panfrost: Distribute masks for FPK selection

We can calculate these at much lower frequencies and just & together at
draw time.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10869>

src/gallium/drivers/panfrost/pan_blend_cso.c

@@ -115,6 +115,11 @@ panfrost_create_blend_state(struct pipe_context *pipe,
 
                 so->pan.rts[c].equation = equation;
 
+                /* Bifrost needs to know if any render target loads its
+                 * destination in the hot draw path, so precompute this */
+                if (so->info[c].load_dest)
+                        so->load_dest_mask |= BITFIELD_BIT(c);
+
                 /* Converting equations to Mali style is expensive, do it at
                  * CSO create time instead of draw-time */
                 if (so->info[c].fixed_function) {

src/gallium/drivers/panfrost/pan_blend_cso.h

@@ -47,6 +47,9 @@ struct panfrost_blend_state {
         struct pan_blend_state pan;
         struct pan_blend_info info[PIPE_MAX_COLOR_BUFS];
         struct mali_blend_equation_packed equation[PIPE_MAX_COLOR_BUFS];
+
+        /* info.load presented as a bitfield for draw call hot paths */
+        unsigned load_dest_mask : PIPE_MAX_COLOR_BUFS;
 };
 
 void

src/gallium/drivers/panfrost/pan_cmdstream.c

@@ -495,15 +495,8 @@ panfrost_prepare_bifrost_fs_state(struct panfrost_context *ctx,
         if (panfrost_fs_required(fs, so, &ctx->pipe_framebuffer)) {
                 /* Track if any colour buffer is reused across draws, either
                  * from reading it directly, or from failing to write it */
-                bool blend_reads_dest = false;
-                unsigned rt_mask = 0;
-
-                for (unsigned i = 0; i < ctx->pipe_framebuffer.nr_cbufs; ++i) {
-                        if (ctx->pipe_framebuffer.cbufs[i]) {
-                                rt_mask |= (1 << i);
-                                blend_reads_dest |= so->info[i].load_dest;
-                        }
-                }
+                unsigned rt_mask = ctx->fb_rt_mask;
+                bool blend_reads_dest = (so->load_dest_mask & rt_mask);
 
                 state->properties.bifrost.allow_forward_pixel_to_kill =
                         fs->info.fs.can_fpk &&

src/gallium/drivers/panfrost/pan_context.c

@@ -1379,6 +1379,14 @@ panfrost_set_framebuffer_state(struct pipe_context *pctx,
         util_copy_framebuffer_state(&ctx->pipe_framebuffer, fb);
         ctx->batch = NULL;
 
+        /* Hot draw call path needs the mask of active render targets */
+        ctx->fb_rt_mask = 0;
+
+        for (unsigned i = 0; i < ctx->pipe_framebuffer.nr_cbufs; ++i) {
+                if (ctx->pipe_framebuffer.cbufs[i])
+                        ctx->fb_rt_mask |= BITFIELD_BIT(i);
+        }
+
         /* We may need to generate a new variant if the fragment shader is
          * keyed to the framebuffer format (due to EXT_framebuffer_fetch) */
         struct panfrost_shader_variants *fs = ctx->shader[PIPE_SHADER_FRAGMENT];

src/gallium/drivers/panfrost/pan_context.h

@@ -195,6 +195,9 @@ struct panfrost_context {
         enum pipe_render_cond_flag cond_mode;
 
         bool is_noop;
+
+        /* Mask of active render targets */
+        uint8_t fb_rt_mask;
 };
 
 /* Corresponds to the CSO */