diff --git a/doc/NOTES.md b/doc/NOTES.md
index eac1007..be91c4c 100644
--- a/doc/NOTES.md
+++ b/doc/NOTES.md
@@ -1,3 +1,5 @@
+Parser generators in Rust:
+--------------------------
 I'm convinced nobody actually uses parser generators in Rust:
 * pomelo can't generate lexer (understandable, as it is a port of lemon and lemon can't do this either)
 * pest can't do parse actions, you have to convert your parse tree to ast manually
@@ -9,15 +11,17 @@ I'm convinced nobody actually uses parser generators in Rust:
 * no library supports island grammars
 
 What to emit?
+-------------
 * SPIR-V
     * Better library support, easier to emit
     * Can by optimized by IGC
     * Can't do some things (not sure what exactly yet)
-    * But we can work around things with inline VISA
+        * But we can work around with inline VISA
 * VISA
     * Quicker compilation
 
 A64 vs BTS
+----------
 * How to force A64: -cl-intel-greater-than-4GB-buffer-required
 * PTX made a baffling desing choice: global pointers are represented as untyped 64bit integers
 * Consequently, there's no 100% certain way to know which argument is a surface and which is a scalar
@@ -30,3 +34,18 @@ A64 vs BTS
 * Potential solution: compile only during the dispatch, when type of arguments is known?
     * Can't do, the set of arguments passed to cuLaunchKernel is untyped
 * Solution: treat all arguments as untyped integers and say goodbye to BTS access
+
+Implicit conversions
+--------------------
+* PTX support for implicit conversions is completely degenerate, docs say:  
+_For convenience, ld, st, and cvt instructions permit source and destination data operands to be wider than the instruction-type size, so that narrow values may be loaded, stored, and converted using regular-width registers. For example, 8-bit or 16-bit values may be held directly in 32-bit or 64-bit registers when being loaded, stored, or converted to other types and sizes_  
+Which is sensible, but completely untrue. In reality ptxas compiles silly code like this:
+    ```
+    param.f32       param_1
+    ...
+    .reg.s32        %r1
+    ld.param.b16 	%r1, [param_1];
+    ```
+* Surprise, surprise, there's two kind of implicit conversions at play in the example above:
+    * "Relaxed type-checking rules": this is the conversion of b16 operation type to s32 dst register
+    * Undocumented type coercion when dereferencing param_1. The PTX behaviour is to coerce **every** type. It's something like `[param_1] = *(b16*)param_1`
\ No newline at end of file
diff --git a/ptx/tools/implicit_ld_dst.py b/ptx/tools/implicit_ld_dst.py
new file mode 100644
index 0000000..ea95199
--- /dev/null
+++ b/ptx/tools/implicit_ld_dst.py
@@ -0,0 +1,31 @@
+import os
+import subprocess
+import tempfile
+
+types = ["b8", "b16", "b32", "b64", "u8", "u16", "u32", "u64", "s8", "s16", "s32", "s64", "f32", "f64"]
+
+for op_type in types:
+    for output_type in types:
+        with tempfile.TemporaryDirectory() as dir:
+            f_name = os.path.join(dir, 'ptx')
+            out_name = os.path.join(dir, 'out')
+            with open(f_name, 'w') as f:
+                f.write(
+                f"""
+                .version 6.5
+                .target sm_30
+                .address_size 64
+                .visible .entry VecAdd_kernel(
+                    .param .{op_type} input
+                )
+                {{
+                    .reg.{output_type}        r1;
+                    ld.param.{op_type} 	r1, [input];
+                    ret;
+                }}
+                """)
+            err = subprocess.run(f"ptxas {f_name} -o {out_name}", capture_output = True)
+            if err.returncode == 0:
+                print(f"{op_type} {output_type}")
+            else:
+                print(f"[INVALID] {op_type} {output_type}")
\ No newline at end of file
diff --git a/ptx/tools/implicit_ld_src.py b/ptx/tools/implicit_ld_src.py
new file mode 100644
index 0000000..cb6780e
--- /dev/null
+++ b/ptx/tools/implicit_ld_src.py
@@ -0,0 +1,31 @@
+import os
+import subprocess
+import tempfile
+
+types = ["b8", "b16", "b32", "b64", "u8", "u16", "u32", "u64", "s8", "s16", "s32", "s64", "f32", "f64"]
+
+for input_type in types:
+    for op_type in types:
+        with tempfile.TemporaryDirectory() as dir:
+            f_name = os.path.join(dir, 'ptx')
+            out_name = os.path.join(dir, 'out')
+            with open(f_name, 'w') as f:
+                f.write(
+                f"""
+                .version 6.5
+                .target sm_30
+                .address_size 64
+                .visible .entry VecAdd_kernel(
+                    .param .{input_type} input
+                )
+                {{
+                    .reg.{op_type}        r1;
+                    ld.param.{op_type} 	r1, [input];
+                    ret;
+                }}
+                """)
+            err = subprocess.run(f"ptxas {f_name} -o {out_name}")
+            if err.returncode == 0:
+                print(f"{op_type} {input_type}")
+            else:
+                print(f"[INVALID] {op_type} {input_type}")
\ No newline at end of file