src/gallium/auxiliary/gallivm/lp_bld_format_soa.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2009 VMware, Inc.
   4  * All Rights Reserved.
   5  *
   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:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * of the Software.
  17  *
  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.
  25  *
  26  **************************************************************************/
  27
  28
  29 #include "pipe/p_defines.h"
  30
  31 #include "util/u_format.h"
  32 #include "util/u_memory.h"
  33 #include "util/u_string.h"
  34
  35 #include "lp_bld_type.h"
  36 #include "lp_bld_const.h"
  37 #include "lp_bld_conv.h"
  38 #include "lp_bld_swizzle.h"
  39 #include "lp_bld_gather.h"
  40 #include "lp_bld_debug.h"
  41 #include "lp_bld_format.h"
  42
  43
  44 void
  45 lp_build_format_swizzle_soa(const struct util_format_description *format_desc,
  46                             struct lp_build_context *bld,
  47                             const LLVMValueRef *unswizzled,
  48                             LLVMValueRef swizzled_out[4])
  49 {
  50    assert(UTIL_FORMAT_SWIZZLE_0 == PIPE_SWIZZLE_ZERO);
  51    assert(UTIL_FORMAT_SWIZZLE_1 == PIPE_SWIZZLE_ONE);
  52
  53    if (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) {
  54       /*
  55        * Return zzz1 for depth-stencil formats.
  56        *
  57        * XXX: Allow to control the depth swizzle with an additional parameter,
  58        * as the caller may wish another depth swizzle, or retain the stencil
  59        * value.
  60        */
  61       enum util_format_swizzle swizzle = format_desc->swizzle[0];
  62       LLVMValueRef depth = lp_build_swizzle_soa_channel(bld, unswizzled, swizzle);
  63       swizzled_out[2] = swizzled_out[1] = swizzled_out[0] = depth;
  64       swizzled_out[3] = bld->one;
  65    }
  66    else {
  67       unsigned chan;
  68       for (chan = 0; chan < 4; ++chan) {
  69          enum util_format_swizzle swizzle = format_desc->swizzle[chan];
  70          swizzled_out[chan] = lp_build_swizzle_soa_channel(bld, unswizzled, swizzle);
  71       }
  72    }
  73 }
  74
  75
  76 /**
  77  * Unpack several pixels in SoA.
  78  *
  79  * It takes a vector of packed pixels:
  80  *
  81  *   packed = {P0, P1, P2, P3, ..., Pn}
  82  *
  83  * And will produce four vectors:
  84  *
  85  *   red    = {R0, R1, R2, R3, ..., Rn}
  86  *   green  = {G0, G1, G2, G3, ..., Gn}
  87  *   blue   = {B0, B1, B2, B3, ..., Bn}
  88  *   alpha  = {A0, A1, A2, A3, ..., An}
  89  *
  90  * It requires that a packed pixel fits into an element of the output
  91  * channels. The common case is when converting pixel with a depth of 32 bit or
  92  * less into floats.
  93  *
  94  * \param format_desc  the format of the 'packed' incoming pixel vector
  95  * \param type  the desired type for rgba_out (type.length = n, above)
  96  * \param packed  the incoming vector of packed pixels
  97  * \param rgba_out  returns the SoA R,G,B,A vectors
  98  */
  99 void
 100 lp_build_unpack_rgba_soa(LLVMBuilderRef builder,
 101                          const struct util_format_description *format_desc,
 102                          struct lp_type type,
 103                          LLVMValueRef packed,
 104                          LLVMValueRef rgba_out[4])
 105 {
 106    struct lp_build_context bld;
 107    LLVMValueRef inputs[4];
 108    unsigned start;
 109    unsigned chan;
 110
 111    assert(format_desc->layout == UTIL_FORMAT_LAYOUT_PLAIN);
 112    assert(format_desc->block.width == 1);
 113    assert(format_desc->block.height == 1);
 114    assert(format_desc->block.bits <= type.width);
 115    /* FIXME: Support more output types */
 116    assert(type.floating);
 117    assert(type.width == 32);
 118
 119    lp_build_context_init(&bld, builder, type);
 120
 121    /* Decode the input vector components */
 122    start = 0;
 123    for (chan = 0; chan < format_desc->nr_channels; ++chan) {
 124       const unsigned width = format_desc->channel[chan].size;
 125       const unsigned stop = start + width;
 126       LLVMValueRef input;
 127
 128       input = packed;
 129
 130       switch(format_desc->channel[chan].type) {
 131       case UTIL_FORMAT_TYPE_VOID:
 132          input = lp_build_undef(type);
 133          break;
 134
 135       case UTIL_FORMAT_TYPE_UNSIGNED:
 136          /*
 137           * Align the LSB
 138           */
 139
 140          if (start) {
 141             input = LLVMBuildLShr(builder, input, lp_build_const_int_vec(type, start), "");
 142          }
 143
 144          /*
 145           * Zero the MSBs
 146           */
 147
 148          if (stop < format_desc->block.bits) {
 149             unsigned mask = ((unsigned long long)1 << width) - 1;
 150             input = LLVMBuildAnd(builder, input, lp_build_const_int_vec(type, mask), "");
 151          }
 152
 153          /*
 154           * Type conversion
 155           */
 156
 157          if (type.floating) {
 158             if(format_desc->channel[chan].normalized)
 159                input = lp_build_unsigned_norm_to_float(builder, width, type, input);
 160             else
 161                input = LLVMBuildSIToFP(builder, input, lp_build_vec_type(type), "");
 162          }
 163          else {
 164             /* FIXME */
 165             assert(0);
 166             input = lp_build_undef(type);
 167          }
 168
 169          break;
 170
 171       case UTIL_FORMAT_TYPE_SIGNED:
 172          /*
 173           * Align the sign bit first.
 174           */
 175
 176          if (stop < type.width) {
 177             unsigned bits = type.width - stop;
 178             LLVMValueRef bits_val = lp_build_const_int_vec(type, bits);
 179             input = LLVMBuildShl(builder, input, bits_val, "");
 180          }
 181
 182          /*
 183           * Align the LSB (with an arithmetic shift to preserve the sign)
 184           */
 185
 186          if (format_desc->channel[chan].size < type.width) {
 187             unsigned bits = type.width - format_desc->channel[chan].size;
 188             LLVMValueRef bits_val = lp_build_const_int_vec(type, bits);
 189             input = LLVMBuildAShr(builder, input, bits_val, "");
 190          }
 191
 192          /*
 193           * Type conversion
 194           */
 195
 196          if (type.floating) {
 197             input = LLVMBuildSIToFP(builder, input, lp_build_vec_type(type), "");
 198             if (format_desc->channel[chan].normalized) {
 199                double scale = 1.0 / ((1 << (format_desc->channel[chan].size - 1)) - 1);
 200                LLVMValueRef scale_val = lp_build_const_vec(type, scale);
 201                input = LLVMBuildFMul(builder, input, scale_val, "");
 202             }
 203          }
 204          else {
 205             /* FIXME */
 206             assert(0);
 207             input = lp_build_undef(type);
 208          }
 209
 210          break;
 211
 212       case UTIL_FORMAT_TYPE_FLOAT:
 213          if (type.floating) {
 214             assert(start == 0);
 215             assert(stop == 32);
 216             assert(type.width == 32);
 217             input = LLVMBuildBitCast(builder, input, lp_build_vec_type(type), "");
 218          }
 219          else {
 220             /* FIXME */
 221             assert(0);
 222             input = lp_build_undef(type);
 223          }
 224          break;
 225
 226       case UTIL_FORMAT_TYPE_FIXED:
 227          if (type.floating) {
 228             double scale = 1.0 / ((1 << (format_desc->channel[chan].size/2)) - 1);
 229             LLVMValueRef scale_val = lp_build_const_vec(type, scale);
 230             input = LLVMBuildSIToFP(builder, input, lp_build_vec_type(type), "");
 231             input = LLVMBuildFMul(builder, input, scale_val, "");
 232          }
 233          else {
 234             /* FIXME */
 235             assert(0);
 236             input = lp_build_undef(type);
 237          }
 238          break;
 239
 240       default:
 241          assert(0);
 242          input = lp_build_undef(type);
 243          break;
 244       }
 245
 246       inputs[chan] = input;
 247
 248       start = stop;
 249    }
 250
 251    lp_build_format_swizzle_soa(format_desc, &bld, inputs, rgba_out);
 252 }
 253
 254
 255 void
 256 lp_build_rgba8_to_f32_soa(LLVMBuilderRef builder,
 257                           struct lp_type dst_type,
 258                           LLVMValueRef packed,
 259                           LLVMValueRef *rgba)
 260 {
 261    LLVMValueRef mask = lp_build_const_int_vec(dst_type, 0xff);
 262    unsigned chan;
 263
 264    packed = LLVMBuildBitCast(builder, packed,
 265                              lp_build_int_vec_type(dst_type), "");
 266
 267    /* Decode the input vector components */
 268    for (chan = 0; chan < 4; ++chan) {
 269       unsigned start = chan*8;
 270       unsigned stop = start + 8;
 271       LLVMValueRef input;
 272
 273       input = packed;
 274
 275       if (start)
 276          input = LLVMBuildLShr(builder, input,
 277                                lp_build_const_int_vec(dst_type, start), "");
 278
 279       if (stop < 32)
 280          input = LLVMBuildAnd(builder, input, mask, "");
 281
 282       input = lp_build_unsigned_norm_to_float(builder, 8, dst_type, input);
 283
 284       rgba[chan] = input;
 285    }
 286 }
 287
 288
 289
 290 /**
 291  * Fetch a texels from a texture, returning them in SoA layout.
 292  *
 293  * \param type  the desired return type for 'rgba'.  The vector length
 294  *              is the number of texels to fetch
 295  *
 296  * \param base_ptr  points to start of the texture image block.  For non-
 297  *                  compressed formats, this simply points to the texel.
 298  *                  For compressed formats, it points to the start of the
 299  *                  compressed data block.
 300  *
 301  * \param i, j  the sub-block pixel coordinates.  For non-compressed formats
 302  *              these will always be (0,0).  For compressed formats, i will
 303  *              be in [0, block_width-1] and j will be in [0, block_height-1].
 304  */
 305 void
 306 lp_build_fetch_rgba_soa(LLVMBuilderRef builder,
 307                         const struct util_format_description *format_desc,
 308                         struct lp_type type,
 309                         LLVMValueRef base_ptr,
 310                         LLVMValueRef offset,
 311                         LLVMValueRef i,
 312                         LLVMValueRef j,
 313                         LLVMValueRef rgba_out[4])
 314 {
 315
 316    if (format_desc->layout == UTIL_FORMAT_LAYOUT_PLAIN &&
 317        (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
 318         format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) &&
 319        format_desc->block.width == 1 &&
 320        format_desc->block.height == 1 &&
 321        format_desc->block.bits <= type.width &&
 322        (format_desc->channel[0].type != UTIL_FORMAT_TYPE_FLOAT ||
 323         format_desc->channel[0].size == 32))
 324    {
 325       /*
 326        * The packed pixel fits into an element of the destination format. Put
 327        * the packed pixels into a vector and extract each component for all
 328        * vector elements in parallel.
 329        */
 330
 331       LLVMValueRef packed;
 332
 333       /*
 334        * gather the texels from the texture
 335        * Ex: packed = {BGRA, BGRA, BGRA, BGRA}.
 336        */
 337       packed = lp_build_gather(builder,
 338                                type.length,
 339                                format_desc->block.bits,
 340                                type.width,
 341                                base_ptr, offset);
 342
 343       /*
 344        * convert texels to float rgba
 345        */
 346       lp_build_unpack_rgba_soa(builder,
 347                                format_desc,
 348                                type,
 349                                packed, rgba_out);
 350       return;
 351    }
 352
 353    /*
 354     * Try calling lp_build_fetch_rgba_aos for all pixels.
 355     */
 356
 357    if (util_format_fits_8unorm(format_desc) &&
 358        type.floating && type.width == 32 && type.length == 4) {
 359       struct lp_type tmp_type;
 360       LLVMValueRef tmp;
 361
 362       memset(&tmp_type, 0, sizeof tmp_type);
 363       tmp_type.width = 8;
 364       tmp_type.length = type.length * 4;
 365       tmp_type.norm = TRUE;
 366
 367       tmp = lp_build_fetch_rgba_aos(builder, format_desc, tmp_type,
 368                                     base_ptr, offset, i, j);
 369
 370       lp_build_rgba8_to_f32_soa(builder,
 371                                 type,
 372                                 tmp,
 373                                 rgba_out);
 374
 375       return;
 376    }
 377
 378    /*
 379     * Fallback to calling lp_build_fetch_rgba_aos for each pixel.
 380     *
 381     * This is not the most efficient way of fetching pixels, as we
 382     * miss some opportunities to do vectorization, but this is
 383     * convenient for formats or scenarios for which there was no
 384     * opportunity or incentive to optimize.
 385     */
 386
 387    {
 388       unsigned k, chan;
 389       struct lp_type tmp_type;
 390
 391       if (gallivm_debug & GALLIVM_DEBUG_PERF) {
 392          debug_printf("%s: scalar unpacking of %s\n",
 393                       __FUNCTION__, format_desc->short_name);
 394       }
 395
 396       tmp_type = type;
 397       tmp_type.length = 4;
 398
 399       for (chan = 0; chan < 4; ++chan) {
 400          rgba_out[chan] = lp_build_undef(type);
 401       }
 402
 403       /* loop over number of pixels */
 404       for(k = 0; k < type.length; ++k) {
 405          LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), k, 0);
 406          LLVMValueRef offset_elem;
 407          LLVMValueRef i_elem, j_elem;
 408          LLVMValueRef tmp;
 409
 410          offset_elem = LLVMBuildExtractElement(builder, offset, index, "");
 411
 412          i_elem = LLVMBuildExtractElement(builder, i, index, "");
 413          j_elem = LLVMBuildExtractElement(builder, j, index, "");
 414
 415          /* Get a single float[4]={R,G,B,A} pixel */
 416          tmp = lp_build_fetch_rgba_aos(builder, format_desc, tmp_type,
 417                                        base_ptr, offset_elem,
 418                                        i_elem, j_elem);
 419
 420          /*
 421           * Insert the AoS tmp value channels into the SoA result vectors at
 422           * position = 'index'.
 423           */
 424          for (chan = 0; chan < 4; ++chan) {
 425             LLVMValueRef chan_val = LLVMConstInt(LLVMInt32Type(), chan, 0),
 426             tmp_chan = LLVMBuildExtractElement(builder, tmp, chan_val, "");
 427             rgba_out[chan] = LLVMBuildInsertElement(builder, rgba_out[chan],
 428                                                     tmp_chan, index, "");
 429          }
 430       }
 431    }
 432 }