summaryrefslogtreecommitdiff
path: root/src/gallium/drivers/panfrost/pan_blend_shaders.c
blob: 33a1f61c941539ce25bba2f45ad2ed8851d1027f (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
/*
 * © Copyright 2018 Alyssa Rosenzweig
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice (including the next
 * paragraph) shall be included in all copies or substantial portions of the
 * Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 *
 */

#include <stdio.h>
#include "pan_blend_shaders.h"
#include "pan_util.h"
#include "midgard/midgard_compile.h"
#include "compiler/nir/nir_builder.h"
#include "nir/nir_lower_blend.h"
#include "gallium/auxiliary/util/u_blend.h"
#include "util/u_memory.h"

/*
 * Implements the command stream portion of programmatic blend shaders.
 *
 * On Midgard, common blending operations are accelerated by the fixed-function
 * blending pipeline. Panfrost supports this fast path via the code in
 * pan_blending.c. Nevertheless, uncommon blend modes (including some seemingly
 * simple modes present in ES2) require "blend shaders", a special internal
 * shader type used for programmable blending.
 *
 * Blend shaders operate during the normal blending time, but they bypass the
 * fixed-function blending pipeline and instead go straight to the Midgard
 * shader cores. The shaders themselves are essentially just fragment shaders,
 * making heavy use of uint8 arithmetic to manipulate RGB values for the
 * framebuffer.
 *
 * As is typical with Midgard, shader binaries must be accompanied by
 * information about the first tag (ORed with the bottom nibble of address,
 * like usual) and work registers. Work register count is specified in the
 * blend descriptor, as well as in the coresponding fragment shader's work
 * count. This suggests that blend shader invocation is tied to fragment shader
 * execution.
 *
 * ---
 *
 * As for blend shaders, they use the standard ISA.
 *
 * The source pixel colour, including alpha, is preloaded into r0 as a vec4 of
 * float32.
 *
 * The destination pixel colour must be loaded explicitly via load/store ops.
 * TODO: Investigate.
 *
 * They use fragment shader writeout; however, instead of writing a vec4 of
 * float32 for RGBA encoding, we writeout a vec4 of uint8, using 8-bit imov
 * instead of 32-bit fmov. The net result is that r0 encodes a single uint32
 * containing all four channels of the color.  Accordingly, the blend shader
 * epilogue has to scale all four channels by 255 and then type convert to a
 * uint8.
 *
 * ---
 *
 * Blend shaders hardcode constants. Naively, this requires recompilation each
 * time the blend color changes, which is a performance risk. Accordingly, we
 * 'cheat' a bit: instead of loading the constant, we compile a shader with a
 * dummy constant, exporting the offset to the immediate in the shader binary,
 * storing this generic binary and metadata in the CSO itself at CSO create
 * time.
 *
 * We then hot patch in the color into this shader at attachment / color change
 * time, allowing for CSO create to be the only expensive operation
 * (compilation).
 */

static nir_lower_blend_options
nir_make_options(const struct pipe_blend_state *blend, unsigned nr_cbufs)
{
        nir_lower_blend_options options;

        for (unsigned i = 0; i < nr_cbufs; ++i) {
                /* If blend is disabled, we just use replace mode */

                nir_lower_blend_channel rgb = {
                        .func = BLEND_FUNC_ADD,
                        .src_factor = BLEND_FACTOR_ZERO,
                        .invert_src_factor = true,
                        .dst_factor = BLEND_FACTOR_ZERO,
                        .invert_dst_factor = false
                };

                nir_lower_blend_channel alpha = rgb;

                if (blend->rt[i].blend_enable) {
                        rgb.func = util_blend_func_to_shader(blend->rt[i].rgb_func);
                        rgb.src_factor = util_blend_factor_to_shader(blend->rt[i].rgb_src_factor);
                        rgb.dst_factor = util_blend_factor_to_shader(blend->rt[i].rgb_dst_factor);
                        rgb.invert_src_factor = util_blend_factor_is_inverted(blend->rt[i].rgb_src_factor);
                        rgb.invert_dst_factor = util_blend_factor_is_inverted(blend->rt[i].rgb_dst_factor);

                        alpha.func = util_blend_func_to_shader(blend->rt[i].alpha_func);
                        alpha.src_factor = util_blend_factor_to_shader(blend->rt[i].alpha_src_factor);
                        alpha.dst_factor = util_blend_factor_to_shader(blend->rt[i].alpha_dst_factor);
                        alpha.invert_src_factor = util_blend_factor_is_inverted(blend->rt[i].alpha_src_factor);
                        alpha.invert_dst_factor = util_blend_factor_is_inverted(blend->rt[i].alpha_dst_factor);
                }

                options.rt[i].rgb = rgb;
                options.rt[i].alpha = alpha;

                options.rt[i].colormask = blend->rt[i].colormask;
        }

        return options;
}

struct panfrost_blend_shader
panfrost_compile_blend_shader(
        struct panfrost_context *ctx,
        struct pipe_blend_state *cso,
        enum pipe_format format)
{
        struct panfrost_screen *screen = pan_screen(ctx->base.screen);
        struct panfrost_blend_shader res;

        res.ctx = ctx;

        /* Build the shader */

        nir_shader *shader = nir_shader_create(NULL, MESA_SHADER_FRAGMENT, &midgard_nir_options, NULL);
        nir_function *fn = nir_function_create(shader, "main");
        nir_function_impl *impl = nir_function_impl_create(fn);

        /* Create the blend variables */

        nir_variable *c_src = nir_variable_create(shader, nir_var_shader_in, glsl_vector_type(GLSL_TYPE_FLOAT, 4), "gl_Color");
        nir_variable *c_out = nir_variable_create(shader, nir_var_shader_out, glsl_vector_type(GLSL_TYPE_FLOAT, 4), "gl_FragColor");

        c_src->data.location = VARYING_SLOT_COL0;
        c_out->data.location = FRAG_RESULT_COLOR;

        /* Setup nir_builder */

        nir_builder _b;
        nir_builder *b = &_b;
        nir_builder_init(b, impl);
        b->cursor = nir_before_block(nir_start_block(impl));

        /* Setup inputs */

        nir_ssa_def *s_src = nir_load_var(b, c_src);

        /* Build a trivial blend shader */
        nir_store_var(b, c_out, s_src, 0xFF);

        nir_lower_blend_options options =
                nir_make_options(cso, 1);
        NIR_PASS_V(shader, nir_lower_blend, options);

        NIR_PASS_V(shader, nir_lower_framebuffer, format, screen->gpu_id);

        /* Compile the built shader */

        midgard_program program;
        midgard_compile_shader_nir(shader, &program, true, screen->gpu_id);

        /* At least two work registers are needed due to an encoding quirk */
        res.work_count = MAX2(program.work_register_count, 2);

        /* Allow us to patch later */
        res.patch_index = program.blend_patch_offset;
        res.first_tag = program.first_tag;
        res.size = program.compiled.size;
        res.buffer = program.compiled.data;

        return res;
}