src/gallium/drivers/cell/ppu/cell_vertex_fetch.c

   1 /*
   2  * (C) Copyright IBM Corporation 2008
   3  * All Rights Reserved.
   4  *
   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:
  11  *
  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
  14  * Software.
  15  *
  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  * AUTHORS, COPYRIGHT HOLDERS, 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.
  23  */
  24
  25 #include <inttypes.h>
  26 #include "pipe/p_defines.h"
  27 #include "pipe/p_context.h"
  28 #include "pipe/p_format.h"
  29
  30 #include "../auxiliary/draw/draw_context.h"
  31 #include "../auxiliary/draw/draw_private.h"
  32
  33 #include "cell_context.h"
  34 #include "rtasm/rtasm_ppc_spe.h"
  35
  36
  37 /**
  38  * Emit a 4x4 matrix transpose operation
  39  *
  40  * \param p         Function that the transpose operation is to be appended to
  41  * \param row0      Register containing row 0 of the source matrix
  42  * \param row1      Register containing row 1 of the source matrix
  43  * \param row2      Register containing row 2 of the source matrix
  44  * \param row3      Register containing row 3 of the source matrix
  45  * \param dest_ptr  Register containing the address of the destination matrix
  46  * \param shuf_ptr  Register containing the address of the shuffled data
  47  * \param count     Number of colums to actually be written to the destination
  48  *
  49  * \note
  50  * This function assumes that the registers named by \c row0, \c row1,
  51  * \c row2, and \c row3 are scratch and can be modified by the generated code.
  52  * Furthermore, these registers will be released, via calls to
  53  * \c release_register, by this function.
  54  *
  55  * \note
  56  * This function requires that four temporary are available on entry.
  57  */
  58 static void
  59 emit_matrix_transpose(struct spe_function *p,
  60                       unsigned row0, unsigned row1, unsigned row2,
  61                       unsigned row3, unsigned dest_ptr,
  62                       unsigned shuf_ptr, unsigned count)
  63 {
  64    int shuf_hi = spe_allocate_available_register(p);
  65    int shuf_lo = spe_allocate_available_register(p);
  66    int t1 = spe_allocate_available_register(p);
  67    int t2 = spe_allocate_available_register(p);
  68    int t3;
  69    int t4;
  70    int col0;
  71    int col1;
  72    int col2;
  73    int col3;
  74
  75
  76    spe_lqd(p, shuf_hi, shuf_ptr, 3);
  77    spe_lqd(p, shuf_lo, shuf_ptr, 4);
  78    spe_shufb(p, t1, row0, row2, shuf_hi);
  79    spe_shufb(p, t2, row0, row2, shuf_lo);
  80
  81
  82    /* row0 and row2 are now no longer needed.  Re-use those registers as
  83     * temporaries.
  84     */
  85    t3 = row0;
  86    t4 = row2;
  87
  88    spe_shufb(p, t3, row1, row3, shuf_hi);
  89    spe_shufb(p, t4, row1, row3, shuf_lo);
  90
  91
  92    /* row1 and row3 are now no longer needed.  Re-use those registers as
  93     * temporaries.
  94     */
  95    col0 = row1;
  96    col1 = row3;
  97
  98    spe_shufb(p, col0, t1, t3, shuf_hi);
  99    if (count > 1) {
 100       spe_shufb(p, col1, t1, t3, shuf_lo);
 101    }
 102
 103    /* t1 and t3 are now no longer needed.  Re-use those registers as
 104     * temporaries.
 105     */
 106    col2 = t1;
 107    col3 = t3;
 108
 109    if (count > 2) {
 110       spe_shufb(p, col2, t2, t4, shuf_hi);
 111    }
 112
 113    if (count > 3) {
 114       spe_shufb(p, col3, t2, t4, shuf_lo);
 115    }
 116
 117
 118    /* Store the results.  Remember that the stqd instruction is encoded using
 119     * the qword offset (stand-alone assemblers to the byte-offset to
 120     * qword-offset conversion for you), so the byte-offset needs be divided by
 121     * 16.
 122     */
 123    switch (count) {
 124    case 4:
 125       spe_stqd(p, col3, dest_ptr, 3);
 126    case 3:
 127       spe_stqd(p, col2, dest_ptr, 2);
 128    case 2:
 129       spe_stqd(p, col1, dest_ptr, 1);
 130    case 1:
 131       spe_stqd(p, col0, dest_ptr, 0);
 132    }
 133
 134
 135    /* Release all of the temporary registers used.
 136     */
 137    spe_release_register(p, col0);
 138    spe_release_register(p, col1);
 139    spe_release_register(p, col2);
 140    spe_release_register(p, col3);
 141    spe_release_register(p, shuf_hi);
 142    spe_release_register(p, shuf_lo);
 143    spe_release_register(p, t2);
 144    spe_release_register(p, t4);
 145 }
 146
 147
 148 static void
 149 emit_fetch(struct spe_function *p,
 150            unsigned in_ptr, unsigned *offset,
 151            unsigned out_ptr, unsigned shuf_ptr,
 152            enum pipe_format format)
 153 {
 154    const unsigned count = (pf_size_x(format) != 0) + (pf_size_y(format) != 0)
 155        + (pf_size_z(format) != 0) + (pf_size_w(format) != 0);
 156    const unsigned type = pf_type(format);
 157    const unsigned bytes = pf_size_x(format);
 158
 159    int v0 = spe_allocate_available_register(p);
 160    int v1 = spe_allocate_available_register(p);
 161    int v2 = spe_allocate_available_register(p);
 162    int v3 = spe_allocate_available_register(p);
 163    int tmp = spe_allocate_available_register(p);
 164    int float_zero = -1;
 165    int float_one = -1;
 166    float scale_signed = 0.0;
 167    float scale_unsigned = 0.0;
 168
 169    spe_lqd(p, v0, in_ptr, 0 + offset[0]);
 170    spe_lqd(p, v1, in_ptr, 1 + offset[0]);
 171    spe_lqd(p, v2, in_ptr, 2 + offset[0]);
 172    spe_lqd(p, v3, in_ptr, 3 + offset[0]);
 173    offset[0] += 4;
 174
 175    switch (bytes) {
 176    case 1:
 177       scale_signed = 1.0f / 127.0f;
 178       scale_unsigned = 1.0f / 255.0f;
 179       spe_lqd(p, tmp, shuf_ptr, 1);
 180       spe_shufb(p, v0, v0, v0, tmp);
 181       spe_shufb(p, v1, v1, v1, tmp);
 182       spe_shufb(p, v2, v2, v2, tmp);
 183       spe_shufb(p, v3, v3, v3, tmp);
 184       break;
 185    case 2:
 186       scale_signed = 1.0f / 32767.0f;
 187       scale_unsigned = 1.0f / 65535.0f;
 188       spe_lqd(p, tmp, shuf_ptr, 2);
 189       spe_shufb(p, v0, v0, v0, tmp);
 190       spe_shufb(p, v1, v1, v1, tmp);
 191       spe_shufb(p, v2, v2, v2, tmp);
 192       spe_shufb(p, v3, v3, v3, tmp);
 193       break;
 194    case 4:
 195       scale_signed = 1.0f / 2147483647.0f;
 196       scale_unsigned = 1.0f / 4294967295.0f;
 197       break;
 198    default:
 199       assert(0);
 200       break;
 201    }
 202
 203    switch (type) {
 204    case PIPE_FORMAT_TYPE_FLOAT:
 205       break;
 206    case PIPE_FORMAT_TYPE_UNORM:
 207       spe_ilhu(p, tmp, ((unsigned) scale_unsigned) >> 16);
 208       spe_iohl(p, tmp, ((unsigned) scale_unsigned) & 0x0ffff);
 209       spe_cuflt(p, v0, v0, 0);
 210       spe_fm(p, v0, v0, tmp);
 211       break;
 212    case PIPE_FORMAT_TYPE_SNORM:
 213       spe_ilhu(p, tmp, ((unsigned) scale_signed) >> 16);
 214       spe_iohl(p, tmp, ((unsigned) scale_signed) & 0x0ffff);
 215       spe_csflt(p, v0, v0, 0);
 216       spe_fm(p, v0, v0, tmp);
 217       break;
 218    case PIPE_FORMAT_TYPE_USCALED:
 219       spe_cuflt(p, v0, v0, 0);
 220       break;
 221    case PIPE_FORMAT_TYPE_SSCALED:
 222       spe_csflt(p, v0, v0, 0);
 223       break;
 224    }
 225
 226
 227    if (count < 4) {
 228       float_one = spe_allocate_available_register(p);
 229       spe_il(p, float_one, 1);
 230       spe_cuflt(p, float_one, float_one, 0);
 231
 232       if (count < 3) {
 233          float_zero = spe_allocate_available_register(p);
 234          spe_il(p, float_zero, 0);
 235       }
 236    }
 237
 238    spe_release_register(p, tmp);
 239
 240    emit_matrix_transpose(p, v0, v1, v2, v3, out_ptr, shuf_ptr, count);
 241
 242    switch (count) {
 243    case 1:
 244       spe_stqd(p, float_zero, out_ptr, 1);
 245    case 2:
 246       spe_stqd(p, float_zero, out_ptr, 2);
 247    case 3:
 248       spe_stqd(p, float_one, out_ptr, 3);
 249    }
 250
 251    if (float_zero != -1) {
 252       spe_release_register(p, float_zero);
 253    }
 254
 255    if (float_one != -1) {
 256       spe_release_register(p, float_one);
 257    }
 258 }
 259
 260
 261 void cell_update_vertex_fetch(struct draw_context *draw)
 262 {
 263 #if 0
 264    struct cell_context *const cell =
 265        (struct cell_context *) draw->driver_private;
 266    struct spe_function *p = &cell->attrib_fetch;
 267    unsigned function_index[PIPE_MAX_ATTRIBS];
 268    unsigned unique_attr_formats;
 269    int out_ptr;
 270    int in_ptr;
 271    int shuf_ptr;
 272    unsigned i;
 273    unsigned j;
 274
 275
 276    /* Determine how many unique input attribute formats there are.  At the
 277     * same time, store the index of the lowest numbered attribute that has
 278     * the same format as any non-unique format.
 279     */
 280    unique_attr_formats = 1;
 281    function_index[0] = 0;
 282    for (i = 1; i < draw->vertex_fetch.nr_attrs; i++) {
 283       const enum pipe_format curr_fmt = draw->vertex_element[i].src_format;
 284
 285       for (j = 0; j < i; j++) {
 286          if (curr_fmt == draw->vertex_element[j].src_format) {
 287             break;
 288          }
 289       }
 290
 291       if (j == i) {
 292          unique_attr_formats++;
 293       }
 294
 295       function_index[i] = j;
 296    }
 297
 298
 299    /* Each fetch function can be a maximum of 34 instructions (note: this is
 300     * actually a slight over-estimate).  That means (34 * 4) = 136 bytes
 301     * each maximum.
 302     */
 303    spe_init_func(p, 136 * unique_attr_formats);
 304
 305
 306    /* Allocate registers for the function's input parameters.
 307     */
 308    out_ptr = spe_allocate_register(p, 3);
 309    in_ptr = spe_allocate_register(p, 4);
 310    shuf_ptr = spe_allocate_register(p, 5);
 311
 312
 313    /* Generate code for the individual attribute fetch functions.
 314     */
 315    for (i = 0; i < draw->vertex_fetch.nr_attrs; i++) {
 316       unsigned offset;
 317
 318       if (function_index[i] == i) {
 319          cell->attrib_fetch_offsets[i] = (unsigned) ((void *) p->csr
 320                                                      - (void *) p->store);
 321
 322          offset = 0;
 323          emit_fetch(p, in_ptr, &offset, out_ptr, shuf_ptr,
 324                     draw->vertex_element[i].src_format);
 325          spe_bi(p, 0, 0, 0);
 326
 327          /* Round up to the next 16-byte boundary.
 328           */
 329          if ((((unsigned) p->store) & 0x0f) != 0) {
 330             const unsigned align = ((unsigned) p->store) & 0x0f;
 331             p->store = (uint32_t *) (((void *) p->store) + align);
 332          }
 333       } else {
 334          /* Use the same function entry-point as a previously seen attribute
 335           * with the same format.
 336           */
 337          cell->attrib_fetch_offsets[i] =
 338              cell->attrib_fetch_offsets[function_index[i]];
 339       }
 340    }
 341 #else
 342    assert(0);
 343 #endif
 344 }