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 "gallium/auxiliary/util/u_blend.h"
  32 #include "util/u_memory.h"
  33
  34 /*
  35  * Implements the command stream portion of programmatic blend shaders.
  36  *
  37  * On Midgard, common blending operations are accelerated by the fixed-function
  38  * blending pipeline. Panfrost supports this fast path via the code in
  39  * pan_blending.c. Nevertheless, uncommon blend modes (including some seemingly
  40  * simple modes present in ES2) require "blend shaders", a special internal
  41  * shader type used for programmable blending.
  42  *
  43  * Blend shaders operate during the normal blending time, but they bypass the
  44  * fixed-function blending pipeline and instead go straight to the Midgard
  45  * shader cores. The shaders themselves are essentially just fragment shaders,
  46  * making heavy use of uint8 arithmetic to manipulate RGB values for the
  47  * framebuffer.
  48  *
  49  * As is typical with Midgard, shader binaries must be accompanied by
  50  * information about the first tag (ORed with the bottom nibble of address,
  51  * like usual) and work registers. Work register count is specified in the
  52  * blend descriptor, as well as in the coresponding fragment shader's work
  53  * count. This suggests that blend shader invocation is tied to fragment shader
  54  * execution.
  55  *
  56  * ---
  57  *
  58  * As for blend shaders, they use the standard ISA.
  59  *
  60  * The source pixel colour, including alpha, is preloaded into r0 as a vec4 of
  61  * float32.
  62  *
  63  * The destination pixel colour must be loaded explicitly via load/store ops.
  64  * TODO: Investigate.
  65  *
  66  * They use fragment shader writeout; however, instead of writing a vec4 of
  67  * float32 for RGBA encoding, we writeout a vec4 of uint8, using 8-bit imov
  68  * instead of 32-bit fmov. The net result is that r0 encodes a single uint32
  69  * containing all four channels of the color.  Accordingly, the blend shader
  70  * epilogue has to scale all four channels by 255 and then type convert to a
  71  * uint8.
  72  *
  73  * ---
  74  *
  75  * Blend shaders hardcode constants. Naively, this requires recompilation each
  76  * time the blend color changes, which is a performance risk. Accordingly, we
  77  * 'cheat' a bit: instead of loading the constant, we compile a shader with a
  78  * dummy constant, exporting the offset to the immediate in the shader binary,
  79  * storing this generic binary and metadata in the CSO itself at CSO create
  80  * time.
  81  *
  82  * We then hot patch in the color into this shader at attachment / color change
  83  * time, allowing for CSO create to be the only expensive operation
  84  * (compilation).
  85  */
  86
  87 static nir_lower_blend_options
  88 nir_make_options(const struct pipe_blend_state *blend, unsigned nr_cbufs)
  89 {
  90         nir_lower_blend_options options;
  91
  92         for (unsigned i = 0; i < nr_cbufs; ++i) {
  93                 /* If blend is disabled, we just use replace mode */
  94
  95                 nir_lower_blend_channel rgb = {
  96                         .func = BLEND_FUNC_ADD,
  97                         .src_factor = BLEND_FACTOR_ZERO,
  98                         .invert_src_factor = true,
  99                         .dst_factor = BLEND_FACTOR_ZERO,
 100                         .invert_dst_factor = false
 101                 };
 102
 103                 nir_lower_blend_channel alpha = rgb;
 104
 105                 if (blend->rt[i].blend_enable) {
 106                         rgb.func = util_blend_func_to_shader(blend->rt[i].rgb_func);
 107                         rgb.src_factor = util_blend_factor_to_shader(blend->rt[i].rgb_src_factor);
 108                         rgb.dst_factor = util_blend_factor_to_shader(blend->rt[i].rgb_dst_factor);
 109                         rgb.invert_src_factor = util_blend_factor_is_inverted(blend->rt[i].rgb_src_factor);
 110                         rgb.invert_dst_factor = util_blend_factor_is_inverted(blend->rt[i].rgb_dst_factor);
 111
 112                         alpha.func = util_blend_func_to_shader(blend->rt[i].alpha_func);
 113                         alpha.src_factor = util_blend_factor_to_shader(blend->rt[i].alpha_src_factor);
 114                         alpha.dst_factor = util_blend_factor_to_shader(blend->rt[i].alpha_dst_factor);
 115                         alpha.invert_src_factor = util_blend_factor_is_inverted(blend->rt[i].alpha_src_factor);
 116                         alpha.invert_dst_factor = util_blend_factor_is_inverted(blend->rt[i].alpha_dst_factor);
 117                 }
 118
 119                 options.rt[i].rgb = rgb;
 120                 options.rt[i].alpha = alpha;
 121
 122                 options.rt[i].colormask = blend->rt[i].colormask;
 123         }
 124
 125         return options;
 126 }
 127
 128 struct panfrost_blend_shader
 129 panfrost_compile_blend_shader(
 130         struct panfrost_context *ctx,
 131         struct pipe_blend_state *cso,
 132         enum pipe_format format)
 133 {
 134         struct panfrost_screen *screen = pan_screen(ctx->base.screen);
 135         struct panfrost_blend_shader res;
 136
 137         res.ctx = ctx;
 138
 139         /* Build the shader */
 140
 141         nir_shader *shader = nir_shader_create(NULL, MESA_SHADER_FRAGMENT, &midgard_nir_options, NULL);
 142         nir_function *fn = nir_function_create(shader, "main");
 143         nir_function_impl *impl = nir_function_impl_create(fn);
 144
 145         /* Create the blend variables */
 146
 147         nir_variable *c_src = nir_variable_create(shader, nir_var_shader_in, glsl_vector_type(GLSL_TYPE_FLOAT, 4), "gl_Color");
 148         nir_variable *c_out = nir_variable_create(shader, nir_var_shader_out, glsl_vector_type(GLSL_TYPE_FLOAT, 4), "gl_FragColor");
 149
 150         c_src->data.location = VARYING_SLOT_COL0;
 151         c_out->data.location = FRAG_RESULT_COLOR;
 152
 153         /* Setup nir_builder */
 154
 155         nir_builder _b;
 156         nir_builder *b = &_b;
 157         nir_builder_init(b, impl);
 158         b->cursor = nir_before_block(nir_start_block(impl));
 159
 160         /* Setup inputs */
 161
 162         nir_ssa_def *s_src = nir_load_var(b, c_src);
 163
 164         /* Build a trivial blend shader */
 165         nir_store_var(b, c_out, s_src, 0xFF);
 166
 167         nir_lower_blend_options options =
 168                 nir_make_options(cso, 1);
 169         NIR_PASS_V(shader, nir_lower_blend, options);
 170
 171         NIR_PASS_V(shader, nir_lower_framebuffer, format, screen->gpu_id);
 172
 173         /* Compile the built shader */
 174
 175         midgard_program program;
 176         midgard_compile_shader_nir(&ctx->compiler, shader, &program, true, screen->gpu_id);
 177
 178         /* At least two work registers are needed due to an encoding quirk */
 179         res.work_count = MAX2(program.work_register_count, 2);
 180
 181         /* Allow us to patch later */
 182         res.patch_index = program.blend_patch_offset;
 183         res.first_tag = program.first_tag;
 184         res.size = program.compiled.size;
 185         res.buffer = program.compiled.data;
 186
 187         return res;
 188 }