diff --git a/libswscale/ops_backend.c b/libswscale/ops_backend.c
index a503139016..449ba8c975 100644
--- a/libswscale/ops_backend.c
+++ b/libswscale/ops_backend.c
@@ -48,29 +48,6 @@ typedef    float f32block_t[SWS_BLOCK_SIZE];
 # include "ops_tmpl_float.c"
 #undef BIT_DEPTH
 
-static void process(const SwsOpExec *exec, const void *priv,
-                    const int bx_start, const int y_start, int bx_end, int y_end)
-{
-    const SwsOpChain *chain = priv;
-    const SwsOpImpl *impl = chain->impl;
-    u32block_t x, y, z, w; /* allocate enough space for any intermediate */
-
-    SwsOpIter iterdata;
-    SwsOpIter *iter = &iterdata; /* for CONTINUE() macro to work */
-
-    for (iter->y = y_start; iter->y < y_end; iter->y++) {
-        for (int i = 0; i < 4; i++) {
-            iter->in[i]  = exec->in[i]  + (iter->y - y_start) * exec->in_stride[i];
-            iter->out[i] = exec->out[i] + (iter->y - y_start) * exec->out_stride[i];
-        }
-
-        for (int block = bx_start; block < bx_end; block++) {
-            iter->x = block * SWS_BLOCK_SIZE;
-            CONTINUE(u32block_t, x, y, z, w);
-        }
-    }
-}
-
 static int compile(SwsContext *ctx, SwsOpList *ops, SwsCompiledOp *out)
 {
     int ret;
@@ -79,6 +56,9 @@ static int compile(SwsContext *ctx, SwsOpList *ops, SwsCompiledOp *out)
     if (!chain)
         return AVERROR(ENOMEM);
 
+    av_assert0(ops->num_ops > 0);
+    const SwsPixelType read_type = ops->ops[0].type;
+
     static const SwsOpTable *const tables[] = {
         &bitfn(op_table_int,    u8),
         &bitfn(op_table_int,   u16),
@@ -96,12 +76,20 @@ static int compile(SwsContext *ctx, SwsOpList *ops, SwsCompiledOp *out)
     }
 
     *out = (SwsCompiledOp) {
-        .func       = process,
         .block_size = SWS_BLOCK_SIZE,
         .cpu_flags  = chain->cpu_flags,
         .priv       = chain,
         .free       = ff_sws_op_chain_free_cb,
     };
+
+    switch (read_type) {
+    case SWS_PIXEL_U8:  out->func = process_u8;  break;
+    case SWS_PIXEL_U16: out->func = process_u16; break;
+    case SWS_PIXEL_U32: out->func = process_u32; break;
+    case SWS_PIXEL_F32: out->func = process_f32; break;
+    default: return AVERROR(EINVAL);
+    }
+
     return 0;
 }
 
diff --git a/libswscale/ops_tmpl_common.c b/libswscale/ops_tmpl_common.c
index 7cfec4e3f6..c0e0d9f3fb 100644
--- a/libswscale/ops_tmpl_common.c
+++ b/libswscale/ops_tmpl_common.c
@@ -175,3 +175,27 @@ WRAP_COMMON_PATTERNS(scale,
     .setup = ff_sws_setup_q,
     .flexible = true,
 );
+
+static void fn(process)(const SwsOpExec *exec, const void *priv,
+                        const int bx_start, const int y_start,
+                        int bx_end, int y_end)
+{
+    const SwsOpChain *chain = priv;
+    const SwsOpImpl *impl = chain->impl;
+    u32block_t x, y, z, w; /* allocate enough space for any intermediate */
+
+    SwsOpIter iterdata;
+    SwsOpIter *iter = &iterdata; /* for CONTINUE() macro to work */
+
+    for (iter->y = y_start; iter->y < y_end; iter->y++) {
+        for (int i = 0; i < 4; i++) {
+            iter->in[i]  = exec->in[i]  + (iter->y - y_start) * exec->in_stride[i];
+            iter->out[i] = exec->out[i] + (iter->y - y_start) * exec->out_stride[i];
+        }
+
+        for (int block = bx_start; block < bx_end; block++) {
+            iter->x = block * SWS_BLOCK_SIZE;
+            CONTINUE(block_t, (void *) x, (void *) y, (void *) z, (void *) w);
+        }
+    }
+}