freedreno/a6xx: Pre-calculate user const state size

We can do this when we construct the program state object, rather than
at draw time.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18646>

src/freedreno/ir3/ir3_shader.h

@@ -128,8 +128,6 @@ struct ir3_ubo_analysis_state {
    struct ir3_ubo_range range[IR3_MAX_UBO_PUSH_RANGES];
    uint32_t num_enabled;
    uint32_t size;
-   uint32_t
-      cmdstream_size; /* for per-gen backend to stash required cmdstream size */
 };
 
 /**

src/gallium/drivers/freedreno/a6xx/fd6_const.c

@@ -238,27 +238,25 @@ fd6_emit_ubos(struct fd_context *ctx, const struct ir3_shader_variant *v,
    }
 }
 
-static unsigned
-user_consts_cmdstream_size(struct ir3_shader_variant *v)
+unsigned
+fd6_user_consts_cmdstream_size(struct ir3_shader_variant *v)
 {
+   if (!v)
+      return 0;
+
    struct ir3_const_state *const_state = ir3_const_state(v);
    struct ir3_ubo_analysis_state *ubo_state = &const_state->ubo_state;
+   unsigned packets, size;
 
-   if (unlikely(!ubo_state->cmdstream_size)) {
-      unsigned packets, size;
+   /* pre-calculate size required for userconst stateobj: */
+   ir3_user_consts_size(ubo_state, &packets, &size);
 
-      /* pre-calculate size required for userconst stateobj: */
-      ir3_user_consts_size(ubo_state, &packets, &size);
+   /* also account for UBO addresses: */
+   packets += 1;
+   size += 2 * const_state->num_ubos;
 
-      /* also account for UBO addresses: */
-      packets += 1;
-      size += 2 * const_state->num_ubos;
-
-      unsigned sizedwords = (4 * packets) + size;
-      ubo_state->cmdstream_size = sizedwords * 4;
-   }
-
-   return ubo_state->cmdstream_size;
+   unsigned sizedwords = (4 * packets) + size;
+   return sizedwords * 4;
 }
 
 struct fd_ringbuffer *
@@ -272,13 +270,7 @@ fd6_build_user_consts(struct fd6_emit *emit)
       emit->vs, emit->hs, emit->ds, emit->gs, emit->fs,
    };
    struct fd_context *ctx = emit->ctx;
-   unsigned sz = 0;
-
-   for (unsigned i = 0; i < ARRAY_SIZE(types); i++) {
-      if (!variants[i])
-         continue;
-      sz += user_consts_cmdstream_size(variants[i]);
-   }
+   unsigned sz = emit->prog->user_consts_cmdstream_size;
 
    struct fd_ringbuffer *constobj =
       fd_submit_new_ringbuffer(ctx->batch->submit, sz, FD_RINGBUFFER_STREAMING);

src/gallium/drivers/freedreno/a6xx/fd6_const.h

@@ -28,6 +28,7 @@
 #include "fd6_emit.h"
 
 struct fd_ringbuffer *fd6_build_tess_consts(struct fd6_emit *emit) assert_dt;
+unsigned fd6_user_consts_cmdstream_size(struct ir3_shader_variant *v);
 struct fd_ringbuffer *fd6_build_user_consts(struct fd6_emit *emit) assert_dt;
 struct fd_ringbuffer *
 fd6_build_driver_params(struct fd6_emit *emit) assert_dt;

src/gallium/drivers/freedreno/a6xx/fd6_program.c

@@ -1288,6 +1288,14 @@ fd6_program_create(void *data, struct ir3_shader_variant *bs,
    if (stream_output->num_outputs > 0)
       state->stream_output = stream_output;
 
+   /* Note that binning pass uses same const state as draw pass: */
+   state->user_consts_cmdstream_size =
+         fd6_user_consts_cmdstream_size(state->vs) +
+         fd6_user_consts_cmdstream_size(state->hs) +
+         fd6_user_consts_cmdstream_size(state->ds) +
+         fd6_user_consts_cmdstream_size(state->gs) +
+         fd6_user_consts_cmdstream_size(state->fs);
+
    return &state->base;
 }
 

src/gallium/drivers/freedreno/a6xx/fd6_program.h

@@ -58,6 +58,12 @@ struct fd6_program_state {
     * render targets.
     */
    uint32_t mrt_components;
+
+   /**
+    * Rather than calculating user consts state size each draw,
+    * calculate it up-front.
+    */
+   uint32_t user_consts_cmdstream_size;
 };
 
 static inline struct fd6_program_state *