freedreno/a6xx: Don't disable LRZ for invalid channels

Similar idea to commit ("c9db71c8ba1 turnip: Don't disable LRZ for color
write mask channels beyond the format's"), but different implementation.
If the blend state masks writes to a color channel that doesn't exist,
that doesn't mean we need to disable LRZ write.  In particular, this
scenario will come up with z-prepasses (doh!)

Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19710>

src/gallium/drivers/freedreno/a6xx/fd6_blend.c

@@ -167,27 +167,16 @@ fd6_blend_state_create(struct pipe_context *pctx,
    so->use_dual_src_blend =
       cso->rt[0].blend_enable && util_blend_state_is_dual(cso, 0);
 
+   STATIC_ASSERT((4 * PIPE_MAX_COLOR_BUFS) == (8 * sizeof(so->all_mrt_write_mask)));
+   so->all_mrt_write_mask = 0;
+
    unsigned nr = cso->independent_blend_enable ? cso->max_rt : 0;
    for (unsigned i = 0; i <= nr; i++) {
       const struct pipe_rt_blend_state *rt = &cso->rt[i];
 
       so->reads_dest |= rt->blend_enable;
 
-      /* From the PoV of LRZ, having masked color channels is
-       * the same as having blend enabled, in that the draw will
-       * care about the fragments from an earlier draw.
-       *
-       * NOTE we actually don't care about masked color channels
-       * that don't actually exist in the render target, but we
-       * don't know the render target format here to determine
-       * that.  It is probably not worth worrying about, but if
-       * we find a game/benchmark that goes out of it's way to
-       * mask off non-existent channels, we should fixup the
-       * pipe_blend_state to give us more info.
-       */
-      if (rt->blend_enable || (rt->colormask != 0xf)) {
-         so->reads_dest = true;
-      }
+      so->all_mrt_write_mask |= rt->colormask << (4 * i);
    }
 
    util_dynarray_init(&so->variants, so);

src/gallium/drivers/freedreno/a6xx/fd6_blend.h

@@ -51,6 +51,7 @@ struct fd6_blend_stateobj {
 
    struct fd_context *ctx;
    bool reads_dest;
+   uint32_t all_mrt_write_mask;
    struct util_dynarray variants;
 };
 

src/gallium/drivers/freedreno/a6xx/fd6_emit.c

@@ -433,6 +433,14 @@ compute_lrz_state(struct fd6_emit *emit) assert_dt
       lrz.write = false;
    }
 
+   /* Unwritten channels *that actually exist* are a form of blending
+    * reading the dest from the PoV of LRZ, but the valid dst channels
+    * isn't known when blend CSO is constructed so we need to handle
+    * that here.
+    */
+   if (ctx->all_mrt_channel_mask & ~blend->all_mrt_write_mask)
+      lrz.write = false;
+
    /* if we change depthfunc direction, bail out on using LRZ.  The
     * LRZ buffer encodes a min/max depth value per block, but if
     * we switch from GT/GE <-> LT/LE, those values cannot be

src/gallium/drivers/freedreno/freedreno_context.h

@@ -399,8 +399,11 @@ struct fd_context {
    struct pipe_stencil_ref stencil_ref dt;
    unsigned sample_mask dt;
    unsigned min_samples dt;
+
    /* local context fb state, for when ctx->batch is null: */
    struct pipe_framebuffer_state framebuffer dt;
+   uint32_t all_mrt_channel_mask dt;
+
    struct pipe_poly_stipple stipple dt;
    struct pipe_viewport_state viewport[PIPE_MAX_VIEWPORTS] dt;
    struct pipe_scissor_state viewport_scissor[PIPE_MAX_VIEWPORTS] dt;

src/gallium/drivers/freedreno/freedreno_state.c

@@ -289,6 +289,25 @@ fd_set_framebuffer_state(struct pipe_context *pctx,
 
    util_copy_framebuffer_state(cso, framebuffer);
 
+   STATIC_ASSERT((4 * PIPE_MAX_COLOR_BUFS) == (8 * sizeof(ctx->all_mrt_channel_mask)));
+   ctx->all_mrt_channel_mask = 0;
+
+   /* Generate a bitmask of all valid channels for all MRTs.  Blend
+    * state with unwritten channels essentially acts as blend enabled,
+    * which disables LRZ write.  But only if the cbuf *has* the masked
+    * channels, which is not known at the time the blend state is
+    * created.
+    */
+   for (unsigned i = 0; i < framebuffer->nr_cbufs; i++) {
+      if (!framebuffer->cbufs[i])
+         continue;
+
+      enum pipe_format format = framebuffer->cbufs[i]->format;
+      unsigned nr = util_format_get_nr_components(format);
+
+      ctx->all_mrt_channel_mask |= BITFIELD_MASK(nr) << (4 * i);
+   }
+
    cso->samples = util_framebuffer_get_num_samples(cso);
 
    if (ctx->screen->reorder) {