v3d: Add a lowering pass for line smoothing
authorNeil Roberts <nroberts@igalia.com>
Mon, 22 Jun 2020 12:26:02 +0000 (14:26 +0200)
committerMarge Bot <eric+marge@anholt.net>
Mon, 6 Jul 2020 21:59:16 +0000 (21:59 +0000)
When line smoothing is enabled, the driver now increases the width of
the line so that it can add some semi-transparent pixels to either side
of the line. A lowering pass is added which modifies the alpha component
of every write to fragment output 0 so that if the fragment is outside
the width of the line then the alpha is reduced. It additionally
discards fragments that are completely invisible. It might seem bad to
use discard on a tiled renderer but the assumption is that any bad
effects from using discard will also happen anyway because of enabling
alpha blending.

v2: Disable the line smoothing pass entirely when the framebuffer
    contains an integer colour output or one with no alpha channel.
    Calculate the coverage once upfront and store in a global variable
    instead of calculating each time an output write is modified. Also
    do the conditional discard once upfront.
v3: Don’t check whether the output buffer has an alpha channel. Only
    look at output 0. Use aa_line_width intrinsic instead of calculating
    the real line width in the shader. Clamp the coverage as part of the
    global variable, not per output write.

Reviewed-by: Eric Anholt <eric@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5624>

src/broadcom/Makefile.sources
src/broadcom/compiler/meson.build
src/broadcom/compiler/v3d_compiler.h
src/broadcom/compiler/v3d_nir_lower_line_smooth.c [new file with mode: 0644]
src/broadcom/compiler/vir.c
src/gallium/drivers/v3d/v3d_context.c
src/gallium/drivers/v3d/v3d_context.h
src/gallium/drivers/v3d/v3d_program.c
src/gallium/drivers/v3d/v3d_uniforms.c
src/gallium/drivers/v3d/v3dx_emit.c

index e759b545496d1dcf5860ed6c9cf43174d09fa17d..e3ed63d63300225555ad8225c512442cdc567583 100644 (file)
@@ -41,6 +41,7 @@ BROADCOM_FILES = \
        compiler/v3d_compiler.h \
        compiler/v3d_nir_lower_image_load_store.c \
        compiler/v3d_nir_lower_io.c \
+       compiler/v3d_nir_lower_line_smooth.c \
        compiler/v3d_nir_lower_scratch.c \
        compiler/v3d_nir_lower_txf_ms.c \
        qpu/qpu_disasm.c \
index 9094f1ac6d067ddca5174c08bc9b5cba52e49876..43a312c6bea428e7e337ed08b4978269070e0ea2 100644 (file)
@@ -37,6 +37,7 @@ libbroadcom_compiler_files = files(
   'v3d_compiler.h',
   'v3d_nir_lower_io.c',
   'v3d_nir_lower_image_load_store.c',
+  'v3d_nir_lower_line_smooth.c',
   'v3d_nir_lower_logic_ops.c',
   'v3d_nir_lower_scratch.c',
   'v3d_nir_lower_txf_ms.c',
index d6de62486d913c54c8b94517b48980b7da7e94d0..10df1af3ce4a624b65e85a502f03aed62a833f48 100644 (file)
@@ -352,6 +352,7 @@ struct v3d_fs_key {
         bool depth_enabled;
         bool is_points;
         bool is_lines;
+        bool line_smoothing;
         bool alpha_test;
         bool point_coord_upper_left;
         bool light_twoside;
@@ -867,6 +868,7 @@ bool vir_opt_small_immediates(struct v3d_compile *c);
 bool vir_opt_vpm(struct v3d_compile *c);
 void v3d_nir_lower_blend(nir_shader *s, struct v3d_compile *c);
 void v3d_nir_lower_io(nir_shader *s, struct v3d_compile *c);
+void v3d_nir_lower_line_smooth(nir_shader *shader);
 void v3d_nir_lower_logic_ops(nir_shader *s, struct v3d_compile *c);
 void v3d_nir_lower_scratch(nir_shader *s);
 void v3d_nir_lower_txf_ms(nir_shader *s, struct v3d_compile *c);
diff --git a/src/broadcom/compiler/v3d_nir_lower_line_smooth.c b/src/broadcom/compiler/v3d_nir_lower_line_smooth.c
new file mode 100644 (file)
index 0000000..ace7d5c
--- /dev/null
@@ -0,0 +1,163 @@
+/*
+ * Copyright © 2020 Raspberry Pi
+ *
+ * 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 "compiler/v3d_compiler.h"
+#include "compiler/nir/nir_builder.h"
+#include <math.h>
+
+/**
+ * Lowers line smoothing by modifying the alpha component of fragment outputs
+ * using the distance from the center of the line.
+ */
+
+struct lower_line_smooth_state {
+        nir_shader *shader;
+        nir_variable *coverage;
+};
+
+static void
+lower_line_smooth_intrinsic(struct lower_line_smooth_state *state,
+                            nir_builder *b,
+                            nir_intrinsic_instr *intr)
+{
+        b->cursor = nir_before_instr(&intr->instr);
+
+        nir_ssa_def *one = nir_imm_float(b, 1.0f);
+
+        nir_ssa_def *coverage = nir_load_var(b, state->coverage);
+
+        nir_ssa_def *new_val = nir_fmul(b, nir_vec4(b, one, one, one, coverage),
+                                        intr->src[0].ssa);
+
+        nir_instr_rewrite_src(&intr->instr,
+                              &intr->src[0],
+                              nir_src_for_ssa(new_val));
+}
+
+static void
+lower_line_smooth_func(struct lower_line_smooth_state *state,
+                       nir_function_impl *impl)
+{
+        nir_builder b;
+
+        nir_builder_init(&b, impl);
+
+        nir_foreach_block(block, impl) {
+                nir_foreach_instr_safe(instr, block) {
+                        if (instr->type != nir_instr_type_intrinsic)
+                                continue;
+
+                        nir_intrinsic_instr *intr =
+                                nir_instr_as_intrinsic(instr);
+
+                        if (intr->intrinsic != nir_intrinsic_store_output ||
+                            nir_intrinsic_base(intr) != 0 ||
+                            intr->num_components != 4 ||
+                            !intr->src[0].is_ssa)
+                                continue;
+
+                        lower_line_smooth_intrinsic(state, &b, intr);
+                }
+        }
+}
+
+static void
+initialise_coverage_var(struct lower_line_smooth_state *state,
+                        nir_function_impl *impl)
+{
+        nir_builder b;
+
+        nir_builder_init(&b, impl);
+
+        b.cursor = nir_before_block(nir_start_block(impl));
+
+        nir_ssa_def *line_width = nir_load_line_width(&b);
+
+        nir_ssa_def *real_line_width = nir_load_aa_line_width(&b);
+
+        /* The line coord varies from 0.0 to 1.0 across the width of the line */
+        nir_ssa_def *line_coord = nir_load_line_coord(&b);
+
+        /* fabs(line_coord - 0.5) * real_line_width */
+        nir_ssa_def *pixels_from_center =
+                nir_fmul(&b, real_line_width,
+                         nir_fabs(&b, nir_fsub(&b, line_coord,
+                                               nir_imm_float(&b, 0.5f))));
+
+        /* 0.5 - 1/√2 * (pixels_from_center - line_width * 0.5) */
+        nir_ssa_def *coverage =
+                nir_fsub(&b,
+                         nir_imm_float(&b, 0.5f),
+                         nir_fmul(&b,
+                                  nir_imm_float(&b, 1.0f / M_SQRT2),
+                                  nir_fsub(&b, pixels_from_center,
+                                           nir_fmul(&b,
+                                                    line_width,
+                                                    nir_imm_float(&b, 0.5f)))));
+
+        /* Discard fragments that aren’t covered at all by the line */
+        nir_ssa_def *outside = nir_fge(&b, nir_imm_float(&b, 0.0f), coverage);
+
+        nir_intrinsic_instr *discard =
+                nir_intrinsic_instr_create(state->shader,
+                                           nir_intrinsic_discard_if);
+        discard->src[0] = nir_src_for_ssa(outside);
+        nir_builder_instr_insert(&b, &discard->instr);
+
+        /* Clamp to at most 1.0. If it was less than 0.0 then the fragment will
+         * be discarded so we don’t need to handle that.
+         */
+        nir_ssa_def *clamped = nir_fmin(&b, coverage, nir_imm_float(&b, 1.0f));
+
+        nir_store_var(&b, state->coverage, clamped, 0x1 /* writemask */);
+}
+
+static nir_variable *
+make_coverage_var(nir_shader *s)
+{
+        nir_variable *var = nir_variable_create(s,
+                                                nir_var_shader_temp,
+                                                glsl_float_type(),
+                                                "line_coverage");
+        var->data.how_declared = nir_var_hidden;
+
+        return var;
+}
+
+void
+v3d_nir_lower_line_smooth(nir_shader *s)
+{
+        assert(s->info.stage == MESA_SHADER_FRAGMENT);
+
+        struct lower_line_smooth_state state = {
+                .shader = s,
+                .coverage = make_coverage_var(s),
+        };
+
+        nir_foreach_function(function, s) {
+                if (function->is_entrypoint)
+                        initialise_coverage_var(&state, function->impl);
+
+                lower_line_smooth_func(&state, function->impl);
+        }
+}
index ff2e0290122f4b56b219cded47da55b6d1a3544d..d26a3250ead0659a91e2d46111699181eade1d89 100644 (file)
@@ -890,6 +890,13 @@ v3d_nir_lower_fs_early(struct v3d_compile *c)
 
         NIR_PASS_V(c->s, v3d_nir_lower_logic_ops, c);
 
+        if (c->fs_key->line_smoothing) {
+                v3d_nir_lower_line_smooth(c->s);
+                NIR_PASS_V(c->s, nir_lower_global_vars_to_local);
+                /* The lowering pass can introduce new sysval reads */
+                nir_shader_gather_info(c->s, nir_shader_get_entrypoint(c->s));
+        }
+
         /* If the shader has no non-TLB side effects, we can promote it to
          * enabling early_fragment_tests even if the user didn't.
          */
index f3dc3a92fec397ff9a4faa5bc5e51fb47f99cc6f..185bdc687b4c06310242166f46e8bdb3e8c88748 100644 (file)
@@ -149,6 +149,51 @@ v3d_update_primitive_counters(struct v3d_context *v3d)
         }
 }
 
+bool
+v3d_line_smoothing_enabled(struct v3d_context *v3d)
+{
+        if (!v3d->rasterizer->base.line_smooth)
+                return false;
+
+        /* According to the OpenGL docs, line smoothing shouldn’t be applied
+         * when multisampling
+         */
+        if (v3d->job->msaa || v3d->rasterizer->base.multisample)
+                return false;
+
+        if (v3d->framebuffer.nr_cbufs <= 0)
+                return false;
+
+        struct pipe_surface *cbuf = v3d->framebuffer.cbufs[0];
+        if (!cbuf)
+                return false;
+
+        /* Modifying the alpha for pure integer formats probably
+         * doesn’t make sense because we don’t know how the application
+         * uses the alpha value.
+         */
+        if (util_format_is_pure_integer(cbuf->format))
+                return false;
+
+        return true;
+}
+
+float
+v3d_get_real_line_width(struct v3d_context *v3d)
+{
+        float width = v3d->rasterizer->base.line_width;
+
+        if (v3d_line_smoothing_enabled(v3d)) {
+                /* If line smoothing is enabled then we want to add some extra
+                 * pixels to the width in order to have some semi-transparent
+                 * edges.
+                 */
+                width = floorf(M_SQRT2 * width) + 3;
+        }
+
+        return width;
+}
+
 static void
 v3d_context_destroy(struct pipe_context *pctx)
 {
index 55bc96bed9d88bacb147d183d863d6fb8ab9c9c5..74df773ca9adc13f664367e9bccdb01fb72268eb 100644 (file)
@@ -699,6 +699,10 @@ struct v3d_fence *v3d_fence_create(struct v3d_context *v3d);
 
 void v3d_update_primitive_counters(struct v3d_context *v3d);
 
+bool v3d_line_smoothing_enabled(struct v3d_context *v3d);
+
+float v3d_get_real_line_width(struct v3d_context *v3d);
+
 #ifdef v3dX
 #  include "v3dx_context.h"
 #else
index 2415b2797544cfd1b4ed4bf55b0eb469c2f35595..07a596d3eaf9d4b7ee232ce4dba19b01095ee59d 100644 (file)
@@ -541,6 +541,8 @@ v3d_update_compiled_fs(struct v3d_context *v3d, uint8_t prim_mode)
         key->is_points = (prim_mode == PIPE_PRIM_POINTS);
         key->is_lines = (prim_mode >= PIPE_PRIM_LINES &&
                          prim_mode <= PIPE_PRIM_LINE_STRIP);
+        key->line_smoothing = (key->is_lines &&
+                               v3d_line_smoothing_enabled(v3d));
         key->clamp_color = v3d->rasterizer->base.clamp_fragment_color;
         if (v3d->blend->base.logicop_enable) {
                 key->logicop_func = v3d->blend->base.logicop_func;
index 64fde274a998dab0063a44a6d06aec20f14c7c21..37fb98daa5748e65bf8c70d77308b8b7d552618a 100644 (file)
@@ -313,11 +313,14 @@ v3d_write_uniforms(struct v3d_context *v3d, struct v3d_job *job,
                         break;
 
                 case QUNIFORM_LINE_WIDTH:
-                case QUNIFORM_AA_LINE_WIDTH:
                         cl_aligned_f(&uniforms,
                                      v3d->rasterizer->base.line_width);
                         break;
 
+                case QUNIFORM_AA_LINE_WIDTH:
+                        cl_aligned_f(&uniforms, v3d_get_real_line_width(v3d));
+                        break;
+
                 case QUNIFORM_UBO_ADDR: {
                         uint32_t unit = v3d_unit_data_get_unit(data);
                         /* Constant buffer 0 may be a system memory pointer,
index bcad6cddac68a10a7d85a7662c573db5db0e5b86..2dad2e0e24626768d729702c6bc33f9fc7e48a45 100644 (file)
@@ -551,7 +551,7 @@ v3dX(emit_state)(struct pipe_context *pctx)
                 }
 
                 cl_emit(&job->bcl, LINE_WIDTH, line_width) {
-                        line_width.line_width = v3d->rasterizer->base.line_width;
+                        line_width.line_width = v3d_get_real_line_width(v3d);
                 }
         }