<html dir="ltr"><head></head><body style="text-align:left; direction:ltr;"><div>On Thu, 2019-02-14 at 16:59 -0600, Jason Ekstrand wrote:</div><blockquote type="cite" style="margin:0 0 0 .8ex; border-left:2px #729fcf solid;padding-left:1ex"><div dir="ltr"><div class="gmail_quote"><div dir="ltr" class="gmail_attr">On Tue, Feb 12, 2019 at 5:56 AM Iago Toral Quiroga <<a href="mailto:itoral@igalia.com">itoral@igalia.com</a>> wrote: </div><blockquote type="cite" style="margin:0 0 0 .8ex; border-left:2px #729fcf solid;padding-left:1ex">Some conversions are not directly supported in hardware and need to be split in two conversion instructions going through an intermediary type. Doing this at the NIR level simplifies a bit the complexity in the backend. v2: - Consider fp16 rounding conversion opcodes - Properly handle swizzles on conversion sources. v3 - Run the pass earlier, right after nir_opt_algebraic_late (Jason) - NIR alu output types already have the bit-size (Jason) - Use the new is_conversion field in nir_op_info to select conversion instructions (Jason) - Use 'is_conversion' to identify conversion operations (Jason) Reviewed-by: Topi Pohjolainen <<a href="mailto:topi.pohjolainen@intel.com" target="_blank">topi.pohjolainen@intel.com</a>> (v1) --- src/intel/Makefile.sources | 1 + src/intel/compiler/brw_nir.c | 2 + src/intel/compiler/brw_nir.h | 2 + .../compiler/brw_nir_lower_conversions.c | 158 ++++++++++++++++++ src/intel/compiler/meson.build | 1 + 5 files changed, 164 insertions(+) create mode 100644 src/intel/compiler/brw_nir_lower_conversions.c diff --git a/src/intel/Makefile.sources b/src/intel/Makefile.sources index 94a28d370e8..9975daa3ad1 100644 --- a/src/intel/Makefile.sources +++ b/src/intel/Makefile.sources @@ -83,6 +83,7 @@ COMPILER_FILES = \ compiler/brw_nir_analyze_boolean_resolves.c \ compiler/brw_nir_analyze_ubo_ranges.c \ compiler/brw_nir_attribute_workarounds.c \ + compiler/brw_nir_lower_conversions.c \ compiler/brw_nir_lower_cs_intrinsics.c \ compiler/brw_nir_lower_image_load_store.c \ compiler/brw_nir_lower_mem_access_bit_sizes.c \ diff --git a/src/intel/compiler/brw_nir.c b/src/intel/compiler/brw_nir.c index 9dbf06004a4..7e3dbc9e447 100644 --- a/src/intel/compiler/brw_nir.c +++ b/src/intel/compiler/brw_nir.c @@ -876,6 +876,8 @@ brw_postprocess_nir(nir_shader *nir, const struct brw_compiler *compiler, OPT(nir_opt_algebraic_late); + OPT(brw_nir_lower_conversions); + OPT(nir_lower_to_source_mods, nir_lower_all_source_mods); OPT(nir_copy_prop); OPT(nir_opt_dce); diff --git a/src/intel/compiler/brw_nir.h b/src/intel/compiler/brw_nir.h index bc81950d47e..662b2627e95 100644 --- a/src/intel/compiler/brw_nir.h +++ b/src/intel/compiler/brw_nir.h @@ -114,6 +114,8 @@ void brw_nir_lower_tcs_outputs(nir_shader *nir, const struct brw_vue_map *vue, GLenum tes_primitive_mode); void brw_nir_lower_fs_outputs(nir_shader *nir); +bool brw_nir_lower_conversions(nir_shader *nir); + bool brw_nir_lower_image_load_store(nir_shader *nir, const struct gen_device_info *devinfo); void brw_nir_rewrite_image_intrinsic(nir_intrinsic_instr *intrin, diff --git a/src/intel/compiler/brw_nir_lower_conversions.c b/src/intel/compiler/brw_nir_lower_conversions.c new file mode 100644 index 00000000000..bb1312ad428 --- /dev/null +++ b/src/intel/compiler/brw_nir_lower_conversions.c @@ -0,0 +1,158 @@ +/* + * Copyright © 2018 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + */ + +#include "brw_nir.h" +#include "compiler/nir/nir_builder.h" + +static nir_op +get_conversion_op(nir_alu_type src_type, + unsigned src_bit_size, + nir_alu_type dst_type, + unsigned dst_bit_size, + nir_rounding_mode rounding_mode) +{ + nir_alu_type src_full_type = (nir_alu_type) (src_type | src_bit_size); + nir_alu_type dst_full_type = (nir_alu_type) (dst_type | dst_bit_size); + + return nir_type_conversion_op(src_full_type, dst_full_type, rounding_mode); +} + +static nir_rounding_mode +get_opcode_rounding_mode(nir_op op) +{ + switch (op) { + case nir_op_f2f16_rtz: + return nir_rounding_mode_rtz; + case nir_op_f2f16_rtne: + return nir_rounding_mode_rtne; + default: + return nir_rounding_mode_undef; + } +} + +static void +split_conversion(nir_builder *b, nir_alu_instr *alu, nir_op op1, nir_op op2) +{ + b->cursor = nir_before_instr(&alu->instr); + assert(alu->dest.write_mask == 1); + nir_ssa_def *src = nir_ssa_for_alu_src(b, alu, 0); + nir_ssa_def *tmp = nir_build_alu(b, op1, src, NULL, NULL, NULL); + nir_ssa_def *res = nir_build_alu(b, op2, tmp, NULL, NULL, NULL); + nir_ssa_def_rewrite_uses(&alu->dest.dest.ssa, nir_src_for_ssa(res)); + nir_instr_remove(&alu->instr); +} + +static bool +lower_instr(nir_builder *b, nir_alu_instr *alu) +{ + unsigned src_bit_size = nir_src_bit_size(alu->src[0].src); + nir_alu_type src_type = nir_op_infos[alu->op].input_types[0]; + nir_alu_type src_full_type = (nir_alu_type) (src_type | src_bit_size); + + unsigned dst_bit_size = nir_dest_bit_size(alu->dest.dest); + nir_alu_type dst_full_type = nir_op_infos[alu->op].output_type; + nir_alu_type dst_type = nir_alu_type_get_base_type(dst_full_type); + + /* BDW PRM, vol02, Command Reference Instructions, mov - MOVE: + * + * "There is no direct conversion from HF to DF or DF to HF. + * Use two instructions and F (Float) as an intermediate type. + * + * There is no direct conversion from HF to Q/UQ or Q/UQ to HF. + * Use two instructions and F (Float) or a word integer type + * or a DWord integer type as an intermediate type." + */ + if ((src_full_type == nir_type_float16 && dst_bit_size == 64) || + (src_bit_size == 64 && dst_full_type == nir_type_float16)) { + nir_op op1 = get_conversion_op(src_type, src_bit_size, src_type, 32, + nir_rounding_mode_undef); + nir_op op2 = get_conversion_op(src_type, 32, dst_type, dst_bit_size, + get_opcode_rounding_mode(alu->op)); </blockquote><div> </div><div>This always goes from the source type to the 32-bit variant of the source type. However, for something like a uint64_t -> float16 cast, this is incorrect because the entire range of uint64_t is representable as float16 and this will cause a uint32_t cast in the middle. I think what you want is to always convert to 32-bit float as the middle type. </div></div></div></blockquote><div> </div><div>You are right, I'll fix that.</div><div> </div><blockquote type="cite" style="margin:0 0 0 .8ex; border-left:2px #729fcf solid;padding-left:1ex"><div dir="ltr"><div class="gmail_quote"><blockquote type="cite" style="margin:0 0 0 .8ex; border-left:2px #729fcf solid;padding-left:1ex"> + split_conversion(b, alu, op1, op2); + return true; + } + + /* SKL PRM, vol 02a, Command Reference: Instructions, Move: + * + * "There is no direct conversion from B/UB to DF or DF to B/UB. Use + * two instructions and a word or DWord intermediate type." + * + * "There is no direct conversion from B/UB to Q/UQ or Q/UQ to B/UB. + * Use two instructions and a word or DWord intermediate integer + * type." + */ + if ((src_bit_size == 8 && dst_bit_size == 64) || + (src_bit_size == 64 && dst_bit_size == 8)) { + nir_op op1 = get_conversion_op(src_type, src_bit_size, src_type, 32, + nir_rounding_mode_undef); + nir_op op2 = get_conversion_op(src_type, 32, dst_type, dst_bit_size, + nir_rounding_mode_undef); </blockquote><div> </div><div>When doing a double -> [u]int8_t conversion, simply going through 32-bit float isn't going to yield the correct results. If, for instance, you have a double that's just below 5.0 but close enough that a rtne conversion to 32-bit float will make it 5.0 exactly, then this will yield 5 where it should yield 4.</div></div></div><div dir="ltr"><div class="gmail_quote"><div> If, on the other hand, we always use a 32-bit (with a matching sign) integer type for both directions of the conversion, we should get the correct results because it will be a double -> [u]int32_t conversion which should round reasonably. For the direction where the source is 8-bit, it doesn't matter. </div></div></div></blockquote><div> </div><div>Oh yes, great catch! I'll fix this too.</div><div> </div><div>Thanks!</div><div> </div><div> </div><div> </div><blockquote type="cite" style="margin:0 0 0 .8ex; border-left:2px #729fcf solid;padding-left:1ex"><div dir="ltr"><div class="gmail_quote"><blockquote type="cite" style="margin:0 0 0 .8ex; border-left:2px #729fcf solid;padding-left:1ex"> + split_conversion(b, alu, op1, op2); + return true; + } + + return false; +} + +static bool +lower_impl(nir_function_impl *impl) +{ + nir_builder b; + nir_builder_init(&b, impl); + bool progress = false; + + nir_foreach_block(block, impl) { + nir_foreach_instr_safe(instr, block) { + if (instr->type != nir_instr_type_alu) + continue; + + nir_alu_instr *alu = nir_instr_as_alu(instr); + assert(alu->dest.dest.is_ssa); + + if (!nir_op_infos[alu->op].is_conversion) + continue; + + progress = lower_instr(&b, alu) || progress; + } + } + + if (progress) { + nir_metadata_preserve(impl, nir_metadata_block_index | + nir_metadata_dominance); + } + + return progress; +} + +bool +brw_nir_lower_conversions(nir_shader *shader) +{ + bool progress = false; + + nir_foreach_function(function, shader) { + if (function->impl) + progress |= lower_impl(function->impl); + } + + return progress; +} diff --git a/src/intel/compiler/meson.build b/src/intel/compiler/meson.build index f8e5e2518fe..9daf2a94260 100644 --- a/src/intel/compiler/meson.build +++ b/src/intel/compiler/meson.build @@ -76,6 +76,7 @@ libintel_compiler_files = files( 'brw_nir_analyze_boolean_resolves.c', 'brw_nir_analyze_ubo_ranges.c', 'brw_nir_attribute_workarounds.c', + 'brw_nir_lower_conversions.c', 'brw_nir_lower_cs_intrinsics.c', 'brw_nir_lower_image_load_store.c', 'brw_nir_lower_mem_access_bit_sizes.c', </blockquote></div></div></blockquote></body></html>