs
, i
);
2929 struct si_ps_exports
{
2931 LLVMValueRef args
[10][9];
2934 unsigned si_get_spi_shader_z_format(bool writes_z
, bool writes_stencil
,
2935 bool writes_samplemask
)
2938 /* Z needs 32 bits. */
2939 if (writes_samplemask
)
2940 return V_028710_SPI_SHADER_32_ABGR
;
2941 else if (writes_stencil
)
2942 return V_028710_SPI_SHADER_32_GR
;
2944 return V_028710_SPI_SHADER_32_R
;
2945 } else if (writes_stencil
|| writes_samplemask
) {
2946 /* Both stencil and sample mask need only 16 bits. */
2947 return V_028710_SPI_SHADER_UINT16_ABGR
;
2949 return V_028710_SPI_SHADER_ZERO
;
2953 static void si_export_mrt_z(struct lp_build_tgsi_context
*bld_base
,
2954 LLVMValueRef depth
, LLVMValueRef stencil
,
2955 LLVMValueRef samplemask
, struct si_ps_exports
*exp
)
2957 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2958 struct lp_build_context
*base
= &bld_base
->base
;
2959 struct lp_build_context
*uint
= &bld_base
->uint_bld
;
2960 LLVMValueRef args
[9];
2962 unsigned format
= si_get_spi_shader_z_format(depth
!= NULL
,
2964 samplemask
!= NULL
);
2966 assert(depth
|| stencil
|| samplemask
);
2968 args
[1] = uint
->one
; /* whether the EXEC mask is valid */
2969 args
[2] = uint
->one
; /* DONE bit */
2971 /* Specify the target we are exporting */
2972 args
[3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_MRTZ
);
2974 args
[4] = uint
->zero
; /* COMP flag */
2975 args
[5] = base
->undef
; /* R, depth */
2976 args
[6] = base
->undef
; /* G, stencil test value[0:7], stencil op value[8:15] */
2977 args
[7] = base
->undef
; /* B, sample mask */
2978 args
[8] = base
->undef
; /* A, alpha to mask */
2980 if (format
== V_028710_SPI_SHADER_UINT16_ABGR
) {
2982 args
[4] = uint
->one
; /* COMPR flag */
2985 /* Stencil should be in X[23:16]. */
2986 stencil
= bitcast(bld_base
, TGSI_TYPE_UNSIGNED
, stencil
);
2987 stencil
= LLVMBuildShl(base
->gallivm
->builder
, stencil
,
2988 LLVMConstInt(ctx
->i32
, 16, 0), "");
2989 args
[5] = bitcast(bld_base
, TGSI_TYPE_FLOAT
, stencil
);
2993 /* SampleMask should be in Y[15:0]. */
2994 args
[6] = samplemask
;
3007 args
[7] = samplemask
;
3012 /* SI (except OLAND) has a bug that it only looks
3013 * at the X writemask component. */
3014 if (ctx
->screen
->b
.chip_class
== SI
&&
3015 ctx
->screen
->b
.family
!= CHIP_OLAND
)
3018 /* Specify which components to enable */
3019 args
[0] = lp_build_const_int32(base
->gallivm
, mask
);
3021 memcpy(exp
->args
[exp
->num
++], args
, sizeof(args
));
3024 static void si_export_mrt_color(struct lp_build_tgsi_context
*bld_base
,
3025 LLVMValueRef
*color
, unsigned index
,
3026 unsigned samplemask_param
,
3027 bool is_last
, struct si_ps_exports
*exp
)
3029 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3030 struct lp_build_context
*base
= &bld_base
->base
;
3034 if (ctx
->shader
->key
.ps
.epilog
.clamp_color
)
3035 for (i
= 0; i
< 4; i
++)
3036 color
[i
] = radeon_llvm_saturate(bld_base
, color
[i
]);
3039 if (ctx
->shader
->key
.ps
.epilog
.alpha_to_one
)
3040 color
[3] = base
->one
;
3044 ctx
->shader
->key
.ps
.epilog
.alpha_func
!= PIPE_FUNC_ALWAYS
)
3045 si_alpha_test(bld_base
, color
[3]);
3047 /* Line & polygon smoothing */
3048 if (ctx
->shader
->key
.ps
.epilog
.poly_line_smoothing
)
3049 color
[3] = si_scale_alpha_by_sample_mask(bld_base
, color
[3],
3052 /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
3053 if (ctx
->shader
->key
.ps
.epilog
.last_cbuf
> 0) {
3054 LLVMValueRef args
[8][9];
3057 /* Get the export arguments, also find out what the last one is. */
3058 for (c
= 0; c
<= ctx
->shader
->key
.ps
.epilog
.last_cbuf
; c
++) {
3059 si_llvm_init_export_args(bld_base
, color
,
3060 V_008DFC_SQ_EXP_MRT
+ c
, args
[c
]);
3061 if (args
[c
][0] != bld_base
->uint_bld
.zero
)
3065 /* Emit all exports. */
3066 for (c
= 0; c
<= ctx
->shader
->key
.ps
.epilog
.last_cbuf
; c
++) {
3067 if (is_last
&& last
== c
) {
3068 args
[c
][1] = bld_base
->uint_bld
.one
; /* whether the EXEC mask is valid */
3069 args
[c
][2] = bld_base
->uint_bld
.one
; /* DONE bit */
3070 } else if (args
[c
][0] == bld_base
->uint_bld
.zero
)
3071 continue; /* unnecessary NULL export */
3073 memcpy(exp
->args
[exp
->num
++], args
[c
], sizeof(args
[c
]));
3076 LLVMValueRef args
[9];
3079 si_llvm_init_export_args(bld_base
, color
, V_008DFC_SQ_EXP_MRT
+ index
,
3082 args
[1] = bld_base
->uint_bld
.one
; /* whether the EXEC mask is valid */
3083 args
[2] = bld_base
->uint_bld
.one
; /* DONE bit */
3084 } else if (args
[0] == bld_base
->uint_bld
.zero
)
3085 return; /* unnecessary NULL export */
3087 memcpy(exp
->args
[exp
->num
++], args
, sizeof(args
));
3091 static void si_emit_ps_exports(struct si_shader_context
*ctx
,
3092 struct si_ps_exports
*exp
)
3094 for (unsigned i
= 0; i
< exp
->num
; i
++)
3095 lp_build_intrinsic(ctx
->radeon_bld
.gallivm
.builder
,
3096 "llvm.SI.export", ctx
->voidt
,
3097 exp
->args
[i
], 9, 0);
3100 static void si_export_null(struct lp_build_tgsi_context
*bld_base
)
3102 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3103 struct lp_build_context
*base
= &bld_base
->base
;
3104 struct lp_build_context
*uint
= &bld_base
->uint_bld
;
3105 LLVMValueRef args
[9];
3107 args
[0] = lp_build_const_int32(base
->gallivm
, 0x0); /* enabled channels */
3108 args
[1] = uint
->one
; /* whether the EXEC mask is valid */
3109 args
[2] = uint
->one
; /* DONE bit */
3110 args
[3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_NULL
);
3111 args
[4] = uint
->zero
; /* COMPR flag (0 = 32-bit export) */
3112 args
[5] = base
->undef
; /* R */
3113 args
[6] = base
->undef
; /* G */
3114 args
[7] = base
->undef
; /* B */
3115 args
[8] = base
->undef
; /* A */
3117 lp_build_intrinsic(base
->gallivm
->builder
, "llvm.SI.export",
3118 ctx
->voidt
, args
, 9, 0);
3121 static void si_llvm_emit_fs_epilogue(struct lp_build_tgsi_context
*bld_base
)
3123 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3124 struct si_shader
*shader
= ctx
->shader
;
3125 struct lp_build_context
*base
= &bld_base
->base
;
3126 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
3127 LLVMBuilderRef builder
= base
->gallivm
->builder
;
3128 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
3129 int last_color_export
= -1;
3131 struct si_ps_exports exp
= {};
3133 /* Determine the last export. If MRTZ is present, it's always last.
3134 * Otherwise, find the last color export.
3136 if (!info
->writes_z
&& !info
->writes_stencil
&& !info
->writes_samplemask
) {
3137 unsigned spi_format
= shader
->key
.ps
.epilog
.spi_shader_col_format
;
3139 /* Don't export NULL and return if alpha-test is enabled. */
3140 if (shader
->key
.ps
.epilog
.alpha_func
!= PIPE_FUNC_ALWAYS
&&
3141 shader
->key
.ps
.epilog
.alpha_func
!= PIPE_FUNC_NEVER
&&
3142 (spi_format
& 0xf) == 0)
3143 spi_format
|= V_028714_SPI_SHADER_32_AR
;
3145 for (i
= 0; i
< info
->num_outputs
; i
++) {
3146 unsigned index
= info
->output_semantic_index
[i
];
3148 if (info
->output_semantic_name
[i
] != TGSI_SEMANTIC_COLOR
)
3151 /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
3152 if (shader
->key
.ps
.epilog
.last_cbuf
> 0) {
3153 /* Just set this if any of the colorbuffers are enabled. */
3155 ((1llu << (4 * (shader
->key
.ps
.epilog
.last_cbuf
+ 1))) - 1))
3156 last_color_export
= i
;
3160 if ((spi_format
>> (index
* 4)) & 0xf)
3161 last_color_export
= i
;
3164 /* If there are no outputs, export NULL. */
3165 if (last_color_export
== -1) {
3166 si_export_null(bld_base
);
3171 for (i
= 0; i
< info
->num_outputs
; i
++) {
3172 unsigned semantic_name
= info
->output_semantic_name
[i
];
3173 unsigned semantic_index
= info
->output_semantic_index
[i
];
3175 LLVMValueRef color
[4] = {};
3177 /* Select the correct target */
3178 switch (semantic_name
) {
3179 case TGSI_SEMANTIC_POSITION
:
3180 depth
= LLVMBuildLoad(builder
,
3181 ctx
->radeon_bld
.soa
.outputs
[i
][2], "");
3183 case TGSI_SEMANTIC_STENCIL
:
3184 stencil
= LLVMBuildLoad(builder
,
3185 ctx
->radeon_bld
.soa
.outputs
[i
][1], "");
3187 case TGSI_SEMANTIC_SAMPLEMASK
:
3188 samplemask
= LLVMBuildLoad(builder
,
3189 ctx
->radeon_bld
.soa
.outputs
[i
][0], "");
3191 case TGSI_SEMANTIC_COLOR
:
3192 for (j
= 0; j
< 4; j
++)
3193 color
[j
] = LLVMBuildLoad(builder
,
3194 ctx
->radeon_bld
.soa
.outputs
[i
][j
], "");
3196 si_export_mrt_color(bld_base
, color
, semantic_index
,
3197 SI_PARAM_SAMPLE_COVERAGE
,
3198 last_color_export
== i
, &exp
);
3202 "Warning: SI unhandled fs output type:%d\n",
3207 if (depth
|| stencil
|| samplemask
)
3208 si_export_mrt_z(bld_base
, depth
, stencil
, samplemask
, &exp
);
3210 si_emit_ps_exports(ctx
, &exp
);
3214 * Return PS outputs in this order:
3216 * v[0:3] = color0.xyzw
3217 * v[4:7] = color1.xyzw
3222 * vN+3 = SampleMaskIn (used for OpenGL smoothing)
3224 * The alpha-ref SGPR is returned via its original location.
3226 static void si_llvm_return_fs_outputs(struct lp_build_tgsi_context
*bld_base
)
3228 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3229 struct si_shader
*shader
= ctx
->shader
;
3230 struct lp_build_context
*base
= &bld_base
->base
;
3231 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
3232 LLVMBuilderRef builder
= base
->gallivm
->builder
;
3233 unsigned i
, j
, first_vgpr
, vgpr
;
3235 LLVMValueRef color
[8][4] = {};
3236 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
3239 /* Read the output values. */
3240 for (i
= 0; i
< info
->num_outputs
; i
++) {
3241 unsigned semantic_name
= info
->output_semantic_name
[i
];
3242 unsigned semantic_index
= info
->output_semantic_index
[i
];
3244 switch (semantic_name
) {
3245 case TGSI_SEMANTIC_COLOR
:
3246 assert(semantic_index
< 8);
3247 for (j
= 0; j
< 4; j
++) {
3248 LLVMValueRef ptr
= ctx
->radeon_bld
.soa
.outputs
[i
][j
];
3249 LLVMValueRef result
= LLVMBuildLoad(builder
, ptr
, "");
3250 color
[semantic_index
][j
] = result
;
3253 case TGSI_SEMANTIC_POSITION
:
3254 depth
= LLVMBuildLoad(builder
,
3255 ctx
->radeon_bld
.soa
.outputs
[i
][2], "");
3257 case TGSI_SEMANTIC_STENCIL
:
3258 stencil
= LLVMBuildLoad(builder
,
3259 ctx
->radeon_bld
.soa
.outputs
[i
][1], "");
3261 case TGSI_SEMANTIC_SAMPLEMASK
:
3262 samplemask
= LLVMBuildLoad(builder
,
3263 ctx
->radeon_bld
.soa
.outputs
[i
][0], "");
3266 fprintf(stderr
, "Warning: SI unhandled fs output type:%d\n",
3271 /* Fill the return structure. */
3272 ret
= ctx
->return_value
;
3275 ret
= LLVMBuildInsertValue(builder
, ret
,
3276 bitcast(bld_base
, TGSI_TYPE_SIGNED
,
3277 LLVMGetParam(ctx
->radeon_bld
.main_fn
,
3278 SI_PARAM_ALPHA_REF
)),
3279 SI_SGPR_ALPHA_REF
, "");
3282 first_vgpr
= vgpr
= SI_SGPR_ALPHA_REF
+ 1;
3283 for (i
= 0; i
< ARRAY_SIZE(color
); i
++) {
3287 for (j
= 0; j
< 4; j
++)
3288 ret
= LLVMBuildInsertValue(builder
, ret
, color
[i
][j
], vgpr
++, "");
3291 ret
= LLVMBuildInsertValue(builder
, ret
, depth
, vgpr
++, "");
3293 ret
= LLVMBuildInsertValue(builder
, ret
, stencil
, vgpr
++, "");
3295 ret
= LLVMBuildInsertValue(builder
, ret
, samplemask
, vgpr
++, "");
3297 /* Add the input sample mask for smoothing at the end. */
3298 if (vgpr
< first_vgpr
+ PS_EPILOG_SAMPLEMASK_MIN_LOC
)
3299 vgpr
= first_vgpr
+ PS_EPILOG_SAMPLEMASK_MIN_LOC
;
3300 ret
= LLVMBuildInsertValue(builder
, ret
,
3301 LLVMGetParam(ctx
->radeon_bld
.main_fn
,
3302 SI_PARAM_SAMPLE_COVERAGE
), vgpr
++, "");
3304 ctx
->return_value
= ret
;
3308 * Given a v8i32 resource descriptor for a buffer, extract the size of the
3309 * buffer in number of elements and return it as an i32.
3311 static LLVMValueRef
get_buffer_size(
3312 struct lp_build_tgsi_context
*bld_base
,
3313 LLVMValueRef descriptor
)
3315 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3316 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3317 LLVMBuilderRef builder
= gallivm
->builder
;
3319 LLVMBuildExtractElement(builder
, descriptor
,
3320 lp_build_const_int32(gallivm
, 6), "");
3322 if (ctx
->screen
->b
.chip_class
>= VI
) {
3323 /* On VI, the descriptor contains the size in bytes,
3324 * but TXQ must return the size in elements.
3325 * The stride is always non-zero for resources using TXQ.
3327 LLVMValueRef stride
=
3328 LLVMBuildExtractElement(builder
, descriptor
,
3329 lp_build_const_int32(gallivm
, 5), "");
3330 stride
= LLVMBuildLShr(builder
, stride
,
3331 lp_build_const_int32(gallivm
, 16), "");
3332 stride
= LLVMBuildAnd(builder
, stride
,
3333 lp_build_const_int32(gallivm
, 0x3FFF), "");
3335 size
= LLVMBuildUDiv(builder
, size
, stride
, "");
3342 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
3345 static void build_int_type_name(
3347 char *buf
, unsigned bufsize
)
3349 assert(bufsize
>= 6);
3351 if (LLVMGetTypeKind(type
) == LLVMVectorTypeKind
)
3352 snprintf(buf
, bufsize
, "v%ui32",
3353 LLVMGetVectorSize(type
));
3358 static void build_tex_intrinsic(const struct lp_build_tgsi_action
*action
,
3359 struct lp_build_tgsi_context
*bld_base
,
3360 struct lp_build_emit_data
*emit_data
);
3362 /* Prevent optimizations (at least of memory accesses) across the current
3363 * point in the program by emitting empty inline assembly that is marked as
3364 * having side effects.
3366 static void emit_optimization_barrier(struct si_shader_context
*ctx
)
3368 LLVMBuilderRef builder
= ctx
->radeon_bld
.gallivm
.builder
;
3369 LLVMTypeRef ftype
= LLVMFunctionType(ctx
->voidt
, NULL
, 0, false);
3370 LLVMValueRef inlineasm
= LLVMConstInlineAsm(ftype
, "", "", true, false);
3371 LLVMBuildCall(builder
, inlineasm
, NULL
, 0, "");
3374 static void emit_waitcnt(struct si_shader_context
*ctx
)
3376 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
3377 LLVMBuilderRef builder
= gallivm
->builder
;
3378 LLVMValueRef args
[1] = {
3379 lp_build_const_int32(gallivm
, 0xf70)
3381 lp_build_intrinsic(builder
, "llvm.amdgcn.s.waitcnt",
3382 ctx
->voidt
, args
, 1, 0);
3385 static void membar_emit(
3386 const struct lp_build_tgsi_action
*action
,
3387 struct lp_build_tgsi_context
*bld_base
,
3388 struct lp_build_emit_data
*emit_data
)
3390 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3396 shader_buffer_fetch_rsrc(struct si_shader_context
*ctx
,
3397 const struct tgsi_full_src_register
*reg
)
3400 LLVMValueRef rsrc_ptr
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
3401 SI_PARAM_SHADER_BUFFERS
);
3403 if (!reg
->Register
.Indirect
)
3404 index
= LLVMConstInt(ctx
->i32
, reg
->Register
.Index
, 0);
3406 index
= get_bounded_indirect_index(ctx
, ®
->Indirect
,
3407 reg
->Register
.Index
,
3408 SI_NUM_SHADER_BUFFERS
);
3410 return build_indexed_load_const(ctx
, rsrc_ptr
, index
);
3413 static bool tgsi_is_array_sampler(unsigned target
)
3415 return target
== TGSI_TEXTURE_1D_ARRAY
||
3416 target
== TGSI_TEXTURE_SHADOW1D_ARRAY
||
3417 target
== TGSI_TEXTURE_2D_ARRAY
||
3418 target
== TGSI_TEXTURE_SHADOW2D_ARRAY
||
3419 target
== TGSI_TEXTURE_CUBE_ARRAY
||
3420 target
== TGSI_TEXTURE_SHADOWCUBE_ARRAY
||
3421 target
== TGSI_TEXTURE_2D_ARRAY_MSAA
;
3424 static bool tgsi_is_array_image(unsigned target
)
3426 return target
== TGSI_TEXTURE_3D
||
3427 target
== TGSI_TEXTURE_CUBE
||
3428 target
== TGSI_TEXTURE_1D_ARRAY
||
3429 target
== TGSI_TEXTURE_2D_ARRAY
||
3430 target
== TGSI_TEXTURE_CUBE_ARRAY
||
3431 target
== TGSI_TEXTURE_2D_ARRAY_MSAA
;
3435 * Given a 256-bit resource descriptor, force the DCC enable bit to off.
3437 * At least on Tonga, executing image stores on images with DCC enabled and
3438 * non-trivial can eventually lead to lockups. This can occur when an
3439 * application binds an image as read-only but then uses a shader that writes
3440 * to it. The OpenGL spec allows almost arbitrarily bad behavior (including
3441 * program termination) in this case, but it doesn't cost much to be a bit
3442 * nicer: disabling DCC in the shader still leads to undefined results but
3443 * avoids the lockup.
3445 static LLVMValueRef
force_dcc_off(struct si_shader_context
*ctx
,
3448 if (ctx
->screen
->b
.chip_class
<= CIK
) {
3451 LLVMBuilderRef builder
= ctx
->radeon_bld
.gallivm
.builder
;
3452 LLVMValueRef i32_6
= LLVMConstInt(ctx
->i32
, 6, 0);
3453 LLVMValueRef i32_C
= LLVMConstInt(ctx
->i32
, C_008F28_COMPRESSION_EN
, 0);
3456 tmp
= LLVMBuildExtractElement(builder
, rsrc
, i32_6
, "");
3457 tmp
= LLVMBuildAnd(builder
, tmp
, i32_C
, "");
3458 return LLVMBuildInsertElement(builder
, rsrc
, tmp
, i32_6
, "");
3463 * Load the resource descriptor for \p image.
3467 struct lp_build_tgsi_context
*bld_base
,
3468 const struct tgsi_full_src_register
*image
,
3472 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3473 LLVMValueRef rsrc_ptr
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
3475 LLVMValueRef index
, tmp
;
3477 assert(image
->Register
.File
== TGSI_FILE_IMAGE
);
3479 if (!image
->Register
.Indirect
) {
3480 const struct tgsi_shader_info
*info
= bld_base
->info
;
3482 index
= LLVMConstInt(ctx
->i32
, image
->Register
.Index
, 0);
3484 if (info
->images_writemask
& (1 << image
->Register
.Index
) &&
3485 !(info
->images_buffers
& (1 << image
->Register
.Index
)))
3488 /* From the GL_ARB_shader_image_load_store extension spec:
3490 * If a shader performs an image load, store, or atomic
3491 * operation using an image variable declared as an array,
3492 * and if the index used to select an individual element is
3493 * negative or greater than or equal to the size of the
3494 * array, the results of the operation are undefined but may
3495 * not lead to termination.
3497 index
= get_bounded_indirect_index(ctx
, &image
->Indirect
,
3498 image
->Register
.Index
,
3502 tmp
= build_indexed_load_const(ctx
, rsrc_ptr
, index
);
3504 tmp
= force_dcc_off(ctx
, tmp
);
3508 static LLVMValueRef
image_fetch_coords(
3509 struct lp_build_tgsi_context
*bld_base
,
3510 const struct tgsi_full_instruction
*inst
,
3513 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3514 LLVMBuilderRef builder
= gallivm
->builder
;
3515 unsigned target
= inst
->Memory
.Texture
;
3516 unsigned num_coords
= tgsi_util_get_texture_coord_dim(target
);
3517 LLVMValueRef coords
[4];
3521 for (chan
= 0; chan
< num_coords
; ++chan
) {
3522 tmp
= lp_build_emit_fetch(bld_base
, inst
, src
, chan
);
3523 tmp
= LLVMBuildBitCast(builder
, tmp
, bld_base
->uint_bld
.elem_type
, "");
3527 if (num_coords
== 1)
3530 if (num_coords
== 3) {
3531 /* LLVM has difficulties lowering 3-element vectors. */
3532 coords
[3] = bld_base
->uint_bld
.undef
;
3536 return lp_build_gather_values(gallivm
, coords
, num_coords
);
3540 * Append the extra mode bits that are used by image load and store.
3542 static void image_append_args(
3543 struct si_shader_context
*ctx
,
3544 struct lp_build_emit_data
* emit_data
,
3548 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3549 LLVMValueRef i1false
= LLVMConstInt(ctx
->i1
, 0, 0);
3550 LLVMValueRef i1true
= LLVMConstInt(ctx
->i1
, 1, 0);
3552 emit_data
->args
[emit_data
->arg_count
++] = i1false
; /* r128 */
3553 emit_data
->args
[emit_data
->arg_count
++] =
3554 tgsi_is_array_image(target
) ? i1true
: i1false
; /* da */
3556 emit_data
->args
[emit_data
->arg_count
++] =
3557 inst
->Memory
.Qualifier
& (TGSI_MEMORY_COHERENT
| TGSI_MEMORY_VOLATILE
) ?
3558 i1true
: i1false
; /* glc */
3560 emit_data
->args
[emit_data
->arg_count
++] = i1false
; /* slc */
3564 * Given a 256 bit resource, extract the top half (which stores the buffer
3565 * resource in the case of textures and images).
3567 static LLVMValueRef
extract_rsrc_top_half(
3568 struct si_shader_context
*ctx
,
3571 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
3572 struct lp_build_tgsi_context
*bld_base
= &ctx
->radeon_bld
.soa
.bld_base
;
3573 LLVMTypeRef v2i128
= LLVMVectorType(ctx
->i128
, 2);
3575 rsrc
= LLVMBuildBitCast(gallivm
->builder
, rsrc
, v2i128
, "");
3576 rsrc
= LLVMBuildExtractElement(gallivm
->builder
, rsrc
, bld_base
->uint_bld
.one
, "");
3577 rsrc
= LLVMBuildBitCast(gallivm
->builder
, rsrc
, ctx
->v4i32
, "");
3583 * Append the resource and indexing arguments for buffer intrinsics.
3585 * \param rsrc the v4i32 buffer resource
3586 * \param index index into the buffer (stride-based)
3587 * \param offset byte offset into the buffer
3589 static void buffer_append_args(
3590 struct si_shader_context
*ctx
,
3591 struct lp_build_emit_data
*emit_data
,
3594 LLVMValueRef offset
,
3597 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3598 LLVMValueRef i1false
= LLVMConstInt(ctx
->i1
, 0, 0);
3599 LLVMValueRef i1true
= LLVMConstInt(ctx
->i1
, 1, 0);
3601 emit_data
->args
[emit_data
->arg_count
++] = rsrc
;
3602 emit_data
->args
[emit_data
->arg_count
++] = index
; /* vindex */
3603 emit_data
->args
[emit_data
->arg_count
++] = offset
; /* voffset */
3605 emit_data
->args
[emit_data
->arg_count
++] =
3606 inst
->Memory
.Qualifier
& (TGSI_MEMORY_COHERENT
| TGSI_MEMORY_VOLATILE
) ?
3607 i1true
: i1false
; /* glc */
3609 emit_data
->args
[emit_data
->arg_count
++] = i1false
; /* slc */
3612 static void load_fetch_args(
3613 struct lp_build_tgsi_context
* bld_base
,
3614 struct lp_build_emit_data
* emit_data
)
3616 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3617 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3618 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
3619 unsigned target
= inst
->Memory
.Texture
;
3622 emit_data
->dst_type
= LLVMVectorType(bld_base
->base
.elem_type
, 4);
3624 if (inst
->Src
[0].Register
.File
== TGSI_FILE_BUFFER
) {
3625 LLVMBuilderRef builder
= gallivm
->builder
;
3626 LLVMValueRef offset
;
3629 rsrc
= shader_buffer_fetch_rsrc(ctx
, &inst
->Src
[0]);
3631 tmp
= lp_build_emit_fetch(bld_base
, inst
, 1, 0);
3632 offset
= LLVMBuildBitCast(builder
, tmp
, bld_base
->uint_bld
.elem_type
, "");
3634 buffer_append_args(ctx
, emit_data
, rsrc
, bld_base
->uint_bld
.zero
,
3636 } else if (inst
->Src
[0].Register
.File
== TGSI_FILE_IMAGE
) {
3637 LLVMValueRef coords
;
3639 image_fetch_rsrc(bld_base
, &inst
->Src
[0], false, &rsrc
);
3640 coords
= image_fetch_coords(bld_base
, inst
, 1);
3642 if (target
== TGSI_TEXTURE_BUFFER
) {
3643 rsrc
= extract_rsrc_top_half(ctx
, rsrc
);
3644 buffer_append_args(ctx
, emit_data
, rsrc
, coords
,
3645 bld_base
->uint_bld
.zero
, false);
3647 emit_data
->args
[0] = coords
;
3648 emit_data
->args
[1] = rsrc
;
3649 emit_data
->args
[2] = lp_build_const_int32(gallivm
, 15); /* dmask */
3650 emit_data
->arg_count
= 3;
3652 image_append_args(ctx
, emit_data
, target
, false);
3657 static void load_emit_buffer(struct si_shader_context
*ctx
,
3658 struct lp_build_emit_data
*emit_data
)
3660 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3661 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
3662 LLVMBuilderRef builder
= gallivm
->builder
;
3663 uint writemask
= inst
->Dst
[0].Register
.WriteMask
;
3664 uint count
= util_last_bit(writemask
);
3665 const char *intrinsic_name
;
3666 LLVMTypeRef dst_type
;
3670 intrinsic_name
= "llvm.amdgcn.buffer.load.f32";
3671 dst_type
= ctx
->f32
;
3674 intrinsic_name
= "llvm.amdgcn.buffer.load.v2f32";
3675 dst_type
= LLVMVectorType(ctx
->f32
, 2);
3678 intrinsic_name
= "llvm.amdgcn.buffer.load.v4f32";
3679 dst_type
= ctx
->v4f32
;
3683 emit_data
->output
[emit_data
->chan
] = lp_build_intrinsic(
3684 builder
, intrinsic_name
, dst_type
,
3685 emit_data
->args
, emit_data
->arg_count
,
3686 LLVMReadOnlyAttribute
);
3689 static LLVMValueRef
get_memory_ptr(struct si_shader_context
*ctx
,
3690 const struct tgsi_full_instruction
*inst
,
3691 LLVMTypeRef type
, int arg
)
3693 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
3694 LLVMBuilderRef builder
= gallivm
->builder
;
3695 LLVMValueRef offset
, ptr
;
3698 offset
= lp_build_emit_fetch(&ctx
->radeon_bld
.soa
.bld_base
, inst
, arg
, 0);
3699 offset
= LLVMBuildBitCast(builder
, offset
, ctx
->i32
, "");
3701 ptr
= ctx
->shared_memory
;
3702 ptr
= LLVMBuildGEP(builder
, ptr
, &offset
, 1, "");
3703 addr_space
= LLVMGetPointerAddressSpace(LLVMTypeOf(ptr
));
3704 ptr
= LLVMBuildBitCast(builder
, ptr
, LLVMPointerType(type
, addr_space
), "");
3709 static void load_emit_memory(
3710 struct si_shader_context
*ctx
,
3711 struct lp_build_emit_data
*emit_data
)
3713 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3714 struct lp_build_context
*base
= &ctx
->radeon_bld
.soa
.bld_base
.base
;
3715 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
3716 LLVMBuilderRef builder
= gallivm
->builder
;
3717 unsigned writemask
= inst
->Dst
[0].Register
.WriteMask
;
3718 LLVMValueRef channels
[4], ptr
, derived_ptr
, index
;
3721 ptr
= get_memory_ptr(ctx
, inst
, base
->elem_type
, 1);
3723 for (chan
= 0; chan
< 4; ++chan
) {
3724 if (!(writemask
& (1 << chan
))) {
3725 channels
[chan
] = LLVMGetUndef(base
->elem_type
);
3729 index
= lp_build_const_int32(gallivm
, chan
);
3730 derived_ptr
= LLVMBuildGEP(builder
, ptr
, &index
, 1, "");
3731 channels
[chan
] = LLVMBuildLoad(builder
, derived_ptr
, "");
3733 emit_data
->output
[emit_data
->chan
] = lp_build_gather_values(gallivm
, channels
, 4);
3736 static void load_emit(
3737 const struct lp_build_tgsi_action
*action
,
3738 struct lp_build_tgsi_context
*bld_base
,
3739 struct lp_build_emit_data
*emit_data
)
3741 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3742 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3743 LLVMBuilderRef builder
= gallivm
->builder
;
3744 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
3745 char intrinsic_name
[32];
3746 char coords_type
[8];
3748 if (inst
->Src
[0].Register
.File
== TGSI_FILE_MEMORY
) {
3749 load_emit_memory(ctx
, emit_data
);
3753 if (inst
->Memory
.Qualifier
& TGSI_MEMORY_VOLATILE
)
3756 if (inst
->Src
[0].Register
.File
== TGSI_FILE_BUFFER
) {
3757 load_emit_buffer(ctx
, emit_data
);
3761 if (inst
->Memory
.Texture
== TGSI_TEXTURE_BUFFER
) {
3762 emit_data
->output
[emit_data
->chan
] =
3764 builder
, "llvm.amdgcn.buffer.load.format.v4f32", emit_data
->dst_type
,
3765 emit_data
->args
, emit_data
->arg_count
,
3766 LLVMReadOnlyAttribute
);
3768 build_int_type_name(LLVMTypeOf(emit_data
->args
[0]),
3769 coords_type
, sizeof(coords_type
));
3771 snprintf(intrinsic_name
, sizeof(intrinsic_name
),
3772 "llvm.amdgcn.image.load.%s", coords_type
);
3774 emit_data
->output
[emit_data
->chan
] =
3776 builder
, intrinsic_name
, emit_data
->dst_type
,
3777 emit_data
->args
, emit_data
->arg_count
,
3778 LLVMReadOnlyAttribute
);
3782 static void store_fetch_args(
3783 struct lp_build_tgsi_context
* bld_base
,
3784 struct lp_build_emit_data
* emit_data
)
3786 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3787 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3788 LLVMBuilderRef builder
= gallivm
->builder
;
3789 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
3790 struct tgsi_full_src_register memory
;
3791 LLVMValueRef chans
[4];
3796 emit_data
->dst_type
= LLVMVoidTypeInContext(gallivm
->context
);
3798 for (chan
= 0; chan
< 4; ++chan
) {
3799 chans
[chan
] = lp_build_emit_fetch(bld_base
, inst
, 1, chan
);
3801 data
= lp_build_gather_values(gallivm
, chans
, 4);
3803 emit_data
->args
[emit_data
->arg_count
++] = data
;
3805 memory
= tgsi_full_src_register_from_dst(&inst
->Dst
[0]);
3807 if (inst
->Dst
[0].Register
.File
== TGSI_FILE_BUFFER
) {
3808 LLVMValueRef offset
;
3811 rsrc
= shader_buffer_fetch_rsrc(ctx
, &memory
);
3813 tmp
= lp_build_emit_fetch(bld_base
, inst
, 0, 0);
3814 offset
= LLVMBuildBitCast(builder
, tmp
, bld_base
->uint_bld
.elem_type
, "");
3816 buffer_append_args(ctx
, emit_data
, rsrc
, bld_base
->uint_bld
.zero
,
3818 } else if (inst
->Dst
[0].Register
.File
== TGSI_FILE_IMAGE
) {
3819 unsigned target
= inst
->Memory
.Texture
;
3820 LLVMValueRef coords
;
3822 coords
= image_fetch_coords(bld_base
, inst
, 0);
3824 if (target
== TGSI_TEXTURE_BUFFER
) {
3825 image_fetch_rsrc(bld_base
, &memory
, false, &rsrc
);
3827 rsrc
= extract_rsrc_top_half(ctx
, rsrc
);
3828 buffer_append_args(ctx
, emit_data
, rsrc
, coords
,
3829 bld_base
->uint_bld
.zero
, false);
3831 emit_data
->args
[1] = coords
;
3832 image_fetch_rsrc(bld_base
, &memory
, true, &emit_data
->args
[2]);
3833 emit_data
->args
[3] = lp_build_const_int32(gallivm
, 15); /* dmask */
3834 emit_data
->arg_count
= 4;
3836 image_append_args(ctx
, emit_data
, target
, false);
3841 static void store_emit_buffer(
3842 struct si_shader_context
*ctx
,
3843 struct lp_build_emit_data
*emit_data
)
3845 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3846 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
3847 LLVMBuilderRef builder
= gallivm
->builder
;
3848 struct lp_build_context
*uint_bld
= &ctx
->radeon_bld
.soa
.bld_base
.uint_bld
;
3849 LLVMValueRef base_data
= emit_data
->args
[0];
3850 LLVMValueRef base_offset
= emit_data
->args
[3];
3851 unsigned writemask
= inst
->Dst
[0].Register
.WriteMask
;
3855 const char *intrinsic_name
;
3857 LLVMValueRef offset
;
3860 u_bit_scan_consecutive_range(&writemask
, &start
, &count
);
3862 /* Due to an LLVM limitation, split 3-element writes
3863 * into a 2-element and a 1-element write. */
3865 writemask
|= 1 << (start
+ 2);
3871 intrinsic_name
= "llvm.amdgcn.buffer.store.v4f32";
3872 } else if (count
== 2) {
3873 LLVMTypeRef v2f32
= LLVMVectorType(ctx
->f32
, 2);
3875 tmp
= LLVMBuildExtractElement(
3877 lp_build_const_int32(gallivm
, start
), "");
3878 data
= LLVMBuildInsertElement(
3879 builder
, LLVMGetUndef(v2f32
), tmp
,
3880 uint_bld
->zero
, "");
3882 tmp
= LLVMBuildExtractElement(
3884 lp_build_const_int32(gallivm
, start
+ 1), "");
3885 data
= LLVMBuildInsertElement(
3886 builder
, data
, tmp
, uint_bld
->one
, "");
3888 intrinsic_name
= "llvm.amdgcn.buffer.store.v2f32";
3891 data
= LLVMBuildExtractElement(
3893 lp_build_const_int32(gallivm
, start
), "");
3894 intrinsic_name
= "llvm.amdgcn.buffer.store.f32";
3897 offset
= base_offset
;
3899 offset
= LLVMBuildAdd(
3901 lp_build_const_int32(gallivm
, start
* 4), "");
3904 emit_data
->args
[0] = data
;
3905 emit_data
->args
[3] = offset
;
3908 builder
, intrinsic_name
, emit_data
->dst_type
,
3909 emit_data
->args
, emit_data
->arg_count
, 0);
3913 static void store_emit_memory(
3914 struct si_shader_context
*ctx
,
3915 struct lp_build_emit_data
*emit_data
)
3917 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3918 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
3919 struct lp_build_context
*base
= &ctx
->radeon_bld
.soa
.bld_base
.base
;
3920 LLVMBuilderRef builder
= gallivm
->builder
;
3921 unsigned writemask
= inst
->Dst
[0].Register
.WriteMask
;
3922 LLVMValueRef ptr
, derived_ptr
, data
, index
;
3925 ptr
= get_memory_ptr(ctx
, inst
, base
->elem_type
, 0);
3927 for (chan
= 0; chan
< 4; ++chan
) {
3928 if (!(writemask
& (1 << chan
))) {
3931 data
= lp_build_emit_fetch(&ctx
->radeon_bld
.soa
.bld_base
, inst
, 1, chan
);
3932 index
= lp_build_const_int32(gallivm
, chan
);
3933 derived_ptr
= LLVMBuildGEP(builder
, ptr
, &index
, 1, "");
3934 LLVMBuildStore(builder
, data
, derived_ptr
);
3938 static void store_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
);
3944 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3945 LLVMBuilderRef builder
= gallivm
->builder
;
3946 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
3947 unsigned target
= inst
->Memory
.Texture
;
3948 char intrinsic_name
[32];
3949 char coords_type
[8];
3951 if (inst
->Dst
[0].Register
.File
== TGSI_FILE_MEMORY
) {
3952 store_emit_memory(ctx
, emit_data
);
3956 if (inst
->Memory
.Qualifier
& TGSI_MEMORY_VOLATILE
)
3959 if (inst
->Dst
[0].Register
.File
== TGSI_FILE_BUFFER
) {
3960 store_emit_buffer(ctx
, emit_data
);
3964 if (target
== TGSI_TEXTURE_BUFFER
) {
3965 emit_data
->output
[emit_data
->chan
] = lp_build_intrinsic(
3966 builder
, "llvm.amdgcn.buffer.store.format.v4f32",
3967 emit_data
->dst_type
, emit_data
->args
,
3968 emit_data
->arg_count
, 0);
3970 build_int_type_name(LLVMTypeOf(emit_data
->args
[1]),
3971 coords_type
, sizeof(coords_type
));
3972 snprintf(intrinsic_name
, sizeof(intrinsic_name
),
3973 "llvm.amdgcn.image.store.%s", coords_type
);
3975 emit_data
->output
[emit_data
->chan
] =
3977 builder
, intrinsic_name
, emit_data
->dst_type
,
3978 emit_data
->args
, emit_data
->arg_count
, 0);
3982 static void atomic_fetch_args(
3983 struct lp_build_tgsi_context
* bld_base
,
3984 struct lp_build_emit_data
* emit_data
)
3986 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3987 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3988 LLVMBuilderRef builder
= gallivm
->builder
;
3989 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
3990 LLVMValueRef data1
, data2
;
3994 emit_data
->dst_type
= bld_base
->base
.elem_type
;
3996 tmp
= lp_build_emit_fetch(bld_base
, inst
, 2, 0);
3997 data1
= LLVMBuildBitCast(builder
, tmp
, bld_base
->uint_bld
.elem_type
, "");
3999 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_ATOMCAS
) {
4000 tmp
= lp_build_emit_fetch(bld_base
, inst
, 3, 0);
4001 data2
= LLVMBuildBitCast(builder
, tmp
, bld_base
->uint_bld
.elem_type
, "");
4004 /* llvm.amdgcn.image/buffer.atomic.cmpswap reflect the hardware order
4005 * of arguments, which is reversed relative to TGSI (and GLSL)
4007 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_ATOMCAS
)
4008 emit_data
->args
[emit_data
->arg_count
++] = data2
;
4009 emit_data
->args
[emit_data
->arg_count
++] = data1
;
4011 if (inst
->Src
[0].Register
.File
== TGSI_FILE_BUFFER
) {
4012 LLVMValueRef offset
;
4014 rsrc
= shader_buffer_fetch_rsrc(ctx
, &inst
->Src
[0]);
4016 tmp
= lp_build_emit_fetch(bld_base
, inst
, 1, 0);
4017 offset
= LLVMBuildBitCast(builder
, tmp
, bld_base
->uint_bld
.elem_type
, "");
4019 buffer_append_args(ctx
, emit_data
, rsrc
, bld_base
->uint_bld
.zero
,
4021 } else if (inst
->Src
[0].Register
.File
== TGSI_FILE_IMAGE
) {
4022 unsigned target
= inst
->Memory
.Texture
;
4023 LLVMValueRef coords
;
4025 image_fetch_rsrc(bld_base
, &inst
->Src
[0],
4026 target
!= TGSI_TEXTURE_BUFFER
, &rsrc
);
4027 coords
= image_fetch_coords(bld_base
, inst
, 1);
4029 if (target
== TGSI_TEXTURE_BUFFER
) {
4030 rsrc
= extract_rsrc_top_half(ctx
, rsrc
);
4031 buffer_append_args(ctx
, emit_data
, rsrc
, coords
,
4032 bld_base
->uint_bld
.zero
, true);
4034 emit_data
->args
[emit_data
->arg_count
++] = coords
;
4035 emit_data
->args
[emit_data
->arg_count
++] = rsrc
;
4037 image_append_args(ctx
, emit_data
, target
, true);
4042 static void atomic_emit_memory(struct si_shader_context
*ctx
,
4043 struct lp_build_emit_data
*emit_data
) {
4044 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
4045 LLVMBuilderRef builder
= gallivm
->builder
;
4046 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
4047 LLVMValueRef ptr
, result
, arg
;
4049 ptr
= get_memory_ptr(ctx
, inst
, ctx
->i32
, 1);
4051 arg
= lp_build_emit_fetch(&ctx
->radeon_bld
.soa
.bld_base
, inst
, 2, 0);
4052 arg
= LLVMBuildBitCast(builder
, arg
, ctx
->i32
, "");
4054 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_ATOMCAS
) {
4055 LLVMValueRef new_data
;
4056 new_data
= lp_build_emit_fetch(&ctx
->radeon_bld
.soa
.bld_base
,
4059 new_data
= LLVMBuildBitCast(builder
, new_data
, ctx
->i32
, "");
4061 #if HAVE_LLVM >= 0x309
4062 result
= LLVMBuildAtomicCmpXchg(builder
, ptr
, arg
, new_data
,
4063 LLVMAtomicOrderingSequentiallyConsistent
,
4064 LLVMAtomicOrderingSequentiallyConsistent
,
4068 result
= LLVMBuildExtractValue(builder
, result
, 0, "");
4070 LLVMAtomicRMWBinOp op
;
4072 switch(inst
->Instruction
.Opcode
) {
4073 case TGSI_OPCODE_ATOMUADD
:
4074 op
= LLVMAtomicRMWBinOpAdd
;
4076 case TGSI_OPCODE_ATOMXCHG
:
4077 op
= LLVMAtomicRMWBinOpXchg
;
4079 case TGSI_OPCODE_ATOMAND
:
4080 op
= LLVMAtomicRMWBinOpAnd
;
4082 case TGSI_OPCODE_ATOMOR
:
4083 op
= LLVMAtomicRMWBinOpOr
;
4085 case TGSI_OPCODE_ATOMXOR
:
4086 op
= LLVMAtomicRMWBinOpXor
;
4088 case TGSI_OPCODE_ATOMUMIN
:
4089 op
= LLVMAtomicRMWBinOpUMin
;
4091 case TGSI_OPCODE_ATOMUMAX
:
4092 op
= LLVMAtomicRMWBinOpUMax
;
4094 case TGSI_OPCODE_ATOMIMIN
:
4095 op
= LLVMAtomicRMWBinOpMin
;
4097 case TGSI_OPCODE_ATOMIMAX
:
4098 op
= LLVMAtomicRMWBinOpMax
;
4101 unreachable("unknown atomic opcode");
4104 result
= LLVMBuildAtomicRMW(builder
, op
, ptr
, arg
,
4105 LLVMAtomicOrderingSequentiallyConsistent
,
4108 emit_data
->output
[emit_data
->chan
] = LLVMBuildBitCast(builder
, result
, emit_data
->dst_type
, "");
4111 static void atomic_emit(
4112 const struct lp_build_tgsi_action
*action
,
4113 struct lp_build_tgsi_context
*bld_base
,
4114 struct lp_build_emit_data
*emit_data
)
4116 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4117 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4118 LLVMBuilderRef builder
= gallivm
->builder
;
4119 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
4120 char intrinsic_name
[40];
4123 if (inst
->Src
[0].Register
.File
== TGSI_FILE_MEMORY
) {
4124 atomic_emit_memory(ctx
, emit_data
);
4128 if (inst
->Src
[0].Register
.File
== TGSI_FILE_BUFFER
||
4129 inst
->Memory
.Texture
== TGSI_TEXTURE_BUFFER
) {
4130 snprintf(intrinsic_name
, sizeof(intrinsic_name
),
4131 "llvm.amdgcn.buffer.atomic.%s", action
->intr_name
);
4133 char coords_type
[8];
4135 build_int_type_name(LLVMTypeOf(emit_data
->args
[1]),
4136 coords_type
, sizeof(coords_type
));
4137 snprintf(intrinsic_name
, sizeof(intrinsic_name
),
4138 "llvm.amdgcn.image.atomic.%s.%s",
4139 action
->intr_name
, coords_type
);
4142 tmp
= lp_build_intrinsic(
4143 builder
, intrinsic_name
, bld_base
->uint_bld
.elem_type
,
4144 emit_data
->args
, emit_data
->arg_count
, 0);
4145 emit_data
->output
[emit_data
->chan
] =
4146 LLVMBuildBitCast(builder
, tmp
, bld_base
->base
.elem_type
, "");
4149 static void resq_fetch_args(
4150 struct lp_build_tgsi_context
* bld_base
,
4151 struct lp_build_emit_data
* emit_data
)
4153 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4154 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4155 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
4156 const struct tgsi_full_src_register
*reg
= &inst
->Src
[0];
4158 emit_data
->dst_type
= ctx
->v4i32
;
4160 if (reg
->Register
.File
== TGSI_FILE_BUFFER
) {
4161 emit_data
->args
[0] = shader_buffer_fetch_rsrc(ctx
, reg
);
4162 emit_data
->arg_count
= 1;
4163 } else if (inst
->Memory
.Texture
== TGSI_TEXTURE_BUFFER
) {
4164 image_fetch_rsrc(bld_base
, reg
, false, &emit_data
->args
[0]);
4165 emit_data
->arg_count
= 1;
4167 emit_data
->args
[0] = bld_base
->uint_bld
.zero
; /* mip level */
4168 image_fetch_rsrc(bld_base
, reg
, false, &emit_data
->args
[1]);
4169 emit_data
->args
[2] = lp_build_const_int32(gallivm
, 15); /* dmask */
4170 emit_data
->args
[3] = bld_base
->uint_bld
.zero
; /* unorm */
4171 emit_data
->args
[4] = bld_base
->uint_bld
.zero
; /* r128 */
4172 emit_data
->args
[5] = tgsi_is_array_image(inst
->Memory
.Texture
) ?
4173 bld_base
->uint_bld
.one
: bld_base
->uint_bld
.zero
; /* da */
4174 emit_data
->args
[6] = bld_base
->uint_bld
.zero
; /* glc */
4175 emit_data
->args
[7] = bld_base
->uint_bld
.zero
; /* slc */
4176 emit_data
->args
[8] = bld_base
->uint_bld
.zero
; /* tfe */
4177 emit_data
->args
[9] = bld_base
->uint_bld
.zero
; /* lwe */
4178 emit_data
->arg_count
= 10;
4182 static void resq_emit(
4183 const struct lp_build_tgsi_action
*action
,
4184 struct lp_build_tgsi_context
*bld_base
,
4185 struct lp_build_emit_data
*emit_data
)
4187 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4188 LLVMBuilderRef builder
= gallivm
->builder
;
4189 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
4192 if (inst
->Src
[0].Register
.File
== TGSI_FILE_BUFFER
) {
4193 out
= LLVMBuildExtractElement(builder
, emit_data
->args
[0],
4194 lp_build_const_int32(gallivm
, 2), "");
4195 } else if (inst
->Memory
.Texture
== TGSI_TEXTURE_BUFFER
) {
4196 out
= get_buffer_size(bld_base
, emit_data
->args
[0]);
4198 out
= lp_build_intrinsic(
4199 builder
, "llvm.SI.getresinfo.i32", emit_data
->dst_type
,
4200 emit_data
->args
, emit_data
->arg_count
,
4201 LLVMReadNoneAttribute
);
4203 /* Divide the number of layers by 6 to get the number of cubes. */
4204 if (inst
->Memory
.Texture
== TGSI_TEXTURE_CUBE_ARRAY
) {
4205 LLVMValueRef imm2
= lp_build_const_int32(gallivm
, 2);
4206 LLVMValueRef imm6
= lp_build_const_int32(gallivm
, 6);
4208 LLVMValueRef z
= LLVMBuildExtractElement(builder
, out
, imm2
, "");
4209 z
= LLVMBuildSDiv(builder
, z
, imm6
, "");
4210 out
= LLVMBuildInsertElement(builder
, out
, z
, imm2
, "");
4214 emit_data
->output
[emit_data
->chan
] = out
;
4217 static void set_tex_fetch_args(struct si_shader_context
*ctx
,
4218 struct lp_build_emit_data
*emit_data
,
4219 unsigned opcode
, unsigned target
,
4220 LLVMValueRef res_ptr
, LLVMValueRef samp_ptr
,
4221 LLVMValueRef
*param
, unsigned count
,
4224 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
4226 unsigned is_rect
= target
== TGSI_TEXTURE_RECT
;
4228 /* Pad to power of two vector */
4229 while (count
< util_next_power_of_two(count
))
4230 param
[count
++] = LLVMGetUndef(ctx
->i32
);
4232 /* Texture coordinates. */
4234 emit_data
->args
[0] = lp_build_gather_values(gallivm
, param
, count
);
4236 emit_data
->args
[0] = param
[0];
4239 emit_data
->args
[1] = res_ptr
;
4242 if (opcode
== TGSI_OPCODE_TXF
|| opcode
== TGSI_OPCODE_TXQ
)
4243 emit_data
->dst_type
= ctx
->v4i32
;
4245 emit_data
->dst_type
= ctx
->v4f32
;
4247 emit_data
->args
[num_args
++] = samp_ptr
;
4250 emit_data
->args
[num_args
++] = lp_build_const_int32(gallivm
, dmask
);
4251 emit_data
->args
[num_args
++] = lp_build_const_int32(gallivm
, is_rect
); /* unorm */
4252 emit_data
->args
[num_args
++] = lp_build_const_int32(gallivm
, 0); /* r128 */
4253 emit_data
->args
[num_args
++] = lp_build_const_int32(gallivm
,
4254 tgsi_is_array_sampler(target
)); /* da */
4255 emit_data
->args
[num_args
++] = lp_build_const_int32(gallivm
, 0); /* glc */
4256 emit_data
->args
[num_args
++] = lp_build_const_int32(gallivm
, 0); /* slc */
4257 emit_data
->args
[num_args
++] = lp_build_const_int32(gallivm
, 0); /* tfe */
4258 emit_data
->args
[num_args
++] = lp_build_const_int32(gallivm
, 0); /* lwe */
4260 emit_data
->arg_count
= num_args
;
4263 static const struct lp_build_tgsi_action tex_action
;
4271 static LLVMTypeRef
const_array(LLVMTypeRef elem_type
, int num_elements
)
4273 return LLVMPointerType(LLVMArrayType(elem_type
, num_elements
),
4278 * Load an image view, fmask view. or sampler state descriptor.
4280 static LLVMValueRef
load_sampler_desc_custom(struct si_shader_context
*ctx
,
4281 LLVMValueRef list
, LLVMValueRef index
,
4282 enum desc_type type
)
4284 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
4285 LLVMBuilderRef builder
= gallivm
->builder
;
4289 /* The image is at [0:7]. */
4290 index
= LLVMBuildMul(builder
, index
, LLVMConstInt(ctx
->i32
, 2, 0), "");
4293 /* The FMASK is at [8:15]. */
4294 index
= LLVMBuildMul(builder
, index
, LLVMConstInt(ctx
->i32
, 2, 0), "");
4295 index
= LLVMBuildAdd(builder
, index
, LLVMConstInt(ctx
->i32
, 1, 0), "");
4298 /* The sampler state is at [12:15]. */
4299 index
= LLVMBuildMul(builder
, index
, LLVMConstInt(ctx
->i32
, 4, 0), "");
4300 index
= LLVMBuildAdd(builder
, index
, LLVMConstInt(ctx
->i32
, 3, 0), "");
4301 list
= LLVMBuildPointerCast(builder
, list
,
4302 const_array(ctx
->v4i32
, 0), "");
4306 return build_indexed_load_const(ctx
, list
, index
);
4309 static LLVMValueRef
load_sampler_desc(struct si_shader_context
*ctx
,
4310 LLVMValueRef index
, enum desc_type type
)
4312 LLVMValueRef list
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
4315 return load_sampler_desc_custom(ctx
, list
, index
, type
);
4318 /* Disable anisotropic filtering if BASE_LEVEL == LAST_LEVEL.
4321 * If BASE_LEVEL == LAST_LEVEL, the shader must disable anisotropic
4322 * filtering manually. The driver sets img7 to a mask clearing
4323 * MAX_ANISO_RATIO if BASE_LEVEL == LAST_LEVEL. The shader must do:
4324 * s_and_b32 samp0, samp0, img7
4327 * The ANISO_OVERRIDE sampler field enables this fix in TA.
4329 static LLVMValueRef
sici_fix_sampler_aniso(struct si_shader_context
*ctx
,
4330 LLVMValueRef res
, LLVMValueRef samp
)
4332 LLVMBuilderRef builder
= ctx
->radeon_bld
.gallivm
.builder
;
4333 LLVMValueRef img7
, samp0
;
4335 if (ctx
->screen
->b
.chip_class
>= VI
)
4338 img7
= LLVMBuildExtractElement(builder
, res
,
4339 LLVMConstInt(ctx
->i32
, 7, 0), "");
4340 samp0
= LLVMBuildExtractElement(builder
, samp
,
4341 LLVMConstInt(ctx
->i32
, 0, 0), "");
4342 samp0
= LLVMBuildAnd(builder
, samp0
, img7
, "");
4343 return LLVMBuildInsertElement(builder
, samp
, samp0
,
4344 LLVMConstInt(ctx
->i32
, 0, 0), "");
4347 static void tex_fetch_ptrs(
4348 struct lp_build_tgsi_context
*bld_base
,
4349 struct lp_build_emit_data
*emit_data
,
4350 LLVMValueRef
*res_ptr
, LLVMValueRef
*samp_ptr
, LLVMValueRef
*fmask_ptr
)
4352 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4353 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
4354 unsigned target
= inst
->Texture
.Texture
;
4355 unsigned sampler_src
;
4356 unsigned sampler_index
;
4359 sampler_src
= emit_data
->inst
->Instruction
.NumSrcRegs
- 1;
4360 sampler_index
= emit_data
->inst
->Src
[sampler_src
].Register
.Index
;
4362 if (emit_data
->inst
->Src
[sampler_src
].Register
.Indirect
) {
4363 const struct tgsi_full_src_register
*reg
= &emit_data
->inst
->Src
[sampler_src
];
4365 index
= get_bounded_indirect_index(ctx
,
4367 reg
->Register
.Index
,
4370 index
= LLVMConstInt(ctx
->i32
, sampler_index
, 0);
4373 *res_ptr
= load_sampler_desc(ctx
, index
, DESC_IMAGE
);
4375 if (target
== TGSI_TEXTURE_2D_MSAA
||
4376 target
== TGSI_TEXTURE_2D_ARRAY_MSAA
) {
4380 *fmask_ptr
= load_sampler_desc(ctx
, index
, DESC_FMASK
);
4383 *samp_ptr
= load_sampler_desc(ctx
, index
, DESC_SAMPLER
);
4384 *samp_ptr
= sici_fix_sampler_aniso(ctx
, *res_ptr
, *samp_ptr
);
4391 static void txq_fetch_args(
4392 struct lp_build_tgsi_context
*bld_base
,
4393 struct lp_build_emit_data
*emit_data
)
4395 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4396 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4397 LLVMBuilderRef builder
= gallivm
->builder
;
4398 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
4399 unsigned target
= inst
->Texture
.Texture
;
4400 LLVMValueRef res_ptr
;
4401 LLVMValueRef address
;
4403 tex_fetch_ptrs(bld_base
, emit_data
, &res_ptr
, NULL
, NULL
);
4405 if (target
== TGSI_TEXTURE_BUFFER
) {
4406 /* Read the size from the buffer descriptor directly. */
4407 LLVMValueRef res
= LLVMBuildBitCast(builder
, res_ptr
, ctx
->v8i32
, "");
4408 emit_data
->args
[0] = get_buffer_size(bld_base
, res
);
4412 /* Textures - set the mip level. */
4413 address
= lp_build_emit_fetch(bld_base
, inst
, 0, TGSI_CHAN_X
);
4415 set_tex_fetch_args(ctx
, emit_data
, TGSI_OPCODE_TXQ
, target
, res_ptr
,
4416 NULL
, &address
, 1, 0xf);
4419 static void txq_emit(const struct lp_build_tgsi_action
*action
,
4420 struct lp_build_tgsi_context
*bld_base
,
4421 struct lp_build_emit_data
*emit_data
)
4423 struct lp_build_context
*base
= &bld_base
->base
;
4424 unsigned target
= emit_data
->inst
->Texture
.Texture
;
4426 if (target
== TGSI_TEXTURE_BUFFER
) {
4427 /* Just return the buffer size. */
4428 emit_data
->output
[emit_data
->chan
] = emit_data
->args
[0];
4432 emit_data
->output
[emit_data
->chan
] = lp_build_intrinsic(
4433 base
->gallivm
->builder
, "llvm.SI.getresinfo.i32",
4434 emit_data
->dst_type
, emit_data
->args
, emit_data
->arg_count
,
4435 LLVMReadNoneAttribute
);
4437 /* Divide the number of layers by 6 to get the number of cubes. */
4438 if (target
== TGSI_TEXTURE_CUBE_ARRAY
||
4439 target
== TGSI_TEXTURE_SHADOWCUBE_ARRAY
) {
4440 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
4441 LLVMValueRef two
= lp_build_const_int32(bld_base
->base
.gallivm
, 2);
4442 LLVMValueRef six
= lp_build_const_int32(bld_base
->base
.gallivm
, 6);
4444 LLVMValueRef v4
= emit_data
->output
[emit_data
->chan
];
4445 LLVMValueRef z
= LLVMBuildExtractElement(builder
, v4
, two
, "");
4446 z
= LLVMBuildSDiv(builder
, z
, six
, "");
4448 emit_data
->output
[emit_data
->chan
] =
4449 LLVMBuildInsertElement(builder
, v4
, z
, two
, "");
4453 static void tex_fetch_args(
4454 struct lp_build_tgsi_context
*bld_base
,
4455 struct lp_build_emit_data
*emit_data
)
4457 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4458 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4459 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
4460 unsigned opcode
= inst
->Instruction
.Opcode
;
4461 unsigned target
= inst
->Texture
.Texture
;
4462 LLVMValueRef coords
[5], derivs
[6];
4463 LLVMValueRef address
[16];
4464 unsigned num_coords
= tgsi_util_get_texture_coord_dim(target
);
4465 int ref_pos
= tgsi_util_get_shadow_ref_src_index(target
);
4468 unsigned num_deriv_channels
= 0;
4469 bool has_offset
= inst
->Texture
.NumOffsets
> 0;
4470 LLVMValueRef res_ptr
, samp_ptr
, fmask_ptr
= NULL
;
4471 unsigned dmask
= 0xf;
4473 tex_fetch_ptrs(bld_base
, emit_data
, &res_ptr
, &samp_ptr
, &fmask_ptr
);
4475 if (target
== TGSI_TEXTURE_BUFFER
) {
4476 LLVMTypeRef v2i128
= LLVMVectorType(ctx
->i128
, 2);
4478 /* Bitcast and truncate v8i32 to v16i8. */
4479 LLVMValueRef res
= res_ptr
;
4480 res
= LLVMBuildBitCast(gallivm
->builder
, res
, v2i128
, "");
4481 res
= LLVMBuildExtractElement(gallivm
->builder
, res
, bld_base
->uint_bld
.one
, "");
4482 res
= LLVMBuildBitCast(gallivm
->builder
, res
, ctx
->v16i8
, "");
4484 emit_data
->dst_type
= ctx
->v4f32
;
4485 emit_data
->args
[0] = res
;
4486 emit_data
->args
[1] = bld_base
->uint_bld
.zero
;
4487 emit_data
->args
[2] = lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, TGSI_CHAN_X
);
4488 emit_data
->arg_count
= 3;
4492 /* Fetch and project texture coordinates */
4493 coords
[3] = lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, TGSI_CHAN_W
);
4494 for (chan
= 0; chan
< 3; chan
++ ) {
4495 coords
[chan
] = lp_build_emit_fetch(bld_base
,
4498 if (opcode
== TGSI_OPCODE_TXP
)
4499 coords
[chan
] = lp_build_emit_llvm_binary(bld_base
,
4505 if (opcode
== TGSI_OPCODE_TXP
)
4506 coords
[3] = bld_base
->base
.one
;
4509 if (has_offset
&& opcode
!= TGSI_OPCODE_TXF
) {
4510 /* The offsets are six-bit signed integers packed like this:
4511 * X=[5:0], Y=[13:8], and Z=[21:16].
4513 LLVMValueRef offset
[3], pack
;
4515 assert(inst
->Texture
.NumOffsets
== 1);
4517 for (chan
= 0; chan
< 3; chan
++) {
4518 offset
[chan
] = lp_build_emit_fetch_texoffset(bld_base
,
4519 emit_data
->inst
, 0, chan
);
4520 offset
[chan
] = LLVMBuildAnd(gallivm
->builder
, offset
[chan
],
4521 lp_build_const_int32(gallivm
, 0x3f), "");
4523 offset
[chan
] = LLVMBuildShl(gallivm
->builder
, offset
[chan
],
4524 lp_build_const_int32(gallivm
, chan
*8), "");
4527 pack
= LLVMBuildOr(gallivm
->builder
, offset
[0], offset
[1], "");
4528 pack
= LLVMBuildOr(gallivm
->builder
, pack
, offset
[2], "");
4529 address
[count
++] = pack
;
4532 /* Pack LOD bias value */
4533 if (opcode
== TGSI_OPCODE_TXB
)
4534 address
[count
++] = coords
[3];
4535 if (opcode
== TGSI_OPCODE_TXB2
)
4536 address
[count
++] = lp_build_emit_fetch(bld_base
, inst
, 1, TGSI_CHAN_X
);
4538 /* Pack depth comparison value */
4539 if (tgsi_is_shadow_target(target
) && opcode
!= TGSI_OPCODE_LODQ
) {
4540 if (target
== TGSI_TEXTURE_SHADOWCUBE_ARRAY
) {
4541 address
[count
++] = lp_build_emit_fetch(bld_base
, inst
, 1, TGSI_CHAN_X
);
4543 assert(ref_pos
>= 0);
4544 address
[count
++] = coords
[ref_pos
];
4548 /* Pack user derivatives */
4549 if (opcode
== TGSI_OPCODE_TXD
) {
4550 int param
, num_src_deriv_channels
;
4553 case TGSI_TEXTURE_3D
:
4554 num_src_deriv_channels
= 3;
4555 num_deriv_channels
= 3;
4557 case TGSI_TEXTURE_2D
:
4558 case TGSI_TEXTURE_SHADOW2D
:
4559 case TGSI_TEXTURE_RECT
:
4560 case TGSI_TEXTURE_SHADOWRECT
:
4561 case TGSI_TEXTURE_2D_ARRAY
:
4562 case TGSI_TEXTURE_SHADOW2D_ARRAY
:
4563 num_src_deriv_channels
= 2;
4564 num_deriv_channels
= 2;
4566 case TGSI_TEXTURE_CUBE
:
4567 case TGSI_TEXTURE_SHADOWCUBE
:
4568 case TGSI_TEXTURE_CUBE_ARRAY
:
4569 case TGSI_TEXTURE_SHADOWCUBE_ARRAY
:
4570 /* Cube derivatives will be converted to 2D. */
4571 num_src_deriv_channels
= 3;
4572 num_deriv_channels
= 2;
4574 case TGSI_TEXTURE_1D
:
4575 case TGSI_TEXTURE_SHADOW1D
:
4576 case TGSI_TEXTURE_1D_ARRAY
:
4577 case TGSI_TEXTURE_SHADOW1D_ARRAY
:
4578 num_src_deriv_channels
= 1;
4579 num_deriv_channels
= 1;
4582 unreachable("invalid target");
4585 for (param
= 0; param
< 2; param
++)
4586 for (chan
= 0; chan
< num_src_deriv_channels
; chan
++)
4587 derivs
[param
* num_src_deriv_channels
+ chan
] =
4588 lp_build_emit_fetch(bld_base
, inst
, param
+1, chan
);
4591 if (target
== TGSI_TEXTURE_CUBE
||
4592 target
== TGSI_TEXTURE_CUBE_ARRAY
||
4593 target
== TGSI_TEXTURE_SHADOWCUBE
||
4594 target
== TGSI_TEXTURE_SHADOWCUBE_ARRAY
)
4595 radeon_llvm_emit_prepare_cube_coords(bld_base
, emit_data
, coords
, derivs
);
4597 if (opcode
== TGSI_OPCODE_TXD
)
4598 for (int i
= 0; i
< num_deriv_channels
* 2; i
++)
4599 address
[count
++] = derivs
[i
];
4601 /* Pack texture coordinates */
4602 address
[count
++] = coords
[0];
4604 address
[count
++] = coords
[1];
4606 address
[count
++] = coords
[2];
4608 /* Pack LOD or sample index */
4609 if (opcode
== TGSI_OPCODE_TXL
|| opcode
== TGSI_OPCODE_TXF
)
4610 address
[count
++] = coords
[3];
4611 else if (opcode
== TGSI_OPCODE_TXL2
)
4612 address
[count
++] = lp_build_emit_fetch(bld_base
, inst
, 1, TGSI_CHAN_X
);
4615 assert(!"Cannot handle more than 16 texture address parameters");
4619 for (chan
= 0; chan
< count
; chan
++ ) {
4620 address
[chan
] = LLVMBuildBitCast(gallivm
->builder
,
4621 address
[chan
], ctx
->i32
, "");
4624 /* Adjust the sample index according to FMASK.
4626 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
4627 * which is the identity mapping. Each nibble says which physical sample
4628 * should be fetched to get that sample.
4630 * For example, 0x11111100 means there are only 2 samples stored and
4631 * the second sample covers 3/4 of the pixel. When reading samples 0
4632 * and 1, return physical sample 0 (determined by the first two 0s
4633 * in FMASK), otherwise return physical sample 1.
4635 * The sample index should be adjusted as follows:
4636 * sample_index = (fmask >> (sample_index * 4)) & 0xF;
4638 if (target
== TGSI_TEXTURE_2D_MSAA
||
4639 target
== TGSI_TEXTURE_2D_ARRAY_MSAA
) {
4640 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
4641 struct lp_build_emit_data txf_emit_data
= *emit_data
;
4642 LLVMValueRef txf_address
[4];
4643 unsigned txf_count
= count
;
4644 struct tgsi_full_instruction inst
= {};
4646 memcpy(txf_address
, address
, sizeof(txf_address
));
4648 if (target
== TGSI_TEXTURE_2D_MSAA
) {
4649 txf_address
[2] = bld_base
->uint_bld
.zero
;
4651 txf_address
[3] = bld_base
->uint_bld
.zero
;
4653 /* Read FMASK using TXF. */
4654 inst
.Instruction
.Opcode
= TGSI_OPCODE_TXF
;
4655 inst
.Texture
.Texture
= target
;
4656 txf_emit_data
.inst
= &inst
;
4657 txf_emit_data
.chan
= 0;
4658 set_tex_fetch_args(ctx
, &txf_emit_data
, TGSI_OPCODE_TXF
,
4659 target
, fmask_ptr
, NULL
,
4660 txf_address
, txf_count
, 0xf);
4661 build_tex_intrinsic(&tex_action
, bld_base
, &txf_emit_data
);
4663 /* Initialize some constants. */
4664 LLVMValueRef four
= LLVMConstInt(ctx
->i32
, 4, 0);
4665 LLVMValueRef F
= LLVMConstInt(ctx
->i32
, 0xF, 0);
4667 /* Apply the formula. */
4668 LLVMValueRef fmask
=
4669 LLVMBuildExtractElement(gallivm
->builder
,
4670 txf_emit_data
.output
[0],
4671 uint_bld
->zero
, "");
4673 unsigned sample_chan
= target
== TGSI_TEXTURE_2D_MSAA
? 2 : 3;
4675 LLVMValueRef sample_index4
=
4676 LLVMBuildMul(gallivm
->builder
, address
[sample_chan
], four
, "");
4678 LLVMValueRef shifted_fmask
=
4679 LLVMBuildLShr(gallivm
->builder
, fmask
, sample_index4
, "");
4681 LLVMValueRef final_sample
=
4682 LLVMBuildAnd(gallivm
->builder
, shifted_fmask
, F
, "");
4684 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
4685 * resource descriptor is 0 (invalid),
4687 LLVMValueRef fmask_desc
=
4688 LLVMBuildBitCast(gallivm
->builder
, fmask_ptr
,
4691 LLVMValueRef fmask_word1
=
4692 LLVMBuildExtractElement(gallivm
->builder
, fmask_desc
,
4695 LLVMValueRef word1_is_nonzero
=
4696 LLVMBuildICmp(gallivm
->builder
, LLVMIntNE
,
4697 fmask_word1
, uint_bld
->zero
, "");
4699 /* Replace the MSAA sample index. */
4700 address
[sample_chan
] =
4701 LLVMBuildSelect(gallivm
->builder
, word1_is_nonzero
,
4702 final_sample
, address
[sample_chan
], "");
4705 if (opcode
== TGSI_OPCODE_TXF
) {
4706 /* add tex offsets */
4707 if (inst
->Texture
.NumOffsets
) {
4708 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
4709 struct lp_build_tgsi_soa_context
*bld
= lp_soa_context(bld_base
);
4710 const struct tgsi_texture_offset
*off
= inst
->TexOffsets
;
4712 assert(inst
->Texture
.NumOffsets
== 1);
4715 case TGSI_TEXTURE_3D
:
4716 address
[2] = lp_build_add(uint_bld
, address
[2],
4717 bld
->immediates
[off
->Index
][off
->SwizzleZ
]);
4719 case TGSI_TEXTURE_2D
:
4720 case TGSI_TEXTURE_SHADOW2D
:
4721 case TGSI_TEXTURE_RECT
:
4722 case TGSI_TEXTURE_SHADOWRECT
:
4723 case TGSI_TEXTURE_2D_ARRAY
:
4724 case TGSI_TEXTURE_SHADOW2D_ARRAY
:
4726 lp_build_add(uint_bld
, address
[1],
4727 bld
->immediates
[off
->Index
][off
->SwizzleY
]);
4729 case TGSI_TEXTURE_1D
:
4730 case TGSI_TEXTURE_SHADOW1D
:
4731 case TGSI_TEXTURE_1D_ARRAY
:
4732 case TGSI_TEXTURE_SHADOW1D_ARRAY
:
4734 lp_build_add(uint_bld
, address
[0],
4735 bld
->immediates
[off
->Index
][off
->SwizzleX
]);
4737 /* texture offsets do not apply to other texture targets */
4742 if (opcode
== TGSI_OPCODE_TG4
) {
4743 unsigned gather_comp
= 0;
4745 /* DMASK was repurposed for GATHER4. 4 components are always
4746 * returned and DMASK works like a swizzle - it selects
4747 * the component to fetch. The only valid DMASK values are
4748 * 1=red, 2=green, 4=blue, 8=alpha. (e.g. 1 returns
4749 * (red,red,red,red) etc.) The ISA document doesn't mention
4753 /* Get the component index from src1.x for Gather4. */
4754 if (!tgsi_is_shadow_target(target
)) {
4755 LLVMValueRef (*imms
)[4] = lp_soa_context(bld_base
)->immediates
;
4756 LLVMValueRef comp_imm
;
4757 struct tgsi_src_register src1
= inst
->Src
[1].Register
;
4759 assert(src1
.File
== TGSI_FILE_IMMEDIATE
);
4761 comp_imm
= imms
[src1
.Index
][src1
.SwizzleX
];
4762 gather_comp
= LLVMConstIntGetZExtValue(comp_imm
);
4763 gather_comp
= CLAMP(gather_comp
, 0, 3);
4766 dmask
= 1 << gather_comp
;
4769 set_tex_fetch_args(ctx
, emit_data
, opcode
, target
, res_ptr
,
4770 samp_ptr
, address
, count
, dmask
);
4773 /* Gather4 should follow the same rules as bilinear filtering, but the hardware
4774 * incorrectly forces nearest filtering if the texture format is integer.
4775 * The only effect it has on Gather4, which always returns 4 texels for
4776 * bilinear filtering, is that the final coordinates are off by 0.5 of
4779 * The workaround is to subtract 0.5 from the unnormalized coordinates,
4780 * or (0.5 / size) from the normalized coordinates.
4782 static void si_lower_gather4_integer(struct si_shader_context
*ctx
,
4783 struct lp_build_emit_data
*emit_data
,
4784 const char *intr_name
,
4785 unsigned coord_vgpr_index
)
4787 LLVMBuilderRef builder
= ctx
->radeon_bld
.gallivm
.builder
;
4788 LLVMValueRef coord
= emit_data
->args
[0];
4789 LLVMValueRef half_texel
[2];
4792 if (emit_data
->inst
->Texture
.Texture
== TGSI_TEXTURE_RECT
||
4793 emit_data
->inst
->Texture
.Texture
== TGSI_TEXTURE_SHADOWRECT
) {
4794 half_texel
[0] = half_texel
[1] = LLVMConstReal(ctx
->f32
, -0.5);
4796 struct tgsi_full_instruction txq_inst
= {};
4797 struct lp_build_emit_data txq_emit_data
= {};
4799 /* Query the texture size. */
4800 txq_inst
.Texture
.Texture
= emit_data
->inst
->Texture
.Texture
;
4801 txq_emit_data
.inst
= &txq_inst
;
4802 txq_emit_data
.dst_type
= ctx
->v4i32
;
4803 set_tex_fetch_args(ctx
, &txq_emit_data
, TGSI_OPCODE_TXQ
,
4804 txq_inst
.Texture
.Texture
,
4805 emit_data
->args
[1], NULL
,
4806 &ctx
->radeon_bld
.soa
.bld_base
.uint_bld
.zero
,
4808 txq_emit(NULL
, &ctx
->radeon_bld
.soa
.bld_base
, &txq_emit_data
);
4810 /* Compute -0.5 / size. */
4811 for (c
= 0; c
< 2; c
++) {
4813 LLVMBuildExtractElement(builder
, txq_emit_data
.output
[0],
4814 LLVMConstInt(ctx
->i32
, c
, 0), "");
4815 half_texel
[c
] = LLVMBuildUIToFP(builder
, half_texel
[c
], ctx
->f32
, "");
4817 lp_build_emit_llvm_unary(&ctx
->radeon_bld
.soa
.bld_base
,
4818 TGSI_OPCODE_RCP
, half_texel
[c
]);
4819 half_texel
[c
] = LLVMBuildFMul(builder
, half_texel
[c
],
4820 LLVMConstReal(ctx
->f32
, -0.5), "");
4824 for (c
= 0; c
< 2; c
++) {
4826 LLVMValueRef index
= LLVMConstInt(ctx
->i32
, coord_vgpr_index
+ c
, 0);
4828 tmp
= LLVMBuildExtractElement(builder
, coord
, index
, "");
4829 tmp
= LLVMBuildBitCast(builder
, tmp
, ctx
->f32
, "");
4830 tmp
= LLVMBuildFAdd(builder
, tmp
, half_texel
[c
], "");
4831 tmp
= LLVMBuildBitCast(builder
, tmp
, ctx
->i32
, "");
4832 coord
= LLVMBuildInsertElement(builder
, coord
, tmp
, index
, "");
4835 emit_data
->args
[0] = coord
;
4836 emit_data
->output
[emit_data
->chan
] =
4837 lp_build_intrinsic(builder
, intr_name
, emit_data
->dst_type
,
4838 emit_data
->args
, emit_data
->arg_count
,
4839 LLVMReadNoneAttribute
);
4842 static void build_tex_intrinsic(const struct lp_build_tgsi_action
*action
,
4843 struct lp_build_tgsi_context
*bld_base
,
4844 struct lp_build_emit_data
*emit_data
)
4846 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4847 struct lp_build_context
*base
= &bld_base
->base
;
4848 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
4849 unsigned opcode
= inst
->Instruction
.Opcode
;
4850 unsigned target
= inst
->Texture
.Texture
;
4851 char intr_name
[127];
4852 bool has_offset
= inst
->Texture
.NumOffsets
> 0;
4853 bool is_shadow
= tgsi_is_shadow_target(target
);
4855 const char *name
= "llvm.SI.image.sample";
4856 const char *infix
= "";
4858 if (target
== TGSI_TEXTURE_BUFFER
) {
4859 emit_data
->output
[emit_data
->chan
] = lp_build_intrinsic(
4860 base
->gallivm
->builder
,
4861 "llvm.SI.vs.load.input", emit_data
->dst_type
,
4862 emit_data
->args
, emit_data
->arg_count
,
4863 LLVMReadNoneAttribute
);
4868 case TGSI_OPCODE_TXF
:
4869 name
= target
== TGSI_TEXTURE_2D_MSAA
||
4870 target
== TGSI_TEXTURE_2D_ARRAY_MSAA
?
4871 "llvm.SI.image.load" :
4872 "llvm.SI.image.load.mip";
4876 case TGSI_OPCODE_LODQ
:
4877 name
= "llvm.SI.getlod";
4881 case TGSI_OPCODE_TEX
:
4882 case TGSI_OPCODE_TEX2
:
4883 case TGSI_OPCODE_TXP
:
4884 if (ctx
->type
!= PIPE_SHADER_FRAGMENT
)
4887 case TGSI_OPCODE_TXB
:
4888 case TGSI_OPCODE_TXB2
:
4889 assert(ctx
->type
== PIPE_SHADER_FRAGMENT
);
4892 case TGSI_OPCODE_TXL
:
4893 case TGSI_OPCODE_TXL2
:
4896 case TGSI_OPCODE_TXD
:
4899 case TGSI_OPCODE_TG4
:
4900 name
= "llvm.SI.gather4";
4908 /* Add the type and suffixes .c, .o if needed. */
4909 build_int_type_name(LLVMTypeOf(emit_data
->args
[0]), type
, sizeof(type
));
4910 sprintf(intr_name
, "%s%s%s%s.%s",
4911 name
, is_shadow
? ".c" : "", infix
,
4912 has_offset
? ".o" : "", type
);
4914 /* The hardware needs special lowering for Gather4 with integer formats. */
4915 if (opcode
== TGSI_OPCODE_TG4
) {
4916 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
4917 /* This will also work with non-constant indexing because of how
4918 * glsl_to_tgsi works and we intent to preserve that behavior.
4920 const unsigned src_idx
= 2;
4921 unsigned sampler
= inst
->Src
[src_idx
].Register
.Index
;
4923 assert(inst
->Src
[src_idx
].Register
.File
== TGSI_FILE_SAMPLER
);
4925 if (info
->sampler_type
[sampler
] == TGSI_RETURN_TYPE_SINT
||
4926 info
->sampler_type
[sampler
] == TGSI_RETURN_TYPE_UINT
) {
4927 /* Texture coordinates start after:
4928 * {offset, bias, z-compare, derivatives}
4929 * Only the offset and z-compare can occur here.
4931 si_lower_gather4_integer(ctx
, emit_data
, intr_name
,
4932 (int)has_offset
+ (int)is_shadow
);
4937 emit_data
->output
[emit_data
->chan
] = lp_build_intrinsic(
4938 base
->gallivm
->builder
, intr_name
, emit_data
->dst_type
,
4939 emit_data
->args
, emit_data
->arg_count
,
4940 LLVMReadNoneAttribute
);
4943 static void si_llvm_emit_txqs(
4944 const struct lp_build_tgsi_action
*action
,
4945 struct lp_build_tgsi_context
*bld_base
,
4946 struct lp_build_emit_data
*emit_data
)
4948 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4949 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4950 LLVMBuilderRef builder
= gallivm
->builder
;
4951 LLVMValueRef res
, samples
;
4952 LLVMValueRef res_ptr
, samp_ptr
, fmask_ptr
= NULL
;
4954 tex_fetch_ptrs(bld_base
, emit_data
, &res_ptr
, &samp_ptr
, &fmask_ptr
);
4957 /* Read the samples from the descriptor directly. */
4958 res
= LLVMBuildBitCast(builder
, res_ptr
, ctx
->v8i32
, "");
4959 samples
= LLVMBuildExtractElement(
4961 lp_build_const_int32(gallivm
, 3), "");
4962 samples
= LLVMBuildLShr(builder
, samples
,
4963 lp_build_const_int32(gallivm
, 16), "");
4964 samples
= LLVMBuildAnd(builder
, samples
,
4965 lp_build_const_int32(gallivm
, 0xf), "");
4966 samples
= LLVMBuildShl(builder
, lp_build_const_int32(gallivm
, 1),
4969 emit_data
->output
[emit_data
->chan
] = samples
;
4973 * SI implements derivatives using the local data store (LDS)
4974 * All writes to the LDS happen in all executing threads at
4975 * the same time. TID is the Thread ID for the current
4976 * thread and is a value between 0 and 63, representing
4977 * the thread's position in the wavefront.
4979 * For the pixel shader threads are grouped into quads of four pixels.
4980 * The TIDs of the pixels of a quad are:
4988 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
4989 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
4990 * the current pixel's column, and masking with 0xfffffffe yields the TID
4991 * of the left pixel of the current pixel's row.
4993 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
4994 * adding 2 yields the TID of the pixel below the top pixel.
4996 /* masks for thread ID. */
4997 #define TID_MASK_TOP_LEFT 0xfffffffc
4998 #define TID_MASK_TOP 0xfffffffd
4999 #define TID_MASK_LEFT 0xfffffffe
5001 static void si_llvm_emit_ddxy(
5002 const struct lp_build_tgsi_action
*action
,
5003 struct lp_build_tgsi_context
*bld_base
,
5004 struct lp_build_emit_data
*emit_data
)
5006 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
5007 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
5008 unsigned opcode
= emit_data
->info
->opcode
;
5009 LLVMValueRef thread_id
, tl
, trbl
, tl_tid
, trbl_tid
, val
, args
[2];
5012 bool has_ds_bpermute
= HAVE_LLVM
>= 0x0309 &&
5013 ctx
->screen
->b
.chip_class
>= VI
;
5015 thread_id
= get_thread_id(ctx
);
5017 if (opcode
== TGSI_OPCODE_DDX_FINE
)
5018 mask
= TID_MASK_LEFT
;
5019 else if (opcode
== TGSI_OPCODE_DDY_FINE
)
5020 mask
= TID_MASK_TOP
;
5022 mask
= TID_MASK_TOP_LEFT
;
5024 tl_tid
= LLVMBuildAnd(gallivm
->builder
, thread_id
,
5025 lp_build_const_int32(gallivm
, mask
), "");
5027 /* for DDX we want to next X pixel, DDY next Y pixel. */
5028 idx
= (opcode
== TGSI_OPCODE_DDX
|| opcode
== TGSI_OPCODE_DDX_FINE
) ? 1 : 2;
5029 trbl_tid
= LLVMBuildAdd(gallivm
->builder
, tl_tid
,
5030 lp_build_const_int32(gallivm
, idx
), "");
5032 val
= LLVMBuildBitCast(gallivm
->builder
, emit_data
->args
[0], ctx
->i32
, "");
5034 if (has_ds_bpermute
) {
5035 args
[0] = LLVMBuildMul(gallivm
->builder
, tl_tid
,
5036 lp_build_const_int32(gallivm
, 4), "");
5038 tl
= lp_build_intrinsic(gallivm
->builder
,
5039 "llvm.amdgcn.ds.bpermute", ctx
->i32
,
5040 args
, 2, LLVMReadNoneAttribute
);
5042 args
[0] = LLVMBuildMul(gallivm
->builder
, trbl_tid
,
5043 lp_build_const_int32(gallivm
, 4), "");
5044 trbl
= lp_build_intrinsic(gallivm
->builder
,
5045 "llvm.amdgcn.ds.bpermute", ctx
->i32
,
5046 args
, 2, LLVMReadNoneAttribute
);
5048 LLVMValueRef store_ptr
, load_ptr0
, load_ptr1
;
5050 store_ptr
= build_gep0(ctx
, ctx
->lds
, thread_id
);
5051 load_ptr0
= build_gep0(ctx
, ctx
->lds
, tl_tid
);
5052 load_ptr1
= build_gep0(ctx
, ctx
->lds
, trbl_tid
);
5054 LLVMBuildStore(gallivm
->builder
, val
, store_ptr
);
5055 tl
= LLVMBuildLoad(gallivm
->builder
, load_ptr0
, "");
5056 trbl
= LLVMBuildLoad(gallivm
->builder
, load_ptr1
, "");
5059 tl
= LLVMBuildBitCast(gallivm
->builder
, tl
, ctx
->f32
, "");
5060 trbl
= LLVMBuildBitCast(gallivm
->builder
, trbl
, ctx
->f32
, "");
5062 emit_data
->output
[emit_data
->chan
] =
5063 LLVMBuildFSub(gallivm
->builder
, trbl
, tl
, "");
5067 * this takes an I,J coordinate pair,
5068 * and works out the X and Y derivatives.
5069 * it returns DDX(I), DDX(J), DDY(I), DDY(J).
5071 static LLVMValueRef
si_llvm_emit_ddxy_interp(
5072 struct lp_build_tgsi_context
*bld_base
,
5073 LLVMValueRef interp_ij
)
5075 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
5076 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
5077 LLVMValueRef store_ptr
, load_ptr_x
, load_ptr_y
, load_ptr_ddx
, load_ptr_ddy
, temp
, temp2
;
5078 LLVMValueRef tl
, tr
, bl
, result
[4], thread_id
;
5081 thread_id
= get_thread_id(ctx
);
5082 store_ptr
= build_gep0(ctx
, ctx
->lds
, thread_id
);
5084 temp
= LLVMBuildAnd(gallivm
->builder
, thread_id
,
5085 lp_build_const_int32(gallivm
, TID_MASK_LEFT
), "");
5087 temp2
= LLVMBuildAnd(gallivm
->builder
, thread_id
,
5088 lp_build_const_int32(gallivm
, TID_MASK_TOP
), "");
5090 load_ptr_x
= build_gep0(ctx
, ctx
->lds
, temp
);
5092 load_ptr_y
= build_gep0(ctx
, ctx
->lds
, temp2
);
5094 load_ptr_ddx
= build_gep0(ctx
, ctx
->lds
,
5095 LLVMBuildAdd(gallivm
->builder
, temp
,
5096 lp_build_const_int32(gallivm
, 1), ""));
5098 load_ptr_ddy
= build_gep0(ctx
, ctx
->lds
,
5099 LLVMBuildAdd(gallivm
->builder
, temp2
,
5100 lp_build_const_int32(gallivm
, 2), ""));
5102 for (c
= 0; c
< 2; ++c
) {
5103 LLVMValueRef store_val
;
5104 LLVMValueRef c_ll
= lp_build_const_int32(gallivm
, c
);
5106 store_val
= LLVMBuildExtractElement(gallivm
->builder
,
5107 interp_ij
, c_ll
, "");
5108 LLVMBuildStore(gallivm
->builder
,
5112 tl
= LLVMBuildLoad(gallivm
->builder
, load_ptr_x
, "");
5113 tl
= LLVMBuildBitCast(gallivm
->builder
, tl
, ctx
->f32
, "");
5115 tr
= LLVMBuildLoad(gallivm
->builder
, load_ptr_ddx
, "");
5116 tr
= LLVMBuildBitCast(gallivm
->builder
, tr
, ctx
->f32
, "");
5118 result
[c
] = LLVMBuildFSub(gallivm
->builder
, tr
, tl
, "");
5120 tl
= LLVMBuildLoad(gallivm
->builder
, load_ptr_y
, "");
5121 tl
= LLVMBuildBitCast(gallivm
->builder
, tl
, ctx
->f32
, "");
5123 bl
= LLVMBuildLoad(gallivm
->builder
, load_ptr_ddy
, "");
5124 bl
= LLVMBuildBitCast(gallivm
->builder
, bl
, ctx
->f32
, "");
5126 result
[c
+ 2] = LLVMBuildFSub(gallivm
->builder
, bl
, tl
, "");
5129 return lp_build_gather_values(gallivm
, result
, 4);
5132 static void interp_fetch_args(
5133 struct lp_build_tgsi_context
*bld_base
,
5134 struct lp_build_emit_data
*emit_data
)
5136 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
5137 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
5138 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
5140 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_OFFSET
) {
5141 /* offset is in second src, first two channels */
5142 emit_data
->args
[0] = lp_build_emit_fetch(bld_base
,
5145 emit_data
->args
[1] = lp_build_emit_fetch(bld_base
,
5148 emit_data
->arg_count
= 2;
5149 } else if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_SAMPLE
) {
5150 LLVMValueRef sample_position
;
5151 LLVMValueRef sample_id
;
5152 LLVMValueRef halfval
= lp_build_const_float(gallivm
, 0.5f
);
5154 /* fetch sample ID, then fetch its sample position,
5155 * and place into first two channels.
5157 sample_id
= lp_build_emit_fetch(bld_base
,
5158 emit_data
->inst
, 1, TGSI_CHAN_X
);
5159 sample_id
= LLVMBuildBitCast(gallivm
->builder
, sample_id
,
5161 sample_position
= load_sample_position(&ctx
->radeon_bld
, sample_id
);
5163 emit_data
->args
[0] = LLVMBuildExtractElement(gallivm
->builder
,
5165 lp_build_const_int32(gallivm
, 0), "");
5167 emit_data
->args
[0] = LLVMBuildFSub(gallivm
->builder
, emit_data
->args
[0], halfval
, "");
5168 emit_data
->args
[1] = LLVMBuildExtractElement(gallivm
->builder
,
5170 lp_build_const_int32(gallivm
, 1), "");
5171 emit_data
->args
[1] = LLVMBuildFSub(gallivm
->builder
, emit_data
->args
[1], halfval
, "");
5172 emit_data
->arg_count
= 2;
5176 static void build_interp_intrinsic(const struct lp_build_tgsi_action
*action
,
5177 struct lp_build_tgsi_context
*bld_base
,
5178 struct lp_build_emit_data
*emit_data
)
5180 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
5181 struct si_shader
*shader
= ctx
->shader
;
5182 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
5183 LLVMValueRef interp_param
;
5184 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
5185 const char *intr_name
;
5186 int input_index
= inst
->Src
[0].Register
.Index
;
5189 LLVMValueRef attr_number
;
5190 LLVMValueRef params
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_PRIM_MASK
);
5191 int interp_param_idx
;
5192 unsigned interp
= shader
->selector
->info
.input_interpolate
[input_index
];
5195 assert(inst
->Src
[0].Register
.File
== TGSI_FILE_INPUT
);
5197 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_OFFSET
||
5198 inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_SAMPLE
)
5199 location
= TGSI_INTERPOLATE_LOC_CENTER
;
5201 location
= TGSI_INTERPOLATE_LOC_CENTROID
;
5203 interp_param_idx
= lookup_interp_param_index(interp
, location
);
5204 if (interp_param_idx
== -1)
5206 else if (interp_param_idx
)
5207 interp_param
= get_interp_param(ctx
, interp_param_idx
);
5209 interp_param
= NULL
;
5211 attr_number
= lp_build_const_int32(gallivm
, input_index
);
5213 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_OFFSET
||
5214 inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_SAMPLE
) {
5215 LLVMValueRef ij_out
[2];
5216 LLVMValueRef ddxy_out
= si_llvm_emit_ddxy_interp(bld_base
, interp_param
);
5219 * take the I then J parameters, and the DDX/Y for it, and
5220 * calculate the IJ inputs for the interpolator.
5221 * temp1 = ddx * offset/sample.x + I;
5222 * interp_param.I = ddy * offset/sample.y + temp1;
5223 * temp1 = ddx * offset/sample.x + J;
5224 * interp_param.J = ddy * offset/sample.y + temp1;
5226 for (i
= 0; i
< 2; i
++) {
5227 LLVMValueRef ix_ll
= lp_build_const_int32(gallivm
, i
);
5228 LLVMValueRef iy_ll
= lp_build_const_int32(gallivm
, i
+ 2);
5229 LLVMValueRef ddx_el
= LLVMBuildExtractElement(gallivm
->builder
,
5230 ddxy_out
, ix_ll
, "");
5231 LLVMValueRef ddy_el
= LLVMBuildExtractElement(gallivm
->builder
,
5232 ddxy_out
, iy_ll
, "");
5233 LLVMValueRef interp_el
= LLVMBuildExtractElement(gallivm
->builder
,
5234 interp_param
, ix_ll
, "");
5235 LLVMValueRef temp1
, temp2
;
5237 interp_el
= LLVMBuildBitCast(gallivm
->builder
, interp_el
,
5240 temp1
= LLVMBuildFMul(gallivm
->builder
, ddx_el
, emit_data
->args
[0], "");
5242 temp1
= LLVMBuildFAdd(gallivm
->builder
, temp1
, interp_el
, "");
5244 temp2
= LLVMBuildFMul(gallivm
->builder
, ddy_el
, emit_data
->args
[1], "");
5246 temp2
= LLVMBuildFAdd(gallivm
->builder
, temp2
, temp1
, "");
5248 ij_out
[i
] = LLVMBuildBitCast(gallivm
->builder
,
5249 temp2
, ctx
->i32
, "");
5251 interp_param
= lp_build_gather_values(bld_base
->base
.gallivm
, ij_out
, 2);
5254 intr_name
= interp_param
? "llvm.SI.fs.interp" : "llvm.SI.fs.constant";
5255 for (chan
= 0; chan
< 2; chan
++) {
5256 LLVMValueRef args
[4];
5257 LLVMValueRef llvm_chan
;
5260 schan
= tgsi_util_get_full_src_register_swizzle(&inst
->Src
[0], chan
);
5261 llvm_chan
= lp_build_const_int32(gallivm
, schan
);
5263 args
[0] = llvm_chan
;
5264 args
[1] = attr_number
;
5266 args
[3] = interp_param
;
5268 emit_data
->output
[chan
] =
5269 lp_build_intrinsic(gallivm
->builder
, intr_name
,
5270 ctx
->f32
, args
, args
[3] ? 4 : 3,
5271 LLVMReadNoneAttribute
);
5275 static unsigned si_llvm_get_stream(struct lp_build_tgsi_context
*bld_base
,
5276 struct lp_build_emit_data
*emit_data
)
5278 LLVMValueRef (*imms
)[4] = lp_soa_context(bld_base
)->immediates
;
5279 struct tgsi_src_register src0
= emit_data
->inst
->Src
[0].Register
;
5282 assert(src0
.File
== TGSI_FILE_IMMEDIATE
);
5284 stream
= LLVMConstIntGetZExtValue(imms
[src0
.Index
][src0
.SwizzleX
]) & 0x3;
5288 /* Emit one vertex from the geometry shader */
5289 static void si_llvm_emit_vertex(
5290 const struct lp_build_tgsi_action
*action
,
5291 struct lp_build_tgsi_context
*bld_base
,
5292 struct lp_build_emit_data
*emit_data
)
5294 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
5295 struct lp_build_context
*uint
= &bld_base
->uint_bld
;
5296 struct si_shader
*shader
= ctx
->shader
;
5297 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
5298 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
5299 LLVMValueRef soffset
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
5300 SI_PARAM_GS2VS_OFFSET
);
5301 LLVMValueRef gs_next_vertex
;
5302 LLVMValueRef can_emit
, kill
;
5303 LLVMValueRef args
[2];
5308 stream
= si_llvm_get_stream(bld_base
, emit_data
);
5310 /* Write vertex attribute values to GSVS ring */
5311 gs_next_vertex
= LLVMBuildLoad(gallivm
->builder
,
5312 ctx
->gs_next_vertex
[stream
],
5315 /* If this thread has already emitted the declared maximum number of
5316 * vertices, kill it: excessive vertex emissions are not supposed to
5317 * have any effect, and GS threads have no externally observable
5318 * effects other than emitting vertices.
5320 can_emit
= LLVMBuildICmp(gallivm
->builder
, LLVMIntULE
, gs_next_vertex
,
5321 lp_build_const_int32(gallivm
,
5322 shader
->selector
->gs_max_out_vertices
), "");
5323 kill
= lp_build_select(&bld_base
->base
, can_emit
,
5324 lp_build_const_float(gallivm
, 1.0f
),
5325 lp_build_const_float(gallivm
, -1.0f
));
5327 lp_build_intrinsic(gallivm
->builder
, "llvm.AMDGPU.kill",
5328 ctx
->voidt
, &kill
, 1, 0);
5330 for (i
= 0; i
< info
->num_outputs
; i
++) {
5331 LLVMValueRef
*out_ptr
=
5332 ctx
->radeon_bld
.soa
.outputs
[i
];
5334 for (chan
= 0; chan
< 4; chan
++) {
5335 LLVMValueRef out_val
= LLVMBuildLoad(gallivm
->builder
, out_ptr
[chan
], "");
5336 LLVMValueRef voffset
=
5337 lp_build_const_int32(gallivm
, (i
* 4 + chan
) *
5338 shader
->selector
->gs_max_out_vertices
);
5340 voffset
= lp_build_add(uint
, voffset
, gs_next_vertex
);
5341 voffset
= lp_build_mul_imm(uint
, voffset
, 4);
5343 out_val
= LLVMBuildBitCast(gallivm
->builder
, out_val
, ctx
->i32
, "");
5345 build_tbuffer_store(ctx
,
5346 ctx
->gsvs_ring
[stream
],
5348 voffset
, soffset
, 0,
5349 V_008F0C_BUF_DATA_FORMAT_32
,
5350 V_008F0C_BUF_NUM_FORMAT_UINT
,
5354 gs_next_vertex
= lp_build_add(uint
, gs_next_vertex
,
5355 lp_build_const_int32(gallivm
, 1));
5357 LLVMBuildStore(gallivm
->builder
, gs_next_vertex
, ctx
->gs_next_vertex
[stream
]);
5359 /* Signal vertex emission */
5360 args
[0] = lp_build_const_int32(gallivm
, SENDMSG_GS_OP_EMIT
| SENDMSG_GS
| (stream
<< 8));
5361 args
[1] = LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_GS_WAVE_ID
);
5362 lp_build_intrinsic(gallivm
->builder
, "llvm.SI.sendmsg",
5363 ctx
->voidt
, args
, 2, 0);
5366 /* Cut one primitive from the geometry shader */
5367 static void si_llvm_emit_primitive(
5368 const struct lp_build_tgsi_action
*action
,
5369 struct lp_build_tgsi_context
*bld_base
,
5370 struct lp_build_emit_data
*emit_data
)
5372 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
5373 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
5374 LLVMValueRef args
[2];
5377 /* Signal primitive cut */
5378 stream
= si_llvm_get_stream(bld_base
, emit_data
);
5379 args
[0] = lp_build_const_int32(gallivm
, SENDMSG_GS_OP_CUT
| SENDMSG_GS
| (stream
<< 8));
5380 args
[1] = LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_GS_WAVE_ID
);
5381 lp_build_intrinsic(gallivm
->builder
, "llvm.SI.sendmsg",
5382 ctx
->voidt
, args
, 2, 0);
5385 static void si_llvm_emit_barrier(const struct lp_build_tgsi_action
*action
,
5386 struct lp_build_tgsi_context
*bld_base
,
5387 struct lp_build_emit_data
*emit_data
)
5389 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
5390 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
5392 /* The real barrier instruction isn’t needed, because an entire patch
5393 * always fits into a single wave.
5395 if (ctx
->type
== PIPE_SHADER_TESS_CTRL
) {
5396 emit_optimization_barrier(ctx
);
5400 lp_build_intrinsic(gallivm
->builder
,
5401 HAVE_LLVM
>= 0x0309 ? "llvm.amdgcn.s.barrier"
5402 : "llvm.AMDGPU.barrier.local",
5403 ctx
->voidt
, NULL
, 0, 0);
5406 static const struct lp_build_tgsi_action tex_action
= {
5407 .fetch_args
= tex_fetch_args
,
5408 .emit
= build_tex_intrinsic
,
5411 static const struct lp_build_tgsi_action interp_action
= {
5412 .fetch_args
= interp_fetch_args
,
5413 .emit
= build_interp_intrinsic
,
5416 static void si_create_function(struct si_shader_context
*ctx
,
5417 LLVMTypeRef
*returns
, unsigned num_returns
,
5418 LLVMTypeRef
*params
, unsigned num_params
,
5423 radeon_llvm_create_func(&ctx
->radeon_bld
, returns
, num_returns
,
5424 params
, num_params
);
5425 radeon_llvm_shader_type(ctx
->radeon_bld
.main_fn
, ctx
->type
);
5426 ctx
->return_value
= LLVMGetUndef(ctx
->radeon_bld
.return_type
);
5428 for (i
= 0; i
<= last_sgpr
; ++i
) {
5429 LLVMValueRef P
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, i
);
5431 /* The combination of:
5435 * allows the optimization passes to move loads and reduces
5436 * SGPR spilling significantly.
5438 if (LLVMGetTypeKind(LLVMTypeOf(P
)) == LLVMPointerTypeKind
) {
5439 LLVMAddAttribute(P
, LLVMByValAttribute
);
5440 lp_add_attr_dereferenceable(P
, UINT64_MAX
);
5442 LLVMAddAttribute(P
, LLVMInRegAttribute
);
5445 if (ctx
->screen
->b
.debug_flags
& DBG_UNSAFE_MATH
) {
5446 /* These were copied from some LLVM test. */
5447 LLVMAddTargetDependentFunctionAttr(ctx
->radeon_bld
.main_fn
,
5448 "less-precise-fpmad",
5450 LLVMAddTargetDependentFunctionAttr(ctx
->radeon_bld
.main_fn
,
5453 LLVMAddTargetDependentFunctionAttr(ctx
->radeon_bld
.main_fn
,
5456 LLVMAddTargetDependentFunctionAttr(ctx
->radeon_bld
.main_fn
,
5462 static void create_meta_data(struct si_shader_context
*ctx
)
5464 struct gallivm_state
*gallivm
= ctx
->radeon_bld
.soa
.bld_base
.base
.gallivm
;
5466 ctx
->invariant_load_md_kind
= LLVMGetMDKindIDInContext(gallivm
->context
,
5467 "invariant.load", 14);
5468 ctx
->range_md_kind
= LLVMGetMDKindIDInContext(gallivm
->context
,
5470 ctx
->uniform_md_kind
= LLVMGetMDKindIDInContext(gallivm
->context
,
5471 "amdgpu.uniform", 14);
5473 ctx
->empty_md
= LLVMMDNodeInContext(gallivm
->context
, NULL
, 0);
5476 static void declare_streamout_params(struct si_shader_context
*ctx
,
5477 struct pipe_stream_output_info
*so
,
5478 LLVMTypeRef
*params
, LLVMTypeRef i32
,
5479 unsigned *num_params
)
5483 /* Streamout SGPRs. */
5484 if (so
->num_outputs
) {
5485 if (ctx
->type
!= PIPE_SHADER_TESS_EVAL
)
5486 params
[ctx
->param_streamout_config
= (*num_params
)++] = i32
;
5488 ctx
->param_streamout_config
= ctx
->param_tess_offchip
;
5490 params
[ctx
->param_streamout_write_index
= (*num_params
)++] = i32
;
5492 /* A streamout buffer offset is loaded if the stride is non-zero. */
5493 for (i
= 0; i
< 4; i
++) {
5497 params
[ctx
->param_streamout_offset
[i
] = (*num_params
)++] = i32
;
5501 static unsigned llvm_get_type_size(LLVMTypeRef type
)
5503 LLVMTypeKind kind
= LLVMGetTypeKind(type
);
5506 case LLVMIntegerTypeKind
:
5507 return LLVMGetIntTypeWidth(type
) / 8;
5508 case LLVMFloatTypeKind
:
5510 case LLVMPointerTypeKind
:
5512 case LLVMVectorTypeKind
:
5513 return LLVMGetVectorSize(type
) *
5514 llvm_get_type_size(LLVMGetElementType(type
));
5521 static void declare_tess_lds(struct si_shader_context
*ctx
)
5523 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
5524 struct lp_build_tgsi_context
*bld_base
= &ctx
->radeon_bld
.soa
.bld_base
;
5525 struct lp_build_context
*uint
= &bld_base
->uint_bld
;
5527 unsigned lds_size
= ctx
->screen
->b
.chip_class
>= CIK
? 65536 : 32768;
5528 ctx
->lds
= LLVMBuildIntToPtr(gallivm
->builder
, uint
->zero
,
5529 LLVMPointerType(LLVMArrayType(ctx
->i32
, lds_size
/ 4), LOCAL_ADDR_SPACE
),
5533 static void create_function(struct si_shader_context
*ctx
)
5535 struct lp_build_tgsi_context
*bld_base
= &ctx
->radeon_bld
.soa
.bld_base
;
5536 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
5537 struct si_shader
*shader
= ctx
->shader
;
5538 LLVMTypeRef params
[SI_NUM_PARAMS
+ SI_NUM_VERTEX_BUFFERS
], v3i32
;
5539 LLVMTypeRef returns
[16+32*4];
5540 unsigned i
, last_sgpr
, num_params
, num_return_sgprs
;
5541 unsigned num_returns
= 0;
5543 v3i32
= LLVMVectorType(ctx
->i32
, 3);
5545 params
[SI_PARAM_RW_BUFFERS
] = const_array(ctx
->v16i8
, SI_NUM_RW_BUFFERS
);
5546 params
[SI_PARAM_CONST_BUFFERS
] = const_array(ctx
->v16i8
, SI_NUM_CONST_BUFFERS
);
5547 params
[SI_PARAM_SAMPLERS
] = const_array(ctx
->v8i32
, SI_NUM_SAMPLERS
);
5548 params
[SI_PARAM_IMAGES
] = const_array(ctx
->v8i32
, SI_NUM_IMAGES
);
5549 params
[SI_PARAM_SHADER_BUFFERS
] = const_array(ctx
->v4i32
, SI_NUM_SHADER_BUFFERS
);
5551 switch (ctx
->type
) {
5552 case PIPE_SHADER_VERTEX
:
5553 params
[SI_PARAM_VERTEX_BUFFERS
] = const_array(ctx
->v16i8
, SI_NUM_VERTEX_BUFFERS
);
5554 params
[SI_PARAM_BASE_VERTEX
] = ctx
->i32
;
5555 params
[SI_PARAM_START_INSTANCE
] = ctx
->i32
;
5556 params
[SI_PARAM_DRAWID
] = ctx
->i32
;
5557 num_params
= SI_PARAM_DRAWID
+1;
5559 if (shader
->key
.vs
.as_es
) {
5560 params
[ctx
->param_es2gs_offset
= num_params
++] = ctx
->i32
;
5561 } else if (shader
->key
.vs
.as_ls
) {
5562 params
[SI_PARAM_LS_OUT_LAYOUT
] = ctx
->i32
;
5563 num_params
= SI_PARAM_LS_OUT_LAYOUT
+1;
5565 if (ctx
->is_gs_copy_shader
) {
5566 num_params
= SI_PARAM_RW_BUFFERS
+1;
5568 params
[SI_PARAM_VS_STATE_BITS
] = ctx
->i32
;
5569 num_params
= SI_PARAM_VS_STATE_BITS
+1;
5572 /* The locations of the other parameters are assigned dynamically. */
5573 declare_streamout_params(ctx
, &shader
->selector
->so
,
5574 params
, ctx
->i32
, &num_params
);
5577 last_sgpr
= num_params
-1;
5580 params
[ctx
->param_vertex_id
= num_params
++] = ctx
->i32
;
5581 params
[ctx
->param_rel_auto_id
= num_params
++] = ctx
->i32
;
5582 params
[ctx
->param_vs_prim_id
= num_params
++] = ctx
->i32
;
5583 params
[ctx
->param_instance_id
= num_params
++] = ctx
->i32
;
5585 if (!ctx
->is_monolithic
&&
5586 !ctx
->is_gs_copy_shader
) {
5587 /* Vertex load indices. */
5588 ctx
->param_vertex_index0
= num_params
;
5590 for (i
= 0; i
< shader
->selector
->info
.num_inputs
; i
++)
5591 params
[num_params
++] = ctx
->i32
;
5593 /* PrimitiveID output. */
5594 if (!shader
->key
.vs
.as_es
&& !shader
->key
.vs
.as_ls
)
5595 for (i
= 0; i
<= VS_EPILOG_PRIMID_LOC
; i
++)
5596 returns
[num_returns
++] = ctx
->f32
;
5600 case PIPE_SHADER_TESS_CTRL
:
5601 params
[SI_PARAM_TCS_OFFCHIP_LAYOUT
] = ctx
->i32
;
5602 params
[SI_PARAM_TCS_OUT_OFFSETS
] = ctx
->i32
;
5603 params
[SI_PARAM_TCS_OUT_LAYOUT
] = ctx
->i32
;
5604 params
[SI_PARAM_TCS_IN_LAYOUT
] = ctx
->i32
;
5605 params
[ctx
->param_oc_lds
= SI_PARAM_TCS_OC_LDS
] = ctx
->i32
;
5606 params
[SI_PARAM_TESS_FACTOR_OFFSET
] = ctx
->i32
;
5607 last_sgpr
= SI_PARAM_TESS_FACTOR_OFFSET
;
5610 params
[SI_PARAM_PATCH_ID
] = ctx
->i32
;
5611 params
[SI_PARAM_REL_IDS
] = ctx
->i32
;
5612 num_params
= SI_PARAM_REL_IDS
+1;
5614 if (!ctx
->is_monolithic
) {
5615 /* SI_PARAM_TCS_OC_LDS and PARAM_TESS_FACTOR_OFFSET are
5616 * placed after the user SGPRs.
5618 for (i
= 0; i
< SI_TCS_NUM_USER_SGPR
+ 2; i
++)
5619 returns
[num_returns
++] = ctx
->i32
; /* SGPRs */
5621 for (i
= 0; i
< 3; i
++)
5622 returns
[num_returns
++] = ctx
->f32
; /* VGPRs */
5626 case PIPE_SHADER_TESS_EVAL
:
5627 params
[SI_PARAM_TCS_OFFCHIP_LAYOUT
] = ctx
->i32
;
5628 num_params
= SI_PARAM_TCS_OFFCHIP_LAYOUT
+1;
5630 if (shader
->key
.tes
.as_es
) {
5631 params
[ctx
->param_oc_lds
= num_params
++] = ctx
->i32
;
5632 params
[ctx
->param_tess_offchip
= num_params
++] = ctx
->i32
;
5633 params
[ctx
->param_es2gs_offset
= num_params
++] = ctx
->i32
;
5635 params
[ctx
->param_tess_offchip
= num_params
++] = ctx
->i32
;
5636 declare_streamout_params(ctx
, &shader
->selector
->so
,
5637 params
, ctx
->i32
, &num_params
);
5638 params
[ctx
->param_oc_lds
= num_params
++] = ctx
->i32
;
5640 last_sgpr
= num_params
- 1;
5643 params
[ctx
->param_tes_u
= num_params
++] = ctx
->f32
;
5644 params
[ctx
->param_tes_v
= num_params
++] = ctx
->f32
;
5645 params
[ctx
->param_tes_rel_patch_id
= num_params
++] = ctx
->i32
;
5646 params
[ctx
->param_tes_patch_id
= num_params
++] = ctx
->i32
;
5648 /* PrimitiveID output. */
5649 if (!ctx
->is_monolithic
&& !shader
->key
.tes
.as_es
)
5650 for (i
= 0; i
<= VS_EPILOG_PRIMID_LOC
; i
++)
5651 returns
[num_returns
++] = ctx
->f32
;
5654 case PIPE_SHADER_GEOMETRY
:
5655 params
[SI_PARAM_GS2VS_OFFSET
] = ctx
->i32
;
5656 params
[SI_PARAM_GS_WAVE_ID
] = ctx
->i32
;
5657 last_sgpr
= SI_PARAM_GS_WAVE_ID
;
5660 params
[SI_PARAM_VTX0_OFFSET
] = ctx
->i32
;
5661 params
[SI_PARAM_VTX1_OFFSET
] = ctx
->i32
;
5662 params
[SI_PARAM_PRIMITIVE_ID
] = ctx
->i32
;
5663 params
[SI_PARAM_VTX2_OFFSET
] = ctx
->i32
;
5664 params
[SI_PARAM_VTX3_OFFSET
] = ctx
->i32
;
5665 params
[SI_PARAM_VTX4_OFFSET
] = ctx
->i32
;
5666 params
[SI_PARAM_VTX5_OFFSET
] = ctx
->i32
;
5667 params
[SI_PARAM_GS_INSTANCE_ID
] = ctx
->i32
;
5668 num_params
= SI_PARAM_GS_INSTANCE_ID
+1;
5671 case PIPE_SHADER_FRAGMENT
:
5672 params
[SI_PARAM_ALPHA_REF
] = ctx
->f32
;
5673 params
[SI_PARAM_PRIM_MASK
] = ctx
->i32
;
5674 last_sgpr
= SI_PARAM_PRIM_MASK
;
5675 params
[SI_PARAM_PERSP_SAMPLE
] = ctx
->v2i32
;
5676 params
[SI_PARAM_PERSP_CENTER
] = ctx
->v2i32
;
5677 params
[SI_PARAM_PERSP_CENTROID
] = ctx
->v2i32
;
5678 params
[SI_PARAM_PERSP_PULL_MODEL
] = v3i32
;
5679 params
[SI_PARAM_LINEAR_SAMPLE
] = ctx
->v2i32
;
5680 params
[SI_PARAM_LINEAR_CENTER
] = ctx
->v2i32
;
5681 params
[SI_PARAM_LINEAR_CENTROID
] = ctx
->v2i32
;
5682 params
[SI_PARAM_LINE_STIPPLE_TEX
] = ctx
->f32
;
5683 params
[SI_PARAM_POS_X_FLOAT
] = ctx
->f32
;
5684 params
[SI_PARAM_POS_Y_FLOAT
] = ctx
->f32
;
5685 params
[SI_PARAM_POS_Z_FLOAT
] = ctx
->f32
;
5686 params
[SI_PARAM_POS_W_FLOAT
] = ctx
->f32
;
5687 params
[SI_PARAM_FRONT_FACE
] = ctx
->i32
;
5688 params
[SI_PARAM_ANCILLARY
] = ctx
->i32
;
5689 params
[SI_PARAM_SAMPLE_COVERAGE
] = ctx
->f32
;
5690 params
[SI_PARAM_POS_FIXED_PT
] = ctx
->i32
;
5691 num_params
= SI_PARAM_POS_FIXED_PT
+1;
5693 if (!ctx
->is_monolithic
) {
5694 /* Color inputs from the prolog. */
5695 if (shader
->selector
->info
.colors_read
) {
5696 unsigned num_color_elements
=
5697 util_bitcount(shader
->selector
->info
.colors_read
);
5699 assert(num_params
+ num_color_elements
<= ARRAY_SIZE(params
));
5700 for (i
= 0; i
< num_color_elements
; i
++)
5701 params
[num_params
++] = ctx
->f32
;
5704 /* Outputs for the epilog. */
5705 num_return_sgprs
= SI_SGPR_ALPHA_REF
+ 1;
5708 util_bitcount(shader
->selector
->info
.colors_written
) * 4 +
5709 shader
->selector
->info
.writes_z
+
5710 shader
->selector
->info
.writes_stencil
+
5711 shader
->selector
->info
.writes_samplemask
+
5712 1 /* SampleMaskIn */;
5714 num_returns
= MAX2(num_returns
,
5716 PS_EPILOG_SAMPLEMASK_MIN_LOC
+ 1);
5718 for (i
= 0; i
< num_return_sgprs
; i
++)
5719 returns
[i
] = ctx
->i32
;
5720 for (; i
< num_returns
; i
++)
5721 returns
[i
] = ctx
->f32
;
5725 case PIPE_SHADER_COMPUTE
:
5726 params
[SI_PARAM_GRID_SIZE
] = v3i32
;
5727 params
[SI_PARAM_BLOCK_ID
] = v3i32
;
5728 last_sgpr
= SI_PARAM_BLOCK_ID
;
5730 params
[SI_PARAM_THREAD_ID
] = v3i32
;
5731 num_params
= SI_PARAM_THREAD_ID
+ 1;
5734 assert(0 && "unimplemented shader");
5738 assert(num_params
<= ARRAY_SIZE(params
));
5740 si_create_function(ctx
, returns
, num_returns
, params
,
5741 num_params
, last_sgpr
);
5743 /* Reserve register locations for VGPR inputs the PS prolog may need. */
5744 if (ctx
->type
== PIPE_SHADER_FRAGMENT
&&
5745 !ctx
->is_monolithic
) {
5746 radeon_llvm_add_attribute(ctx
->radeon_bld
.main_fn
,
5747 "InitialPSInputAddr",
5748 S_0286D0_PERSP_SAMPLE_ENA(1) |
5749 S_0286D0_PERSP_CENTER_ENA(1) |
5750 S_0286D0_PERSP_CENTROID_ENA(1) |
5751 S_0286D0_LINEAR_SAMPLE_ENA(1) |
5752 S_0286D0_LINEAR_CENTER_ENA(1) |
5753 S_0286D0_LINEAR_CENTROID_ENA(1) |
5754 S_0286D0_FRONT_FACE_ENA(1) |
5755 S_0286D0_POS_FIXED_PT_ENA(1));
5756 } else if (ctx
->type
== PIPE_SHADER_COMPUTE
) {
5757 const unsigned *properties
= shader
->selector
->info
.properties
;
5758 unsigned max_work_group_size
=
5759 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
] *
5760 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT
] *
5761 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH
];
5763 assert(max_work_group_size
);
5765 radeon_llvm_add_attribute(ctx
->radeon_bld
.main_fn
,
5766 "amdgpu-max-work-group-size",
5767 max_work_group_size
);
5770 shader
->info
.num_input_sgprs
= 0;
5771 shader
->info
.num_input_vgprs
= 0;
5773 for (i
= 0; i
<= last_sgpr
; ++i
)
5774 shader
->info
.num_input_sgprs
+= llvm_get_type_size(params
[i
]) / 4;
5776 /* Unused fragment shader inputs are eliminated by the compiler,
5777 * so we don't know yet how many there will be.
5779 if (ctx
->type
!= PIPE_SHADER_FRAGMENT
)
5780 for (; i
< num_params
; ++i
)
5781 shader
->info
.num_input_vgprs
+= llvm_get_type_size(params
[i
]) / 4;
5783 if (bld_base
->info
&&
5784 (bld_base
->info
->opcode_count
[TGSI_OPCODE_DDX
] > 0 ||
5785 bld_base
->info
->opcode_count
[TGSI_OPCODE_DDY
] > 0 ||
5786 bld_base
->info
->opcode_count
[TGSI_OPCODE_DDX_FINE
] > 0 ||
5787 bld_base
->info
->opcode_count
[TGSI_OPCODE_DDY_FINE
] > 0 ||
5788 bld_base
->info
->opcode_count
[TGSI_OPCODE_INTERP_OFFSET
] > 0 ||
5789 bld_base
->info
->opcode_count
[TGSI_OPCODE_INTERP_SAMPLE
] > 0))
5791 LLVMAddGlobalInAddressSpace(gallivm
->module
,
5792 LLVMArrayType(ctx
->i32
, 64),
5796 if ((ctx
->type
== PIPE_SHADER_VERTEX
&& shader
->key
.vs
.as_ls
) ||
5797 ctx
->type
== PIPE_SHADER_TESS_CTRL
||
5798 ctx
->type
== PIPE_SHADER_TESS_EVAL
)
5799 declare_tess_lds(ctx
);
5803 * Load ESGS and GSVS ring buffer resource descriptors and save the variables
5806 static void preload_ring_buffers(struct si_shader_context
*ctx
)
5808 struct gallivm_state
*gallivm
=
5809 ctx
->radeon_bld
.soa
.bld_base
.base
.gallivm
;
5811 LLVMValueRef buf_ptr
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
5812 SI_PARAM_RW_BUFFERS
);
5814 if ((ctx
->type
== PIPE_SHADER_VERTEX
&&
5815 ctx
->shader
->key
.vs
.as_es
) ||
5816 (ctx
->type
== PIPE_SHADER_TESS_EVAL
&&
5817 ctx
->shader
->key
.tes
.as_es
) ||
5818 ctx
->type
== PIPE_SHADER_GEOMETRY
) {
5820 ctx
->type
== PIPE_SHADER_GEOMETRY
? SI_GS_RING_ESGS
5822 LLVMValueRef offset
= lp_build_const_int32(gallivm
, ring
);
5825 build_indexed_load_const(ctx
, buf_ptr
, offset
);
5828 if (ctx
->is_gs_copy_shader
) {
5829 LLVMValueRef offset
= lp_build_const_int32(gallivm
, SI_VS_RING_GSVS
);
5832 build_indexed_load_const(ctx
, buf_ptr
, offset
);
5834 if (ctx
->type
== PIPE_SHADER_GEOMETRY
) {
5836 for (i
= 0; i
< 4; i
++) {
5837 LLVMValueRef offset
= lp_build_const_int32(gallivm
, SI_GS_RING_GSVS0
+ i
);
5840 build_indexed_load_const(ctx
, buf_ptr
, offset
);
5845 static void si_llvm_emit_polygon_stipple(struct si_shader_context
*ctx
,
5846 LLVMValueRef param_rw_buffers
,
5847 unsigned param_pos_fixed_pt
)
5849 struct lp_build_tgsi_context
*bld_base
=
5850 &ctx
->radeon_bld
.soa
.bld_base
;
5851 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
5852 LLVMBuilderRef builder
= gallivm
->builder
;
5853 LLVMValueRef slot
, desc
, offset
, row
, bit
, address
[2];
5855 /* Use the fixed-point gl_FragCoord input.
5856 * Since the stipple pattern is 32x32 and it repeats, just get 5 bits
5857 * per coordinate to get the repeating effect.
5859 address
[0] = unpack_param(ctx
, param_pos_fixed_pt
, 0, 5);
5860 address
[1] = unpack_param(ctx
, param_pos_fixed_pt
, 16, 5);
5862 /* Load the buffer descriptor. */
5863 slot
= lp_build_const_int32(gallivm
, SI_PS_CONST_POLY_STIPPLE
);
5864 desc
= build_indexed_load_const(ctx
, param_rw_buffers
, slot
);
5866 /* The stipple pattern is 32x32, each row has 32 bits. */
5867 offset
= LLVMBuildMul(builder
, address
[1],
5868 LLVMConstInt(ctx
->i32
, 4, 0), "");
5869 row
= buffer_load_const(ctx
, desc
, offset
);
5870 row
= LLVMBuildBitCast(builder
, row
, ctx
->i32
, "");
5871 bit
= LLVMBuildLShr(builder
, row
, address
[0], "");
5872 bit
= LLVMBuildTrunc(builder
, bit
, ctx
->i1
, "");
5874 /* The intrinsic kills the thread if arg < 0. */
5875 bit
= LLVMBuildSelect(builder
, bit
, LLVMConstReal(ctx
->f32
, 0),
5876 LLVMConstReal(ctx
->f32
, -1), "");
5877 lp_build_intrinsic(builder
, "llvm.AMDGPU.kill", ctx
->voidt
, &bit
, 1, 0);
5880 void si_shader_binary_read_config(struct radeon_shader_binary
*binary
,
5881 struct si_shader_config
*conf
,
5882 unsigned symbol_offset
)
5885 const unsigned char *config
=
5886 radeon_shader_binary_config_start(binary
, symbol_offset
);
5887 bool really_needs_scratch
= false;
5889 /* LLVM adds SGPR spills to the scratch size.
5890 * Find out if we really need the scratch buffer.
5892 for (i
= 0; i
< binary
->reloc_count
; i
++) {
5893 const struct radeon_shader_reloc
*reloc
= &binary
->relocs
[i
];
5895 if (!strcmp(scratch_rsrc_dword0_symbol
, reloc
->name
) ||
5896 !strcmp(scratch_rsrc_dword1_symbol
, reloc
->name
)) {
5897 really_needs_scratch
= true;
5902 /* XXX: We may be able to emit some of these values directly rather than
5903 * extracting fields to be emitted later.
5906 for (i
= 0; i
< binary
->config_size_per_symbol
; i
+= 8) {
5907 unsigned reg
= util_le32_to_cpu(*(uint32_t*)(config
+ i
));
5908 unsigned value
= util_le32_to_cpu(*(uint32_t*)(config
+ i
+ 4));
5910 case R_00B028_SPI_SHADER_PGM_RSRC1_PS
:
5911 case R_00B128_SPI_SHADER_PGM_RSRC1_VS
:
5912 case R_00B228_SPI_SHADER_PGM_RSRC1_GS
:
5913 case R_00B848_COMPUTE_PGM_RSRC1
:
5914 conf
->num_sgprs
= MAX2(conf
->num_sgprs
, (G_00B028_SGPRS(value
) + 1) * 8);
5915 conf
->num_vgprs
= MAX2(conf
->num_vgprs
, (G_00B028_VGPRS(value
) + 1) * 4);
5916 conf
->float_mode
= G_00B028_FLOAT_MODE(value
);
5917 conf
->rsrc1
= value
;
5919 case R_00B02C_SPI_SHADER_PGM_RSRC2_PS
:
5920 conf
->lds_size
= MAX2(conf
->lds_size
, G_00B02C_EXTRA_LDS_SIZE(value
));
5922 case R_00B84C_COMPUTE_PGM_RSRC2
:
5923 conf
->lds_size
= MAX2(conf
->lds_size
, G_00B84C_LDS_SIZE(value
));
5924 conf
->rsrc2
= value
;
5926 case R_0286CC_SPI_PS_INPUT_ENA
:
5927 conf
->spi_ps_input_ena
= value
;
5929 case R_0286D0_SPI_PS_INPUT_ADDR
:
5930 conf
->spi_ps_input_addr
= value
;
5932 case R_0286E8_SPI_TMPRING_SIZE
:
5933 case R_00B860_COMPUTE_TMPRING_SIZE
:
5934 /* WAVESIZE is in units of 256 dwords. */
5935 if (really_needs_scratch
)
5936 conf
->scratch_bytes_per_wave
=
5937 G_00B860_WAVESIZE(value
) * 256 * 4;
5939 case 0x4: /* SPILLED_SGPRS */
5940 conf
->spilled_sgprs
= value
;
5942 case 0x8: /* SPILLED_VGPRS */
5943 conf
->spilled_vgprs
= value
;
5947 static bool printed
;
5950 fprintf(stderr
, "Warning: LLVM emitted unknown "
5951 "config register: 0x%x\n", reg
);
5959 if (!conf
->spi_ps_input_addr
)
5960 conf
->spi_ps_input_addr
= conf
->spi_ps_input_ena
;
5963 void si_shader_apply_scratch_relocs(struct si_context
*sctx
,
5964 struct si_shader
*shader
,
5965 struct si_shader_config
*config
,
5966 uint64_t scratch_va
)
5969 uint32_t scratch_rsrc_dword0
= scratch_va
;
5970 uint32_t scratch_rsrc_dword1
=
5971 S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32);
5973 /* Enable scratch coalescing if LLVM sets ELEMENT_SIZE & INDEX_STRIDE
5976 if (HAVE_LLVM
>= 0x0309)
5977 scratch_rsrc_dword1
|= S_008F04_SWIZZLE_ENABLE(1);
5979 scratch_rsrc_dword1
|=
5980 S_008F04_STRIDE(config
->scratch_bytes_per_wave
/ 64);
5982 for (i
= 0 ; i
< shader
->binary
.reloc_count
; i
++) {
5983 const struct radeon_shader_reloc
*reloc
=
5984 &shader
->binary
.relocs
[i
];
5985 if (!strcmp(scratch_rsrc_dword0_symbol
, reloc
->name
)) {
5986 util_memcpy_cpu_to_le32(shader
->binary
.code
+ reloc
->offset
,
5987 &scratch_rsrc_dword0
, 4);
5988 } else if (!strcmp(scratch_rsrc_dword1_symbol
, reloc
->name
)) {
5989 util_memcpy_cpu_to_le32(shader
->binary
.code
+ reloc
->offset
,
5990 &scratch_rsrc_dword1
, 4);
5995 static unsigned si_get_shader_binary_size(struct si_shader
*shader
)
5997 unsigned size
= shader
->binary
.code_size
;
6000 size
+= shader
->prolog
->binary
.code_size
;
6002 size
+= shader
->epilog
->binary
.code_size
;
6006 int si_shader_binary_upload(struct si_screen
*sscreen
, struct si_shader
*shader
)
6008 const struct radeon_shader_binary
*prolog
=
6009 shader
->prolog
? &shader
->prolog
->binary
: NULL
;
6010 const struct radeon_shader_binary
*epilog
=
6011 shader
->epilog
? &shader
->epilog
->binary
: NULL
;
6012 const struct radeon_shader_binary
*mainb
= &shader
->binary
;
6013 unsigned bo_size
= si_get_shader_binary_size(shader
) +
6014 (!epilog
? mainb
->rodata_size
: 0);
6017 assert(!prolog
|| !prolog
->rodata_size
);
6018 assert((!prolog
&& !epilog
) || !mainb
->rodata_size
);
6019 assert(!epilog
|| !epilog
->rodata_size
);
6021 r600_resource_reference(&shader
->bo
, NULL
);
6022 shader
->bo
= si_resource_create_custom(&sscreen
->b
.b
,
6023 PIPE_USAGE_IMMUTABLE
,
6029 ptr
= sscreen
->b
.ws
->buffer_map(shader
->bo
->buf
, NULL
,
6030 PIPE_TRANSFER_READ_WRITE
);
6033 util_memcpy_cpu_to_le32(ptr
, prolog
->code
, prolog
->code_size
);
6034 ptr
+= prolog
->code_size
;
6037 util_memcpy_cpu_to_le32(ptr
, mainb
->code
, mainb
->code_size
);
6038 ptr
+= mainb
->code_size
;
6041 util_memcpy_cpu_to_le32(ptr
, epilog
->code
, epilog
->code_size
);
6042 else if (mainb
->rodata_size
> 0)
6043 util_memcpy_cpu_to_le32(ptr
, mainb
->rodata
, mainb
->rodata_size
);
6045 sscreen
->b
.ws
->buffer_unmap(shader
->bo
->buf
);
6049 static void si_shader_dump_disassembly(const struct radeon_shader_binary
*binary
,
6050 struct pipe_debug_callback
*debug
,
6051 const char *name
, FILE *file
)
6056 if (binary
->disasm_string
) {
6057 fprintf(file
, "Shader %s disassembly:\n", name
);
6058 fprintf(file
, "%s", binary
->disasm_string
);
6060 if (debug
&& debug
->debug_message
) {
6061 /* Very long debug messages are cut off, so send the
6062 * disassembly one line at a time. This causes more
6063 * overhead, but on the plus side it simplifies
6064 * parsing of resulting logs.
6066 pipe_debug_message(debug
, SHADER_INFO
,
6067 "Shader Disassembly Begin");
6069 line
= binary
->disasm_string
;
6071 p
= util_strchrnul(line
, '\n');
6075 pipe_debug_message(debug
, SHADER_INFO
,
6076 "%.*s", count
, line
);
6084 pipe_debug_message(debug
, SHADER_INFO
,
6085 "Shader Disassembly End");
6088 fprintf(file
, "Shader %s binary:\n", name
);
6089 for (i
= 0; i
< binary
->code_size
; i
+= 4) {
6090 fprintf(file
, "@0x%x: %02x%02x%02x%02x\n", i
,
6091 binary
->code
[i
+ 3], binary
->code
[i
+ 2],
6092 binary
->code
[i
+ 1], binary
->code
[i
]);
6097 static void si_shader_dump_stats(struct si_screen
*sscreen
,
6098 struct si_shader_config
*conf
,
6099 unsigned num_inputs
,
6101 struct pipe_debug_callback
*debug
,
6105 unsigned lds_increment
= sscreen
->b
.chip_class
>= CIK
? 512 : 256;
6106 unsigned lds_per_wave
= 0;
6107 unsigned max_simd_waves
= 10;
6109 /* Compute LDS usage for PS. */
6110 if (processor
== PIPE_SHADER_FRAGMENT
) {
6111 /* The minimum usage per wave is (num_inputs * 48). The maximum
6112 * usage is (num_inputs * 48 * 16).
6113 * We can get anything in between and it varies between waves.
6115 * The 48 bytes per input for a single primitive is equal to
6116 * 4 bytes/component * 4 components/input * 3 points.
6118 * Other stages don't know the size at compile time or don't
6119 * allocate LDS per wave, but instead they do it per thread group.
6121 lds_per_wave
= conf
->lds_size
* lds_increment
+
6122 align(num_inputs
* 48, lds_increment
);
6125 /* Compute the per-SIMD wave counts. */
6126 if (conf
->num_sgprs
) {
6127 if (sscreen
->b
.chip_class
>= VI
)
6128 max_simd_waves
= MIN2(max_simd_waves
, 800 / conf
->num_sgprs
);
6130 max_simd_waves
= MIN2(max_simd_waves
, 512 / conf
->num_sgprs
);
6133 if (conf
->num_vgprs
)
6134 max_simd_waves
= MIN2(max_simd_waves
, 256 / conf
->num_vgprs
);
6136 /* LDS is 64KB per CU (4 SIMDs), divided into 16KB blocks per SIMD
6140 max_simd_waves
= MIN2(max_simd_waves
, 16384 / lds_per_wave
);
6142 if (file
!= stderr
||
6143 r600_can_dump_shader(&sscreen
->b
, processor
)) {
6144 if (processor
== PIPE_SHADER_FRAGMENT
) {
6145 fprintf(file
, "*** SHADER CONFIG ***\n"
6146 "SPI_PS_INPUT_ADDR = 0x%04x\n"
6147 "SPI_PS_INPUT_ENA = 0x%04x\n",
6148 conf
->spi_ps_input_addr
, conf
->spi_ps_input_ena
);
6151 fprintf(file
, "*** SHADER STATS ***\n"
6154 "Spilled SGPRs: %d\n"
6155 "Spilled VGPRs: %d\n"
6156 "Code Size: %d bytes\n"
6158 "Scratch: %d bytes per wave\n"
6160 "********************\n\n\n",
6161 conf
->num_sgprs
, conf
->num_vgprs
,
6162 conf
->spilled_sgprs
, conf
->spilled_vgprs
, code_size
,
6163 conf
->lds_size
, conf
->scratch_bytes_per_wave
,
6167 pipe_debug_message(debug
, SHADER_INFO
,
6168 "Shader Stats: SGPRS: %d VGPRS: %d Code Size: %d "
6169 "LDS: %d Scratch: %d Max Waves: %d Spilled SGPRs: %d "
6170 "Spilled VGPRs: %d",
6171 conf
->num_sgprs
, conf
->num_vgprs
, code_size
,
6172 conf
->lds_size
, conf
->scratch_bytes_per_wave
,
6173 max_simd_waves
, conf
->spilled_sgprs
,
6174 conf
->spilled_vgprs
);
6177 static const char *si_get_shader_name(struct si_shader
*shader
,
6180 switch (processor
) {
6181 case PIPE_SHADER_VERTEX
:
6182 if (shader
->key
.vs
.as_es
)
6183 return "Vertex Shader as ES";
6184 else if (shader
->key
.vs
.as_ls
)
6185 return "Vertex Shader as LS";
6187 return "Vertex Shader as VS";
6188 case PIPE_SHADER_TESS_CTRL
:
6189 return "Tessellation Control Shader";
6190 case PIPE_SHADER_TESS_EVAL
:
6191 if (shader
->key
.tes
.as_es
)
6192 return "Tessellation Evaluation Shader as ES";
6194 return "Tessellation Evaluation Shader as VS";
6195 case PIPE_SHADER_GEOMETRY
:
6196 if (shader
->gs_copy_shader
== NULL
)
6197 return "GS Copy Shader as VS";
6199 return "Geometry Shader";
6200 case PIPE_SHADER_FRAGMENT
:
6201 return "Pixel Shader";
6202 case PIPE_SHADER_COMPUTE
:
6203 return "Compute Shader";
6205 return "Unknown Shader";
6209 void si_shader_dump(struct si_screen
*sscreen
, struct si_shader
*shader
,
6210 struct pipe_debug_callback
*debug
, unsigned processor
,
6213 if (file
!= stderr
||
6214 r600_can_dump_shader(&sscreen
->b
, processor
))
6215 si_dump_shader_key(processor
, &shader
->key
, file
);
6217 if (file
!= stderr
&& shader
->binary
.llvm_ir_string
) {
6218 fprintf(file
, "\n%s - main shader part - LLVM IR:\n\n",
6219 si_get_shader_name(shader
, processor
));
6220 fprintf(file
, "%s\n", shader
->binary
.llvm_ir_string
);
6223 if (file
!= stderr
||
6224 (r600_can_dump_shader(&sscreen
->b
, processor
) &&
6225 !(sscreen
->b
.debug_flags
& DBG_NO_ASM
))) {
6226 fprintf(file
, "\n%s:\n", si_get_shader_name(shader
, processor
));
6229 si_shader_dump_disassembly(&shader
->prolog
->binary
,
6230 debug
, "prolog", file
);
6232 si_shader_dump_disassembly(&shader
->binary
, debug
, "main", file
);
6235 si_shader_dump_disassembly(&shader
->epilog
->binary
,
6236 debug
, "epilog", file
);
6237 fprintf(file
, "\n");
6240 si_shader_dump_stats(sscreen
, &shader
->config
,
6241 shader
->selector
? shader
->selector
->info
.num_inputs
: 0,
6242 si_get_shader_binary_size(shader
), debug
, processor
,
6246 int si_compile_llvm(struct si_screen
*sscreen
,
6247 struct radeon_shader_binary
*binary
,
6248 struct si_shader_config
*conf
,
6249 LLVMTargetMachineRef tm
,
6251 struct pipe_debug_callback
*debug
,
6256 unsigned count
= p_atomic_inc_return(&sscreen
->b
.num_compilations
);
6258 if (r600_can_dump_shader(&sscreen
->b
, processor
)) {
6259 fprintf(stderr
, "radeonsi: Compiling shader %d\n", count
);
6261 if (!(sscreen
->b
.debug_flags
& (DBG_NO_IR
| DBG_PREOPT_IR
))) {
6262 fprintf(stderr
, "%s LLVM IR:\n\n", name
);
6263 LLVMDumpModule(mod
);
6264 fprintf(stderr
, "\n");
6268 if (sscreen
->record_llvm_ir
) {
6269 char *ir
= LLVMPrintModuleToString(mod
);
6270 binary
->llvm_ir_string
= strdup(ir
);
6271 LLVMDisposeMessage(ir
);
6274 if (!si_replace_shader(count
, binary
)) {
6275 r
= radeon_llvm_compile(mod
, binary
, tm
, debug
);
6280 si_shader_binary_read_config(binary
, conf
, 0);
6282 /* Enable 64-bit and 16-bit denormals, because there is no performance
6285 * If denormals are enabled, all floating-point output modifiers are
6288 * Don't enable denormals for 32-bit floats, because:
6289 * - Floating-point output modifiers would be ignored by the hw.
6290 * - Some opcodes don't support denormals, such as v_mad_f32. We would
6291 * have to stop using those.
6292 * - SI & CI would be very slow.
6294 conf
->float_mode
|= V_00B028_FP_64_DENORMS
;
6296 FREE(binary
->config
);
6297 FREE(binary
->global_symbol_offsets
);
6298 binary
->config
= NULL
;
6299 binary
->global_symbol_offsets
= NULL
;
6301 /* Some shaders can't have rodata because their binaries can be
6304 if (binary
->rodata_size
&&
6305 (processor
== PIPE_SHADER_VERTEX
||
6306 processor
== PIPE_SHADER_TESS_CTRL
||
6307 processor
== PIPE_SHADER_TESS_EVAL
||
6308 processor
== PIPE_SHADER_FRAGMENT
)) {
6309 fprintf(stderr
, "radeonsi: The shader can't have rodata.");
6316 static void si_llvm_build_ret(struct si_shader_context
*ctx
, LLVMValueRef ret
)
6318 if (LLVMGetTypeKind(LLVMTypeOf(ret
)) == LLVMVoidTypeKind
)
6319 LLVMBuildRetVoid(ctx
->radeon_bld
.gallivm
.builder
);
6321 LLVMBuildRet(ctx
->radeon_bld
.gallivm
.builder
, ret
);
6324 /* Generate code for the hardware VS shader stage to go with a geometry shader */
6325 static int si_generate_gs_copy_shader(struct si_screen
*sscreen
,
6326 struct si_shader_context
*ctx
,
6327 struct si_shader
*gs
,
6328 struct pipe_debug_callback
*debug
)
6330 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
6331 struct lp_build_tgsi_context
*bld_base
= &ctx
->radeon_bld
.soa
.bld_base
;
6332 struct lp_build_context
*uint
= &bld_base
->uint_bld
;
6333 struct si_shader_output_values
*outputs
;
6334 struct tgsi_shader_info
*gsinfo
= &gs
->selector
->info
;
6335 LLVMValueRef args
[9];
6338 outputs
= MALLOC(gsinfo
->num_outputs
* sizeof(outputs
[0]));
6340 si_init_shader_ctx(ctx
, sscreen
, ctx
->shader
, ctx
->tm
);
6341 ctx
->type
= PIPE_SHADER_VERTEX
;
6342 ctx
->is_gs_copy_shader
= true;
6344 create_meta_data(ctx
);
6345 create_function(ctx
);
6346 preload_ring_buffers(ctx
);
6348 args
[0] = ctx
->gsvs_ring
[0];
6349 args
[1] = lp_build_mul_imm(uint
,
6350 LLVMGetParam(ctx
->radeon_bld
.main_fn
,
6351 ctx
->param_vertex_id
),
6353 args
[3] = uint
->zero
;
6354 args
[4] = uint
->one
; /* OFFEN */
6355 args
[5] = uint
->zero
; /* IDXEN */
6356 args
[6] = uint
->one
; /* GLC */
6357 args
[7] = uint
->one
; /* SLC */
6358 args
[8] = uint
->zero
; /* TFE */
6360 /* Fetch vertex data from GSVS ring */
6361 for (i
= 0; i
< gsinfo
->num_outputs
; ++i
) {
6364 outputs
[i
].name
= gsinfo
->output_semantic_name
[i
];
6365 outputs
[i
].sid
= gsinfo
->output_semantic_index
[i
];
6367 for (chan
= 0; chan
< 4; chan
++) {
6368 args
[2] = lp_build_const_int32(gallivm
,
6370 gs
->selector
->gs_max_out_vertices
* 16 * 4);
6372 outputs
[i
].values
[chan
] =
6373 LLVMBuildBitCast(gallivm
->builder
,
6374 lp_build_intrinsic(gallivm
->builder
,
6375 "llvm.SI.buffer.load.dword.i32.i32",
6377 LLVMReadOnlyAttribute
),
6382 si_llvm_export_vs(bld_base
, outputs
, gsinfo
->num_outputs
);
6384 LLVMBuildRetVoid(gallivm
->builder
);
6386 /* Dump LLVM IR before any optimization passes */
6387 if (sscreen
->b
.debug_flags
& DBG_PREOPT_IR
&&
6388 r600_can_dump_shader(&sscreen
->b
, PIPE_SHADER_GEOMETRY
))
6389 LLVMDumpModule(bld_base
->base
.gallivm
->module
);
6391 radeon_llvm_finalize_module(&ctx
->radeon_bld
);
6393 r
= si_compile_llvm(sscreen
, &ctx
->shader
->binary
,
6394 &ctx
->shader
->config
, ctx
->tm
,
6395 bld_base
->base
.gallivm
->module
,
6396 debug
, PIPE_SHADER_GEOMETRY
,
6399 if (r600_can_dump_shader(&sscreen
->b
, PIPE_SHADER_GEOMETRY
))
6400 fprintf(stderr
, "GS Copy Shader:\n");
6401 si_shader_dump(sscreen
, ctx
->shader
, debug
,
6402 PIPE_SHADER_GEOMETRY
, stderr
);
6403 r
= si_shader_binary_upload(sscreen
, ctx
->shader
);
6406 radeon_llvm_dispose(&ctx
->radeon_bld
);
6412 static void si_dump_shader_key(unsigned shader
, union si_shader_key
*key
,
6417 fprintf(f
, "SHADER KEY\n");
6420 case PIPE_SHADER_VERTEX
:
6421 fprintf(f
, " instance_divisors = {");
6422 for (i
= 0; i
< ARRAY_SIZE(key
->vs
.prolog
.instance_divisors
); i
++)
6423 fprintf(f
, !i
? "%u" : ", %u",
6424 key
->vs
.prolog
.instance_divisors
[i
]);
6426 fprintf(f
, " as_es = %u\n", key
->vs
.as_es
);
6427 fprintf(f
, " as_ls = %u\n", key
->vs
.as_ls
);
6428 fprintf(f
, " export_prim_id = %u\n", key
->vs
.epilog
.export_prim_id
);
6431 case PIPE_SHADER_TESS_CTRL
:
6432 fprintf(f
, " prim_mode = %u\n", key
->tcs
.epilog
.prim_mode
);
6435 case PIPE_SHADER_TESS_EVAL
:
6436 fprintf(f
, " as_es = %u\n", key
->tes
.as_es
);
6437 fprintf(f
, " export_prim_id = %u\n", key
->tes
.epilog
.export_prim_id
);
6440 case PIPE_SHADER_GEOMETRY
:
6441 case PIPE_SHADER_COMPUTE
:
6444 case PIPE_SHADER_FRAGMENT
:
6445 fprintf(f
, " prolog.color_two_side = %u\n", key
->ps
.prolog
.color_two_side
);
6446 fprintf(f
, " prolog.flatshade_colors = %u\n", key
->ps
.prolog
.flatshade_colors
);
6447 fprintf(f
, " prolog.poly_stipple = %u\n", key
->ps
.prolog
.poly_stipple
);
6448 fprintf(f
, " prolog.force_persp_sample_interp = %u\n", key
->ps
.prolog
.force_persp_sample_interp
);
6449 fprintf(f
, " prolog.force_linear_sample_interp = %u\n", key
->ps
.prolog
.force_linear_sample_interp
);
6450 fprintf(f
, " prolog.force_persp_center_interp = %u\n", key
->ps
.prolog
.force_persp_center_interp
);
6451 fprintf(f
, " prolog.force_linear_center_interp = %u\n", key
->ps
.prolog
.force_linear_center_interp
);
6452 fprintf(f
, " prolog.bc_optimize_for_persp = %u\n", key
->ps
.prolog
.bc_optimize_for_persp
);
6453 fprintf(f
, " prolog.bc_optimize_for_linear = %u\n", key
->ps
.prolog
.bc_optimize_for_linear
);
6454 fprintf(f
, " epilog.spi_shader_col_format = 0x%x\n", key
->ps
.epilog
.spi_shader_col_format
);
6455 fprintf(f
, " epilog.color_is_int8 = 0x%X\n", key
->ps
.epilog
.color_is_int8
);
6456 fprintf(f
, " epilog.last_cbuf = %u\n", key
->ps
.epilog
.last_cbuf
);
6457 fprintf(f
, " epilog.alpha_func = %u\n", key
->ps
.epilog
.alpha_func
);
6458 fprintf(f
, " epilog.alpha_to_one = %u\n", key
->ps
.epilog
.alpha_to_one
);
6459 fprintf(f
, " epilog.poly_line_smoothing = %u\n", key
->ps
.epilog
.poly_line_smoothing
);
6460 fprintf(f
, " epilog.clamp_color = %u\n", key
->ps
.epilog
.clamp_color
);
6468 static void si_init_shader_ctx(struct si_shader_context
*ctx
,
6469 struct si_screen
*sscreen
,
6470 struct si_shader
*shader
,
6471 LLVMTargetMachineRef tm
)
6473 struct lp_build_tgsi_context
*bld_base
;
6474 struct lp_build_tgsi_action tmpl
= {};
6476 memset(ctx
, 0, sizeof(*ctx
));
6477 radeon_llvm_context_init(
6478 &ctx
->radeon_bld
, "amdgcn--",
6479 (shader
&& shader
->selector
) ? &shader
->selector
->info
: NULL
,
6480 (shader
&& shader
->selector
) ? shader
->selector
->tokens
: NULL
);
6482 ctx
->screen
= sscreen
;
6483 if (shader
&& shader
->selector
)
6484 ctx
->type
= shader
->selector
->info
.processor
;
6487 ctx
->shader
= shader
;
6489 ctx
->voidt
= LLVMVoidTypeInContext(ctx
->radeon_bld
.gallivm
.context
);
6490 ctx
->i1
= LLVMInt1TypeInContext(ctx
->radeon_bld
.gallivm
.context
);
6491 ctx
->i8
= LLVMInt8TypeInContext(ctx
->radeon_bld
.gallivm
.context
);
6492 ctx
->i32
= LLVMInt32TypeInContext(ctx
->radeon_bld
.gallivm
.context
);
6493 ctx
->i64
= LLVMInt64TypeInContext(ctx
->radeon_bld
.gallivm
.context
);
6494 ctx
->i128
= LLVMIntTypeInContext(ctx
->radeon_bld
.gallivm
.context
, 128);
6495 ctx
->f32
= LLVMFloatTypeInContext(ctx
->radeon_bld
.gallivm
.context
);
6496 ctx
->v16i8
= LLVMVectorType(ctx
->i8
, 16);
6497 ctx
->v2i32
= LLVMVectorType(ctx
->i32
, 2);
6498 ctx
->v4i32
= LLVMVectorType(ctx
->i32
, 4);
6499 ctx
->v4f32
= LLVMVectorType(ctx
->f32
, 4);
6500 ctx
->v8i32
= LLVMVectorType(ctx
->i32
, 8);
6502 bld_base
= &ctx
->radeon_bld
.soa
.bld_base
;
6503 bld_base
->emit_fetch_funcs
[TGSI_FILE_CONSTANT
] = fetch_constant
;
6505 bld_base
->op_actions
[TGSI_OPCODE_INTERP_CENTROID
] = interp_action
;
6506 bld_base
->op_actions
[TGSI_OPCODE_INTERP_SAMPLE
] = interp_action
;
6507 bld_base
->op_actions
[TGSI_OPCODE_INTERP_OFFSET
] = interp_action
;
6509 bld_base
->op_actions
[TGSI_OPCODE_TEX
] = tex_action
;
6510 bld_base
->op_actions
[TGSI_OPCODE_TEX2
] = tex_action
;
6511 bld_base
->op_actions
[TGSI_OPCODE_TXB
] = tex_action
;
6512 bld_base
->op_actions
[TGSI_OPCODE_TXB2
] = tex_action
;
6513 bld_base
->op_actions
[TGSI_OPCODE_TXD
] = tex_action
;
6514 bld_base
->op_actions
[TGSI_OPCODE_TXF
] = tex_action
;
6515 bld_base
->op_actions
[TGSI_OPCODE_TXL
] = tex_action
;
6516 bld_base
->op_actions
[TGSI_OPCODE_TXL2
] = tex_action
;
6517 bld_base
->op_actions
[TGSI_OPCODE_TXP
] = tex_action
;
6518 bld_base
->op_actions
[TGSI_OPCODE_TXQ
].fetch_args
= txq_fetch_args
;
6519 bld_base
->op_actions
[TGSI_OPCODE_TXQ
].emit
= txq_emit
;
6520 bld_base
->op_actions
[TGSI_OPCODE_TG4
] = tex_action
;
6521 bld_base
->op_actions
[TGSI_OPCODE_LODQ
] = tex_action
;
6522 bld_base
->op_actions
[TGSI_OPCODE_TXQS
].emit
= si_llvm_emit_txqs
;
6524 bld_base
->op_actions
[TGSI_OPCODE_LOAD
].fetch_args
= load_fetch_args
;
6525 bld_base
->op_actions
[TGSI_OPCODE_LOAD
].emit
= load_emit
;
6526 bld_base
->op_actions
[TGSI_OPCODE_STORE
].fetch_args
= store_fetch_args
;
6527 bld_base
->op_actions
[TGSI_OPCODE_STORE
].emit
= store_emit
;
6528 bld_base
->op_actions
[TGSI_OPCODE_RESQ
].fetch_args
= resq_fetch_args
;
6529 bld_base
->op_actions
[TGSI_OPCODE_RESQ
].emit
= resq_emit
;
6531 tmpl
.fetch_args
= atomic_fetch_args
;
6532 tmpl
.emit
= atomic_emit
;
6533 bld_base
->op_actions
[TGSI_OPCODE_ATOMUADD
] = tmpl
;
6534 bld_base
->op_actions
[TGSI_OPCODE_ATOMUADD
].intr_name
= "add";
6535 bld_base
->op_actions
[TGSI_OPCODE_ATOMXCHG
] = tmpl
;
6536 bld_base
->op_actions
[TGSI_OPCODE_ATOMXCHG
].intr_name
= "swap";
6537 bld_base
->op_actions
[TGSI_OPCODE_ATOMCAS
] = tmpl
;
6538 bld_base
->op_actions
[TGSI_OPCODE_ATOMCAS
].intr_name
= "cmpswap";
6539 bld_base
->op_actions
[TGSI_OPCODE_ATOMAND
] = tmpl
;
6540 bld_base
->op_actions
[TGSI_OPCODE_ATOMAND
].intr_name
= "and";
6541 bld_base
->op_actions
[TGSI_OPCODE_ATOMOR
] = tmpl
;
6542 bld_base
->op_actions
[TGSI_OPCODE_ATOMOR
].intr_name
= "or";
6543 bld_base
->op_actions
[TGSI_OPCODE_ATOMXOR
] = tmpl
;
6544 bld_base
->op_actions
[TGSI_OPCODE_ATOMXOR
].intr_name
= "xor";
6545 bld_base
->op_actions
[TGSI_OPCODE_ATOMUMIN
] = tmpl
;
6546 bld_base
->op_actions
[TGSI_OPCODE_ATOMUMIN
].intr_name
= "umin";
6547 bld_base
->op_actions
[TGSI_OPCODE_ATOMUMAX
] = tmpl
;
6548 bld_base
->op_actions
[TGSI_OPCODE_ATOMUMAX
].intr_name
= "umax";
6549 bld_base
->op_actions
[TGSI_OPCODE_ATOMIMIN
] = tmpl
;
6550 bld_base
->op_actions
[TGSI_OPCODE_ATOMIMIN
].intr_name
= "smin";
6551 bld_base
->op_actions
[TGSI_OPCODE_ATOMIMAX
] = tmpl
;
6552 bld_base
->op_actions
[TGSI_OPCODE_ATOMIMAX
].intr_name
= "smax";
6554 bld_base
->op_actions
[TGSI_OPCODE_MEMBAR
].emit
= membar_emit
;
6556 bld_base
->op_actions
[TGSI_OPCODE_DDX
].emit
= si_llvm_emit_ddxy
;
6557 bld_base
->op_actions
[TGSI_OPCODE_DDY
].emit
= si_llvm_emit_ddxy
;
6558 bld_base
->op_actions
[TGSI_OPCODE_DDX_FINE
].emit
= si_llvm_emit_ddxy
;
6559 bld_base
->op_actions
[TGSI_OPCODE_DDY_FINE
].emit
= si_llvm_emit_ddxy
;
6561 bld_base
->op_actions
[TGSI_OPCODE_EMIT
].emit
= si_llvm_emit_vertex
;
6562 bld_base
->op_actions
[TGSI_OPCODE_ENDPRIM
].emit
= si_llvm_emit_primitive
;
6563 bld_base
->op_actions
[TGSI_OPCODE_BARRIER
].emit
= si_llvm_emit_barrier
;
6565 bld_base
->op_actions
[TGSI_OPCODE_MAX
].emit
= build_tgsi_intrinsic_nomem
;
6566 bld_base
->op_actions
[TGSI_OPCODE_MAX
].intr_name
= "llvm.maxnum.f32";
6567 bld_base
->op_actions
[TGSI_OPCODE_MIN
].emit
= build_tgsi_intrinsic_nomem
;
6568 bld_base
->op_actions
[TGSI_OPCODE_MIN
].intr_name
= "llvm.minnum.f32";
6571 int si_compile_tgsi_shader(struct si_screen
*sscreen
,
6572 LLVMTargetMachineRef tm
,
6573 struct si_shader
*shader
,
6575 struct pipe_debug_callback
*debug
)
6577 struct si_shader_selector
*sel
= shader
->selector
;
6578 struct si_shader_context ctx
;
6579 struct lp_build_tgsi_context
*bld_base
;
6583 /* Dump TGSI code before doing TGSI->LLVM conversion in case the
6584 * conversion fails. */
6585 if (r600_can_dump_shader(&sscreen
->b
, sel
->info
.processor
) &&
6586 !(sscreen
->b
.debug_flags
& DBG_NO_TGSI
)) {
6587 tgsi_dump(sel
->tokens
, 0);
6588 si_dump_streamout(&sel
->so
);
6591 si_init_shader_ctx(&ctx
, sscreen
, shader
, tm
);
6592 ctx
.is_monolithic
= is_monolithic
;
6594 shader
->info
.uses_instanceid
= sel
->info
.uses_instanceid
;
6596 bld_base
= &ctx
.radeon_bld
.soa
.bld_base
;
6597 ctx
.radeon_bld
.load_system_value
= declare_system_value
;
6600 case PIPE_SHADER_VERTEX
:
6601 ctx
.radeon_bld
.load_input
= declare_input_vs
;
6602 if (shader
->key
.vs
.as_ls
)
6603 bld_base
->emit_epilogue
= si_llvm_emit_ls_epilogue
;
6604 else if (shader
->key
.vs
.as_es
)
6605 bld_base
->emit_epilogue
= si_llvm_emit_es_epilogue
;
6607 bld_base
->emit_epilogue
= si_llvm_emit_vs_epilogue
;
6609 case PIPE_SHADER_TESS_CTRL
:
6610 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = fetch_input_tcs
;
6611 bld_base
->emit_fetch_funcs
[TGSI_FILE_OUTPUT
] = fetch_output_tcs
;
6612 bld_base
->emit_store
= store_output_tcs
;
6613 bld_base
->emit_epilogue
= si_llvm_emit_tcs_epilogue
;
6615 case PIPE_SHADER_TESS_EVAL
:
6616 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = fetch_input_tes
;
6617 if (shader
->key
.tes
.as_es
)
6618 bld_base
->emit_epilogue
= si_llvm_emit_es_epilogue
;
6620 bld_base
->emit_epilogue
= si_llvm_emit_vs_epilogue
;
6622 case PIPE_SHADER_GEOMETRY
:
6623 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = fetch_input_gs
;
6624 bld_base
->emit_epilogue
= si_llvm_emit_gs_epilogue
;
6626 case PIPE_SHADER_FRAGMENT
:
6627 ctx
.radeon_bld
.load_input
= declare_input_fs
;
6629 bld_base
->emit_epilogue
= si_llvm_emit_fs_epilogue
;
6631 bld_base
->emit_epilogue
= si_llvm_return_fs_outputs
;
6633 case PIPE_SHADER_COMPUTE
:
6634 ctx
.radeon_bld
.declare_memory_region
= declare_compute_memory
;
6637 assert(!"Unsupported shader type");
6641 create_meta_data(&ctx
);
6642 create_function(&ctx
);
6643 preload_ring_buffers(&ctx
);
6645 if (ctx
.is_monolithic
&& sel
->type
== PIPE_SHADER_FRAGMENT
&&
6646 shader
->key
.ps
.prolog
.poly_stipple
) {
6647 LLVMValueRef list
= LLVMGetParam(ctx
.radeon_bld
.main_fn
,
6648 SI_PARAM_RW_BUFFERS
);
6649 si_llvm_emit_polygon_stipple(&ctx
, list
,
6650 SI_PARAM_POS_FIXED_PT
);
6653 if (ctx
.type
== PIPE_SHADER_GEOMETRY
) {
6655 for (i
= 0; i
< 4; i
++) {
6656 ctx
.gs_next_vertex
[i
] =
6657 lp_build_alloca(bld_base
->base
.gallivm
,
6662 if (!lp_build_tgsi_llvm(bld_base
, sel
->tokens
)) {
6663 fprintf(stderr
, "Failed to translate shader from TGSI to LLVM\n");
6667 si_llvm_build_ret(&ctx
, ctx
.return_value
);
6668 mod
= bld_base
->base
.gallivm
->module
;
6670 /* Dump LLVM IR before any optimization passes */
6671 if (sscreen
->b
.debug_flags
& DBG_PREOPT_IR
&&
6672 r600_can_dump_shader(&sscreen
->b
, ctx
.type
))
6673 LLVMDumpModule(mod
);
6675 radeon_llvm_finalize_module(&ctx
.radeon_bld
);
6677 r
= si_compile_llvm(sscreen
, &shader
->binary
, &shader
->config
, tm
,
6678 mod
, debug
, ctx
.type
, "TGSI shader");
6680 fprintf(stderr
, "LLVM failed to compile shader\n");
6684 radeon_llvm_dispose(&ctx
.radeon_bld
);
6686 /* Validate SGPR and VGPR usage for compute to detect compiler bugs.
6687 * LLVM 3.9svn has this bug.
6689 if (sel
->type
== PIPE_SHADER_COMPUTE
) {
6690 unsigned *props
= sel
->info
.properties
;
6691 unsigned wave_size
= 64;
6692 unsigned max_vgprs
= 256;
6693 unsigned max_sgprs
= sscreen
->b
.chip_class
>= VI
? 800 : 512;
6694 unsigned max_sgprs_per_wave
= 128;
6695 unsigned min_waves_per_cu
=
6696 DIV_ROUND_UP(props
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
] *
6697 props
[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT
] *
6698 props
[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH
],
6700 unsigned min_waves_per_simd
= DIV_ROUND_UP(min_waves_per_cu
, 4);
6702 max_vgprs
= max_vgprs
/ min_waves_per_simd
;
6703 max_sgprs
= MIN2(max_sgprs
/ min_waves_per_simd
, max_sgprs_per_wave
);
6705 if (shader
->config
.num_sgprs
> max_sgprs
||
6706 shader
->config
.num_vgprs
> max_vgprs
) {
6707 fprintf(stderr
, "LLVM failed to compile a shader correctly: "
6708 "SGPR:VGPR usage is %u:%u, but the hw limit is %u:%u\n",
6709 shader
->config
.num_sgprs
, shader
->config
.num_vgprs
,
6710 max_sgprs
, max_vgprs
);
6712 /* Just terminate the process, because dependent
6713 * shaders can hang due to bad input data, but use
6714 * the env var to allow shader-db to work.
6716 if (!debug_get_bool_option("SI_PASS_BAD_SHADERS", false))
6721 /* Add the scratch offset to input SGPRs. */
6722 if (shader
->config
.scratch_bytes_per_wave
)
6723 shader
->info
.num_input_sgprs
+= 1; /* scratch byte offset */
6725 /* Calculate the number of fragment input VGPRs. */
6726 if (ctx
.type
== PIPE_SHADER_FRAGMENT
) {
6727 shader
->info
.num_input_vgprs
= 0;
6728 shader
->info
.face_vgpr_index
= -1;
6730 if (G_0286CC_PERSP_SAMPLE_ENA(shader
->config
.spi_ps_input_addr
))
6731 shader
->info
.num_input_vgprs
+= 2;
6732 if (G_0286CC_PERSP_CENTER_ENA(shader
->config
.spi_ps_input_addr
))
6733 shader
->info
.num_input_vgprs
+= 2;
6734 if (G_0286CC_PERSP_CENTROID_ENA(shader
->config
.spi_ps_input_addr
))
6735 shader
->info
.num_input_vgprs
+= 2;
6736 if (G_0286CC_PERSP_PULL_MODEL_ENA(shader
->config
.spi_ps_input_addr
))
6737 shader
->info
.num_input_vgprs
+= 3;
6738 if (G_0286CC_LINEAR_SAMPLE_ENA(shader
->config
.spi_ps_input_addr
))
6739 shader
->info
.num_input_vgprs
+= 2;
6740 if (G_0286CC_LINEAR_CENTER_ENA(shader
->config
.spi_ps_input_addr
))
6741 shader
->info
.num_input_vgprs
+= 2;
6742 if (G_0286CC_LINEAR_CENTROID_ENA(shader
->config
.spi_ps_input_addr
))
6743 shader
->info
.num_input_vgprs
+= 2;
6744 if (G_0286CC_LINE_STIPPLE_TEX_ENA(shader
->config
.spi_ps_input_addr
))
6745 shader
->info
.num_input_vgprs
+= 1;
6746 if (G_0286CC_POS_X_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
6747 shader
->info
.num_input_vgprs
+= 1;
6748 if (G_0286CC_POS_Y_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
6749 shader
->info
.num_input_vgprs
+= 1;
6750 if (G_0286CC_POS_Z_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
6751 shader
->info
.num_input_vgprs
+= 1;
6752 if (G_0286CC_POS_W_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
6753 shader
->info
.num_input_vgprs
+= 1;
6754 if (G_0286CC_FRONT_FACE_ENA(shader
->config
.spi_ps_input_addr
)) {
6755 shader
->info
.face_vgpr_index
= shader
->info
.num_input_vgprs
;
6756 shader
->info
.num_input_vgprs
+= 1;
6758 if (G_0286CC_ANCILLARY_ENA(shader
->config
.spi_ps_input_addr
))
6759 shader
->info
.num_input_vgprs
+= 1;
6760 if (G_0286CC_SAMPLE_COVERAGE_ENA(shader
->config
.spi_ps_input_addr
))
6761 shader
->info
.num_input_vgprs
+= 1;
6762 if (G_0286CC_POS_FIXED_PT_ENA(shader
->config
.spi_ps_input_addr
))
6763 shader
->info
.num_input_vgprs
+= 1;
6766 if (ctx
.type
== PIPE_SHADER_GEOMETRY
) {
6767 shader
->gs_copy_shader
= CALLOC_STRUCT(si_shader
);
6768 shader
->gs_copy_shader
->selector
= shader
->selector
;
6769 ctx
.shader
= shader
->gs_copy_shader
;
6770 if ((r
= si_generate_gs_copy_shader(sscreen
, &ctx
,
6772 free(shader
->gs_copy_shader
);
6773 shader
->gs_copy_shader
= NULL
;
6783 * Create, compile and return a shader part (prolog or epilog).
6785 * \param sscreen screen
6786 * \param list list of shader parts of the same category
6787 * \param key shader part key
6788 * \param tm LLVM target machine
6789 * \param debug debug callback
6790 * \param compile the callback responsible for compilation
6791 * \return non-NULL on success
6793 static struct si_shader_part
*
6794 si_get_shader_part(struct si_screen
*sscreen
,
6795 struct si_shader_part
**list
,
6796 union si_shader_part_key
*key
,
6797 LLVMTargetMachineRef tm
,
6798 struct pipe_debug_callback
*debug
,
6799 bool (*compile
)(struct si_screen
*,
6800 LLVMTargetMachineRef
,
6801 struct pipe_debug_callback
*,
6802 struct si_shader_part
*))
6804 struct si_shader_part
*result
;
6806 pipe_mutex_lock(sscreen
->shader_parts_mutex
);
6808 /* Find existing. */
6809 for (result
= *list
; result
; result
= result
->next
) {
6810 if (memcmp(&result
->key
, key
, sizeof(*key
)) == 0) {
6811 pipe_mutex_unlock(sscreen
->shader_parts_mutex
);
6816 /* Compile a new one. */
6817 result
= CALLOC_STRUCT(si_shader_part
);
6819 if (!compile(sscreen
, tm
, debug
, result
)) {
6821 pipe_mutex_unlock(sscreen
->shader_parts_mutex
);
6825 result
->next
= *list
;
6827 pipe_mutex_unlock(sscreen
->shader_parts_mutex
);
6832 * Create a vertex shader prolog.
6834 * The inputs are the same as VS (a lot of SGPRs and 4 VGPR system values).
6835 * All inputs are returned unmodified. The vertex load indices are
6836 * stored after them, which will used by the API VS for fetching inputs.
6838 * For example, the expected outputs for instance_divisors[] = {0, 1, 2} are:
6843 * (VertexID + BaseVertex),
6844 * (InstanceID + StartInstance),
6845 * (InstanceID / 2 + StartInstance)
6847 static bool si_compile_vs_prolog(struct si_screen
*sscreen
,
6848 LLVMTargetMachineRef tm
,
6849 struct pipe_debug_callback
*debug
,
6850 struct si_shader_part
*out
)
6852 union si_shader_part_key
*key
= &out
->key
;
6853 struct si_shader shader
= {};
6854 struct si_shader_context ctx
;
6855 struct gallivm_state
*gallivm
= &ctx
.radeon_bld
.gallivm
;
6856 LLVMTypeRef
*params
, *returns
;
6857 LLVMValueRef ret
, func
;
6858 int last_sgpr
, num_params
, num_returns
, i
;
6861 si_init_shader_ctx(&ctx
, sscreen
, &shader
, tm
);
6862 ctx
.type
= PIPE_SHADER_VERTEX
;
6863 ctx
.param_vertex_id
= key
->vs_prolog
.num_input_sgprs
;
6864 ctx
.param_instance_id
= key
->vs_prolog
.num_input_sgprs
+ 3;
6866 /* 4 preloaded VGPRs + vertex load indices as prolog outputs */
6867 params
= alloca((key
->vs_prolog
.num_input_sgprs
+ 4) *
6868 sizeof(LLVMTypeRef
));
6869 returns
= alloca((key
->vs_prolog
.num_input_sgprs
+ 4 +
6870 key
->vs_prolog
.last_input
+ 1) *
6871 sizeof(LLVMTypeRef
));
6875 /* Declare input and output SGPRs. */
6877 for (i
= 0; i
< key
->vs_prolog
.num_input_sgprs
; i
++) {
6878 params
[num_params
++] = ctx
.i32
;
6879 returns
[num_returns
++] = ctx
.i32
;
6881 last_sgpr
= num_params
- 1;
6883 /* 4 preloaded VGPRs (outputs must be floats) */
6884 for (i
= 0; i
< 4; i
++) {
6885 params
[num_params
++] = ctx
.i32
;
6886 returns
[num_returns
++] = ctx
.f32
;
6889 /* Vertex load indices. */
6890 for (i
= 0; i
<= key
->vs_prolog
.last_input
; i
++)
6891 returns
[num_returns
++] = ctx
.f32
;
6893 /* Create the function. */
6894 si_create_function(&ctx
, returns
, num_returns
, params
,
6895 num_params
, last_sgpr
);
6896 func
= ctx
.radeon_bld
.main_fn
;
6898 /* Copy inputs to outputs. This should be no-op, as the registers match,
6899 * but it will prevent the compiler from overwriting them unintentionally.
6901 ret
= ctx
.return_value
;
6902 for (i
= 0; i
< key
->vs_prolog
.num_input_sgprs
; i
++) {
6903 LLVMValueRef p
= LLVMGetParam(func
, i
);
6904 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
, p
, i
, "");
6906 for (i
= num_params
- 4; i
< num_params
; i
++) {
6907 LLVMValueRef p
= LLVMGetParam(func
, i
);
6908 p
= LLVMBuildBitCast(gallivm
->builder
, p
, ctx
.f32
, "");
6909 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
, p
, i
, "");
6912 /* Compute vertex load indices from instance divisors. */
6913 for (i
= 0; i
<= key
->vs_prolog
.last_input
; i
++) {
6914 unsigned divisor
= key
->vs_prolog
.states
.instance_divisors
[i
];
6918 /* InstanceID / Divisor + StartInstance */
6919 index
= get_instance_index_for_fetch(&ctx
.radeon_bld
,
6920 SI_SGPR_START_INSTANCE
,
6923 /* VertexID + BaseVertex */
6924 index
= LLVMBuildAdd(gallivm
->builder
,
6925 LLVMGetParam(func
, ctx
.param_vertex_id
),
6926 LLVMGetParam(func
, SI_SGPR_BASE_VERTEX
), "");
6929 index
= LLVMBuildBitCast(gallivm
->builder
, index
, ctx
.f32
, "");
6930 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
, index
,
6935 si_llvm_build_ret(&ctx
, ret
);
6936 radeon_llvm_finalize_module(&ctx
.radeon_bld
);
6938 if (si_compile_llvm(sscreen
, &out
->binary
, &out
->config
, tm
,
6939 gallivm
->module
, debug
, ctx
.type
,
6940 "Vertex Shader Prolog"))
6943 radeon_llvm_dispose(&ctx
.radeon_bld
);
6948 * Compile the vertex shader epilog. This is also used by the tessellation
6949 * evaluation shader compiled as VS.
6951 * The input is PrimitiveID.
6953 * If PrimitiveID is required by the pixel shader, export it.
6954 * Otherwise, do nothing.
6956 static bool si_compile_vs_epilog(struct si_screen
*sscreen
,
6957 LLVMTargetMachineRef tm
,
6958 struct pipe_debug_callback
*debug
,
6959 struct si_shader_part
*out
)
6961 union si_shader_part_key
*key
= &out
->key
;
6962 struct si_shader_context ctx
;
6963 struct gallivm_state
*gallivm
= &ctx
.radeon_bld
.gallivm
;
6964 struct lp_build_tgsi_context
*bld_base
= &ctx
.radeon_bld
.soa
.bld_base
;
6965 LLVMTypeRef params
[5];
6969 si_init_shader_ctx(&ctx
, sscreen
, NULL
, tm
);
6970 ctx
.type
= PIPE_SHADER_VERTEX
;
6972 /* Declare input VGPRs. */
6973 num_params
= key
->vs_epilog
.states
.export_prim_id
?
6974 (VS_EPILOG_PRIMID_LOC
+ 1) : 0;
6975 assert(num_params
<= ARRAY_SIZE(params
));
6977 for (i
= 0; i
< num_params
; i
++)
6978 params
[i
] = ctx
.f32
;
6980 /* Create the function. */
6981 si_create_function(&ctx
, NULL
, 0, params
, num_params
, -1);
6984 if (key
->vs_epilog
.states
.export_prim_id
) {
6985 struct lp_build_context
*base
= &bld_base
->base
;
6986 struct lp_build_context
*uint
= &bld_base
->uint_bld
;
6987 LLVMValueRef args
[9];
6989 args
[0] = lp_build_const_int32(base
->gallivm
, 0x0); /* enabled channels */
6990 args
[1] = uint
->zero
; /* whether the EXEC mask is valid */
6991 args
[2] = uint
->zero
; /* DONE bit */
6992 args
[3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_PARAM
+
6993 key
->vs_epilog
.prim_id_param_offset
);
6994 args
[4] = uint
->zero
; /* COMPR flag (0 = 32-bit export) */
6995 args
[5] = LLVMGetParam(ctx
.radeon_bld
.main_fn
,
6996 VS_EPILOG_PRIMID_LOC
); /* X */
6997 args
[6] = uint
->undef
; /* Y */
6998 args
[7] = uint
->undef
; /* Z */
6999 args
[8] = uint
->undef
; /* W */
7001 lp_build_intrinsic(base
->gallivm
->builder
, "llvm.SI.export",
7002 LLVMVoidTypeInContext(base
->gallivm
->context
),
7007 LLVMBuildRetVoid(gallivm
->builder
);
7008 radeon_llvm_finalize_module(&ctx
.radeon_bld
);
7010 if (si_compile_llvm(sscreen
, &out
->binary
, &out
->config
, tm
,
7011 gallivm
->module
, debug
, ctx
.type
,
7012 "Vertex Shader Epilog"))
7015 radeon_llvm_dispose(&ctx
.radeon_bld
);
7020 * Create & compile a vertex shader epilog. This a helper used by VS and TES.
7022 static bool si_get_vs_epilog(struct si_screen
*sscreen
,
7023 LLVMTargetMachineRef tm
,
7024 struct si_shader
*shader
,
7025 struct pipe_debug_callback
*debug
,
7026 struct si_vs_epilog_bits
*states
)
7028 union si_shader_part_key epilog_key
;
7030 memset(&epilog_key
, 0, sizeof(epilog_key
));
7031 epilog_key
.vs_epilog
.states
= *states
;
7033 /* Set up the PrimitiveID output. */
7034 if (shader
->key
.vs
.epilog
.export_prim_id
) {
7035 unsigned index
= shader
->selector
->info
.num_outputs
;
7036 unsigned offset
= shader
->info
.nr_param_exports
++;
7038 epilog_key
.vs_epilog
.prim_id_param_offset
= offset
;
7039 assert(index
< ARRAY_SIZE(shader
->info
.vs_output_param_offset
));
7040 shader
->info
.vs_output_param_offset
[index
] = offset
;
7043 shader
->epilog
= si_get_shader_part(sscreen
, &sscreen
->vs_epilogs
,
7044 &epilog_key
, tm
, debug
,
7045 si_compile_vs_epilog
);
7046 return shader
->epilog
!= NULL
;
7050 * Select and compile (or reuse) vertex shader parts (prolog & epilog).
7052 static bool si_shader_select_vs_parts(struct si_screen
*sscreen
,
7053 LLVMTargetMachineRef tm
,
7054 struct si_shader
*shader
,
7055 struct pipe_debug_callback
*debug
)
7057 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
7058 union si_shader_part_key prolog_key
;
7061 /* Get the prolog. */
7062 memset(&prolog_key
, 0, sizeof(prolog_key
));
7063 prolog_key
.vs_prolog
.states
= shader
->key
.vs
.prolog
;
7064 prolog_key
.vs_prolog
.num_input_sgprs
= shader
->info
.num_input_sgprs
;
7065 prolog_key
.vs_prolog
.last_input
= MAX2(1, info
->num_inputs
) - 1;
7067 /* The prolog is a no-op if there are no inputs. */
7068 if (info
->num_inputs
) {
7070 si_get_shader_part(sscreen
, &sscreen
->vs_prologs
,
7071 &prolog_key
, tm
, debug
,
7072 si_compile_vs_prolog
);
7073 if (!shader
->prolog
)
7077 /* Get the epilog. */
7078 if (!shader
->key
.vs
.as_es
&& !shader
->key
.vs
.as_ls
&&
7079 !si_get_vs_epilog(sscreen
, tm
, shader
, debug
,
7080 &shader
->key
.vs
.epilog
))
7083 /* Set the instanceID flag. */
7084 for (i
= 0; i
< info
->num_inputs
; i
++)
7085 if (prolog_key
.vs_prolog
.states
.instance_divisors
[i
])
7086 shader
->info
.uses_instanceid
= true;
7092 * Select and compile (or reuse) TES parts (epilog).
7094 static bool si_shader_select_tes_parts(struct si_screen
*sscreen
,
7095 LLVMTargetMachineRef tm
,
7096 struct si_shader
*shader
,
7097 struct pipe_debug_callback
*debug
)
7099 if (shader
->key
.tes
.as_es
)
7102 /* TES compiled as VS. */
7103 return si_get_vs_epilog(sscreen
, tm
, shader
, debug
,
7104 &shader
->key
.tes
.epilog
);
7108 * Compile the TCS epilog. This writes tesselation factors to memory based on
7109 * the output primitive type of the tesselator (determined by TES).
7111 static bool si_compile_tcs_epilog(struct si_screen
*sscreen
,
7112 LLVMTargetMachineRef tm
,
7113 struct pipe_debug_callback
*debug
,
7114 struct si_shader_part
*out
)
7116 union si_shader_part_key
*key
= &out
->key
;
7117 struct si_shader shader
= {};
7118 struct si_shader_context ctx
;
7119 struct gallivm_state
*gallivm
= &ctx
.radeon_bld
.gallivm
;
7120 struct lp_build_tgsi_context
*bld_base
= &ctx
.radeon_bld
.soa
.bld_base
;
7121 LLVMTypeRef params
[16];
7123 int last_sgpr
, num_params
;
7126 si_init_shader_ctx(&ctx
, sscreen
, &shader
, tm
);
7127 ctx
.type
= PIPE_SHADER_TESS_CTRL
;
7128 shader
.key
.tcs
.epilog
= key
->tcs_epilog
.states
;
7130 /* Declare inputs. Only RW_BUFFERS and TESS_FACTOR_OFFSET are used. */
7131 params
[SI_PARAM_RW_BUFFERS
] = const_array(ctx
.v16i8
, SI_NUM_RW_BUFFERS
);
7132 params
[SI_PARAM_CONST_BUFFERS
] = ctx
.i64
;
7133 params
[SI_PARAM_SAMPLERS
] = ctx
.i64
;
7134 params
[SI_PARAM_IMAGES
] = ctx
.i64
;
7135 params
[SI_PARAM_SHADER_BUFFERS
] = ctx
.i64
;
7136 params
[SI_PARAM_TCS_OFFCHIP_LAYOUT
] = ctx
.i32
;
7137 params
[SI_PARAM_TCS_OUT_OFFSETS
] = ctx
.i32
;
7138 params
[SI_PARAM_TCS_OUT_LAYOUT
] = ctx
.i32
;
7139 params
[SI_PARAM_TCS_IN_LAYOUT
] = ctx
.i32
;
7140 params
[ctx
.param_oc_lds
= SI_PARAM_TCS_OC_LDS
] = ctx
.i32
;
7141 params
[SI_PARAM_TESS_FACTOR_OFFSET
] = ctx
.i32
;
7142 last_sgpr
= SI_PARAM_TESS_FACTOR_OFFSET
;
7143 num_params
= last_sgpr
+ 1;
7145 params
[num_params
++] = ctx
.i32
; /* patch index within the wave (REL_PATCH_ID) */
7146 params
[num_params
++] = ctx
.i32
; /* invocation ID within the patch */
7147 params
[num_params
++] = ctx
.i32
; /* LDS offset where tess factors should be loaded from */
7149 /* Create the function. */
7150 si_create_function(&ctx
, NULL
, 0, params
, num_params
, last_sgpr
);
7151 declare_tess_lds(&ctx
);
7152 func
= ctx
.radeon_bld
.main_fn
;
7154 si_write_tess_factors(bld_base
,
7155 LLVMGetParam(func
, last_sgpr
+ 1),
7156 LLVMGetParam(func
, last_sgpr
+ 2),
7157 LLVMGetParam(func
, last_sgpr
+ 3));
7160 LLVMBuildRetVoid(gallivm
->builder
);
7161 radeon_llvm_finalize_module(&ctx
.radeon_bld
);
7163 if (si_compile_llvm(sscreen
, &out
->binary
, &out
->config
, tm
,
7164 gallivm
->module
, debug
, ctx
.type
,
7165 "Tessellation Control Shader Epilog"))
7168 radeon_llvm_dispose(&ctx
.radeon_bld
);
7173 * Select and compile (or reuse) TCS parts (epilog).
7175 static bool si_shader_select_tcs_parts(struct si_screen
*sscreen
,
7176 LLVMTargetMachineRef tm
,
7177 struct si_shader
*shader
,
7178 struct pipe_debug_callback
*debug
)
7180 union si_shader_part_key epilog_key
;
7182 /* Get the epilog. */
7183 memset(&epilog_key
, 0, sizeof(epilog_key
));
7184 epilog_key
.tcs_epilog
.states
= shader
->key
.tcs
.epilog
;
7186 shader
->epilog
= si_get_shader_part(sscreen
, &sscreen
->tcs_epilogs
,
7187 &epilog_key
, tm
, debug
,
7188 si_compile_tcs_epilog
);
7189 return shader
->epilog
!= NULL
;
7193 * Compile the pixel shader prolog. This handles:
7194 * - two-side color selection and interpolation
7195 * - overriding interpolation parameters for the API PS
7196 * - polygon stippling
7198 * All preloaded SGPRs and VGPRs are passed through unmodified unless they are
7199 * overriden by other states. (e.g. per-sample interpolation)
7200 * Interpolated colors are stored after the preloaded VGPRs.
7202 static bool si_compile_ps_prolog(struct si_screen
*sscreen
,
7203 LLVMTargetMachineRef tm
,
7204 struct pipe_debug_callback
*debug
,
7205 struct si_shader_part
*out
)
7207 union si_shader_part_key
*key
= &out
->key
;
7208 struct si_shader shader
= {};
7209 struct si_shader_context ctx
;
7210 struct gallivm_state
*gallivm
= &ctx
.radeon_bld
.gallivm
;
7211 LLVMTypeRef
*params
;
7212 LLVMValueRef ret
, func
;
7213 int last_sgpr
, num_params
, num_returns
, i
, num_color_channels
;
7216 si_init_shader_ctx(&ctx
, sscreen
, &shader
, tm
);
7217 ctx
.type
= PIPE_SHADER_FRAGMENT
;
7218 shader
.key
.ps
.prolog
= key
->ps_prolog
.states
;
7220 /* Number of inputs + 8 color elements. */
7221 params
= alloca((key
->ps_prolog
.num_input_sgprs
+
7222 key
->ps_prolog
.num_input_vgprs
+ 8) *
7223 sizeof(LLVMTypeRef
));
7225 /* Declare inputs. */
7227 for (i
= 0; i
< key
->ps_prolog
.num_input_sgprs
; i
++)
7228 params
[num_params
++] = ctx
.i32
;
7229 last_sgpr
= num_params
- 1;
7231 for (i
= 0; i
< key
->ps_prolog
.num_input_vgprs
; i
++)
7232 params
[num_params
++] = ctx
.f32
;
7234 /* Declare outputs (same as inputs + add colors if needed) */
7235 num_returns
= num_params
;
7236 num_color_channels
= util_bitcount(key
->ps_prolog
.colors_read
);
7237 for (i
= 0; i
< num_color_channels
; i
++)
7238 params
[num_returns
++] = ctx
.f32
;
7240 /* Create the function. */
7241 si_create_function(&ctx
, params
, num_returns
, params
,
7242 num_params
, last_sgpr
);
7243 func
= ctx
.radeon_bld
.main_fn
;
7245 /* Copy inputs to outputs. This should be no-op, as the registers match,
7246 * but it will prevent the compiler from overwriting them unintentionally.
7248 ret
= ctx
.return_value
;
7249 for (i
= 0; i
< num_params
; i
++) {
7250 LLVMValueRef p
= LLVMGetParam(func
, i
);
7251 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
, p
, i
, "");
7254 /* Polygon stippling. */
7255 if (key
->ps_prolog
.states
.poly_stipple
) {
7256 /* POS_FIXED_PT is always last. */
7257 unsigned pos
= key
->ps_prolog
.num_input_sgprs
+
7258 key
->ps_prolog
.num_input_vgprs
- 1;
7259 LLVMValueRef ptr
[2], list
;
7261 /* Get the pointer to rw buffers. */
7262 ptr
[0] = LLVMGetParam(func
, SI_SGPR_RW_BUFFERS
);
7263 ptr
[1] = LLVMGetParam(func
, SI_SGPR_RW_BUFFERS_HI
);
7264 list
= lp_build_gather_values(gallivm
, ptr
, 2);
7265 list
= LLVMBuildBitCast(gallivm
->builder
, list
, ctx
.i64
, "");
7266 list
= LLVMBuildIntToPtr(gallivm
->builder
, list
,
7267 const_array(ctx
.v16i8
, SI_NUM_RW_BUFFERS
), "");
7269 si_llvm_emit_polygon_stipple(&ctx
, list
, pos
);
7272 if (key
->ps_prolog
.states
.bc_optimize_for_persp
||
7273 key
->ps_prolog
.states
.bc_optimize_for_linear
) {
7274 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7275 LLVMValueRef center
[2], centroid
[2], tmp
, bc_optimize
;
7277 /* The shader should do: if (PRIM_MASK[31]) CENTROID = CENTER;
7278 * The hw doesn't compute CENTROID if the whole wave only
7279 * contains fully-covered quads.
7281 * PRIM_MASK is after user SGPRs.
7283 bc_optimize
= LLVMGetParam(func
, SI_PS_NUM_USER_SGPR
);
7284 bc_optimize
= LLVMBuildLShr(gallivm
->builder
, bc_optimize
,
7285 LLVMConstInt(ctx
.i32
, 31, 0), "");
7286 bc_optimize
= LLVMBuildTrunc(gallivm
->builder
, bc_optimize
,
7289 if (key
->ps_prolog
.states
.bc_optimize_for_persp
) {
7290 /* Read PERSP_CENTER. */
7291 for (i
= 0; i
< 2; i
++)
7292 center
[i
] = LLVMGetParam(func
, base
+ 2 + i
);
7293 /* Read PERSP_CENTROID. */
7294 for (i
= 0; i
< 2; i
++)
7295 centroid
[i
] = LLVMGetParam(func
, base
+ 4 + i
);
7296 /* Select PERSP_CENTROID. */
7297 for (i
= 0; i
< 2; i
++) {
7298 tmp
= LLVMBuildSelect(gallivm
->builder
, bc_optimize
,
7299 center
[i
], centroid
[i
], "");
7300 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
,
7301 tmp
, base
+ 4 + i
, "");
7304 if (key
->ps_prolog
.states
.bc_optimize_for_linear
) {
7305 /* Read LINEAR_CENTER. */
7306 for (i
= 0; i
< 2; i
++)
7307 center
[i
] = LLVMGetParam(func
, base
+ 8 + i
);
7308 /* Read LINEAR_CENTROID. */
7309 for (i
= 0; i
< 2; i
++)
7310 centroid
[i
] = LLVMGetParam(func
, base
+ 10 + i
);
7311 /* Select LINEAR_CENTROID. */
7312 for (i
= 0; i
< 2; i
++) {
7313 tmp
= LLVMBuildSelect(gallivm
->builder
, bc_optimize
,
7314 center
[i
], centroid
[i
], "");
7315 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
,
7316 tmp
, base
+ 10 + i
, "");
7321 /* Interpolate colors. */
7322 for (i
= 0; i
< 2; i
++) {
7323 unsigned writemask
= (key
->ps_prolog
.colors_read
>> (i
* 4)) & 0xf;
7324 unsigned face_vgpr
= key
->ps_prolog
.num_input_sgprs
+
7325 key
->ps_prolog
.face_vgpr_index
;
7326 LLVMValueRef interp
[2], color
[4];
7327 LLVMValueRef interp_ij
= NULL
, prim_mask
= NULL
, face
= NULL
;
7332 /* If the interpolation qualifier is not CONSTANT (-1). */
7333 if (key
->ps_prolog
.color_interp_vgpr_index
[i
] != -1) {
7334 unsigned interp_vgpr
= key
->ps_prolog
.num_input_sgprs
+
7335 key
->ps_prolog
.color_interp_vgpr_index
[i
];
7337 /* Get the (i,j) updated by bc_optimize handling. */
7338 interp
[0] = LLVMBuildExtractValue(gallivm
->builder
, ret
,
7340 interp
[1] = LLVMBuildExtractValue(gallivm
->builder
, ret
,
7341 interp_vgpr
+ 1, "");
7342 interp_ij
= lp_build_gather_values(gallivm
, interp
, 2);
7343 interp_ij
= LLVMBuildBitCast(gallivm
->builder
, interp_ij
,
7347 /* Use the absolute location of the input. */
7348 prim_mask
= LLVMGetParam(func
, SI_PS_NUM_USER_SGPR
);
7350 if (key
->ps_prolog
.states
.color_two_side
) {
7351 face
= LLVMGetParam(func
, face_vgpr
);
7352 face
= LLVMBuildBitCast(gallivm
->builder
, face
, ctx
.i32
, "");
7355 interp_fs_input(&ctx
,
7356 key
->ps_prolog
.color_attr_index
[i
],
7357 TGSI_SEMANTIC_COLOR
, i
,
7358 key
->ps_prolog
.num_interp_inputs
,
7359 key
->ps_prolog
.colors_read
, interp_ij
,
7360 prim_mask
, face
, color
);
7363 unsigned chan
= u_bit_scan(&writemask
);
7364 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
, color
[chan
],
7369 /* Force per-sample interpolation. */
7370 if (key
->ps_prolog
.states
.force_persp_sample_interp
) {
7371 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7372 LLVMValueRef persp_sample
[2];
7374 /* Read PERSP_SAMPLE. */
7375 for (i
= 0; i
< 2; i
++)
7376 persp_sample
[i
] = LLVMGetParam(func
, base
+ i
);
7377 /* Overwrite PERSP_CENTER. */
7378 for (i
= 0; i
< 2; i
++)
7379 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
,
7380 persp_sample
[i
], base
+ 2 + i
, "");
7381 /* Overwrite PERSP_CENTROID. */
7382 for (i
= 0; i
< 2; i
++)
7383 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
,
7384 persp_sample
[i
], base
+ 4 + i
, "");
7386 if (key
->ps_prolog
.states
.force_linear_sample_interp
) {
7387 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7388 LLVMValueRef linear_sample
[2];
7390 /* Read LINEAR_SAMPLE. */
7391 for (i
= 0; i
< 2; i
++)
7392 linear_sample
[i
] = LLVMGetParam(func
, base
+ 6 + i
);
7393 /* Overwrite LINEAR_CENTER. */
7394 for (i
= 0; i
< 2; i
++)
7395 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
,
7396 linear_sample
[i
], base
+ 8 + i
, "");
7397 /* Overwrite LINEAR_CENTROID. */
7398 for (i
= 0; i
< 2; i
++)
7399 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
,
7400 linear_sample
[i
], base
+ 10 + i
, "");
7403 /* Force center interpolation. */
7404 if (key
->ps_prolog
.states
.force_persp_center_interp
) {
7405 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7406 LLVMValueRef persp_center
[2];
7408 /* Read PERSP_CENTER. */
7409 for (i
= 0; i
< 2; i
++)
7410 persp_center
[i
] = LLVMGetParam(func
, base
+ 2 + i
);
7411 /* Overwrite PERSP_SAMPLE. */
7412 for (i
= 0; i
< 2; i
++)
7413 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
,
7414 persp_center
[i
], base
+ i
, "");
7415 /* Overwrite PERSP_CENTROID. */
7416 for (i
= 0; i
< 2; i
++)
7417 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
,
7418 persp_center
[i
], base
+ 4 + i
, "");
7420 if (key
->ps_prolog
.states
.force_linear_center_interp
) {
7421 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
7422 LLVMValueRef linear_center
[2];
7424 /* Read LINEAR_CENTER. */
7425 for (i
= 0; i
< 2; i
++)
7426 linear_center
[i
] = LLVMGetParam(func
, base
+ 8 + i
);
7427 /* Overwrite LINEAR_SAMPLE. */
7428 for (i
= 0; i
< 2; i
++)
7429 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
,
7430 linear_center
[i
], base
+ 6 + i
, "");
7431 /* Overwrite LINEAR_CENTROID. */
7432 for (i
= 0; i
< 2; i
++)
7433 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
,
7434 linear_center
[i
], base
+ 10 + i
, "");
7437 /* Tell LLVM to insert WQM instruction sequence when needed. */
7438 if (key
->ps_prolog
.wqm
) {
7439 LLVMAddTargetDependentFunctionAttr(func
,
7440 "amdgpu-ps-wqm-outputs", "");
7444 si_llvm_build_ret(&ctx
, ret
);
7445 radeon_llvm_finalize_module(&ctx
.radeon_bld
);
7447 if (si_compile_llvm(sscreen
, &out
->binary
, &out
->config
, tm
,
7448 gallivm
->module
, debug
, ctx
.type
,
7449 "Fragment Shader Prolog"))
7452 radeon_llvm_dispose(&ctx
.radeon_bld
);
7457 * Compile the pixel shader epilog. This handles everything that must be
7458 * emulated for pixel shader exports. (alpha-test, format conversions, etc)
7460 static bool si_compile_ps_epilog(struct si_screen
*sscreen
,
7461 LLVMTargetMachineRef tm
,
7462 struct pipe_debug_callback
*debug
,
7463 struct si_shader_part
*out
)
7465 union si_shader_part_key
*key
= &out
->key
;
7466 struct si_shader shader
= {};
7467 struct si_shader_context ctx
;
7468 struct gallivm_state
*gallivm
= &ctx
.radeon_bld
.gallivm
;
7469 struct lp_build_tgsi_context
*bld_base
= &ctx
.radeon_bld
.soa
.bld_base
;
7470 LLVMTypeRef params
[16+8*4+3];
7471 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
7472 int last_sgpr
, num_params
, i
;
7474 struct si_ps_exports exp
= {};
7476 si_init_shader_ctx(&ctx
, sscreen
, &shader
, tm
);
7477 ctx
.type
= PIPE_SHADER_FRAGMENT
;
7478 shader
.key
.ps
.epilog
= key
->ps_epilog
.states
;
7480 /* Declare input SGPRs. */
7481 params
[SI_PARAM_RW_BUFFERS
] = ctx
.i64
;
7482 params
[SI_PARAM_CONST_BUFFERS
] = ctx
.i64
;
7483 params
[SI_PARAM_SAMPLERS
] = ctx
.i64
;
7484 params
[SI_PARAM_IMAGES
] = ctx
.i64
;
7485 params
[SI_PARAM_SHADER_BUFFERS
] = ctx
.i64
;
7486 params
[SI_PARAM_ALPHA_REF
] = ctx
.f32
;
7487 last_sgpr
= SI_PARAM_ALPHA_REF
;
7489 /* Declare input VGPRs. */
7490 num_params
= (last_sgpr
+ 1) +
7491 util_bitcount(key
->ps_epilog
.colors_written
) * 4 +
7492 key
->ps_epilog
.writes_z
+
7493 key
->ps_epilog
.writes_stencil
+
7494 key
->ps_epilog
.writes_samplemask
;
7496 num_params
= MAX2(num_params
,
7497 last_sgpr
+ 1 + PS_EPILOG_SAMPLEMASK_MIN_LOC
+ 1);
7499 assert(num_params
<= ARRAY_SIZE(params
));
7501 for (i
= last_sgpr
+ 1; i
< num_params
; i
++)
7502 params
[i
] = ctx
.f32
;
7504 /* Create the function. */
7505 si_create_function(&ctx
, NULL
, 0, params
, num_params
, last_sgpr
);
7506 /* Disable elimination of unused inputs. */
7507 radeon_llvm_add_attribute(ctx
.radeon_bld
.main_fn
,
7508 "InitialPSInputAddr", 0xffffff);
7510 /* Process colors. */
7511 unsigned vgpr
= last_sgpr
+ 1;
7512 unsigned colors_written
= key
->ps_epilog
.colors_written
;
7513 int last_color_export
= -1;
7515 /* Find the last color export. */
7516 if (!key
->ps_epilog
.writes_z
&&
7517 !key
->ps_epilog
.writes_stencil
&&
7518 !key
->ps_epilog
.writes_samplemask
) {
7519 unsigned spi_format
= key
->ps_epilog
.states
.spi_shader_col_format
;
7521 /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
7522 if (colors_written
== 0x1 && key
->ps_epilog
.states
.last_cbuf
> 0) {
7523 /* Just set this if any of the colorbuffers are enabled. */
7525 ((1llu << (4 * (key
->ps_epilog
.states
.last_cbuf
+ 1))) - 1))
7526 last_color_export
= 0;
7528 for (i
= 0; i
< 8; i
++)
7529 if (colors_written
& (1 << i
) &&
7530 (spi_format
>> (i
* 4)) & 0xf)
7531 last_color_export
= i
;
7535 while (colors_written
) {
7536 LLVMValueRef color
[4];
7537 int mrt
= u_bit_scan(&colors_written
);
7539 for (i
= 0; i
< 4; i
++)
7540 color
[i
] = LLVMGetParam(ctx
.radeon_bld
.main_fn
, vgpr
++);
7542 si_export_mrt_color(bld_base
, color
, mrt
,
7544 mrt
== last_color_export
, &exp
);
7547 /* Process depth, stencil, samplemask. */
7548 if (key
->ps_epilog
.writes_z
)
7549 depth
= LLVMGetParam(ctx
.radeon_bld
.main_fn
, vgpr
++);
7550 if (key
->ps_epilog
.writes_stencil
)
7551 stencil
= LLVMGetParam(ctx
.radeon_bld
.main_fn
, vgpr
++);
7552 if (key
->ps_epilog
.writes_samplemask
)
7553 samplemask
= LLVMGetParam(ctx
.radeon_bld
.main_fn
, vgpr
++);
7555 if (depth
|| stencil
|| samplemask
)
7556 si_export_mrt_z(bld_base
, depth
, stencil
, samplemask
, &exp
);
7557 else if (last_color_export
== -1)
7558 si_export_null(bld_base
);
7561 si_emit_ps_exports(&ctx
, &exp
);
7564 LLVMBuildRetVoid(gallivm
->builder
);
7565 radeon_llvm_finalize_module(&ctx
.radeon_bld
);
7567 if (si_compile_llvm(sscreen
, &out
->binary
, &out
->config
, tm
,
7568 gallivm
->module
, debug
, ctx
.type
,
7569 "Fragment Shader Epilog"))
7572 radeon_llvm_dispose(&ctx
.radeon_bld
);
7577 * Select and compile (or reuse) pixel shader parts (prolog & epilog).
7579 static bool si_shader_select_ps_parts(struct si_screen
*sscreen
,
7580 LLVMTargetMachineRef tm
,
7581 struct si_shader
*shader
,
7582 struct pipe_debug_callback
*debug
)
7584 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
7585 union si_shader_part_key prolog_key
;
7586 union si_shader_part_key epilog_key
;
7589 /* Get the prolog. */
7590 memset(&prolog_key
, 0, sizeof(prolog_key
));
7591 prolog_key
.ps_prolog
.states
= shader
->key
.ps
.prolog
;
7592 prolog_key
.ps_prolog
.colors_read
= info
->colors_read
;
7593 prolog_key
.ps_prolog
.num_input_sgprs
= shader
->info
.num_input_sgprs
;
7594 prolog_key
.ps_prolog
.num_input_vgprs
= shader
->info
.num_input_vgprs
;
7595 prolog_key
.ps_prolog
.wqm
= info
->uses_derivatives
&&
7596 (prolog_key
.ps_prolog
.colors_read
||
7597 prolog_key
.ps_prolog
.states
.force_persp_sample_interp
||
7598 prolog_key
.ps_prolog
.states
.force_linear_sample_interp
||
7599 prolog_key
.ps_prolog
.states
.force_persp_center_interp
||
7600 prolog_key
.ps_prolog
.states
.force_linear_center_interp
||
7601 prolog_key
.ps_prolog
.states
.bc_optimize_for_persp
||
7602 prolog_key
.ps_prolog
.states
.bc_optimize_for_linear
);
7604 if (info
->colors_read
) {
7605 unsigned *color
= shader
->selector
->color_attr_index
;
7607 if (shader
->key
.ps
.prolog
.color_two_side
) {
7608 /* BCOLORs are stored after the last input. */
7609 prolog_key
.ps_prolog
.num_interp_inputs
= info
->num_inputs
;
7610 prolog_key
.ps_prolog
.face_vgpr_index
= shader
->info
.face_vgpr_index
;
7611 shader
->config
.spi_ps_input_ena
|= S_0286CC_FRONT_FACE_ENA(1);
7614 for (i
= 0; i
< 2; i
++) {
7615 unsigned interp
= info
->input_interpolate
[color
[i
]];
7616 unsigned location
= info
->input_interpolate_loc
[color
[i
]];
7618 if (!(info
->colors_read
& (0xf << i
*4)))
7621 prolog_key
.ps_prolog
.color_attr_index
[i
] = color
[i
];
7623 if (shader
->key
.ps
.prolog
.flatshade_colors
&&
7624 interp
== TGSI_INTERPOLATE_COLOR
)
7625 interp
= TGSI_INTERPOLATE_CONSTANT
;
7628 case TGSI_INTERPOLATE_CONSTANT
:
7629 prolog_key
.ps_prolog
.color_interp_vgpr_index
[i
] = -1;
7631 case TGSI_INTERPOLATE_PERSPECTIVE
:
7632 case TGSI_INTERPOLATE_COLOR
:
7633 /* Force the interpolation location for colors here. */
7634 if (shader
->key
.ps
.prolog
.force_persp_sample_interp
)
7635 location
= TGSI_INTERPOLATE_LOC_SAMPLE
;
7636 if (shader
->key
.ps
.prolog
.force_persp_center_interp
)
7637 location
= TGSI_INTERPOLATE_LOC_CENTER
;
7640 case TGSI_INTERPOLATE_LOC_SAMPLE
:
7641 prolog_key
.ps_prolog
.color_interp_vgpr_index
[i
] = 0;
7642 shader
->config
.spi_ps_input_ena
|=
7643 S_0286CC_PERSP_SAMPLE_ENA(1);
7645 case TGSI_INTERPOLATE_LOC_CENTER
:
7646 prolog_key
.ps_prolog
.color_interp_vgpr_index
[i
] = 2;
7647 shader
->config
.spi_ps_input_ena
|=
7648 S_0286CC_PERSP_CENTER_ENA(1);
7650 case TGSI_INTERPOLATE_LOC_CENTROID
:
7651 prolog_key
.ps_prolog
.color_interp_vgpr_index
[i
] = 4;
7652 shader
->config
.spi_ps_input_ena
|=
7653 S_0286CC_PERSP_CENTROID_ENA(1);
7659 case TGSI_INTERPOLATE_LINEAR
:
7660 /* Force the interpolation location for colors here. */
7661 if (shader
->key
.ps
.prolog
.force_linear_sample_interp
)
7662 location
= TGSI_INTERPOLATE_LOC_SAMPLE
;
7663 if (shader
->key
.ps
.prolog
.force_linear_center_interp
)
7664 location
= TGSI_INTERPOLATE_LOC_CENTER
;
7667 case TGSI_INTERPOLATE_LOC_SAMPLE
:
7668 prolog_key
.ps_prolog
.color_interp_vgpr_index
[i
] = 6;
7669 shader
->config
.spi_ps_input_ena
|=
7670 S_0286CC_LINEAR_SAMPLE_ENA(1);
7672 case TGSI_INTERPOLATE_LOC_CENTER
:
7673 prolog_key
.ps_prolog
.color_interp_vgpr_index
[i
] = 8;
7674 shader
->config
.spi_ps_input_ena
|=
7675 S_0286CC_LINEAR_CENTER_ENA(1);
7677 case TGSI_INTERPOLATE_LOC_CENTROID
:
7678 prolog_key
.ps_prolog
.color_interp_vgpr_index
[i
] = 10;
7679 shader
->config
.spi_ps_input_ena
|=
7680 S_0286CC_LINEAR_CENTROID_ENA(1);
7692 /* The prolog is a no-op if these aren't set. */
7693 if (prolog_key
.ps_prolog
.colors_read
||
7694 prolog_key
.ps_prolog
.states
.force_persp_sample_interp
||
7695 prolog_key
.ps_prolog
.states
.force_linear_sample_interp
||
7696 prolog_key
.ps_prolog
.states
.force_persp_center_interp
||
7697 prolog_key
.ps_prolog
.states
.force_linear_center_interp
||
7698 prolog_key
.ps_prolog
.states
.bc_optimize_for_persp
||
7699 prolog_key
.ps_prolog
.states
.bc_optimize_for_linear
||
7700 prolog_key
.ps_prolog
.states
.poly_stipple
) {
7702 si_get_shader_part(sscreen
, &sscreen
->ps_prologs
,
7703 &prolog_key
, tm
, debug
,
7704 si_compile_ps_prolog
);
7705 if (!shader
->prolog
)
7709 /* Get the epilog. */
7710 memset(&epilog_key
, 0, sizeof(epilog_key
));
7711 epilog_key
.ps_epilog
.colors_written
= info
->colors_written
;
7712 epilog_key
.ps_epilog
.writes_z
= info
->writes_z
;
7713 epilog_key
.ps_epilog
.writes_stencil
= info
->writes_stencil
;
7714 epilog_key
.ps_epilog
.writes_samplemask
= info
->writes_samplemask
;
7715 epilog_key
.ps_epilog
.states
= shader
->key
.ps
.epilog
;
7718 si_get_shader_part(sscreen
, &sscreen
->ps_epilogs
,
7719 &epilog_key
, tm
, debug
,
7720 si_compile_ps_epilog
);
7721 if (!shader
->epilog
)
7724 /* Enable POS_FIXED_PT if polygon stippling is enabled. */
7725 if (shader
->key
.ps
.prolog
.poly_stipple
) {
7726 shader
->config
.spi_ps_input_ena
|= S_0286CC_POS_FIXED_PT_ENA(1);
7727 assert(G_0286CC_POS_FIXED_PT_ENA(shader
->config
.spi_ps_input_addr
));
7730 /* Set up the enable bits for per-sample shading if needed. */
7731 if (shader
->key
.ps
.prolog
.force_persp_sample_interp
&&
7732 (G_0286CC_PERSP_CENTER_ENA(shader
->config
.spi_ps_input_ena
) ||
7733 G_0286CC_PERSP_CENTROID_ENA(shader
->config
.spi_ps_input_ena
))) {
7734 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_CENTER_ENA
;
7735 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_CENTROID_ENA
;
7736 shader
->config
.spi_ps_input_ena
|= S_0286CC_PERSP_SAMPLE_ENA(1);
7738 if (shader
->key
.ps
.prolog
.force_linear_sample_interp
&&
7739 (G_0286CC_LINEAR_CENTER_ENA(shader
->config
.spi_ps_input_ena
) ||
7740 G_0286CC_LINEAR_CENTROID_ENA(shader
->config
.spi_ps_input_ena
))) {
7741 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_CENTER_ENA
;
7742 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_CENTROID_ENA
;
7743 shader
->config
.spi_ps_input_ena
|= S_0286CC_LINEAR_SAMPLE_ENA(1);
7745 if (shader
->key
.ps
.prolog
.force_persp_center_interp
&&
7746 (G_0286CC_PERSP_SAMPLE_ENA(shader
->config
.spi_ps_input_ena
) ||
7747 G_0286CC_PERSP_CENTROID_ENA(shader
->config
.spi_ps_input_ena
))) {
7748 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_SAMPLE_ENA
;
7749 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_CENTROID_ENA
;
7750 shader
->config
.spi_ps_input_ena
|= S_0286CC_PERSP_CENTER_ENA(1);
7752 if (shader
->key
.ps
.prolog
.force_linear_center_interp
&&
7753 (G_0286CC_LINEAR_SAMPLE_ENA(shader
->config
.spi_ps_input_ena
) ||
7754 G_0286CC_LINEAR_CENTROID_ENA(shader
->config
.spi_ps_input_ena
))) {
7755 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_SAMPLE_ENA
;
7756 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_CENTROID_ENA
;
7757 shader
->config
.spi_ps_input_ena
|= S_0286CC_LINEAR_CENTER_ENA(1);
7760 /* POW_W_FLOAT requires that one of the perspective weights is enabled. */
7761 if (G_0286CC_POS_W_FLOAT_ENA(shader
->config
.spi_ps_input_ena
) &&
7762 !(shader
->config
.spi_ps_input_ena
& 0xf)) {
7763 shader
->config
.spi_ps_input_ena
|= S_0286CC_PERSP_CENTER_ENA(1);
7764 assert(G_0286CC_PERSP_CENTER_ENA(shader
->config
.spi_ps_input_addr
));
7767 /* At least one pair of interpolation weights must be enabled. */
7768 if (!(shader
->config
.spi_ps_input_ena
& 0x7f)) {
7769 shader
->config
.spi_ps_input_ena
|= S_0286CC_LINEAR_CENTER_ENA(1);
7770 assert(G_0286CC_LINEAR_CENTER_ENA(shader
->config
.spi_ps_input_addr
));
7773 /* The sample mask input is always enabled, because the API shader always
7774 * passes it through to the epilog. Disable it here if it's unused.
7776 if (!shader
->key
.ps
.epilog
.poly_line_smoothing
&&
7777 !shader
->selector
->info
.reads_samplemask
)
7778 shader
->config
.spi_ps_input_ena
&= C_0286CC_SAMPLE_COVERAGE_ENA
;
7783 static void si_fix_num_sgprs(struct si_shader
*shader
)
7785 unsigned min_sgprs
= shader
->info
.num_input_sgprs
+ 2; /* VCC */
7787 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
, min_sgprs
);
7790 int si_shader_create(struct si_screen
*sscreen
, LLVMTargetMachineRef tm
,
7791 struct si_shader
*shader
,
7792 struct pipe_debug_callback
*debug
)
7794 struct si_shader
*mainp
= shader
->selector
->main_shader_part
;
7797 /* LS, ES, VS are compiled on demand if the main part hasn't been
7798 * compiled for that stage.
7801 (shader
->selector
->type
== PIPE_SHADER_VERTEX
&&
7802 (shader
->key
.vs
.as_es
!= mainp
->key
.vs
.as_es
||
7803 shader
->key
.vs
.as_ls
!= mainp
->key
.vs
.as_ls
)) ||
7804 (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
&&
7805 shader
->key
.tes
.as_es
!= mainp
->key
.tes
.as_es
) ||
7806 (shader
->selector
->type
== PIPE_SHADER_TESS_CTRL
&&
7807 shader
->key
.tcs
.epilog
.inputs_to_copy
) ||
7808 shader
->selector
->type
== PIPE_SHADER_COMPUTE
) {
7809 /* Monolithic shader (compiled as a whole, has many variants,
7810 * may take a long time to compile).
7812 r
= si_compile_tgsi_shader(sscreen
, tm
, shader
, true, debug
);
7816 /* The shader consists of 2-3 parts:
7818 * - the middle part is the user shader, it has 1 variant only
7819 * and it was compiled during the creation of the shader
7821 * - the prolog part is inserted at the beginning
7822 * - the epilog part is inserted at the end
7824 * The prolog and epilog have many (but simple) variants.
7827 /* Copy the compiled TGSI shader data over. */
7828 shader
->is_binary_shared
= true;
7829 shader
->binary
= mainp
->binary
;
7830 shader
->config
= mainp
->config
;
7831 shader
->info
.num_input_sgprs
= mainp
->info
.num_input_sgprs
;
7832 shader
->info
.num_input_vgprs
= mainp
->info
.num_input_vgprs
;
7833 shader
->info
.face_vgpr_index
= mainp
->info
.face_vgpr_index
;
7834 memcpy(shader
->info
.vs_output_param_offset
,
7835 mainp
->info
.vs_output_param_offset
,
7836 sizeof(mainp
->info
.vs_output_param_offset
));
7837 shader
->info
.uses_instanceid
= mainp
->info
.uses_instanceid
;
7838 shader
->info
.nr_pos_exports
= mainp
->info
.nr_pos_exports
;
7839 shader
->info
.nr_param_exports
= mainp
->info
.nr_param_exports
;
7841 /* Select prologs and/or epilogs. */
7842 switch (shader
->selector
->type
) {
7843 case PIPE_SHADER_VERTEX
:
7844 if (!si_shader_select_vs_parts(sscreen
, tm
, shader
, debug
))
7847 case PIPE_SHADER_TESS_CTRL
:
7848 if (!si_shader_select_tcs_parts(sscreen
, tm
, shader
, debug
))
7851 case PIPE_SHADER_TESS_EVAL
:
7852 if (!si_shader_select_tes_parts(sscreen
, tm
, shader
, debug
))
7855 case PIPE_SHADER_FRAGMENT
:
7856 if (!si_shader_select_ps_parts(sscreen
, tm
, shader
, debug
))
7859 /* Make sure we have at least as many VGPRs as there
7860 * are allocated inputs.
7862 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
7863 shader
->info
.num_input_vgprs
);
7867 /* Update SGPR and VGPR counts. */
7868 if (shader
->prolog
) {
7869 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
,
7870 shader
->prolog
->config
.num_sgprs
);
7871 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
7872 shader
->prolog
->config
.num_vgprs
);
7874 if (shader
->epilog
) {
7875 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
,
7876 shader
->epilog
->config
.num_sgprs
);
7877 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
7878 shader
->epilog
->config
.num_vgprs
);
7882 si_fix_num_sgprs(shader
);
7883 si_shader_dump(sscreen
, shader
, debug
, shader
->selector
->info
.processor
,
7887 r
= si_shader_binary_upload(sscreen
, shader
);
7889 fprintf(stderr
, "LLVM failed to upload shader\n");
7896 void si_shader_destroy(struct si_shader
*shader
)
7898 if (shader
->gs_copy_shader
) {
7899 si_shader_destroy(shader
->gs_copy_shader
);
7900 FREE(shader
->gs_copy_shader
);
7903 if (shader
->scratch_bo
)
7904 r600_resource_reference(&shader
->scratch_bo
, NULL
);
7906 r600_resource_reference(&shader
->bo
, NULL
);
7908 if (!shader
->is_binary_shared
)
7909 radeon_shader_binary_clean(&shader
->binary
);
7911 free(shader
->shader_log
);