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
; /* TGSI_PROCESSOR_* 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_USER_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 TGSI_PROCESSOR_TESS_CTRL
:
244 return unpack_param(ctx
, SI_PARAM_REL_IDS
, 0, 8);
246 case TGSI_PROCESSOR_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
== TGSI_PROCESSOR_VERTEX
)
281 return unpack_param(ctx
, SI_PARAM_LS_OUT_LAYOUT
, 0, 13);
282 else if (ctx
->type
== TGSI_PROCESSOR_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 TGSI_PROCESSOR_VERTEX
:
503 return LLVMGetParam(ctx
->radeon_bld
.main_fn
,
504 ctx
->param_vs_prim_id
);
505 case TGSI_PROCESSOR_TESS_CTRL
:
506 return LLVMGetParam(ctx
->radeon_bld
.main_fn
,
508 case TGSI_PROCESSOR_TESS_EVAL
:
509 return LLVMGetParam(ctx
->radeon_bld
.main_fn
,
510 ctx
->param_tes_patch_id
);
511 case TGSI_PROCESSOR_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 * Load a dword from a constant buffer.
1122 static LLVMValueRef
buffer_load_const(LLVMBuilderRef builder
, LLVMValueRef resource
,
1123 LLVMValueRef offset
, LLVMTypeRef return_type
)
1125 LLVMValueRef args
[2] = {resource
, offset
};
1127 return lp_build_intrinsic(builder
, "llvm.SI.load.const", return_type
, args
, 2,
1128 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
1131 static LLVMValueRef
load_sample_position(struct radeon_llvm_context
*radeon_bld
, LLVMValueRef sample_id
)
1133 struct si_shader_context
*ctx
=
1134 si_shader_context(&radeon_bld
->soa
.bld_base
);
1135 struct lp_build_context
*uint_bld
= &radeon_bld
->soa
.bld_base
.uint_bld
;
1136 struct gallivm_state
*gallivm
= &radeon_bld
->gallivm
;
1137 LLVMBuilderRef builder
= gallivm
->builder
;
1138 LLVMValueRef desc
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_CONST_BUFFERS
);
1139 LLVMValueRef buf_index
= lp_build_const_int32(gallivm
, SI_DRIVER_STATE_CONST_BUF
);
1140 LLVMValueRef resource
= build_indexed_load_const(ctx
, desc
, buf_index
);
1142 /* offset = sample_id * 8 (8 = 2 floats containing samplepos.xy) */
1143 LLVMValueRef offset0
= lp_build_mul_imm(uint_bld
, sample_id
, 8);
1144 LLVMValueRef offset1
= LLVMBuildAdd(builder
, offset0
, lp_build_const_int32(gallivm
, 4), "");
1146 LLVMValueRef pos
[4] = {
1147 buffer_load_const(builder
, resource
, offset0
, ctx
->f32
),
1148 buffer_load_const(builder
, resource
, offset1
, ctx
->f32
),
1149 lp_build_const_float(gallivm
, 0),
1150 lp_build_const_float(gallivm
, 0)
1153 return lp_build_gather_values(gallivm
, pos
, 4);
1156 static void declare_system_value(
1157 struct radeon_llvm_context
*radeon_bld
,
1159 const struct tgsi_full_declaration
*decl
)
1161 struct si_shader_context
*ctx
=
1162 si_shader_context(&radeon_bld
->soa
.bld_base
);
1163 struct lp_build_context
*bld
= &radeon_bld
->soa
.bld_base
.base
;
1164 struct gallivm_state
*gallivm
= &radeon_bld
->gallivm
;
1165 LLVMValueRef value
= 0;
1167 switch (decl
->Semantic
.Name
) {
1168 case TGSI_SEMANTIC_INSTANCEID
:
1169 value
= LLVMGetParam(radeon_bld
->main_fn
,
1170 ctx
->param_instance_id
);
1173 case TGSI_SEMANTIC_VERTEXID
:
1174 value
= LLVMBuildAdd(gallivm
->builder
,
1175 LLVMGetParam(radeon_bld
->main_fn
,
1176 ctx
->param_vertex_id
),
1177 LLVMGetParam(radeon_bld
->main_fn
,
1178 SI_PARAM_BASE_VERTEX
), "");
1181 case TGSI_SEMANTIC_VERTEXID_NOBASE
:
1182 value
= LLVMGetParam(radeon_bld
->main_fn
,
1183 ctx
->param_vertex_id
);
1186 case TGSI_SEMANTIC_BASEVERTEX
:
1187 value
= LLVMGetParam(radeon_bld
->main_fn
,
1188 SI_PARAM_BASE_VERTEX
);
1191 case TGSI_SEMANTIC_INVOCATIONID
:
1192 if (ctx
->type
== TGSI_PROCESSOR_TESS_CTRL
)
1193 value
= unpack_param(ctx
, SI_PARAM_REL_IDS
, 8, 5);
1194 else if (ctx
->type
== TGSI_PROCESSOR_GEOMETRY
)
1195 value
= LLVMGetParam(radeon_bld
->main_fn
,
1196 SI_PARAM_GS_INSTANCE_ID
);
1198 assert(!"INVOCATIONID not implemented");
1201 case TGSI_SEMANTIC_POSITION
:
1203 LLVMValueRef pos
[4] = {
1204 LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_POS_X_FLOAT
),
1205 LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_POS_Y_FLOAT
),
1206 LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_POS_Z_FLOAT
),
1207 lp_build_emit_llvm_unary(&radeon_bld
->soa
.bld_base
, TGSI_OPCODE_RCP
,
1208 LLVMGetParam(radeon_bld
->main_fn
,
1209 SI_PARAM_POS_W_FLOAT
)),
1211 value
= lp_build_gather_values(gallivm
, pos
, 4);
1215 case TGSI_SEMANTIC_FACE
:
1216 value
= LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_FRONT_FACE
);
1219 case TGSI_SEMANTIC_SAMPLEID
:
1220 value
= get_sample_id(radeon_bld
);
1223 case TGSI_SEMANTIC_SAMPLEPOS
: {
1224 LLVMValueRef pos
[4] = {
1225 LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_POS_X_FLOAT
),
1226 LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_POS_Y_FLOAT
),
1227 lp_build_const_float(gallivm
, 0),
1228 lp_build_const_float(gallivm
, 0)
1230 pos
[0] = lp_build_emit_llvm_unary(&radeon_bld
->soa
.bld_base
,
1231 TGSI_OPCODE_FRC
, pos
[0]);
1232 pos
[1] = lp_build_emit_llvm_unary(&radeon_bld
->soa
.bld_base
,
1233 TGSI_OPCODE_FRC
, pos
[1]);
1234 value
= lp_build_gather_values(gallivm
, pos
, 4);
1238 case TGSI_SEMANTIC_SAMPLEMASK
:
1239 /* This can only occur with the OpenGL Core profile, which
1240 * doesn't support smoothing.
1242 value
= LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_SAMPLE_COVERAGE
);
1245 case TGSI_SEMANTIC_TESSCOORD
:
1247 LLVMValueRef coord
[4] = {
1248 LLVMGetParam(radeon_bld
->main_fn
, ctx
->param_tes_u
),
1249 LLVMGetParam(radeon_bld
->main_fn
, ctx
->param_tes_v
),
1254 /* For triangles, the vector should be (u, v, 1-u-v). */
1255 if (ctx
->shader
->selector
->info
.properties
[TGSI_PROPERTY_TES_PRIM_MODE
] ==
1256 PIPE_PRIM_TRIANGLES
)
1257 coord
[2] = lp_build_sub(bld
, bld
->one
,
1258 lp_build_add(bld
, coord
[0], coord
[1]));
1260 value
= lp_build_gather_values(gallivm
, coord
, 4);
1264 case TGSI_SEMANTIC_VERTICESIN
:
1265 value
= unpack_param(ctx
, SI_PARAM_TCS_OUT_LAYOUT
, 26, 6);
1268 case TGSI_SEMANTIC_TESSINNER
:
1269 case TGSI_SEMANTIC_TESSOUTER
:
1271 LLVMValueRef dw_addr
;
1272 int param
= si_shader_io_get_unique_index(decl
->Semantic
.Name
, 0);
1274 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
1275 dw_addr
= LLVMBuildAdd(gallivm
->builder
, dw_addr
,
1276 lp_build_const_int32(gallivm
, param
* 4), "");
1278 value
= lds_load(&radeon_bld
->soa
.bld_base
, TGSI_TYPE_FLOAT
,
1283 case TGSI_SEMANTIC_PRIMID
:
1284 value
= get_primitive_id(&radeon_bld
->soa
.bld_base
, 0);
1287 case TGSI_SEMANTIC_GRID_SIZE
:
1288 value
= LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_GRID_SIZE
);
1291 case TGSI_SEMANTIC_BLOCK_SIZE
:
1293 LLVMValueRef values
[3];
1295 unsigned *properties
= ctx
->shader
->selector
->info
.properties
;
1296 unsigned sizes
[3] = {
1297 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
],
1298 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT
],
1299 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH
]
1302 for (i
= 0; i
< 3; ++i
)
1303 values
[i
] = lp_build_const_int32(gallivm
, sizes
[i
]);
1305 value
= lp_build_gather_values(gallivm
, values
, 3);
1309 case TGSI_SEMANTIC_BLOCK_ID
:
1310 value
= LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_BLOCK_ID
);
1313 case TGSI_SEMANTIC_THREAD_ID
:
1314 value
= LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_THREAD_ID
);
1318 assert(!"unknown system value");
1322 radeon_bld
->system_values
[index
] = value
;
1325 static void declare_compute_memory(struct radeon_llvm_context
*radeon_bld
,
1326 const struct tgsi_full_declaration
*decl
)
1328 struct si_shader_context
*ctx
=
1329 si_shader_context(&radeon_bld
->soa
.bld_base
);
1330 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
1331 struct gallivm_state
*gallivm
= &radeon_bld
->gallivm
;
1333 LLVMTypeRef i8p
= LLVMPointerType(ctx
->i8
, LOCAL_ADDR_SPACE
);
1336 assert(decl
->Declaration
.MemType
== TGSI_MEMORY_TYPE_SHARED
);
1337 assert(decl
->Range
.First
== decl
->Range
.Last
);
1338 assert(!ctx
->shared_memory
);
1340 var
= LLVMAddGlobalInAddressSpace(gallivm
->module
,
1341 LLVMArrayType(ctx
->i8
, sel
->local_size
),
1344 LLVMSetAlignment(var
, 4);
1346 ctx
->shared_memory
= LLVMBuildBitCast(gallivm
->builder
, var
, i8p
, "");
1349 static LLVMValueRef
fetch_constant(
1350 struct lp_build_tgsi_context
*bld_base
,
1351 const struct tgsi_full_src_register
*reg
,
1352 enum tgsi_opcode_type type
,
1355 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1356 struct lp_build_context
*base
= &bld_base
->base
;
1357 const struct tgsi_ind_register
*ireg
= ®
->Indirect
;
1360 LLVMValueRef addr
, bufp
;
1361 LLVMValueRef result
;
1363 if (swizzle
== LP_CHAN_ALL
) {
1365 LLVMValueRef values
[4];
1366 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
)
1367 values
[chan
] = fetch_constant(bld_base
, reg
, type
, chan
);
1369 return lp_build_gather_values(bld_base
->base
.gallivm
, values
, 4);
1372 buf
= reg
->Register
.Dimension
? reg
->Dimension
.Index
: 0;
1373 idx
= reg
->Register
.Index
* 4 + swizzle
;
1375 if (!reg
->Register
.Indirect
&& !reg
->Dimension
.Indirect
) {
1376 if (type
!= TGSI_TYPE_DOUBLE
)
1377 return bitcast(bld_base
, type
, ctx
->constants
[buf
][idx
]);
1379 return radeon_llvm_emit_fetch_double(bld_base
,
1380 ctx
->constants
[buf
][idx
],
1381 ctx
->constants
[buf
][idx
+ 1]);
1385 if (reg
->Register
.Dimension
&& reg
->Dimension
.Indirect
) {
1386 LLVMValueRef ptr
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_CONST_BUFFERS
);
1388 index
= get_bounded_indirect_index(ctx
, ®
->DimIndirect
,
1389 reg
->Dimension
.Index
,
1390 SI_NUM_USER_CONST_BUFFERS
);
1391 bufp
= build_indexed_load_const(ctx
, ptr
, index
);
1393 bufp
= ctx
->const_buffers
[buf
];
1395 addr
= ctx
->radeon_bld
.soa
.addr
[ireg
->Index
][ireg
->Swizzle
];
1396 addr
= LLVMBuildLoad(base
->gallivm
->builder
, addr
, "load addr reg");
1397 addr
= lp_build_mul_imm(&bld_base
->uint_bld
, addr
, 16);
1398 addr
= lp_build_add(&bld_base
->uint_bld
, addr
,
1399 lp_build_const_int32(base
->gallivm
, idx
* 4));
1401 result
= buffer_load_const(base
->gallivm
->builder
, bufp
,
1404 if (type
!= TGSI_TYPE_DOUBLE
)
1405 result
= bitcast(bld_base
, type
, result
);
1407 LLVMValueRef addr2
, result2
;
1408 addr2
= ctx
->radeon_bld
.soa
.addr
[ireg
->Index
][ireg
->Swizzle
+ 1];
1409 addr2
= LLVMBuildLoad(base
->gallivm
->builder
, addr2
, "load addr reg2");
1410 addr2
= lp_build_mul_imm(&bld_base
->uint_bld
, addr2
, 16);
1411 addr2
= lp_build_add(&bld_base
->uint_bld
, addr2
,
1412 lp_build_const_int32(base
->gallivm
, idx
* 4));
1414 result2
= buffer_load_const(base
->gallivm
->builder
, ctx
->const_buffers
[buf
],
1417 result
= radeon_llvm_emit_fetch_double(bld_base
,
1423 /* Upper 16 bits must be zero. */
1424 static LLVMValueRef
si_llvm_pack_two_int16(struct gallivm_state
*gallivm
,
1425 LLVMValueRef val
[2])
1427 return LLVMBuildOr(gallivm
->builder
, val
[0],
1428 LLVMBuildShl(gallivm
->builder
, val
[1],
1429 lp_build_const_int32(gallivm
, 16),
1433 /* Upper 16 bits are ignored and will be dropped. */
1434 static LLVMValueRef
si_llvm_pack_two_int32_as_int16(struct gallivm_state
*gallivm
,
1435 LLVMValueRef val
[2])
1437 LLVMValueRef v
[2] = {
1438 LLVMBuildAnd(gallivm
->builder
, val
[0],
1439 lp_build_const_int32(gallivm
, 0xffff), ""),
1442 return si_llvm_pack_two_int16(gallivm
, v
);
1445 /* Initialize arguments for the shader export intrinsic */
1446 static void si_llvm_init_export_args(struct lp_build_tgsi_context
*bld_base
,
1447 LLVMValueRef
*values
,
1451 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1452 struct lp_build_context
*uint
=
1453 &ctx
->radeon_bld
.soa
.bld_base
.uint_bld
;
1454 struct lp_build_context
*base
= &bld_base
->base
;
1455 struct gallivm_state
*gallivm
= base
->gallivm
;
1456 LLVMBuilderRef builder
= base
->gallivm
->builder
;
1457 LLVMValueRef val
[4];
1458 unsigned spi_shader_col_format
= V_028714_SPI_SHADER_32_ABGR
;
1462 /* Default is 0xf. Adjusted below depending on the format. */
1463 args
[0] = lp_build_const_int32(base
->gallivm
, 0xf); /* writemask */
1465 /* Specify whether the EXEC mask represents the valid mask */
1466 args
[1] = uint
->zero
;
1468 /* Specify whether this is the last export */
1469 args
[2] = uint
->zero
;
1471 /* Specify the target we are exporting */
1472 args
[3] = lp_build_const_int32(base
->gallivm
, target
);
1474 if (ctx
->type
== TGSI_PROCESSOR_FRAGMENT
) {
1475 const union si_shader_key
*key
= &ctx
->shader
->key
;
1476 unsigned col_formats
= key
->ps
.epilog
.spi_shader_col_format
;
1477 int cbuf
= target
- V_008DFC_SQ_EXP_MRT
;
1479 assert(cbuf
>= 0 && cbuf
< 8);
1480 spi_shader_col_format
= (col_formats
>> (cbuf
* 4)) & 0xf;
1481 is_int8
= (key
->ps
.epilog
.color_is_int8
>> cbuf
) & 0x1;
1484 args
[4] = uint
->zero
; /* COMPR flag */
1485 args
[5] = base
->undef
;
1486 args
[6] = base
->undef
;
1487 args
[7] = base
->undef
;
1488 args
[8] = base
->undef
;
1490 switch (spi_shader_col_format
) {
1491 case V_028714_SPI_SHADER_ZERO
:
1492 args
[0] = uint
->zero
; /* writemask */
1493 args
[3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_NULL
);
1496 case V_028714_SPI_SHADER_32_R
:
1497 args
[0] = uint
->one
; /* writemask */
1498 args
[5] = values
[0];
1501 case V_028714_SPI_SHADER_32_GR
:
1502 args
[0] = lp_build_const_int32(base
->gallivm
, 0x3); /* writemask */
1503 args
[5] = values
[0];
1504 args
[6] = values
[1];
1507 case V_028714_SPI_SHADER_32_AR
:
1508 args
[0] = lp_build_const_int32(base
->gallivm
, 0x9); /* writemask */
1509 args
[5] = values
[0];
1510 args
[8] = values
[3];
1513 case V_028714_SPI_SHADER_FP16_ABGR
:
1514 args
[4] = uint
->one
; /* COMPR flag */
1516 for (chan
= 0; chan
< 2; chan
++) {
1517 LLVMValueRef pack_args
[2] = {
1519 values
[2 * chan
+ 1]
1521 LLVMValueRef packed
;
1523 packed
= lp_build_intrinsic(base
->gallivm
->builder
,
1525 ctx
->i32
, pack_args
, 2,
1526 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
1528 LLVMBuildBitCast(base
->gallivm
->builder
,
1529 packed
, ctx
->f32
, "");
1533 case V_028714_SPI_SHADER_UNORM16_ABGR
:
1534 for (chan
= 0; chan
< 4; chan
++) {
1535 val
[chan
] = radeon_llvm_saturate(bld_base
, values
[chan
]);
1536 val
[chan
] = LLVMBuildFMul(builder
, val
[chan
],
1537 lp_build_const_float(gallivm
, 65535), "");
1538 val
[chan
] = LLVMBuildFAdd(builder
, val
[chan
],
1539 lp_build_const_float(gallivm
, 0.5), "");
1540 val
[chan
] = LLVMBuildFPToUI(builder
, val
[chan
],
1544 args
[4] = uint
->one
; /* COMPR flag */
1545 args
[5] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
1546 si_llvm_pack_two_int16(gallivm
, val
));
1547 args
[6] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
1548 si_llvm_pack_two_int16(gallivm
, val
+2));
1551 case V_028714_SPI_SHADER_SNORM16_ABGR
:
1552 for (chan
= 0; chan
< 4; chan
++) {
1553 /* Clamp between [-1, 1]. */
1554 val
[chan
] = lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_MIN
,
1556 lp_build_const_float(gallivm
, 1));
1557 val
[chan
] = lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_MAX
,
1559 lp_build_const_float(gallivm
, -1));
1560 /* Convert to a signed integer in [-32767, 32767]. */
1561 val
[chan
] = LLVMBuildFMul(builder
, val
[chan
],
1562 lp_build_const_float(gallivm
, 32767), "");
1563 /* If positive, add 0.5, else add -0.5. */
1564 val
[chan
] = LLVMBuildFAdd(builder
, val
[chan
],
1565 LLVMBuildSelect(builder
,
1566 LLVMBuildFCmp(builder
, LLVMRealOGE
,
1567 val
[chan
], base
->zero
, ""),
1568 lp_build_const_float(gallivm
, 0.5),
1569 lp_build_const_float(gallivm
, -0.5), ""), "");
1570 val
[chan
] = LLVMBuildFPToSI(builder
, val
[chan
], ctx
->i32
, "");
1573 args
[4] = uint
->one
; /* COMPR flag */
1574 args
[5] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
1575 si_llvm_pack_two_int32_as_int16(gallivm
, val
));
1576 args
[6] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
1577 si_llvm_pack_two_int32_as_int16(gallivm
, val
+2));
1580 case V_028714_SPI_SHADER_UINT16_ABGR
: {
1581 LLVMValueRef max
= lp_build_const_int32(gallivm
, is_int8
?
1584 for (chan
= 0; chan
< 4; chan
++) {
1585 val
[chan
] = bitcast(bld_base
, TGSI_TYPE_UNSIGNED
, values
[chan
]);
1586 val
[chan
] = lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_UMIN
,
1590 args
[4] = uint
->one
; /* COMPR flag */
1591 args
[5] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
1592 si_llvm_pack_two_int16(gallivm
, val
));
1593 args
[6] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
1594 si_llvm_pack_two_int16(gallivm
, val
+2));
1598 case V_028714_SPI_SHADER_SINT16_ABGR
: {
1599 LLVMValueRef max
= lp_build_const_int32(gallivm
, is_int8
?
1601 LLVMValueRef min
= lp_build_const_int32(gallivm
, is_int8
?
1604 for (chan
= 0; chan
< 4; chan
++) {
1605 val
[chan
] = bitcast(bld_base
, TGSI_TYPE_UNSIGNED
, values
[chan
]);
1606 val
[chan
] = lp_build_emit_llvm_binary(bld_base
,
1609 val
[chan
] = lp_build_emit_llvm_binary(bld_base
,
1614 args
[4] = uint
->one
; /* COMPR flag */
1615 args
[5] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
1616 si_llvm_pack_two_int32_as_int16(gallivm
, val
));
1617 args
[6] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
1618 si_llvm_pack_two_int32_as_int16(gallivm
, val
+2));
1622 case V_028714_SPI_SHADER_32_ABGR
:
1623 memcpy(&args
[5], values
, sizeof(values
[0]) * 4);
1628 static void si_alpha_test(struct lp_build_tgsi_context
*bld_base
,
1631 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1632 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
1634 if (ctx
->shader
->key
.ps
.epilog
.alpha_func
!= PIPE_FUNC_NEVER
) {
1635 LLVMValueRef alpha_ref
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
1636 SI_PARAM_ALPHA_REF
);
1638 LLVMValueRef alpha_pass
=
1639 lp_build_cmp(&bld_base
->base
,
1640 ctx
->shader
->key
.ps
.epilog
.alpha_func
,
1643 lp_build_select(&bld_base
->base
,
1645 lp_build_const_float(gallivm
, 1.0f
),
1646 lp_build_const_float(gallivm
, -1.0f
));
1648 lp_build_intrinsic(gallivm
->builder
, "llvm.AMDGPU.kill",
1649 ctx
->voidt
, &arg
, 1, 0);
1651 lp_build_intrinsic(gallivm
->builder
, "llvm.AMDGPU.kilp",
1652 ctx
->voidt
, NULL
, 0, 0);
1656 static LLVMValueRef
si_scale_alpha_by_sample_mask(struct lp_build_tgsi_context
*bld_base
,
1658 unsigned samplemask_param
)
1660 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1661 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
1662 LLVMValueRef coverage
;
1664 /* alpha = alpha * popcount(coverage) / SI_NUM_SMOOTH_AA_SAMPLES */
1665 coverage
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
1667 coverage
= bitcast(bld_base
, TGSI_TYPE_SIGNED
, coverage
);
1669 coverage
= lp_build_intrinsic(gallivm
->builder
, "llvm.ctpop.i32",
1671 &coverage
, 1, LLVMReadNoneAttribute
);
1673 coverage
= LLVMBuildUIToFP(gallivm
->builder
, coverage
,
1676 coverage
= LLVMBuildFMul(gallivm
->builder
, coverage
,
1677 lp_build_const_float(gallivm
,
1678 1.0 / SI_NUM_SMOOTH_AA_SAMPLES
), "");
1680 return LLVMBuildFMul(gallivm
->builder
, alpha
, coverage
, "");
1683 static void si_llvm_emit_clipvertex(struct lp_build_tgsi_context
*bld_base
,
1684 LLVMValueRef (*pos
)[9], LLVMValueRef
*out_elts
)
1686 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1687 struct lp_build_context
*base
= &bld_base
->base
;
1688 struct lp_build_context
*uint
= &ctx
->radeon_bld
.soa
.bld_base
.uint_bld
;
1691 unsigned const_chan
;
1692 LLVMValueRef base_elt
;
1693 LLVMValueRef ptr
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_CONST_BUFFERS
);
1694 LLVMValueRef constbuf_index
= lp_build_const_int32(base
->gallivm
, SI_DRIVER_STATE_CONST_BUF
);
1695 LLVMValueRef const_resource
= build_indexed_load_const(ctx
, ptr
, constbuf_index
);
1697 for (reg_index
= 0; reg_index
< 2; reg_index
++) {
1698 LLVMValueRef
*args
= pos
[2 + reg_index
];
1703 args
[8] = lp_build_const_float(base
->gallivm
, 0.0f
);
1705 /* Compute dot products of position and user clip plane vectors */
1706 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
1707 for (const_chan
= 0; const_chan
< TGSI_NUM_CHANNELS
; const_chan
++) {
1708 args
[1] = lp_build_const_int32(base
->gallivm
,
1709 ((reg_index
* 4 + chan
) * 4 +
1711 base_elt
= buffer_load_const(base
->gallivm
->builder
, const_resource
,
1714 lp_build_add(base
, args
[5 + chan
],
1715 lp_build_mul(base
, base_elt
,
1716 out_elts
[const_chan
]));
1720 args
[0] = lp_build_const_int32(base
->gallivm
, 0xf);
1721 args
[1] = uint
->zero
;
1722 args
[2] = uint
->zero
;
1723 args
[3] = lp_build_const_int32(base
->gallivm
,
1724 V_008DFC_SQ_EXP_POS
+ 2 + reg_index
);
1725 args
[4] = uint
->zero
;
1729 static void si_dump_streamout(struct pipe_stream_output_info
*so
)
1733 if (so
->num_outputs
)
1734 fprintf(stderr
, "STREAMOUT\n");
1736 for (i
= 0; i
< so
->num_outputs
; i
++) {
1737 unsigned mask
= ((1 << so
->output
[i
].num_components
) - 1) <<
1738 so
->output
[i
].start_component
;
1739 fprintf(stderr
, " %i: BUF%i[%i..%i] <- OUT[%i].%s%s%s%s\n",
1740 i
, so
->output
[i
].output_buffer
,
1741 so
->output
[i
].dst_offset
, so
->output
[i
].dst_offset
+ so
->output
[i
].num_components
- 1,
1742 so
->output
[i
].register_index
,
1743 mask
& 1 ? "x" : "",
1744 mask
& 2 ? "y" : "",
1745 mask
& 4 ? "z" : "",
1746 mask
& 8 ? "w" : "");
1750 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
1751 * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
1752 * or v4i32 (num_channels=3,4). */
1753 static void build_tbuffer_store(struct si_shader_context
*ctx
,
1756 unsigned num_channels
,
1758 LLVMValueRef soffset
,
1759 unsigned inst_offset
,
1768 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
1769 LLVMValueRef args
[] = {
1772 LLVMConstInt(ctx
->i32
, num_channels
, 0),
1775 LLVMConstInt(ctx
->i32
, inst_offset
, 0),
1776 LLVMConstInt(ctx
->i32
, dfmt
, 0),
1777 LLVMConstInt(ctx
->i32
, nfmt
, 0),
1778 LLVMConstInt(ctx
->i32
, offen
, 0),
1779 LLVMConstInt(ctx
->i32
, idxen
, 0),
1780 LLVMConstInt(ctx
->i32
, glc
, 0),
1781 LLVMConstInt(ctx
->i32
, slc
, 0),
1782 LLVMConstInt(ctx
->i32
, tfe
, 0)
1785 /* The instruction offset field has 12 bits */
1786 assert(offen
|| inst_offset
< (1 << 12));
1788 /* The intrinsic is overloaded, we need to add a type suffix for overloading to work. */
1789 unsigned func
= CLAMP(num_channels
, 1, 3) - 1;
1790 const char *types
[] = {"i32", "v2i32", "v4i32"};
1792 snprintf(name
, sizeof(name
), "llvm.SI.tbuffer.store.%s", types
[func
]);
1794 lp_build_intrinsic(gallivm
->builder
, name
, ctx
->voidt
,
1795 args
, Elements(args
), 0);
1798 static void build_tbuffer_store_dwords(struct si_shader_context
*ctx
,
1801 unsigned num_channels
,
1803 LLVMValueRef soffset
,
1804 unsigned inst_offset
)
1806 static unsigned dfmt
[] = {
1807 V_008F0C_BUF_DATA_FORMAT_32
,
1808 V_008F0C_BUF_DATA_FORMAT_32_32
,
1809 V_008F0C_BUF_DATA_FORMAT_32_32_32
,
1810 V_008F0C_BUF_DATA_FORMAT_32_32_32_32
1812 assert(num_channels
>= 1 && num_channels
<= 4);
1814 build_tbuffer_store(ctx
, rsrc
, vdata
, num_channels
, vaddr
, soffset
,
1815 inst_offset
, dfmt
[num_channels
-1],
1816 V_008F0C_BUF_NUM_FORMAT_UINT
, 1, 0, 1, 1, 0);
1819 /* On SI, the vertex shader is responsible for writing streamout data
1821 static void si_llvm_emit_streamout(struct si_shader_context
*ctx
,
1822 struct si_shader_output_values
*outputs
,
1825 struct pipe_stream_output_info
*so
= &ctx
->shader
->selector
->so
;
1826 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
1827 LLVMBuilderRef builder
= gallivm
->builder
;
1829 struct lp_build_if_state if_ctx
;
1831 /* Get bits [22:16], i.e. (so_param >> 16) & 127; */
1832 LLVMValueRef so_vtx_count
=
1833 unpack_param(ctx
, ctx
->param_streamout_config
, 16, 7);
1835 LLVMValueRef tid
= lp_build_intrinsic(builder
, "llvm.SI.tid", ctx
->i32
,
1836 NULL
, 0, LLVMReadNoneAttribute
);
1838 /* can_emit = tid < so_vtx_count; */
1839 LLVMValueRef can_emit
=
1840 LLVMBuildICmp(builder
, LLVMIntULT
, tid
, so_vtx_count
, "");
1842 LLVMValueRef stream_id
=
1843 unpack_param(ctx
, ctx
->param_streamout_config
, 24, 2);
1845 /* Emit the streamout code conditionally. This actually avoids
1846 * out-of-bounds buffer access. The hw tells us via the SGPR
1847 * (so_vtx_count) which threads are allowed to emit streamout data. */
1848 lp_build_if(&if_ctx
, gallivm
, can_emit
);
1850 /* The buffer offset is computed as follows:
1851 * ByteOffset = streamout_offset[buffer_id]*4 +
1852 * (streamout_write_index + thread_id)*stride[buffer_id] +
1856 LLVMValueRef so_write_index
=
1857 LLVMGetParam(ctx
->radeon_bld
.main_fn
,
1858 ctx
->param_streamout_write_index
);
1860 /* Compute (streamout_write_index + thread_id). */
1861 so_write_index
= LLVMBuildAdd(builder
, so_write_index
, tid
, "");
1863 /* Compute the write offset for each enabled buffer. */
1864 LLVMValueRef so_write_offset
[4] = {};
1865 for (i
= 0; i
< 4; i
++) {
1869 LLVMValueRef so_offset
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
1870 ctx
->param_streamout_offset
[i
]);
1871 so_offset
= LLVMBuildMul(builder
, so_offset
, LLVMConstInt(ctx
->i32
, 4, 0), "");
1873 so_write_offset
[i
] = LLVMBuildMul(builder
, so_write_index
,
1874 LLVMConstInt(ctx
->i32
, so
->stride
[i
]*4, 0), "");
1875 so_write_offset
[i
] = LLVMBuildAdd(builder
, so_write_offset
[i
], so_offset
, "");
1878 /* Write streamout data. */
1879 for (i
= 0; i
< so
->num_outputs
; i
++) {
1880 unsigned buf_idx
= so
->output
[i
].output_buffer
;
1881 unsigned reg
= so
->output
[i
].register_index
;
1882 unsigned start
= so
->output
[i
].start_component
;
1883 unsigned num_comps
= so
->output
[i
].num_components
;
1884 unsigned stream
= so
->output
[i
].stream
;
1885 LLVMValueRef out
[4];
1886 struct lp_build_if_state if_ctx_stream
;
1888 assert(num_comps
&& num_comps
<= 4);
1889 if (!num_comps
|| num_comps
> 4)
1895 /* Load the output as int. */
1896 for (j
= 0; j
< num_comps
; j
++) {
1897 out
[j
] = LLVMBuildBitCast(builder
,
1898 outputs
[reg
].values
[start
+j
],
1902 /* Pack the output. */
1903 LLVMValueRef vdata
= NULL
;
1905 switch (num_comps
) {
1906 case 1: /* as i32 */
1909 case 2: /* as v2i32 */
1910 case 3: /* as v4i32 (aligned to 4) */
1911 case 4: /* as v4i32 */
1912 vdata
= LLVMGetUndef(LLVMVectorType(ctx
->i32
, util_next_power_of_two(num_comps
)));
1913 for (j
= 0; j
< num_comps
; j
++) {
1914 vdata
= LLVMBuildInsertElement(builder
, vdata
, out
[j
],
1915 LLVMConstInt(ctx
->i32
, j
, 0), "");
1920 LLVMValueRef can_emit_stream
=
1921 LLVMBuildICmp(builder
, LLVMIntEQ
,
1923 lp_build_const_int32(gallivm
, stream
), "");
1925 lp_build_if(&if_ctx_stream
, gallivm
, can_emit_stream
);
1926 build_tbuffer_store_dwords(ctx
, ctx
->so_buffers
[buf_idx
],
1928 so_write_offset
[buf_idx
],
1929 LLVMConstInt(ctx
->i32
, 0, 0),
1930 so
->output
[i
].dst_offset
*4);
1931 lp_build_endif(&if_ctx_stream
);
1934 lp_build_endif(&if_ctx
);
1938 /* Generate export instructions for hardware VS shader stage */
1939 static void si_llvm_export_vs(struct lp_build_tgsi_context
*bld_base
,
1940 struct si_shader_output_values
*outputs
,
1943 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
1944 struct si_shader
*shader
= ctx
->shader
;
1945 struct lp_build_context
*base
= &bld_base
->base
;
1946 struct lp_build_context
*uint
=
1947 &ctx
->radeon_bld
.soa
.bld_base
.uint_bld
;
1948 LLVMValueRef args
[9];
1949 LLVMValueRef pos_args
[4][9] = { { 0 } };
1950 LLVMValueRef psize_value
= NULL
, edgeflag_value
= NULL
, layer_value
= NULL
, viewport_index_value
= NULL
;
1951 unsigned semantic_name
, semantic_index
;
1953 unsigned param_count
= 0;
1957 if (outputs
&& ctx
->shader
->selector
->so
.num_outputs
) {
1958 si_llvm_emit_streamout(ctx
, outputs
, noutput
);
1961 for (i
= 0; i
< noutput
; i
++) {
1962 semantic_name
= outputs
[i
].name
;
1963 semantic_index
= outputs
[i
].sid
;
1966 /* Select the correct target */
1967 switch(semantic_name
) {
1968 case TGSI_SEMANTIC_PSIZE
:
1969 psize_value
= outputs
[i
].values
[0];
1971 case TGSI_SEMANTIC_EDGEFLAG
:
1972 edgeflag_value
= outputs
[i
].values
[0];
1974 case TGSI_SEMANTIC_LAYER
:
1975 layer_value
= outputs
[i
].values
[0];
1976 semantic_name
= TGSI_SEMANTIC_GENERIC
;
1977 goto handle_semantic
;
1978 case TGSI_SEMANTIC_VIEWPORT_INDEX
:
1979 viewport_index_value
= outputs
[i
].values
[0];
1980 semantic_name
= TGSI_SEMANTIC_GENERIC
;
1981 goto handle_semantic
;
1982 case TGSI_SEMANTIC_POSITION
:
1983 target
= V_008DFC_SQ_EXP_POS
;
1985 case TGSI_SEMANTIC_COLOR
:
1986 case TGSI_SEMANTIC_BCOLOR
:
1987 target
= V_008DFC_SQ_EXP_PARAM
+ param_count
;
1988 assert(i
< ARRAY_SIZE(shader
->info
.vs_output_param_offset
));
1989 shader
->info
.vs_output_param_offset
[i
] = param_count
;
1992 case TGSI_SEMANTIC_CLIPDIST
:
1993 target
= V_008DFC_SQ_EXP_POS
+ 2 + semantic_index
;
1995 case TGSI_SEMANTIC_CLIPVERTEX
:
1996 si_llvm_emit_clipvertex(bld_base
, pos_args
, outputs
[i
].values
);
1998 case TGSI_SEMANTIC_PRIMID
:
1999 case TGSI_SEMANTIC_FOG
:
2000 case TGSI_SEMANTIC_TEXCOORD
:
2001 case TGSI_SEMANTIC_GENERIC
:
2002 target
= V_008DFC_SQ_EXP_PARAM
+ param_count
;
2003 assert(i
< ARRAY_SIZE(shader
->info
.vs_output_param_offset
));
2004 shader
->info
.vs_output_param_offset
[i
] = param_count
;
2010 "Warning: SI unhandled vs output type:%d\n",
2014 si_llvm_init_export_args(bld_base
, outputs
[i
].values
, target
, args
);
2016 if (target
>= V_008DFC_SQ_EXP_POS
&&
2017 target
<= (V_008DFC_SQ_EXP_POS
+ 3)) {
2018 memcpy(pos_args
[target
- V_008DFC_SQ_EXP_POS
],
2019 args
, sizeof(args
));
2021 lp_build_intrinsic(base
->gallivm
->builder
,
2022 "llvm.SI.export", ctx
->voidt
,
2026 if (semantic_name
== TGSI_SEMANTIC_CLIPDIST
) {
2027 semantic_name
= TGSI_SEMANTIC_GENERIC
;
2028 goto handle_semantic
;
2032 shader
->info
.nr_param_exports
= param_count
;
2034 /* We need to add the position output manually if it's missing. */
2035 if (!pos_args
[0][0]) {
2036 pos_args
[0][0] = lp_build_const_int32(base
->gallivm
, 0xf); /* writemask */
2037 pos_args
[0][1] = uint
->zero
; /* EXEC mask */
2038 pos_args
[0][2] = uint
->zero
; /* last export? */
2039 pos_args
[0][3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_POS
);
2040 pos_args
[0][4] = uint
->zero
; /* COMPR flag */
2041 pos_args
[0][5] = base
->zero
; /* X */
2042 pos_args
[0][6] = base
->zero
; /* Y */
2043 pos_args
[0][7] = base
->zero
; /* Z */
2044 pos_args
[0][8] = base
->one
; /* W */
2047 /* Write the misc vector (point size, edgeflag, layer, viewport). */
2048 if (shader
->selector
->info
.writes_psize
||
2049 shader
->selector
->info
.writes_edgeflag
||
2050 shader
->selector
->info
.writes_viewport_index
||
2051 shader
->selector
->info
.writes_layer
) {
2052 pos_args
[1][0] = lp_build_const_int32(base
->gallivm
, /* writemask */
2053 shader
->selector
->info
.writes_psize
|
2054 (shader
->selector
->info
.writes_edgeflag
<< 1) |
2055 (shader
->selector
->info
.writes_layer
<< 2) |
2056 (shader
->selector
->info
.writes_viewport_index
<< 3));
2057 pos_args
[1][1] = uint
->zero
; /* EXEC mask */
2058 pos_args
[1][2] = uint
->zero
; /* last export? */
2059 pos_args
[1][3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_POS
+ 1);
2060 pos_args
[1][4] = uint
->zero
; /* COMPR flag */
2061 pos_args
[1][5] = base
->zero
; /* X */
2062 pos_args
[1][6] = base
->zero
; /* Y */
2063 pos_args
[1][7] = base
->zero
; /* Z */
2064 pos_args
[1][8] = base
->zero
; /* W */
2066 if (shader
->selector
->info
.writes_psize
)
2067 pos_args
[1][5] = psize_value
;
2069 if (shader
->selector
->info
.writes_edgeflag
) {
2070 /* The output is a float, but the hw expects an integer
2071 * with the first bit containing the edge flag. */
2072 edgeflag_value
= LLVMBuildFPToUI(base
->gallivm
->builder
,
2075 edgeflag_value
= lp_build_min(&bld_base
->int_bld
,
2077 bld_base
->int_bld
.one
);
2079 /* The LLVM intrinsic expects a float. */
2080 pos_args
[1][6] = LLVMBuildBitCast(base
->gallivm
->builder
,
2085 if (shader
->selector
->info
.writes_layer
)
2086 pos_args
[1][7] = layer_value
;
2088 if (shader
->selector
->info
.writes_viewport_index
)
2089 pos_args
[1][8] = viewport_index_value
;
2092 for (i
= 0; i
< 4; i
++)
2094 shader
->info
.nr_pos_exports
++;
2097 for (i
= 0; i
< 4; i
++) {
2098 if (!pos_args
[i
][0])
2101 /* Specify the target we are exporting */
2102 pos_args
[i
][3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_POS
+ pos_idx
++);
2104 if (pos_idx
== shader
->info
.nr_pos_exports
)
2105 /* Specify that this is the last export */
2106 pos_args
[i
][2] = uint
->one
;
2108 lp_build_intrinsic(base
->gallivm
->builder
, "llvm.SI.export",
2109 ctx
->voidt
, pos_args
[i
], 9, 0);
2113 static void si_write_tess_factors(struct lp_build_tgsi_context
*bld_base
,
2114 LLVMValueRef rel_patch_id
,
2115 LLVMValueRef invocation_id
,
2116 LLVMValueRef tcs_out_current_patch_data_offset
)
2118 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2119 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2120 struct si_shader
*shader
= ctx
->shader
;
2121 unsigned tess_inner_index
, tess_outer_index
;
2122 LLVMValueRef lds_base
, lds_inner
, lds_outer
, byteoffset
, buffer
;
2123 LLVMValueRef out
[6], vec0
, vec1
, rw_buffers
, tf_base
;
2124 unsigned stride
, outer_comps
, inner_comps
, i
;
2125 struct lp_build_if_state if_ctx
;
2127 /* Do this only for invocation 0, because the tess levels are per-patch,
2130 * This can't jump, because invocation 0 executes this. It should
2131 * at least mask out the loads and stores for other invocations.
2133 lp_build_if(&if_ctx
, gallivm
,
2134 LLVMBuildICmp(gallivm
->builder
, LLVMIntEQ
,
2135 invocation_id
, bld_base
->uint_bld
.zero
, ""));
2137 /* Determine the layout of one tess factor element in the buffer. */
2138 switch (shader
->key
.tcs
.epilog
.prim_mode
) {
2139 case PIPE_PRIM_LINES
:
2140 stride
= 2; /* 2 dwords, 1 vec2 store */
2144 case PIPE_PRIM_TRIANGLES
:
2145 stride
= 4; /* 4 dwords, 1 vec4 store */
2149 case PIPE_PRIM_QUADS
:
2150 stride
= 6; /* 6 dwords, 2 stores (vec4 + vec2) */
2159 /* Load tess_inner and tess_outer from LDS.
2160 * Any invocation can write them, so we can't get them from a temporary.
2162 tess_inner_index
= si_shader_io_get_unique_index(TGSI_SEMANTIC_TESSINNER
, 0);
2163 tess_outer_index
= si_shader_io_get_unique_index(TGSI_SEMANTIC_TESSOUTER
, 0);
2165 lds_base
= tcs_out_current_patch_data_offset
;
2166 lds_inner
= LLVMBuildAdd(gallivm
->builder
, lds_base
,
2167 lp_build_const_int32(gallivm
,
2168 tess_inner_index
* 4), "");
2169 lds_outer
= LLVMBuildAdd(gallivm
->builder
, lds_base
,
2170 lp_build_const_int32(gallivm
,
2171 tess_outer_index
* 4), "");
2173 for (i
= 0; i
< outer_comps
; i
++)
2174 out
[i
] = lds_load(bld_base
, TGSI_TYPE_SIGNED
, i
, lds_outer
);
2175 for (i
= 0; i
< inner_comps
; i
++)
2176 out
[outer_comps
+i
] = lds_load(bld_base
, TGSI_TYPE_SIGNED
, i
, lds_inner
);
2178 /* Convert the outputs to vectors for stores. */
2179 vec0
= lp_build_gather_values(gallivm
, out
, MIN2(stride
, 4));
2183 vec1
= lp_build_gather_values(gallivm
, out
+4, stride
- 4);
2185 /* Get the buffer. */
2186 rw_buffers
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
2187 SI_PARAM_RW_BUFFERS
);
2188 buffer
= build_indexed_load_const(ctx
, rw_buffers
,
2189 lp_build_const_int32(gallivm
, SI_HS_RING_TESS_FACTOR
));
2191 /* Get the offset. */
2192 tf_base
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
2193 SI_PARAM_TESS_FACTOR_OFFSET
);
2194 byteoffset
= LLVMBuildMul(gallivm
->builder
, rel_patch_id
,
2195 lp_build_const_int32(gallivm
, 4 * stride
), "");
2197 /* Store the outputs. */
2198 build_tbuffer_store_dwords(ctx
, buffer
, vec0
,
2199 MIN2(stride
, 4), byteoffset
, tf_base
, 0);
2201 build_tbuffer_store_dwords(ctx
, buffer
, vec1
,
2202 stride
- 4, byteoffset
, tf_base
, 16);
2203 lp_build_endif(&if_ctx
);
2206 /* This only writes the tessellation factor levels. */
2207 static void si_llvm_emit_tcs_epilogue(struct lp_build_tgsi_context
*bld_base
)
2209 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2210 LLVMValueRef rel_patch_id
, invocation_id
, tf_lds_offset
;
2212 rel_patch_id
= get_rel_patch_id(ctx
);
2213 invocation_id
= unpack_param(ctx
, SI_PARAM_REL_IDS
, 8, 5);
2214 tf_lds_offset
= get_tcs_out_current_patch_data_offset(ctx
);
2216 if (!ctx
->is_monolithic
) {
2217 /* Return epilog parameters from this function. */
2218 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
2219 LLVMValueRef ret
= ctx
->return_value
;
2220 LLVMValueRef rw_buffers
, rw0
, rw1
, tf_soffset
;
2223 /* RW_BUFFERS pointer */
2224 rw_buffers
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
2225 SI_PARAM_RW_BUFFERS
);
2226 rw_buffers
= LLVMBuildPtrToInt(builder
, rw_buffers
, ctx
->i64
, "");
2227 rw_buffers
= LLVMBuildBitCast(builder
, rw_buffers
, ctx
->v2i32
, "");
2228 rw0
= LLVMBuildExtractElement(builder
, rw_buffers
,
2229 bld_base
->uint_bld
.zero
, "");
2230 rw1
= LLVMBuildExtractElement(builder
, rw_buffers
,
2231 bld_base
->uint_bld
.one
, "");
2232 ret
= LLVMBuildInsertValue(builder
, ret
, rw0
, 0, "");
2233 ret
= LLVMBuildInsertValue(builder
, ret
, rw1
, 1, "");
2235 /* Tess factor buffer soffset is after user SGPRs. */
2236 tf_soffset
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
2237 SI_PARAM_TESS_FACTOR_OFFSET
);
2238 ret
= LLVMBuildInsertValue(builder
, ret
, tf_soffset
,
2239 SI_TCS_NUM_USER_SGPR
, "");
2242 rel_patch_id
= bitcast(bld_base
, TGSI_TYPE_FLOAT
, rel_patch_id
);
2243 invocation_id
= bitcast(bld_base
, TGSI_TYPE_FLOAT
, invocation_id
);
2244 tf_lds_offset
= bitcast(bld_base
, TGSI_TYPE_FLOAT
, tf_lds_offset
);
2246 vgpr
= SI_TCS_NUM_USER_SGPR
+ 1;
2247 ret
= LLVMBuildInsertValue(builder
, ret
, rel_patch_id
, vgpr
++, "");
2248 ret
= LLVMBuildInsertValue(builder
, ret
, invocation_id
, vgpr
++, "");
2249 ret
= LLVMBuildInsertValue(builder
, ret
, tf_lds_offset
, vgpr
++, "");
2250 ctx
->return_value
= ret
;
2254 si_write_tess_factors(bld_base
, rel_patch_id
, invocation_id
, tf_lds_offset
);
2257 static void si_llvm_emit_ls_epilogue(struct lp_build_tgsi_context
*bld_base
)
2259 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2260 struct si_shader
*shader
= ctx
->shader
;
2261 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
2262 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2264 LLVMValueRef vertex_id
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
2265 ctx
->param_rel_auto_id
);
2266 LLVMValueRef vertex_dw_stride
=
2267 unpack_param(ctx
, SI_PARAM_LS_OUT_LAYOUT
, 13, 8);
2268 LLVMValueRef base_dw_addr
= LLVMBuildMul(gallivm
->builder
, vertex_id
,
2269 vertex_dw_stride
, "");
2271 /* Write outputs to LDS. The next shader (TCS aka HS) will read
2272 * its inputs from it. */
2273 for (i
= 0; i
< info
->num_outputs
; i
++) {
2274 LLVMValueRef
*out_ptr
= ctx
->radeon_bld
.soa
.outputs
[i
];
2275 unsigned name
= info
->output_semantic_name
[i
];
2276 unsigned index
= info
->output_semantic_index
[i
];
2277 int param
= si_shader_io_get_unique_index(name
, index
);
2278 LLVMValueRef dw_addr
= LLVMBuildAdd(gallivm
->builder
, base_dw_addr
,
2279 lp_build_const_int32(gallivm
, param
* 4), "");
2281 for (chan
= 0; chan
< 4; chan
++) {
2282 lds_store(bld_base
, chan
, dw_addr
,
2283 LLVMBuildLoad(gallivm
->builder
, out_ptr
[chan
], ""));
2288 static void si_llvm_emit_es_epilogue(struct lp_build_tgsi_context
*bld_base
)
2290 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2291 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2292 struct si_shader
*es
= ctx
->shader
;
2293 struct tgsi_shader_info
*info
= &es
->selector
->info
;
2294 LLVMValueRef soffset
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
2295 ctx
->param_es2gs_offset
);
2299 for (i
= 0; i
< info
->num_outputs
; i
++) {
2300 LLVMValueRef
*out_ptr
=
2301 ctx
->radeon_bld
.soa
.outputs
[i
];
2304 if (info
->output_semantic_name
[i
] == TGSI_SEMANTIC_VIEWPORT_INDEX
||
2305 info
->output_semantic_name
[i
] == TGSI_SEMANTIC_LAYER
)
2308 param_index
= si_shader_io_get_unique_index(info
->output_semantic_name
[i
],
2309 info
->output_semantic_index
[i
]);
2311 for (chan
= 0; chan
< 4; chan
++) {
2312 LLVMValueRef out_val
= LLVMBuildLoad(gallivm
->builder
, out_ptr
[chan
], "");
2313 out_val
= LLVMBuildBitCast(gallivm
->builder
, out_val
, ctx
->i32
, "");
2315 build_tbuffer_store(ctx
,
2318 LLVMGetUndef(ctx
->i32
), soffset
,
2319 (4 * param_index
+ chan
) * 4,
2320 V_008F0C_BUF_DATA_FORMAT_32
,
2321 V_008F0C_BUF_NUM_FORMAT_UINT
,
2327 static void si_llvm_emit_gs_epilogue(struct lp_build_tgsi_context
*bld_base
)
2329 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2330 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2331 LLVMValueRef args
[2];
2333 args
[0] = lp_build_const_int32(gallivm
, SENDMSG_GS_OP_NOP
| SENDMSG_GS_DONE
);
2334 args
[1] = LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_GS_WAVE_ID
);
2335 lp_build_intrinsic(gallivm
->builder
, "llvm.SI.sendmsg",
2336 ctx
->voidt
, args
, 2, LLVMNoUnwindAttribute
);
2339 static void si_llvm_emit_vs_epilogue(struct lp_build_tgsi_context
*bld_base
)
2341 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2342 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2343 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
2344 struct si_shader_output_values
*outputs
= NULL
;
2347 assert(!ctx
->is_gs_copy_shader
);
2349 outputs
= MALLOC((info
->num_outputs
+ 1) * sizeof(outputs
[0]));
2351 /* Vertex color clamping.
2353 * This uses a state constant loaded in a user data SGPR and
2354 * an IF statement is added that clamps all colors if the constant
2357 if (ctx
->type
== TGSI_PROCESSOR_VERTEX
) {
2358 struct lp_build_if_state if_ctx
;
2359 LLVMValueRef cond
= NULL
;
2360 LLVMValueRef addr
, val
;
2362 for (i
= 0; i
< info
->num_outputs
; i
++) {
2363 if (info
->output_semantic_name
[i
] != TGSI_SEMANTIC_COLOR
&&
2364 info
->output_semantic_name
[i
] != TGSI_SEMANTIC_BCOLOR
)
2367 /* We've found a color. */
2369 /* The state is in the first bit of the user SGPR. */
2370 cond
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
2371 SI_PARAM_VS_STATE_BITS
);
2372 cond
= LLVMBuildTrunc(gallivm
->builder
, cond
,
2374 lp_build_if(&if_ctx
, gallivm
, cond
);
2377 for (j
= 0; j
< 4; j
++) {
2378 addr
= ctx
->radeon_bld
.soa
.outputs
[i
][j
];
2379 val
= LLVMBuildLoad(gallivm
->builder
, addr
, "");
2380 val
= radeon_llvm_saturate(bld_base
, val
);
2381 LLVMBuildStore(gallivm
->builder
, val
, addr
);
2386 lp_build_endif(&if_ctx
);
2389 for (i
= 0; i
< info
->num_outputs
; i
++) {
2390 outputs
[i
].name
= info
->output_semantic_name
[i
];
2391 outputs
[i
].sid
= info
->output_semantic_index
[i
];
2393 for (j
= 0; j
< 4; j
++)
2394 outputs
[i
].values
[j
] =
2395 LLVMBuildLoad(gallivm
->builder
,
2396 ctx
->radeon_bld
.soa
.outputs
[i
][j
],
2400 if (ctx
->is_monolithic
) {
2401 /* Export PrimitiveID when PS needs it. */
2402 if (si_vs_exports_prim_id(ctx
->shader
)) {
2403 outputs
[i
].name
= TGSI_SEMANTIC_PRIMID
;
2405 outputs
[i
].values
[0] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
2406 get_primitive_id(bld_base
, 0));
2407 outputs
[i
].values
[1] = bld_base
->base
.undef
;
2408 outputs
[i
].values
[2] = bld_base
->base
.undef
;
2409 outputs
[i
].values
[3] = bld_base
->base
.undef
;
2413 /* Return the primitive ID from the LLVM function. */
2415 LLVMBuildInsertValue(gallivm
->builder
,
2417 bitcast(bld_base
, TGSI_TYPE_FLOAT
,
2418 get_primitive_id(bld_base
, 0)),
2419 VS_EPILOG_PRIMID_LOC
, "");
2422 si_llvm_export_vs(bld_base
, outputs
, i
);
2426 static void si_export_mrt_z(struct lp_build_tgsi_context
*bld_base
,
2427 LLVMValueRef depth
, LLVMValueRef stencil
,
2428 LLVMValueRef samplemask
)
2430 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2431 struct lp_build_context
*base
= &bld_base
->base
;
2432 struct lp_build_context
*uint
= &bld_base
->uint_bld
;
2433 LLVMValueRef args
[9];
2436 assert(depth
|| stencil
|| samplemask
);
2438 args
[1] = uint
->one
; /* whether the EXEC mask is valid */
2439 args
[2] = uint
->one
; /* DONE bit */
2441 /* Specify the target we are exporting */
2442 args
[3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_MRTZ
);
2444 args
[4] = uint
->zero
; /* COMP flag */
2445 args
[5] = base
->undef
; /* R, depth */
2446 args
[6] = base
->undef
; /* G, stencil test value[0:7], stencil op value[8:15] */
2447 args
[7] = base
->undef
; /* B, sample mask */
2448 args
[8] = base
->undef
; /* A, alpha to mask */
2461 args
[7] = samplemask
;
2465 /* SI (except OLAND) has a bug that it only looks
2466 * at the X writemask component. */
2467 if (ctx
->screen
->b
.chip_class
== SI
&&
2468 ctx
->screen
->b
.family
!= CHIP_OLAND
)
2471 /* Specify which components to enable */
2472 args
[0] = lp_build_const_int32(base
->gallivm
, mask
);
2474 lp_build_intrinsic(base
->gallivm
->builder
, "llvm.SI.export",
2475 ctx
->voidt
, args
, 9, 0);
2478 static void si_export_mrt_color(struct lp_build_tgsi_context
*bld_base
,
2479 LLVMValueRef
*color
, unsigned index
,
2480 unsigned samplemask_param
,
2483 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2484 struct lp_build_context
*base
= &bld_base
->base
;
2488 if (ctx
->shader
->key
.ps
.epilog
.clamp_color
)
2489 for (i
= 0; i
< 4; i
++)
2490 color
[i
] = radeon_llvm_saturate(bld_base
, color
[i
]);
2493 if (ctx
->shader
->key
.ps
.epilog
.alpha_to_one
)
2494 color
[3] = base
->one
;
2498 ctx
->shader
->key
.ps
.epilog
.alpha_func
!= PIPE_FUNC_ALWAYS
)
2499 si_alpha_test(bld_base
, color
[3]);
2501 /* Line & polygon smoothing */
2502 if (ctx
->shader
->key
.ps
.epilog
.poly_line_smoothing
)
2503 color
[3] = si_scale_alpha_by_sample_mask(bld_base
, color
[3],
2506 /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
2507 if (ctx
->shader
->key
.ps
.epilog
.last_cbuf
> 0) {
2508 LLVMValueRef args
[8][9];
2511 /* Get the export arguments, also find out what the last one is. */
2512 for (c
= 0; c
<= ctx
->shader
->key
.ps
.epilog
.last_cbuf
; c
++) {
2513 si_llvm_init_export_args(bld_base
, color
,
2514 V_008DFC_SQ_EXP_MRT
+ c
, args
[c
]);
2515 if (args
[c
][0] != bld_base
->uint_bld
.zero
)
2519 /* Emit all exports. */
2520 for (c
= 0; c
<= ctx
->shader
->key
.ps
.epilog
.last_cbuf
; c
++) {
2521 if (is_last
&& last
== c
) {
2522 args
[c
][1] = bld_base
->uint_bld
.one
; /* whether the EXEC mask is valid */
2523 args
[c
][2] = bld_base
->uint_bld
.one
; /* DONE bit */
2524 } else if (args
[c
][0] == bld_base
->uint_bld
.zero
)
2525 continue; /* unnecessary NULL export */
2527 lp_build_intrinsic(base
->gallivm
->builder
, "llvm.SI.export",
2528 ctx
->voidt
, args
[c
], 9, 0);
2531 LLVMValueRef args
[9];
2534 si_llvm_init_export_args(bld_base
, color
, V_008DFC_SQ_EXP_MRT
+ index
,
2537 args
[1] = bld_base
->uint_bld
.one
; /* whether the EXEC mask is valid */
2538 args
[2] = bld_base
->uint_bld
.one
; /* DONE bit */
2539 } else if (args
[0] == bld_base
->uint_bld
.zero
)
2540 return; /* unnecessary NULL export */
2542 lp_build_intrinsic(base
->gallivm
->builder
, "llvm.SI.export",
2543 ctx
->voidt
, args
, 9, 0);
2547 static void si_export_null(struct lp_build_tgsi_context
*bld_base
)
2549 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2550 struct lp_build_context
*base
= &bld_base
->base
;
2551 struct lp_build_context
*uint
= &bld_base
->uint_bld
;
2552 LLVMValueRef args
[9];
2554 args
[0] = lp_build_const_int32(base
->gallivm
, 0x0); /* enabled channels */
2555 args
[1] = uint
->one
; /* whether the EXEC mask is valid */
2556 args
[2] = uint
->one
; /* DONE bit */
2557 args
[3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_NULL
);
2558 args
[4] = uint
->zero
; /* COMPR flag (0 = 32-bit export) */
2559 args
[5] = uint
->undef
; /* R */
2560 args
[6] = uint
->undef
; /* G */
2561 args
[7] = uint
->undef
; /* B */
2562 args
[8] = uint
->undef
; /* A */
2564 lp_build_intrinsic(base
->gallivm
->builder
, "llvm.SI.export",
2565 ctx
->voidt
, args
, 9, 0);
2568 static void si_llvm_emit_fs_epilogue(struct lp_build_tgsi_context
*bld_base
)
2570 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2571 struct si_shader
*shader
= ctx
->shader
;
2572 struct lp_build_context
*base
= &bld_base
->base
;
2573 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
2574 LLVMBuilderRef builder
= base
->gallivm
->builder
;
2575 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
2576 int last_color_export
= -1;
2579 /* Determine the last export. If MRTZ is present, it's always last.
2580 * Otherwise, find the last color export.
2582 if (!info
->writes_z
&& !info
->writes_stencil
&& !info
->writes_samplemask
) {
2583 unsigned spi_format
= shader
->key
.ps
.epilog
.spi_shader_col_format
;
2585 /* Don't export NULL and return if alpha-test is enabled. */
2586 if (shader
->key
.ps
.epilog
.alpha_func
!= PIPE_FUNC_ALWAYS
&&
2587 shader
->key
.ps
.epilog
.alpha_func
!= PIPE_FUNC_NEVER
&&
2588 (spi_format
& 0xf) == 0)
2589 spi_format
|= V_028714_SPI_SHADER_32_AR
;
2591 for (i
= 0; i
< info
->num_outputs
; i
++) {
2592 unsigned index
= info
->output_semantic_index
[i
];
2594 if (info
->output_semantic_name
[i
] != TGSI_SEMANTIC_COLOR
)
2597 /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
2598 if (shader
->key
.ps
.epilog
.last_cbuf
> 0) {
2599 /* Just set this if any of the colorbuffers are enabled. */
2601 ((1llu << (4 * (shader
->key
.ps
.epilog
.last_cbuf
+ 1))) - 1))
2602 last_color_export
= i
;
2606 if ((spi_format
>> (index
* 4)) & 0xf)
2607 last_color_export
= i
;
2610 /* If there are no outputs, export NULL. */
2611 if (last_color_export
== -1) {
2612 si_export_null(bld_base
);
2617 for (i
= 0; i
< info
->num_outputs
; i
++) {
2618 unsigned semantic_name
= info
->output_semantic_name
[i
];
2619 unsigned semantic_index
= info
->output_semantic_index
[i
];
2621 LLVMValueRef color
[4] = {};
2623 /* Select the correct target */
2624 switch (semantic_name
) {
2625 case TGSI_SEMANTIC_POSITION
:
2626 depth
= LLVMBuildLoad(builder
,
2627 ctx
->radeon_bld
.soa
.outputs
[i
][2], "");
2629 case TGSI_SEMANTIC_STENCIL
:
2630 stencil
= LLVMBuildLoad(builder
,
2631 ctx
->radeon_bld
.soa
.outputs
[i
][1], "");
2633 case TGSI_SEMANTIC_SAMPLEMASK
:
2634 samplemask
= LLVMBuildLoad(builder
,
2635 ctx
->radeon_bld
.soa
.outputs
[i
][0], "");
2637 case TGSI_SEMANTIC_COLOR
:
2638 for (j
= 0; j
< 4; j
++)
2639 color
[j
] = LLVMBuildLoad(builder
,
2640 ctx
->radeon_bld
.soa
.outputs
[i
][j
], "");
2642 si_export_mrt_color(bld_base
, color
, semantic_index
,
2643 SI_PARAM_SAMPLE_COVERAGE
,
2644 last_color_export
== i
);
2648 "Warning: SI unhandled fs output type:%d\n",
2653 if (depth
|| stencil
|| samplemask
)
2654 si_export_mrt_z(bld_base
, depth
, stencil
, samplemask
);
2658 * Return PS outputs in this order:
2660 * v[0:3] = color0.xyzw
2661 * v[4:7] = color1.xyzw
2666 * vN+3 = SampleMaskIn (used for OpenGL smoothing)
2668 * The alpha-ref SGPR is returned via its original location.
2670 static void si_llvm_return_fs_outputs(struct lp_build_tgsi_context
*bld_base
)
2672 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2673 struct si_shader
*shader
= ctx
->shader
;
2674 struct lp_build_context
*base
= &bld_base
->base
;
2675 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
2676 LLVMBuilderRef builder
= base
->gallivm
->builder
;
2677 unsigned i
, j
, first_vgpr
, vgpr
;
2679 LLVMValueRef color
[8][4] = {};
2680 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
2683 /* Read the output values. */
2684 for (i
= 0; i
< info
->num_outputs
; i
++) {
2685 unsigned semantic_name
= info
->output_semantic_name
[i
];
2686 unsigned semantic_index
= info
->output_semantic_index
[i
];
2688 switch (semantic_name
) {
2689 case TGSI_SEMANTIC_COLOR
:
2690 assert(semantic_index
< 8);
2691 for (j
= 0; j
< 4; j
++) {
2692 LLVMValueRef ptr
= ctx
->radeon_bld
.soa
.outputs
[i
][j
];
2693 LLVMValueRef result
= LLVMBuildLoad(builder
, ptr
, "");
2694 color
[semantic_index
][j
] = result
;
2697 case TGSI_SEMANTIC_POSITION
:
2698 depth
= LLVMBuildLoad(builder
,
2699 ctx
->radeon_bld
.soa
.outputs
[i
][2], "");
2701 case TGSI_SEMANTIC_STENCIL
:
2702 stencil
= LLVMBuildLoad(builder
,
2703 ctx
->radeon_bld
.soa
.outputs
[i
][1], "");
2705 case TGSI_SEMANTIC_SAMPLEMASK
:
2706 samplemask
= LLVMBuildLoad(builder
,
2707 ctx
->radeon_bld
.soa
.outputs
[i
][0], "");
2710 fprintf(stderr
, "Warning: SI unhandled fs output type:%d\n",
2715 /* Fill the return structure. */
2716 ret
= ctx
->return_value
;
2719 ret
= LLVMBuildInsertValue(builder
, ret
,
2720 bitcast(bld_base
, TGSI_TYPE_SIGNED
,
2721 LLVMGetParam(ctx
->radeon_bld
.main_fn
,
2722 SI_PARAM_ALPHA_REF
)),
2723 SI_SGPR_ALPHA_REF
, "");
2726 first_vgpr
= vgpr
= SI_SGPR_ALPHA_REF
+ 1;
2727 for (i
= 0; i
< ARRAY_SIZE(color
); i
++) {
2731 for (j
= 0; j
< 4; j
++)
2732 ret
= LLVMBuildInsertValue(builder
, ret
, color
[i
][j
], vgpr
++, "");
2735 ret
= LLVMBuildInsertValue(builder
, ret
, depth
, vgpr
++, "");
2737 ret
= LLVMBuildInsertValue(builder
, ret
, stencil
, vgpr
++, "");
2739 ret
= LLVMBuildInsertValue(builder
, ret
, samplemask
, vgpr
++, "");
2741 /* Add the input sample mask for smoothing at the end. */
2742 if (vgpr
< first_vgpr
+ PS_EPILOG_SAMPLEMASK_MIN_LOC
)
2743 vgpr
= first_vgpr
+ PS_EPILOG_SAMPLEMASK_MIN_LOC
;
2744 ret
= LLVMBuildInsertValue(builder
, ret
,
2745 LLVMGetParam(ctx
->radeon_bld
.main_fn
,
2746 SI_PARAM_SAMPLE_COVERAGE
), vgpr
++, "");
2748 ctx
->return_value
= ret
;
2752 * Given a v8i32 resource descriptor for a buffer, extract the size of the
2753 * buffer in number of elements and return it as an i32.
2755 static LLVMValueRef
get_buffer_size(
2756 struct lp_build_tgsi_context
*bld_base
,
2757 LLVMValueRef descriptor
)
2759 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2760 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2761 LLVMBuilderRef builder
= gallivm
->builder
;
2763 LLVMBuildExtractElement(builder
, descriptor
,
2764 lp_build_const_int32(gallivm
, 6), "");
2766 if (ctx
->screen
->b
.chip_class
>= VI
) {
2767 /* On VI, the descriptor contains the size in bytes,
2768 * but TXQ must return the size in elements.
2769 * The stride is always non-zero for resources using TXQ.
2771 LLVMValueRef stride
=
2772 LLVMBuildExtractElement(builder
, descriptor
,
2773 lp_build_const_int32(gallivm
, 5), "");
2774 stride
= LLVMBuildLShr(builder
, stride
,
2775 lp_build_const_int32(gallivm
, 16), "");
2776 stride
= LLVMBuildAnd(builder
, stride
,
2777 lp_build_const_int32(gallivm
, 0x3FFF), "");
2779 size
= LLVMBuildUDiv(builder
, size
, stride
, "");
2786 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
2789 static void build_int_type_name(
2791 char *buf
, unsigned bufsize
)
2793 assert(bufsize
>= 6);
2795 if (LLVMGetTypeKind(type
) == LLVMVectorTypeKind
)
2796 snprintf(buf
, bufsize
, "v%ui32",
2797 LLVMGetVectorSize(type
));
2802 static void build_tex_intrinsic(const struct lp_build_tgsi_action
*action
,
2803 struct lp_build_tgsi_context
*bld_base
,
2804 struct lp_build_emit_data
*emit_data
);
2806 /* Prevent optimizations (at least of memory accesses) across the current
2807 * point in the program by emitting empty inline assembly that is marked as
2808 * having side effects.
2810 static void emit_optimization_barrier(struct si_shader_context
*ctx
)
2812 LLVMBuilderRef builder
= ctx
->radeon_bld
.gallivm
.builder
;
2813 LLVMTypeRef ftype
= LLVMFunctionType(ctx
->voidt
, NULL
, 0, false);
2814 LLVMValueRef inlineasm
= LLVMConstInlineAsm(ftype
, "", "", true, false);
2815 LLVMBuildCall(builder
, inlineasm
, NULL
, 0, "");
2818 static void membar_emit(
2819 const struct lp_build_tgsi_action
*action
,
2820 struct lp_build_tgsi_context
*bld_base
,
2821 struct lp_build_emit_data
*emit_data
)
2823 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2825 /* Since memoryBarrier only makes guarantees about atomics and
2826 * coherent image accesses (which bypass TC L1), we do not need to emit
2827 * any special cache handling here.
2829 * We do have to prevent LLVM from re-ordering loads across
2830 * the barrier though.
2832 emit_optimization_barrier(ctx
);
2836 shader_buffer_fetch_rsrc(struct si_shader_context
*ctx
,
2837 const struct tgsi_full_src_register
*reg
)
2839 LLVMValueRef ind_index
;
2840 LLVMValueRef rsrc_ptr
;
2842 if (!reg
->Register
.Indirect
)
2843 return ctx
->shader_buffers
[reg
->Register
.Index
];
2845 ind_index
= get_bounded_indirect_index(ctx
, ®
->Indirect
,
2846 reg
->Register
.Index
,
2847 SI_NUM_SHADER_BUFFERS
);
2849 rsrc_ptr
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_SHADER_BUFFERS
);
2850 return build_indexed_load_const(ctx
, rsrc_ptr
, ind_index
);
2853 static bool tgsi_is_array_sampler(unsigned target
)
2855 return target
== TGSI_TEXTURE_1D_ARRAY
||
2856 target
== TGSI_TEXTURE_SHADOW1D_ARRAY
||
2857 target
== TGSI_TEXTURE_2D_ARRAY
||
2858 target
== TGSI_TEXTURE_SHADOW2D_ARRAY
||
2859 target
== TGSI_TEXTURE_CUBE_ARRAY
||
2860 target
== TGSI_TEXTURE_SHADOWCUBE_ARRAY
||
2861 target
== TGSI_TEXTURE_2D_ARRAY_MSAA
;
2864 static bool tgsi_is_array_image(unsigned target
)
2866 return target
== TGSI_TEXTURE_3D
||
2867 target
== TGSI_TEXTURE_CUBE
||
2868 target
== TGSI_TEXTURE_1D_ARRAY
||
2869 target
== TGSI_TEXTURE_2D_ARRAY
||
2870 target
== TGSI_TEXTURE_CUBE_ARRAY
||
2871 target
== TGSI_TEXTURE_2D_ARRAY_MSAA
;
2875 * Given a 256-bit resource descriptor, force the DCC enable bit to off.
2877 * At least on Tonga, executing image stores on images with DCC enabled and
2878 * non-trivial can eventually lead to lockups. This can occur when an
2879 * application binds an image as read-only but then uses a shader that writes
2880 * to it. The OpenGL spec allows almost arbitrarily bad behavior (including
2881 * program termination) in this case, but it doesn't cost much to be a bit
2882 * nicer: disabling DCC in the shader still leads to undefined results but
2883 * avoids the lockup.
2885 static LLVMValueRef
force_dcc_off(struct si_shader_context
*ctx
,
2888 if (ctx
->screen
->b
.chip_class
<= CIK
) {
2891 LLVMBuilderRef builder
= ctx
->radeon_bld
.gallivm
.builder
;
2892 LLVMValueRef i32_6
= LLVMConstInt(ctx
->i32
, 6, 0);
2893 LLVMValueRef i32_C
= LLVMConstInt(ctx
->i32
, C_008F28_COMPRESSION_EN
, 0);
2896 tmp
= LLVMBuildExtractElement(builder
, rsrc
, i32_6
, "");
2897 tmp
= LLVMBuildAnd(builder
, tmp
, i32_C
, "");
2898 return LLVMBuildInsertElement(builder
, rsrc
, tmp
, i32_6
, "");
2903 * Load the resource descriptor for \p image.
2907 struct lp_build_tgsi_context
*bld_base
,
2908 const struct tgsi_full_src_register
*image
,
2912 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
2914 assert(image
->Register
.File
== TGSI_FILE_IMAGE
);
2916 if (!image
->Register
.Indirect
) {
2917 /* Fast path: use preloaded resources */
2918 *rsrc
= ctx
->images
[image
->Register
.Index
];
2920 /* Indexing and manual load */
2921 LLVMValueRef ind_index
;
2922 LLVMValueRef rsrc_ptr
;
2925 /* From the GL_ARB_shader_image_load_store extension spec:
2927 * If a shader performs an image load, store, or atomic
2928 * operation using an image variable declared as an array,
2929 * and if the index used to select an individual element is
2930 * negative or greater than or equal to the size of the
2931 * array, the results of the operation are undefined but may
2932 * not lead to termination.
2934 ind_index
= get_bounded_indirect_index(ctx
, &image
->Indirect
,
2935 image
->Register
.Index
,
2938 rsrc_ptr
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_IMAGES
);
2939 tmp
= build_indexed_load_const(ctx
, rsrc_ptr
, ind_index
);
2941 tmp
= force_dcc_off(ctx
, tmp
);
2946 static LLVMValueRef
image_fetch_coords(
2947 struct lp_build_tgsi_context
*bld_base
,
2948 const struct tgsi_full_instruction
*inst
,
2951 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2952 LLVMBuilderRef builder
= gallivm
->builder
;
2953 unsigned target
= inst
->Memory
.Texture
;
2954 unsigned num_coords
= tgsi_util_get_texture_coord_dim(target
);
2955 LLVMValueRef coords
[4];
2959 for (chan
= 0; chan
< num_coords
; ++chan
) {
2960 tmp
= lp_build_emit_fetch(bld_base
, inst
, src
, chan
);
2961 tmp
= LLVMBuildBitCast(builder
, tmp
, bld_base
->uint_bld
.elem_type
, "");
2965 if (num_coords
== 1)
2968 if (num_coords
== 3) {
2969 /* LLVM has difficulties lowering 3-element vectors. */
2970 coords
[3] = bld_base
->uint_bld
.undef
;
2974 return lp_build_gather_values(gallivm
, coords
, num_coords
);
2978 * Append the extra mode bits that are used by image load and store.
2980 static void image_append_args(
2981 struct si_shader_context
*ctx
,
2982 struct lp_build_emit_data
* emit_data
,
2986 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
2987 LLVMValueRef i1false
= LLVMConstInt(ctx
->i1
, 0, 0);
2988 LLVMValueRef i1true
= LLVMConstInt(ctx
->i1
, 1, 0);
2990 emit_data
->args
[emit_data
->arg_count
++] = i1false
; /* r128 */
2991 emit_data
->args
[emit_data
->arg_count
++] =
2992 tgsi_is_array_image(target
) ? i1true
: i1false
; /* da */
2994 emit_data
->args
[emit_data
->arg_count
++] =
2995 inst
->Memory
.Qualifier
& (TGSI_MEMORY_COHERENT
| TGSI_MEMORY_VOLATILE
) ?
2996 i1true
: i1false
; /* glc */
2998 emit_data
->args
[emit_data
->arg_count
++] = i1false
; /* slc */
3002 * Given a 256 bit resource, extract the top half (which stores the buffer
3003 * resource in the case of textures and images).
3005 static LLVMValueRef
extract_rsrc_top_half(
3006 struct si_shader_context
*ctx
,
3009 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
3010 struct lp_build_tgsi_context
*bld_base
= &ctx
->radeon_bld
.soa
.bld_base
;
3011 LLVMTypeRef v2i128
= LLVMVectorType(ctx
->i128
, 2);
3013 rsrc
= LLVMBuildBitCast(gallivm
->builder
, rsrc
, v2i128
, "");
3014 rsrc
= LLVMBuildExtractElement(gallivm
->builder
, rsrc
, bld_base
->uint_bld
.one
, "");
3015 rsrc
= LLVMBuildBitCast(gallivm
->builder
, rsrc
, ctx
->v4i32
, "");
3021 * Append the resource and indexing arguments for buffer intrinsics.
3023 * \param rsrc the v4i32 buffer resource
3024 * \param index index into the buffer (stride-based)
3025 * \param offset byte offset into the buffer
3027 static void buffer_append_args(
3028 struct si_shader_context
*ctx
,
3029 struct lp_build_emit_data
*emit_data
,
3032 LLVMValueRef offset
,
3035 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3036 LLVMValueRef i1false
= LLVMConstInt(ctx
->i1
, 0, 0);
3037 LLVMValueRef i1true
= LLVMConstInt(ctx
->i1
, 1, 0);
3039 emit_data
->args
[emit_data
->arg_count
++] = rsrc
;
3040 emit_data
->args
[emit_data
->arg_count
++] = index
; /* vindex */
3041 emit_data
->args
[emit_data
->arg_count
++] = offset
; /* voffset */
3043 emit_data
->args
[emit_data
->arg_count
++] =
3044 inst
->Memory
.Qualifier
& (TGSI_MEMORY_COHERENT
| TGSI_MEMORY_VOLATILE
) ?
3045 i1true
: i1false
; /* glc */
3047 emit_data
->args
[emit_data
->arg_count
++] = i1false
; /* slc */
3050 static void load_fetch_args(
3051 struct lp_build_tgsi_context
* bld_base
,
3052 struct lp_build_emit_data
* emit_data
)
3054 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3055 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3056 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
3057 unsigned target
= inst
->Memory
.Texture
;
3060 emit_data
->dst_type
= LLVMVectorType(bld_base
->base
.elem_type
, 4);
3062 if (inst
->Src
[0].Register
.File
== TGSI_FILE_BUFFER
) {
3063 LLVMBuilderRef builder
= gallivm
->builder
;
3064 LLVMValueRef offset
;
3067 rsrc
= shader_buffer_fetch_rsrc(ctx
, &inst
->Src
[0]);
3069 tmp
= lp_build_emit_fetch(bld_base
, inst
, 1, 0);
3070 offset
= LLVMBuildBitCast(builder
, tmp
, bld_base
->uint_bld
.elem_type
, "");
3072 buffer_append_args(ctx
, emit_data
, rsrc
, bld_base
->uint_bld
.zero
,
3074 } else if (inst
->Src
[0].Register
.File
== TGSI_FILE_IMAGE
) {
3075 LLVMValueRef coords
;
3077 image_fetch_rsrc(bld_base
, &inst
->Src
[0], false, &rsrc
);
3078 coords
= image_fetch_coords(bld_base
, inst
, 1);
3080 if (target
== TGSI_TEXTURE_BUFFER
) {
3081 rsrc
= extract_rsrc_top_half(ctx
, rsrc
);
3082 buffer_append_args(ctx
, emit_data
, rsrc
, coords
,
3083 bld_base
->uint_bld
.zero
, false);
3085 emit_data
->args
[0] = coords
;
3086 emit_data
->args
[1] = rsrc
;
3087 emit_data
->args
[2] = lp_build_const_int32(gallivm
, 15); /* dmask */
3088 emit_data
->arg_count
= 3;
3090 image_append_args(ctx
, emit_data
, target
, false);
3095 static void load_emit_buffer(struct si_shader_context
*ctx
,
3096 struct lp_build_emit_data
*emit_data
)
3098 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3099 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
3100 LLVMBuilderRef builder
= gallivm
->builder
;
3101 uint writemask
= inst
->Dst
[0].Register
.WriteMask
;
3102 uint count
= util_last_bit(writemask
);
3103 const char *intrinsic_name
;
3104 LLVMTypeRef dst_type
;
3108 intrinsic_name
= "llvm.amdgcn.buffer.load.f32";
3109 dst_type
= ctx
->f32
;
3112 intrinsic_name
= "llvm.amdgcn.buffer.load.v2f32";
3113 dst_type
= LLVMVectorType(ctx
->f32
, 2);
3116 intrinsic_name
= "llvm.amdgcn.buffer.load.v4f32";
3117 dst_type
= ctx
->v4f32
;
3121 emit_data
->output
[emit_data
->chan
] = lp_build_intrinsic(
3122 builder
, intrinsic_name
, dst_type
,
3123 emit_data
->args
, emit_data
->arg_count
,
3124 LLVMReadOnlyAttribute
| LLVMNoUnwindAttribute
);
3127 static LLVMValueRef
get_memory_ptr(struct si_shader_context
*ctx
,
3128 const struct tgsi_full_instruction
*inst
,
3129 LLVMTypeRef type
, int arg
)
3131 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
3132 LLVMBuilderRef builder
= gallivm
->builder
;
3133 LLVMValueRef offset
, ptr
;
3136 offset
= lp_build_emit_fetch(&ctx
->radeon_bld
.soa
.bld_base
, inst
, arg
, 0);
3137 offset
= LLVMBuildBitCast(builder
, offset
, ctx
->i32
, "");
3139 ptr
= ctx
->shared_memory
;
3140 ptr
= LLVMBuildGEP(builder
, ptr
, &offset
, 1, "");
3141 addr_space
= LLVMGetPointerAddressSpace(LLVMTypeOf(ptr
));
3142 ptr
= LLVMBuildBitCast(builder
, ptr
, LLVMPointerType(type
, addr_space
), "");
3147 static void load_emit_memory(
3148 struct si_shader_context
*ctx
,
3149 struct lp_build_emit_data
*emit_data
)
3151 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3152 struct lp_build_context
*base
= &ctx
->radeon_bld
.soa
.bld_base
.base
;
3153 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
3154 LLVMBuilderRef builder
= gallivm
->builder
;
3155 unsigned writemask
= inst
->Dst
[0].Register
.WriteMask
;
3156 LLVMValueRef channels
[4], ptr
, derived_ptr
, index
;
3159 ptr
= get_memory_ptr(ctx
, inst
, base
->elem_type
, 1);
3161 for (chan
= 0; chan
< 4; ++chan
) {
3162 if (!(writemask
& (1 << chan
))) {
3163 channels
[chan
] = LLVMGetUndef(base
->elem_type
);
3167 index
= lp_build_const_int32(gallivm
, chan
);
3168 derived_ptr
= LLVMBuildGEP(builder
, ptr
, &index
, 1, "");
3169 channels
[chan
] = LLVMBuildLoad(builder
, derived_ptr
, "");
3171 emit_data
->output
[emit_data
->chan
] = lp_build_gather_values(gallivm
, channels
, 4);
3174 static void load_emit(
3175 const struct lp_build_tgsi_action
*action
,
3176 struct lp_build_tgsi_context
*bld_base
,
3177 struct lp_build_emit_data
*emit_data
)
3179 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3180 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3181 LLVMBuilderRef builder
= gallivm
->builder
;
3182 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
3183 char intrinsic_name
[32];
3184 char coords_type
[8];
3186 if (inst
->Src
[0].Register
.File
== TGSI_FILE_MEMORY
) {
3187 load_emit_memory(ctx
, emit_data
);
3191 if (inst
->Memory
.Qualifier
& TGSI_MEMORY_VOLATILE
)
3192 emit_optimization_barrier(ctx
);
3194 if (inst
->Src
[0].Register
.File
== TGSI_FILE_BUFFER
) {
3195 load_emit_buffer(ctx
, emit_data
);
3199 if (inst
->Memory
.Texture
== TGSI_TEXTURE_BUFFER
) {
3200 emit_data
->output
[emit_data
->chan
] =
3202 builder
, "llvm.amdgcn.buffer.load.format.v4f32", emit_data
->dst_type
,
3203 emit_data
->args
, emit_data
->arg_count
,
3204 LLVMReadOnlyAttribute
| LLVMNoUnwindAttribute
);
3206 build_int_type_name(LLVMTypeOf(emit_data
->args
[0]),
3207 coords_type
, sizeof(coords_type
));
3209 snprintf(intrinsic_name
, sizeof(intrinsic_name
),
3210 "llvm.amdgcn.image.load.%s", coords_type
);
3212 emit_data
->output
[emit_data
->chan
] =
3214 builder
, intrinsic_name
, emit_data
->dst_type
,
3215 emit_data
->args
, emit_data
->arg_count
,
3216 LLVMReadOnlyAttribute
| LLVMNoUnwindAttribute
);
3220 static void store_fetch_args(
3221 struct lp_build_tgsi_context
* bld_base
,
3222 struct lp_build_emit_data
* emit_data
)
3224 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3225 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3226 LLVMBuilderRef builder
= gallivm
->builder
;
3227 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
3228 struct tgsi_full_src_register memory
;
3229 LLVMValueRef chans
[4];
3234 emit_data
->dst_type
= LLVMVoidTypeInContext(gallivm
->context
);
3236 for (chan
= 0; chan
< 4; ++chan
) {
3237 chans
[chan
] = lp_build_emit_fetch(bld_base
, inst
, 1, chan
);
3239 data
= lp_build_gather_values(gallivm
, chans
, 4);
3241 emit_data
->args
[emit_data
->arg_count
++] = data
;
3243 memory
= tgsi_full_src_register_from_dst(&inst
->Dst
[0]);
3245 if (inst
->Dst
[0].Register
.File
== TGSI_FILE_BUFFER
) {
3246 LLVMValueRef offset
;
3249 rsrc
= shader_buffer_fetch_rsrc(ctx
, &memory
);
3251 tmp
= lp_build_emit_fetch(bld_base
, inst
, 0, 0);
3252 offset
= LLVMBuildBitCast(builder
, tmp
, bld_base
->uint_bld
.elem_type
, "");
3254 buffer_append_args(ctx
, emit_data
, rsrc
, bld_base
->uint_bld
.zero
,
3256 } else if (inst
->Dst
[0].Register
.File
== TGSI_FILE_IMAGE
) {
3257 unsigned target
= inst
->Memory
.Texture
;
3258 LLVMValueRef coords
;
3260 coords
= image_fetch_coords(bld_base
, inst
, 0);
3262 if (target
== TGSI_TEXTURE_BUFFER
) {
3263 image_fetch_rsrc(bld_base
, &memory
, false, &rsrc
);
3265 rsrc
= extract_rsrc_top_half(ctx
, rsrc
);
3266 buffer_append_args(ctx
, emit_data
, rsrc
, coords
,
3267 bld_base
->uint_bld
.zero
, false);
3269 emit_data
->args
[1] = coords
;
3270 image_fetch_rsrc(bld_base
, &memory
, true, &emit_data
->args
[2]);
3271 emit_data
->args
[3] = lp_build_const_int32(gallivm
, 15); /* dmask */
3272 emit_data
->arg_count
= 4;
3274 image_append_args(ctx
, emit_data
, target
, false);
3279 static void store_emit_buffer(
3280 struct si_shader_context
*ctx
,
3281 struct lp_build_emit_data
*emit_data
)
3283 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3284 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
3285 LLVMBuilderRef builder
= gallivm
->builder
;
3286 struct lp_build_context
*uint_bld
= &ctx
->radeon_bld
.soa
.bld_base
.uint_bld
;
3287 LLVMValueRef base_data
= emit_data
->args
[0];
3288 LLVMValueRef base_offset
= emit_data
->args
[3];
3289 unsigned writemask
= inst
->Dst
[0].Register
.WriteMask
;
3293 const char *intrinsic_name
;
3295 LLVMValueRef offset
;
3298 u_bit_scan_consecutive_range(&writemask
, &start
, &count
);
3300 /* Due to an LLVM limitation, split 3-element writes
3301 * into a 2-element and a 1-element write. */
3303 writemask
|= 1 << (start
+ 2);
3309 intrinsic_name
= "llvm.amdgcn.buffer.store.v4f32";
3310 } else if (count
== 2) {
3311 LLVMTypeRef v2f32
= LLVMVectorType(ctx
->f32
, 2);
3313 tmp
= LLVMBuildExtractElement(
3315 lp_build_const_int32(gallivm
, start
), "");
3316 data
= LLVMBuildInsertElement(
3317 builder
, LLVMGetUndef(v2f32
), tmp
,
3318 uint_bld
->zero
, "");
3320 tmp
= LLVMBuildExtractElement(
3322 lp_build_const_int32(gallivm
, start
+ 1), "");
3323 data
= LLVMBuildInsertElement(
3324 builder
, data
, tmp
, uint_bld
->one
, "");
3326 intrinsic_name
= "llvm.amdgcn.buffer.store.v2f32";
3329 data
= LLVMBuildExtractElement(
3331 lp_build_const_int32(gallivm
, start
), "");
3332 intrinsic_name
= "llvm.amdgcn.buffer.store.f32";
3335 offset
= base_offset
;
3337 offset
= LLVMBuildAdd(
3339 lp_build_const_int32(gallivm
, start
* 4), "");
3342 emit_data
->args
[0] = data
;
3343 emit_data
->args
[3] = offset
;
3346 builder
, intrinsic_name
, emit_data
->dst_type
,
3347 emit_data
->args
, emit_data
->arg_count
,
3348 LLVMNoUnwindAttribute
);
3352 static void store_emit_memory(
3353 struct si_shader_context
*ctx
,
3354 struct lp_build_emit_data
*emit_data
)
3356 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3357 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
3358 struct lp_build_context
*base
= &ctx
->radeon_bld
.soa
.bld_base
.base
;
3359 LLVMBuilderRef builder
= gallivm
->builder
;
3360 unsigned writemask
= inst
->Dst
[0].Register
.WriteMask
;
3361 LLVMValueRef ptr
, derived_ptr
, data
, index
;
3364 ptr
= get_memory_ptr(ctx
, inst
, base
->elem_type
, 0);
3366 for (chan
= 0; chan
< 4; ++chan
) {
3367 if (!(writemask
& (1 << chan
))) {
3370 data
= lp_build_emit_fetch(&ctx
->radeon_bld
.soa
.bld_base
, inst
, 1, chan
);
3371 index
= lp_build_const_int32(gallivm
, chan
);
3372 derived_ptr
= LLVMBuildGEP(builder
, ptr
, &index
, 1, "");
3373 LLVMBuildStore(builder
, data
, derived_ptr
);
3377 static void store_emit(
3378 const struct lp_build_tgsi_action
*action
,
3379 struct lp_build_tgsi_context
*bld_base
,
3380 struct lp_build_emit_data
*emit_data
)
3382 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3383 LLVMBuilderRef builder
= gallivm
->builder
;
3384 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
3385 unsigned target
= inst
->Memory
.Texture
;
3386 char intrinsic_name
[32];
3387 char coords_type
[8];
3389 if (inst
->Dst
[0].Register
.File
== TGSI_FILE_BUFFER
) {
3390 store_emit_buffer(si_shader_context(bld_base
), emit_data
);
3392 } else if (inst
->Dst
[0].Register
.File
== TGSI_FILE_MEMORY
) {
3393 store_emit_memory(si_shader_context(bld_base
), emit_data
);
3397 if (target
== TGSI_TEXTURE_BUFFER
) {
3398 emit_data
->output
[emit_data
->chan
] = lp_build_intrinsic(
3399 builder
, "llvm.amdgcn.buffer.store.format.v4f32",
3400 emit_data
->dst_type
, emit_data
->args
, emit_data
->arg_count
,
3401 LLVMNoUnwindAttribute
);
3403 build_int_type_name(LLVMTypeOf(emit_data
->args
[1]),
3404 coords_type
, sizeof(coords_type
));
3405 snprintf(intrinsic_name
, sizeof(intrinsic_name
),
3406 "llvm.amdgcn.image.store.%s", coords_type
);
3408 emit_data
->output
[emit_data
->chan
] =
3410 builder
, intrinsic_name
, emit_data
->dst_type
,
3411 emit_data
->args
, emit_data
->arg_count
,
3412 LLVMNoUnwindAttribute
);
3416 static void atomic_fetch_args(
3417 struct lp_build_tgsi_context
* bld_base
,
3418 struct lp_build_emit_data
* emit_data
)
3420 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3421 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3422 LLVMBuilderRef builder
= gallivm
->builder
;
3423 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
3424 LLVMValueRef data1
, data2
;
3428 emit_data
->dst_type
= bld_base
->base
.elem_type
;
3430 tmp
= lp_build_emit_fetch(bld_base
, inst
, 2, 0);
3431 data1
= LLVMBuildBitCast(builder
, tmp
, bld_base
->uint_bld
.elem_type
, "");
3433 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_ATOMCAS
) {
3434 tmp
= lp_build_emit_fetch(bld_base
, inst
, 3, 0);
3435 data2
= LLVMBuildBitCast(builder
, tmp
, bld_base
->uint_bld
.elem_type
, "");
3438 /* llvm.amdgcn.image/buffer.atomic.cmpswap reflect the hardware order
3439 * of arguments, which is reversed relative to TGSI (and GLSL)
3441 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_ATOMCAS
)
3442 emit_data
->args
[emit_data
->arg_count
++] = data2
;
3443 emit_data
->args
[emit_data
->arg_count
++] = data1
;
3445 if (inst
->Src
[0].Register
.File
== TGSI_FILE_BUFFER
) {
3446 LLVMValueRef offset
;
3448 rsrc
= shader_buffer_fetch_rsrc(ctx
, &inst
->Src
[0]);
3450 tmp
= lp_build_emit_fetch(bld_base
, inst
, 1, 0);
3451 offset
= LLVMBuildBitCast(builder
, tmp
, bld_base
->uint_bld
.elem_type
, "");
3453 buffer_append_args(ctx
, emit_data
, rsrc
, bld_base
->uint_bld
.zero
,
3455 } else if (inst
->Src
[0].Register
.File
== TGSI_FILE_IMAGE
) {
3456 unsigned target
= inst
->Memory
.Texture
;
3457 LLVMValueRef coords
;
3459 image_fetch_rsrc(bld_base
, &inst
->Src
[0],
3460 target
!= TGSI_TEXTURE_BUFFER
, &rsrc
);
3461 coords
= image_fetch_coords(bld_base
, inst
, 1);
3463 if (target
== TGSI_TEXTURE_BUFFER
) {
3464 rsrc
= extract_rsrc_top_half(ctx
, rsrc
);
3465 buffer_append_args(ctx
, emit_data
, rsrc
, coords
,
3466 bld_base
->uint_bld
.zero
, true);
3468 emit_data
->args
[emit_data
->arg_count
++] = coords
;
3469 emit_data
->args
[emit_data
->arg_count
++] = rsrc
;
3471 image_append_args(ctx
, emit_data
, target
, true);
3476 static void atomic_emit_memory(struct si_shader_context
*ctx
,
3477 struct lp_build_emit_data
*emit_data
) {
3478 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
3479 LLVMBuilderRef builder
= gallivm
->builder
;
3480 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
3481 LLVMValueRef ptr
, result
, arg
;
3483 ptr
= get_memory_ptr(ctx
, inst
, ctx
->i32
, 1);
3485 arg
= lp_build_emit_fetch(&ctx
->radeon_bld
.soa
.bld_base
, inst
, 2, 0);
3486 arg
= LLVMBuildBitCast(builder
, arg
, ctx
->i32
, "");
3488 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_ATOMCAS
) {
3489 LLVMValueRef new_data
;
3490 new_data
= lp_build_emit_fetch(&ctx
->radeon_bld
.soa
.bld_base
,
3493 new_data
= LLVMBuildBitCast(builder
, new_data
, ctx
->i32
, "");
3495 #if HAVE_LLVM >= 0x309
3496 result
= LLVMBuildAtomicCmpXchg(builder
, ptr
, arg
, new_data
,
3497 LLVMAtomicOrderingSequentiallyConsistent
,
3498 LLVMAtomicOrderingSequentiallyConsistent
,
3502 result
= LLVMBuildExtractValue(builder
, result
, 0, "");
3504 LLVMAtomicRMWBinOp op
;
3506 switch(inst
->Instruction
.Opcode
) {
3507 case TGSI_OPCODE_ATOMUADD
:
3508 op
= LLVMAtomicRMWBinOpAdd
;
3510 case TGSI_OPCODE_ATOMXCHG
:
3511 op
= LLVMAtomicRMWBinOpXchg
;
3513 case TGSI_OPCODE_ATOMAND
:
3514 op
= LLVMAtomicRMWBinOpAnd
;
3516 case TGSI_OPCODE_ATOMOR
:
3517 op
= LLVMAtomicRMWBinOpOr
;
3519 case TGSI_OPCODE_ATOMXOR
:
3520 op
= LLVMAtomicRMWBinOpXor
;
3522 case TGSI_OPCODE_ATOMUMIN
:
3523 op
= LLVMAtomicRMWBinOpUMin
;
3525 case TGSI_OPCODE_ATOMUMAX
:
3526 op
= LLVMAtomicRMWBinOpUMax
;
3528 case TGSI_OPCODE_ATOMIMIN
:
3529 op
= LLVMAtomicRMWBinOpMin
;
3531 case TGSI_OPCODE_ATOMIMAX
:
3532 op
= LLVMAtomicRMWBinOpMax
;
3535 unreachable("unknown atomic opcode");
3538 result
= LLVMBuildAtomicRMW(builder
, op
, ptr
, arg
,
3539 LLVMAtomicOrderingSequentiallyConsistent
,
3542 emit_data
->output
[emit_data
->chan
] = LLVMBuildBitCast(builder
, result
, emit_data
->dst_type
, "");
3545 static void atomic_emit(
3546 const struct lp_build_tgsi_action
*action
,
3547 struct lp_build_tgsi_context
*bld_base
,
3548 struct lp_build_emit_data
*emit_data
)
3550 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3551 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3552 LLVMBuilderRef builder
= gallivm
->builder
;
3553 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
3554 char intrinsic_name
[40];
3557 if (inst
->Src
[0].Register
.File
== TGSI_FILE_MEMORY
) {
3558 atomic_emit_memory(ctx
, emit_data
);
3562 if (inst
->Src
[0].Register
.File
== TGSI_FILE_BUFFER
||
3563 inst
->Memory
.Texture
== TGSI_TEXTURE_BUFFER
) {
3564 snprintf(intrinsic_name
, sizeof(intrinsic_name
),
3565 "llvm.amdgcn.buffer.atomic.%s", action
->intr_name
);
3567 char coords_type
[8];
3569 build_int_type_name(LLVMTypeOf(emit_data
->args
[1]),
3570 coords_type
, sizeof(coords_type
));
3571 snprintf(intrinsic_name
, sizeof(intrinsic_name
),
3572 "llvm.amdgcn.image.atomic.%s.%s",
3573 action
->intr_name
, coords_type
);
3576 tmp
= lp_build_intrinsic(
3577 builder
, intrinsic_name
, bld_base
->uint_bld
.elem_type
,
3578 emit_data
->args
, emit_data
->arg_count
,
3579 LLVMNoUnwindAttribute
);
3580 emit_data
->output
[emit_data
->chan
] =
3581 LLVMBuildBitCast(builder
, tmp
, bld_base
->base
.elem_type
, "");
3584 static void resq_fetch_args(
3585 struct lp_build_tgsi_context
* bld_base
,
3586 struct lp_build_emit_data
* emit_data
)
3588 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3589 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3590 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3591 const struct tgsi_full_src_register
*reg
= &inst
->Src
[0];
3593 emit_data
->dst_type
= LLVMVectorType(bld_base
->base
.elem_type
, 4);
3595 if (reg
->Register
.File
== TGSI_FILE_BUFFER
) {
3596 emit_data
->args
[0] = shader_buffer_fetch_rsrc(ctx
, reg
);
3597 emit_data
->arg_count
= 1;
3598 } else if (inst
->Memory
.Texture
== TGSI_TEXTURE_BUFFER
) {
3599 image_fetch_rsrc(bld_base
, reg
, false, &emit_data
->args
[0]);
3600 emit_data
->arg_count
= 1;
3602 emit_data
->args
[0] = bld_base
->uint_bld
.zero
; /* mip level */
3603 image_fetch_rsrc(bld_base
, reg
, false, &emit_data
->args
[1]);
3604 emit_data
->args
[2] = lp_build_const_int32(gallivm
, 15); /* dmask */
3605 emit_data
->args
[3] = bld_base
->uint_bld
.zero
; /* unorm */
3606 emit_data
->args
[4] = bld_base
->uint_bld
.zero
; /* r128 */
3607 emit_data
->args
[5] = tgsi_is_array_image(inst
->Memory
.Texture
) ?
3608 bld_base
->uint_bld
.one
: bld_base
->uint_bld
.zero
; /* da */
3609 emit_data
->args
[6] = bld_base
->uint_bld
.zero
; /* glc */
3610 emit_data
->args
[7] = bld_base
->uint_bld
.zero
; /* slc */
3611 emit_data
->args
[8] = bld_base
->uint_bld
.zero
; /* tfe */
3612 emit_data
->args
[9] = bld_base
->uint_bld
.zero
; /* lwe */
3613 emit_data
->arg_count
= 10;
3617 static void resq_emit(
3618 const struct lp_build_tgsi_action
*action
,
3619 struct lp_build_tgsi_context
*bld_base
,
3620 struct lp_build_emit_data
*emit_data
)
3622 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3623 LLVMBuilderRef builder
= gallivm
->builder
;
3624 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3627 if (inst
->Src
[0].Register
.File
== TGSI_FILE_BUFFER
) {
3628 out
= LLVMBuildExtractElement(builder
, emit_data
->args
[0],
3629 lp_build_const_int32(gallivm
, 2), "");
3630 } else if (inst
->Memory
.Texture
== TGSI_TEXTURE_BUFFER
) {
3631 out
= get_buffer_size(bld_base
, emit_data
->args
[0]);
3633 out
= lp_build_intrinsic(
3634 builder
, "llvm.SI.getresinfo.i32", emit_data
->dst_type
,
3635 emit_data
->args
, emit_data
->arg_count
,
3636 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
3638 /* Divide the number of layers by 6 to get the number of cubes. */
3639 if (inst
->Memory
.Texture
== TGSI_TEXTURE_CUBE_ARRAY
) {
3640 LLVMValueRef imm2
= lp_build_const_int32(gallivm
, 2);
3641 LLVMValueRef imm6
= lp_build_const_int32(gallivm
, 6);
3643 LLVMValueRef z
= LLVMBuildExtractElement(builder
, out
, imm2
, "");
3644 z
= LLVMBuildBitCast(builder
, z
, bld_base
->uint_bld
.elem_type
, "");
3645 z
= LLVMBuildSDiv(builder
, z
, imm6
, "");
3646 z
= LLVMBuildBitCast(builder
, z
, bld_base
->base
.elem_type
, "");
3647 out
= LLVMBuildInsertElement(builder
, out
, z
, imm2
, "");
3651 emit_data
->output
[emit_data
->chan
] = out
;
3654 static void set_tex_fetch_args(struct si_shader_context
*ctx
,
3655 struct lp_build_emit_data
*emit_data
,
3656 unsigned opcode
, unsigned target
,
3657 LLVMValueRef res_ptr
, LLVMValueRef samp_ptr
,
3658 LLVMValueRef
*param
, unsigned count
,
3661 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
3663 unsigned is_rect
= target
== TGSI_TEXTURE_RECT
;
3665 /* Pad to power of two vector */
3666 while (count
< util_next_power_of_two(count
))
3667 param
[count
++] = LLVMGetUndef(ctx
->i32
);
3669 /* Texture coordinates. */
3671 emit_data
->args
[0] = lp_build_gather_values(gallivm
, param
, count
);
3673 emit_data
->args
[0] = param
[0];
3676 emit_data
->args
[1] = res_ptr
;
3679 if (opcode
== TGSI_OPCODE_TXF
|| opcode
== TGSI_OPCODE_TXQ
)
3680 emit_data
->dst_type
= ctx
->v4i32
;
3682 emit_data
->dst_type
= ctx
->v4f32
;
3684 emit_data
->args
[num_args
++] = samp_ptr
;
3687 emit_data
->args
[num_args
++] = lp_build_const_int32(gallivm
, dmask
);
3688 emit_data
->args
[num_args
++] = lp_build_const_int32(gallivm
, is_rect
); /* unorm */
3689 emit_data
->args
[num_args
++] = lp_build_const_int32(gallivm
, 0); /* r128 */
3690 emit_data
->args
[num_args
++] = lp_build_const_int32(gallivm
,
3691 tgsi_is_array_sampler(target
)); /* da */
3692 emit_data
->args
[num_args
++] = lp_build_const_int32(gallivm
, 0); /* glc */
3693 emit_data
->args
[num_args
++] = lp_build_const_int32(gallivm
, 0); /* slc */
3694 emit_data
->args
[num_args
++] = lp_build_const_int32(gallivm
, 0); /* tfe */
3695 emit_data
->args
[num_args
++] = lp_build_const_int32(gallivm
, 0); /* lwe */
3697 emit_data
->arg_count
= num_args
;
3700 static const struct lp_build_tgsi_action tex_action
;
3708 static LLVMTypeRef
const_array(LLVMTypeRef elem_type
, int num_elements
)
3710 return LLVMPointerType(LLVMArrayType(elem_type
, num_elements
),
3715 * Load an image view, fmask view. or sampler state descriptor.
3717 static LLVMValueRef
get_sampler_desc_custom(struct si_shader_context
*ctx
,
3718 LLVMValueRef list
, LLVMValueRef index
,
3719 enum desc_type type
)
3721 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
3722 LLVMBuilderRef builder
= gallivm
->builder
;
3726 /* The image is at [0:7]. */
3727 index
= LLVMBuildMul(builder
, index
, LLVMConstInt(ctx
->i32
, 2, 0), "");
3730 /* The FMASK is at [8:15]. */
3731 index
= LLVMBuildMul(builder
, index
, LLVMConstInt(ctx
->i32
, 2, 0), "");
3732 index
= LLVMBuildAdd(builder
, index
, LLVMConstInt(ctx
->i32
, 1, 0), "");
3735 /* The sampler state is at [12:15]. */
3736 index
= LLVMBuildMul(builder
, index
, LLVMConstInt(ctx
->i32
, 4, 0), "");
3737 index
= LLVMBuildAdd(builder
, index
, LLVMConstInt(ctx
->i32
, 3, 0), "");
3738 list
= LLVMBuildPointerCast(builder
, list
,
3739 const_array(ctx
->v4i32
, 0), "");
3743 return build_indexed_load_const(ctx
, list
, index
);
3746 static LLVMValueRef
get_sampler_desc(struct si_shader_context
*ctx
,
3747 LLVMValueRef index
, enum desc_type type
)
3749 LLVMValueRef list
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
3752 return get_sampler_desc_custom(ctx
, list
, index
, type
);
3755 /* Disable anisotropic filtering if BASE_LEVEL == LAST_LEVEL.
3758 * If BASE_LEVEL == LAST_LEVEL, the shader must disable anisotropic
3759 * filtering manually. The driver sets img7 to a mask clearing
3760 * MAX_ANISO_RATIO if BASE_LEVEL == LAST_LEVEL. The shader must do:
3761 * s_and_b32 samp0, samp0, img7
3764 * The ANISO_OVERRIDE sampler field enables this fix in TA.
3766 static LLVMValueRef
sici_fix_sampler_aniso(struct si_shader_context
*ctx
,
3767 LLVMValueRef res
, LLVMValueRef samp
)
3769 LLVMBuilderRef builder
= ctx
->radeon_bld
.gallivm
.builder
;
3770 LLVMValueRef img7
, samp0
;
3772 if (ctx
->screen
->b
.chip_class
>= VI
)
3775 img7
= LLVMBuildExtractElement(builder
, res
,
3776 LLVMConstInt(ctx
->i32
, 7, 0), "");
3777 samp0
= LLVMBuildExtractElement(builder
, samp
,
3778 LLVMConstInt(ctx
->i32
, 0, 0), "");
3779 samp0
= LLVMBuildAnd(builder
, samp0
, img7
, "");
3780 return LLVMBuildInsertElement(builder
, samp
, samp0
,
3781 LLVMConstInt(ctx
->i32
, 0, 0), "");
3784 static void tex_fetch_ptrs(
3785 struct lp_build_tgsi_context
*bld_base
,
3786 struct lp_build_emit_data
*emit_data
,
3787 LLVMValueRef
*res_ptr
, LLVMValueRef
*samp_ptr
, LLVMValueRef
*fmask_ptr
)
3789 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3790 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3791 unsigned target
= inst
->Texture
.Texture
;
3792 unsigned sampler_src
;
3793 unsigned sampler_index
;
3795 sampler_src
= emit_data
->inst
->Instruction
.NumSrcRegs
- 1;
3796 sampler_index
= emit_data
->inst
->Src
[sampler_src
].Register
.Index
;
3798 if (emit_data
->inst
->Src
[sampler_src
].Register
.Indirect
) {
3799 const struct tgsi_full_src_register
*reg
= &emit_data
->inst
->Src
[sampler_src
];
3800 LLVMValueRef ind_index
;
3802 ind_index
= get_bounded_indirect_index(ctx
,
3804 reg
->Register
.Index
,
3805 SI_NUM_USER_SAMPLERS
);
3807 *res_ptr
= get_sampler_desc(ctx
, ind_index
, DESC_IMAGE
);
3809 if (target
== TGSI_TEXTURE_2D_MSAA
||
3810 target
== TGSI_TEXTURE_2D_ARRAY_MSAA
) {
3812 *fmask_ptr
= get_sampler_desc(ctx
, ind_index
, DESC_FMASK
);
3814 *samp_ptr
= get_sampler_desc(ctx
, ind_index
, DESC_SAMPLER
);
3815 *samp_ptr
= sici_fix_sampler_aniso(ctx
, *res_ptr
, *samp_ptr
);
3819 *res_ptr
= ctx
->sampler_views
[sampler_index
];
3820 *samp_ptr
= ctx
->sampler_states
[sampler_index
];
3821 *fmask_ptr
= ctx
->fmasks
[sampler_index
];
3825 static void tex_fetch_args(
3826 struct lp_build_tgsi_context
*bld_base
,
3827 struct lp_build_emit_data
*emit_data
)
3829 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
3830 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3831 LLVMBuilderRef builder
= gallivm
->builder
;
3832 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3833 unsigned opcode
= inst
->Instruction
.Opcode
;
3834 unsigned target
= inst
->Texture
.Texture
;
3835 LLVMValueRef coords
[5], derivs
[6];
3836 LLVMValueRef address
[16];
3837 unsigned num_coords
= tgsi_util_get_texture_coord_dim(target
);
3838 int ref_pos
= tgsi_util_get_shadow_ref_src_index(target
);
3841 unsigned num_deriv_channels
= 0;
3842 bool has_offset
= inst
->Texture
.NumOffsets
> 0;
3843 LLVMValueRef res_ptr
, samp_ptr
, fmask_ptr
= NULL
;
3844 unsigned dmask
= 0xf;
3846 tex_fetch_ptrs(bld_base
, emit_data
, &res_ptr
, &samp_ptr
, &fmask_ptr
);
3848 if (opcode
== TGSI_OPCODE_TXQ
) {
3849 if (target
== TGSI_TEXTURE_BUFFER
) {
3850 /* Read the size from the buffer descriptor directly. */
3851 LLVMValueRef res
= LLVMBuildBitCast(builder
, res_ptr
, ctx
->v8i32
, "");
3852 emit_data
->args
[0] = get_buffer_size(bld_base
, res
);
3856 /* Textures - set the mip level. */
3857 address
[count
++] = lp_build_emit_fetch(bld_base
, inst
, 0, TGSI_CHAN_X
);
3859 set_tex_fetch_args(ctx
, emit_data
, opcode
, target
, res_ptr
,
3860 NULL
, address
, count
, 0xf);
3864 if (target
== TGSI_TEXTURE_BUFFER
) {
3865 LLVMTypeRef v2i128
= LLVMVectorType(ctx
->i128
, 2);
3867 /* Bitcast and truncate v8i32 to v16i8. */
3868 LLVMValueRef res
= res_ptr
;
3869 res
= LLVMBuildBitCast(gallivm
->builder
, res
, v2i128
, "");
3870 res
= LLVMBuildExtractElement(gallivm
->builder
, res
, bld_base
->uint_bld
.one
, "");
3871 res
= LLVMBuildBitCast(gallivm
->builder
, res
, ctx
->v16i8
, "");
3873 emit_data
->dst_type
= ctx
->v4f32
;
3874 emit_data
->args
[0] = res
;
3875 emit_data
->args
[1] = bld_base
->uint_bld
.zero
;
3876 emit_data
->args
[2] = lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, TGSI_CHAN_X
);
3877 emit_data
->arg_count
= 3;
3881 /* Fetch and project texture coordinates */
3882 coords
[3] = lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, TGSI_CHAN_W
);
3883 for (chan
= 0; chan
< 3; chan
++ ) {
3884 coords
[chan
] = lp_build_emit_fetch(bld_base
,
3887 if (opcode
== TGSI_OPCODE_TXP
)
3888 coords
[chan
] = lp_build_emit_llvm_binary(bld_base
,
3894 if (opcode
== TGSI_OPCODE_TXP
)
3895 coords
[3] = bld_base
->base
.one
;
3898 if (has_offset
&& opcode
!= TGSI_OPCODE_TXF
) {
3899 /* The offsets are six-bit signed integers packed like this:
3900 * X=[5:0], Y=[13:8], and Z=[21:16].
3902 LLVMValueRef offset
[3], pack
;
3904 assert(inst
->Texture
.NumOffsets
== 1);
3906 for (chan
= 0; chan
< 3; chan
++) {
3907 offset
[chan
] = lp_build_emit_fetch_texoffset(bld_base
,
3908 emit_data
->inst
, 0, chan
);
3909 offset
[chan
] = LLVMBuildAnd(gallivm
->builder
, offset
[chan
],
3910 lp_build_const_int32(gallivm
, 0x3f), "");
3912 offset
[chan
] = LLVMBuildShl(gallivm
->builder
, offset
[chan
],
3913 lp_build_const_int32(gallivm
, chan
*8), "");
3916 pack
= LLVMBuildOr(gallivm
->builder
, offset
[0], offset
[1], "");
3917 pack
= LLVMBuildOr(gallivm
->builder
, pack
, offset
[2], "");
3918 address
[count
++] = pack
;
3921 /* Pack LOD bias value */
3922 if (opcode
== TGSI_OPCODE_TXB
)
3923 address
[count
++] = coords
[3];
3924 if (opcode
== TGSI_OPCODE_TXB2
)
3925 address
[count
++] = lp_build_emit_fetch(bld_base
, inst
, 1, TGSI_CHAN_X
);
3927 /* Pack depth comparison value */
3928 if (tgsi_is_shadow_target(target
) && opcode
!= TGSI_OPCODE_LODQ
) {
3929 if (target
== TGSI_TEXTURE_SHADOWCUBE_ARRAY
) {
3930 address
[count
++] = lp_build_emit_fetch(bld_base
, inst
, 1, TGSI_CHAN_X
);
3932 assert(ref_pos
>= 0);
3933 address
[count
++] = coords
[ref_pos
];
3937 /* Pack user derivatives */
3938 if (opcode
== TGSI_OPCODE_TXD
) {
3939 int param
, num_src_deriv_channels
;
3942 case TGSI_TEXTURE_3D
:
3943 num_src_deriv_channels
= 3;
3944 num_deriv_channels
= 3;
3946 case TGSI_TEXTURE_2D
:
3947 case TGSI_TEXTURE_SHADOW2D
:
3948 case TGSI_TEXTURE_RECT
:
3949 case TGSI_TEXTURE_SHADOWRECT
:
3950 case TGSI_TEXTURE_2D_ARRAY
:
3951 case TGSI_TEXTURE_SHADOW2D_ARRAY
:
3952 num_src_deriv_channels
= 2;
3953 num_deriv_channels
= 2;
3955 case TGSI_TEXTURE_CUBE
:
3956 case TGSI_TEXTURE_SHADOWCUBE
:
3957 case TGSI_TEXTURE_CUBE_ARRAY
:
3958 case TGSI_TEXTURE_SHADOWCUBE_ARRAY
:
3959 /* Cube derivatives will be converted to 2D. */
3960 num_src_deriv_channels
= 3;
3961 num_deriv_channels
= 2;
3963 case TGSI_TEXTURE_1D
:
3964 case TGSI_TEXTURE_SHADOW1D
:
3965 case TGSI_TEXTURE_1D_ARRAY
:
3966 case TGSI_TEXTURE_SHADOW1D_ARRAY
:
3967 num_src_deriv_channels
= 1;
3968 num_deriv_channels
= 1;
3971 unreachable("invalid target");
3974 for (param
= 0; param
< 2; param
++)
3975 for (chan
= 0; chan
< num_src_deriv_channels
; chan
++)
3976 derivs
[param
* num_src_deriv_channels
+ chan
] =
3977 lp_build_emit_fetch(bld_base
, inst
, param
+1, chan
);
3980 if (target
== TGSI_TEXTURE_CUBE
||
3981 target
== TGSI_TEXTURE_CUBE_ARRAY
||
3982 target
== TGSI_TEXTURE_SHADOWCUBE
||
3983 target
== TGSI_TEXTURE_SHADOWCUBE_ARRAY
)
3984 radeon_llvm_emit_prepare_cube_coords(bld_base
, emit_data
, coords
, derivs
);
3986 if (opcode
== TGSI_OPCODE_TXD
)
3987 for (int i
= 0; i
< num_deriv_channels
* 2; i
++)
3988 address
[count
++] = derivs
[i
];
3990 /* Pack texture coordinates */
3991 address
[count
++] = coords
[0];
3993 address
[count
++] = coords
[1];
3995 address
[count
++] = coords
[2];
3997 /* Pack LOD or sample index */
3998 if (opcode
== TGSI_OPCODE_TXL
|| opcode
== TGSI_OPCODE_TXF
)
3999 address
[count
++] = coords
[3];
4000 else if (opcode
== TGSI_OPCODE_TXL2
)
4001 address
[count
++] = lp_build_emit_fetch(bld_base
, inst
, 1, TGSI_CHAN_X
);
4004 assert(!"Cannot handle more than 16 texture address parameters");
4008 for (chan
= 0; chan
< count
; chan
++ ) {
4009 address
[chan
] = LLVMBuildBitCast(gallivm
->builder
,
4010 address
[chan
], ctx
->i32
, "");
4013 /* Adjust the sample index according to FMASK.
4015 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
4016 * which is the identity mapping. Each nibble says which physical sample
4017 * should be fetched to get that sample.
4019 * For example, 0x11111100 means there are only 2 samples stored and
4020 * the second sample covers 3/4 of the pixel. When reading samples 0
4021 * and 1, return physical sample 0 (determined by the first two 0s
4022 * in FMASK), otherwise return physical sample 1.
4024 * The sample index should be adjusted as follows:
4025 * sample_index = (fmask >> (sample_index * 4)) & 0xF;
4027 if (target
== TGSI_TEXTURE_2D_MSAA
||
4028 target
== TGSI_TEXTURE_2D_ARRAY_MSAA
) {
4029 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
4030 struct lp_build_emit_data txf_emit_data
= *emit_data
;
4031 LLVMValueRef txf_address
[4];
4032 unsigned txf_count
= count
;
4033 struct tgsi_full_instruction inst
= {};
4035 memcpy(txf_address
, address
, sizeof(txf_address
));
4037 if (target
== TGSI_TEXTURE_2D_MSAA
) {
4038 txf_address
[2] = bld_base
->uint_bld
.zero
;
4040 txf_address
[3] = bld_base
->uint_bld
.zero
;
4042 /* Read FMASK using TXF. */
4043 inst
.Instruction
.Opcode
= TGSI_OPCODE_TXF
;
4044 inst
.Texture
.Texture
= target
;
4045 txf_emit_data
.inst
= &inst
;
4046 txf_emit_data
.chan
= 0;
4047 set_tex_fetch_args(ctx
, &txf_emit_data
, TGSI_OPCODE_TXF
,
4048 target
, fmask_ptr
, NULL
,
4049 txf_address
, txf_count
, 0xf);
4050 build_tex_intrinsic(&tex_action
, bld_base
, &txf_emit_data
);
4052 /* Initialize some constants. */
4053 LLVMValueRef four
= LLVMConstInt(ctx
->i32
, 4, 0);
4054 LLVMValueRef F
= LLVMConstInt(ctx
->i32
, 0xF, 0);
4056 /* Apply the formula. */
4057 LLVMValueRef fmask
=
4058 LLVMBuildExtractElement(gallivm
->builder
,
4059 txf_emit_data
.output
[0],
4060 uint_bld
->zero
, "");
4062 unsigned sample_chan
= target
== TGSI_TEXTURE_2D_MSAA
? 2 : 3;
4064 LLVMValueRef sample_index4
=
4065 LLVMBuildMul(gallivm
->builder
, address
[sample_chan
], four
, "");
4067 LLVMValueRef shifted_fmask
=
4068 LLVMBuildLShr(gallivm
->builder
, fmask
, sample_index4
, "");
4070 LLVMValueRef final_sample
=
4071 LLVMBuildAnd(gallivm
->builder
, shifted_fmask
, F
, "");
4073 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
4074 * resource descriptor is 0 (invalid),
4076 LLVMValueRef fmask_desc
=
4077 LLVMBuildBitCast(gallivm
->builder
, fmask_ptr
,
4080 LLVMValueRef fmask_word1
=
4081 LLVMBuildExtractElement(gallivm
->builder
, fmask_desc
,
4084 LLVMValueRef word1_is_nonzero
=
4085 LLVMBuildICmp(gallivm
->builder
, LLVMIntNE
,
4086 fmask_word1
, uint_bld
->zero
, "");
4088 /* Replace the MSAA sample index. */
4089 address
[sample_chan
] =
4090 LLVMBuildSelect(gallivm
->builder
, word1_is_nonzero
,
4091 final_sample
, address
[sample_chan
], "");
4094 if (opcode
== TGSI_OPCODE_TXF
) {
4095 /* add tex offsets */
4096 if (inst
->Texture
.NumOffsets
) {
4097 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
4098 struct lp_build_tgsi_soa_context
*bld
= lp_soa_context(bld_base
);
4099 const struct tgsi_texture_offset
*off
= inst
->TexOffsets
;
4101 assert(inst
->Texture
.NumOffsets
== 1);
4104 case TGSI_TEXTURE_3D
:
4105 address
[2] = lp_build_add(uint_bld
, address
[2],
4106 bld
->immediates
[off
->Index
][off
->SwizzleZ
]);
4108 case TGSI_TEXTURE_2D
:
4109 case TGSI_TEXTURE_SHADOW2D
:
4110 case TGSI_TEXTURE_RECT
:
4111 case TGSI_TEXTURE_SHADOWRECT
:
4112 case TGSI_TEXTURE_2D_ARRAY
:
4113 case TGSI_TEXTURE_SHADOW2D_ARRAY
:
4115 lp_build_add(uint_bld
, address
[1],
4116 bld
->immediates
[off
->Index
][off
->SwizzleY
]);
4118 case TGSI_TEXTURE_1D
:
4119 case TGSI_TEXTURE_SHADOW1D
:
4120 case TGSI_TEXTURE_1D_ARRAY
:
4121 case TGSI_TEXTURE_SHADOW1D_ARRAY
:
4123 lp_build_add(uint_bld
, address
[0],
4124 bld
->immediates
[off
->Index
][off
->SwizzleX
]);
4126 /* texture offsets do not apply to other texture targets */
4131 if (opcode
== TGSI_OPCODE_TG4
) {
4132 unsigned gather_comp
= 0;
4134 /* DMASK was repurposed for GATHER4. 4 components are always
4135 * returned and DMASK works like a swizzle - it selects
4136 * the component to fetch. The only valid DMASK values are
4137 * 1=red, 2=green, 4=blue, 8=alpha. (e.g. 1 returns
4138 * (red,red,red,red) etc.) The ISA document doesn't mention
4142 /* Get the component index from src1.x for Gather4. */
4143 if (!tgsi_is_shadow_target(target
)) {
4144 LLVMValueRef (*imms
)[4] = lp_soa_context(bld_base
)->immediates
;
4145 LLVMValueRef comp_imm
;
4146 struct tgsi_src_register src1
= inst
->Src
[1].Register
;
4148 assert(src1
.File
== TGSI_FILE_IMMEDIATE
);
4150 comp_imm
= imms
[src1
.Index
][src1
.SwizzleX
];
4151 gather_comp
= LLVMConstIntGetZExtValue(comp_imm
);
4152 gather_comp
= CLAMP(gather_comp
, 0, 3);
4155 dmask
= 1 << gather_comp
;
4158 set_tex_fetch_args(ctx
, emit_data
, opcode
, target
, res_ptr
,
4159 samp_ptr
, address
, count
, dmask
);
4162 static void build_tex_intrinsic(const struct lp_build_tgsi_action
*action
,
4163 struct lp_build_tgsi_context
*bld_base
,
4164 struct lp_build_emit_data
*emit_data
)
4166 struct lp_build_context
*base
= &bld_base
->base
;
4167 unsigned opcode
= emit_data
->inst
->Instruction
.Opcode
;
4168 unsigned target
= emit_data
->inst
->Texture
.Texture
;
4169 char intr_name
[127];
4170 bool has_offset
= emit_data
->inst
->Texture
.NumOffsets
> 0;
4171 bool is_shadow
= tgsi_is_shadow_target(target
);
4173 const char *name
= "llvm.SI.image.sample";
4174 const char *infix
= "";
4176 if (opcode
== TGSI_OPCODE_TXQ
&& target
== TGSI_TEXTURE_BUFFER
) {
4177 /* Just return the buffer size. */
4178 emit_data
->output
[emit_data
->chan
] = emit_data
->args
[0];
4182 if (target
== TGSI_TEXTURE_BUFFER
) {
4183 emit_data
->output
[emit_data
->chan
] = lp_build_intrinsic(
4184 base
->gallivm
->builder
,
4185 "llvm.SI.vs.load.input", emit_data
->dst_type
,
4186 emit_data
->args
, emit_data
->arg_count
,
4187 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
4192 case TGSI_OPCODE_TXF
:
4193 name
= target
== TGSI_TEXTURE_2D_MSAA
||
4194 target
== TGSI_TEXTURE_2D_ARRAY_MSAA
?
4195 "llvm.SI.image.load" :
4196 "llvm.SI.image.load.mip";
4200 case TGSI_OPCODE_TXQ
:
4201 name
= "llvm.SI.getresinfo";
4205 case TGSI_OPCODE_LODQ
:
4206 name
= "llvm.SI.getlod";
4210 case TGSI_OPCODE_TEX
:
4211 case TGSI_OPCODE_TEX2
:
4212 case TGSI_OPCODE_TXP
:
4214 case TGSI_OPCODE_TXB
:
4215 case TGSI_OPCODE_TXB2
:
4218 case TGSI_OPCODE_TXL
:
4219 case TGSI_OPCODE_TXL2
:
4222 case TGSI_OPCODE_TXD
:
4225 case TGSI_OPCODE_TG4
:
4226 name
= "llvm.SI.gather4";
4233 /* Add the type and suffixes .c, .o if needed. */
4234 build_int_type_name(LLVMTypeOf(emit_data
->args
[0]), type
, sizeof(type
));
4235 sprintf(intr_name
, "%s%s%s%s.%s",
4236 name
, is_shadow
? ".c" : "", infix
,
4237 has_offset
? ".o" : "", type
);
4239 emit_data
->output
[emit_data
->chan
] = lp_build_intrinsic(
4240 base
->gallivm
->builder
, intr_name
, emit_data
->dst_type
,
4241 emit_data
->args
, emit_data
->arg_count
,
4242 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
4244 /* Divide the number of layers by 6 to get the number of cubes. */
4245 if (opcode
== TGSI_OPCODE_TXQ
&&
4246 (target
== TGSI_TEXTURE_CUBE_ARRAY
||
4247 target
== TGSI_TEXTURE_SHADOWCUBE_ARRAY
)) {
4248 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
4249 LLVMValueRef two
= lp_build_const_int32(bld_base
->base
.gallivm
, 2);
4250 LLVMValueRef six
= lp_build_const_int32(bld_base
->base
.gallivm
, 6);
4252 LLVMValueRef v4
= emit_data
->output
[emit_data
->chan
];
4253 LLVMValueRef z
= LLVMBuildExtractElement(builder
, v4
, two
, "");
4254 z
= LLVMBuildSDiv(builder
, z
, six
, "");
4256 emit_data
->output
[emit_data
->chan
] =
4257 LLVMBuildInsertElement(builder
, v4
, z
, two
, "");
4261 static void si_llvm_emit_txqs(
4262 const struct lp_build_tgsi_action
*action
,
4263 struct lp_build_tgsi_context
*bld_base
,
4264 struct lp_build_emit_data
*emit_data
)
4266 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4267 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4268 LLVMBuilderRef builder
= gallivm
->builder
;
4269 LLVMValueRef res
, samples
;
4270 LLVMValueRef res_ptr
, samp_ptr
, fmask_ptr
= NULL
;
4272 tex_fetch_ptrs(bld_base
, emit_data
, &res_ptr
, &samp_ptr
, &fmask_ptr
);
4275 /* Read the samples from the descriptor directly. */
4276 res
= LLVMBuildBitCast(builder
, res_ptr
, ctx
->v8i32
, "");
4277 samples
= LLVMBuildExtractElement(
4279 lp_build_const_int32(gallivm
, 3), "");
4280 samples
= LLVMBuildLShr(builder
, samples
,
4281 lp_build_const_int32(gallivm
, 16), "");
4282 samples
= LLVMBuildAnd(builder
, samples
,
4283 lp_build_const_int32(gallivm
, 0xf), "");
4284 samples
= LLVMBuildShl(builder
, lp_build_const_int32(gallivm
, 1),
4287 emit_data
->output
[emit_data
->chan
] = samples
;
4291 * SI implements derivatives using the local data store (LDS)
4292 * All writes to the LDS happen in all executing threads at
4293 * the same time. TID is the Thread ID for the current
4294 * thread and is a value between 0 and 63, representing
4295 * the thread's position in the wavefront.
4297 * For the pixel shader threads are grouped into quads of four pixels.
4298 * The TIDs of the pixels of a quad are:
4306 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
4307 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
4308 * the current pixel's column, and masking with 0xfffffffe yields the TID
4309 * of the left pixel of the current pixel's row.
4311 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
4312 * adding 2 yields the TID of the pixel below the top pixel.
4314 /* masks for thread ID. */
4315 #define TID_MASK_TOP_LEFT 0xfffffffc
4316 #define TID_MASK_TOP 0xfffffffd
4317 #define TID_MASK_LEFT 0xfffffffe
4319 static void si_llvm_emit_ddxy(
4320 const struct lp_build_tgsi_action
*action
,
4321 struct lp_build_tgsi_context
*bld_base
,
4322 struct lp_build_emit_data
*emit_data
)
4324 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4325 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4326 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
4327 unsigned opcode
= inst
->Instruction
.Opcode
;
4328 LLVMValueRef indices
[2];
4329 LLVMValueRef store_ptr
, load_ptr0
, load_ptr1
;
4330 LLVMValueRef tl
, trbl
, result
[4];
4331 unsigned swizzle
[4];
4336 indices
[0] = bld_base
->uint_bld
.zero
;
4337 indices
[1] = lp_build_intrinsic(gallivm
->builder
, "llvm.SI.tid", ctx
->i32
,
4338 NULL
, 0, LLVMReadNoneAttribute
);
4339 store_ptr
= LLVMBuildGEP(gallivm
->builder
, ctx
->lds
,
4342 if (opcode
== TGSI_OPCODE_DDX_FINE
)
4343 mask
= TID_MASK_LEFT
;
4344 else if (opcode
== TGSI_OPCODE_DDY_FINE
)
4345 mask
= TID_MASK_TOP
;
4347 mask
= TID_MASK_TOP_LEFT
;
4349 indices
[1] = LLVMBuildAnd(gallivm
->builder
, indices
[1],
4350 lp_build_const_int32(gallivm
, mask
), "");
4351 load_ptr0
= LLVMBuildGEP(gallivm
->builder
, ctx
->lds
,
4354 /* for DDX we want to next X pixel, DDY next Y pixel. */
4355 idx
= (opcode
== TGSI_OPCODE_DDX
|| opcode
== TGSI_OPCODE_DDX_FINE
) ? 1 : 2;
4356 indices
[1] = LLVMBuildAdd(gallivm
->builder
, indices
[1],
4357 lp_build_const_int32(gallivm
, idx
), "");
4358 load_ptr1
= LLVMBuildGEP(gallivm
->builder
, ctx
->lds
,
4361 for (c
= 0; c
< 4; ++c
) {
4364 swizzle
[c
] = tgsi_util_get_full_src_register_swizzle(&inst
->Src
[0], c
);
4365 for (i
= 0; i
< c
; ++i
) {
4366 if (swizzle
[i
] == swizzle
[c
]) {
4367 result
[c
] = result
[i
];
4374 LLVMBuildStore(gallivm
->builder
,
4375 LLVMBuildBitCast(gallivm
->builder
,
4376 lp_build_emit_fetch(bld_base
, inst
, 0, c
),
4380 tl
= LLVMBuildLoad(gallivm
->builder
, load_ptr0
, "");
4381 tl
= LLVMBuildBitCast(gallivm
->builder
, tl
, ctx
->f32
, "");
4383 trbl
= LLVMBuildLoad(gallivm
->builder
, load_ptr1
, "");
4384 trbl
= LLVMBuildBitCast(gallivm
->builder
, trbl
, ctx
->f32
, "");
4386 result
[c
] = LLVMBuildFSub(gallivm
->builder
, trbl
, tl
, "");
4389 emit_data
->output
[0] = lp_build_gather_values(gallivm
, result
, 4);
4393 * this takes an I,J coordinate pair,
4394 * and works out the X and Y derivatives.
4395 * it returns DDX(I), DDX(J), DDY(I), DDY(J).
4397 static LLVMValueRef
si_llvm_emit_ddxy_interp(
4398 struct lp_build_tgsi_context
*bld_base
,
4399 LLVMValueRef interp_ij
)
4401 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4402 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4403 LLVMValueRef indices
[2];
4404 LLVMValueRef store_ptr
, load_ptr_x
, load_ptr_y
, load_ptr_ddx
, load_ptr_ddy
, temp
, temp2
;
4405 LLVMValueRef tl
, tr
, bl
, result
[4];
4408 indices
[0] = bld_base
->uint_bld
.zero
;
4409 indices
[1] = lp_build_intrinsic(gallivm
->builder
, "llvm.SI.tid", ctx
->i32
,
4410 NULL
, 0, LLVMReadNoneAttribute
);
4411 store_ptr
= LLVMBuildGEP(gallivm
->builder
, ctx
->lds
,
4414 temp
= LLVMBuildAnd(gallivm
->builder
, indices
[1],
4415 lp_build_const_int32(gallivm
, TID_MASK_LEFT
), "");
4417 temp2
= LLVMBuildAnd(gallivm
->builder
, indices
[1],
4418 lp_build_const_int32(gallivm
, TID_MASK_TOP
), "");
4421 load_ptr_x
= LLVMBuildGEP(gallivm
->builder
, ctx
->lds
,
4425 load_ptr_y
= LLVMBuildGEP(gallivm
->builder
, ctx
->lds
,
4428 indices
[1] = LLVMBuildAdd(gallivm
->builder
, temp
,
4429 lp_build_const_int32(gallivm
, 1), "");
4430 load_ptr_ddx
= LLVMBuildGEP(gallivm
->builder
, ctx
->lds
,
4433 indices
[1] = LLVMBuildAdd(gallivm
->builder
, temp2
,
4434 lp_build_const_int32(gallivm
, 2), "");
4435 load_ptr_ddy
= LLVMBuildGEP(gallivm
->builder
, ctx
->lds
,
4438 for (c
= 0; c
< 2; ++c
) {
4439 LLVMValueRef store_val
;
4440 LLVMValueRef c_ll
= lp_build_const_int32(gallivm
, c
);
4442 store_val
= LLVMBuildExtractElement(gallivm
->builder
,
4443 interp_ij
, c_ll
, "");
4444 LLVMBuildStore(gallivm
->builder
,
4448 tl
= LLVMBuildLoad(gallivm
->builder
, load_ptr_x
, "");
4449 tl
= LLVMBuildBitCast(gallivm
->builder
, tl
, ctx
->f32
, "");
4451 tr
= LLVMBuildLoad(gallivm
->builder
, load_ptr_ddx
, "");
4452 tr
= LLVMBuildBitCast(gallivm
->builder
, tr
, ctx
->f32
, "");
4454 result
[c
] = LLVMBuildFSub(gallivm
->builder
, tr
, tl
, "");
4456 tl
= LLVMBuildLoad(gallivm
->builder
, load_ptr_y
, "");
4457 tl
= LLVMBuildBitCast(gallivm
->builder
, tl
, ctx
->f32
, "");
4459 bl
= LLVMBuildLoad(gallivm
->builder
, load_ptr_ddy
, "");
4460 bl
= LLVMBuildBitCast(gallivm
->builder
, bl
, ctx
->f32
, "");
4462 result
[c
+ 2] = LLVMBuildFSub(gallivm
->builder
, bl
, tl
, "");
4465 return lp_build_gather_values(gallivm
, result
, 4);
4468 static void interp_fetch_args(
4469 struct lp_build_tgsi_context
*bld_base
,
4470 struct lp_build_emit_data
*emit_data
)
4472 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4473 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4474 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
4476 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_OFFSET
) {
4477 /* offset is in second src, first two channels */
4478 emit_data
->args
[0] = lp_build_emit_fetch(bld_base
,
4481 emit_data
->args
[1] = lp_build_emit_fetch(bld_base
,
4484 emit_data
->arg_count
= 2;
4485 } else if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_SAMPLE
) {
4486 LLVMValueRef sample_position
;
4487 LLVMValueRef sample_id
;
4488 LLVMValueRef halfval
= lp_build_const_float(gallivm
, 0.5f
);
4490 /* fetch sample ID, then fetch its sample position,
4491 * and place into first two channels.
4493 sample_id
= lp_build_emit_fetch(bld_base
,
4494 emit_data
->inst
, 1, TGSI_CHAN_X
);
4495 sample_id
= LLVMBuildBitCast(gallivm
->builder
, sample_id
,
4497 sample_position
= load_sample_position(&ctx
->radeon_bld
, sample_id
);
4499 emit_data
->args
[0] = LLVMBuildExtractElement(gallivm
->builder
,
4501 lp_build_const_int32(gallivm
, 0), "");
4503 emit_data
->args
[0] = LLVMBuildFSub(gallivm
->builder
, emit_data
->args
[0], halfval
, "");
4504 emit_data
->args
[1] = LLVMBuildExtractElement(gallivm
->builder
,
4506 lp_build_const_int32(gallivm
, 1), "");
4507 emit_data
->args
[1] = LLVMBuildFSub(gallivm
->builder
, emit_data
->args
[1], halfval
, "");
4508 emit_data
->arg_count
= 2;
4512 static void build_interp_intrinsic(const struct lp_build_tgsi_action
*action
,
4513 struct lp_build_tgsi_context
*bld_base
,
4514 struct lp_build_emit_data
*emit_data
)
4516 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4517 struct si_shader
*shader
= ctx
->shader
;
4518 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4519 LLVMValueRef interp_param
;
4520 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
4521 const char *intr_name
;
4522 int input_index
= inst
->Src
[0].Register
.Index
;
4525 LLVMValueRef attr_number
;
4526 LLVMValueRef params
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_PRIM_MASK
);
4527 int interp_param_idx
;
4528 unsigned interp
= shader
->selector
->info
.input_interpolate
[input_index
];
4531 assert(inst
->Src
[0].Register
.File
== TGSI_FILE_INPUT
);
4533 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_OFFSET
||
4534 inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_SAMPLE
)
4535 location
= TGSI_INTERPOLATE_LOC_CENTER
;
4537 location
= TGSI_INTERPOLATE_LOC_CENTROID
;
4539 interp_param_idx
= lookup_interp_param_index(interp
, location
);
4540 if (interp_param_idx
== -1)
4542 else if (interp_param_idx
)
4543 interp_param
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, interp_param_idx
);
4545 interp_param
= NULL
;
4547 attr_number
= lp_build_const_int32(gallivm
, input_index
);
4549 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_OFFSET
||
4550 inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_SAMPLE
) {
4551 LLVMValueRef ij_out
[2];
4552 LLVMValueRef ddxy_out
= si_llvm_emit_ddxy_interp(bld_base
, interp_param
);
4555 * take the I then J parameters, and the DDX/Y for it, and
4556 * calculate the IJ inputs for the interpolator.
4557 * temp1 = ddx * offset/sample.x + I;
4558 * interp_param.I = ddy * offset/sample.y + temp1;
4559 * temp1 = ddx * offset/sample.x + J;
4560 * interp_param.J = ddy * offset/sample.y + temp1;
4562 for (i
= 0; i
< 2; i
++) {
4563 LLVMValueRef ix_ll
= lp_build_const_int32(gallivm
, i
);
4564 LLVMValueRef iy_ll
= lp_build_const_int32(gallivm
, i
+ 2);
4565 LLVMValueRef ddx_el
= LLVMBuildExtractElement(gallivm
->builder
,
4566 ddxy_out
, ix_ll
, "");
4567 LLVMValueRef ddy_el
= LLVMBuildExtractElement(gallivm
->builder
,
4568 ddxy_out
, iy_ll
, "");
4569 LLVMValueRef interp_el
= LLVMBuildExtractElement(gallivm
->builder
,
4570 interp_param
, ix_ll
, "");
4571 LLVMValueRef temp1
, temp2
;
4573 interp_el
= LLVMBuildBitCast(gallivm
->builder
, interp_el
,
4576 temp1
= LLVMBuildFMul(gallivm
->builder
, ddx_el
, emit_data
->args
[0], "");
4578 temp1
= LLVMBuildFAdd(gallivm
->builder
, temp1
, interp_el
, "");
4580 temp2
= LLVMBuildFMul(gallivm
->builder
, ddy_el
, emit_data
->args
[1], "");
4582 temp2
= LLVMBuildFAdd(gallivm
->builder
, temp2
, temp1
, "");
4584 ij_out
[i
] = LLVMBuildBitCast(gallivm
->builder
,
4585 temp2
, ctx
->i32
, "");
4587 interp_param
= lp_build_gather_values(bld_base
->base
.gallivm
, ij_out
, 2);
4590 intr_name
= interp_param
? "llvm.SI.fs.interp" : "llvm.SI.fs.constant";
4591 for (chan
= 0; chan
< 2; chan
++) {
4592 LLVMValueRef args
[4];
4593 LLVMValueRef llvm_chan
;
4596 schan
= tgsi_util_get_full_src_register_swizzle(&inst
->Src
[0], chan
);
4597 llvm_chan
= lp_build_const_int32(gallivm
, schan
);
4599 args
[0] = llvm_chan
;
4600 args
[1] = attr_number
;
4602 args
[3] = interp_param
;
4604 emit_data
->output
[chan
] =
4605 lp_build_intrinsic(gallivm
->builder
, intr_name
,
4606 ctx
->f32
, args
, args
[3] ? 4 : 3,
4607 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
4611 static unsigned si_llvm_get_stream(struct lp_build_tgsi_context
*bld_base
,
4612 struct lp_build_emit_data
*emit_data
)
4614 LLVMValueRef (*imms
)[4] = lp_soa_context(bld_base
)->immediates
;
4615 struct tgsi_src_register src0
= emit_data
->inst
->Src
[0].Register
;
4618 assert(src0
.File
== TGSI_FILE_IMMEDIATE
);
4620 stream
= LLVMConstIntGetZExtValue(imms
[src0
.Index
][src0
.SwizzleX
]) & 0x3;
4624 /* Emit one vertex from the geometry shader */
4625 static void si_llvm_emit_vertex(
4626 const struct lp_build_tgsi_action
*action
,
4627 struct lp_build_tgsi_context
*bld_base
,
4628 struct lp_build_emit_data
*emit_data
)
4630 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4631 struct lp_build_context
*uint
= &bld_base
->uint_bld
;
4632 struct si_shader
*shader
= ctx
->shader
;
4633 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
4634 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4635 LLVMValueRef soffset
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
4636 SI_PARAM_GS2VS_OFFSET
);
4637 LLVMValueRef gs_next_vertex
;
4638 LLVMValueRef can_emit
, kill
;
4639 LLVMValueRef args
[2];
4644 stream
= si_llvm_get_stream(bld_base
, emit_data
);
4646 /* Write vertex attribute values to GSVS ring */
4647 gs_next_vertex
= LLVMBuildLoad(gallivm
->builder
,
4648 ctx
->gs_next_vertex
[stream
],
4651 /* If this thread has already emitted the declared maximum number of
4652 * vertices, kill it: excessive vertex emissions are not supposed to
4653 * have any effect, and GS threads have no externally observable
4654 * effects other than emitting vertices.
4656 can_emit
= LLVMBuildICmp(gallivm
->builder
, LLVMIntULE
, gs_next_vertex
,
4657 lp_build_const_int32(gallivm
,
4658 shader
->selector
->gs_max_out_vertices
), "");
4659 kill
= lp_build_select(&bld_base
->base
, can_emit
,
4660 lp_build_const_float(gallivm
, 1.0f
),
4661 lp_build_const_float(gallivm
, -1.0f
));
4663 lp_build_intrinsic(gallivm
->builder
, "llvm.AMDGPU.kill",
4664 ctx
->voidt
, &kill
, 1, 0);
4666 for (i
= 0; i
< info
->num_outputs
; i
++) {
4667 LLVMValueRef
*out_ptr
=
4668 ctx
->radeon_bld
.soa
.outputs
[i
];
4670 for (chan
= 0; chan
< 4; chan
++) {
4671 LLVMValueRef out_val
= LLVMBuildLoad(gallivm
->builder
, out_ptr
[chan
], "");
4672 LLVMValueRef voffset
=
4673 lp_build_const_int32(gallivm
, (i
* 4 + chan
) *
4674 shader
->selector
->gs_max_out_vertices
);
4676 voffset
= lp_build_add(uint
, voffset
, gs_next_vertex
);
4677 voffset
= lp_build_mul_imm(uint
, voffset
, 4);
4679 out_val
= LLVMBuildBitCast(gallivm
->builder
, out_val
, ctx
->i32
, "");
4681 build_tbuffer_store(ctx
,
4682 ctx
->gsvs_ring
[stream
],
4684 voffset
, soffset
, 0,
4685 V_008F0C_BUF_DATA_FORMAT_32
,
4686 V_008F0C_BUF_NUM_FORMAT_UINT
,
4690 gs_next_vertex
= lp_build_add(uint
, gs_next_vertex
,
4691 lp_build_const_int32(gallivm
, 1));
4693 LLVMBuildStore(gallivm
->builder
, gs_next_vertex
, ctx
->gs_next_vertex
[stream
]);
4695 /* Signal vertex emission */
4696 args
[0] = lp_build_const_int32(gallivm
, SENDMSG_GS_OP_EMIT
| SENDMSG_GS
| (stream
<< 8));
4697 args
[1] = LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_GS_WAVE_ID
);
4698 lp_build_intrinsic(gallivm
->builder
, "llvm.SI.sendmsg",
4699 ctx
->voidt
, args
, 2, LLVMNoUnwindAttribute
);
4702 /* Cut one primitive from the geometry shader */
4703 static void si_llvm_emit_primitive(
4704 const struct lp_build_tgsi_action
*action
,
4705 struct lp_build_tgsi_context
*bld_base
,
4706 struct lp_build_emit_data
*emit_data
)
4708 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4709 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4710 LLVMValueRef args
[2];
4713 /* Signal primitive cut */
4714 stream
= si_llvm_get_stream(bld_base
, emit_data
);
4715 args
[0] = lp_build_const_int32(gallivm
, SENDMSG_GS_OP_CUT
| SENDMSG_GS
| (stream
<< 8));
4716 args
[1] = LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_GS_WAVE_ID
);
4717 lp_build_intrinsic(gallivm
->builder
, "llvm.SI.sendmsg",
4718 ctx
->voidt
, args
, 2, LLVMNoUnwindAttribute
);
4721 static void si_llvm_emit_barrier(const struct lp_build_tgsi_action
*action
,
4722 struct lp_build_tgsi_context
*bld_base
,
4723 struct lp_build_emit_data
*emit_data
)
4725 struct si_shader_context
*ctx
= si_shader_context(bld_base
);
4726 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4728 /* The real barrier instruction isn’t needed, because an entire patch
4729 * always fits into a single wave.
4731 if (ctx
->type
== TGSI_PROCESSOR_TESS_CTRL
) {
4732 emit_optimization_barrier(ctx
);
4736 lp_build_intrinsic(gallivm
->builder
,
4737 HAVE_LLVM
>= 0x0309 ? "llvm.amdgcn.s.barrier"
4738 : "llvm.AMDGPU.barrier.local",
4739 ctx
->voidt
, NULL
, 0, LLVMNoUnwindAttribute
);
4742 static const struct lp_build_tgsi_action tex_action
= {
4743 .fetch_args
= tex_fetch_args
,
4744 .emit
= build_tex_intrinsic
,
4747 static const struct lp_build_tgsi_action interp_action
= {
4748 .fetch_args
= interp_fetch_args
,
4749 .emit
= build_interp_intrinsic
,
4752 static void si_create_function(struct si_shader_context
*ctx
,
4753 LLVMTypeRef
*returns
, unsigned num_returns
,
4754 LLVMTypeRef
*params
, unsigned num_params
,
4755 int last_array_pointer
, int last_sgpr
)
4759 radeon_llvm_create_func(&ctx
->radeon_bld
, returns
, num_returns
,
4760 params
, num_params
);
4761 radeon_llvm_shader_type(ctx
->radeon_bld
.main_fn
, ctx
->type
);
4762 ctx
->return_value
= LLVMGetUndef(ctx
->radeon_bld
.return_type
);
4764 for (i
= 0; i
<= last_sgpr
; ++i
) {
4765 LLVMValueRef P
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, i
);
4767 /* We tell llvm that array inputs are passed by value to allow Sinking pass
4768 * to move load. Inputs are constant so this is fine. */
4769 if (i
<= last_array_pointer
)
4770 LLVMAddAttribute(P
, LLVMByValAttribute
);
4772 LLVMAddAttribute(P
, LLVMInRegAttribute
);
4776 static void create_meta_data(struct si_shader_context
*ctx
)
4778 struct gallivm_state
*gallivm
= ctx
->radeon_bld
.soa
.bld_base
.base
.gallivm
;
4779 LLVMValueRef args
[3];
4781 args
[0] = LLVMMDStringInContext(gallivm
->context
, "const", 5);
4783 args
[2] = lp_build_const_int32(gallivm
, 1);
4785 ctx
->const_md
= LLVMMDNodeInContext(gallivm
->context
, args
, 3);
4788 static void declare_streamout_params(struct si_shader_context
*ctx
,
4789 struct pipe_stream_output_info
*so
,
4790 LLVMTypeRef
*params
, LLVMTypeRef i32
,
4791 unsigned *num_params
)
4795 /* Streamout SGPRs. */
4796 if (so
->num_outputs
) {
4797 params
[ctx
->param_streamout_config
= (*num_params
)++] = i32
;
4798 params
[ctx
->param_streamout_write_index
= (*num_params
)++] = i32
;
4800 /* A streamout buffer offset is loaded if the stride is non-zero. */
4801 for (i
= 0; i
< 4; i
++) {
4805 params
[ctx
->param_streamout_offset
[i
] = (*num_params
)++] = i32
;
4809 static unsigned llvm_get_type_size(LLVMTypeRef type
)
4811 LLVMTypeKind kind
= LLVMGetTypeKind(type
);
4814 case LLVMIntegerTypeKind
:
4815 return LLVMGetIntTypeWidth(type
) / 8;
4816 case LLVMFloatTypeKind
:
4818 case LLVMPointerTypeKind
:
4820 case LLVMVectorTypeKind
:
4821 return LLVMGetVectorSize(type
) *
4822 llvm_get_type_size(LLVMGetElementType(type
));
4829 static void declare_tess_lds(struct si_shader_context
*ctx
)
4831 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
4832 LLVMTypeRef i32
= ctx
->radeon_bld
.soa
.bld_base
.uint_bld
.elem_type
;
4834 /* This is the upper bound, maximum is 32 inputs times 32 vertices */
4835 unsigned vertex_data_dw_size
= 32*32*4;
4836 unsigned patch_data_dw_size
= 32*4;
4837 /* The formula is: TCS inputs + TCS outputs + TCS patch outputs. */
4838 unsigned patch_dw_size
= vertex_data_dw_size
*2 + patch_data_dw_size
;
4839 unsigned lds_dwords
= patch_dw_size
;
4841 /* The actual size is computed outside of the shader to reduce
4842 * the number of shader variants. */
4844 LLVMAddGlobalInAddressSpace(gallivm
->module
,
4845 LLVMArrayType(i32
, lds_dwords
),
4850 static void create_function(struct si_shader_context
*ctx
)
4852 struct lp_build_tgsi_context
*bld_base
= &ctx
->radeon_bld
.soa
.bld_base
;
4853 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
4854 struct si_shader
*shader
= ctx
->shader
;
4855 LLVMTypeRef params
[SI_NUM_PARAMS
+ SI_NUM_VERTEX_BUFFERS
], v3i32
;
4856 LLVMTypeRef returns
[16+32*4];
4857 unsigned i
, last_array_pointer
, last_sgpr
, num_params
, num_return_sgprs
;
4858 unsigned num_returns
= 0;
4860 v3i32
= LLVMVectorType(ctx
->i32
, 3);
4862 params
[SI_PARAM_RW_BUFFERS
] = const_array(ctx
->v16i8
, SI_NUM_RW_BUFFERS
);
4863 params
[SI_PARAM_CONST_BUFFERS
] = const_array(ctx
->v16i8
, SI_NUM_CONST_BUFFERS
);
4864 params
[SI_PARAM_SAMPLERS
] = const_array(ctx
->v8i32
, SI_NUM_SAMPLERS
);
4865 params
[SI_PARAM_IMAGES
] = const_array(ctx
->v8i32
, SI_NUM_IMAGES
);
4866 params
[SI_PARAM_SHADER_BUFFERS
] = const_array(ctx
->v4i32
, SI_NUM_SHADER_BUFFERS
);
4867 last_array_pointer
= SI_PARAM_SHADER_BUFFERS
;
4869 switch (ctx
->type
) {
4870 case TGSI_PROCESSOR_VERTEX
:
4871 params
[SI_PARAM_VERTEX_BUFFERS
] = const_array(ctx
->v16i8
, SI_NUM_VERTEX_BUFFERS
);
4872 last_array_pointer
= SI_PARAM_VERTEX_BUFFERS
;
4873 params
[SI_PARAM_BASE_VERTEX
] = ctx
->i32
;
4874 params
[SI_PARAM_START_INSTANCE
] = ctx
->i32
;
4875 num_params
= SI_PARAM_START_INSTANCE
+1;
4877 if (shader
->key
.vs
.as_es
) {
4878 params
[ctx
->param_es2gs_offset
= num_params
++] = ctx
->i32
;
4879 } else if (shader
->key
.vs
.as_ls
) {
4880 params
[SI_PARAM_LS_OUT_LAYOUT
] = ctx
->i32
;
4881 num_params
= SI_PARAM_LS_OUT_LAYOUT
+1;
4883 if (ctx
->is_gs_copy_shader
) {
4884 last_array_pointer
= SI_PARAM_CONST_BUFFERS
;
4885 num_params
= SI_PARAM_CONST_BUFFERS
+1;
4887 params
[SI_PARAM_VS_STATE_BITS
] = ctx
->i32
;
4888 num_params
= SI_PARAM_VS_STATE_BITS
+1;
4891 /* The locations of the other parameters are assigned dynamically. */
4892 declare_streamout_params(ctx
, &shader
->selector
->so
,
4893 params
, ctx
->i32
, &num_params
);
4896 last_sgpr
= num_params
-1;
4899 params
[ctx
->param_vertex_id
= num_params
++] = ctx
->i32
;
4900 params
[ctx
->param_rel_auto_id
= num_params
++] = ctx
->i32
;
4901 params
[ctx
->param_vs_prim_id
= num_params
++] = ctx
->i32
;
4902 params
[ctx
->param_instance_id
= num_params
++] = ctx
->i32
;
4904 if (!ctx
->is_monolithic
&&
4905 !ctx
->is_gs_copy_shader
) {
4906 /* Vertex load indices. */
4907 ctx
->param_vertex_index0
= num_params
;
4909 for (i
= 0; i
< shader
->selector
->info
.num_inputs
; i
++)
4910 params
[num_params
++] = ctx
->i32
;
4912 /* PrimitiveID output. */
4913 if (!shader
->key
.vs
.as_es
&& !shader
->key
.vs
.as_ls
)
4914 for (i
= 0; i
<= VS_EPILOG_PRIMID_LOC
; i
++)
4915 returns
[num_returns
++] = ctx
->f32
;
4919 case TGSI_PROCESSOR_TESS_CTRL
:
4920 params
[SI_PARAM_TCS_OUT_OFFSETS
] = ctx
->i32
;
4921 params
[SI_PARAM_TCS_OUT_LAYOUT
] = ctx
->i32
;
4922 params
[SI_PARAM_TCS_IN_LAYOUT
] = ctx
->i32
;
4923 params
[SI_PARAM_TESS_FACTOR_OFFSET
] = ctx
->i32
;
4924 last_sgpr
= SI_PARAM_TESS_FACTOR_OFFSET
;
4927 params
[SI_PARAM_PATCH_ID
] = ctx
->i32
;
4928 params
[SI_PARAM_REL_IDS
] = ctx
->i32
;
4929 num_params
= SI_PARAM_REL_IDS
+1;
4931 if (!ctx
->is_monolithic
) {
4932 /* PARAM_TESS_FACTOR_OFFSET is after user SGPRs. */
4933 for (i
= 0; i
<= SI_TCS_NUM_USER_SGPR
; i
++)
4934 returns
[num_returns
++] = ctx
->i32
; /* SGPRs */
4936 for (i
= 0; i
< 3; i
++)
4937 returns
[num_returns
++] = ctx
->f32
; /* VGPRs */
4941 case TGSI_PROCESSOR_TESS_EVAL
:
4942 params
[SI_PARAM_TCS_OUT_OFFSETS
] = ctx
->i32
;
4943 params
[SI_PARAM_TCS_OUT_LAYOUT
] = ctx
->i32
;
4944 num_params
= SI_PARAM_TCS_OUT_LAYOUT
+1;
4946 if (shader
->key
.tes
.as_es
) {
4947 params
[ctx
->param_es2gs_offset
= num_params
++] = ctx
->i32
;
4949 declare_streamout_params(ctx
, &shader
->selector
->so
,
4950 params
, ctx
->i32
, &num_params
);
4952 last_sgpr
= num_params
- 1;
4955 params
[ctx
->param_tes_u
= num_params
++] = ctx
->f32
;
4956 params
[ctx
->param_tes_v
= num_params
++] = ctx
->f32
;
4957 params
[ctx
->param_tes_rel_patch_id
= num_params
++] = ctx
->i32
;
4958 params
[ctx
->param_tes_patch_id
= num_params
++] = ctx
->i32
;
4960 /* PrimitiveID output. */
4961 if (!ctx
->is_monolithic
&& !shader
->key
.tes
.as_es
)
4962 for (i
= 0; i
<= VS_EPILOG_PRIMID_LOC
; i
++)
4963 returns
[num_returns
++] = ctx
->f32
;
4966 case TGSI_PROCESSOR_GEOMETRY
:
4967 params
[SI_PARAM_GS2VS_OFFSET
] = ctx
->i32
;
4968 params
[SI_PARAM_GS_WAVE_ID
] = ctx
->i32
;
4969 last_sgpr
= SI_PARAM_GS_WAVE_ID
;
4972 params
[SI_PARAM_VTX0_OFFSET
] = ctx
->i32
;
4973 params
[SI_PARAM_VTX1_OFFSET
] = ctx
->i32
;
4974 params
[SI_PARAM_PRIMITIVE_ID
] = ctx
->i32
;
4975 params
[SI_PARAM_VTX2_OFFSET
] = ctx
->i32
;
4976 params
[SI_PARAM_VTX3_OFFSET
] = ctx
->i32
;
4977 params
[SI_PARAM_VTX4_OFFSET
] = ctx
->i32
;
4978 params
[SI_PARAM_VTX5_OFFSET
] = ctx
->i32
;
4979 params
[SI_PARAM_GS_INSTANCE_ID
] = ctx
->i32
;
4980 num_params
= SI_PARAM_GS_INSTANCE_ID
+1;
4983 case TGSI_PROCESSOR_FRAGMENT
:
4984 params
[SI_PARAM_ALPHA_REF
] = ctx
->f32
;
4985 params
[SI_PARAM_PRIM_MASK
] = ctx
->i32
;
4986 last_sgpr
= SI_PARAM_PRIM_MASK
;
4987 params
[SI_PARAM_PERSP_SAMPLE
] = ctx
->v2i32
;
4988 params
[SI_PARAM_PERSP_CENTER
] = ctx
->v2i32
;
4989 params
[SI_PARAM_PERSP_CENTROID
] = ctx
->v2i32
;
4990 params
[SI_PARAM_PERSP_PULL_MODEL
] = v3i32
;
4991 params
[SI_PARAM_LINEAR_SAMPLE
] = ctx
->v2i32
;
4992 params
[SI_PARAM_LINEAR_CENTER
] = ctx
->v2i32
;
4993 params
[SI_PARAM_LINEAR_CENTROID
] = ctx
->v2i32
;
4994 params
[SI_PARAM_LINE_STIPPLE_TEX
] = ctx
->f32
;
4995 params
[SI_PARAM_POS_X_FLOAT
] = ctx
->f32
;
4996 params
[SI_PARAM_POS_Y_FLOAT
] = ctx
->f32
;
4997 params
[SI_PARAM_POS_Z_FLOAT
] = ctx
->f32
;
4998 params
[SI_PARAM_POS_W_FLOAT
] = ctx
->f32
;
4999 params
[SI_PARAM_FRONT_FACE
] = ctx
->i32
;
5000 params
[SI_PARAM_ANCILLARY
] = ctx
->i32
;
5001 params
[SI_PARAM_SAMPLE_COVERAGE
] = ctx
->f32
;
5002 params
[SI_PARAM_POS_FIXED_PT
] = ctx
->i32
;
5003 num_params
= SI_PARAM_POS_FIXED_PT
+1;
5005 if (!ctx
->is_monolithic
) {
5006 /* Color inputs from the prolog. */
5007 if (shader
->selector
->info
.colors_read
) {
5008 unsigned num_color_elements
=
5009 util_bitcount(shader
->selector
->info
.colors_read
);
5011 assert(num_params
+ num_color_elements
<= ARRAY_SIZE(params
));
5012 for (i
= 0; i
< num_color_elements
; i
++)
5013 params
[num_params
++] = ctx
->f32
;
5016 /* Outputs for the epilog. */
5017 num_return_sgprs
= SI_SGPR_ALPHA_REF
+ 1;
5020 util_bitcount(shader
->selector
->info
.colors_written
) * 4 +
5021 shader
->selector
->info
.writes_z
+
5022 shader
->selector
->info
.writes_stencil
+
5023 shader
->selector
->info
.writes_samplemask
+
5024 1 /* SampleMaskIn */;
5026 num_returns
= MAX2(num_returns
,
5028 PS_EPILOG_SAMPLEMASK_MIN_LOC
+ 1);
5030 for (i
= 0; i
< num_return_sgprs
; i
++)
5031 returns
[i
] = ctx
->i32
;
5032 for (; i
< num_returns
; i
++)
5033 returns
[i
] = ctx
->f32
;
5037 case TGSI_PROCESSOR_COMPUTE
:
5038 params
[SI_PARAM_GRID_SIZE
] = v3i32
;
5039 params
[SI_PARAM_BLOCK_ID
] = v3i32
;
5040 last_sgpr
= SI_PARAM_BLOCK_ID
;
5042 params
[SI_PARAM_THREAD_ID
] = v3i32
;
5043 num_params
= SI_PARAM_THREAD_ID
+ 1;
5046 assert(0 && "unimplemented shader");
5050 assert(num_params
<= Elements(params
));
5052 si_create_function(ctx
, returns
, num_returns
, params
,
5053 num_params
, last_array_pointer
, last_sgpr
);
5055 /* Reserve register locations for VGPR inputs the PS prolog may need. */
5056 if (ctx
->type
== TGSI_PROCESSOR_FRAGMENT
&&
5057 !ctx
->is_monolithic
) {
5058 radeon_llvm_add_attribute(ctx
->radeon_bld
.main_fn
,
5059 "InitialPSInputAddr",
5060 S_0286D0_PERSP_SAMPLE_ENA(1) |
5061 S_0286D0_PERSP_CENTER_ENA(1) |
5062 S_0286D0_PERSP_CENTROID_ENA(1) |
5063 S_0286D0_LINEAR_SAMPLE_ENA(1) |
5064 S_0286D0_LINEAR_CENTER_ENA(1) |
5065 S_0286D0_LINEAR_CENTROID_ENA(1) |
5066 S_0286D0_FRONT_FACE_ENA(1) |
5067 S_0286D0_POS_FIXED_PT_ENA(1));
5068 } else if (ctx
->type
== TGSI_PROCESSOR_COMPUTE
) {
5069 const unsigned *properties
= shader
->selector
->info
.properties
;
5070 unsigned max_work_group_size
=
5071 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
] *
5072 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT
] *
5073 properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH
];
5075 assert(max_work_group_size
);
5077 radeon_llvm_add_attribute(ctx
->radeon_bld
.main_fn
,
5078 "amdgpu-max-work-group-size",
5079 max_work_group_size
);
5082 shader
->info
.num_input_sgprs
= 0;
5083 shader
->info
.num_input_vgprs
= 0;
5085 for (i
= 0; i
<= last_sgpr
; ++i
)
5086 shader
->info
.num_input_sgprs
+= llvm_get_type_size(params
[i
]) / 4;
5088 /* Unused fragment shader inputs are eliminated by the compiler,
5089 * so we don't know yet how many there will be.
5091 if (ctx
->type
!= TGSI_PROCESSOR_FRAGMENT
)
5092 for (; i
< num_params
; ++i
)
5093 shader
->info
.num_input_vgprs
+= llvm_get_type_size(params
[i
]) / 4;
5095 if (bld_base
->info
&&
5096 (bld_base
->info
->opcode_count
[TGSI_OPCODE_DDX
] > 0 ||
5097 bld_base
->info
->opcode_count
[TGSI_OPCODE_DDY
] > 0 ||
5098 bld_base
->info
->opcode_count
[TGSI_OPCODE_DDX_FINE
] > 0 ||
5099 bld_base
->info
->opcode_count
[TGSI_OPCODE_DDY_FINE
] > 0 ||
5100 bld_base
->info
->opcode_count
[TGSI_OPCODE_INTERP_OFFSET
] > 0 ||
5101 bld_base
->info
->opcode_count
[TGSI_OPCODE_INTERP_SAMPLE
] > 0))
5103 LLVMAddGlobalInAddressSpace(gallivm
->module
,
5104 LLVMArrayType(ctx
->i32
, 64),
5108 if ((ctx
->type
== TGSI_PROCESSOR_VERTEX
&& shader
->key
.vs
.as_ls
) ||
5109 ctx
->type
== TGSI_PROCESSOR_TESS_CTRL
||
5110 ctx
->type
== TGSI_PROCESSOR_TESS_EVAL
)
5111 declare_tess_lds(ctx
);
5114 static void preload_constants(struct si_shader_context
*ctx
)
5116 struct lp_build_tgsi_context
*bld_base
= &ctx
->radeon_bld
.soa
.bld_base
;
5117 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
5118 const struct tgsi_shader_info
*info
= bld_base
->info
;
5120 LLVMValueRef ptr
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_CONST_BUFFERS
);
5122 for (buf
= 0; buf
< SI_NUM_CONST_BUFFERS
; buf
++) {
5123 unsigned i
, num_const
= info
->const_file_max
[buf
] + 1;
5128 /* Allocate space for the constant values */
5129 ctx
->constants
[buf
] = CALLOC(num_const
* 4, sizeof(LLVMValueRef
));
5131 /* Load the resource descriptor */
5132 ctx
->const_buffers
[buf
] =
5133 build_indexed_load_const(ctx
, ptr
, lp_build_const_int32(gallivm
, buf
));
5135 /* Load the constants, we rely on the code sinking to do the rest */
5136 for (i
= 0; i
< num_const
* 4; ++i
) {
5137 ctx
->constants
[buf
][i
] =
5138 buffer_load_const(gallivm
->builder
,
5139 ctx
->const_buffers
[buf
],
5140 lp_build_const_int32(gallivm
, i
* 4),
5146 static void preload_shader_buffers(struct si_shader_context
*ctx
)
5148 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
5149 LLVMValueRef ptr
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_SHADER_BUFFERS
);
5152 maxbuf
= MIN2(ctx
->shader
->selector
->info
.file_max
[TGSI_FILE_BUFFER
],
5153 SI_NUM_SHADER_BUFFERS
- 1);
5154 for (buf
= 0; buf
<= maxbuf
; ++buf
) {
5155 ctx
->shader_buffers
[buf
] =
5156 build_indexed_load_const(
5157 ctx
, ptr
, lp_build_const_int32(gallivm
, buf
));
5161 static void preload_samplers(struct si_shader_context
*ctx
)
5163 struct lp_build_tgsi_context
*bld_base
= &ctx
->radeon_bld
.soa
.bld_base
;
5164 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
5165 const struct tgsi_shader_info
*info
= bld_base
->info
;
5166 unsigned i
, num_samplers
= info
->file_max
[TGSI_FILE_SAMPLER
] + 1;
5167 LLVMValueRef offset
;
5169 if (num_samplers
== 0)
5172 /* Load the resources and samplers, we rely on the code sinking to do the rest */
5173 for (i
= 0; i
< num_samplers
; ++i
) {
5175 offset
= lp_build_const_int32(gallivm
, i
);
5176 ctx
->sampler_views
[i
] =
5177 get_sampler_desc(ctx
, offset
, DESC_IMAGE
);
5179 /* FMASK resource */
5180 if (info
->is_msaa_sampler
[i
])
5182 get_sampler_desc(ctx
, offset
, DESC_FMASK
);
5184 ctx
->sampler_states
[i
] =
5185 get_sampler_desc(ctx
, offset
, DESC_SAMPLER
);
5186 ctx
->sampler_states
[i
] =
5187 sici_fix_sampler_aniso(ctx
, ctx
->sampler_views
[i
],
5188 ctx
->sampler_states
[i
]);
5193 static void preload_images(struct si_shader_context
*ctx
)
5195 struct lp_build_tgsi_context
*bld_base
= &ctx
->radeon_bld
.soa
.bld_base
;
5196 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
5197 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
5198 unsigned num_images
= bld_base
->info
->file_max
[TGSI_FILE_IMAGE
] + 1;
5199 LLVMValueRef res_ptr
;
5202 if (num_images
== 0)
5205 res_ptr
= LLVMGetParam(ctx
->radeon_bld
.main_fn
, SI_PARAM_IMAGES
);
5207 for (i
= 0; i
< num_images
; ++i
) {
5208 /* Rely on LLVM to shrink the load for buffer resources. */
5210 build_indexed_load_const(ctx
, res_ptr
,
5211 lp_build_const_int32(gallivm
, i
));
5213 if (info
->images_writemask
& (1 << i
) &&
5214 !(info
->images_buffers
& (1 << i
)))
5215 rsrc
= force_dcc_off(ctx
, rsrc
);
5217 ctx
->images
[i
] = rsrc
;
5221 static void preload_streamout_buffers(struct si_shader_context
*ctx
)
5223 struct lp_build_tgsi_context
*bld_base
= &ctx
->radeon_bld
.soa
.bld_base
;
5224 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
5227 /* Streamout can only be used if the shader is compiled as VS. */
5228 if (!ctx
->shader
->selector
->so
.num_outputs
||
5229 (ctx
->type
== TGSI_PROCESSOR_VERTEX
&&
5230 (ctx
->shader
->key
.vs
.as_es
||
5231 ctx
->shader
->key
.vs
.as_ls
)) ||
5232 (ctx
->type
== TGSI_PROCESSOR_TESS_EVAL
&&
5233 ctx
->shader
->key
.tes
.as_es
))
5236 LLVMValueRef buf_ptr
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
5237 SI_PARAM_RW_BUFFERS
);
5239 /* Load the resources, we rely on the code sinking to do the rest */
5240 for (i
= 0; i
< 4; ++i
) {
5241 if (ctx
->shader
->selector
->so
.stride
[i
]) {
5242 LLVMValueRef offset
= lp_build_const_int32(gallivm
,
5243 SI_VS_STREAMOUT_BUF0
+ i
);
5245 ctx
->so_buffers
[i
] = build_indexed_load_const(ctx
, buf_ptr
, offset
);
5251 * Load ESGS and GSVS ring buffer resource descriptors and save the variables
5254 static void preload_ring_buffers(struct si_shader_context
*ctx
)
5256 struct gallivm_state
*gallivm
=
5257 ctx
->radeon_bld
.soa
.bld_base
.base
.gallivm
;
5259 LLVMValueRef buf_ptr
= LLVMGetParam(ctx
->radeon_bld
.main_fn
,
5260 SI_PARAM_RW_BUFFERS
);
5262 if ((ctx
->type
== TGSI_PROCESSOR_VERTEX
&&
5263 ctx
->shader
->key
.vs
.as_es
) ||
5264 (ctx
->type
== TGSI_PROCESSOR_TESS_EVAL
&&
5265 ctx
->shader
->key
.tes
.as_es
) ||
5266 ctx
->type
== TGSI_PROCESSOR_GEOMETRY
) {
5268 ctx
->type
== TGSI_PROCESSOR_GEOMETRY
? SI_GS_RING_ESGS
5270 LLVMValueRef offset
= lp_build_const_int32(gallivm
, ring
);
5273 build_indexed_load_const(ctx
, buf_ptr
, offset
);
5276 if (ctx
->is_gs_copy_shader
) {
5277 LLVMValueRef offset
= lp_build_const_int32(gallivm
, SI_VS_RING_GSVS
);
5280 build_indexed_load_const(ctx
, buf_ptr
, offset
);
5282 if (ctx
->type
== TGSI_PROCESSOR_GEOMETRY
) {
5284 for (i
= 0; i
< 4; i
++) {
5285 LLVMValueRef offset
= lp_build_const_int32(gallivm
, SI_GS_RING_GSVS0
+ i
);
5288 build_indexed_load_const(ctx
, buf_ptr
, offset
);
5293 static void si_llvm_emit_polygon_stipple(struct si_shader_context
*ctx
,
5294 LLVMValueRef param_sampler_views
,
5295 unsigned param_pos_fixed_pt
)
5297 struct lp_build_tgsi_context
*bld_base
=
5298 &ctx
->radeon_bld
.soa
.bld_base
;
5299 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
5300 struct lp_build_emit_data result
= {};
5301 struct tgsi_full_instruction inst
= {};
5302 LLVMValueRef desc
, sampler_index
, address
[2], pix
;
5304 /* Use the fixed-point gl_FragCoord input.
5305 * Since the stipple pattern is 32x32 and it repeats, just get 5 bits
5306 * per coordinate to get the repeating effect.
5308 address
[0] = unpack_param(ctx
, param_pos_fixed_pt
, 0, 5);
5309 address
[1] = unpack_param(ctx
, param_pos_fixed_pt
, 16, 5);
5311 /* Load the sampler view descriptor. */
5312 sampler_index
= lp_build_const_int32(gallivm
, SI_POLY_STIPPLE_SAMPLER
);
5313 desc
= get_sampler_desc_custom(ctx
, param_sampler_views
,
5314 sampler_index
, DESC_IMAGE
);
5316 /* Load the texel. */
5317 inst
.Instruction
.Opcode
= TGSI_OPCODE_TXF
;
5318 inst
.Texture
.Texture
= TGSI_TEXTURE_2D_MSAA
; /* = use load, not load_mip */
5319 result
.inst
= &inst
;
5320 set_tex_fetch_args(ctx
, &result
, TGSI_OPCODE_TXF
,
5321 inst
.Texture
.Texture
,
5322 desc
, NULL
, address
, ARRAY_SIZE(address
), 0xf);
5323 build_tex_intrinsic(&tex_action
, bld_base
, &result
);
5325 /* Kill the thread accordingly. */
5326 pix
= LLVMBuildExtractElement(gallivm
->builder
, result
.output
[0],
5327 lp_build_const_int32(gallivm
, 3), "");
5328 pix
= bitcast(bld_base
, TGSI_TYPE_FLOAT
, pix
);
5329 pix
= LLVMBuildFNeg(gallivm
->builder
, pix
, "");
5331 lp_build_intrinsic(gallivm
->builder
, "llvm.AMDGPU.kill",
5332 LLVMVoidTypeInContext(gallivm
->context
),
5336 void si_shader_binary_read_config(struct radeon_shader_binary
*binary
,
5337 struct si_shader_config
*conf
,
5338 unsigned symbol_offset
)
5341 const unsigned char *config
=
5342 radeon_shader_binary_config_start(binary
, symbol_offset
);
5344 /* XXX: We may be able to emit some of these values directly rather than
5345 * extracting fields to be emitted later.
5348 for (i
= 0; i
< binary
->config_size_per_symbol
; i
+= 8) {
5349 unsigned reg
= util_le32_to_cpu(*(uint32_t*)(config
+ i
));
5350 unsigned value
= util_le32_to_cpu(*(uint32_t*)(config
+ i
+ 4));
5352 case R_00B028_SPI_SHADER_PGM_RSRC1_PS
:
5353 case R_00B128_SPI_SHADER_PGM_RSRC1_VS
:
5354 case R_00B228_SPI_SHADER_PGM_RSRC1_GS
:
5355 case R_00B848_COMPUTE_PGM_RSRC1
:
5356 conf
->num_sgprs
= MAX2(conf
->num_sgprs
, (G_00B028_SGPRS(value
) + 1) * 8);
5357 conf
->num_vgprs
= MAX2(conf
->num_vgprs
, (G_00B028_VGPRS(value
) + 1) * 4);
5358 conf
->float_mode
= G_00B028_FLOAT_MODE(value
);
5359 conf
->rsrc1
= value
;
5361 case R_00B02C_SPI_SHADER_PGM_RSRC2_PS
:
5362 conf
->lds_size
= MAX2(conf
->lds_size
, G_00B02C_EXTRA_LDS_SIZE(value
));
5364 case R_00B84C_COMPUTE_PGM_RSRC2
:
5365 conf
->lds_size
= MAX2(conf
->lds_size
, G_00B84C_LDS_SIZE(value
));
5366 conf
->rsrc2
= value
;
5368 case R_0286CC_SPI_PS_INPUT_ENA
:
5369 conf
->spi_ps_input_ena
= value
;
5371 case R_0286D0_SPI_PS_INPUT_ADDR
:
5372 conf
->spi_ps_input_addr
= value
;
5374 case R_0286E8_SPI_TMPRING_SIZE
:
5375 case R_00B860_COMPUTE_TMPRING_SIZE
:
5376 /* WAVESIZE is in units of 256 dwords. */
5377 conf
->scratch_bytes_per_wave
=
5378 G_00B860_WAVESIZE(value
) * 256 * 4 * 1;
5382 static bool printed
;
5385 fprintf(stderr
, "Warning: LLVM emitted unknown "
5386 "config register: 0x%x\n", reg
);
5393 if (!conf
->spi_ps_input_addr
)
5394 conf
->spi_ps_input_addr
= conf
->spi_ps_input_ena
;
5398 void si_shader_apply_scratch_relocs(struct si_context
*sctx
,
5399 struct si_shader
*shader
,
5400 struct si_shader_config
*config
,
5401 uint64_t scratch_va
)
5404 uint32_t scratch_rsrc_dword0
= scratch_va
;
5405 uint32_t scratch_rsrc_dword1
=
5406 S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32)
5407 | S_008F04_STRIDE(config
->scratch_bytes_per_wave
/ 64);
5409 for (i
= 0 ; i
< shader
->binary
.reloc_count
; i
++) {
5410 const struct radeon_shader_reloc
*reloc
=
5411 &shader
->binary
.relocs
[i
];
5412 if (!strcmp(scratch_rsrc_dword0_symbol
, reloc
->name
)) {
5413 util_memcpy_cpu_to_le32(shader
->binary
.code
+ reloc
->offset
,
5414 &scratch_rsrc_dword0
, 4);
5415 } else if (!strcmp(scratch_rsrc_dword1_symbol
, reloc
->name
)) {
5416 util_memcpy_cpu_to_le32(shader
->binary
.code
+ reloc
->offset
,
5417 &scratch_rsrc_dword1
, 4);
5422 static unsigned si_get_shader_binary_size(struct si_shader
*shader
)
5424 unsigned size
= shader
->binary
.code_size
;
5427 size
+= shader
->prolog
->binary
.code_size
;
5429 size
+= shader
->epilog
->binary
.code_size
;
5433 int si_shader_binary_upload(struct si_screen
*sscreen
, struct si_shader
*shader
)
5435 const struct radeon_shader_binary
*prolog
=
5436 shader
->prolog
? &shader
->prolog
->binary
: NULL
;
5437 const struct radeon_shader_binary
*epilog
=
5438 shader
->epilog
? &shader
->epilog
->binary
: NULL
;
5439 const struct radeon_shader_binary
*mainb
= &shader
->binary
;
5440 unsigned bo_size
= si_get_shader_binary_size(shader
) +
5441 (!epilog
? mainb
->rodata_size
: 0);
5444 assert(!prolog
|| !prolog
->rodata_size
);
5445 assert((!prolog
&& !epilog
) || !mainb
->rodata_size
);
5446 assert(!epilog
|| !epilog
->rodata_size
);
5448 r600_resource_reference(&shader
->bo
, NULL
);
5449 shader
->bo
= si_resource_create_custom(&sscreen
->b
.b
,
5450 PIPE_USAGE_IMMUTABLE
,
5456 ptr
= sscreen
->b
.ws
->buffer_map(shader
->bo
->buf
, NULL
,
5457 PIPE_TRANSFER_READ_WRITE
);
5460 util_memcpy_cpu_to_le32(ptr
, prolog
->code
, prolog
->code_size
);
5461 ptr
+= prolog
->code_size
;
5464 util_memcpy_cpu_to_le32(ptr
, mainb
->code
, mainb
->code_size
);
5465 ptr
+= mainb
->code_size
;
5468 util_memcpy_cpu_to_le32(ptr
, epilog
->code
, epilog
->code_size
);
5469 else if (mainb
->rodata_size
> 0)
5470 util_memcpy_cpu_to_le32(ptr
, mainb
->rodata
, mainb
->rodata_size
);
5472 sscreen
->b
.ws
->buffer_unmap(shader
->bo
->buf
);
5476 static void si_shader_dump_disassembly(const struct radeon_shader_binary
*binary
,
5477 struct pipe_debug_callback
*debug
,
5478 const char *name
, FILE *file
)
5483 if (binary
->disasm_string
) {
5484 fprintf(file
, "Shader %s disassembly:\n", name
);
5485 fprintf(file
, "%s", binary
->disasm_string
);
5487 if (debug
&& debug
->debug_message
) {
5488 /* Very long debug messages are cut off, so send the
5489 * disassembly one line at a time. This causes more
5490 * overhead, but on the plus side it simplifies
5491 * parsing of resulting logs.
5493 pipe_debug_message(debug
, SHADER_INFO
,
5494 "Shader Disassembly Begin");
5496 line
= binary
->disasm_string
;
5498 p
= util_strchrnul(line
, '\n');
5502 pipe_debug_message(debug
, SHADER_INFO
,
5503 "%.*s", count
, line
);
5511 pipe_debug_message(debug
, SHADER_INFO
,
5512 "Shader Disassembly End");
5515 fprintf(file
, "Shader %s binary:\n", name
);
5516 for (i
= 0; i
< binary
->code_size
; i
+= 4) {
5517 fprintf(file
, "@0x%x: %02x%02x%02x%02x\n", i
,
5518 binary
->code
[i
+ 3], binary
->code
[i
+ 2],
5519 binary
->code
[i
+ 1], binary
->code
[i
]);
5524 static void si_shader_dump_stats(struct si_screen
*sscreen
,
5525 struct si_shader_config
*conf
,
5526 unsigned num_inputs
,
5528 struct pipe_debug_callback
*debug
,
5532 unsigned lds_increment
= sscreen
->b
.chip_class
>= CIK
? 512 : 256;
5533 unsigned lds_per_wave
= 0;
5534 unsigned max_simd_waves
= 10;
5536 /* Compute LDS usage for PS. */
5537 if (processor
== TGSI_PROCESSOR_FRAGMENT
) {
5538 /* The minimum usage per wave is (num_inputs * 36). The maximum
5539 * usage is (num_inputs * 36 * 16).
5540 * We can get anything in between and it varies between waves.
5542 * Other stages don't know the size at compile time or don't
5543 * allocate LDS per wave, but instead they do it per thread group.
5545 lds_per_wave
= conf
->lds_size
* lds_increment
+
5546 align(num_inputs
* 36, lds_increment
);
5549 /* Compute the per-SIMD wave counts. */
5550 if (conf
->num_sgprs
) {
5551 if (sscreen
->b
.chip_class
>= VI
)
5552 max_simd_waves
= MIN2(max_simd_waves
, 800 / conf
->num_sgprs
);
5554 max_simd_waves
= MIN2(max_simd_waves
, 512 / conf
->num_sgprs
);
5557 if (conf
->num_vgprs
)
5558 max_simd_waves
= MIN2(max_simd_waves
, 256 / conf
->num_vgprs
);
5560 /* LDS is 64KB per CU (4 SIMDs), divided into 16KB blocks per SIMD
5564 max_simd_waves
= MIN2(max_simd_waves
, 16384 / lds_per_wave
);
5566 if (file
!= stderr
||
5567 r600_can_dump_shader(&sscreen
->b
, processor
)) {
5568 if (processor
== TGSI_PROCESSOR_FRAGMENT
) {
5569 fprintf(file
, "*** SHADER CONFIG ***\n"
5570 "SPI_PS_INPUT_ADDR = 0x%04x\n"
5571 "SPI_PS_INPUT_ENA = 0x%04x\n",
5572 conf
->spi_ps_input_addr
, conf
->spi_ps_input_ena
);
5575 fprintf(file
, "*** SHADER STATS ***\n"
5578 "Code Size: %d bytes\n"
5580 "Scratch: %d bytes per wave\n"
5582 "********************\n",
5583 conf
->num_sgprs
, conf
->num_vgprs
, code_size
,
5584 conf
->lds_size
, conf
->scratch_bytes_per_wave
,
5588 pipe_debug_message(debug
, SHADER_INFO
,
5589 "Shader Stats: SGPRS: %d VGPRS: %d Code Size: %d "
5590 "LDS: %d Scratch: %d Max Waves: %d",
5591 conf
->num_sgprs
, conf
->num_vgprs
, code_size
,
5592 conf
->lds_size
, conf
->scratch_bytes_per_wave
,
5596 static const char *si_get_shader_name(struct si_shader
*shader
,
5599 switch (processor
) {
5600 case TGSI_PROCESSOR_VERTEX
:
5601 if (shader
->key
.vs
.as_es
)
5602 return "Vertex Shader as ES";
5603 else if (shader
->key
.vs
.as_ls
)
5604 return "Vertex Shader as LS";
5606 return "Vertex Shader as VS";
5607 case TGSI_PROCESSOR_TESS_CTRL
:
5608 return "Tessellation Control Shader";
5609 case TGSI_PROCESSOR_TESS_EVAL
:
5610 if (shader
->key
.tes
.as_es
)
5611 return "Tessellation Evaluation Shader as ES";
5613 return "Tessellation Evaluation Shader as VS";
5614 case TGSI_PROCESSOR_GEOMETRY
:
5615 if (shader
->gs_copy_shader
== NULL
)
5616 return "GS Copy Shader as VS";
5618 return "Geometry Shader";
5619 case TGSI_PROCESSOR_FRAGMENT
:
5620 return "Pixel Shader";
5621 case TGSI_PROCESSOR_COMPUTE
:
5622 return "Compute Shader";
5624 return "Unknown Shader";
5628 void si_shader_dump(struct si_screen
*sscreen
, struct si_shader
*shader
,
5629 struct pipe_debug_callback
*debug
, unsigned processor
,
5632 if (file
!= stderr
||
5633 (r600_can_dump_shader(&sscreen
->b
, processor
) &&
5634 !(sscreen
->b
.debug_flags
& DBG_NO_ASM
))) {
5635 fprintf(file
, "\n%s:\n", si_get_shader_name(shader
, processor
));
5638 si_shader_dump_disassembly(&shader
->prolog
->binary
,
5639 debug
, "prolog", file
);
5641 si_shader_dump_disassembly(&shader
->binary
, debug
, "main", file
);
5644 si_shader_dump_disassembly(&shader
->epilog
->binary
,
5645 debug
, "epilog", file
);
5646 fprintf(file
, "\n");
5649 si_shader_dump_stats(sscreen
, &shader
->config
,
5650 shader
->selector
? shader
->selector
->info
.num_inputs
: 0,
5651 si_get_shader_binary_size(shader
), debug
, processor
,
5655 int si_compile_llvm(struct si_screen
*sscreen
,
5656 struct radeon_shader_binary
*binary
,
5657 struct si_shader_config
*conf
,
5658 LLVMTargetMachineRef tm
,
5660 struct pipe_debug_callback
*debug
,
5665 unsigned count
= p_atomic_inc_return(&sscreen
->b
.num_compilations
);
5667 if (r600_can_dump_shader(&sscreen
->b
, processor
)) {
5668 fprintf(stderr
, "radeonsi: Compiling shader %d\n", count
);
5670 if (!(sscreen
->b
.debug_flags
& (DBG_NO_IR
| DBG_PREOPT_IR
))) {
5671 fprintf(stderr
, "%s LLVM IR:\n\n", name
);
5672 LLVMDumpModule(mod
);
5673 fprintf(stderr
, "\n");
5677 if (!si_replace_shader(count
, binary
)) {
5678 r
= radeon_llvm_compile(mod
, binary
,
5679 r600_get_llvm_processor_name(sscreen
->b
.family
), tm
,
5685 si_shader_binary_read_config(binary
, conf
, 0);
5687 /* Enable 64-bit and 16-bit denormals, because there is no performance
5690 * If denormals are enabled, all floating-point output modifiers are
5693 * Don't enable denormals for 32-bit floats, because:
5694 * - Floating-point output modifiers would be ignored by the hw.
5695 * - Some opcodes don't support denormals, such as v_mad_f32. We would
5696 * have to stop using those.
5697 * - SI & CI would be very slow.
5699 conf
->float_mode
|= V_00B028_FP_64_DENORMS
;
5701 FREE(binary
->config
);
5702 FREE(binary
->global_symbol_offsets
);
5703 binary
->config
= NULL
;
5704 binary
->global_symbol_offsets
= NULL
;
5706 /* Some shaders can't have rodata because their binaries can be
5709 if (binary
->rodata_size
&&
5710 (processor
== TGSI_PROCESSOR_VERTEX
||
5711 processor
== TGSI_PROCESSOR_TESS_CTRL
||
5712 processor
== TGSI_PROCESSOR_TESS_EVAL
||
5713 processor
== TGSI_PROCESSOR_FRAGMENT
)) {
5714 fprintf(stderr
, "radeonsi: The shader can't have rodata.");
5721 /* Generate code for the hardware VS shader stage to go with a geometry shader */
5722 static int si_generate_gs_copy_shader(struct si_screen
*sscreen
,
5723 struct si_shader_context
*ctx
,
5724 struct si_shader
*gs
,
5725 struct pipe_debug_callback
*debug
)
5727 struct gallivm_state
*gallivm
= &ctx
->radeon_bld
.gallivm
;
5728 struct lp_build_tgsi_context
*bld_base
= &ctx
->radeon_bld
.soa
.bld_base
;
5729 struct lp_build_context
*uint
= &bld_base
->uint_bld
;
5730 struct si_shader_output_values
*outputs
;
5731 struct tgsi_shader_info
*gsinfo
= &gs
->selector
->info
;
5732 LLVMValueRef args
[9];
5735 outputs
= MALLOC(gsinfo
->num_outputs
* sizeof(outputs
[0]));
5737 si_init_shader_ctx(ctx
, sscreen
, ctx
->shader
, ctx
->tm
);
5738 ctx
->type
= TGSI_PROCESSOR_VERTEX
;
5739 ctx
->is_gs_copy_shader
= true;
5741 create_meta_data(ctx
);
5742 create_function(ctx
);
5743 preload_streamout_buffers(ctx
);
5744 preload_ring_buffers(ctx
);
5746 args
[0] = ctx
->gsvs_ring
[0];
5747 args
[1] = lp_build_mul_imm(uint
,
5748 LLVMGetParam(ctx
->radeon_bld
.main_fn
,
5749 ctx
->param_vertex_id
),
5751 args
[3] = uint
->zero
;
5752 args
[4] = uint
->one
; /* OFFEN */
5753 args
[5] = uint
->zero
; /* IDXEN */
5754 args
[6] = uint
->one
; /* GLC */
5755 args
[7] = uint
->one
; /* SLC */
5756 args
[8] = uint
->zero
; /* TFE */
5758 /* Fetch vertex data from GSVS ring */
5759 for (i
= 0; i
< gsinfo
->num_outputs
; ++i
) {
5762 outputs
[i
].name
= gsinfo
->output_semantic_name
[i
];
5763 outputs
[i
].sid
= gsinfo
->output_semantic_index
[i
];
5765 for (chan
= 0; chan
< 4; chan
++) {
5766 args
[2] = lp_build_const_int32(gallivm
,
5768 gs
->selector
->gs_max_out_vertices
* 16 * 4);
5770 outputs
[i
].values
[chan
] =
5771 LLVMBuildBitCast(gallivm
->builder
,
5772 lp_build_intrinsic(gallivm
->builder
,
5773 "llvm.SI.buffer.load.dword.i32.i32",
5775 LLVMReadOnlyAttribute
| LLVMNoUnwindAttribute
),
5780 si_llvm_export_vs(bld_base
, outputs
, gsinfo
->num_outputs
);
5782 LLVMBuildRet(gallivm
->builder
, ctx
->return_value
);
5784 /* Dump LLVM IR before any optimization passes */
5785 if (sscreen
->b
.debug_flags
& DBG_PREOPT_IR
&&
5786 r600_can_dump_shader(&sscreen
->b
, TGSI_PROCESSOR_GEOMETRY
))
5787 LLVMDumpModule(bld_base
->base
.gallivm
->module
);
5789 radeon_llvm_finalize_module(&ctx
->radeon_bld
);
5791 r
= si_compile_llvm(sscreen
, &ctx
->shader
->binary
,
5792 &ctx
->shader
->config
, ctx
->tm
,
5793 bld_base
->base
.gallivm
->module
,
5794 debug
, TGSI_PROCESSOR_GEOMETRY
,
5797 if (r600_can_dump_shader(&sscreen
->b
, TGSI_PROCESSOR_GEOMETRY
))
5798 fprintf(stderr
, "GS Copy Shader:\n");
5799 si_shader_dump(sscreen
, ctx
->shader
, debug
,
5800 TGSI_PROCESSOR_GEOMETRY
, stderr
);
5801 r
= si_shader_binary_upload(sscreen
, ctx
->shader
);
5804 radeon_llvm_dispose(&ctx
->radeon_bld
);
5810 void si_dump_shader_key(unsigned shader
, union si_shader_key
*key
, FILE *f
)
5814 fprintf(f
, "SHADER KEY\n");
5817 case PIPE_SHADER_VERTEX
:
5818 fprintf(f
, " instance_divisors = {");
5819 for (i
= 0; i
< Elements(key
->vs
.prolog
.instance_divisors
); i
++)
5820 fprintf(f
, !i
? "%u" : ", %u",
5821 key
->vs
.prolog
.instance_divisors
[i
]);
5823 fprintf(f
, " as_es = %u\n", key
->vs
.as_es
);
5824 fprintf(f
, " as_ls = %u\n", key
->vs
.as_ls
);
5825 fprintf(f
, " export_prim_id = %u\n", key
->vs
.epilog
.export_prim_id
);
5828 case PIPE_SHADER_TESS_CTRL
:
5829 fprintf(f
, " prim_mode = %u\n", key
->tcs
.epilog
.prim_mode
);
5832 case PIPE_SHADER_TESS_EVAL
:
5833 fprintf(f
, " as_es = %u\n", key
->tes
.as_es
);
5834 fprintf(f
, " export_prim_id = %u\n", key
->tes
.epilog
.export_prim_id
);
5837 case PIPE_SHADER_GEOMETRY
:
5838 case PIPE_SHADER_COMPUTE
:
5841 case PIPE_SHADER_FRAGMENT
:
5842 fprintf(f
, " prolog.color_two_side = %u\n", key
->ps
.prolog
.color_two_side
);
5843 fprintf(f
, " prolog.poly_stipple = %u\n", key
->ps
.prolog
.poly_stipple
);
5844 fprintf(f
, " prolog.force_persample_interp = %u\n", key
->ps
.prolog
.force_persample_interp
);
5845 fprintf(f
, " epilog.spi_shader_col_format = 0x%x\n", key
->ps
.epilog
.spi_shader_col_format
);
5846 fprintf(f
, " epilog.color_is_int8 = 0x%X\n", key
->ps
.epilog
.color_is_int8
);
5847 fprintf(f
, " epilog.last_cbuf = %u\n", key
->ps
.epilog
.last_cbuf
);
5848 fprintf(f
, " epilog.alpha_func = %u\n", key
->ps
.epilog
.alpha_func
);
5849 fprintf(f
, " epilog.alpha_to_one = %u\n", key
->ps
.epilog
.alpha_to_one
);
5850 fprintf(f
, " epilog.poly_line_smoothing = %u\n", key
->ps
.epilog
.poly_line_smoothing
);
5851 fprintf(f
, " epilog.clamp_color = %u\n", key
->ps
.epilog
.clamp_color
);
5859 static void si_init_shader_ctx(struct si_shader_context
*ctx
,
5860 struct si_screen
*sscreen
,
5861 struct si_shader
*shader
,
5862 LLVMTargetMachineRef tm
)
5864 struct lp_build_tgsi_context
*bld_base
;
5865 struct lp_build_tgsi_action tmpl
= {};
5867 memset(ctx
, 0, sizeof(*ctx
));
5868 radeon_llvm_context_init(&ctx
->radeon_bld
, "amdgcn--");
5870 ctx
->screen
= sscreen
;
5871 if (shader
&& shader
->selector
)
5872 ctx
->type
= shader
->selector
->info
.processor
;
5875 ctx
->shader
= shader
;
5877 ctx
->voidt
= LLVMVoidTypeInContext(ctx
->radeon_bld
.gallivm
.context
);
5878 ctx
->i1
= LLVMInt1TypeInContext(ctx
->radeon_bld
.gallivm
.context
);
5879 ctx
->i8
= LLVMInt8TypeInContext(ctx
->radeon_bld
.gallivm
.context
);
5880 ctx
->i32
= LLVMInt32TypeInContext(ctx
->radeon_bld
.gallivm
.context
);
5881 ctx
->i64
= LLVMInt64TypeInContext(ctx
->radeon_bld
.gallivm
.context
);
5882 ctx
->i128
= LLVMIntTypeInContext(ctx
->radeon_bld
.gallivm
.context
, 128);
5883 ctx
->f32
= LLVMFloatTypeInContext(ctx
->radeon_bld
.gallivm
.context
);
5884 ctx
->v16i8
= LLVMVectorType(ctx
->i8
, 16);
5885 ctx
->v2i32
= LLVMVectorType(ctx
->i32
, 2);
5886 ctx
->v4i32
= LLVMVectorType(ctx
->i32
, 4);
5887 ctx
->v4f32
= LLVMVectorType(ctx
->f32
, 4);
5888 ctx
->v8i32
= LLVMVectorType(ctx
->i32
, 8);
5890 bld_base
= &ctx
->radeon_bld
.soa
.bld_base
;
5891 if (shader
&& shader
->selector
)
5892 bld_base
->info
= &shader
->selector
->info
;
5893 bld_base
->emit_fetch_funcs
[TGSI_FILE_CONSTANT
] = fetch_constant
;
5895 bld_base
->op_actions
[TGSI_OPCODE_INTERP_CENTROID
] = interp_action
;
5896 bld_base
->op_actions
[TGSI_OPCODE_INTERP_SAMPLE
] = interp_action
;
5897 bld_base
->op_actions
[TGSI_OPCODE_INTERP_OFFSET
] = interp_action
;
5899 bld_base
->op_actions
[TGSI_OPCODE_TEX
] = tex_action
;
5900 bld_base
->op_actions
[TGSI_OPCODE_TEX2
] = tex_action
;
5901 bld_base
->op_actions
[TGSI_OPCODE_TXB
] = tex_action
;
5902 bld_base
->op_actions
[TGSI_OPCODE_TXB2
] = tex_action
;
5903 bld_base
->op_actions
[TGSI_OPCODE_TXD
] = tex_action
;
5904 bld_base
->op_actions
[TGSI_OPCODE_TXF
] = tex_action
;
5905 bld_base
->op_actions
[TGSI_OPCODE_TXL
] = tex_action
;
5906 bld_base
->op_actions
[TGSI_OPCODE_TXL2
] = tex_action
;
5907 bld_base
->op_actions
[TGSI_OPCODE_TXP
] = tex_action
;
5908 bld_base
->op_actions
[TGSI_OPCODE_TXQ
] = tex_action
;
5909 bld_base
->op_actions
[TGSI_OPCODE_TG4
] = tex_action
;
5910 bld_base
->op_actions
[TGSI_OPCODE_LODQ
] = tex_action
;
5911 bld_base
->op_actions
[TGSI_OPCODE_TXQS
].emit
= si_llvm_emit_txqs
;
5913 bld_base
->op_actions
[TGSI_OPCODE_LOAD
].fetch_args
= load_fetch_args
;
5914 bld_base
->op_actions
[TGSI_OPCODE_LOAD
].emit
= load_emit
;
5915 bld_base
->op_actions
[TGSI_OPCODE_STORE
].fetch_args
= store_fetch_args
;
5916 bld_base
->op_actions
[TGSI_OPCODE_STORE
].emit
= store_emit
;
5917 bld_base
->op_actions
[TGSI_OPCODE_RESQ
].fetch_args
= resq_fetch_args
;
5918 bld_base
->op_actions
[TGSI_OPCODE_RESQ
].emit
= resq_emit
;
5920 tmpl
.fetch_args
= atomic_fetch_args
;
5921 tmpl
.emit
= atomic_emit
;
5922 bld_base
->op_actions
[TGSI_OPCODE_ATOMUADD
] = tmpl
;
5923 bld_base
->op_actions
[TGSI_OPCODE_ATOMUADD
].intr_name
= "add";
5924 bld_base
->op_actions
[TGSI_OPCODE_ATOMXCHG
] = tmpl
;
5925 bld_base
->op_actions
[TGSI_OPCODE_ATOMXCHG
].intr_name
= "swap";
5926 bld_base
->op_actions
[TGSI_OPCODE_ATOMCAS
] = tmpl
;
5927 bld_base
->op_actions
[TGSI_OPCODE_ATOMCAS
].intr_name
= "cmpswap";
5928 bld_base
->op_actions
[TGSI_OPCODE_ATOMAND
] = tmpl
;
5929 bld_base
->op_actions
[TGSI_OPCODE_ATOMAND
].intr_name
= "and";
5930 bld_base
->op_actions
[TGSI_OPCODE_ATOMOR
] = tmpl
;
5931 bld_base
->op_actions
[TGSI_OPCODE_ATOMOR
].intr_name
= "or";
5932 bld_base
->op_actions
[TGSI_OPCODE_ATOMXOR
] = tmpl
;
5933 bld_base
->op_actions
[TGSI_OPCODE_ATOMXOR
].intr_name
= "xor";
5934 bld_base
->op_actions
[TGSI_OPCODE_ATOMUMIN
] = tmpl
;
5935 bld_base
->op_actions
[TGSI_OPCODE_ATOMUMIN
].intr_name
= "umin";
5936 bld_base
->op_actions
[TGSI_OPCODE_ATOMUMAX
] = tmpl
;
5937 bld_base
->op_actions
[TGSI_OPCODE_ATOMUMAX
].intr_name
= "umax";
5938 bld_base
->op_actions
[TGSI_OPCODE_ATOMIMIN
] = tmpl
;
5939 bld_base
->op_actions
[TGSI_OPCODE_ATOMIMIN
].intr_name
= "smin";
5940 bld_base
->op_actions
[TGSI_OPCODE_ATOMIMAX
] = tmpl
;
5941 bld_base
->op_actions
[TGSI_OPCODE_ATOMIMAX
].intr_name
= "smax";
5943 bld_base
->op_actions
[TGSI_OPCODE_MEMBAR
].emit
= membar_emit
;
5945 bld_base
->op_actions
[TGSI_OPCODE_DDX
].emit
= si_llvm_emit_ddxy
;
5946 bld_base
->op_actions
[TGSI_OPCODE_DDY
].emit
= si_llvm_emit_ddxy
;
5947 bld_base
->op_actions
[TGSI_OPCODE_DDX_FINE
].emit
= si_llvm_emit_ddxy
;
5948 bld_base
->op_actions
[TGSI_OPCODE_DDY_FINE
].emit
= si_llvm_emit_ddxy
;
5950 bld_base
->op_actions
[TGSI_OPCODE_EMIT
].emit
= si_llvm_emit_vertex
;
5951 bld_base
->op_actions
[TGSI_OPCODE_ENDPRIM
].emit
= si_llvm_emit_primitive
;
5952 bld_base
->op_actions
[TGSI_OPCODE_BARRIER
].emit
= si_llvm_emit_barrier
;
5954 bld_base
->op_actions
[TGSI_OPCODE_MAX
].emit
= build_tgsi_intrinsic_nomem
;
5955 bld_base
->op_actions
[TGSI_OPCODE_MAX
].intr_name
= "llvm.maxnum.f32";
5956 bld_base
->op_actions
[TGSI_OPCODE_MIN
].emit
= build_tgsi_intrinsic_nomem
;
5957 bld_base
->op_actions
[TGSI_OPCODE_MIN
].intr_name
= "llvm.minnum.f32";
5960 int si_compile_tgsi_shader(struct si_screen
*sscreen
,
5961 LLVMTargetMachineRef tm
,
5962 struct si_shader
*shader
,
5964 struct pipe_debug_callback
*debug
)
5966 struct si_shader_selector
*sel
= shader
->selector
;
5967 struct si_shader_context ctx
;
5968 struct lp_build_tgsi_context
*bld_base
;
5972 /* Dump TGSI code before doing TGSI->LLVM conversion in case the
5973 * conversion fails. */
5974 if (r600_can_dump_shader(&sscreen
->b
, sel
->info
.processor
) &&
5975 !(sscreen
->b
.debug_flags
& DBG_NO_TGSI
)) {
5976 si_dump_shader_key(sel
->type
, &shader
->key
, stderr
);
5977 tgsi_dump(sel
->tokens
, 0);
5978 si_dump_streamout(&sel
->so
);
5981 si_init_shader_ctx(&ctx
, sscreen
, shader
, tm
);
5982 ctx
.is_monolithic
= is_monolithic
;
5984 shader
->info
.uses_instanceid
= sel
->info
.uses_instanceid
;
5986 bld_base
= &ctx
.radeon_bld
.soa
.bld_base
;
5987 ctx
.radeon_bld
.load_system_value
= declare_system_value
;
5990 case TGSI_PROCESSOR_VERTEX
:
5991 ctx
.radeon_bld
.load_input
= declare_input_vs
;
5992 if (shader
->key
.vs
.as_ls
)
5993 bld_base
->emit_epilogue
= si_llvm_emit_ls_epilogue
;
5994 else if (shader
->key
.vs
.as_es
)
5995 bld_base
->emit_epilogue
= si_llvm_emit_es_epilogue
;
5997 bld_base
->emit_epilogue
= si_llvm_emit_vs_epilogue
;
5999 case TGSI_PROCESSOR_TESS_CTRL
:
6000 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = fetch_input_tcs
;
6001 bld_base
->emit_fetch_funcs
[TGSI_FILE_OUTPUT
] = fetch_output_tcs
;
6002 bld_base
->emit_store
= store_output_tcs
;
6003 bld_base
->emit_epilogue
= si_llvm_emit_tcs_epilogue
;
6005 case TGSI_PROCESSOR_TESS_EVAL
:
6006 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = fetch_input_tes
;
6007 if (shader
->key
.tes
.as_es
)
6008 bld_base
->emit_epilogue
= si_llvm_emit_es_epilogue
;
6010 bld_base
->emit_epilogue
= si_llvm_emit_vs_epilogue
;
6012 case TGSI_PROCESSOR_GEOMETRY
:
6013 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = fetch_input_gs
;
6014 bld_base
->emit_epilogue
= si_llvm_emit_gs_epilogue
;
6016 case TGSI_PROCESSOR_FRAGMENT
:
6017 ctx
.radeon_bld
.load_input
= declare_input_fs
;
6019 bld_base
->emit_epilogue
= si_llvm_emit_fs_epilogue
;
6021 bld_base
->emit_epilogue
= si_llvm_return_fs_outputs
;
6023 case TGSI_PROCESSOR_COMPUTE
:
6024 ctx
.radeon_bld
.declare_memory_region
= declare_compute_memory
;
6027 assert(!"Unsupported shader type");
6031 create_meta_data(&ctx
);
6032 create_function(&ctx
);
6033 preload_constants(&ctx
);
6034 preload_shader_buffers(&ctx
);
6035 preload_samplers(&ctx
);
6036 preload_images(&ctx
);
6037 preload_streamout_buffers(&ctx
);
6038 preload_ring_buffers(&ctx
);
6040 if (ctx
.is_monolithic
&& sel
->type
== PIPE_SHADER_FRAGMENT
&&
6041 shader
->key
.ps
.prolog
.poly_stipple
) {
6042 LLVMValueRef views
= LLVMGetParam(ctx
.radeon_bld
.main_fn
,
6044 si_llvm_emit_polygon_stipple(&ctx
, views
,
6045 SI_PARAM_POS_FIXED_PT
);
6048 if (ctx
.type
== TGSI_PROCESSOR_GEOMETRY
) {
6050 for (i
= 0; i
< 4; i
++) {
6051 ctx
.gs_next_vertex
[i
] =
6052 lp_build_alloca(bld_base
->base
.gallivm
,
6057 if (!lp_build_tgsi_llvm(bld_base
, sel
->tokens
)) {
6058 fprintf(stderr
, "Failed to translate shader from TGSI to LLVM\n");
6062 LLVMBuildRet(bld_base
->base
.gallivm
->builder
, ctx
.return_value
);
6063 mod
= bld_base
->base
.gallivm
->module
;
6065 /* Dump LLVM IR before any optimization passes */
6066 if (sscreen
->b
.debug_flags
& DBG_PREOPT_IR
&&
6067 r600_can_dump_shader(&sscreen
->b
, ctx
.type
))
6068 LLVMDumpModule(mod
);
6070 radeon_llvm_finalize_module(&ctx
.radeon_bld
);
6072 r
= si_compile_llvm(sscreen
, &shader
->binary
, &shader
->config
, tm
,
6073 mod
, debug
, ctx
.type
, "TGSI shader");
6075 fprintf(stderr
, "LLVM failed to compile shader\n");
6079 radeon_llvm_dispose(&ctx
.radeon_bld
);
6081 /* Add the scratch offset to input SGPRs. */
6082 if (shader
->config
.scratch_bytes_per_wave
)
6083 shader
->info
.num_input_sgprs
+= 1; /* scratch byte offset */
6085 /* Calculate the number of fragment input VGPRs. */
6086 if (ctx
.type
== TGSI_PROCESSOR_FRAGMENT
) {
6087 shader
->info
.num_input_vgprs
= 0;
6088 shader
->info
.face_vgpr_index
= -1;
6090 if (G_0286CC_PERSP_SAMPLE_ENA(shader
->config
.spi_ps_input_addr
))
6091 shader
->info
.num_input_vgprs
+= 2;
6092 if (G_0286CC_PERSP_CENTER_ENA(shader
->config
.spi_ps_input_addr
))
6093 shader
->info
.num_input_vgprs
+= 2;
6094 if (G_0286CC_PERSP_CENTROID_ENA(shader
->config
.spi_ps_input_addr
))
6095 shader
->info
.num_input_vgprs
+= 2;
6096 if (G_0286CC_PERSP_PULL_MODEL_ENA(shader
->config
.spi_ps_input_addr
))
6097 shader
->info
.num_input_vgprs
+= 3;
6098 if (G_0286CC_LINEAR_SAMPLE_ENA(shader
->config
.spi_ps_input_addr
))
6099 shader
->info
.num_input_vgprs
+= 2;
6100 if (G_0286CC_LINEAR_CENTER_ENA(shader
->config
.spi_ps_input_addr
))
6101 shader
->info
.num_input_vgprs
+= 2;
6102 if (G_0286CC_LINEAR_CENTROID_ENA(shader
->config
.spi_ps_input_addr
))
6103 shader
->info
.num_input_vgprs
+= 2;
6104 if (G_0286CC_LINE_STIPPLE_TEX_ENA(shader
->config
.spi_ps_input_addr
))
6105 shader
->info
.num_input_vgprs
+= 1;
6106 if (G_0286CC_POS_X_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
6107 shader
->info
.num_input_vgprs
+= 1;
6108 if (G_0286CC_POS_Y_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
6109 shader
->info
.num_input_vgprs
+= 1;
6110 if (G_0286CC_POS_Z_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
6111 shader
->info
.num_input_vgprs
+= 1;
6112 if (G_0286CC_POS_W_FLOAT_ENA(shader
->config
.spi_ps_input_addr
))
6113 shader
->info
.num_input_vgprs
+= 1;
6114 if (G_0286CC_FRONT_FACE_ENA(shader
->config
.spi_ps_input_addr
)) {
6115 shader
->info
.face_vgpr_index
= shader
->info
.num_input_vgprs
;
6116 shader
->info
.num_input_vgprs
+= 1;
6118 if (G_0286CC_ANCILLARY_ENA(shader
->config
.spi_ps_input_addr
))
6119 shader
->info
.num_input_vgprs
+= 1;
6120 if (G_0286CC_SAMPLE_COVERAGE_ENA(shader
->config
.spi_ps_input_addr
))
6121 shader
->info
.num_input_vgprs
+= 1;
6122 if (G_0286CC_POS_FIXED_PT_ENA(shader
->config
.spi_ps_input_addr
))
6123 shader
->info
.num_input_vgprs
+= 1;
6126 if (ctx
.type
== TGSI_PROCESSOR_GEOMETRY
) {
6127 shader
->gs_copy_shader
= CALLOC_STRUCT(si_shader
);
6128 shader
->gs_copy_shader
->selector
= shader
->selector
;
6129 ctx
.shader
= shader
->gs_copy_shader
;
6130 if ((r
= si_generate_gs_copy_shader(sscreen
, &ctx
,
6132 free(shader
->gs_copy_shader
);
6133 shader
->gs_copy_shader
= NULL
;
6139 for (int i
= 0; i
< SI_NUM_CONST_BUFFERS
; i
++)
6140 FREE(ctx
.constants
[i
]);
6145 * Create, compile and return a shader part (prolog or epilog).
6147 * \param sscreen screen
6148 * \param list list of shader parts of the same category
6149 * \param key shader part key
6150 * \param tm LLVM target machine
6151 * \param debug debug callback
6152 * \param compile the callback responsible for compilation
6153 * \return non-NULL on success
6155 static struct si_shader_part
*
6156 si_get_shader_part(struct si_screen
*sscreen
,
6157 struct si_shader_part
**list
,
6158 union si_shader_part_key
*key
,
6159 LLVMTargetMachineRef tm
,
6160 struct pipe_debug_callback
*debug
,
6161 bool (*compile
)(struct si_screen
*,
6162 LLVMTargetMachineRef
,
6163 struct pipe_debug_callback
*,
6164 struct si_shader_part
*))
6166 struct si_shader_part
*result
;
6168 pipe_mutex_lock(sscreen
->shader_parts_mutex
);
6170 /* Find existing. */
6171 for (result
= *list
; result
; result
= result
->next
) {
6172 if (memcmp(&result
->key
, key
, sizeof(*key
)) == 0) {
6173 pipe_mutex_unlock(sscreen
->shader_parts_mutex
);
6178 /* Compile a new one. */
6179 result
= CALLOC_STRUCT(si_shader_part
);
6181 if (!compile(sscreen
, tm
, debug
, result
)) {
6183 pipe_mutex_unlock(sscreen
->shader_parts_mutex
);
6187 result
->next
= *list
;
6189 pipe_mutex_unlock(sscreen
->shader_parts_mutex
);
6194 * Create a vertex shader prolog.
6196 * The inputs are the same as VS (a lot of SGPRs and 4 VGPR system values).
6197 * All inputs are returned unmodified. The vertex load indices are
6198 * stored after them, which will used by the API VS for fetching inputs.
6200 * For example, the expected outputs for instance_divisors[] = {0, 1, 2} are:
6205 * (VertexID + BaseVertex),
6206 * (InstanceID + StartInstance),
6207 * (InstanceID / 2 + StartInstance)
6209 static bool si_compile_vs_prolog(struct si_screen
*sscreen
,
6210 LLVMTargetMachineRef tm
,
6211 struct pipe_debug_callback
*debug
,
6212 struct si_shader_part
*out
)
6214 union si_shader_part_key
*key
= &out
->key
;
6215 struct si_shader shader
= {};
6216 struct si_shader_context ctx
;
6217 struct gallivm_state
*gallivm
= &ctx
.radeon_bld
.gallivm
;
6218 LLVMTypeRef
*params
, *returns
;
6219 LLVMValueRef ret
, func
;
6220 int last_sgpr
, num_params
, num_returns
, i
;
6223 si_init_shader_ctx(&ctx
, sscreen
, &shader
, tm
);
6224 ctx
.type
= TGSI_PROCESSOR_VERTEX
;
6225 ctx
.param_vertex_id
= key
->vs_prolog
.num_input_sgprs
;
6226 ctx
.param_instance_id
= key
->vs_prolog
.num_input_sgprs
+ 3;
6228 /* 4 preloaded VGPRs + vertex load indices as prolog outputs */
6229 params
= alloca((key
->vs_prolog
.num_input_sgprs
+ 4) *
6230 sizeof(LLVMTypeRef
));
6231 returns
= alloca((key
->vs_prolog
.num_input_sgprs
+ 4 +
6232 key
->vs_prolog
.last_input
+ 1) *
6233 sizeof(LLVMTypeRef
));
6237 /* Declare input and output SGPRs. */
6239 for (i
= 0; i
< key
->vs_prolog
.num_input_sgprs
; i
++) {
6240 params
[num_params
++] = ctx
.i32
;
6241 returns
[num_returns
++] = ctx
.i32
;
6243 last_sgpr
= num_params
- 1;
6245 /* 4 preloaded VGPRs (outputs must be floats) */
6246 for (i
= 0; i
< 4; i
++) {
6247 params
[num_params
++] = ctx
.i32
;
6248 returns
[num_returns
++] = ctx
.f32
;
6251 /* Vertex load indices. */
6252 for (i
= 0; i
<= key
->vs_prolog
.last_input
; i
++)
6253 returns
[num_returns
++] = ctx
.f32
;
6255 /* Create the function. */
6256 si_create_function(&ctx
, returns
, num_returns
, params
,
6257 num_params
, -1, last_sgpr
);
6258 func
= ctx
.radeon_bld
.main_fn
;
6260 /* Copy inputs to outputs. This should be no-op, as the registers match,
6261 * but it will prevent the compiler from overwriting them unintentionally.
6263 ret
= ctx
.return_value
;
6264 for (i
= 0; i
< key
->vs_prolog
.num_input_sgprs
; i
++) {
6265 LLVMValueRef p
= LLVMGetParam(func
, i
);
6266 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
, p
, i
, "");
6268 for (i
= num_params
- 4; i
< num_params
; i
++) {
6269 LLVMValueRef p
= LLVMGetParam(func
, i
);
6270 p
= LLVMBuildBitCast(gallivm
->builder
, p
, ctx
.f32
, "");
6271 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
, p
, i
, "");
6274 /* Compute vertex load indices from instance divisors. */
6275 for (i
= 0; i
<= key
->vs_prolog
.last_input
; i
++) {
6276 unsigned divisor
= key
->vs_prolog
.states
.instance_divisors
[i
];
6280 /* InstanceID / Divisor + StartInstance */
6281 index
= get_instance_index_for_fetch(&ctx
.radeon_bld
,
6282 SI_SGPR_START_INSTANCE
,
6285 /* VertexID + BaseVertex */
6286 index
= LLVMBuildAdd(gallivm
->builder
,
6287 LLVMGetParam(func
, ctx
.param_vertex_id
),
6288 LLVMGetParam(func
, SI_SGPR_BASE_VERTEX
), "");
6291 index
= LLVMBuildBitCast(gallivm
->builder
, index
, ctx
.f32
, "");
6292 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
, index
,
6297 LLVMBuildRet(gallivm
->builder
, ret
);
6298 radeon_llvm_finalize_module(&ctx
.radeon_bld
);
6300 if (si_compile_llvm(sscreen
, &out
->binary
, &out
->config
, tm
,
6301 gallivm
->module
, debug
, ctx
.type
,
6302 "Vertex Shader Prolog"))
6305 radeon_llvm_dispose(&ctx
.radeon_bld
);
6310 * Compile the vertex shader epilog. This is also used by the tessellation
6311 * evaluation shader compiled as VS.
6313 * The input is PrimitiveID.
6315 * If PrimitiveID is required by the pixel shader, export it.
6316 * Otherwise, do nothing.
6318 static bool si_compile_vs_epilog(struct si_screen
*sscreen
,
6319 LLVMTargetMachineRef tm
,
6320 struct pipe_debug_callback
*debug
,
6321 struct si_shader_part
*out
)
6323 union si_shader_part_key
*key
= &out
->key
;
6324 struct si_shader_context ctx
;
6325 struct gallivm_state
*gallivm
= &ctx
.radeon_bld
.gallivm
;
6326 struct lp_build_tgsi_context
*bld_base
= &ctx
.radeon_bld
.soa
.bld_base
;
6327 LLVMTypeRef params
[5];
6331 si_init_shader_ctx(&ctx
, sscreen
, NULL
, tm
);
6332 ctx
.type
= TGSI_PROCESSOR_VERTEX
;
6334 /* Declare input VGPRs. */
6335 num_params
= key
->vs_epilog
.states
.export_prim_id
?
6336 (VS_EPILOG_PRIMID_LOC
+ 1) : 0;
6337 assert(num_params
<= ARRAY_SIZE(params
));
6339 for (i
= 0; i
< num_params
; i
++)
6340 params
[i
] = ctx
.f32
;
6342 /* Create the function. */
6343 si_create_function(&ctx
, NULL
, 0, params
, num_params
,
6347 if (key
->vs_epilog
.states
.export_prim_id
) {
6348 struct lp_build_context
*base
= &bld_base
->base
;
6349 struct lp_build_context
*uint
= &bld_base
->uint_bld
;
6350 LLVMValueRef args
[9];
6352 args
[0] = lp_build_const_int32(base
->gallivm
, 0x0); /* enabled channels */
6353 args
[1] = uint
->zero
; /* whether the EXEC mask is valid */
6354 args
[2] = uint
->zero
; /* DONE bit */
6355 args
[3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_PARAM
+
6356 key
->vs_epilog
.prim_id_param_offset
);
6357 args
[4] = uint
->zero
; /* COMPR flag (0 = 32-bit export) */
6358 args
[5] = LLVMGetParam(ctx
.radeon_bld
.main_fn
,
6359 VS_EPILOG_PRIMID_LOC
); /* X */
6360 args
[6] = uint
->undef
; /* Y */
6361 args
[7] = uint
->undef
; /* Z */
6362 args
[8] = uint
->undef
; /* W */
6364 lp_build_intrinsic(base
->gallivm
->builder
, "llvm.SI.export",
6365 LLVMVoidTypeInContext(base
->gallivm
->context
),
6370 LLVMBuildRet(gallivm
->builder
, ctx
.return_value
);
6371 radeon_llvm_finalize_module(&ctx
.radeon_bld
);
6373 if (si_compile_llvm(sscreen
, &out
->binary
, &out
->config
, tm
,
6374 gallivm
->module
, debug
, ctx
.type
,
6375 "Vertex Shader Epilog"))
6378 radeon_llvm_dispose(&ctx
.radeon_bld
);
6383 * Create & compile a vertex shader epilog. This a helper used by VS and TES.
6385 static bool si_get_vs_epilog(struct si_screen
*sscreen
,
6386 LLVMTargetMachineRef tm
,
6387 struct si_shader
*shader
,
6388 struct pipe_debug_callback
*debug
,
6389 struct si_vs_epilog_bits
*states
)
6391 union si_shader_part_key epilog_key
;
6393 memset(&epilog_key
, 0, sizeof(epilog_key
));
6394 epilog_key
.vs_epilog
.states
= *states
;
6396 /* Set up the PrimitiveID output. */
6397 if (shader
->key
.vs
.epilog
.export_prim_id
) {
6398 unsigned index
= shader
->selector
->info
.num_outputs
;
6399 unsigned offset
= shader
->info
.nr_param_exports
++;
6401 epilog_key
.vs_epilog
.prim_id_param_offset
= offset
;
6402 assert(index
< ARRAY_SIZE(shader
->info
.vs_output_param_offset
));
6403 shader
->info
.vs_output_param_offset
[index
] = offset
;
6406 shader
->epilog
= si_get_shader_part(sscreen
, &sscreen
->vs_epilogs
,
6407 &epilog_key
, tm
, debug
,
6408 si_compile_vs_epilog
);
6409 return shader
->epilog
!= NULL
;
6413 * Select and compile (or reuse) vertex shader parts (prolog & epilog).
6415 static bool si_shader_select_vs_parts(struct si_screen
*sscreen
,
6416 LLVMTargetMachineRef tm
,
6417 struct si_shader
*shader
,
6418 struct pipe_debug_callback
*debug
)
6420 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
6421 union si_shader_part_key prolog_key
;
6424 /* Get the prolog. */
6425 memset(&prolog_key
, 0, sizeof(prolog_key
));
6426 prolog_key
.vs_prolog
.states
= shader
->key
.vs
.prolog
;
6427 prolog_key
.vs_prolog
.num_input_sgprs
= shader
->info
.num_input_sgprs
;
6428 prolog_key
.vs_prolog
.last_input
= MAX2(1, info
->num_inputs
) - 1;
6430 /* The prolog is a no-op if there are no inputs. */
6431 if (info
->num_inputs
) {
6433 si_get_shader_part(sscreen
, &sscreen
->vs_prologs
,
6434 &prolog_key
, tm
, debug
,
6435 si_compile_vs_prolog
);
6436 if (!shader
->prolog
)
6440 /* Get the epilog. */
6441 if (!shader
->key
.vs
.as_es
&& !shader
->key
.vs
.as_ls
&&
6442 !si_get_vs_epilog(sscreen
, tm
, shader
, debug
,
6443 &shader
->key
.vs
.epilog
))
6446 /* Set the instanceID flag. */
6447 for (i
= 0; i
< info
->num_inputs
; i
++)
6448 if (prolog_key
.vs_prolog
.states
.instance_divisors
[i
])
6449 shader
->info
.uses_instanceid
= true;
6455 * Select and compile (or reuse) TES parts (epilog).
6457 static bool si_shader_select_tes_parts(struct si_screen
*sscreen
,
6458 LLVMTargetMachineRef tm
,
6459 struct si_shader
*shader
,
6460 struct pipe_debug_callback
*debug
)
6462 if (shader
->key
.tes
.as_es
)
6465 /* TES compiled as VS. */
6466 return si_get_vs_epilog(sscreen
, tm
, shader
, debug
,
6467 &shader
->key
.tes
.epilog
);
6471 * Compile the TCS epilog. This writes tesselation factors to memory based on
6472 * the output primitive type of the tesselator (determined by TES).
6474 static bool si_compile_tcs_epilog(struct si_screen
*sscreen
,
6475 LLVMTargetMachineRef tm
,
6476 struct pipe_debug_callback
*debug
,
6477 struct si_shader_part
*out
)
6479 union si_shader_part_key
*key
= &out
->key
;
6480 struct si_shader shader
= {};
6481 struct si_shader_context ctx
;
6482 struct gallivm_state
*gallivm
= &ctx
.radeon_bld
.gallivm
;
6483 struct lp_build_tgsi_context
*bld_base
= &ctx
.radeon_bld
.soa
.bld_base
;
6484 LLVMTypeRef params
[16];
6486 int last_array_pointer
, last_sgpr
, num_params
;
6489 si_init_shader_ctx(&ctx
, sscreen
, &shader
, tm
);
6490 ctx
.type
= TGSI_PROCESSOR_TESS_CTRL
;
6491 shader
.key
.tcs
.epilog
= key
->tcs_epilog
.states
;
6493 /* Declare inputs. Only RW_BUFFERS and TESS_FACTOR_OFFSET are used. */
6494 params
[SI_PARAM_RW_BUFFERS
] = const_array(ctx
.v16i8
, SI_NUM_RW_BUFFERS
);
6495 last_array_pointer
= SI_PARAM_RW_BUFFERS
;
6496 params
[SI_PARAM_CONST_BUFFERS
] = ctx
.i64
;
6497 params
[SI_PARAM_SAMPLERS
] = ctx
.i64
;
6498 params
[SI_PARAM_IMAGES
] = ctx
.i64
;
6499 params
[SI_PARAM_SHADER_BUFFERS
] = ctx
.i64
;
6500 params
[SI_PARAM_TCS_OUT_OFFSETS
] = ctx
.i32
;
6501 params
[SI_PARAM_TCS_OUT_LAYOUT
] = ctx
.i32
;
6502 params
[SI_PARAM_TCS_IN_LAYOUT
] = ctx
.i32
;
6503 params
[SI_PARAM_TESS_FACTOR_OFFSET
] = ctx
.i32
;
6504 last_sgpr
= SI_PARAM_TESS_FACTOR_OFFSET
;
6505 num_params
= last_sgpr
+ 1;
6507 params
[num_params
++] = ctx
.i32
; /* patch index within the wave (REL_PATCH_ID) */
6508 params
[num_params
++] = ctx
.i32
; /* invocation ID within the patch */
6509 params
[num_params
++] = ctx
.i32
; /* LDS offset where tess factors should be loaded from */
6511 /* Create the function. */
6512 si_create_function(&ctx
, NULL
, 0, params
, num_params
,
6513 last_array_pointer
, last_sgpr
);
6514 declare_tess_lds(&ctx
);
6515 func
= ctx
.radeon_bld
.main_fn
;
6517 si_write_tess_factors(bld_base
,
6518 LLVMGetParam(func
, last_sgpr
+ 1),
6519 LLVMGetParam(func
, last_sgpr
+ 2),
6520 LLVMGetParam(func
, last_sgpr
+ 3));
6523 LLVMBuildRet(gallivm
->builder
, ctx
.return_value
);
6524 radeon_llvm_finalize_module(&ctx
.radeon_bld
);
6526 if (si_compile_llvm(sscreen
, &out
->binary
, &out
->config
, tm
,
6527 gallivm
->module
, debug
, ctx
.type
,
6528 "Tessellation Control Shader Epilog"))
6531 radeon_llvm_dispose(&ctx
.radeon_bld
);
6536 * Select and compile (or reuse) TCS parts (epilog).
6538 static bool si_shader_select_tcs_parts(struct si_screen
*sscreen
,
6539 LLVMTargetMachineRef tm
,
6540 struct si_shader
*shader
,
6541 struct pipe_debug_callback
*debug
)
6543 union si_shader_part_key epilog_key
;
6545 /* Get the epilog. */
6546 memset(&epilog_key
, 0, sizeof(epilog_key
));
6547 epilog_key
.tcs_epilog
.states
= shader
->key
.tcs
.epilog
;
6549 shader
->epilog
= si_get_shader_part(sscreen
, &sscreen
->tcs_epilogs
,
6550 &epilog_key
, tm
, debug
,
6551 si_compile_tcs_epilog
);
6552 return shader
->epilog
!= NULL
;
6556 * Compile the pixel shader prolog. This handles:
6557 * - two-side color selection and interpolation
6558 * - overriding interpolation parameters for the API PS
6559 * - polygon stippling
6561 * All preloaded SGPRs and VGPRs are passed through unmodified unless they are
6562 * overriden by other states. (e.g. per-sample interpolation)
6563 * Interpolated colors are stored after the preloaded VGPRs.
6565 static bool si_compile_ps_prolog(struct si_screen
*sscreen
,
6566 LLVMTargetMachineRef tm
,
6567 struct pipe_debug_callback
*debug
,
6568 struct si_shader_part
*out
)
6570 union si_shader_part_key
*key
= &out
->key
;
6571 struct si_shader shader
= {};
6572 struct si_shader_context ctx
;
6573 struct gallivm_state
*gallivm
= &ctx
.radeon_bld
.gallivm
;
6574 LLVMTypeRef
*params
;
6575 LLVMValueRef ret
, func
;
6576 int last_sgpr
, num_params
, num_returns
, i
, num_color_channels
;
6579 si_init_shader_ctx(&ctx
, sscreen
, &shader
, tm
);
6580 ctx
.type
= TGSI_PROCESSOR_FRAGMENT
;
6581 shader
.key
.ps
.prolog
= key
->ps_prolog
.states
;
6583 /* Number of inputs + 8 color elements. */
6584 params
= alloca((key
->ps_prolog
.num_input_sgprs
+
6585 key
->ps_prolog
.num_input_vgprs
+ 8) *
6586 sizeof(LLVMTypeRef
));
6588 /* Declare inputs. */
6590 for (i
= 0; i
< key
->ps_prolog
.num_input_sgprs
; i
++)
6591 params
[num_params
++] = ctx
.i32
;
6592 last_sgpr
= num_params
- 1;
6594 for (i
= 0; i
< key
->ps_prolog
.num_input_vgprs
; i
++)
6595 params
[num_params
++] = ctx
.f32
;
6597 /* Declare outputs (same as inputs + add colors if needed) */
6598 num_returns
= num_params
;
6599 num_color_channels
= util_bitcount(key
->ps_prolog
.colors_read
);
6600 for (i
= 0; i
< num_color_channels
; i
++)
6601 params
[num_returns
++] = ctx
.f32
;
6603 /* Create the function. */
6604 si_create_function(&ctx
, params
, num_returns
, params
,
6605 num_params
, -1, last_sgpr
);
6606 func
= ctx
.radeon_bld
.main_fn
;
6608 /* Copy inputs to outputs. This should be no-op, as the registers match,
6609 * but it will prevent the compiler from overwriting them unintentionally.
6611 ret
= ctx
.return_value
;
6612 for (i
= 0; i
< num_params
; i
++) {
6613 LLVMValueRef p
= LLVMGetParam(func
, i
);
6614 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
, p
, i
, "");
6617 /* Polygon stippling. */
6618 if (key
->ps_prolog
.states
.poly_stipple
) {
6619 /* POS_FIXED_PT is always last. */
6620 unsigned pos
= key
->ps_prolog
.num_input_sgprs
+
6621 key
->ps_prolog
.num_input_vgprs
- 1;
6622 LLVMValueRef ptr
[2], views
;
6624 /* Get the pointer to sampler views. */
6625 ptr
[0] = LLVMGetParam(func
, SI_SGPR_SAMPLERS
);
6626 ptr
[1] = LLVMGetParam(func
, SI_SGPR_SAMPLERS
+1);
6627 views
= lp_build_gather_values(gallivm
, ptr
, 2);
6628 views
= LLVMBuildBitCast(gallivm
->builder
, views
, ctx
.i64
, "");
6629 views
= LLVMBuildIntToPtr(gallivm
->builder
, views
,
6630 const_array(ctx
.v8i32
, SI_NUM_SAMPLERS
), "");
6632 si_llvm_emit_polygon_stipple(&ctx
, views
, pos
);
6635 /* Interpolate colors. */
6636 for (i
= 0; i
< 2; i
++) {
6637 unsigned writemask
= (key
->ps_prolog
.colors_read
>> (i
* 4)) & 0xf;
6638 unsigned face_vgpr
= key
->ps_prolog
.num_input_sgprs
+
6639 key
->ps_prolog
.face_vgpr_index
;
6640 LLVMValueRef interp
[2], color
[4];
6641 LLVMValueRef interp_ij
= NULL
, prim_mask
= NULL
, face
= NULL
;
6646 /* If the interpolation qualifier is not CONSTANT (-1). */
6647 if (key
->ps_prolog
.color_interp_vgpr_index
[i
] != -1) {
6648 unsigned interp_vgpr
= key
->ps_prolog
.num_input_sgprs
+
6649 key
->ps_prolog
.color_interp_vgpr_index
[i
];
6651 interp
[0] = LLVMGetParam(func
, interp_vgpr
);
6652 interp
[1] = LLVMGetParam(func
, interp_vgpr
+ 1);
6653 interp_ij
= lp_build_gather_values(gallivm
, interp
, 2);
6654 interp_ij
= LLVMBuildBitCast(gallivm
->builder
, interp_ij
,
6658 /* Use the absolute location of the input. */
6659 prim_mask
= LLVMGetParam(func
, SI_PS_NUM_USER_SGPR
);
6661 if (key
->ps_prolog
.states
.color_two_side
) {
6662 face
= LLVMGetParam(func
, face_vgpr
);
6663 face
= LLVMBuildBitCast(gallivm
->builder
, face
, ctx
.i32
, "");
6666 interp_fs_input(&ctx
,
6667 key
->ps_prolog
.color_attr_index
[i
],
6668 TGSI_SEMANTIC_COLOR
, i
,
6669 key
->ps_prolog
.num_interp_inputs
,
6670 key
->ps_prolog
.colors_read
, interp_ij
,
6671 prim_mask
, face
, color
);
6674 unsigned chan
= u_bit_scan(&writemask
);
6675 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
, color
[chan
],
6680 /* Force per-sample interpolation. */
6681 if (key
->ps_prolog
.states
.force_persample_interp
) {
6682 unsigned i
, base
= key
->ps_prolog
.num_input_sgprs
;
6683 LLVMValueRef persp_sample
[2], linear_sample
[2];
6685 /* Read PERSP_SAMPLE. */
6686 for (i
= 0; i
< 2; i
++)
6687 persp_sample
[i
] = LLVMGetParam(func
, base
+ i
);
6688 /* Overwrite PERSP_CENTER. */
6689 for (i
= 0; i
< 2; i
++)
6690 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
,
6691 persp_sample
[i
], base
+ 2 + i
, "");
6692 /* Overwrite PERSP_CENTROID. */
6693 for (i
= 0; i
< 2; i
++)
6694 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
,
6695 persp_sample
[i
], base
+ 4 + i
, "");
6696 /* Read LINEAR_SAMPLE. */
6697 for (i
= 0; i
< 2; i
++)
6698 linear_sample
[i
] = LLVMGetParam(func
, base
+ 6 + i
);
6699 /* Overwrite LINEAR_CENTER. */
6700 for (i
= 0; i
< 2; i
++)
6701 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
,
6702 linear_sample
[i
], base
+ 8 + i
, "");
6703 /* Overwrite LINEAR_CENTROID. */
6704 for (i
= 0; i
< 2; i
++)
6705 ret
= LLVMBuildInsertValue(gallivm
->builder
, ret
,
6706 linear_sample
[i
], base
+ 10 + i
, "");
6710 LLVMBuildRet(gallivm
->builder
, ret
);
6711 radeon_llvm_finalize_module(&ctx
.radeon_bld
);
6713 if (si_compile_llvm(sscreen
, &out
->binary
, &out
->config
, tm
,
6714 gallivm
->module
, debug
, ctx
.type
,
6715 "Fragment Shader Prolog"))
6718 radeon_llvm_dispose(&ctx
.radeon_bld
);
6723 * Compile the pixel shader epilog. This handles everything that must be
6724 * emulated for pixel shader exports. (alpha-test, format conversions, etc)
6726 static bool si_compile_ps_epilog(struct si_screen
*sscreen
,
6727 LLVMTargetMachineRef tm
,
6728 struct pipe_debug_callback
*debug
,
6729 struct si_shader_part
*out
)
6731 union si_shader_part_key
*key
= &out
->key
;
6732 struct si_shader shader
= {};
6733 struct si_shader_context ctx
;
6734 struct gallivm_state
*gallivm
= &ctx
.radeon_bld
.gallivm
;
6735 struct lp_build_tgsi_context
*bld_base
= &ctx
.radeon_bld
.soa
.bld_base
;
6736 LLVMTypeRef params
[16+8*4+3];
6737 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
6738 int last_array_pointer
, last_sgpr
, num_params
, i
;
6741 si_init_shader_ctx(&ctx
, sscreen
, &shader
, tm
);
6742 ctx
.type
= TGSI_PROCESSOR_FRAGMENT
;
6743 shader
.key
.ps
.epilog
= key
->ps_epilog
.states
;
6745 /* Declare input SGPRs. */
6746 params
[SI_PARAM_RW_BUFFERS
] = ctx
.i64
;
6747 params
[SI_PARAM_CONST_BUFFERS
] = ctx
.i64
;
6748 params
[SI_PARAM_SAMPLERS
] = ctx
.i64
;
6749 params
[SI_PARAM_IMAGES
] = ctx
.i64
;
6750 params
[SI_PARAM_SHADER_BUFFERS
] = ctx
.i64
;
6751 params
[SI_PARAM_ALPHA_REF
] = ctx
.f32
;
6752 last_array_pointer
= -1;
6753 last_sgpr
= SI_PARAM_ALPHA_REF
;
6755 /* Declare input VGPRs. */
6756 num_params
= (last_sgpr
+ 1) +
6757 util_bitcount(key
->ps_epilog
.colors_written
) * 4 +
6758 key
->ps_epilog
.writes_z
+
6759 key
->ps_epilog
.writes_stencil
+
6760 key
->ps_epilog
.writes_samplemask
;
6762 num_params
= MAX2(num_params
,
6763 last_sgpr
+ 1 + PS_EPILOG_SAMPLEMASK_MIN_LOC
+ 1);
6765 assert(num_params
<= ARRAY_SIZE(params
));
6767 for (i
= last_sgpr
+ 1; i
< num_params
; i
++)
6768 params
[i
] = ctx
.f32
;
6770 /* Create the function. */
6771 si_create_function(&ctx
, NULL
, 0, params
, num_params
,
6772 last_array_pointer
, last_sgpr
);
6773 /* Disable elimination of unused inputs. */
6774 radeon_llvm_add_attribute(ctx
.radeon_bld
.main_fn
,
6775 "InitialPSInputAddr", 0xffffff);
6777 /* Process colors. */
6778 unsigned vgpr
= last_sgpr
+ 1;
6779 unsigned colors_written
= key
->ps_epilog
.colors_written
;
6780 int last_color_export
= -1;
6782 /* Find the last color export. */
6783 if (!key
->ps_epilog
.writes_z
&&
6784 !key
->ps_epilog
.writes_stencil
&&
6785 !key
->ps_epilog
.writes_samplemask
) {
6786 unsigned spi_format
= key
->ps_epilog
.states
.spi_shader_col_format
;
6788 /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
6789 if (colors_written
== 0x1 && key
->ps_epilog
.states
.last_cbuf
> 0) {
6790 /* Just set this if any of the colorbuffers are enabled. */
6792 ((1llu << (4 * (key
->ps_epilog
.states
.last_cbuf
+ 1))) - 1))
6793 last_color_export
= 0;
6795 for (i
= 0; i
< 8; i
++)
6796 if (colors_written
& (1 << i
) &&
6797 (spi_format
>> (i
* 4)) & 0xf)
6798 last_color_export
= i
;
6802 while (colors_written
) {
6803 LLVMValueRef color
[4];
6804 int mrt
= u_bit_scan(&colors_written
);
6806 for (i
= 0; i
< 4; i
++)
6807 color
[i
] = LLVMGetParam(ctx
.radeon_bld
.main_fn
, vgpr
++);
6809 si_export_mrt_color(bld_base
, color
, mrt
,
6811 mrt
== last_color_export
);
6814 /* Process depth, stencil, samplemask. */
6815 if (key
->ps_epilog
.writes_z
)
6816 depth
= LLVMGetParam(ctx
.radeon_bld
.main_fn
, vgpr
++);
6817 if (key
->ps_epilog
.writes_stencil
)
6818 stencil
= LLVMGetParam(ctx
.radeon_bld
.main_fn
, vgpr
++);
6819 if (key
->ps_epilog
.writes_samplemask
)
6820 samplemask
= LLVMGetParam(ctx
.radeon_bld
.main_fn
, vgpr
++);
6822 if (depth
|| stencil
|| samplemask
)
6823 si_export_mrt_z(bld_base
, depth
, stencil
, samplemask
);
6824 else if (last_color_export
== -1)
6825 si_export_null(bld_base
);
6828 LLVMBuildRetVoid(gallivm
->builder
);
6829 radeon_llvm_finalize_module(&ctx
.radeon_bld
);
6831 if (si_compile_llvm(sscreen
, &out
->binary
, &out
->config
, tm
,
6832 gallivm
->module
, debug
, ctx
.type
,
6833 "Fragment Shader Epilog"))
6836 radeon_llvm_dispose(&ctx
.radeon_bld
);
6841 * Select and compile (or reuse) pixel shader parts (prolog & epilog).
6843 static bool si_shader_select_ps_parts(struct si_screen
*sscreen
,
6844 LLVMTargetMachineRef tm
,
6845 struct si_shader
*shader
,
6846 struct pipe_debug_callback
*debug
)
6848 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
6849 union si_shader_part_key prolog_key
;
6850 union si_shader_part_key epilog_key
;
6853 /* Get the prolog. */
6854 memset(&prolog_key
, 0, sizeof(prolog_key
));
6855 prolog_key
.ps_prolog
.states
= shader
->key
.ps
.prolog
;
6856 prolog_key
.ps_prolog
.colors_read
= info
->colors_read
;
6857 prolog_key
.ps_prolog
.num_input_sgprs
= shader
->info
.num_input_sgprs
;
6858 prolog_key
.ps_prolog
.num_input_vgprs
= shader
->info
.num_input_vgprs
;
6860 if (info
->colors_read
) {
6861 unsigned *color
= shader
->selector
->color_attr_index
;
6863 if (shader
->key
.ps
.prolog
.color_two_side
) {
6864 /* BCOLORs are stored after the last input. */
6865 prolog_key
.ps_prolog
.num_interp_inputs
= info
->num_inputs
;
6866 prolog_key
.ps_prolog
.face_vgpr_index
= shader
->info
.face_vgpr_index
;
6867 shader
->config
.spi_ps_input_ena
|= S_0286CC_FRONT_FACE_ENA(1);
6870 for (i
= 0; i
< 2; i
++) {
6871 unsigned location
= info
->input_interpolate_loc
[color
[i
]];
6873 if (!(info
->colors_read
& (0xf << i
*4)))
6876 prolog_key
.ps_prolog
.color_attr_index
[i
] = color
[i
];
6878 /* Force per-sample interpolation for the colors here. */
6879 if (shader
->key
.ps
.prolog
.force_persample_interp
)
6880 location
= TGSI_INTERPOLATE_LOC_SAMPLE
;
6882 switch (info
->input_interpolate
[color
[i
]]) {
6883 case TGSI_INTERPOLATE_CONSTANT
:
6884 prolog_key
.ps_prolog
.color_interp_vgpr_index
[i
] = -1;
6886 case TGSI_INTERPOLATE_PERSPECTIVE
:
6887 case TGSI_INTERPOLATE_COLOR
:
6889 case TGSI_INTERPOLATE_LOC_SAMPLE
:
6890 prolog_key
.ps_prolog
.color_interp_vgpr_index
[i
] = 0;
6891 shader
->config
.spi_ps_input_ena
|=
6892 S_0286CC_PERSP_SAMPLE_ENA(1);
6894 case TGSI_INTERPOLATE_LOC_CENTER
:
6895 prolog_key
.ps_prolog
.color_interp_vgpr_index
[i
] = 2;
6896 shader
->config
.spi_ps_input_ena
|=
6897 S_0286CC_PERSP_CENTER_ENA(1);
6899 case TGSI_INTERPOLATE_LOC_CENTROID
:
6900 prolog_key
.ps_prolog
.color_interp_vgpr_index
[i
] = 4;
6901 shader
->config
.spi_ps_input_ena
|=
6902 S_0286CC_PERSP_CENTROID_ENA(1);
6908 case TGSI_INTERPOLATE_LINEAR
:
6910 case TGSI_INTERPOLATE_LOC_SAMPLE
:
6911 prolog_key
.ps_prolog
.color_interp_vgpr_index
[i
] = 6;
6912 shader
->config
.spi_ps_input_ena
|=
6913 S_0286CC_LINEAR_SAMPLE_ENA(1);
6915 case TGSI_INTERPOLATE_LOC_CENTER
:
6916 prolog_key
.ps_prolog
.color_interp_vgpr_index
[i
] = 8;
6917 shader
->config
.spi_ps_input_ena
|=
6918 S_0286CC_LINEAR_CENTER_ENA(1);
6920 case TGSI_INTERPOLATE_LOC_CENTROID
:
6921 prolog_key
.ps_prolog
.color_interp_vgpr_index
[i
] = 10;
6922 shader
->config
.spi_ps_input_ena
|=
6923 S_0286CC_LINEAR_CENTROID_ENA(1);
6935 /* The prolog is a no-op if these aren't set. */
6936 if (prolog_key
.ps_prolog
.colors_read
||
6937 prolog_key
.ps_prolog
.states
.force_persample_interp
||
6938 prolog_key
.ps_prolog
.states
.poly_stipple
) {
6940 si_get_shader_part(sscreen
, &sscreen
->ps_prologs
,
6941 &prolog_key
, tm
, debug
,
6942 si_compile_ps_prolog
);
6943 if (!shader
->prolog
)
6947 /* Get the epilog. */
6948 memset(&epilog_key
, 0, sizeof(epilog_key
));
6949 epilog_key
.ps_epilog
.colors_written
= info
->colors_written
;
6950 epilog_key
.ps_epilog
.writes_z
= info
->writes_z
;
6951 epilog_key
.ps_epilog
.writes_stencil
= info
->writes_stencil
;
6952 epilog_key
.ps_epilog
.writes_samplemask
= info
->writes_samplemask
;
6953 epilog_key
.ps_epilog
.states
= shader
->key
.ps
.epilog
;
6956 si_get_shader_part(sscreen
, &sscreen
->ps_epilogs
,
6957 &epilog_key
, tm
, debug
,
6958 si_compile_ps_epilog
);
6959 if (!shader
->epilog
)
6962 /* Enable POS_FIXED_PT if polygon stippling is enabled. */
6963 if (shader
->key
.ps
.prolog
.poly_stipple
) {
6964 shader
->config
.spi_ps_input_ena
|= S_0286CC_POS_FIXED_PT_ENA(1);
6965 assert(G_0286CC_POS_FIXED_PT_ENA(shader
->config
.spi_ps_input_addr
));
6968 /* Set up the enable bits for per-sample shading if needed. */
6969 if (shader
->key
.ps
.prolog
.force_persample_interp
) {
6970 if (G_0286CC_PERSP_CENTER_ENA(shader
->config
.spi_ps_input_ena
) ||
6971 G_0286CC_PERSP_CENTROID_ENA(shader
->config
.spi_ps_input_ena
)) {
6972 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_CENTER_ENA
;
6973 shader
->config
.spi_ps_input_ena
&= C_0286CC_PERSP_CENTROID_ENA
;
6974 shader
->config
.spi_ps_input_ena
|= S_0286CC_PERSP_SAMPLE_ENA(1);
6976 if (G_0286CC_LINEAR_CENTER_ENA(shader
->config
.spi_ps_input_ena
) ||
6977 G_0286CC_LINEAR_CENTROID_ENA(shader
->config
.spi_ps_input_ena
)) {
6978 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_CENTER_ENA
;
6979 shader
->config
.spi_ps_input_ena
&= C_0286CC_LINEAR_CENTROID_ENA
;
6980 shader
->config
.spi_ps_input_ena
|= S_0286CC_LINEAR_SAMPLE_ENA(1);
6984 /* POW_W_FLOAT requires that one of the perspective weights is enabled. */
6985 if (G_0286CC_POS_W_FLOAT_ENA(shader
->config
.spi_ps_input_ena
) &&
6986 !(shader
->config
.spi_ps_input_ena
& 0xf)) {
6987 shader
->config
.spi_ps_input_ena
|= S_0286CC_PERSP_CENTER_ENA(1);
6988 assert(G_0286CC_PERSP_CENTER_ENA(shader
->config
.spi_ps_input_addr
));
6991 /* At least one pair of interpolation weights must be enabled. */
6992 if (!(shader
->config
.spi_ps_input_ena
& 0x7f)) {
6993 shader
->config
.spi_ps_input_ena
|= S_0286CC_LINEAR_CENTER_ENA(1);
6994 assert(G_0286CC_LINEAR_CENTER_ENA(shader
->config
.spi_ps_input_addr
));
6997 /* The sample mask input is always enabled, because the API shader always
6998 * passes it through to the epilog. Disable it here if it's unused.
7000 if (!shader
->key
.ps
.epilog
.poly_line_smoothing
&&
7001 !shader
->selector
->info
.reads_samplemask
)
7002 shader
->config
.spi_ps_input_ena
&= C_0286CC_SAMPLE_COVERAGE_ENA
;
7007 static void si_fix_num_sgprs(struct si_shader
*shader
)
7009 unsigned min_sgprs
= shader
->info
.num_input_sgprs
+ 2; /* VCC */
7011 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
, min_sgprs
);
7014 int si_shader_create(struct si_screen
*sscreen
, LLVMTargetMachineRef tm
,
7015 struct si_shader
*shader
,
7016 struct pipe_debug_callback
*debug
)
7018 struct si_shader
*mainp
= shader
->selector
->main_shader_part
;
7021 /* LS, ES, VS are compiled on demand if the main part hasn't been
7022 * compiled for that stage.
7025 (shader
->selector
->type
== PIPE_SHADER_VERTEX
&&
7026 (shader
->key
.vs
.as_es
!= mainp
->key
.vs
.as_es
||
7027 shader
->key
.vs
.as_ls
!= mainp
->key
.vs
.as_ls
)) ||
7028 (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
&&
7029 shader
->key
.tes
.as_es
!= mainp
->key
.tes
.as_es
) ||
7030 shader
->selector
->type
== PIPE_SHADER_COMPUTE
) {
7031 /* Monolithic shader (compiled as a whole, has many variants,
7032 * may take a long time to compile).
7034 r
= si_compile_tgsi_shader(sscreen
, tm
, shader
, true, debug
);
7038 /* The shader consists of 2-3 parts:
7040 * - the middle part is the user shader, it has 1 variant only
7041 * and it was compiled during the creation of the shader
7043 * - the prolog part is inserted at the beginning
7044 * - the epilog part is inserted at the end
7046 * The prolog and epilog have many (but simple) variants.
7049 /* Copy the compiled TGSI shader data over. */
7050 shader
->is_binary_shared
= true;
7051 shader
->binary
= mainp
->binary
;
7052 shader
->config
= mainp
->config
;
7053 shader
->info
.num_input_sgprs
= mainp
->info
.num_input_sgprs
;
7054 shader
->info
.num_input_vgprs
= mainp
->info
.num_input_vgprs
;
7055 shader
->info
.face_vgpr_index
= mainp
->info
.face_vgpr_index
;
7056 memcpy(shader
->info
.vs_output_param_offset
,
7057 mainp
->info
.vs_output_param_offset
,
7058 sizeof(mainp
->info
.vs_output_param_offset
));
7059 shader
->info
.uses_instanceid
= mainp
->info
.uses_instanceid
;
7060 shader
->info
.nr_pos_exports
= mainp
->info
.nr_pos_exports
;
7061 shader
->info
.nr_param_exports
= mainp
->info
.nr_param_exports
;
7063 /* Select prologs and/or epilogs. */
7064 switch (shader
->selector
->type
) {
7065 case PIPE_SHADER_VERTEX
:
7066 if (!si_shader_select_vs_parts(sscreen
, tm
, shader
, debug
))
7069 case PIPE_SHADER_TESS_CTRL
:
7070 if (!si_shader_select_tcs_parts(sscreen
, tm
, shader
, debug
))
7073 case PIPE_SHADER_TESS_EVAL
:
7074 if (!si_shader_select_tes_parts(sscreen
, tm
, shader
, debug
))
7077 case PIPE_SHADER_FRAGMENT
:
7078 if (!si_shader_select_ps_parts(sscreen
, tm
, shader
, debug
))
7081 /* Make sure we have at least as many VGPRs as there
7082 * are allocated inputs.
7084 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
7085 shader
->info
.num_input_vgprs
);
7089 /* Update SGPR and VGPR counts. */
7090 if (shader
->prolog
) {
7091 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
,
7092 shader
->prolog
->config
.num_sgprs
);
7093 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
7094 shader
->prolog
->config
.num_vgprs
);
7096 if (shader
->epilog
) {
7097 shader
->config
.num_sgprs
= MAX2(shader
->config
.num_sgprs
,
7098 shader
->epilog
->config
.num_sgprs
);
7099 shader
->config
.num_vgprs
= MAX2(shader
->config
.num_vgprs
,
7100 shader
->epilog
->config
.num_vgprs
);
7104 si_fix_num_sgprs(shader
);
7105 si_shader_dump(sscreen
, shader
, debug
, shader
->selector
->info
.processor
,
7109 r
= si_shader_binary_upload(sscreen
, shader
);
7111 fprintf(stderr
, "LLVM failed to upload shader\n");
7118 void si_shader_destroy(struct si_shader
*shader
)
7120 if (shader
->gs_copy_shader
) {
7121 si_shader_destroy(shader
->gs_copy_shader
);
7122 FREE(shader
->gs_copy_shader
);
7125 if (shader
->scratch_bo
)
7126 r600_resource_reference(&shader
->scratch_bo
, NULL
);
7128 r600_resource_reference(&shader
->bo
, NULL
);
7130 if (!shader
->is_binary_shared
)
7131 radeon_shader_binary_clean(&shader
->binary
);