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 "radeon_llvm.h"
37 #include "radeon_llvm_emit.h"
38 #include "tgsi/tgsi_info.h"
39 #include "tgsi/tgsi_parse.h"
40 #include "tgsi/tgsi_scan.h"
41 #include "tgsi/tgsi_dump.h"
43 #include "radeonsi_pipe.h"
44 #include "radeonsi_shader.h"
52 struct si_shader_context
54 struct radeon_llvm_context radeon_bld
;
55 struct r600_context
*rctx
;
56 struct tgsi_parse_context parse
;
57 struct tgsi_token
* tokens
;
58 struct si_pipe_shader
*shader
;
59 struct si_shader_key key
;
60 unsigned type
; /* TGSI_PROCESSOR_* specifies the type of shader. */
61 unsigned ninput_emitted
;
62 /* struct list_head inputs; */
63 /* unsigned * input_mappings *//* From TGSI to SI hw */
64 /* struct tgsi_shader_info info;*/
67 static struct si_shader_context
* si_shader_context(
68 struct lp_build_tgsi_context
* bld_base
)
70 return (struct si_shader_context
*)bld_base
;
74 #define PERSPECTIVE_BASE 0
77 #define SAMPLE_OFFSET 0
78 #define CENTER_OFFSET 2
79 #define CENTROID_OFSET 4
81 #define USE_SGPR_MAX_SUFFIX_LEN 5
82 #define CONST_ADDR_SPACE 2
83 #define USER_SGPR_ADDR_SPACE 8
86 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad
88 * @param offset The offset parameter specifies the number of
89 * elements to offset, not the number of bytes or dwords. An element is the
90 * the type pointed to by the base_ptr parameter (e.g. int is the element of
93 * When LLVM lowers the load instruction, it will convert the element offset
94 * into a dword offset automatically.
97 static LLVMValueRef
build_indexed_load(
98 struct gallivm_state
* gallivm
,
99 LLVMValueRef base_ptr
,
102 LLVMValueRef computed_ptr
= LLVMBuildGEP(
103 gallivm
->builder
, base_ptr
, &offset
, 1, "");
105 return LLVMBuildLoad(gallivm
->builder
, computed_ptr
, "");
108 static void declare_input_vs(
109 struct si_shader_context
* si_shader_ctx
,
110 unsigned input_index
,
111 const struct tgsi_full_declaration
*decl
)
113 LLVMValueRef t_list_ptr
;
114 LLVMValueRef t_offset
;
116 LLVMValueRef attribute_offset
;
117 LLVMValueRef buffer_index_reg
;
118 LLVMValueRef args
[3];
119 LLVMTypeRef vec4_type
;
121 struct lp_build_context
* base
= &si_shader_ctx
->radeon_bld
.soa
.bld_base
.base
;
122 //struct pipe_vertex_element *velem = &rctx->vertex_elements->elements[input_index];
125 /* Load the T list */
126 t_list_ptr
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, SI_PARAM_VERTEX_BUFFER
);
128 t_offset
= lp_build_const_int32(base
->gallivm
, input_index
);
130 t_list
= build_indexed_load(base
->gallivm
, t_list_ptr
, t_offset
);
132 /* Build the attribute offset */
133 attribute_offset
= lp_build_const_int32(base
->gallivm
, 0);
135 /* Load the buffer index, which is always stored in VGPR0
136 * for Vertex Shaders */
137 buffer_index_reg
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, SI_PARAM_VERTEX_INDEX
);
139 vec4_type
= LLVMVectorType(base
->elem_type
, 4);
141 args
[1] = attribute_offset
;
142 args
[2] = buffer_index_reg
;
143 input
= lp_build_intrinsic(base
->gallivm
->builder
,
144 "llvm.SI.vs.load.input", vec4_type
, args
, 3);
146 /* Break up the vec4 into individual components */
147 for (chan
= 0; chan
< 4; chan
++) {
148 LLVMValueRef llvm_chan
= lp_build_const_int32(base
->gallivm
, chan
);
149 /* XXX: Use a helper function for this. There is one in
151 si_shader_ctx
->radeon_bld
.inputs
[radeon_llvm_reg_index_soa(input_index
, chan
)] =
152 LLVMBuildExtractElement(base
->gallivm
->builder
,
153 input
, llvm_chan
, "");
157 static void declare_input_fs(
158 struct si_shader_context
* si_shader_ctx
,
159 unsigned input_index
,
160 const struct tgsi_full_declaration
*decl
)
162 const char * intr_name
;
164 struct si_shader
*shader
= &si_shader_ctx
->shader
->shader
;
165 struct lp_build_context
* base
=
166 &si_shader_ctx
->radeon_bld
.soa
.bld_base
.base
;
167 struct gallivm_state
* gallivm
= base
->gallivm
;
168 LLVMTypeRef input_type
= LLVMFloatTypeInContext(gallivm
->context
);
171 * [15:0] NewPrimMask (Bit mask for each quad. It is set it the
172 * quad begins a new primitive. Bit 0 always needs
174 * [32:16] ParamOffset
177 LLVMValueRef params
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, SI_PARAM_PRIM_MASK
);
178 LLVMValueRef attr_number
;
180 if (decl
->Semantic
.Name
== TGSI_SEMANTIC_POSITION
) {
181 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
182 LLVMValueRef args
[1];
184 radeon_llvm_reg_index_soa(input_index
, chan
);
185 args
[0] = lp_build_const_int32(gallivm
, chan
);
186 si_shader_ctx
->radeon_bld
.inputs
[soa_index
] =
187 build_intrinsic(base
->gallivm
->builder
,
188 "llvm.SI.fs.read.pos", input_type
,
189 args
, 1, LLVMReadNoneAttribute
);
192 /* RCP for fragcoord.w */
193 si_shader_ctx
->radeon_bld
.inputs
[soa_index
] =
194 LLVMBuildFDiv(gallivm
->builder
,
195 lp_build_const_float(gallivm
, 1.0f
),
196 si_shader_ctx
->radeon_bld
.inputs
[soa_index
],
202 if (decl
->Semantic
.Name
== TGSI_SEMANTIC_FACE
) {
203 LLVMValueRef face
, is_face_positive
;
205 face
= build_intrinsic(gallivm
->builder
,
206 "llvm.SI.fs.read.face",
208 NULL
, 0, LLVMReadNoneAttribute
);
209 is_face_positive
= LLVMBuildFCmp(gallivm
->builder
,
211 lp_build_const_float(gallivm
, 0.0f
),
214 si_shader_ctx
->radeon_bld
.inputs
[radeon_llvm_reg_index_soa(input_index
, 0)] =
215 LLVMBuildSelect(gallivm
->builder
,
217 lp_build_const_float(gallivm
, 1.0f
),
218 lp_build_const_float(gallivm
, 0.0f
),
220 si_shader_ctx
->radeon_bld
.inputs
[radeon_llvm_reg_index_soa(input_index
, 1)] =
221 si_shader_ctx
->radeon_bld
.inputs
[radeon_llvm_reg_index_soa(input_index
, 2)] =
222 lp_build_const_float(gallivm
, 0.0f
);
223 si_shader_ctx
->radeon_bld
.inputs
[radeon_llvm_reg_index_soa(input_index
, 3)] =
224 lp_build_const_float(gallivm
, 1.0f
);
229 shader
->input
[input_index
].param_offset
= shader
->ninterp
++;
230 attr_number
= lp_build_const_int32(gallivm
,
231 shader
->input
[input_index
].param_offset
);
233 /* XXX: Handle all possible interpolation modes */
234 switch (decl
->Interp
.Interpolate
) {
235 case TGSI_INTERPOLATE_COLOR
:
236 if (si_shader_ctx
->key
.flatshade
) {
237 intr_name
= "llvm.SI.fs.interp.constant";
239 if (decl
->Interp
.Centroid
)
240 intr_name
= "llvm.SI.fs.interp.persp.centroid";
242 intr_name
= "llvm.SI.fs.interp.persp.center";
245 case TGSI_INTERPOLATE_CONSTANT
:
246 intr_name
= "llvm.SI.fs.interp.constant";
248 case TGSI_INTERPOLATE_LINEAR
:
249 if (decl
->Interp
.Centroid
)
250 intr_name
= "llvm.SI.fs.interp.linear.centroid";
252 intr_name
= "llvm.SI.fs.interp.linear.center";
254 case TGSI_INTERPOLATE_PERSPECTIVE
:
255 if (decl
->Interp
.Centroid
)
256 intr_name
= "llvm.SI.fs.interp.persp.centroid";
258 intr_name
= "llvm.SI.fs.interp.persp.center";
261 fprintf(stderr
, "Warning: Unhandled interpolation mode.\n");
265 if (!si_shader_ctx
->ninput_emitted
++) {
266 /* Enable whole quad mode */
267 lp_build_intrinsic(gallivm
->builder
,
269 LLVMVoidTypeInContext(gallivm
->context
),
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
[3];
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
= build_intrinsic(gallivm
->builder
,
283 "llvm.SI.fs.read.face",
285 NULL
, 0, LLVMReadNoneAttribute
);
286 is_face_positive
= LLVMBuildFCmp(gallivm
->builder
,
288 lp_build_const_float(gallivm
, 0.0f
),
292 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
293 LLVMValueRef llvm_chan
= lp_build_const_int32(gallivm
, chan
);
294 unsigned soa_index
= radeon_llvm_reg_index_soa(input_index
, chan
);
295 LLVMValueRef front
, back
;
298 args
[1] = attr_number
;
299 front
= build_intrinsic(base
->gallivm
->builder
, intr_name
,
300 input_type
, args
, 3, LLVMReadOnlyAttribute
);
302 args
[1] = back_attr_number
;
303 back
= build_intrinsic(base
->gallivm
->builder
, intr_name
,
304 input_type
, args
, 3, LLVMReadOnlyAttribute
);
306 si_shader_ctx
->radeon_bld
.inputs
[soa_index
] =
307 LLVMBuildSelect(gallivm
->builder
,
316 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
317 LLVMValueRef args
[3];
318 LLVMValueRef llvm_chan
= lp_build_const_int32(gallivm
, chan
);
319 unsigned soa_index
= radeon_llvm_reg_index_soa(input_index
, chan
);
321 args
[1] = attr_number
;
323 si_shader_ctx
->radeon_bld
.inputs
[soa_index
] =
324 build_intrinsic(base
->gallivm
->builder
, intr_name
,
325 input_type
, args
, 3, LLVMReadOnlyAttribute
);
330 static void declare_input(
331 struct radeon_llvm_context
* radeon_bld
,
332 unsigned input_index
,
333 const struct tgsi_full_declaration
*decl
)
335 struct si_shader_context
* si_shader_ctx
=
336 si_shader_context(&radeon_bld
->soa
.bld_base
);
337 if (si_shader_ctx
->type
== TGSI_PROCESSOR_VERTEX
) {
338 declare_input_vs(si_shader_ctx
, input_index
, decl
);
339 } else if (si_shader_ctx
->type
== TGSI_PROCESSOR_FRAGMENT
) {
340 declare_input_fs(si_shader_ctx
, input_index
, decl
);
342 fprintf(stderr
, "Warning: Unsupported shader type,\n");
346 static LLVMValueRef
fetch_constant(
347 struct lp_build_tgsi_context
* bld_base
,
348 const struct tgsi_full_src_register
*reg
,
349 enum tgsi_opcode_type type
,
352 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
353 struct lp_build_context
* base
= &bld_base
->base
;
356 LLVMValueRef const_ptr
;
360 if (swizzle
== LP_CHAN_ALL
) {
362 LLVMValueRef values
[4];
363 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
)
364 values
[chan
] = fetch_constant(bld_base
, reg
, type
, chan
);
366 return lp_build_gather_values(bld_base
->base
.gallivm
, values
, 4);
369 /* currently not supported */
370 if (reg
->Register
.Indirect
) {
372 load
= lp_build_const_int32(base
->gallivm
, 0);
373 return bitcast(bld_base
, type
, load
);
376 const_ptr
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, SI_PARAM_CONST
);
378 /* XXX: This assumes that the constant buffer is not packed, so
379 * CONST[0].x will have an offset of 0 and CONST[1].x will have an
381 idx
= (reg
->Register
.Index
* 4) + swizzle
;
382 offset
= lp_build_const_int32(base
->gallivm
, idx
);
384 load
= build_indexed_load(base
->gallivm
, const_ptr
, offset
);
385 return bitcast(bld_base
, type
, load
);
388 /* Initialize arguments for the shader export intrinsic */
389 static void si_llvm_init_export_args(struct lp_build_tgsi_context
*bld_base
,
390 struct tgsi_full_declaration
*d
,
395 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
396 struct lp_build_context
*uint
=
397 &si_shader_ctx
->radeon_bld
.soa
.bld_base
.uint_bld
;
398 struct lp_build_context
*base
= &bld_base
->base
;
399 unsigned compressed
= 0;
402 if (si_shader_ctx
->type
== TGSI_PROCESSOR_FRAGMENT
) {
403 int cbuf
= target
- V_008DFC_SQ_EXP_MRT
;
405 if (cbuf
>= 0 && cbuf
< 8) {
406 compressed
= (si_shader_ctx
->key
.export_16bpc
>> cbuf
) & 0x1;
409 si_shader_ctx
->shader
->spi_shader_col_format
|=
410 V_028714_SPI_SHADER_FP16_ABGR
<< (4 * cbuf
);
412 si_shader_ctx
->shader
->spi_shader_col_format
|=
413 V_028714_SPI_SHADER_32_ABGR
<< (4 * cbuf
);
418 /* Pixel shader needs to pack output values before export */
419 for (chan
= 0; chan
< 2; chan
++ ) {
420 LLVMValueRef
*out_ptr
=
421 si_shader_ctx
->radeon_bld
.soa
.outputs
[index
];
422 args
[0] = LLVMBuildLoad(base
->gallivm
->builder
,
423 out_ptr
[2 * chan
], "");
424 args
[1] = LLVMBuildLoad(base
->gallivm
->builder
,
425 out_ptr
[2 * chan
+ 1], "");
427 build_intrinsic(base
->gallivm
->builder
,
429 LLVMInt32TypeInContext(base
->gallivm
->context
),
431 LLVMReadNoneAttribute
);
432 args
[chan
+ 7] = args
[chan
+ 5] =
433 LLVMBuildBitCast(base
->gallivm
->builder
,
435 LLVMFloatTypeInContext(base
->gallivm
->context
),
442 for (chan
= 0; chan
< 4; chan
++ ) {
443 LLVMValueRef out_ptr
=
444 si_shader_ctx
->radeon_bld
.soa
.outputs
[index
][chan
];
445 /* +5 because the first output value will be
446 * the 6th argument to the intrinsic. */
447 args
[chan
+ 5] = LLVMBuildLoad(base
->gallivm
->builder
,
451 /* Clear COMPR flag */
452 args
[4] = uint
->zero
;
455 /* XXX: This controls which components of the output
456 * registers actually get exported. (e.g bit 0 means export
457 * X component, bit 1 means export Y component, etc.) I'm
458 * hard coding this to 0xf for now. In the future, we might
459 * want to do something else. */
460 args
[0] = lp_build_const_int32(base
->gallivm
, 0xf);
462 /* Specify whether the EXEC mask represents the valid mask */
463 args
[1] = uint
->zero
;
465 /* Specify whether this is the last export */
466 args
[2] = uint
->zero
;
468 /* Specify the target we are exporting */
469 args
[3] = lp_build_const_int32(base
->gallivm
, target
);
471 /* XXX: We probably need to keep track of the output
472 * values, so we know what we are passing to the next
476 static void si_alpha_test(struct lp_build_tgsi_context
*bld_base
,
479 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
480 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
482 if (si_shader_ctx
->key
.alpha_func
!= PIPE_FUNC_NEVER
) {
483 LLVMValueRef out_ptr
= si_shader_ctx
->radeon_bld
.soa
.outputs
[index
][3];
484 LLVMValueRef alpha_pass
=
485 lp_build_cmp(&bld_base
->base
,
486 si_shader_ctx
->key
.alpha_func
,
487 LLVMBuildLoad(gallivm
->builder
, out_ptr
, ""),
488 lp_build_const_float(gallivm
, si_shader_ctx
->key
.alpha_ref
));
490 lp_build_select(&bld_base
->base
,
492 lp_build_const_float(gallivm
, 1.0f
),
493 lp_build_const_float(gallivm
, -1.0f
));
495 build_intrinsic(gallivm
->builder
,
497 LLVMVoidTypeInContext(gallivm
->context
),
500 build_intrinsic(gallivm
->builder
,
502 LLVMVoidTypeInContext(gallivm
->context
),
507 /* XXX: This is partially implemented for VS only at this point. It is not complete */
508 static void si_llvm_emit_epilogue(struct lp_build_tgsi_context
* bld_base
)
510 struct si_shader_context
* si_shader_ctx
= si_shader_context(bld_base
);
511 struct si_shader
* shader
= &si_shader_ctx
->shader
->shader
;
512 struct lp_build_context
* base
= &bld_base
->base
;
513 struct lp_build_context
* uint
=
514 &si_shader_ctx
->radeon_bld
.soa
.bld_base
.uint_bld
;
515 struct tgsi_parse_context
*parse
= &si_shader_ctx
->parse
;
516 LLVMValueRef args
[9];
517 LLVMValueRef last_args
[9] = { 0 };
518 unsigned color_count
= 0;
519 unsigned param_count
= 0;
520 int depth_index
= -1, stencil_index
= -1;
522 while (!tgsi_parse_end_of_tokens(parse
)) {
523 struct tgsi_full_declaration
*d
=
524 &parse
->FullToken
.FullDeclaration
;
529 tgsi_parse_token(parse
);
531 if (parse
->FullToken
.Token
.Type
== TGSI_TOKEN_TYPE_PROPERTY
&&
532 parse
->FullToken
.FullProperty
.Property
.PropertyName
==
533 TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
)
534 shader
->fs_write_all
= TRUE
;
536 if (parse
->FullToken
.Token
.Type
!= TGSI_TOKEN_TYPE_DECLARATION
)
539 switch (d
->Declaration
.File
) {
540 case TGSI_FILE_INPUT
:
541 i
= shader
->ninput
++;
542 shader
->input
[i
].name
= d
->Semantic
.Name
;
543 shader
->input
[i
].sid
= d
->Semantic
.Index
;
544 shader
->input
[i
].interpolate
= d
->Interp
.Interpolate
;
545 shader
->input
[i
].centroid
= d
->Interp
.Centroid
;
548 case TGSI_FILE_OUTPUT
:
549 i
= shader
->noutput
++;
550 shader
->output
[i
].name
= d
->Semantic
.Name
;
551 shader
->output
[i
].sid
= d
->Semantic
.Index
;
552 shader
->output
[i
].interpolate
= d
->Interp
.Interpolate
;
559 for (index
= d
->Range
.First
; index
<= d
->Range
.Last
; index
++) {
560 /* Select the correct target */
561 switch(d
->Semantic
.Name
) {
562 case TGSI_SEMANTIC_PSIZE
:
563 target
= V_008DFC_SQ_EXP_POS
;
565 case TGSI_SEMANTIC_POSITION
:
566 if (si_shader_ctx
->type
== TGSI_PROCESSOR_VERTEX
) {
567 target
= V_008DFC_SQ_EXP_POS
;
573 case TGSI_SEMANTIC_STENCIL
:
574 stencil_index
= index
;
576 case TGSI_SEMANTIC_COLOR
:
577 if (si_shader_ctx
->type
== TGSI_PROCESSOR_VERTEX
) {
578 case TGSI_SEMANTIC_BCOLOR
:
579 target
= V_008DFC_SQ_EXP_PARAM
+ param_count
;
580 shader
->output
[i
].param_offset
= param_count
;
583 target
= V_008DFC_SQ_EXP_MRT
+ color_count
;
584 if (color_count
== 0 &&
585 si_shader_ctx
->key
.alpha_func
!= PIPE_FUNC_ALWAYS
)
586 si_alpha_test(bld_base
, index
);
591 case TGSI_SEMANTIC_FOG
:
592 case TGSI_SEMANTIC_GENERIC
:
593 target
= V_008DFC_SQ_EXP_PARAM
+ param_count
;
594 shader
->output
[i
].param_offset
= param_count
;
600 "Warning: SI unhandled output type:%d\n",
604 si_llvm_init_export_args(bld_base
, d
, index
, target
, args
);
606 if (si_shader_ctx
->type
== TGSI_PROCESSOR_VERTEX
?
607 (d
->Semantic
.Name
== TGSI_SEMANTIC_POSITION
) :
608 (d
->Semantic
.Name
== TGSI_SEMANTIC_COLOR
)) {
610 lp_build_intrinsic(base
->gallivm
->builder
,
612 LLVMVoidTypeInContext(base
->gallivm
->context
),
616 memcpy(last_args
, args
, sizeof(args
));
618 lp_build_intrinsic(base
->gallivm
->builder
,
620 LLVMVoidTypeInContext(base
->gallivm
->context
),
627 if (depth_index
>= 0 || stencil_index
>= 0) {
628 LLVMValueRef out_ptr
;
631 /* Specify the target we are exporting */
632 args
[3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_MRTZ
);
634 if (depth_index
>= 0) {
635 out_ptr
= si_shader_ctx
->radeon_bld
.soa
.outputs
[depth_index
][2];
636 args
[5] = LLVMBuildLoad(base
->gallivm
->builder
, out_ptr
, "");
639 if (stencil_index
< 0) {
646 if (stencil_index
>= 0) {
647 out_ptr
= si_shader_ctx
->radeon_bld
.soa
.outputs
[stencil_index
][1];
650 args
[6] = LLVMBuildLoad(base
->gallivm
->builder
, out_ptr
, "");
657 /* Specify which components to enable */
658 args
[0] = lp_build_const_int32(base
->gallivm
, mask
);
662 args
[4] = uint
->zero
;
665 lp_build_intrinsic(base
->gallivm
->builder
,
667 LLVMVoidTypeInContext(base
->gallivm
->context
),
670 memcpy(last_args
, args
, sizeof(args
));
674 assert(si_shader_ctx
->type
== TGSI_PROCESSOR_FRAGMENT
);
676 /* Specify which components to enable */
677 last_args
[0] = lp_build_const_int32(base
->gallivm
, 0x0);
679 /* Specify the target we are exporting */
680 last_args
[3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_MRT
);
682 /* Set COMPR flag to zero to export data as 32-bit */
683 last_args
[4] = uint
->zero
;
686 last_args
[5]= uint
->zero
;
687 last_args
[6]= uint
->zero
;
688 last_args
[7]= uint
->zero
;
689 last_args
[8]= uint
->zero
;
691 si_shader_ctx
->shader
->spi_shader_col_format
|=
692 V_028714_SPI_SHADER_32_ABGR
;
695 /* Specify whether the EXEC mask represents the valid mask */
696 last_args
[1] = lp_build_const_int32(base
->gallivm
,
697 si_shader_ctx
->type
== TGSI_PROCESSOR_FRAGMENT
);
699 if (shader
->fs_write_all
&& shader
->nr_cbufs
> 1) {
702 /* Specify that this is not yet the last export */
703 last_args
[2] = lp_build_const_int32(base
->gallivm
, 0);
705 for (i
= 1; i
< shader
->nr_cbufs
; i
++) {
706 /* Specify the target we are exporting */
707 last_args
[3] = lp_build_const_int32(base
->gallivm
,
708 V_008DFC_SQ_EXP_MRT
+ i
);
710 lp_build_intrinsic(base
->gallivm
->builder
,
712 LLVMVoidTypeInContext(base
->gallivm
->context
),
715 si_shader_ctx
->shader
->spi_shader_col_format
|=
716 si_shader_ctx
->shader
->spi_shader_col_format
<< 4;
719 last_args
[3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_MRT
);
722 /* Specify that this is the last export */
723 last_args
[2] = lp_build_const_int32(base
->gallivm
, 1);
725 lp_build_intrinsic(base
->gallivm
->builder
,
727 LLVMVoidTypeInContext(base
->gallivm
->context
),
730 /* XXX: Look up what this function does */
731 /* ctx->shader->output[i].spi_sid = r600_spi_sid(&ctx->shader->output[i]);*/
734 static void tex_fetch_args(
735 struct lp_build_tgsi_context
* bld_base
,
736 struct lp_build_emit_data
* emit_data
)
738 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
739 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
740 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
741 unsigned opcode
= inst
->Instruction
.Opcode
;
742 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 ptr
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, SI_PARAM_RESOURCE
);
863 offset
= lp_build_const_int32(bld_base
->base
.gallivm
,
864 emit_data
->inst
->Src
[1].Register
.Index
);
865 emit_data
->args
[2] = build_indexed_load(bld_base
->base
.gallivm
,
869 ptr
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, SI_PARAM_SAMPLER
);
870 offset
= lp_build_const_int32(bld_base
->base
.gallivm
,
871 emit_data
->inst
->Src
[1].Register
.Index
);
872 emit_data
->args
[3] = build_indexed_load(bld_base
->base
.gallivm
,
876 emit_data
->args
[4] = lp_build_const_int32(bld_base
->base
.gallivm
, target
);
878 emit_data
->arg_count
= 5;
879 /* XXX: To optimize, we could use a float or v2f32, if the last bits of
880 * the writemask are clear */
881 emit_data
->dst_type
= LLVMVectorType(
882 LLVMFloatTypeInContext(bld_base
->base
.gallivm
->context
),
886 static void build_tex_intrinsic(const struct lp_build_tgsi_action
* action
,
887 struct lp_build_tgsi_context
* bld_base
,
888 struct lp_build_emit_data
* emit_data
)
890 struct lp_build_context
* base
= &bld_base
->base
;
893 sprintf(intr_name
, "%sv%ui32", action
->intr_name
,
894 LLVMGetVectorSize(LLVMTypeOf(emit_data
->args
[1])));
896 emit_data
->output
[emit_data
->chan
] = lp_build_intrinsic(
897 base
->gallivm
->builder
, intr_name
, emit_data
->dst_type
,
898 emit_data
->args
, emit_data
->arg_count
);
901 static const struct lp_build_tgsi_action tex_action
= {
902 .fetch_args
= tex_fetch_args
,
903 .emit
= build_tex_intrinsic
,
904 .intr_name
= "llvm.SI.sample."
907 static const struct lp_build_tgsi_action txb_action
= {
908 .fetch_args
= tex_fetch_args
,
909 .emit
= build_tex_intrinsic
,
910 .intr_name
= "llvm.SI.sampleb."
913 static const struct lp_build_tgsi_action txl_action
= {
914 .fetch_args
= tex_fetch_args
,
915 .emit
= build_tex_intrinsic
,
916 .intr_name
= "llvm.SI.samplel."
919 static void create_function(struct si_shader_context
*si_shader_ctx
)
921 struct gallivm_state
*gallivm
= si_shader_ctx
->radeon_bld
.soa
.bld_base
.base
.gallivm
;
922 LLVMTypeRef params
[5], f
, i8
, i32
;
925 f
= LLVMFloatTypeInContext(gallivm
->context
);
926 i8
= LLVMInt8TypeInContext(gallivm
->context
);
927 i32
= LLVMInt32TypeInContext(gallivm
->context
);
929 params
[SI_PARAM_CONST
] = LLVMPointerType(f
, CONST_ADDR_SPACE
);
930 params
[SI_PARAM_SAMPLER
] = LLVMPointerType(LLVMVectorType(i8
, 16), CONST_ADDR_SPACE
);
931 params
[SI_PARAM_RESOURCE
] = LLVMPointerType(LLVMVectorType(i8
, 32), CONST_ADDR_SPACE
);
933 if (si_shader_ctx
->type
== TGSI_PROCESSOR_VERTEX
) {
934 params
[SI_PARAM_VERTEX_BUFFER
] = params
[SI_PARAM_SAMPLER
];
935 params
[SI_PARAM_VERTEX_INDEX
] = i32
;
936 radeon_llvm_create_func(&si_shader_ctx
->radeon_bld
, params
, 5);
938 params
[SI_PARAM_PRIM_MASK
] = LLVMInt32TypeInContext(gallivm
->context
);
939 radeon_llvm_create_func(&si_shader_ctx
->radeon_bld
, params
, 4);
942 radeon_llvm_shader_type(si_shader_ctx
->radeon_bld
.main_fn
, si_shader_ctx
->type
);
943 for (i
= SI_PARAM_CONST
; i
<= SI_PARAM_VERTEX_BUFFER
; ++i
) {
944 LLVMValueRef P
= LLVMGetParam(si_shader_ctx
->radeon_bld
.main_fn
, i
);
945 LLVMAddAttribute(P
, LLVMInRegAttribute
);
949 int si_pipe_shader_create(
950 struct pipe_context
*ctx
,
951 struct si_pipe_shader
*shader
,
952 struct si_shader_key key
)
954 struct r600_context
*rctx
= (struct r600_context
*)ctx
;
955 struct si_pipe_shader_selector
*sel
= shader
->selector
;
956 struct si_shader_context si_shader_ctx
;
957 struct tgsi_shader_info shader_info
;
958 struct lp_build_tgsi_context
* bld_base
;
960 unsigned char * inst_bytes
;
961 unsigned inst_byte_count
;
966 dump
= debug_get_bool_option("RADEON_DUMP_SHADERS", FALSE
);
968 assert(shader
->shader
.noutput
== 0);
969 assert(shader
->shader
.ninterp
== 0);
970 assert(shader
->shader
.ninput
== 0);
972 memset(&si_shader_ctx
, 0, sizeof(si_shader_ctx
));
973 radeon_llvm_context_init(&si_shader_ctx
.radeon_bld
);
974 bld_base
= &si_shader_ctx
.radeon_bld
.soa
.bld_base
;
976 tgsi_scan_shader(sel
->tokens
, &shader_info
);
977 if (shader_info
.indirect_files
!= 0) {
978 fprintf(stderr
, "Indirect addressing not fully handled yet\n");
982 shader
->shader
.uses_kill
= shader_info
.uses_kill
;
983 bld_base
->info
= &shader_info
;
984 bld_base
->emit_fetch_funcs
[TGSI_FILE_CONSTANT
] = fetch_constant
;
985 bld_base
->emit_epilogue
= si_llvm_emit_epilogue
;
987 bld_base
->op_actions
[TGSI_OPCODE_TEX
] = tex_action
;
988 bld_base
->op_actions
[TGSI_OPCODE_TXB
] = txb_action
;
989 bld_base
->op_actions
[TGSI_OPCODE_TXL
] = txl_action
;
990 bld_base
->op_actions
[TGSI_OPCODE_TXP
] = tex_action
;
992 si_shader_ctx
.radeon_bld
.load_input
= declare_input
;
993 si_shader_ctx
.tokens
= sel
->tokens
;
994 tgsi_parse_init(&si_shader_ctx
.parse
, si_shader_ctx
.tokens
);
995 si_shader_ctx
.shader
= shader
;
996 si_shader_ctx
.key
= key
;
997 si_shader_ctx
.type
= si_shader_ctx
.parse
.FullHeader
.Processor
.Processor
;
998 si_shader_ctx
.rctx
= rctx
;
1000 create_function(&si_shader_ctx
);
1002 shader
->shader
.nr_cbufs
= rctx
->framebuffer
.nr_cbufs
;
1004 /* Dump TGSI code before doing TGSI->LLVM conversion in case the
1005 * conversion fails. */
1007 tgsi_dump(sel
->tokens
, 0);
1010 if (!lp_build_tgsi_llvm(bld_base
, sel
->tokens
)) {
1011 fprintf(stderr
, "Failed to translate shader from TGSI to LLVM\n");
1015 radeon_llvm_finalize_module(&si_shader_ctx
.radeon_bld
);
1017 mod
= bld_base
->base
.gallivm
->module
;
1019 LLVMDumpModule(mod
);
1021 radeon_llvm_compile(mod
, &inst_bytes
, &inst_byte_count
, "SI", dump
);
1023 fprintf(stderr
, "SI CODE:\n");
1024 for (i
= 0; i
< inst_byte_count
; i
+=4 ) {
1025 fprintf(stderr
, "%02x%02x%02x%02x\n", inst_bytes
[i
+ 3],
1026 inst_bytes
[i
+ 2], inst_bytes
[i
+ 1],
1031 shader
->num_sgprs
= util_le32_to_cpu(*(uint32_t*)inst_bytes
);
1032 shader
->num_vgprs
= util_le32_to_cpu(*(uint32_t*)(inst_bytes
+ 4));
1033 shader
->spi_ps_input_ena
= util_le32_to_cpu(*(uint32_t*)(inst_bytes
+ 8));
1035 radeon_llvm_dispose(&si_shader_ctx
.radeon_bld
);
1036 tgsi_parse_free(&si_shader_ctx
.parse
);
1038 /* copy new shader */
1039 si_resource_reference(&shader
->bo
, NULL
);
1040 shader
->bo
= si_resource_create_custom(ctx
->screen
, PIPE_USAGE_IMMUTABLE
,
1041 inst_byte_count
- 12);
1042 if (shader
->bo
== NULL
) {
1046 ptr
= (uint32_t*)rctx
->ws
->buffer_map(shader
->bo
->cs_buf
, rctx
->cs
, PIPE_TRANSFER_WRITE
);
1047 if (0 /*R600_BIG_ENDIAN*/) {
1048 for (i
= 0; i
< (inst_byte_count
-12)/4; ++i
) {
1049 ptr
[i
] = util_bswap32(*(uint32_t*)(inst_bytes
+12 + i
*4));
1052 memcpy(ptr
, inst_bytes
+ 12, inst_byte_count
- 12);
1054 rctx
->ws
->buffer_unmap(shader
->bo
->cs_buf
);
1061 void si_pipe_shader_destroy(struct pipe_context
*ctx
, struct si_pipe_shader
*shader
)
1063 si_resource_reference(&shader
->bo
, NULL
);