1 /**************************************************************************
3 * Copyright 2009 VMware, Inc.
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * 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, sub license, 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 portions
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 **************************************************************************/
30 * AoS pixel format manipulation.
32 * @author Jose Fonseca <jfonseca@vmware.com>
36 #include "util/u_format.h"
37 #include "util/u_memory.h"
38 #include "util/u_math.h"
39 #include "util/u_pointer.h"
40 #include "util/u_string.h"
42 #include "lp_bld_arit.h"
43 #include "lp_bld_init.h"
44 #include "lp_bld_type.h"
45 #include "lp_bld_flow.h"
46 #include "lp_bld_const.h"
47 #include "lp_bld_conv.h"
48 #include "lp_bld_swizzle.h"
49 #include "lp_bld_gather.h"
50 #include "lp_bld_debug.h"
51 #include "lp_bld_format.h"
55 * Basic swizzling. Rearrange the order of the unswizzled array elements
56 * according to the format description. PIPE_SWIZZLE_ZERO/ONE are supported
58 * Ex: if unswizzled[4] = {B, G, R, x}, then swizzled_out[4] = {R, G, B, 1}.
61 lp_build_format_swizzle_aos(const struct util_format_description
*desc
,
62 struct lp_build_context
*bld
,
63 LLVMValueRef unswizzled
)
65 unsigned char swizzles
[4];
68 assert(bld
->type
.length
% 4 == 0);
70 for (chan
= 0; chan
< 4; ++chan
) {
71 enum util_format_swizzle swizzle
;
73 if (desc
->colorspace
== UTIL_FORMAT_COLORSPACE_ZS
) {
75 * For ZS formats do RGBA = ZZZ1
78 swizzle
= UTIL_FORMAT_SWIZZLE_1
;
79 } else if (desc
->swizzle
[0] == UTIL_FORMAT_SWIZZLE_NONE
) {
80 swizzle
= UTIL_FORMAT_SWIZZLE_0
;
82 swizzle
= desc
->swizzle
[0];
85 swizzle
= desc
->swizzle
[chan
];
87 swizzles
[chan
] = swizzle
;
90 return lp_build_swizzle_aos(bld
, unswizzled
, swizzles
);
95 * Whether the format matches the vector type, apart of swizzles.
98 format_matches_type(const struct util_format_description
*desc
,
101 enum util_format_type chan_type
;
104 assert(type
.length
% 4 == 0);
106 if (desc
->layout
!= UTIL_FORMAT_LAYOUT_PLAIN
||
107 desc
->colorspace
!= UTIL_FORMAT_COLORSPACE_RGB
||
108 desc
->block
.width
!= 1 ||
109 desc
->block
.height
!= 1) {
114 chan_type
= UTIL_FORMAT_TYPE_FLOAT
;
115 } else if (type
.fixed
) {
116 chan_type
= UTIL_FORMAT_TYPE_FIXED
;
117 } else if (type
.sign
) {
118 chan_type
= UTIL_FORMAT_TYPE_SIGNED
;
120 chan_type
= UTIL_FORMAT_TYPE_UNSIGNED
;
123 for (chan
= 0; chan
< desc
->nr_channels
; ++chan
) {
124 if (desc
->channel
[chan
].size
!= type
.width
) {
128 if (desc
->channel
[chan
].type
!= UTIL_FORMAT_TYPE_VOID
) {
129 if (desc
->channel
[chan
].type
!= chan_type
||
130 desc
->channel
[chan
].normalized
!= type
.norm
) {
141 * Unpack a single pixel into its RGBA components.
143 * @param desc the pixel format for the packed pixel value
144 * @param packed integer pixel in a format such as PIPE_FORMAT_B8G8R8A8_UNORM
146 * @return RGBA in a float[4] or ubyte[4] or ushort[4] vector.
148 static INLINE LLVMValueRef
149 lp_build_unpack_arith_rgba_aos(struct gallivm_state
*gallivm
,
150 const struct util_format_description
*desc
,
153 LLVMBuilderRef builder
= gallivm
->builder
;
154 LLVMValueRef shifted
, casted
, scaled
, masked
;
155 LLVMValueRef shifts
[4];
156 LLVMValueRef masks
[4];
157 LLVMValueRef scales
[4];
160 boolean needs_uitofp
;
164 /* TODO: Support more formats */
165 assert(desc
->layout
== UTIL_FORMAT_LAYOUT_PLAIN
);
166 assert(desc
->block
.width
== 1);
167 assert(desc
->block
.height
== 1);
168 assert(desc
->block
.bits
<= 32);
170 /* Do the intermediate integer computations with 32bit integers since it
171 * matches floating point size */
172 assert (LLVMTypeOf(packed
) == LLVMInt32TypeInContext(gallivm
->context
));
174 /* Broadcast the packed value to all four channels
175 * before: packed = BGRA
176 * after: packed = {BGRA, BGRA, BGRA, BGRA}
178 packed
= LLVMBuildInsertElement(builder
,
179 LLVMGetUndef(LLVMVectorType(LLVMInt32TypeInContext(gallivm
->context
), 4)),
181 LLVMConstNull(LLVMInt32TypeInContext(gallivm
->context
)),
183 packed
= LLVMBuildShuffleVector(builder
,
185 LLVMGetUndef(LLVMVectorType(LLVMInt32TypeInContext(gallivm
->context
), 4)),
186 LLVMConstNull(LLVMVectorType(LLVMInt32TypeInContext(gallivm
->context
), 4)),
189 /* Initialize vector constants */
191 needs_uitofp
= FALSE
;
194 /* Loop over 4 color components */
195 for (i
= 0; i
< 4; ++i
) {
196 unsigned bits
= desc
->channel
[i
].size
;
198 if (desc
->channel
[i
].type
== UTIL_FORMAT_TYPE_VOID
) {
199 shifts
[i
] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
200 masks
[i
] = LLVMConstNull(LLVMInt32TypeInContext(gallivm
->context
));
201 scales
[i
] = LLVMConstNull(LLVMFloatTypeInContext(gallivm
->context
));
204 unsigned long long mask
= (1ULL << bits
) - 1;
206 assert(desc
->channel
[i
].type
== UTIL_FORMAT_TYPE_UNSIGNED
);
212 shifts
[i
] = lp_build_const_int32(gallivm
, shift
);
213 masks
[i
] = lp_build_const_int32(gallivm
, mask
);
215 if (desc
->channel
[i
].normalized
) {
216 scales
[i
] = lp_build_const_float(gallivm
, 1.0 / mask
);
220 scales
[i
] = lp_build_const_float(gallivm
, 1.0);
226 /* Ex: convert packed = {BGRA, BGRA, BGRA, BGRA}
227 * into masked = {B, G, R, A}
229 shifted
= LLVMBuildLShr(builder
, packed
, LLVMConstVector(shifts
, 4), "");
230 masked
= LLVMBuildAnd(builder
, shifted
, LLVMConstVector(masks
, 4), "");
234 /* UIToFP can't be expressed in SSE2 */
235 casted
= LLVMBuildSIToFP(builder
, masked
, LLVMVectorType(LLVMFloatTypeInContext(gallivm
->context
), 4), "");
237 casted
= LLVMBuildUIToFP(builder
, masked
, LLVMVectorType(LLVMFloatTypeInContext(gallivm
->context
), 4), "");
240 /* At this point 'casted' may be a vector of floats such as
241 * {255.0, 255.0, 255.0, 255.0}. Next, if the pixel values are normalized
242 * we'll scale this to {1.0, 1.0, 1.0, 1.0}.
246 scaled
= LLVMBuildFMul(builder
, casted
, LLVMConstVector(scales
, 4), "");
255 * Pack a single pixel.
257 * @param rgba 4 float vector with the unpacked components.
259 * XXX: This is mostly for reference and testing -- operating a single pixel at
260 * a time is rarely if ever needed.
263 lp_build_pack_rgba_aos(struct gallivm_state
*gallivm
,
264 const struct util_format_description
*desc
,
267 LLVMBuilderRef builder
= gallivm
->builder
;
269 LLVMValueRef packed
= NULL
;
270 LLVMValueRef swizzles
[4];
271 LLVMValueRef shifted
, casted
, scaled
, unswizzled
;
272 LLVMValueRef shifts
[4];
273 LLVMValueRef scales
[4];
278 assert(desc
->layout
== UTIL_FORMAT_LAYOUT_PLAIN
);
279 assert(desc
->block
.width
== 1);
280 assert(desc
->block
.height
== 1);
282 type
= LLVMIntTypeInContext(gallivm
->context
, desc
->block
.bits
);
284 /* Unswizzle the color components into the source vector. */
285 for (i
= 0; i
< 4; ++i
) {
286 for (j
= 0; j
< 4; ++j
) {
287 if (desc
->swizzle
[j
] == i
)
291 swizzles
[i
] = lp_build_const_int32(gallivm
, j
);
293 swizzles
[i
] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
296 unswizzled
= LLVMBuildShuffleVector(builder
, rgba
,
297 LLVMGetUndef(LLVMVectorType(LLVMFloatTypeInContext(gallivm
->context
), 4)),
298 LLVMConstVector(swizzles
, 4), "");
302 for (i
= 0; i
< 4; ++i
) {
303 unsigned bits
= desc
->channel
[i
].size
;
305 if (desc
->channel
[i
].type
== UTIL_FORMAT_TYPE_VOID
) {
306 shifts
[i
] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
307 scales
[i
] = LLVMGetUndef(LLVMFloatTypeInContext(gallivm
->context
));
310 unsigned mask
= (1 << bits
) - 1;
312 assert(desc
->channel
[i
].type
== UTIL_FORMAT_TYPE_UNSIGNED
);
315 shifts
[i
] = lp_build_const_int32(gallivm
, shift
);
317 if (desc
->channel
[i
].normalized
) {
318 scales
[i
] = lp_build_const_float(gallivm
, mask
);
322 scales
[i
] = lp_build_const_float(gallivm
, 1.0);
329 scaled
= LLVMBuildFMul(builder
, unswizzled
, LLVMConstVector(scales
, 4), "");
333 casted
= LLVMBuildFPToSI(builder
, scaled
, LLVMVectorType(LLVMInt32TypeInContext(gallivm
->context
), 4), "");
335 shifted
= LLVMBuildShl(builder
, casted
, LLVMConstVector(shifts
, 4), "");
337 /* Bitwise or all components */
338 for (i
= 0; i
< 4; ++i
) {
339 if (desc
->channel
[i
].type
== UTIL_FORMAT_TYPE_UNSIGNED
) {
340 LLVMValueRef component
= LLVMBuildExtractElement(builder
, shifted
,
341 lp_build_const_int32(gallivm
, i
), "");
343 packed
= LLVMBuildOr(builder
, packed
, component
, "");
350 packed
= LLVMGetUndef(LLVMInt32TypeInContext(gallivm
->context
));
352 if (desc
->block
.bits
< 32)
353 packed
= LLVMBuildTrunc(builder
, packed
, type
, "");
362 * Fetch a pixel into a 4 float AoS.
364 * \param format_desc describes format of the image we're fetching from
365 * \param ptr address of the pixel block (or the texel if uncompressed)
366 * \param i, j the sub-block pixel coordinates. For non-compressed formats
367 * these will always be (0, 0).
368 * \return a 4 element vector with the pixel's RGBA values.
371 lp_build_fetch_rgba_aos(struct gallivm_state
*gallivm
,
372 const struct util_format_description
*format_desc
,
374 LLVMValueRef base_ptr
,
379 LLVMBuilderRef builder
= gallivm
->builder
;
380 unsigned num_pixels
= type
.length
/ 4;
381 struct lp_build_context bld
;
383 assert(type
.length
<= LP_MAX_VECTOR_LENGTH
);
384 assert(type
.length
% 4 == 0);
386 lp_build_context_init(&bld
, gallivm
, type
);
391 * The format matches the type (apart of a swizzle) so no need for
392 * scaling or converting.
395 if (format_matches_type(format_desc
, type
) &&
396 format_desc
->block
.bits
<= type
.width
* 4 &&
397 util_is_power_of_two(format_desc
->block
.bits
)) {
401 * The format matches the type (apart of a swizzle) so no need for
402 * scaling or converting.
405 packed
= lp_build_gather(gallivm
, type
.length
/4,
406 format_desc
->block
.bits
, type
.width
*4,
409 assert(format_desc
->block
.bits
<= type
.width
* type
.length
);
411 packed
= LLVMBuildBitCast(gallivm
->builder
, packed
,
412 lp_build_vec_type(gallivm
, type
), "");
414 return lp_build_format_swizzle_aos(format_desc
, &bld
, packed
);
421 if (format_desc
->layout
== UTIL_FORMAT_LAYOUT_PLAIN
&&
422 (format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_RGB
||
423 format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_ZS
) &&
424 format_desc
->block
.width
== 1 &&
425 format_desc
->block
.height
== 1 &&
426 util_is_power_of_two(format_desc
->block
.bits
) &&
427 format_desc
->block
.bits
<= 32 &&
428 format_desc
->is_bitmask
&&
429 !format_desc
->is_mixed
&&
430 (format_desc
->channel
[0].type
== UTIL_FORMAT_TYPE_UNSIGNED
||
431 format_desc
->channel
[1].type
== UTIL_FORMAT_TYPE_UNSIGNED
)) {
433 LLVMValueRef tmps
[LP_MAX_VECTOR_LENGTH
/4];
438 * Unpack a pixel at a time into a <4 x float> RGBA vector
441 for (k
= 0; k
< num_pixels
; ++k
) {
444 packed
= lp_build_gather_elem(gallivm
, num_pixels
,
445 format_desc
->block
.bits
, 32,
446 base_ptr
, offset
, k
);
448 tmps
[k
] = lp_build_unpack_arith_rgba_aos(gallivm
,
456 * TODO: We could avoid floating conversion for integer to
457 * integer conversions.
460 if (gallivm_debug
& GALLIVM_DEBUG_PERF
&& !type
.floating
) {
461 debug_printf("%s: unpacking %s with floating point\n",
462 __FUNCTION__
, format_desc
->short_name
);
465 lp_build_conv(gallivm
,
466 lp_float32_vec4_type(),
468 tmps
, num_pixels
, &res
, 1);
470 return lp_build_format_swizzle_aos(format_desc
, &bld
, res
);
474 * YUV / subsampled formats
477 if (format_desc
->layout
== UTIL_FORMAT_LAYOUT_SUBSAMPLED
) {
478 struct lp_type tmp_type
;
481 memset(&tmp_type
, 0, sizeof tmp_type
);
483 tmp_type
.length
= num_pixels
* 4;
484 tmp_type
.norm
= TRUE
;
486 tmp
= lp_build_fetch_subsampled_rgba_aos(gallivm
,
493 lp_build_conv(gallivm
,
501 * Fallback to util_format_description::fetch_rgba_8unorm().
504 if (format_desc
->fetch_rgba_8unorm
&&
505 !type
.floating
&& type
.width
== 8 && !type
.sign
&& type
.norm
) {
507 * Fallback to calling util_format_description::fetch_rgba_8unorm.
509 * This is definitely not the most efficient way of fetching pixels, as
510 * we miss the opportunity to do vectorization, but this it is a
511 * convenient for formats or scenarios for which there was no opportunity
512 * or incentive to optimize.
515 LLVMTypeRef i8t
= LLVMInt8TypeInContext(gallivm
->context
);
516 LLVMTypeRef pi8t
= LLVMPointerType(i8t
, 0);
517 LLVMTypeRef i32t
= LLVMInt32TypeInContext(gallivm
->context
);
518 LLVMValueRef function
;
519 LLVMValueRef tmp_ptr
;
524 if (gallivm_debug
& GALLIVM_DEBUG_PERF
) {
525 debug_printf("%s: falling back to util_format_%s_fetch_rgba_8unorm\n",
526 __FUNCTION__
, format_desc
->short_name
);
530 * Declare and bind format_desc->fetch_rgba_8unorm().
535 * Function to call looks like:
536 * fetch(uint8_t *dst, const uint8_t *src, unsigned i, unsigned j)
538 LLVMTypeRef ret_type
;
539 LLVMTypeRef arg_types
[4];
540 LLVMTypeRef function_type
;
542 ret_type
= LLVMVoidTypeInContext(gallivm
->context
);
547 function_type
= LLVMFunctionType(ret_type
, arg_types
,
548 Elements(arg_types
), 0);
550 /* make const pointer for the C fetch_rgba_8unorm function */
551 function
= lp_build_const_int_pointer(gallivm
,
552 func_to_pointer((func_pointer
) format_desc
->fetch_rgba_8unorm
));
554 /* cast the callee pointer to the function's type */
555 function
= LLVMBuildBitCast(builder
, function
,
556 LLVMPointerType(function_type
, 0),
560 tmp_ptr
= lp_build_alloca(gallivm
, i32t
, "");
562 res
= LLVMGetUndef(LLVMVectorType(i32t
, num_pixels
));
565 * Invoke format_desc->fetch_rgba_8unorm() for each pixel and insert the result
566 * in the SoA vectors.
569 for (k
= 0; k
< num_pixels
; ++k
) {
570 LLVMValueRef index
= lp_build_const_int32(gallivm
, k
);
571 LLVMValueRef args
[4];
573 args
[0] = LLVMBuildBitCast(builder
, tmp_ptr
, pi8t
, "");
574 args
[1] = lp_build_gather_elem_ptr(gallivm
, num_pixels
,
575 base_ptr
, offset
, k
);
577 if (num_pixels
== 1) {
582 args
[2] = LLVMBuildExtractElement(builder
, i
, index
, "");
583 args
[3] = LLVMBuildExtractElement(builder
, j
, index
, "");
586 LLVMBuildCall(builder
, function
, args
, Elements(args
), "");
588 tmp
= LLVMBuildLoad(builder
, tmp_ptr
, "");
590 if (num_pixels
== 1) {
594 res
= LLVMBuildInsertElement(builder
, res
, tmp
, index
, "");
598 /* Bitcast from <n x i32> to <4n x i8> */
599 res
= LLVMBuildBitCast(builder
, res
, bld
.vec_type
, "");
606 * Fallback to util_format_description::fetch_rgba_float().
609 if (format_desc
->fetch_rgba_float
) {
611 * Fallback to calling util_format_description::fetch_rgba_float.
613 * This is definitely not the most efficient way of fetching pixels, as
614 * we miss the opportunity to do vectorization, but this it is a
615 * convenient for formats or scenarios for which there was no opportunity
616 * or incentive to optimize.
619 LLVMTypeRef f32t
= LLVMFloatTypeInContext(gallivm
->context
);
620 LLVMTypeRef f32x4t
= LLVMVectorType(f32t
, 4);
621 LLVMTypeRef pf32t
= LLVMPointerType(f32t
, 0);
622 LLVMTypeRef pi8t
= LLVMPointerType(LLVMInt8TypeInContext(gallivm
->context
), 0);
623 LLVMTypeRef i32t
= LLVMInt32TypeInContext(gallivm
->context
);
624 LLVMValueRef function
;
625 LLVMValueRef tmp_ptr
;
626 LLVMValueRef tmps
[LP_MAX_VECTOR_LENGTH
/4];
630 if (gallivm_debug
& GALLIVM_DEBUG_PERF
) {
631 debug_printf("%s: falling back to util_format_%s_fetch_rgba_float\n",
632 __FUNCTION__
, format_desc
->short_name
);
636 * Declare and bind format_desc->fetch_rgba_float().
641 * Function to call looks like:
642 * fetch(float *dst, const uint8_t *src, unsigned i, unsigned j)
644 LLVMTypeRef ret_type
;
645 LLVMTypeRef arg_types
[4];
646 LLVMTypeRef function_type
;
648 ret_type
= LLVMVoidTypeInContext(gallivm
->context
);
649 arg_types
[0] = pf32t
;
653 function_type
= LLVMFunctionType(ret_type
, arg_types
,
654 Elements(arg_types
), 0);
656 /* Note: we're using this casting here instead of LLVMAddGlobalMapping()
657 * to work around a bug in LLVM 2.6, and for efficiency/simplicity.
660 /* make const pointer for the C fetch_rgba_float function */
661 function
= lp_build_const_int_pointer(gallivm
,
662 func_to_pointer((func_pointer
) format_desc
->fetch_rgba_float
));
664 /* cast the callee pointer to the function's type */
665 function
= LLVMBuildBitCast(builder
, function
,
666 LLVMPointerType(function_type
, 0),
670 tmp_ptr
= lp_build_alloca(gallivm
, f32x4t
, "");
673 * Invoke format_desc->fetch_rgba_float() for each pixel and insert the result
674 * in the SoA vectors.
677 for (k
= 0; k
< num_pixels
; ++k
) {
678 LLVMValueRef args
[4];
680 args
[0] = LLVMBuildBitCast(builder
, tmp_ptr
, pf32t
, "");
681 args
[1] = lp_build_gather_elem_ptr(gallivm
, num_pixels
,
682 base_ptr
, offset
, k
);
684 if (num_pixels
== 1) {
689 LLVMValueRef index
= lp_build_const_int32(gallivm
, k
);
690 args
[2] = LLVMBuildExtractElement(builder
, i
, index
, "");
691 args
[3] = LLVMBuildExtractElement(builder
, j
, index
, "");
694 LLVMBuildCall(builder
, function
, args
, Elements(args
), "");
696 tmps
[k
] = LLVMBuildLoad(builder
, tmp_ptr
, "");
699 lp_build_conv(gallivm
,
700 lp_float32_vec4_type(),
702 tmps
, num_pixels
, &res
, 1);
708 return lp_build_undef(gallivm
, type
);