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_bitarit.h"
35 #include "gallivm/lp_bld_flow.h"
36 #include "radeon/r600_cs.h"
37 #include "radeon/radeon_llvm.h"
38 #include "radeon/radeon_elf_util.h"
39 #include "radeon/radeon_llvm_emit.h"
40 #include "util/u_memory.h"
41 #include "util/u_pstipple.h"
42 #include "util/u_string.h"
43 #include "tgsi/tgsi_parse.h"
44 #include "tgsi/tgsi_build.h"
45 #include "tgsi/tgsi_util.h"
46 #include "tgsi/tgsi_dump.h"
49 #include "si_shader.h"
54 static const char *scratch_rsrc_dword0_symbol
=
55 "SCRATCH_RSRC_DWORD0";
57 static const char *scratch_rsrc_dword1_symbol
=
58 "SCRATCH_RSRC_DWORD1";
60 struct si_shader_output_values
62 LLVMValueRef values
[4];
67 struct si_shader_context
69 struct radeon_llvm_context radeon_bld
;
70 struct si_shader
*shader
;
71 struct si_screen
*screen
;
73 unsigned type
; /* PIPE_SHADER_* specifies the type of shader. */
74 bool is_gs_copy_shader
;
76 /* Whether to generate the optimized shader variant compiled as a whole
77 * (without a prolog and epilog)
81 int param_streamout_config
;
82 int param_streamout_write_index
;
83 int param_streamout_offset
[4];
85 int param_rel_auto_id
;
87 int param_instance_id
;
88 int param_vertex_index0
;
91 int param_tes_rel_patch_id
;
92 int param_tes_patch_id
;
93 int param_es2gs_offset
;
95 LLVMTargetMachineRef tm
;
97 LLVMValueRef const_md
;
98 LLVMValueRef const_buffers
[SI_NUM_CONST_BUFFERS
];
100 LLVMValueRef
*constants
[SI_NUM_CONST_BUFFERS
];
101 LLVMValueRef shader_buffers
[SI_NUM_SHADER_BUFFERS
];
102 LLVMValueRef sampler_views
[SI_NUM_SAMPLERS
];
103 LLVMValueRef sampler_states
[SI_NUM_SAMPLERS
];
104 LLVMValueRef fmasks
[SI_NUM_SAMPLERS
];
105 LLVMValueRef images
[SI_NUM_IMAGES
];
106 LLVMValueRef so_buffers
[4];
107 LLVMValueRef esgs_ring
;
108 LLVMValueRef gsvs_ring
[4];
109 LLVMValueRef gs_next_vertex
[4];
110 LLVMValueRef return_value
;
125 LLVMValueRef shared_memory
;
128 static struct si_shader_context
*si_shader_context(
129 struct lp_build_tgsi_context
*bld_base
)
131 return (struct si_shader_context
*)bld_base
;
134 static void si_init_shader_ctx(struct si_shader_context
*ctx
,
135 struct si_screen
*sscreen
,
136 struct si_shader
*shader
,
137 LLVMTargetMachineRef tm
);
139 /* Ideally pass the sample mask input to the PS epilog as v13, which
140 * is its usual location, so that the shader doesn't have to add v_mov.
142 #define PS_EPILOG_SAMPLEMASK_MIN_LOC 13
144 /* The VS location of the PrimitiveID input is the same in the epilog,
145 * so that the main shader part doesn't have to move it.
147 #define VS_EPILOG_PRIMID_LOC 2
149 #define PERSPECTIVE_BASE 0
150 #define LINEAR_BASE 9
152 #define SAMPLE_OFFSET 0
153 #define CENTER_OFFSET 2
154 #define CENTROID_OFSET 4
156 #define USE_SGPR_MAX_SUFFIX_LEN 5
157 #define CONST_ADDR_SPACE 2
158 #define LOCAL_ADDR_SPACE 3
159 #define USER_SGPR_ADDR_SPACE 8
163 #define SENDMSG_GS_DONE 3
165 #define SENDMSG_GS_OP_NOP (0 << 4)
166 #define SENDMSG_GS_OP_CUT (1 << 4)
167 #define SENDMSG_GS_OP_EMIT (2 << 4)
168 #define SENDMSG_GS_OP_EMIT_CUT (3 << 4)
171 * Returns a unique index for a semantic name and index. The index must be
172 * less than 64, so that a 64-bit bitmask of used inputs or outputs can be
175 unsigned si_shader_io_get_unique_index(unsigned semantic_name
, unsigned index
)
177 switch (semantic_name
) {
178 case TGSI_SEMANTIC_POSITION
:
180 case TGSI_SEMANTIC_PSIZE
:
182 case TGSI_SEMANTIC_CLIPDIST
:
185 case TGSI_SEMANTIC_GENERIC
:
189 /* same explanation as in the default statement,
190 * the only user hitting this is st/nine.
194 /* patch indices are completely separate and thus start from 0 */
195 case TGSI_SEMANTIC_TESSOUTER
:
197 case TGSI_SEMANTIC_TESSINNER
:
199 case TGSI_SEMANTIC_PATCH
:
203 /* Don't fail here. The result of this function is only used
204 * for LS, TCS, TES, and GS, where legacy GL semantics can't
205 * occur, but this function is called for all vertex shaders
206 * before it's known whether LS will be compiled or not.
213 * Get the value of a shader input parameter and extract a bitfield.
215 static LLVMValueRef
unpack_param(struct si_shader_context
*ctx
,
216 unsigned param
, unsigned rshift
,
219 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
220 LLVMValueRef value
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
223 if (LLVMGetTypeKind(LLVMTypeOf(value
)) == LLVMFloatTypeKind
)
224 value
= bitcast(&ctx
->radeon_bld
.soa
.bld_base
,
225 TGSI_TYPE_UNSIGNED
, value
);
228 value
= LLVMBuildLShr(gallivm
->builder
, value
,
229 lp_build_const_int32(gallivm
, rshift
), "");
231 if (rshift
+ bitwidth
< 32) {
232 unsigned mask
= (1 << bitwidth
) - 1;
233 value
= LLVMBuildAnd(gallivm
->builder
, value
,
234 lp_build_const_int32(gallivm
, mask
), "");
240 static LLVMValueRef
get_rel_patch_id(struct si_shader_context
*ctx
)
243 case PIPE_SHADER_TESS_CTRL
:
244 return unpack_param(ctx
, SI_PARAM_REL_IDS
, 0, 8);
246 case PIPE_SHADER_TESS_EVAL
:
247 return LLVMGetParam(ctx
->radeon_bld
.main_fn
,
248 ctx
->param_tes_rel_patch_id
);
256 /* Tessellation shaders pass outputs to the next shader using LDS.
258 * LS outputs = TCS inputs
259 * TCS outputs = TES inputs
262 * - TCS inputs for patch 0
263 * - TCS inputs for patch 1
264 * - TCS inputs for patch 2 = get_tcs_in_current_patch_offset (if RelPatchID==2)
266 * - TCS outputs for patch 0 = get_tcs_out_patch0_offset
267 * - Per-patch TCS outputs for patch 0 = get_tcs_out_patch0_patch_data_offset
268 * - TCS outputs for patch 1
269 * - Per-patch TCS outputs for patch 1
270 * - TCS outputs for patch 2 = get_tcs_out_current_patch_offset (if RelPatchID==2)
271 * - Per-patch TCS outputs for patch 2 = get_tcs_out_current_patch_data_offset (if RelPatchID==2)
274 * All three shaders VS(LS), TCS, TES share the same LDS space.
278 get_tcs_in_patch_stride(struct si_shader_context
*ctx
)
280 if (ctx
->type
== PIPE_SHADER_VERTEX
)
281 return unpack_param(ctx
, SI_PARAM_LS_OUT_LAYOUT
, 0, 13);
282 else if (ctx
->type
== PIPE_SHADER_TESS_CTRL
)
283 return unpack_param(ctx
, SI_PARAM_TCS_IN_LAYOUT
, 0, 13);
291 get_tcs_out_patch_stride(struct si_shader_context
*ctx
)
293 return unpack_param(ctx
, SI_PARAM_TCS_OUT_LAYOUT
, 0, 13);
297 get_tcs_out_patch0_offset(struct si_shader_context
*ctx
)
299 return lp_build_mul_imm(&ctx
->radeon_bld
.soa
.bld_base
.uint_bld
,
301 SI_PARAM_TCS_OUT_OFFSETS
,
307 get_tcs_out_patch0_patch_data_offset(struct si_shader_context
*ctx
)
309 return lp_build_mul_imm(&ctx
->radeon_bld
.soa
.bld_base
.uint_bld
,
311 SI_PARAM_TCS_OUT_OFFSETS
,
317 get_tcs_in_current_patch_offset(struct si_shader_context
*ctx
)
319 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
320 LLVMValueRef patch_stride
= get_tcs_in_patch_stride(ctx
);
321 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
323 return LLVMBuildMul(gallivm
->builder
, patch_stride
, rel_patch_id
, "");
327 get_tcs_out_current_patch_offset(struct si_shader_context
*ctx
)
329 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
330 LLVMValueRef patch0_offset
= get_tcs_out_patch0_offset(ctx
);
331 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(ctx
);
332 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
334 return LLVMBuildAdd(gallivm
->builder
, patch0_offset
,
335 LLVMBuildMul(gallivm
->builder
, patch_stride
,
341 get_tcs_out_current_patch_data_offset(struct si_shader_context
*ctx
)
343 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
344 LLVMValueRef patch0_patch_data_offset
=
345 get_tcs_out_patch0_patch_data_offset(ctx
);
346 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(ctx
);
347 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
349 return LLVMBuildAdd(gallivm
->builder
, patch0_patch_data_offset
,
350 LLVMBuildMul(gallivm
->builder
, patch_stride
,
355 static void build_indexed_store(struct si_shader_context
*ctx
,
356 LLVMValueRef base_ptr
, LLVMValueRef index
,
359 struct lp_build_tgsi_context
*bld_base
= &ctx
->radeon_bld
.soa
.bld_base
;
360 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
361 LLVMValueRef indices
[2], pointer
;
363 indices
[0] = bld_base
->uint_bld
.zero
;
366 pointer
= LLVMBuildGEP(gallivm
->builder
, base_ptr
, indices
, 2, "");
367 LLVMBuildStore(gallivm
->builder
, value
, pointer
);
371 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
372 * It's equivalent to doing a load from &base_ptr[index].
374 * \param base_ptr Where the array starts.
375 * \param index The element index into the array.
377 static LLVMValueRef
build_indexed_load(struct si_shader_context
*ctx
,
378 LLVMValueRef base_ptr
, LLVMValueRef index
)
380 struct lp_build_tgsi_context
*bld_base
= &ctx
->radeon_bld
.soa
.bld_base
;
381 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
382 LLVMValueRef indices
[2], pointer
;
384 indices
[0] = bld_base
->uint_bld
.zero
;
387 pointer
= LLVMBuildGEP(gallivm
->builder
, base_ptr
, indices
, 2, "");
388 return LLVMBuildLoad(gallivm
->builder
, pointer
, "");
392 * Do a load from &base_ptr[index], but also add a flag that it's loading
395 static LLVMValueRef
build_indexed_load_const(
396 struct si_shader_context
*ctx
,
397 LLVMValueRef base_ptr
, LLVMValueRef index
)
399 LLVMValueRef result
= build_indexed_load(ctx
, base_ptr
, index
);
400 LLVMSetMetadata(result
, 1, ctx
->const_md
);
404 static LLVMValueRef
get_instance_index_for_fetch(
405 struct radeon_llvm_context
*radeon_bld
,
406 unsigned param_start_instance
, unsigned divisor
)
408 struct si_shader_context
*ctx
=
409 si_shader_context(&radeon_bld
->soa
.bld_base
);
410 struct gallivm_state
*gallivm
= radeon_bld
->soa
.bld_base
.base
.gallivm
;
412 LLVMValueRef result
= LLVMGetParam(radeon_bld
->main_fn
,
413 ctx
->param_instance_id
);
415 /* The division must be done before START_INSTANCE is added. */
417 result
= LLVMBuildUDiv(gallivm
->builder
, result
,
418 lp_build_const_int32(gallivm
, divisor
), "");
420 return LLVMBuildAdd(gallivm
->builder
, result
,
421 LLVMGetParam(radeon_bld
->main_fn
, param_start_instance
), "");
424 static void declare_input_vs(
425 struct radeon_llvm_context
*radeon_bld
,
426 unsigned input_index
,
427 const struct tgsi_full_declaration
*decl
)
429 struct lp_build_context
*base
= &radeon_bld
->soa
.bld_base
.base
;
430 struct gallivm_state
*gallivm
= base
->gallivm
;
431 struct si_shader_context
*ctx
=
432 si_shader_context(&radeon_bld
->soa
.bld_base
);
434 ctx
->shader
->key
.vs
.prolog
.instance_divisors
[input_index
];
438 LLVMValueRef t_list_ptr
;
439 LLVMValueRef t_offset
;
441 LLVMValueRef attribute_offset
;
442 LLVMValueRef buffer_index
;
443 LLVMValueRef args
[3];
446 /* Load the T list */
447 t_list_ptr
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_VERTEX_BUFFERS
);
449 t_offset
= lp_build_const_int32(gallivm
, input_index
);
451 t_list
= build_indexed_load_const(ctx
, t_list_ptr
, t_offset
);
453 /* Build the attribute offset */
454 attribute_offset
= lp_build_const_int32(gallivm
, 0);
456 if (!ctx
->is_monolithic
) {
457 buffer_index
= LLVMGetParam(radeon_bld
->main_fn
,
458 ctx
->param_vertex_index0
+
460 } else if (divisor
) {
461 /* Build index from instance ID, start instance and divisor */
462 ctx
->shader
->info
.uses_instanceid
= true;
463 buffer_index
= get_instance_index_for_fetch(&ctx
->radeon_bld
,
464 SI_PARAM_START_INSTANCE
,
467 /* Load the buffer index for vertices. */
468 LLVMValueRef vertex_id
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
469 ctx
->param_vertex_id
);
470 LLVMValueRef base_vertex
= LLVMGetParam(radeon_bld
->main_fn
,
471 SI_PARAM_BASE_VERTEX
);
472 buffer_index
= LLVMBuildAdd(gallivm
->builder
, base_vertex
, vertex_id
, "");
476 args
[1] = attribute_offset
;
477 args
[2] = buffer_index
;
478 input
= lp_build_intrinsic(gallivm
->builder
,
479 "llvm.SI.vs.load.input", ctx
->v4f32
, args
, 3,
480 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
482 /* Break up the vec4 into individual components */
483 for (chan
= 0; chan
< 4; chan
++) {
484 LLVMValueRef llvm_chan
= lp_build_const_int32(gallivm
, chan
);
485 /* XXX: Use a helper function for this. There is one in
487 ctx
->radeon_bld
.inputs
[radeon_llvm_reg_index_soa(input_index
, chan
)] =
488 LLVMBuildExtractElement(gallivm
->builder
,
489 input
, llvm_chan
, "");
493 static LLVMValueRef
get_primitive_id(struct lp_build_tgsi_context
*bld_base
,
496 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
499 return bld_base
->uint_bld
.zero
;
502 case PIPE_SHADER_VERTEX
:
503 return LLVMGetParam(ctx
->radeon_bld
.main_fn
,
504 ctx
->param_vs_prim_id
);
505 case PIPE_SHADER_TESS_CTRL
:
506 return LLVMGetParam(ctx
->radeon_bld
.main_fn
,
508 case PIPE_SHADER_TESS_EVAL
:
509 return LLVMGetParam(ctx
->radeon_bld
.main_fn
,
510 ctx
->param_tes_patch_id
);
511 case PIPE_SHADER_GEOMETRY
:
512 return LLVMGetParam(ctx
->radeon_bld
.main_fn
,
513 SI_PARAM_PRIMITIVE_ID
);
516 return bld_base
->uint_bld
.zero
;
521 * Return the value of tgsi_ind_register for indexing.
522 * This is the indirect index with the constant offset added to it.
524 static LLVMValueRef
get_indirect_index(struct si_shader_context
*ctx
,
525 const struct tgsi_ind_register
*ind
,
528 struct gallivm_state
*gallivm
= ctx
->radeon_bld
.soa
.bld_base
.base
.gallivm
;
531 result
= ctx
->radeon_bld
.soa
.addr
[ind
->Index
][ind
->Swizzle
];
532 result
= LLVMBuildLoad(gallivm
->builder
, result
, "");
533 result
= LLVMBuildAdd(gallivm
->builder
, result
,
534 lp_build_const_int32(gallivm
, rel_index
), "");
539 * Like get_indirect_index, but restricts the return value to a (possibly
540 * undefined) value inside [0..num).
542 static LLVMValueRef
get_bounded_indirect_index(struct si_shader_context
*ctx
,
543 const struct tgsi_ind_register
*ind
,
544 int rel_index
, unsigned num
)
546 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
547 LLVMBuilderRef builder
= gallivm
->builder
;
548 LLVMValueRef result
= get_indirect_index(ctx
, ind
, rel_index
);
549 LLVMValueRef c_max
= LLVMConstInt(ctx
->i32
, num
- 1, 0);
552 if (util_is_power_of_two(num
)) {
553 result
= LLVMBuildAnd(builder
, result
, c_max
, "");
555 /* In theory, this MAX pattern should result in code that is
556 * as good as the bit-wise AND above.
558 * In practice, LLVM generates worse code (at the time of
559 * writing), because its value tracking is not strong enough.
561 cc
= LLVMBuildICmp(builder
, LLVMIntULE
, result
, c_max
, "");
562 result
= LLVMBuildSelect(builder
, cc
, result
, c_max
, "");
570 * Calculate a dword address given an input or output register and a stride.
572 static LLVMValueRef
get_dw_address(struct si_shader_context
*ctx
,
573 const struct tgsi_full_dst_register
*dst
,
574 const struct tgsi_full_src_register
*src
,
575 LLVMValueRef vertex_dw_stride
,
576 LLVMValueRef base_addr
)
578 struct gallivm_state
*gallivm
= ctx
->radeon_bld
.soa
.bld_base
.base
.gallivm
;
579 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
580 ubyte
*name
, *index
, *array_first
;
582 struct tgsi_full_dst_register reg
;
584 /* Set the register description. The address computation is the same
585 * for sources and destinations. */
587 reg
.Register
.File
= src
->Register
.File
;
588 reg
.Register
.Index
= src
->Register
.Index
;
589 reg
.Register
.Indirect
= src
->Register
.Indirect
;
590 reg
.Register
.Dimension
= src
->Register
.Dimension
;
591 reg
.Indirect
= src
->Indirect
;
592 reg
.Dimension
= src
->Dimension
;
593 reg
.DimIndirect
= src
->DimIndirect
;
597 /* If the register is 2-dimensional (e.g. an array of vertices
598 * in a primitive), calculate the base address of the vertex. */
599 if (reg
.Register
.Dimension
) {
602 if (reg
.Dimension
.Indirect
)
603 index
= get_indirect_index(ctx
, ®
.DimIndirect
,
604 reg
.Dimension
.Index
);
606 index
= lp_build_const_int32(gallivm
, reg
.Dimension
.Index
);
608 base_addr
= LLVMBuildAdd(gallivm
->builder
, base_addr
,
609 LLVMBuildMul(gallivm
->builder
, index
,
610 vertex_dw_stride
, ""), "");
613 /* Get information about the register. */
614 if (reg
.Register
.File
== TGSI_FILE_INPUT
) {
615 name
= info
->input_semantic_name
;
616 index
= info
->input_semantic_index
;
617 array_first
= info
->input_array_first
;
618 } else if (reg
.Register
.File
== TGSI_FILE_OUTPUT
) {
619 name
= info
->output_semantic_name
;
620 index
= info
->output_semantic_index
;
621 array_first
= info
->output_array_first
;
627 if (reg
.Register
.Indirect
) {
628 /* Add the relative address of the element. */
629 LLVMValueRef ind_index
;
631 if (reg
.Indirect
.ArrayID
)
632 first
= array_first
[reg
.Indirect
.ArrayID
];
634 first
= reg
.Register
.Index
;
636 ind_index
= get_indirect_index(ctx
, ®
.Indirect
,
637 reg
.Register
.Index
- first
);
639 base_addr
= LLVMBuildAdd(gallivm
->builder
, base_addr
,
640 LLVMBuildMul(gallivm
->builder
, ind_index
,
641 lp_build_const_int32(gallivm
, 4), ""), "");
643 param
= si_shader_io_get_unique_index(name
[first
], index
[first
]);
645 param
= si_shader_io_get_unique_index(name
[reg
.Register
.Index
],
646 index
[reg
.Register
.Index
]);
649 /* Add the base address of the element. */
650 return LLVMBuildAdd(gallivm
->builder
, base_addr
,
651 lp_build_const_int32(gallivm
, param
* 4), "");
657 * \param type output value type
658 * \param swizzle offset (typically 0..3); it can be ~0, which loads a vec4
659 * \param dw_addr address in dwords
661 static LLVMValueRef
lds_load(struct lp_build_tgsi_context
*bld_base
,
662 enum tgsi_opcode_type type
, unsigned swizzle
,
663 LLVMValueRef dw_addr
)
665 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
666 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
670 LLVMValueRef values
[TGSI_NUM_CHANNELS
];
672 for (unsigned chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++)
673 values
[chan
] = lds_load(bld_base
, type
, chan
, dw_addr
);
675 return lp_build_gather_values(bld_base
->base
.gallivm
, values
,
679 dw_addr
= lp_build_add(&bld_base
->uint_bld
, dw_addr
,
680 lp_build_const_int32(gallivm
, swizzle
));
682 value
= build_indexed_load(ctx
, ctx
->lds
, dw_addr
);
683 if (type
== TGSI_TYPE_DOUBLE
) {
685 dw_addr
= lp_build_add(&bld_base
->uint_bld
, dw_addr
,
686 lp_build_const_int32(gallivm
, swizzle
+ 1));
687 value2
= build_indexed_load(ctx
, ctx
->lds
, dw_addr
);
688 return radeon_llvm_emit_fetch_double(bld_base
, value
, value2
);
691 return LLVMBuildBitCast(gallivm
->builder
, value
,
692 tgsi2llvmtype(bld_base
, type
), "");
698 * \param swizzle offset (typically 0..3)
699 * \param dw_addr address in dwords
700 * \param value value to store
702 static void lds_store(struct lp_build_tgsi_context
*bld_base
,
703 unsigned swizzle
, LLVMValueRef dw_addr
,
706 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
707 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
709 dw_addr
= lp_build_add(&bld_base
->uint_bld
, dw_addr
,
710 lp_build_const_int32(gallivm
, swizzle
));
712 value
= LLVMBuildBitCast(gallivm
->builder
, value
, ctx
->i32
, "");
713 build_indexed_store(ctx
, ctx
->lds
,
717 static LLVMValueRef
fetch_input_tcs(
718 struct lp_build_tgsi_context
*bld_base
,
719 const struct tgsi_full_src_register
*reg
,
720 enum tgsi_opcode_type type
, unsigned swizzle
)
722 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
723 LLVMValueRef dw_addr
, stride
;
725 stride
= unpack_param(ctx
, SI_PARAM_TCS_IN_LAYOUT
, 13, 8);
726 dw_addr
= get_tcs_in_current_patch_offset(ctx
);
727 dw_addr
= get_dw_address(ctx
, NULL
, reg
, stride
, dw_addr
);
729 return lds_load(bld_base
, type
, swizzle
, dw_addr
);
732 static LLVMValueRef
fetch_output_tcs(
733 struct lp_build_tgsi_context
*bld_base
,
734 const struct tgsi_full_src_register
*reg
,
735 enum tgsi_opcode_type type
, unsigned swizzle
)
737 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
738 LLVMValueRef dw_addr
, stride
;
740 if (reg
->Register
.Dimension
) {
741 stride
= unpack_param(ctx
, SI_PARAM_TCS_OUT_LAYOUT
, 13, 8);
742 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
743 dw_addr
= get_dw_address(ctx
, NULL
, reg
, stride
, dw_addr
);
745 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
746 dw_addr
= get_dw_address(ctx
, NULL
, reg
, NULL
, dw_addr
);
749 return lds_load(bld_base
, type
, swizzle
, dw_addr
);
752 static LLVMValueRef
fetch_input_tes(
753 struct lp_build_tgsi_context
*bld_base
,
754 const struct tgsi_full_src_register
*reg
,
755 enum tgsi_opcode_type type
, unsigned swizzle
)
757 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
758 LLVMValueRef dw_addr
, stride
;
760 if (reg
->Register
.Dimension
) {
761 stride
= unpack_param(ctx
, SI_PARAM_TCS_OUT_LAYOUT
, 13, 8);
762 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
763 dw_addr
= get_dw_address(ctx
, NULL
, reg
, stride
, dw_addr
);
765 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
766 dw_addr
= get_dw_address(ctx
, NULL
, reg
, NULL
, dw_addr
);
769 return lds_load(bld_base
, type
, swizzle
, dw_addr
);
772 static void store_output_tcs(struct lp_build_tgsi_context
*bld_base
,
773 const struct tgsi_full_instruction
*inst
,
774 const struct tgsi_opcode_info
*info
,
777 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
778 const struct tgsi_full_dst_register
*reg
= &inst
->Dst
[0];
780 LLVMValueRef dw_addr
, stride
;
782 /* Only handle per-patch and per-vertex outputs here.
783 * Vectors will be lowered to scalars and this function will be called again.
785 if (reg
->Register
.File
!= TGSI_FILE_OUTPUT
||
786 (dst
[0] && LLVMGetTypeKind(LLVMTypeOf(dst
[0])) == LLVMVectorTypeKind
)) {
787 radeon_llvm_emit_store(bld_base
, inst
, info
, dst
);
791 if (reg
->Register
.Dimension
) {
792 stride
= unpack_param(ctx
, SI_PARAM_TCS_OUT_LAYOUT
, 13, 8);
793 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
794 dw_addr
= get_dw_address(ctx
, reg
, NULL
, stride
, dw_addr
);
796 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
797 dw_addr
= get_dw_address(ctx
, reg
, NULL
, NULL
, dw_addr
);
800 TGSI_FOR_EACH_DST0_ENABLED_CHANNEL(inst
, chan_index
) {
801 LLVMValueRef value
= dst
[chan_index
];
803 if (inst
->Instruction
.Saturate
)
804 value
= radeon_llvm_saturate(bld_base
, value
);
806 lds_store(bld_base
, chan_index
, dw_addr
, value
);
810 static LLVMValueRef
fetch_input_gs(
811 struct lp_build_tgsi_context
*bld_base
,
812 const struct tgsi_full_src_register
*reg
,
813 enum tgsi_opcode_type type
,
816 struct lp_build_context
*base
= &bld_base
->base
;
817 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
818 struct si_shader
*shader
= ctx
->shader
;
819 struct lp_build_context
*uint
= &ctx
->radeon_bld
.soa
.bld_base
.uint_bld
;
820 struct gallivm_state
*gallivm
= base
->gallivm
;
821 LLVMValueRef vtx_offset
;
822 LLVMValueRef args
[9];
823 unsigned vtx_offset_param
;
824 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
825 unsigned semantic_name
= info
->input_semantic_name
[reg
->Register
.Index
];
826 unsigned semantic_index
= info
->input_semantic_index
[reg
->Register
.Index
];
830 if (swizzle
!= ~0 && semantic_name
== TGSI_SEMANTIC_PRIMID
)
831 return get_primitive_id(bld_base
, swizzle
);
833 if (!reg
->Register
.Dimension
)
837 LLVMValueRef values
[TGSI_NUM_CHANNELS
];
839 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
840 values
[chan
] = fetch_input_gs(bld_base
, reg
, type
, chan
);
842 return lp_build_gather_values(bld_base
->base
.gallivm
, values
,
846 /* Get the vertex offset parameter */
847 vtx_offset_param
= reg
->Dimension
.Index
;
848 if (vtx_offset_param
< 2) {
849 vtx_offset_param
+= SI_PARAM_VTX0_OFFSET
;
851 assert(vtx_offset_param
< 6);
852 vtx_offset_param
+= SI_PARAM_VTX2_OFFSET
- 2;
854 vtx_offset
= lp_build_mul_imm(uint
,
855 LLVMGetParam(ctx
->radeon_bld
.main_fn
,
859 param
= si_shader_io_get_unique_index(semantic_name
, semantic_index
);
860 args
[0] = ctx
->esgs_ring
;
861 args
[1] = vtx_offset
;
862 args
[2] = lp_build_const_int32(gallivm
, (param
* 4 + swizzle
) * 256);
863 args
[3] = uint
->zero
;
864 args
[4] = uint
->one
; /* OFFEN */
865 args
[5] = uint
->zero
; /* IDXEN */
866 args
[6] = uint
->one
; /* GLC */
867 args
[7] = uint
->zero
; /* SLC */
868 args
[8] = uint
->zero
; /* TFE */
870 value
= lp_build_intrinsic(gallivm
->builder
,
871 "llvm.SI.buffer.load.dword.i32.i32",
873 LLVMReadOnlyAttribute
| LLVMNoUnwindAttribute
);
874 if (type
== TGSI_TYPE_DOUBLE
) {
876 args
[2] = lp_build_const_int32(gallivm
, (param
* 4 + swizzle
+ 1) * 256);
877 value2
= lp_build_intrinsic(gallivm
->builder
,
878 "llvm.SI.buffer.load.dword.i32.i32",
880 LLVMReadOnlyAttribute
| LLVMNoUnwindAttribute
);
881 return radeon_llvm_emit_fetch_double(bld_base
,
884 return LLVMBuildBitCast(gallivm
->builder
,
886 tgsi2llvmtype(bld_base
, type
), "");
889 static int lookup_interp_param_index(unsigned interpolate
, unsigned location
)
891 switch (interpolate
) {
892 case TGSI_INTERPOLATE_CONSTANT
:
895 case TGSI_INTERPOLATE_LINEAR
:
896 if (location
== TGSI_INTERPOLATE_LOC_SAMPLE
)
897 return SI_PARAM_LINEAR_SAMPLE
;
898 else if (location
== TGSI_INTERPOLATE_LOC_CENTROID
)
899 return SI_PARAM_LINEAR_CENTROID
;
901 return SI_PARAM_LINEAR_CENTER
;
903 case TGSI_INTERPOLATE_COLOR
:
904 case TGSI_INTERPOLATE_PERSPECTIVE
:
905 if (location
== TGSI_INTERPOLATE_LOC_SAMPLE
)
906 return SI_PARAM_PERSP_SAMPLE
;
907 else if (location
== TGSI_INTERPOLATE_LOC_CENTROID
)
908 return SI_PARAM_PERSP_CENTROID
;
910 return SI_PARAM_PERSP_CENTER
;
913 fprintf(stderr
, "Warning: Unhandled interpolation mode.\n");
918 /* This shouldn't be used by explicit INTERP opcodes. */
919 static unsigned select_interp_param(struct si_shader_context
*ctx
,
922 if (!ctx
->shader
->key
.ps
.prolog
.force_persample_interp
||
926 /* If the shader doesn't use center/centroid, just return the parameter.
928 * If the shader only uses one set of (i,j), "si_emit_spi_ps_input" can
929 * switch between center/centroid and sample without shader changes.
932 case SI_PARAM_PERSP_CENTROID
:
933 case SI_PARAM_PERSP_CENTER
:
934 return SI_PARAM_PERSP_SAMPLE
;
936 case SI_PARAM_LINEAR_CENTROID
:
937 case SI_PARAM_LINEAR_CENTER
:
938 return SI_PARAM_LINEAR_SAMPLE
;
946 * Interpolate a fragment shader input.
949 * @param input_index index of the input in hardware
950 * @param semantic_name TGSI_SEMANTIC_*
951 * @param semantic_index semantic index
952 * @param num_interp_inputs number of all interpolated inputs (= BCOLOR offset)
953 * @param colors_read_mask color components read (4 bits for each color, 8 bits in total)
954 * @param interp_param interpolation weights (i,j)
955 * @param prim_mask SI_PARAM_PRIM_MASK
956 * @param face SI_PARAM_FRONT_FACE
957 * @param result the return value (4 components)
959 static void interp_fs_input(struct si_shader_context
*ctx
,
960 unsigned input_index
,
961 unsigned semantic_name
,
962 unsigned semantic_index
,
963 unsigned num_interp_inputs
,
964 unsigned colors_read_mask
,
965 LLVMValueRef interp_param
,
966 LLVMValueRef prim_mask
,
968 LLVMValueRef result
[4])
970 struct lp_build_context
*base
= &ctx
->radeon_bld
.soa
.bld_base
.base
;
971 struct lp_build_context
*uint
= &ctx
->radeon_bld
.soa
.bld_base
.uint_bld
;
972 struct gallivm_state
*gallivm
= base
->gallivm
;
973 const char *intr_name
;
974 LLVMValueRef attr_number
;
978 attr_number
= lp_build_const_int32(gallivm
, input_index
);
980 /* fs.constant returns the param from the middle vertex, so it's not
981 * really useful for flat shading. It's meant to be used for custom
982 * interpolation (but the intrinsic can't fetch from the other two
985 * Luckily, it doesn't matter, because we rely on the FLAT_SHADE state
986 * to do the right thing. The only reason we use fs.constant is that
987 * fs.interp cannot be used on integers, because they can be equal
990 intr_name
= interp_param
? "llvm.SI.fs.interp" : "llvm.SI.fs.constant";
992 if (semantic_name
== TGSI_SEMANTIC_COLOR
&&
993 ctx
->shader
->key
.ps
.prolog
.color_two_side
) {
994 LLVMValueRef args
[4];
995 LLVMValueRef is_face_positive
;
996 LLVMValueRef back_attr_number
;
998 /* If BCOLOR0 is used, BCOLOR1 is at offset "num_inputs + 1",
999 * otherwise it's at offset "num_inputs".
1001 unsigned back_attr_offset
= num_interp_inputs
;
1002 if (semantic_index
== 1 && colors_read_mask
& 0xf)
1003 back_attr_offset
+= 1;
1005 back_attr_number
= lp_build_const_int32(gallivm
, back_attr_offset
);
1007 is_face_positive
= LLVMBuildICmp(gallivm
->builder
, LLVMIntNE
,
1008 face
, uint
->zero
, "");
1010 args
[2] = prim_mask
;
1011 args
[3] = interp_param
;
1012 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
1013 LLVMValueRef llvm_chan
= lp_build_const_int32(gallivm
, chan
);
1014 LLVMValueRef front
, back
;
1016 args
[0] = llvm_chan
;
1017 args
[1] = attr_number
;
1018 front
= lp_build_intrinsic(gallivm
->builder
, intr_name
,
1019 ctx
->f32
, args
, args
[3] ? 4 : 3,
1020 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
1022 args
[1] = back_attr_number
;
1023 back
= lp_build_intrinsic(gallivm
->builder
, intr_name
,
1024 ctx
->f32
, args
, args
[3] ? 4 : 3,
1025 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
1027 result
[chan
] = LLVMBuildSelect(gallivm
->builder
,
1033 } else if (semantic_name
== TGSI_SEMANTIC_FOG
) {
1034 LLVMValueRef args
[4];
1036 args
[0] = uint
->zero
;
1037 args
[1] = attr_number
;
1038 args
[2] = prim_mask
;
1039 args
[3] = interp_param
;
1040 result
[0] = lp_build_intrinsic(gallivm
->builder
, intr_name
,
1041 ctx
->f32
, args
, args
[3] ? 4 : 3,
1042 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
1044 result
[2] = lp_build_const_float(gallivm
, 0.0f
);
1045 result
[3] = lp_build_const_float(gallivm
, 1.0f
);
1047 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
1048 LLVMValueRef args
[4];
1049 LLVMValueRef llvm_chan
= lp_build_const_int32(gallivm
, chan
);
1051 args
[0] = llvm_chan
;
1052 args
[1] = attr_number
;
1053 args
[2] = prim_mask
;
1054 args
[3] = interp_param
;
1055 result
[chan
] = lp_build_intrinsic(gallivm
->builder
, intr_name
,
1056 ctx
->f32
, args
, args
[3] ? 4 : 3,
1057 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
1062 static void declare_input_fs(
1063 struct radeon_llvm_context
*radeon_bld
,
1064 unsigned input_index
,
1065 const struct tgsi_full_declaration
*decl
)
1067 struct lp_build_context
*base
= &radeon_bld
->soa
.bld_base
.base
;
1068 struct si_shader_context
*ctx
=
1069 si_shader_context(&radeon_bld
->soa
.bld_base
);
1070 struct si_shader
*shader
= ctx
->shader
;
1071 LLVMValueRef main_fn
= radeon_bld
->main_fn
;
1072 LLVMValueRef interp_param
= NULL
;
1073 int interp_param_idx
;
1075 /* Get colors from input VGPRs (set by the prolog). */
1076 if (!ctx
->is_monolithic
&&
1077 decl
->Semantic
.Name
== TGSI_SEMANTIC_COLOR
) {
1078 unsigned i
= decl
->Semantic
.Index
;
1079 unsigned colors_read
= shader
->selector
->info
.colors_read
;
1080 unsigned mask
= colors_read
>> (i
* 4);
1081 unsigned offset
= SI_PARAM_POS_FIXED_PT
+ 1 +
1082 (i
? util_bitcount(colors_read
& 0xf) : 0);
1084 radeon_bld
->inputs
[radeon_llvm_reg_index_soa(input_index
, 0)] =
1085 mask
& 0x1 ? LLVMGetParam(main_fn
, offset
++) : base
->undef
;
1086 radeon_bld
->inputs
[radeon_llvm_reg_index_soa(input_index
, 1)] =
1087 mask
& 0x2 ? LLVMGetParam(main_fn
, offset
++) : base
->undef
;
1088 radeon_bld
->inputs
[radeon_llvm_reg_index_soa(input_index
, 2)] =
1089 mask
& 0x4 ? LLVMGetParam(main_fn
, offset
++) : base
->undef
;
1090 radeon_bld
->inputs
[radeon_llvm_reg_index_soa(input_index
, 3)] =
1091 mask
& 0x8 ? LLVMGetParam(main_fn
, offset
++) : base
->undef
;
1095 interp_param_idx
= lookup_interp_param_index(decl
->Interp
.Interpolate
,
1096 decl
->Interp
.Location
);
1097 if (interp_param_idx
== -1)
1099 else if (interp_param_idx
) {
1100 interp_param_idx
= select_interp_param(ctx
,
1102 interp_param
= LLVMGetParam(main_fn
, interp_param_idx
);
1105 interp_fs_input(ctx
, input_index
, decl
->Semantic
.Name
,
1106 decl
->Semantic
.Index
, shader
->selector
->info
.num_inputs
,
1107 shader
->selector
->info
.colors_read
, interp_param
,
1108 LLVMGetParam(main_fn
, SI_PARAM_PRIM_MASK
),
1109 LLVMGetParam(main_fn
, SI_PARAM_FRONT_FACE
),
1110 &radeon_bld
->inputs
[radeon_llvm_reg_index_soa(input_index
, 0)]);
1113 static LLVMValueRef
get_sample_id(struct radeon_llvm_context
*radeon_bld
)
1115 return unpack_param(si_shader_context(&radeon_bld
->soa
.bld_base
),
1116 SI_PARAM_ANCILLARY
, 8, 4);
1120 * Set range metadata on an instruction. This can only be used on load and
1121 * call instructions. If you know an instruction can only produce the values
1122 * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
1123 * \p lo is the minimum value inclusive.
1124 * \p hi is the maximum value exclusive.
1126 static void set_range_metadata(LLVMValueRef value
, unsigned lo
, unsigned hi
)
1128 const char *range_md_string
= "range";
1129 LLVMValueRef range_md
, md_args
[2];
1130 LLVMTypeRef type
= LLVMTypeOf(value
);
1131 LLVMContextRef context
= LLVMGetTypeContext(type
);
1132 unsigned md_range_id
= LLVMGetMDKindIDInContext(context
,
1133 range_md_string
, strlen(range_md_string
));
1135 md_args
[0] = LLVMConstInt(type
, lo
, false);
1136 md_args
[1] = LLVMConstInt(type
, hi
, false);
1137 range_md
= LLVMMDNodeInContext(context
, md_args
, 2);
1138 LLVMSetMetadata(value
, md_range_id
, range_md
);
1141 static LLVMValueRef
get_thread_id(struct si_shader_context
*ctx
)
1143 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
1146 if (HAVE_LLVM
< 0x0308) {
1147 tid
= lp_build_intrinsic(gallivm
->builder
, "llvm.SI.tid",
1148 ctx
->i32
, NULL
, 0, LLVMReadNoneAttribute
);
1150 LLVMValueRef tid_args
[2];
1151 tid_args
[0] = lp_build_const_int32(gallivm
, 0xffffffff);
1152 tid_args
[1] = lp_build_const_int32(gallivm
, 0);
1153 tid_args
[1] = lp_build_intrinsic(gallivm
->builder
,
1154 "llvm.amdgcn.mbcnt.lo", ctx
->i32
,
1155 tid_args
, 2, LLVMReadNoneAttribute
);
1157 tid
= lp_build_intrinsic(gallivm
->builder
,
1158 "llvm.amdgcn.mbcnt.hi", ctx
->i32
,
1159 tid_args
, 2, LLVMReadNoneAttribute
);
1161 set_range_metadata(tid
, 0, 64);
1166 * Load a dword from a constant buffer.
1168 static LLVMValueRef
buffer_load_const(LLVMBuilderRef builder
, LLVMValueRef resource
,
1169 LLVMValueRef offset
, LLVMTypeRef return_type
)
1171 LLVMValueRef args
[2] = {resource
, offset
};
1173 return lp_build_intrinsic(builder
, "llvm.SI.load.const", return_type
, args
, 2,
1174 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
1177 static LLVMValueRef
load_sample_position(struct radeon_llvm_context
*radeon_bld
, LLVMValueRef sample_id
)
1179 struct si_shader_context
*ctx
=
1180 si_shader_context(&radeon_bld
->soa
.bld_base
);
1181 struct lp_build_context
*uint_bld
= &radeon_bld
->soa
.bld_base
.uint_bld
;
1182 struct gallivm_state
*gallivm
= &radeon_bld
->gallivm
;
1183 LLVMBuilderRef builder
= gallivm
->builder
;
1184 LLVMValueRef desc
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_RW_BUFFERS
);
1185 LLVMValueRef buf_index
= lp_build_const_int32(gallivm
, SI_PS_CONST_SAMPLE_POSITIONS
);
1186 LLVMValueRef resource
= build_indexed_load_const(ctx
, desc
, buf_index
);
1188 /* offset = sample_id * 8 (8 = 2 floats containing samplepos.xy) */
1189 LLVMValueRef offset0
= lp_build_mul_imm(uint_bld
, sample_id
, 8);
1190 LLVMValueRef offset1
= LLVMBuildAdd(builder
, offset0
, lp_build_const_int32(gallivm
, 4), "");
1192 LLVMValueRef pos
[4] = {
1193 buffer_load_const(builder
, resource
, offset0
, ctx
->f32
),
1194 buffer_load_const(builder
, resource
, offset1
, ctx
->f32
),
1195 lp_build_const_float(gallivm
, 0),
1196 lp_build_const_float(gallivm
, 0)
1199 return lp_build_gather_values(gallivm
, pos
, 4);
1202 static void declare_system_value(
1203 struct radeon_llvm_context
*radeon_bld
,
1205 const struct tgsi_full_declaration
*decl
)
1207 struct si_shader_context
*ctx
=
1208 si_shader_context(&radeon_bld
->soa
.bld_base
);
1209 struct lp_build_context
*bld
= &radeon_bld
->soa
.bld_base
.base
;
1210 struct gallivm_state
*gallivm
= &radeon_bld
->gallivm
;
1211 LLVMValueRef value
= 0;
1213 switch (decl
->Semantic
.Name
) {
1214 case TGSI_SEMANTIC_INSTANCEID
:
1215 value
= LLVMGetParam(radeon_bld
->main_fn
,
1216 ctx
->param_instance_id
);
1219 case TGSI_SEMANTIC_VERTEXID
:
1220 value
= LLVMBuildAdd(gallivm
->builder
,
1221 LLVMGetParam(radeon_bld
->main_fn
,
1222 ctx
->param_vertex_id
),
1223 LLVMGetParam(radeon_bld
->main_fn
,
1224 SI_PARAM_BASE_VERTEX
), "");
1227 case TGSI_SEMANTIC_VERTEXID_NOBASE
:
1228 value
= LLVMGetParam(radeon_bld
->main_fn
,
1229 ctx
->param_vertex_id
);
1232 case TGSI_SEMANTIC_BASEVERTEX
:
1233 value
= LLVMGetParam(radeon_bld
->main_fn
,
1234 SI_PARAM_BASE_VERTEX
);
1237 case TGSI_SEMANTIC_INVOCATIONID
:
1238 if (ctx
->type
== PIPE_SHADER_TESS_CTRL
)
1239 value
= unpack_param(ctx
, SI_PARAM_REL_IDS
, 8, 5);
1240 else if (ctx
->type
== PIPE_SHADER_GEOMETRY
)
1241 value
= LLVMGetParam(radeon_bld
->main_fn
,
1242 SI_PARAM_GS_INSTANCE_ID
);
1244 assert(!"INVOCATIONID not implemented");
1247 case TGSI_SEMANTIC_POSITION
:
1249 LLVMValueRef pos
[4] = {
1250 LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_POS_X_FLOAT
),
1251 LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_POS_Y_FLOAT
),
1252 LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_POS_Z_FLOAT
),
1253 lp_build_emit_llvm_unary(&radeon_bld
->soa
.bld_base
, TGSI_OPCODE_RCP
,
1254 LLVMGetParam(radeon_bld
->main_fn
,
1255 SI_PARAM_POS_W_FLOAT
)),
1257 value
= lp_build_gather_values(gallivm
, pos
, 4);
1261 case TGSI_SEMANTIC_FACE
:
1262 value
= LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_FRONT_FACE
);
1265 case TGSI_SEMANTIC_SAMPLEID
:
1266 value
= get_sample_id(radeon_bld
);
1269 case TGSI_SEMANTIC_SAMPLEPOS
: {
1270 LLVMValueRef pos
[4] = {
1271 LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_POS_X_FLOAT
),
1272 LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_POS_Y_FLOAT
),
1273 lp_build_const_float(gallivm
, 0),
1274 lp_build_const_float(gallivm
, 0)
1276 pos
[0] = lp_build_emit_llvm_unary(&radeon_bld
->soa
.bld_base
,
1277 TGSI_OPCODE_FRC
, pos
[0]);
1278 pos
[1] = lp_build_emit_llvm_unary(&radeon_bld
->soa
.bld_base
,
1279 TGSI_OPCODE_FRC
, pos
[1]);
1280 value
= lp_build_gather_values(gallivm
, pos
, 4);
1284 case TGSI_SEMANTIC_SAMPLEMASK
:
1285 /* This can only occur with the OpenGL Core profile, which
1286 * doesn't support smoothing.
1288 value
= LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_SAMPLE_COVERAGE
);
1291 case TGSI_SEMANTIC_TESSCOORD
:
1293 LLVMValueRef coord
[4] = {
1294 LLVMGetParam(radeon_bld
->main_fn
, ctx
->param_tes_u
),
1295 LLVMGetParam(radeon_bld
->main_fn
, ctx
->param_tes_v
),
1300 /* For triangles, the vector should be (u, v, 1-u-v). */
1301 if (ctx
->shader
->selector
->info
.properties
[TGSI_PROPERTY_TES_PRIM_MODE
] ==
1302 PIPE_PRIM_TRIANGLES
)
1303 coord
[2] = lp_build_sub(bld
, bld
->one
,
1304 lp_build_add(bld
, coord
[0], coord
[1]));
1306 value
= lp_build_gather_values(gallivm
, coord
, 4);
1310 case TGSI_SEMANTIC_VERTICESIN
:
1311 value
= unpack_param(ctx
, SI_PARAM_TCS_OUT_LAYOUT
, 26, 6);
1314 case TGSI_SEMANTIC_TESSINNER
:
1315 case TGSI_SEMANTIC_TESSOUTER
:
1317 LLVMValueRef dw_addr
;
1318 int param
= si_shader_io_get_unique_index(decl
->Semantic
.Name
, 0);
1320 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
1321 dw_addr
= LLVMBuildAdd(gallivm
->builder
, dw_addr
,
1322 lp_build_const_int32(gallivm
, param
* 4), "");
1324 value
= lds_load(&radeon_bld
->soa
.bld_base
, TGSI_TYPE_FLOAT
,
1329 case TGSI_SEMANTIC_DEFAULT_TESSOUTER_SI
:
1330 case TGSI_SEMANTIC_DEFAULT_TESSINNER_SI
:
1332 LLVMValueRef buf
, slot
, val
[4];
1335 slot
= lp_build_const_int32(gallivm
, SI_HS_CONST_DEFAULT_TESS_LEVELS
);
1336 buf
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_RW_BUFFERS
);
1337 buf
= build_indexed_load_const(ctx
, buf
, slot
);
1338 offset
= decl
->Semantic
.Name
== TGSI_SEMANTIC_DEFAULT_TESSINNER_SI
? 4 : 0;
1340 for (i
= 0; i
< 4; i
++)
1341 val
[i
] = buffer_load_const(gallivm
->builder
, buf
,
1342 lp_build_const_int32(gallivm
, (offset
+ i
) * 4),
1344 value
= lp_build_gather_values(gallivm
, val
, 4);
1348 case TGSI_SEMANTIC_PRIMID
:
1349 value
= get_primitive_id(&radeon_bld
->soa
.bld_base
, 0);
1352 case TGSI_SEMANTIC_GRID_SIZE
:
1353 value
= LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_GRID_SIZE
);
1356 case TGSI_SEMANTIC_BLOCK_SIZE
:
1358 LLVMValueRef values
[3];
1360 unsigned *properties
= ctx
->shader
->selector
->info
.properties
;
1361 unsigned sizes
[3] = {
1362 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
],
1363 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT
],
1364 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH
]
1367 for (i
= 0; i
< 3; ++i
)
1368 values
[i
] = lp_build_const_int32(gallivm
, sizes
[i
]);
1370 value
= lp_build_gather_values(gallivm
, values
, 3);
1374 case TGSI_SEMANTIC_BLOCK_ID
:
1375 value
= LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_BLOCK_ID
);
1378 case TGSI_SEMANTIC_THREAD_ID
:
1379 value
= LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_THREAD_ID
);
1382 #if HAVE_LLVM >= 0x0309
1383 case TGSI_SEMANTIC_HELPER_INVOCATION
:
1384 value
= lp_build_intrinsic(gallivm
->builder
,
1385 "llvm.amdgcn.ps.live",
1387 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
1388 value
= LLVMBuildNot(gallivm
->builder
, value
, "");
1389 value
= LLVMBuildSExt(gallivm
->builder
, value
, ctx
->i32
, "");
1394 assert(!"unknown system value");
1398 radeon_bld
->system_values
[index
] = value
;
1401 static void declare_compute_memory(struct radeon_llvm_context
*radeon_bld
,
1402 const struct tgsi_full_declaration
*decl
)
1404 struct si_shader_context
*ctx
=
1405 si_shader_context(&radeon_bld
->soa
.bld_base
);
1406 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
1407 struct gallivm_state
*gallivm
= &radeon_bld
->gallivm
;
1409 LLVMTypeRef i8p
= LLVMPointerType(ctx
->i8
, LOCAL_ADDR_SPACE
);
1412 assert(decl
->Declaration
.MemType
== TGSI_MEMORY_TYPE_SHARED
);
1413 assert(decl
->Range
.First
== decl
->Range
.Last
);
1414 assert(!ctx
->shared_memory
);
1416 var
= LLVMAddGlobalInAddressSpace(gallivm
->module
,
1417 LLVMArrayType(ctx
->i8
, sel
->local_size
),
1420 LLVMSetAlignment(var
, 4);
1422 ctx
->shared_memory
= LLVMBuildBitCast(gallivm
->builder
, var
, i8p
, "");
1425 static LLVMValueRef
fetch_constant(
1426 struct lp_build_tgsi_context
*bld_base
,
1427 const struct tgsi_full_src_register
*reg
,
1428 enum tgsi_opcode_type type
,
1431 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1432 struct lp_build_context
*base
= &bld_base
->base
;
1433 const struct tgsi_ind_register
*ireg
= ®
->Indirect
;
1436 LLVMValueRef addr
, bufp
;
1437 LLVMValueRef result
;
1439 if (swizzle
== LP_CHAN_ALL
) {
1441 LLVMValueRef values
[4];
1442 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
)
1443 values
[chan
] = fetch_constant(bld_base
, reg
, type
, chan
);
1445 return lp_build_gather_values(bld_base
->base
.gallivm
, values
, 4);
1448 buf
= reg
->Register
.Dimension
? reg
->Dimension
.Index
: 0;
1449 idx
= reg
->Register
.Index
* 4 + swizzle
;
1451 if (!reg
->Register
.Indirect
&& !reg
->Dimension
.Indirect
) {
1452 if (type
!= TGSI_TYPE_DOUBLE
)
1453 return bitcast(bld_base
, type
, ctx
->constants
[buf
][idx
]);
1455 return radeon_llvm_emit_fetch_double(bld_base
,
1456 ctx
->constants
[buf
][idx
],
1457 ctx
->constants
[buf
][idx
+ 1]);
1461 if (reg
->Register
.Dimension
&& reg
->Dimension
.Indirect
) {
1462 LLVMValueRef ptr
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_CONST_BUFFERS
);
1464 index
= get_bounded_indirect_index(ctx
, ®
->DimIndirect
,
1465 reg
->Dimension
.Index
,
1466 SI_NUM_CONST_BUFFERS
);
1467 bufp
= build_indexed_load_const(ctx
, ptr
, index
);
1469 bufp
= ctx
->const_buffers
[buf
];
1471 addr
= ctx
->radeon_bld
.soa
.addr
[ireg
->Index
][ireg
->Swizzle
];
1472 addr
= LLVMBuildLoad(base
->gallivm
->builder
, addr
, "load addr reg");
1473 addr
= lp_build_mul_imm(&bld_base
->uint_bld
, addr
, 16);
1474 addr
= lp_build_add(&bld_base
->uint_bld
, addr
,
1475 lp_build_const_int32(base
->gallivm
, idx
* 4));
1477 result
= buffer_load_const(base
->gallivm
->builder
, bufp
,
1480 if (type
!= TGSI_TYPE_DOUBLE
)
1481 result
= bitcast(bld_base
, type
, result
);
1483 LLVMValueRef addr2
, result2
;
1484 addr2
= ctx
->radeon_bld
.soa
.addr
[ireg
->Index
][ireg
->Swizzle
+ 1];
1485 addr2
= LLVMBuildLoad(base
->gallivm
->builder
, addr2
, "load addr reg2");
1486 addr2
= lp_build_mul_imm(&bld_base
->uint_bld
, addr2
, 16);
1487 addr2
= lp_build_add(&bld_base
->uint_bld
, addr2
,
1488 lp_build_const_int32(base
->gallivm
, idx
* 4));
1490 result2
= buffer_load_const(base
->gallivm
->builder
, ctx
->const_buffers
[buf
],
1493 result
= radeon_llvm_emit_fetch_double(bld_base
,
1499 /* Upper 16 bits must be zero. */
1500 static LLVMValueRef
si_llvm_pack_two_int16(struct gallivm_state
*gallivm
,
1501 LLVMValueRef val
[2])
1503 return LLVMBuildOr(gallivm
->builder
, val
[0],
1504 LLVMBuildShl(gallivm
->builder
, val
[1],
1505 lp_build_const_int32(gallivm
, 16),
1509 /* Upper 16 bits are ignored and will be dropped. */
1510 static LLVMValueRef
si_llvm_pack_two_int32_as_int16(struct gallivm_state
*gallivm
,
1511 LLVMValueRef val
[2])
1513 LLVMValueRef v
[2] = {
1514 LLVMBuildAnd(gallivm
->builder
, val
[0],
1515 lp_build_const_int32(gallivm
, 0xffff), ""),
1518 return si_llvm_pack_two_int16(gallivm
, v
);
1521 /* Initialize arguments for the shader export intrinsic */
1522 static void si_llvm_init_export_args(struct lp_build_tgsi_context
*bld_base
,
1523 LLVMValueRef
*values
,
1527 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1528 struct lp_build_context
*uint
=
1529 &ctx
->radeon_bld
.soa
.bld_base
.uint_bld
;
1530 struct lp_build_context
*base
= &bld_base
->base
;
1531 struct gallivm_state
*gallivm
= base
->gallivm
;
1532 LLVMBuilderRef builder
= base
->gallivm
->builder
;
1533 LLVMValueRef val
[4];
1534 unsigned spi_shader_col_format
= V_028714_SPI_SHADER_32_ABGR
;
1538 /* Default is 0xf. Adjusted below depending on the format. */
1539 args
[0] = lp_build_const_int32(base
->gallivm
, 0xf); /* writemask */
1541 /* Specify whether the EXEC mask represents the valid mask */
1542 args
[1] = uint
->zero
;
1544 /* Specify whether this is the last export */
1545 args
[2] = uint
->zero
;
1547 /* Specify the target we are exporting */
1548 args
[3] = lp_build_const_int32(base
->gallivm
, target
);
1550 if (ctx
->type
== PIPE_SHADER_FRAGMENT
) {
1551 const union si_shader_key
*key
= &ctx
->shader
->key
;
1552 unsigned col_formats
= key
->ps
.epilog
.spi_shader_col_format
;
1553 int cbuf
= target
- V_008DFC_SQ_EXP_MRT
;
1555 assert(cbuf
>= 0 && cbuf
< 8);
1556 spi_shader_col_format
= (col_formats
>> (cbuf
* 4)) & 0xf;
1557 is_int8
= (key
->ps
.epilog
.color_is_int8
>> cbuf
) & 0x1;
1560 args
[4] = uint
->zero
; /* COMPR flag */
1561 args
[5] = base
->undef
;
1562 args
[6] = base
->undef
;
1563 args
[7] = base
->undef
;
1564 args
[8] = base
->undef
;
1566 switch (spi_shader_col_format
) {
1567 case V_028714_SPI_SHADER_ZERO
:
1568 args
[0] = uint
->zero
; /* writemask */
1569 args
[3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_NULL
);
1572 case V_028714_SPI_SHADER_32_R
:
1573 args
[0] = uint
->one
; /* writemask */
1574 args
[5] = values
[0];
1577 case V_028714_SPI_SHADER_32_GR
:
1578 args
[0] = lp_build_const_int32(base
->gallivm
, 0x3); /* writemask */
1579 args
[5] = values
[0];
1580 args
[6] = values
[1];
1583 case V_028714_SPI_SHADER_32_AR
:
1584 args
[0] = lp_build_const_int32(base
->gallivm
, 0x9); /* writemask */
1585 args
[5] = values
[0];
1586 args
[8] = values
[3];
1589 case V_028714_SPI_SHADER_FP16_ABGR
:
1590 args
[4] = uint
->one
; /* COMPR flag */
1592 for (chan
= 0; chan
< 2; chan
++) {
1593 LLVMValueRef pack_args
[2] = {
1595 values
[2 * chan
+ 1]
1597 LLVMValueRef packed
;
1599 packed
= lp_build_intrinsic(base
->gallivm
->builder
,
1601 ctx
->i32
, pack_args
, 2,
1602 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
1604 LLVMBuildBitCast(base
->gallivm
->builder
,
1605 packed
, ctx
->f32
, "");
1609 case V_028714_SPI_SHADER_UNORM16_ABGR
:
1610 for (chan
= 0; chan
< 4; chan
++) {
1611 val
[chan
] = radeon_llvm_saturate(bld_base
, values
[chan
]);
1612 val
[chan
] = LLVMBuildFMul(builder
, val
[chan
],
1613 lp_build_const_float(gallivm
, 65535), "");
1614 val
[chan
] = LLVMBuildFAdd(builder
, val
[chan
],
1615 lp_build_const_float(gallivm
, 0.5), "");
1616 val
[chan
] = LLVMBuildFPToUI(builder
, val
[chan
],
1620 args
[4] = uint
->one
; /* COMPR flag */
1621 args
[5] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
1622 si_llvm_pack_two_int16(gallivm
, val
));
1623 args
[6] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
1624 si_llvm_pack_two_int16(gallivm
, val
+2));
1627 case V_028714_SPI_SHADER_SNORM16_ABGR
:
1628 for (chan
= 0; chan
< 4; chan
++) {
1629 /* Clamp between [-1, 1]. */
1630 val
[chan
] = lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_MIN
,
1632 lp_build_const_float(gallivm
, 1));
1633 val
[chan
] = lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_MAX
,
1635 lp_build_const_float(gallivm
, -1));
1636 /* Convert to a signed integer in [-32767, 32767]. */
1637 val
[chan
] = LLVMBuildFMul(builder
, val
[chan
],
1638 lp_build_const_float(gallivm
, 32767), "");
1639 /* If positive, add 0.5, else add -0.5. */
1640 val
[chan
] = LLVMBuildFAdd(builder
, val
[chan
],
1641 LLVMBuildSelect(builder
,
1642 LLVMBuildFCmp(builder
, LLVMRealOGE
,
1643 val
[chan
], base
->zero
, ""),
1644 lp_build_const_float(gallivm
, 0.5),
1645 lp_build_const_float(gallivm
, -0.5), ""), "");
1646 val
[chan
] = LLVMBuildFPToSI(builder
, val
[chan
], ctx
->i32
, "");
1649 args
[4] = uint
->one
; /* COMPR flag */
1650 args
[5] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
1651 si_llvm_pack_two_int32_as_int16(gallivm
, val
));
1652 args
[6] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
1653 si_llvm_pack_two_int32_as_int16(gallivm
, val
+2));
1656 case V_028714_SPI_SHADER_UINT16_ABGR
: {
1657 LLVMValueRef max
= lp_build_const_int32(gallivm
, is_int8
?
1660 for (chan
= 0; chan
< 4; chan
++) {
1661 val
[chan
] = bitcast(bld_base
, TGSI_TYPE_UNSIGNED
, values
[chan
]);
1662 val
[chan
] = lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_UMIN
,
1666 args
[4] = uint
->one
; /* COMPR flag */
1667 args
[5] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
1668 si_llvm_pack_two_int16(gallivm
, val
));
1669 args
[6] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
1670 si_llvm_pack_two_int16(gallivm
, val
+2));
1674 case V_028714_SPI_SHADER_SINT16_ABGR
: {
1675 LLVMValueRef max
= lp_build_const_int32(gallivm
, is_int8
?
1677 LLVMValueRef min
= lp_build_const_int32(gallivm
, is_int8
?
1680 for (chan
= 0; chan
< 4; chan
++) {
1681 val
[chan
] = bitcast(bld_base
, TGSI_TYPE_UNSIGNED
, values
[chan
]);
1682 val
[chan
] = lp_build_emit_llvm_binary(bld_base
,
1685 val
[chan
] = lp_build_emit_llvm_binary(bld_base
,
1690 args
[4] = uint
->one
; /* COMPR flag */
1691 args
[5] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
1692 si_llvm_pack_two_int32_as_int16(gallivm
, val
));
1693 args
[6] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
1694 si_llvm_pack_two_int32_as_int16(gallivm
, val
+2));
1698 case V_028714_SPI_SHADER_32_ABGR
:
1699 memcpy(&args
[5], values
, sizeof(values
[0]) * 4);
1704 static void si_alpha_test(struct lp_build_tgsi_context
*bld_base
,
1707 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1708 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
1710 if (ctx
->shader
->key
.ps
.epilog
.alpha_func
!= PIPE_FUNC_NEVER
) {
1711 LLVMValueRef alpha_ref
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
1712 SI_PARAM_ALPHA_REF
);
1714 LLVMValueRef alpha_pass
=
1715 lp_build_cmp(&bld_base
->base
,
1716 ctx
->shader
->key
.ps
.epilog
.alpha_func
,
1719 lp_build_select(&bld_base
->base
,
1721 lp_build_const_float(gallivm
, 1.0f
),
1722 lp_build_const_float(gallivm
, -1.0f
));
1724 lp_build_intrinsic(gallivm
->builder
, "llvm.AMDGPU.kill",
1725 ctx
->voidt
, &arg
, 1, 0);
1727 lp_build_intrinsic(gallivm
->builder
, "llvm.AMDGPU.kilp",
1728 ctx
->voidt
, NULL
, 0, 0);
1732 static LLVMValueRef
si_scale_alpha_by_sample_mask(struct lp_build_tgsi_context
*bld_base
,
1734 unsigned samplemask_param
)
1736 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1737 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
1738 LLVMValueRef coverage
;
1740 /* alpha = alpha * popcount(coverage) / SI_NUM_SMOOTH_AA_SAMPLES */
1741 coverage
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
1743 coverage
= bitcast(bld_base
, TGSI_TYPE_SIGNED
, coverage
);
1745 coverage
= lp_build_intrinsic(gallivm
->builder
, "llvm.ctpop.i32",
1747 &coverage
, 1, LLVMReadNoneAttribute
);
1749 coverage
= LLVMBuildUIToFP(gallivm
->builder
, coverage
,
1752 coverage
= LLVMBuildFMul(gallivm
->builder
, coverage
,
1753 lp_build_const_float(gallivm
,
1754 1.0 / SI_NUM_SMOOTH_AA_SAMPLES
), "");
1756 return LLVMBuildFMul(gallivm
->builder
, alpha
, coverage
, "");
1759 static void si_llvm_emit_clipvertex(struct lp_build_tgsi_context
*bld_base
,
1760 LLVMValueRef (*pos
)[9], LLVMValueRef
*out_elts
)
1762 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1763 struct lp_build_context
*base
= &bld_base
->base
;
1764 struct lp_build_context
*uint
= &ctx
->radeon_bld
.soa
.bld_base
.uint_bld
;
1767 unsigned const_chan
;
1768 LLVMValueRef base_elt
;
1769 LLVMValueRef ptr
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_RW_BUFFERS
);
1770 LLVMValueRef constbuf_index
= lp_build_const_int32(base
->gallivm
,
1771 SI_VS_CONST_CLIP_PLANES
);
1772 LLVMValueRef const_resource
= build_indexed_load_const(ctx
, ptr
, constbuf_index
);
1774 for (reg_index
= 0; reg_index
< 2; reg_index
++) {
1775 LLVMValueRef
*args
= pos
[2 + reg_index
];
1780 args
[8] = lp_build_const_float(base
->gallivm
, 0.0f
);
1782 /* Compute dot products of position and user clip plane vectors */
1783 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
1784 for (const_chan
= 0; const_chan
< TGSI_NUM_CHANNELS
; const_chan
++) {
1785 args
[1] = lp_build_const_int32(base
->gallivm
,
1786 ((reg_index
* 4 + chan
) * 4 +
1788 base_elt
= buffer_load_const(base
->gallivm
->builder
, const_resource
,
1791 lp_build_add(base
, args
[5 + chan
],
1792 lp_build_mul(base
, base_elt
,
1793 out_elts
[const_chan
]));
1797 args
[0] = lp_build_const_int32(base
->gallivm
, 0xf);
1798 args
[1] = uint
->zero
;
1799 args
[2] = uint
->zero
;
1800 args
[3] = lp_build_const_int32(base
->gallivm
,
1801 V_008DFC_SQ_EXP_POS
+ 2 + reg_index
);
1802 args
[4] = uint
->zero
;
1806 static void si_dump_streamout(struct pipe_stream_output_info
*so
)
1810 if (so
->num_outputs
)
1811 fprintf(stderr
, "STREAMOUT\n");
1813 for (i
= 0; i
< so
->num_outputs
; i
++) {
1814 unsigned mask
= ((1 << so
->output
[i
].num_components
) - 1) <<
1815 so
->output
[i
].start_component
;
1816 fprintf(stderr
, " %i: BUF%i[%i..%i] <- OUT[%i].%s%s%s%s\n",
1817 i
, so
->output
[i
].output_buffer
,
1818 so
->output
[i
].dst_offset
, so
->output
[i
].dst_offset
+ so
->output
[i
].num_components
- 1,
1819 so
->output
[i
].register_index
,
1820 mask
& 1 ? "x" : "",
1821 mask
& 2 ? "y" : "",
1822 mask
& 4 ? "z" : "",
1823 mask
& 8 ? "w" : "");
1827 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
1828 * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
1829 * or v4i32 (num_channels=3,4). */
1830 static void build_tbuffer_store(struct si_shader_context
*ctx
,
1833 unsigned num_channels
,
1835 LLVMValueRef soffset
,
1836 unsigned inst_offset
,
1845 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
1846 LLVMValueRef args
[] = {
1849 LLVMConstInt(ctx
->i32
, num_channels
, 0),
1852 LLVMConstInt(ctx
->i32
, inst_offset
, 0),
1853 LLVMConstInt(ctx
->i32
, dfmt
, 0),
1854 LLVMConstInt(ctx
->i32
, nfmt
, 0),
1855 LLVMConstInt(ctx
->i32
, offen
, 0),
1856 LLVMConstInt(ctx
->i32
, idxen
, 0),
1857 LLVMConstInt(ctx
->i32
, glc
, 0),
1858 LLVMConstInt(ctx
->i32
, slc
, 0),
1859 LLVMConstInt(ctx
->i32
, tfe
, 0)
1862 /* The instruction offset field has 12 bits */
1863 assert(offen
|| inst_offset
< (1 << 12));
1865 /* The intrinsic is overloaded, we need to add a type suffix for overloading to work. */
1866 unsigned func
= CLAMP(num_channels
, 1, 3) - 1;
1867 const char *types
[] = {"i32", "v2i32", "v4i32"};
1869 snprintf(name
, sizeof(name
), "llvm.SI.tbuffer.store.%s", types
[func
]);
1871 lp_build_intrinsic(gallivm
->builder
, name
, ctx
->voidt
,
1872 args
, Elements(args
), 0);
1875 static void build_tbuffer_store_dwords(struct si_shader_context
*ctx
,
1878 unsigned num_channels
,
1880 LLVMValueRef soffset
,
1881 unsigned inst_offset
)
1883 static unsigned dfmt
[] = {
1884 V_008F0C_BUF_DATA_FORMAT_32
,
1885 V_008F0C_BUF_DATA_FORMAT_32_32
,
1886 V_008F0C_BUF_DATA_FORMAT_32_32_32
,
1887 V_008F0C_BUF_DATA_FORMAT_32_32_32_32
1889 assert(num_channels
>= 1 && num_channels
<= 4);
1891 build_tbuffer_store(ctx
, rsrc
, vdata
, num_channels
, vaddr
, soffset
,
1892 inst_offset
, dfmt
[num_channels
-1],
1893 V_008F0C_BUF_NUM_FORMAT_UINT
, 1, 0, 1, 1, 0);
1896 /* On SI, the vertex shader is responsible for writing streamout data
1898 static void si_llvm_emit_streamout(struct si_shader_context
*ctx
,
1899 struct si_shader_output_values
*outputs
,
1902 struct pipe_stream_output_info
*so
= &ctx
->shader
->selector
->so
;
1903 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
1904 LLVMBuilderRef builder
= gallivm
->builder
;
1906 struct lp_build_if_state if_ctx
;
1908 /* Get bits [22:16], i.e. (so_param >> 16) & 127; */
1909 LLVMValueRef so_vtx_count
=
1910 unpack_param(ctx
, ctx
->param_streamout_config
, 16, 7);
1912 LLVMValueRef tid
= get_thread_id(ctx
);
1914 /* can_emit = tid < so_vtx_count; */
1915 LLVMValueRef can_emit
=
1916 LLVMBuildICmp(builder
, LLVMIntULT
, tid
, so_vtx_count
, "");
1918 LLVMValueRef stream_id
=
1919 unpack_param(ctx
, ctx
->param_streamout_config
, 24, 2);
1921 /* Emit the streamout code conditionally. This actually avoids
1922 * out-of-bounds buffer access. The hw tells us via the SGPR
1923 * (so_vtx_count) which threads are allowed to emit streamout data. */
1924 lp_build_if(&if_ctx
, gallivm
, can_emit
);
1926 /* The buffer offset is computed as follows:
1927 * ByteOffset = streamout_offset[buffer_id]*4 +
1928 * (streamout_write_index + thread_id)*stride[buffer_id] +
1932 LLVMValueRef so_write_index
=
1933 LLVMGetParam(ctx
->radeon_bld
.main_fn
,
1934 ctx
->param_streamout_write_index
);
1936 /* Compute (streamout_write_index + thread_id). */
1937 so_write_index
= LLVMBuildAdd(builder
, so_write_index
, tid
, "");
1939 /* Compute the write offset for each enabled buffer. */
1940 LLVMValueRef so_write_offset
[4] = {};
1941 for (i
= 0; i
< 4; i
++) {
1945 LLVMValueRef so_offset
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
1946 ctx
->param_streamout_offset
[i
]);
1947 so_offset
= LLVMBuildMul(builder
, so_offset
, LLVMConstInt(ctx
->i32
, 4, 0), "");
1949 so_write_offset
[i
] = LLVMBuildMul(builder
, so_write_index
,
1950 LLVMConstInt(ctx
->i32
, so
->stride
[i
]*4, 0), "");
1951 so_write_offset
[i
] = LLVMBuildAdd(builder
, so_write_offset
[i
], so_offset
, "");
1954 /* Write streamout data. */
1955 for (i
= 0; i
< so
->num_outputs
; i
++) {
1956 unsigned buf_idx
= so
->output
[i
].output_buffer
;
1957 unsigned reg
= so
->output
[i
].register_index
;
1958 unsigned start
= so
->output
[i
].start_component
;
1959 unsigned num_comps
= so
->output
[i
].num_components
;
1960 unsigned stream
= so
->output
[i
].stream
;
1961 LLVMValueRef out
[4];
1962 struct lp_build_if_state if_ctx_stream
;
1964 assert(num_comps
&& num_comps
<= 4);
1965 if (!num_comps
|| num_comps
> 4)
1971 /* Load the output as int. */
1972 for (j
= 0; j
< num_comps
; j
++) {
1973 out
[j
] = LLVMBuildBitCast(builder
,
1974 outputs
[reg
].values
[start
+j
],
1978 /* Pack the output. */
1979 LLVMValueRef vdata
= NULL
;
1981 switch (num_comps
) {
1982 case 1: /* as i32 */
1985 case 2: /* as v2i32 */
1986 case 3: /* as v4i32 (aligned to 4) */
1987 case 4: /* as v4i32 */
1988 vdata
= LLVMGetUndef(LLVMVectorType(ctx
->i32
, util_next_power_of_two(num_comps
)));
1989 for (j
= 0; j
< num_comps
; j
++) {
1990 vdata
= LLVMBuildInsertElement(builder
, vdata
, out
[j
],
1991 LLVMConstInt(ctx
->i32
, j
, 0), "");
1996 LLVMValueRef can_emit_stream
=
1997 LLVMBuildICmp(builder
, LLVMIntEQ
,
1999 lp_build_const_int32(gallivm
, stream
), "");
2001 lp_build_if(&if_ctx_stream
, gallivm
, can_emit_stream
);
2002 build_tbuffer_store_dwords(ctx
, ctx
->so_buffers
[buf_idx
],
2004 so_write_offset
[buf_idx
],
2005 LLVMConstInt(ctx
->i32
, 0, 0),
2006 so
->output
[i
].dst_offset
*4);
2007 lp_build_endif(&if_ctx_stream
);
2010 lp_build_endif(&if_ctx
);
2014 /* Generate export instructions for hardware VS shader stage */
2015 static void si_llvm_export_vs(struct lp_build_tgsi_context
*bld_base
,
2016 struct si_shader_output_values
*outputs
,
2019 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2020 struct si_shader
*shader
= ctx
->shader
;
2021 struct lp_build_context
*base
= &bld_base
->base
;
2022 struct lp_build_context
*uint
=
2023 &ctx
->radeon_bld
.soa
.bld_base
.uint_bld
;
2024 LLVMValueRef args
[9];
2025 LLVMValueRef pos_args
[4][9] = { { 0 } };
2026 LLVMValueRef psize_value
= NULL
, edgeflag_value
= NULL
, layer_value
= NULL
, viewport_index_value
= NULL
;
2027 unsigned semantic_name
, semantic_index
;
2029 unsigned param_count
= 0;
2033 if (outputs
&& ctx
->shader
->selector
->so
.num_outputs
) {
2034 si_llvm_emit_streamout(ctx
, outputs
, noutput
);
2037 for (i
= 0; i
< noutput
; i
++) {
2038 semantic_name
= outputs
[i
].name
;
2039 semantic_index
= outputs
[i
].sid
;
2042 /* Select the correct target */
2043 switch(semantic_name
) {
2044 case TGSI_SEMANTIC_PSIZE
:
2045 psize_value
= outputs
[i
].values
[0];
2047 case TGSI_SEMANTIC_EDGEFLAG
:
2048 edgeflag_value
= outputs
[i
].values
[0];
2050 case TGSI_SEMANTIC_LAYER
:
2051 layer_value
= outputs
[i
].values
[0];
2052 semantic_name
= TGSI_SEMANTIC_GENERIC
;
2053 goto handle_semantic
;
2054 case TGSI_SEMANTIC_VIEWPORT_INDEX
:
2055 viewport_index_value
= outputs
[i
].values
[0];
2056 semantic_name
= TGSI_SEMANTIC_GENERIC
;
2057 goto handle_semantic
;
2058 case TGSI_SEMANTIC_POSITION
:
2059 target
= V_008DFC_SQ_EXP_POS
;
2061 case TGSI_SEMANTIC_COLOR
:
2062 case TGSI_SEMANTIC_BCOLOR
:
2063 target
= V_008DFC_SQ_EXP_PARAM
+ param_count
;
2064 assert(i
< ARRAY_SIZE(shader
->info
.vs_output_param_offset
));
2065 shader
->info
.vs_output_param_offset
[i
] = param_count
;
2068 case TGSI_SEMANTIC_CLIPDIST
:
2069 target
= V_008DFC_SQ_EXP_POS
+ 2 + semantic_index
;
2071 case TGSI_SEMANTIC_CLIPVERTEX
:
2072 si_llvm_emit_clipvertex(bld_base
, pos_args
, outputs
[i
].values
);
2074 case TGSI_SEMANTIC_PRIMID
:
2075 case TGSI_SEMANTIC_FOG
:
2076 case TGSI_SEMANTIC_TEXCOORD
:
2077 case TGSI_SEMANTIC_GENERIC
:
2078 target
= V_008DFC_SQ_EXP_PARAM
+ param_count
;
2079 assert(i
< ARRAY_SIZE(shader
->info
.vs_output_param_offset
));
2080 shader
->info
.vs_output_param_offset
[i
] = param_count
;
2086 "Warning: SI unhandled vs output type:%d\n",
2090 si_llvm_init_export_args(bld_base
, outputs
[i
].values
, target
, args
);
2092 if (target
>= V_008DFC_SQ_EXP_POS
&&
2093 target
<= (V_008DFC_SQ_EXP_POS
+ 3)) {
2094 memcpy(pos_args
[target
- V_008DFC_SQ_EXP_POS
],
2095 args
, sizeof(args
));
2097 lp_build_intrinsic(base
->gallivm
->builder
,
2098 "llvm.SI.export", ctx
->voidt
,
2102 if (semantic_name
== TGSI_SEMANTIC_CLIPDIST
) {
2103 semantic_name
= TGSI_SEMANTIC_GENERIC
;
2104 goto handle_semantic
;
2108 shader
->info
.nr_param_exports
= param_count
;
2110 /* We need to add the position output manually if it's missing. */
2111 if (!pos_args
[0][0]) {
2112 pos_args
[0][0] = lp_build_const_int32(base
->gallivm
, 0xf); /* writemask */
2113 pos_args
[0][1] = uint
->zero
; /* EXEC mask */
2114 pos_args
[0][2] = uint
->zero
; /* last export? */
2115 pos_args
[0][3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_POS
);
2116 pos_args
[0][4] = uint
->zero
; /* COMPR flag */
2117 pos_args
[0][5] = base
->zero
; /* X */
2118 pos_args
[0][6] = base
->zero
; /* Y */
2119 pos_args
[0][7] = base
->zero
; /* Z */
2120 pos_args
[0][8] = base
->one
; /* W */
2123 /* Write the misc vector (point size, edgeflag, layer, viewport). */
2124 if (shader
->selector
->info
.writes_psize
||
2125 shader
->selector
->info
.writes_edgeflag
||
2126 shader
->selector
->info
.writes_viewport_index
||
2127 shader
->selector
->info
.writes_layer
) {
2128 pos_args
[1][0] = lp_build_const_int32(base
->gallivm
, /* writemask */
2129 shader
->selector
->info
.writes_psize
|
2130 (shader
->selector
->info
.writes_edgeflag
<< 1) |
2131 (shader
->selector
->info
.writes_layer
<< 2) |
2132 (shader
->selector
->info
.writes_viewport_index
<< 3));
2133 pos_args
[1][1] = uint
->zero
; /* EXEC mask */
2134 pos_args
[1][2] = uint
->zero
; /* last export? */
2135 pos_args
[1][3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_POS
+ 1);
2136 pos_args
[1][4] = uint
->zero
; /* COMPR flag */
2137 pos_args
[1][5] = base
->zero
; /* X */
2138 pos_args
[1][6] = base
->zero
; /* Y */
2139 pos_args
[1][7] = base
->zero
; /* Z */
2140 pos_args
[1][8] = base
->zero
; /* W */
2142 if (shader
->selector
->info
.writes_psize
)
2143 pos_args
[1][5] = psize_value
;
2145 if (shader
->selector
->info
.writes_edgeflag
) {
2146 /* The output is a float, but the hw expects an integer
2147 * with the first bit containing the edge flag. */
2148 edgeflag_value
= LLVMBuildFPToUI(base
->gallivm
->builder
,
2151 edgeflag_value
= lp_build_min(&bld_base
->int_bld
,
2153 bld_base
->int_bld
.one
);
2155 /* The LLVM intrinsic expects a float. */
2156 pos_args
[1][6] = LLVMBuildBitCast(base
->gallivm
->builder
,
2161 if (shader
->selector
->info
.writes_layer
)
2162 pos_args
[1][7] = layer_value
;
2164 if (shader
->selector
->info
.writes_viewport_index
)
2165 pos_args
[1][8] = viewport_index_value
;
2168 for (i
= 0; i
< 4; i
++)
2170 shader
->info
.nr_pos_exports
++;
2173 for (i
= 0; i
< 4; i
++) {
2174 if (!pos_args
[i
][0])
2177 /* Specify the target we are exporting */
2178 pos_args
[i
][3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_POS
+ pos_idx
++);
2180 if (pos_idx
== shader
->info
.nr_pos_exports
)
2181 /* Specify that this is the last export */
2182 pos_args
[i
][2] = uint
->one
;
2184 lp_build_intrinsic(base
->gallivm
->builder
, "llvm.SI.export",
2185 ctx
->voidt
, pos_args
[i
], 9, 0);
2189 static void si_write_tess_factors(struct lp_build_tgsi_context
*bld_base
,
2190 LLVMValueRef rel_patch_id
,
2191 LLVMValueRef invocation_id
,
2192 LLVMValueRef tcs_out_current_patch_data_offset
)
2194 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2195 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2196 struct si_shader
*shader
= ctx
->shader
;
2197 unsigned tess_inner_index
, tess_outer_index
;
2198 LLVMValueRef lds_base
, lds_inner
, lds_outer
, byteoffset
, buffer
;
2199 LLVMValueRef out
[6], vec0
, vec1
, rw_buffers
, tf_base
;
2200 unsigned stride
, outer_comps
, inner_comps
, i
;
2201 struct lp_build_if_state if_ctx
;
2203 /* Do this only for invocation 0, because the tess levels are per-patch,
2206 * This can't jump, because invocation 0 executes this. It should
2207 * at least mask out the loads and stores for other invocations.
2209 lp_build_if(&if_ctx
, gallivm
,
2210 LLVMBuildICmp(gallivm
->builder
, LLVMIntEQ
,
2211 invocation_id
, bld_base
->uint_bld
.zero
, ""));
2213 /* Determine the layout of one tess factor element in the buffer. */
2214 switch (shader
->key
.tcs
.epilog
.prim_mode
) {
2215 case PIPE_PRIM_LINES
:
2216 stride
= 2; /* 2 dwords, 1 vec2 store */
2220 case PIPE_PRIM_TRIANGLES
:
2221 stride
= 4; /* 4 dwords, 1 vec4 store */
2225 case PIPE_PRIM_QUADS
:
2226 stride
= 6; /* 6 dwords, 2 stores (vec4 + vec2) */
2235 /* Load tess_inner and tess_outer from LDS.
2236 * Any invocation can write them, so we can't get them from a temporary.
2238 tess_inner_index
= si_shader_io_get_unique_index(TGSI_SEMANTIC_TESSINNER
, 0);
2239 tess_outer_index
= si_shader_io_get_unique_index(TGSI_SEMANTIC_TESSOUTER
, 0);
2241 lds_base
= tcs_out_current_patch_data_offset
;
2242 lds_inner
= LLVMBuildAdd(gallivm
->builder
, lds_base
,
2243 lp_build_const_int32(gallivm
,
2244 tess_inner_index
* 4), "");
2245 lds_outer
= LLVMBuildAdd(gallivm
->builder
, lds_base
,
2246 lp_build_const_int32(gallivm
,
2247 tess_outer_index
* 4), "");
2249 for (i
= 0; i
< outer_comps
; i
++)
2250 out
[i
] = lds_load(bld_base
, TGSI_TYPE_SIGNED
, i
, lds_outer
);
2251 for (i
= 0; i
< inner_comps
; i
++)
2252 out
[outer_comps
+i
] = lds_load(bld_base
, TGSI_TYPE_SIGNED
, i
, lds_inner
);
2254 /* Convert the outputs to vectors for stores. */
2255 vec0
= lp_build_gather_values(gallivm
, out
, MIN2(stride
, 4));
2259 vec1
= lp_build_gather_values(gallivm
, out
+4, stride
- 4);
2261 /* Get the buffer. */
2262 rw_buffers
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
2263 SI_PARAM_RW_BUFFERS
);
2264 buffer
= build_indexed_load_const(ctx
, rw_buffers
,
2265 lp_build_const_int32(gallivm
, SI_HS_RING_TESS_FACTOR
));
2267 /* Get the offset. */
2268 tf_base
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
2269 SI_PARAM_TESS_FACTOR_OFFSET
);
2270 byteoffset
= LLVMBuildMul(gallivm
->builder
, rel_patch_id
,
2271 lp_build_const_int32(gallivm
, 4 * stride
), "");
2273 /* Store the outputs. */
2274 build_tbuffer_store_dwords(ctx
, buffer
, vec0
,
2275 MIN2(stride
, 4), byteoffset
, tf_base
, 0);
2277 build_tbuffer_store_dwords(ctx
, buffer
, vec1
,
2278 stride
- 4, byteoffset
, tf_base
, 16);
2279 lp_build_endif(&if_ctx
);
2282 /* This only writes the tessellation factor levels. */
2283 static void si_llvm_emit_tcs_epilogue(struct lp_build_tgsi_context
*bld_base
)
2285 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2286 LLVMValueRef rel_patch_id
, invocation_id
, tf_lds_offset
;
2288 rel_patch_id
= get_rel_patch_id(ctx
);
2289 invocation_id
= unpack_param(ctx
, SI_PARAM_REL_IDS
, 8, 5);
2290 tf_lds_offset
= get_tcs_out_current_patch_data_offset(ctx
);
2292 if (!ctx
->is_monolithic
) {
2293 /* Return epilog parameters from this function. */
2294 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
2295 LLVMValueRef ret
= ctx
->return_value
;
2296 LLVMValueRef rw_buffers
, rw0
, rw1
, tf_soffset
;
2299 /* RW_BUFFERS pointer */
2300 rw_buffers
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
2301 SI_PARAM_RW_BUFFERS
);
2302 rw_buffers
= LLVMBuildPtrToInt(builder
, rw_buffers
, ctx
->i64
, "");
2303 rw_buffers
= LLVMBuildBitCast(builder
, rw_buffers
, ctx
->v2i32
, "");
2304 rw0
= LLVMBuildExtractElement(builder
, rw_buffers
,
2305 bld_base
->uint_bld
.zero
, "");
2306 rw1
= LLVMBuildExtractElement(builder
, rw_buffers
,
2307 bld_base
->uint_bld
.one
, "");
2308 ret
= LLVMBuildInsertValue(builder
, ret
, rw0
, 0, "");
2309 ret
= LLVMBuildInsertValue(builder
, ret
, rw1
, 1, "");
2311 /* Tess factor buffer soffset is after user SGPRs. */
2312 tf_soffset
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
2313 SI_PARAM_TESS_FACTOR_OFFSET
);
2314 ret
= LLVMBuildInsertValue(builder
, ret
, tf_soffset
,
2315 SI_TCS_NUM_USER_SGPR
, "");
2318 rel_patch_id
= bitcast(bld_base
, TGSI_TYPE_FLOAT
, rel_patch_id
);
2319 invocation_id
= bitcast(bld_base
, TGSI_TYPE_FLOAT
, invocation_id
);
2320 tf_lds_offset
= bitcast(bld_base
, TGSI_TYPE_FLOAT
, tf_lds_offset
);
2322 vgpr
= SI_TCS_NUM_USER_SGPR
+ 1;
2323 ret
= LLVMBuildInsertValue(builder
, ret
, rel_patch_id
, vgpr
++, "");
2324 ret
= LLVMBuildInsertValue(builder
, ret
, invocation_id
, vgpr
++, "");
2325 ret
= LLVMBuildInsertValue(builder
, ret
, tf_lds_offset
, vgpr
++, "");
2326 ctx
->return_value
= ret
;
2330 si_write_tess_factors(bld_base
, rel_patch_id
, invocation_id
, tf_lds_offset
);
2333 static void si_llvm_emit_ls_epilogue(struct lp_build_tgsi_context
*bld_base
)
2335 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2336 struct si_shader
*shader
= ctx
->shader
;
2337 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
2338 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2340 LLVMValueRef vertex_id
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
2341 ctx
->param_rel_auto_id
);
2342 LLVMValueRef vertex_dw_stride
=
2343 unpack_param(ctx
, SI_PARAM_LS_OUT_LAYOUT
, 13, 8);
2344 LLVMValueRef base_dw_addr
= LLVMBuildMul(gallivm
->builder
, vertex_id
,
2345 vertex_dw_stride
, "");
2347 /* Write outputs to LDS. The next shader (TCS aka HS) will read
2348 * its inputs from it. */
2349 for (i
= 0; i
< info
->num_outputs
; i
++) {
2350 LLVMValueRef
*out_ptr
= ctx
->radeon_bld
.soa
.outputs
[i
];
2351 unsigned name
= info
->output_semantic_name
[i
];
2352 unsigned index
= info
->output_semantic_index
[i
];
2353 int param
= si_shader_io_get_unique_index(name
, index
);
2354 LLVMValueRef dw_addr
= LLVMBuildAdd(gallivm
->builder
, base_dw_addr
,
2355 lp_build_const_int32(gallivm
, param
* 4), "");
2357 for (chan
= 0; chan
< 4; chan
++) {
2358 lds_store(bld_base
, chan
, dw_addr
,
2359 LLVMBuildLoad(gallivm
->builder
, out_ptr
[chan
], ""));
2364 static void si_llvm_emit_es_epilogue(struct lp_build_tgsi_context
*bld_base
)
2366 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2367 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2368 struct si_shader
*es
= ctx
->shader
;
2369 struct tgsi_shader_info
*info
= &es
->selector
->info
;
2370 LLVMValueRef soffset
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
2371 ctx
->param_es2gs_offset
);
2375 for (i
= 0; i
< info
->num_outputs
; i
++) {
2376 LLVMValueRef
*out_ptr
=
2377 ctx
->radeon_bld
.soa
.outputs
[i
];
2380 if (info
->output_semantic_name
[i
] == TGSI_SEMANTIC_VIEWPORT_INDEX
||
2381 info
->output_semantic_name
[i
] == TGSI_SEMANTIC_LAYER
)
2384 param_index
= si_shader_io_get_unique_index(info
->output_semantic_name
[i
],
2385 info
->output_semantic_index
[i
]);
2387 for (chan
= 0; chan
< 4; chan
++) {
2388 LLVMValueRef out_val
= LLVMBuildLoad(gallivm
->builder
, out_ptr
[chan
], "");
2389 out_val
= LLVMBuildBitCast(gallivm
->builder
, out_val
, ctx
->i32
, "");
2391 build_tbuffer_store(ctx
,
2394 LLVMGetUndef(ctx
->i32
), soffset
,
2395 (4 * param_index
+ chan
) * 4,
2396 V_008F0C_BUF_DATA_FORMAT_32
,
2397 V_008F0C_BUF_NUM_FORMAT_UINT
,
2403 static void si_llvm_emit_gs_epilogue(struct lp_build_tgsi_context
*bld_base
)
2405 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2406 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2407 LLVMValueRef args
[2];
2409 args
[0] = lp_build_const_int32(gallivm
, SENDMSG_GS_OP_NOP
| SENDMSG_GS_DONE
);
2410 args
[1] = LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_GS_WAVE_ID
);
2411 lp_build_intrinsic(gallivm
->builder
, "llvm.SI.sendmsg",
2412 ctx
->voidt
, args
, 2, LLVMNoUnwindAttribute
);
2415 static void si_llvm_emit_vs_epilogue(struct lp_build_tgsi_context
*bld_base
)
2417 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2418 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2419 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
2420 struct si_shader_output_values
*outputs
= NULL
;
2423 assert(!ctx
->is_gs_copy_shader
);
2425 outputs
= MALLOC((info
->num_outputs
+ 1) * sizeof(outputs
[0]));
2427 /* Vertex color clamping.
2429 * This uses a state constant loaded in a user data SGPR and
2430 * an IF statement is added that clamps all colors if the constant
2433 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
2434 struct lp_build_if_state if_ctx
;
2435 LLVMValueRef cond
= NULL
;
2436 LLVMValueRef addr
, val
;
2438 for (i
= 0; i
< info
->num_outputs
; i
++) {
2439 if (info
->output_semantic_name
[i
] != TGSI_SEMANTIC_COLOR
&&
2440 info
->output_semantic_name
[i
] != TGSI_SEMANTIC_BCOLOR
)
2443 /* We've found a color. */
2445 /* The state is in the first bit of the user SGPR. */
2446 cond
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
2447 SI_PARAM_VS_STATE_BITS
);
2448 cond
= LLVMBuildTrunc(gallivm
->builder
, cond
,
2450 lp_build_if(&if_ctx
, gallivm
, cond
);
2453 for (j
= 0; j
< 4; j
++) {
2454 addr
= ctx
->radeon_bld
.soa
.outputs
[i
][j
];
2455 val
= LLVMBuildLoad(gallivm
->builder
, addr
, "");
2456 val
= radeon_llvm_saturate(bld_base
, val
);
2457 LLVMBuildStore(gallivm
->builder
, val
, addr
);
2462 lp_build_endif(&if_ctx
);
2465 for (i
= 0; i
< info
->num_outputs
; i
++) {
2466 outputs
[i
].name
= info
->output_semantic_name
[i
];
2467 outputs
[i
].sid
= info
->output_semantic_index
[i
];
2469 for (j
= 0; j
< 4; j
++)
2470 outputs
[i
].values
[j
] =
2471 LLVMBuildLoad(gallivm
->builder
,
2472 ctx
->radeon_bld
.soa
.outputs
[i
][j
],
2476 if (ctx
->is_monolithic
) {
2477 /* Export PrimitiveID when PS needs it. */
2478 if (si_vs_exports_prim_id(ctx
->shader
)) {
2479 outputs
[i
].name
= TGSI_SEMANTIC_PRIMID
;
2481 outputs
[i
].values
[0] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
2482 get_primitive_id(bld_base
, 0));
2483 outputs
[i
].values
[1] = bld_base
->base
.undef
;
2484 outputs
[i
].values
[2] = bld_base
->base
.undef
;
2485 outputs
[i
].values
[3] = bld_base
->base
.undef
;
2489 /* Return the primitive ID from the LLVM function. */
2491 LLVMBuildInsertValue(gallivm
->builder
,
2493 bitcast(bld_base
, TGSI_TYPE_FLOAT
,
2494 get_primitive_id(bld_base
, 0)),
2495 VS_EPILOG_PRIMID_LOC
, "");
2498 si_llvm_export_vs(bld_base
, outputs
, i
);
2502 static void si_export_mrt_z(struct lp_build_tgsi_context
*bld_base
,
2503 LLVMValueRef depth
, LLVMValueRef stencil
,
2504 LLVMValueRef samplemask
)
2506 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2507 struct lp_build_context
*base
= &bld_base
->base
;
2508 struct lp_build_context
*uint
= &bld_base
->uint_bld
;
2509 LLVMValueRef args
[9];
2512 assert(depth
|| stencil
|| samplemask
);
2514 args
[1] = uint
->one
; /* whether the EXEC mask is valid */
2515 args
[2] = uint
->one
; /* DONE bit */
2517 /* Specify the target we are exporting */
2518 args
[3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_MRTZ
);
2520 args
[4] = uint
->zero
; /* COMP flag */
2521 args
[5] = base
->undef
; /* R, depth */
2522 args
[6] = base
->undef
; /* G, stencil test value[0:7], stencil op value[8:15] */
2523 args
[7] = base
->undef
; /* B, sample mask */
2524 args
[8] = base
->undef
; /* A, alpha to mask */
2537 args
[7] = samplemask
;
2541 /* SI (except OLAND) has a bug that it only looks
2542 * at the X writemask component. */
2543 if (ctx
->screen
->b
.chip_class
== SI
&&
2544 ctx
->screen
->b
.family
!= CHIP_OLAND
)
2547 /* Specify which components to enable */
2548 args
[0] = lp_build_const_int32(base
->gallivm
, mask
);
2550 lp_build_intrinsic(base
->gallivm
->builder
, "llvm.SI.export",
2551 ctx
->voidt
, args
, 9, 0);
2554 static void si_export_mrt_color(struct lp_build_tgsi_context
*bld_base
,
2555 LLVMValueRef
*color
, unsigned index
,
2556 unsigned samplemask_param
,
2559 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2560 struct lp_build_context
*base
= &bld_base
->base
;
2564 if (ctx
->shader
->key
.ps
.epilog
.clamp_color
)
2565 for (i
= 0; i
< 4; i
++)
2566 color
[i
] = radeon_llvm_saturate(bld_base
, color
[i
]);
2569 if (ctx
->shader
->key
.ps
.epilog
.alpha_to_one
)
2570 color
[3] = base
->one
;
2574 ctx
->shader
->key
.ps
.epilog
.alpha_func
!= PIPE_FUNC_ALWAYS
)
2575 si_alpha_test(bld_base
, color
[3]);
2577 /* Line & polygon smoothing */
2578 if (ctx
->shader
->key
.ps
.epilog
.poly_line_smoothing
)
2579 color
[3] = si_scale_alpha_by_sample_mask(bld_base
, color
[3],
2582 /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
2583 if (ctx
->shader
->key
.ps
.epilog
.last_cbuf
> 0) {
2584 LLVMValueRef args
[8][9];
2587 /* Get the export arguments, also find out what the last one is. */
2588 for (c
= 0; c
<= ctx
->shader
->key
.ps
.epilog
.last_cbuf
; c
++) {
2589 si_llvm_init_export_args(bld_base
, color
,
2590 V_008DFC_SQ_EXP_MRT
+ c
, args
[c
]);
2591 if (args
[c
][0] != bld_base
->uint_bld
.zero
)
2595 /* Emit all exports. */
2596 for (c
= 0; c
<= ctx
->shader
->key
.ps
.epilog
.last_cbuf
; c
++) {
2597 if (is_last
&& last
== c
) {
2598 args
[c
][1] = bld_base
->uint_bld
.one
; /* whether the EXEC mask is valid */
2599 args
[c
][2] = bld_base
->uint_bld
.one
; /* DONE bit */
2600 } else if (args
[c
][0] == bld_base
->uint_bld
.zero
)
2601 continue; /* unnecessary NULL export */
2603 lp_build_intrinsic(base
->gallivm
->builder
, "llvm.SI.export",
2604 ctx
->voidt
, args
[c
], 9, 0);
2607 LLVMValueRef args
[9];
2610 si_llvm_init_export_args(bld_base
, color
, V_008DFC_SQ_EXP_MRT
+ index
,
2613 args
[1] = bld_base
->uint_bld
.one
; /* whether the EXEC mask is valid */
2614 args
[2] = bld_base
->uint_bld
.one
; /* DONE bit */
2615 } else if (args
[0] == bld_base
->uint_bld
.zero
)
2616 return; /* unnecessary NULL export */
2618 lp_build_intrinsic(base
->gallivm
->builder
, "llvm.SI.export",
2619 ctx
->voidt
, args
, 9, 0);
2623 static void si_export_null(struct lp_build_tgsi_context
*bld_base
)
2625 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2626 struct lp_build_context
*base
= &bld_base
->base
;
2627 struct lp_build_context
*uint
= &bld_base
->uint_bld
;
2628 LLVMValueRef args
[9];
2630 args
[0] = lp_build_const_int32(base
->gallivm
, 0x0); /* enabled channels */
2631 args
[1] = uint
->one
; /* whether the EXEC mask is valid */
2632 args
[2] = uint
->one
; /* DONE bit */
2633 args
[3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_NULL
);
2634 args
[4] = uint
->zero
; /* COMPR flag (0 = 32-bit export) */
2635 args
[5] = uint
->undef
; /* R */
2636 args
[6] = uint
->undef
; /* G */
2637 args
[7] = uint
->undef
; /* B */
2638 args
[8] = uint
->undef
; /* A */
2640 lp_build_intrinsic(base
->gallivm
->builder
, "llvm.SI.export",
2641 ctx
->voidt
, args
, 9, 0);
2644 static void si_llvm_emit_fs_epilogue(struct lp_build_tgsi_context
*bld_base
)
2646 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2647 struct si_shader
*shader
= ctx
->shader
;
2648 struct lp_build_context
*base
= &bld_base
->base
;
2649 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
2650 LLVMBuilderRef builder
= base
->gallivm
->builder
;
2651 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
2652 int last_color_export
= -1;
2655 /* Determine the last export. If MRTZ is present, it's always last.
2656 * Otherwise, find the last color export.
2658 if (!info
->writes_z
&& !info
->writes_stencil
&& !info
->writes_samplemask
) {
2659 unsigned spi_format
= shader
->key
.ps
.epilog
.spi_shader_col_format
;
2661 /* Don't export NULL and return if alpha-test is enabled. */
2662 if (shader
->key
.ps
.epilog
.alpha_func
!= PIPE_FUNC_ALWAYS
&&
2663 shader
->key
.ps
.epilog
.alpha_func
!= PIPE_FUNC_NEVER
&&
2664 (spi_format
& 0xf) == 0)
2665 spi_format
|= V_028714_SPI_SHADER_32_AR
;
2667 for (i
= 0; i
< info
->num_outputs
; i
++) {
2668 unsigned index
= info
->output_semantic_index
[i
];
2670 if (info
->output_semantic_name
[i
] != TGSI_SEMANTIC_COLOR
)
2673 /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
2674 if (shader
->key
.ps
.epilog
.last_cbuf
> 0) {
2675 /* Just set this if any of the colorbuffers are enabled. */
2677 ((1llu << (4 * (shader
->key
.ps
.epilog
.last_cbuf
+ 1))) - 1))
2678 last_color_export
= i
;
2682 if ((spi_format
>> (index
* 4)) & 0xf)
2683 last_color_export
= i
;
2686 /* If there are no outputs, export NULL. */
2687 if (last_color_export
== -1) {
2688 si_export_null(bld_base
);
2693 for (i
= 0; i
< info
->num_outputs
; i
++) {
2694 unsigned semantic_name
= info
->output_semantic_name
[i
];
2695 unsigned semantic_index
= info
->output_semantic_index
[i
];
2697 LLVMValueRef color
[4] = {};
2699 /* Select the correct target */
2700 switch (semantic_name
) {
2701 case TGSI_SEMANTIC_POSITION
:
2702 depth
= LLVMBuildLoad(builder
,
2703 ctx
->radeon_bld
.soa
.outputs
[i
][2], "");
2705 case TGSI_SEMANTIC_STENCIL
:
2706 stencil
= LLVMBuildLoad(builder
,
2707 ctx
->radeon_bld
.soa
.outputs
[i
][1], "");
2709 case TGSI_SEMANTIC_SAMPLEMASK
:
2710 samplemask
= LLVMBuildLoad(builder
,
2711 ctx
->radeon_bld
.soa
.outputs
[i
][0], "");
2713 case TGSI_SEMANTIC_COLOR
:
2714 for (j
= 0; j
< 4; j
++)
2715 color
[j
] = LLVMBuildLoad(builder
,
2716 ctx
->radeon_bld
.soa
.outputs
[i
][j
], "");
2718 si_export_mrt_color(bld_base
, color
, semantic_index
,
2719 SI_PARAM_SAMPLE_COVERAGE
,
2720 last_color_export
== i
);
2724 "Warning: SI unhandled fs output type:%d\n",
2729 if (depth
|| stencil
|| samplemask
)
2730 si_export_mrt_z(bld_base
, depth
, stencil
, samplemask
);
2734 * Return PS outputs in this order:
2736 * v[0:3] = color0.xyzw
2737 * v[4:7] = color1.xyzw
2742 * vN+3 = SampleMaskIn (used for OpenGL smoothing)
2744 * The alpha-ref SGPR is returned via its original location.
2746 static void si_llvm_return_fs_outputs(struct lp_build_tgsi_context
*bld_base
)
2748 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2749 struct si_shader
*shader
= ctx
->shader
;
2750 struct lp_build_context
*base
= &bld_base
->base
;
2751 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
2752 LLVMBuilderRef builder
= base
->gallivm
->builder
;
2753 unsigned i
, j
, first_vgpr
, vgpr
;
2755 LLVMValueRef color
[8][4] = {};
2756 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
2759 /* Read the output values. */
2760 for (i
= 0; i
< info
->num_outputs
; i
++) {
2761 unsigned semantic_name
= info
->output_semantic_name
[i
];
2762 unsigned semantic_index
= info
->output_semantic_index
[i
];
2764 switch (semantic_name
) {
2765 case TGSI_SEMANTIC_COLOR
:
2766 assert(semantic_index
< 8);
2767 for (j
= 0; j
< 4; j
++) {
2768 LLVMValueRef ptr
= ctx
->radeon_bld
.soa
.outputs
[i
][j
];
2769 LLVMValueRef result
= LLVMBuildLoad(builder
, ptr
, "");
2770 color
[semantic_index
][j
] = result
;
2773 case TGSI_SEMANTIC_POSITION
:
2774 depth
= LLVMBuildLoad(builder
,
2775 ctx
->radeon_bld
.soa
.outputs
[i
][2], "");
2777 case TGSI_SEMANTIC_STENCIL
:
2778 stencil
= LLVMBuildLoad(builder
,
2779 ctx
->radeon_bld
.soa
.outputs
[i
][1], "");
2781 case TGSI_SEMANTIC_SAMPLEMASK
:
2782 samplemask
= LLVMBuildLoad(builder
,
2783 ctx
->radeon_bld
.soa
.outputs
[i
][0], "");
2786 fprintf(stderr
, "Warning: SI unhandled fs output type:%d\n",
2791 /* Fill the return structure. */
2792 ret
= ctx
->return_value
;
2795 ret
= LLVMBuildInsertValue(builder
, ret
,
2796 bitcast(bld_base
, TGSI_TYPE_SIGNED
,
2797 LLVMGetParam(ctx
->radeon_bld
.main_fn
,
2798 SI_PARAM_ALPHA_REF
)),
2799 SI_SGPR_ALPHA_REF
, "");
2802 first_vgpr
= vgpr
= SI_SGPR_ALPHA_REF
+ 1;
2803 for (i
= 0; i
< ARRAY_SIZE(color
); i
++) {
2807 for (j
= 0; j
< 4; j
++)
2808 ret
= LLVMBuildInsertValue(builder
, ret
, color
[i
][j
], vgpr
++, "");
2811 ret
= LLVMBuildInsertValue(builder
, ret
, depth
, vgpr
++, "");
2813 ret
= LLVMBuildInsertValue(builder
, ret
, stencil
, vgpr
++, "");
2815 ret
= LLVMBuildInsertValue(builder
, ret
, samplemask
, vgpr
++, "");
2817 /* Add the input sample mask for smoothing at the end. */
2818 if (vgpr
< first_vgpr
+ PS_EPILOG_SAMPLEMASK_MIN_LOC
)
2819 vgpr
= first_vgpr
+ PS_EPILOG_SAMPLEMASK_MIN_LOC
;
2820 ret
= LLVMBuildInsertValue(builder
, ret
,
2821 LLVMGetParam(ctx
->radeon_bld
.main_fn
,
2822 SI_PARAM_SAMPLE_COVERAGE
), vgpr
++, "");
2824 ctx
->return_value
= ret
;
2828 * Given a v8i32 resource descriptor for a buffer, extract the size of the
2829 * buffer in number of elements and return it as an i32.
2831 static LLVMValueRef
get_buffer_size(
2832 struct lp_build_tgsi_context
*bld_base
,
2833 LLVMValueRef descriptor
)
2835 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2836 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2837 LLVMBuilderRef builder
= gallivm
->builder
;
2839 LLVMBuildExtractElement(builder
, descriptor
,
2840 lp_build_const_int32(gallivm
, 6), "");
2842 if (ctx
->screen
->b
.chip_class
>= VI
) {
2843 /* On VI, the descriptor contains the size in bytes,
2844 * but TXQ must return the size in elements.
2845 * The stride is always non-zero for resources using TXQ.
2847 LLVMValueRef stride
=
2848 LLVMBuildExtractElement(builder
, descriptor
,
2849 lp_build_const_int32(gallivm
, 5), "");
2850 stride
= LLVMBuildLShr(builder
, stride
,
2851 lp_build_const_int32(gallivm
, 16), "");
2852 stride
= LLVMBuildAnd(builder
, stride
,
2853 lp_build_const_int32(gallivm
, 0x3FFF), "");
2855 size
= LLVMBuildUDiv(builder
, size
, stride
, "");
2862 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
2865 static void build_int_type_name(
2867 char *buf
, unsigned bufsize
)
2869 assert(bufsize
>= 6);
2871 if (LLVMGetTypeKind(type
) == LLVMVectorTypeKind
)
2872 snprintf(buf
, bufsize
, "v%ui32",
2873 LLVMGetVectorSize(type
));
2878 static void build_tex_intrinsic(const struct lp_build_tgsi_action
*action
,
2879 struct lp_build_tgsi_context
*bld_base
,
2880 struct lp_build_emit_data
*emit_data
);
2882 /* Prevent optimizations (at least of memory accesses) across the current
2883 * point in the program by emitting empty inline assembly that is marked as
2884 * having side effects.
2886 static void emit_optimization_barrier(struct si_shader_context
*ctx
)
2888 LLVMBuilderRef builder
= ctx
->radeon_bld
.gallivm
.builder
;
2889 LLVMTypeRef ftype
= LLVMFunctionType(ctx
->voidt
, NULL
, 0, false);
2890 LLVMValueRef inlineasm
= LLVMConstInlineAsm(ftype
, "", "", true, false);
2891 LLVMBuildCall(builder
, inlineasm
, NULL
, 0, "");
2894 static void membar_emit(
2895 const struct lp_build_tgsi_action
*action
,
2896 struct lp_build_tgsi_context
*bld_base
,
2897 struct lp_build_emit_data
*emit_data
)
2899 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2901 /* Since memoryBarrier only makes guarantees about atomics and
2902 * coherent image accesses (which bypass TC L1), we do not need to emit
2903 * any special cache handling here.
2905 * We do have to prevent LLVM from re-ordering loads across
2906 * the barrier though.
2908 emit_optimization_barrier(ctx
);
2912 shader_buffer_fetch_rsrc(struct si_shader_context
*ctx
,
2913 const struct tgsi_full_src_register
*reg
)
2915 LLVMValueRef ind_index
;
2916 LLVMValueRef rsrc_ptr
;
2918 if (!reg
->Register
.Indirect
)
2919 return ctx
->shader_buffers
[reg
->Register
.Index
];
2921 ind_index
= get_bounded_indirect_index(ctx
, ®
->Indirect
,
2922 reg
->Register
.Index
,
2923 SI_NUM_SHADER_BUFFERS
);
2925 rsrc_ptr
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_SHADER_BUFFERS
);
2926 return build_indexed_load_const(ctx
, rsrc_ptr
, ind_index
);
2929 static bool tgsi_is_array_sampler(unsigned target
)
2931 return target
== TGSI_TEXTURE_1D_ARRAY
||
2932 target
== TGSI_TEXTURE_SHADOW1D_ARRAY
||
2933 target
== TGSI_TEXTURE_2D_ARRAY
||
2934 target
== TGSI_TEXTURE_SHADOW2D_ARRAY
||
2935 target
== TGSI_TEXTURE_CUBE_ARRAY
||
2936 target
== TGSI_TEXTURE_SHADOWCUBE_ARRAY
||
2937 target
== TGSI_TEXTURE_2D_ARRAY_MSAA
;
2940 static bool tgsi_is_array_image(unsigned target
)
2942 return target
== TGSI_TEXTURE_3D
||
2943 target
== TGSI_TEXTURE_CUBE
||
2944 target
== TGSI_TEXTURE_1D_ARRAY
||
2945 target
== TGSI_TEXTURE_2D_ARRAY
||
2946 target
== TGSI_TEXTURE_CUBE_ARRAY
||
2947 target
== TGSI_TEXTURE_2D_ARRAY_MSAA
;
2951 * Given a 256-bit resource descriptor, force the DCC enable bit to off.
2953 * At least on Tonga, executing image stores on images with DCC enabled and
2954 * non-trivial can eventually lead to lockups. This can occur when an
2955 * application binds an image as read-only but then uses a shader that writes
2956 * to it. The OpenGL spec allows almost arbitrarily bad behavior (including
2957 * program termination) in this case, but it doesn't cost much to be a bit
2958 * nicer: disabling DCC in the shader still leads to undefined results but
2959 * avoids the lockup.
2961 static LLVMValueRef
force_dcc_off(struct si_shader_context
*ctx
,
2964 if (ctx
->screen
->b
.chip_class
<= CIK
) {
2967 LLVMBuilderRef builder
= ctx
->radeon_bld
.gallivm
.builder
;
2968 LLVMValueRef i32_6
= LLVMConstInt(ctx
->i32
, 6, 0);
2969 LLVMValueRef i32_C
= LLVMConstInt(ctx
->i32
, C_008F28_COMPRESSION_EN
, 0);
2972 tmp
= LLVMBuildExtractElement(builder
, rsrc
, i32_6
, "");
2973 tmp
= LLVMBuildAnd(builder
, tmp
, i32_C
, "");
2974 return LLVMBuildInsertElement(builder
, rsrc
, tmp
, i32_6
, "");
2979 * Load the resource descriptor for \p image.
2983 struct lp_build_tgsi_context
*bld_base
,
2984 const struct tgsi_full_src_register
*image
,
2988 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2990 assert(image
->Register
.File
== TGSI_FILE_IMAGE
);
2992 if (!image
->Register
.Indirect
) {
2993 /* Fast path: use preloaded resources */
2994 *rsrc
= ctx
->images
[image
->Register
.Index
];
2996 /* Indexing and manual load */
2997 LLVMValueRef ind_index
;
2998 LLVMValueRef rsrc_ptr
;
3001 /* From the GL_ARB_shader_image_load_store extension spec:
3003 * If a shader performs an image load, store, or atomic
3004 * operation using an image variable declared as an array,
3005 * and if the index used to select an individual element is
3006 * negative or greater than or equal to the size of the
3007 * array, the results of the operation are undefined but may
3008 * not lead to termination.
3010 ind_index
= get_bounded_indirect_index(ctx
, &image
->Indirect
,
3011 image
->Register
.Index
,
3014 rsrc_ptr
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_IMAGES
);
3015 tmp
= build_indexed_load_const(ctx
, rsrc_ptr
, ind_index
);
3017 tmp
= force_dcc_off(ctx
, tmp
);
3022 static LLVMValueRef
image_fetch_coords(
3023 struct lp_build_tgsi_context
*bld_base
,
3024 const struct tgsi_full_instruction
*inst
,
3027 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3028 LLVMBuilderRef builder
= gallivm
->builder
;
3029 unsigned target
= inst
->Memory
.Texture
;
3030 unsigned num_coords
= tgsi_util_get_texture_coord_dim(target
);
3031 LLVMValueRef coords
[4];
3035 for (chan
= 0; chan
< num_coords
; ++chan
) {
3036 tmp
= lp_build_emit_fetch(bld_base
, inst
, src
, chan
);
3037 tmp
= LLVMBuildBitCast(builder
, tmp
, bld_base
->uint_bld
.elem_type
, "");
3041 if (num_coords
== 1)
3044 if (num_coords
== 3) {
3045 /* LLVM has difficulties lowering 3-element vectors. */
3046 coords
[3] = bld_base
->uint_bld
.undef
;
3050 return lp_build_gather_values(gallivm
, coords
, num_coords
);
3054 * Append the extra mode bits that are used by image load and store.
3056 static void image_append_args(
3057 struct si_shader_context
*ctx
,
3058 struct lp_build_emit_data
* emit_data
,
3062 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3063 LLVMValueRef i1false
= LLVMConstInt(ctx
->i1
, 0, 0);
3064 LLVMValueRef i1true
= LLVMConstInt(ctx
->i1
, 1, 0);
3066 emit_data
->args
[emit_data
->arg_count
++] = i1false
; /* r128 */
3067 emit_data
->args
[emit_data
->arg_count
++] =
3068 tgsi_is_array_image(target
) ? i1true
: i1false
; /* da */
3070 emit_data
->args
[emit_data
->arg_count
++] =
3071 inst
->Memory
.Qualifier
& (TGSI_MEMORY_COHERENT
| TGSI_MEMORY_VOLATILE
) ?
3072 i1true
: i1false
; /* glc */
3074 emit_data
->args
[emit_data
->arg_count
++] = i1false
; /* slc */
3078 * Given a 256 bit resource, extract the top half (which stores the buffer
3079 * resource in the case of textures and images).
3081 static LLVMValueRef
extract_rsrc_top_half(
3082 struct si_shader_context
*ctx
,
3085 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
3086 struct lp_build_tgsi_context
*bld_base
= &ctx
->radeon_bld
.soa
.bld_base
;
3087 LLVMTypeRef v2i128
= LLVMVectorType(ctx
->i128
, 2);
3089 rsrc
= LLVMBuildBitCast(gallivm
->builder
, rsrc
, v2i128
, "");
3090 rsrc
= LLVMBuildExtractElement(gallivm
->builder
, rsrc
, bld_base
->uint_bld
.one
, "");
3091 rsrc
= LLVMBuildBitCast(gallivm
->builder
, rsrc
, ctx
->v4i32
, "");
3097 * Append the resource and indexing arguments for buffer intrinsics.
3099 * \param rsrc the v4i32 buffer resource
3100 * \param index index into the buffer (stride-based)
3101 * \param offset byte offset into the buffer
3103 static void buffer_append_args(
3104 struct si_shader_context
*ctx
,
3105 struct lp_build_emit_data
*emit_data
,
3108 LLVMValueRef offset
,
3111 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3112 LLVMValueRef i1false
= LLVMConstInt(ctx
->i1
, 0, 0);
3113 LLVMValueRef i1true
= LLVMConstInt(ctx
->i1
, 1, 0);
3115 emit_data
->args
[emit_data
->arg_count
++] = rsrc
;
3116 emit_data
->args
[emit_data
->arg_count
++] = index
; /* vindex */
3117 emit_data
->args
[emit_data
->arg_count
++] = offset
; /* voffset */
3119 emit_data
->args
[emit_data
->arg_count
++] =
3120 inst
->Memory
.Qualifier
& (TGSI_MEMORY_COHERENT
| TGSI_MEMORY_VOLATILE
) ?
3121 i1true
: i1false
; /* glc */
3123 emit_data
->args
[emit_data
->arg_count
++] = i1false
; /* slc */
3126 static void load_fetch_args(
3127 struct lp_build_tgsi_context
* bld_base
,
3128 struct lp_build_emit_data
* emit_data
)
3130 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3131 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3132 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
3133 unsigned target
= inst
->Memory
.Texture
;
3136 emit_data
->dst_type
= LLVMVectorType(bld_base
->base
.elem_type
, 4);
3138 if (inst
->Src
[0].Register
.File
== TGSI_FILE_BUFFER
) {
3139 LLVMBuilderRef builder
= gallivm
->builder
;
3140 LLVMValueRef offset
;
3143 rsrc
= shader_buffer_fetch_rsrc(ctx
, &inst
->Src
[0]);
3145 tmp
= lp_build_emit_fetch(bld_base
, inst
, 1, 0);
3146 offset
= LLVMBuildBitCast(builder
, tmp
, bld_base
->uint_bld
.elem_type
, "");
3148 buffer_append_args(ctx
, emit_data
, rsrc
, bld_base
->uint_bld
.zero
,
3150 } else if (inst
->Src
[0].Register
.File
== TGSI_FILE_IMAGE
) {
3151 LLVMValueRef coords
;
3153 image_fetch_rsrc(bld_base
, &inst
->Src
[0], false, &rsrc
);
3154 coords
= image_fetch_coords(bld_base
, inst
, 1);
3156 if (target
== TGSI_TEXTURE_BUFFER
) {
3157 rsrc
= extract_rsrc_top_half(ctx
, rsrc
);
3158 buffer_append_args(ctx
, emit_data
, rsrc
, coords
,
3159 bld_base
->uint_bld
.zero
, false);
3161 emit_data
->args
[0] = coords
;
3162 emit_data
->args
[1] = rsrc
;
3163 emit_data
->args
[2] = lp_build_const_int32(gallivm
, 15); /* dmask */
3164 emit_data
->arg_count
= 3;
3166 image_append_args(ctx
, emit_data
, target
, false);
3171 static void load_emit_buffer(struct si_shader_context
*ctx
,
3172 struct lp_build_emit_data
*emit_data
)
3174 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3175 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
3176 LLVMBuilderRef builder
= gallivm
->builder
;
3177 uint writemask
= inst
->Dst
[0].Register
.WriteMask
;
3178 uint count
= util_last_bit(writemask
);
3179 const char *intrinsic_name
;
3180 LLVMTypeRef dst_type
;
3184 intrinsic_name
= "llvm.amdgcn.buffer.load.f32";
3185 dst_type
= ctx
->f32
;
3188 intrinsic_name
= "llvm.amdgcn.buffer.load.v2f32";
3189 dst_type
= LLVMVectorType(ctx
->f32
, 2);
3192 intrinsic_name
= "llvm.amdgcn.buffer.load.v4f32";
3193 dst_type
= ctx
->v4f32
;
3197 emit_data
->output
[emit_data
->chan
] = lp_build_intrinsic(
3198 builder
, intrinsic_name
, dst_type
,
3199 emit_data
->args
, emit_data
->arg_count
,
3200 LLVMReadOnlyAttribute
| LLVMNoUnwindAttribute
);
3203 static LLVMValueRef
get_memory_ptr(struct si_shader_context
*ctx
,
3204 const struct tgsi_full_instruction
*inst
,
3205 LLVMTypeRef type
, int arg
)
3207 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
3208 LLVMBuilderRef builder
= gallivm
->builder
;
3209 LLVMValueRef offset
, ptr
;
3212 offset
= lp_build_emit_fetch(&ctx
->radeon_bld
.soa
.bld_base
, inst
, arg
, 0);
3213 offset
= LLVMBuildBitCast(builder
, offset
, ctx
->i32
, "");
3215 ptr
= ctx
->shared_memory
;
3216 ptr
= LLVMBuildGEP(builder
, ptr
, &offset
, 1, "");
3217 addr_space
= LLVMGetPointerAddressSpace(LLVMTypeOf(ptr
));
3218 ptr
= LLVMBuildBitCast(builder
, ptr
, LLVMPointerType(type
, addr_space
), "");
3223 static void load_emit_memory(
3224 struct si_shader_context
*ctx
,
3225 struct lp_build_emit_data
*emit_data
)
3227 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3228 struct lp_build_context
*base
= &ctx
->radeon_bld
.soa
.bld_base
.base
;
3229 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
3230 LLVMBuilderRef builder
= gallivm
->builder
;
3231 unsigned writemask
= inst
->Dst
[0].Register
.WriteMask
;
3232 LLVMValueRef channels
[4], ptr
, derived_ptr
, index
;
3235 ptr
= get_memory_ptr(ctx
, inst
, base
->elem_type
, 1);
3237 for (chan
= 0; chan
< 4; ++chan
) {
3238 if (!(writemask
& (1 << chan
))) {
3239 channels
[chan
] = LLVMGetUndef(base
->elem_type
);
3243 index
= lp_build_const_int32(gallivm
, chan
);
3244 derived_ptr
= LLVMBuildGEP(builder
, ptr
, &index
, 1, "");
3245 channels
[chan
] = LLVMBuildLoad(builder
, derived_ptr
, "");
3247 emit_data
->output
[emit_data
->chan
] = lp_build_gather_values(gallivm
, channels
, 4);
3250 static void load_emit(
3251 const struct lp_build_tgsi_action
*action
,
3252 struct lp_build_tgsi_context
*bld_base
,
3253 struct lp_build_emit_data
*emit_data
)
3255 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3256 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3257 LLVMBuilderRef builder
= gallivm
->builder
;
3258 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
3259 char intrinsic_name
[32];
3260 char coords_type
[8];
3262 if (inst
->Src
[0].Register
.File
== TGSI_FILE_MEMORY
) {
3263 load_emit_memory(ctx
, emit_data
);
3267 if (inst
->Memory
.Qualifier
& TGSI_MEMORY_VOLATILE
)
3268 emit_optimization_barrier(ctx
);
3270 if (inst
->Src
[0].Register
.File
== TGSI_FILE_BUFFER
) {
3271 load_emit_buffer(ctx
, emit_data
);
3275 if (inst
->Memory
.Texture
== TGSI_TEXTURE_BUFFER
) {
3276 emit_data
->output
[emit_data
->chan
] =
3278 builder
, "llvm.amdgcn.buffer.load.format.v4f32", emit_data
->dst_type
,
3279 emit_data
->args
, emit_data
->arg_count
,
3280 LLVMReadOnlyAttribute
| LLVMNoUnwindAttribute
);
3282 build_int_type_name(LLVMTypeOf(emit_data
->args
[0]),
3283 coords_type
, sizeof(coords_type
));
3285 snprintf(intrinsic_name
, sizeof(intrinsic_name
),
3286 "llvm.amdgcn.image.load.%s", coords_type
);
3288 emit_data
->output
[emit_data
->chan
] =
3290 builder
, intrinsic_name
, emit_data
->dst_type
,
3291 emit_data
->args
, emit_data
->arg_count
,
3292 LLVMReadOnlyAttribute
| LLVMNoUnwindAttribute
);
3296 static void store_fetch_args(
3297 struct lp_build_tgsi_context
* bld_base
,
3298 struct lp_build_emit_data
* emit_data
)
3300 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3301 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3302 LLVMBuilderRef builder
= gallivm
->builder
;
3303 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
3304 struct tgsi_full_src_register memory
;
3305 LLVMValueRef chans
[4];
3310 emit_data
->dst_type
= LLVMVoidTypeInContext(gallivm
->context
);
3312 for (chan
= 0; chan
< 4; ++chan
) {
3313 chans
[chan
] = lp_build_emit_fetch(bld_base
, inst
, 1, chan
);
3315 data
= lp_build_gather_values(gallivm
, chans
, 4);
3317 emit_data
->args
[emit_data
->arg_count
++] = data
;
3319 memory
= tgsi_full_src_register_from_dst(&inst
->Dst
[0]);
3321 if (inst
->Dst
[0].Register
.File
== TGSI_FILE_BUFFER
) {
3322 LLVMValueRef offset
;
3325 rsrc
= shader_buffer_fetch_rsrc(ctx
, &memory
);
3327 tmp
= lp_build_emit_fetch(bld_base
, inst
, 0, 0);
3328 offset
= LLVMBuildBitCast(builder
, tmp
, bld_base
->uint_bld
.elem_type
, "");
3330 buffer_append_args(ctx
, emit_data
, rsrc
, bld_base
->uint_bld
.zero
,
3332 } else if (inst
->Dst
[0].Register
.File
== TGSI_FILE_IMAGE
) {
3333 unsigned target
= inst
->Memory
.Texture
;
3334 LLVMValueRef coords
;
3336 coords
= image_fetch_coords(bld_base
, inst
, 0);
3338 if (target
== TGSI_TEXTURE_BUFFER
) {
3339 image_fetch_rsrc(bld_base
, &memory
, false, &rsrc
);
3341 rsrc
= extract_rsrc_top_half(ctx
, rsrc
);
3342 buffer_append_args(ctx
, emit_data
, rsrc
, coords
,
3343 bld_base
->uint_bld
.zero
, false);
3345 emit_data
->args
[1] = coords
;
3346 image_fetch_rsrc(bld_base
, &memory
, true, &emit_data
->args
[2]);
3347 emit_data
->args
[3] = lp_build_const_int32(gallivm
, 15); /* dmask */
3348 emit_data
->arg_count
= 4;
3350 image_append_args(ctx
, emit_data
, target
, false);
3355 static void store_emit_buffer(
3356 struct si_shader_context
*ctx
,
3357 struct lp_build_emit_data
*emit_data
)
3359 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3360 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
3361 LLVMBuilderRef builder
= gallivm
->builder
;
3362 struct lp_build_context
*uint_bld
= &ctx
->radeon_bld
.soa
.bld_base
.uint_bld
;
3363 LLVMValueRef base_data
= emit_data
->args
[0];
3364 LLVMValueRef base_offset
= emit_data
->args
[3];
3365 unsigned writemask
= inst
->Dst
[0].Register
.WriteMask
;
3369 const char *intrinsic_name
;
3371 LLVMValueRef offset
;
3374 u_bit_scan_consecutive_range(&writemask
, &start
, &count
);
3376 /* Due to an LLVM limitation, split 3-element writes
3377 * into a 2-element and a 1-element write. */
3379 writemask
|= 1 << (start
+ 2);
3385 intrinsic_name
= "llvm.amdgcn.buffer.store.v4f32";
3386 } else if (count
== 2) {
3387 LLVMTypeRef v2f32
= LLVMVectorType(ctx
->f32
, 2);
3389 tmp
= LLVMBuildExtractElement(
3391 lp_build_const_int32(gallivm
, start
), "");
3392 data
= LLVMBuildInsertElement(
3393 builder
, LLVMGetUndef(v2f32
), tmp
,
3394 uint_bld
->zero
, "");
3396 tmp
= LLVMBuildExtractElement(
3398 lp_build_const_int32(gallivm
, start
+ 1), "");
3399 data
= LLVMBuildInsertElement(
3400 builder
, data
, tmp
, uint_bld
->one
, "");
3402 intrinsic_name
= "llvm.amdgcn.buffer.store.v2f32";
3405 data
= LLVMBuildExtractElement(
3407 lp_build_const_int32(gallivm
, start
), "");
3408 intrinsic_name
= "llvm.amdgcn.buffer.store.f32";
3411 offset
= base_offset
;
3413 offset
= LLVMBuildAdd(
3415 lp_build_const_int32(gallivm
, start
* 4), "");
3418 emit_data
->args
[0] = data
;
3419 emit_data
->args
[3] = offset
;
3422 builder
, intrinsic_name
, emit_data
->dst_type
,
3423 emit_data
->args
, emit_data
->arg_count
,
3424 LLVMNoUnwindAttribute
);
3428 static void store_emit_memory(
3429 struct si_shader_context
*ctx
,
3430 struct lp_build_emit_data
*emit_data
)
3432 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3433 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
3434 struct lp_build_context
*base
= &ctx
->radeon_bld
.soa
.bld_base
.base
;
3435 LLVMBuilderRef builder
= gallivm
->builder
;
3436 unsigned writemask
= inst
->Dst
[0].Register
.WriteMask
;
3437 LLVMValueRef ptr
, derived_ptr
, data
, index
;
3440 ptr
= get_memory_ptr(ctx
, inst
, base
->elem_type
, 0);
3442 for (chan
= 0; chan
< 4; ++chan
) {
3443 if (!(writemask
& (1 << chan
))) {
3446 data
= lp_build_emit_fetch(&ctx
->radeon_bld
.soa
.bld_base
, inst
, 1, chan
);
3447 index
= lp_build_const_int32(gallivm
, chan
);
3448 derived_ptr
= LLVMBuildGEP(builder
, ptr
, &index
, 1, "");
3449 LLVMBuildStore(builder
, data
, derived_ptr
);
3453 static void store_emit(
3454 const struct lp_build_tgsi_action
*action
,
3455 struct lp_build_tgsi_context
*bld_base
,
3456 struct lp_build_emit_data
*emit_data
)
3458 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3459 LLVMBuilderRef builder
= gallivm
->builder
;
3460 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
3461 unsigned target
= inst
->Memory
.Texture
;
3462 char intrinsic_name
[32];
3463 char coords_type
[8];
3465 if (inst
->Dst
[0].Register
.File
== TGSI_FILE_BUFFER
) {
3466 store_emit_buffer(si_shader_context(bld_base
), emit_data
);
3468 } else if (inst
->Dst
[0].Register
.File
== TGSI_FILE_MEMORY
) {
3469 store_emit_memory(si_shader_context(bld_base
), emit_data
);
3473 if (target
== TGSI_TEXTURE_BUFFER
) {
3474 emit_data
->output
[emit_data
->chan
] = lp_build_intrinsic(
3475 builder
, "llvm.amdgcn.buffer.store.format.v4f32",
3476 emit_data
->dst_type
, emit_data
->args
, emit_data
->arg_count
,
3477 LLVMNoUnwindAttribute
);
3479 build_int_type_name(LLVMTypeOf(emit_data
->args
[1]),
3480 coords_type
, sizeof(coords_type
));
3481 snprintf(intrinsic_name
, sizeof(intrinsic_name
),
3482 "llvm.amdgcn.image.store.%s", coords_type
);
3484 emit_data
->output
[emit_data
->chan
] =
3486 builder
, intrinsic_name
, emit_data
->dst_type
,
3487 emit_data
->args
, emit_data
->arg_count
,
3488 LLVMNoUnwindAttribute
);
3492 static void atomic_fetch_args(
3493 struct lp_build_tgsi_context
* bld_base
,
3494 struct lp_build_emit_data
* emit_data
)
3496 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3497 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3498 LLVMBuilderRef builder
= gallivm
->builder
;
3499 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
3500 LLVMValueRef data1
, data2
;
3504 emit_data
->dst_type
= bld_base
->base
.elem_type
;
3506 tmp
= lp_build_emit_fetch(bld_base
, inst
, 2, 0);
3507 data1
= LLVMBuildBitCast(builder
, tmp
, bld_base
->uint_bld
.elem_type
, "");
3509 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_ATOMCAS
) {
3510 tmp
= lp_build_emit_fetch(bld_base
, inst
, 3, 0);
3511 data2
= LLVMBuildBitCast(builder
, tmp
, bld_base
->uint_bld
.elem_type
, "");
3514 /* llvm.amdgcn.image/buffer.atomic.cmpswap reflect the hardware order
3515 * of arguments, which is reversed relative to TGSI (and GLSL)
3517 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_ATOMCAS
)
3518 emit_data
->args
[emit_data
->arg_count
++] = data2
;
3519 emit_data
->args
[emit_data
->arg_count
++] = data1
;
3521 if (inst
->Src
[0].Register
.File
== TGSI_FILE_BUFFER
) {
3522 LLVMValueRef offset
;
3524 rsrc
= shader_buffer_fetch_rsrc(ctx
, &inst
->Src
[0]);
3526 tmp
= lp_build_emit_fetch(bld_base
, inst
, 1, 0);
3527 offset
= LLVMBuildBitCast(builder
, tmp
, bld_base
->uint_bld
.elem_type
, "");
3529 buffer_append_args(ctx
, emit_data
, rsrc
, bld_base
->uint_bld
.zero
,
3531 } else if (inst
->Src
[0].Register
.File
== TGSI_FILE_IMAGE
) {
3532 unsigned target
= inst
->Memory
.Texture
;
3533 LLVMValueRef coords
;
3535 image_fetch_rsrc(bld_base
, &inst
->Src
[0],
3536 target
!= TGSI_TEXTURE_BUFFER
, &rsrc
);
3537 coords
= image_fetch_coords(bld_base
, inst
, 1);
3539 if (target
== TGSI_TEXTURE_BUFFER
) {
3540 rsrc
= extract_rsrc_top_half(ctx
, rsrc
);
3541 buffer_append_args(ctx
, emit_data
, rsrc
, coords
,
3542 bld_base
->uint_bld
.zero
, true);
3544 emit_data
->args
[emit_data
->arg_count
++] = coords
;
3545 emit_data
->args
[emit_data
->arg_count
++] = rsrc
;
3547 image_append_args(ctx
, emit_data
, target
, true);
3552 static void atomic_emit_memory(struct si_shader_context
*ctx
,
3553 struct lp_build_emit_data
*emit_data
) {
3554 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
3555 LLVMBuilderRef builder
= gallivm
->builder
;
3556 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
3557 LLVMValueRef ptr
, result
, arg
;
3559 ptr
= get_memory_ptr(ctx
, inst
, ctx
->i32
, 1);
3561 arg
= lp_build_emit_fetch(&ctx
->radeon_bld
.soa
.bld_base
, inst
, 2, 0);
3562 arg
= LLVMBuildBitCast(builder
, arg
, ctx
->i32
, "");
3564 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_ATOMCAS
) {
3565 LLVMValueRef new_data
;
3566 new_data
= lp_build_emit_fetch(&ctx
->radeon_bld
.soa
.bld_base
,
3569 new_data
= LLVMBuildBitCast(builder
, new_data
, ctx
->i32
, "");
3571 #if HAVE_LLVM >= 0x309
3572 result
= LLVMBuildAtomicCmpXchg(builder
, ptr
, arg
, new_data
,
3573 LLVMAtomicOrderingSequentiallyConsistent
,
3574 LLVMAtomicOrderingSequentiallyConsistent
,
3578 result
= LLVMBuildExtractValue(builder
, result
, 0, "");
3580 LLVMAtomicRMWBinOp op
;
3582 switch(inst
->Instruction
.Opcode
) {
3583 case TGSI_OPCODE_ATOMUADD
:
3584 op
= LLVMAtomicRMWBinOpAdd
;
3586 case TGSI_OPCODE_ATOMXCHG
:
3587 op
= LLVMAtomicRMWBinOpXchg
;
3589 case TGSI_OPCODE_ATOMAND
:
3590 op
= LLVMAtomicRMWBinOpAnd
;
3592 case TGSI_OPCODE_ATOMOR
:
3593 op
= LLVMAtomicRMWBinOpOr
;
3595 case TGSI_OPCODE_ATOMXOR
:
3596 op
= LLVMAtomicRMWBinOpXor
;
3598 case TGSI_OPCODE_ATOMUMIN
:
3599 op
= LLVMAtomicRMWBinOpUMin
;
3601 case TGSI_OPCODE_ATOMUMAX
:
3602 op
= LLVMAtomicRMWBinOpUMax
;
3604 case TGSI_OPCODE_ATOMIMIN
:
3605 op
= LLVMAtomicRMWBinOpMin
;
3607 case TGSI_OPCODE_ATOMIMAX
:
3608 op
= LLVMAtomicRMWBinOpMax
;
3611 unreachable("unknown atomic opcode");
3614 result
= LLVMBuildAtomicRMW(builder
, op
, ptr
, arg
,
3615 LLVMAtomicOrderingSequentiallyConsistent
,
3618 emit_data
->output
[emit_data
->chan
] = LLVMBuildBitCast(builder
, result
, emit_data
->dst_type
, "");
3621 static void atomic_emit(
3622 const struct lp_build_tgsi_action
*action
,
3623 struct lp_build_tgsi_context
*bld_base
,
3624 struct lp_build_emit_data
*emit_data
)
3626 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3627 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3628 LLVMBuilderRef builder
= gallivm
->builder
;
3629 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
3630 char intrinsic_name
[40];
3633 if (inst
->Src
[0].Register
.File
== TGSI_FILE_MEMORY
) {
3634 atomic_emit_memory(ctx
, emit_data
);
3638 if (inst
->Src
[0].Register
.File
== TGSI_FILE_BUFFER
||
3639 inst
->Memory
.Texture
== TGSI_TEXTURE_BUFFER
) {
3640 snprintf(intrinsic_name
, sizeof(intrinsic_name
),
3641 "llvm.amdgcn.buffer.atomic.%s", action
->intr_name
);
3643 char coords_type
[8];
3645 build_int_type_name(LLVMTypeOf(emit_data
->args
[1]),
3646 coords_type
, sizeof(coords_type
));
3647 snprintf(intrinsic_name
, sizeof(intrinsic_name
),
3648 "llvm.amdgcn.image.atomic.%s.%s",
3649 action
->intr_name
, coords_type
);
3652 tmp
= lp_build_intrinsic(
3653 builder
, intrinsic_name
, bld_base
->uint_bld
.elem_type
,
3654 emit_data
->args
, emit_data
->arg_count
,
3655 LLVMNoUnwindAttribute
);
3656 emit_data
->output
[emit_data
->chan
] =
3657 LLVMBuildBitCast(builder
, tmp
, bld_base
->base
.elem_type
, "");
3660 static void resq_fetch_args(
3661 struct lp_build_tgsi_context
* bld_base
,
3662 struct lp_build_emit_data
* emit_data
)
3664 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3665 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3666 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3667 const struct tgsi_full_src_register
*reg
= &inst
->Src
[0];
3669 emit_data
->dst_type
= LLVMVectorType(bld_base
->base
.elem_type
, 4);
3671 if (reg
->Register
.File
== TGSI_FILE_BUFFER
) {
3672 emit_data
->args
[0] = shader_buffer_fetch_rsrc(ctx
, reg
);
3673 emit_data
->arg_count
= 1;
3674 } else if (inst
->Memory
.Texture
== TGSI_TEXTURE_BUFFER
) {
3675 image_fetch_rsrc(bld_base
, reg
, false, &emit_data
->args
[0]);
3676 emit_data
->arg_count
= 1;
3678 emit_data
->args
[0] = bld_base
->uint_bld
.zero
; /* mip level */
3679 image_fetch_rsrc(bld_base
, reg
, false, &emit_data
->args
[1]);
3680 emit_data
->args
[2] = lp_build_const_int32(gallivm
, 15); /* dmask */
3681 emit_data
->args
[3] = bld_base
->uint_bld
.zero
; /* unorm */
3682 emit_data
->args
[4] = bld_base
->uint_bld
.zero
; /* r128 */
3683 emit_data
->args
[5] = tgsi_is_array_image(inst
->Memory
.Texture
) ?
3684 bld_base
->uint_bld
.one
: bld_base
->uint_bld
.zero
; /* da */
3685 emit_data
->args
[6] = bld_base
->uint_bld
.zero
; /* glc */
3686 emit_data
->args
[7] = bld_base
->uint_bld
.zero
; /* slc */
3687 emit_data
->args
[8] = bld_base
->uint_bld
.zero
; /* tfe */
3688 emit_data
->args
[9] = bld_base
->uint_bld
.zero
; /* lwe */
3689 emit_data
->arg_count
= 10;
3693 static void resq_emit(
3694 const struct lp_build_tgsi_action
*action
,
3695 struct lp_build_tgsi_context
*bld_base
,
3696 struct lp_build_emit_data
*emit_data
)
3698 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3699 LLVMBuilderRef builder
= gallivm
->builder
;
3700 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3703 if (inst
->Src
[0].Register
.File
== TGSI_FILE_BUFFER
) {
3704 out
= LLVMBuildExtractElement(builder
, emit_data
->args
[0],
3705 lp_build_const_int32(gallivm
, 2), "");
3706 } else if (inst
->Memory
.Texture
== TGSI_TEXTURE_BUFFER
) {
3707 out
= get_buffer_size(bld_base
, emit_data
->args
[0]);
3709 out
= lp_build_intrinsic(
3710 builder
, "llvm.SI.getresinfo.i32", emit_data
->dst_type
,
3711 emit_data
->args
, emit_data
->arg_count
,
3712 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
3714 /* Divide the number of layers by 6 to get the number of cubes. */
3715 if (inst
->Memory
.Texture
== TGSI_TEXTURE_CUBE_ARRAY
) {
3716 LLVMValueRef imm2
= lp_build_const_int32(gallivm
, 2);
3717 LLVMValueRef imm6
= lp_build_const_int32(gallivm
, 6);
3719 LLVMValueRef z
= LLVMBuildExtractElement(builder
, out
, imm2
, "");
3720 z
= LLVMBuildBitCast(builder
, z
, bld_base
->uint_bld
.elem_type
, "");
3721 z
= LLVMBuildSDiv(builder
, z
, imm6
, "");
3722 z
= LLVMBuildBitCast(builder
, z
, bld_base
->base
.elem_type
, "");
3723 out
= LLVMBuildInsertElement(builder
, out
, z
, imm2
, "");
3727 emit_data
->output
[emit_data
->chan
] = out
;
3730 static void set_tex_fetch_args(struct si_shader_context
*ctx
,
3731 struct lp_build_emit_data
*emit_data
,
3732 unsigned opcode
, unsigned target
,
3733 LLVMValueRef res_ptr
, LLVMValueRef samp_ptr
,
3734 LLVMValueRef
*param
, unsigned count
,
3737 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
3739 unsigned is_rect
= target
== TGSI_TEXTURE_RECT
;
3741 /* Pad to power of two vector */
3742 while (count
< util_next_power_of_two(count
))
3743 param
[count
++] = LLVMGetUndef(ctx
->i32
);
3745 /* Texture coordinates. */
3747 emit_data
->args
[0] = lp_build_gather_values(gallivm
, param
, count
);
3749 emit_data
->args
[0] = param
[0];
3752 emit_data
->args
[1] = res_ptr
;
3755 if (opcode
== TGSI_OPCODE_TXF
|| opcode
== TGSI_OPCODE_TXQ
)
3756 emit_data
->dst_type
= ctx
->v4i32
;
3758 emit_data
->dst_type
= ctx
->v4f32
;
3760 emit_data
->args
[num_args
++] = samp_ptr
;
3763 emit_data
->args
[num_args
++] = lp_build_const_int32(gallivm
, dmask
);
3764 emit_data
->args
[num_args
++] = lp_build_const_int32(gallivm
, is_rect
); /* unorm */
3765 emit_data
->args
[num_args
++] = lp_build_const_int32(gallivm
, 0); /* r128 */
3766 emit_data
->args
[num_args
++] = lp_build_const_int32(gallivm
,
3767 tgsi_is_array_sampler(target
)); /* da */
3768 emit_data
->args
[num_args
++] = lp_build_const_int32(gallivm
, 0); /* glc */
3769 emit_data
->args
[num_args
++] = lp_build_const_int32(gallivm
, 0); /* slc */
3770 emit_data
->args
[num_args
++] = lp_build_const_int32(gallivm
, 0); /* tfe */
3771 emit_data
->args
[num_args
++] = lp_build_const_int32(gallivm
, 0); /* lwe */
3773 emit_data
->arg_count
= num_args
;
3776 static const struct lp_build_tgsi_action tex_action
;
3784 static LLVMTypeRef
const_array(LLVMTypeRef elem_type
, int num_elements
)
3786 return LLVMPointerType(LLVMArrayType(elem_type
, num_elements
),
3791 * Load an image view, fmask view. or sampler state descriptor.
3793 static LLVMValueRef
get_sampler_desc_custom(struct si_shader_context
*ctx
,
3794 LLVMValueRef list
, LLVMValueRef index
,
3795 enum desc_type type
)
3797 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
3798 LLVMBuilderRef builder
= gallivm
->builder
;
3802 /* The image is at [0:7]. */
3803 index
= LLVMBuildMul(builder
, index
, LLVMConstInt(ctx
->i32
, 2, 0), "");
3806 /* The FMASK is at [8:15]. */
3807 index
= LLVMBuildMul(builder
, index
, LLVMConstInt(ctx
->i32
, 2, 0), "");
3808 index
= LLVMBuildAdd(builder
, index
, LLVMConstInt(ctx
->i32
, 1, 0), "");
3811 /* The sampler state is at [12:15]. */
3812 index
= LLVMBuildMul(builder
, index
, LLVMConstInt(ctx
->i32
, 4, 0), "");
3813 index
= LLVMBuildAdd(builder
, index
, LLVMConstInt(ctx
->i32
, 3, 0), "");
3814 list
= LLVMBuildPointerCast(builder
, list
,
3815 const_array(ctx
->v4i32
, 0), "");
3819 return build_indexed_load_const(ctx
, list
, index
);
3822 static LLVMValueRef
get_sampler_desc(struct si_shader_context
*ctx
,
3823 LLVMValueRef index
, enum desc_type type
)
3825 LLVMValueRef list
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
3828 return get_sampler_desc_custom(ctx
, list
, index
, type
);
3831 /* Disable anisotropic filtering if BASE_LEVEL == LAST_LEVEL.
3834 * If BASE_LEVEL == LAST_LEVEL, the shader must disable anisotropic
3835 * filtering manually. The driver sets img7 to a mask clearing
3836 * MAX_ANISO_RATIO if BASE_LEVEL == LAST_LEVEL. The shader must do:
3837 * s_and_b32 samp0, samp0, img7
3840 * The ANISO_OVERRIDE sampler field enables this fix in TA.
3842 static LLVMValueRef
sici_fix_sampler_aniso(struct si_shader_context
*ctx
,
3843 LLVMValueRef res
, LLVMValueRef samp
)
3845 LLVMBuilderRef builder
= ctx
->radeon_bld
.gallivm
.builder
;
3846 LLVMValueRef img7
, samp0
;
3848 if (ctx
->screen
->b
.chip_class
>= VI
)
3851 img7
= LLVMBuildExtractElement(builder
, res
,
3852 LLVMConstInt(ctx
->i32
, 7, 0), "");
3853 samp0
= LLVMBuildExtractElement(builder
, samp
,
3854 LLVMConstInt(ctx
->i32
, 0, 0), "");
3855 samp0
= LLVMBuildAnd(builder
, samp0
, img7
, "");
3856 return LLVMBuildInsertElement(builder
, samp
, samp0
,
3857 LLVMConstInt(ctx
->i32
, 0, 0), "");
3860 static void tex_fetch_ptrs(
3861 struct lp_build_tgsi_context
*bld_base
,
3862 struct lp_build_emit_data
*emit_data
,
3863 LLVMValueRef
*res_ptr
, LLVMValueRef
*samp_ptr
, LLVMValueRef
*fmask_ptr
)
3865 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3866 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3867 unsigned target
= inst
->Texture
.Texture
;
3868 unsigned sampler_src
;
3869 unsigned sampler_index
;
3871 sampler_src
= emit_data
->inst
->Instruction
.NumSrcRegs
- 1;
3872 sampler_index
= emit_data
->inst
->Src
[sampler_src
].Register
.Index
;
3874 if (emit_data
->inst
->Src
[sampler_src
].Register
.Indirect
) {
3875 const struct tgsi_full_src_register
*reg
= &emit_data
->inst
->Src
[sampler_src
];
3876 LLVMValueRef ind_index
;
3878 ind_index
= get_bounded_indirect_index(ctx
,
3880 reg
->Register
.Index
,
3883 *res_ptr
= get_sampler_desc(ctx
, ind_index
, DESC_IMAGE
);
3885 if (target
== TGSI_TEXTURE_2D_MSAA
||
3886 target
== TGSI_TEXTURE_2D_ARRAY_MSAA
) {
3888 *fmask_ptr
= get_sampler_desc(ctx
, ind_index
, DESC_FMASK
);
3890 *samp_ptr
= get_sampler_desc(ctx
, ind_index
, DESC_SAMPLER
);
3891 *samp_ptr
= sici_fix_sampler_aniso(ctx
, *res_ptr
, *samp_ptr
);
3895 *res_ptr
= ctx
->sampler_views
[sampler_index
];
3896 *samp_ptr
= ctx
->sampler_states
[sampler_index
];
3897 *fmask_ptr
= ctx
->fmasks
[sampler_index
];
3901 static void tex_fetch_args(
3902 struct lp_build_tgsi_context
*bld_base
,
3903 struct lp_build_emit_data
*emit_data
)
3905 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3906 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3907 LLVMBuilderRef builder
= gallivm
->builder
;
3908 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3909 unsigned opcode
= inst
->Instruction
.Opcode
;
3910 unsigned target
= inst
->Texture
.Texture
;
3911 LLVMValueRef coords
[5], derivs
[6];
3912 LLVMValueRef address
[16];
3913 unsigned num_coords
= tgsi_util_get_texture_coord_dim(target
);
3914 int ref_pos
= tgsi_util_get_shadow_ref_src_index(target
);
3917 unsigned num_deriv_channels
= 0;
3918 bool has_offset
= inst
->Texture
.NumOffsets
> 0;
3919 LLVMValueRef res_ptr
, samp_ptr
, fmask_ptr
= NULL
;
3920 unsigned dmask
= 0xf;
3922 tex_fetch_ptrs(bld_base
, emit_data
, &res_ptr
, &samp_ptr
, &fmask_ptr
);
3924 if (opcode
== TGSI_OPCODE_TXQ
) {
3925 if (target
== TGSI_TEXTURE_BUFFER
) {
3926 /* Read the size from the buffer descriptor directly. */
3927 LLVMValueRef res
= LLVMBuildBitCast(builder
, res_ptr
, ctx
->v8i32
, "");
3928 emit_data
->args
[0] = get_buffer_size(bld_base
, res
);
3932 /* Textures - set the mip level. */
3933 address
[count
++] = lp_build_emit_fetch(bld_base
, inst
, 0, TGSI_CHAN_X
);
3935 set_tex_fetch_args(ctx
, emit_data
, opcode
, target
, res_ptr
,
3936 NULL
, address
, count
, 0xf);
3940 if (target
== TGSI_TEXTURE_BUFFER
) {
3941 LLVMTypeRef v2i128
= LLVMVectorType(ctx
->i128
, 2);
3943 /* Bitcast and truncate v8i32 to v16i8. */
3944 LLVMValueRef res
= res_ptr
;
3945 res
= LLVMBuildBitCast(gallivm
->builder
, res
, v2i128
, "");
3946 res
= LLVMBuildExtractElement(gallivm
->builder
, res
, bld_base
->uint_bld
.one
, "");
3947 res
= LLVMBuildBitCast(gallivm
->builder
, res
, ctx
->v16i8
, "");
3949 emit_data
->dst_type
= ctx
->v4f32
;
3950 emit_data
->args
[0] = res
;
3951 emit_data
->args
[1] = bld_base
->uint_bld
.zero
;
3952 emit_data
->args
[2] = lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, TGSI_CHAN_X
);
3953 emit_data
->arg_count
= 3;
3957 /* Fetch and project texture coordinates */
3958 coords
[3] = lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, TGSI_CHAN_W
);
3959 for (chan
= 0; chan
< 3; chan
++ ) {
3960 coords
[chan
] = lp_build_emit_fetch(bld_base
,
3963 if (opcode
== TGSI_OPCODE_TXP
)
3964 coords
[chan
] = lp_build_emit_llvm_binary(bld_base
,
3970 if (opcode
== TGSI_OPCODE_TXP
)
3971 coords
[3] = bld_base
->base
.one
;
3974 if (has_offset
&& opcode
!= TGSI_OPCODE_TXF
) {
3975 /* The offsets are six-bit signed integers packed like this:
3976 * X=[5:0], Y=[13:8], and Z=[21:16].
3978 LLVMValueRef offset
[3], pack
;
3980 assert(inst
->Texture
.NumOffsets
== 1);
3982 for (chan
= 0; chan
< 3; chan
++) {
3983 offset
[chan
] = lp_build_emit_fetch_texoffset(bld_base
,
3984 emit_data
->inst
, 0, chan
);
3985 offset
[chan
] = LLVMBuildAnd(gallivm
->builder
, offset
[chan
],
3986 lp_build_const_int32(gallivm
, 0x3f), "");
3988 offset
[chan
] = LLVMBuildShl(gallivm
->builder
, offset
[chan
],
3989 lp_build_const_int32(gallivm
, chan
*8), "");
3992 pack
= LLVMBuildOr(gallivm
->builder
, offset
[0], offset
[1], "");
3993 pack
= LLVMBuildOr(gallivm
->builder
, pack
, offset
[2], "");
3994 address
[count
++] = pack
;
3997 /* Pack LOD bias value */
3998 if (opcode
== TGSI_OPCODE_TXB
)
3999 address
[count
++] = coords
[3];
4000 if (opcode
== TGSI_OPCODE_TXB2
)
4001 address
[count
++] = lp_build_emit_fetch(bld_base
, inst
, 1, TGSI_CHAN_X
);
4003 /* Pack depth comparison value */
4004 if (tgsi_is_shadow_target(target
) && opcode
!= TGSI_OPCODE_LODQ
) {
4005 if (target
== TGSI_TEXTURE_SHADOWCUBE_ARRAY
) {
4006 address
[count
++] = lp_build_emit_fetch(bld_base
, inst
, 1, TGSI_CHAN_X
);
4008 assert(ref_pos
>= 0);
4009 address
[count
++] = coords
[ref_pos
];
4013 /* Pack user derivatives */
4014 if (opcode
== TGSI_OPCODE_TXD
) {
4015 int param
, num_src_deriv_channels
;
4018 case TGSI_TEXTURE_3D
:
4019 num_src_deriv_channels
= 3;
4020 num_deriv_channels
= 3;
4022 case TGSI_TEXTURE_2D
:
4023 case TGSI_TEXTURE_SHADOW2D
:
4024 case TGSI_TEXTURE_RECT
:
4025 case TGSI_TEXTURE_SHADOWRECT
:
4026 case TGSI_TEXTURE_2D_ARRAY
:
4027 case TGSI_TEXTURE_SHADOW2D_ARRAY
:
4028 num_src_deriv_channels
= 2;
4029 num_deriv_channels
= 2;
4031 case TGSI_TEXTURE_CUBE
:
4032 case TGSI_TEXTURE_SHADOWCUBE
:
4033 case TGSI_TEXTURE_CUBE_ARRAY
:
4034 case TGSI_TEXTURE_SHADOWCUBE_ARRAY
:
4035 /* Cube derivatives will be converted to 2D. */
4036 num_src_deriv_channels
= 3;
4037 num_deriv_channels
= 2;
4039 case TGSI_TEXTURE_1D
:
4040 case TGSI_TEXTURE_SHADOW1D
:
4041 case TGSI_TEXTURE_1D_ARRAY
:
4042 case TGSI_TEXTURE_SHADOW1D_ARRAY
:
4043 num_src_deriv_channels
= 1;
4044 num_deriv_channels
= 1;
4047 unreachable("invalid target");
4050 for (param
= 0; param
< 2; param
++)
4051 for (chan
= 0; chan
< num_src_deriv_channels
; chan
++)
4052 derivs
[param
* num_src_deriv_channels
+ chan
] =
4053 lp_build_emit_fetch(bld_base
, inst
, param
+1, chan
);
4056 if (target
== TGSI_TEXTURE_CUBE
||
4057 target
== TGSI_TEXTURE_CUBE_ARRAY
||
4058 target
== TGSI_TEXTURE_SHADOWCUBE
||
4059 target
== TGSI_TEXTURE_SHADOWCUBE_ARRAY
)
4060 radeon_llvm_emit_prepare_cube_coords(bld_base
, emit_data
, coords
, derivs
);
4062 if (opcode
== TGSI_OPCODE_TXD
)
4063 for (int i
= 0; i
< num_deriv_channels
* 2; i
++)
4064 address
[count
++] = derivs
[i
];
4066 /* Pack texture coordinates */
4067 address
[count
++] = coords
[0];
4069 address
[count
++] = coords
[1];
4071 address
[count
++] = coords
[2];
4073 /* Pack LOD or sample index */
4074 if (opcode
== TGSI_OPCODE_TXL
|| opcode
== TGSI_OPCODE_TXF
)
4075 address
[count
++] = coords
[3];
4076 else if (opcode
== TGSI_OPCODE_TXL2
)
4077 address
[count
++] = lp_build_emit_fetch(bld_base
, inst
, 1, TGSI_CHAN_X
);
4080 assert(!"Cannot handle more than 16 texture address parameters");
4084 for (chan
= 0; chan
< count
; chan
++ ) {
4085 address
[chan
] = LLVMBuildBitCast(gallivm
->builder
,
4086 address
[chan
], ctx
->i32
, "");
4089 /* Adjust the sample index according to FMASK.
4091 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
4092 * which is the identity mapping. Each nibble says which physical sample
4093 * should be fetched to get that sample.
4095 * For example, 0x11111100 means there are only 2 samples stored and
4096 * the second sample covers 3/4 of the pixel. When reading samples 0
4097 * and 1, return physical sample 0 (determined by the first two 0s
4098 * in FMASK), otherwise return physical sample 1.
4100 * The sample index should be adjusted as follows:
4101 * sample_index = (fmask >> (sample_index * 4)) & 0xF;
4103 if (target
== TGSI_TEXTURE_2D_MSAA
||
4104 target
== TGSI_TEXTURE_2D_ARRAY_MSAA
) {
4105 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
4106 struct lp_build_emit_data txf_emit_data
= *emit_data
;
4107 LLVMValueRef txf_address
[4];
4108 unsigned txf_count
= count
;
4109 struct tgsi_full_instruction inst
= {};
4111 memcpy(txf_address
, address
, sizeof(txf_address
));
4113 if (target
== TGSI_TEXTURE_2D_MSAA
) {
4114 txf_address
[2] = bld_base
->uint_bld
.zero
;
4116 txf_address
[3] = bld_base
->uint_bld
.zero
;
4118 /* Read FMASK using TXF. */
4119 inst
.Instruction
.Opcode
= TGSI_OPCODE_TXF
;
4120 inst
.Texture
.Texture
= target
;
4121 txf_emit_data
.inst
= &inst
;
4122 txf_emit_data
.chan
= 0;
4123 set_tex_fetch_args(ctx
, &txf_emit_data
, TGSI_OPCODE_TXF
,
4124 target
, fmask_ptr
, NULL
,
4125 txf_address
, txf_count
, 0xf);
4126 build_tex_intrinsic(&tex_action
, bld_base
, &txf_emit_data
);
4128 /* Initialize some constants. */
4129 LLVMValueRef four
= LLVMConstInt(ctx
->i32
, 4, 0);
4130 LLVMValueRef F
= LLVMConstInt(ctx
->i32
, 0xF, 0);
4132 /* Apply the formula. */
4133 LLVMValueRef fmask
=
4134 LLVMBuildExtractElement(gallivm
->builder
,
4135 txf_emit_data
.output
[0],
4136 uint_bld
->zero
, "");
4138 unsigned sample_chan
= target
== TGSI_TEXTURE_2D_MSAA
? 2 : 3;
4140 LLVMValueRef sample_index4
=
4141 LLVMBuildMul(gallivm
->builder
, address
[sample_chan
], four
, "");
4143 LLVMValueRef shifted_fmask
=
4144 LLVMBuildLShr(gallivm
->builder
, fmask
, sample_index4
, "");
4146 LLVMValueRef final_sample
=
4147 LLVMBuildAnd(gallivm
->builder
, shifted_fmask
, F
, "");
4149 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
4150 * resource descriptor is 0 (invalid),
4152 LLVMValueRef fmask_desc
=
4153 LLVMBuildBitCast(gallivm
->builder
, fmask_ptr
,
4156 LLVMValueRef fmask_word1
=
4157 LLVMBuildExtractElement(gallivm
->builder
, fmask_desc
,
4160 LLVMValueRef word1_is_nonzero
=
4161 LLVMBuildICmp(gallivm
->builder
, LLVMIntNE
,
4162 fmask_word1
, uint_bld
->zero
, "");
4164 /* Replace the MSAA sample index. */
4165 address
[sample_chan
] =
4166 LLVMBuildSelect(gallivm
->builder
, word1_is_nonzero
,
4167 final_sample
, address
[sample_chan
], "");
4170 if (opcode
== TGSI_OPCODE_TXF
) {
4171 /* add tex offsets */
4172 if (inst
->Texture
.NumOffsets
) {
4173 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
4174 struct lp_build_tgsi_soa_context
*bld
= lp_soa_context(bld_base
);
4175 const struct tgsi_texture_offset
*off
= inst
->TexOffsets
;
4177 assert(inst
->Texture
.NumOffsets
== 1);
4180 case TGSI_TEXTURE_3D
:
4181 address
[2] = lp_build_add(uint_bld
, address
[2],
4182 bld
->immediates
[off
->Index
][off
->SwizzleZ
]);
4184 case TGSI_TEXTURE_2D
:
4185 case TGSI_TEXTURE_SHADOW2D
:
4186 case TGSI_TEXTURE_RECT
:
4187 case TGSI_TEXTURE_SHADOWRECT
:
4188 case TGSI_TEXTURE_2D_ARRAY
:
4189 case TGSI_TEXTURE_SHADOW2D_ARRAY
:
4191 lp_build_add(uint_bld
, address
[1],
4192 bld
->immediates
[off
->Index
][off
->SwizzleY
]);
4194 case TGSI_TEXTURE_1D
:
4195 case TGSI_TEXTURE_SHADOW1D
:
4196 case TGSI_TEXTURE_1D_ARRAY
:
4197 case TGSI_TEXTURE_SHADOW1D_ARRAY
:
4199 lp_build_add(uint_bld
, address
[0],
4200 bld
->immediates
[off
->Index
][off
->SwizzleX
]);
4202 /* texture offsets do not apply to other texture targets */
4207 if (opcode
== TGSI_OPCODE_TG4
) {
4208 unsigned gather_comp
= 0;
4210 /* DMASK was repurposed for GATHER4. 4 components are always
4211 * returned and DMASK works like a swizzle - it selects
4212 * the component to fetch. The only valid DMASK values are
4213 * 1=red, 2=green, 4=blue, 8=alpha. (e.g. 1 returns
4214 * (red,red,red,red) etc.) The ISA document doesn't mention
4218 /* Get the component index from src1.x for Gather4. */
4219 if (!tgsi_is_shadow_target(target
)) {
4220 LLVMValueRef (*imms
)[4] = lp_soa_context(bld_base
)->immediates
;
4221 LLVMValueRef comp_imm
;
4222 struct tgsi_src_register src1
= inst
->Src
[1].Register
;
4224 assert(src1
.File
== TGSI_FILE_IMMEDIATE
);
4226 comp_imm
= imms
[src1
.Index
][src1
.SwizzleX
];
4227 gather_comp
= LLVMConstIntGetZExtValue(comp_imm
);
4228 gather_comp
= CLAMP(gather_comp
, 0, 3);
4231 dmask
= 1 << gather_comp
;
4234 set_tex_fetch_args(ctx
, emit_data
, opcode
, target
, res_ptr
,
4235 samp_ptr
, address
, count
, dmask
);
4238 static void build_tex_intrinsic(const struct lp_build_tgsi_action
*action
,
4239 struct lp_build_tgsi_context
*bld_base
,
4240 struct lp_build_emit_data
*emit_data
)
4242 struct lp_build_context
*base
= &bld_base
->base
;
4243 unsigned opcode
= emit_data
->inst
->Instruction
.Opcode
;
4244 unsigned target
= emit_data
->inst
->Texture
.Texture
;
4245 char intr_name
[127];
4246 bool has_offset
= emit_data
->inst
->Texture
.NumOffsets
> 0;
4247 bool is_shadow
= tgsi_is_shadow_target(target
);
4249 const char *name
= "llvm.SI.image.sample";
4250 const char *infix
= "";
4252 if (opcode
== TGSI_OPCODE_TXQ
&& target
== TGSI_TEXTURE_BUFFER
) {
4253 /* Just return the buffer size. */
4254 emit_data
->output
[emit_data
->chan
] = emit_data
->args
[0];
4258 if (target
== TGSI_TEXTURE_BUFFER
) {
4259 emit_data
->output
[emit_data
->chan
] = lp_build_intrinsic(
4260 base
->gallivm
->builder
,
4261 "llvm.SI.vs.load.input", emit_data
->dst_type
,
4262 emit_data
->args
, emit_data
->arg_count
,
4263 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
4268 case TGSI_OPCODE_TXF
:
4269 name
= target
== TGSI_TEXTURE_2D_MSAA
||
4270 target
== TGSI_TEXTURE_2D_ARRAY_MSAA
?
4271 "llvm.SI.image.load" :
4272 "llvm.SI.image.load.mip";
4276 case TGSI_OPCODE_TXQ
:
4277 name
= "llvm.SI.getresinfo";
4281 case TGSI_OPCODE_LODQ
:
4282 name
= "llvm.SI.getlod";
4286 case TGSI_OPCODE_TEX
:
4287 case TGSI_OPCODE_TEX2
:
4288 case TGSI_OPCODE_TXP
:
4290 case TGSI_OPCODE_TXB
:
4291 case TGSI_OPCODE_TXB2
:
4294 case TGSI_OPCODE_TXL
:
4295 case TGSI_OPCODE_TXL2
:
4298 case TGSI_OPCODE_TXD
:
4301 case TGSI_OPCODE_TG4
:
4302 name
= "llvm.SI.gather4";
4309 /* Add the type and suffixes .c, .o if needed. */
4310 build_int_type_name(LLVMTypeOf(emit_data
->args
[0]), type
, sizeof(type
));
4311 sprintf(intr_name
, "%s%s%s%s.%s",
4312 name
, is_shadow
? ".c" : "", infix
,
4313 has_offset
? ".o" : "", type
);
4315 emit_data
->output
[emit_data
->chan
] = lp_build_intrinsic(
4316 base
->gallivm
->builder
, intr_name
, emit_data
->dst_type
,
4317 emit_data
->args
, emit_data
->arg_count
,
4318 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
4320 /* Divide the number of layers by 6 to get the number of cubes. */
4321 if (opcode
== TGSI_OPCODE_TXQ
&&
4322 (target
== TGSI_TEXTURE_CUBE_ARRAY
||
4323 target
== TGSI_TEXTURE_SHADOWCUBE_ARRAY
)) {
4324 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
4325 LLVMValueRef two
= lp_build_const_int32(bld_base
->base
.gallivm
, 2);
4326 LLVMValueRef six
= lp_build_const_int32(bld_base
->base
.gallivm
, 6);
4328 LLVMValueRef v4
= emit_data
->output
[emit_data
->chan
];
4329 LLVMValueRef z
= LLVMBuildExtractElement(builder
, v4
, two
, "");
4330 z
= LLVMBuildSDiv(builder
, z
, six
, "");
4332 emit_data
->output
[emit_data
->chan
] =
4333 LLVMBuildInsertElement(builder
, v4
, z
, two
, "");
4337 static void si_llvm_emit_txqs(
4338 const struct lp_build_tgsi_action
*action
,
4339 struct lp_build_tgsi_context
*bld_base
,
4340 struct lp_build_emit_data
*emit_data
)
4342 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4343 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4344 LLVMBuilderRef builder
= gallivm
->builder
;
4345 LLVMValueRef res
, samples
;
4346 LLVMValueRef res_ptr
, samp_ptr
, fmask_ptr
= NULL
;
4348 tex_fetch_ptrs(bld_base
, emit_data
, &res_ptr
, &samp_ptr
, &fmask_ptr
);
4351 /* Read the samples from the descriptor directly. */
4352 res
= LLVMBuildBitCast(builder
, res_ptr
, ctx
->v8i32
, "");
4353 samples
= LLVMBuildExtractElement(
4355 lp_build_const_int32(gallivm
, 3), "");
4356 samples
= LLVMBuildLShr(builder
, samples
,
4357 lp_build_const_int32(gallivm
, 16), "");
4358 samples
= LLVMBuildAnd(builder
, samples
,
4359 lp_build_const_int32(gallivm
, 0xf), "");
4360 samples
= LLVMBuildShl(builder
, lp_build_const_int32(gallivm
, 1),
4363 emit_data
->output
[emit_data
->chan
] = samples
;
4367 * SI implements derivatives using the local data store (LDS)
4368 * All writes to the LDS happen in all executing threads at
4369 * the same time. TID is the Thread ID for the current
4370 * thread and is a value between 0 and 63, representing
4371 * the thread's position in the wavefront.
4373 * For the pixel shader threads are grouped into quads of four pixels.
4374 * The TIDs of the pixels of a quad are:
4382 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
4383 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
4384 * the current pixel's column, and masking with 0xfffffffe yields the TID
4385 * of the left pixel of the current pixel's row.
4387 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
4388 * adding 2 yields the TID of the pixel below the top pixel.
4390 /* masks for thread ID. */
4391 #define TID_MASK_TOP_LEFT 0xfffffffc
4392 #define TID_MASK_TOP 0xfffffffd
4393 #define TID_MASK_LEFT 0xfffffffe
4395 static void si_llvm_emit_ddxy(
4396 const struct lp_build_tgsi_action
*action
,
4397 struct lp_build_tgsi_context
*bld_base
,
4398 struct lp_build_emit_data
*emit_data
)
4400 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4401 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4402 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
4403 unsigned opcode
= inst
->Instruction
.Opcode
;
4404 LLVMValueRef indices
[2];
4405 LLVMValueRef store_ptr
, load_ptr0
, load_ptr1
;
4406 LLVMValueRef tl
, trbl
, result
[4];
4407 LLVMValueRef tl_tid
, trbl_tid
;
4408 unsigned swizzle
[4];
4413 indices
[0] = bld_base
->uint_bld
.zero
;
4414 indices
[1] = get_thread_id(ctx
);
4415 store_ptr
= LLVMBuildGEP(gallivm
->builder
, ctx
->lds
,
4418 if (opcode
== TGSI_OPCODE_DDX_FINE
)
4419 mask
= TID_MASK_LEFT
;
4420 else if (opcode
== TGSI_OPCODE_DDY_FINE
)
4421 mask
= TID_MASK_TOP
;
4423 mask
= TID_MASK_TOP_LEFT
;
4425 tl_tid
= LLVMBuildAnd(gallivm
->builder
, indices
[1],
4426 lp_build_const_int32(gallivm
, mask
), "");
4427 indices
[1] = tl_tid
;
4428 load_ptr0
= LLVMBuildGEP(gallivm
->builder
, ctx
->lds
,
4431 /* for DDX we want to next X pixel, DDY next Y pixel. */
4432 idx
= (opcode
== TGSI_OPCODE_DDX
|| opcode
== TGSI_OPCODE_DDX_FINE
) ? 1 : 2;
4433 trbl_tid
= LLVMBuildAdd(gallivm
->builder
, indices
[1],
4434 lp_build_const_int32(gallivm
, idx
), "");
4435 indices
[1] = trbl_tid
;
4436 load_ptr1
= LLVMBuildGEP(gallivm
->builder
, ctx
->lds
,
4439 for (c
= 0; c
< 4; ++c
) {
4442 LLVMValueRef args
[2];
4444 swizzle
[c
] = tgsi_util_get_full_src_register_swizzle(&inst
->Src
[0], c
);
4445 for (i
= 0; i
< c
; ++i
) {
4446 if (swizzle
[i
] == swizzle
[c
]) {
4447 result
[c
] = result
[i
];
4454 val
= LLVMBuildBitCast(gallivm
->builder
,
4455 lp_build_emit_fetch(bld_base
, inst
, 0, c
),
4458 if ((HAVE_LLVM
>= 0x0309) && ctx
->screen
->b
.family
>= CHIP_TONGA
) {
4460 args
[0] = LLVMBuildMul(gallivm
->builder
, tl_tid
,
4461 lp_build_const_int32(gallivm
, 4), "");
4463 tl
= lp_build_intrinsic(gallivm
->builder
,
4464 "llvm.amdgcn.ds.bpermute", ctx
->i32
,
4465 args
, 2, LLVMReadNoneAttribute
);
4467 args
[0] = LLVMBuildMul(gallivm
->builder
, trbl_tid
,
4468 lp_build_const_int32(gallivm
, 4), "");
4469 trbl
= lp_build_intrinsic(gallivm
->builder
,
4470 "llvm.amdgcn.ds.bpermute", ctx
->i32
,
4471 args
, 2, LLVMReadNoneAttribute
);
4473 LLVMBuildStore(gallivm
->builder
, val
, store_ptr
);
4474 tl
= LLVMBuildLoad(gallivm
->builder
, load_ptr0
, "");
4475 trbl
= LLVMBuildLoad(gallivm
->builder
, load_ptr1
, "");
4477 tl
= LLVMBuildBitCast(gallivm
->builder
, tl
, ctx
->f32
, "");
4478 trbl
= LLVMBuildBitCast(gallivm
->builder
, trbl
, ctx
->f32
, "");
4479 result
[c
] = LLVMBuildFSub(gallivm
->builder
, trbl
, tl
, "");
4482 emit_data
->output
[0] = lp_build_gather_values(gallivm
, result
, 4);
4486 * this takes an I,J coordinate pair,
4487 * and works out the X and Y derivatives.
4488 * it returns DDX(I), DDX(J), DDY(I), DDY(J).
4490 static LLVMValueRef
si_llvm_emit_ddxy_interp(
4491 struct lp_build_tgsi_context
*bld_base
,
4492 LLVMValueRef interp_ij
)
4494 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4495 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4496 LLVMValueRef indices
[2];
4497 LLVMValueRef store_ptr
, load_ptr_x
, load_ptr_y
, load_ptr_ddx
, load_ptr_ddy
, temp
, temp2
;
4498 LLVMValueRef tl
, tr
, bl
, result
[4];
4501 indices
[0] = bld_base
->uint_bld
.zero
;
4502 indices
[1] = get_thread_id(ctx
);
4503 store_ptr
= LLVMBuildGEP(gallivm
->builder
, ctx
->lds
,
4506 temp
= LLVMBuildAnd(gallivm
->builder
, indices
[1],
4507 lp_build_const_int32(gallivm
, TID_MASK_LEFT
), "");
4509 temp2
= LLVMBuildAnd(gallivm
->builder
, indices
[1],
4510 lp_build_const_int32(gallivm
, TID_MASK_TOP
), "");
4513 load_ptr_x
= LLVMBuildGEP(gallivm
->builder
, ctx
->lds
,
4517 load_ptr_y
= LLVMBuildGEP(gallivm
->builder
, ctx
->lds
,
4520 indices
[1] = LLVMBuildAdd(gallivm
->builder
, temp
,
4521 lp_build_const_int32(gallivm
, 1), "");
4522 load_ptr_ddx
= LLVMBuildGEP(gallivm
->builder
, ctx
->lds
,
4525 indices
[1] = LLVMBuildAdd(gallivm
->builder
, temp2
,
4526 lp_build_const_int32(gallivm
, 2), "");
4527 load_ptr_ddy
= LLVMBuildGEP(gallivm
->builder
, ctx
->lds
,
4530 for (c
= 0; c
< 2; ++c
) {
4531 LLVMValueRef store_val
;
4532 LLVMValueRef c_ll
= lp_build_const_int32(gallivm
, c
);
4534 store_val
= LLVMBuildExtractElement(gallivm
->builder
,
4535 interp_ij
, c_ll
, "");
4536 LLVMBuildStore(gallivm
->builder
,
4540 tl
= LLVMBuildLoad(gallivm
->builder
, load_ptr_x
, "");
4541 tl
= LLVMBuildBitCast(gallivm
->builder
, tl
, ctx
->f32
, "");
4543 tr
= LLVMBuildLoad(gallivm
->builder
, load_ptr_ddx
, "");
4544 tr
= LLVMBuildBitCast(gallivm
->builder
, tr
, ctx
->f32
, "");
4546 result
[c
] = LLVMBuildFSub(gallivm
->builder
, tr
, tl
, "");
4548 tl
= LLVMBuildLoad(gallivm
->builder
, load_ptr_y
, "");
4549 tl
= LLVMBuildBitCast(gallivm
->builder
, tl
, ctx
->f32
, "");
4551 bl
= LLVMBuildLoad(gallivm
->builder
, load_ptr_ddy
, "");
4552 bl
= LLVMBuildBitCast(gallivm
->builder
, bl
, ctx
->f32
, "");
4554 result
[c
+ 2] = LLVMBuildFSub(gallivm
->builder
, bl
, tl
, "");
4557 return lp_build_gather_values(gallivm
, result
, 4);
4560 static void interp_fetch_args(
4561 struct lp_build_tgsi_context
*bld_base
,
4562 struct lp_build_emit_data
*emit_data
)
4564 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4565 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4566 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
4568 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_OFFSET
) {
4569 /* offset is in second src, first two channels */
4570 emit_data
->args
[0] = lp_build_emit_fetch(bld_base
,
4573 emit_data
->args
[1] = lp_build_emit_fetch(bld_base
,
4576 emit_data
->arg_count
= 2;
4577 } else if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_SAMPLE
) {
4578 LLVMValueRef sample_position
;
4579 LLVMValueRef sample_id
;
4580 LLVMValueRef halfval
= lp_build_const_float(gallivm
, 0.5f
);
4582 /* fetch sample ID, then fetch its sample position,
4583 * and place into first two channels.
4585 sample_id
= lp_build_emit_fetch(bld_base
,
4586 emit_data
->inst
, 1, TGSI_CHAN_X
);
4587 sample_id
= LLVMBuildBitCast(gallivm
->builder
, sample_id
,
4589 sample_position
= load_sample_position(&ctx
->radeon_bld
, sample_id
);
4591 emit_data
->args
[0] = LLVMBuildExtractElement(gallivm
->builder
,
4593 lp_build_const_int32(gallivm
, 0), "");
4595 emit_data
->args
[0] = LLVMBuildFSub(gallivm
->builder
, emit_data
->args
[0], halfval
, "");
4596 emit_data
->args
[1] = LLVMBuildExtractElement(gallivm
->builder
,
4598 lp_build_const_int32(gallivm
, 1), "");
4599 emit_data
->args
[1] = LLVMBuildFSub(gallivm
->builder
, emit_data
->args
[1], halfval
, "");
4600 emit_data
->arg_count
= 2;
4604 static void build_interp_intrinsic(const struct lp_build_tgsi_action
*action
,
4605 struct lp_build_tgsi_context
*bld_base
,
4606 struct lp_build_emit_data
*emit_data
)
4608 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4609 struct si_shader
*shader
= ctx
->shader
;
4610 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4611 LLVMValueRef interp_param
;
4612 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
4613 const char *intr_name
;
4614 int input_index
= inst
->Src
[0].Register
.Index
;
4617 LLVMValueRef attr_number
;
4618 LLVMValueRef params
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_PRIM_MASK
);
4619 int interp_param_idx
;
4620 unsigned interp
= shader
->selector
->info
.input_interpolate
[input_index
];
4623 assert(inst
->Src
[0].Register
.File
== TGSI_FILE_INPUT
);
4625 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_OFFSET
||
4626 inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_SAMPLE
)
4627 location
= TGSI_INTERPOLATE_LOC_CENTER
;
4629 location
= TGSI_INTERPOLATE_LOC_CENTROID
;
4631 interp_param_idx
= lookup_interp_param_index(interp
, location
);
4632 if (interp_param_idx
== -1)
4634 else if (interp_param_idx
)
4635 interp_param
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, interp_param_idx
);
4637 interp_param
= NULL
;
4639 attr_number
= lp_build_const_int32(gallivm
, input_index
);
4641 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_OFFSET
||
4642 inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_SAMPLE
) {
4643 LLVMValueRef ij_out
[2];
4644 LLVMValueRef ddxy_out
= si_llvm_emit_ddxy_interp(bld_base
, interp_param
);
4647 * take the I then J parameters, and the DDX/Y for it, and
4648 * calculate the IJ inputs for the interpolator.
4649 * temp1 = ddx * offset/sample.x + I;
4650 * interp_param.I = ddy * offset/sample.y + temp1;
4651 * temp1 = ddx * offset/sample.x + J;
4652 * interp_param.J = ddy * offset/sample.y + temp1;
4654 for (i
= 0; i
< 2; i
++) {
4655 LLVMValueRef ix_ll
= lp_build_const_int32(gallivm
, i
);
4656 LLVMValueRef iy_ll
= lp_build_const_int32(gallivm
, i
+ 2);
4657 LLVMValueRef ddx_el
= LLVMBuildExtractElement(gallivm
->builder
,
4658 ddxy_out
, ix_ll
, "");
4659 LLVMValueRef ddy_el
= LLVMBuildExtractElement(gallivm
->builder
,
4660 ddxy_out
, iy_ll
, "");
4661 LLVMValueRef interp_el
= LLVMBuildExtractElement(gallivm
->builder
,
4662 interp_param
, ix_ll
, "");
4663 LLVMValueRef temp1
, temp2
;
4665 interp_el
= LLVMBuildBitCast(gallivm
->builder
, interp_el
,
4668 temp1
= LLVMBuildFMul(gallivm
->builder
, ddx_el
, emit_data
->args
[0], "");
4670 temp1
= LLVMBuildFAdd(gallivm
->builder
, temp1
, interp_el
, "");
4672 temp2
= LLVMBuildFMul(gallivm
->builder
, ddy_el
, emit_data
->args
[1], "");
4674 temp2
= LLVMBuildFAdd(gallivm
->builder
, temp2
, temp1
, "");
4676 ij_out
[i
] = LLVMBuildBitCast(gallivm
->builder
,
4677 temp2
, ctx
->i32
, "");
4679 interp_param
= lp_build_gather_values(bld_base
->base
.gallivm
, ij_out
, 2);
4682 intr_name
= interp_param
? "llvm.SI.fs.interp" : "llvm.SI.fs.constant";
4683 for (chan
= 0; chan
< 2; chan
++) {
4684 LLVMValueRef args
[4];
4685 LLVMValueRef llvm_chan
;
4688 schan
= tgsi_util_get_full_src_register_swizzle(&inst
->Src
[0], chan
);
4689 llvm_chan
= lp_build_const_int32(gallivm
, schan
);
4691 args
[0] = llvm_chan
;
4692 args
[1] = attr_number
;
4694 args
[3] = interp_param
;
4696 emit_data
->output
[chan
] =
4697 lp_build_intrinsic(gallivm
->builder
, intr_name
,
4698 ctx
->f32
, args
, args
[3] ? 4 : 3,
4699 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
4703 static unsigned si_llvm_get_stream(struct lp_build_tgsi_context
*bld_base
,
4704 struct lp_build_emit_data
*emit_data
)
4706 LLVMValueRef (*imms
)[4] = lp_soa_context(bld_base
)->immediates
;
4707 struct tgsi_src_register src0
= emit_data
->inst
->Src
[0].Register
;
4710 assert(src0
.File
== TGSI_FILE_IMMEDIATE
);
4712 stream
= LLVMConstIntGetZExtValue(imms
[src0
.Index
][src0
.SwizzleX
]) & 0x3;
4716 /* Emit one vertex from the geometry shader */
4717 static void si_llvm_emit_vertex(
4718 const struct lp_build_tgsi_action
*action
,
4719 struct lp_build_tgsi_context
*bld_base
,
4720 struct lp_build_emit_data
*emit_data
)
4722 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4723 struct lp_build_context
*uint
= &bld_base
->uint_bld
;
4724 struct si_shader
*shader
= ctx
->shader
;
4725 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
4726 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4727 LLVMValueRef soffset
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
4728 SI_PARAM_GS2VS_OFFSET
);
4729 LLVMValueRef gs_next_vertex
;
4730 LLVMValueRef can_emit
, kill
;
4731 LLVMValueRef args
[2];
4736 stream
= si_llvm_get_stream(bld_base
, emit_data
);
4738 /* Write vertex attribute values to GSVS ring */
4739 gs_next_vertex
= LLVMBuildLoad(gallivm
->builder
,
4740 ctx
->gs_next_vertex
[stream
],
4743 /* If this thread has already emitted the declared maximum number of
4744 * vertices, kill it: excessive vertex emissions are not supposed to
4745 * have any effect, and GS threads have no externally observable
4746 * effects other than emitting vertices.
4748 can_emit
= LLVMBuildICmp(gallivm
->builder
, LLVMIntULE
, gs_next_vertex
,
4749 lp_build_const_int32(gallivm
,
4750 shader
->selector
->gs_max_out_vertices
), "");
4751 kill
= lp_build_select(&bld_base
->base
, can_emit
,
4752 lp_build_const_float(gallivm
, 1.0f
),
4753 lp_build_const_float(gallivm
, -1.0f
));
4755 lp_build_intrinsic(gallivm
->builder
, "llvm.AMDGPU.kill",
4756 ctx
->voidt
, &kill
, 1, 0);
4758 for (i
= 0; i
< info
->num_outputs
; i
++) {
4759 LLVMValueRef
*out_ptr
=
4760 ctx
->radeon_bld
.soa
.outputs
[i
];
4762 for (chan
= 0; chan
< 4; chan
++) {
4763 LLVMValueRef out_val
= LLVMBuildLoad(gallivm
->builder
, out_ptr
[chan
], "");
4764 LLVMValueRef voffset
=
4765 lp_build_const_int32(gallivm
, (i
* 4 + chan
) *
4766 shader
->selector
->gs_max_out_vertices
);
4768 voffset
= lp_build_add(uint
, voffset
, gs_next_vertex
);
4769 voffset
= lp_build_mul_imm(uint
, voffset
, 4);
4771 out_val
= LLVMBuildBitCast(gallivm
->builder
, out_val
, ctx
->i32
, "");
4773 build_tbuffer_store(ctx
,
4774 ctx
->gsvs_ring
[stream
],
4776 voffset
, soffset
, 0,
4777 V_008F0C_BUF_DATA_FORMAT_32
,
4778 V_008F0C_BUF_NUM_FORMAT_UINT
,
4782 gs_next_vertex
= lp_build_add(uint
, gs_next_vertex
,
4783 lp_build_const_int32(gallivm
, 1));
4785 LLVMBuildStore(gallivm
->builder
, gs_next_vertex
, ctx
->gs_next_vertex
[stream
]);
4787 /* Signal vertex emission */
4788 args
[0] = lp_build_const_int32(gallivm
, SENDMSG_GS_OP_EMIT
| SENDMSG_GS
| (stream
<< 8));
4789 args
[1] = LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_GS_WAVE_ID
);
4790 lp_build_intrinsic(gallivm
->builder
, "llvm.SI.sendmsg",
4791 ctx
->voidt
, args
, 2, LLVMNoUnwindAttribute
);
4794 /* Cut one primitive from the geometry shader */
4795 static void si_llvm_emit_primitive(
4796 const struct lp_build_tgsi_action
*action
,
4797 struct lp_build_tgsi_context
*bld_base
,
4798 struct lp_build_emit_data
*emit_data
)
4800 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4801 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4802 LLVMValueRef args
[2];
4805 /* Signal primitive cut */
4806 stream
= si_llvm_get_stream(bld_base
, emit_data
);
4807 args
[0] = lp_build_const_int32(gallivm
, SENDMSG_GS_OP_CUT
| SENDMSG_GS
| (stream
<< 8));
4808 args
[1] = LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_GS_WAVE_ID
);
4809 lp_build_intrinsic(gallivm
->builder
, "llvm.SI.sendmsg",
4810 ctx
->voidt
, args
, 2, LLVMNoUnwindAttribute
);
4813 static void si_llvm_emit_barrier(const struct lp_build_tgsi_action
*action
,
4814 struct lp_build_tgsi_context
*bld_base
,
4815 struct lp_build_emit_data
*emit_data
)
4817 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4818 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4820 /* The real barrier instruction isn’t needed, because an entire patch
4821 * always fits into a single wave.
4823 if (ctx
->type
== PIPE_SHADER_TESS_CTRL
) {
4824 emit_optimization_barrier(ctx
);
4828 lp_build_intrinsic(gallivm
->builder
,
4829 HAVE_LLVM
>= 0x0309 ? "llvm.amdgcn.s.barrier"
4830 : "llvm.AMDGPU.barrier.local",
4831 ctx
->voidt
, NULL
, 0, LLVMNoUnwindAttribute
);
4834 static const struct lp_build_tgsi_action tex_action
= {
4835 .fetch_args
= tex_fetch_args
,
4836 .emit
= build_tex_intrinsic
,
4839 static const struct lp_build_tgsi_action interp_action
= {
4840 .fetch_args
= interp_fetch_args
,
4841 .emit
= build_interp_intrinsic
,
4844 static void si_create_function(struct si_shader_context
*ctx
,
4845 LLVMTypeRef
*returns
, unsigned num_returns
,
4846 LLVMTypeRef
*params
, unsigned num_params
,
4847 int last_array_pointer
, int last_sgpr
)
4851 radeon_llvm_create_func(&ctx
->radeon_bld
, returns
, num_returns
,
4852 params
, num_params
);
4853 radeon_llvm_shader_type(ctx
->radeon_bld
.main_fn
, ctx
->type
);
4854 ctx
->return_value
= LLVMGetUndef(ctx
->radeon_bld
.return_type
);
4856 for (i
= 0; i
<= last_sgpr
; ++i
) {
4857 LLVMValueRef P
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, i
);
4859 /* We tell llvm that array inputs are passed by value to allow Sinking pass
4860 * to move load. Inputs are constant so this is fine. */
4861 if (i
<= last_array_pointer
)
4862 LLVMAddAttribute(P
, LLVMByValAttribute
);
4864 LLVMAddAttribute(P
, LLVMInRegAttribute
);
4868 static void create_meta_data(struct si_shader_context
*ctx
)
4870 struct gallivm_state
*gallivm
= ctx
->radeon_bld
.soa
.bld_base
.base
.gallivm
;
4871 LLVMValueRef args
[3];
4873 args
[0] = LLVMMDStringInContext(gallivm
->context
, "const", 5);
4875 args
[2] = lp_build_const_int32(gallivm
, 1);
4877 ctx
->const_md
= LLVMMDNodeInContext(gallivm
->context
, args
, 3);
4880 static void declare_streamout_params(struct si_shader_context
*ctx
,
4881 struct pipe_stream_output_info
*so
,
4882 LLVMTypeRef
*params
, LLVMTypeRef i32
,
4883 unsigned *num_params
)
4887 /* Streamout SGPRs. */
4888 if (so
->num_outputs
) {
4889 params
[ctx
->param_streamout_config
= (*num_params
)++] = i32
;
4890 params
[ctx
->param_streamout_write_index
= (*num_params
)++] = i32
;
4892 /* A streamout buffer offset is loaded if the stride is non-zero. */
4893 for (i
= 0; i
< 4; i
++) {
4897 params
[ctx
->param_streamout_offset
[i
] = (*num_params
)++] = i32
;
4901 static unsigned llvm_get_type_size(LLVMTypeRef type
)
4903 LLVMTypeKind kind
= LLVMGetTypeKind(type
);
4906 case LLVMIntegerTypeKind
:
4907 return LLVMGetIntTypeWidth(type
) / 8;
4908 case LLVMFloatTypeKind
:
4910 case LLVMPointerTypeKind
:
4912 case LLVMVectorTypeKind
:
4913 return LLVMGetVectorSize(type
) *
4914 llvm_get_type_size(LLVMGetElementType(type
));
4921 static void declare_tess_lds(struct si_shader_context
*ctx
)
4923 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
4924 LLVMTypeRef i32
= ctx
->radeon_bld
.soa
.bld_base
.uint_bld
.elem_type
;
4926 /* This is the upper bound, maximum is 32 inputs times 32 vertices */
4927 unsigned vertex_data_dw_size
= 32*32*4;
4928 unsigned patch_data_dw_size
= 32*4;
4929 /* The formula is: TCS inputs + TCS outputs + TCS patch outputs. */
4930 unsigned patch_dw_size
= vertex_data_dw_size
*2 + patch_data_dw_size
;
4931 unsigned lds_dwords
= patch_dw_size
;
4933 /* The actual size is computed outside of the shader to reduce
4934 * the number of shader variants. */
4936 LLVMAddGlobalInAddressSpace(gallivm
->module
,
4937 LLVMArrayType(i32
, lds_dwords
),
4942 static void create_function(struct si_shader_context
*ctx
)
4944 struct lp_build_tgsi_context
*bld_base
= &ctx
->radeon_bld
.soa
.bld_base
;
4945 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4946 struct si_shader
*shader
= ctx
->shader
;
4947 LLVMTypeRef params
[SI_NUM_PARAMS
+ SI_NUM_VERTEX_BUFFERS
], v3i32
;
4948 LLVMTypeRef returns
[16+32*4];
4949 unsigned i
, last_array_pointer
, last_sgpr
, num_params
, num_return_sgprs
;
4950 unsigned num_returns
= 0;
4952 v3i32
= LLVMVectorType(ctx
->i32
, 3);
4954 params
[SI_PARAM_RW_BUFFERS
] = const_array(ctx
->v16i8
, SI_NUM_RW_BUFFERS
);
4955 params
[SI_PARAM_CONST_BUFFERS
] = const_array(ctx
->v16i8
, SI_NUM_CONST_BUFFERS
);
4956 params
[SI_PARAM_SAMPLERS
] = const_array(ctx
->v8i32
, SI_NUM_SAMPLERS
);
4957 params
[SI_PARAM_IMAGES
] = const_array(ctx
->v8i32
, SI_NUM_IMAGES
);
4958 params
[SI_PARAM_SHADER_BUFFERS
] = const_array(ctx
->v4i32
, SI_NUM_SHADER_BUFFERS
);
4959 last_array_pointer
= SI_PARAM_SHADER_BUFFERS
;
4961 switch (ctx
->type
) {
4962 case PIPE_SHADER_VERTEX
:
4963 params
[SI_PARAM_VERTEX_BUFFERS
] = const_array(ctx
->v16i8
, SI_NUM_VERTEX_BUFFERS
);
4964 last_array_pointer
= SI_PARAM_VERTEX_BUFFERS
;
4965 params
[SI_PARAM_BASE_VERTEX
] = ctx
->i32
;
4966 params
[SI_PARAM_START_INSTANCE
] = ctx
->i32
;
4967 num_params
= SI_PARAM_START_INSTANCE
+1;
4969 if (shader
->key
.vs
.as_es
) {
4970 params
[ctx
->param_es2gs_offset
= num_params
++] = ctx
->i32
;
4971 } else if (shader
->key
.vs
.as_ls
) {
4972 params
[SI_PARAM_LS_OUT_LAYOUT
] = ctx
->i32
;
4973 num_params
= SI_PARAM_LS_OUT_LAYOUT
+1;
4975 if (ctx
->is_gs_copy_shader
) {
4976 last_array_pointer
= SI_PARAM_RW_BUFFERS
;
4977 num_params
= SI_PARAM_RW_BUFFERS
+1;
4979 params
[SI_PARAM_VS_STATE_BITS
] = ctx
->i32
;
4980 num_params
= SI_PARAM_VS_STATE_BITS
+1;
4983 /* The locations of the other parameters are assigned dynamically. */
4984 declare_streamout_params(ctx
, &shader
->selector
->so
,
4985 params
, ctx
->i32
, &num_params
);
4988 last_sgpr
= num_params
-1;
4991 params
[ctx
->param_vertex_id
= num_params
++] = ctx
->i32
;
4992 params
[ctx
->param_rel_auto_id
= num_params
++] = ctx
->i32
;
4993 params
[ctx
->param_vs_prim_id
= num_params
++] = ctx
->i32
;
4994 params
[ctx
->param_instance_id
= num_params
++] = ctx
->i32
;
4996 if (!ctx
->is_monolithic
&&
4997 !ctx
->is_gs_copy_shader
) {
4998 /* Vertex load indices. */
4999 ctx
->param_vertex_index0
= num_params
;
5001 for (i
= 0; i
< shader
->selector
->info
.num_inputs
; i
++)
5002 params
[num_params
++] = ctx
->i32
;
5004 /* PrimitiveID output. */
5005 if (!shader
->key
.vs
.as_es
&& !shader
->key
.vs
.as_ls
)
5006 for (i
= 0; i
<= VS_EPILOG_PRIMID_LOC
; i
++)
5007 returns
[num_returns
++] = ctx
->f32
;
5011 case PIPE_SHADER_TESS_CTRL
:
5012 params
[SI_PARAM_TCS_OUT_OFFSETS
] = ctx
->i32
;
5013 params
[SI_PARAM_TCS_OUT_LAYOUT
] = ctx
->i32
;
5014 params
[SI_PARAM_TCS_IN_LAYOUT
] = ctx
->i32
;
5015 params
[SI_PARAM_TESS_FACTOR_OFFSET
] = ctx
->i32
;
5016 last_sgpr
= SI_PARAM_TESS_FACTOR_OFFSET
;
5019 params
[SI_PARAM_PATCH_ID
] = ctx
->i32
;
5020 params
[SI_PARAM_REL_IDS
] = ctx
->i32
;
5021 num_params
= SI_PARAM_REL_IDS
+1;
5023 if (!ctx
->is_monolithic
) {
5024 /* PARAM_TESS_FACTOR_OFFSET is after user SGPRs. */
5025 for (i
= 0; i
<= SI_TCS_NUM_USER_SGPR
; i
++)
5026 returns
[num_returns
++] = ctx
->i32
; /* SGPRs */
5028 for (i
= 0; i
< 3; i
++)
5029 returns
[num_returns
++] = ctx
->f32
; /* VGPRs */
5033 case PIPE_SHADER_TESS_EVAL
:
5034 params
[SI_PARAM_TCS_OUT_OFFSETS
] = ctx
->i32
;
5035 params
[SI_PARAM_TCS_OUT_LAYOUT
] = ctx
->i32
;
5036 num_params
= SI_PARAM_TCS_OUT_LAYOUT
+1;
5038 if (shader
->key
.tes
.as_es
) {
5039 params
[ctx
->param_es2gs_offset
= num_params
++] = ctx
->i32
;
5041 declare_streamout_params(ctx
, &shader
->selector
->so
,
5042 params
, ctx
->i32
, &num_params
);
5044 last_sgpr
= num_params
- 1;
5047 params
[ctx
->param_tes_u
= num_params
++] = ctx
->f32
;
5048 params
[ctx
->param_tes_v
= num_params
++] = ctx
->f32
;
5049 params
[ctx
->param_tes_rel_patch_id
= num_params
++] = ctx
->i32
;
5050 params
[ctx
->param_tes_patch_id
= num_params
++] = ctx
->i32
;
5052 /* PrimitiveID output. */
5053 if (!ctx
->is_monolithic
&& !shader
->key
.tes
.as_es
)
5054 for (i
= 0; i
<= VS_EPILOG_PRIMID_LOC
; i
++)
5055 returns
[num_returns
++] = ctx
->f32
;
5058 case PIPE_SHADER_GEOMETRY
:
5059 params
[SI_PARAM_GS2VS_OFFSET
] = ctx
->i32
;
5060 params
[SI_PARAM_GS_WAVE_ID
] = ctx
->i32
;
5061 last_sgpr
= SI_PARAM_GS_WAVE_ID
;
5064 params
[SI_PARAM_VTX0_OFFSET
] = ctx
->i32
;
5065 params
[SI_PARAM_VTX1_OFFSET
] = ctx
->i32
;
5066 params
[SI_PARAM_PRIMITIVE_ID
] = ctx
->i32
;
5067 params
[SI_PARAM_VTX2_OFFSET
] = ctx
->i32
;
5068 params
[SI_PARAM_VTX3_OFFSET
] = ctx
->i32
;
5069 params
[SI_PARAM_VTX4_OFFSET
] = ctx
->i32
;
5070 params
[SI_PARAM_VTX5_OFFSET
] = ctx
->i32
;
5071 params
[SI_PARAM_GS_INSTANCE_ID
] = ctx
->i32
;
5072 num_params
= SI_PARAM_GS_INSTANCE_ID
+1;
5075 case PIPE_SHADER_FRAGMENT
:
5076 params
[SI_PARAM_ALPHA_REF
] = ctx
->f32
;
5077 params
[SI_PARAM_PRIM_MASK
] = ctx
->i32
;
5078 last_sgpr
= SI_PARAM_PRIM_MASK
;
5079 params
[SI_PARAM_PERSP_SAMPLE
] = ctx
->v2i32
;
5080 params
[SI_PARAM_PERSP_CENTER
] = ctx
->v2i32
;
5081 params
[SI_PARAM_PERSP_CENTROID
] = ctx
->v2i32
;
5082 params
[SI_PARAM_PERSP_PULL_MODEL
] = v3i32
;
5083 params
[SI_PARAM_LINEAR_SAMPLE
] = ctx
->v2i32
;
5084 params
[SI_PARAM_LINEAR_CENTER
] = ctx
->v2i32
;
5085 params
[SI_PARAM_LINEAR_CENTROID
] = ctx
->v2i32
;
5086 params
[SI_PARAM_LINE_STIPPLE_TEX
] = ctx
->f32
;
5087 params
[SI_PARAM_POS_X_FLOAT
] = ctx
->f32
;
5088 params
[SI_PARAM_POS_Y_FLOAT
] = ctx
->f32
;
5089 params
[SI_PARAM_POS_Z_FLOAT
] = ctx
->f32
;
5090 params
[SI_PARAM_POS_W_FLOAT
] = ctx
->f32
;
5091 params
[SI_PARAM_FRONT_FACE
] = ctx
->i32
;
5092 params
[SI_PARAM_ANCILLARY
] = ctx
->i32
;
5093 params
[SI_PARAM_SAMPLE_COVERAGE
] = ctx
->f32
;
5094 params
[SI_PARAM_POS_FIXED_PT
] = ctx
->i32
;
5095 num_params
= SI_PARAM_POS_FIXED_PT
+1;
5097 if (!ctx
->is_monolithic
) {
5098 /* Color inputs from the prolog. */
5099 if (shader
->selector
->info
.colors_read
) {
5100 unsigned num_color_elements
=
5101 util_bitcount(shader
->selector
->info
.colors_read
);
5103 assert(num_params
+ num_color_elements
<= ARRAY_SIZE(params
));
5104 for (i
= 0; i
< num_color_elements
; i
++)
5105 params
[num_params
++] = ctx
->f32
;
5108 /* Outputs for the epilog. */
5109 num_return_sgprs
= SI_SGPR_ALPHA_REF
+ 1;
5112 util_bitcount(shader
->selector
->info
.colors_written
) * 4 +
5113 shader
->selector
->info
.writes_z
+
5114 shader
->selector
->info
.writes_stencil
+
5115 shader
->selector
->info
.writes_samplemask
+
5116 1 /* SampleMaskIn */;
5118 num_returns
= MAX2(num_returns
,
5120 PS_EPILOG_SAMPLEMASK_MIN_LOC
+ 1);
5122 for (i
= 0; i
< num_return_sgprs
; i
++)
5123 returns
[i
] = ctx
->i32
;
5124 for (; i
< num_returns
; i
++)
5125 returns
[i
] = ctx
->f32
;
5129 case PIPE_SHADER_COMPUTE
:
5130 params
[SI_PARAM_GRID_SIZE
] = v3i32
;
5131 params
[SI_PARAM_BLOCK_ID
] = v3i32
;
5132 last_sgpr
= SI_PARAM_BLOCK_ID
;
5134 params
[SI_PARAM_THREAD_ID
] = v3i32
;
5135 num_params
= SI_PARAM_THREAD_ID
+ 1;
5138 assert(0 && "unimplemented shader");
5142 assert(num_params
<= Elements(params
));
5144 si_create_function(ctx
, returns
, num_returns
, params
,
5145 num_params
, last_array_pointer
, last_sgpr
);
5147 /* Reserve register locations for VGPR inputs the PS prolog may need. */
5148 if (ctx
->type
== PIPE_SHADER_FRAGMENT
&&
5149 !ctx
->is_monolithic
) {
5150 radeon_llvm_add_attribute(ctx
->radeon_bld
.main_fn
,
5151 "InitialPSInputAddr",
5152 S_0286D0_PERSP_SAMPLE_ENA(1) |
5153 S_0286D0_PERSP_CENTER_ENA(1) |
5154 S_0286D0_PERSP_CENTROID_ENA(1) |
5155 S_0286D0_LINEAR_SAMPLE_ENA(1) |
5156 S_0286D0_LINEAR_CENTER_ENA(1) |
5157 S_0286D0_LINEAR_CENTROID_ENA(1) |
5158 S_0286D0_FRONT_FACE_ENA(1) |
5159 S_0286D0_POS_FIXED_PT_ENA(1));
5160 } else if (ctx
->type
== PIPE_SHADER_COMPUTE
) {
5161 const unsigned *properties
= shader
->selector
->info
.properties
;
5162 unsigned max_work_group_size
=
5163 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
] *
5164 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT
] *
5165 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH
];
5167 assert(max_work_group_size
);
5169 radeon_llvm_add_attribute(ctx
->radeon_bld
.main_fn
,
5170 "amdgpu-max-work-group-size",
5171 max_work_group_size
);
5174 shader
->info
.num_input_sgprs
= 0;
5175 shader
->info
.num_input_vgprs
= 0;
5177 for (i
= 0; i
<= last_sgpr
; ++i
)
5178 shader
->info
.num_input_sgprs
+= llvm_get_type_size(params
[i
]) / 4;
5180 /* Unused fragment shader inputs are eliminated by the compiler,
5181 * so we don't know yet how many there will be.
5183 if (ctx
->type
!= PIPE_SHADER_FRAGMENT
)
5184 for (; i
< num_params
; ++i
)
5185 shader
->info
.num_input_vgprs
+= llvm_get_type_size(params
[i
]) / 4;
5187 if (bld_base
->info
&&
5188 (bld_base
->info
->opcode_count
[TGSI_OPCODE_DDX
] > 0 ||
5189 bld_base
->info
->opcode_count
[TGSI_OPCODE_DDY
] > 0 ||
5190 bld_base
->info
->opcode_count
[TGSI_OPCODE_DDX_FINE
] > 0 ||
5191 bld_base
->info
->opcode_count
[TGSI_OPCODE_DDY_FINE
] > 0 ||
5192 bld_base
->info
->opcode_count
[TGSI_OPCODE_INTERP_OFFSET
] > 0 ||
5193 bld_base
->info
->opcode_count
[TGSI_OPCODE_INTERP_SAMPLE
] > 0))
5195 LLVMAddGlobalInAddressSpace(gallivm
->module
,
5196 LLVMArrayType(ctx
->i32
, 64),
5200 if ((ctx
->type
== PIPE_SHADER_VERTEX
&& shader
->key
.vs
.as_ls
) ||
5201 ctx
->type
== PIPE_SHADER_TESS_CTRL
||
5202 ctx
->type
== PIPE_SHADER_TESS_EVAL
)
5203 declare_tess_lds(ctx
);
5206 static void preload_constants(struct si_shader_context
*ctx
)
5208 struct lp_build_tgsi_context
*bld_base
= &ctx
->radeon_bld
.soa
.bld_base
;
5209 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
5210 const struct tgsi_shader_info
*info
= bld_base
->info
;
5212 LLVMValueRef ptr
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_CONST_BUFFERS
);
5214 for (buf
= 0; buf
< SI_NUM_CONST_BUFFERS
; buf
++) {
5215 unsigned i
, num_const
= info
->const_file_max
[buf
] + 1;
5220 /* Allocate space for the constant values */
5221 ctx
->constants
[buf
] = CALLOC(num_const
* 4, sizeof(LLVMValueRef
));
5223 /* Load the resource descriptor */
5224 ctx
->const_buffers
[buf
] =
5225 build_indexed_load_const(ctx
, ptr
, lp_build_const_int32(gallivm
, buf
));
5227 /* Load the constants, we rely on the code sinking to do the rest */
5228 for (i
= 0; i
< num_const
* 4; ++i
) {
5229 ctx
->constants
[buf
][i
] =
5230 buffer_load_const(gallivm
->builder
,
5231 ctx
->const_buffers
[buf
],
5232 lp_build_const_int32(gallivm
, i
* 4),
5238 static void preload_shader_buffers(struct si_shader_context
*ctx
)
5240 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
5241 LLVMValueRef ptr
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_SHADER_BUFFERS
);
5244 maxbuf
= MIN2(ctx
->shader
->selector
->info
.file_max
[TGSI_FILE_BUFFER
],
5245 SI_NUM_SHADER_BUFFERS
- 1);
5246 for (buf
= 0; buf
<= maxbuf
; ++buf
) {
5247 ctx
->shader_buffers
[buf
] =
5248 build_indexed_load_const(
5249 ctx
, ptr
, lp_build_const_int32(gallivm
, buf
));
5253 static void preload_samplers(struct si_shader_context
*ctx
)
5255 struct lp_build_tgsi_context
*bld_base
= &ctx
->radeon_bld
.soa
.bld_base
;
5256 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
5257 const struct tgsi_shader_info
*info
= bld_base
->info
;
5258 unsigned i
, num_samplers
= info
->file_max
[TGSI_FILE_SAMPLER
] + 1;
5259 LLVMValueRef offset
;
5261 if (num_samplers
== 0)
5264 /* Load the resources and samplers, we rely on the code sinking to do the rest */
5265 for (i
= 0; i
< num_samplers
; ++i
) {
5267 offset
= lp_build_const_int32(gallivm
, i
);
5268 ctx
->sampler_views
[i
] =
5269 get_sampler_desc(ctx
, offset
, DESC_IMAGE
);
5271 /* FMASK resource */
5272 if (info
->is_msaa_sampler
[i
])
5274 get_sampler_desc(ctx
, offset
, DESC_FMASK
);
5276 ctx
->sampler_states
[i
] =
5277 get_sampler_desc(ctx
, offset
, DESC_SAMPLER
);
5278 ctx
->sampler_states
[i
] =
5279 sici_fix_sampler_aniso(ctx
, ctx
->sampler_views
[i
],
5280 ctx
->sampler_states
[i
]);
5285 static void preload_images(struct si_shader_context
*ctx
)
5287 struct lp_build_tgsi_context
*bld_base
= &ctx
->radeon_bld
.soa
.bld_base
;
5288 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
5289 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
5290 unsigned num_images
= bld_base
->info
->file_max
[TGSI_FILE_IMAGE
] + 1;
5291 LLVMValueRef res_ptr
;
5294 if (num_images
== 0)
5297 res_ptr
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_IMAGES
);
5299 for (i
= 0; i
< num_images
; ++i
) {
5300 /* Rely on LLVM to shrink the load for buffer resources. */
5302 build_indexed_load_const(ctx
, res_ptr
,
5303 lp_build_const_int32(gallivm
, i
));
5305 if (info
->images_writemask
& (1 << i
) &&
5306 !(info
->images_buffers
& (1 << i
)))
5307 rsrc
= force_dcc_off(ctx
, rsrc
);
5309 ctx
->images
[i
] = rsrc
;
5313 static void preload_streamout_buffers(struct si_shader_context
*ctx
)
5315 struct lp_build_tgsi_context
*bld_base
= &ctx
->radeon_bld
.soa
.bld_base
;
5316 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
5319 /* Streamout can only be used if the shader is compiled as VS. */
5320 if (!ctx
->shader
->selector
->so
.num_outputs
||
5321 (ctx
->type
== PIPE_SHADER_VERTEX
&&
5322 (ctx
->shader
->key
.vs
.as_es
||
5323 ctx
->shader
->key
.vs
.as_ls
)) ||
5324 (ctx
->type
== PIPE_SHADER_TESS_EVAL
&&
5325 ctx
->shader
->key
.tes
.as_es
))
5328 LLVMValueRef buf_ptr
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
5329 SI_PARAM_RW_BUFFERS
);
5331 /* Load the resources, we rely on the code sinking to do the rest */
5332 for (i
= 0; i
< 4; ++i
) {
5333 if (ctx
->shader
->selector
->so
.stride
[i
]) {
5334 LLVMValueRef offset
= lp_build_const_int32(gallivm
,
5335 SI_VS_STREAMOUT_BUF0
+ i
);
5337 ctx
->so_buffers
[i
] = build_indexed_load_const(ctx
, buf_ptr
, offset
);
5343 * Load ESGS and GSVS ring buffer resource descriptors and save the variables
5346 static void preload_ring_buffers(struct si_shader_context
*ctx
)
5348 struct gallivm_state
*gallivm
=
5349 ctx
->radeon_bld
.soa
.bld_base
.base
.gallivm
;
5351 LLVMValueRef buf_ptr
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
5352 SI_PARAM_RW_BUFFERS
);
5354 if ((ctx
->type
== PIPE_SHADER_VERTEX
&&
5355 ctx
->shader
->key
.vs
.as_es
) ||
5356 (ctx
->type
== PIPE_SHADER_TESS_EVAL
&&
5357 ctx
->shader
->key
.tes
.as_es
) ||
5358 ctx
->type
== PIPE_SHADER_GEOMETRY
) {
5360 ctx
->type
== PIPE_SHADER_GEOMETRY
? SI_GS_RING_ESGS
5362 LLVMValueRef offset
= lp_build_const_int32(gallivm
, ring
);
5365 build_indexed_load_const(ctx
, buf_ptr
, offset
);
5368 if (ctx
->is_gs_copy_shader
) {
5369 LLVMValueRef offset
= lp_build_const_int32(gallivm
, SI_VS_RING_GSVS
);
5372 build_indexed_load_const(ctx
, buf_ptr
, offset
);
5374 if (ctx
->type
== PIPE_SHADER_GEOMETRY
) {
5376 for (i
= 0; i
< 4; i
++) {
5377 LLVMValueRef offset
= lp_build_const_int32(gallivm
, SI_GS_RING_GSVS0
+ i
);
5380 build_indexed_load_const(ctx
, buf_ptr
, offset
);
5385 static void si_llvm_emit_polygon_stipple(struct si_shader_context
*ctx
,
5386 LLVMValueRef param_rw_buffers
,
5387 unsigned param_pos_fixed_pt
)
5389 struct lp_build_tgsi_context
*bld_base
=
5390 &ctx
->radeon_bld
.soa
.bld_base
;
5391 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
5392 LLVMBuilderRef builder
= gallivm
->builder
;
5393 LLVMValueRef slot
, desc
, offset
, row
, bit
, address
[2];
5395 /* Use the fixed-point gl_FragCoord input.
5396 * Since the stipple pattern is 32x32 and it repeats, just get 5 bits
5397 * per coordinate to get the repeating effect.
5399 address
[0] = unpack_param(ctx
, param_pos_fixed_pt
, 0, 5);
5400 address
[1] = unpack_param(ctx
, param_pos_fixed_pt
, 16, 5);
5402 /* Load the buffer descriptor. */
5403 slot
= lp_build_const_int32(gallivm
, SI_PS_CONST_POLY_STIPPLE
);
5404 desc
= build_indexed_load_const(ctx
, param_rw_buffers
, slot
);
5406 /* The stipple pattern is 32x32, each row has 32 bits. */
5407 offset
= LLVMBuildMul(builder
, address
[1],
5408 LLVMConstInt(ctx
->i32
, 4, 0), "");
5409 row
= buffer_load_const(builder
, desc
, offset
, ctx
->i32
);
5410 bit
= LLVMBuildLShr(builder
, row
, address
[0], "");
5411 bit
= LLVMBuildTrunc(builder
, bit
, ctx
->i1
, "");
5413 /* The intrinsic kills the thread if arg < 0. */
5414 bit
= LLVMBuildSelect(builder
, bit
, LLVMConstReal(ctx
->f32
, 0),
5415 LLVMConstReal(ctx
->f32
, -1), "");
5416 lp_build_intrinsic(builder
, "llvm.AMDGPU.kill", ctx
->voidt
, &bit
, 1, 0);
5419 void si_shader_binary_read_config(struct radeon_shader_binary
*binary
,
5420 struct si_shader_config
*conf
,
5421 unsigned symbol_offset
)
5424 const unsigned char *config
=
5425 radeon_shader_binary_config_start(binary
, symbol_offset
);
5427 /* XXX: We may be able to emit some of these values directly rather than
5428 * extracting fields to be emitted later.
5431 for (i
= 0; i
< binary
->config_size_per_symbol
; i
+= 8) {
5432 unsigned reg
= util_le32_to_cpu(*(uint32_t*)(config
+ i
));
5433 unsigned value
= util_le32_to_cpu(*(uint32_t*)(config
+ i
+ 4));
5435 case R_00B028_SPI_SHADER_PGM_RSRC1_PS
:
5436 case R_00B128_SPI_SHADER_PGM_RSRC1_VS
:
5437 case R_00B228_SPI_SHADER_PGM_RSRC1_GS
:
5438 case R_00B848_COMPUTE_PGM_RSRC1
:
5439 conf
->num_sgprs
= MAX2(conf
->num_sgprs
, (G_00B028_SGPRS(value
) + 1) * 8);
5440 conf
->num_vgprs
= MAX2(conf
->num_vgprs
, (G_00B028_VGPRS(value
) + 1) * 4);
5441 conf
->float_mode
= G_00B028_FLOAT_MODE(value
);
5442 conf
->rsrc1
= value
;
5444 case R_00B02C_SPI_SHADER_PGM_RSRC2_PS
:
5445 conf
->lds_size
= MAX2(conf
->lds_size
, G_00B02C_EXTRA_LDS_SIZE(value
));
5447 case R_00B84C_COMPUTE_PGM_RSRC2
:
5448 conf
->lds_size
= MAX2(conf
->lds_size
, G_00B84C_LDS_SIZE(value
));
5449 conf
->rsrc2
= value
;
5451 case R_0286CC_SPI_PS_INPUT_ENA
:
5452 conf
->spi_ps_input_ena
= value
;
5454 case R_0286D0_SPI_PS_INPUT_ADDR
:
5455 conf
->spi_ps_input_addr
= value
;
5457 case R_0286E8_SPI_TMPRING_SIZE
:
5458 case R_00B860_COMPUTE_TMPRING_SIZE
:
5459 /* WAVESIZE is in units of 256 dwords. */
5460 conf
->scratch_bytes_per_wave
=
5461 G_00B860_WAVESIZE(value
) * 256 * 4 * 1;
5465 static bool printed
;
5468 fprintf(stderr
, "Warning: LLVM emitted unknown "
5469 "config register: 0x%x\n", reg
);
5476 if (!conf
->spi_ps_input_addr
)
5477 conf
->spi_ps_input_addr
= conf
->spi_ps_input_ena
;
5481 void si_shader_apply_scratch_relocs(struct si_context
*sctx
,
5482 struct si_shader
*shader
,
5483 struct si_shader_config
*config
,
5484 uint64_t scratch_va
)
5487 uint32_t scratch_rsrc_dword0
= scratch_va
;
5488 uint32_t scratch_rsrc_dword1
=
5489 S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32)
5490 | S_008F04_STRIDE(config
->scratch_bytes_per_wave
/ 64);
5492 for (i
= 0 ; i
< shader
->binary
.reloc_count
; i
++) {
5493 const struct radeon_shader_reloc
*reloc
=
5494 &shader
->binary
.relocs
[i
];
5495 if (!strcmp(scratch_rsrc_dword0_symbol
, reloc
->name
)) {
5496 util_memcpy_cpu_to_le32(shader
->binary
.code
+ reloc
->offset
,
5497 &scratch_rsrc_dword0
, 4);
5498 } else if (!strcmp(scratch_rsrc_dword1_symbol
, reloc
->name
)) {
5499 util_memcpy_cpu_to_le32(shader
->binary
.code
+ reloc
->offset
,
5500 &scratch_rsrc_dword1
, 4);
5505 static unsigned si_get_shader_binary_size(struct si_shader
*shader
)
5507 unsigned size
= shader
->binary
.code_size
;
5510 size
+= shader
->prolog
->binary
.code_size
;
5512 size
+= shader
->epilog
->binary
.code_size
;
5516 int si_shader_binary_upload(struct si_screen
*sscreen
, struct si_shader
*shader
)
5518 const struct radeon_shader_binary
*prolog
=
5519 shader
->prolog
? &shader
->prolog
->binary
: NULL
;
5520 const struct radeon_shader_binary
*epilog
=
5521 shader
->epilog
? &shader
->epilog
->binary
: NULL
;
5522 const struct radeon_shader_binary
*mainb
= &shader
->binary
;
5523 unsigned bo_size
= si_get_shader_binary_size(shader
) +
5524 (!epilog
? mainb
->rodata_size
: 0);
5527 assert(!prolog
|| !prolog
->rodata_size
);
5528 assert((!prolog
&& !epilog
) || !mainb
->rodata_size
);
5529 assert(!epilog
|| !epilog
->rodata_size
);
5531 r600_resource_reference(&shader
->bo
, NULL
);
5532 shader
->bo
= si_resource_create_custom(&sscreen
->b
.b
,
5533 PIPE_USAGE_IMMUTABLE
,
5539 ptr
= sscreen
->b
.ws
->buffer_map(shader
->bo
->buf
, NULL
,
5540 PIPE_TRANSFER_READ_WRITE
);
5543 util_memcpy_cpu_to_le32(ptr
, prolog
->code
, prolog
->code_size
);
5544 ptr
+= prolog
->code_size
;
5547 util_memcpy_cpu_to_le32(ptr
, mainb
->code
, mainb
->code_size
);
5548 ptr
+= mainb
->code_size
;
5551 util_memcpy_cpu_to_le32(ptr
, epilog
->code
, epilog
->code_size
);
5552 else if (mainb
->rodata_size
> 0)
5553 util_memcpy_cpu_to_le32(ptr
, mainb
->rodata
, mainb
->rodata_size
);
5555 sscreen
->b
.ws
->buffer_unmap(shader
->bo
->buf
);
5559 static void si_shader_dump_disassembly(const struct radeon_shader_binary
*binary
,
5560 struct pipe_debug_callback
*debug
,
5561 const char *name
, FILE *file
)
5566 if (binary
->disasm_string
) {
5567 fprintf(file
, "Shader %s disassembly:\n", name
);
5568 fprintf(file
, "%s", binary
->disasm_string
);
5570 if (debug
&& debug
->debug_message
) {
5571 /* Very long debug messages are cut off, so send the
5572 * disassembly one line at a time. This causes more
5573 * overhead, but on the plus side it simplifies
5574 * parsing of resulting logs.
5576 pipe_debug_message(debug
, SHADER_INFO
,
5577 "Shader Disassembly Begin");
5579 line
= binary
->disasm_string
;
5581 p
= util_strchrnul(line
, '\n');
5585 pipe_debug_message(debug
, SHADER_INFO
,
5586 "%.*s", count
, line
);
5594 pipe_debug_message(debug
, SHADER_INFO
,
5595 "Shader Disassembly End");
5598 fprintf(file
, "Shader %s binary:\n", name
);
5599 for (i
= 0; i
< binary
->code_size
; i
+= 4) {
5600 fprintf(file
, "@0x%x: %02x%02x%02x%02x\n", i
,
5601 binary
->code
[i
+ 3], binary
->code
[i
+ 2],
5602 binary
->code
[i
+ 1], binary
->code
[i
]);
5607 static void si_shader_dump_stats(struct si_screen
*sscreen
,
5608 struct si_shader_config
*conf
,
5609 unsigned num_inputs
,
5611 struct pipe_debug_callback
*debug
,
5615 unsigned lds_increment
= sscreen
->b
.chip_class
>= CIK
? 512 : 256;
5616 unsigned lds_per_wave
= 0;
5617 unsigned max_simd_waves
= 10;
5619 /* Compute LDS usage for PS. */
5620 if (processor
== PIPE_SHADER_FRAGMENT
) {
5621 /* The minimum usage per wave is (num_inputs * 36). The maximum
5622 * usage is (num_inputs * 36 * 16).
5623 * We can get anything in between and it varies between waves.
5625 * Other stages don't know the size at compile time or don't
5626 * allocate LDS per wave, but instead they do it per thread group.
5628 lds_per_wave
= conf
->lds_size
* lds_increment
+
5629 align(num_inputs
* 36, lds_increment
);
5632 /* Compute the per-SIMD wave counts. */
5633 if (conf
->num_sgprs
) {
5634 if (sscreen
->b
.chip_class
>= VI
)
5635 max_simd_waves
= MIN2(max_simd_waves
, 800 / conf
->num_sgprs
);
5637 max_simd_waves
= MIN2(max_simd_waves
, 512 / conf
->num_sgprs
);
5640 if (conf
->num_vgprs
)
5641 max_simd_waves
= MIN2(max_simd_waves
, 256 / conf
->num_vgprs
);
5643 /* LDS is 64KB per CU (4 SIMDs), divided into 16KB blocks per SIMD
5647 max_simd_waves
= MIN2(max_simd_waves
, 16384 / lds_per_wave
);
5649 if (file
!= stderr
||
5650 r600_can_dump_shader(&sscreen
->b
, processor
)) {
5651 if (processor
== PIPE_SHADER_FRAGMENT
) {
5652 fprintf(file
, "*** SHADER CONFIG ***\n"
5653 "SPI_PS_INPUT_ADDR = 0x%04x\n"
5654 "SPI_PS_INPUT_ENA = 0x%04x\n",
5655 conf
->spi_ps_input_addr
, conf
->spi_ps_input_ena
);
5658 fprintf(file
, "*** SHADER STATS ***\n"
5661 "Code Size: %d bytes\n"
5663 "Scratch: %d bytes per wave\n"
5665 "********************\n",
5666 conf
->num_sgprs
, conf
->num_vgprs
, code_size
,
5667 conf
->lds_size
, conf
->scratch_bytes_per_wave
,
5671 pipe_debug_message(debug
, SHADER_INFO
,
5672 "Shader Stats: SGPRS: %d VGPRS: %d Code Size: %d "
5673 "LDS: %d Scratch: %d Max Waves: %d",
5674 conf
->num_sgprs
, conf
->num_vgprs
, code_size
,
5675 conf
->lds_size
, conf
->scratch_bytes_per_wave
,
5679 static const char *si_get_shader_name(struct si_shader
*shader
,
5682 switch (processor
) {
5683 case PIPE_SHADER_VERTEX
:
5684 if (shader
->key
.vs
.as_es
)
5685 return "Vertex Shader as ES";
5686 else if (shader
->key
.vs
.as_ls
)
5687 return "Vertex Shader as LS";
5689 return "Vertex Shader as VS";
5690 case PIPE_SHADER_TESS_CTRL
:
5691 return "Tessellation Control Shader";
5692 case PIPE_SHADER_TESS_EVAL
:
5693 if (shader
->key
.tes
.as_es
)
5694 return "Tessellation Evaluation Shader as ES";
5696 return "Tessellation Evaluation Shader as VS";
5697 case PIPE_SHADER_GEOMETRY
:
5698 if (shader
->gs_copy_shader
== NULL
)
5699 return "GS Copy Shader as VS";
5701 return "Geometry Shader";
5702 case PIPE_SHADER_FRAGMENT
:
5703 return "Pixel Shader";
5704 case PIPE_SHADER_COMPUTE
:
5705 return "Compute Shader";
5707 return "Unknown Shader";
5711 void si_shader_dump(struct si_screen
*sscreen
, struct si_shader
*shader
,
5712 struct pipe_debug_callback
*debug
, unsigned processor
,
5715 if (file
!= stderr
||
5716 (r600_can_dump_shader(&sscreen
->b
, processor
) &&
5717 !(sscreen
->b
.debug_flags
& DBG_NO_ASM
))) {
5718 fprintf(file
, "\n%s:\n", si_get_shader_name(shader
, processor
));
5721 si_shader_dump_disassembly(&shader
->prolog
->binary
,
5722 debug
, "prolog", file
);
5724 si_shader_dump_disassembly(&shader
->binary
, debug
, "main", file
);
5727 si_shader_dump_disassembly(&shader
->epilog
->binary
,
5728 debug
, "epilog", file
);
5729 fprintf(file
, "\n");
5732 si_shader_dump_stats(sscreen
, &shader
->config
,
5733 shader
->selector
? shader
->selector
->info
.num_inputs
: 0,
5734 si_get_shader_binary_size(shader
), debug
, processor
,
5738 int si_compile_llvm(struct si_screen
*sscreen
,
5739 struct radeon_shader_binary
*binary
,
5740 struct si_shader_config
*conf
,
5741 LLVMTargetMachineRef tm
,
5743 struct pipe_debug_callback
*debug
,
5748 unsigned count
= p_atomic_inc_return(&sscreen
->b
.num_compilations
);
5750 if (r600_can_dump_shader(&sscreen
->b
, processor
)) {
5751 fprintf(stderr
, "radeonsi: Compiling shader %d\n", count
);
5753 if (!(sscreen
->b
.debug_flags
& (DBG_NO_IR
| DBG_PREOPT_IR
))) {
5754 fprintf(stderr
, "%s LLVM IR:\n\n", name
);
5755 LLVMDumpModule(mod
);
5756 fprintf(stderr
, "\n");
5760 if (!si_replace_shader(count
, binary
)) {
5761 r
= radeon_llvm_compile(mod
, binary
,
5762 r600_get_llvm_processor_name(sscreen
->b
.family
), tm
,
5768 si_shader_binary_read_config(binary
, conf
, 0);
5770 /* Enable 64-bit and 16-bit denormals, because there is no performance
5773 * If denormals are enabled, all floating-point output modifiers are
5776 * Don't enable denormals for 32-bit floats, because:
5777 * - Floating-point output modifiers would be ignored by the hw.
5778 * - Some opcodes don't support denormals, such as v_mad_f32. We would
5779 * have to stop using those.
5780 * - SI & CI would be very slow.
5782 conf
->float_mode
|= V_00B028_FP_64_DENORMS
;
5784 FREE(binary
->config
);
5785 FREE(binary
->global_symbol_offsets
);
5786 binary
->config
= NULL
;
5787 binary
->global_symbol_offsets
= NULL
;
5789 /* Some shaders can't have rodata because their binaries can be
5792 if (binary
->rodata_size
&&
5793 (processor
== PIPE_SHADER_VERTEX
||
5794 processor
== PIPE_SHADER_TESS_CTRL
||
5795 processor
== PIPE_SHADER_TESS_EVAL
||
5796 processor
== PIPE_SHADER_FRAGMENT
)) {
5797 fprintf(stderr
, "radeonsi: The shader can't have rodata.");
5804 /* Generate code for the hardware VS shader stage to go with a geometry shader */
5805 static int si_generate_gs_copy_shader(struct si_screen
*sscreen
,
5806 struct si_shader_context
*ctx
,
5807 struct si_shader
*gs
,
5808 struct pipe_debug_callback
*debug
)
5810 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
5811 struct lp_build_tgsi_context
*bld_base
= &ctx
->radeon_bld
.soa
.bld_base
;
5812 struct lp_build_context
*uint
= &bld_base
->uint_bld
;
5813 struct si_shader_output_values
*outputs
;
5814 struct tgsi_shader_info
*gsinfo
= &gs
->selector
->info
;
5815 LLVMValueRef args
[9];
5818 outputs
= MALLOC(gsinfo
->num_outputs
* sizeof(outputs
[0]));
5820 si_init_shader_ctx(ctx
, sscreen
, ctx
->shader
, ctx
->tm
);
5821 ctx
->type
= PIPE_SHADER_VERTEX
;
5822 ctx
->is_gs_copy_shader
= true;
5824 create_meta_data(ctx
);
5825 create_function(ctx
);
5826 preload_streamout_buffers(ctx
);
5827 preload_ring_buffers(ctx
);
5829 args
[0] = ctx
->gsvs_ring
[0];
5830 args
[1] = lp_build_mul_imm(uint
,
5831 LLVMGetParam(ctx
->radeon_bld
.main_fn
,
5832 ctx
->param_vertex_id
),
5834 args
[3] = uint
->zero
;
5835 args
[4] = uint
->one
; /* OFFEN */
5836 args
[5] = uint
->zero
; /* IDXEN */
5837 args
[6] = uint
->one
; /* GLC */
5838 args
[7] = uint
->one
; /* SLC */
5839 args
[8] = uint
->zero
; /* TFE */
5841 /* Fetch vertex data from GSVS ring */
5842 for (i
= 0; i
< gsinfo
->num_outputs
; ++i
) {
5845 outputs
[i
].name
= gsinfo
->output_semantic_name
[i
];
5846 outputs
[i
].sid
= gsinfo
->output_semantic_index
[i
];
5848 for (chan
= 0; chan
< 4; chan
++) {
5849 args
[2] = lp_build_const_int32(gallivm
,
5851 gs
->selector
->gs_max_out_vertices
* 16 * 4);
5853 outputs
[i
].values
[chan
] =
5854 LLVMBuildBitCast(gallivm
->builder
,
5855 lp_build_intrinsic(gallivm
->builder
,
5856 "llvm.SI.buffer.load.dword.i32.i32",
5858 LLVMReadOnlyAttribute
| LLVMNoUnwindAttribute
),
5863 si_llvm_export_vs(bld_base
, outputs
, gsinfo
->num_outputs
);
5865 LLVMBuildRet(gallivm
->builder
, ctx
->return_value
);
5867 /* Dump LLVM IR before any optimization passes */
5868 if (sscreen
->b
.debug_flags
& DBG_PREOPT_IR
&&
5869 r600_can_dump_shader(&sscreen
->b
, PIPE_SHADER_GEOMETRY
))
5870 LLVMDumpModule(bld_base
->base
.gallivm
->module
);
5872 radeon_llvm_finalize_module(&ctx
->radeon_bld
);
5874 r
= si_compile_llvm(sscreen
, &ctx
->shader
->binary
,
5875 &ctx
->shader
->config
, ctx
->tm
,
5876 bld_base
->base
.gallivm
->module
,
5877 debug
, PIPE_SHADER_GEOMETRY
,
5880 if (r600_can_dump_shader(&sscreen
->b
, PIPE_SHADER_GEOMETRY
))
5881 fprintf(stderr
, "GS Copy Shader:\n");
5882 si_shader_dump(sscreen
, ctx
->shader
, debug
,
5883 PIPE_SHADER_GEOMETRY
, stderr
);
5884 r
= si_shader_binary_upload(sscreen
, ctx
->shader
);
5887 radeon_llvm_dispose(&ctx
->radeon_bld
);
5893 void si_dump_shader_key(unsigned shader
, union si_shader_key
*key
, FILE *f
)
5897 fprintf(f
, "SHADER KEY\n");
5900 case PIPE_SHADER_VERTEX
:
5901 fprintf(f
, " instance_divisors = {");
5902 for (i
= 0; i
< Elements(key
->vs
.prolog
.instance_divisors
); i
++)
5903 fprintf(f
, !i
? "%u" : ", %u",
5904 key
->vs
.prolog
.instance_divisors
[i
]);
5906 fprintf(f
, " as_es = %u\n", key
->vs
.as_es
);
5907 fprintf(f
, " as_ls = %u\n", key
->vs
.as_ls
);
5908 fprintf(f
, " export_prim_id = %u\n", key
->vs
.epilog
.export_prim_id
);
5911 case PIPE_SHADER_TESS_CTRL
:
5912 fprintf(f
, " prim_mode = %u\n", key
->tcs
.epilog
.prim_mode
);
5915 case PIPE_SHADER_TESS_EVAL
:
5916 fprintf(f
, " as_es = %u\n", key
->tes
.as_es
);
5917 fprintf(f
, " export_prim_id = %u\n", key
->tes
.epilog
.export_prim_id
);
5920 case PIPE_SHADER_GEOMETRY
:
5921 case PIPE_SHADER_COMPUTE
:
5924 case PIPE_SHADER_FRAGMENT
:
5925 fprintf(f
, " prolog.color_two_side = %u\n", key
->ps
.prolog
.color_two_side
);
5926 fprintf(f
, " prolog.poly_stipple = %u\n", key
->ps
.prolog
.poly_stipple
);
5927 fprintf(f
, " prolog.force_persample_interp = %u\n", key
->ps
.prolog
.force_persample_interp
);
5928 fprintf(f
, " epilog.spi_shader_col_format = 0x%x\n", key
->ps
.epilog
.spi_shader_col_format
);
5929 fprintf(f
, " epilog.color_is_int8 = 0x%X\n", key
->ps
.epilog
.color_is_int8
);
5930 fprintf(f
, " epilog.last_cbuf = %u\n", key
->ps
.epilog
.last_cbuf
);
5931 fprintf(f
, " epilog.alpha_func = %u\n", key
->ps
.epilog
.alpha_func
);
5932 fprintf(f
, " epilog.alpha_to_one = %u\n", key
->ps
.epilog
.alpha_to_one
);
5933 fprintf(f
, " epilog.poly_line_smoothing = %u\n", key
->ps
.epilog
.poly_line_smoothing
);
5934 fprintf(f
, " epilog.clamp_color = %u\n", key
->ps
.epilog
.clamp_color
);
5942 static void si_init_shader_ctx(struct si_shader_context
*ctx
,
5943 struct si_screen
*sscreen
,
5944 struct si_shader
*shader
,
5945 LLVMTargetMachineRef tm
)
5947 struct lp_build_tgsi_context
*bld_base
;
5948 struct lp_build_tgsi_action tmpl
= {};
5950 memset(ctx
, 0, sizeof(*ctx
));
5951 radeon_llvm_context_init(&ctx
->radeon_bld
, "amdgcn--");
5953 ctx
->screen
= sscreen
;
5954 if (shader
&& shader
->selector
)
5955 ctx
->type
= shader
->selector
->info
.processor
;
5958 ctx
->shader
= shader
;
5960 ctx
->voidt
= LLVMVoidTypeInContext(ctx
->radeon_bld
.gallivm
.context
);
5961 ctx
->i1
= LLVMInt1TypeInContext(ctx
->radeon_bld
.gallivm
.context
);
5962 ctx
->i8
= LLVMInt8TypeInContext(ctx
->radeon_bld
.gallivm
.context
);
5963 ctx
->i32
= LLVMInt32TypeInContext(ctx
->radeon_bld
.gallivm
.context
);
5964 ctx
->i64
= LLVMInt64TypeInContext(ctx
->radeon_bld
.gallivm
.context
);
5965 ctx
->i128
= LLVMIntTypeInContext(ctx
->radeon_bld
.gallivm
.context
, 128);
5966 ctx
->f32
= LLVMFloatTypeInContext(ctx
->radeon_bld
.gallivm
.context
);
5967 ctx
->v16i8
= LLVMVectorType(ctx
->i8
, 16);
5968 ctx
->v2i32
= LLVMVectorType(ctx
->i32
, 2);
5969 ctx
->v4i32
= LLVMVectorType(ctx
->i32
, 4);
5970 ctx
->v4f32
= LLVMVectorType(ctx
->f32
, 4);
5971 ctx
->v8i32
= LLVMVectorType(ctx
->i32
, 8);
5973 bld_base
= &ctx
->radeon_bld
.soa
.bld_base
;
5974 if (shader
&& shader
->selector
)
5975 bld_base
->info
= &shader
->selector
->info
;
5976 bld_base
->emit_fetch_funcs
[TGSI_FILE_CONSTANT
] = fetch_constant
;
5978 bld_base
->op_actions
[TGSI_OPCODE_INTERP_CENTROID
] = interp_action
;
5979 bld_base
->op_actions
[TGSI_OPCODE_INTERP_SAMPLE
] = interp_action
;
5980 bld_base
->op_actions
[TGSI_OPCODE_INTERP_OFFSET
] = interp_action
;
5982 bld_base
->op_actions
[TGSI_OPCODE_TEX
] = tex_action
;
5983 bld_base
->op_actions
[TGSI_OPCODE_TEX2
] = tex_action
;
5984 bld_base
->op_actions
[TGSI_OPCODE_TXB
] = tex_action
;
5985 bld_base
->op_actions
[TGSI_OPCODE_TXB2
] = tex_action
;
5986 bld_base
->op_actions
[TGSI_OPCODE_TXD
] = tex_action
;
5987 bld_base
->op_actions
[TGSI_OPCODE_TXF
] = tex_action
;
5988 bld_base
->op_actions
[TGSI_OPCODE_TXL
] = tex_action
;
5989 bld_base
->op_actions
[TGSI_OPCODE_TXL2
] = tex_action
;
5990 bld_base
->op_actions
[TGSI_OPCODE_TXP
] = tex_action
;
5991 bld_base
->op_actions
[TGSI_OPCODE_TXQ
] = tex_action
;
5992 bld_base
->op_actions
[TGSI_OPCODE_TG4
] = tex_action
;
5993 bld_base
->op_actions
[TGSI_OPCODE_LODQ
] = tex_action
;
5994 bld_base
->op_actions
[TGSI_OPCODE_TXQS
].emit
= si_llvm_emit_txqs
;
5996 bld_base
->op_actions
[TGSI_OPCODE_LOAD
].fetch_args
= load_fetch_args
;
5997 bld_base
->op_actions
[TGSI_OPCODE_LOAD
].emit
= load_emit
;
5998 bld_base
->op_actions
[TGSI_OPCODE_STORE
].fetch_args
= store_fetch_args
;
5999 bld_base
->op_actions
[TGSI_OPCODE_STORE
].emit
= store_emit
;
6000 bld_base
->op_actions
[TGSI_OPCODE_RESQ
].fetch_args
= resq_fetch_args
;
6001 bld_base
->op_actions
[TGSI_OPCODE_RESQ
].emit
= resq_emit
;
6003 tmpl
.fetch_args
= atomic_fetch_args
;
6004 tmpl
.emit
= atomic_emit
;
6005 bld_base
->op_actions
[TGSI_OPCODE_ATOMUADD
] = tmpl
;
6006 bld_base
->op_actions
[TGSI_OPCODE_ATOMUADD
].intr_name
= "add";
6007 bld_base
->op_actions
[TGSI_OPCODE_ATOMXCHG
] = tmpl
;
6008 bld_base
->op_actions
[TGSI_OPCODE_ATOMXCHG
].intr_name
= "swap";
6009 bld_base
->op_actions
[TGSI_OPCODE_ATOMCAS
] = tmpl
;
6010 bld_base
->op_actions
[TGSI_OPCODE_ATOMCAS
].intr_name
= "cmpswap";
6011 bld_base
->op_actions
[TGSI_OPCODE_ATOMAND
] = tmpl
;
6012 bld_base
->op_actions
[TGSI_OPCODE_ATOMAND
].intr_name
= "and";
6013 bld_base
->op_actions
[TGSI_OPCODE_ATOMOR
] = tmpl
;
6014 bld_base
->op_actions
[TGSI_OPCODE_ATOMOR
].intr_name
= "or";
6015 bld_base
->op_actions
[TGSI_OPCODE_ATOMXOR
] = tmpl
;
6016 bld_base
->op_actions
[TGSI_OPCODE_ATOMXOR
].intr_name
= "xor";
6017 bld_base
->op_actions
[TGSI_OPCODE_ATOMUMIN
] = tmpl
;
6018 bld_base
->op_actions
[TGSI_OPCODE_ATOMUMIN
].intr_name
= "umin";
6019 bld_base
->op_actions
[TGSI_OPCODE_ATOMUMAX
] = tmpl
;
6020 bld_base
->op_actions
[TGSI_OPCODE_ATOMUMAX
].intr_name
= "umax";
6021 bld_base
->op_actions
[TGSI_OPCODE_ATOMIMIN
] = tmpl
;
6022 bld_base
->op_actions
[TGSI_OPCODE_ATOMIMIN
].intr_name
= "smin";
6023 bld_base
->op_actions
[TGSI_OPCODE_ATOMIMAX
] = tmpl
;
6024 bld_base
->op_actions
[TGSI_OPCODE_ATOMIMAX
].intr_name
= "smax";
6026 bld_base
->op_actions
[TGSI_OPCODE_MEMBAR
].emit
= membar_emit
;
6028 bld_base
->op_actions
[TGSI_OPCODE_DDX
].emit
= si_llvm_emit_ddxy
;
6029 bld_base
->op_actions
[TGSI_OPCODE_DDY
].emit
= si_llvm_emit_ddxy
;
6030 bld_base
->op_actions
[TGSI_OPCODE_DDX_FINE
].emit
= si_llvm_emit_ddxy
;
6031 bld_base
->op_actions
[TGSI_OPCODE_DDY_FINE
].emit
= si_llvm_emit_ddxy
;
6033 bld_base
->op_actions
[TGSI_OPCODE_EMIT
].emit
= si_llvm_emit_vertex
;
6034 bld_base
->op_actions
[TGSI_OPCODE_ENDPRIM
].emit
= si_llvm_emit_primitive
;
6035 bld_base
->op_actions
[TGSI_OPCODE_BARRIER
].emit
= si_llvm_emit_barrier
;
6037 bld_base
->op_actions
[TGSI_OPCODE_MAX
].emit
= build_tgsi_intrinsic_nomem
;
6038 bld_base
->op_actions
[TGSI_OPCODE_MAX
].intr_name
= "llvm.maxnum.f32";
6039 bld_base
->op_actions
[TGSI_OPCODE_MIN
].emit
= build_tgsi_intrinsic_nomem
;
6040 bld_base
->op_actions
[TGSI_OPCODE_MIN
].intr_name
= "llvm.minnum.f32";
6043 int si_compile_tgsi_shader(struct si_screen
*sscreen
,
6044 LLVMTargetMachineRef tm
,
6045 struct si_shader
*shader
,
6047 struct pipe_debug_callback
*debug
)
6049 struct si_shader_selector
*sel
= shader
->selector
;
6050 struct si_shader_context ctx
;
6051 struct lp_build_tgsi_context
*bld_base
;
6055 /* Dump TGSI code before doing TGSI->LLVM conversion in case the
6056 * conversion fails. */
6057 if (r600_can_dump_shader(&sscreen
->b
, sel
->info
.processor
) &&
6058 !(sscreen
->b
.debug_flags
& DBG_NO_TGSI
)) {
6059 si_dump_shader_key(sel
->type
, &shader
->key
, stderr
);
6060 tgsi_dump(sel
->tokens
, 0);
6061 si_dump_streamout(&sel
->so
);
6064 si_init_shader_ctx(&ctx
, sscreen
, shader
, tm
);
6065 ctx
.is_monolithic
= is_monolithic
;
6067 shader
->info
.uses_instanceid
= sel
->info
.uses_instanceid
;
6069 bld_base
= &ctx
.radeon_bld
.soa
.bld_base
;
6070 ctx
.radeon_bld
.load_system_value
= declare_system_value
;
6073 case PIPE_SHADER_VERTEX
:
6074 ctx
.radeon_bld
.load_input
= declare_input_vs
;
6075 if (shader
->key
.vs
.as_ls
)
6076 bld_base
->emit_epilogue
= si_llvm_emit_ls_epilogue
;
6077 else if (shader
->key
.vs
.as_es
)
6078 bld_base
->emit_epilogue
= si_llvm_emit_es_epilogue
;
6080 bld_base
->emit_epilogue
= si_llvm_emit_vs_epilogue
;
6082 case PIPE_SHADER_TESS_CTRL
:
6083 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = fetch_input_tcs
;
6084 bld_base
->emit_fetch_funcs
[TGSI_FILE_OUTPUT
] = fetch_output_tcs
;
6085 bld_base
->emit_store
= store_output_tcs
;
6086 bld_base
->emit_epilogue
= si_llvm_emit_tcs_epilogue
;
6088 case PIPE_SHADER_TESS_EVAL
:
6089 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = fetch_input_tes
;
6090 if (shader
->key
.tes
.as_es
)
6091 bld_base
->emit_epilogue
= si_llvm_emit_es_epilogue
;
6093 bld_base
->emit_epilogue
= si_llvm_emit_vs_epilogue
;
6095 case PIPE_SHADER_GEOMETRY
:
6096 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = fetch_input_gs
;
6097 bld_base
->emit_epilogue
= si_llvm_emit_gs_epilogue
;
6099 case PIPE_SHADER_FRAGMENT
:
6100 ctx
.radeon_bld
.load_input
= declare_input_fs
;
6102 bld_base
->emit_epilogue
= si_llvm_emit_fs_epilogue
;
6104 bld_base
->emit_epilogue
= si_llvm_return_fs_outputs
;
6106 case PIPE_SHADER_COMPUTE
:
6107 ctx
.radeon_bld
.declare_memory_region
= declare_compute_memory
;
6110 assert(!"Unsupported shader type");
6114 create_meta_data(&ctx
);
6115 create_function(&ctx
);
6116 preload_constants(&ctx
);
6117 preload_shader_buffers(&ctx
);
6118 preload_samplers(&ctx
);
6119 preload_images(&ctx
);
6120 preload_streamout_buffers(&ctx
);
6121 preload_ring_buffers(&ctx
);
6123 if (ctx
.is_monolithic
&& sel
->type
== PIPE_SHADER_FRAGMENT
&&
6124 shader
->key
.ps
.prolog
.poly_stipple
) {
6125 LLVMValueRef list
= LLVMGetParam(ctx
.radeon_bld
.main_fn
,
6126 SI_PARAM_RW_BUFFERS
);
6127 si_llvm_emit_polygon_stipple(&ctx
, list
,
6128 SI_PARAM_POS_FIXED_PT
);
6131 if (ctx
.type
== PIPE_SHADER_GEOMETRY
) {
6133 for (i
= 0; i
< 4; i
++) {
6134 ctx
.gs_next_vertex
[i
] =
6135 lp_build_alloca(bld_base
->base
.gallivm
,
6140 if (!lp_build_tgsi_llvm(bld_base
, sel
->tokens
)) {
6141 fprintf(stderr
, "Failed to translate shader from TGSI to LLVM\n");
6145 LLVMBuildRet(bld_base
->base
.gallivm
->builder
, ctx
.return_value
);
6146 mod
= bld_base
->base
.gallivm
->module
;
6148 /* Dump LLVM IR before any optimization passes */
6149 if (sscreen
->b
.debug_flags
& DBG_PREOPT_IR
&&
6150 r600_can_dump_shader(&sscreen
->b
, ctx
.type
))
6151 LLVMDumpModule(mod
);
6153 radeon_llvm_finalize_module(&ctx
.radeon_bld
);
6155 r
= si_compile_llvm(sscreen
, &shader
->binary
, &shader
->config
, tm
,
6156 mod
, debug
, ctx
.type
, "TGSI shader");
6158 fprintf(stderr
, "LLVM failed to compile shader\n");
6162 radeon_llvm_dispose(&ctx
.radeon_bld
);
6164 /* Add the scratch offset to input SGPRs. */
6165 if (shader
->config
.scratch_bytes_per_wave
)
6166 shader
->info
.num_input_sgprs
+= 1; /* scratch byte offset */
6168 /* Calculate the number of fragment input VGPRs. */
6169 if (ctx
.type
== PIPE_SHADER_FRAGMENT
) {
6170 shader
->info
.num_input_vgprs
= 0;
6171 shader
->info
.face_vgpr_index
= -1;
6173 if (G_0286CC_PERSP_SAMPLE_ENA(shader
->config
.spi_ps_input_addr
))
6174 shader
->info
.num_input_vgprs
+= 2;
6175 if (G_0286CC_PERSP_CENTER_ENA(shader
->config
.spi_ps_input_addr
))
6176 shader
->info
.num_input_vgprs
+= 2;
6177 if (G_0286CC_PERSP_CENTROID_ENA(shader
->config
.spi_ps_input_addr
))
6178 shader
->info
.num_input_vgprs
+= 2;
6179 if (G_0286CC_PERSP_PULL_MODEL_ENA(shader
->config
.spi_ps_input_addr
))
6180 shader
->info
.num_input_vgprs
+= 3;
6181 if (G_0286CC_LINEAR_SAMPLE_ENA(shader
->config
.spi_ps_input_addr
))
6182 shader
->info
.num_input_vgprs
+= 2;
6183 if (G_0286CC_LINEAR_CENTER_ENA(shader
->config
.spi_ps_input_addr
))
6184 shader
->info
.num_input_vgprs
+= 2;
6185 if (G_0286CC_LINEAR_CENTROID_ENA(shader
->config
.spi_ps_input_addr
))
6186 shader
->info
.num_input_vgprs
+= 2;
6187 if (G_0286CC_LINE_STIPPLE_TEX_ENA(shader
->config
.spi_ps_input_addr
))
6188 shader
->info
.num_input_vgprs
+= 1;
6189 if (G_0286CC_POS_X_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
6190 shader
->info
.num_input_vgprs
+= 1;
6191 if (G_0286CC_POS_Y_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
6192 shader
->info
.num_input_vgprs
+= 1;
6193 if (G_0286CC_POS_Z_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
6194 shader
->info
.num_input_vgprs
+= 1;
6195 if (G_0286CC_POS_W_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
6196 shader
->info
.num_input_vgprs
+= 1;
6197 if (G_0286CC_FRONT_FACE_ENA(shader
->config
.spi_ps_input_addr
)) {
6198 shader
->info
.face_vgpr_index
= shader
->info
.num_input_vgprs
;
6199 shader
->info
.num_input_vgprs
+= 1;
6201 if (G_0286CC_ANCILLARY_ENA(shader
->config
.spi_ps_input_addr
))
6202 shader
->info
.num_input_vgprs
+= 1;
6203 if (G_0286CC_SAMPLE_COVERAGE_ENA(shader
->config
.spi_ps_input_addr
))
6204 shader
->info
.num_input_vgprs
+= 1;
6205 if (G_0286CC_POS_FIXED_PT_ENA(shader
->config
.spi_ps_input_addr
))
6206 shader
->info
.num_input_vgprs
+= 1;
6209 if (ctx
.type
== PIPE_SHADER_GEOMETRY
) {
6210 shader
->gs_copy_shader
= CALLOC_STRUCT(si_shader
);
6211 shader
->gs_copy_shader
->selector
= shader
->selector
;
6212 ctx
.shader
= shader
->gs_copy_shader
;
6213 if ((r
= si_generate_gs_copy_shader(sscreen
, &ctx
,
6215 free(shader
->gs_copy_shader
);
6216 shader
->gs_copy_shader
= NULL
;
6222 for (int i
= 0; i
< SI_NUM_CONST_BUFFERS
; i
++)
6223 FREE(ctx
.constants
[i
]);
6228 * Create, compile and return a shader part (prolog or epilog).
6230 * \param sscreen screen
6231 * \param list list of shader parts of the same category
6232 * \param key shader part key
6233 * \param tm LLVM target machine
6234 * \param debug debug callback
6235 * \param compile the callback responsible for compilation
6236 * \return non-NULL on success
6238 static struct si_shader_part
*
6239 si_get_shader_part(struct si_screen
*sscreen
,
6240 struct si_shader_part
**list
,
6241 union si_shader_part_key
*key
,
6242 LLVMTargetMachineRef tm
,
6243 struct pipe_debug_callback
*debug
,
6244 bool (*compile
)(struct si_screen
*,
6245 LLVMTargetMachineRef
,
6246 struct pipe_debug_callback
*,
6247 struct si_shader_part
*))
6249 struct si_shader_part
*result
;
6251 pipe_mutex_lock(sscreen
->shader_parts_mutex
);
6253 /* Find existing. */
6254 for (result
= *list
; result
; result
= result
->next
) {
6255 if (memcmp(&result
->key
, key
, sizeof(*key
)) == 0) {
6256 pipe_mutex_unlock(sscreen
->shader_parts_mutex
);
6261 /* Compile a new one. */
6262 result
= CALLOC_STRUCT(si_shader_part
);
6264 if (!compile(sscreen
, tm
, debug
, result
)) {
6266 pipe_mutex_unlock(sscreen
->shader_parts_mutex
);
6270 result
->next
= *list
;
6272 pipe_mutex_unlock(sscreen
->shader_parts_mutex
);
6277 * Create a vertex shader prolog.
6279 * The inputs are the same as VS (a lot of SGPRs and 4 VGPR system values).
6280 * All inputs are returned unmodified. The vertex load indices are
6281 * stored after them, which will used by the API VS for fetching inputs.
6283 * For example, the expected outputs for instance_divisors[] = {0, 1, 2} are:
6288 * (VertexID + BaseVertex),
6289 * (InstanceID + StartInstance),
6290 * (InstanceID / 2 + StartInstance)
6292 static bool si_compile_vs_prolog(struct si_screen
*sscreen
,
6293 LLVMTargetMachineRef tm
,
6294 struct pipe_debug_callback
*debug
,
6295 struct si_shader_part
*out
)
6297 union si_shader_part_key
*key
= &out
->key
;
6298 struct si_shader shader
= {};
6299 struct si_shader_context ctx
;
6300 struct gallivm_state
*gallivm
= &ctx
.radeon_bld
.gallivm
;
6301 LLVMTypeRef
*params
, *returns
;
6302 LLVMValueRef ret
, func
;
6303 int last_sgpr
, num_params
, num_returns
, i
;
6306 si_init_shader_ctx(&ctx
, sscreen
, &shader
, tm
);
6307 ctx
.type
= PIPE_SHADER_VERTEX
;
6308 ctx
.param_vertex_id
= key
->vs_prolog
.num_input_sgprs
;
6309 ctx
.param_instance_id
= key
->vs_prolog
.num_input_sgprs
+ 3;
6311 /* 4 preloaded VGPRs + vertex load indices as prolog outputs */
6312 params
= alloca((key
->vs_prolog
.num_input_sgprs
+ 4) *
6313 sizeof(LLVMTypeRef
));
6314 returns
= alloca((key
->vs_prolog
.num_input_sgprs
+ 4 +
6315 key
->vs_prolog
.last_input
+ 1) *
6316 sizeof(LLVMTypeRef
));
6320 /* Declare input and output SGPRs. */
6322 for (i
= 0; i
< key
->vs_prolog
.num_input_sgprs
; i
++) {
6323 params
[num_params
++] = ctx
.i32
;
6324 returns
[num_returns
++] = ctx
.i32
;
6326 last_sgpr
= num_params
- 1;
6328 /* 4 preloaded VGPRs (outputs must be floats) */
6329 for (i
= 0; i
< 4; i
++) {
6330 params
[num_params
++] = ctx
.i32
;
6331 returns
[num_returns
++] = ctx
.f32
;
6334 /* Vertex load indices. */
6335 for (i
= 0; i
<= key
->vs_prolog
.last_input
; i
++)
6336 returns
[num_returns
++] = ctx
.f32
;
6338 /* Create the function. */
6339 si_create_function(&ctx
, returns
, num_returns
, params
,
6340 num_params
, -1, last_sgpr
);
6341 func
= ctx
.radeon_bld
.main_fn
;
6343 /* Copy inputs to outputs. This should be no-op, as the registers match,
6344 * but it will prevent the compiler from overwriting them unintentionally.
6346 ret
= ctx
.return_value
;
6347 for (i
= 0; i
< key
->vs_prolog
.num_input_sgprs
; i
++) {
6348 LLVMValueRef p
= LLVMGetParam(func
, i
);
6349 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
, p
, i
, "");
6351 for (i
= num_params
- 4; i
< num_params
; i
++) {
6352 LLVMValueRef p
= LLVMGetParam(func
, i
);
6353 p
= LLVMBuildBitCast(gallivm
->builder
, p
, ctx
.f32
, "");
6354 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
, p
, i
, "");
6357 /* Compute vertex load indices from instance divisors. */
6358 for (i
= 0; i
<= key
->vs_prolog
.last_input
; i
++) {
6359 unsigned divisor
= key
->vs_prolog
.states
.instance_divisors
[i
];
6363 /* InstanceID / Divisor + StartInstance */
6364 index
= get_instance_index_for_fetch(&ctx
.radeon_bld
,
6365 SI_SGPR_START_INSTANCE
,
6368 /* VertexID + BaseVertex */
6369 index
= LLVMBuildAdd(gallivm
->builder
,
6370 LLVMGetParam(func
, ctx
.param_vertex_id
),
6371 LLVMGetParam(func
, SI_SGPR_BASE_VERTEX
), "");
6374 index
= LLVMBuildBitCast(gallivm
->builder
, index
, ctx
.f32
, "");
6375 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
, index
,
6380 LLVMBuildRet(gallivm
->builder
, ret
);
6381 radeon_llvm_finalize_module(&ctx
.radeon_bld
);
6383 if (si_compile_llvm(sscreen
, &out
->binary
, &out
->config
, tm
,
6384 gallivm
->module
, debug
, ctx
.type
,
6385 "Vertex Shader Prolog"))
6388 radeon_llvm_dispose(&ctx
.radeon_bld
);
6393 * Compile the vertex shader epilog. This is also used by the tessellation
6394 * evaluation shader compiled as VS.
6396 * The input is PrimitiveID.
6398 * If PrimitiveID is required by the pixel shader, export it.
6399 * Otherwise, do nothing.
6401 static bool si_compile_vs_epilog(struct si_screen
*sscreen
,
6402 LLVMTargetMachineRef tm
,
6403 struct pipe_debug_callback
*debug
,
6404 struct si_shader_part
*out
)
6406 union si_shader_part_key
*key
= &out
->key
;
6407 struct si_shader_context ctx
;
6408 struct gallivm_state
*gallivm
= &ctx
.radeon_bld
.gallivm
;
6409 struct lp_build_tgsi_context
*bld_base
= &ctx
.radeon_bld
.soa
.bld_base
;
6410 LLVMTypeRef params
[5];
6414 si_init_shader_ctx(&ctx
, sscreen
, NULL
, tm
);
6415 ctx
.type
= PIPE_SHADER_VERTEX
;
6417 /* Declare input VGPRs. */
6418 num_params
= key
->vs_epilog
.states
.export_prim_id
?
6419 (VS_EPILOG_PRIMID_LOC
+ 1) : 0;
6420 assert(num_params
<= ARRAY_SIZE(params
));
6422 for (i
= 0; i
< num_params
; i
++)
6423 params
[i
] = ctx
.f32
;
6425 /* Create the function. */
6426 si_create_function(&ctx
, NULL
, 0, params
, num_params
,
6430 if (key
->vs_epilog
.states
.export_prim_id
) {
6431 struct lp_build_context
*base
= &bld_base
->base
;
6432 struct lp_build_context
*uint
= &bld_base
->uint_bld
;
6433 LLVMValueRef args
[9];
6435 args
[0] = lp_build_const_int32(base
->gallivm
, 0x0); /* enabled channels */
6436 args
[1] = uint
->zero
; /* whether the EXEC mask is valid */
6437 args
[2] = uint
->zero
; /* DONE bit */
6438 args
[3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_PARAM
+
6439 key
->vs_epilog
.prim_id_param_offset
);
6440 args
[4] = uint
->zero
; /* COMPR flag (0 = 32-bit export) */
6441 args
[5] = LLVMGetParam(ctx
.radeon_bld
.main_fn
,
6442 VS_EPILOG_PRIMID_LOC
); /* X */
6443 args
[6] = uint
->undef
; /* Y */
6444 args
[7] = uint
->undef
; /* Z */
6445 args
[8] = uint
->undef
; /* W */
6447 lp_build_intrinsic(base
->gallivm
->builder
, "llvm.SI.export",
6448 LLVMVoidTypeInContext(base
->gallivm
->context
),
6453 LLVMBuildRet(gallivm
->builder
, ctx
.return_value
);
6454 radeon_llvm_finalize_module(&ctx
.radeon_bld
);
6456 if (si_compile_llvm(sscreen
, &out
->binary
, &out
->config
, tm
,
6457 gallivm
->module
, debug
, ctx
.type
,
6458 "Vertex Shader Epilog"))
6461 radeon_llvm_dispose(&ctx
.radeon_bld
);
6466 * Create & compile a vertex shader epilog. This a helper used by VS and TES.
6468 static bool si_get_vs_epilog(struct si_screen
*sscreen
,
6469 LLVMTargetMachineRef tm
,
6470 struct si_shader
*shader
,
6471 struct pipe_debug_callback
*debug
,
6472 struct si_vs_epilog_bits
*states
)
6474 union si_shader_part_key epilog_key
;
6476 memset(&epilog_key
, 0, sizeof(epilog_key
));
6477 epilog_key
.vs_epilog
.states
= *states
;
6479 /* Set up the PrimitiveID output. */
6480 if (shader
->key
.vs
.epilog
.export_prim_id
) {
6481 unsigned index
= shader
->selector
->info
.num_outputs
;
6482 unsigned offset
= shader
->info
.nr_param_exports
++;
6484 epilog_key
.vs_epilog
.prim_id_param_offset
= offset
;
6485 assert(index
< ARRAY_SIZE(shader
->info
.vs_output_param_offset
));
6486 shader
->info
.vs_output_param_offset
[index
] = offset
;
6489 shader
->epilog
= si_get_shader_part(sscreen
, &sscreen
->vs_epilogs
,
6490 &epilog_key
, tm
, debug
,
6491 si_compile_vs_epilog
);
6492 return shader
->epilog
!= NULL
;
6496 * Select and compile (or reuse) vertex shader parts (prolog & epilog).
6498 static bool si_shader_select_vs_parts(struct si_screen
*sscreen
,
6499 LLVMTargetMachineRef tm
,
6500 struct si_shader
*shader
,
6501 struct pipe_debug_callback
*debug
)
6503 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
6504 union si_shader_part_key prolog_key
;
6507 /* Get the prolog. */
6508 memset(&prolog_key
, 0, sizeof(prolog_key
));
6509 prolog_key
.vs_prolog
.states
= shader
->key
.vs
.prolog
;
6510 prolog_key
.vs_prolog
.num_input_sgprs
= shader
->info
.num_input_sgprs
;
6511 prolog_key
.vs_prolog
.last_input
= MAX2(1, info
->num_inputs
) - 1;
6513 /* The prolog is a no-op if there are no inputs. */
6514 if (info
->num_inputs
) {
6516 si_get_shader_part(sscreen
, &sscreen
->vs_prologs
,
6517 &prolog_key
, tm
, debug
,
6518 si_compile_vs_prolog
);
6519 if (!shader
->prolog
)
6523 /* Get the epilog. */
6524 if (!shader
->key
.vs
.as_es
&& !shader
->key
.vs
.as_ls
&&
6525 !si_get_vs_epilog(sscreen
, tm
, shader
, debug
,
6526 &shader
->key
.vs
.epilog
))
6529 /* Set the instanceID flag. */
6530 for (i
= 0; i
< info
->num_inputs
; i
++)
6531 if (prolog_key
.vs_prolog
.states
.instance_divisors
[i
])
6532 shader
->info
.uses_instanceid
= true;
6538 * Select and compile (or reuse) TES parts (epilog).
6540 static bool si_shader_select_tes_parts(struct si_screen
*sscreen
,
6541 LLVMTargetMachineRef tm
,
6542 struct si_shader
*shader
,
6543 struct pipe_debug_callback
*debug
)
6545 if (shader
->key
.tes
.as_es
)
6548 /* TES compiled as VS. */
6549 return si_get_vs_epilog(sscreen
, tm
, shader
, debug
,
6550 &shader
->key
.tes
.epilog
);
6554 * Compile the TCS epilog. This writes tesselation factors to memory based on
6555 * the output primitive type of the tesselator (determined by TES).
6557 static bool si_compile_tcs_epilog(struct si_screen
*sscreen
,
6558 LLVMTargetMachineRef tm
,
6559 struct pipe_debug_callback
*debug
,
6560 struct si_shader_part
*out
)
6562 union si_shader_part_key
*key
= &out
->key
;
6563 struct si_shader shader
= {};
6564 struct si_shader_context ctx
;
6565 struct gallivm_state
*gallivm
= &ctx
.radeon_bld
.gallivm
;
6566 struct lp_build_tgsi_context
*bld_base
= &ctx
.radeon_bld
.soa
.bld_base
;
6567 LLVMTypeRef params
[16];
6569 int last_array_pointer
, last_sgpr
, num_params
;
6572 si_init_shader_ctx(&ctx
, sscreen
, &shader
, tm
);
6573 ctx
.type
= PIPE_SHADER_TESS_CTRL
;
6574 shader
.key
.tcs
.epilog
= key
->tcs_epilog
.states
;
6576 /* Declare inputs. Only RW_BUFFERS and TESS_FACTOR_OFFSET are used. */
6577 params
[SI_PARAM_RW_BUFFERS
] = const_array(ctx
.v16i8
, SI_NUM_RW_BUFFERS
);
6578 last_array_pointer
= SI_PARAM_RW_BUFFERS
;
6579 params
[SI_PARAM_CONST_BUFFERS
] = ctx
.i64
;
6580 params
[SI_PARAM_SAMPLERS
] = ctx
.i64
;
6581 params
[SI_PARAM_IMAGES
] = ctx
.i64
;
6582 params
[SI_PARAM_SHADER_BUFFERS
] = ctx
.i64
;
6583 params
[SI_PARAM_TCS_OUT_OFFSETS
] = ctx
.i32
;
6584 params
[SI_PARAM_TCS_OUT_LAYOUT
] = ctx
.i32
;
6585 params
[SI_PARAM_TCS_IN_LAYOUT
] = ctx
.i32
;
6586 params
[SI_PARAM_TESS_FACTOR_OFFSET
] = ctx
.i32
;
6587 last_sgpr
= SI_PARAM_TESS_FACTOR_OFFSET
;
6588 num_params
= last_sgpr
+ 1;
6590 params
[num_params
++] = ctx
.i32
; /* patch index within the wave (REL_PATCH_ID) */
6591 params
[num_params
++] = ctx
.i32
; /* invocation ID within the patch */
6592 params
[num_params
++] = ctx
.i32
; /* LDS offset where tess factors should be loaded from */
6594 /* Create the function. */
6595 si_create_function(&ctx
, NULL
, 0, params
, num_params
,
6596 last_array_pointer
, last_sgpr
);
6597 declare_tess_lds(&ctx
);
6598 func
= ctx
.radeon_bld
.main_fn
;
6600 si_write_tess_factors(bld_base
,
6601 LLVMGetParam(func
, last_sgpr
+ 1),
6602 LLVMGetParam(func
, last_sgpr
+ 2),
6603 LLVMGetParam(func
, last_sgpr
+ 3));
6606 LLVMBuildRet(gallivm
->builder
, ctx
.return_value
);
6607 radeon_llvm_finalize_module(&ctx
.radeon_bld
);
6609 if (si_compile_llvm(sscreen
, &out
->binary
, &out
->config
, tm
,
6610 gallivm
->module
, debug
, ctx
.type
,
6611 "Tessellation Control Shader Epilog"))
6614 radeon_llvm_dispose(&ctx
.radeon_bld
);
6619 * Select and compile (or reuse) TCS parts (epilog).
6621 static bool si_shader_select_tcs_parts(struct si_screen
*sscreen
,
6622 LLVMTargetMachineRef tm
,
6623 struct si_shader
*shader
,
6624 struct pipe_debug_callback
*debug
)
6626 union si_shader_part_key epilog_key
;
6628 /* Get the epilog. */
6629 memset(&epilog_key
, 0, sizeof(epilog_key
));
6630 epilog_key
.tcs_epilog
.states
= shader
->key
.tcs
.epilog
;
6632 shader
->epilog
= si_get_shader_part(sscreen
, &sscreen
->tcs_epilogs
,
6633 &epilog_key
, tm
, debug
,
6634 si_compile_tcs_epilog
);
6635 return shader
->epilog
!= NULL
;
6639 * Compile the pixel shader prolog. This handles:
6640 * - two-side color selection and interpolation
6641 * - overriding interpolation parameters for the API PS
6642 * - polygon stippling
6644 * All preloaded SGPRs and VGPRs are passed through unmodified unless they are
6645 * overriden by other states. (e.g. per-sample interpolation)
6646 * Interpolated colors are stored after the preloaded VGPRs.
6648 static bool si_compile_ps_prolog(struct si_screen
*sscreen
,
6649 LLVMTargetMachineRef tm
,
6650 struct pipe_debug_callback
*debug
,
6651 struct si_shader_part
*out
)
6653 union si_shader_part_key
*key
= &out
->key
;
6654 struct si_shader shader
= {};
6655 struct si_shader_context ctx
;
6656 struct gallivm_state
*gallivm
= &ctx
.radeon_bld
.gallivm
;
6657 LLVMTypeRef
*params
;
6658 LLVMValueRef ret
, func
;
6659 int last_sgpr
, num_params
, num_returns
, i
, num_color_channels
;
6662 si_init_shader_ctx(&ctx
, sscreen
, &shader
, tm
);
6663 ctx
.type
= PIPE_SHADER_FRAGMENT
;
6664 shader
.key
.ps
.prolog
= key
->ps_prolog
.states
;
6666 /* Number of inputs + 8 color elements. */
6667 params
= alloca((key
->ps_prolog
.num_input_sgprs
+
6668 key
->ps_prolog
.num_input_vgprs
+ 8) *
6669 sizeof(LLVMTypeRef
));
6671 /* Declare inputs. */
6673 for (i
= 0; i
< key
->ps_prolog
.num_input_sgprs
; i
++)
6674 params
[num_params
++] = ctx
.i32
;
6675 last_sgpr
= num_params
- 1;
6677 for (i
= 0; i
< key
->ps_prolog
.num_input_vgprs
; i
++)
6678 params
[num_params
++] = ctx
.f32
;
6680 /* Declare outputs (same as inputs + add colors if needed) */
6681 num_returns
= num_params
;
6682 num_color_channels
= util_bitcount(key
->ps_prolog
.colors_read
);
6683 for (i
= 0; i
< num_color_channels
; i
++)
6684 params
[num_returns
++] = ctx
.f32
;
6686 /* Create the function. */
6687 si_create_function(&ctx
, params
, num_returns
, params
,
6688 num_params
, -1, last_sgpr
);
6689 func
= ctx
.radeon_bld
.main_fn
;
6691 /* Copy inputs to outputs. This should be no-op, as the registers match,
6692 * but it will prevent the compiler from overwriting them unintentionally.
6694 ret
= ctx
.return_value
;
6695 for (i
= 0; i
< num_params
; i
++) {
6696 LLVMValueRef p
= LLVMGetParam(func
, i
);
6697 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
, p
, i
, "");
6700 /* Polygon stippling. */
6701 if (key
->ps_prolog
.states
.poly_stipple
) {
6702 /* POS_FIXED_PT is always last. */
6703 unsigned pos
= key
->ps_prolog
.num_input_sgprs
+
6704 key
->ps_prolog
.num_input_vgprs
- 1;
6705 LLVMValueRef ptr
[2], list
;
6707 /* Get the pointer to rw buffers. */
6708 ptr
[0] = LLVMGetParam(func
, SI_SGPR_RW_BUFFERS
);
6709 ptr
[1] = LLVMGetParam(func
, SI_SGPR_RW_BUFFERS_HI
);
6710 list
= lp_build_gather_values(gallivm
, ptr
, 2);
6711 list
= LLVMBuildBitCast(gallivm
->builder
, list
, ctx
.i64
, "");
6712 list
= LLVMBuildIntToPtr(gallivm
->builder
, list
,
6713 const_array(ctx
.v16i8
, SI_NUM_RW_BUFFERS
), "");
6715 si_llvm_emit_polygon_stipple(&ctx
, list
, pos
);
6718 /* Interpolate colors. */
6719 for (i
= 0; i
< 2; i
++) {
6720 unsigned writemask
= (key
->ps_prolog
.colors_read
>> (i
* 4)) & 0xf;
6721 unsigned face_vgpr
= key
->ps_prolog
.num_input_sgprs
+
6722 key
->ps_prolog
.face_vgpr_index
;
6723 LLVMValueRef interp
[2], color
[4];
6724 LLVMValueRef interp_ij
= NULL
, prim_mask
= NULL
, face
= NULL
;
6729 /* If the interpolation qualifier is not CONSTANT (-1). */
6730 if (key
->ps_prolog
.color_interp_vgpr_index
[i
] != -1) {
6731 unsigned interp_vgpr
= key
->ps_prolog
.num_input_sgprs
+
6732 key
->ps_prolog
.color_interp_vgpr_index
[i
];
6734 interp
[0] = LLVMGetParam(func
, interp_vgpr
);
6735 interp
[1] = LLVMGetParam(func
, interp_vgpr
+ 1);
6736 interp_ij
= lp_build_gather_values(gallivm
, interp
, 2);
6737 interp_ij
= LLVMBuildBitCast(gallivm
->builder
, interp_ij
,
6741 /* Use the absolute location of the input. */
6742 prim_mask
= LLVMGetParam(func
, SI_PS_NUM_USER_SGPR
);
6744 if (key
->ps_prolog
.states
.color_two_side
) {
6745 face
= LLVMGetParam(func
, face_vgpr
);
6746 face
= LLVMBuildBitCast(gallivm
->builder
, face
, ctx
.i32
, "");
6749 interp_fs_input(&ctx
,
6750 key
->ps_prolog
.color_attr_index
[i
],
6751 TGSI_SEMANTIC_COLOR
, i
,
6752 key
->ps_prolog
.num_interp_inputs
,
6753 key
->ps_prolog
.colors_read
, interp_ij
,
6754 prim_mask
, face
, color
);
6757 unsigned chan
= u_bit_scan(&writemask
);
6758 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
, color
[chan
],
6763 /* Force per-sample interpolation. */
6764 if (key
->ps_prolog
.states
.force_persample_interp
) {
6765 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
6766 LLVMValueRef persp_sample
[2], linear_sample
[2];
6768 /* Read PERSP_SAMPLE. */
6769 for (i
= 0; i
< 2; i
++)
6770 persp_sample
[i
] = LLVMGetParam(func
, base
+ i
);
6771 /* Overwrite PERSP_CENTER. */
6772 for (i
= 0; i
< 2; i
++)
6773 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
,
6774 persp_sample
[i
], base
+ 2 + i
, "");
6775 /* Overwrite PERSP_CENTROID. */
6776 for (i
= 0; i
< 2; i
++)
6777 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
,
6778 persp_sample
[i
], base
+ 4 + i
, "");
6779 /* Read LINEAR_SAMPLE. */
6780 for (i
= 0; i
< 2; i
++)
6781 linear_sample
[i
] = LLVMGetParam(func
, base
+ 6 + i
);
6782 /* Overwrite LINEAR_CENTER. */
6783 for (i
= 0; i
< 2; i
++)
6784 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
,
6785 linear_sample
[i
], base
+ 8 + i
, "");
6786 /* Overwrite LINEAR_CENTROID. */
6787 for (i
= 0; i
< 2; i
++)
6788 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
,
6789 linear_sample
[i
], base
+ 10 + i
, "");
6793 LLVMBuildRet(gallivm
->builder
, ret
);
6794 radeon_llvm_finalize_module(&ctx
.radeon_bld
);
6796 if (si_compile_llvm(sscreen
, &out
->binary
, &out
->config
, tm
,
6797 gallivm
->module
, debug
, ctx
.type
,
6798 "Fragment Shader Prolog"))
6801 radeon_llvm_dispose(&ctx
.radeon_bld
);
6806 * Compile the pixel shader epilog. This handles everything that must be
6807 * emulated for pixel shader exports. (alpha-test, format conversions, etc)
6809 static bool si_compile_ps_epilog(struct si_screen
*sscreen
,
6810 LLVMTargetMachineRef tm
,
6811 struct pipe_debug_callback
*debug
,
6812 struct si_shader_part
*out
)
6814 union si_shader_part_key
*key
= &out
->key
;
6815 struct si_shader shader
= {};
6816 struct si_shader_context ctx
;
6817 struct gallivm_state
*gallivm
= &ctx
.radeon_bld
.gallivm
;
6818 struct lp_build_tgsi_context
*bld_base
= &ctx
.radeon_bld
.soa
.bld_base
;
6819 LLVMTypeRef params
[16+8*4+3];
6820 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
6821 int last_array_pointer
, last_sgpr
, num_params
, i
;
6824 si_init_shader_ctx(&ctx
, sscreen
, &shader
, tm
);
6825 ctx
.type
= PIPE_SHADER_FRAGMENT
;
6826 shader
.key
.ps
.epilog
= key
->ps_epilog
.states
;
6828 /* Declare input SGPRs. */
6829 params
[SI_PARAM_RW_BUFFERS
] = ctx
.i64
;
6830 params
[SI_PARAM_CONST_BUFFERS
] = ctx
.i64
;
6831 params
[SI_PARAM_SAMPLERS
] = ctx
.i64
;
6832 params
[SI_PARAM_IMAGES
] = ctx
.i64
;
6833 params
[SI_PARAM_SHADER_BUFFERS
] = ctx
.i64
;
6834 params
[SI_PARAM_ALPHA_REF
] = ctx
.f32
;
6835 last_array_pointer
= -1;
6836 last_sgpr
= SI_PARAM_ALPHA_REF
;
6838 /* Declare input VGPRs. */
6839 num_params
= (last_sgpr
+ 1) +
6840 util_bitcount(key
->ps_epilog
.colors_written
) * 4 +
6841 key
->ps_epilog
.writes_z
+
6842 key
->ps_epilog
.writes_stencil
+
6843 key
->ps_epilog
.writes_samplemask
;
6845 num_params
= MAX2(num_params
,
6846 last_sgpr
+ 1 + PS_EPILOG_SAMPLEMASK_MIN_LOC
+ 1);
6848 assert(num_params
<= ARRAY_SIZE(params
));
6850 for (i
= last_sgpr
+ 1; i
< num_params
; i
++)
6851 params
[i
] = ctx
.f32
;
6853 /* Create the function. */
6854 si_create_function(&ctx
, NULL
, 0, params
, num_params
,
6855 last_array_pointer
, last_sgpr
);
6856 /* Disable elimination of unused inputs. */
6857 radeon_llvm_add_attribute(ctx
.radeon_bld
.main_fn
,
6858 "InitialPSInputAddr", 0xffffff);
6860 /* Process colors. */
6861 unsigned vgpr
= last_sgpr
+ 1;
6862 unsigned colors_written
= key
->ps_epilog
.colors_written
;
6863 int last_color_export
= -1;
6865 /* Find the last color export. */
6866 if (!key
->ps_epilog
.writes_z
&&
6867 !key
->ps_epilog
.writes_stencil
&&
6868 !key
->ps_epilog
.writes_samplemask
) {
6869 unsigned spi_format
= key
->ps_epilog
.states
.spi_shader_col_format
;
6871 /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
6872 if (colors_written
== 0x1 && key
->ps_epilog
.states
.last_cbuf
> 0) {
6873 /* Just set this if any of the colorbuffers are enabled. */
6875 ((1llu << (4 * (key
->ps_epilog
.states
.last_cbuf
+ 1))) - 1))
6876 last_color_export
= 0;
6878 for (i
= 0; i
< 8; i
++)
6879 if (colors_written
& (1 << i
) &&
6880 (spi_format
>> (i
* 4)) & 0xf)
6881 last_color_export
= i
;
6885 while (colors_written
) {
6886 LLVMValueRef color
[4];
6887 int mrt
= u_bit_scan(&colors_written
);
6889 for (i
= 0; i
< 4; i
++)
6890 color
[i
] = LLVMGetParam(ctx
.radeon_bld
.main_fn
, vgpr
++);
6892 si_export_mrt_color(bld_base
, color
, mrt
,
6894 mrt
== last_color_export
);
6897 /* Process depth, stencil, samplemask. */
6898 if (key
->ps_epilog
.writes_z
)
6899 depth
= LLVMGetParam(ctx
.radeon_bld
.main_fn
, vgpr
++);
6900 if (key
->ps_epilog
.writes_stencil
)
6901 stencil
= LLVMGetParam(ctx
.radeon_bld
.main_fn
, vgpr
++);
6902 if (key
->ps_epilog
.writes_samplemask
)
6903 samplemask
= LLVMGetParam(ctx
.radeon_bld
.main_fn
, vgpr
++);
6905 if (depth
|| stencil
|| samplemask
)
6906 si_export_mrt_z(bld_base
, depth
, stencil
, samplemask
);
6907 else if (last_color_export
== -1)
6908 si_export_null(bld_base
);
6911 LLVMBuildRetVoid(gallivm
->builder
);
6912 radeon_llvm_finalize_module(&ctx
.radeon_bld
);
6914 if (si_compile_llvm(sscreen
, &out
->binary
, &out
->config
, tm
,
6915 gallivm
->module
, debug
, ctx
.type
,
6916 "Fragment Shader Epilog"))
6919 radeon_llvm_dispose(&ctx
.radeon_bld
);
6924 * Select and compile (or reuse) pixel shader parts (prolog & epilog).
6926 static bool si_shader_select_ps_parts(struct si_screen
*sscreen
,
6927 LLVMTargetMachineRef tm
,
6928 struct si_shader
*shader
,
6929 struct pipe_debug_callback
*debug
)
6931 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
6932 union si_shader_part_key prolog_key
;
6933 union si_shader_part_key epilog_key
;
6936 /* Get the prolog. */
6937 memset(&prolog_key
, 0, sizeof(prolog_key
));
6938 prolog_key
.ps_prolog
.states
= shader
->key
.ps
.prolog
;
6939 prolog_key
.ps_prolog
.colors_read
= info
->colors_read
;
6940 prolog_key
.ps_prolog
.num_input_sgprs
= shader
->info
.num_input_sgprs
;
6941 prolog_key
.ps_prolog
.num_input_vgprs
= shader
->info
.num_input_vgprs
;
6943 if (info
->colors_read
) {
6944 unsigned *color
= shader
->selector
->color_attr_index
;
6946 if (shader
->key
.ps
.prolog
.color_two_side
) {
6947 /* BCOLORs are stored after the last input. */
6948 prolog_key
.ps_prolog
.num_interp_inputs
= info
->num_inputs
;
6949 prolog_key
.ps_prolog
.face_vgpr_index
= shader
->info
.face_vgpr_index
;
6950 shader
->config
.spi_ps_input_ena
|= S_0286CC_FRONT_FACE_ENA(1);
6953 for (i
= 0; i
< 2; i
++) {
6954 unsigned location
= info
->input_interpolate_loc
[color
[i
]];
6956 if (!(info
->colors_read
& (0xf << i
*4)))
6959 prolog_key
.ps_prolog
.color_attr_index
[i
] = color
[i
];
6961 /* Force per-sample interpolation for the colors here. */
6962 if (shader
->key
.ps
.prolog
.force_persample_interp
)
6963 location
= TGSI_INTERPOLATE_LOC_SAMPLE
;
6965 switch (info
->input_interpolate
[color
[i
]]) {
6966 case TGSI_INTERPOLATE_CONSTANT
:
6967 prolog_key
.ps_prolog
.color_interp_vgpr_index
[i
] = -1;
6969 case TGSI_INTERPOLATE_PERSPECTIVE
:
6970 case TGSI_INTERPOLATE_COLOR
:
6972 case TGSI_INTERPOLATE_LOC_SAMPLE
:
6973 prolog_key
.ps_prolog
.color_interp_vgpr_index
[i
] = 0;
6974 shader
->config
.spi_ps_input_ena
|=
6975 S_0286CC_PERSP_SAMPLE_ENA(1);
6977 case TGSI_INTERPOLATE_LOC_CENTER
:
6978 prolog_key
.ps_prolog
.color_interp_vgpr_index
[i
] = 2;
6979 shader
->config
.spi_ps_input_ena
|=
6980 S_0286CC_PERSP_CENTER_ENA(1);
6982 case TGSI_INTERPOLATE_LOC_CENTROID
:
6983 prolog_key
.ps_prolog
.color_interp_vgpr_index
[i
] = 4;
6984 shader
->config
.spi_ps_input_ena
|=
6985 S_0286CC_PERSP_CENTROID_ENA(1);
6991 case TGSI_INTERPOLATE_LINEAR
:
6993 case TGSI_INTERPOLATE_LOC_SAMPLE
:
6994 prolog_key
.ps_prolog
.color_interp_vgpr_index
[i
] = 6;
6995 shader
->config
.spi_ps_input_ena
|=
6996 S_0286CC_LINEAR_SAMPLE_ENA(1);
6998 case TGSI_INTERPOLATE_LOC_CENTER
:
6999 prolog_key
.ps_prolog
.color_interp_vgpr_index
[i
] = 8;
7000 shader
->config
.spi_ps_input_ena
|=
7001 S_0286CC_LINEAR_CENTER_ENA(1);
7003 case TGSI_INTERPOLATE_LOC_CENTROID
:
7004 prolog_key
.ps_prolog
.color_interp_vgpr_index
[i
] = 10;
7005 shader
->config
.spi_ps_input_ena
|=
7006 S_0286CC_LINEAR_CENTROID_ENA(1);
7018 /* The prolog is a no-op if these aren't set. */
7019 if (prolog_key
.ps_prolog
.colors_read
||
7020 prolog_key
.ps_prolog
.states
.force_persample_interp
||
7021 prolog_key
.ps_prolog
.states
.poly_stipple
) {
7023 si_get_shader_part(sscreen
, &sscreen
->ps_prologs
,
7024 &prolog_key
, tm
, debug
,
7025 si_compile_ps_prolog
);
7026 if (!shader
->prolog
)
7030 /* Get the epilog. */
7031 memset(&epilog_key
, 0, sizeof(epilog_key
));
7032 epilog_key
.ps_epilog
.colors_written
= info
->colors_written
;
7033 epilog_key
.ps_epilog
.writes_z
= info
->writes_z
;
7034 epilog_key
.ps_epilog
.writes_stencil
= info
->writes_stencil
;
7035 epilog_key
.ps_epilog
.writes_samplemask
= info
->writes_samplemask
;
7036 epilog_key
.ps_epilog
.states
= shader
->key
.ps
.epilog
;
7039 si_get_shader_part(sscreen
, &sscreen
->ps_epilogs
,
7040 &epilog_key
, tm
, debug
,
7041 si_compile_ps_epilog
);
7042 if (!shader
->epilog
)
7045 /* Enable POS_FIXED_PT if polygon stippling is enabled. */
7046 if (shader
->key
.ps
.prolog
.poly_stipple
) {
7047 shader
->config
.spi_ps_input_ena
|= S_0286CC_POS_FIXED_PT_ENA(1);
7048 assert(G_0286CC_POS_FIXED_PT_ENA(shader
->config
.spi_ps_input_addr
));
7051 /* Set up the enable bits for per-sample shading if needed. */
7052 if (shader
->key
.ps
.prolog
.force_persample_interp
) {
7053 if (G_0286CC_PERSP_CENTER_ENA(shader
->config
.spi_ps_input_ena
) ||
7054 G_0286CC_PERSP_CENTROID_ENA(shader
->config
.spi_ps_input_ena
)) {
7055 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_CENTER_ENA
;
7056 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_CENTROID_ENA
;
7057 shader
->config
.spi_ps_input_ena
|= S_0286CC_PERSP_SAMPLE_ENA(1);
7059 if (G_0286CC_LINEAR_CENTER_ENA(shader
->config
.spi_ps_input_ena
) ||
7060 G_0286CC_LINEAR_CENTROID_ENA(shader
->config
.spi_ps_input_ena
)) {
7061 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_CENTER_ENA
;
7062 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_CENTROID_ENA
;
7063 shader
->config
.spi_ps_input_ena
|= S_0286CC_LINEAR_SAMPLE_ENA(1);
7067 /* POW_W_FLOAT requires that one of the perspective weights is enabled. */
7068 if (G_0286CC_POS_W_FLOAT_ENA(shader
->config
.spi_ps_input_ena
) &&
7069 !(shader
->config
.spi_ps_input_ena
& 0xf)) {
7070 shader
->config
.spi_ps_input_ena
|= S_0286CC_PERSP_CENTER_ENA(1);
7071 assert(G_0286CC_PERSP_CENTER_ENA(shader
->config
.spi_ps_input_addr
));
7074 /* At least one pair of interpolation weights must be enabled. */
7075 if (!(shader
->config
.spi_ps_input_ena
& 0x7f)) {
7076 shader
->config
.spi_ps_input_ena
|= S_0286CC_LINEAR_CENTER_ENA(1);
7077 assert(G_0286CC_LINEAR_CENTER_ENA(shader
->config
.spi_ps_input_addr
));
7080 /* The sample mask input is always enabled, because the API shader always
7081 * passes it through to the epilog. Disable it here if it's unused.
7083 if (!shader
->key
.ps
.epilog
.poly_line_smoothing
&&
7084 !shader
->selector
->info
.reads_samplemask
)
7085 shader
->config
.spi_ps_input_ena
&= C_0286CC_SAMPLE_COVERAGE_ENA
;
7090 static void si_fix_num_sgprs(struct si_shader
*shader
)
7092 unsigned min_sgprs
= shader
->info
.num_input_sgprs
+ 2; /* VCC */
7094 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
, min_sgprs
);
7097 int si_shader_create(struct si_screen
*sscreen
, LLVMTargetMachineRef tm
,
7098 struct si_shader
*shader
,
7099 struct pipe_debug_callback
*debug
)
7101 struct si_shader
*mainp
= shader
->selector
->main_shader_part
;
7104 /* LS, ES, VS are compiled on demand if the main part hasn't been
7105 * compiled for that stage.
7108 (shader
->selector
->type
== PIPE_SHADER_VERTEX
&&
7109 (shader
->key
.vs
.as_es
!= mainp
->key
.vs
.as_es
||
7110 shader
->key
.vs
.as_ls
!= mainp
->key
.vs
.as_ls
)) ||
7111 (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
&&
7112 shader
->key
.tes
.as_es
!= mainp
->key
.tes
.as_es
) ||
7113 shader
->selector
->type
== PIPE_SHADER_COMPUTE
) {
7114 /* Monolithic shader (compiled as a whole, has many variants,
7115 * may take a long time to compile).
7117 r
= si_compile_tgsi_shader(sscreen
, tm
, shader
, true, debug
);
7121 /* The shader consists of 2-3 parts:
7123 * - the middle part is the user shader, it has 1 variant only
7124 * and it was compiled during the creation of the shader
7126 * - the prolog part is inserted at the beginning
7127 * - the epilog part is inserted at the end
7129 * The prolog and epilog have many (but simple) variants.
7132 /* Copy the compiled TGSI shader data over. */
7133 shader
->is_binary_shared
= true;
7134 shader
->binary
= mainp
->binary
;
7135 shader
->config
= mainp
->config
;
7136 shader
->info
.num_input_sgprs
= mainp
->info
.num_input_sgprs
;
7137 shader
->info
.num_input_vgprs
= mainp
->info
.num_input_vgprs
;
7138 shader
->info
.face_vgpr_index
= mainp
->info
.face_vgpr_index
;
7139 memcpy(shader
->info
.vs_output_param_offset
,
7140 mainp
->info
.vs_output_param_offset
,
7141 sizeof(mainp
->info
.vs_output_param_offset
));
7142 shader
->info
.uses_instanceid
= mainp
->info
.uses_instanceid
;
7143 shader
->info
.nr_pos_exports
= mainp
->info
.nr_pos_exports
;
7144 shader
->info
.nr_param_exports
= mainp
->info
.nr_param_exports
;
7146 /* Select prologs and/or epilogs. */
7147 switch (shader
->selector
->type
) {
7148 case PIPE_SHADER_VERTEX
:
7149 if (!si_shader_select_vs_parts(sscreen
, tm
, shader
, debug
))
7152 case PIPE_SHADER_TESS_CTRL
:
7153 if (!si_shader_select_tcs_parts(sscreen
, tm
, shader
, debug
))
7156 case PIPE_SHADER_TESS_EVAL
:
7157 if (!si_shader_select_tes_parts(sscreen
, tm
, shader
, debug
))
7160 case PIPE_SHADER_FRAGMENT
:
7161 if (!si_shader_select_ps_parts(sscreen
, tm
, shader
, debug
))
7164 /* Make sure we have at least as many VGPRs as there
7165 * are allocated inputs.
7167 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
7168 shader
->info
.num_input_vgprs
);
7172 /* Update SGPR and VGPR counts. */
7173 if (shader
->prolog
) {
7174 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
,
7175 shader
->prolog
->config
.num_sgprs
);
7176 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
7177 shader
->prolog
->config
.num_vgprs
);
7179 if (shader
->epilog
) {
7180 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
,
7181 shader
->epilog
->config
.num_sgprs
);
7182 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
7183 shader
->epilog
->config
.num_vgprs
);
7187 si_fix_num_sgprs(shader
);
7188 si_shader_dump(sscreen
, shader
, debug
, shader
->selector
->info
.processor
,
7192 r
= si_shader_binary_upload(sscreen
, shader
);
7194 fprintf(stderr
, "LLVM failed to upload shader\n");
7201 void si_shader_destroy(struct si_shader
*shader
)
7203 if (shader
->gs_copy_shader
) {
7204 si_shader_destroy(shader
->gs_copy_shader
);
7205 FREE(shader
->gs_copy_shader
);
7208 if (shader
->scratch_bo
)
7209 r600_resource_reference(&shader
->scratch_bo
, NULL
);
7211 r600_resource_reference(&shader
->bo
, NULL
);
7213 if (!shader
->is_binary_shared
)
7214 radeon_shader_binary_clean(&shader
->binary
);