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 "tgsi/tgsi_parse.h"
43 #include "tgsi/tgsi_util.h"
44 #include "tgsi/tgsi_dump.h"
47 #include "si_shader.h"
52 static const char *scratch_rsrc_dword0_symbol
=
53 "SCRATCH_RSRC_DWORD0";
55 static const char *scratch_rsrc_dword1_symbol
=
56 "SCRATCH_RSRC_DWORD1";
58 struct si_shader_output_values
60 LLVMValueRef values
[4];
65 struct si_shader_context
67 struct radeon_llvm_context radeon_bld
;
68 struct si_shader
*shader
;
69 struct si_screen
*screen
;
70 unsigned type
; /* TGSI_PROCESSOR_* specifies the type of shader. */
71 int param_streamout_config
;
72 int param_streamout_write_index
;
73 int param_streamout_offset
[4];
75 int param_rel_auto_id
;
77 int param_instance_id
;
80 int param_tes_rel_patch_id
;
81 int param_tes_patch_id
;
82 int param_es2gs_offset
;
83 LLVMTargetMachineRef tm
;
84 LLVMValueRef const_md
;
85 LLVMValueRef const_resource
[SI_NUM_CONST_BUFFERS
];
87 LLVMValueRef
*constants
[SI_NUM_CONST_BUFFERS
];
88 LLVMValueRef resources
[SI_NUM_SAMPLER_VIEWS
];
89 LLVMValueRef samplers
[SI_NUM_SAMPLER_STATES
];
90 LLVMValueRef so_buffers
[4];
91 LLVMValueRef esgs_ring
;
92 LLVMValueRef gsvs_ring
[4];
93 LLVMValueRef gs_next_vertex
[4];
96 static struct si_shader_context
* si_shader_context(
97 struct lp_build_tgsi_context
* bld_base
)
99 return (struct si_shader_context
*)bld_base
;
103 #define PERSPECTIVE_BASE 0
104 #define LINEAR_BASE 9
106 #define SAMPLE_OFFSET 0
107 #define CENTER_OFFSET 2
108 #define CENTROID_OFSET 4
110 #define USE_SGPR_MAX_SUFFIX_LEN 5
111 #define CONST_ADDR_SPACE 2
112 #define LOCAL_ADDR_SPACE 3
113 #define USER_SGPR_ADDR_SPACE 8
117 #define SENDMSG_GS_DONE 3
119 #define SENDMSG_GS_OP_NOP (0 << 4)
120 #define SENDMSG_GS_OP_CUT (1 << 4)
121 #define SENDMSG_GS_OP_EMIT (2 << 4)
122 #define SENDMSG_GS_OP_EMIT_CUT (3 << 4)
125 * Returns a unique index for a semantic name and index. The index must be
126 * less than 64, so that a 64-bit bitmask of used inputs or outputs can be
129 unsigned si_shader_io_get_unique_index(unsigned semantic_name
, unsigned index
)
131 switch (semantic_name
) {
132 case TGSI_SEMANTIC_POSITION
:
134 case TGSI_SEMANTIC_PSIZE
:
136 case TGSI_SEMANTIC_CLIPDIST
:
139 case TGSI_SEMANTIC_GENERIC
:
143 /* same explanation as in the default statement,
144 * the only user hitting this is st/nine.
148 /* patch indices are completely separate and thus start from 0 */
149 case TGSI_SEMANTIC_TESSOUTER
:
151 case TGSI_SEMANTIC_TESSINNER
:
153 case TGSI_SEMANTIC_PATCH
:
157 /* Don't fail here. The result of this function is only used
158 * for LS, TCS, TES, and GS, where legacy GL semantics can't
159 * occur, but this function is called for all vertex shaders
160 * before it's known whether LS will be compiled or not.
167 * Given a semantic name and index of a parameter and a mask of used parameters
168 * (inputs or outputs), return the index of the parameter in the list of all
171 * For example, assume this list of parameters:
172 * POSITION, PSIZE, GENERIC0, GENERIC2
173 * which has the mask:
176 * querying POSITION returns 0,
177 * querying PSIZE returns 1,
178 * querying GENERIC0 returns 2,
179 * querying GENERIC2 returns 3.
181 * Which can be used as an offset to a parameter buffer in units of vec4s.
183 static int get_param_index(unsigned semantic_name
, unsigned index
,
186 unsigned unique_index
= si_shader_io_get_unique_index(semantic_name
, index
);
187 int i
, param_index
= 0;
189 /* If not present... */
190 if (!((1llu << unique_index
) & mask
))
193 for (i
= 0; mask
; i
++) {
194 uint64_t bit
= 1llu << i
;
197 if (i
== unique_index
)
205 assert(!"unreachable");
210 * Get the value of a shader input parameter and extract a bitfield.
212 static LLVMValueRef
unpack_param(struct si_shader_context
*si_shader_ctx
,
213 unsigned param
, unsigned rshift
,
216 struct gallivm_state
*gallivm
= &si_shader_ctx
->radeon_bld
.gallivm
;
217 LLVMValueRef value
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
,
221 value
= LLVMBuildLShr(gallivm
->builder
, value
,
222 lp_build_const_int32(gallivm
, rshift
), "");
224 if (rshift
+ bitwidth
< 32) {
225 unsigned mask
= (1 << bitwidth
) - 1;
226 value
= LLVMBuildAnd(gallivm
->builder
, value
,
227 lp_build_const_int32(gallivm
, mask
), "");
233 static LLVMValueRef
get_rel_patch_id(struct si_shader_context
*si_shader_ctx
)
235 switch (si_shader_ctx
->type
) {
236 case TGSI_PROCESSOR_TESS_CTRL
:
237 return unpack_param(si_shader_ctx
, SI_PARAM_REL_IDS
, 0, 8);
239 case TGSI_PROCESSOR_TESS_EVAL
:
240 return LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
,
241 si_shader_ctx
->param_tes_rel_patch_id
);
249 /* Tessellation shaders pass outputs to the next shader using LDS.
251 * LS outputs = TCS inputs
252 * TCS outputs = TES inputs
255 * - TCS inputs for patch 0
256 * - TCS inputs for patch 1
257 * - TCS inputs for patch 2 = get_tcs_in_current_patch_offset (if RelPatchID==2)
259 * - TCS outputs for patch 0 = get_tcs_out_patch0_offset
260 * - Per-patch TCS outputs for patch 0 = get_tcs_out_patch0_patch_data_offset
261 * - TCS outputs for patch 1
262 * - Per-patch TCS outputs for patch 1
263 * - TCS outputs for patch 2 = get_tcs_out_current_patch_offset (if RelPatchID==2)
264 * - Per-patch TCS outputs for patch 2 = get_tcs_out_current_patch_data_offset (if RelPatchID==2)
267 * All three shaders VS(LS), TCS, TES share the same LDS space.
271 get_tcs_in_patch_stride(struct si_shader_context
*si_shader_ctx
)
273 if (si_shader_ctx
->type
== TGSI_PROCESSOR_VERTEX
)
274 return unpack_param(si_shader_ctx
, SI_PARAM_LS_OUT_LAYOUT
, 0, 13);
275 else if (si_shader_ctx
->type
== TGSI_PROCESSOR_TESS_CTRL
)
276 return unpack_param(si_shader_ctx
, SI_PARAM_TCS_IN_LAYOUT
, 0, 13);
284 get_tcs_out_patch_stride(struct si_shader_context
*si_shader_ctx
)
286 return unpack_param(si_shader_ctx
, SI_PARAM_TCS_OUT_LAYOUT
, 0, 13);
290 get_tcs_out_patch0_offset(struct si_shader_context
*si_shader_ctx
)
292 return lp_build_mul_imm(&si_shader_ctx
->radeon_bld
.soa
.bld_base
.uint_bld
,
293 unpack_param(si_shader_ctx
,
294 SI_PARAM_TCS_OUT_OFFSETS
,
300 get_tcs_out_patch0_patch_data_offset(struct si_shader_context
*si_shader_ctx
)
302 return lp_build_mul_imm(&si_shader_ctx
->radeon_bld
.soa
.bld_base
.uint_bld
,
303 unpack_param(si_shader_ctx
,
304 SI_PARAM_TCS_OUT_OFFSETS
,
310 get_tcs_in_current_patch_offset(struct si_shader_context
*si_shader_ctx
)
312 struct gallivm_state
*gallivm
= &si_shader_ctx
->radeon_bld
.gallivm
;
313 LLVMValueRef patch_stride
= get_tcs_in_patch_stride(si_shader_ctx
);
314 LLVMValueRef rel_patch_id
= get_rel_patch_id(si_shader_ctx
);
316 return LLVMBuildMul(gallivm
->builder
, patch_stride
, rel_patch_id
, "");
320 get_tcs_out_current_patch_offset(struct si_shader_context
*si_shader_ctx
)
322 struct gallivm_state
*gallivm
= &si_shader_ctx
->radeon_bld
.gallivm
;
323 LLVMValueRef patch0_offset
= get_tcs_out_patch0_offset(si_shader_ctx
);
324 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(si_shader_ctx
);
325 LLVMValueRef rel_patch_id
= get_rel_patch_id(si_shader_ctx
);
327 return LLVMBuildAdd(gallivm
->builder
, patch0_offset
,
328 LLVMBuildMul(gallivm
->builder
, patch_stride
,
334 get_tcs_out_current_patch_data_offset(struct si_shader_context
*si_shader_ctx
)
336 struct gallivm_state
*gallivm
= &si_shader_ctx
->radeon_bld
.gallivm
;
337 LLVMValueRef patch0_patch_data_offset
=
338 get_tcs_out_patch0_patch_data_offset(si_shader_ctx
);
339 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(si_shader_ctx
);
340 LLVMValueRef rel_patch_id
= get_rel_patch_id(si_shader_ctx
);
342 return LLVMBuildAdd(gallivm
->builder
, patch0_patch_data_offset
,
343 LLVMBuildMul(gallivm
->builder
, patch_stride
,
348 static void build_indexed_store(struct si_shader_context
*si_shader_ctx
,
349 LLVMValueRef base_ptr
, LLVMValueRef index
,
352 struct lp_build_tgsi_context
*bld_base
= &si_shader_ctx
->radeon_bld
.soa
.bld_base
;
353 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
354 LLVMValueRef indices
[2], pointer
;
356 indices
[0] = bld_base
->uint_bld
.zero
;
359 pointer
= LLVMBuildGEP(gallivm
->builder
, base_ptr
, indices
, 2, "");
360 LLVMBuildStore(gallivm
->builder
, value
, pointer
);
364 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
365 * It's equivalent to doing a load from &base_ptr[index].
367 * \param base_ptr Where the array starts.
368 * \param index The element index into the array.
370 static LLVMValueRef
build_indexed_load(struct si_shader_context
*si_shader_ctx
,
371 LLVMValueRef base_ptr
, LLVMValueRef index
)
373 struct lp_build_tgsi_context
*bld_base
= &si_shader_ctx
->radeon_bld
.soa
.bld_base
;
374 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
375 LLVMValueRef indices
[2], pointer
;
377 indices
[0] = bld_base
->uint_bld
.zero
;
380 pointer
= LLVMBuildGEP(gallivm
->builder
, base_ptr
, indices
, 2, "");
381 return LLVMBuildLoad(gallivm
->builder
, pointer
, "");
385 * Do a load from &base_ptr[index], but also add a flag that it's loading
388 static LLVMValueRef
build_indexed_load_const(
389 struct si_shader_context
* si_shader_ctx
,
390 LLVMValueRef base_ptr
, LLVMValueRef index
)
392 LLVMValueRef result
= build_indexed_load(si_shader_ctx
, base_ptr
, index
);
393 LLVMSetMetadata(result
, 1, si_shader_ctx
->const_md
);
397 static LLVMValueRef
get_instance_index_for_fetch(
398 struct radeon_llvm_context
* radeon_bld
,
401 struct si_shader_context
*si_shader_ctx
=
402 si_shader_context(&radeon_bld
->soa
.bld_base
);
403 struct gallivm_state
* gallivm
= radeon_bld
->soa
.bld_base
.base
.gallivm
;
405 LLVMValueRef result
= LLVMGetParam(radeon_bld
->main_fn
,
406 si_shader_ctx
->param_instance_id
);
408 /* The division must be done before START_INSTANCE is added. */
410 result
= LLVMBuildUDiv(gallivm
->builder
, result
,
411 lp_build_const_int32(gallivm
, divisor
), "");
413 return LLVMBuildAdd(gallivm
->builder
, result
, LLVMGetParam(
414 radeon_bld
->main_fn
, SI_PARAM_START_INSTANCE
), "");
417 static void declare_input_vs(
418 struct radeon_llvm_context
*radeon_bld
,
419 unsigned input_index
,
420 const struct tgsi_full_declaration
*decl
)
422 struct lp_build_context
*base
= &radeon_bld
->soa
.bld_base
.base
;
423 struct gallivm_state
*gallivm
= base
->gallivm
;
424 struct si_shader_context
*si_shader_ctx
=
425 si_shader_context(&radeon_bld
->soa
.bld_base
);
426 unsigned divisor
= si_shader_ctx
->shader
->key
.vs
.instance_divisors
[input_index
];
430 LLVMValueRef t_list_ptr
;
431 LLVMValueRef t_offset
;
433 LLVMValueRef attribute_offset
;
434 LLVMValueRef buffer_index
;
435 LLVMValueRef args
[3];
436 LLVMTypeRef vec4_type
;
439 /* Load the T list */
440 t_list_ptr
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, SI_PARAM_VERTEX_BUFFER
);
442 t_offset
= lp_build_const_int32(gallivm
, input_index
);
444 t_list
= build_indexed_load_const(si_shader_ctx
, t_list_ptr
, t_offset
);
446 /* Build the attribute offset */
447 attribute_offset
= lp_build_const_int32(gallivm
, 0);
450 /* Build index from instance ID, start instance and divisor */
451 si_shader_ctx
->shader
->uses_instanceid
= true;
452 buffer_index
= get_instance_index_for_fetch(&si_shader_ctx
->radeon_bld
, divisor
);
454 /* Load the buffer index for vertices. */
455 LLVMValueRef vertex_id
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
,
456 si_shader_ctx
->param_vertex_id
);
457 LLVMValueRef base_vertex
= LLVMGetParam(radeon_bld
->main_fn
,
458 SI_PARAM_BASE_VERTEX
);
459 buffer_index
= LLVMBuildAdd(gallivm
->builder
, base_vertex
, vertex_id
, "");
462 vec4_type
= LLVMVectorType(base
->elem_type
, 4);
464 args
[1] = attribute_offset
;
465 args
[2] = buffer_index
;
466 input
= lp_build_intrinsic(gallivm
->builder
,
467 "llvm.SI.vs.load.input", vec4_type
, args
, 3,
468 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
470 /* Break up the vec4 into individual components */
471 for (chan
= 0; chan
< 4; chan
++) {
472 LLVMValueRef llvm_chan
= lp_build_const_int32(gallivm
, chan
);
473 /* XXX: Use a helper function for this. There is one in
475 si_shader_ctx
->radeon_bld
.inputs
[radeon_llvm_reg_index_soa(input_index
, chan
)] =
476 LLVMBuildExtractElement(gallivm
->builder
,
477 input
, llvm_chan
, "");
481 static LLVMValueRef
get_primitive_id(struct lp_build_tgsi_context
*bld_base
,
484 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
487 return bld_base
->uint_bld
.zero
;
489 switch (si_shader_ctx
->type
) {
490 case TGSI_PROCESSOR_VERTEX
:
491 return LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
,
492 si_shader_ctx
->param_vs_prim_id
);
493 case TGSI_PROCESSOR_TESS_CTRL
:
494 return LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
,
496 case TGSI_PROCESSOR_TESS_EVAL
:
497 return LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
,
498 si_shader_ctx
->param_tes_patch_id
);
499 case TGSI_PROCESSOR_GEOMETRY
:
500 return LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
,
501 SI_PARAM_PRIMITIVE_ID
);
504 return bld_base
->uint_bld
.zero
;
509 * Return the value of tgsi_ind_register for indexing.
510 * This is the indirect index with the constant offset added to it.
512 static LLVMValueRef
get_indirect_index(struct si_shader_context
*si_shader_ctx
,
513 const struct tgsi_ind_register
*ind
,
516 struct gallivm_state
*gallivm
= si_shader_ctx
->radeon_bld
.soa
.bld_base
.base
.gallivm
;
519 result
= si_shader_ctx
->radeon_bld
.soa
.addr
[ind
->Index
][ind
->Swizzle
];
520 result
= LLVMBuildLoad(gallivm
->builder
, result
, "");
521 result
= LLVMBuildAdd(gallivm
->builder
, result
,
522 lp_build_const_int32(gallivm
, rel_index
), "");
527 * Calculate a dword address given an input or output register and a stride.
529 static LLVMValueRef
get_dw_address(struct si_shader_context
*si_shader_ctx
,
530 const struct tgsi_full_dst_register
*dst
,
531 const struct tgsi_full_src_register
*src
,
532 LLVMValueRef vertex_dw_stride
,
533 LLVMValueRef base_addr
)
535 struct gallivm_state
*gallivm
= si_shader_ctx
->radeon_bld
.soa
.bld_base
.base
.gallivm
;
536 struct tgsi_shader_info
*info
= &si_shader_ctx
->shader
->selector
->info
;
537 ubyte
*name
, *index
, *array_first
;
539 struct tgsi_full_dst_register reg
;
541 /* Set the register description. The address computation is the same
542 * for sources and destinations. */
544 reg
.Register
.File
= src
->Register
.File
;
545 reg
.Register
.Index
= src
->Register
.Index
;
546 reg
.Register
.Indirect
= src
->Register
.Indirect
;
547 reg
.Register
.Dimension
= src
->Register
.Dimension
;
548 reg
.Indirect
= src
->Indirect
;
549 reg
.Dimension
= src
->Dimension
;
550 reg
.DimIndirect
= src
->DimIndirect
;
554 /* If the register is 2-dimensional (e.g. an array of vertices
555 * in a primitive), calculate the base address of the vertex. */
556 if (reg
.Register
.Dimension
) {
559 if (reg
.Dimension
.Indirect
)
560 index
= get_indirect_index(si_shader_ctx
, ®
.DimIndirect
,
561 reg
.Dimension
.Index
);
563 index
= lp_build_const_int32(gallivm
, reg
.Dimension
.Index
);
565 base_addr
= LLVMBuildAdd(gallivm
->builder
, base_addr
,
566 LLVMBuildMul(gallivm
->builder
, index
,
567 vertex_dw_stride
, ""), "");
570 /* Get information about the register. */
571 if (reg
.Register
.File
== TGSI_FILE_INPUT
) {
572 name
= info
->input_semantic_name
;
573 index
= info
->input_semantic_index
;
574 array_first
= info
->input_array_first
;
575 } else if (reg
.Register
.File
== TGSI_FILE_OUTPUT
) {
576 name
= info
->output_semantic_name
;
577 index
= info
->output_semantic_index
;
578 array_first
= info
->output_array_first
;
584 if (reg
.Register
.Indirect
) {
585 /* Add the relative address of the element. */
586 LLVMValueRef ind_index
;
588 if (reg
.Indirect
.ArrayID
)
589 first
= array_first
[reg
.Indirect
.ArrayID
];
591 first
= reg
.Register
.Index
;
593 ind_index
= get_indirect_index(si_shader_ctx
, ®
.Indirect
,
594 reg
.Register
.Index
- first
);
596 base_addr
= LLVMBuildAdd(gallivm
->builder
, base_addr
,
597 LLVMBuildMul(gallivm
->builder
, ind_index
,
598 lp_build_const_int32(gallivm
, 4), ""), "");
600 param
= si_shader_io_get_unique_index(name
[first
], index
[first
]);
602 param
= si_shader_io_get_unique_index(name
[reg
.Register
.Index
],
603 index
[reg
.Register
.Index
]);
606 /* Add the base address of the element. */
607 return LLVMBuildAdd(gallivm
->builder
, base_addr
,
608 lp_build_const_int32(gallivm
, param
* 4), "");
614 * \param type output value type
615 * \param swizzle offset (typically 0..3); it can be ~0, which loads a vec4
616 * \param dw_addr address in dwords
618 static LLVMValueRef
lds_load(struct lp_build_tgsi_context
*bld_base
,
619 enum tgsi_opcode_type type
, unsigned swizzle
,
620 LLVMValueRef dw_addr
)
622 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
623 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
627 LLVMValueRef values
[TGSI_NUM_CHANNELS
];
629 for (unsigned chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++)
630 values
[chan
] = lds_load(bld_base
, type
, chan
, dw_addr
);
632 return lp_build_gather_values(bld_base
->base
.gallivm
, values
,
636 dw_addr
= lp_build_add(&bld_base
->uint_bld
, dw_addr
,
637 lp_build_const_int32(gallivm
, swizzle
));
639 value
= build_indexed_load(si_shader_ctx
, si_shader_ctx
->lds
, dw_addr
);
640 return LLVMBuildBitCast(gallivm
->builder
, value
,
641 tgsi2llvmtype(bld_base
, type
), "");
647 * \param swizzle offset (typically 0..3)
648 * \param dw_addr address in dwords
649 * \param value value to store
651 static void lds_store(struct lp_build_tgsi_context
* bld_base
,
652 unsigned swizzle
, LLVMValueRef dw_addr
,
655 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
656 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
658 dw_addr
= lp_build_add(&bld_base
->uint_bld
, dw_addr
,
659 lp_build_const_int32(gallivm
, swizzle
));
661 value
= LLVMBuildBitCast(gallivm
->builder
, value
,
662 LLVMInt32TypeInContext(gallivm
->context
), "");
663 build_indexed_store(si_shader_ctx
, si_shader_ctx
->lds
,
667 static LLVMValueRef
fetch_input_tcs(
668 struct lp_build_tgsi_context
*bld_base
,
669 const struct tgsi_full_src_register
*reg
,
670 enum tgsi_opcode_type type
, unsigned swizzle
)
672 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
673 LLVMValueRef dw_addr
, stride
;
675 stride
= unpack_param(si_shader_ctx
, SI_PARAM_TCS_IN_LAYOUT
, 13, 8);
676 dw_addr
= get_tcs_in_current_patch_offset(si_shader_ctx
);
677 dw_addr
= get_dw_address(si_shader_ctx
, NULL
, reg
, stride
, dw_addr
);
679 return lds_load(bld_base
, type
, swizzle
, dw_addr
);
682 static LLVMValueRef
fetch_output_tcs(
683 struct lp_build_tgsi_context
*bld_base
,
684 const struct tgsi_full_src_register
*reg
,
685 enum tgsi_opcode_type type
, unsigned swizzle
)
687 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
688 LLVMValueRef dw_addr
, stride
;
690 if (reg
->Register
.Dimension
) {
691 stride
= unpack_param(si_shader_ctx
, SI_PARAM_TCS_OUT_LAYOUT
, 13, 8);
692 dw_addr
= get_tcs_out_current_patch_offset(si_shader_ctx
);
693 dw_addr
= get_dw_address(si_shader_ctx
, NULL
, reg
, stride
, dw_addr
);
695 dw_addr
= get_tcs_out_current_patch_data_offset(si_shader_ctx
);
696 dw_addr
= get_dw_address(si_shader_ctx
, NULL
, reg
, NULL
, dw_addr
);
699 return lds_load(bld_base
, type
, swizzle
, dw_addr
);
702 static LLVMValueRef
fetch_input_tes(
703 struct lp_build_tgsi_context
*bld_base
,
704 const struct tgsi_full_src_register
*reg
,
705 enum tgsi_opcode_type type
, unsigned swizzle
)
707 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
708 LLVMValueRef dw_addr
, stride
;
710 if (reg
->Register
.Dimension
) {
711 stride
= unpack_param(si_shader_ctx
, SI_PARAM_TCS_OUT_LAYOUT
, 13, 8);
712 dw_addr
= get_tcs_out_current_patch_offset(si_shader_ctx
);
713 dw_addr
= get_dw_address(si_shader_ctx
, NULL
, reg
, stride
, dw_addr
);
715 dw_addr
= get_tcs_out_current_patch_data_offset(si_shader_ctx
);
716 dw_addr
= get_dw_address(si_shader_ctx
, NULL
, reg
, NULL
, dw_addr
);
719 return lds_load(bld_base
, type
, swizzle
, dw_addr
);
722 static void store_output_tcs(struct lp_build_tgsi_context
* bld_base
,
723 const struct tgsi_full_instruction
* inst
,
724 const struct tgsi_opcode_info
* info
,
727 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
728 const struct tgsi_full_dst_register
*reg
= &inst
->Dst
[0];
730 LLVMValueRef dw_addr
, stride
;
732 /* Only handle per-patch and per-vertex outputs here.
733 * Vectors will be lowered to scalars and this function will be called again.
735 if (reg
->Register
.File
!= TGSI_FILE_OUTPUT
||
736 (dst
[0] && LLVMGetTypeKind(LLVMTypeOf(dst
[0])) == LLVMVectorTypeKind
)) {
737 radeon_llvm_emit_store(bld_base
, inst
, info
, dst
);
741 if (reg
->Register
.Dimension
) {
742 stride
= unpack_param(si_shader_ctx
, SI_PARAM_TCS_OUT_LAYOUT
, 13, 8);
743 dw_addr
= get_tcs_out_current_patch_offset(si_shader_ctx
);
744 dw_addr
= get_dw_address(si_shader_ctx
, reg
, NULL
, stride
, dw_addr
);
746 dw_addr
= get_tcs_out_current_patch_data_offset(si_shader_ctx
);
747 dw_addr
= get_dw_address(si_shader_ctx
, reg
, NULL
, NULL
, dw_addr
);
750 TGSI_FOR_EACH_DST0_ENABLED_CHANNEL(inst
, chan_index
) {
751 LLVMValueRef value
= dst
[chan_index
];
753 if (inst
->Instruction
.Saturate
)
754 value
= radeon_llvm_saturate(bld_base
, value
);
756 lds_store(bld_base
, chan_index
, dw_addr
, value
);
760 static LLVMValueRef
fetch_input_gs(
761 struct lp_build_tgsi_context
*bld_base
,
762 const struct tgsi_full_src_register
*reg
,
763 enum tgsi_opcode_type type
,
766 struct lp_build_context
*base
= &bld_base
->base
;
767 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
768 struct si_shader
*shader
= si_shader_ctx
->shader
;
769 struct lp_build_context
*uint
= &si_shader_ctx
->radeon_bld
.soa
.bld_base
.uint_bld
;
770 struct gallivm_state
*gallivm
= base
->gallivm
;
771 LLVMTypeRef i32
= LLVMInt32TypeInContext(gallivm
->context
);
772 LLVMValueRef vtx_offset
;
773 LLVMValueRef args
[9];
774 unsigned vtx_offset_param
;
775 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
776 unsigned semantic_name
= info
->input_semantic_name
[reg
->Register
.Index
];
777 unsigned semantic_index
= info
->input_semantic_index
[reg
->Register
.Index
];
779 if (swizzle
!= ~0 && semantic_name
== TGSI_SEMANTIC_PRIMID
)
780 return get_primitive_id(bld_base
, swizzle
);
782 if (!reg
->Register
.Dimension
)
786 LLVMValueRef values
[TGSI_NUM_CHANNELS
];
788 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
789 values
[chan
] = fetch_input_gs(bld_base
, reg
, type
, chan
);
791 return lp_build_gather_values(bld_base
->base
.gallivm
, values
,
795 /* Get the vertex offset parameter */
796 vtx_offset_param
= reg
->Dimension
.Index
;
797 if (vtx_offset_param
< 2) {
798 vtx_offset_param
+= SI_PARAM_VTX0_OFFSET
;
800 assert(vtx_offset_param
< 6);
801 vtx_offset_param
+= SI_PARAM_VTX2_OFFSET
- 2;
803 vtx_offset
= lp_build_mul_imm(uint
,
804 LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
,
808 args
[0] = si_shader_ctx
->esgs_ring
;
809 args
[1] = vtx_offset
;
810 args
[2] = lp_build_const_int32(gallivm
,
811 (get_param_index(semantic_name
, semantic_index
,
812 shader
->selector
->inputs_read
) * 4 +
814 args
[3] = uint
->zero
;
815 args
[4] = uint
->one
; /* OFFEN */
816 args
[5] = uint
->zero
; /* IDXEN */
817 args
[6] = uint
->one
; /* GLC */
818 args
[7] = uint
->zero
; /* SLC */
819 args
[8] = uint
->zero
; /* TFE */
821 return LLVMBuildBitCast(gallivm
->builder
,
822 lp_build_intrinsic(gallivm
->builder
,
823 "llvm.SI.buffer.load.dword.i32.i32",
825 LLVMReadOnlyAttribute
| LLVMNoUnwindAttribute
),
826 tgsi2llvmtype(bld_base
, type
), "");
829 static int lookup_interp_param_index(unsigned interpolate
, unsigned location
)
831 switch (interpolate
) {
832 case TGSI_INTERPOLATE_CONSTANT
:
835 case TGSI_INTERPOLATE_LINEAR
:
836 if (location
== TGSI_INTERPOLATE_LOC_SAMPLE
)
837 return SI_PARAM_LINEAR_SAMPLE
;
838 else if (location
== TGSI_INTERPOLATE_LOC_CENTROID
)
839 return SI_PARAM_LINEAR_CENTROID
;
841 return SI_PARAM_LINEAR_CENTER
;
843 case TGSI_INTERPOLATE_COLOR
:
844 case TGSI_INTERPOLATE_PERSPECTIVE
:
845 if (location
== TGSI_INTERPOLATE_LOC_SAMPLE
)
846 return SI_PARAM_PERSP_SAMPLE
;
847 else if (location
== TGSI_INTERPOLATE_LOC_CENTROID
)
848 return SI_PARAM_PERSP_CENTROID
;
850 return SI_PARAM_PERSP_CENTER
;
853 fprintf(stderr
, "Warning: Unhandled interpolation mode.\n");
858 static void declare_input_fs(
859 struct radeon_llvm_context
*radeon_bld
,
860 unsigned input_index
,
861 const struct tgsi_full_declaration
*decl
)
863 struct lp_build_context
*base
= &radeon_bld
->soa
.bld_base
.base
;
864 struct si_shader_context
*si_shader_ctx
=
865 si_shader_context(&radeon_bld
->soa
.bld_base
);
866 struct si_shader
*shader
= si_shader_ctx
->shader
;
867 struct lp_build_context
*uint
= &radeon_bld
->soa
.bld_base
.uint_bld
;
868 struct gallivm_state
*gallivm
= base
->gallivm
;
869 LLVMTypeRef input_type
= LLVMFloatTypeInContext(gallivm
->context
);
870 LLVMValueRef main_fn
= radeon_bld
->main_fn
;
872 LLVMValueRef interp_param
= NULL
;
873 int interp_param_idx
;
874 const char * intr_name
;
877 * [15:0] NewPrimMask (Bit mask for each quad. It is set it the
878 * quad begins a new primitive. Bit 0 always needs
880 * [32:16] ParamOffset
883 LLVMValueRef params
= LLVMGetParam(main_fn
, SI_PARAM_PRIM_MASK
);
884 LLVMValueRef attr_number
;
888 if (decl
->Semantic
.Name
== TGSI_SEMANTIC_POSITION
) {
889 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
891 radeon_llvm_reg_index_soa(input_index
, chan
);
892 radeon_bld
->inputs
[soa_index
] =
893 LLVMGetParam(main_fn
, SI_PARAM_POS_X_FLOAT
+ chan
);
896 /* RCP for fragcoord.w */
897 radeon_bld
->inputs
[soa_index
] =
898 LLVMBuildFDiv(gallivm
->builder
,
899 lp_build_const_float(gallivm
, 1.0f
),
900 radeon_bld
->inputs
[soa_index
],
906 if (decl
->Semantic
.Name
== TGSI_SEMANTIC_FACE
) {
907 radeon_bld
->inputs
[radeon_llvm_reg_index_soa(input_index
, 0)] =
908 LLVMGetParam(main_fn
, SI_PARAM_FRONT_FACE
);
909 radeon_bld
->inputs
[radeon_llvm_reg_index_soa(input_index
, 1)] =
910 radeon_bld
->inputs
[radeon_llvm_reg_index_soa(input_index
, 2)] =
911 lp_build_const_float(gallivm
, 0.0f
);
912 radeon_bld
->inputs
[radeon_llvm_reg_index_soa(input_index
, 3)] =
913 lp_build_const_float(gallivm
, 1.0f
);
918 shader
->ps_input_param_offset
[input_index
] = shader
->nparam
++;
919 attr_number
= lp_build_const_int32(gallivm
,
920 shader
->ps_input_param_offset
[input_index
]);
922 shader
->ps_input_interpolate
[input_index
] = decl
->Interp
.Interpolate
;
923 interp_param_idx
= lookup_interp_param_index(decl
->Interp
.Interpolate
,
924 decl
->Interp
.Location
);
925 if (interp_param_idx
== -1)
927 else if (interp_param_idx
)
928 interp_param
= LLVMGetParam(main_fn
, interp_param_idx
);
930 /* fs.constant returns the param from the middle vertex, so it's not
931 * really useful for flat shading. It's meant to be used for custom
932 * interpolation (but the intrinsic can't fetch from the other two
935 * Luckily, it doesn't matter, because we rely on the FLAT_SHADE state
936 * to do the right thing. The only reason we use fs.constant is that
937 * fs.interp cannot be used on integers, because they can be equal
940 intr_name
= interp_param
? "llvm.SI.fs.interp" : "llvm.SI.fs.constant";
942 if (decl
->Semantic
.Name
== TGSI_SEMANTIC_COLOR
&&
943 si_shader_ctx
->shader
->key
.ps
.color_two_side
) {
944 LLVMValueRef args
[4];
945 LLVMValueRef face
, is_face_positive
;
946 LLVMValueRef back_attr_number
=
947 lp_build_const_int32(gallivm
,
948 shader
->ps_input_param_offset
[input_index
] + 1);
950 face
= LLVMGetParam(main_fn
, SI_PARAM_FRONT_FACE
);
952 is_face_positive
= LLVMBuildFCmp(gallivm
->builder
,
954 lp_build_const_float(gallivm
, 0.0f
),
958 args
[3] = interp_param
;
959 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
960 LLVMValueRef llvm_chan
= lp_build_const_int32(gallivm
, chan
);
961 unsigned soa_index
= radeon_llvm_reg_index_soa(input_index
, chan
);
962 LLVMValueRef front
, back
;
965 args
[1] = attr_number
;
966 front
= lp_build_intrinsic(gallivm
->builder
, intr_name
,
967 input_type
, args
, args
[3] ? 4 : 3,
968 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
970 args
[1] = back_attr_number
;
971 back
= lp_build_intrinsic(gallivm
->builder
, intr_name
,
972 input_type
, args
, args
[3] ? 4 : 3,
973 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
975 radeon_bld
->inputs
[soa_index
] =
976 LLVMBuildSelect(gallivm
->builder
,
984 } else if (decl
->Semantic
.Name
== TGSI_SEMANTIC_FOG
) {
985 LLVMValueRef args
[4];
987 args
[0] = uint
->zero
;
988 args
[1] = attr_number
;
990 args
[3] = interp_param
;
991 radeon_bld
->inputs
[radeon_llvm_reg_index_soa(input_index
, 0)] =
992 lp_build_intrinsic(gallivm
->builder
, intr_name
,
993 input_type
, args
, args
[3] ? 4 : 3,
994 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
995 radeon_bld
->inputs
[radeon_llvm_reg_index_soa(input_index
, 1)] =
996 radeon_bld
->inputs
[radeon_llvm_reg_index_soa(input_index
, 2)] =
997 lp_build_const_float(gallivm
, 0.0f
);
998 radeon_bld
->inputs
[radeon_llvm_reg_index_soa(input_index
, 3)] =
999 lp_build_const_float(gallivm
, 1.0f
);
1001 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
1002 LLVMValueRef args
[4];
1003 LLVMValueRef llvm_chan
= lp_build_const_int32(gallivm
, chan
);
1004 unsigned soa_index
= radeon_llvm_reg_index_soa(input_index
, chan
);
1005 args
[0] = llvm_chan
;
1006 args
[1] = attr_number
;
1008 args
[3] = interp_param
;
1009 radeon_bld
->inputs
[soa_index
] =
1010 lp_build_intrinsic(gallivm
->builder
, intr_name
,
1011 input_type
, args
, args
[3] ? 4 : 3,
1012 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
1017 static LLVMValueRef
get_sample_id(struct radeon_llvm_context
*radeon_bld
)
1019 return unpack_param(si_shader_context(&radeon_bld
->soa
.bld_base
),
1020 SI_PARAM_ANCILLARY
, 8, 4);
1024 * Load a dword from a constant buffer.
1026 static LLVMValueRef
buffer_load_const(LLVMBuilderRef builder
, LLVMValueRef resource
,
1027 LLVMValueRef offset
, LLVMTypeRef return_type
)
1029 LLVMValueRef args
[2] = {resource
, offset
};
1031 return lp_build_intrinsic(builder
, "llvm.SI.load.const", return_type
, args
, 2,
1032 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
1035 static LLVMValueRef
load_sample_position(struct radeon_llvm_context
*radeon_bld
, LLVMValueRef sample_id
)
1037 struct si_shader_context
*si_shader_ctx
=
1038 si_shader_context(&radeon_bld
->soa
.bld_base
);
1039 struct lp_build_context
*uint_bld
= &radeon_bld
->soa
.bld_base
.uint_bld
;
1040 struct gallivm_state
*gallivm
= &radeon_bld
->gallivm
;
1041 LLVMBuilderRef builder
= gallivm
->builder
;
1042 LLVMValueRef desc
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, SI_PARAM_CONST
);
1043 LLVMValueRef buf_index
= lp_build_const_int32(gallivm
, SI_DRIVER_STATE_CONST_BUF
);
1044 LLVMValueRef resource
= build_indexed_load_const(si_shader_ctx
, desc
, buf_index
);
1046 /* offset = sample_id * 8 (8 = 2 floats containing samplepos.xy) */
1047 LLVMValueRef offset0
= lp_build_mul_imm(uint_bld
, sample_id
, 8);
1048 LLVMValueRef offset1
= LLVMBuildAdd(builder
, offset0
, lp_build_const_int32(gallivm
, 4), "");
1050 LLVMValueRef pos
[4] = {
1051 buffer_load_const(builder
, resource
, offset0
, radeon_bld
->soa
.bld_base
.base
.elem_type
),
1052 buffer_load_const(builder
, resource
, offset1
, radeon_bld
->soa
.bld_base
.base
.elem_type
),
1053 lp_build_const_float(gallivm
, 0),
1054 lp_build_const_float(gallivm
, 0)
1057 return lp_build_gather_values(gallivm
, pos
, 4);
1060 static void declare_system_value(
1061 struct radeon_llvm_context
* radeon_bld
,
1063 const struct tgsi_full_declaration
*decl
)
1065 struct si_shader_context
*si_shader_ctx
=
1066 si_shader_context(&radeon_bld
->soa
.bld_base
);
1067 struct lp_build_context
*bld
= &radeon_bld
->soa
.bld_base
.base
;
1068 struct lp_build_context
*uint_bld
= &radeon_bld
->soa
.bld_base
.uint_bld
;
1069 struct gallivm_state
*gallivm
= &radeon_bld
->gallivm
;
1070 LLVMValueRef value
= 0;
1072 switch (decl
->Semantic
.Name
) {
1073 case TGSI_SEMANTIC_INSTANCEID
:
1074 value
= LLVMGetParam(radeon_bld
->main_fn
,
1075 si_shader_ctx
->param_instance_id
);
1078 case TGSI_SEMANTIC_VERTEXID
:
1079 value
= LLVMBuildAdd(gallivm
->builder
,
1080 LLVMGetParam(radeon_bld
->main_fn
,
1081 si_shader_ctx
->param_vertex_id
),
1082 LLVMGetParam(radeon_bld
->main_fn
,
1083 SI_PARAM_BASE_VERTEX
), "");
1086 case TGSI_SEMANTIC_VERTEXID_NOBASE
:
1087 value
= LLVMGetParam(radeon_bld
->main_fn
,
1088 si_shader_ctx
->param_vertex_id
);
1091 case TGSI_SEMANTIC_BASEVERTEX
:
1092 value
= LLVMGetParam(radeon_bld
->main_fn
,
1093 SI_PARAM_BASE_VERTEX
);
1096 case TGSI_SEMANTIC_INVOCATIONID
:
1097 if (si_shader_ctx
->type
== TGSI_PROCESSOR_TESS_CTRL
)
1098 value
= unpack_param(si_shader_ctx
, SI_PARAM_REL_IDS
, 8, 5);
1099 else if (si_shader_ctx
->type
== TGSI_PROCESSOR_GEOMETRY
)
1100 value
= LLVMGetParam(radeon_bld
->main_fn
,
1101 SI_PARAM_GS_INSTANCE_ID
);
1103 assert(!"INVOCATIONID not implemented");
1106 case TGSI_SEMANTIC_SAMPLEID
:
1107 value
= get_sample_id(radeon_bld
);
1110 case TGSI_SEMANTIC_SAMPLEPOS
:
1111 value
= load_sample_position(radeon_bld
, get_sample_id(radeon_bld
));
1114 case TGSI_SEMANTIC_SAMPLEMASK
:
1115 /* Smoothing isn't MSAA in GL, but it's MSAA in hardware.
1116 * Therefore, force gl_SampleMaskIn to 1 for GL. */
1117 if (si_shader_ctx
->shader
->key
.ps
.poly_line_smoothing
)
1118 value
= uint_bld
->one
;
1120 value
= LLVMGetParam(radeon_bld
->main_fn
, SI_PARAM_SAMPLE_COVERAGE
);
1123 case TGSI_SEMANTIC_TESSCOORD
:
1125 LLVMValueRef coord
[4] = {
1126 LLVMGetParam(radeon_bld
->main_fn
, si_shader_ctx
->param_tes_u
),
1127 LLVMGetParam(radeon_bld
->main_fn
, si_shader_ctx
->param_tes_v
),
1132 /* For triangles, the vector should be (u, v, 1-u-v). */
1133 if (si_shader_ctx
->shader
->selector
->info
.properties
[TGSI_PROPERTY_TES_PRIM_MODE
] ==
1134 PIPE_PRIM_TRIANGLES
)
1135 coord
[2] = lp_build_sub(bld
, bld
->one
,
1136 lp_build_add(bld
, coord
[0], coord
[1]));
1138 value
= lp_build_gather_values(gallivm
, coord
, 4);
1142 case TGSI_SEMANTIC_VERTICESIN
:
1143 value
= unpack_param(si_shader_ctx
, SI_PARAM_TCS_OUT_LAYOUT
, 26, 6);
1146 case TGSI_SEMANTIC_TESSINNER
:
1147 case TGSI_SEMANTIC_TESSOUTER
:
1149 LLVMValueRef dw_addr
;
1150 int param
= si_shader_io_get_unique_index(decl
->Semantic
.Name
, 0);
1152 dw_addr
= get_tcs_out_current_patch_data_offset(si_shader_ctx
);
1153 dw_addr
= LLVMBuildAdd(gallivm
->builder
, dw_addr
,
1154 lp_build_const_int32(gallivm
, param
* 4), "");
1156 value
= lds_load(&radeon_bld
->soa
.bld_base
, TGSI_TYPE_FLOAT
,
1161 case TGSI_SEMANTIC_PRIMID
:
1162 value
= get_primitive_id(&radeon_bld
->soa
.bld_base
, 0);
1166 assert(!"unknown system value");
1170 radeon_bld
->system_values
[index
] = value
;
1173 static LLVMValueRef
fetch_constant(
1174 struct lp_build_tgsi_context
* bld_base
,
1175 const struct tgsi_full_src_register
*reg
,
1176 enum tgsi_opcode_type type
,
1179 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
1180 struct lp_build_context
* base
= &bld_base
->base
;
1181 const struct tgsi_ind_register
*ireg
= ®
->Indirect
;
1184 LLVMValueRef addr
, bufp
;
1185 LLVMValueRef result
;
1187 if (swizzle
== LP_CHAN_ALL
) {
1189 LLVMValueRef values
[4];
1190 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
)
1191 values
[chan
] = fetch_constant(bld_base
, reg
, type
, chan
);
1193 return lp_build_gather_values(bld_base
->base
.gallivm
, values
, 4);
1196 buf
= reg
->Register
.Dimension
? reg
->Dimension
.Index
: 0;
1197 idx
= reg
->Register
.Index
* 4 + swizzle
;
1199 if (!reg
->Register
.Indirect
&& !reg
->Dimension
.Indirect
) {
1200 if (type
!= TGSI_TYPE_DOUBLE
)
1201 return bitcast(bld_base
, type
, si_shader_ctx
->constants
[buf
][idx
]);
1203 return radeon_llvm_emit_fetch_double(bld_base
,
1204 si_shader_ctx
->constants
[buf
][idx
],
1205 si_shader_ctx
->constants
[buf
][idx
+ 1]);
1209 if (reg
->Register
.Dimension
&& reg
->Dimension
.Indirect
) {
1210 LLVMValueRef ptr
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, SI_PARAM_CONST
);
1212 index
= get_indirect_index(si_shader_ctx
, ®
->DimIndirect
,
1213 reg
->Dimension
.Index
);
1214 bufp
= build_indexed_load_const(si_shader_ctx
, ptr
, index
);
1216 bufp
= si_shader_ctx
->const_resource
[buf
];
1218 addr
= si_shader_ctx
->radeon_bld
.soa
.addr
[ireg
->Index
][ireg
->Swizzle
];
1219 addr
= LLVMBuildLoad(base
->gallivm
->builder
, addr
, "load addr reg");
1220 addr
= lp_build_mul_imm(&bld_base
->uint_bld
, addr
, 16);
1221 addr
= lp_build_add(&bld_base
->uint_bld
, addr
,
1222 lp_build_const_int32(base
->gallivm
, idx
* 4));
1224 result
= buffer_load_const(base
->gallivm
->builder
, bufp
,
1225 addr
, bld_base
->base
.elem_type
);
1227 if (type
!= TGSI_TYPE_DOUBLE
)
1228 result
= bitcast(bld_base
, type
, result
);
1230 LLVMValueRef addr2
, result2
;
1231 addr2
= si_shader_ctx
->radeon_bld
.soa
.addr
[ireg
->Index
][ireg
->Swizzle
+ 1];
1232 addr2
= LLVMBuildLoad(base
->gallivm
->builder
, addr2
, "load addr reg2");
1233 addr2
= lp_build_mul_imm(&bld_base
->uint_bld
, addr2
, 16);
1234 addr2
= lp_build_add(&bld_base
->uint_bld
, addr2
,
1235 lp_build_const_int32(base
->gallivm
, idx
* 4));
1237 result2
= buffer_load_const(base
->gallivm
->builder
, si_shader_ctx
->const_resource
[buf
],
1238 addr2
, bld_base
->base
.elem_type
);
1240 result
= radeon_llvm_emit_fetch_double(bld_base
,
1246 /* Initialize arguments for the shader export intrinsic */
1247 static void si_llvm_init_export_args(struct lp_build_tgsi_context
*bld_base
,
1248 LLVMValueRef
*values
,
1252 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
1253 struct lp_build_context
*uint
=
1254 &si_shader_ctx
->radeon_bld
.soa
.bld_base
.uint_bld
;
1255 struct lp_build_context
*base
= &bld_base
->base
;
1256 unsigned compressed
= 0;
1259 if (si_shader_ctx
->type
== TGSI_PROCESSOR_FRAGMENT
) {
1260 int cbuf
= target
- V_008DFC_SQ_EXP_MRT
;
1262 if (cbuf
>= 0 && cbuf
< 8) {
1263 compressed
= (si_shader_ctx
->shader
->key
.ps
.export_16bpc
>> cbuf
) & 0x1;
1266 si_shader_ctx
->shader
->spi_shader_col_format
|=
1267 V_028714_SPI_SHADER_FP16_ABGR
<< (4 * cbuf
);
1269 si_shader_ctx
->shader
->spi_shader_col_format
|=
1270 V_028714_SPI_SHADER_32_ABGR
<< (4 * cbuf
);
1272 si_shader_ctx
->shader
->cb_shader_mask
|= 0xf << (4 * cbuf
);
1277 /* Pixel shader needs to pack output values before export */
1278 for (chan
= 0; chan
< 2; chan
++ ) {
1279 args
[0] = values
[2 * chan
];
1280 args
[1] = values
[2 * chan
+ 1];
1282 lp_build_intrinsic(base
->gallivm
->builder
,
1284 LLVMInt32TypeInContext(base
->gallivm
->context
),
1286 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
1287 args
[chan
+ 7] = args
[chan
+ 5] =
1288 LLVMBuildBitCast(base
->gallivm
->builder
,
1290 LLVMFloatTypeInContext(base
->gallivm
->context
),
1294 /* Set COMPR flag */
1295 args
[4] = uint
->one
;
1297 for (chan
= 0; chan
< 4; chan
++ )
1298 /* +5 because the first output value will be
1299 * the 6th argument to the intrinsic. */
1300 args
[chan
+ 5] = values
[chan
];
1302 /* Clear COMPR flag */
1303 args
[4] = uint
->zero
;
1306 /* XXX: This controls which components of the output
1307 * registers actually get exported. (e.g bit 0 means export
1308 * X component, bit 1 means export Y component, etc.) I'm
1309 * hard coding this to 0xf for now. In the future, we might
1310 * want to do something else. */
1311 args
[0] = lp_build_const_int32(base
->gallivm
, 0xf);
1313 /* Specify whether the EXEC mask represents the valid mask */
1314 args
[1] = uint
->zero
;
1316 /* Specify whether this is the last export */
1317 args
[2] = uint
->zero
;
1319 /* Specify the target we are exporting */
1320 args
[3] = lp_build_const_int32(base
->gallivm
, target
);
1322 /* XXX: We probably need to keep track of the output
1323 * values, so we know what we are passing to the next
1327 /* Load from output pointers and initialize arguments for the shader export intrinsic */
1328 static void si_llvm_init_export_args_load(struct lp_build_tgsi_context
*bld_base
,
1329 LLVMValueRef
*out_ptr
,
1333 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
1334 LLVMValueRef values
[4];
1337 for (i
= 0; i
< 4; i
++)
1338 values
[i
] = LLVMBuildLoad(gallivm
->builder
, out_ptr
[i
], "");
1340 si_llvm_init_export_args(bld_base
, values
, target
, args
);
1343 static void si_alpha_test(struct lp_build_tgsi_context
*bld_base
,
1344 LLVMValueRef alpha_ptr
)
1346 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
1347 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
1349 if (si_shader_ctx
->shader
->key
.ps
.alpha_func
!= PIPE_FUNC_NEVER
) {
1350 LLVMValueRef alpha_ref
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
,
1351 SI_PARAM_ALPHA_REF
);
1353 LLVMValueRef alpha_pass
=
1354 lp_build_cmp(&bld_base
->base
,
1355 si_shader_ctx
->shader
->key
.ps
.alpha_func
,
1356 LLVMBuildLoad(gallivm
->builder
, alpha_ptr
, ""),
1359 lp_build_select(&bld_base
->base
,
1361 lp_build_const_float(gallivm
, 1.0f
),
1362 lp_build_const_float(gallivm
, -1.0f
));
1364 lp_build_intrinsic(gallivm
->builder
,
1366 LLVMVoidTypeInContext(gallivm
->context
),
1369 lp_build_intrinsic(gallivm
->builder
,
1371 LLVMVoidTypeInContext(gallivm
->context
),
1375 si_shader_ctx
->shader
->db_shader_control
|= S_02880C_KILL_ENABLE(1);
1378 static void si_scale_alpha_by_sample_mask(struct lp_build_tgsi_context
*bld_base
,
1379 LLVMValueRef alpha_ptr
)
1381 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
1382 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
1383 LLVMValueRef coverage
, alpha
;
1385 /* alpha = alpha * popcount(coverage) / SI_NUM_SMOOTH_AA_SAMPLES */
1386 coverage
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
,
1387 SI_PARAM_SAMPLE_COVERAGE
);
1388 coverage
= bitcast(bld_base
, TGSI_TYPE_SIGNED
, coverage
);
1390 coverage
= lp_build_intrinsic(gallivm
->builder
, "llvm.ctpop.i32",
1391 bld_base
->int_bld
.elem_type
,
1392 &coverage
, 1, LLVMReadNoneAttribute
);
1394 coverage
= LLVMBuildUIToFP(gallivm
->builder
, coverage
,
1395 bld_base
->base
.elem_type
, "");
1397 coverage
= LLVMBuildFMul(gallivm
->builder
, coverage
,
1398 lp_build_const_float(gallivm
,
1399 1.0 / SI_NUM_SMOOTH_AA_SAMPLES
), "");
1401 alpha
= LLVMBuildLoad(gallivm
->builder
, alpha_ptr
, "");
1402 alpha
= LLVMBuildFMul(gallivm
->builder
, alpha
, coverage
, "");
1403 LLVMBuildStore(gallivm
->builder
, alpha
, alpha_ptr
);
1406 static void si_llvm_emit_clipvertex(struct lp_build_tgsi_context
* bld_base
,
1407 LLVMValueRef (*pos
)[9], LLVMValueRef
*out_elts
)
1409 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
1410 struct lp_build_context
*base
= &bld_base
->base
;
1411 struct lp_build_context
*uint
= &si_shader_ctx
->radeon_bld
.soa
.bld_base
.uint_bld
;
1414 unsigned const_chan
;
1415 LLVMValueRef base_elt
;
1416 LLVMValueRef ptr
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, SI_PARAM_CONST
);
1417 LLVMValueRef constbuf_index
= lp_build_const_int32(base
->gallivm
, SI_DRIVER_STATE_CONST_BUF
);
1418 LLVMValueRef const_resource
= build_indexed_load_const(si_shader_ctx
, ptr
, constbuf_index
);
1420 for (reg_index
= 0; reg_index
< 2; reg_index
++) {
1421 LLVMValueRef
*args
= pos
[2 + reg_index
];
1426 args
[8] = lp_build_const_float(base
->gallivm
, 0.0f
);
1428 /* Compute dot products of position and user clip plane vectors */
1429 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
1430 for (const_chan
= 0; const_chan
< TGSI_NUM_CHANNELS
; const_chan
++) {
1431 args
[1] = lp_build_const_int32(base
->gallivm
,
1432 ((reg_index
* 4 + chan
) * 4 +
1434 base_elt
= buffer_load_const(base
->gallivm
->builder
, const_resource
,
1435 args
[1], base
->elem_type
);
1437 lp_build_add(base
, args
[5 + chan
],
1438 lp_build_mul(base
, base_elt
,
1439 out_elts
[const_chan
]));
1443 args
[0] = lp_build_const_int32(base
->gallivm
, 0xf);
1444 args
[1] = uint
->zero
;
1445 args
[2] = uint
->zero
;
1446 args
[3] = lp_build_const_int32(base
->gallivm
,
1447 V_008DFC_SQ_EXP_POS
+ 2 + reg_index
);
1448 args
[4] = uint
->zero
;
1452 static void si_dump_streamout(struct pipe_stream_output_info
*so
)
1456 if (so
->num_outputs
)
1457 fprintf(stderr
, "STREAMOUT\n");
1459 for (i
= 0; i
< so
->num_outputs
; i
++) {
1460 unsigned mask
= ((1 << so
->output
[i
].num_components
) - 1) <<
1461 so
->output
[i
].start_component
;
1462 fprintf(stderr
, " %i: BUF%i[%i..%i] <- OUT[%i].%s%s%s%s\n",
1463 i
, so
->output
[i
].output_buffer
,
1464 so
->output
[i
].dst_offset
, so
->output
[i
].dst_offset
+ so
->output
[i
].num_components
- 1,
1465 so
->output
[i
].register_index
,
1466 mask
& 1 ? "x" : "",
1467 mask
& 2 ? "y" : "",
1468 mask
& 4 ? "z" : "",
1469 mask
& 8 ? "w" : "");
1473 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
1474 * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
1475 * or v4i32 (num_channels=3,4). */
1476 static void build_tbuffer_store(struct si_shader_context
*shader
,
1479 unsigned num_channels
,
1481 LLVMValueRef soffset
,
1482 unsigned inst_offset
,
1491 struct gallivm_state
*gallivm
= &shader
->radeon_bld
.gallivm
;
1492 LLVMTypeRef i32
= LLVMInt32TypeInContext(gallivm
->context
);
1493 LLVMValueRef args
[] = {
1496 LLVMConstInt(i32
, num_channels
, 0),
1499 LLVMConstInt(i32
, inst_offset
, 0),
1500 LLVMConstInt(i32
, dfmt
, 0),
1501 LLVMConstInt(i32
, nfmt
, 0),
1502 LLVMConstInt(i32
, offen
, 0),
1503 LLVMConstInt(i32
, idxen
, 0),
1504 LLVMConstInt(i32
, glc
, 0),
1505 LLVMConstInt(i32
, slc
, 0),
1506 LLVMConstInt(i32
, tfe
, 0)
1509 /* The instruction offset field has 12 bits */
1510 assert(offen
|| inst_offset
< (1 << 12));
1512 /* The intrinsic is overloaded, we need to add a type suffix for overloading to work. */
1513 unsigned func
= CLAMP(num_channels
, 1, 3) - 1;
1514 const char *types
[] = {"i32", "v2i32", "v4i32"};
1516 snprintf(name
, sizeof(name
), "llvm.SI.tbuffer.store.%s", types
[func
]);
1518 lp_build_intrinsic(gallivm
->builder
, name
,
1519 LLVMVoidTypeInContext(gallivm
->context
),
1520 args
, Elements(args
), 0);
1523 static void build_tbuffer_store_dwords(struct si_shader_context
*shader
,
1526 unsigned num_channels
,
1528 LLVMValueRef soffset
,
1529 unsigned inst_offset
)
1531 static unsigned dfmt
[] = {
1532 V_008F0C_BUF_DATA_FORMAT_32
,
1533 V_008F0C_BUF_DATA_FORMAT_32_32
,
1534 V_008F0C_BUF_DATA_FORMAT_32_32_32
,
1535 V_008F0C_BUF_DATA_FORMAT_32_32_32_32
1537 assert(num_channels
>= 1 && num_channels
<= 4);
1539 build_tbuffer_store(shader
, rsrc
, vdata
, num_channels
, vaddr
, soffset
,
1540 inst_offset
, dfmt
[num_channels
-1],
1541 V_008F0C_BUF_NUM_FORMAT_UINT
, 1, 0, 1, 1, 0);
1544 /* On SI, the vertex shader is responsible for writing streamout data
1546 static void si_llvm_emit_streamout(struct si_shader_context
*shader
,
1547 struct si_shader_output_values
*outputs
,
1550 struct pipe_stream_output_info
*so
= &shader
->shader
->selector
->so
;
1551 struct gallivm_state
*gallivm
= &shader
->radeon_bld
.gallivm
;
1552 LLVMBuilderRef builder
= gallivm
->builder
;
1554 struct lp_build_if_state if_ctx
;
1556 LLVMTypeRef i32
= LLVMInt32TypeInContext(gallivm
->context
);
1558 /* Get bits [22:16], i.e. (so_param >> 16) & 127; */
1559 LLVMValueRef so_vtx_count
=
1560 unpack_param(shader
, shader
->param_streamout_config
, 16, 7);
1562 LLVMValueRef tid
= lp_build_intrinsic(builder
, "llvm.SI.tid", i32
,
1563 NULL
, 0, LLVMReadNoneAttribute
);
1565 /* can_emit = tid < so_vtx_count; */
1566 LLVMValueRef can_emit
=
1567 LLVMBuildICmp(builder
, LLVMIntULT
, tid
, so_vtx_count
, "");
1569 LLVMValueRef stream_id
=
1570 unpack_param(shader
, shader
->param_streamout_config
, 24, 2);
1572 /* Emit the streamout code conditionally. This actually avoids
1573 * out-of-bounds buffer access. The hw tells us via the SGPR
1574 * (so_vtx_count) which threads are allowed to emit streamout data. */
1575 lp_build_if(&if_ctx
, gallivm
, can_emit
);
1577 /* The buffer offset is computed as follows:
1578 * ByteOffset = streamout_offset[buffer_id]*4 +
1579 * (streamout_write_index + thread_id)*stride[buffer_id] +
1583 LLVMValueRef so_write_index
=
1584 LLVMGetParam(shader
->radeon_bld
.main_fn
,
1585 shader
->param_streamout_write_index
);
1587 /* Compute (streamout_write_index + thread_id). */
1588 so_write_index
= LLVMBuildAdd(builder
, so_write_index
, tid
, "");
1590 /* Compute the write offset for each enabled buffer. */
1591 LLVMValueRef so_write_offset
[4] = {};
1592 for (i
= 0; i
< 4; i
++) {
1596 LLVMValueRef so_offset
= LLVMGetParam(shader
->radeon_bld
.main_fn
,
1597 shader
->param_streamout_offset
[i
]);
1598 so_offset
= LLVMBuildMul(builder
, so_offset
, LLVMConstInt(i32
, 4, 0), "");
1600 so_write_offset
[i
] = LLVMBuildMul(builder
, so_write_index
,
1601 LLVMConstInt(i32
, so
->stride
[i
]*4, 0), "");
1602 so_write_offset
[i
] = LLVMBuildAdd(builder
, so_write_offset
[i
], so_offset
, "");
1605 /* Write streamout data. */
1606 for (i
= 0; i
< so
->num_outputs
; i
++) {
1607 unsigned buf_idx
= so
->output
[i
].output_buffer
;
1608 unsigned reg
= so
->output
[i
].register_index
;
1609 unsigned start
= so
->output
[i
].start_component
;
1610 unsigned num_comps
= so
->output
[i
].num_components
;
1611 unsigned stream
= so
->output
[i
].stream
;
1612 LLVMValueRef out
[4];
1613 struct lp_build_if_state if_ctx_stream
;
1615 assert(num_comps
&& num_comps
<= 4);
1616 if (!num_comps
|| num_comps
> 4)
1622 /* Load the output as int. */
1623 for (j
= 0; j
< num_comps
; j
++) {
1624 out
[j
] = LLVMBuildBitCast(builder
,
1625 outputs
[reg
].values
[start
+j
],
1629 /* Pack the output. */
1630 LLVMValueRef vdata
= NULL
;
1632 switch (num_comps
) {
1633 case 1: /* as i32 */
1636 case 2: /* as v2i32 */
1637 case 3: /* as v4i32 (aligned to 4) */
1638 case 4: /* as v4i32 */
1639 vdata
= LLVMGetUndef(LLVMVectorType(i32
, util_next_power_of_two(num_comps
)));
1640 for (j
= 0; j
< num_comps
; j
++) {
1641 vdata
= LLVMBuildInsertElement(builder
, vdata
, out
[j
],
1642 LLVMConstInt(i32
, j
, 0), "");
1647 LLVMValueRef can_emit_stream
=
1648 LLVMBuildICmp(builder
, LLVMIntEQ
,
1650 lp_build_const_int32(gallivm
, stream
), "");
1652 lp_build_if(&if_ctx_stream
, gallivm
, can_emit_stream
);
1653 build_tbuffer_store_dwords(shader
, shader
->so_buffers
[buf_idx
],
1655 so_write_offset
[buf_idx
],
1656 LLVMConstInt(i32
, 0, 0),
1657 so
->output
[i
].dst_offset
*4);
1658 lp_build_endif(&if_ctx_stream
);
1661 lp_build_endif(&if_ctx
);
1665 /* Generate export instructions for hardware VS shader stage */
1666 static void si_llvm_export_vs(struct lp_build_tgsi_context
*bld_base
,
1667 struct si_shader_output_values
*outputs
,
1670 struct si_shader_context
* si_shader_ctx
= si_shader_context(bld_base
);
1671 struct si_shader
* shader
= si_shader_ctx
->shader
;
1672 struct lp_build_context
* base
= &bld_base
->base
;
1673 struct lp_build_context
* uint
=
1674 &si_shader_ctx
->radeon_bld
.soa
.bld_base
.uint_bld
;
1675 LLVMValueRef args
[9];
1676 LLVMValueRef pos_args
[4][9] = { { 0 } };
1677 LLVMValueRef psize_value
= NULL
, edgeflag_value
= NULL
, layer_value
= NULL
, viewport_index_value
= NULL
;
1678 unsigned semantic_name
, semantic_index
;
1680 unsigned param_count
= 0;
1684 if (outputs
&& si_shader_ctx
->shader
->selector
->so
.num_outputs
) {
1685 si_llvm_emit_streamout(si_shader_ctx
, outputs
, noutput
);
1688 for (i
= 0; i
< noutput
; i
++) {
1689 semantic_name
= outputs
[i
].name
;
1690 semantic_index
= outputs
[i
].sid
;
1693 /* Select the correct target */
1694 switch(semantic_name
) {
1695 case TGSI_SEMANTIC_PSIZE
:
1696 psize_value
= outputs
[i
].values
[0];
1698 case TGSI_SEMANTIC_EDGEFLAG
:
1699 edgeflag_value
= outputs
[i
].values
[0];
1701 case TGSI_SEMANTIC_LAYER
:
1702 layer_value
= outputs
[i
].values
[0];
1703 semantic_name
= TGSI_SEMANTIC_GENERIC
;
1704 goto handle_semantic
;
1705 case TGSI_SEMANTIC_VIEWPORT_INDEX
:
1706 viewport_index_value
= outputs
[i
].values
[0];
1707 semantic_name
= TGSI_SEMANTIC_GENERIC
;
1708 goto handle_semantic
;
1709 case TGSI_SEMANTIC_POSITION
:
1710 target
= V_008DFC_SQ_EXP_POS
;
1712 case TGSI_SEMANTIC_COLOR
:
1713 case TGSI_SEMANTIC_BCOLOR
:
1714 target
= V_008DFC_SQ_EXP_PARAM
+ param_count
;
1715 shader
->vs_output_param_offset
[i
] = param_count
;
1718 case TGSI_SEMANTIC_CLIPDIST
:
1719 target
= V_008DFC_SQ_EXP_POS
+ 2 + semantic_index
;
1721 case TGSI_SEMANTIC_CLIPVERTEX
:
1722 si_llvm_emit_clipvertex(bld_base
, pos_args
, outputs
[i
].values
);
1724 case TGSI_SEMANTIC_PRIMID
:
1725 case TGSI_SEMANTIC_FOG
:
1726 case TGSI_SEMANTIC_TEXCOORD
:
1727 case TGSI_SEMANTIC_GENERIC
:
1728 target
= V_008DFC_SQ_EXP_PARAM
+ param_count
;
1729 shader
->vs_output_param_offset
[i
] = param_count
;
1735 "Warning: SI unhandled vs output type:%d\n",
1739 si_llvm_init_export_args(bld_base
, outputs
[i
].values
, target
, args
);
1741 if (target
>= V_008DFC_SQ_EXP_POS
&&
1742 target
<= (V_008DFC_SQ_EXP_POS
+ 3)) {
1743 memcpy(pos_args
[target
- V_008DFC_SQ_EXP_POS
],
1744 args
, sizeof(args
));
1746 lp_build_intrinsic(base
->gallivm
->builder
,
1748 LLVMVoidTypeInContext(base
->gallivm
->context
),
1752 if (semantic_name
== TGSI_SEMANTIC_CLIPDIST
) {
1753 semantic_name
= TGSI_SEMANTIC_GENERIC
;
1754 goto handle_semantic
;
1758 shader
->nr_param_exports
= param_count
;
1760 /* We need to add the position output manually if it's missing. */
1761 if (!pos_args
[0][0]) {
1762 pos_args
[0][0] = lp_build_const_int32(base
->gallivm
, 0xf); /* writemask */
1763 pos_args
[0][1] = uint
->zero
; /* EXEC mask */
1764 pos_args
[0][2] = uint
->zero
; /* last export? */
1765 pos_args
[0][3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_POS
);
1766 pos_args
[0][4] = uint
->zero
; /* COMPR flag */
1767 pos_args
[0][5] = base
->zero
; /* X */
1768 pos_args
[0][6] = base
->zero
; /* Y */
1769 pos_args
[0][7] = base
->zero
; /* Z */
1770 pos_args
[0][8] = base
->one
; /* W */
1773 /* Write the misc vector (point size, edgeflag, layer, viewport). */
1774 if (shader
->selector
->info
.writes_psize
||
1775 shader
->selector
->info
.writes_edgeflag
||
1776 shader
->selector
->info
.writes_viewport_index
||
1777 shader
->selector
->info
.writes_layer
) {
1778 pos_args
[1][0] = lp_build_const_int32(base
->gallivm
, /* writemask */
1779 shader
->selector
->info
.writes_psize
|
1780 (shader
->selector
->info
.writes_edgeflag
<< 1) |
1781 (shader
->selector
->info
.writes_layer
<< 2) |
1782 (shader
->selector
->info
.writes_viewport_index
<< 3));
1783 pos_args
[1][1] = uint
->zero
; /* EXEC mask */
1784 pos_args
[1][2] = uint
->zero
; /* last export? */
1785 pos_args
[1][3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_POS
+ 1);
1786 pos_args
[1][4] = uint
->zero
; /* COMPR flag */
1787 pos_args
[1][5] = base
->zero
; /* X */
1788 pos_args
[1][6] = base
->zero
; /* Y */
1789 pos_args
[1][7] = base
->zero
; /* Z */
1790 pos_args
[1][8] = base
->zero
; /* W */
1792 if (shader
->selector
->info
.writes_psize
)
1793 pos_args
[1][5] = psize_value
;
1795 if (shader
->selector
->info
.writes_edgeflag
) {
1796 /* The output is a float, but the hw expects an integer
1797 * with the first bit containing the edge flag. */
1798 edgeflag_value
= LLVMBuildFPToUI(base
->gallivm
->builder
,
1800 bld_base
->uint_bld
.elem_type
, "");
1801 edgeflag_value
= lp_build_min(&bld_base
->int_bld
,
1803 bld_base
->int_bld
.one
);
1805 /* The LLVM intrinsic expects a float. */
1806 pos_args
[1][6] = LLVMBuildBitCast(base
->gallivm
->builder
,
1808 base
->elem_type
, "");
1811 if (shader
->selector
->info
.writes_layer
)
1812 pos_args
[1][7] = layer_value
;
1814 if (shader
->selector
->info
.writes_viewport_index
)
1815 pos_args
[1][8] = viewport_index_value
;
1818 for (i
= 0; i
< 4; i
++)
1820 shader
->nr_pos_exports
++;
1823 for (i
= 0; i
< 4; i
++) {
1824 if (!pos_args
[i
][0])
1827 /* Specify the target we are exporting */
1828 pos_args
[i
][3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_POS
+ pos_idx
++);
1830 if (pos_idx
== shader
->nr_pos_exports
)
1831 /* Specify that this is the last export */
1832 pos_args
[i
][2] = uint
->one
;
1834 lp_build_intrinsic(base
->gallivm
->builder
,
1836 LLVMVoidTypeInContext(base
->gallivm
->context
),
1841 /* This only writes the tessellation factor levels. */
1842 static void si_llvm_emit_tcs_epilogue(struct lp_build_tgsi_context
*bld_base
)
1844 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
1845 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
1846 struct si_shader
*shader
= si_shader_ctx
->shader
;
1847 unsigned tess_inner_index
, tess_outer_index
;
1848 LLVMValueRef lds_base
, lds_inner
, lds_outer
;
1849 LLVMValueRef tf_base
, rel_patch_id
, byteoffset
, buffer
, rw_buffers
;
1850 LLVMValueRef out
[6], vec0
, vec1
, invocation_id
;
1851 unsigned stride
, outer_comps
, inner_comps
, i
;
1852 struct lp_build_if_state if_ctx
;
1854 invocation_id
= unpack_param(si_shader_ctx
, SI_PARAM_REL_IDS
, 8, 5);
1856 /* Do this only for invocation 0, because the tess levels are per-patch,
1859 * This can't jump, because invocation 0 executes this. It should
1860 * at least mask out the loads and stores for other invocations.
1862 lp_build_if(&if_ctx
, gallivm
,
1863 LLVMBuildICmp(gallivm
->builder
, LLVMIntEQ
,
1864 invocation_id
, bld_base
->uint_bld
.zero
, ""));
1866 /* Determine the layout of one tess factor element in the buffer. */
1867 switch (shader
->key
.tcs
.prim_mode
) {
1868 case PIPE_PRIM_LINES
:
1869 stride
= 2; /* 2 dwords, 1 vec2 store */
1873 case PIPE_PRIM_TRIANGLES
:
1874 stride
= 4; /* 4 dwords, 1 vec4 store */
1878 case PIPE_PRIM_QUADS
:
1879 stride
= 6; /* 6 dwords, 2 stores (vec4 + vec2) */
1888 /* Load tess_inner and tess_outer from LDS.
1889 * Any invocation can write them, so we can't get them from a temporary.
1891 tess_inner_index
= si_shader_io_get_unique_index(TGSI_SEMANTIC_TESSINNER
, 0);
1892 tess_outer_index
= si_shader_io_get_unique_index(TGSI_SEMANTIC_TESSOUTER
, 0);
1894 lds_base
= get_tcs_out_current_patch_data_offset(si_shader_ctx
);
1895 lds_inner
= LLVMBuildAdd(gallivm
->builder
, lds_base
,
1896 lp_build_const_int32(gallivm
,
1897 tess_inner_index
* 4), "");
1898 lds_outer
= LLVMBuildAdd(gallivm
->builder
, lds_base
,
1899 lp_build_const_int32(gallivm
,
1900 tess_outer_index
* 4), "");
1902 for (i
= 0; i
< outer_comps
; i
++)
1903 out
[i
] = lds_load(bld_base
, TGSI_TYPE_SIGNED
, i
, lds_outer
);
1904 for (i
= 0; i
< inner_comps
; i
++)
1905 out
[outer_comps
+i
] = lds_load(bld_base
, TGSI_TYPE_SIGNED
, i
, lds_inner
);
1907 /* Convert the outputs to vectors for stores. */
1908 vec0
= lp_build_gather_values(gallivm
, out
, MIN2(stride
, 4));
1912 vec1
= lp_build_gather_values(gallivm
, out
+4, stride
- 4);
1914 /* Get the buffer. */
1915 rw_buffers
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
,
1916 SI_PARAM_RW_BUFFERS
);
1917 buffer
= build_indexed_load_const(si_shader_ctx
, rw_buffers
,
1918 lp_build_const_int32(gallivm
, SI_RING_TESS_FACTOR
));
1920 /* Get the offset. */
1921 tf_base
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
,
1922 SI_PARAM_TESS_FACTOR_OFFSET
);
1923 rel_patch_id
= get_rel_patch_id(si_shader_ctx
);
1924 byteoffset
= LLVMBuildMul(gallivm
->builder
, rel_patch_id
,
1925 lp_build_const_int32(gallivm
, 4 * stride
), "");
1927 /* Store the outputs. */
1928 build_tbuffer_store_dwords(si_shader_ctx
, buffer
, vec0
,
1929 MIN2(stride
, 4), byteoffset
, tf_base
, 0);
1931 build_tbuffer_store_dwords(si_shader_ctx
, buffer
, vec1
,
1932 stride
- 4, byteoffset
, tf_base
, 16);
1933 lp_build_endif(&if_ctx
);
1936 static void si_llvm_emit_ls_epilogue(struct lp_build_tgsi_context
* bld_base
)
1938 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
1939 struct si_shader
*shader
= si_shader_ctx
->shader
;
1940 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
1941 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
1943 LLVMValueRef vertex_id
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
,
1944 si_shader_ctx
->param_rel_auto_id
);
1945 LLVMValueRef vertex_dw_stride
=
1946 unpack_param(si_shader_ctx
, SI_PARAM_LS_OUT_LAYOUT
, 13, 8);
1947 LLVMValueRef base_dw_addr
= LLVMBuildMul(gallivm
->builder
, vertex_id
,
1948 vertex_dw_stride
, "");
1950 /* Write outputs to LDS. The next shader (TCS aka HS) will read
1951 * its inputs from it. */
1952 for (i
= 0; i
< info
->num_outputs
; i
++) {
1953 LLVMValueRef
*out_ptr
= si_shader_ctx
->radeon_bld
.soa
.outputs
[i
];
1954 unsigned name
= info
->output_semantic_name
[i
];
1955 unsigned index
= info
->output_semantic_index
[i
];
1956 int param
= si_shader_io_get_unique_index(name
, index
);
1957 LLVMValueRef dw_addr
= LLVMBuildAdd(gallivm
->builder
, base_dw_addr
,
1958 lp_build_const_int32(gallivm
, param
* 4), "");
1960 for (chan
= 0; chan
< 4; chan
++) {
1961 lds_store(bld_base
, chan
, dw_addr
,
1962 LLVMBuildLoad(gallivm
->builder
, out_ptr
[chan
], ""));
1967 static void si_llvm_emit_es_epilogue(struct lp_build_tgsi_context
* bld_base
)
1969 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
1970 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
1971 struct si_shader
*es
= si_shader_ctx
->shader
;
1972 struct tgsi_shader_info
*info
= &es
->selector
->info
;
1973 LLVMTypeRef i32
= LLVMInt32TypeInContext(gallivm
->context
);
1974 LLVMValueRef soffset
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
,
1975 si_shader_ctx
->param_es2gs_offset
);
1976 uint64_t enabled_outputs
= si_shader_ctx
->type
== TGSI_PROCESSOR_TESS_EVAL
?
1977 es
->key
.tes
.es_enabled_outputs
:
1978 es
->key
.vs
.es_enabled_outputs
;
1982 for (i
= 0; i
< info
->num_outputs
; i
++) {
1983 LLVMValueRef
*out_ptr
=
1984 si_shader_ctx
->radeon_bld
.soa
.outputs
[i
];
1987 if (info
->output_semantic_name
[i
] == TGSI_SEMANTIC_VIEWPORT_INDEX
||
1988 info
->output_semantic_name
[i
] == TGSI_SEMANTIC_LAYER
)
1991 param_index
= get_param_index(info
->output_semantic_name
[i
],
1992 info
->output_semantic_index
[i
],
1994 if (param_index
< 0)
1997 for (chan
= 0; chan
< 4; chan
++) {
1998 LLVMValueRef out_val
= LLVMBuildLoad(gallivm
->builder
, out_ptr
[chan
], "");
1999 out_val
= LLVMBuildBitCast(gallivm
->builder
, out_val
, i32
, "");
2001 build_tbuffer_store(si_shader_ctx
,
2002 si_shader_ctx
->esgs_ring
,
2004 LLVMGetUndef(i32
), soffset
,
2005 (4 * param_index
+ chan
) * 4,
2006 V_008F0C_BUF_DATA_FORMAT_32
,
2007 V_008F0C_BUF_NUM_FORMAT_UINT
,
2013 static void si_llvm_emit_gs_epilogue(struct lp_build_tgsi_context
*bld_base
)
2015 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
2016 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2017 LLVMValueRef args
[2];
2019 args
[0] = lp_build_const_int32(gallivm
, SENDMSG_GS_OP_NOP
| SENDMSG_GS_DONE
);
2020 args
[1] = LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, SI_PARAM_GS_WAVE_ID
);
2021 lp_build_intrinsic(gallivm
->builder
, "llvm.SI.sendmsg",
2022 LLVMVoidTypeInContext(gallivm
->context
), args
, 2,
2023 LLVMNoUnwindAttribute
);
2026 static void si_llvm_emit_vs_epilogue(struct lp_build_tgsi_context
* bld_base
)
2028 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
2029 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2030 struct tgsi_shader_info
*info
= &si_shader_ctx
->shader
->selector
->info
;
2031 struct si_shader_output_values
*outputs
= NULL
;
2034 outputs
= MALLOC((info
->num_outputs
+ 1) * sizeof(outputs
[0]));
2036 for (i
= 0; i
< info
->num_outputs
; i
++) {
2037 outputs
[i
].name
= info
->output_semantic_name
[i
];
2038 outputs
[i
].sid
= info
->output_semantic_index
[i
];
2040 for (j
= 0; j
< 4; j
++)
2041 outputs
[i
].values
[j
] =
2042 LLVMBuildLoad(gallivm
->builder
,
2043 si_shader_ctx
->radeon_bld
.soa
.outputs
[i
][j
],
2047 /* Export PrimitiveID when PS needs it. */
2048 if (si_vs_exports_prim_id(si_shader_ctx
->shader
)) {
2049 outputs
[i
].name
= TGSI_SEMANTIC_PRIMID
;
2051 outputs
[i
].values
[0] = bitcast(bld_base
, TGSI_TYPE_FLOAT
,
2052 get_primitive_id(bld_base
, 0));
2053 outputs
[i
].values
[1] = bld_base
->base
.undef
;
2054 outputs
[i
].values
[2] = bld_base
->base
.undef
;
2055 outputs
[i
].values
[3] = bld_base
->base
.undef
;
2059 si_llvm_export_vs(bld_base
, outputs
, i
);
2063 static void si_llvm_emit_fs_epilogue(struct lp_build_tgsi_context
* bld_base
)
2065 struct si_shader_context
* si_shader_ctx
= si_shader_context(bld_base
);
2066 struct si_shader
* shader
= si_shader_ctx
->shader
;
2067 struct lp_build_context
* base
= &bld_base
->base
;
2068 struct lp_build_context
* uint
= &bld_base
->uint_bld
;
2069 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
2070 LLVMValueRef args
[9];
2071 LLVMValueRef last_args
[9] = { 0 };
2072 int depth_index
= -1, stencil_index
= -1, samplemask_index
= -1;
2075 for (i
= 0; i
< info
->num_outputs
; i
++) {
2076 unsigned semantic_name
= info
->output_semantic_name
[i
];
2077 unsigned semantic_index
= info
->output_semantic_index
[i
];
2079 LLVMValueRef alpha_ptr
;
2081 /* Select the correct target */
2082 switch (semantic_name
) {
2083 case TGSI_SEMANTIC_POSITION
:
2086 case TGSI_SEMANTIC_STENCIL
:
2089 case TGSI_SEMANTIC_SAMPLEMASK
:
2090 samplemask_index
= i
;
2092 case TGSI_SEMANTIC_COLOR
:
2093 target
= V_008DFC_SQ_EXP_MRT
+ semantic_index
;
2094 alpha_ptr
= si_shader_ctx
->radeon_bld
.soa
.outputs
[i
][3];
2096 if (si_shader_ctx
->shader
->key
.ps
.alpha_to_one
)
2097 LLVMBuildStore(base
->gallivm
->builder
,
2098 base
->one
, alpha_ptr
);
2100 if (semantic_index
== 0 &&
2101 si_shader_ctx
->shader
->key
.ps
.alpha_func
!= PIPE_FUNC_ALWAYS
)
2102 si_alpha_test(bld_base
, alpha_ptr
);
2104 if (si_shader_ctx
->shader
->key
.ps
.poly_line_smoothing
)
2105 si_scale_alpha_by_sample_mask(bld_base
, alpha_ptr
);
2110 "Warning: SI unhandled fs output type:%d\n",
2114 si_llvm_init_export_args_load(bld_base
,
2115 si_shader_ctx
->radeon_bld
.soa
.outputs
[i
],
2118 if (semantic_name
== TGSI_SEMANTIC_COLOR
) {
2119 /* If there is an export instruction waiting to be emitted, do so now. */
2121 lp_build_intrinsic(base
->gallivm
->builder
,
2123 LLVMVoidTypeInContext(base
->gallivm
->context
),
2127 /* This instruction will be emitted at the end of the shader. */
2128 memcpy(last_args
, args
, sizeof(args
));
2130 /* Handle FS_COLOR0_WRITES_ALL_CBUFS. */
2131 if (shader
->selector
->info
.properties
[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
] &&
2132 semantic_index
== 0 &&
2133 si_shader_ctx
->shader
->key
.ps
.last_cbuf
> 0) {
2134 for (int c
= 1; c
<= si_shader_ctx
->shader
->key
.ps
.last_cbuf
; c
++) {
2135 si_llvm_init_export_args_load(bld_base
,
2136 si_shader_ctx
->radeon_bld
.soa
.outputs
[i
],
2137 V_008DFC_SQ_EXP_MRT
+ c
, args
);
2138 lp_build_intrinsic(base
->gallivm
->builder
,
2140 LLVMVoidTypeInContext(base
->gallivm
->context
),
2145 lp_build_intrinsic(base
->gallivm
->builder
,
2147 LLVMVoidTypeInContext(base
->gallivm
->context
),
2152 if (depth_index
>= 0 || stencil_index
>= 0 || samplemask_index
>= 0) {
2153 LLVMValueRef out_ptr
;
2156 /* Specify the target we are exporting */
2157 args
[3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_MRTZ
);
2159 args
[5] = base
->zero
; /* R, depth */
2160 args
[6] = base
->zero
; /* G, stencil test value[0:7], stencil op value[8:15] */
2161 args
[7] = base
->zero
; /* B, sample mask */
2162 args
[8] = base
->zero
; /* A, alpha to mask */
2164 if (depth_index
>= 0) {
2165 out_ptr
= si_shader_ctx
->radeon_bld
.soa
.outputs
[depth_index
][2];
2166 args
[5] = LLVMBuildLoad(base
->gallivm
->builder
, out_ptr
, "");
2168 si_shader_ctx
->shader
->db_shader_control
|= S_02880C_Z_EXPORT_ENABLE(1);
2171 if (stencil_index
>= 0) {
2172 out_ptr
= si_shader_ctx
->radeon_bld
.soa
.outputs
[stencil_index
][1];
2173 args
[6] = LLVMBuildLoad(base
->gallivm
->builder
, out_ptr
, "");
2175 si_shader_ctx
->shader
->db_shader_control
|=
2176 S_02880C_STENCIL_TEST_VAL_EXPORT_ENABLE(1);
2179 if (samplemask_index
>= 0) {
2180 out_ptr
= si_shader_ctx
->radeon_bld
.soa
.outputs
[samplemask_index
][0];
2181 args
[7] = LLVMBuildLoad(base
->gallivm
->builder
, out_ptr
, "");
2183 si_shader_ctx
->shader
->db_shader_control
|= S_02880C_MASK_EXPORT_ENABLE(1);
2186 /* SI (except OLAND) has a bug that it only looks
2187 * at the X writemask component. */
2188 if (si_shader_ctx
->screen
->b
.chip_class
== SI
&&
2189 si_shader_ctx
->screen
->b
.family
!= CHIP_OLAND
)
2192 if (samplemask_index
>= 0)
2193 si_shader_ctx
->shader
->spi_shader_z_format
= V_028710_SPI_SHADER_32_ABGR
;
2194 else if (stencil_index
>= 0)
2195 si_shader_ctx
->shader
->spi_shader_z_format
= V_028710_SPI_SHADER_32_GR
;
2197 si_shader_ctx
->shader
->spi_shader_z_format
= V_028710_SPI_SHADER_32_R
;
2199 /* Specify which components to enable */
2200 args
[0] = lp_build_const_int32(base
->gallivm
, mask
);
2204 args
[4] = uint
->zero
;
2207 lp_build_intrinsic(base
->gallivm
->builder
,
2209 LLVMVoidTypeInContext(base
->gallivm
->context
),
2212 memcpy(last_args
, args
, sizeof(args
));
2215 if (!last_args
[0]) {
2216 /* Specify which components to enable */
2217 last_args
[0] = lp_build_const_int32(base
->gallivm
, 0x0);
2219 /* Specify the target we are exporting */
2220 last_args
[3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_MRT
);
2222 /* Set COMPR flag to zero to export data as 32-bit */
2223 last_args
[4] = uint
->zero
;
2226 last_args
[5]= uint
->zero
;
2227 last_args
[6]= uint
->zero
;
2228 last_args
[7]= uint
->zero
;
2229 last_args
[8]= uint
->zero
;
2232 /* Specify whether the EXEC mask represents the valid mask */
2233 last_args
[1] = uint
->one
;
2235 /* Specify that this is the last export */
2236 last_args
[2] = lp_build_const_int32(base
->gallivm
, 1);
2238 lp_build_intrinsic(base
->gallivm
->builder
,
2240 LLVMVoidTypeInContext(base
->gallivm
->context
),
2244 static void build_tex_intrinsic(const struct lp_build_tgsi_action
* action
,
2245 struct lp_build_tgsi_context
* bld_base
,
2246 struct lp_build_emit_data
* emit_data
);
2248 static bool tgsi_is_shadow_sampler(unsigned target
)
2250 return target
== TGSI_TEXTURE_SHADOW1D
||
2251 target
== TGSI_TEXTURE_SHADOW1D_ARRAY
||
2252 target
== TGSI_TEXTURE_SHADOW2D
||
2253 target
== TGSI_TEXTURE_SHADOW2D_ARRAY
||
2254 target
== TGSI_TEXTURE_SHADOWCUBE
||
2255 target
== TGSI_TEXTURE_SHADOWCUBE_ARRAY
||
2256 target
== TGSI_TEXTURE_SHADOWRECT
;
2259 static const struct lp_build_tgsi_action tex_action
;
2261 static void tex_fetch_args(
2262 struct lp_build_tgsi_context
* bld_base
,
2263 struct lp_build_emit_data
* emit_data
)
2265 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
2266 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2267 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
2268 unsigned opcode
= inst
->Instruction
.Opcode
;
2269 unsigned target
= inst
->Texture
.Texture
;
2270 LLVMValueRef coords
[5], derivs
[6];
2271 LLVMValueRef address
[16];
2273 unsigned num_coords
= tgsi_util_get_texture_coord_dim(target
, &ref_pos
);
2276 unsigned sampler_src
;
2277 unsigned sampler_index
;
2278 unsigned num_deriv_channels
= 0;
2279 bool has_offset
= HAVE_LLVM
>= 0x0305 ? inst
->Texture
.NumOffsets
> 0 : false;
2280 LLVMValueRef res_ptr
, samp_ptr
, fmask_ptr
= NULL
;
2282 sampler_src
= emit_data
->inst
->Instruction
.NumSrcRegs
- 1;
2283 sampler_index
= emit_data
->inst
->Src
[sampler_src
].Register
.Index
;
2285 if (emit_data
->inst
->Src
[sampler_src
].Register
.Indirect
) {
2286 const struct tgsi_full_src_register
*reg
= &emit_data
->inst
->Src
[sampler_src
];
2287 LLVMValueRef ind_index
;
2289 ind_index
= get_indirect_index(si_shader_ctx
, ®
->Indirect
, reg
->Register
.Index
);
2291 res_ptr
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, SI_PARAM_RESOURCE
);
2292 res_ptr
= build_indexed_load_const(si_shader_ctx
, res_ptr
, ind_index
);
2294 samp_ptr
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, SI_PARAM_SAMPLER
);
2295 samp_ptr
= build_indexed_load_const(si_shader_ctx
, samp_ptr
, ind_index
);
2297 if (target
== TGSI_TEXTURE_2D_MSAA
||
2298 target
== TGSI_TEXTURE_2D_ARRAY_MSAA
) {
2299 ind_index
= LLVMBuildAdd(gallivm
->builder
, ind_index
,
2300 lp_build_const_int32(gallivm
,
2301 SI_FMASK_TEX_OFFSET
), "");
2302 fmask_ptr
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, SI_PARAM_RESOURCE
);
2303 fmask_ptr
= build_indexed_load_const(si_shader_ctx
, res_ptr
, ind_index
);
2306 res_ptr
= si_shader_ctx
->resources
[sampler_index
];
2307 samp_ptr
= si_shader_ctx
->samplers
[sampler_index
];
2308 fmask_ptr
= si_shader_ctx
->resources
[SI_FMASK_TEX_OFFSET
+ sampler_index
];
2311 if (target
== TGSI_TEXTURE_BUFFER
) {
2312 LLVMTypeRef i128
= LLVMIntTypeInContext(gallivm
->context
, 128);
2313 LLVMTypeRef v2i128
= LLVMVectorType(i128
, 2);
2314 LLVMTypeRef i8
= LLVMInt8TypeInContext(gallivm
->context
);
2315 LLVMTypeRef v16i8
= LLVMVectorType(i8
, 16);
2317 /* Bitcast and truncate v8i32 to v16i8. */
2318 LLVMValueRef res
= res_ptr
;
2319 res
= LLVMBuildBitCast(gallivm
->builder
, res
, v2i128
, "");
2320 res
= LLVMBuildExtractElement(gallivm
->builder
, res
, bld_base
->uint_bld
.one
, "");
2321 res
= LLVMBuildBitCast(gallivm
->builder
, res
, v16i8
, "");
2323 emit_data
->dst_type
= LLVMVectorType(bld_base
->base
.elem_type
, 4);
2324 emit_data
->args
[0] = res
;
2325 emit_data
->args
[1] = bld_base
->uint_bld
.zero
;
2326 emit_data
->args
[2] = lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, 0);
2327 emit_data
->arg_count
= 3;
2331 /* Fetch and project texture coordinates */
2332 coords
[3] = lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, TGSI_CHAN_W
);
2333 for (chan
= 0; chan
< 3; chan
++ ) {
2334 coords
[chan
] = lp_build_emit_fetch(bld_base
,
2337 if (opcode
== TGSI_OPCODE_TXP
)
2338 coords
[chan
] = lp_build_emit_llvm_binary(bld_base
,
2344 if (opcode
== TGSI_OPCODE_TXP
)
2345 coords
[3] = bld_base
->base
.one
;
2348 if (has_offset
&& opcode
!= TGSI_OPCODE_TXF
) {
2349 /* The offsets are six-bit signed integers packed like this:
2350 * X=[5:0], Y=[13:8], and Z=[21:16].
2352 LLVMValueRef offset
[3], pack
;
2354 assert(inst
->Texture
.NumOffsets
== 1);
2356 for (chan
= 0; chan
< 3; chan
++) {
2357 offset
[chan
] = lp_build_emit_fetch_texoffset(bld_base
,
2358 emit_data
->inst
, 0, chan
);
2359 offset
[chan
] = LLVMBuildAnd(gallivm
->builder
, offset
[chan
],
2360 lp_build_const_int32(gallivm
, 0x3f), "");
2362 offset
[chan
] = LLVMBuildShl(gallivm
->builder
, offset
[chan
],
2363 lp_build_const_int32(gallivm
, chan
*8), "");
2366 pack
= LLVMBuildOr(gallivm
->builder
, offset
[0], offset
[1], "");
2367 pack
= LLVMBuildOr(gallivm
->builder
, pack
, offset
[2], "");
2368 address
[count
++] = pack
;
2371 /* Pack LOD bias value */
2372 if (opcode
== TGSI_OPCODE_TXB
)
2373 address
[count
++] = coords
[3];
2374 if (opcode
== TGSI_OPCODE_TXB2
)
2375 address
[count
++] = lp_build_emit_fetch(bld_base
, inst
, 1, 0);
2377 /* Pack depth comparison value */
2378 if (tgsi_is_shadow_sampler(target
) && opcode
!= TGSI_OPCODE_LODQ
) {
2379 if (target
== TGSI_TEXTURE_SHADOWCUBE_ARRAY
) {
2380 address
[count
++] = lp_build_emit_fetch(bld_base
, inst
, 1, 0);
2382 assert(ref_pos
>= 0);
2383 address
[count
++] = coords
[ref_pos
];
2387 /* Pack user derivatives */
2388 if (opcode
== TGSI_OPCODE_TXD
) {
2389 int param
, num_src_deriv_channels
;
2392 case TGSI_TEXTURE_3D
:
2393 num_src_deriv_channels
= 3;
2394 num_deriv_channels
= 3;
2396 case TGSI_TEXTURE_2D
:
2397 case TGSI_TEXTURE_SHADOW2D
:
2398 case TGSI_TEXTURE_RECT
:
2399 case TGSI_TEXTURE_SHADOWRECT
:
2400 case TGSI_TEXTURE_2D_ARRAY
:
2401 case TGSI_TEXTURE_SHADOW2D_ARRAY
:
2402 num_src_deriv_channels
= 2;
2403 num_deriv_channels
= 2;
2405 case TGSI_TEXTURE_CUBE
:
2406 case TGSI_TEXTURE_SHADOWCUBE
:
2407 case TGSI_TEXTURE_CUBE_ARRAY
:
2408 case TGSI_TEXTURE_SHADOWCUBE_ARRAY
:
2409 /* Cube derivatives will be converted to 2D. */
2410 num_src_deriv_channels
= 3;
2411 num_deriv_channels
= 2;
2413 case TGSI_TEXTURE_1D
:
2414 case TGSI_TEXTURE_SHADOW1D
:
2415 case TGSI_TEXTURE_1D_ARRAY
:
2416 case TGSI_TEXTURE_SHADOW1D_ARRAY
:
2417 num_src_deriv_channels
= 1;
2418 num_deriv_channels
= 1;
2421 assert(0); /* no other targets are valid here */
2424 for (param
= 0; param
< 2; param
++)
2425 for (chan
= 0; chan
< num_src_deriv_channels
; chan
++)
2426 derivs
[param
* num_src_deriv_channels
+ chan
] =
2427 lp_build_emit_fetch(bld_base
, inst
, param
+1, chan
);
2430 if (target
== TGSI_TEXTURE_CUBE
||
2431 target
== TGSI_TEXTURE_CUBE_ARRAY
||
2432 target
== TGSI_TEXTURE_SHADOWCUBE
||
2433 target
== TGSI_TEXTURE_SHADOWCUBE_ARRAY
)
2434 radeon_llvm_emit_prepare_cube_coords(bld_base
, emit_data
, coords
, derivs
);
2436 if (opcode
== TGSI_OPCODE_TXD
)
2437 for (int i
= 0; i
< num_deriv_channels
* 2; i
++)
2438 address
[count
++] = derivs
[i
];
2440 /* Pack texture coordinates */
2441 address
[count
++] = coords
[0];
2443 address
[count
++] = coords
[1];
2445 address
[count
++] = coords
[2];
2447 /* Pack LOD or sample index */
2448 if (opcode
== TGSI_OPCODE_TXL
|| opcode
== TGSI_OPCODE_TXF
)
2449 address
[count
++] = coords
[3];
2450 else if (opcode
== TGSI_OPCODE_TXL2
)
2451 address
[count
++] = lp_build_emit_fetch(bld_base
, inst
, 1, 0);
2454 assert(!"Cannot handle more than 16 texture address parameters");
2458 for (chan
= 0; chan
< count
; chan
++ ) {
2459 address
[chan
] = LLVMBuildBitCast(gallivm
->builder
,
2461 LLVMInt32TypeInContext(gallivm
->context
),
2465 /* Adjust the sample index according to FMASK.
2467 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
2468 * which is the identity mapping. Each nibble says which physical sample
2469 * should be fetched to get that sample.
2471 * For example, 0x11111100 means there are only 2 samples stored and
2472 * the second sample covers 3/4 of the pixel. When reading samples 0
2473 * and 1, return physical sample 0 (determined by the first two 0s
2474 * in FMASK), otherwise return physical sample 1.
2476 * The sample index should be adjusted as follows:
2477 * sample_index = (fmask >> (sample_index * 4)) & 0xF;
2479 if (target
== TGSI_TEXTURE_2D_MSAA
||
2480 target
== TGSI_TEXTURE_2D_ARRAY_MSAA
) {
2481 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
2482 struct lp_build_emit_data txf_emit_data
= *emit_data
;
2483 LLVMValueRef txf_address
[4];
2484 unsigned txf_count
= count
;
2485 struct tgsi_full_instruction inst
= {};
2487 memcpy(txf_address
, address
, sizeof(txf_address
));
2489 if (target
== TGSI_TEXTURE_2D_MSAA
) {
2490 txf_address
[2] = bld_base
->uint_bld
.zero
;
2492 txf_address
[3] = bld_base
->uint_bld
.zero
;
2494 /* Pad to a power-of-two size. */
2495 while (txf_count
< util_next_power_of_two(txf_count
))
2496 txf_address
[txf_count
++] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
2498 /* Read FMASK using TXF. */
2499 inst
.Instruction
.Opcode
= TGSI_OPCODE_TXF
;
2500 inst
.Texture
.Texture
= target
== TGSI_TEXTURE_2D_MSAA
? TGSI_TEXTURE_2D
: TGSI_TEXTURE_2D_ARRAY
;
2501 txf_emit_data
.inst
= &inst
;
2502 txf_emit_data
.chan
= 0;
2503 txf_emit_data
.dst_type
= LLVMVectorType(
2504 LLVMInt32TypeInContext(gallivm
->context
), 4);
2505 txf_emit_data
.args
[0] = lp_build_gather_values(gallivm
, txf_address
, txf_count
);
2506 txf_emit_data
.args
[1] = fmask_ptr
;
2507 txf_emit_data
.args
[2] = lp_build_const_int32(gallivm
, inst
.Texture
.Texture
);
2508 txf_emit_data
.arg_count
= 3;
2510 build_tex_intrinsic(&tex_action
, bld_base
, &txf_emit_data
);
2512 /* Initialize some constants. */
2513 LLVMValueRef four
= LLVMConstInt(uint_bld
->elem_type
, 4, 0);
2514 LLVMValueRef F
= LLVMConstInt(uint_bld
->elem_type
, 0xF, 0);
2516 /* Apply the formula. */
2517 LLVMValueRef fmask
=
2518 LLVMBuildExtractElement(gallivm
->builder
,
2519 txf_emit_data
.output
[0],
2520 uint_bld
->zero
, "");
2522 unsigned sample_chan
= target
== TGSI_TEXTURE_2D_MSAA
? 2 : 3;
2524 LLVMValueRef sample_index4
=
2525 LLVMBuildMul(gallivm
->builder
, address
[sample_chan
], four
, "");
2527 LLVMValueRef shifted_fmask
=
2528 LLVMBuildLShr(gallivm
->builder
, fmask
, sample_index4
, "");
2530 LLVMValueRef final_sample
=
2531 LLVMBuildAnd(gallivm
->builder
, shifted_fmask
, F
, "");
2533 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
2534 * resource descriptor is 0 (invalid),
2536 LLVMValueRef fmask_desc
=
2537 LLVMBuildBitCast(gallivm
->builder
, fmask_ptr
,
2538 LLVMVectorType(uint_bld
->elem_type
, 8), "");
2540 LLVMValueRef fmask_word1
=
2541 LLVMBuildExtractElement(gallivm
->builder
, fmask_desc
,
2544 LLVMValueRef word1_is_nonzero
=
2545 LLVMBuildICmp(gallivm
->builder
, LLVMIntNE
,
2546 fmask_word1
, uint_bld
->zero
, "");
2548 /* Replace the MSAA sample index. */
2549 address
[sample_chan
] =
2550 LLVMBuildSelect(gallivm
->builder
, word1_is_nonzero
,
2551 final_sample
, address
[sample_chan
], "");
2555 emit_data
->args
[1] = res_ptr
;
2557 if (opcode
== TGSI_OPCODE_TXF
) {
2558 /* add tex offsets */
2559 if (inst
->Texture
.NumOffsets
) {
2560 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
2561 struct lp_build_tgsi_soa_context
*bld
= lp_soa_context(bld_base
);
2562 const struct tgsi_texture_offset
* off
= inst
->TexOffsets
;
2564 assert(inst
->Texture
.NumOffsets
== 1);
2567 case TGSI_TEXTURE_3D
:
2568 address
[2] = lp_build_add(uint_bld
, address
[2],
2569 bld
->immediates
[off
->Index
][off
->SwizzleZ
]);
2571 case TGSI_TEXTURE_2D
:
2572 case TGSI_TEXTURE_SHADOW2D
:
2573 case TGSI_TEXTURE_RECT
:
2574 case TGSI_TEXTURE_SHADOWRECT
:
2575 case TGSI_TEXTURE_2D_ARRAY
:
2576 case TGSI_TEXTURE_SHADOW2D_ARRAY
:
2578 lp_build_add(uint_bld
, address
[1],
2579 bld
->immediates
[off
->Index
][off
->SwizzleY
]);
2581 case TGSI_TEXTURE_1D
:
2582 case TGSI_TEXTURE_SHADOW1D
:
2583 case TGSI_TEXTURE_1D_ARRAY
:
2584 case TGSI_TEXTURE_SHADOW1D_ARRAY
:
2586 lp_build_add(uint_bld
, address
[0],
2587 bld
->immediates
[off
->Index
][off
->SwizzleX
]);
2589 /* texture offsets do not apply to other texture targets */
2593 emit_data
->args
[2] = lp_build_const_int32(gallivm
, target
);
2594 emit_data
->arg_count
= 3;
2596 emit_data
->dst_type
= LLVMVectorType(
2597 LLVMInt32TypeInContext(gallivm
->context
),
2599 } else if (opcode
== TGSI_OPCODE_TG4
||
2600 opcode
== TGSI_OPCODE_LODQ
||
2602 unsigned is_array
= target
== TGSI_TEXTURE_1D_ARRAY
||
2603 target
== TGSI_TEXTURE_SHADOW1D_ARRAY
||
2604 target
== TGSI_TEXTURE_2D_ARRAY
||
2605 target
== TGSI_TEXTURE_SHADOW2D_ARRAY
||
2606 target
== TGSI_TEXTURE_CUBE_ARRAY
||
2607 target
== TGSI_TEXTURE_SHADOWCUBE_ARRAY
;
2608 unsigned is_rect
= target
== TGSI_TEXTURE_RECT
;
2609 unsigned dmask
= 0xf;
2611 if (opcode
== TGSI_OPCODE_TG4
) {
2612 unsigned gather_comp
= 0;
2614 /* DMASK was repurposed for GATHER4. 4 components are always
2615 * returned and DMASK works like a swizzle - it selects
2616 * the component to fetch. The only valid DMASK values are
2617 * 1=red, 2=green, 4=blue, 8=alpha. (e.g. 1 returns
2618 * (red,red,red,red) etc.) The ISA document doesn't mention
2622 /* Get the component index from src1.x for Gather4. */
2623 if (!tgsi_is_shadow_sampler(target
)) {
2624 LLVMValueRef (*imms
)[4] = lp_soa_context(bld_base
)->immediates
;
2625 LLVMValueRef comp_imm
;
2626 struct tgsi_src_register src1
= inst
->Src
[1].Register
;
2628 assert(src1
.File
== TGSI_FILE_IMMEDIATE
);
2630 comp_imm
= imms
[src1
.Index
][src1
.SwizzleX
];
2631 gather_comp
= LLVMConstIntGetZExtValue(comp_imm
);
2632 gather_comp
= CLAMP(gather_comp
, 0, 3);
2635 dmask
= 1 << gather_comp
;
2638 emit_data
->args
[2] = samp_ptr
;
2639 emit_data
->args
[3] = lp_build_const_int32(gallivm
, dmask
);
2640 emit_data
->args
[4] = lp_build_const_int32(gallivm
, is_rect
); /* unorm */
2641 emit_data
->args
[5] = lp_build_const_int32(gallivm
, 0); /* r128 */
2642 emit_data
->args
[6] = lp_build_const_int32(gallivm
, is_array
); /* da */
2643 emit_data
->args
[7] = lp_build_const_int32(gallivm
, 0); /* glc */
2644 emit_data
->args
[8] = lp_build_const_int32(gallivm
, 0); /* slc */
2645 emit_data
->args
[9] = lp_build_const_int32(gallivm
, 0); /* tfe */
2646 emit_data
->args
[10] = lp_build_const_int32(gallivm
, 0); /* lwe */
2648 emit_data
->arg_count
= 11;
2650 emit_data
->dst_type
= LLVMVectorType(
2651 LLVMFloatTypeInContext(gallivm
->context
),
2654 emit_data
->args
[2] = samp_ptr
;
2655 emit_data
->args
[3] = lp_build_const_int32(gallivm
, target
);
2656 emit_data
->arg_count
= 4;
2658 emit_data
->dst_type
= LLVMVectorType(
2659 LLVMFloatTypeInContext(gallivm
->context
),
2663 /* The fetch opcode has been converted to a 2D array fetch.
2664 * This simplifies the LLVM backend. */
2665 if (target
== TGSI_TEXTURE_CUBE_ARRAY
)
2666 target
= TGSI_TEXTURE_2D_ARRAY
;
2667 else if (target
== TGSI_TEXTURE_SHADOWCUBE_ARRAY
)
2668 target
= TGSI_TEXTURE_SHADOW2D_ARRAY
;
2670 /* Pad to power of two vector */
2671 while (count
< util_next_power_of_two(count
))
2672 address
[count
++] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
2674 emit_data
->args
[0] = lp_build_gather_values(gallivm
, address
, count
);
2677 static void build_tex_intrinsic(const struct lp_build_tgsi_action
* action
,
2678 struct lp_build_tgsi_context
* bld_base
,
2679 struct lp_build_emit_data
* emit_data
)
2681 struct lp_build_context
* base
= &bld_base
->base
;
2682 unsigned opcode
= emit_data
->inst
->Instruction
.Opcode
;
2683 unsigned target
= emit_data
->inst
->Texture
.Texture
;
2684 char intr_name
[127];
2685 bool has_offset
= HAVE_LLVM
>= 0x0305 ?
2686 emit_data
->inst
->Texture
.NumOffsets
> 0 : false;
2688 if (target
== TGSI_TEXTURE_BUFFER
) {
2689 emit_data
->output
[emit_data
->chan
] = lp_build_intrinsic(
2690 base
->gallivm
->builder
,
2691 "llvm.SI.vs.load.input", emit_data
->dst_type
,
2692 emit_data
->args
, emit_data
->arg_count
,
2693 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
2697 if (opcode
== TGSI_OPCODE_TG4
||
2698 opcode
== TGSI_OPCODE_LODQ
||
2699 (opcode
!= TGSI_OPCODE_TXF
&& has_offset
)) {
2700 bool is_shadow
= tgsi_is_shadow_sampler(target
);
2701 const char *name
= "llvm.SI.image.sample";
2702 const char *infix
= "";
2705 case TGSI_OPCODE_TEX
:
2706 case TGSI_OPCODE_TEX2
:
2707 case TGSI_OPCODE_TXP
:
2709 case TGSI_OPCODE_TXB
:
2710 case TGSI_OPCODE_TXB2
:
2713 case TGSI_OPCODE_TXL
:
2714 case TGSI_OPCODE_TXL2
:
2717 case TGSI_OPCODE_TXD
:
2720 case TGSI_OPCODE_TG4
:
2721 name
= "llvm.SI.gather4";
2723 case TGSI_OPCODE_LODQ
:
2724 name
= "llvm.SI.getlod";
2733 /* Add the type and suffixes .c, .o if needed. */
2734 sprintf(intr_name
, "%s%s%s%s.v%ui32", name
,
2735 is_shadow
? ".c" : "", infix
, has_offset
? ".o" : "",
2736 LLVMGetVectorSize(LLVMTypeOf(emit_data
->args
[0])));
2738 emit_data
->output
[emit_data
->chan
] = lp_build_intrinsic(
2739 base
->gallivm
->builder
, intr_name
, emit_data
->dst_type
,
2740 emit_data
->args
, emit_data
->arg_count
,
2741 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
2743 LLVMTypeRef i8
, v16i8
, v32i8
;
2747 case TGSI_OPCODE_TEX
:
2748 case TGSI_OPCODE_TEX2
:
2749 case TGSI_OPCODE_TXP
:
2750 name
= "llvm.SI.sample";
2752 case TGSI_OPCODE_TXB
:
2753 case TGSI_OPCODE_TXB2
:
2754 name
= "llvm.SI.sampleb";
2756 case TGSI_OPCODE_TXD
:
2757 name
= "llvm.SI.sampled";
2759 case TGSI_OPCODE_TXF
:
2760 name
= "llvm.SI.imageload";
2762 case TGSI_OPCODE_TXL
:
2763 case TGSI_OPCODE_TXL2
:
2764 name
= "llvm.SI.samplel";
2771 i8
= LLVMInt8TypeInContext(base
->gallivm
->context
);
2772 v16i8
= LLVMVectorType(i8
, 16);
2773 v32i8
= LLVMVectorType(i8
, 32);
2775 emit_data
->args
[1] = LLVMBuildBitCast(base
->gallivm
->builder
,
2776 emit_data
->args
[1], v32i8
, "");
2777 if (opcode
!= TGSI_OPCODE_TXF
) {
2778 emit_data
->args
[2] = LLVMBuildBitCast(base
->gallivm
->builder
,
2779 emit_data
->args
[2], v16i8
, "");
2782 sprintf(intr_name
, "%s.v%ui32", name
,
2783 LLVMGetVectorSize(LLVMTypeOf(emit_data
->args
[0])));
2785 emit_data
->output
[emit_data
->chan
] = lp_build_intrinsic(
2786 base
->gallivm
->builder
, intr_name
, emit_data
->dst_type
,
2787 emit_data
->args
, emit_data
->arg_count
,
2788 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
2792 static void txq_fetch_args(
2793 struct lp_build_tgsi_context
* bld_base
,
2794 struct lp_build_emit_data
* emit_data
)
2796 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
2797 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
2798 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2799 LLVMBuilderRef builder
= gallivm
->builder
;
2800 unsigned target
= inst
->Texture
.Texture
;
2801 LLVMValueRef res_ptr
;
2803 if (inst
->Src
[1].Register
.Indirect
) {
2804 const struct tgsi_full_src_register
*reg
= &inst
->Src
[1];
2805 LLVMValueRef ind_index
;
2807 ind_index
= get_indirect_index(si_shader_ctx
, ®
->Indirect
, reg
->Register
.Index
);
2809 res_ptr
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, SI_PARAM_RESOURCE
);
2810 res_ptr
= build_indexed_load_const(si_shader_ctx
, res_ptr
,
2813 res_ptr
= si_shader_ctx
->resources
[inst
->Src
[1].Register
.Index
];
2815 if (target
== TGSI_TEXTURE_BUFFER
) {
2816 LLVMTypeRef i32
= LLVMInt32TypeInContext(gallivm
->context
);
2817 LLVMTypeRef v8i32
= LLVMVectorType(i32
, 8);
2819 /* Read the size from the buffer descriptor directly. */
2820 LLVMValueRef res
= LLVMBuildBitCast(builder
, res_ptr
, v8i32
, "");
2821 LLVMValueRef size
= LLVMBuildExtractElement(builder
, res
,
2822 lp_build_const_int32(gallivm
, 6), "");
2824 if (si_shader_ctx
->screen
->b
.chip_class
>= VI
) {
2825 /* On VI, the descriptor contains the size in bytes,
2826 * but TXQ must return the size in elements.
2827 * The stride is always non-zero for resources using TXQ.
2829 LLVMValueRef stride
=
2830 LLVMBuildExtractElement(builder
, res
,
2831 lp_build_const_int32(gallivm
, 5), "");
2832 stride
= LLVMBuildLShr(builder
, stride
,
2833 lp_build_const_int32(gallivm
, 16), "");
2834 stride
= LLVMBuildAnd(builder
, stride
,
2835 lp_build_const_int32(gallivm
, 0x3FFF), "");
2837 size
= LLVMBuildUDiv(builder
, size
, stride
, "");
2840 emit_data
->args
[0] = size
;
2845 emit_data
->args
[0] = lp_build_emit_fetch(bld_base
, inst
, 0, TGSI_CHAN_X
);
2848 emit_data
->args
[1] = res_ptr
;
2850 /* Texture target */
2851 if (target
== TGSI_TEXTURE_CUBE_ARRAY
||
2852 target
== TGSI_TEXTURE_SHADOWCUBE_ARRAY
)
2853 target
= TGSI_TEXTURE_2D_ARRAY
;
2855 emit_data
->args
[2] = lp_build_const_int32(bld_base
->base
.gallivm
,
2858 emit_data
->arg_count
= 3;
2860 emit_data
->dst_type
= LLVMVectorType(
2861 LLVMInt32TypeInContext(bld_base
->base
.gallivm
->context
),
2865 static void build_txq_intrinsic(const struct lp_build_tgsi_action
* action
,
2866 struct lp_build_tgsi_context
* bld_base
,
2867 struct lp_build_emit_data
* emit_data
)
2869 unsigned target
= emit_data
->inst
->Texture
.Texture
;
2871 if (target
== TGSI_TEXTURE_BUFFER
) {
2872 /* Just return the buffer size. */
2873 emit_data
->output
[emit_data
->chan
] = emit_data
->args
[0];
2877 build_tgsi_intrinsic_nomem(action
, bld_base
, emit_data
);
2879 /* Divide the number of layers by 6 to get the number of cubes. */
2880 if (target
== TGSI_TEXTURE_CUBE_ARRAY
||
2881 target
== TGSI_TEXTURE_SHADOWCUBE_ARRAY
) {
2882 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
2883 LLVMValueRef two
= lp_build_const_int32(bld_base
->base
.gallivm
, 2);
2884 LLVMValueRef six
= lp_build_const_int32(bld_base
->base
.gallivm
, 6);
2886 LLVMValueRef v4
= emit_data
->output
[emit_data
->chan
];
2887 LLVMValueRef z
= LLVMBuildExtractElement(builder
, v4
, two
, "");
2888 z
= LLVMBuildSDiv(builder
, z
, six
, "");
2890 emit_data
->output
[emit_data
->chan
] =
2891 LLVMBuildInsertElement(builder
, v4
, z
, two
, "");
2896 * SI implements derivatives using the local data store (LDS)
2897 * All writes to the LDS happen in all executing threads at
2898 * the same time. TID is the Thread ID for the current
2899 * thread and is a value between 0 and 63, representing
2900 * the thread's position in the wavefront.
2902 * For the pixel shader threads are grouped into quads of four pixels.
2903 * The TIDs of the pixels of a quad are:
2911 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
2912 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
2913 * the current pixel's column, and masking with 0xfffffffe yields the TID
2914 * of the left pixel of the current pixel's row.
2916 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
2917 * adding 2 yields the TID of the pixel below the top pixel.
2919 /* masks for thread ID. */
2920 #define TID_MASK_TOP_LEFT 0xfffffffc
2921 #define TID_MASK_TOP 0xfffffffd
2922 #define TID_MASK_LEFT 0xfffffffe
2924 static void si_llvm_emit_ddxy(
2925 const struct lp_build_tgsi_action
* action
,
2926 struct lp_build_tgsi_context
* bld_base
,
2927 struct lp_build_emit_data
* emit_data
)
2929 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
2930 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2931 struct lp_build_context
* base
= &bld_base
->base
;
2932 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
2933 unsigned opcode
= inst
->Instruction
.Opcode
;
2934 LLVMValueRef indices
[2];
2935 LLVMValueRef store_ptr
, load_ptr0
, load_ptr1
;
2936 LLVMValueRef tl
, trbl
, result
[4];
2938 unsigned swizzle
[4];
2943 i32
= LLVMInt32TypeInContext(gallivm
->context
);
2945 indices
[0] = bld_base
->uint_bld
.zero
;
2946 indices
[1] = lp_build_intrinsic(gallivm
->builder
, "llvm.SI.tid", i32
,
2947 NULL
, 0, LLVMReadNoneAttribute
);
2948 store_ptr
= LLVMBuildGEP(gallivm
->builder
, si_shader_ctx
->lds
,
2951 if (opcode
== TGSI_OPCODE_DDX_FINE
)
2952 mask
= TID_MASK_LEFT
;
2953 else if (opcode
== TGSI_OPCODE_DDY_FINE
)
2954 mask
= TID_MASK_TOP
;
2956 mask
= TID_MASK_TOP_LEFT
;
2958 indices
[1] = LLVMBuildAnd(gallivm
->builder
, indices
[1],
2959 lp_build_const_int32(gallivm
, mask
), "");
2960 load_ptr0
= LLVMBuildGEP(gallivm
->builder
, si_shader_ctx
->lds
,
2963 /* for DDX we want to next X pixel, DDY next Y pixel. */
2964 idx
= (opcode
== TGSI_OPCODE_DDX
|| opcode
== TGSI_OPCODE_DDX_FINE
) ? 1 : 2;
2965 indices
[1] = LLVMBuildAdd(gallivm
->builder
, indices
[1],
2966 lp_build_const_int32(gallivm
, idx
), "");
2967 load_ptr1
= LLVMBuildGEP(gallivm
->builder
, si_shader_ctx
->lds
,
2970 for (c
= 0; c
< 4; ++c
) {
2973 swizzle
[c
] = tgsi_util_get_full_src_register_swizzle(&inst
->Src
[0], c
);
2974 for (i
= 0; i
< c
; ++i
) {
2975 if (swizzle
[i
] == swizzle
[c
]) {
2976 result
[c
] = result
[i
];
2983 LLVMBuildStore(gallivm
->builder
,
2984 LLVMBuildBitCast(gallivm
->builder
,
2985 lp_build_emit_fetch(bld_base
, inst
, 0, c
),
2989 tl
= LLVMBuildLoad(gallivm
->builder
, load_ptr0
, "");
2990 tl
= LLVMBuildBitCast(gallivm
->builder
, tl
, base
->elem_type
, "");
2992 trbl
= LLVMBuildLoad(gallivm
->builder
, load_ptr1
, "");
2993 trbl
= LLVMBuildBitCast(gallivm
->builder
, trbl
, base
->elem_type
, "");
2995 result
[c
] = LLVMBuildFSub(gallivm
->builder
, trbl
, tl
, "");
2998 emit_data
->output
[0] = lp_build_gather_values(gallivm
, result
, 4);
3002 * this takes an I,J coordinate pair,
3003 * and works out the X and Y derivatives.
3004 * it returns DDX(I), DDX(J), DDY(I), DDY(J).
3006 static LLVMValueRef
si_llvm_emit_ddxy_interp(
3007 struct lp_build_tgsi_context
*bld_base
,
3008 LLVMValueRef interp_ij
)
3010 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
3011 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3012 struct lp_build_context
*base
= &bld_base
->base
;
3013 LLVMValueRef indices
[2];
3014 LLVMValueRef store_ptr
, load_ptr_x
, load_ptr_y
, load_ptr_ddx
, load_ptr_ddy
, temp
, temp2
;
3015 LLVMValueRef tl
, tr
, bl
, result
[4];
3019 i32
= LLVMInt32TypeInContext(gallivm
->context
);
3021 indices
[0] = bld_base
->uint_bld
.zero
;
3022 indices
[1] = lp_build_intrinsic(gallivm
->builder
, "llvm.SI.tid", i32
,
3023 NULL
, 0, LLVMReadNoneAttribute
);
3024 store_ptr
= LLVMBuildGEP(gallivm
->builder
, si_shader_ctx
->lds
,
3027 temp
= LLVMBuildAnd(gallivm
->builder
, indices
[1],
3028 lp_build_const_int32(gallivm
, TID_MASK_LEFT
), "");
3030 temp2
= LLVMBuildAnd(gallivm
->builder
, indices
[1],
3031 lp_build_const_int32(gallivm
, TID_MASK_TOP
), "");
3034 load_ptr_x
= LLVMBuildGEP(gallivm
->builder
, si_shader_ctx
->lds
,
3038 load_ptr_y
= LLVMBuildGEP(gallivm
->builder
, si_shader_ctx
->lds
,
3041 indices
[1] = LLVMBuildAdd(gallivm
->builder
, temp
,
3042 lp_build_const_int32(gallivm
, 1), "");
3043 load_ptr_ddx
= LLVMBuildGEP(gallivm
->builder
, si_shader_ctx
->lds
,
3046 indices
[1] = LLVMBuildAdd(gallivm
->builder
, temp2
,
3047 lp_build_const_int32(gallivm
, 2), "");
3048 load_ptr_ddy
= LLVMBuildGEP(gallivm
->builder
, si_shader_ctx
->lds
,
3051 for (c
= 0; c
< 2; ++c
) {
3052 LLVMValueRef store_val
;
3053 LLVMValueRef c_ll
= lp_build_const_int32(gallivm
, c
);
3055 store_val
= LLVMBuildExtractElement(gallivm
->builder
,
3056 interp_ij
, c_ll
, "");
3057 LLVMBuildStore(gallivm
->builder
,
3061 tl
= LLVMBuildLoad(gallivm
->builder
, load_ptr_x
, "");
3062 tl
= LLVMBuildBitCast(gallivm
->builder
, tl
, base
->elem_type
, "");
3064 tr
= LLVMBuildLoad(gallivm
->builder
, load_ptr_ddx
, "");
3065 tr
= LLVMBuildBitCast(gallivm
->builder
, tr
, base
->elem_type
, "");
3067 result
[c
] = LLVMBuildFSub(gallivm
->builder
, tr
, tl
, "");
3069 tl
= LLVMBuildLoad(gallivm
->builder
, load_ptr_y
, "");
3070 tl
= LLVMBuildBitCast(gallivm
->builder
, tl
, base
->elem_type
, "");
3072 bl
= LLVMBuildLoad(gallivm
->builder
, load_ptr_ddy
, "");
3073 bl
= LLVMBuildBitCast(gallivm
->builder
, bl
, base
->elem_type
, "");
3075 result
[c
+ 2] = LLVMBuildFSub(gallivm
->builder
, bl
, tl
, "");
3078 return lp_build_gather_values(gallivm
, result
, 4);
3081 static void interp_fetch_args(
3082 struct lp_build_tgsi_context
*bld_base
,
3083 struct lp_build_emit_data
*emit_data
)
3085 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
3086 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3087 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3089 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_OFFSET
) {
3090 /* offset is in second src, first two channels */
3091 emit_data
->args
[0] = lp_build_emit_fetch(bld_base
,
3094 emit_data
->args
[1] = lp_build_emit_fetch(bld_base
,
3097 emit_data
->arg_count
= 2;
3098 } else if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_SAMPLE
) {
3099 LLVMValueRef sample_position
;
3100 LLVMValueRef sample_id
;
3101 LLVMValueRef halfval
= lp_build_const_float(gallivm
, 0.5f
);
3103 /* fetch sample ID, then fetch its sample position,
3104 * and place into first two channels.
3106 sample_id
= lp_build_emit_fetch(bld_base
,
3107 emit_data
->inst
, 1, 0);
3108 sample_id
= LLVMBuildBitCast(gallivm
->builder
, sample_id
,
3109 LLVMInt32TypeInContext(gallivm
->context
),
3111 sample_position
= load_sample_position(&si_shader_ctx
->radeon_bld
, sample_id
);
3113 emit_data
->args
[0] = LLVMBuildExtractElement(gallivm
->builder
,
3115 lp_build_const_int32(gallivm
, 0), "");
3117 emit_data
->args
[0] = LLVMBuildFSub(gallivm
->builder
, emit_data
->args
[0], halfval
, "");
3118 emit_data
->args
[1] = LLVMBuildExtractElement(gallivm
->builder
,
3120 lp_build_const_int32(gallivm
, 1), "");
3121 emit_data
->args
[1] = LLVMBuildFSub(gallivm
->builder
, emit_data
->args
[1], halfval
, "");
3122 emit_data
->arg_count
= 2;
3126 static void build_interp_intrinsic(const struct lp_build_tgsi_action
*action
,
3127 struct lp_build_tgsi_context
*bld_base
,
3128 struct lp_build_emit_data
*emit_data
)
3130 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
3131 struct si_shader
*shader
= si_shader_ctx
->shader
;
3132 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3133 LLVMValueRef interp_param
;
3134 const struct tgsi_full_instruction
*inst
= emit_data
->inst
;
3135 const char *intr_name
;
3139 LLVMValueRef attr_number
;
3140 LLVMTypeRef input_type
= LLVMFloatTypeInContext(gallivm
->context
);
3141 LLVMValueRef params
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, SI_PARAM_PRIM_MASK
);
3142 int interp_param_idx
;
3145 assert(inst
->Src
[0].Register
.File
== TGSI_FILE_INPUT
);
3146 input_index
= inst
->Src
[0].Register
.Index
;
3148 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_OFFSET
||
3149 inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_SAMPLE
)
3150 location
= TGSI_INTERPOLATE_LOC_CENTER
;
3152 location
= TGSI_INTERPOLATE_LOC_CENTROID
;
3154 interp_param_idx
= lookup_interp_param_index(shader
->ps_input_interpolate
[input_index
],
3156 if (interp_param_idx
== -1)
3158 else if (interp_param_idx
)
3159 interp_param
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, interp_param_idx
);
3161 interp_param
= NULL
;
3163 attr_number
= lp_build_const_int32(gallivm
,
3164 shader
->ps_input_param_offset
[input_index
]);
3166 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_OFFSET
||
3167 inst
->Instruction
.Opcode
== TGSI_OPCODE_INTERP_SAMPLE
) {
3168 LLVMValueRef ij_out
[2];
3169 LLVMValueRef ddxy_out
= si_llvm_emit_ddxy_interp(bld_base
, interp_param
);
3172 * take the I then J parameters, and the DDX/Y for it, and
3173 * calculate the IJ inputs for the interpolator.
3174 * temp1 = ddx * offset/sample.x + I;
3175 * interp_param.I = ddy * offset/sample.y + temp1;
3176 * temp1 = ddx * offset/sample.x + J;
3177 * interp_param.J = ddy * offset/sample.y + temp1;
3179 for (i
= 0; i
< 2; i
++) {
3180 LLVMValueRef ix_ll
= lp_build_const_int32(gallivm
, i
);
3181 LLVMValueRef iy_ll
= lp_build_const_int32(gallivm
, i
+ 2);
3182 LLVMValueRef ddx_el
= LLVMBuildExtractElement(gallivm
->builder
,
3183 ddxy_out
, ix_ll
, "");
3184 LLVMValueRef ddy_el
= LLVMBuildExtractElement(gallivm
->builder
,
3185 ddxy_out
, iy_ll
, "");
3186 LLVMValueRef interp_el
= LLVMBuildExtractElement(gallivm
->builder
,
3187 interp_param
, ix_ll
, "");
3188 LLVMValueRef temp1
, temp2
;
3190 interp_el
= LLVMBuildBitCast(gallivm
->builder
, interp_el
,
3191 LLVMFloatTypeInContext(gallivm
->context
), "");
3193 temp1
= LLVMBuildFMul(gallivm
->builder
, ddx_el
, emit_data
->args
[0], "");
3195 temp1
= LLVMBuildFAdd(gallivm
->builder
, temp1
, interp_el
, "");
3197 temp2
= LLVMBuildFMul(gallivm
->builder
, ddy_el
, emit_data
->args
[1], "");
3199 temp2
= LLVMBuildFAdd(gallivm
->builder
, temp2
, temp1
, "");
3201 ij_out
[i
] = LLVMBuildBitCast(gallivm
->builder
,
3203 LLVMIntTypeInContext(gallivm
->context
, 32), "");
3205 interp_param
= lp_build_gather_values(bld_base
->base
.gallivm
, ij_out
, 2);
3208 intr_name
= interp_param
? "llvm.SI.fs.interp" : "llvm.SI.fs.constant";
3209 for (chan
= 0; chan
< 2; chan
++) {
3210 LLVMValueRef args
[4];
3211 LLVMValueRef llvm_chan
;
3214 schan
= tgsi_util_get_full_src_register_swizzle(&inst
->Src
[0], chan
);
3215 llvm_chan
= lp_build_const_int32(gallivm
, schan
);
3217 args
[0] = llvm_chan
;
3218 args
[1] = attr_number
;
3220 args
[3] = interp_param
;
3222 emit_data
->output
[chan
] =
3223 lp_build_intrinsic(gallivm
->builder
, intr_name
,
3224 input_type
, args
, args
[3] ? 4 : 3,
3225 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
3229 static unsigned si_llvm_get_stream(struct lp_build_tgsi_context
*bld_base
,
3230 struct lp_build_emit_data
*emit_data
)
3232 LLVMValueRef (*imms
)[4] = lp_soa_context(bld_base
)->immediates
;
3233 struct tgsi_src_register src0
= emit_data
->inst
->Src
[0].Register
;
3236 assert(src0
.File
== TGSI_FILE_IMMEDIATE
);
3238 stream
= LLVMConstIntGetZExtValue(imms
[src0
.Index
][src0
.SwizzleX
]) & 0x3;
3242 /* Emit one vertex from the geometry shader */
3243 static void si_llvm_emit_vertex(
3244 const struct lp_build_tgsi_action
*action
,
3245 struct lp_build_tgsi_context
*bld_base
,
3246 struct lp_build_emit_data
*emit_data
)
3248 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
3249 struct lp_build_context
*uint
= &bld_base
->uint_bld
;
3250 struct si_shader
*shader
= si_shader_ctx
->shader
;
3251 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
3252 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3253 LLVMTypeRef i32
= LLVMInt32TypeInContext(gallivm
->context
);
3254 LLVMValueRef soffset
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
,
3255 SI_PARAM_GS2VS_OFFSET
);
3256 LLVMValueRef gs_next_vertex
;
3257 LLVMValueRef can_emit
, kill
;
3258 LLVMValueRef args
[2];
3263 stream
= si_llvm_get_stream(bld_base
, emit_data
);
3265 /* Write vertex attribute values to GSVS ring */
3266 gs_next_vertex
= LLVMBuildLoad(gallivm
->builder
,
3267 si_shader_ctx
->gs_next_vertex
[stream
],
3270 /* If this thread has already emitted the declared maximum number of
3271 * vertices, kill it: excessive vertex emissions are not supposed to
3272 * have any effect, and GS threads have no externally observable
3273 * effects other than emitting vertices.
3275 can_emit
= LLVMBuildICmp(gallivm
->builder
, LLVMIntULE
, gs_next_vertex
,
3276 lp_build_const_int32(gallivm
,
3277 shader
->selector
->gs_max_out_vertices
), "");
3278 kill
= lp_build_select(&bld_base
->base
, can_emit
,
3279 lp_build_const_float(gallivm
, 1.0f
),
3280 lp_build_const_float(gallivm
, -1.0f
));
3282 lp_build_intrinsic(gallivm
->builder
, "llvm.AMDGPU.kill",
3283 LLVMVoidTypeInContext(gallivm
->context
), &kill
, 1, 0);
3285 for (i
= 0; i
< info
->num_outputs
; i
++) {
3286 LLVMValueRef
*out_ptr
=
3287 si_shader_ctx
->radeon_bld
.soa
.outputs
[i
];
3289 for (chan
= 0; chan
< 4; chan
++) {
3290 LLVMValueRef out_val
= LLVMBuildLoad(gallivm
->builder
, out_ptr
[chan
], "");
3291 LLVMValueRef voffset
=
3292 lp_build_const_int32(gallivm
, (i
* 4 + chan
) *
3293 shader
->selector
->gs_max_out_vertices
);
3295 voffset
= lp_build_add(uint
, voffset
, gs_next_vertex
);
3296 voffset
= lp_build_mul_imm(uint
, voffset
, 4);
3298 out_val
= LLVMBuildBitCast(gallivm
->builder
, out_val
, i32
, "");
3300 build_tbuffer_store(si_shader_ctx
,
3301 si_shader_ctx
->gsvs_ring
[stream
],
3303 voffset
, soffset
, 0,
3304 V_008F0C_BUF_DATA_FORMAT_32
,
3305 V_008F0C_BUF_NUM_FORMAT_UINT
,
3309 gs_next_vertex
= lp_build_add(uint
, gs_next_vertex
,
3310 lp_build_const_int32(gallivm
, 1));
3312 LLVMBuildStore(gallivm
->builder
, gs_next_vertex
, si_shader_ctx
->gs_next_vertex
[stream
]);
3314 /* Signal vertex emission */
3315 args
[0] = lp_build_const_int32(gallivm
, SENDMSG_GS_OP_EMIT
| SENDMSG_GS
| (stream
<< 8));
3316 args
[1] = LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, SI_PARAM_GS_WAVE_ID
);
3317 lp_build_intrinsic(gallivm
->builder
, "llvm.SI.sendmsg",
3318 LLVMVoidTypeInContext(gallivm
->context
), args
, 2,
3319 LLVMNoUnwindAttribute
);
3322 /* Cut one primitive from the geometry shader */
3323 static void si_llvm_emit_primitive(
3324 const struct lp_build_tgsi_action
*action
,
3325 struct lp_build_tgsi_context
*bld_base
,
3326 struct lp_build_emit_data
*emit_data
)
3328 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
3329 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3330 LLVMValueRef args
[2];
3333 /* Signal primitive cut */
3334 stream
= si_llvm_get_stream(bld_base
, emit_data
);
3335 args
[0] = lp_build_const_int32(gallivm
, SENDMSG_GS_OP_CUT
| SENDMSG_GS
| (stream
<< 8));
3336 args
[1] = LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, SI_PARAM_GS_WAVE_ID
);
3337 lp_build_intrinsic(gallivm
->builder
, "llvm.SI.sendmsg",
3338 LLVMVoidTypeInContext(gallivm
->context
), args
, 2,
3339 LLVMNoUnwindAttribute
);
3342 static void si_llvm_emit_barrier(const struct lp_build_tgsi_action
*action
,
3343 struct lp_build_tgsi_context
*bld_base
,
3344 struct lp_build_emit_data
*emit_data
)
3346 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3348 lp_build_intrinsic(gallivm
->builder
, "llvm.AMDGPU.barrier.local",
3349 LLVMVoidTypeInContext(gallivm
->context
), NULL
, 0,
3350 LLVMNoUnwindAttribute
);
3353 static const struct lp_build_tgsi_action tex_action
= {
3354 .fetch_args
= tex_fetch_args
,
3355 .emit
= build_tex_intrinsic
,
3358 static const struct lp_build_tgsi_action txq_action
= {
3359 .fetch_args
= txq_fetch_args
,
3360 .emit
= build_txq_intrinsic
,
3361 .intr_name
= "llvm.SI.resinfo"
3364 static const struct lp_build_tgsi_action interp_action
= {
3365 .fetch_args
= interp_fetch_args
,
3366 .emit
= build_interp_intrinsic
,
3369 static void create_meta_data(struct si_shader_context
*si_shader_ctx
)
3371 struct gallivm_state
*gallivm
= si_shader_ctx
->radeon_bld
.soa
.bld_base
.base
.gallivm
;
3372 LLVMValueRef args
[3];
3374 args
[0] = LLVMMDStringInContext(gallivm
->context
, "const", 5);
3376 args
[2] = lp_build_const_int32(gallivm
, 1);
3378 si_shader_ctx
->const_md
= LLVMMDNodeInContext(gallivm
->context
, args
, 3);
3381 static LLVMTypeRef
const_array(LLVMTypeRef elem_type
, int num_elements
)
3383 return LLVMPointerType(LLVMArrayType(elem_type
, num_elements
),
3387 static void declare_streamout_params(struct si_shader_context
*si_shader_ctx
,
3388 struct pipe_stream_output_info
*so
,
3389 LLVMTypeRef
*params
, LLVMTypeRef i32
,
3390 unsigned *num_params
)
3394 /* Streamout SGPRs. */
3395 if (so
->num_outputs
) {
3396 params
[si_shader_ctx
->param_streamout_config
= (*num_params
)++] = i32
;
3397 params
[si_shader_ctx
->param_streamout_write_index
= (*num_params
)++] = i32
;
3399 /* A streamout buffer offset is loaded if the stride is non-zero. */
3400 for (i
= 0; i
< 4; i
++) {
3404 params
[si_shader_ctx
->param_streamout_offset
[i
] = (*num_params
)++] = i32
;
3408 static void create_function(struct si_shader_context
*si_shader_ctx
)
3410 struct lp_build_tgsi_context
*bld_base
= &si_shader_ctx
->radeon_bld
.soa
.bld_base
;
3411 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
3412 struct si_shader
*shader
= si_shader_ctx
->shader
;
3413 LLVMTypeRef params
[SI_NUM_PARAMS
], f32
, i8
, i32
, v2i32
, v3i32
, v16i8
, v4i32
, v8i32
;
3414 unsigned i
, last_array_pointer
, last_sgpr
, num_params
;
3416 i8
= LLVMInt8TypeInContext(gallivm
->context
);
3417 i32
= LLVMInt32TypeInContext(gallivm
->context
);
3418 f32
= LLVMFloatTypeInContext(gallivm
->context
);
3419 v2i32
= LLVMVectorType(i32
, 2);
3420 v3i32
= LLVMVectorType(i32
, 3);
3421 v4i32
= LLVMVectorType(i32
, 4);
3422 v8i32
= LLVMVectorType(i32
, 8);
3423 v16i8
= LLVMVectorType(i8
, 16);
3425 params
[SI_PARAM_RW_BUFFERS
] = const_array(v16i8
, SI_NUM_RW_BUFFERS
);
3426 params
[SI_PARAM_CONST
] = const_array(v16i8
, SI_NUM_CONST_BUFFERS
);
3427 params
[SI_PARAM_SAMPLER
] = const_array(v4i32
, SI_NUM_SAMPLER_STATES
);
3428 params
[SI_PARAM_RESOURCE
] = const_array(v8i32
, SI_NUM_SAMPLER_VIEWS
);
3429 last_array_pointer
= SI_PARAM_RESOURCE
;
3431 switch (si_shader_ctx
->type
) {
3432 case TGSI_PROCESSOR_VERTEX
:
3433 params
[SI_PARAM_VERTEX_BUFFER
] = const_array(v16i8
, SI_NUM_VERTEX_BUFFERS
);
3434 last_array_pointer
= SI_PARAM_VERTEX_BUFFER
;
3435 params
[SI_PARAM_BASE_VERTEX
] = i32
;
3436 params
[SI_PARAM_START_INSTANCE
] = i32
;
3437 num_params
= SI_PARAM_START_INSTANCE
+1;
3439 if (shader
->key
.vs
.as_es
) {
3440 params
[si_shader_ctx
->param_es2gs_offset
= num_params
++] = i32
;
3441 } else if (shader
->key
.vs
.as_ls
) {
3442 params
[SI_PARAM_LS_OUT_LAYOUT
] = i32
;
3443 num_params
= SI_PARAM_LS_OUT_LAYOUT
+1;
3445 if (shader
->is_gs_copy_shader
) {
3446 last_array_pointer
= SI_PARAM_CONST
;
3447 num_params
= SI_PARAM_CONST
+1;
3450 /* The locations of the other parameters are assigned dynamically. */
3451 declare_streamout_params(si_shader_ctx
, &shader
->selector
->so
,
3452 params
, i32
, &num_params
);
3455 last_sgpr
= num_params
-1;
3458 params
[si_shader_ctx
->param_vertex_id
= num_params
++] = i32
;
3459 params
[si_shader_ctx
->param_rel_auto_id
= num_params
++] = i32
;
3460 params
[si_shader_ctx
->param_vs_prim_id
= num_params
++] = i32
;
3461 params
[si_shader_ctx
->param_instance_id
= num_params
++] = i32
;
3464 case TGSI_PROCESSOR_TESS_CTRL
:
3465 params
[SI_PARAM_TCS_OUT_OFFSETS
] = i32
;
3466 params
[SI_PARAM_TCS_OUT_LAYOUT
] = i32
;
3467 params
[SI_PARAM_TCS_IN_LAYOUT
] = i32
;
3468 params
[SI_PARAM_TESS_FACTOR_OFFSET
] = i32
;
3469 last_sgpr
= SI_PARAM_TESS_FACTOR_OFFSET
;
3472 params
[SI_PARAM_PATCH_ID
] = i32
;
3473 params
[SI_PARAM_REL_IDS
] = i32
;
3474 num_params
= SI_PARAM_REL_IDS
+1;
3477 case TGSI_PROCESSOR_TESS_EVAL
:
3478 params
[SI_PARAM_TCS_OUT_OFFSETS
] = i32
;
3479 params
[SI_PARAM_TCS_OUT_LAYOUT
] = i32
;
3480 num_params
= SI_PARAM_TCS_OUT_LAYOUT
+1;
3482 if (shader
->key
.tes
.as_es
) {
3483 params
[si_shader_ctx
->param_es2gs_offset
= num_params
++] = i32
;
3485 declare_streamout_params(si_shader_ctx
, &shader
->selector
->so
,
3486 params
, i32
, &num_params
);
3488 last_sgpr
= num_params
- 1;
3491 params
[si_shader_ctx
->param_tes_u
= num_params
++] = f32
;
3492 params
[si_shader_ctx
->param_tes_v
= num_params
++] = f32
;
3493 params
[si_shader_ctx
->param_tes_rel_patch_id
= num_params
++] = i32
;
3494 params
[si_shader_ctx
->param_tes_patch_id
= num_params
++] = i32
;
3497 case TGSI_PROCESSOR_GEOMETRY
:
3498 params
[SI_PARAM_GS2VS_OFFSET
] = i32
;
3499 params
[SI_PARAM_GS_WAVE_ID
] = i32
;
3500 last_sgpr
= SI_PARAM_GS_WAVE_ID
;
3503 params
[SI_PARAM_VTX0_OFFSET
] = i32
;
3504 params
[SI_PARAM_VTX1_OFFSET
] = i32
;
3505 params
[SI_PARAM_PRIMITIVE_ID
] = i32
;
3506 params
[SI_PARAM_VTX2_OFFSET
] = i32
;
3507 params
[SI_PARAM_VTX3_OFFSET
] = i32
;
3508 params
[SI_PARAM_VTX4_OFFSET
] = i32
;
3509 params
[SI_PARAM_VTX5_OFFSET
] = i32
;
3510 params
[SI_PARAM_GS_INSTANCE_ID
] = i32
;
3511 num_params
= SI_PARAM_GS_INSTANCE_ID
+1;
3514 case TGSI_PROCESSOR_FRAGMENT
:
3515 params
[SI_PARAM_ALPHA_REF
] = f32
;
3516 params
[SI_PARAM_PRIM_MASK
] = i32
;
3517 last_sgpr
= SI_PARAM_PRIM_MASK
;
3518 params
[SI_PARAM_PERSP_SAMPLE
] = v2i32
;
3519 params
[SI_PARAM_PERSP_CENTER
] = v2i32
;
3520 params
[SI_PARAM_PERSP_CENTROID
] = v2i32
;
3521 params
[SI_PARAM_PERSP_PULL_MODEL
] = v3i32
;
3522 params
[SI_PARAM_LINEAR_SAMPLE
] = v2i32
;
3523 params
[SI_PARAM_LINEAR_CENTER
] = v2i32
;
3524 params
[SI_PARAM_LINEAR_CENTROID
] = v2i32
;
3525 params
[SI_PARAM_LINE_STIPPLE_TEX
] = f32
;
3526 params
[SI_PARAM_POS_X_FLOAT
] = f32
;
3527 params
[SI_PARAM_POS_Y_FLOAT
] = f32
;
3528 params
[SI_PARAM_POS_Z_FLOAT
] = f32
;
3529 params
[SI_PARAM_POS_W_FLOAT
] = f32
;
3530 params
[SI_PARAM_FRONT_FACE
] = f32
;
3531 params
[SI_PARAM_ANCILLARY
] = i32
;
3532 params
[SI_PARAM_SAMPLE_COVERAGE
] = f32
;
3533 params
[SI_PARAM_POS_FIXED_PT
] = f32
;
3534 num_params
= SI_PARAM_POS_FIXED_PT
+1;
3538 assert(0 && "unimplemented shader");
3542 assert(num_params
<= Elements(params
));
3543 radeon_llvm_create_func(&si_shader_ctx
->radeon_bld
, params
, num_params
);
3544 radeon_llvm_shader_type(si_shader_ctx
->radeon_bld
.main_fn
, si_shader_ctx
->type
);
3546 if (shader
->dx10_clamp_mode
)
3547 LLVMAddTargetDependentFunctionAttr(si_shader_ctx
->radeon_bld
.main_fn
,
3548 "enable-no-nans-fp-math", "true");
3550 for (i
= 0; i
<= last_sgpr
; ++i
) {
3551 LLVMValueRef P
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, i
);
3553 /* We tell llvm that array inputs are passed by value to allow Sinking pass
3554 * to move load. Inputs are constant so this is fine. */
3555 if (i
<= last_array_pointer
)
3556 LLVMAddAttribute(P
, LLVMByValAttribute
);
3558 LLVMAddAttribute(P
, LLVMInRegAttribute
);
3561 if (bld_base
->info
&&
3562 (bld_base
->info
->opcode_count
[TGSI_OPCODE_DDX
] > 0 ||
3563 bld_base
->info
->opcode_count
[TGSI_OPCODE_DDY
] > 0 ||
3564 bld_base
->info
->opcode_count
[TGSI_OPCODE_DDX_FINE
] > 0 ||
3565 bld_base
->info
->opcode_count
[TGSI_OPCODE_DDY_FINE
] > 0 ||
3566 bld_base
->info
->opcode_count
[TGSI_OPCODE_INTERP_OFFSET
] > 0 ||
3567 bld_base
->info
->opcode_count
[TGSI_OPCODE_INTERP_SAMPLE
] > 0))
3568 si_shader_ctx
->lds
=
3569 LLVMAddGlobalInAddressSpace(gallivm
->module
,
3570 LLVMArrayType(i32
, 64),
3574 if ((si_shader_ctx
->type
== TGSI_PROCESSOR_VERTEX
&& shader
->key
.vs
.as_ls
) ||
3575 si_shader_ctx
->type
== TGSI_PROCESSOR_TESS_CTRL
||
3576 si_shader_ctx
->type
== TGSI_PROCESSOR_TESS_EVAL
) {
3577 /* This is the upper bound, maximum is 32 inputs times 32 vertices */
3578 unsigned vertex_data_dw_size
= 32*32*4;
3579 unsigned patch_data_dw_size
= 32*4;
3580 /* The formula is: TCS inputs + TCS outputs + TCS patch outputs. */
3581 unsigned patch_dw_size
= vertex_data_dw_size
*2 + patch_data_dw_size
;
3582 unsigned lds_dwords
= patch_dw_size
;
3584 /* The actual size is computed outside of the shader to reduce
3585 * the number of shader variants. */
3586 si_shader_ctx
->lds
=
3587 LLVMAddGlobalInAddressSpace(gallivm
->module
,
3588 LLVMArrayType(i32
, lds_dwords
),
3594 static void preload_constants(struct si_shader_context
*si_shader_ctx
)
3596 struct lp_build_tgsi_context
* bld_base
= &si_shader_ctx
->radeon_bld
.soa
.bld_base
;
3597 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
3598 const struct tgsi_shader_info
* info
= bld_base
->info
;
3600 LLVMValueRef ptr
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, SI_PARAM_CONST
);
3602 for (buf
= 0; buf
< SI_NUM_CONST_BUFFERS
; buf
++) {
3603 unsigned i
, num_const
= info
->const_file_max
[buf
] + 1;
3608 /* Allocate space for the constant values */
3609 si_shader_ctx
->constants
[buf
] = CALLOC(num_const
* 4, sizeof(LLVMValueRef
));
3611 /* Load the resource descriptor */
3612 si_shader_ctx
->const_resource
[buf
] =
3613 build_indexed_load_const(si_shader_ctx
, ptr
, lp_build_const_int32(gallivm
, buf
));
3615 /* Load the constants, we rely on the code sinking to do the rest */
3616 for (i
= 0; i
< num_const
* 4; ++i
) {
3617 si_shader_ctx
->constants
[buf
][i
] =
3618 buffer_load_const(gallivm
->builder
,
3619 si_shader_ctx
->const_resource
[buf
],
3620 lp_build_const_int32(gallivm
, i
* 4),
3621 bld_base
->base
.elem_type
);
3626 static void preload_samplers(struct si_shader_context
*si_shader_ctx
)
3628 struct lp_build_tgsi_context
* bld_base
= &si_shader_ctx
->radeon_bld
.soa
.bld_base
;
3629 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
3630 const struct tgsi_shader_info
* info
= bld_base
->info
;
3632 unsigned i
, num_samplers
= info
->file_max
[TGSI_FILE_SAMPLER
] + 1;
3634 LLVMValueRef res_ptr
, samp_ptr
;
3635 LLVMValueRef offset
;
3637 if (num_samplers
== 0)
3640 res_ptr
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, SI_PARAM_RESOURCE
);
3641 samp_ptr
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, SI_PARAM_SAMPLER
);
3643 /* Load the resources and samplers, we rely on the code sinking to do the rest */
3644 for (i
= 0; i
< num_samplers
; ++i
) {
3646 offset
= lp_build_const_int32(gallivm
, i
);
3647 si_shader_ctx
->resources
[i
] = build_indexed_load_const(si_shader_ctx
, res_ptr
, offset
);
3650 offset
= lp_build_const_int32(gallivm
, i
);
3651 si_shader_ctx
->samplers
[i
] = build_indexed_load_const(si_shader_ctx
, samp_ptr
, offset
);
3653 /* FMASK resource */
3654 if (info
->is_msaa_sampler
[i
]) {
3655 offset
= lp_build_const_int32(gallivm
, SI_FMASK_TEX_OFFSET
+ i
);
3656 si_shader_ctx
->resources
[SI_FMASK_TEX_OFFSET
+ i
] =
3657 build_indexed_load_const(si_shader_ctx
, res_ptr
, offset
);
3662 static void preload_streamout_buffers(struct si_shader_context
*si_shader_ctx
)
3664 struct lp_build_tgsi_context
* bld_base
= &si_shader_ctx
->radeon_bld
.soa
.bld_base
;
3665 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
3668 /* Streamout can only be used if the shader is compiled as VS. */
3669 if (!si_shader_ctx
->shader
->selector
->so
.num_outputs
||
3670 (si_shader_ctx
->type
== TGSI_PROCESSOR_VERTEX
&&
3671 (si_shader_ctx
->shader
->key
.vs
.as_es
||
3672 si_shader_ctx
->shader
->key
.vs
.as_ls
)) ||
3673 (si_shader_ctx
->type
== TGSI_PROCESSOR_TESS_EVAL
&&
3674 si_shader_ctx
->shader
->key
.tes
.as_es
))
3677 LLVMValueRef buf_ptr
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
,
3678 SI_PARAM_RW_BUFFERS
);
3680 /* Load the resources, we rely on the code sinking to do the rest */
3681 for (i
= 0; i
< 4; ++i
) {
3682 if (si_shader_ctx
->shader
->selector
->so
.stride
[i
]) {
3683 LLVMValueRef offset
= lp_build_const_int32(gallivm
,
3684 SI_SO_BUF_OFFSET
+ i
);
3686 si_shader_ctx
->so_buffers
[i
] = build_indexed_load_const(si_shader_ctx
, buf_ptr
, offset
);
3692 * Load ESGS and GSVS ring buffer resource descriptors and save the variables
3695 static void preload_ring_buffers(struct si_shader_context
*si_shader_ctx
)
3697 struct gallivm_state
*gallivm
=
3698 si_shader_ctx
->radeon_bld
.soa
.bld_base
.base
.gallivm
;
3700 LLVMValueRef buf_ptr
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
,
3701 SI_PARAM_RW_BUFFERS
);
3703 if ((si_shader_ctx
->type
== TGSI_PROCESSOR_VERTEX
&&
3704 si_shader_ctx
->shader
->key
.vs
.as_es
) ||
3705 (si_shader_ctx
->type
== TGSI_PROCESSOR_TESS_EVAL
&&
3706 si_shader_ctx
->shader
->key
.tes
.as_es
) ||
3707 si_shader_ctx
->type
== TGSI_PROCESSOR_GEOMETRY
) {
3708 LLVMValueRef offset
= lp_build_const_int32(gallivm
, SI_RING_ESGS
);
3710 si_shader_ctx
->esgs_ring
=
3711 build_indexed_load_const(si_shader_ctx
, buf_ptr
, offset
);
3714 if (si_shader_ctx
->shader
->is_gs_copy_shader
) {
3715 LLVMValueRef offset
= lp_build_const_int32(gallivm
, SI_RING_GSVS
);
3717 si_shader_ctx
->gsvs_ring
[0] =
3718 build_indexed_load_const(si_shader_ctx
, buf_ptr
, offset
);
3720 if (si_shader_ctx
->type
== TGSI_PROCESSOR_GEOMETRY
) {
3722 for (i
= 0; i
< 4; i
++) {
3723 LLVMValueRef offset
= lp_build_const_int32(gallivm
, SI_RING_GSVS
+ i
);
3725 si_shader_ctx
->gsvs_ring
[i
] =
3726 build_indexed_load_const(si_shader_ctx
, buf_ptr
, offset
);
3731 void si_shader_binary_read_config(const struct si_screen
*sscreen
,
3732 struct si_shader
*shader
,
3733 unsigned symbol_offset
)
3736 const unsigned char *config
=
3737 radeon_shader_binary_config_start(&shader
->binary
,
3740 /* XXX: We may be able to emit some of these values directly rather than
3741 * extracting fields to be emitted later.
3744 for (i
= 0; i
< shader
->binary
.config_size_per_symbol
; i
+= 8) {
3745 unsigned reg
= util_le32_to_cpu(*(uint32_t*)(config
+ i
));
3746 unsigned value
= util_le32_to_cpu(*(uint32_t*)(config
+ i
+ 4));
3748 case R_00B028_SPI_SHADER_PGM_RSRC1_PS
:
3749 case R_00B128_SPI_SHADER_PGM_RSRC1_VS
:
3750 case R_00B228_SPI_SHADER_PGM_RSRC1_GS
:
3751 case R_00B848_COMPUTE_PGM_RSRC1
:
3752 shader
->num_sgprs
= MAX2(shader
->num_sgprs
, (G_00B028_SGPRS(value
) + 1) * 8);
3753 shader
->num_vgprs
= MAX2(shader
->num_vgprs
, (G_00B028_VGPRS(value
) + 1) * 4);
3754 shader
->float_mode
= G_00B028_FLOAT_MODE(value
);
3756 case R_00B02C_SPI_SHADER_PGM_RSRC2_PS
:
3757 shader
->lds_size
= MAX2(shader
->lds_size
, G_00B02C_EXTRA_LDS_SIZE(value
));
3759 case R_00B84C_COMPUTE_PGM_RSRC2
:
3760 shader
->lds_size
= MAX2(shader
->lds_size
, G_00B84C_LDS_SIZE(value
));
3762 case R_0286CC_SPI_PS_INPUT_ENA
:
3763 shader
->spi_ps_input_ena
= value
;
3765 case R_0286E8_SPI_TMPRING_SIZE
:
3766 case R_00B860_COMPUTE_TMPRING_SIZE
:
3767 /* WAVESIZE is in units of 256 dwords. */
3768 shader
->scratch_bytes_per_wave
=
3769 G_00B860_WAVESIZE(value
) * 256 * 4 * 1;
3772 fprintf(stderr
, "Warning: Compiler emitted unknown "
3773 "config register: 0x%x\n", reg
);
3779 void si_shader_apply_scratch_relocs(struct si_context
*sctx
,
3780 struct si_shader
*shader
,
3781 uint64_t scratch_va
)
3784 uint32_t scratch_rsrc_dword0
= scratch_va
& 0xffffffff;
3785 uint32_t scratch_rsrc_dword1
=
3786 S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32)
3787 | S_008F04_STRIDE(shader
->scratch_bytes_per_wave
/ 64);
3789 for (i
= 0 ; i
< shader
->binary
.reloc_count
; i
++) {
3790 const struct radeon_shader_reloc
*reloc
=
3791 &shader
->binary
.relocs
[i
];
3792 if (!strcmp(scratch_rsrc_dword0_symbol
, reloc
->name
)) {
3793 util_memcpy_cpu_to_le32(shader
->binary
.code
+ reloc
->offset
,
3794 &scratch_rsrc_dword0
, 4);
3795 } else if (!strcmp(scratch_rsrc_dword1_symbol
, reloc
->name
)) {
3796 util_memcpy_cpu_to_le32(shader
->binary
.code
+ reloc
->offset
,
3797 &scratch_rsrc_dword1
, 4);
3802 int si_shader_binary_upload(struct si_screen
*sscreen
, struct si_shader
*shader
)
3804 const struct radeon_shader_binary
*binary
= &shader
->binary
;
3805 unsigned code_size
= binary
->code_size
+ binary
->rodata_size
;
3808 r600_resource_reference(&shader
->bo
, NULL
);
3809 shader
->bo
= si_resource_create_custom(&sscreen
->b
.b
,
3810 PIPE_USAGE_IMMUTABLE
,
3815 ptr
= sscreen
->b
.ws
->buffer_map(shader
->bo
->cs_buf
, NULL
,
3816 PIPE_TRANSFER_READ_WRITE
);
3817 util_memcpy_cpu_to_le32(ptr
, binary
->code
, binary
->code_size
);
3818 if (binary
->rodata_size
> 0) {
3819 ptr
+= binary
->code_size
;
3820 util_memcpy_cpu_to_le32(ptr
, binary
->rodata
,
3821 binary
->rodata_size
);
3824 sscreen
->b
.ws
->buffer_unmap(shader
->bo
->cs_buf
);
3828 int si_shader_binary_read(struct si_screen
*sscreen
, struct si_shader
*shader
)
3830 const struct radeon_shader_binary
*binary
= &shader
->binary
;
3832 bool dump
= r600_can_dump_shader(&sscreen
->b
,
3833 shader
->selector
? shader
->selector
->tokens
: NULL
);
3835 si_shader_binary_read_config(sscreen
, shader
, 0);
3836 si_shader_binary_upload(sscreen
, shader
);
3839 if (!(sscreen
->b
.debug_flags
& DBG_NO_ASM
)) {
3840 if (binary
->disasm_string
) {
3841 fprintf(stderr
, "\nShader Disassembly:\n\n");
3842 fprintf(stderr
, "%s\n", binary
->disasm_string
);
3844 fprintf(stderr
, "SI CODE:\n");
3845 for (i
= 0; i
< binary
->code_size
; i
+=4 ) {
3846 fprintf(stderr
, "@0x%x: %02x%02x%02x%02x\n", i
, binary
->code
[i
+ 3],
3847 binary
->code
[i
+ 2], binary
->code
[i
+ 1],
3853 fprintf(stderr
, "*** SHADER STATS ***\n"
3854 "SGPRS: %d\nVGPRS: %d\nCode Size: %d bytes\nLDS: %d blocks\n"
3855 "Scratch: %d bytes per wave\n********************\n",
3856 shader
->num_sgprs
, shader
->num_vgprs
, binary
->code_size
,
3857 shader
->lds_size
, shader
->scratch_bytes_per_wave
);
3862 int si_compile_llvm(struct si_screen
*sscreen
, struct si_shader
*shader
,
3863 LLVMTargetMachineRef tm
, LLVMModuleRef mod
)
3866 bool dump_asm
= r600_can_dump_shader(&sscreen
->b
,
3867 shader
->selector
? shader
->selector
->tokens
: NULL
);
3868 bool dump_ir
= dump_asm
&& !(sscreen
->b
.debug_flags
& DBG_NO_IR
);
3870 r
= radeon_llvm_compile(mod
, &shader
->binary
,
3871 r600_get_llvm_processor_name(sscreen
->b
.family
), dump_ir
, dump_asm
, tm
);
3875 r
= si_shader_binary_read(sscreen
, shader
);
3877 FREE(shader
->binary
.config
);
3878 FREE(shader
->binary
.rodata
);
3879 FREE(shader
->binary
.global_symbol_offsets
);
3880 if (shader
->scratch_bytes_per_wave
== 0) {
3881 FREE(shader
->binary
.code
);
3882 FREE(shader
->binary
.relocs
);
3883 memset(&shader
->binary
, 0,
3884 offsetof(struct radeon_shader_binary
, disasm_string
));
3889 /* Generate code for the hardware VS shader stage to go with a geometry shader */
3890 static int si_generate_gs_copy_shader(struct si_screen
*sscreen
,
3891 struct si_shader_context
*si_shader_ctx
,
3892 struct si_shader
*gs
, bool dump
)
3894 struct gallivm_state
*gallivm
= &si_shader_ctx
->radeon_bld
.gallivm
;
3895 struct lp_build_tgsi_context
*bld_base
= &si_shader_ctx
->radeon_bld
.soa
.bld_base
;
3896 struct lp_build_context
*base
= &bld_base
->base
;
3897 struct lp_build_context
*uint
= &bld_base
->uint_bld
;
3898 struct si_shader
*shader
= si_shader_ctx
->shader
;
3899 struct si_shader_output_values
*outputs
;
3900 struct tgsi_shader_info
*gsinfo
= &gs
->selector
->info
;
3901 LLVMValueRef args
[9];
3904 outputs
= MALLOC(gsinfo
->num_outputs
* sizeof(outputs
[0]));
3906 si_shader_ctx
->type
= TGSI_PROCESSOR_VERTEX
;
3907 shader
->is_gs_copy_shader
= true;
3909 radeon_llvm_context_init(&si_shader_ctx
->radeon_bld
);
3911 create_meta_data(si_shader_ctx
);
3912 create_function(si_shader_ctx
);
3913 preload_streamout_buffers(si_shader_ctx
);
3914 preload_ring_buffers(si_shader_ctx
);
3916 args
[0] = si_shader_ctx
->gsvs_ring
[0];
3917 args
[1] = lp_build_mul_imm(uint
,
3918 LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
,
3919 si_shader_ctx
->param_vertex_id
),
3921 args
[3] = uint
->zero
;
3922 args
[4] = uint
->one
; /* OFFEN */
3923 args
[5] = uint
->zero
; /* IDXEN */
3924 args
[6] = uint
->one
; /* GLC */
3925 args
[7] = uint
->one
; /* SLC */
3926 args
[8] = uint
->zero
; /* TFE */
3928 /* Fetch vertex data from GSVS ring */
3929 for (i
= 0; i
< gsinfo
->num_outputs
; ++i
) {
3932 outputs
[i
].name
= gsinfo
->output_semantic_name
[i
];
3933 outputs
[i
].sid
= gsinfo
->output_semantic_index
[i
];
3935 for (chan
= 0; chan
< 4; chan
++) {
3936 args
[2] = lp_build_const_int32(gallivm
,
3938 gs
->selector
->gs_max_out_vertices
* 16 * 4);
3940 outputs
[i
].values
[chan
] =
3941 LLVMBuildBitCast(gallivm
->builder
,
3942 lp_build_intrinsic(gallivm
->builder
,
3943 "llvm.SI.buffer.load.dword.i32.i32",
3944 LLVMInt32TypeInContext(gallivm
->context
),
3946 LLVMReadOnlyAttribute
| LLVMNoUnwindAttribute
),
3947 base
->elem_type
, "");
3951 si_llvm_export_vs(bld_base
, outputs
, gsinfo
->num_outputs
);
3953 radeon_llvm_finalize_module(&si_shader_ctx
->radeon_bld
);
3956 fprintf(stderr
, "Copy Vertex Shader for Geometry Shader:\n\n");
3958 r
= si_compile_llvm(sscreen
, si_shader_ctx
->shader
,
3959 si_shader_ctx
->tm
, bld_base
->base
.gallivm
->module
);
3961 radeon_llvm_dispose(&si_shader_ctx
->radeon_bld
);
3967 static void si_dump_key(unsigned shader
, union si_shader_key
*key
)
3971 fprintf(stderr
, "SHADER KEY\n");
3974 case PIPE_SHADER_VERTEX
:
3975 fprintf(stderr
, " instance_divisors = {");
3976 for (i
= 0; i
< Elements(key
->vs
.instance_divisors
); i
++)
3977 fprintf(stderr
, !i
? "%u" : ", %u",
3978 key
->vs
.instance_divisors
[i
]);
3979 fprintf(stderr
, "}\n");
3982 fprintf(stderr
, " es_enabled_outputs = 0x%"PRIx64
"\n",
3983 key
->vs
.es_enabled_outputs
);
3984 fprintf(stderr
, " as_es = %u\n", key
->vs
.as_es
);
3985 fprintf(stderr
, " as_ls = %u\n", key
->vs
.as_ls
);
3988 case PIPE_SHADER_TESS_CTRL
:
3989 fprintf(stderr
, " prim_mode = %u\n", key
->tcs
.prim_mode
);
3992 case PIPE_SHADER_TESS_EVAL
:
3994 fprintf(stderr
, " es_enabled_outputs = 0x%"PRIx64
"\n",
3995 key
->tes
.es_enabled_outputs
);
3996 fprintf(stderr
, " as_es = %u\n", key
->tes
.as_es
);
3999 case PIPE_SHADER_GEOMETRY
:
4002 case PIPE_SHADER_FRAGMENT
:
4003 fprintf(stderr
, " export_16bpc = 0x%X\n", key
->ps
.export_16bpc
);
4004 fprintf(stderr
, " last_cbuf = %u\n", key
->ps
.last_cbuf
);
4005 fprintf(stderr
, " color_two_side = %u\n", key
->ps
.color_two_side
);
4006 fprintf(stderr
, " alpha_func = %u\n", key
->ps
.alpha_func
);
4007 fprintf(stderr
, " alpha_to_one = %u\n", key
->ps
.alpha_to_one
);
4008 fprintf(stderr
, " poly_stipple = %u\n", key
->ps
.poly_stipple
);
4016 int si_shader_create(struct si_screen
*sscreen
, LLVMTargetMachineRef tm
,
4017 struct si_shader
*shader
)
4019 struct si_shader_selector
*sel
= shader
->selector
;
4020 struct tgsi_token
*tokens
= sel
->tokens
;
4021 struct si_shader_context si_shader_ctx
;
4022 struct lp_build_tgsi_context
* bld_base
;
4023 struct tgsi_shader_info stipple_shader_info
;
4026 bool poly_stipple
= sel
->type
== PIPE_SHADER_FRAGMENT
&&
4027 shader
->key
.ps
.poly_stipple
;
4028 bool dump
= r600_can_dump_shader(&sscreen
->b
, sel
->tokens
);
4031 tokens
= util_pstipple_create_fragment_shader(tokens
, NULL
,
4032 SI_POLY_STIPPLE_SAMPLER
);
4033 tgsi_scan_shader(tokens
, &stipple_shader_info
);
4036 /* Dump TGSI code before doing TGSI->LLVM conversion in case the
4037 * conversion fails. */
4038 if (dump
&& !(sscreen
->b
.debug_flags
& DBG_NO_TGSI
)) {
4039 si_dump_key(sel
->type
, &shader
->key
);
4040 tgsi_dump(tokens
, 0);
4041 si_dump_streamout(&sel
->so
);
4044 assert(shader
->nparam
== 0);
4046 memset(&si_shader_ctx
, 0, sizeof(si_shader_ctx
));
4047 radeon_llvm_context_init(&si_shader_ctx
.radeon_bld
);
4048 bld_base
= &si_shader_ctx
.radeon_bld
.soa
.bld_base
;
4050 if (sel
->type
!= PIPE_SHADER_COMPUTE
)
4051 shader
->dx10_clamp_mode
= true;
4053 if (sel
->info
.uses_kill
)
4054 shader
->db_shader_control
|= S_02880C_KILL_ENABLE(1);
4056 shader
->uses_instanceid
= sel
->info
.uses_instanceid
;
4057 bld_base
->info
= poly_stipple
? &stipple_shader_info
: &sel
->info
;
4058 bld_base
->emit_fetch_funcs
[TGSI_FILE_CONSTANT
] = fetch_constant
;
4060 bld_base
->op_actions
[TGSI_OPCODE_INTERP_CENTROID
] = interp_action
;
4061 bld_base
->op_actions
[TGSI_OPCODE_INTERP_SAMPLE
] = interp_action
;
4062 bld_base
->op_actions
[TGSI_OPCODE_INTERP_OFFSET
] = interp_action
;
4064 bld_base
->op_actions
[TGSI_OPCODE_TEX
] = tex_action
;
4065 bld_base
->op_actions
[TGSI_OPCODE_TEX2
] = tex_action
;
4066 bld_base
->op_actions
[TGSI_OPCODE_TXB
] = tex_action
;
4067 bld_base
->op_actions
[TGSI_OPCODE_TXB2
] = tex_action
;
4068 bld_base
->op_actions
[TGSI_OPCODE_TXD
] = tex_action
;
4069 bld_base
->op_actions
[TGSI_OPCODE_TXF
] = tex_action
;
4070 bld_base
->op_actions
[TGSI_OPCODE_TXL
] = tex_action
;
4071 bld_base
->op_actions
[TGSI_OPCODE_TXL2
] = tex_action
;
4072 bld_base
->op_actions
[TGSI_OPCODE_TXP
] = tex_action
;
4073 bld_base
->op_actions
[TGSI_OPCODE_TXQ
] = txq_action
;
4074 bld_base
->op_actions
[TGSI_OPCODE_TG4
] = tex_action
;
4075 bld_base
->op_actions
[TGSI_OPCODE_LODQ
] = tex_action
;
4077 bld_base
->op_actions
[TGSI_OPCODE_DDX
].emit
= si_llvm_emit_ddxy
;
4078 bld_base
->op_actions
[TGSI_OPCODE_DDY
].emit
= si_llvm_emit_ddxy
;
4079 bld_base
->op_actions
[TGSI_OPCODE_DDX_FINE
].emit
= si_llvm_emit_ddxy
;
4080 bld_base
->op_actions
[TGSI_OPCODE_DDY_FINE
].emit
= si_llvm_emit_ddxy
;
4082 bld_base
->op_actions
[TGSI_OPCODE_EMIT
].emit
= si_llvm_emit_vertex
;
4083 bld_base
->op_actions
[TGSI_OPCODE_ENDPRIM
].emit
= si_llvm_emit_primitive
;
4084 bld_base
->op_actions
[TGSI_OPCODE_BARRIER
].emit
= si_llvm_emit_barrier
;
4086 if (HAVE_LLVM
>= 0x0306) {
4087 bld_base
->op_actions
[TGSI_OPCODE_MAX
].emit
= build_tgsi_intrinsic_nomem
;
4088 bld_base
->op_actions
[TGSI_OPCODE_MAX
].intr_name
= "llvm.maxnum.f32";
4089 bld_base
->op_actions
[TGSI_OPCODE_MIN
].emit
= build_tgsi_intrinsic_nomem
;
4090 bld_base
->op_actions
[TGSI_OPCODE_MIN
].intr_name
= "llvm.minnum.f32";
4093 si_shader_ctx
.radeon_bld
.load_system_value
= declare_system_value
;
4094 si_shader_ctx
.shader
= shader
;
4095 si_shader_ctx
.type
= tgsi_get_processor_type(tokens
);
4096 si_shader_ctx
.screen
= sscreen
;
4097 si_shader_ctx
.tm
= tm
;
4099 switch (si_shader_ctx
.type
) {
4100 case TGSI_PROCESSOR_VERTEX
:
4101 si_shader_ctx
.radeon_bld
.load_input
= declare_input_vs
;
4102 if (shader
->key
.vs
.as_ls
)
4103 bld_base
->emit_epilogue
= si_llvm_emit_ls_epilogue
;
4104 else if (shader
->key
.vs
.as_es
)
4105 bld_base
->emit_epilogue
= si_llvm_emit_es_epilogue
;
4107 bld_base
->emit_epilogue
= si_llvm_emit_vs_epilogue
;
4109 case TGSI_PROCESSOR_TESS_CTRL
:
4110 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = fetch_input_tcs
;
4111 bld_base
->emit_fetch_funcs
[TGSI_FILE_OUTPUT
] = fetch_output_tcs
;
4112 bld_base
->emit_store
= store_output_tcs
;
4113 bld_base
->emit_epilogue
= si_llvm_emit_tcs_epilogue
;
4115 case TGSI_PROCESSOR_TESS_EVAL
:
4116 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = fetch_input_tes
;
4117 if (shader
->key
.tes
.as_es
)
4118 bld_base
->emit_epilogue
= si_llvm_emit_es_epilogue
;
4120 bld_base
->emit_epilogue
= si_llvm_emit_vs_epilogue
;
4122 case TGSI_PROCESSOR_GEOMETRY
:
4123 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = fetch_input_gs
;
4124 bld_base
->emit_epilogue
= si_llvm_emit_gs_epilogue
;
4126 case TGSI_PROCESSOR_FRAGMENT
:
4127 si_shader_ctx
.radeon_bld
.load_input
= declare_input_fs
;
4128 bld_base
->emit_epilogue
= si_llvm_emit_fs_epilogue
;
4130 switch (sel
->info
.properties
[TGSI_PROPERTY_FS_DEPTH_LAYOUT
]) {
4131 case TGSI_FS_DEPTH_LAYOUT_GREATER
:
4132 shader
->db_shader_control
|=
4133 S_02880C_CONSERVATIVE_Z_EXPORT(V_02880C_EXPORT_GREATER_THAN_Z
);
4135 case TGSI_FS_DEPTH_LAYOUT_LESS
:
4136 shader
->db_shader_control
|=
4137 S_02880C_CONSERVATIVE_Z_EXPORT(V_02880C_EXPORT_LESS_THAN_Z
);
4142 assert(!"Unsupported shader type");
4146 create_meta_data(&si_shader_ctx
);
4147 create_function(&si_shader_ctx
);
4148 preload_constants(&si_shader_ctx
);
4149 preload_samplers(&si_shader_ctx
);
4150 preload_streamout_buffers(&si_shader_ctx
);
4151 preload_ring_buffers(&si_shader_ctx
);
4153 if (si_shader_ctx
.type
== TGSI_PROCESSOR_GEOMETRY
) {
4155 for (i
= 0; i
< 4; i
++) {
4156 si_shader_ctx
.gs_next_vertex
[i
] =
4157 lp_build_alloca(bld_base
->base
.gallivm
,
4158 bld_base
->uint_bld
.elem_type
, "");
4162 if (!lp_build_tgsi_llvm(bld_base
, tokens
)) {
4163 fprintf(stderr
, "Failed to translate shader from TGSI to LLVM\n");
4167 radeon_llvm_finalize_module(&si_shader_ctx
.radeon_bld
);
4169 mod
= bld_base
->base
.gallivm
->module
;
4170 r
= si_compile_llvm(sscreen
, shader
, tm
, mod
);
4172 fprintf(stderr
, "LLVM failed to compile shader\n");
4176 radeon_llvm_dispose(&si_shader_ctx
.radeon_bld
);
4178 if (si_shader_ctx
.type
== TGSI_PROCESSOR_GEOMETRY
) {
4179 shader
->gs_copy_shader
= CALLOC_STRUCT(si_shader
);
4180 shader
->gs_copy_shader
->selector
= shader
->selector
;
4181 shader
->gs_copy_shader
->key
= shader
->key
;
4182 si_shader_ctx
.shader
= shader
->gs_copy_shader
;
4183 if ((r
= si_generate_gs_copy_shader(sscreen
, &si_shader_ctx
,
4185 free(shader
->gs_copy_shader
);
4186 shader
->gs_copy_shader
= NULL
;
4192 for (int i
= 0; i
< SI_NUM_CONST_BUFFERS
; i
++)
4193 FREE(si_shader_ctx
.constants
[i
]);
4195 tgsi_free_tokens(tokens
);
4199 void si_shader_destroy(struct pipe_context
*ctx
, struct si_shader
*shader
)
4201 if (shader
->gs_copy_shader
)
4202 si_shader_destroy(ctx
, shader
->gs_copy_shader
);
4204 if (shader
->scratch_bo
)
4205 r600_resource_reference(&shader
->scratch_bo
, NULL
);
4207 r600_resource_reference(&shader
->bo
, NULL
);
4209 FREE(shader
->binary
.code
);
4210 FREE(shader
->binary
.relocs
);
4211 FREE(shader
->binary
.disasm_string
);