X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fcompiler%2Fnir%2Fnir_lower_alu_to_scalar.c;h=138318fbf60b0bd0b02e56a7ac9421b009558d01;hb=7665398e6c4fa903405d3daefcc93ddb0a37b488;hp=8a967c56831b8e63287dba62dd0e95d181b23d2b;hpb=3460d05a718f3859a77fe100f3972095d194be26;p=mesa.git diff --git a/src/compiler/nir/nir_lower_alu_to_scalar.c b/src/compiler/nir/nir_lower_alu_to_scalar.c index 8a967c56831..138318fbf60 100644 --- a/src/compiler/nir/nir_lower_alu_to_scalar.c +++ b/src/compiler/nir/nir_lower_alu_to_scalar.c @@ -24,39 +24,61 @@ #include "nir.h" #include "nir_builder.h" +struct alu_to_scalar_data { + nir_instr_filter_cb cb; + const void *data; +}; + /** @file nir_lower_alu_to_scalar.c * * Replaces nir_alu_instr operations with more than one channel used in the * arguments with individual per-channel operations. */ +static bool +inst_is_vector_alu(const nir_instr *instr, const void *_state) +{ + if (instr->type != nir_instr_type_alu) + return false; + + nir_alu_instr *alu = nir_instr_as_alu(instr); + + /* There is no ALU instruction which has a scalar destination, scalar + * src[0], and some other vector source. + */ + assert(alu->dest.dest.is_ssa); + assert(alu->src[0].src.is_ssa); + return alu->dest.dest.ssa.num_components > 1 || + nir_op_infos[alu->op].input_sizes[0] > 1; +} + static void -nir_alu_ssa_dest_init(nir_alu_instr *instr, unsigned num_components, +nir_alu_ssa_dest_init(nir_alu_instr *alu, unsigned num_components, unsigned bit_size) { - nir_ssa_dest_init(&instr->instr, &instr->dest.dest, num_components, + nir_ssa_dest_init(&alu->instr, &alu->dest.dest, num_components, bit_size, NULL); - instr->dest.write_mask = (1 << num_components) - 1; + alu->dest.write_mask = (1 << num_components) - 1; } -static void -lower_reduction(nir_alu_instr *instr, nir_op chan_op, nir_op merge_op, +static nir_ssa_def * +lower_reduction(nir_alu_instr *alu, nir_op chan_op, nir_op merge_op, nir_builder *builder) { - unsigned num_components = nir_op_infos[instr->op].input_sizes[0]; + unsigned num_components = nir_op_infos[alu->op].input_sizes[0]; nir_ssa_def *last = NULL; for (unsigned i = 0; i < num_components; i++) { nir_alu_instr *chan = nir_alu_instr_create(builder->shader, chan_op); - nir_alu_ssa_dest_init(chan, 1, instr->dest.dest.ssa.bit_size); - nir_alu_src_copy(&chan->src[0], &instr->src[0], chan); + nir_alu_ssa_dest_init(chan, 1, alu->dest.dest.ssa.bit_size); + nir_alu_src_copy(&chan->src[0], &alu->src[0], chan); chan->src[0].swizzle[0] = chan->src[0].swizzle[i]; if (nir_op_infos[chan_op].num_inputs > 1) { assert(nir_op_infos[chan_op].num_inputs == 2); - nir_alu_src_copy(&chan->src[1], &instr->src[1], chan); + nir_alu_src_copy(&chan->src[1], &alu->src[1], chan); chan->src[1].swizzle[0] = chan->src[1].swizzle[i]; } - chan->exact = instr->exact; + chan->exact = alu->exact; nir_builder_instr_insert(builder, &chan->instr); @@ -68,52 +90,55 @@ lower_reduction(nir_alu_instr *instr, nir_op chan_op, nir_op merge_op, } } - assert(instr->dest.write_mask == 1); - nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(last)); - nir_instr_remove(&instr->instr); + return last; } -static bool -lower_alu_instr_scalar(nir_alu_instr *instr, nir_builder *b) +static nir_ssa_def * +lower_alu_instr_scalar(nir_builder *b, nir_instr *instr, void *_data) { - unsigned num_src = nir_op_infos[instr->op].num_inputs; + struct alu_to_scalar_data *data = _data; + nir_alu_instr *alu = nir_instr_as_alu(instr); + unsigned num_src = nir_op_infos[alu->op].num_inputs; unsigned i, chan; - assert(instr->dest.dest.is_ssa); - assert(instr->dest.write_mask != 0); + assert(alu->dest.dest.is_ssa); + assert(alu->dest.write_mask != 0); - b->cursor = nir_before_instr(&instr->instr); - b->exact = instr->exact; + b->cursor = nir_before_instr(&alu->instr); + b->exact = alu->exact; + + if (data->cb && !data->cb(instr, data->data)) + return NULL; #define LOWER_REDUCTION(name, chan, merge) \ case name##2: \ case name##3: \ case name##4: \ - lower_reduction(instr, chan, merge, b); \ - return true; + case name##8: \ + case name##16: \ + return lower_reduction(alu, chan, merge, b); \ - switch (instr->op) { + switch (alu->op) { + case nir_op_vec16: + case nir_op_vec8: case nir_op_vec4: case nir_op_vec3: case nir_op_vec2: + case nir_op_cube_face_coord: + case nir_op_cube_face_index: /* We don't need to scalarize these ops, they're the ones generated to * group up outputs into a value that can be SSAed. */ - return false; + return NULL; - case nir_op_pack_half_2x16: + case nir_op_pack_half_2x16: { if (!b->shader->options->lower_pack_half_2x16) - return false; - - nir_ssa_def *val = - nir_pack_half_2x16_split(b, nir_channel(b, instr->src[0].src.ssa, - instr->src[0].swizzle[0]), - nir_channel(b, instr->src[0].src.ssa, - instr->src[0].swizzle[1])); + return NULL; - nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(val)); - nir_instr_remove(&instr->instr); - return true; + nir_ssa_def *src_vec2 = nir_ssa_for_alu_src(b, alu, 0); + return nir_pack_half_2x16_split(b, nir_channel(b, src_vec2, 0), + nir_channel(b, src_vec2, 1)); + } case nir_op_unpack_unorm_4x8: case nir_op_unpack_snorm_4x8: @@ -122,159 +147,142 @@ lower_alu_instr_scalar(nir_alu_instr *instr, nir_builder *b) /* There is no scalar version of these ops, unless we were to break it * down to bitshifts and math (which is definitely not intended). */ - return false; + return NULL; + case nir_op_unpack_half_2x16_flush_to_zero: case nir_op_unpack_half_2x16: { if (!b->shader->options->lower_unpack_half_2x16) - return false; - - nir_ssa_def *comps[2]; - comps[0] = nir_unpack_half_2x16_split_x(b, instr->src[0].src.ssa); - comps[1] = nir_unpack_half_2x16_split_y(b, instr->src[0].src.ssa); - nir_ssa_def *vec = nir_vec(b, comps, 2); - - nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(vec)); - nir_instr_remove(&instr->instr); - return true; + return NULL; + + nir_ssa_def *packed = nir_ssa_for_alu_src(b, alu, 0); + if (alu->op == nir_op_unpack_half_2x16_flush_to_zero) { + return nir_vec2(b, + nir_unpack_half_2x16_split_x_flush_to_zero(b, + packed), + nir_unpack_half_2x16_split_y_flush_to_zero(b, + packed)); + } else { + return nir_vec2(b, + nir_unpack_half_2x16_split_x(b, packed), + nir_unpack_half_2x16_split_y(b, packed)); + } } case nir_op_pack_uvec2_to_uint: { assert(b->shader->options->lower_pack_snorm_2x16 || b->shader->options->lower_pack_unorm_2x16); - nir_ssa_def *word = - nir_extract_u16(b, instr->src[0].src.ssa, nir_imm_int(b, 0)); - nir_ssa_def *val = - nir_ior(b, nir_ishl(b, nir_channel(b, word, 1), nir_imm_int(b, 16)), - nir_channel(b, word, 0)); - - nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(val)); - nir_instr_remove(&instr->instr); - break; + nir_ssa_def *word = nir_extract_u16(b, nir_ssa_for_alu_src(b, alu, 0), + nir_imm_int(b, 0)); + return nir_ior(b, nir_ishl(b, nir_channel(b, word, 1), + nir_imm_int(b, 16)), + nir_channel(b, word, 0)); } case nir_op_pack_uvec4_to_uint: { assert(b->shader->options->lower_pack_snorm_4x8 || b->shader->options->lower_pack_unorm_4x8); - nir_ssa_def *byte = - nir_extract_u8(b, instr->src[0].src.ssa, nir_imm_int(b, 0)); - nir_ssa_def *val = - nir_ior(b, nir_ior(b, nir_ishl(b, nir_channel(b, byte, 3), nir_imm_int(b, 24)), - nir_ishl(b, nir_channel(b, byte, 2), nir_imm_int(b, 16))), - nir_ior(b, nir_ishl(b, nir_channel(b, byte, 1), nir_imm_int(b, 8)), - nir_channel(b, byte, 0))); - - nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(val)); - nir_instr_remove(&instr->instr); - break; + nir_ssa_def *byte = nir_extract_u8(b, nir_ssa_for_alu_src(b, alu, 0), + nir_imm_int(b, 0)); + return nir_ior(b, nir_ior(b, nir_ishl(b, nir_channel(b, byte, 3), + nir_imm_int(b, 24)), + nir_ishl(b, nir_channel(b, byte, 2), + nir_imm_int(b, 16))), + nir_ior(b, nir_ishl(b, nir_channel(b, byte, 1), + nir_imm_int(b, 8)), + nir_channel(b, byte, 0))); } case nir_op_fdph: { + nir_ssa_def *src0_vec = nir_ssa_for_alu_src(b, alu, 0); + nir_ssa_def *src1_vec = nir_ssa_for_alu_src(b, alu, 1); + nir_ssa_def *sum[4]; for (unsigned i = 0; i < 3; i++) { - sum[i] = nir_fmul(b, nir_channel(b, instr->src[0].src.ssa, - instr->src[0].swizzle[i]), - nir_channel(b, instr->src[1].src.ssa, - instr->src[1].swizzle[i])); + sum[i] = nir_fmul(b, nir_channel(b, src0_vec, i), + nir_channel(b, src1_vec, i)); } - sum[3] = nir_channel(b, instr->src[1].src.ssa, instr->src[1].swizzle[3]); - - nir_ssa_def *val = nir_fadd(b, nir_fadd(b, sum[0], sum[1]), - nir_fadd(b, sum[2], sum[3])); + sum[3] = nir_channel(b, src1_vec, 3); - nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(val)); - nir_instr_remove(&instr->instr); - return true; + return nir_fadd(b, nir_fadd(b, sum[0], sum[1]), + nir_fadd(b, sum[2], sum[3])); } - case nir_op_unpack_double_2x32: - case nir_op_unpack_int_2x32: - return false; + case nir_op_unpack_64_2x32: + case nir_op_unpack_64_4x16: + case nir_op_unpack_32_2x16: + return NULL; LOWER_REDUCTION(nir_op_fdot, nir_op_fmul, nir_op_fadd); LOWER_REDUCTION(nir_op_ball_fequal, nir_op_feq, nir_op_iand); LOWER_REDUCTION(nir_op_ball_iequal, nir_op_ieq, nir_op_iand); LOWER_REDUCTION(nir_op_bany_fnequal, nir_op_fne, nir_op_ior); LOWER_REDUCTION(nir_op_bany_inequal, nir_op_ine, nir_op_ior); - LOWER_REDUCTION(nir_op_fall_equal, nir_op_seq, nir_op_fand); - LOWER_REDUCTION(nir_op_fany_nequal, nir_op_sne, nir_op_for); + LOWER_REDUCTION(nir_op_b8all_fequal, nir_op_feq8, nir_op_iand); + LOWER_REDUCTION(nir_op_b8all_iequal, nir_op_ieq8, nir_op_iand); + LOWER_REDUCTION(nir_op_b8any_fnequal, nir_op_fne8, nir_op_ior); + LOWER_REDUCTION(nir_op_b8any_inequal, nir_op_ine8, nir_op_ior); + LOWER_REDUCTION(nir_op_b16all_fequal, nir_op_feq16, nir_op_iand); + LOWER_REDUCTION(nir_op_b16all_iequal, nir_op_ieq16, nir_op_iand); + LOWER_REDUCTION(nir_op_b16any_fnequal, nir_op_fne16, nir_op_ior); + LOWER_REDUCTION(nir_op_b16any_inequal, nir_op_ine16, nir_op_ior); + LOWER_REDUCTION(nir_op_b32all_fequal, nir_op_feq32, nir_op_iand); + LOWER_REDUCTION(nir_op_b32all_iequal, nir_op_ieq32, nir_op_iand); + LOWER_REDUCTION(nir_op_b32any_fnequal, nir_op_fne32, nir_op_ior); + LOWER_REDUCTION(nir_op_b32any_inequal, nir_op_ine32, nir_op_ior); + LOWER_REDUCTION(nir_op_fall_equal, nir_op_seq, nir_op_fmin); + LOWER_REDUCTION(nir_op_fany_nequal, nir_op_sne, nir_op_fmax); default: break; } - if (instr->dest.dest.ssa.num_components == 1) - return false; + if (alu->dest.dest.ssa.num_components == 1) + return NULL; - unsigned num_components = instr->dest.dest.ssa.num_components; - nir_ssa_def *comps[] = { NULL, NULL, NULL, NULL }; + unsigned num_components = alu->dest.dest.ssa.num_components; + nir_ssa_def *comps[NIR_MAX_VEC_COMPONENTS] = { NULL }; - for (chan = 0; chan < 4; chan++) { - if (!(instr->dest.write_mask & (1 << chan))) + for (chan = 0; chan < NIR_MAX_VEC_COMPONENTS; chan++) { + if (!(alu->dest.write_mask & (1 << chan))) continue; - nir_alu_instr *lower = nir_alu_instr_create(b->shader, instr->op); + nir_alu_instr *lower = nir_alu_instr_create(b->shader, alu->op); for (i = 0; i < num_src; i++) { /* We only handle same-size-as-dest (input_sizes[] == 0) or scalar * args (input_sizes[] == 1). */ - assert(nir_op_infos[instr->op].input_sizes[i] < 2); - unsigned src_chan = (nir_op_infos[instr->op].input_sizes[i] == 1 ? + assert(nir_op_infos[alu->op].input_sizes[i] < 2); + unsigned src_chan = (nir_op_infos[alu->op].input_sizes[i] == 1 ? 0 : chan); - nir_alu_src_copy(&lower->src[i], &instr->src[i], lower); - for (int j = 0; j < 4; j++) - lower->src[i].swizzle[j] = instr->src[i].swizzle[src_chan]; + nir_alu_src_copy(&lower->src[i], &alu->src[i], lower); + for (int j = 0; j < NIR_MAX_VEC_COMPONENTS; j++) + lower->src[i].swizzle[j] = alu->src[i].swizzle[src_chan]; } - nir_alu_ssa_dest_init(lower, 1, instr->dest.dest.ssa.bit_size); - lower->dest.saturate = instr->dest.saturate; + nir_alu_ssa_dest_init(lower, 1, alu->dest.dest.ssa.bit_size); + lower->dest.saturate = alu->dest.saturate; comps[chan] = &lower->dest.dest.ssa; - lower->exact = instr->exact; + lower->exact = alu->exact; nir_builder_instr_insert(b, &lower->instr); } - nir_ssa_def *vec = nir_vec(b, comps, num_components); - - nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(vec)); - - nir_instr_remove(&instr->instr); - return true; -} - -static bool -nir_lower_alu_to_scalar_impl(nir_function_impl *impl) -{ - nir_builder builder; - nir_builder_init(&builder, impl); - bool progress = false; - - nir_foreach_block(block, impl) { - nir_foreach_instr_safe(instr, block) { - if (instr->type == nir_instr_type_alu) { - progress = lower_alu_instr_scalar(nir_instr_as_alu(instr), - &builder) || progress; - } - } - } - - nir_metadata_preserve(impl, nir_metadata_block_index | - nir_metadata_dominance); - - return progress; + return nir_vec(b, comps, num_components); } bool -nir_lower_alu_to_scalar(nir_shader *shader) +nir_lower_alu_to_scalar(nir_shader *shader, nir_instr_filter_cb cb, const void *_data) { - bool progress = false; - - nir_foreach_function(function, shader) { - if (function->impl) - progress = nir_lower_alu_to_scalar_impl(function->impl) || progress; - } - - return progress; + struct alu_to_scalar_data data = { + .cb = cb, + .data = _data, + }; + + return nir_shader_lower_instructions(shader, + inst_is_vector_alu, + lower_alu_instr_scalar, + &data); }