nir/opcodes: Add nir_op_f2fmp

This opcode is the same as the f2f16 opcode except that it comes with
a promise that it is safe to optimise it out if the result is
immediately converted back to a 32-bit float again. Normally this
would be a lossy conversion and so it would be visible to the
application, but if the conversion is generated as part of the mediump
lowering process then this removal doesn’t matter. The opcode is
eventually replaced with a regular f2f16 in the late optimisations so
the backends don’t need to handle it.

Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3822>

diff --git a/src/compiler/nir/nir_opcodes.py b/src/compiler/nir/nir_opcodes.py
index d5e3108..f83b830 100644 (file)
--- a/src/compiler/nir/nir_opcodes.py
+++ b/src/compiler/nir/nir_opcodes.py
@@ -100,6 +100,7 @@ tbool16 = "bool16"
 tbool32 = "bool32"
 tuint = "uint"
 tuint16 = "uint16"
+tfloat16 = "float16"
 tfloat32 = "float32"
 tint32 = "int32"
 tuint32 = "uint32"
@@ -265,6 +266,11 @@ for src_t in [tint, tuint, tfloat, tbool]:
                                                        dst_bit_size),
                                    dst_t + str(dst_bit_size), src_t, conv_expr)
 
+# Special opcode that is the same as f2f16 except that it is safe to remove it
+# if the result is immediately converted back to float32 again. This is
+# generated as part of the precision lowering pass. mp stands for medium
+# precision.
+unop_numeric_convert("f2fmp", tfloat16, tfloat, opcodes["f2f16"].const_expr)
 
 # Unary floating-point rounding operations.
 

diff --git a/src/compiler/nir/nir_opt_algebraic.py b/src/compiler/nir/nir_opt_algebraic.py
index 94c8836..281b3a6 100644 (file)
--- a/src/compiler/nir/nir_opt_algebraic.py
+++ b/src/compiler/nir/nir_opt_algebraic.py
@@ -1737,6 +1737,11 @@ late_optimizations = [
     ('ffma', a, b, ('ffma', c, d, ('ffma', e, 'f', 'g'))), '(info->stage != MESA_SHADER_VERTEX && info->stage != MESA_SHADER_GEOMETRY) && !options->intel_vec4'),
    (('~fadd', ('ffma(is_used_once)', a, b, ('fmul', 'c(is_not_const_and_not_fsign)', 'd(is_not_const_and_not_fsign)') ), 'e(is_not_const)'),
     ('ffma', a, b, ('ffma', c, d, e)), '(info->stage != MESA_SHADER_VERTEX && info->stage != MESA_SHADER_GEOMETRY) && !options->intel_vec4'),
+
+   # Convert f2fmp instructions to concrete f2f16 instructions. At this point
+   # any conversions that could have been removed will have been removed in
+   # nir_opt_algebraic so any remaining ones are required.
+   (('f2fmp', a), ('f2f16', a)),
 ]
 
 for op in ['fadd']: