From df645fa369d12be4d5e0fd9e4f6d4455caf2f4c3 Mon Sep 17 00:00:00 2001 From: Rhys Perry Date: Sat, 9 Nov 2019 20:51:45 +0000 Subject: [PATCH] aco: implement VK_KHR_shader_float_controls MIME-Version: 1.0 Content-Type: text/plain; charset=utf8 Content-Transfer-Encoding: 8bit This actually supports more of the extension than the LLVM backend but we can't enable it because ACO doesn't work with all stages yet. With more of it enabled, some CTS tests fail because our 64-bit sqrt is very imprecise. I can't find any precision requirements for it anywhere, so I'm thinking it might be a CTS issue. Signed-off-by: Rhys Perry Reviewed-by: Daniel Schürmann --- src/amd/compiler/aco_assembler.cpp | 2 +- .../compiler/aco_instruction_selection.cpp | 143 +++++++++++++++++- .../aco_instruction_selection_setup.cpp | 1 - src/amd/compiler/aco_ir.h | 51 +++++++ src/amd/compiler/aco_lower_to_hw_instr.cpp | 16 ++ src/amd/compiler/aco_opt_value_numbering.cpp | 3 +- src/amd/compiler/aco_optimizer.cpp | 32 ++-- src/amd/vulkan/radv_device.c | 2 + src/amd/vulkan/radv_extensions.py | 2 +- 9 files changed, 227 insertions(+), 25 deletions(-) diff --git a/src/amd/compiler/aco_assembler.cpp b/src/amd/compiler/aco_assembler.cpp index ee575e882c9..9b76ba740dd 100644 --- a/src/amd/compiler/aco_assembler.cpp +++ b/src/amd/compiler/aco_assembler.cpp @@ -105,7 +105,7 @@ void emit_instruction(asm_context& ctx, std::vector& out, Instruction* encoding |= !instr->definitions.empty() && !(instr->definitions[0].physReg() == scc) ? instr->definitions[0].physReg() << 16 : - !instr->operands.empty() && !(instr->operands[0].physReg() == scc) ? + !instr->operands.empty() && instr->operands[0].physReg() <= 127 ? instr->operands[0].physReg() << 16 : 0; encoding |= sopk->imm; out.push_back(encoding); diff --git a/src/amd/compiler/aco_instruction_selection.cpp b/src/amd/compiler/aco_instruction_selection.cpp index b094340b02f..63fd36f3724 100644 --- a/src/amd/compiler/aco_instruction_selection.cpp +++ b/src/amd/compiler/aco_instruction_selection.cpp @@ -647,6 +647,61 @@ void emit_bcsel(isel_context *ctx, nir_alu_instr *instr, Temp dst) bld.sop2(aco_opcode::s_andn2_b64, bld.def(s2), bld.def(s1, scc), els, cond)); } +void emit_scaled_op(isel_context *ctx, Builder& bld, Definition dst, Temp val, + aco_opcode op, uint32_t undo) +{ + /* multiply by 16777216 to handle denormals */ + Temp is_denormal = bld.vopc(aco_opcode::v_cmp_class_f32, bld.hint_vcc(bld.def(s2)), + as_vgpr(ctx, val), bld.copy(bld.def(v1), Operand((1u << 7) | (1u << 4)))); + Temp scaled = bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), Operand(0x4b800000u), val); + scaled = bld.vop1(op, bld.def(v1), scaled); + scaled = bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), Operand(undo), scaled); + + Temp not_scaled = bld.vop1(op, bld.def(v1), val); + + bld.vop2(aco_opcode::v_cndmask_b32, dst, not_scaled, scaled, is_denormal); +} + +void emit_rcp(isel_context *ctx, Builder& bld, Definition dst, Temp val) +{ + if (ctx->block->fp_mode.denorm32 == 0) { + bld.vop1(aco_opcode::v_rcp_f32, dst, val); + return; + } + + emit_scaled_op(ctx, bld, dst, val, aco_opcode::v_rcp_f32, 0x4b800000u); +} + +void emit_rsq(isel_context *ctx, Builder& bld, Definition dst, Temp val) +{ + if (ctx->block->fp_mode.denorm32 == 0) { + bld.vop1(aco_opcode::v_rsq_f32, dst, val); + return; + } + + emit_scaled_op(ctx, bld, dst, val, aco_opcode::v_rsq_f32, 0x45800000u); +} + +void emit_sqrt(isel_context *ctx, Builder& bld, Definition dst, Temp val) +{ + if (ctx->block->fp_mode.denorm32 == 0) { + bld.vop1(aco_opcode::v_sqrt_f32, dst, val); + return; + } + + emit_scaled_op(ctx, bld, dst, val, aco_opcode::v_sqrt_f32, 0x39800000u); +} + +void emit_log2(isel_context *ctx, Builder& bld, Definition dst, Temp val) +{ + if (ctx->block->fp_mode.denorm32 == 0) { + bld.vop1(aco_opcode::v_log_f32, dst, val); + return; + } + + emit_scaled_op(ctx, bld, dst, val, aco_opcode::v_log_f32, 0xc1c00000u); +} + void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr) { if (!instr->dest.dest.is_ssa) { @@ -1399,7 +1454,7 @@ void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr) } case nir_op_frsq: { if (dst.size() == 1) { - emit_vop1_instruction(ctx, instr, aco_opcode::v_rsq_f32, dst); + emit_rsq(ctx, bld, Definition(dst), get_alu_src(ctx, instr->src[0])); } else if (dst.size() == 2) { emit_vop1_instruction(ctx, instr, aco_opcode::v_rsq_f64, dst); } else { @@ -1412,8 +1467,12 @@ void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr) case nir_op_fneg: { Temp src = get_alu_src(ctx, instr->src[0]); if (dst.size() == 1) { + if (ctx->block->fp_mode.must_flush_denorms32) + src = bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), Operand(0x3f800000u), as_vgpr(ctx, src)); bld.vop2(aco_opcode::v_xor_b32, Definition(dst), Operand(0x80000000u), as_vgpr(ctx, src)); } else if (dst.size() == 2) { + if (ctx->block->fp_mode.must_flush_denorms16_64) + src = bld.vop3(aco_opcode::v_mul_f64, bld.def(v2), Operand(0x3FF0000000000000lu), as_vgpr(ctx, src)); Temp upper = bld.tmp(v1), lower = bld.tmp(v1); bld.pseudo(aco_opcode::p_split_vector, Definition(lower), Definition(upper), src); upper = bld.vop2(aco_opcode::v_xor_b32, bld.def(v1), Operand(0x80000000u), upper); @@ -1428,8 +1487,12 @@ void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr) case nir_op_fabs: { Temp src = get_alu_src(ctx, instr->src[0]); if (dst.size() == 1) { + if (ctx->block->fp_mode.must_flush_denorms32) + src = bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), Operand(0x3f800000u), as_vgpr(ctx, src)); bld.vop2(aco_opcode::v_and_b32, Definition(dst), Operand(0x7FFFFFFFu), as_vgpr(ctx, src)); } else if (dst.size() == 2) { + if (ctx->block->fp_mode.must_flush_denorms16_64) + src = bld.vop3(aco_opcode::v_mul_f64, bld.def(v2), Operand(0x3FF0000000000000lu), as_vgpr(ctx, src)); Temp upper = bld.tmp(v1), lower = bld.tmp(v1); bld.pseudo(aco_opcode::p_split_vector, Definition(lower), Definition(upper), src); upper = bld.vop2(aco_opcode::v_and_b32, bld.def(v1), Operand(0x7FFFFFFFu), upper); @@ -1458,7 +1521,7 @@ void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr) } case nir_op_flog2: { if (dst.size() == 1) { - emit_vop1_instruction(ctx, instr, aco_opcode::v_log_f32, dst); + emit_log2(ctx, bld, Definition(dst), get_alu_src(ctx, instr->src[0])); } else { fprintf(stderr, "Unimplemented NIR instr bit size: "); nir_print_instr(&instr->instr, stderr); @@ -1468,7 +1531,7 @@ void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr) } case nir_op_frcp: { if (dst.size() == 1) { - emit_vop1_instruction(ctx, instr, aco_opcode::v_rcp_f32, dst); + emit_rcp(ctx, bld, Definition(dst), get_alu_src(ctx, instr->src[0])); } else if (dst.size() == 2) { emit_vop1_instruction(ctx, instr, aco_opcode::v_rcp_f64, dst); } else { @@ -1490,7 +1553,7 @@ void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr) } case nir_op_fsqrt: { if (dst.size() == 1) { - emit_vop1_instruction(ctx, instr, aco_opcode::v_sqrt_f32, dst); + emit_sqrt(ctx, bld, Definition(dst), get_alu_src(ctx, instr->src[0])); } else if (dst.size() == 2) { emit_vop1_instruction(ctx, instr, aco_opcode::v_sqrt_f64, dst); } else { @@ -2040,8 +2103,12 @@ void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr) Temp src0 = bld.tmp(v1); Temp src1 = bld.tmp(v1); bld.pseudo(aco_opcode::p_split_vector, Definition(src0), Definition(src1), src); - bld.vop3(aco_opcode::v_cvt_pkrtz_f16_f32, Definition(dst), src0, src1); - + if (!ctx->block->fp_mode.care_about_round32 || ctx->block->fp_mode.round32 == fp_round_tz) + bld.vop3(aco_opcode::v_cvt_pkrtz_f16_f32, Definition(dst), src0, src1); + else + bld.vop3(aco_opcode::v_cvt_pk_u16_u32, Definition(dst), + bld.vop1(aco_opcode::v_cvt_f32_f16, bld.def(v1), src0), + bld.vop1(aco_opcode::v_cvt_f32_f16, bld.def(v1), src1)); } else { fprintf(stderr, "Unimplemented NIR instr bit size: "); nir_print_instr(&instr->instr, stderr); @@ -2074,7 +2141,8 @@ void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr) break; } case nir_op_fquantize2f16: { - Temp f16 = bld.vop1(aco_opcode::v_cvt_f16_f32, bld.def(v1), get_alu_src(ctx, instr->src[0])); + Temp src = get_alu_src(ctx, instr->src[0]); + Temp f16 = bld.vop1(aco_opcode::v_cvt_f16_f32, bld.def(v1), src); Temp mask = bld.copy(bld.def(s1), Operand(0x36Fu)); /* value is NOT negative/positive denormal value */ @@ -2083,7 +2151,12 @@ void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr) Temp f32 = bld.vop1(aco_opcode::v_cvt_f32_f16, bld.def(v1), f16); - bld.vop2(aco_opcode::v_cndmask_b32, Definition(dst), Operand(0u), f32, cmp_res); + if (ctx->block->fp_mode.preserve_signed_zero_inf_nan32) { + Temp copysign_0 = bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), Operand(0u), as_vgpr(ctx, src)); + bld.vop2(aco_opcode::v_cndmask_b32, Definition(dst), copysign_0, f32, cmp_res); + } else { + bld.vop2(aco_opcode::v_cndmask_b32, Definition(dst), Operand(0u), f32, cmp_res); + } break; } case nir_op_bfm: { @@ -7593,6 +7666,56 @@ void handle_bc_optimize(isel_context *ctx) } } +void setup_fp_mode(isel_context *ctx, nir_shader *shader) +{ + Program *program = ctx->program; + + unsigned float_controls = shader->info.float_controls_execution_mode; + + program->next_fp_mode.preserve_signed_zero_inf_nan32 = + float_controls & FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP32; + program->next_fp_mode.preserve_signed_zero_inf_nan16_64 = + float_controls & (FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP16 | + FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP64); + + program->next_fp_mode.must_flush_denorms32 = + float_controls & FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP32; + program->next_fp_mode.must_flush_denorms16_64 = + float_controls & (FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP16 | + FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP64); + + program->next_fp_mode.care_about_round32 = + float_controls & (FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32 | FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32); + + program->next_fp_mode.care_about_round16_64 = + float_controls & (FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16 | FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64 | + FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16 | FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64); + + /* default to preserving fp16 and fp64 denorms, since it's free */ + if (program->next_fp_mode.must_flush_denorms16_64) + program->next_fp_mode.denorm16_64 = 0; + else + program->next_fp_mode.denorm16_64 = fp_denorm_keep; + + /* preserving fp32 denorms is expensive, so only do it if asked */ + if (float_controls & FLOAT_CONTROLS_DENORM_PRESERVE_FP32) + program->next_fp_mode.denorm32 = fp_denorm_keep; + else + program->next_fp_mode.denorm32 = 0; + + if (float_controls & FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32) + program->next_fp_mode.round32 = fp_round_tz; + else + program->next_fp_mode.round32 = fp_round_ne; + + if (float_controls & (FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16 | FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64)) + program->next_fp_mode.round16_64 = fp_round_tz; + else + program->next_fp_mode.round16_64 = fp_round_ne; + + ctx->block->fp_mode = program->next_fp_mode; +} + void select_program(Program *program, unsigned shader_count, struct nir_shader *const *shaders, @@ -7606,6 +7729,8 @@ void select_program(Program *program, nir_shader *nir = shaders[i]; init_context(&ctx, nir); + setup_fp_mode(&ctx, nir); + if (!i) { add_startpgm(&ctx); /* needs to be after init_context() for FS */ append_logical_start(ctx.block); @@ -7648,6 +7773,8 @@ void select_program(Program *program, ralloc_free(ctx.divergent_vals); } + program->config->float_mode = program->blocks[0].fp_mode.val; + append_logical_end(ctx.block); ctx.block->kind |= block_kind_uniform; Builder bld(ctx.program, ctx.block); diff --git a/src/amd/compiler/aco_instruction_selection_setup.cpp b/src/amd/compiler/aco_instruction_selection_setup.cpp index cdc8103497b..807ce746868 100644 --- a/src/amd/compiler/aco_instruction_selection_setup.cpp +++ b/src/amd/compiler/aco_instruction_selection_setup.cpp @@ -1360,7 +1360,6 @@ setup_isel_context(Program* program, scratch_size = std::max(scratch_size, shaders[i]->scratch_size); ctx.scratch_enabled = scratch_size > 0; ctx.program->config->scratch_bytes_per_wave = align(scratch_size * ctx.program->wave_size, 1024); - ctx.program->config->float_mode = V_00B028_FP_64_DENORMS; ctx.block = ctx.program->create_and_insert_block(); ctx.block->loop_nest_depth = 0; diff --git a/src/amd/compiler/aco_ir.h b/src/amd/compiler/aco_ir.h index a6fe846c74d..59e77feffe5 100644 --- a/src/amd/compiler/aco_ir.h +++ b/src/amd/compiler/aco_ir.h @@ -110,6 +110,53 @@ enum barrier_interaction { barrier_count = 4, }; +enum fp_round { + fp_round_ne = 0, + fp_round_pi = 1, + fp_round_ni = 2, + fp_round_tz = 3, +}; + +enum fp_denorm { + /* Note that v_rcp_f32, v_exp_f32, v_log_f32, v_sqrt_f32, v_rsq_f32 and + * v_mad_f32/v_madak_f32/v_madmk_f32/v_mac_f32 always flush denormals. */ + fp_denorm_flush = 0x0, + fp_denorm_keep = 0x3, +}; + +struct float_mode { + /* matches encoding of the MODE register */ + union { + struct { + fp_round round32:2; + fp_round round16_64:2; + unsigned denorm32:2; + unsigned denorm16_64:2; + }; + uint8_t val = 0; + }; + /* if false, optimizations which may remove infs/nan/-0.0 can be done */ + bool preserve_signed_zero_inf_nan32:1; + bool preserve_signed_zero_inf_nan16_64:1; + /* if false, optimizations which may remove denormal flushing can be done */ + bool must_flush_denorms32:1; + bool must_flush_denorms16_64:1; + bool care_about_round32:1; + bool care_about_round16_64:1; + + /* Returns true if instructions using the mode "other" can safely use the + * current one instead. */ + bool canReplace(float_mode other) const noexcept { + return val == other.val && + (preserve_signed_zero_inf_nan32 || !other.preserve_signed_zero_inf_nan32) && + (preserve_signed_zero_inf_nan16_64 || !other.preserve_signed_zero_inf_nan16_64) && + (must_flush_denorms32 || !other.must_flush_denorms32) && + (must_flush_denorms16_64 || !other.must_flush_denorms16_64) && + (care_about_round32 || !other.care_about_round32) && + (care_about_round16_64 || !other.care_about_round16_64); + } +}; + constexpr Format asVOP3(Format format) { return (Format) ((uint32_t) Format::VOP3 | (uint32_t) format); }; @@ -1019,6 +1066,7 @@ struct RegisterDemand { /* CFG */ struct Block { + float_mode fp_mode; unsigned index; unsigned offset = 0; std::vector> instructions; @@ -1086,6 +1134,7 @@ static constexpr Stage geometry_gs = sw_gs | hw_gs; class Program final { public: + float_mode next_fp_mode; std::vector blocks; RegisterDemand max_reg_demand = RegisterDemand(); uint16_t num_waves = 0; @@ -1133,11 +1182,13 @@ public: Block* create_and_insert_block() { blocks.emplace_back(blocks.size()); + blocks.back().fp_mode = next_fp_mode; return &blocks.back(); } Block* insert_block(Block&& block) { block.index = blocks.size(); + block.fp_mode = next_fp_mode; blocks.emplace_back(std::move(block)); return &blocks.back(); } diff --git a/src/amd/compiler/aco_lower_to_hw_instr.cpp b/src/amd/compiler/aco_lower_to_hw_instr.cpp index 1502619b9db..8db54064202 100644 --- a/src/amd/compiler/aco_lower_to_hw_instr.cpp +++ b/src/amd/compiler/aco_lower_to_hw_instr.cpp @@ -592,6 +592,22 @@ void lower_to_hw_instr(Program* program) ctx.program = program; Builder bld(program, &ctx.instructions); + bool set_mode = i == 0 && block->fp_mode.val != program->config->float_mode; + for (unsigned pred : block->linear_preds) { + if (program->blocks[pred].fp_mode.val != block->fp_mode.val) { + set_mode = true; + break; + } + } + if (set_mode) { + /* only allow changing modes at top-level blocks so this doesn't break + * the "jump over empty blocks" optimization */ + assert(block->kind & block_kind_top_level); + uint32_t mode = block->fp_mode.val; + /* "((size - 1) << 11) | register" (MODE is encoded as register 1) */ + bld.sopk(aco_opcode::s_setreg_imm32_b32, Operand(mode), (7 << 11) | 1); + } + for (size_t j = 0; j < block->instructions.size(); j++) { aco_ptr& instr = block->instructions[j]; aco_ptr mov; diff --git a/src/amd/compiler/aco_opt_value_numbering.cpp b/src/amd/compiler/aco_opt_value_numbering.cpp index 803249637d5..40823da3c36 100644 --- a/src/amd/compiler/aco_opt_value_numbering.cpp +++ b/src/amd/compiler/aco_opt_value_numbering.cpp @@ -303,7 +303,8 @@ void process_block(vn_ctx& ctx, Block& block) Instruction* orig_instr = res.first->first; assert(instr->definitions.size() == orig_instr->definitions.size()); /* check if the original instruction dominates the current one */ - if (dominates(ctx, res.first->second, block.index)) { + if (dominates(ctx, res.first->second, block.index) && + ctx.program->blocks[res.first->second].fp_mode.canReplace(block.fp_mode)) { for (unsigned i = 0; i < instr->definitions.size(); i++) { assert(instr->definitions[i].regClass() == orig_instr->definitions[i].regClass()); ctx.renames[instr->definitions[i].tempId()] = orig_instr->definitions[i].getTemp(); diff --git a/src/amd/compiler/aco_optimizer.cpp b/src/amd/compiler/aco_optimizer.cpp index 5b4fcf75126..7b66aa1eeb3 100644 --- a/src/amd/compiler/aco_optimizer.cpp +++ b/src/amd/compiler/aco_optimizer.cpp @@ -548,7 +548,7 @@ bool parse_base_offset(opt_ctx &ctx, Instruction* instr, unsigned op_index, Temp return false; } -void label_instruction(opt_ctx &ctx, aco_ptr& instr) +void label_instruction(opt_ctx &ctx, Block& block, aco_ptr& instr) { if (instr->isSALU() || instr->isVALU() || instr->format == Format::PSEUDO) { ASSERTED bool all_const = false; @@ -888,7 +888,8 @@ void label_instruction(opt_ctx &ctx, aco_ptr& instr) ctx.info[instr->operands[i].tempId()].set_omod4(); } else if (instr->operands[!i].constantValue() == 0x3f000000) { /* 0.5 */ ctx.info[instr->operands[i].tempId()].set_omod5(); - } else if (instr->operands[!i].constantValue() == 0x3f800000) { /* 1.0 */ + } else if (instr->operands[!i].constantValue() == 0x3f800000 && + !block.fp_mode.must_flush_denorms32) { /* 1.0 */ ctx.info[instr->definitions[0].tempId()].set_temp(instr->operands[i].getTemp()); } else { continue; @@ -1892,7 +1893,7 @@ void apply_sgprs(opt_ctx &ctx, aco_ptr& instr) } } -bool apply_omod_clamp(opt_ctx &ctx, aco_ptr& instr) +bool apply_omod_clamp(opt_ctx &ctx, Block& block, aco_ptr& instr) { /* check if we could apply omod on predecessor */ if (instr->opcode == aco_opcode::v_mul_f32) { @@ -1959,18 +1960,21 @@ bool apply_omod_clamp(opt_ctx &ctx, aco_ptr& instr) } } + /* omod has no effect if denormals are enabled */ + bool can_use_omod = block.fp_mode.denorm32 == 0; + /* apply omod / clamp modifiers if the def is used only once and the instruction can have modifiers */ if (!instr->definitions.empty() && ctx.uses[instr->definitions[0].tempId()] == 1 && can_use_VOP3(instr) && instr_info.can_use_output_modifiers[(int)instr->opcode]) { - if(ctx.info[instr->definitions[0].tempId()].is_omod2()) { + if (can_use_omod && ctx.info[instr->definitions[0].tempId()].is_omod2()) { to_VOP3(ctx, instr); static_cast(instr.get())->omod = 1; ctx.info[instr->definitions[0].tempId()].set_omod_success(instr.get()); - } else if (ctx.info[instr->definitions[0].tempId()].is_omod4()) { + } else if (can_use_omod && ctx.info[instr->definitions[0].tempId()].is_omod4()) { to_VOP3(ctx, instr); static_cast(instr.get())->omod = 2; ctx.info[instr->definitions[0].tempId()].set_omod_success(instr.get()); - } else if (ctx.info[instr->definitions[0].tempId()].is_omod5()) { + } else if (can_use_omod && ctx.info[instr->definitions[0].tempId()].is_omod5()) { to_VOP3(ctx, instr); static_cast(instr.get())->omod = 3; ctx.info[instr->definitions[0].tempId()].set_omod_success(instr.get()); @@ -1987,7 +1991,7 @@ bool apply_omod_clamp(opt_ctx &ctx, aco_ptr& instr) // TODO: we could possibly move the whole label_instruction pass to combine_instruction: // this would mean that we'd have to fix the instruction uses while value propagation -void combine_instruction(opt_ctx &ctx, aco_ptr& instr) +void combine_instruction(opt_ctx &ctx, Block& block, aco_ptr& instr) { if (instr->definitions.empty() || !ctx.uses[instr->definitions[0].tempId()]) return; @@ -1995,7 +1999,7 @@ void combine_instruction(opt_ctx &ctx, aco_ptr& instr) if (instr->isVALU()) { if (can_apply_sgprs(instr)) apply_sgprs(ctx, instr); - if (apply_omod_clamp(ctx, instr)) + if (apply_omod_clamp(ctx, block, instr)) return; } @@ -2048,9 +2052,11 @@ void combine_instruction(opt_ctx &ctx, aco_ptr& instr) return; } /* combine mul+add -> mad */ - else if (instr->opcode == aco_opcode::v_add_f32 || - instr->opcode == aco_opcode::v_sub_f32 || - instr->opcode == aco_opcode::v_subrev_f32) { + else if ((instr->opcode == aco_opcode::v_add_f32 || + instr->opcode == aco_opcode::v_sub_f32 || + instr->opcode == aco_opcode::v_subrev_f32) && + block.fp_mode.denorm32 == 0 && !block.fp_mode.preserve_signed_zero_inf_nan32) { + //TODO: we could use fma instead when denormals are enabled if the NIR isn't marked as precise uint32_t uses_src0 = UINT32_MAX; uint32_t uses_src1 = UINT32_MAX; @@ -2394,7 +2400,7 @@ void optimize(Program* program) /* 1. Bottom-Up DAG pass (forward) to label all ssa-defs */ for (Block& block : program->blocks) { for (aco_ptr& instr : block.instructions) - label_instruction(ctx, instr); + label_instruction(ctx, block, instr); } ctx.uses = std::move(dead_code_analysis(program)); @@ -2402,7 +2408,7 @@ void optimize(Program* program) /* 2. Combine v_mad, omod, clamp and propagate sgpr on VALU instructions */ for (Block& block : program->blocks) { for (aco_ptr& instr : block.instructions) - combine_instruction(ctx, instr); + combine_instruction(ctx, block, instr); } /* 3. Top-Down DAG pass (backward) to select instructions (includes DCE) */ diff --git a/src/amd/vulkan/radv_device.c b/src/amd/vulkan/radv_device.c index 4775609629f..b561980c123 100644 --- a/src/amd/vulkan/radv_device.c +++ b/src/amd/vulkan/radv_device.c @@ -1557,6 +1557,8 @@ void radv_GetPhysicalDeviceProperties2( * support for changing the register. The same logic * applies for the rounding modes because they are * configured with the same config register. + * TODO: we can enable a lot of these for ACO when it + * supports all stages */ properties->shaderDenormFlushToZeroFloat32 = true; properties->shaderDenormPreserveFloat32 = false; diff --git a/src/amd/vulkan/radv_extensions.py b/src/amd/vulkan/radv_extensions.py index 587e9820844..a4983ba0f61 100644 --- a/src/amd/vulkan/radv_extensions.py +++ b/src/amd/vulkan/radv_extensions.py @@ -89,7 +89,7 @@ EXTENSIONS = [ Extension('VK_KHR_shader_atomic_int64', 1, 'LLVM_VERSION_MAJOR >= 9'), Extension('VK_KHR_shader_clock', 1, True), Extension('VK_KHR_shader_draw_parameters', 1, True), - Extension('VK_KHR_shader_float_controls', 1, '!device->use_aco'), + Extension('VK_KHR_shader_float_controls', 1, True), Extension('VK_KHR_shader_float16_int8', 1, '!device->use_aco'), Extension('VK_KHR_spirv_1_4', 1, True), Extension('VK_KHR_storage_buffer_storage_class', 1, True), -- 2.30.2