From 224ff0afaaa64b2cac63c9f2a393ee9c205197ef Mon Sep 17 00:00:00 2001 From: Lioncash Date: Wed, 20 Mar 2019 14:05:41 -0400 Subject: [PATCH 1/3] A64: Implement SQRDMULH's by-index vector variant --- src/frontend/A64/decoder/a64.inc | 2 +- .../impl/simd_vector_x_indexed_element.cpp | 26 +++++++++++++++++++ 2 files changed, 27 insertions(+), 1 deletion(-) diff --git a/src/frontend/A64/decoder/a64.inc b/src/frontend/A64/decoder/a64.inc index 707f1a0d..550c87c6 100644 --- a/src/frontend/A64/decoder/a64.inc +++ b/src/frontend/A64/decoder/a64.inc @@ -874,7 +874,7 @@ INST(MUL_elt, "MUL (by element)", "0Q001 INST(SMULL_elt, "SMULL, SMULL2 (by element)", "0Q001111zzLMmmmm1010H0nnnnnddddd") //INST(SQDMULL_elt_2, "SQDMULL, SQDMULL2 (by element)", "0Q001111zzLMmmmm1011H0nnnnnddddd") INST(SQDMULH_elt_2, "SQDMULH (by element)", "0Q001111zzLMmmmm1100H0nnnnnddddd") -//INST(SQRDMULH_elt_2, "SQRDMULH (by element)", "0Q001111zzLMmmmm1101H0nnnnnddddd") +INST(SQRDMULH_elt_2, "SQRDMULH (by element)", "0Q001111zzLMmmmm1101H0nnnnnddddd") INST(SDOT_elt, "SDOT (by element)", "0Q001111zzLMmmmm1110H0nnnnnddddd") //INST(FMLA_elt_3, "FMLA (by element)", "0Q00111100LMmmmm0001H0nnnnnddddd") INST(FMLA_elt_4, "FMLA (by element)", "0Q0011111zLMmmmm0001H0nnnnnddddd") diff --git a/src/frontend/A64/translate/impl/simd_vector_x_indexed_element.cpp b/src/frontend/A64/translate/impl/simd_vector_x_indexed_element.cpp index d57fad04..83da7f74 100644 --- a/src/frontend/A64/translate/impl/simd_vector_x_indexed_element.cpp +++ b/src/frontend/A64/translate/impl/simd_vector_x_indexed_element.cpp @@ -246,6 +246,32 @@ bool TranslatorVisitor::SQDMULH_elt_2(bool Q, Imm<2> size, Imm<1> L, Imm<1> M, I return true; } +bool TranslatorVisitor::SQRDMULH_elt_2(bool Q, Imm<2> size, Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H, Vec Vn, Vec Vd) { + if (size == 0b00 || size == 0b11) { + return UnallocatedEncoding(); + } + + const size_t idxsize = H == 1 ? 128 : 64; + const size_t esize = 8 << size.ZeroExtend(); + const size_t datasize = Q ? 128 : 64; + const auto [index, Vmhi] = [=] { + if (size == 0b01) { + return std::make_pair(concatenate(H, L, M).ZeroExtend(), Imm<1>{0}); + } + + return std::make_pair(concatenate(H, L).ZeroExtend(), M); + }(); + + const IR::U128 operand1 = V(datasize, Vn); + const IR::U128 operand2 = V(idxsize, concatenate(Vmhi, Vmlo).ZeroExtend()); + const IR::U128 index_vector = ir.VectorBroadcast(esize, ir.VectorGetElement(esize, operand2, index)); + const IR::UpperAndLower multiply = ir.VectorSignedSaturatedDoublingMultiply(esize, operand1, index_vector); + const IR::U128 result = ir.VectorAdd(esize, multiply.upper, ir.VectorLogicalShiftRight(esize, multiply.lower, static_cast(esize - 1))); + + V(datasize, Vd, result); + return true; +} + bool TranslatorVisitor::SDOT_elt(bool Q, Imm<2> size, Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H, Vec Vn, Vec Vd) { return DotProduct(*this, Q, size, L, M, Vmlo, H, Vn, Vd, &IREmitter::SignExtendToWord); } From c043b831d51786784ea37f7a1f1467fc7da04dd0 Mon Sep 17 00:00:00 2001 From: Lioncash Date: Wed, 20 Mar 2019 15:19:49 -0400 Subject: [PATCH 2/3] A64: Implement SQDMULL{2}'s by-element variant --- src/frontend/A64/decoder/a64.inc | 2 +- src/frontend/A64/translate/impl/impl.h | 6 ++--- .../impl/simd_vector_x_indexed_element.cpp | 26 +++++++++++++++++++ 3 files changed, 30 insertions(+), 4 deletions(-) diff --git a/src/frontend/A64/decoder/a64.inc b/src/frontend/A64/decoder/a64.inc index 550c87c6..cfaf7e9b 100644 --- a/src/frontend/A64/decoder/a64.inc +++ b/src/frontend/A64/decoder/a64.inc @@ -872,7 +872,7 @@ INST(SMLSL_elt, "SMLSL, SMLSL2 (by element)", "0Q001 //INST(SQDMLSL_elt_2, "SQDMLSL, SQDMLSL2 (by element)", "0Q001111zzLMmmmm0111H0nnnnnddddd") INST(MUL_elt, "MUL (by element)", "0Q001111zzLMmmmm1000H0nnnnnddddd") INST(SMULL_elt, "SMULL, SMULL2 (by element)", "0Q001111zzLMmmmm1010H0nnnnnddddd") -//INST(SQDMULL_elt_2, "SQDMULL, SQDMULL2 (by element)", "0Q001111zzLMmmmm1011H0nnnnnddddd") +INST(SQDMULL_elt_2, "SQDMULL, SQDMULL2 (by element)", "0Q001111zzLMmmmm1011H0nnnnnddddd") INST(SQDMULH_elt_2, "SQDMULH (by element)", "0Q001111zzLMmmmm1100H0nnnnnddddd") INST(SQRDMULH_elt_2, "SQRDMULH (by element)", "0Q001111zzLMmmmm1101H0nnnnnddddd") INST(SDOT_elt, "SDOT (by element)", "0Q001111zzLMmmmm1110H0nnnnnddddd") diff --git a/src/frontend/A64/translate/impl/impl.h b/src/frontend/A64/translate/impl/impl.h index 1f502aa1..e9b2f024 100644 --- a/src/frontend/A64/translate/impl/impl.h +++ b/src/frontend/A64/translate/impl/impl.h @@ -937,12 +937,12 @@ struct TranslatorVisitor final { // Data Processing - FP and SIMD - SIMD vector x indexed element bool SMLAL_elt(bool Q, Imm<2> size, Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H, Vec Vn, Vec Vd); - bool SQDMLAL_elt_2(bool Q, Imm<2> size, Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H, Reg Rn, Vec Vd); + bool SQDMLAL_elt_2(bool Q, Imm<2> size, Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H, Vec Vn, Vec Vd); bool SMLSL_elt(bool Q, Imm<2> size, Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H, Vec Vn, Vec Vd); - bool SQDMLSL_elt_2(bool Q, Imm<2> size, Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H, Reg Rn, Vec Vd); + bool SQDMLSL_elt_2(bool Q, Imm<2> size, Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H, Vec Vn, Vec Vd); bool MUL_elt(bool Q, Imm<2> size, Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H, Vec Vn, Vec Vd); bool SMULL_elt(bool Q, Imm<2> size, Imm<1> L, Imm<1> M, Imm<4> Vm, Imm<1> H, Vec Vn, Vec Vd); - bool SQDMULL_elt_2(bool Q, Imm<2> size, Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H, Reg Rn, Vec Vd); + bool SQDMULL_elt_2(bool Q, Imm<2> size, Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H, Vec Vn, Vec Vd); bool SQDMULH_elt_2(bool Q, Imm<2> size, Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H, Vec Vn, Vec Vd); bool SQRDMULH_elt_2(bool Q, Imm<2> size, Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H, Vec Vn, Vec Vd); bool SDOT_elt(bool Q, Imm<2> size, Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H, Vec Vn, Vec Vd); diff --git a/src/frontend/A64/translate/impl/simd_vector_x_indexed_element.cpp b/src/frontend/A64/translate/impl/simd_vector_x_indexed_element.cpp index 83da7f74..dabcbc2c 100644 --- a/src/frontend/A64/translate/impl/simd_vector_x_indexed_element.cpp +++ b/src/frontend/A64/translate/impl/simd_vector_x_indexed_element.cpp @@ -221,6 +221,32 @@ bool TranslatorVisitor::SMULL_elt(bool Q, Imm<2> size, Imm<1> L, Imm<1> M, Imm<4 return MultiplyLong(*this, Q, size, L, M, Vmlo, H, Vn, Vd, ExtraBehavior::None, Signedness::Signed); } +bool TranslatorVisitor::SQDMULL_elt_2(bool Q, Imm<2> size, Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H, Vec Vn, Vec Vd) { + if (size == 0b00 || size == 0b11) { + return UnallocatedEncoding(); + } + + const size_t part = Q ? 1 : 0; + const size_t idxsize = H == 1 ? 128 : 64; + const size_t esize = 8 << size.ZeroExtend(); + const size_t datasize = 64; + const auto [index, Vmhi] = [=] { + if (size == 0b01) { + return std::make_pair(concatenate(H, L, M).ZeroExtend(), Imm<1>{0}); + } + + return std::make_pair(concatenate(H, L).ZeroExtend(), M); + }(); + + const IR::U128 operand1 = Vpart(datasize, Vn, part); + const IR::U128 operand2 = V(idxsize, concatenate(Vmhi, Vmlo).ZeroExtend()); + const IR::U128 index_vector = ir.VectorBroadcast(esize, ir.VectorGetElement(esize, operand2, index)); + const IR::U128 result = ir.VectorSignedSaturatedDoublingMultiplyLong(esize, operand1, index_vector); + + V(128, Vd, result); + return true; +} + bool TranslatorVisitor::SQDMULH_elt_2(bool Q, Imm<2> size, Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H, Vec Vn, Vec Vd) { if (size == 0b00 || size == 0b11) { return UnallocatedEncoding(); From 49b51e34f1772facb948d2e412196897252af9bf Mon Sep 17 00:00:00 2001 From: Lioncash Date: Wed, 20 Mar 2019 16:21:48 -0400 Subject: [PATCH 3/3] simd_vector_x_indexed_element: Deduplicate index and Vm operand construction --- .../impl/simd_vector_x_indexed_element.cpp | 40 ++++--------------- 1 file changed, 8 insertions(+), 32 deletions(-) diff --git a/src/frontend/A64/translate/impl/simd_vector_x_indexed_element.cpp b/src/frontend/A64/translate/impl/simd_vector_x_indexed_element.cpp index dabcbc2c..2fc62bab 100644 --- a/src/frontend/A64/translate/impl/simd_vector_x_indexed_element.cpp +++ b/src/frontend/A64/translate/impl/simd_vector_x_indexed_element.cpp @@ -139,13 +139,7 @@ bool MultiplyLong(TranslatorVisitor& v, bool Q, Imm<2> size, Imm<1> L, Imm<1> M, const size_t idxsize = H == 1 ? 128 : 64; const size_t esize = 8 << size.ZeroExtend(); const size_t datasize = 64; - const auto [index, Vmhi] = [=] { - if (size == 0b01) { - return std::make_pair(concatenate(H, L, M).ZeroExtend(), Imm<1>{0}); - } - - return std::make_pair(concatenate(H, L).ZeroExtend(), M); - }(); + const auto [index, Vm] = Combine(size, H, L, M, Vmlo); const auto extend_operands = [&](const IR::U128& lhs, const IR::U128& rhs) { if (sign == Signedness::Signed) { @@ -158,7 +152,7 @@ bool MultiplyLong(TranslatorVisitor& v, bool Q, Imm<2> size, Imm<1> L, Imm<1> M, }; const IR::U128 operand1 = v.Vpart(datasize, Vn, Q); - const IR::U128 operand2 = v.V(idxsize, concatenate(Vmhi, Vmlo).ZeroExtend()); + const IR::U128 operand2 = v.V(idxsize, Vm); const IR::U128 index_vector = v.ir.VectorBroadcast(esize, v.ir.VectorGetElement(esize, operand2, index)); const IR::U128 result = [&] { @@ -230,16 +224,10 @@ bool TranslatorVisitor::SQDMULL_elt_2(bool Q, Imm<2> size, Imm<1> L, Imm<1> M, I const size_t idxsize = H == 1 ? 128 : 64; const size_t esize = 8 << size.ZeroExtend(); const size_t datasize = 64; - const auto [index, Vmhi] = [=] { - if (size == 0b01) { - return std::make_pair(concatenate(H, L, M).ZeroExtend(), Imm<1>{0}); - } - - return std::make_pair(concatenate(H, L).ZeroExtend(), M); - }(); + const auto [index, Vm] = Combine(size, H, L, M, Vmlo); const IR::U128 operand1 = Vpart(datasize, Vn, part); - const IR::U128 operand2 = V(idxsize, concatenate(Vmhi, Vmlo).ZeroExtend()); + const IR::U128 operand2 = V(idxsize, Vm); const IR::U128 index_vector = ir.VectorBroadcast(esize, ir.VectorGetElement(esize, operand2, index)); const IR::U128 result = ir.VectorSignedSaturatedDoublingMultiplyLong(esize, operand1, index_vector); @@ -255,16 +243,10 @@ bool TranslatorVisitor::SQDMULH_elt_2(bool Q, Imm<2> size, Imm<1> L, Imm<1> M, I const size_t idxsize = H == 1 ? 128 : 64; const size_t esize = 8 << size.ZeroExtend(); const size_t datasize = Q ? 128 : 64; - const auto [index, Vmhi] = [=] { - if (size == 0b01) { - return std::make_pair(concatenate(H, L, M).ZeroExtend(), Imm<1>{0}); - } - - return std::make_pair(concatenate(H, L).ZeroExtend(), M); - }(); + const auto [index, Vm] = Combine(size, H, L, M, Vmlo); const IR::U128 operand1 = V(datasize, Vn); - const IR::U128 operand2 = V(idxsize, concatenate(Vmhi, Vmlo).ZeroExtend()); + const IR::U128 operand2 = V(idxsize, Vm); const IR::U128 index_vector = ir.VectorBroadcast(esize, ir.VectorGetElement(esize, operand2, index)); const IR::U128 result = ir.VectorSignedSaturatedDoublingMultiply(esize, operand1, index_vector).upper; @@ -280,16 +262,10 @@ bool TranslatorVisitor::SQRDMULH_elt_2(bool Q, Imm<2> size, Imm<1> L, Imm<1> M, const size_t idxsize = H == 1 ? 128 : 64; const size_t esize = 8 << size.ZeroExtend(); const size_t datasize = Q ? 128 : 64; - const auto [index, Vmhi] = [=] { - if (size == 0b01) { - return std::make_pair(concatenate(H, L, M).ZeroExtend(), Imm<1>{0}); - } - - return std::make_pair(concatenate(H, L).ZeroExtend(), M); - }(); + const auto [index, Vm] = Combine(size, H, L, M, Vmlo); const IR::U128 operand1 = V(datasize, Vn); - const IR::U128 operand2 = V(idxsize, concatenate(Vmhi, Vmlo).ZeroExtend()); + const IR::U128 operand2 = V(idxsize, Vm); const IR::U128 index_vector = ir.VectorBroadcast(esize, ir.VectorGetElement(esize, operand2, index)); const IR::UpperAndLower multiply = ir.VectorSignedSaturatedDoublingMultiply(esize, operand1, index_vector); const IR::U128 result = ir.VectorAdd(esize, multiply.upper, ir.VectorLogicalShiftRight(esize, multiply.lower, static_cast(esize - 1)));