2 Copyright (C) Intel Corp. 2006. All Rights Reserved.
3 Intel funded Tungsten Graphics to
4 develop this 3D driver.
6 Permission is hereby granted, free of charge, to any person obtaining
7 a copy of this software and associated documentation files (the
8 "Software"), to deal in the Software without restriction, including
9 without limitation the rights to use, copy, modify, merge, publish,
10 distribute, sublicense, and/or sell copies of the Software, and to
11 permit persons to whom the Software is furnished to do so, subject to
12 the following conditions:
14 The above copyright notice and this permission notice (including the
15 next paragraph) shall be included in all copies or substantial
16 portions of the Software.
18 THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
19 EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
21 IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
22 LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
23 OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
24 WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 **********************************************************************/
29 * Keith Whitwell <keithw@vmware.com>
32 #include "brw_context.h"
34 #include "brw_state.h"
35 #include "main/enums.h"
36 #include "main/formats.h"
37 #include "main/fbobject.h"
38 #include "main/samplerobj.h"
39 #include "program/prog_parameter.h"
40 #include "program/program.h"
41 #include "intel_mipmap_tree.h"
43 #include "glsl/ralloc.h"
46 * Return a bitfield where bit n is set if barycentric interpolation mode n
47 * (see enum brw_wm_barycentric_interp_mode) is needed by the fragment shader.
50 brw_compute_barycentric_interp_modes(struct brw_context
*brw
,
51 bool shade_model_flat
,
52 bool persample_shading
,
53 const struct gl_fragment_program
*fprog
)
55 unsigned barycentric_interp_modes
= 0;
58 /* Loop through all fragment shader inputs to figure out what interpolation
59 * modes are in use, and set the appropriate bits in
60 * barycentric_interp_modes.
62 for (attr
= 0; attr
< VARYING_SLOT_MAX
; ++attr
) {
63 enum glsl_interp_qualifier interp_qualifier
=
64 fprog
->InterpQualifier
[attr
];
65 bool is_centroid
= (fprog
->IsCentroid
& BITFIELD64_BIT(attr
)) &&
67 bool is_sample
= (fprog
->IsSample
& BITFIELD64_BIT(attr
)) ||
69 bool is_gl_Color
= attr
== VARYING_SLOT_COL0
|| attr
== VARYING_SLOT_COL1
;
71 /* Ignore unused inputs. */
72 if (!(fprog
->Base
.InputsRead
& BITFIELD64_BIT(attr
)))
75 /* Ignore WPOS and FACE, because they don't require interpolation. */
76 if (attr
== VARYING_SLOT_POS
|| attr
== VARYING_SLOT_FACE
)
79 /* Determine the set (or sets) of barycentric coordinates needed to
80 * interpolate this variable. Note that when
81 * brw->needs_unlit_centroid_workaround is set, centroid interpolation
82 * uses PIXEL interpolation for unlit pixels and CENTROID interpolation
83 * for lit pixels, so we need both sets of barycentric coordinates.
85 if (interp_qualifier
== INTERP_QUALIFIER_NOPERSPECTIVE
) {
87 barycentric_interp_modes
|=
88 1 << BRW_WM_NONPERSPECTIVE_CENTROID_BARYCENTRIC
;
89 } else if (is_sample
) {
90 barycentric_interp_modes
|=
91 1 << BRW_WM_NONPERSPECTIVE_SAMPLE_BARYCENTRIC
;
93 if ((!is_centroid
&& !is_sample
) ||
94 brw
->needs_unlit_centroid_workaround
) {
95 barycentric_interp_modes
|=
96 1 << BRW_WM_NONPERSPECTIVE_PIXEL_BARYCENTRIC
;
98 } else if (interp_qualifier
== INTERP_QUALIFIER_SMOOTH
||
99 (!(shade_model_flat
&& is_gl_Color
) &&
100 interp_qualifier
== INTERP_QUALIFIER_NONE
)) {
102 barycentric_interp_modes
|=
103 1 << BRW_WM_PERSPECTIVE_CENTROID_BARYCENTRIC
;
104 } else if (is_sample
) {
105 barycentric_interp_modes
|=
106 1 << BRW_WM_PERSPECTIVE_SAMPLE_BARYCENTRIC
;
108 if ((!is_centroid
&& !is_sample
) ||
109 brw
->needs_unlit_centroid_workaround
) {
110 barycentric_interp_modes
|=
111 1 << BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC
;
116 return barycentric_interp_modes
;
120 brw_wm_prog_data_compare(const void *in_a
, const void *in_b
)
122 const struct brw_wm_prog_data
*a
= in_a
;
123 const struct brw_wm_prog_data
*b
= in_b
;
125 /* Compare the base structure. */
126 if (!brw_stage_prog_data_compare(&a
->base
, &b
->base
))
129 /* Compare the rest of the structure. */
130 const unsigned offset
= sizeof(struct brw_stage_prog_data
);
131 if (memcmp(((char *) a
) + offset
, ((char *) b
) + offset
,
132 sizeof(struct brw_wm_prog_data
) - offset
))
139 * All Mesa program -> GPU code generation goes through this function.
140 * Depending on the instructions used (i.e. flow control instructions)
141 * we'll use one of two code generators.
143 bool do_wm_prog(struct brw_context
*brw
,
144 struct gl_shader_program
*prog
,
145 struct brw_fragment_program
*fp
,
146 struct brw_wm_prog_key
*key
)
148 struct gl_context
*ctx
= &brw
->ctx
;
149 struct brw_wm_compile
*c
;
150 const GLuint
*program
;
151 struct gl_shader
*fs
= NULL
;
155 fs
= prog
->_LinkedShaders
[MESA_SHADER_FRAGMENT
];
157 c
= rzalloc(NULL
, struct brw_wm_compile
);
159 /* Allocate the references to the uniforms that will end up in the
160 * prog_data associated with the compiled program, and which will be freed
161 * by the state cache.
165 param_count
= fs
->num_uniform_components
;
167 param_count
= fp
->program
.Base
.Parameters
->NumParameters
* 4;
169 /* The backend also sometimes adds params for texture size. */
170 param_count
+= 2 * ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MaxTextureImageUnits
;
171 c
->prog_data
.base
.param
= rzalloc_array(NULL
, const float *, param_count
);
172 c
->prog_data
.base
.pull_param
=
173 rzalloc_array(NULL
, const float *, param_count
);
174 c
->prog_data
.base
.nr_params
= param_count
;
176 memcpy(&c
->key
, key
, sizeof(*key
));
178 c
->prog_data
.barycentric_interp_modes
=
179 brw_compute_barycentric_interp_modes(brw
, c
->key
.flat_shade
,
180 c
->key
.persample_shading
,
183 program
= brw_wm_fs_emit(brw
, c
, &fp
->program
, prog
, &program_size
);
187 /* Scratch space is used for register spilling */
188 if (c
->last_scratch
) {
189 perf_debug("Fragment shader triggered register spilling. "
190 "Try reducing the number of live scalar values to "
191 "improve performance.\n");
193 c
->prog_data
.total_scratch
= brw_get_scratch_size(c
->last_scratch
);
195 brw_get_scratch_bo(brw
, &brw
->wm
.base
.scratch_bo
,
196 c
->prog_data
.total_scratch
* brw
->max_wm_threads
);
199 if (unlikely(INTEL_DEBUG
& DEBUG_WM
))
200 fprintf(stderr
, "\n");
202 brw_upload_cache(&brw
->cache
, BRW_WM_PROG
,
203 &c
->key
, sizeof(c
->key
),
204 program
, program_size
,
205 &c
->prog_data
, sizeof(c
->prog_data
),
206 &brw
->wm
.base
.prog_offset
, &brw
->wm
.prog_data
);
214 key_debug(struct brw_context
*brw
, const char *name
, int a
, int b
)
217 perf_debug(" %s %d->%d\n", name
, a
, b
);
225 brw_debug_recompile_sampler_key(struct brw_context
*brw
,
226 const struct brw_sampler_prog_key_data
*old_key
,
227 const struct brw_sampler_prog_key_data
*key
)
231 for (unsigned int i
= 0; i
< MAX_SAMPLERS
; i
++) {
232 found
|= key_debug(brw
, "EXT_texture_swizzle or DEPTH_TEXTURE_MODE",
233 old_key
->swizzles
[i
], key
->swizzles
[i
]);
235 found
|= key_debug(brw
, "GL_CLAMP enabled on any texture unit's 1st coordinate",
236 old_key
->gl_clamp_mask
[0], key
->gl_clamp_mask
[0]);
237 found
|= key_debug(brw
, "GL_CLAMP enabled on any texture unit's 2nd coordinate",
238 old_key
->gl_clamp_mask
[1], key
->gl_clamp_mask
[1]);
239 found
|= key_debug(brw
, "GL_CLAMP enabled on any texture unit's 3rd coordinate",
240 old_key
->gl_clamp_mask
[2], key
->gl_clamp_mask
[2]);
241 found
|= key_debug(brw
, "gather channel quirk on any texture unit",
242 old_key
->gather_channel_quirk_mask
, key
->gather_channel_quirk_mask
);
248 brw_wm_debug_recompile(struct brw_context
*brw
,
249 struct gl_shader_program
*prog
,
250 const struct brw_wm_prog_key
*key
)
252 struct brw_cache_item
*c
= NULL
;
253 const struct brw_wm_prog_key
*old_key
= NULL
;
256 perf_debug("Recompiling fragment shader for program %d\n", prog
->Name
);
258 for (unsigned int i
= 0; i
< brw
->cache
.size
; i
++) {
259 for (c
= brw
->cache
.items
[i
]; c
; c
= c
->next
) {
260 if (c
->cache_id
== BRW_WM_PROG
) {
263 if (old_key
->program_string_id
== key
->program_string_id
)
272 perf_debug(" Didn't find previous compile in the shader cache for debug\n");
276 found
|= key_debug(brw
, "alphatest, computed depth, depth test, or "
278 old_key
->iz_lookup
, key
->iz_lookup
);
279 found
|= key_debug(brw
, "depth statistics",
280 old_key
->stats_wm
, key
->stats_wm
);
281 found
|= key_debug(brw
, "flat shading",
282 old_key
->flat_shade
, key
->flat_shade
);
283 found
|= key_debug(brw
, "number of color buffers",
284 old_key
->nr_color_regions
, key
->nr_color_regions
);
285 found
|= key_debug(brw
, "MRT alpha test or alpha-to-coverage",
286 old_key
->replicate_alpha
, key
->replicate_alpha
);
287 found
|= key_debug(brw
, "rendering to FBO",
288 old_key
->render_to_fbo
, key
->render_to_fbo
);
289 found
|= key_debug(brw
, "fragment color clamping",
290 old_key
->clamp_fragment_color
, key
->clamp_fragment_color
);
291 found
|= key_debug(brw
, "line smoothing",
292 old_key
->line_aa
, key
->line_aa
);
293 found
|= key_debug(brw
, "renderbuffer height",
294 old_key
->drawable_height
, key
->drawable_height
);
295 found
|= key_debug(brw
, "input slots valid",
296 old_key
->input_slots_valid
, key
->input_slots_valid
);
297 found
|= key_debug(brw
, "mrt alpha test function",
298 old_key
->alpha_test_func
, key
->alpha_test_func
);
299 found
|= key_debug(brw
, "mrt alpha test reference value",
300 old_key
->alpha_test_ref
, key
->alpha_test_ref
);
302 found
|= brw_debug_recompile_sampler_key(brw
, &old_key
->tex
, &key
->tex
);
305 perf_debug(" Something else\n");
310 gen6_gather_workaround(GLenum internalformat
)
312 switch (internalformat
) {
313 case GL_R8I
: return WA_SIGN
| WA_8BIT
;
314 case GL_R8UI
: return WA_8BIT
;
315 case GL_R16I
: return WA_SIGN
| WA_16BIT
;
316 case GL_R16UI
: return WA_16BIT
;
317 /* note that even though GL_R32I and GL_R32UI have format overrides
318 * in the surface state, there is no shader w/a required */
324 brw_populate_sampler_prog_key_data(struct gl_context
*ctx
,
325 const struct gl_program
*prog
,
326 unsigned sampler_count
,
327 struct brw_sampler_prog_key_data
*key
)
329 struct brw_context
*brw
= brw_context(ctx
);
331 for (int s
= 0; s
< sampler_count
; s
++) {
332 key
->swizzles
[s
] = SWIZZLE_NOOP
;
334 if (!(prog
->SamplersUsed
& (1 << s
)))
337 int unit_id
= prog
->SamplerUnits
[s
];
338 const struct gl_texture_unit
*unit
= &ctx
->Texture
.Unit
[unit_id
];
340 if (unit
->_ReallyEnabled
&& unit
->_Current
->Target
!= GL_TEXTURE_BUFFER
) {
341 const struct gl_texture_object
*t
= unit
->_Current
;
342 const struct gl_texture_image
*img
= t
->Image
[0][t
->BaseLevel
];
343 struct gl_sampler_object
*sampler
= _mesa_get_samplerobj(ctx
, unit_id
);
345 const bool alpha_depth
= t
->DepthMode
== GL_ALPHA
&&
346 (img
->_BaseFormat
== GL_DEPTH_COMPONENT
||
347 img
->_BaseFormat
== GL_DEPTH_STENCIL
);
349 /* Haswell handles texture swizzling as surface format overrides
350 * (except for GL_ALPHA); all other platforms need MOVs in the shader.
352 if (alpha_depth
|| (brw
->gen
< 8 && !brw
->is_haswell
))
353 key
->swizzles
[s
] = brw_get_texture_swizzle(ctx
, t
);
355 if (sampler
->MinFilter
!= GL_NEAREST
&&
356 sampler
->MagFilter
!= GL_NEAREST
) {
357 if (sampler
->WrapS
== GL_CLAMP
)
358 key
->gl_clamp_mask
[0] |= 1 << s
;
359 if (sampler
->WrapT
== GL_CLAMP
)
360 key
->gl_clamp_mask
[1] |= 1 << s
;
361 if (sampler
->WrapR
== GL_CLAMP
)
362 key
->gl_clamp_mask
[2] |= 1 << s
;
365 /* gather4's channel select for green from RG32F is broken;
366 * requires a shader w/a on IVB; fixable with just SCS on HSW. */
367 if (brw
->gen
== 7 && !brw
->is_haswell
&& prog
->UsesGather
) {
368 if (img
->InternalFormat
== GL_RG32F
)
369 key
->gather_channel_quirk_mask
|= 1 << s
;
372 /* Gen6's gather4 is broken for UINT/SINT; we treat them as
373 * UNORM/FLOAT instead and fix it in the shader.
375 if (brw
->gen
== 6 && prog
->UsesGather
) {
376 key
->gen6_gather_wa
[s
] = gen6_gather_workaround(img
->InternalFormat
);
379 /* If this is a multisample sampler, and uses the CMS MSAA layout,
380 * then we need to emit slightly different code to first sample the
383 struct intel_texture_object
*intel_tex
=
384 intel_texture_object((struct gl_texture_object
*)t
);
387 intel_tex
->mt
->msaa_layout
== INTEL_MSAA_LAYOUT_CMS
) {
388 key
->compressed_multisample_layout_mask
|= 1 << s
;
394 static void brw_wm_populate_key( struct brw_context
*brw
,
395 struct brw_wm_prog_key
*key
)
397 struct gl_context
*ctx
= &brw
->ctx
;
398 /* BRW_NEW_FRAGMENT_PROGRAM */
399 const struct brw_fragment_program
*fp
=
400 (struct brw_fragment_program
*)brw
->fragment_program
;
401 const struct gl_program
*prog
= (struct gl_program
*) brw
->fragment_program
;
404 bool program_uses_dfdy
= fp
->program
.UsesDFdy
;
405 bool multisample_fbo
= ctx
->DrawBuffer
->Visual
.samples
> 1;
407 memset(key
, 0, sizeof(*key
));
409 /* Build the index for table lookup
413 if (fp
->program
.UsesKill
|| ctx
->Color
.AlphaEnabled
)
414 lookup
|= IZ_PS_KILL_ALPHATEST_BIT
;
416 if (fp
->program
.Base
.OutputsWritten
& BITFIELD64_BIT(FRAG_RESULT_DEPTH
))
417 lookup
|= IZ_PS_COMPUTES_DEPTH_BIT
;
421 lookup
|= IZ_DEPTH_TEST_ENABLE_BIT
;
423 if (ctx
->Depth
.Test
&& ctx
->Depth
.Mask
) /* ?? */
424 lookup
|= IZ_DEPTH_WRITE_ENABLE_BIT
;
426 /* _NEW_STENCIL | _NEW_BUFFERS */
427 if (ctx
->Stencil
._Enabled
) {
428 lookup
|= IZ_STENCIL_TEST_ENABLE_BIT
;
430 if (ctx
->Stencil
.WriteMask
[0] ||
431 ctx
->Stencil
.WriteMask
[ctx
->Stencil
._BackFace
])
432 lookup
|= IZ_STENCIL_WRITE_ENABLE_BIT
;
434 key
->iz_lookup
= lookup
;
439 /* _NEW_LINE, _NEW_POLYGON, BRW_NEW_REDUCED_PRIMITIVE */
440 if (ctx
->Line
.SmoothFlag
) {
441 if (brw
->reduced_primitive
== GL_LINES
) {
444 else if (brw
->reduced_primitive
== GL_TRIANGLES
) {
445 if (ctx
->Polygon
.FrontMode
== GL_LINE
) {
446 line_aa
= AA_SOMETIMES
;
448 if (ctx
->Polygon
.BackMode
== GL_LINE
||
449 (ctx
->Polygon
.CullFlag
&&
450 ctx
->Polygon
.CullFaceMode
== GL_BACK
))
453 else if (ctx
->Polygon
.BackMode
== GL_LINE
) {
454 line_aa
= AA_SOMETIMES
;
456 if ((ctx
->Polygon
.CullFlag
&&
457 ctx
->Polygon
.CullFaceMode
== GL_FRONT
))
463 key
->line_aa
= line_aa
;
466 if (brw
->disable_derivative_optimization
) {
467 key
->high_quality_derivatives
=
468 ctx
->Hint
.FragmentShaderDerivative
!= GL_FASTEST
;
470 key
->high_quality_derivatives
=
471 ctx
->Hint
.FragmentShaderDerivative
== GL_NICEST
;
475 key
->stats_wm
= brw
->stats_wm
;
478 key
->flat_shade
= (ctx
->Light
.ShadeModel
== GL_FLAT
);
480 /* _NEW_FRAG_CLAMP | _NEW_BUFFERS */
481 key
->clamp_fragment_color
= ctx
->Color
._ClampFragmentColor
;
484 brw_populate_sampler_prog_key_data(ctx
, prog
, brw
->wm
.base
.sampler_count
,
489 * Include the draw buffer origin and height so that we can calculate
490 * fragment position values relative to the bottom left of the drawable,
491 * from the incoming screen origin relative position we get as part of our
494 * This is only needed for the WM_WPOSXY opcode when the fragment program
495 * uses the gl_FragCoord input.
497 * We could avoid recompiling by including this as a constant referenced by
498 * our program, but if we were to do that it would also be nice to handle
499 * getting that constant updated at batchbuffer submit time (when we
500 * hold the lock and know where the buffer really is) rather than at emit
501 * time when we don't hold the lock and are just guessing. We could also
502 * just avoid using this as key data if the program doesn't use
505 * For DRI2 the origin_x/y will always be (0,0) but we still need the
506 * drawable height in order to invert the Y axis.
508 if (fp
->program
.Base
.InputsRead
& VARYING_BIT_POS
) {
509 key
->drawable_height
= ctx
->DrawBuffer
->Height
;
512 if ((fp
->program
.Base
.InputsRead
& VARYING_BIT_POS
) || program_uses_dfdy
) {
513 key
->render_to_fbo
= _mesa_is_user_fbo(ctx
->DrawBuffer
);
517 key
->nr_color_regions
= ctx
->DrawBuffer
->_NumColorDrawBuffers
;
519 /* _NEW_MULTISAMPLE, _NEW_COLOR, _NEW_BUFFERS */
520 key
->replicate_alpha
= ctx
->DrawBuffer
->_NumColorDrawBuffers
> 1 &&
521 (ctx
->Multisample
.SampleAlphaToCoverage
|| ctx
->Color
.AlphaEnabled
);
523 /* _NEW_BUFFERS _NEW_MULTISAMPLE */
524 /* Ignore sample qualifier while computing this flag. */
525 key
->persample_shading
=
526 _mesa_get_min_invocations_per_fragment(ctx
, &fp
->program
, true) > 1;
528 key
->compute_pos_offset
=
529 _mesa_get_min_invocations_per_fragment(ctx
, &fp
->program
, false) > 1 &&
530 fp
->program
.Base
.SystemValuesRead
& SYSTEM_BIT_SAMPLE_POS
;
532 key
->compute_sample_id
=
534 ctx
->Multisample
.Enabled
&&
535 (fp
->program
.Base
.SystemValuesRead
& SYSTEM_BIT_SAMPLE_ID
);
537 /* BRW_NEW_VUE_MAP_GEOM_OUT */
538 if (brw
->gen
< 6 || _mesa_bitcount_64(fp
->program
.Base
.InputsRead
&
539 BRW_FS_VARYING_INPUT_MASK
) > 16)
540 key
->input_slots_valid
= brw
->vue_map_geom_out
.slots_valid
;
543 /* _NEW_COLOR | _NEW_BUFFERS */
544 /* Pre-gen6, the hardware alpha test always used each render
545 * target's alpha to do alpha test, as opposed to render target 0's alpha
546 * like GL requires. Fix that by building the alpha test into the
547 * shader, and we'll skip enabling the fixed function alpha test.
549 if (brw
->gen
< 6 && ctx
->DrawBuffer
->_NumColorDrawBuffers
> 1 && ctx
->Color
.AlphaEnabled
) {
550 key
->alpha_test_func
= ctx
->Color
.AlphaFunc
;
551 key
->alpha_test_ref
= ctx
->Color
.AlphaRef
;
554 /* The unique fragment program ID */
555 key
->program_string_id
= fp
->id
;
560 brw_upload_wm_prog(struct brw_context
*brw
)
562 struct gl_context
*ctx
= &brw
->ctx
;
563 struct brw_wm_prog_key key
;
564 struct brw_fragment_program
*fp
= (struct brw_fragment_program
*)
565 brw
->fragment_program
;
567 brw_wm_populate_key(brw
, &key
);
569 if (!brw_search_cache(&brw
->cache
, BRW_WM_PROG
,
571 &brw
->wm
.base
.prog_offset
, &brw
->wm
.prog_data
)) {
572 bool success
= do_wm_prog(brw
, ctx
->_Shader
->_CurrentFragmentProgram
, fp
,
577 brw
->wm
.base
.prog_data
= &brw
->wm
.prog_data
->base
;
581 const struct brw_tracked_state brw_wm_prog
= {
583 .mesa
= (_NEW_COLOR
|
594 .brw
= (BRW_NEW_FRAGMENT_PROGRAM
|
595 BRW_NEW_REDUCED_PRIMITIVE
|
596 BRW_NEW_VUE_MAP_GEOM_OUT
|
599 .emit
= brw_upload_wm_prog