freedreno/a6xx: Use LATE_Z with OC + discard

If occlusion query is used, and the FS discards samples, we need to use
LATE_Z in order to get the sample count after the FS.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32869>

src/gallium/drivers/freedreno/a5xx/fd5_context.h

@@ -36,9 +36,6 @@ struct fd5_context {
    /* storage for ctx->last.key: */
    struct ir3_shader_key last_key;
 
-   /* number of active samples-passed queries: */
-   int samples_passed_queries;
-
    /* cached state about current emitted shader program (3d): */
    unsigned max_loc;
 };

src/gallium/drivers/freedreno/a5xx/fd5_emit.h

@@ -154,7 +154,7 @@ fd5_emit_render_cntl(struct fd_context *ctx, bool blit, bool binning) assert_dt
     * Other bits seem to depend on query state, like if samples-passed
     * query is active.
     */
-   bool samples_passed = (fd5_context(ctx)->samples_passed_queries > 0);
+   bool samples_passed = (ctx->occlusion_queries_active > 0);
    OUT_PKT4(ring, REG_A5XX_RB_RENDER_CNTL, 1);
    OUT_RING(ring, 0x00000000 | /* RB_RENDER_CNTL */
                      COND(binning, A5XX_RB_RENDER_CNTL_BINNING_PASS) |

src/gallium/drivers/freedreno/a5xx/fd5_query.c

@@ -47,7 +47,9 @@ FD_DEFINE_CAST(fd_acc_query_sample, fd5_query_sample);
 
 static void
 occlusion_resume(struct fd_acc_query *aq, struct fd_batch *batch)
+   assert_dt
 {
+   struct fd_context *ctx = batch->ctx;
    struct fd_ringbuffer *ring = batch->draw;
 
    OUT_PKT4(ring, REG_A5XX_RB_SAMPLE_COUNT_CONTROL, 1);
@@ -61,12 +63,14 @@ occlusion_resume(struct fd_acc_query *aq, struct fd_batch *batch)
    fd5_event_write(batch, ring, ZPASS_DONE, false);
    fd_reset_wfi(batch);
 
-   fd5_context(batch->ctx)->samples_passed_queries++;
+   ctx->occlusion_queries_active++;
 }
 
 static void
 occlusion_pause(struct fd_acc_query *aq, struct fd_batch *batch)
+   assert_dt
 {
+   struct fd_context *ctx = batch->ctx;
    struct fd_ringbuffer *ring = batch->draw;
 
    OUT_PKT7(ring, CP_MEM_WRITE, 4);
@@ -102,7 +106,8 @@ occlusion_pause(struct fd_acc_query *aq, struct fd_batch *batch)
    OUT_RELOC(ring, query_sample(aq, stop));   /* srcB */
    OUT_RELOC(ring, query_sample(aq, start));  /* srcC */
 
-   fd5_context(batch->ctx)->samples_passed_queries--;
+   assert(ctx->occlusion_queries_active > 0);
+   ctx->occlusion_queries_active--;
 }
 
 static void

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

@@ -84,19 +84,19 @@ compute_ztest_mode(struct fd6_emit *emit, bool lrz_valid) assert_dt
       return emit->prog->lrz_mask.z_mode;
 
    struct fd_context *ctx = emit->ctx;
-   struct pipe_framebuffer_state *pfb = &ctx->batch->framebuffer;
    struct fd6_zsa_stateobj *zsa = fd6_zsa_stateobj(ctx->zsa);
    const struct ir3_shader_variant *fs = emit->fs;
 
    if (!zsa->base.depth_enabled) {
       return A6XX_LATE_Z;
    } else if ((fs->has_kill || zsa->alpha_test) &&
-              (zsa->writes_zs || !pfb->zsbuf)) {
-      /* Slightly odd, but seems like the hw wants us to select
-       * LATE_Z mode if there is no depth buffer + discard.  Either
-       * that, or when occlusion query is enabled.  See:
+              (zsa->writes_zs || ctx->occlusion_queries_active)) {
+      /* If occlusion queries are active, we don't want to use EARLY_Z
+       * since that will count samples that are discarded by fs
        *
-       * dEQP-GLES31.functional.fbo.no_attachments.*
+       * I'm not entirely sure about the interaction with LRZ, since
+       * that could discard samples that would otherwise only be
+       * hidden by a later draw.
        */
       return lrz_valid ? A6XX_EARLY_LRZ_LATE_Z : A6XX_LATE_Z;
    } else {

src/gallium/drivers/freedreno/a6xx/fd6_query.cc

@@ -96,6 +96,12 @@ occlusion_resume(struct fd_acc_query *aq, struct fd_batch *batch)
       );
    }
 
+   ctx->occlusion_queries_active++;
+
+   /* Just directly bash the gen specific LRZ dirty bit, since we don't
+    * need to re-emit any other LRZ related state:
+    */
+   ctx->gen_dirty |= FD6_GROUP_LRZ;
 }
 
 template <chip CHIP>
@@ -170,6 +176,14 @@ occlusion_pause(struct fd_acc_query *aq, struct fd_batch *batch) assert_dt
          EV_DST_RAM_CP_EVENT_WRITE7_1(query_sample(aq, start)),
       );
    }
+
+   assert(ctx->occlusion_queries_active > 0);
+   ctx->occlusion_queries_active--;
+
+   /* Just directly bash the gen specific LRZ dirty bit, since we don't
+    * need to re-emit any other LRZ related state:
+    */
+   ctx->gen_dirty |= FD6_GROUP_LRZ;
 }
 
 static void

src/gallium/drivers/freedreno/freedreno_context.h

@@ -319,6 +319,8 @@ struct fd_context {
     */
    bool active_queries dt;
 
+   uint32_t occlusion_queries_active dt;
+
    /* shaders used by clear, and gmem->mem blits: */
    struct fd_program_stateobj solid_prog; // TODO move to screen?
    struct fd_program_stateobj solid_layered_prog;