From fc94f4d23abd4bc427b701f54ad85282e9ec7872 Mon Sep 17 00:00:00 2001 From: Michael Tyler Date: Tue, 4 Jun 2024 15:47:37 +0100 Subject: Update CPU kernels and add mixed sign GEMM support - Add support for mixed sign quantized convolution. - Add support for mixed sign dequantized GEMM. - Add SME FP16 GEMV kernel. - Change SME vector length function to use RDSVL instead of static variable. - Add GEMM dilation support internally (not exposed yet). - Remove unused "get_default_activation_values" functions. - Add SVE fixed format interleaved BF16 DOT kernel. - Updates and optimizations to assembly kernels. Resolves COMPMID-6926 Change-Id: I227f502502611d4cc4111c89e30c53ce94079544 Signed-off-by: Michael Tyler Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/11570 Tested-by: Arm Jenkins Reviewed-by: Gunes Bayir Comments-Addressed: Arm Jenkins Benchmark: Arm Jenkins --- .../sme_transpose_interleave_1VL_2x2_fp32bf16.hpp | 182 ++++++++++----------- 1 file changed, 91 insertions(+), 91 deletions(-) (limited to 'src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_1VL_2x2_fp32bf16.hpp') diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_1VL_2x2_fp32bf16.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_1VL_2x2_fp32bf16.hpp index b90063028d..2fafefbbc5 100644 --- a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_1VL_2x2_fp32bf16.hpp +++ b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_1VL_2x2_fp32bf16.hpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2022-2023 Arm Limited. + * Copyright (c) 2022-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -45,88 +45,88 @@ void sme_transpose_interleave_1VL_2x2_fp32bf16(bfloat16 *out, const float *in, s "blt 6f\n" "1:" // Main row loop: Head "mov x26, %x[in]\n" - "add x25, x26, %x[in_stride]\n" - "add x24, x25, %x[in_stride]\n" - "mov x23, %x[width]\n" - "cnth x21, ALL, MUL #2\n" - "add x20, x24, %x[in_stride]\n" - "cmp x23, x21\n" - "add %x[in], x20, %x[in_stride]\n" + "mov x25, %x[width]\n" + "add x24, x26, %x[in_stride]\n" + "cnth x23, ALL, MUL #2\n" + "add x21, x24, %x[in_stride]\n" + "cmp x25, x23\n" + "add x20, x21, %x[in_stride]\n" "mov x22, %x[out]\n" + "add %x[in], x20, %x[in_stride]\n" "sub %x[height], %x[height], #0x4\n" "blt 3f\n" "2:" // Main row loop: Unroll column loop - "ld1w { z16.s }, p1/Z, [x26]\n" + "ld1w { z19.s }, p1/Z, [x26]\n" + "sub x25, x25, x23\n" + "ld1w { z18.s }, p1/Z, [x21]\n" + "cmp x25, x23\n" + "ld1w { z17.s }, p1/Z, [x26, #1, MUL VL]\n" + "ld1w { z16.s }, p1/Z, [x21, #1, MUL VL]\n" + ".inst 0x658aa67b // bfcvt z27.h, p1/M, z19.s\n" + "ld1w { z19.s }, p1/Z, [x26, #2, MUL VL]\n" + ".inst 0x658aa65a // bfcvt z26.h, p1/M, z18.s\n" + "ld1w { z18.s }, p1/Z, [x21, #2, MUL VL]\n" + ".inst 0x658aa639 // bfcvt z25.h, p1/M, z17.s\n" + "ld1w { z17.s }, p1/Z, [x26, #3, MUL VL]\n" ".inst 0x658aa618 // bfcvt z24.h, p1/M, z16.s\n" - "sub x23, x23, x21\n" - "cmp x23, x21\n" - "ld1w { z16.s }, p1/Z, [x24]\n" - ".inst 0x658aa617 // bfcvt z23.h, p1/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x26, #1, MUL VL]\n" - ".inst 0x658aa616 // bfcvt z22.h, p1/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x24, #1, MUL VL]\n" - ".inst 0x658aa615 // bfcvt z21.h, p1/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x26, #2, MUL VL]\n" - ".inst 0x658aa614 // bfcvt z20.h, p1/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x24, #2, MUL VL]\n" - ".inst 0x658aa613 // bfcvt z19.h, p1/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x26, #3, MUL VL]\n" - ".inst 0x658aa612 // bfcvt z18.h, p1/M, z16.s\n" "addvl x26, x26, #4\n" - "ld1w { z16.s }, p1/Z, [x24, #3, MUL VL]\n" - ".inst 0x658aa611 // bfcvt z17.h, p1/M, z16.s\n" - "addvl x24, x24, #4\n" - "ld1w { z16.s }, p1/Z, [x25]\n" - ".inst 0x648aa618 // bfcvtnt z24.h, p1/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x20]\n" - ".inst 0x648aa617 // bfcvtnt z23.h, p1/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x25, #1, MUL VL]\n" - ".inst 0x648aa616 // bfcvtnt z22.h, p1/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x21, #3, MUL VL]\n" + ".inst 0x658aa677 // bfcvt z23.h, p1/M, z19.s\n" + "addvl x21, x21, #4\n" + "ld1w { z19.s }, p1/Z, [x24]\n" + ".inst 0x658aa656 // bfcvt z22.h, p1/M, z18.s\n" + "ld1w { z18.s }, p1/Z, [x20]\n" + ".inst 0x658aa635 // bfcvt z21.h, p1/M, z17.s\n" + "ld1w { z17.s }, p1/Z, [x24, #1, MUL VL]\n" + ".inst 0x658aa614 // bfcvt z20.h, p1/M, z16.s\n" "ld1w { z16.s }, p1/Z, [x20, #1, MUL VL]\n" - ".inst 0x648aa615 // bfcvtnt z21.h, p1/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x25, #2, MUL VL]\n" - ".inst 0x648aa614 // bfcvtnt z20.h, p1/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x20, #2, MUL VL]\n" - ".inst 0x648aa613 // bfcvtnt z19.h, p1/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x25, #3, MUL VL]\n" - "addvl x25, x25, #4\n" - ".inst 0x648aa612 // bfcvtnt z18.h, p1/M, z16.s\n" + ".inst 0x648aa67b // bfcvtnt z27.h, p1/M, z19.s\n" + "ld1w { z19.s }, p1/Z, [x24, #2, MUL VL]\n" + ".inst 0x648aa65a // bfcvtnt z26.h, p1/M, z18.s\n" + "ld1w { z18.s }, p1/Z, [x20, #2, MUL VL]\n" + ".inst 0x648aa639 // bfcvtnt z25.h, p1/M, z17.s\n" + "ld1w { z17.s }, p1/Z, [x24, #3, MUL VL]\n" + ".inst 0x648aa618 // bfcvtnt z24.h, p1/M, z16.s\n" + "addvl x24, x24, #4\n" "ld1w { z16.s }, p1/Z, [x20, #3, MUL VL]\n" - "st1h { z24.h }, p1, [x22]\n" + "st1h { z27.h }, p1, [x22]\n" + ".inst 0x648aa677 // bfcvtnt z23.h, p1/M, z19.s\n" "addvl x20, x20, #4\n" - ".inst 0x648aa611 // bfcvtnt z17.h, p1/M, z16.s\n" - "st1h { z23.h }, p1, [x22, #1, MUL VL]\n" + "st1h { z26.h }, p1, [x22, #1, MUL VL]\n" "add x22, x22, %x[out_stride]\n" - "st1h { z22.h }, p1, [x22]\n" - "st1h { z21.h }, p1, [x22, #1, MUL VL]\n" + ".inst 0x648aa656 // bfcvtnt z22.h, p1/M, z18.s\n" + "st1h { z25.h }, p1, [x22]\n" + ".inst 0x648aa635 // bfcvtnt z21.h, p1/M, z17.s\n" + "st1h { z24.h }, p1, [x22, #1, MUL VL]\n" "add x22, x22, %x[out_stride]\n" - "st1h { z20.h }, p1, [x22]\n" - "st1h { z19.h }, p1, [x22, #1, MUL VL]\n" + ".inst 0x648aa614 // bfcvtnt z20.h, p1/M, z16.s\n" + "st1h { z23.h }, p1, [x22]\n" + "st1h { z22.h }, p1, [x22, #1, MUL VL]\n" "add x22, x22, %x[out_stride]\n" - "st1h { z18.h }, p1, [x22]\n" - "st1h { z17.h }, p1, [x22, #1, MUL VL]\n" + "st1h { z21.h }, p1, [x22]\n" + "st1h { z20.h }, p1, [x22, #1, MUL VL]\n" "add x22, x22, %x[out_stride]\n" "bge 2b\n" "3:" // Main row loop: Unroll column loop skip - "cbz x23, 5f\n" + "cbz x25, 5f\n" "4:" // Main row loop: Column loop - "whilelt p0.s, XZR, x23\n" - "ld1w { z16.s }, p0/Z, [x26]\n" - ".inst 0x658aa612 // bfcvt z18.h, p1/M, z16.s\n" - "decw x23\n" - "ld1w { z16.s }, p0/Z, [x24]\n" - ".inst 0x658aa611 // bfcvt z17.h, p1/M, z16.s\n" - "cmp x23, #0x0\n" + "whilelt p0.s, XZR, x25\n" + "decw x25\n" + "ld1w { z17.s }, p0/Z, [x26]\n" + "cmp x25, #0x0\n" "addvl x26, x26, #1\n" - "ld1w { z16.s }, p0/Z, [x25]\n" - "addvl x25, x25, #1\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + "addvl x21, x21, #1\n" + "ld1w { z19.s }, p0/Z, [x24]\n" "addvl x24, x24, #1\n" - ".inst 0x648aa612 // bfcvtnt z18.h, p1/M, z16.s\n" - "ld1w { z16.s }, p0/Z, [x20]\n" + ".inst 0x658aa632 // bfcvt z18.h, p1/M, z17.s\n" + "ld1w { z17.s }, p0/Z, [x20]\n" "addvl x20, x20, #1\n" - ".inst 0x648aa611 // bfcvtnt z17.h, p1/M, z16.s\n" + ".inst 0x658aa610 // bfcvt z16.h, p1/M, z16.s\n" + ".inst 0x648aa672 // bfcvtnt z18.h, p1/M, z19.s\n" + ".inst 0x648aa630 // bfcvtnt z16.h, p1/M, z17.s\n" "st1h { z18.h }, p1, [x22]\n" - "st1h { z17.h }, p1, [x22, #1, MUL VL]\n" + "st1h { z16.h }, p1, [x22, #1, MUL VL]\n" "add x22, x22, %x[out_stride]\n" "bgt 4b\n" "5:" // Main row loop: Column loop skip @@ -137,42 +137,42 @@ void sme_transpose_interleave_1VL_2x2_fp32bf16(bfloat16 *out, const float *in, s "6:" // Main loop skip "7:" // Tail row loop: Head "mov x26, %x[in]\n" - "add x25, x26, %x[in_stride]\n" "cmp %x[height], #0x1\n" + "add x24, x26, %x[in_stride]\n" "mov x21, %x[width]\n" "cnth x20, ALL, MUL #2\n" - "add %x[in], x25, %x[in_stride]\n" - "csel x25, x25, %x[pad_row], GT\n" + "add %x[in], x24, %x[in_stride]\n" + "csel x24, x24, %x[pad_row], GT\n" "cmp x21, x20\n" "mov x22, %x[out]\n" "sub %x[height], %x[height], #0x2\n" "blt 9f\n" "8:" // Tail row loop: Unroll column loop - "ld1w { z16.s }, p1/Z, [x26]\n" - ".inst 0x658aa614 // bfcvt z20.h, p1/M, z16.s\n" + "ld1w { z19.s }, p1/Z, [x26]\n" "sub x21, x21, x20\n" + "ld1w { z18.s }, p1/Z, [x26, #1, MUL VL]\n" "cmp x21, x20\n" - "ld1w { z16.s }, p1/Z, [x26, #1, MUL VL]\n" - ".inst 0x658aa613 // bfcvt z19.h, p1/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x26, #2, MUL VL]\n" - ".inst 0x658aa612 // bfcvt z18.h, p1/M, z16.s\n" + "ld1w { z17.s }, p1/Z, [x26, #2, MUL VL]\n" "ld1w { z16.s }, p1/Z, [x26, #3, MUL VL]\n" - ".inst 0x658aa611 // bfcvt z17.h, p1/M, z16.s\n" + ".inst 0x658aa677 // bfcvt z23.h, p1/M, z19.s\n" "addvl x26, x26, #4\n" - "ld1w { z16.s }, p1/Z, [x25]\n" - ".inst 0x648aa614 // bfcvtnt z20.h, p1/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x25, #1, MUL VL]\n" - ".inst 0x648aa613 // bfcvtnt z19.h, p1/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x25, #2, MUL VL]\n" - ".inst 0x648aa612 // bfcvtnt z18.h, p1/M, z16.s\n" - "ld1w { z16.s }, p1/Z, [x25, #3, MUL VL]\n" - "st1h { z20.h }, p1, [x22]\n" + "ld1w { z22.s }, p1/Z, [x24]\n" + ".inst 0x658aa655 // bfcvt z21.h, p1/M, z18.s\n" + "ld1w { z20.s }, p1/Z, [x24, #1, MUL VL]\n" + ".inst 0x658aa633 // bfcvt z19.h, p1/M, z17.s\n" + "ld1w { z18.s }, p1/Z, [x24, #2, MUL VL]\n" + ".inst 0x658aa611 // bfcvt z17.h, p1/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x24, #3, MUL VL]\n" + ".inst 0x648aa6d7 // bfcvtnt z23.h, p1/M, z22.s\n" + "addvl x24, x24, #4\n" + ".inst 0x648aa695 // bfcvtnt z21.h, p1/M, z20.s\n" + ".inst 0x648aa653 // bfcvtnt z19.h, p1/M, z18.s\n" + ".inst 0x648aa611 // bfcvtnt z17.h, p1/M, z16.s\n" + "st1h { z23.h }, p1, [x22]\n" "add x22, x22, %x[out_stride]\n" - "addvl x25, x25, #4\n" - "st1h { z19.h }, p1, [x22]\n" + "st1h { z21.h }, p1, [x22]\n" "add x22, x22, %x[out_stride]\n" - ".inst 0x648aa611 // bfcvtnt z17.h, p1/M, z16.s\n" - "st1h { z18.h }, p1, [x22]\n" + "st1h { z19.h }, p1, [x22]\n" "add x22, x22, %x[out_stride]\n" "st1h { z17.h }, p1, [x22]\n" "add x22, x22, %x[out_stride]\n" @@ -181,15 +181,15 @@ void sme_transpose_interleave_1VL_2x2_fp32bf16(bfloat16 *out, const float *in, s "cbz x21, 11f\n" "10:" // Tail row loop: Column loop "whilelt p0.s, XZR, x21\n" - "ld1w { z16.s }, p0/Z, [x26]\n" - ".inst 0x658aa611 // bfcvt z17.h, p1/M, z16.s\n" "decw x21\n" - "ld1w { z16.s }, p0/Z, [x25]\n" + "ld1w { z16.s }, p0/Z, [x26]\n" "cmp x21, #0x0\n" "addvl x26, x26, #1\n" - ".inst 0x648aa611 // bfcvtnt z17.h, p1/M, z16.s\n" - "addvl x25, x25, #1\n" - "st1h { z17.h }, p1, [x22]\n" + "ld1w { z17.s }, p0/Z, [x24]\n" + "addvl x24, x24, #1\n" + ".inst 0x658aa610 // bfcvt z16.h, p1/M, z16.s\n" + ".inst 0x648aa630 // bfcvtnt z16.h, p1/M, z17.s\n" + "st1h { z16.h }, p1, [x22]\n" "add x22, x22, %x[out_stride]\n" "bgt 10b\n" "11:" // Tail row loop: Column loop skip -- cgit v1.2.1