<div dir="ltr"><div class="gmail_extra"><div class="gmail_quote">On Thu, Jul 13, 2017 at 7:35 AM, Alejandro Piñeiro <<a href="mailto:apinheiro@igalia.com" target="_blank">apinheiro@igalia.com</a>> wrote: <blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">From: Jose Maria Casanova Crespo <<a href="mailto:jmcasanova@igalia.com">jmcasanova@igalia.com</a>> 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. --- src/compiler/glsl/glsl_to_nir.cpp | 3 ++- src/compiler/nir/nir.h | 3 ++- src/compiler/nir/nir_opcodes.py | 10 ++++++++-- src/compiler/nir/nir_opcodes_c.py | 15 ++++++++++++++- src/compiler/spirv/vtn_alu.c | 2 +- 5 files changed, 27 insertions(+), 6 deletions(-) diff --git a/src/compiler/glsl/glsl_to_nir.cpp b/src/compiler/glsl/glsl_to_nir.cpp index 2153004..23e6121 100644 --- a/src/compiler/glsl/glsl_to_nir.cpp +++ b/src/compiler/glsl/glsl_to_nir.cpp @@ -1509,7 +1509,8 @@ 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 115ec1b..7e48e18 100644 --- a/src/compiler/nir/nir.h +++ b/src/compiler/nir/nir.h @@ -741,7 +741,8 @@ 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 39c01a7..bd342de 100644 --- a/src/compiler/nir/nir_opcodes.py +++ b/src/compiler/nir/nir_opcodes.py @@ -179,8 +179,14 @@ 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'] + 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") + unop_convert("{0}2{1}{2}".format(src_t[0], dst_t[0], bit_size), + dst_t + str(bit_size), src_t, "src0") </blockquote><div> </div><div>You could do this a bit shorter if you make rnd_modes = ['', '_rtne', 'rtz'] and then just loop over them. </div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"> # 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 a1db54f..a7721d3 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 [32, 64]: case ${dst_bits}: +% if src_t == 'float' and dst_t == 'float' and dst_bits == 16: + switch(rnd) { +% for rnd_t in ['rtne' , 'rtz']: + case nir_rounding_mode_${rnd_t}: + return ${'nir_op_{0}2{1}{2}_{3}'.format(src_t[0], dst_t[0], + dst_bits, rnd_t)}; +% endfor + default: </blockquote><div> </div><div>This default makes me a bit uncomfortable. I think I'd rather assert-fail if it's a rounding mode we don't have an opcode for. Otherwise, this function is just silently ignoring rounding modes which seems bad. </div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"> + return ${'nir_op_{0}2{1}{2}'.format(src_t[0], dst_t[0], + dst_bits)}; + } +% else: </blockquote><div> </div><div>The indentation on the above is a bit weird. I'm not really sure what to do with it. What I would say is that you shouldn't worry too much about making the resulting C nicely indented. Just make the generator readable. </div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"> 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 ecf9cbc..7ec30b8 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; -- 2.9.3 _______________________________________________ mesa-dev mailing list <a href="mailto:mesa-dev@lists.freedesktop.org">mesa-dev@lists.freedesktop.org</a> <a href="https://lists.freedesktop.org/mailman/listinfo/mesa-dev" rel="noreferrer" target="_blank">https://lists.freedesktop.org/mailman/listinfo/mesa-dev</a> </blockquote></div> </div></div>