2 * Copyright © 2013 Intel Corporation
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 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * 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 NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
25 #include "main/teximage.h"
26 #include "main/blend.h"
27 #include "main/fbobject.h"
28 #include "main/renderbuffer.h"
31 #include "glsl/ralloc.h"
33 #include "intel_fbo.h"
35 #include "brw_blorp.h"
36 #include "brw_context.h"
38 #include "brw_state.h"
40 #define FILE_DEBUG_FLAG DEBUG_BLORP
42 struct brw_blorp_const_color_prog_key
44 bool use_simd16_replicated_data
;
49 * Parameters for a blorp operation where the fragment shader outputs a
50 * constant color. This is used for both fast color clears and color
53 class brw_blorp_const_color_params
: public brw_blorp_params
56 virtual uint32_t get_wm_prog(struct brw_context
*brw
,
57 brw_blorp_prog_data
**prog_data
) const;
60 brw_blorp_const_color_prog_key wm_prog_key
;
63 class brw_blorp_clear_params
: public brw_blorp_const_color_params
66 brw_blorp_clear_params(struct brw_context
*brw
,
67 struct gl_framebuffer
*fb
,
68 struct gl_renderbuffer
*rb
,
75 * Parameters for a blorp operation that performs a "render target resolve".
76 * This is used to resolve pending fast clear pixels before a color buffer is
77 * used for texturing, ReadPixels, or scanout.
79 class brw_blorp_rt_resolve_params
: public brw_blorp_const_color_params
82 brw_blorp_rt_resolve_params(struct brw_context
*brw
,
83 struct intel_mipmap_tree
*mt
);
87 class brw_blorp_const_color_program
90 brw_blorp_const_color_program(struct brw_context
*brw
,
91 const brw_blorp_const_color_prog_key
*key
);
92 ~brw_blorp_const_color_program();
94 const GLuint
*compile(struct brw_context
*brw
, GLuint
*program_size
);
96 brw_blorp_prog_data prog_data
;
102 struct brw_context
*brw
;
103 const brw_blorp_const_color_prog_key
*key
;
104 struct brw_compile func
;
106 /* Thread dispatch header */
109 /* Pixel X/Y coordinates (always in R1). */
112 /* Register with push constants (a single vec4) */
113 struct brw_reg clear_rgba
;
115 /* MRF used for render target writes */
119 brw_blorp_const_color_program::brw_blorp_const_color_program(
120 struct brw_context
*brw
,
121 const brw_blorp_const_color_prog_key
*key
)
122 : mem_ctx(ralloc_context(NULL
)),
130 brw_init_compile(brw
, &func
, mem_ctx
);
133 brw_blorp_const_color_program::~brw_blorp_const_color_program()
135 ralloc_free(mem_ctx
);
140 * Determine if fast color clear supports the given clear color.
142 * Fast color clear can only clear to color values of 1.0 or 0.0. At the
143 * moment we only support floating point, unorm, and snorm buffers.
146 is_color_fast_clear_compatible(struct brw_context
*brw
,
148 const union gl_color_union
*color
)
150 if (_mesa_is_format_integer_color(format
))
153 for (int i
= 0; i
< 4; i
++) {
154 if (color
->f
[i
] != 0.0 && color
->f
[i
] != 1.0) {
155 perf_debug("Clear color unsupported by fast color clear. "
156 "Falling back to slow clear.\n");
165 * Convert the given color to a bitfield suitable for ORing into DWORD 7 of
169 compute_fast_clear_color_bits(const union gl_color_union
*color
)
172 for (int i
= 0; i
< 4; i
++) {
173 if (color
->f
[i
] != 0.0)
174 bits
|= 1 << (GEN7_SURFACE_CLEAR_COLOR_SHIFT
+ (3 - i
));
180 brw_blorp_clear_params::brw_blorp_clear_params(struct brw_context
*brw
,
181 struct gl_framebuffer
*fb
,
182 struct gl_renderbuffer
*rb
,
186 struct gl_context
*ctx
= &brw
->ctx
;
187 struct intel_renderbuffer
*irb
= intel_renderbuffer(rb
);
189 dst
.set(brw
, irb
->mt
, irb
->mt_level
, irb
->mt_layer
);
191 /* Override the surface format according to the context's sRGB rules. */
192 gl_format format
= _mesa_get_render_format(ctx
, irb
->mt
->format
);
193 dst
.brw_surfaceformat
= brw
->render_target_format
[format
];
201 y0
= rb
->Height
- fb
->_Ymax
;
202 y1
= rb
->Height
- fb
->_Ymin
;
205 float *push_consts
= (float *)&wm_push_consts
;
207 push_consts
[0] = ctx
->Color
.ClearColor
.f
[0];
208 push_consts
[1] = ctx
->Color
.ClearColor
.f
[1];
209 push_consts
[2] = ctx
->Color
.ClearColor
.f
[2];
210 push_consts
[3] = ctx
->Color
.ClearColor
.f
[3];
214 memset(&wm_prog_key
, 0, sizeof(wm_prog_key
));
216 wm_prog_key
.use_simd16_replicated_data
= true;
218 /* From the SNB PRM (Vol4_Part1):
220 * "Replicated data (Message Type = 111) is only supported when
221 * accessing tiled memory. Using this Message Type to access linear
222 * (untiled) memory is UNDEFINED."
224 if (irb
->mt
->region
->tiling
== I915_TILING_NONE
)
225 wm_prog_key
.use_simd16_replicated_data
= false;
227 /* Constant color writes ignore everyting in blend and color calculator
228 * state. This is not documented.
230 for (int i
= 0; i
< 4; i
++) {
231 if (!color_mask
[i
]) {
232 color_write_disable
[i
] = true;
233 wm_prog_key
.use_simd16_replicated_data
= false;
237 /* If we can do this as a fast color clear, do so. */
238 if (irb
->mt
->mcs_state
!= INTEL_MCS_STATE_NONE
&& !partial_clear
&&
239 wm_prog_key
.use_simd16_replicated_data
&&
240 is_color_fast_clear_compatible(brw
, format
, &ctx
->Color
.ClearColor
)) {
241 memset(push_consts
, 0xff, 4*sizeof(float));
242 fast_clear_op
= GEN7_FAST_CLEAR_OP_FAST_CLEAR
;
244 /* From the Ivy Bridge PRM, Vol2 Part1 11.7 "MCS Buffer for Render
245 * Target(s)", beneath the "Fast Color Clear" bullet (p327):
247 * Clear pass must have a clear rectangle that must follow alignment
248 * rules in terms of pixels and lines as shown in the table
249 * below. Further, the clear-rectangle height and width must be
250 * multiple of the following dimensions. If the height and width of
251 * the render target being cleared do not meet these requirements,
252 * an MCS buffer can be created such that it follows the requirement
255 * The alignment size in the table that follows is related to the
256 * alignment size returned by intel_get_non_msrt_mcs_alignment(), but
257 * with X alignment multiplied by 16 and Y alignment multiplied by 32.
259 unsigned x_align
, y_align
;
260 intel_get_non_msrt_mcs_alignment(brw
, irb
->mt
, &x_align
, &y_align
);
264 if (brw
->is_haswell
&& brw
->gt
== 3) {
265 /* From BSpec: 3D-Media-GPGPU Engine > 3D Pipeline > Pixel > Pixel
266 * Backend > MCS Buffer for Render Target(s) [DevIVB+]:
267 * [DevHSW:GT3]: Clear rectangle must be aligned to two times the
268 * number of pixels in the table shown below...
269 * x_align, y_align values computed above are the relevant entries
270 * in the referred table.
272 x0
= ROUND_DOWN_TO(x0
, 2 * x_align
);
273 y0
= ROUND_DOWN_TO(y0
, 2 * y_align
);
274 x1
= ALIGN(x1
, 2 * x_align
);
275 y1
= ALIGN(y1
, 2 * y_align
);
277 x0
= ROUND_DOWN_TO(x0
, x_align
);
278 y0
= ROUND_DOWN_TO(y0
, y_align
);
279 x1
= ALIGN(x1
, x_align
);
280 y1
= ALIGN(y1
, y_align
);
283 /* From the Ivy Bridge PRM, Vol2 Part1 11.7 "MCS Buffer for Render
284 * Target(s)", beneath the "Fast Color Clear" bullet (p327):
286 * In order to optimize the performance MCS buffer (when bound to 1X
287 * RT) clear similarly to MCS buffer clear for MSRT case, clear rect
288 * is required to be scaled by the following factors in the
289 * horizontal and vertical directions:
291 * The X and Y scale down factors in the table that follows are each
292 * equal to half the alignment value computed above.
294 unsigned x_scaledown
= x_align
/ 2;
295 unsigned y_scaledown
= y_align
/ 2;
304 brw_blorp_rt_resolve_params::brw_blorp_rt_resolve_params(
305 struct brw_context
*brw
,
306 struct intel_mipmap_tree
*mt
)
308 dst
.set(brw
, mt
, 0 /* level */, 0 /* layer */);
310 /* From the Ivy Bridge PRM, Vol2 Part1 11.9 "Render Target Resolve":
312 * A rectangle primitive must be scaled down by the following factors
313 * with respect to render target being resolved.
315 * The scaledown factors in the table that follows are related to the
316 * alignment size returned by intel_get_non_msrt_mcs_alignment(), but with
317 * X and Y alignment each divided by 2.
319 unsigned x_align
, y_align
;
320 intel_get_non_msrt_mcs_alignment(brw
, mt
, &x_align
, &y_align
);
321 unsigned x_scaledown
= x_align
/ 2;
322 unsigned y_scaledown
= y_align
/ 2;
324 x1
= ALIGN(mt
->logical_width0
, x_scaledown
) / x_scaledown
;
325 y1
= ALIGN(mt
->logical_height0
, y_scaledown
) / y_scaledown
;
327 fast_clear_op
= GEN7_FAST_CLEAR_OP_RESOLVE
;
329 /* Note: there is no need to initialize push constants because it doesn't
330 * matter what data gets dispatched to the render target. However, we must
331 * ensure that the fragment shader delivers the data using the "replicated
335 memset(&wm_prog_key
, 0, sizeof(wm_prog_key
));
336 wm_prog_key
.use_simd16_replicated_data
= true;
341 brw_blorp_const_color_params::get_wm_prog(struct brw_context
*brw
,
342 brw_blorp_prog_data
**prog_data
)
345 uint32_t prog_offset
= 0;
346 if (!brw_search_cache(&brw
->cache
, BRW_BLORP_CONST_COLOR_PROG
,
347 &this->wm_prog_key
, sizeof(this->wm_prog_key
),
348 &prog_offset
, prog_data
)) {
349 brw_blorp_const_color_program
prog(brw
, &this->wm_prog_key
);
351 const GLuint
*program
= prog
.compile(brw
, &program_size
);
352 brw_upload_cache(&brw
->cache
, BRW_BLORP_CONST_COLOR_PROG
,
353 &this->wm_prog_key
, sizeof(this->wm_prog_key
),
354 program
, program_size
,
355 &prog
.prog_data
, sizeof(prog
.prog_data
),
356 &prog_offset
, prog_data
);
362 brw_blorp_const_color_program::alloc_regs()
365 this->R0
= retype(brw_vec8_grf(reg
++, 0), BRW_REGISTER_TYPE_UW
);
366 this->R1
= retype(brw_vec8_grf(reg
++, 0), BRW_REGISTER_TYPE_UW
);
368 prog_data
.first_curbe_grf
= reg
;
369 clear_rgba
= retype(brw_vec4_grf(reg
++, 0), BRW_REGISTER_TYPE_F
);
370 reg
+= BRW_BLORP_NUM_PUSH_CONST_REGS
;
372 /* Make sure we didn't run out of registers */
373 assert(reg
<= GEN7_MRF_HACK_START
);
379 brw_blorp_const_color_program::compile(struct brw_context
*brw
,
380 GLuint
*program_size
)
382 /* Set up prog_data */
383 memset(&prog_data
, 0, sizeof(prog_data
));
384 prog_data
.persample_msaa_dispatch
= false;
388 brw_set_compression_control(&func
, BRW_COMPRESSION_NONE
);
390 struct brw_reg mrf_rt_write
=
391 retype(vec16(brw_message_reg(base_mrf
)), BRW_REGISTER_TYPE_F
);
393 uint32_t mlen
, msg_type
;
394 if (key
->use_simd16_replicated_data
) {
395 /* The message payload is a single register with the low 4 floats/ints
396 * filled with the constant clear color.
398 brw_set_mask_control(&func
, BRW_MASK_DISABLE
);
399 brw_MOV(&func
, vec4(brw_message_reg(base_mrf
)), clear_rgba
);
400 brw_set_mask_control(&func
, BRW_MASK_ENABLE
);
402 msg_type
= BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE_REPLICATED
;
405 for (int i
= 0; i
< 4; i
++) {
406 /* The message payload is pairs of registers for 16 pixels each of r,
409 brw_set_compression_control(&func
, BRW_COMPRESSION_COMPRESSED
);
411 brw_message_reg(base_mrf
+ i
* 2),
412 brw_vec1_grf(clear_rgba
.nr
, i
));
413 brw_set_compression_control(&func
, BRW_COMPRESSION_NONE
);
416 msg_type
= BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE
;
420 /* Now write to the render target and terminate the thread */
422 16 /* dispatch_width */,
423 base_mrf
/* msg_reg_nr */,
424 mrf_rt_write
/* src0 */,
426 BRW_BLORP_RENDERBUFFER_BINDING_TABLE_INDEX
,
428 0 /* response_length */,
430 false /* header present */);
432 if (unlikely(INTEL_DEBUG
& DEBUG_BLORP
)) {
433 printf("Native code for BLORP clear:\n");
434 brw_dump_compile(&func
, stdout
, 0, func
.next_insn_offset
);
437 return brw_get_program(&func
, program_size
);
442 brw_blorp_clear_color(struct brw_context
*brw
, struct gl_framebuffer
*fb
,
445 struct gl_context
*ctx
= &brw
->ctx
;
447 /* The constant color clear code doesn't work for multisampled surfaces, so
448 * we need to support falling back to other clear mechanisms.
449 * Unfortunately, our clear code is based on a bitmask that doesn't
450 * distinguish individual color attachments, so we walk the attachments to
451 * see if any require fallback, and fall back for all if any of them need
454 for (unsigned buf
= 0; buf
< ctx
->DrawBuffer
->_NumColorDrawBuffers
; buf
++) {
455 struct gl_renderbuffer
*rb
= ctx
->DrawBuffer
->_ColorDrawBuffers
[buf
];
456 struct intel_renderbuffer
*irb
= intel_renderbuffer(rb
);
458 if (irb
&& irb
->mt
->msaa_layout
!= INTEL_MSAA_LAYOUT_NONE
)
462 for (unsigned buf
= 0; buf
< ctx
->DrawBuffer
->_NumColorDrawBuffers
; buf
++) {
463 struct gl_renderbuffer
*rb
= ctx
->DrawBuffer
->_ColorDrawBuffers
[buf
];
464 struct intel_renderbuffer
*irb
= intel_renderbuffer(rb
);
466 /* If this is an ES2 context or GL_ARB_ES2_compatibility is supported,
467 * the framebuffer can be complete with some attachments missing. In
468 * this case the _ColorDrawBuffers pointer will be NULL.
473 brw_blorp_clear_params
params(brw
, fb
, rb
, ctx
->Color
.ColorMask
[buf
],
477 (params
.fast_clear_op
== GEN7_FAST_CLEAR_OP_FAST_CLEAR
);
479 /* Record the clear color in the miptree so that it will be
480 * programmed in SURFACE_STATE by later rendering and resolve
483 uint32_t new_color_value
=
484 compute_fast_clear_color_bits(&ctx
->Color
.ClearColor
);
485 if (irb
->mt
->fast_clear_color_value
!= new_color_value
) {
486 irb
->mt
->fast_clear_color_value
= new_color_value
;
487 brw
->state
.dirty
.brw
|= BRW_NEW_SURFACES
;
490 /* If the buffer is already in INTEL_MCS_STATE_CLEAR, the clear is
491 * redundant and can be skipped.
493 if (irb
->mt
->mcs_state
== INTEL_MCS_STATE_CLEAR
)
496 /* If the MCS buffer hasn't been allocated yet, we need to allocate
499 if (!irb
->mt
->mcs_mt
) {
500 if (!intel_miptree_alloc_non_msrt_mcs(brw
, irb
->mt
)) {
501 /* MCS allocation failed--probably this will only happen in
502 * out-of-memory conditions. But in any case, try to recover
503 * by falling back to a non-blorp clear technique.
507 brw
->state
.dirty
.brw
|= BRW_NEW_SURFACES
;
511 DBG("%s to mt %p level %d layer %d\n", __FUNCTION__
,
512 irb
->mt
, irb
->mt_level
, irb
->mt_layer
);
514 brw_blorp_exec(brw
, ¶ms
);
517 /* Now that the fast clear has occurred, put the buffer in
518 * INTEL_MCS_STATE_CLEAR so that we won't waste time doing redundant
521 irb
->mt
->mcs_state
= INTEL_MCS_STATE_CLEAR
;
529 brw_blorp_resolve_color(struct brw_context
*brw
, struct intel_mipmap_tree
*mt
)
531 DBG("%s to mt %p\n", __FUNCTION__
, mt
);
533 brw_blorp_rt_resolve_params
params(brw
, mt
);
534 brw_blorp_exec(brw
, ¶ms
);
535 mt
->mcs_state
= INTEL_MCS_STATE_RESOLVED
;