src/panfrost/util/pan_lower_framebuffer.c

   1 /*
   2  * Copyright (C) 2020 Collabora, Ltd.
   3  *
   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:
  10  *
  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
  13  * Software.
  14  *
  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
  21  * IN THE SOFTWARE.
  22  *
  23  * Authors (Collabora):
  24  *      Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
  25  */
  26
  27 /**
  28  * Implements framebuffer format conversions in software for Midgard/Bifrost
  29  * blend shaders. This pass is designed for a single render target; Midgard
  30  * duplicates blend shaders for MRT to simplify everything. A particular
  31  * framebuffer format may be categorized as 1) typed load available, 2) typed
  32  * unpack available, or 3) software unpack only, and likewise for stores. The
  33  * first two types are handled in the compiler backend directly, so this module
  34  * is responsible for identifying type 3 formats (hardware dependent) and
  35  * inserting appropriate ALU code to perform the conversion from the packed
  36  * type to a designated unpacked type, and vice versa.
  37  *
  38  * The unpacked type depends on the format:
  39  *
  40  *      - For 32-bit float formats, 32-bit floats.
  41  *      - For other floats, 16-bit floats.
  42  *      - For 32-bit ints, 32-bit ints.
  43  *      - For 8-bit ints, 8-bit ints.
  44  *      - For other ints, 16-bit ints.
  45  *
  46  * The rationale is to optimize blending and logic op instructions by using the
  47  * smallest precision necessary to store the pixel losslessly.
  48  */
  49
  50 #include "compiler/nir/nir.h"
  51 #include "compiler/nir/nir_builder.h"
  52 #include "compiler/nir/nir_format_convert.h"
  53 #include "util/format/u_format.h"
  54 #include "pan_lower_framebuffer.h"
  55 #include "panfrost-quirks.h"
  56
  57 /* Determines the unpacked type best suiting a given format, so the rest of the
  58  * pipeline may be adjusted accordingly */
  59
  60 nir_alu_type
  61 pan_unpacked_type_for_format(const struct util_format_description *desc)
  62 {
  63         int c = util_format_get_first_non_void_channel(desc->format);
  64
  65         if (c == -1)
  66                 unreachable("Void format not renderable");
  67
  68         bool large = (desc->channel[c].size > 16);
  69         bool bit8 = (desc->channel[c].size == 8);
  70         assert(desc->channel[c].size <= 32);
  71
  72         if (desc->channel[c].normalized)
  73                 return large ? nir_type_float32 : nir_type_float16;
  74
  75         switch (desc->channel[c].type) {
  76         case UTIL_FORMAT_TYPE_UNSIGNED:
  77                 return bit8 ? nir_type_uint8 :
  78                         large ? nir_type_uint32 : nir_type_uint16;
  79         case UTIL_FORMAT_TYPE_SIGNED:
  80                 return bit8 ? nir_type_int8 :
  81                         large ? nir_type_int32 : nir_type_int16;
  82         case UTIL_FORMAT_TYPE_FLOAT:
  83                 return large ? nir_type_float32 : nir_type_float16;
  84         default:
  85                 unreachable("Format not renderable");
  86         }
  87 }
  88
  89 enum pan_format_class
  90 pan_format_class_load(const struct util_format_description *desc, unsigned quirks)
  91 {
  92         /* Check if we can do anything better than software architecturally */
  93         if (quirks & MIDGARD_NO_TYPED_BLEND_LOADS) {
  94                 return (quirks & NO_BLEND_PACKS)
  95                         ? PAN_FORMAT_SOFTWARE : PAN_FORMAT_PACK;
  96         }
  97
  98         /* Some formats are missing as typed on some GPUs but have unpacks */
  99         if (quirks & MIDGARD_MISSING_LOADS) {
 100                 switch (desc->format) {
 101                 case PIPE_FORMAT_R11G11B10_FLOAT:
 102                 case PIPE_FORMAT_R10G10B10A2_UNORM:
 103                 case PIPE_FORMAT_B10G10R10A2_UNORM:
 104                 case PIPE_FORMAT_R10G10B10X2_UNORM:
 105                 case PIPE_FORMAT_B10G10R10X2_UNORM:
 106                 case PIPE_FORMAT_R10G10B10A2_UINT:
 107                         return PAN_FORMAT_PACK;
 108                 default:
 109                         return PAN_FORMAT_NATIVE;
 110                 }
 111         }
 112
 113         /* Otherwise, we can do native */
 114         return PAN_FORMAT_NATIVE;
 115 }
 116
 117 enum pan_format_class
 118 pan_format_class_store(const struct util_format_description *desc, unsigned quirks)
 119 {
 120         /* Check if we can do anything better than software architecturally */
 121         if (quirks & MIDGARD_NO_TYPED_BLEND_STORES) {
 122                 return (quirks & NO_BLEND_PACKS)
 123                         ? PAN_FORMAT_SOFTWARE : PAN_FORMAT_PACK;
 124         }
 125
 126         return PAN_FORMAT_NATIVE;
 127 }
 128
 129 /* Software packs/unpacks, by format class. Packs take in the pixel value typed
 130  * as `pan_unpacked_type_for_format` of the format and return an i32vec4
 131  * suitable for storing (with components replicated to fill). Unpacks do the
 132  * reverse but cannot rely on replication.
 133  *
 134  * Pure 32 formats (R32F ... RGBA32F) are 32 unpacked, so just need to
 135  * replicate to fill */
 136
 137 static nir_ssa_def *
 138 pan_pack_pure_32(nir_builder *b, nir_ssa_def *v)
 139 {
 140         nir_ssa_def *replicated[4];
 141
 142         for (unsigned i = 0; i < 4; ++i)
 143                 replicated[i] = nir_channel(b, v, i % v->num_components);
 144
 145         return nir_vec(b, replicated, 4);
 146 }
 147
 148 static nir_ssa_def *
 149 pan_unpack_pure_32(nir_builder *b, nir_ssa_def *pack, unsigned num_components)
 150 {
 151         return nir_channels(b, pack, (1 << num_components) - 1);
 152 }
 153
 154 /* Pure x16 formats are x16 unpacked, so it's similar, but we need to pack
 155  * upper/lower halves of course */
 156
 157 static nir_ssa_def *
 158 pan_pack_pure_16(nir_builder *b, nir_ssa_def *v)
 159 {
 160         nir_ssa_def *replicated[4];
 161
 162         for (unsigned i = 0; i < 4; ++i) {
 163                 unsigned c = 2 * i;
 164
 165                 nir_ssa_def *parts[2] = {
 166                         nir_channel(b, v, (c + 0) % v->num_components),
 167                         nir_channel(b, v, (c + 1) % v->num_components)
 168                 };
 169
 170                 replicated[i] = nir_pack_32_2x16(b, nir_vec(b, parts, 2));
 171         }
 172
 173         return nir_vec(b, replicated, 4);
 174 }
 175
 176 static nir_ssa_def *
 177 pan_unpack_pure_16(nir_builder *b, nir_ssa_def *pack, unsigned num_components)
 178 {
 179         nir_ssa_def *unpacked[4];
 180
 181         assert(num_components <= 4);
 182
 183         for (unsigned i = 0; i < num_components; i += 2) {
 184                 nir_ssa_def *halves =
 185                         nir_unpack_32_2x16(b, nir_channel(b, pack, i >> 1));
 186
 187                 unpacked[i + 0] = nir_channel(b, halves, 0);
 188                 unpacked[i + 1] = nir_channel(b, halves, 1);
 189         }
 190
 191         for (unsigned i = num_components; i < 4; ++i)
 192                 unpacked[i] = nir_imm_intN_t(b, 0, 16);
 193
 194         return nir_vec(b, unpacked, 4);
 195 }
 196
 197 /* Generic dispatches for un/pack regardless of format */
 198
 199 static nir_ssa_def *
 200 pan_unpack(nir_builder *b,
 201                 const struct util_format_description *desc,
 202                 nir_ssa_def *packed)
 203 {
 204         /* Stub */
 205         return packed;
 206 }
 207
 208 static nir_ssa_def *
 209 pan_pack(nir_builder *b,
 210                 const struct util_format_description *desc,
 211                 nir_ssa_def *unpacked)
 212 {
 213         /* Stub */
 214         return unpacked;
 215 }
 216
 217 static void
 218 pan_lower_fb_store(nir_shader *shader,
 219                 nir_builder *b,
 220                 nir_intrinsic_instr *intr,
 221                 const struct util_format_description *desc,
 222                 unsigned quirks)
 223 {
 224         /* For stores, add conversion before */
 225         nir_ssa_def *unpacked = nir_ssa_for_src(b, intr->src[1], 4);
 226         nir_ssa_def *packed = pan_pack(b, desc, unpacked);
 227
 228         nir_intrinsic_instr *new =
 229                 nir_intrinsic_instr_create(shader, nir_intrinsic_store_raw_output_pan);
 230         new->src[0] = nir_src_for_ssa(packed);
 231         new->num_components = 4;
 232         nir_builder_instr_insert(b, &new->instr);
 233 }
 234
 235 static void
 236 pan_lower_fb_load(nir_shader *shader,
 237                 nir_builder *b,
 238                 nir_intrinsic_instr *intr,
 239                 const struct util_format_description *desc,
 240                 unsigned quirks)
 241 {
 242         nir_intrinsic_instr *new = nir_intrinsic_instr_create(shader,
 243                        nir_intrinsic_load_raw_output_pan);
 244         new->num_components = 4;
 245
 246         nir_ssa_dest_init(&new->instr, &new->dest, 4, 32, NULL);
 247         nir_builder_instr_insert(b, &new->instr);
 248
 249         /* Convert the raw value */
 250         nir_ssa_def *packed = &new->dest.ssa;
 251         nir_ssa_def *unpacked = pan_unpack(b, desc, packed);
 252
 253         nir_src rewritten = nir_src_for_ssa(unpacked);
 254         nir_ssa_def_rewrite_uses_after(&intr->dest.ssa, rewritten, &intr->instr);
 255 }
 256
 257 void
 258 pan_lower_framebuffer(nir_shader *shader,
 259                 const struct util_format_description *desc,
 260                 unsigned quirks)
 261 {
 262         /* Blend shaders are represented as special fragment shaders */
 263         assert(shader->info.stage == MESA_SHADER_FRAGMENT);
 264
 265         nir_foreach_function(func, shader) {
 266                 nir_foreach_block(block, func->impl) {
 267                         nir_foreach_instr_safe(instr, block) {
 268                                 if (instr->type != nir_instr_type_intrinsic)
 269                                         continue;
 270
 271                                 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
 272
 273                                 bool is_load = intr->intrinsic == nir_intrinsic_load_deref;
 274                                 bool is_store = intr->intrinsic == nir_intrinsic_store_deref;
 275
 276                                 if (!(is_load || is_store))
 277                                         continue;
 278
 279                                 /* Don't worry about MRT */
 280                                 nir_variable *var = nir_intrinsic_get_var(intr, 0);
 281
 282                                 if (var->data.location != FRAG_RESULT_COLOR)
 283                                         continue;
 284
 285                                 nir_builder b;
 286                                 nir_builder_init(&b, func->impl);
 287
 288                                 if (is_store) {
 289                                         b.cursor = nir_before_instr(instr);
 290                                         pan_lower_fb_store(shader, &b, intr, desc, quirks);
 291                                 } else {
 292                                         b.cursor = nir_after_instr(instr);
 293                                         pan_lower_fb_load(shader, &b, intr, desc, quirks);
 294                                 }
 295
 296                                 nir_instr_remove(instr);
 297                         }
 298                 }
 299
 300                 nir_metadata_preserve(func->impl, nir_metadata_block_index |
 301                                 nir_metadata_dominance);
 302         }
 303 }