3 * Copyright 2012 Advanced Micro Devices, Inc.
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * on the rights to use, copy, modify, merge, publish, distribute, sub
9 * license, and/or sell copies of the Software, and to permit persons to whom
10 * the Software is furnished to do so, subject to the following conditions:
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22 * USE OR OTHER DEALINGS IN THE SOFTWARE.
25 * Tom Stellard <thomas.stellard@amd.com>
26 * Michel Dänzer <michel.daenzer@amd.com>
27 * Christian König <christian.koenig@amd.com>
30 #include "gallivm/lp_bld_tgsi_action.h"
31 #include "gallivm/lp_bld_const.h"
32 #include "gallivm/lp_bld_gather.h"
33 #include "gallivm/lp_bld_intr.h"
34 #include "gallivm/lp_bld_logic.h"
35 #include "gallivm/lp_bld_tgsi.h"
36 #include "gallivm/lp_bld_arit.h"
37 #include "radeon_llvm.h"
38 #include "radeon_llvm_emit.h"
39 #include "util/u_memory.h"
40 #include "tgsi/tgsi_info.h"
41 #include "tgsi/tgsi_parse.h"
42 #include "tgsi/tgsi_scan.h"
43 #include "tgsi/tgsi_dump.h"
45 #include "radeonsi_pipe.h"
46 #include "radeonsi_shader.h"
54 struct si_shader_context
56 struct radeon_llvm_context radeon_bld
;
57 struct r600_context
*rctx
;
58 struct tgsi_parse_context parse
;
59 struct tgsi_token
* tokens
;
60 struct si_pipe_shader
*shader
;
61 struct si_shader_key key
;
62 unsigned type
; /* TGSI_PROCESSOR_* specifies the type of shader. */
63 LLVMValueRef const_md
;
64 LLVMValueRef const_resource
;
65 LLVMValueRef
*constants
;
66 LLVMValueRef
*resources
;
67 LLVMValueRef
*samplers
;
70 static struct si_shader_context
* si_shader_context(
71 struct lp_build_tgsi_context
* bld_base
)
73 return (struct si_shader_context
*)bld_base
;
77 #define PERSPECTIVE_BASE 0
80 #define SAMPLE_OFFSET 0
81 #define CENTER_OFFSET 2
82 #define CENTROID_OFSET 4
84 #define USE_SGPR_MAX_SUFFIX_LEN 5
85 #define CONST_ADDR_SPACE 2
86 #define USER_SGPR_ADDR_SPACE 8
89 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad
91 * @param offset The offset parameter specifies the number of
92 * elements to offset, not the number of bytes or dwords. An element is the
93 * the type pointed to by the base_ptr parameter (e.g. int is the element of
96 * When LLVM lowers the load instruction, it will convert the element offset
97 * into a dword offset automatically.
100 static LLVMValueRef
build_indexed_load(
101 struct si_shader_context
* si_shader_ctx
,
102 LLVMValueRef base_ptr
,
105 struct lp_build_context
* base
= &si_shader_ctx
->radeon_bld
.soa
.bld_base
.base
;
107 LLVMValueRef computed_ptr
= LLVMBuildGEP(
108 base
->gallivm
->builder
, base_ptr
, &offset
, 1, "");
110 LLVMValueRef result
= LLVMBuildLoad(base
->gallivm
->builder
, computed_ptr
, "");
111 LLVMSetMetadata(result
, 1, si_shader_ctx
->const_md
);
115 static void declare_input_vs(
116 struct si_shader_context
* si_shader_ctx
,
117 unsigned input_index
,
118 const struct tgsi_full_declaration
*decl
)
120 LLVMValueRef t_list_ptr
;
121 LLVMValueRef t_offset
;
123 LLVMValueRef attribute_offset
;
124 LLVMValueRef buffer_index_reg
;
125 LLVMValueRef args
[3];
126 LLVMTypeRef vec4_type
;
128 struct lp_build_context
* base
= &si_shader_ctx
->radeon_bld
.soa
.bld_base
.base
;
129 //struct pipe_vertex_element *velem = &rctx->vertex_elements->elements[input_index];
132 /* Load the T list */
133 t_list_ptr
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, SI_PARAM_VERTEX_BUFFER
);
135 t_offset
= lp_build_const_int32(base
->gallivm
, input_index
);
137 t_list
= build_indexed_load(si_shader_ctx
, t_list_ptr
, t_offset
);
139 /* Build the attribute offset */
140 attribute_offset
= lp_build_const_int32(base
->gallivm
, 0);
142 /* Load the buffer index, which is always stored in VGPR0
143 * for Vertex Shaders */
144 buffer_index_reg
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, SI_PARAM_VERTEX_INDEX
);
146 vec4_type
= LLVMVectorType(base
->elem_type
, 4);
148 args
[1] = attribute_offset
;
149 args
[2] = buffer_index_reg
;
150 input
= build_intrinsic(base
->gallivm
->builder
,
151 "llvm.SI.vs.load.input", vec4_type
, args
, 3,
152 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
154 /* Break up the vec4 into individual components */
155 for (chan
= 0; chan
< 4; chan
++) {
156 LLVMValueRef llvm_chan
= lp_build_const_int32(base
->gallivm
, chan
);
157 /* XXX: Use a helper function for this. There is one in
159 si_shader_ctx
->radeon_bld
.inputs
[radeon_llvm_reg_index_soa(input_index
, chan
)] =
160 LLVMBuildExtractElement(base
->gallivm
->builder
,
161 input
, llvm_chan
, "");
165 static void declare_input_fs(
166 struct si_shader_context
* si_shader_ctx
,
167 unsigned input_index
,
168 const struct tgsi_full_declaration
*decl
)
170 struct si_shader
*shader
= &si_shader_ctx
->shader
->shader
;
171 struct lp_build_context
* base
=
172 &si_shader_ctx
->radeon_bld
.soa
.bld_base
.base
;
173 struct gallivm_state
* gallivm
= base
->gallivm
;
174 LLVMTypeRef input_type
= LLVMFloatTypeInContext(gallivm
->context
);
175 LLVMValueRef main_fn
= si_shader_ctx
->radeon_bld
.main_fn
;
177 LLVMValueRef interp_param
;
178 const char * intr_name
;
181 * [15:0] NewPrimMask (Bit mask for each quad. It is set it the
182 * quad begins a new primitive. Bit 0 always needs
184 * [32:16] ParamOffset
187 LLVMValueRef params
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, SI_PARAM_PRIM_MASK
);
188 LLVMValueRef attr_number
;
192 if (decl
->Semantic
.Name
== TGSI_SEMANTIC_POSITION
) {
193 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
195 radeon_llvm_reg_index_soa(input_index
, chan
);
196 si_shader_ctx
->radeon_bld
.inputs
[soa_index
] =
197 LLVMGetParam(main_fn
, SI_PARAM_POS_X_FLOAT
+ chan
);
200 /* RCP for fragcoord.w */
201 si_shader_ctx
->radeon_bld
.inputs
[soa_index
] =
202 LLVMBuildFDiv(gallivm
->builder
,
203 lp_build_const_float(gallivm
, 1.0f
),
204 si_shader_ctx
->radeon_bld
.inputs
[soa_index
],
210 if (decl
->Semantic
.Name
== TGSI_SEMANTIC_FACE
) {
211 LLVMValueRef face
, is_face_positive
;
213 face
= LLVMGetParam(main_fn
, SI_PARAM_FRONT_FACE
);
215 is_face_positive
= LLVMBuildFCmp(gallivm
->builder
,
217 lp_build_const_float(gallivm
, 0.0f
),
220 si_shader_ctx
->radeon_bld
.inputs
[radeon_llvm_reg_index_soa(input_index
, 0)] =
221 LLVMBuildSelect(gallivm
->builder
,
223 lp_build_const_float(gallivm
, 1.0f
),
224 lp_build_const_float(gallivm
, 0.0f
),
226 si_shader_ctx
->radeon_bld
.inputs
[radeon_llvm_reg_index_soa(input_index
, 1)] =
227 si_shader_ctx
->radeon_bld
.inputs
[radeon_llvm_reg_index_soa(input_index
, 2)] =
228 lp_build_const_float(gallivm
, 0.0f
);
229 si_shader_ctx
->radeon_bld
.inputs
[radeon_llvm_reg_index_soa(input_index
, 3)] =
230 lp_build_const_float(gallivm
, 1.0f
);
235 shader
->input
[input_index
].param_offset
= shader
->ninterp
++;
236 attr_number
= lp_build_const_int32(gallivm
,
237 shader
->input
[input_index
].param_offset
);
239 /* XXX: Handle all possible interpolation modes */
240 switch (decl
->Interp
.Interpolate
) {
241 case TGSI_INTERPOLATE_COLOR
:
242 if (si_shader_ctx
->key
.flatshade
) {
245 if (decl
->Interp
.Centroid
)
246 interp_param
= LLVMGetParam(main_fn
, SI_PARAM_PERSP_CENTROID
);
248 interp_param
= LLVMGetParam(main_fn
, SI_PARAM_PERSP_CENTER
);
251 case TGSI_INTERPOLATE_CONSTANT
:
254 case TGSI_INTERPOLATE_LINEAR
:
255 if (decl
->Interp
.Centroid
)
256 interp_param
= LLVMGetParam(main_fn
, SI_PARAM_LINEAR_CENTROID
);
258 interp_param
= LLVMGetParam(main_fn
, SI_PARAM_LINEAR_CENTER
);
260 case TGSI_INTERPOLATE_PERSPECTIVE
:
261 if (decl
->Interp
.Centroid
)
262 interp_param
= LLVMGetParam(main_fn
, SI_PARAM_PERSP_CENTROID
);
264 interp_param
= LLVMGetParam(main_fn
, SI_PARAM_PERSP_CENTER
);
267 fprintf(stderr
, "Warning: Unhandled interpolation mode.\n");
271 intr_name
= interp_param
? "llvm.SI.fs.interp" : "llvm.SI.fs.constant";
273 /* XXX: Could there be more than TGSI_NUM_CHANNELS (4) ? */
274 if (decl
->Semantic
.Name
== TGSI_SEMANTIC_COLOR
&&
275 si_shader_ctx
->key
.color_two_side
) {
276 LLVMValueRef args
[4];
277 LLVMValueRef face
, is_face_positive
;
278 LLVMValueRef back_attr_number
=
279 lp_build_const_int32(gallivm
,
280 shader
->input
[input_index
].param_offset
+ 1);
282 face
= LLVMGetParam(main_fn
, SI_PARAM_FRONT_FACE
);
284 is_face_positive
= LLVMBuildFCmp(gallivm
->builder
,
286 lp_build_const_float(gallivm
, 0.0f
),
290 args
[3] = interp_param
;
291 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
292 LLVMValueRef llvm_chan
= lp_build_const_int32(gallivm
, chan
);
293 unsigned soa_index
= radeon_llvm_reg_index_soa(input_index
, chan
);
294 LLVMValueRef front
, back
;
297 args
[1] = attr_number
;
298 front
= build_intrinsic(base
->gallivm
->builder
, intr_name
,
299 input_type
, args
, args
[3] ? 4 : 3,
300 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
302 args
[1] = back_attr_number
;
303 back
= build_intrinsic(base
->gallivm
->builder
, intr_name
,
304 input_type
, args
, args
[3] ? 4 : 3,
305 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
307 si_shader_ctx
->radeon_bld
.inputs
[soa_index
] =
308 LLVMBuildSelect(gallivm
->builder
,
317 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
318 LLVMValueRef args
[4];
319 LLVMValueRef llvm_chan
= lp_build_const_int32(gallivm
, chan
);
320 unsigned soa_index
= radeon_llvm_reg_index_soa(input_index
, chan
);
322 args
[1] = attr_number
;
324 args
[3] = interp_param
;
325 si_shader_ctx
->radeon_bld
.inputs
[soa_index
] =
326 build_intrinsic(base
->gallivm
->builder
, intr_name
,
327 input_type
, args
, args
[3] ? 4 : 3,
328 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
333 static void declare_input(
334 struct radeon_llvm_context
* radeon_bld
,
335 unsigned input_index
,
336 const struct tgsi_full_declaration
*decl
)
338 struct si_shader_context
* si_shader_ctx
=
339 si_shader_context(&radeon_bld
->soa
.bld_base
);
340 if (si_shader_ctx
->type
== TGSI_PROCESSOR_VERTEX
) {
341 declare_input_vs(si_shader_ctx
, input_index
, decl
);
342 } else if (si_shader_ctx
->type
== TGSI_PROCESSOR_FRAGMENT
) {
343 declare_input_fs(si_shader_ctx
, input_index
, decl
);
345 fprintf(stderr
, "Warning: Unsupported shader type,\n");
349 static LLVMValueRef
fetch_constant(
350 struct lp_build_tgsi_context
* bld_base
,
351 const struct tgsi_full_src_register
*reg
,
352 enum tgsi_opcode_type type
,
355 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
356 struct lp_build_context
* base
= &bld_base
->base
;
357 const struct tgsi_ind_register
*ireg
= ®
->Indirect
;
360 LLVMValueRef args
[2];
364 if (swizzle
== LP_CHAN_ALL
) {
366 LLVMValueRef values
[4];
367 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
)
368 values
[chan
] = fetch_constant(bld_base
, reg
, type
, chan
);
370 return lp_build_gather_values(bld_base
->base
.gallivm
, values
, 4);
373 idx
= reg
->Register
.Index
* 4 + swizzle
;
374 if (!reg
->Register
.Indirect
)
375 return bitcast(bld_base
, type
, si_shader_ctx
->constants
[idx
]);
377 args
[0] = si_shader_ctx
->const_resource
;
378 args
[1] = lp_build_const_int32(base
->gallivm
, idx
* 4);
379 addr
= si_shader_ctx
->radeon_bld
.soa
.addr
[ireg
->Index
][ireg
->Swizzle
];
380 addr
= LLVMBuildLoad(base
->gallivm
->builder
, addr
, "load addr reg");
381 addr
= lp_build_mul_imm(&bld_base
->uint_bld
, addr
, 16);
382 args
[1] = lp_build_add(&bld_base
->uint_bld
, addr
, args
[1]);
384 result
= build_intrinsic(base
->gallivm
->builder
, "llvm.SI.load.const", base
->elem_type
,
385 args
, 2, LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
387 return bitcast(bld_base
, type
, result
);
390 /* Initialize arguments for the shader export intrinsic */
391 static void si_llvm_init_export_args(struct lp_build_tgsi_context
*bld_base
,
392 struct tgsi_full_declaration
*d
,
397 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
398 struct lp_build_context
*uint
=
399 &si_shader_ctx
->radeon_bld
.soa
.bld_base
.uint_bld
;
400 struct lp_build_context
*base
= &bld_base
->base
;
401 unsigned compressed
= 0;
404 if (si_shader_ctx
->type
== TGSI_PROCESSOR_FRAGMENT
) {
405 int cbuf
= target
- V_008DFC_SQ_EXP_MRT
;
407 if (cbuf
>= 0 && cbuf
< 8) {
408 compressed
= (si_shader_ctx
->key
.export_16bpc
>> cbuf
) & 0x1;
411 si_shader_ctx
->shader
->spi_shader_col_format
|=
412 V_028714_SPI_SHADER_FP16_ABGR
<< (4 * cbuf
);
414 si_shader_ctx
->shader
->spi_shader_col_format
|=
415 V_028714_SPI_SHADER_32_ABGR
<< (4 * cbuf
);
420 /* Pixel shader needs to pack output values before export */
421 for (chan
= 0; chan
< 2; chan
++ ) {
422 LLVMValueRef
*out_ptr
=
423 si_shader_ctx
->radeon_bld
.soa
.outputs
[index
];
424 args
[0] = LLVMBuildLoad(base
->gallivm
->builder
,
425 out_ptr
[2 * chan
], "");
426 args
[1] = LLVMBuildLoad(base
->gallivm
->builder
,
427 out_ptr
[2 * chan
+ 1], "");
429 build_intrinsic(base
->gallivm
->builder
,
431 LLVMInt32TypeInContext(base
->gallivm
->context
),
433 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
434 args
[chan
+ 7] = args
[chan
+ 5] =
435 LLVMBuildBitCast(base
->gallivm
->builder
,
437 LLVMFloatTypeInContext(base
->gallivm
->context
),
444 for (chan
= 0; chan
< 4; chan
++ ) {
445 LLVMValueRef out_ptr
=
446 si_shader_ctx
->radeon_bld
.soa
.outputs
[index
][chan
];
447 /* +5 because the first output value will be
448 * the 6th argument to the intrinsic. */
449 args
[chan
+ 5] = LLVMBuildLoad(base
->gallivm
->builder
,
453 /* Clear COMPR flag */
454 args
[4] = uint
->zero
;
457 /* XXX: This controls which components of the output
458 * registers actually get exported. (e.g bit 0 means export
459 * X component, bit 1 means export Y component, etc.) I'm
460 * hard coding this to 0xf for now. In the future, we might
461 * want to do something else. */
462 args
[0] = lp_build_const_int32(base
->gallivm
, 0xf);
464 /* Specify whether the EXEC mask represents the valid mask */
465 args
[1] = uint
->zero
;
467 /* Specify whether this is the last export */
468 args
[2] = uint
->zero
;
470 /* Specify the target we are exporting */
471 args
[3] = lp_build_const_int32(base
->gallivm
, target
);
473 /* XXX: We probably need to keep track of the output
474 * values, so we know what we are passing to the next
478 static void si_alpha_test(struct lp_build_tgsi_context
*bld_base
,
481 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
482 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
484 if (si_shader_ctx
->key
.alpha_func
!= PIPE_FUNC_NEVER
) {
485 LLVMValueRef out_ptr
= si_shader_ctx
->radeon_bld
.soa
.outputs
[index
][3];
486 LLVMValueRef alpha_pass
=
487 lp_build_cmp(&bld_base
->base
,
488 si_shader_ctx
->key
.alpha_func
,
489 LLVMBuildLoad(gallivm
->builder
, out_ptr
, ""),
490 lp_build_const_float(gallivm
, si_shader_ctx
->key
.alpha_ref
));
492 lp_build_select(&bld_base
->base
,
494 lp_build_const_float(gallivm
, 1.0f
),
495 lp_build_const_float(gallivm
, -1.0f
));
497 build_intrinsic(gallivm
->builder
,
499 LLVMVoidTypeInContext(gallivm
->context
),
502 build_intrinsic(gallivm
->builder
,
504 LLVMVoidTypeInContext(gallivm
->context
),
509 /* XXX: This is partially implemented for VS only at this point. It is not complete */
510 static void si_llvm_emit_epilogue(struct lp_build_tgsi_context
* bld_base
)
512 struct si_shader_context
* si_shader_ctx
= si_shader_context(bld_base
);
513 struct si_shader
* shader
= &si_shader_ctx
->shader
->shader
;
514 struct lp_build_context
* base
= &bld_base
->base
;
515 struct lp_build_context
* uint
=
516 &si_shader_ctx
->radeon_bld
.soa
.bld_base
.uint_bld
;
517 struct tgsi_parse_context
*parse
= &si_shader_ctx
->parse
;
518 LLVMValueRef args
[9];
519 LLVMValueRef last_args
[9] = { 0 };
520 unsigned color_count
= 0;
521 unsigned param_count
= 0;
522 int depth_index
= -1, stencil_index
= -1;
524 while (!tgsi_parse_end_of_tokens(parse
)) {
525 struct tgsi_full_declaration
*d
=
526 &parse
->FullToken
.FullDeclaration
;
531 tgsi_parse_token(parse
);
533 if (parse
->FullToken
.Token
.Type
== TGSI_TOKEN_TYPE_PROPERTY
&&
534 parse
->FullToken
.FullProperty
.Property
.PropertyName
==
535 TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
)
536 shader
->fs_write_all
= TRUE
;
538 if (parse
->FullToken
.Token
.Type
!= TGSI_TOKEN_TYPE_DECLARATION
)
541 switch (d
->Declaration
.File
) {
542 case TGSI_FILE_INPUT
:
543 i
= shader
->ninput
++;
544 shader
->input
[i
].name
= d
->Semantic
.Name
;
545 shader
->input
[i
].sid
= d
->Semantic
.Index
;
546 shader
->input
[i
].interpolate
= d
->Interp
.Interpolate
;
547 shader
->input
[i
].centroid
= d
->Interp
.Centroid
;
550 case TGSI_FILE_OUTPUT
:
551 i
= shader
->noutput
++;
552 shader
->output
[i
].name
= d
->Semantic
.Name
;
553 shader
->output
[i
].sid
= d
->Semantic
.Index
;
554 shader
->output
[i
].interpolate
= d
->Interp
.Interpolate
;
561 for (index
= d
->Range
.First
; index
<= d
->Range
.Last
; index
++) {
562 /* Select the correct target */
563 switch(d
->Semantic
.Name
) {
564 case TGSI_SEMANTIC_PSIZE
:
565 target
= V_008DFC_SQ_EXP_POS
;
567 case TGSI_SEMANTIC_POSITION
:
568 if (si_shader_ctx
->type
== TGSI_PROCESSOR_VERTEX
) {
569 target
= V_008DFC_SQ_EXP_POS
;
575 case TGSI_SEMANTIC_STENCIL
:
576 stencil_index
= index
;
578 case TGSI_SEMANTIC_COLOR
:
579 if (si_shader_ctx
->type
== TGSI_PROCESSOR_VERTEX
) {
580 case TGSI_SEMANTIC_BCOLOR
:
581 target
= V_008DFC_SQ_EXP_PARAM
+ param_count
;
582 shader
->output
[i
].param_offset
= param_count
;
585 target
= V_008DFC_SQ_EXP_MRT
+ color_count
;
586 if (color_count
== 0 &&
587 si_shader_ctx
->key
.alpha_func
!= PIPE_FUNC_ALWAYS
)
588 si_alpha_test(bld_base
, index
);
593 case TGSI_SEMANTIC_FOG
:
594 case TGSI_SEMANTIC_GENERIC
:
595 target
= V_008DFC_SQ_EXP_PARAM
+ param_count
;
596 shader
->output
[i
].param_offset
= param_count
;
602 "Warning: SI unhandled output type:%d\n",
606 si_llvm_init_export_args(bld_base
, d
, index
, target
, args
);
608 if (si_shader_ctx
->type
== TGSI_PROCESSOR_VERTEX
?
609 (d
->Semantic
.Name
== TGSI_SEMANTIC_POSITION
) :
610 (d
->Semantic
.Name
== TGSI_SEMANTIC_COLOR
)) {
612 lp_build_intrinsic(base
->gallivm
->builder
,
614 LLVMVoidTypeInContext(base
->gallivm
->context
),
618 memcpy(last_args
, args
, sizeof(args
));
620 lp_build_intrinsic(base
->gallivm
->builder
,
622 LLVMVoidTypeInContext(base
->gallivm
->context
),
629 if (depth_index
>= 0 || stencil_index
>= 0) {
630 LLVMValueRef out_ptr
;
633 /* Specify the target we are exporting */
634 args
[3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_MRTZ
);
636 if (depth_index
>= 0) {
637 out_ptr
= si_shader_ctx
->radeon_bld
.soa
.outputs
[depth_index
][2];
638 args
[5] = LLVMBuildLoad(base
->gallivm
->builder
, out_ptr
, "");
641 if (stencil_index
< 0) {
648 if (stencil_index
>= 0) {
649 out_ptr
= si_shader_ctx
->radeon_bld
.soa
.outputs
[stencil_index
][1];
652 args
[6] = LLVMBuildLoad(base
->gallivm
->builder
, out_ptr
, "");
659 /* Specify which components to enable */
660 args
[0] = lp_build_const_int32(base
->gallivm
, mask
);
664 args
[4] = uint
->zero
;
667 lp_build_intrinsic(base
->gallivm
->builder
,
669 LLVMVoidTypeInContext(base
->gallivm
->context
),
672 memcpy(last_args
, args
, sizeof(args
));
676 assert(si_shader_ctx
->type
== TGSI_PROCESSOR_FRAGMENT
);
678 /* Specify which components to enable */
679 last_args
[0] = lp_build_const_int32(base
->gallivm
, 0x0);
681 /* Specify the target we are exporting */
682 last_args
[3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_MRT
);
684 /* Set COMPR flag to zero to export data as 32-bit */
685 last_args
[4] = uint
->zero
;
688 last_args
[5]= uint
->zero
;
689 last_args
[6]= uint
->zero
;
690 last_args
[7]= uint
->zero
;
691 last_args
[8]= uint
->zero
;
693 si_shader_ctx
->shader
->spi_shader_col_format
|=
694 V_028714_SPI_SHADER_32_ABGR
;
697 /* Specify whether the EXEC mask represents the valid mask */
698 last_args
[1] = lp_build_const_int32(base
->gallivm
,
699 si_shader_ctx
->type
== TGSI_PROCESSOR_FRAGMENT
);
701 if (shader
->fs_write_all
&& shader
->nr_cbufs
> 1) {
704 /* Specify that this is not yet the last export */
705 last_args
[2] = lp_build_const_int32(base
->gallivm
, 0);
707 for (i
= 1; i
< shader
->nr_cbufs
; i
++) {
708 /* Specify the target we are exporting */
709 last_args
[3] = lp_build_const_int32(base
->gallivm
,
710 V_008DFC_SQ_EXP_MRT
+ i
);
712 lp_build_intrinsic(base
->gallivm
->builder
,
714 LLVMVoidTypeInContext(base
->gallivm
->context
),
717 si_shader_ctx
->shader
->spi_shader_col_format
|=
718 si_shader_ctx
->shader
->spi_shader_col_format
<< 4;
721 last_args
[3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_MRT
);
724 /* Specify that this is the last export */
725 last_args
[2] = lp_build_const_int32(base
->gallivm
, 1);
727 lp_build_intrinsic(base
->gallivm
->builder
,
729 LLVMVoidTypeInContext(base
->gallivm
->context
),
732 /* XXX: Look up what this function does */
733 /* ctx->shader->output[i].spi_sid = r600_spi_sid(&ctx->shader->output[i]);*/
736 static void tex_fetch_args(
737 struct lp_build_tgsi_context
* bld_base
,
738 struct lp_build_emit_data
* emit_data
)
740 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
741 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
742 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
743 unsigned opcode
= inst
->Instruction
.Opcode
;
744 unsigned target
= inst
->Texture
.Texture
;
745 LLVMValueRef coords
[4];
746 LLVMValueRef address
[16];
751 /* XXX: should be optimized using emit_data->inst->Dst[0].Register.WriteMask*/
752 emit_data
->args
[0] = lp_build_const_int32(bld_base
->base
.gallivm
, 0xf);
754 /* Fetch and project texture coordinates */
755 coords
[3] = lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, TGSI_CHAN_W
);
756 for (chan
= 0; chan
< 3; chan
++ ) {
757 coords
[chan
] = lp_build_emit_fetch(bld_base
,
760 if (opcode
== TGSI_OPCODE_TXP
)
761 coords
[chan
] = lp_build_emit_llvm_binary(bld_base
,
767 if (opcode
== TGSI_OPCODE_TXP
)
768 coords
[3] = bld_base
->base
.one
;
770 /* Pack LOD bias value */
771 if (opcode
== TGSI_OPCODE_TXB
)
772 address
[count
++] = coords
[3];
774 if ((target
== TGSI_TEXTURE_CUBE
|| target
== TGSI_TEXTURE_SHADOWCUBE
) &&
775 opcode
!= TGSI_OPCODE_TXQ
)
776 radeon_llvm_emit_prepare_cube_coords(bld_base
, emit_data
, coords
);
778 /* Pack depth comparison value */
780 case TGSI_TEXTURE_SHADOW1D
:
781 case TGSI_TEXTURE_SHADOW1D_ARRAY
:
782 case TGSI_TEXTURE_SHADOW2D
:
783 case TGSI_TEXTURE_SHADOWRECT
:
784 address
[count
++] = coords
[2];
786 case TGSI_TEXTURE_SHADOWCUBE
:
787 case TGSI_TEXTURE_SHADOW2D_ARRAY
:
788 address
[count
++] = coords
[3];
790 case TGSI_TEXTURE_SHADOWCUBE_ARRAY
:
791 address
[count
++] = lp_build_emit_fetch(bld_base
, inst
, 1, 0);
794 /* Pack texture coordinates */
795 address
[count
++] = coords
[0];
797 case TGSI_TEXTURE_2D
:
798 case TGSI_TEXTURE_2D_ARRAY
:
799 case TGSI_TEXTURE_3D
:
800 case TGSI_TEXTURE_CUBE
:
801 case TGSI_TEXTURE_RECT
:
802 case TGSI_TEXTURE_SHADOW2D
:
803 case TGSI_TEXTURE_SHADOWRECT
:
804 case TGSI_TEXTURE_SHADOW2D_ARRAY
:
805 case TGSI_TEXTURE_SHADOWCUBE
:
806 case TGSI_TEXTURE_2D_MSAA
:
807 case TGSI_TEXTURE_2D_ARRAY_MSAA
:
808 case TGSI_TEXTURE_CUBE_ARRAY
:
809 case TGSI_TEXTURE_SHADOWCUBE_ARRAY
:
810 address
[count
++] = coords
[1];
813 case TGSI_TEXTURE_3D
:
814 case TGSI_TEXTURE_CUBE
:
815 case TGSI_TEXTURE_SHADOWCUBE
:
816 case TGSI_TEXTURE_CUBE_ARRAY
:
817 case TGSI_TEXTURE_SHADOWCUBE_ARRAY
:
818 address
[count
++] = coords
[2];
821 /* Pack array slice */
823 case TGSI_TEXTURE_1D_ARRAY
:
824 address
[count
++] = coords
[1];
827 case TGSI_TEXTURE_2D_ARRAY
:
828 case TGSI_TEXTURE_2D_ARRAY_MSAA
:
829 case TGSI_TEXTURE_SHADOW2D_ARRAY
:
830 address
[count
++] = coords
[2];
833 case TGSI_TEXTURE_CUBE_ARRAY
:
834 case TGSI_TEXTURE_SHADOW1D_ARRAY
:
835 case TGSI_TEXTURE_SHADOWCUBE_ARRAY
:
836 address
[count
++] = coords
[3];
840 if (opcode
== TGSI_OPCODE_TXL
)
841 address
[count
++] = coords
[3];
844 assert(!"Cannot handle more than 16 texture address parameters");
848 for (chan
= 0; chan
< count
; chan
++ ) {
849 address
[chan
] = LLVMBuildBitCast(gallivm
->builder
,
851 LLVMInt32TypeInContext(gallivm
->context
),
855 /* Pad to power of two vector */
856 while (count
< util_next_power_of_two(count
))
857 address
[count
++] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
859 emit_data
->args
[1] = lp_build_gather_values(gallivm
, address
, count
);
862 emit_data
->args
[2] = si_shader_ctx
->resources
[emit_data
->inst
->Src
[1].Register
.Index
];
865 emit_data
->args
[3] = si_shader_ctx
->samplers
[emit_data
->inst
->Src
[1].Register
.Index
];
868 emit_data
->args
[4] = lp_build_const_int32(bld_base
->base
.gallivm
, target
);
870 emit_data
->arg_count
= 5;
871 /* XXX: To optimize, we could use a float or v2f32, if the last bits of
872 * the writemask are clear */
873 emit_data
->dst_type
= LLVMVectorType(
874 LLVMFloatTypeInContext(bld_base
->base
.gallivm
->context
),
878 static void build_tex_intrinsic(const struct lp_build_tgsi_action
* action
,
879 struct lp_build_tgsi_context
* bld_base
,
880 struct lp_build_emit_data
* emit_data
)
882 struct lp_build_context
* base
= &bld_base
->base
;
885 sprintf(intr_name
, "%sv%ui32", action
->intr_name
,
886 LLVMGetVectorSize(LLVMTypeOf(emit_data
->args
[1])));
888 emit_data
->output
[emit_data
->chan
] = build_intrinsic(
889 base
->gallivm
->builder
, intr_name
, emit_data
->dst_type
,
890 emit_data
->args
, emit_data
->arg_count
,
891 LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
894 static const struct lp_build_tgsi_action tex_action
= {
895 .fetch_args
= tex_fetch_args
,
896 .emit
= build_tex_intrinsic
,
897 .intr_name
= "llvm.SI.sample."
900 static const struct lp_build_tgsi_action txb_action
= {
901 .fetch_args
= tex_fetch_args
,
902 .emit
= build_tex_intrinsic
,
903 .intr_name
= "llvm.SI.sampleb."
906 static const struct lp_build_tgsi_action txl_action
= {
907 .fetch_args
= tex_fetch_args
,
908 .emit
= build_tex_intrinsic
,
909 .intr_name
= "llvm.SI.samplel."
912 static void create_meta_data(struct si_shader_context
*si_shader_ctx
)
914 struct gallivm_state
*gallivm
= si_shader_ctx
->radeon_bld
.soa
.bld_base
.base
.gallivm
;
915 LLVMValueRef args
[3];
917 args
[0] = LLVMMDStringInContext(gallivm
->context
, "const", 5);
919 args
[2] = lp_build_const_int32(gallivm
, 1);
921 si_shader_ctx
->const_md
= LLVMMDNodeInContext(gallivm
->context
, args
, 3);
924 static void create_function(struct si_shader_context
*si_shader_ctx
)
926 struct gallivm_state
*gallivm
= si_shader_ctx
->radeon_bld
.soa
.bld_base
.base
.gallivm
;
927 LLVMTypeRef params
[20], f32
, i8
, i32
, v2i32
, v3i32
;
930 i8
= LLVMInt8TypeInContext(gallivm
->context
);
931 i32
= LLVMInt32TypeInContext(gallivm
->context
);
932 f32
= LLVMFloatTypeInContext(gallivm
->context
);
933 v2i32
= LLVMVectorType(i32
, 2);
934 v3i32
= LLVMVectorType(i32
, 3);
936 params
[SI_PARAM_CONST
] = LLVMPointerType(LLVMVectorType(i8
, 16), CONST_ADDR_SPACE
);
937 params
[SI_PARAM_SAMPLER
] = params
[SI_PARAM_CONST
];
938 params
[SI_PARAM_RESOURCE
] = LLVMPointerType(LLVMVectorType(i8
, 32), CONST_ADDR_SPACE
);
940 if (si_shader_ctx
->type
== TGSI_PROCESSOR_VERTEX
) {
941 params
[SI_PARAM_VERTEX_BUFFER
] = params
[SI_PARAM_SAMPLER
];
942 params
[SI_PARAM_VERTEX_INDEX
] = i32
;
943 radeon_llvm_create_func(&si_shader_ctx
->radeon_bld
, params
, 5);
946 params
[SI_PARAM_PRIM_MASK
] = i32
;
947 params
[SI_PARAM_PERSP_SAMPLE
] = v2i32
;
948 params
[SI_PARAM_PERSP_CENTER
] = v2i32
;
949 params
[SI_PARAM_PERSP_CENTROID
] = v2i32
;
950 params
[SI_PARAM_PERSP_PULL_MODEL
] = v3i32
;
951 params
[SI_PARAM_LINEAR_SAMPLE
] = v2i32
;
952 params
[SI_PARAM_LINEAR_CENTER
] = v2i32
;
953 params
[SI_PARAM_LINEAR_CENTROID
] = v2i32
;
954 params
[SI_PARAM_LINE_STIPPLE_TEX
] = f32
;
955 params
[SI_PARAM_POS_X_FLOAT
] = f32
;
956 params
[SI_PARAM_POS_Y_FLOAT
] = f32
;
957 params
[SI_PARAM_POS_Z_FLOAT
] = f32
;
958 params
[SI_PARAM_POS_W_FLOAT
] = f32
;
959 params
[SI_PARAM_FRONT_FACE
] = f32
;
960 params
[SI_PARAM_ANCILLARY
] = f32
;
961 params
[SI_PARAM_SAMPLE_COVERAGE
] = f32
;
962 params
[SI_PARAM_POS_FIXED_PT
] = f32
;
963 radeon_llvm_create_func(&si_shader_ctx
->radeon_bld
, params
, 20);
966 radeon_llvm_shader_type(si_shader_ctx
->radeon_bld
.main_fn
, si_shader_ctx
->type
);
967 for (i
= SI_PARAM_CONST
; i
<= SI_PARAM_VERTEX_BUFFER
; ++i
) {
968 LLVMValueRef P
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, i
);
969 LLVMAddAttribute(P
, LLVMInRegAttribute
);
973 static void preload_constants(struct si_shader_context
*si_shader_ctx
)
975 struct lp_build_tgsi_context
* bld_base
= &si_shader_ctx
->radeon_bld
.soa
.bld_base
;
976 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
977 const struct tgsi_shader_info
* info
= bld_base
->info
;
979 unsigned i
, num_const
= info
->file_max
[TGSI_FILE_CONSTANT
] + 1;
986 /* Allocate space for the constant values */
987 si_shader_ctx
->constants
= CALLOC(num_const
* 4, sizeof(LLVMValueRef
));
989 /* Load the resource descriptor */
990 ptr
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, SI_PARAM_CONST
);
991 si_shader_ctx
->const_resource
= build_indexed_load(si_shader_ctx
, ptr
, bld_base
->uint_bld
.zero
);
993 /* Load the constants, we rely on the code sinking to do the rest */
994 for (i
= 0; i
< num_const
* 4; ++i
) {
995 LLVMValueRef args
[2] = {
996 si_shader_ctx
->const_resource
,
997 lp_build_const_int32(gallivm
, i
* 4)
999 si_shader_ctx
->constants
[i
] = build_intrinsic(gallivm
->builder
, "llvm.SI.load.const",
1000 bld_base
->base
.elem_type
, args
, 2, LLVMReadNoneAttribute
| LLVMNoUnwindAttribute
);
1004 static void preload_samplers(struct si_shader_context
*si_shader_ctx
)
1006 struct lp_build_tgsi_context
* bld_base
= &si_shader_ctx
->radeon_bld
.soa
.bld_base
;
1007 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
1008 const struct tgsi_shader_info
* info
= bld_base
->info
;
1010 unsigned i
, num_samplers
= info
->file_max
[TGSI_FILE_SAMPLER
] + 1;
1012 LLVMValueRef res_ptr
, samp_ptr
;
1013 LLVMValueRef offset
;
1015 if (num_samplers
== 0)
1018 /* Allocate space for the values */
1019 si_shader_ctx
->resources
= CALLOC(num_samplers
, sizeof(LLVMValueRef
));
1020 si_shader_ctx
->samplers
= CALLOC(num_samplers
, sizeof(LLVMValueRef
));
1022 res_ptr
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, SI_PARAM_RESOURCE
);
1023 samp_ptr
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, SI_PARAM_SAMPLER
);
1025 /* Load the resources and samplers, we rely on the code sinking to do the rest */
1026 for (i
= 0; i
< num_samplers
; ++i
) {
1029 offset
= lp_build_const_int32(gallivm
, i
);
1030 si_shader_ctx
->resources
[i
] = build_indexed_load(si_shader_ctx
, res_ptr
, offset
);
1033 offset
= lp_build_const_int32(gallivm
, i
);
1034 si_shader_ctx
->samplers
[i
] = build_indexed_load(si_shader_ctx
, samp_ptr
, offset
);
1038 int si_pipe_shader_create(
1039 struct pipe_context
*ctx
,
1040 struct si_pipe_shader
*shader
,
1041 struct si_shader_key key
)
1043 struct r600_context
*rctx
= (struct r600_context
*)ctx
;
1044 struct si_pipe_shader_selector
*sel
= shader
->selector
;
1045 struct si_shader_context si_shader_ctx
;
1046 struct tgsi_shader_info shader_info
;
1047 struct lp_build_tgsi_context
* bld_base
;
1049 unsigned char * inst_bytes
;
1050 unsigned inst_byte_count
;
1055 dump
= debug_get_bool_option("RADEON_DUMP_SHADERS", FALSE
);
1057 assert(shader
->shader
.noutput
== 0);
1058 assert(shader
->shader
.ninterp
== 0);
1059 assert(shader
->shader
.ninput
== 0);
1061 memset(&si_shader_ctx
, 0, sizeof(si_shader_ctx
));
1062 radeon_llvm_context_init(&si_shader_ctx
.radeon_bld
);
1063 bld_base
= &si_shader_ctx
.radeon_bld
.soa
.bld_base
;
1065 tgsi_scan_shader(sel
->tokens
, &shader_info
);
1066 shader
->shader
.uses_kill
= shader_info
.uses_kill
;
1067 bld_base
->info
= &shader_info
;
1068 bld_base
->emit_fetch_funcs
[TGSI_FILE_CONSTANT
] = fetch_constant
;
1069 bld_base
->emit_epilogue
= si_llvm_emit_epilogue
;
1071 bld_base
->op_actions
[TGSI_OPCODE_TEX
] = tex_action
;
1072 bld_base
->op_actions
[TGSI_OPCODE_TXB
] = txb_action
;
1073 bld_base
->op_actions
[TGSI_OPCODE_TXL
] = txl_action
;
1074 bld_base
->op_actions
[TGSI_OPCODE_TXP
] = tex_action
;
1076 si_shader_ctx
.radeon_bld
.load_input
= declare_input
;
1077 si_shader_ctx
.tokens
= sel
->tokens
;
1078 tgsi_parse_init(&si_shader_ctx
.parse
, si_shader_ctx
.tokens
);
1079 si_shader_ctx
.shader
= shader
;
1080 si_shader_ctx
.key
= key
;
1081 si_shader_ctx
.type
= si_shader_ctx
.parse
.FullHeader
.Processor
.Processor
;
1082 si_shader_ctx
.rctx
= rctx
;
1084 create_meta_data(&si_shader_ctx
);
1085 create_function(&si_shader_ctx
);
1086 preload_constants(&si_shader_ctx
);
1087 preload_samplers(&si_shader_ctx
);
1089 shader
->shader
.nr_cbufs
= rctx
->framebuffer
.nr_cbufs
;
1091 /* Dump TGSI code before doing TGSI->LLVM conversion in case the
1092 * conversion fails. */
1094 tgsi_dump(sel
->tokens
, 0);
1097 if (!lp_build_tgsi_llvm(bld_base
, sel
->tokens
)) {
1098 fprintf(stderr
, "Failed to translate shader from TGSI to LLVM\n");
1099 FREE(si_shader_ctx
.constants
);
1100 FREE(si_shader_ctx
.resources
);
1101 FREE(si_shader_ctx
.samplers
);
1105 radeon_llvm_finalize_module(&si_shader_ctx
.radeon_bld
);
1107 mod
= bld_base
->base
.gallivm
->module
;
1109 LLVMDumpModule(mod
);
1111 radeon_llvm_compile(mod
, &inst_bytes
, &inst_byte_count
, "SI", dump
);
1113 fprintf(stderr
, "SI CODE:\n");
1114 for (i
= 0; i
< inst_byte_count
; i
+=4 ) {
1115 fprintf(stderr
, "%02x%02x%02x%02x\n", inst_bytes
[i
+ 3],
1116 inst_bytes
[i
+ 2], inst_bytes
[i
+ 1],
1121 shader
->num_sgprs
= util_le32_to_cpu(*(uint32_t*)inst_bytes
);
1122 shader
->num_vgprs
= util_le32_to_cpu(*(uint32_t*)(inst_bytes
+ 4));
1123 shader
->spi_ps_input_ena
= util_le32_to_cpu(*(uint32_t*)(inst_bytes
+ 8));
1125 radeon_llvm_dispose(&si_shader_ctx
.radeon_bld
);
1126 tgsi_parse_free(&si_shader_ctx
.parse
);
1128 /* copy new shader */
1129 si_resource_reference(&shader
->bo
, NULL
);
1130 shader
->bo
= si_resource_create_custom(ctx
->screen
, PIPE_USAGE_IMMUTABLE
,
1131 inst_byte_count
- 12);
1132 if (shader
->bo
== NULL
) {
1133 FREE(si_shader_ctx
.constants
);
1134 FREE(si_shader_ctx
.resources
);
1135 FREE(si_shader_ctx
.samplers
);
1139 ptr
= (uint32_t*)rctx
->ws
->buffer_map(shader
->bo
->cs_buf
, rctx
->cs
, PIPE_TRANSFER_WRITE
);
1140 if (0 /*R600_BIG_ENDIAN*/) {
1141 for (i
= 0; i
< (inst_byte_count
-12)/4; ++i
) {
1142 ptr
[i
] = util_bswap32(*(uint32_t*)(inst_bytes
+12 + i
*4));
1145 memcpy(ptr
, inst_bytes
+ 12, inst_byte_count
- 12);
1147 rctx
->ws
->buffer_unmap(shader
->bo
->cs_buf
);
1149 FREE(si_shader_ctx
.constants
);
1150 FREE(si_shader_ctx
.resources
);
1151 FREE(si_shader_ctx
.samplers
);
1157 void si_pipe_shader_destroy(struct pipe_context
*ctx
, struct si_pipe_shader
*shader
)
1159 si_resource_reference(&shader
->bo
, NULL
);