LGTM, we would like to improve that on the combine pass, but it could
be improved later.
On Tue, Jun 6, 2023 at 8:04 PM <juzhe.zh...@rivai.ai> wrote:
>
> From: Juzhe-Zhong <juzhe.zh...@rivai.ai>
>
> Fix according to comments from Robin of V1 patch.
>
> This patch add combine optimization for following case:
> __attribute__ ((noipa)) void
> vwmaccsu (int16_t *__restrict dst, int8_t *__restrict a, uint8_t *__restrict
> b,
> int n)
> {
> for (int i = 0; i < n; i++)
> dst[i] += (int16_t) a[i] * (int16_t) b[i];
> }
>
> Before this patch:
> ...
> vsext.vf2
> vzext.vf2
> vmadd.vv
> ..
>
> After this patch:
> ...
> vwmaccsu.vv
> ...
>
> gcc/ChangeLog:
>
> * config/riscv/autovec-opt.md (*<optab>_fma<mode>): New pattern.
> (*single_<optab>mult_plus<mode>): Ditto.
> (*double_<optab>mult_plus<mode>): Ditto.
> (*sign_zero_extend_fma): Ditto.
> (*zero_sign_extend_fma): Ditto.
> * config/riscv/riscv-protos.h (enum insn_type): New enum.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target/riscv/rvv/autovec/widen/widen-8.c: New test.
> * gcc.target/riscv/rvv/autovec/widen/widen-9.c: New test.
> * gcc.target/riscv/rvv/autovec/widen/widen-complicate-5.c: New test.
> * gcc.target/riscv/rvv/autovec/widen/widen-complicate-6.c: New test.
> * gcc.target/riscv/rvv/autovec/widen/widen_run-8.c: New test.
> * gcc.target/riscv/rvv/autovec/widen/widen_run-9.c: New test.
>
> ---
> gcc/config/riscv/autovec-opt.md | 160 ++++++++++++++++++
> gcc/config/riscv/riscv-protos.h | 1 +
> .../riscv/rvv/autovec/widen/widen-8.c | 27 +++
> .../riscv/rvv/autovec/widen/widen-9.c | 23 +++
> .../rvv/autovec/widen/widen-complicate-5.c | 32 ++++
> .../rvv/autovec/widen/widen-complicate-6.c | 30 ++++
> .../riscv/rvv/autovec/widen/widen_run-8.c | 38 +++++
> .../riscv/rvv/autovec/widen/widen_run-9.c | 35 ++++
> 8 files changed, 346 insertions(+)
> create mode 100644 gcc/testsuite/gcc.target/riscv/rvv/autovec/widen/widen-8.c
> create mode 100644 gcc/testsuite/gcc.target/riscv/rvv/autovec/widen/widen-9.c
> create mode 100644
> gcc/testsuite/gcc.target/riscv/rvv/autovec/widen/widen-complicate-5.c
> create mode 100644
> gcc/testsuite/gcc.target/riscv/rvv/autovec/widen/widen-complicate-6.c
> create mode 100644
> gcc/testsuite/gcc.target/riscv/rvv/autovec/widen/widen_run-8.c
> create mode 100644
> gcc/testsuite/gcc.target/riscv/rvv/autovec/widen/widen_run-9.c
>
> diff --git a/gcc/config/riscv/autovec-opt.md b/gcc/config/riscv/autovec-opt.md
> index f6052b50572..7bb93eed220 100644
> --- a/gcc/config/riscv/autovec-opt.md
> +++ b/gcc/config/riscv/autovec-opt.md
> @@ -170,3 +170,163 @@
> }
> [(set_attr "type" "vmalu")
> (set_attr "mode" "<MODE>")])
> +
> +;; =========================================================================
> +;; == Widening Ternary arithmetic
> +;; =========================================================================
> +
> +;; -------------------------------------------------------------------------
> +;; ---- [INT] VWMACC
> +;; -------------------------------------------------------------------------
> +;; Includes:
> +;; - vwmacc.vv
> +;; - vwmaccu.vv
> +;; -------------------------------------------------------------------------
> +
> +;; Combine ext + ext + fma ===> widen fma.
> +;; Most of circumstantces, LoopVectorizer will generate the following IR:
> +;; vect__8.64_40 = (vector([4,4]) int) vect__7.63_41;
> +;; vect__11.68_35 = (vector([4,4]) int) vect__10.67_36;
> +;; vect__13.70_33 = .FMA (vect__11.68_35, vect__8.64_40, vect__4.60_45);
> +(define_insn_and_split "*<optab>_fma<mode>"
> + [(set (match_operand:VWEXTI 0 "register_operand")
> + (plus:VWEXTI
> + (mult:VWEXTI
> + (any_extend:VWEXTI
> + (match_operand:<V_DOUBLE_TRUNC> 2 "register_operand"))
> + (any_extend:VWEXTI
> + (match_operand:<V_DOUBLE_TRUNC> 3 "register_operand")))
> + (match_operand:VWEXTI 1 "register_operand")))]
> + "TARGET_VECTOR && can_create_pseudo_p ()"
> + "#"
> + "&& 1"
> + [(const_int 0)]
> + {
> + riscv_vector::emit_vlmax_ternary_insn (code_for_pred_widen_mul_plus
> (<CODE>, <MODE>mode),
> + riscv_vector::RVV_WIDEN_TERNOP,
> operands);
> + DONE;
> + }
> + [(set_attr "type" "viwmuladd")
> + (set_attr "mode" "<V_DOUBLE_TRUNC>")])
> +
> +;; This helps to match ext + fma.
> +(define_insn_and_split "*single_<optab>mult_plus<mode>"
> + [(set (match_operand:VWEXTI 0 "register_operand")
> + (plus:VWEXTI
> + (mult:VWEXTI
> + (any_extend:VWEXTI
> + (match_operand:<V_DOUBLE_TRUNC> 2 "register_operand"))
> + (match_operand:VWEXTI 3 "register_operand"))
> + (match_operand:VWEXTI 1 "register_operand")))]
> + "TARGET_VECTOR && can_create_pseudo_p ()"
> + "#"
> + "&& 1"
> + [(const_int 0)]
> + {
> + insn_code icode = code_for_pred_vf2 (<CODE>, <MODE>mode);
> + rtx tmp = gen_reg_rtx (<MODE>mode);
> + rtx ext_ops[] = {tmp, operands[2]};
> + riscv_vector::emit_vlmax_insn (icode, riscv_vector::RVV_UNOP, ext_ops);
> +
> + rtx dst = expand_ternary_op (<MODE>mode, fma_optab, tmp, operands[3],
> + operands[1], operands[0], 0);
> + emit_move_insn (operands[0], dst);
> + DONE;
> + }
> + [(set_attr "type" "viwmuladd")
> + (set_attr "mode" "<V_DOUBLE_TRUNC>")])
> +
> +;; Combine ext + ext + mult + plus ===> widen fma.
> +;; We have some special cases generated by LoopVectorizer:
> +;; vect__8.18_46 = (vector([8,8]) signed short) vect__7.17_47;
> +;; vect__11.22_41 = (vector([8,8]) signed short) vect__10.21_42;
> +;; vect__12.23_40 = vect__11.22_41 * vect__8.18_46;
> +;; vect__14.25_38 = vect__13.24_39 + vect__5.14_51;
> +;; This situation doesn't generate FMA IR.
> +(define_insn_and_split "*double_<optab>mult_plus<mode>"
> + [(set (match_operand:VWEXTI 0 "register_operand")
> + (if_then_else:VWEXTI
> + (unspec:<VM>
> + [(match_operand:<VM> 1 "vector_mask_operand")
> + (match_operand 6 "vector_length_operand")
> + (match_operand 7 "const_int_operand")
> + (match_operand 8 "const_int_operand")
> + (match_operand 9 "const_int_operand")
> + (reg:SI VL_REGNUM)
> + (reg:SI VTYPE_REGNUM)] UNSPEC_VPREDICATE)
> + (plus:VWEXTI
> + (if_then_else:VWEXTI
> + (unspec:<VM>
> + [(match_dup 1)
> + (match_dup 6)
> + (match_dup 7)
> + (match_dup 8)
> + (match_dup 9)
> + (reg:SI VL_REGNUM)
> + (reg:SI VTYPE_REGNUM)] UNSPEC_VPREDICATE)
> + (mult:VWEXTI
> + (any_extend:VWEXTI
> + (match_operand:<V_DOUBLE_TRUNC> 4 "register_operand"))
> + (any_extend:VWEXTI
> + (match_operand:<V_DOUBLE_TRUNC> 5 "register_operand")))
> + (match_operand:VWEXTI 2 "vector_undef_operand"))
> + (match_operand:VWEXTI 3 "register_operand"))
> + (match_dup 2)))]
> + "TARGET_VECTOR && can_create_pseudo_p ()"
> + "#"
> + "&& 1"
> + [(const_int 0)]
> + {
> + emit_insn (gen_pred_widen_mul_plus (<CODE>, <MODE>mode, operands[0],
> + operands[1], operands[3], operands[4],
> + operands[5], operands[6], operands[7],
> + operands[8], operands[9]));
> + DONE;
> + }
> + [(set_attr "type" "viwmuladd")
> + (set_attr "mode" "<V_DOUBLE_TRUNC>")])
> +
> +;; Combine sign_extend + zero_extend + fma ===> widen fma (su).
> +(define_insn_and_split "*sign_zero_extend_fma"
> + [(set (match_operand:VWEXTI 0 "register_operand")
> + (plus:VWEXTI
> + (mult:VWEXTI
> + (sign_extend:VWEXTI
> + (match_operand:<V_DOUBLE_TRUNC> 2 "register_operand"))
> + (zero_extend:VWEXTI
> + (match_operand:<V_DOUBLE_TRUNC> 3 "register_operand")))
> + (match_operand:VWEXTI 1 "register_operand")))]
> + "TARGET_VECTOR && can_create_pseudo_p ()"
> + "#"
> + "&& 1"
> + [(const_int 0)]
> + {
> + riscv_vector::emit_vlmax_ternary_insn (code_for_pred_widen_mul_plussu
> (<MODE>mode),
> + riscv_vector::RVV_WIDEN_TERNOP,
> operands);
> + DONE;
> + }
> + [(set_attr "type" "viwmuladd")
> + (set_attr "mode" "<V_DOUBLE_TRUNC>")])
> +
> +;; This helps to match zero_extend + sign_extend + fma.
> +(define_insn_and_split "*zero_sign_extend_fma"
> + [(set (match_operand:VWEXTI 0 "register_operand")
> + (plus:VWEXTI
> + (mult:VWEXTI
> + (zero_extend:VWEXTI
> + (match_operand:<V_DOUBLE_TRUNC> 2 "register_operand"))
> + (sign_extend:VWEXTI
> + (match_operand:<V_DOUBLE_TRUNC> 3 "register_operand")))
> + (match_operand:VWEXTI 1 "register_operand")))]
> + "TARGET_VECTOR && can_create_pseudo_p ()"
> + "#"
> + "&& 1"
> + [(const_int 0)]
> + {
> + rtx ops[] = {operands[0], operands[1], operands[3], operands[2]};
> + riscv_vector::emit_vlmax_ternary_insn (code_for_pred_widen_mul_plussu
> (<MODE>mode),
> + riscv_vector::RVV_WIDEN_TERNOP,
> ops);
> + DONE;
> + }
> + [(set_attr "type" "viwmuladd")
> + (set_attr "mode" "<V_DOUBLE_TRUNC>")])
> diff --git a/gcc/config/riscv/riscv-protos.h b/gcc/config/riscv/riscv-protos.h
> index 27ecd16e496..b311b937f17 100644
> --- a/gcc/config/riscv/riscv-protos.h
> +++ b/gcc/config/riscv/riscv-protos.h
> @@ -143,6 +143,7 @@ enum insn_type
> RVV_CMP_MU_OP = RVV_CMP_OP + 2, /* +2 means mask and maskoff operand. */
> RVV_UNOP_MU = RVV_UNOP + 2, /* Likewise. */
> RVV_TERNOP = 5,
> + RVV_WIDEN_TERNOP = 4,
> RVV_SCALAR_MOV_OP = 4, /* +1 for VUNDEF according to vector.md. */
> };
> enum vlmul_type
> diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/widen/widen-8.c
> b/gcc/testsuite/gcc.target/riscv/rvv/autovec/widen/widen-8.c
> new file mode 100644
> index 00000000000..f3ca07c02e0
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/widen/widen-8.c
> @@ -0,0 +1,27 @@
> +/* { dg-do compile } */
> +/* { dg-additional-options "-march=rv32gcv -mabi=ilp32d
> --param=riscv-autovec-preference=scalable" } */
> +
> +#include <stdint-gcc.h>
> +
> +#define TEST_TYPE(TYPE1, TYPE2)
> \
> + __attribute__ ((noipa)) void vwmacc_##TYPE1_##TYPE2 (TYPE1 *__restrict
> dst, \
> + TYPE2 *__restrict a,
> \
> + TYPE2 *__restrict b,
> \
> + int n)
> \
> + {
> \
> + for (int i = 0; i < n; i++)
> \
> + dst[i] += (TYPE1) a[i] * (TYPE1) b[i];
> \
> + }
> +
> +#define TEST_ALL()
> \
> + TEST_TYPE (int16_t, int8_t)
> \
> + TEST_TYPE (uint16_t, uint8_t)
> \
> + TEST_TYPE (int32_t, int16_t)
> \
> + TEST_TYPE (uint32_t, uint16_t)
> \
> + TEST_TYPE (int64_t, int32_t)
> \
> + TEST_TYPE (uint64_t, uint32_t)
> +
> +TEST_ALL ()
> +
> +/* { dg-final { scan-assembler-times {\tvwmacc\.vv} 3 } } */
> +/* { dg-final { scan-assembler-times {\tvwmaccu\.vv} 3 } } */
> diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/widen/widen-9.c
> b/gcc/testsuite/gcc.target/riscv/rvv/autovec/widen/widen-9.c
> new file mode 100644
> index 00000000000..969a1e8f80c
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/widen/widen-9.c
> @@ -0,0 +1,23 @@
> +/* { dg-do compile } */
> +/* { dg-additional-options "-march=rv32gcv -mabi=ilp32d
> --param=riscv-autovec-preference=scalable" } */
> +
> +#include <stdint-gcc.h>
> +
> +#define TEST_TYPE(TYPE1, TYPE2, TYPE3)
> \
> + __attribute__ ((noipa)) void vwmacc_##TYPE1_##TYPE2 (TYPE1 *__restrict
> dst, \
> + TYPE2 *__restrict a,
> \
> + TYPE3 *__restrict b,
> \
> + int n)
> \
> + {
> \
> + for (int i = 0; i < n; i++)
> \
> + dst[i] += (TYPE1) a[i] * (TYPE1) b[i];
> \
> + }
> +
> +#define TEST_ALL()
> \
> + TEST_TYPE (int16_t, int8_t, uint8_t)
> \
> + TEST_TYPE (int32_t, int16_t, uint16_t)
> \
> + TEST_TYPE (int64_t, int32_t, uint32_t)
> +
> +TEST_ALL ()
> +
> +/* { dg-final { scan-assembler-times {\tvwmaccsu\.vv} 3 } } */
> diff --git
> a/gcc/testsuite/gcc.target/riscv/rvv/autovec/widen/widen-complicate-5.c
> b/gcc/testsuite/gcc.target/riscv/rvv/autovec/widen/widen-complicate-5.c
> new file mode 100644
> index 00000000000..187b6db21fd
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/widen/widen-complicate-5.c
> @@ -0,0 +1,32 @@
> +/* { dg-do compile } */
> +/* { dg-additional-options "-march=rv32gcv -mabi=ilp32d
> --param=riscv-autovec-preference=scalable" } */
> +
> +#include <stdint-gcc.h>
> +
> +#define TEST_TYPE(TYPE1, TYPE2)
> \
> + __attribute__ ((noipa)) void vwadd_##TYPE1_##TYPE2 (
> \
> + TYPE1 *__restrict dst, TYPE1 *__restrict dst2, TYPE1 *__restrict dst3,
> \
> + TYPE1 *__restrict dst4, TYPE2 *__restrict a, TYPE2 *__restrict b,
> \
> + TYPE2 *__restrict a2, TYPE2 *__restrict b2, int n)
> \
> + {
> \
> + for (int i = 0; i < n; i++)
> \
> + {
> \
> + dst[i] += (TYPE1) a[i] * (TYPE1) b[i];
> \
> + dst2[i] += (TYPE1) a2[i] * (TYPE1) b[i];
> \
> + dst3[i] += (TYPE1) a2[i] * (TYPE1) a[i];
> \
> + dst4[i] += (TYPE1) a[i] * (TYPE1) b2[i];
> \
> + }
> \
> + }
> +
> +#define TEST_ALL()
> \
> + TEST_TYPE (int16_t, int8_t)
> \
> + TEST_TYPE (uint16_t, uint8_t)
> \
> + TEST_TYPE (int32_t, int16_t)
> \
> + TEST_TYPE (uint32_t, uint16_t)
> \
> + TEST_TYPE (int64_t, int32_t)
> \
> + TEST_TYPE (uint64_t, uint32_t)
> +
> +TEST_ALL ()
> +
> +/* { dg-final { scan-assembler-times {\tvwmacc\.vv} 12 } } */
> +/* { dg-final { scan-assembler-times {\tvwmaccu\.vv} 12 } } */
> diff --git
> a/gcc/testsuite/gcc.target/riscv/rvv/autovec/widen/widen-complicate-6.c
> b/gcc/testsuite/gcc.target/riscv/rvv/autovec/widen/widen-complicate-6.c
> new file mode 100644
> index 00000000000..fa56f21aa81
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/widen/widen-complicate-6.c
> @@ -0,0 +1,30 @@
> +/* { dg-do compile } */
> +/* { dg-additional-options "-march=rv32gcv -mabi=ilp32d
> --param=riscv-autovec-preference=scalable" } */
> +
> +#include <stdint-gcc.h>
> +
> +#define TEST_TYPE(TYPE1, TYPE2, TYPE3)
> \
> + __attribute__ ((noipa)) void vwadd_##TYPE1_##TYPE2 (
> \
> + TYPE1 *__restrict dst, TYPE1 *__restrict dst2, TYPE1 *__restrict dst3,
> \
> + TYPE1 *__restrict dst4, TYPE2 *__restrict a, TYPE3 *__restrict b,
> \
> + TYPE3 *__restrict a2, TYPE2 *__restrict b2, int n)
> \
> + {
> \
> + for (int i = 0; i < n; i++)
> \
> + {
> \
> + dst[i] += (TYPE1) a[i] * (TYPE1) b[i];
> \
> + dst2[i] += (TYPE1) a2[i] * (TYPE1) b[i];
> \
> + dst3[i] += (TYPE1) a2[i] * (TYPE1) a[i];
> \
> + dst4[i] += (TYPE1) a[i] * (TYPE1) b2[i];
> \
> + }
> \
> + }
> +
> +#define TEST_ALL()
> \
> + TEST_TYPE (int16_t, int8_t, uint8_t)
> \
> + TEST_TYPE (int32_t, int16_t, uint16_t)
> \
> + TEST_TYPE (int64_t, int32_t, uint32_t)
> +
> +TEST_ALL ()
> +
> +/* { dg-final { scan-assembler-times {\tvwmaccsu\.vv} 6 } } */
> +/* { dg-final { scan-assembler-times {\tvwmacc\.vv} 3 } } */
> +/* { dg-final { scan-assembler-times {\tvwmaccu\.vv} 3 } } */
> diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/widen/widen_run-8.c
> b/gcc/testsuite/gcc.target/riscv/rvv/autovec/widen/widen_run-8.c
> new file mode 100644
> index 00000000000..f4840d30dc2
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/widen/widen_run-8.c
> @@ -0,0 +1,38 @@
> +/* { dg-do run { target { riscv_vector } } } */
> +/* { dg-additional-options "--param=riscv-autovec-preference=scalable" } */
> +
> +#include <assert.h>
> +#include "widen-8.c"
> +
> +#define SZ 512
> +
> +#define RUN(TYPE1, TYPE2, LIMIT)
> \
> + TYPE2 a##TYPE2[SZ];
> \
> + TYPE2 b##TYPE2[SZ];
> \
> + TYPE1 dst##TYPE1[SZ];
> \
> + TYPE1 dst2##TYPE1[SZ];
> \
> + for (int i = 0; i < SZ; i++)
> \
> + {
> \
> + a##TYPE2[i] = LIMIT + i % 8723;
> \
> + b##TYPE2[i] = LIMIT + i & 1964;
> \
> + dst##TYPE1[i] = LIMIT + i & 628;
> \
> + dst2##TYPE1[i] = LIMIT + i & 628;
> \
> + }
> \
> + vwmacc_##TYPE1_##TYPE2 (dst##TYPE1, a##TYPE2, b##TYPE2, SZ);
> \
> + for (int i = 0; i < SZ; i++)
> \
> + assert (dst##TYPE1[i]
> \
> + == ((TYPE1) a##TYPE2[i] * (TYPE1) b##TYPE2[i]) + dst2##TYPE1[i]);
> +
> +#define RUN_ALL()
> \
> + RUN (int16_t, int8_t, -128)
> \
> + RUN (uint16_t, uint8_t, 255)
> \
> + RUN (int32_t, int16_t, -32768)
> \
> + RUN (uint32_t, uint16_t, 65535)
> \
> + RUN (int64_t, int32_t, -2147483648)
> \
> + RUN (uint64_t, uint32_t, 4294967295)
> +
> +int
> +main ()
> +{
> + RUN_ALL ()
> +}
> diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/widen/widen_run-9.c
> b/gcc/testsuite/gcc.target/riscv/rvv/autovec/widen/widen_run-9.c
> new file mode 100644
> index 00000000000..2caa09a2c5a
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/widen/widen_run-9.c
> @@ -0,0 +1,35 @@
> +/* { dg-do run { target { riscv_vector } } } */
> +/* { dg-additional-options "--param=riscv-autovec-preference=scalable" } */
> +
> +#include <assert.h>
> +#include "widen-9.c"
> +
> +#define SZ 512
> +
> +#define RUN(TYPE1, TYPE2, TYPE3, LIMIT)
> \
> + TYPE2 a##TYPE2[SZ];
> \
> + TYPE3 b##TYPE3[SZ];
> \
> + TYPE1 dst##TYPE1[SZ];
> \
> + TYPE1 dst2##TYPE1[SZ];
> \
> + for (int i = 0; i < SZ; i++)
> \
> + {
> \
> + a##TYPE2[i] = LIMIT + i % 8723;
> \
> + b##TYPE3[i] = LIMIT + i & 1964;
> \
> + dst##TYPE1[i] = LIMIT + i & 728;
> \
> + dst2##TYPE1[i] = LIMIT + i & 728;
> \
> + }
> \
> + vwmacc_##TYPE1_##TYPE2 (dst##TYPE1, a##TYPE2, b##TYPE3, SZ);
> \
> + for (int i = 0; i < SZ; i++)
> \
> + assert (dst##TYPE1[i]
> \
> + == ((TYPE1) a##TYPE2[i] * (TYPE1) b##TYPE3[i]) + dst2##TYPE1[i]);
> +
> +#define RUN_ALL()
> \
> + RUN (int16_t, int8_t, uint8_t, -128)
> \
> + RUN (int32_t, int16_t, uint16_t, -32768)
> \
> + RUN (int64_t, int32_t, uint32_t, -2147483648)
> +
> +int
> +main ()
> +{
> + RUN_ALL ()
> +}
> --
> 2.36.3
>
