src/gallium/drivers/panfrost/midgard/nir_lower_framebuffer.c

   1 /*
   2  * Copyright (C) 2019 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  * @file
  29  *
  30  * Implements framebuffer format conversions in software, specifically for
  31  * blend shaders on Midgard/Bifrost. load_output/store_output (derefs more
  32  * correctly -- pre I/O lowering) normally for the fragment stage within the
  33  * blend shader will operate with purely vec4 float ("nir") encodings. This
  34  * lowering stage, to be run before I/O is lowered, converts the native
  35  * framebuffer format to a NIR encoding after loads and vice versa before
  36  * stores. This pass is designed for a single render target; Midgard duplicates
  37  * blend shaders for MRT to simplify everything.
  38  */
  39
  40 #include "compiler/nir/nir.h"
  41 #include "compiler/nir/nir_builder.h"
  42 #include "nir_lower_blend.h"
  43 #include "util/u_format.h"
  44
  45 static nir_ssa_def *
  46 nir_float_to_unorm8(nir_builder *b, nir_ssa_def *c_float)
  47 {
  48    /* First, we degrade quality to fp16; we don't need the extra bits */
  49    nir_ssa_def *degraded = nir_f2f16(b, c_float);
  50
  51    /* Scale from [0, 1] to [0, 255.0] */
  52    nir_ssa_def *scaled = nir_fmul_imm(b, nir_fsat(b, degraded), 255.0);
  53
  54    /* Next, we type convert */
  55    nir_ssa_def *converted = nir_u2u8(b, nir_f2u16(b,
  56             nir_fround_even(b, scaled)));
  57
  58    return converted;
  59 }
  60
  61 static nir_ssa_def *
  62 nir_unorm8_to_float(nir_builder *b, nir_ssa_def *c_native)
  63 {
  64    /* First, we convert up from u8 to f16 */
  65    nir_ssa_def *converted = nir_u2f16(b, nir_u2u16(b, c_native));
  66
  67    /* Next, we scale down from [0, 255.0] to [0, 1] */
  68    nir_ssa_def *scaled = nir_fsat(b, nir_fmul_imm(b, converted, 1.0/255.0));
  69
  70    return scaled;
  71 }
  72
  73
  74
  75 static nir_ssa_def *
  76 nir_float_to_native(nir_builder *b,
  77       nir_ssa_def *c_float,
  78       const struct util_format_description *desc)
  79 {
  80    if (util_format_is_unorm8(desc))
  81       return nir_float_to_unorm8(b, c_float);
  82    else {
  83       printf("%s\n", desc->name);
  84       unreachable("Unknown format name");
  85    }
  86 }
  87
  88 static nir_ssa_def *
  89 nir_native_to_float(nir_builder *b,
  90    nir_ssa_def *c_native,
  91    const struct util_format_description *desc)
  92 {
  93    if (util_format_is_unorm8(desc))
  94       return nir_unorm8_to_float(b, c_native);
  95    else {
  96       printf("%s\n", desc->name);
  97       unreachable("Unknown format name");
  98    }
  99 }
 100
 101 void
 102 nir_lower_framebuffer(nir_shader *shader)
 103 {
 104    /* Blend shaders are represented as special fragment shaders */
 105    assert(shader->info.stage == MESA_SHADER_FRAGMENT);
 106
 107    enum pipe_format format = PIPE_FORMAT_R8G8B8A8_UNORM;
 108    const struct util_format_description *format_desc =
 109       util_format_description(format);
 110
 111    nir_foreach_function(func, shader) {
 112       nir_foreach_block(block, func->impl) {
 113          nir_foreach_instr_safe(instr, block) {
 114             if (instr->type != nir_instr_type_intrinsic)
 115                continue;
 116
 117             nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
 118
 119             bool is_load = intr->intrinsic == nir_intrinsic_load_deref;
 120             bool is_store = intr->intrinsic == nir_intrinsic_store_deref;
 121
 122             if (!(is_load || is_store))
 123                continue;
 124
 125             /* Don't worry about MRT */
 126             nir_variable *var = nir_intrinsic_get_var(intr, 0);
 127
 128             if (var->data.location != FRAG_RESULT_COLOR)
 129                continue;
 130
 131             nir_builder b;
 132             nir_builder_init(&b, func->impl);
 133
 134             if (is_store) {
 135                /* For stores, add conversion before */
 136                b.cursor = nir_before_instr(instr);
 137
 138                /* Grab the input color */
 139                nir_ssa_def *c_nir = nir_ssa_for_src(&b, intr->src[1], 4);
 140
 141                /* Format convert */
 142                nir_ssa_def *converted = nir_float_to_native(&b, c_nir, format_desc);
 143
 144                /* Rewrite to use a native store by creating a new intrinsic */
 145                nir_intrinsic_instr *new =
 146                   nir_intrinsic_instr_create(shader, nir_intrinsic_store_raw_output_pan);
 147                new->src[0] = nir_src_for_ssa(converted);
 148
 149                /* TODO: What about non-RGBA? Is that different? */
 150                new->num_components = 4;
 151
 152                nir_builder_instr_insert(&b, &new->instr);
 153
 154                /* (And finally removing the old) */
 155                nir_instr_remove(instr);
 156             } else {
 157                /* For loads, add conversion after */
 158                b.cursor = nir_after_instr(instr);
 159
 160                /* Rewrite to use a native load by creating a new intrinsic */
 161
 162                nir_intrinsic_instr *new =
 163                   nir_intrinsic_instr_create(shader, nir_intrinsic_load_raw_output_pan);
 164
 165                new->num_components = 4;
 166
 167                unsigned bitsize = 8;
 168                nir_ssa_dest_init(&new->instr, &new->dest, 4, bitsize, NULL);
 169                nir_builder_instr_insert(&b, &new->instr);
 170
 171                /* Convert the raw value */
 172                nir_ssa_def *raw = &new->dest.ssa;
 173                nir_ssa_def *converted = nir_native_to_float(&b, raw, format_desc);
 174
 175                /* Rewrite to use the converted value */
 176                nir_src rewritten = nir_src_for_ssa(converted);
 177                nir_ssa_def_rewrite_uses_after(&intr->dest.ssa, rewritten, instr);
 178
 179                /* Finally, remove the old load */
 180                nir_instr_remove(instr);
 181             }
 182          }
 183       }
 184
 185       nir_metadata_preserve(func->impl, nir_metadata_block_index |
 186                             nir_metadata_dominance);
 187    }
 188 }