diff --git a/src/panfrost/compiler/bifrost_compile.c b/src/panfrost/compiler/bifrost_compile.c
index 4bf803f7d20..96d4353734b 100644
--- a/src/panfrost/compiler/bifrost_compile.c
+++ b/src/panfrost/compiler/bifrost_compile.c
@@ -556,6 +556,7 @@ bi_emit_load_vary(bi_builder *b, nir_intrinsic_instr *instr)
    enum bi_sample sample = BI_SAMPLE_CENTER;
    enum bi_update update = BI_UPDATE_STORE;
    enum bi_register_format regfmt = BI_REGISTER_FORMAT_AUTO;
+   enum bi_source_format source_format;
    bool smooth = instr->intrinsic == nir_intrinsic_load_interpolated_input;
    bi_index src0 = bi_null();
 
@@ -575,9 +576,11 @@ bi_emit_load_vary(bi_builder *b, nir_intrinsic_instr *instr)
 
       assert(sz == 16 || sz == 32);
       regfmt = (sz == 16) ? BI_REGISTER_FORMAT_F16 : BI_REGISTER_FORMAT_F32;
+      source_format = BI_SOURCE_FORMAT_F32;
    } else {
       assert(sz == 32);
       regfmt = BI_REGISTER_FORMAT_U32;
+      source_format = BI_SOURCE_FORMAT_FLAT32;
 
       /* Valhall can't have bi_null() here, although the source is
        * logically unused for flat varyings
@@ -601,9 +604,6 @@ bi_emit_load_vary(bi_builder *b, nir_intrinsic_instr *instr)
       b->shader->malloc_idvs && b->shader->inputs->valhall.use_ld_var_buf;
 
    if (use_ld_var_buf) {
-      enum bi_source_format source_format =
-         smooth ? BI_SOURCE_FORMAT_F32 : BI_SOURCE_FORMAT_FLAT32;
-
       if (immediate) {
          /* Immediate index given in bytes. */
          bi_ld_var_buf_imm_to(b, sz, dest, src0, regfmt, sample, source_format,
diff --git a/src/panfrost/util/pan_collect_varyings.c b/src/panfrost/util/pan_collect_varyings.c
index 03b43bfd79f..ea6c74f534f 100644
--- a/src/panfrost/util/pan_collect_varyings.c
+++ b/src/panfrost/util/pan_collect_varyings.c
@@ -48,6 +48,7 @@ varying_format(nir_alu_type t, unsigned ncomps)
       VARYING_FORMAT(float, 32, FLOAT, 32),
       VARYING_FORMAT(uint, 32, UINT, 32),
       VARYING_FORMAT(float, 16, FLOAT, 16),
+      VARYING_FORMAT(uint, 16, UINT, 16),
    };
 #undef VARYING_FORMAT
 
@@ -84,6 +85,7 @@ walk_varyings(UNUSED nir_builder *b, nir_instr *instr, void *data)
 
    nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
    unsigned count;
+   unsigned size;
 
    /* Only consider intrinsics that access varyings */
    switch (intr->intrinsic) {
@@ -92,6 +94,7 @@ walk_varyings(UNUSED nir_builder *b, nir_instr *instr, void *data)
          return false;
 
       count = nir_src_num_components(intr->src[0]);
+      size = nir_alu_type_get_type_size(nir_intrinsic_src_type(intr));
       break;
 
    case nir_intrinsic_load_input:
@@ -100,6 +103,7 @@ walk_varyings(UNUSED nir_builder *b, nir_instr *instr, void *data)
          return false;
 
       count = intr->def.num_components;
+      size = intr->def.bit_size;
       break;
 
    default:
@@ -130,9 +134,14 @@ walk_varyings(UNUSED nir_builder *b, nir_instr *instr, void *data)
     * as having no demonstrable benefit in practice.
     */
    if (type == nir_type_float && sem.medium_precision)
-      type |= 16;
-   else
-      type |= 32;
+      size = 16;
+   else {
+      /* We never emit 16-bit user varyings except for lowered mediump floats */
+      if (sem.location >= VARYING_SLOT_VAR0)
+         assert(size == 32);
+   }
+
+   type |= size;
 
    /* Count currently contains the number of components accessed by this
     * intrinsics. However, we may be accessing a fractional location,