aco/isel: Fix large inputs being truncated in int32->f16 conversions

The previous code produced incorrect results for inputs outside the
range [INT16_MIN, INT16_MAX].

A problematic case is e.g. i2f16 32768, which previously would be
converted to -32768.0 instead of returning the exactly representable
floating point result.

Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9597>

src/amd/compiler/aco_instruction_selection.cpp

@@ -2453,10 +2453,17 @@ void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr)
          src = convert_int(ctx, bld, src, 64, 32, false);
       }
 
-      if (ctx->program->chip_class >= GFX8) {
+      if (ctx->program->chip_class >= GFX8 && input_size <= 16) {
          bld.vop1(aco_opcode::v_cvt_f16_i16, Definition(dst), src);
       } else {
-         /* GFX7 and earlier do not support direct f16⟷i16 conversions */
+         /* Convert to f32 and then down to f16. This is needed to handle
+          * inputs slightly outside the range [INT16_MIN, INT16_MAX],
+          * which are representable via f16 but wouldn't be converted
+          * correctly by v_cvt_f16_i16.
+          *
+          * This is also the fallback-path taken on GFX7 and earlier, which
+          * do not support direct f16⟷i16 conversions.
+          */
          src = bld.vop1(aco_opcode::v_cvt_f32_i32, bld.def(v1), src);
          bld.vop1(aco_opcode::v_cvt_f16_f32, Definition(dst), src);
       }
@@ -2522,6 +2529,8 @@ void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr)
       }
 
       if (ctx->program->chip_class >= GFX8) {
+         /* float16 has a range of [0, 65519]. Converting from larger
+          * inputs is UB, so we just need to consider the lower 16 bits */
          bld.vop1(aco_opcode::v_cvt_f16_u16, Definition(dst), src);
       } else {
          /* GFX7 and earlier do not support direct f16⟷u16 conversions */