From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: by sourceware.org (Postfix, from userid 7924) id 90984384DD00; Wed, 27 Sep 2023 08:28:58 +0000 (GMT) DKIM-Filter: OpenDKIM Filter v2.11.0 sourceware.org 90984384DD00 DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gcc.gnu.org; s=default; t=1695803338; bh=IEzUqgvml3fFErUDj+tPtARJCFs3KSgM2MD/roSqrj8=; h=From:To:Subject:Date:From; b=GetyU6S+1r8fohJ27zPyh2OlY85c5rXJfwPKdmOCDq7e4Nz62JlzEa6fgA739je2J CvFbFzrEm55EktkUHa9inofasyX+n90wtFS58vIl0svf+ggq6JoqhP45OQ+uSKb6X2 v9+Syp1ysQTmRQrZJGYFSEqT5nQAZia3y0LQON7c= MIME-Version: 1.0 Content-Transfer-Encoding: 7bit Content-Type: text/plain; charset="utf-8" From: Pan Li To: gcc-cvs@gcc.gnu.org Subject: [gcc r14-4292] RISC-V: Support FP roundeven auto-vectorization X-Act-Checkin: gcc X-Git-Author: Pan Li X-Git-Refname: refs/heads/master X-Git-Oldrev: 073849da3dfd5cabbfd4492a40a17b207b4a7630 X-Git-Newrev: fcbbf158ee90400877f81185bb40b5870dca1558 Message-Id: <20230927082858.90984384DD00@sourceware.org> Date: Wed, 27 Sep 2023 08:28:58 +0000 (GMT) List-Id: https://gcc.gnu.org/g:fcbbf158ee90400877f81185bb40b5870dca1558 commit r14-4292-gfcbbf158ee90400877f81185bb40b5870dca1558 Author: Pan Li Date: Wed Sep 27 16:13:43 2023 +0800 RISC-V: Support FP roundeven auto-vectorization This patch would like to support auto-vectorization for the roundeven API in math.h. It depends on the -ffast-math option. When we would like to call roundeven like v2 = roundeven (v1), we will convert it into below insns (reference the implementation of llvm). * vfcvt.x.f v3, v1, RNE * vfcvt.f.x v2, v3 However, the floating point value may not need the cvt as above if its mantissa is zero. For example single precision floating point below. +-----------+---------------+-----------------+ | raw float | binary layout | after roundeven | +-----------+---------------+-----------------+ | 8388607.5 | 0x4affffff | 8388608.0 | | 8388608.0 | 0x4b000000 | 8388608.0 | | 8388609.0 | 0x4b000001 | 8388609.0 | +-----------+---------------+-----------------+ All single floating point glte 8388608.0 will have all zero mantisaa. We leverage vmflt and mask to filter them out in vector and only do the cvt on mask. Befor this patch: math-roundeven-1.c:21:1: missed: couldn't vectorize loop ... .L3: flw fa0,0(s0) addi s0,s0,4 addi s1,s1,4 call roundeven fsw fa0,-4(s1) bne s0,s2,.L3 After this patch: ... fsrmi 0 // Rounding to nearest, ties to even .L4: vfabs.v v1,v2 vmflt.vf v0,v1,fa5 vfcvt.x.f.v v3,v2,v0.t vfcvt.f.x.v v1,v3,v0.t vfsgnj.vv v1,v1,v2 bne .L4 .L14: fsrm a6 ret Please note VLS mode is also involved in this patch and covered by the test cases. We will add more run test with zfa support later. gcc/ChangeLog: * config/riscv/autovec.md (roundeven2): New pattern. * config/riscv/riscv-protos.h (enum insn_flags): New enum type. (enum insn_type): Ditto. (expand_vec_roundeven): New func decl. * config/riscv/riscv-v.cc (expand_vec_roundeven): New func impl. gcc/testsuite/ChangeLog: * gcc.target/riscv/rvv/autovec/unop/math-roundeven-0.c: New test. * gcc.target/riscv/rvv/autovec/unop/math-roundeven-1.c: New test. * gcc.target/riscv/rvv/autovec/unop/math-roundeven-2.c: New test. * gcc.target/riscv/rvv/autovec/unop/math-roundeven-3.c: New test. * gcc.target/riscv/rvv/autovec/vls/math-roundeven-1.c: New test. Signed-off-by: Pan Li Diff: --- gcc/config/riscv/autovec.md | 10 ++++ gcc/config/riscv/riscv-protos.h | 5 ++ gcc/config/riscv/riscv-v.cc | 24 ++++++++++ .../riscv/rvv/autovec/unop/math-roundeven-0.c | 23 +++++++++ .../riscv/rvv/autovec/unop/math-roundeven-1.c | 23 +++++++++ .../riscv/rvv/autovec/unop/math-roundeven-2.c | 23 +++++++++ .../riscv/rvv/autovec/unop/math-roundeven-3.c | 25 ++++++++++ .../riscv/rvv/autovec/vls/math-roundeven-1.c | 56 ++++++++++++++++++++++ 8 files changed, 189 insertions(+) diff --git a/gcc/config/riscv/autovec.md b/gcc/config/riscv/autovec.md index 680a3374972..cd0cbdd2889 100644 --- a/gcc/config/riscv/autovec.md +++ b/gcc/config/riscv/autovec.md @@ -2271,3 +2271,13 @@ DONE; } ) + +(define_expand "roundeven2" + [(match_operand:V_VLSF 0 "register_operand") + (match_operand:V_VLSF 1 "register_operand")] + "TARGET_VECTOR && !flag_trapping_math && !flag_rounding_math" + { + riscv_vector::expand_vec_roundeven (operands[0], operands[1], mode, mode); + DONE; + } +) diff --git a/gcc/config/riscv/riscv-protos.h b/gcc/config/riscv/riscv-protos.h index 536e70bdcd3..368982a447b 100644 --- a/gcc/config/riscv/riscv-protos.h +++ b/gcc/config/riscv/riscv-protos.h @@ -259,6 +259,9 @@ enum insn_flags : unsigned int /* Means INSN has FRM operand and the value is FRM_RMM. */ FRM_RMM_P = 1 << 18, + + /* Means INSN has FRM operand and the value is FRM_RNE. */ + FRM_RNE_P = 1 << 19, }; enum insn_type : unsigned int @@ -303,6 +306,7 @@ enum insn_type : unsigned int UNARY_OP_TAMU_FRM_RUP = UNARY_OP_TAMU | FRM_RUP_P, UNARY_OP_TAMU_FRM_RDN = UNARY_OP_TAMU | FRM_RDN_P, UNARY_OP_TAMU_FRM_RMM = UNARY_OP_TAMU | FRM_RMM_P, + UNARY_OP_TAMU_FRM_RNE = UNARY_OP_TAMU | FRM_RNE_P, /* Binary operator. */ BINARY_OP = __NORMAL_OP | BINARY_OP_P, @@ -469,6 +473,7 @@ void expand_vec_nearbyint (rtx, rtx, machine_mode, machine_mode); void expand_vec_rint (rtx, rtx, machine_mode, machine_mode); void expand_vec_round (rtx, rtx, machine_mode, machine_mode); void expand_vec_trunc (rtx, rtx, machine_mode, machine_mode); +void expand_vec_roundeven (rtx, rtx, machine_mode, machine_mode); #endif bool sew64_scalar_helper (rtx *, rtx *, rtx, machine_mode, bool, void (*)(rtx *, rtx)); diff --git a/gcc/config/riscv/riscv-v.cc b/gcc/config/riscv/riscv-v.cc index 8992977a51d..359fb2ced8b 100644 --- a/gcc/config/riscv/riscv-v.cc +++ b/gcc/config/riscv/riscv-v.cc @@ -332,6 +332,8 @@ public: add_rounding_mode_operand (FRM_RDN); else if (m_insn_flags & FRM_RMM_P) add_rounding_mode_operand (FRM_RMM); + else if (m_insn_flags & FRM_RNE_P) + add_rounding_mode_operand (FRM_RNE); gcc_assert (insn_data[(int) icode].n_operands == m_opno); expand (icode, any_mem_p); @@ -3776,4 +3778,26 @@ expand_vec_trunc (rtx op_0, rtx op_1, machine_mode vec_fp_mode, emit_vec_copysign (op_0, op_0, op_1, vec_fp_mode); } +void +expand_vec_roundeven (rtx op_0, rtx op_1, machine_mode vec_fp_mode, + machine_mode vec_int_mode) +{ + /* Step-1: Get the abs float value for mask generation. */ + emit_vec_abs (op_0, op_1, vec_fp_mode); + + /* Step-2: Generate the mask on const fp. */ + rtx const_fp = get_fp_rounding_coefficient (GET_MODE_INNER (vec_fp_mode)); + rtx mask = emit_vec_float_cmp_mask (op_0, LT, const_fp, vec_fp_mode); + + /* Step-3: Convert to integer on mask, rounding to nearest, ties to even. */ + rtx tmp = gen_reg_rtx (vec_int_mode); + emit_vec_cvt_x_f (tmp, op_1, mask, UNARY_OP_TAMU_FRM_RNE, vec_fp_mode); + + /* Step-4: Convert to floating-point on mask for the rint result. */ + emit_vec_cvt_f_x (op_0, tmp, mask, UNARY_OP_TAMU_FRM_RNE, vec_fp_mode); + + /* Step-5: Retrieve the sign bit for -0.0. */ + emit_vec_copysign (op_0, op_0, op_1, vec_fp_mode); +} + } // namespace riscv_vector diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/unop/math-roundeven-0.c b/gcc/testsuite/gcc.target/riscv/rvv/autovec/unop/math-roundeven-0.c new file mode 100644 index 00000000000..ab65e372f0e --- /dev/null +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/unop/math-roundeven-0.c @@ -0,0 +1,23 @@ +/* { dg-do compile } */ +/* { dg-options "-march=rv64gcv_zvfh -mabi=lp64d -O3 -ftree-vectorize -fno-vect-cost-model -ffast-math -fno-schedule-insns -fno-schedule-insns2" } */ +/* { dg-final { check-function-bodies "**" "" } } */ + +#include "test-math.h" + +/* +** test__Float16___builtin_roundevenf16: +** frrm\s+[atx][0-9]+ +** ... +** fsrmi\s+0 +** ... +** vsetvli\s+[atx][0-9]+,\s*zero,\s*e16,\s*m1,\s*ta,\s*mu +** vfabs\.v\s+v[0-9]+,\s*v[0-9]+ +** vmflt\.vf\s+v0,\s*v[0-9]+,\s*[fa]+[0-9]+ +** vfcvt\.x\.f\.v\s+v[0-9]+,\s*v[0-9]+,\s*v0\.t +** vfcvt\.f\.x\.v\s+v[0-9]+,\s*v[0-9]+,\s*v0\.t +** vfsgnj\.vv\s+v[0-9]+,v[0-9]+,v[0-9]+ +** ... +** fsrm\s+[atx][0-9]+ +** ... +*/ +TEST_UNARY_CALL (_Float16, __builtin_roundevenf16) diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/unop/math-roundeven-1.c b/gcc/testsuite/gcc.target/riscv/rvv/autovec/unop/math-roundeven-1.c new file mode 100644 index 00000000000..fac85ed0895 --- /dev/null +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/unop/math-roundeven-1.c @@ -0,0 +1,23 @@ +/* { dg-do compile } */ +/* { dg-options "-march=rv64gcv -mabi=lp64d -O3 -ftree-vectorize -fno-vect-cost-model -ffast-math -fno-schedule-insns -fno-schedule-insns2" } */ +/* { dg-final { check-function-bodies "**" "" } } */ + +#include "test-math.h" + +/* +** test_float___builtin_roundevenf: +** frrm\s+[atx][0-9]+ +** ... +** fsrmi\s+0 +** ... +** vsetvli\s+[atx][0-9]+,\s*zero,\s*e32,\s*m1,\s*ta,\s*mu +** vfabs\.v\s+v[0-9]+,\s*v[0-9]+ +** vmflt\.vf\s+v0,\s*v[0-9]+,\s*[fa]+[0-9]+ +** vfcvt\.x\.f\.v\s+v[0-9]+,\s*v[0-9]+,\s*v0\.t +** vfcvt\.f\.x\.v\s+v[0-9]+,\s*v[0-9]+,\s*v0\.t +** vfsgnj\.vv\s+v[0-9]+,v[0-9]+,v[0-9]+ +** ... +** fsrm\s+[atx][0-9]+ +** ... +*/ +TEST_UNARY_CALL (float, __builtin_roundevenf) diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/unop/math-roundeven-2.c b/gcc/testsuite/gcc.target/riscv/rvv/autovec/unop/math-roundeven-2.c new file mode 100644 index 00000000000..074f1b4a1ae --- /dev/null +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/unop/math-roundeven-2.c @@ -0,0 +1,23 @@ +/* { dg-do compile } */ +/* { dg-options "-march=rv64gcv -mabi=lp64d -O3 -ftree-vectorize -fno-vect-cost-model -ffast-math -fno-schedule-insns -fno-schedule-insns2" } */ +/* { dg-final { check-function-bodies "**" "" } } */ + +#include "test-math.h" + +/* +** test_double___builtin_roundeven: +** frrm\s+[atx][0-9]+ +** ... +** fsrmi\s+0 +** ... +** vsetvli\s+[atx][0-9]+,\s*zero,\s*e64,\s*m1,\s*ta,\s*mu +** vfabs\.v\s+v[0-9]+,\s*v[0-9]+ +** vmflt\.vf\s+v0,\s*v[0-9]+,\s*[fa]+[0-9]+ +** vfcvt\.x\.f\.v\s+v[0-9]+,\s*v[0-9]+,\s*v0\.t +** vfcvt\.f\.x\.v\s+v[0-9]+,\s*v[0-9]+,\s*v0\.t +** vfsgnj\.vv\s+v[0-9]+,v[0-9]+,v[0-9]+ +** ... +** fsrm\s+[atx][0-9]+ +** ... +*/ +TEST_UNARY_CALL (double, __builtin_roundeven) diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/unop/math-roundeven-3.c b/gcc/testsuite/gcc.target/riscv/rvv/autovec/unop/math-roundeven-3.c new file mode 100644 index 00000000000..c95e8eca007 --- /dev/null +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/unop/math-roundeven-3.c @@ -0,0 +1,25 @@ +/* { dg-do compile } */ +/* { dg-options "-march=rv64gcv -mabi=lp64d -O3 -ftree-vectorize -fno-vect-cost-model -ffast-math -fno-schedule-insns -fno-schedule-insns2" } */ +/* { dg-final { check-function-bodies "**" "" } } */ + +#include "test-math.h" + +/* +** test_float___builtin_roundevenf: +** frrm\s+[atx][0-9]+ +** ... +** fsrmi\s+0 +** ... +** vsetvli\s+[atx][0-9]+,\s*zero,\s*e32,\s*m1,\s*ta,\s*mu +** vfabs\.v\s+v[0-9]+,\s*v[0-9]+ +** vmflt\.vf\s+v0,\s*v[0-9]+,\s*[fa]+[0-9]+ +** vfcvt\.x\.f\.v\s+v[0-9]+,\s*v[0-9]+,\s*v0\.t +** vfcvt\.f\.x\.v\s+v[0-9]+,\s*v[0-9]+,\s*v0\.t +** vfsgnj\.vv\s+v[0-9]+,v[0-9]+,v[0-9]+ +** ... +** vmerge\.vvm\s+v[0-9]+,\s*v[0-9]+,\s*v[0-9]+,\s*v0 +** ... +** fsrm\s+[atx][0-9]+ +** ... +*/ +TEST_COND_UNARY_CALL (float, __builtin_roundevenf) diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/vls/math-roundeven-1.c b/gcc/testsuite/gcc.target/riscv/rvv/autovec/vls/math-roundeven-1.c new file mode 100644 index 00000000000..8489d39481f --- /dev/null +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/vls/math-roundeven-1.c @@ -0,0 +1,56 @@ +/* { dg-do compile } */ +/* { dg-options "-march=rv64gcv_zvfh_zvl4096b -mabi=lp64d -O3 --param=riscv-autovec-lmul=m8 -ffast-math -fdump-tree-optimized" } */ + +#include "def.h" + +DEF_OP_V (roundevenf16, 1, _Float16, __builtin_roundevenf16) +DEF_OP_V (roundevenf16, 2, _Float16, __builtin_roundevenf16) +DEF_OP_V (roundevenf16, 4, _Float16, __builtin_roundevenf16) +DEF_OP_V (roundevenf16, 8, _Float16, __builtin_roundevenf16) +DEF_OP_V (roundevenf16, 16, _Float16, __builtin_roundevenf16) +DEF_OP_V (roundevenf16, 32, _Float16, __builtin_roundevenf16) +DEF_OP_V (roundevenf16, 64, _Float16, __builtin_roundevenf16) +DEF_OP_V (roundevenf16, 128, _Float16, __builtin_roundevenf16) +DEF_OP_V (roundevenf16, 256, _Float16, __builtin_roundevenf16) +DEF_OP_V (roundevenf16, 512, _Float16, __builtin_roundevenf16) +DEF_OP_V (roundevenf16, 1024, _Float16, __builtin_roundevenf16) +DEF_OP_V (roundevenf16, 2048, _Float16, __builtin_roundevenf16) + +DEF_OP_V (roundevenf, 1, float, __builtin_roundevenf) +DEF_OP_V (roundevenf, 2, float, __builtin_roundevenf) +DEF_OP_V (roundevenf, 4, float, __builtin_roundevenf) +DEF_OP_V (roundevenf, 8, float, __builtin_roundevenf) +DEF_OP_V (roundevenf, 16, float, __builtin_roundevenf) +DEF_OP_V (roundevenf, 32, float, __builtin_roundevenf) +DEF_OP_V (roundevenf, 64, float, __builtin_roundevenf) +DEF_OP_V (roundevenf, 128, float, __builtin_roundevenf) +DEF_OP_V (roundevenf, 256, float, __builtin_roundevenf) +DEF_OP_V (roundevenf, 512, float, __builtin_roundevenf) +DEF_OP_V (roundevenf, 1024, float, __builtin_roundevenf) + +DEF_OP_V (roundeven, 1, double, __builtin_roundeven) +DEF_OP_V (roundeven, 2, double, __builtin_roundeven) +DEF_OP_V (roundeven, 4, double, __builtin_roundeven) +DEF_OP_V (roundeven, 8, double, __builtin_roundeven) +DEF_OP_V (roundeven, 16, double, __builtin_roundeven) +DEF_OP_V (roundeven, 32, double, __builtin_roundeven) +DEF_OP_V (roundeven, 64, double, __builtin_roundeven) +DEF_OP_V (roundeven, 128, double, __builtin_roundeven) +DEF_OP_V (roundeven, 256, double, __builtin_roundeven) +DEF_OP_V (roundeven, 512, double, __builtin_roundeven) + +/* { dg-final { scan-assembler-not {csrr} } } */ +/* { dg-final { scan-tree-dump-not "1,1" "optimized" } } */ +/* { dg-final { scan-tree-dump-not "2,2" "optimized" } } */ +/* { dg-final { scan-tree-dump-not "4,4" "optimized" } } */ +/* { dg-final { scan-tree-dump-not "16,16" "optimized" } } */ +/* { dg-final { scan-tree-dump-not "32,32" "optimized" } } */ +/* { dg-final { scan-tree-dump-not "64,64" "optimized" } } */ +/* { dg-final { scan-tree-dump-not "128,128" "optimized" } } */ +/* { dg-final { scan-tree-dump-not "256,256" "optimized" } } */ +/* { dg-final { scan-tree-dump-not "512,512" "optimized" } } */ +/* { dg-final { scan-tree-dump-not "1024,1024" "optimized" } } */ +/* { dg-final { scan-tree-dump-not "2048,2048" "optimized" } } */ +/* { dg-final { scan-tree-dump-not "4096,4096" "optimized" } } */ +/* { dg-final { scan-assembler-times {vfcvt\.x\.f\.v\s+v[0-9]+,\s*v[0-9]+,\s*v0\.t} 30 } } */ +/* { dg-final { scan-assembler-times {vfcvt\.f\.x\.v\s+v[0-9]+,\s*v[0-9]+,\s*v0\.t} 30 } } */