[all-commits] [llvm/llvm-project] 2eed88: [AArch64] Implement FP8 SVE intrinsics for fused m...

[Momchil Velikov via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 2eed88da6a100216bf542e0c16762d336791876b
      https://github.com/llvm/llvm-project/commit/2eed88da6a100216bf542e0c16762d336791876b
  Author: Momchil Velikov <momchil.velikov at arm.com>
  Date:   2024-12-13 (Fri, 13 Dec 2024)

  Changed paths:
    M clang/include/clang/Basic/arm_sve.td
    A clang/test/CodeGen/AArch64/fp8-intrinsics/acle_sve2_fp8_fmla.c
    M clang/test/Sema/aarch64-sve2-intrinsics/acle_sve2_fp8.c
    M llvm/include/llvm/IR/IntrinsicsAArch64.td
    M llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td
    M llvm/lib/Target/AArch64/SVEInstrFormats.td
    A llvm/test/CodeGen/AArch64/fp8-sve-fmla.ll

  Log Message:
  -----------
  [AArch64] Implement FP8 SVE intrinsics for fused multiply-add (#118126)

This patch adds the following intrinsics:

* 8-bit floating-point multiply-add long to half-precision (bottom).

// Only if (__ARM_FEATURE_SVE2 && __ARM_FEATURE_FP8FMA) ||
__ARM_FEATURE_SSVE_FP8FMA
svfloat16_t svmlalb[_f16_mf8]_fpm(svfloat16_t zda, svmfloat8_t zn,
svmfloat8_t zm, fpm_t fpm);
svfloat16_t svmlalb[_n_f16_mf8]_fpm(svfloat16_t zda, svmfloat8_t zn,
mfloat8_t zm, fpm_t fpm);

* 8-bit floating-point multiply-add long to half-precision (bottom,
indexed).

// Only if (__ARM_FEATURE_SVE2 && __ARM_FEATURE_FP8FMA) ||
__ARM_FEATURE_SSVE_FP8FMA
svfloat16_t svmlalb_lane[_f16_mf8]_fpm(svfloat16_t zda, svmfloat8_t zn,
svmfloat8_t zm,
                                         uint64_t imm0_15, fpm_t fpm);

* 8-bit floating-point multiply-add long to half-precision (top).

// Only if (__ARM_FEATURE_SVE2 && __ARM_FEATURE_FP8FMA) ||
__ARM_FEATURE_SSVE_FP8FMA
svfloat16_t svmlalt[_f16_mf8]_fpm(svfloat16_t zda, svmfloat8_t zn,
svmfloat8_t zm, fpm_t fpm);
svfloat16_t svmlalt[_n_f16_mf8]_fpm(svfloat16_t zda, svmfloat8_t zn,
mfloat8_t zm, fpm_t fpm);

* 8-bit floating-point multiply-add long to half-precision (top,
indexed).

// Only if (__ARM_FEATURE_SVE2 && __ARM_FEATURE_FP8FMA) ||
__ARM_FEATURE_SSVE_FP8FMA
svfloat16_t svmlalt_lane[_f16_mf8]_fpm(svfloat16_t zda, svmfloat8_t zn,
svmfloat8_t zm,
                                         uint64_t imm0_15, fpm_t fpm);

* 8-bit floating-point multiply-add long long to single-precision
(bottom bottom).

// Only if (__ARM_FEATURE_SVE2 && __ARM_FEATURE_FP8FMA) ||
__ARM_FEATURE_SSVE_FP8FMA
svfloat32_t svmlallbb[_f32_mf8]_fpm(svfloat32_t zda, svmfloat8_t zn,
svmfloat8_t zm, fpm_t fpm);
svfloat32_t svmlallbb[_n_f32_mf8]_fpm(svfloat32_t zda, svmfloat8_t zn,
mfloat8_t zm, fpm_t fpm);

* 8-bit floating-point multiply-add long long to single-precision
(bottom bottom, indexed).

// Only if (__ARM_FEATURE_SVE2 && __ARM_FEATURE_FP8FMA) ||
__ARM_FEATURE_SSVE_FP8FMA
svfloat32_t svmlallbb_lane[_f32_mf8]_fpm(svfloat32_t zda, svmfloat8_t
zn, svmfloat8_t zm,
                                           uint64_t imm0_15, fpm_t fpm);

* 8-bit floating-point multiply-add long long to single-precision
(bottom top).

// Only if (__ARM_FEATURE_SVE2 && __ARM_FEATURE_FP8FMA) ||
__ARM_FEATURE_SSVE_FP8FMA
svfloat32_t svmlallbt[_f32_mf8]_fpm(svfloat32_t zda, svmfloat8_t zn,
svmfloat8_t zm, fpm_t fpm);
svfloat32_t svmlallbt[_n_f32_mf8]_fpm(svfloat32_t zda, svmfloat8_t zn,
mfloat8_t zm, fpm_t fpm);

* 8-bit floating-point multiply-add long long to single-precision
(bottom top, indexed).

// Only if (__ARM_FEATURE_SVE2 && __ARM_FEATURE_FP8FMA) ||
__ARM_FEATURE_SSVE_FP8FMA
svfloat32_t svmlallbt_lane[_f32_mf8]_fpm(svfloat32_t zda, svmfloat8_t
zn, svmfloat8_t zm,
                                           uint64_t imm0_15, fpm_t fpm);
* 8-bit floating-point multiply-add long long to single-precision (top
bottom).

// Only if (__ARM_FEATURE_SVE2 && __ARM_FEATURE_FP8FMA) ||
__ARM_FEATURE_SSVE_FP8FMA
svfloat32_t svmlalltb[_f32_mf8]_fpm(svfloat32_t zda, svmfloat8_t zn,
svmfloat8_t zm, fpm_t fpm);
svfloat32_t svmlalltb[_n_f32_mf8]_fpm(svfloat32_t zda, svmfloat8_t zn,
mfloat8_t zm, fpm_t fpm);

* 8-bit floating-point multiply-add long long to single-precision (top
bottom, indexed).

// Only if (__ARM_FEATURE_SVE2 && __ARM_FEATURE_FP8FMA) ||
__ARM_FEATURE_SSVE_FP8FMA
svfloat32_t svmlalltb_lane[_f32_mf8]_fpm(svfloat32_t zda, svmfloat8_t
zn, svmfloat8_t zm,
                                           uint64_t imm0_15, fpm_t fpm);
* 8-bit floating-point multiply-add long long to single-precision (top
top).

// Only if (__ARM_FEATURE_SVE2 && __ARM_FEATURE_FP8FMA) ||
__ARM_FEATURE_SSVE_FP8FMA
svfloat32_t svmlalltt[_f32_mf8]_fpm(svfloat32_t zda, svmfloat8_t zn,
svmfloat8_t zm, fpm_t fpm);
svfloat32_t svmlalltt[_n_f32_mf8]_fpm(svfloat32_t zda, svmfloat8_t zn,
mfloat8_t zm, fpm_t fpm);

* 8-bit floating-point multiply-add long long to single-precision (top
top, indexed).

// Only if (__ARM_FEATURE_SVE2 && __ARM_FEATURE_FP8FMA) ||
__ARM_FEATURE_SSVE_FP8FMA
svfloat32_t svmlalltt_lane[_f32_mf8]_fpm(svfloat32_t zda, svmfloat8_t
zn, svmfloat8_t zm,
                                           uint64_t imm0_15, fpm_t fpm);



To unsubscribe from these emails, change your notification settings at https://github.com/llvm/llvm-project/settings/notifications