diff --git a/src/asahi/lib/agx_nir_lower_gs.c b/src/asahi/lib/agx_nir_lower_gs.c
index 991ce2f477b..914614b239c 100644
--- a/src/asahi/lib/agx_nir_lower_gs.c
+++ b/src/asahi/lib/agx_nir_lower_gs.c
@@ -1628,3 +1628,16 @@ agx_nir_predicate_indirect(nir_builder *b, const void *data)
 
    libagx_predicate_indirect(b, params, thread, indexed);
 }
+
+void
+agx_nir_decompress(nir_builder *b, const void *data)
+{
+   const struct agx_decompress_key *key = data;
+
+   nir_def *params = nir_load_preamble(b, 1, 64, .base = 0);
+   nir_def *tile = nir_load_workgroup_id(b);
+   nir_def *local = nir_channel(b, nir_load_local_invocation_id(b), 0);
+   nir_def *samples = nir_imm_int(b, key->nr_samples);
+
+   libagx_decompress(b, params, tile, local, samples);
+}
diff --git a/src/asahi/lib/agx_nir_lower_gs.h b/src/asahi/lib/agx_nir_lower_gs.h
index 46b32c8f6b6..301344cca40 100644
--- a/src/asahi/lib/agx_nir_lower_gs.h
+++ b/src/asahi/lib/agx_nir_lower_gs.h
@@ -97,3 +97,10 @@ struct agx_predicate_indirect_key {
 static_assert(sizeof(struct agx_predicate_indirect_key) == 1, "padded");
 
 void agx_nir_predicate_indirect(struct nir_builder *b, const void *data);
+
+struct agx_decompress_key {
+   uint8_t nr_samples;
+};
+static_assert(sizeof(struct agx_decompress_key) == 1, "padded");
+
+void agx_nir_decompress(struct nir_builder *b, const void *data);
diff --git a/src/asahi/lib/meson.build b/src/asahi/lib/meson.build
index 04e8d577f95..1cea9742515 100644
--- a/src/asahi/lib/meson.build
+++ b/src/asahi/lib/meson.build
@@ -36,6 +36,7 @@ libasahi_decode_files = files(
 
 libagx_shader_files = files(
   'shaders/libagx.h',
+  'shaders/compression.cl',
   'shaders/draws.cl',
   'shaders/draws.h',
   'shaders/geometry.cl',
diff --git a/src/asahi/lib/shaders/compression.cl b/src/asahi/lib/shaders/compression.cl
new file mode 100644
index 00000000000..73bbb2df40e
--- /dev/null
+++ b/src/asahi/lib/shaders/compression.cl
@@ -0,0 +1,126 @@
+/*
+ * Copyright 2024 Valve Corporation
+ * SPDX-License-Identifier: MIT
+ */
+#include "agx_pack.h"
+#include "compression.h"
+#include "libagx.h"
+
+/*
+ * Decompress in place. The metadata is updated, so other processes can read the
+ * image with a compressed texture descriptor.
+ *
+ * Each workgroup processes one 16x16 tile, avoiding races. We use 32x1
+ * workgroups, matching the warp size, meaning each work-item must process
+ * (16*16)/(32*1) = 8 sampels. Matching the warp size eliminates cross-warp
+ * barriers. It also minimizes launched threads, accelerating the early exit.
+ */
+
+/* Our compiler represents a bindless handle as a uint2 of a uniform base and an
+ * offset in bytes. Since the descriptors are all in the u0_u1 push, the former
+ * is hardcoded and the latter is an offsetof.
+ */
+#define HANDLE(field) (uint2)(0, offsetof(struct libagx_decompress_push, field))
+
+/*
+ * The metadata buffer is fully twiddled, so interleave the X/Y coordinate bits.
+ * While dimensions are padded to powers-of-two, they are not padded to a
+ * square. If the width is more than 2x the height or vice versa, the additional
+ * bits are linear. So we interleave as much as possible, and then add what's
+ * remaining. Finally, layers are strided linear and added at the end.
+ */
+uint
+index_metadata(uint3 c, uint width, uint height, uint layer_stride)
+{
+   uint major_coord = width > height ? c.x : c.y;
+   uint minor_dim = min(width, height);
+
+   uint intl_bits = libagx_logbase2_ceil(minor_dim);
+   uint intl_mask = (1 << intl_bits) - 1;
+   uint2 intl_coords = c.xy & intl_mask;
+
+   return nir_interleave_agx(intl_coords.x, intl_coords.y) +
+          ((major_coord & ~intl_mask) << intl_bits) + (layer_stride * c.z);
+}
+
+/*
+ * For multisampled images, a 2x2 or 1x2 group of samples form a single pixel.
+ * The following two helpers convert a coordinate in samples into a coordinate
+ * in pixels and a sample ID, respectively. They each assume that samples > 1.
+ */
+int4
+decompose_px(int4 c, uint samples)
+{
+   if (samples == 4)
+      c.xy >>= 1;
+   else
+      c.y >>= 1;
+
+   return c;
+}
+
+uint
+sample_id(int4 c, uint samples)
+{
+   if (samples == 4)
+      return (c.x & 1) | ((c.y & 1) << 1);
+   else
+      return c.y & 1;
+}
+
+void
+libagx_decompress(constant struct libagx_decompress_push *push, uint3 coord_tl,
+                  uint local_id, uint samples)
+{
+   /* Index into the metadata buffer */
+   uint index_tl =
+      index_metadata(coord_tl, push->metadata_width_tl,
+                     push->metadata_height_tl, push->metadata_layer_stride_tl);
+
+   /* If the tile is already uncompressed, there's nothing to do. */
+   if (push->metadata[index_tl] == push->tile_uncompressed)
+      return;
+
+   /* Tiles are 16x16 */
+   uint2 coord_sa = (coord_tl.xy * 16);
+   uint layer = coord_tl.z;
+
+   /* Since we use a 32x1 workgroup, each work-item handles half of a row. */
+   uint offs_y_sa = local_id >> 1;
+   uint offs_x_sa = (local_id & 1) ? 8 : 0;
+
+   int2 img_coord_sa_2d = convert_int2(coord_sa) + (int2)(offs_x_sa, offs_y_sa);
+   int4 img_coord_sa = (int4)(img_coord_sa_2d.x, img_coord_sa_2d.y, layer, 0);
+
+   /* Read our half-row into registers. */
+   uint4 texels[8];
+   for (uint i = 0; i < 8; ++i) {
+      int4 c_sa = img_coord_sa + (int4)(i, 0, 0, 0);
+      if (samples == 1) {
+         texels[i] = nir_bindless_image_load_array(HANDLE(compressed), c_sa);
+      } else {
+         int4 dec_px = decompose_px(c_sa, samples);
+         texels[i] = nir_bindless_image_load_ms_array(
+            HANDLE(compressed), dec_px, sample_id(c_sa, samples));
+      }
+   }
+
+   sub_group_barrier(CLK_LOCAL_MEM_FENCE);
+
+   /* Now that the whole tile is read, we write without racing. */
+   for (uint i = 0; i < 8; ++i) {
+      int4 c_sa = img_coord_sa + (int4)(i, 0, 0, 0);
+      if (samples == 1) {
+         nir_bindless_image_store_array(HANDLE(uncompressed), c_sa, texels[i]);
+      } else {
+         int4 dec_px = decompose_px(c_sa, samples);
+         nir_bindless_image_store_ms_array(HANDLE(uncompressed), dec_px,
+                                           sample_id(c_sa, samples), texels[i]);
+      }
+   }
+
+   /* We've replaced the body buffer. Mark the tile as uncompressed. */
+   if (local_id == 0) {
+      push->metadata[index_tl] = push->tile_uncompressed;
+   }
+}
diff --git a/src/asahi/lib/shaders/compression.h b/src/asahi/lib/shaders/compression.h
new file mode 100644
index 00000000000..b750b54b4ba
--- /dev/null
+++ b/src/asahi/lib/shaders/compression.h
@@ -0,0 +1,19 @@
+/*
+ * Copyright 2024 Valve Corporation
+ * SPDX-License-Identifier: MIT
+ */
+#include "agx_pack.h"
+#include "libagx.h"
+
+#pragma once
+
+struct libagx_decompress_push {
+   struct agx_texture_packed compressed;
+   struct agx_pbe_packed uncompressed;
+   GLOBAL(uint64_t) metadata;
+   uint64_t tile_uncompressed;
+   uint32_t metadata_layer_stride_tl;
+   uint16_t metadata_width_tl;
+   uint16_t metadata_height_tl;
+};
+AGX_STATIC_ASSERT(sizeof(struct libagx_decompress_push) == 72);
diff --git a/src/asahi/lib/shaders/libagx.h b/src/asahi/lib/shaders/libagx.h
index 13f63e88e7e..050425b6dde 100644
--- a/src/asahi/lib/shaders/libagx.h
+++ b/src/asahi/lib/shaders/libagx.h
@@ -40,6 +40,12 @@ uint32_t nir_load_helper_arg_lo_agx(void);
 uint32_t nir_load_helper_arg_hi_agx(void);
 uint32_t nir_fence_helper_exit_agx(void);
 
+uint4 nir_bindless_image_load_array(uint2 handle, int4 coord);
+void nir_bindless_image_store_array(uint2 handle, int4 coord, uint4 datum);
+uint4 nir_bindless_image_load_ms_array(uint2 handle, int4 coord, uint sample);
+void nir_bindless_image_store_ms_array(uint2 handle, int4 coord, uint sample,
+                                       uint4 datum);
+
 uint libagx_load_index_buffer_internal(uintptr_t index_buffer,
                                        uint32_t index_buffer_range_el, uint id,
                                        uint index_size);
@@ -59,6 +65,14 @@ align(uint x, uint y)
    return (x + y - 1) & ~(y - 1);
 }
 
+static inline uint32_t
+libagx_logbase2_ceil(uint32_t n)
+{
+   return (n <= 1) ? 0 : 32 - clz(n - 1);
+}
+
+#define offsetof(x, y) __builtin_offsetof(x, y)
+
 #endif
 
 #endif