[Mesa-dev] [PATCH v4 06/44] nir: Handle fp16 rounding modes at nir_type_conversion_op

[Jose Maria Casanova Crespo]

nir_type_conversion enables new operations to handle rounding modes to
convert to fp16 values. Two new opcodes are enabled nir_op_f2f16_rtne
and nir_op_f2f16_rtz.

The undefined behaviour doesn't has any effect and uses the original
nir_op_f2f16 operation.

v2: Indentation fixed (Jason Ekstrand)

v3: Use explicit case for undefined rounding and assert if
    rounding mode is used for non 16-bit float conversions
    (Jason Ekstrand)
---
 src/compiler/glsl/glsl_to_nir.cpp |  3 ++-
 src/compiler/nir/nir.h            |  3 ++-
 src/compiler/nir/nir_opcodes.py   | 11 +++++++++--
 src/compiler/nir/nir_opcodes_c.py | 15 ++++++++++++++-
 src/compiler/spirv/vtn_alu.c      |  2 +-
 5 files changed, 28 insertions(+), 6 deletions(-)

diff --git a/src/compiler/glsl/glsl_to_nir.cpp b/src/compiler/glsl/glsl_to_nir.cpp
index 33ebc13edb..1e636225c1 100644
--- a/src/compiler/glsl/glsl_to_nir.cpp
+++ b/src/compiler/glsl/glsl_to_nir.cpp
@@ -1575,7 +1575,8 @@ nir_visitor::visit(ir_expression *ir)
    case ir_unop_u642i64: {
       nir_alu_type src_type = nir_get_nir_type_for_glsl_base_type(types[0]);
       nir_alu_type dst_type = nir_get_nir_type_for_glsl_base_type(out_type);
-      result = nir_build_alu(&b, nir_type_conversion_op(src_type, dst_type),
+      result = nir_build_alu(&b, nir_type_conversion_op(src_type, dst_type,
+                                 nir_rounding_mode_undef),
                                  srcs[0], NULL, NULL, NULL);
       /* b2i and b2f don't have fixed bit-size versions so the builder will
        * just assume 32 and we have to fix it up here.
diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
index 883f371d1f..c70c0e0220 100644
--- a/src/compiler/nir/nir.h
+++ b/src/compiler/nir/nir.h
@@ -761,7 +761,8 @@ nir_get_nir_type_for_glsl_type(const struct glsl_type *type)
    return nir_get_nir_type_for_glsl_base_type(glsl_get_base_type(type));
 }
 
-nir_op nir_type_conversion_op(nir_alu_type src, nir_alu_type dst);
+nir_op nir_type_conversion_op(nir_alu_type src, nir_alu_type dst,
+                              nir_rounding_mode rnd);
 
 typedef enum {
    NIR_OP_IS_COMMUTATIVE = (1 << 0),
diff --git a/src/compiler/nir/nir_opcodes.py b/src/compiler/nir/nir_opcodes.py
index 28a0467228..ac7333fe78 100644
--- a/src/compiler/nir/nir_opcodes.py
+++ b/src/compiler/nir/nir_opcodes.py
@@ -179,8 +179,15 @@ for src_t in [tint, tuint, tfloat]:
       else:
          bit_sizes = [8, 16, 32, 64]
       for bit_size in bit_sizes:
-         unop_convert("{0}2{1}{2}".format(src_t[0], dst_t[0], bit_size),
-                      dst_t + str(bit_size), src_t, "src0")
+          if bit_size == 16 and dst_t == tfloat and src_t == tfloat:
+              rnd_modes = ['rtne', 'rtz', 'undef']
+              for rnd_mode in rnd_modes:
+                  unop_convert("{0}2{1}{2}_{3}".format(src_t[0], dst_t[0],
+                                                       bit_size, rnd_mode),
+                               dst_t + str(bit_size), src_t, "src0")
+          else:
+              unop_convert("{0}2{1}{2}".format(src_t[0], dst_t[0], bit_size),
+                           dst_t + str(bit_size), src_t, "src0")
 
 # We'll hand-code the to/from bool conversion opcodes.  Because bool doesn't
 # have multiple bit-sizes, we can always infer the size from the other type.
diff --git a/src/compiler/nir/nir_opcodes_c.py b/src/compiler/nir/nir_opcodes_c.py
index 02bb4738ed..c19185534a 100644
--- a/src/compiler/nir/nir_opcodes_c.py
+++ b/src/compiler/nir/nir_opcodes_c.py
@@ -30,7 +30,7 @@ template = Template("""
 #include "nir.h"
 
 nir_op
-nir_type_conversion_op(nir_alu_type src, nir_alu_type dst)
+nir_type_conversion_op(nir_alu_type src, nir_alu_type dst, nir_rounding_mode rnd)
 {
    nir_alu_type src_base = (nir_alu_type) nir_alu_type_get_base_type(src);
    nir_alu_type dst_base = (nir_alu_type) nir_alu_type_get_base_type(dst);
@@ -64,7 +64,20 @@ nir_type_conversion_op(nir_alu_type src, nir_alu_type dst)
                switch (dst_bit_size) {
 %                 for dst_bits in [16, 32, 64]:
                   case ${dst_bits}:
+%                    if src_t == 'float' and dst_t == 'float' and dst_bits == 16:
+                     switch(rnd) {
+%                       for rnd_t in ['rtne', 'rtz', 'undef']:
+                        case nir_rounding_mode_${rnd_t}:
+                           return ${'nir_op_{0}2{1}{2}_{3}'.format(src_t[0], dst_t[0],
+                                                                   dst_bits, rnd_t)};
+%                       endfor
+                        default:
+                           unreachable("Invalid 16-bit nir rounding mode");
+                     }
+%                    else:
+                     assert(rnd == nir_rounding_mode_undef);
                      return ${'nir_op_{0}2{1}{2}'.format(src_t[0], dst_t[0], dst_bits)};
+%                    endif
 %                 endfor
                   default:
                      unreachable("Invalid nir alu bit size");
diff --git a/src/compiler/spirv/vtn_alu.c b/src/compiler/spirv/vtn_alu.c
index ecf9cbc34d..7ec30b8a63 100644
--- a/src/compiler/spirv/vtn_alu.c
+++ b/src/compiler/spirv/vtn_alu.c
@@ -355,7 +355,7 @@ vtn_nir_alu_op_for_spirv_opcode(SpvOp opcode, bool *swap,
    case SpvOpConvertUToF:
    case SpvOpSConvert:
    case SpvOpFConvert:
-      return nir_type_conversion_op(src, dst);
+      return nir_type_conversion_op(src, dst, nir_rounding_mode_undef);
 
    /* Derivatives: */
    case SpvOpDPdx:         return nir_op_fddx;
-- 
2.14.3