diff --git a/src/gallium/drivers/asahi/agx_state.c b/src/gallium/drivers/asahi/agx_state.c
index 3ffdd30b0e0..ef692f223eb 100644
--- a/src/gallium/drivers/asahi/agx_state.c
+++ b/src/gallium/drivers/asahi/agx_state.c
@@ -55,6 +55,7 @@
 #include "nir_builder_opcodes.h"
 #include "nir_intrinsics.h"
 #include "nir_intrinsics_indices.h"
+#include "nir_xfb_info.h"
 #include "pool.h"
 
 void
@@ -2059,6 +2060,18 @@ agx_shader_initialize(struct agx_device *dev, struct agx_uncompiled_shader *so,
                       so->nir_sha1);
 
    so->has_xfb_info = (nir->xfb_info != NULL);
+
+   static_assert(
+      ARRAY_SIZE(so->xfb_strides) == ARRAY_SIZE(nir->info.xfb_stride),
+      "known target count");
+
+   if (so->has_xfb_info) {
+      struct nir_xfb_info *xfb = nir->xfb_info;
+
+      for (unsigned i = 0; i < ARRAY_SIZE(so->xfb_strides); ++i) {
+         so->xfb_strides[i] = xfb->buffers[i].stride;
+      }
+   }
 }
 
 static void *
@@ -4202,6 +4215,22 @@ agx_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info,
               (ctx->dirty & AGX_DIRTY_VERTEX))
       ctx->dirty |= AGX_DIRTY_VS;
 
+   /* Transform feedback always happens via the geometry shader, so look there
+    * to get the XFB strides.
+    */
+   if (ctx->stage[PIPE_SHADER_GEOMETRY].shader) {
+      struct agx_uncompiled_shader *gs =
+         ctx->stage[PIPE_SHADER_GEOMETRY].shader;
+
+      for (unsigned i = 0; i < ctx->streamout.num_targets; ++i) {
+         struct agx_streamout_target *tgt =
+            agx_so_target(ctx->streamout.targets[i]);
+
+         if (tgt != NULL)
+            tgt->stride = gs->xfb_strides[i];
+      }
+   }
+
    agx_update_gs(ctx, info, indirect);
 
    if (ctx->gs) {
diff --git a/src/gallium/drivers/asahi/agx_state.h b/src/gallium/drivers/asahi/agx_state.h
index 184c870de13..ef0486c8298 100644
--- a/src/gallium/drivers/asahi/agx_state.h
+++ b/src/gallium/drivers/asahi/agx_state.h
@@ -42,6 +42,9 @@
 struct agx_streamout_target {
    struct pipe_stream_output_target base;
    struct pipe_resource *offset;
+
+   /* Current stride (bytes per vertex) */
+   uint32_t stride;
 };
 
 static inline struct agx_streamout_target *
@@ -206,6 +209,8 @@ struct agx_uncompiled_shader {
    struct agx_uncompiled_shader_info info;
    struct hash_table *variants;
    struct agx_uncompiled_shader *passthrough_progs[MESA_PRIM_COUNT][3][2];
+
+   uint32_t xfb_strides[4];
    bool has_xfb_info;
 
    /* Whether the shader accesses indexed samplers via the bindless heap */
diff --git a/src/gallium/drivers/asahi/agx_streamout.c b/src/gallium/drivers/asahi/agx_streamout.c
index 9dde9cee916..2e8e19bebd3 100644
--- a/src/gallium/drivers/asahi/agx_streamout.c
+++ b/src/gallium/drivers/asahi/agx_streamout.c
@@ -121,10 +121,13 @@ agx_draw_vbo_from_xfb(struct pipe_context *pctx,
 {
    perf_debug_ctx(agx_context(pctx), "draw auto");
 
-   unsigned count = 0;
-   pipe_buffer_read(pctx,
-                    agx_so_target(indirect->count_from_stream_output)->offset,
-                    0, 4, &count);
+   struct agx_streamout_target *so =
+      agx_so_target(indirect->count_from_stream_output);
+
+   unsigned offset_B;
+   pipe_buffer_read(pctx, so->offset, 0, 4, &offset_B);
+
+   unsigned count = offset_B / so->stride;
 
    /* XXX: Probably need to divide here */