src/broadcom/compiler/v3d_nir_lower_image_load_store.c

   1 /*
   2  * Copyright © 2018 Intel Corporation
   3  * Copyright © 2018 Broadcom
   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  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9  * and/or sell copies of the Software, and to permit persons to whom the
  10  * 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 NONINFRINGEMENT.  IN NO EVENT SHALL
  19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  21  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  22  * IN THE SOFTWARE.
  23  */
  24
  25 #include "v3d_compiler.h"
  26 #include "compiler/nir/nir_builder.h"
  27 #include "compiler/nir/nir_format_convert.h"
  28
  29 /** @file v3d_nir_lower_image_load_store.c
  30  *
  31  * Performs any necessary lowering of GL_ARB_shader_image_load_store
  32  * operations.
  33  *
  34  * On V3D 4.x, we just need to do format conversion for stores such that the
  35  * GPU can effectively memcpy the arguments (in increments of 32-bit words)
  36  * into the texel.  Loads are the same as texturing, where we may need to
  37  * unpack from 16-bit ints or floats.
  38  *
  39  * On V3D 3.x, to implement image load store we would need to do manual tiling
  40  * calculations and load/store using the TMU general memory access path.
  41  */
  42
  43 bool
  44 v3d_gl_format_is_return_32(enum pipe_format format)
  45 {
  46         const struct util_format_description *desc =
  47                 util_format_description(format);
  48         const struct util_format_channel_description *chan = &desc->channel[0];
  49
  50         return chan->size > 16 || (chan->size == 16 && chan->normalized);
  51 }
  52
  53 /* Packs a 32-bit vector of colors in the range [0, (1 << bits[i]) - 1] to a
  54  * 32-bit SSA value, with as many channels as necessary to store all the bits
  55  */
  56 static nir_ssa_def *
  57 pack_bits(nir_builder *b, nir_ssa_def *color, const unsigned *bits,
  58           int num_components, bool mask)
  59 {
  60         nir_ssa_def *results[4];
  61         int offset = 0;
  62         for (int i = 0; i < num_components; i++) {
  63                 nir_ssa_def *chan = nir_channel(b, color, i);
  64
  65                 /* Channels being stored shouldn't cross a 32-bit boundary. */
  66                 assert((offset & ~31) == ((offset + bits[i] - 1) & ~31));
  67
  68                 if (mask) {
  69                         chan = nir_iand(b, chan,
  70                                         nir_imm_int(b, (1 << bits[i]) - 1));
  71                 }
  72
  73                 if (offset % 32 == 0) {
  74                         results[offset / 32] = chan;
  75                 } else {
  76                         results[offset / 32] =
  77                                 nir_ior(b, results[offset / 32],
  78                                         nir_ishl(b, chan,
  79                                                  nir_imm_int(b, offset % 32)));
  80                 }
  81                 offset += bits[i];
  82         }
  83
  84         return nir_vec(b, results, DIV_ROUND_UP(offset, 32));
  85 }
  86
  87 static void
  88 v3d_nir_lower_image_store(nir_builder *b, nir_intrinsic_instr *instr)
  89 {
  90         enum pipe_format format = nir_intrinsic_format(instr);
  91         const struct util_format_description *desc =
  92                 util_format_description(format);
  93         const struct util_format_channel_description *r_chan = &desc->channel[0];
  94         unsigned num_components = util_format_get_nr_components(format);
  95
  96         b->cursor = nir_before_instr(&instr->instr);
  97
  98         nir_ssa_def *color = nir_channels(b,
  99                                           nir_ssa_for_src(b, instr->src[3], 4),
 100                                           (1 << num_components) - 1);
 101         nir_ssa_def *formatted = NULL;
 102
 103         if (format == PIPE_FORMAT_R11G11B10_FLOAT) {
 104                 formatted = nir_format_pack_11f11f10f(b, color);
 105         } else if (format == PIPE_FORMAT_R9G9B9E5_FLOAT) {
 106                 formatted = nir_format_pack_r9g9b9e5(b, color);
 107         } else if (r_chan->size == 32) {
 108                 /* For 32-bit formats, we just have to move the vector
 109                  * across (possibly reducing the number of channels).
 110                  */
 111                 formatted = color;
 112         } else {
 113                 static const unsigned bits_8[4] = {8, 8, 8, 8};
 114                 static const unsigned bits_16[4] = {16, 16, 16, 16};
 115                 static const unsigned bits_1010102[4] = {10, 10, 10, 2};
 116                 const unsigned *bits;
 117
 118                 switch (r_chan->size) {
 119                 case 8:
 120                         bits = bits_8;
 121                         break;
 122                 case 10:
 123                         bits = bits_1010102;
 124                         break;
 125                 case 16:
 126                         bits = bits_16;
 127                         break;
 128                 default:
 129                         unreachable("unrecognized bits");
 130                 }
 131
 132                 bool pack_mask = false;
 133                 if (r_chan->pure_integer &&
 134                     r_chan->type == UTIL_FORMAT_TYPE_SIGNED) {
 135                         formatted = nir_format_clamp_sint(b, color, bits);
 136                         pack_mask = true;
 137                 } else if (r_chan->pure_integer &&
 138                            r_chan->type == UTIL_FORMAT_TYPE_UNSIGNED) {
 139                         formatted = nir_format_clamp_uint(b, color, bits);
 140                 } else if (r_chan->normalized &&
 141                            r_chan->type == UTIL_FORMAT_TYPE_SIGNED) {
 142                         formatted = nir_format_float_to_snorm(b, color, bits);
 143                         pack_mask = true;
 144                 } else if (r_chan->normalized &&
 145                            r_chan->type == UTIL_FORMAT_TYPE_UNSIGNED) {
 146                         formatted = nir_format_float_to_unorm(b, color, bits);
 147                 } else {
 148                         assert(r_chan->size == 16);
 149                         assert(r_chan->type == UTIL_FORMAT_TYPE_FLOAT);
 150                         formatted = nir_format_float_to_half(b, color);
 151                 }
 152
 153                 formatted = pack_bits(b, formatted, bits, num_components,
 154                                       pack_mask);
 155         }
 156
 157         nir_instr_rewrite_src(&instr->instr, &instr->src[3],
 158                               nir_src_for_ssa(formatted));
 159         instr->num_components = formatted->num_components;
 160 }
 161
 162 static void
 163 v3d_nir_lower_image_load(nir_builder *b, nir_intrinsic_instr *instr)
 164 {
 165         static const unsigned bits16[] = {16, 16, 16, 16};
 166         enum pipe_format format = nir_intrinsic_format(instr);
 167
 168         if (v3d_gl_format_is_return_32(format))
 169                 return;
 170
 171         b->cursor = nir_after_instr(&instr->instr);
 172
 173         assert(instr->dest.is_ssa);
 174         nir_ssa_def *result = &instr->dest.ssa;
 175         if (util_format_is_pure_uint(format)) {
 176                 result = nir_format_unpack_uint(b, result, bits16, 4);
 177         } else if (util_format_is_pure_sint(format)) {
 178                 result = nir_format_unpack_sint(b, result, bits16, 4);
 179         } else {
 180             nir_ssa_def *rg = nir_channel(b, result, 0);
 181             nir_ssa_def *ba = nir_channel(b, result, 1);
 182             result = nir_vec4(b,
 183                               nir_unpack_half_2x16_split_x(b, rg),
 184                               nir_unpack_half_2x16_split_y(b, rg),
 185                               nir_unpack_half_2x16_split_x(b, ba),
 186                               nir_unpack_half_2x16_split_y(b, ba));
 187         }
 188
 189         nir_ssa_def_rewrite_uses_after(&instr->dest.ssa, nir_src_for_ssa(result),
 190                                        result->parent_instr);
 191 }
 192
 193 void
 194 v3d_nir_lower_image_load_store(nir_shader *s)
 195 {
 196         nir_foreach_function(function, s) {
 197                 if (!function->impl)
 198                         continue;
 199
 200                 nir_builder b;
 201                 nir_builder_init(&b, function->impl);
 202
 203                 nir_foreach_block(block, function->impl) {
 204                         nir_foreach_instr_safe(instr, block) {
 205                                 if (instr->type != nir_instr_type_intrinsic)
 206                                         continue;
 207
 208                                 nir_intrinsic_instr *intr =
 209                                         nir_instr_as_intrinsic(instr);
 210
 211                                 switch (intr->intrinsic) {
 212                                 case nir_intrinsic_image_load:
 213                                         v3d_nir_lower_image_load(&b, intr);
 214                                         break;
 215                                 case nir_intrinsic_image_store:
 216                                         v3d_nir_lower_image_store(&b, intr);
 217                                         break;
 218                                 default:
 219                                         break;
 220                                 }
 221                         }
 222                 }
 223
 224                 nir_metadata_preserve(function->impl,
 225                                       nir_metadata_block_index |
 226                                       nir_metadata_dominance);
 227         }
 228 }