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 "midgard/nir_lower_blend.h"
  31 #include "gallium/auxiliary/util/u_blend.h"
  32
  33 /*
  34  * Implements the command stream portion of programmatic blend shaders.
  35  *
  36  * On Midgard, common blending operations are accelerated by the fixed-function
  37  * blending pipeline. Panfrost supports this fast path via the code in
  38  * pan_blending.c. Nevertheless, uncommon blend modes (including some seemingly
  39  * simple modes present in ES2) require "blend shaders", a special internal
  40  * shader type used for programmable blending.
  41  *
  42  * Blend shaders operate during the normal blending time, but they bypass the
  43  * fixed-function blending pipeline and instead go straight to the Midgard
  44  * shader cores. The shaders themselves are essentially just fragment shaders,
  45  * making heavy use of uint8 arithmetic to manipulate RGB values for the
  46  * framebuffer.
  47  *
  48  * As is typical with Midgard, shader binaries must be accompanied by
  49  * information about the first tag (ORed with the bottom nibble of address,
  50  * like usual) and work registers. Work register count is specified in the
  51  * blend descriptor, as well as in the coresponding fragment shader's work
  52  * count. This suggests that blend shader invocation is tied to fragment shader
  53  * execution.
  54  *
  55  * ---
  56  *
  57  * As for blend shaders, they use the standard ISA.
  58  *
  59  * The source pixel colour, including alpha, is preloaded into r0 as a vec4 of
  60  * float32.
  61  *
  62  * The destination pixel colour must be loaded explicitly via load/store ops.
  63  * TODO: Investigate.
  64  *
  65  * They use fragment shader writeout; however, instead of writing a vec4 of
  66  * float32 for RGBA encoding, we writeout a vec4 of uint8, using 8-bit imov
  67  * instead of 32-bit fmov. The net result is that r0 encodes a single uint32
  68  * containing all four channels of the color.  Accordingly, the blend shader
  69  * epilogue has to scale all four channels by 255 and then type convert to a
  70  * uint8.
  71  *
  72  * ---
  73  *
  74  * Blend shaders hardcode constants. Naively, this requires recompilation each
  75  * time the blend color changes, which is a performance risk. Accordingly, we
  76  * 'cheat' a bit: instead of loading the constant, we compile a shader with a
  77  * dummy constant, exporting the offset to the immediate in the shader binary,
  78  * storing this generic binary and metadata in the CSO itself at CSO create
  79  * time.
  80  *
  81  * We then hot patch in the color into this shader at attachment / color change
  82  * time, allowing for CSO create to be the only expensive operation
  83  * (compilation).
  84  */
  85
  86 static nir_lower_blend_options
  87 nir_make_options(const struct pipe_blend_state *blend, unsigned nr_cbufs)
  88 {
  89         nir_lower_blend_options options;
  90
  91         for (unsigned i = 0; i < nr_cbufs; ++i) {
  92                 nir_lower_blend_channel rgb = {
  93                         .func = util_blend_func_to_shader(blend->rt[i].rgb_func),
  94                         .src_factor = util_blend_factor_to_shader(blend->rt[i].rgb_src_factor),
  95                         .dst_factor = util_blend_factor_to_shader(blend->rt[i].rgb_dst_factor),
  96                         .invert_src_factor = util_blend_factor_is_inverted(blend->rt[i].rgb_src_factor),
  97                         .invert_dst_factor = util_blend_factor_is_inverted(blend->rt[i].rgb_dst_factor)
  98                 };
  99
 100                 nir_lower_blend_channel alpha = {
 101                         .func = util_blend_func_to_shader(blend->rt[i].alpha_func),
 102                         .src_factor = util_blend_factor_to_shader(blend->rt[i].alpha_src_factor),
 103                         .dst_factor = util_blend_factor_to_shader(blend->rt[i].alpha_dst_factor),
 104                         .invert_src_factor = util_blend_factor_is_inverted(blend->rt[i].alpha_src_factor),
 105                         .invert_dst_factor = util_blend_factor_is_inverted(blend->rt[i].alpha_dst_factor)
 106                 };
 107
 108                 options.rt[i].rgb = rgb;
 109                 options.rt[i].alpha = alpha;
 110
 111                 options.rt[i].colormask = blend->rt[i].colormask;
 112         }
 113
 114         return options;
 115 }
 116
 117 void
 118 panfrost_make_blend_shader(struct panfrost_context *ctx, struct panfrost_blend_state *cso, const struct pipe_blend_color *blend_color)
 119 {
 120         /* Build the shader */
 121
 122         nir_shader *shader = nir_shader_create(NULL, MESA_SHADER_FRAGMENT, &midgard_nir_options, NULL);
 123         nir_function *fn = nir_function_create(shader, "main");
 124         nir_function_impl *impl = nir_function_impl_create(fn);
 125
 126         /* Create the blend variables */
 127
 128         nir_variable *c_src = nir_variable_create(shader, nir_var_shader_in, glsl_vector_type(GLSL_TYPE_FLOAT, 4), "gl_Color");
 129         nir_variable *c_out = nir_variable_create(shader, nir_var_shader_out, glsl_vector_type(GLSL_TYPE_FLOAT, 4), "gl_FragColor");
 130
 131         c_src->data.location = VARYING_SLOT_COL0;
 132         c_out->data.location = FRAG_RESULT_COLOR;
 133
 134         /* Setup nir_builder */
 135
 136         nir_builder _b;
 137         nir_builder *b = &_b;
 138         nir_builder_init(b, impl);
 139         b->cursor = nir_before_block(nir_start_block(impl));
 140
 141         /* Setup inputs */
 142
 143         nir_ssa_def *s_src = nir_load_var(b, c_src);
 144
 145         /* Build a trivial blend shader */
 146         nir_store_var(b, c_out, s_src, 0xFF);
 147
 148         nir_lower_blend_options options =
 149                 nir_make_options(&cso->base, 1);
 150         NIR_PASS_V(shader, nir_lower_blend, options);
 151
 152         /* Compile the built shader */
 153
 154         midgard_program program;
 155         midgard_compile_shader_nir(shader, &program, true);
 156
 157         /* Upload the shader */
 158
 159         int size = program.compiled.size;
 160         uint8_t *dst = program.compiled.data;
 161
 162         /* Hot patch in constant color */
 163
 164         if (program.blend_patch_offset >= 0) {
 165                 float *hot_color = (float *) (dst + program.blend_patch_offset);
 166
 167                 for (int c = 0; c < 4; ++c)
 168                         hot_color[c] = blend_color->color[c];
 169         }
 170
 171         cso->blend_shader = panfrost_upload(&ctx->shaders, dst, size, true) | program.first_tag;
 172
 173         /* We need to switch to shader mode */
 174         cso->has_blend_shader = true;
 175
 176         /* At least two work registers are needed due to an encoding quirk */
 177         cso->blend_work_count = MAX2(program.work_register_count, 2);
 178 }