From 1fdd3ef8a05e3281ca295137bddb00819f83ed52 Mon Sep 17 00:00:00 2001 From: Lioncash Date: Sat, 13 Apr 2019 02:08:20 -0400 Subject: [PATCH] A64: Implement FMLA/FMLS' half-precision scalar indexed variants --- src/frontend/A64/decoder/a64.inc | 4 +- .../impl/simd_scalar_x_indexed_element.cpp | 42 +++++++++++++++++++ 2 files changed, 44 insertions(+), 2 deletions(-) diff --git a/src/frontend/A64/decoder/a64.inc b/src/frontend/A64/decoder/a64.inc index ee133667..720dd45d 100644 --- a/src/frontend/A64/decoder/a64.inc +++ b/src/frontend/A64/decoder/a64.inc @@ -538,9 +538,9 @@ INST(FCVTZU_fix_1, "FCVTZU (vector, fixed-point)", "01111 INST(SQDMULL_elt_1, "SQDMULL, SQDMULL2 (by element)", "01011111zzLMmmmm1011H0nnnnnddddd") INST(SQDMULH_elt_1, "SQDMULH (by element)", "01011111zzLMmmmm1100H0nnnnnddddd") INST(SQRDMULH_elt_1, "SQRDMULH (by element)", "01011111zzLMmmmm1101H0nnnnnddddd") -//INST(FMLA_elt_1, "FMLA (by element)", "0101111100LMmmmm0001H0nnnnnddddd") +INST(FMLA_elt_1, "FMLA (by element)", "0101111100LMmmmm0001H0nnnnnddddd") INST(FMLA_elt_2, "FMLA (by element)", "010111111zLMmmmm0001H0nnnnnddddd") -//INST(FMLS_elt_1, "FMLS (by element)", "0101111100LMmmmm0101H0nnnnnddddd") +INST(FMLS_elt_1, "FMLS (by element)", "0101111100LMmmmm0101H0nnnnnddddd") INST(FMLS_elt_2, "FMLS (by element)", "010111111zLMmmmm0101H0nnnnnddddd") //INST(FMUL_elt_1, "FMUL (by element)", "0101111100LMmmmm1001H0nnnnnddddd") INST(FMUL_elt_2, "FMUL (by element)", "010111111zLMmmmm1001H0nnnnnddddd") diff --git a/src/frontend/A64/translate/impl/simd_scalar_x_indexed_element.cpp b/src/frontend/A64/translate/impl/simd_scalar_x_indexed_element.cpp index 0d475c40..350eb2fe 100644 --- a/src/frontend/A64/translate/impl/simd_scalar_x_indexed_element.cpp +++ b/src/frontend/A64/translate/impl/simd_scalar_x_indexed_element.cpp @@ -58,12 +58,54 @@ bool MultiplyByElement(TranslatorVisitor& v, bool sz, Imm<1> L, Imm<1> M, Imm<4> v.V_scalar(esize, Vd, result); return true; } + +bool MultiplyByElementHalfPrecision(TranslatorVisitor& v, Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H, + Vec Vn, Vec Vd, ExtraBehavior extra_behavior) { + const size_t esize = 16; + const size_t idxsize = H == 1 ? 128 : 64; + const size_t index = concatenate(H, L, M).ZeroExtend(); + + const auto Vm = Vmlo.ZeroExtend(); + const IR::U16 element = v.ir.VectorGetElement(esize, v.V(idxsize, Vm), index); + const IR::U16 result = [&]() -> IR::U16 { + IR::U16 operand1 = v.V_scalar(esize, Vn); + + // TODO: Currently we don't implement half-precision paths + // for regular multiplication and extended multiplication. + + if (extra_behavior == ExtraBehavior::None) { + UNIMPLEMENTED(); + } + + if (extra_behavior == ExtraBehavior::MultiplyExtended) { + UNIMPLEMENTED(); + } + + if (extra_behavior == ExtraBehavior::Subtract) { + operand1 = v.ir.FPNeg(operand1); + } + + const IR::U16 operand2 = v.V_scalar(esize, Vd); + return v.ir.FPMulAdd(operand2, operand1, element, true); + }(); + + v.V_scalar(esize, Vd, result); + return true; +} } // Anonymous namespace +bool TranslatorVisitor::FMLA_elt_1(Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H, Vec Vn, Vec Vd) { + return MultiplyByElementHalfPrecision(*this, L, M, Vmlo, H, Vn, Vd, ExtraBehavior::Accumulate); +} + bool TranslatorVisitor::FMLA_elt_2(bool sz, Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H, Vec Vn, Vec Vd) { return MultiplyByElement(*this, sz, L, M, Vmlo, H, Vn, Vd, ExtraBehavior::Accumulate); } +bool TranslatorVisitor::FMLS_elt_1(Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H, Vec Vn, Vec Vd) { + return MultiplyByElementHalfPrecision(*this, L, M, Vmlo, H, Vn, Vd, ExtraBehavior::Subtract); +} + bool TranslatorVisitor::FMLS_elt_2(bool sz, Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H, Vec Vn, Vec Vd) { return MultiplyByElement(*this, sz, L, M, Vmlo, H, Vn, Vd, ExtraBehavior::Subtract); }