src/gallium/drivers/panfrost/pan_blend_shaders.c

   1 /*
   2  * © Copyright 2018 Alyssa Rosenzweig
   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 FROM,
  20  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  21  * SOFTWARE.
  22  *
  23  */
  24
  25 #include <stdio.h>
  26 #include "pan_blend_shaders.h"
  27 #include "pan_util.h"
  28 #include "midgard/midgard_compile.h"
  29 #include "compiler/nir/nir_builder.h"
  30 #include "nir/nir_lower_blend.h"
  31 #include "panfrost/util/pan_lower_framebuffer.h"
  32 #include "gallium/auxiliary/util/u_blend.h"
  33 #include "util/u_memory.h"
  34
  35 /*
  36  * Implements the command stream portion of programmatic blend shaders.
  37  *
  38  * On Midgard, common blending operations are accelerated by the fixed-function
  39  * blending pipeline. Panfrost supports this fast path via the code in
  40  * pan_blending.c. Nevertheless, uncommon blend modes (including some seemingly
  41  * simple modes present in ES2) require "blend shaders", a special internal
  42  * shader type used for programmable blending.
  43  *
  44  * Blend shaders operate during the normal blending time, but they bypass the
  45  * fixed-function blending pipeline and instead go straight to the Midgard
  46  * shader cores. The shaders themselves are essentially just fragment shaders,
  47  * making heavy use of uint8 arithmetic to manipulate RGB values for the
  48  * framebuffer.
  49  *
  50  * As is typical with Midgard, shader binaries must be accompanied by
  51  * information about the first tag (ORed with the bottom nibble of address,
  52  * like usual) and work registers. Work register count is assumed to be less
  53  * than or equal to the coresponding fragment shader's work count. This
  54  * suggests that blend shader invocation is tied to fragment shader
  55  * execution.
  56  *
  57  * The shaders themselves use the standard ISA. The source pixel colour,
  58  * including alpha, is preloaded into r0 as a vec4 of float32. The destination
  59  * pixel colour must be loaded explicitly via load/store ops, possibly
  60  * performing conversions in software. The blended colour must be stored with a
  61  * fragment writeout in the correct framebuffer format, either in software or
  62  * via conversion opcodes on the load/store pipe.
  63  *
  64  * Blend shaders hardcode constants. Naively, this requires recompilation each
  65  * time the blend color changes, which is a performance risk. Accordingly, we
  66  * 'cheat' a bit: instead of loading the constant, we compile a shader with a
  67  * dummy constant, exporting the offset to the immediate in the shader binary,
  68  * storing this generic binary and metadata in the CSO itself at CSO create
  69  * time.
  70  *
  71  * We then hot patch in the color into this shader at attachment / color change
  72  * time, allowing for CSO create to be the only expensive operation
  73  * (compilation).
  74  */
  75
  76 static nir_lower_blend_options
  77 nir_make_options(const struct pipe_blend_state *blend, unsigned i)
  78 {
  79         nir_lower_blend_options options = { 0 };
  80
  81         if (blend->logicop_enable) {
  82             options.logicop_enable = true;
  83             options.logicop_func = blend->logicop_func;
  84             return options;
  85         }
  86
  87         options.logicop_enable = false;
  88
  89         if (!blend->independent_blend_enable)
  90                 i = 0;
  91
  92         /* If blend is disabled, we just use replace mode */
  93
  94         nir_lower_blend_channel rgb = {
  95                 .func = BLEND_FUNC_ADD,
  96                 .src_factor = BLEND_FACTOR_ZERO,
  97                 .invert_src_factor = true,
  98                 .dst_factor = BLEND_FACTOR_ZERO,
  99                 .invert_dst_factor = false
 100         };
 101
 102         nir_lower_blend_channel alpha = rgb;
 103
 104         if (blend->rt[i].blend_enable) {
 105                 rgb.func = util_blend_func_to_shader(blend->rt[i].rgb_func);
 106                 rgb.src_factor = util_blend_factor_to_shader(blend->rt[i].rgb_src_factor);
 107                 rgb.dst_factor = util_blend_factor_to_shader(blend->rt[i].rgb_dst_factor);
 108                 rgb.invert_src_factor = util_blend_factor_is_inverted(blend->rt[i].rgb_src_factor);
 109                 rgb.invert_dst_factor = util_blend_factor_is_inverted(blend->rt[i].rgb_dst_factor);
 110
 111                 alpha.func = util_blend_func_to_shader(blend->rt[i].alpha_func);
 112                 alpha.src_factor = util_blend_factor_to_shader(blend->rt[i].alpha_src_factor);
 113                 alpha.dst_factor = util_blend_factor_to_shader(blend->rt[i].alpha_dst_factor);
 114                 alpha.invert_src_factor = util_blend_factor_is_inverted(blend->rt[i].alpha_src_factor);
 115                 alpha.invert_dst_factor = util_blend_factor_is_inverted(blend->rt[i].alpha_dst_factor);
 116         }
 117
 118         options.rgb = rgb;
 119         options.alpha = alpha;
 120
 121         options.colormask = blend->rt[i].colormask;
 122
 123         return options;
 124 }
 125
 126 struct panfrost_blend_shader
 127 panfrost_compile_blend_shader(
 128         struct panfrost_context *ctx,
 129         struct pipe_blend_state *cso,
 130         enum pipe_format format,
 131         unsigned rt)
 132 {
 133         struct panfrost_device *dev = pan_device(ctx->base.screen);
 134         struct panfrost_blend_shader res;
 135
 136         res.ctx = ctx;
 137
 138         /* Build the shader */
 139
 140         nir_shader *shader = nir_shader_create(NULL, MESA_SHADER_FRAGMENT, &midgard_nir_options, NULL);
 141         nir_function *fn = nir_function_create(shader, "main");
 142         nir_function_impl *impl = nir_function_impl_create(fn);
 143
 144         const struct util_format_description *format_desc =
 145                 util_format_description(format);
 146
 147         nir_alu_type T = pan_unpacked_type_for_format(format_desc);
 148         enum glsl_base_type g =
 149                 (T == nir_type_float16) ? GLSL_TYPE_FLOAT16 :
 150                 (T == nir_type_float32) ? GLSL_TYPE_FLOAT :
 151                 (T == nir_type_int8) ? GLSL_TYPE_INT8 :
 152                 (T == nir_type_int16) ? GLSL_TYPE_INT16 :
 153                 (T == nir_type_int32) ? GLSL_TYPE_INT :
 154                 (T == nir_type_uint8) ? GLSL_TYPE_UINT8 :
 155                 (T == nir_type_uint16) ? GLSL_TYPE_UINT16 :
 156                 (T == nir_type_uint32) ? GLSL_TYPE_UINT :
 157                 GLSL_TYPE_FLOAT;
 158
 159         /* Create the blend variables */
 160
 161         nir_variable *c_src = nir_variable_create(shader, nir_var_shader_in, glsl_vector_type(GLSL_TYPE_FLOAT, 4), "gl_Color");
 162         nir_variable *c_out = nir_variable_create(shader, nir_var_shader_out, glsl_vector_type(g, 4), "gl_FragColor");
 163
 164         c_src->data.location = VARYING_SLOT_COL0;
 165         c_out->data.location = FRAG_RESULT_COLOR;
 166
 167         /* Setup nir_builder */
 168
 169         nir_builder _b;
 170         nir_builder *b = &_b;
 171         nir_builder_init(b, impl);
 172         b->cursor = nir_before_block(nir_start_block(impl));
 173
 174         /* Setup inputs */
 175
 176         nir_ssa_def *s_src = nir_load_var(b, c_src);
 177
 178         if (T == nir_type_float16)
 179                 s_src = nir_f2f16(b, s_src);
 180         else if (T == nir_type_int16)
 181                 s_src = nir_i2i16(b, s_src);
 182         else if (T == nir_type_uint16)
 183                 s_src = nir_u2u16(b, s_src);
 184         else if (T == nir_type_int8)
 185                 s_src = nir_i2i8(b, s_src);
 186         else if (T == nir_type_uint8)
 187                 s_src = nir_u2u8(b, s_src);
 188
 189         /* Build a trivial blend shader */
 190         nir_store_var(b, c_out, s_src, 0xFF);
 191
 192         nir_lower_blend_options options =
 193                 nir_make_options(cso, rt);
 194         options.format = format;
 195
 196         if (T == nir_type_float16)
 197                 options.half = true;
 198
 199         NIR_PASS_V(shader, nir_lower_blend, options);
 200         NIR_PASS_V(shader, pan_lower_framebuffer, format_desc, dev->quirks);
 201
 202         /* Compile the built shader */
 203
 204         panfrost_program program;
 205         midgard_compile_shader_nir(shader, &program, true, rt, dev->gpu_id, false);
 206
 207         /* Allow us to patch later */
 208         res.patch_index = program.blend_patch_offset;
 209         res.first_tag = program.first_tag;
 210         res.size = program.compiled.size;
 211         res.buffer = program.compiled.data;
 212
 213         return res;
 214 }