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 --- .../sme2_interleave4VL_block2_fp32_bf16.hpp | 140 ++++++++++----------- 1 file changed, 70 insertions(+), 70 deletions(-) (limited to 'src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme2_interleave4VL_block2_fp32_bf16.hpp') diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme2_interleave4VL_block2_fp32_bf16.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme2_interleave4VL_block2_fp32_bf16.hpp index 03575d7ff2..a8187f78e8 100644 --- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme2_interleave4VL_block2_fp32_bf16.hpp +++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme2_interleave4VL_block2_fp32_bf16.hpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2022-2023 Arm Limited. + * Copyright (c) 2022-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -32,63 +32,63 @@ void interleave_block<4, 2, VLType::SME, false>( { __asm__ __volatile__( ".inst 0xd503477f // SMSTART ZA\n" + "sub x14, %x[width], #0x1\n" + "mov x13, %x[in]\n" "cntw x23, ALL, MUL #2\n" - "cntw x10\n" + "cntw x11\n" "cntw x22, ALL, MUL #2\n" "cntw x20, ALL, MUL #3\n" "sub x21, x23, #0x1\n" ".inst 0x25207817 // ptrue pn15.b\n" - "whilelt p1.s, XZR, %x[height]\n" - "whilelt p14.s, x10, %x[height]\n" - "whilelt p13.s, x22, %x[height]\n" - "whilelt p12.s, x20, %x[height]\n" - "sub x9, %x[width], #0x1\n" + "whilelt p2.s, XZR, %x[height]\n" + "whilelt p1.s, x11, %x[height]\n" + "whilelt p14.s, x22, %x[height]\n" + "whilelt p13.s, x20, %x[height]\n" "cntw x20, ALL, MUL #2\n" - "ands x28, %x[width], x21\n" - "mov x27, %x[in]\n" - "add x9, x9, x20\n" - "csel x28, x28, x23, NE\n" - "add x26, x27, x10, LSL #3\n" - "mov x25, #0x0\n" - "udiv x9, x9, x20\n" - "add x28, x28, #0x1\n" + "ands x10, %x[width], x21\n" + "add x14, x14, x20\n" + "csel x10, x10, x23, NE\n" + "add x9, x13, x11, LSL #3\n" + "mov x28, #0x0\n" + "udiv x14, x14, x20\n" + "add x10, x10, #0x1\n" "mov x20, %x[width]\n" - "add x24, x26, x10, LSL #3\n" + "add x27, x9, x11, LSL #3\n" "ptrue p0.b\n" - "mov x23, %x[outptr_raw]\n" - "mov x22, %x[row_offset]\n" - "sub x9, x9, #0x1\n" - "lsr x28, x28, #0x1\n" + "mov x26, %x[outptr_raw]\n" + "mov x25, %x[row_offset]\n" + "sub x14, x14, #0x1\n" + "lsr x10, x10, #0x1\n" "mov x12, #0x0\n" - ".inst 0x25b44733 // whilelt pn11.s, x25, x20, VLx2\n" - "add x21, x24, x10, LSL #3\n" + ".inst 0x25b44794 // whilelt pn12.s, x28, x20, VLx2\n" + "add x24, x27, x11, LSL #3\n" "1:" // Width loop: Preamble: Loop - "ldr x20, [x27], #0x8\n" - ".inst 0x25306c28 // psel p8.s, p11.s/Z, p1.s[w12]\n" - ".inst 0x25306dca // psel p10.s, p11.s/Z, p14.s[w12]\n" - ".inst 0xa0164298 // ld1w { z24.s-z25.s }, pn8.s/Z, [x20, x22, LSL #2]\n" - "ldr x20, [x26], #0x8\n" - ".inst 0x25306da9 // psel p9.s, p11.s/Z, p13.s[w12]\n" - ".inst 0x25306d88 // psel p8.s, p11.s/Z, p12.s[w12]\n" - ".inst 0xa0164a82 // ld1w { z2.s-z3.s }, pn10.s/Z, [x20, x22, LSL #2]\n" + "ldr x23, [x13], #0x8\n" + ".inst 0x2530704b // psel p11.s, p12.s/Z, p2.s[w12]\n" + ".inst 0x2530702a // psel p10.s, p12.s/Z, p1.s[w12]\n" + "ldr x22, [x9], #0x8\n" + ".inst 0x253071c9 // psel p9.s, p12.s/Z, p14.s[w12]\n" + ".inst 0x253071a8 // psel p8.s, p12.s/Z, p13.s[w12]\n" + "ldr x21, [x27], #0x8\n" "ldr x20, [x24], #0x8\n" - ".inst 0xa016468a // ld1w { z10.s-z11.s }, pn9.s/Z, [x20, x22, LSL #2]\n" - ".inst 0xc160e318 // bfcvt z24.h, { z24.s-z25.s }\n" - ".inst 0xc160e042 // bfcvt z2.h, { z2.s-z3.s }\n" - "ldr x20, [x21], #0x8\n" - ".inst 0xa016428c // ld1w { z12.s-z13.s }, pn8.s/Z, [x20, x22, LSL #2]\n" + ".inst 0xa0194eea // ld1w { z10.s-z11.s }, pn11.s/Z, [x23, x25, LSL #2]\n" + ".inst 0xa0194ada // ld1w { z26.s-z27.s }, pn10.s/Z, [x22, x25, LSL #2]\n" + ".inst 0xa01946be // ld1w { z30.s-z31.s }, pn9.s/Z, [x21, x25, LSL #2]\n" + ".inst 0xa019428c // ld1w { z12.s-z13.s }, pn8.s/Z, [x20, x25, LSL #2]\n" ".inst 0xc160e14a // bfcvt z10.h, { z10.s-z11.s }\n" + ".inst 0xc160e35a // bfcvt z26.h, { z26.s-z27.s }\n" + ".inst 0xc0800140 // mova za0h.s[x12], p0/M, z10.s\n" + ".inst 0xc160e3de // bfcvt z30.h, { z30.s-z31.s }\n" + ".inst 0xc0800344 // mova za1h.s[x12], p0/M, z26.s\n" ".inst 0xc160e18c // bfcvt z12.h, { z12.s-z13.s }\n" - ".inst 0xc0800300 // mova za0h.s[x12], p0/M, z24.s\n" - ".inst 0xc0800044 // mova za1h.s[x12], p0/M, z2.s\n" - ".inst 0xc0800148 // mova za2h.s[x12], p0/M, z10.s\n" + ".inst 0xc08003c8 // mova za2h.s[x12], p0/M, z30.s\n" ".inst 0xc080018c // mova za3h.s[x12], p0/M, z12.s\n" "add x12, x12, #0x1\n" - "cmp x12, x10\n" + "cmp x12, x11\n" "blt 1b\n" - "incw x22, ALL, MUL #2\n" "incw x25, ALL, MUL #2\n" - "cbz x9, 5f\n" + "incw x28, ALL, MUL #2\n" + "cbz x14, 5f\n" "2:" // Width loop "mov x12, #0x0\n" "3:" // Width loop: Store: Loop @@ -97,44 +97,44 @@ void interleave_block<4, 2, VLType::SME, false>( ".inst 0xc0828119 // mova z25.s, p0/M, za2v.s[x12]\n" ".inst 0xc082819d // mova z29.s, p0/M, za3v.s[x12]\n" "add x12, x12, #0x1\n" - "cmp x12, x10\n" - ".inst 0xa160def1 // st1w { z17.s, z21.s, z25.s, z29.s }, pn15.b, [x23]\n" - "addvl x23, x23, #4\n" + "cmp x12, x11\n" + ".inst 0xa160df51 // st1w { z17.s, z21.s, z25.s, z29.s }, pn15.b, [x26]\n" + "addvl x26, x26, #4\n" "blt 3b\n" - "mov x27, %x[in]\n" - "add x26, x27, x10, LSL #3\n" + "mov x13, %x[in]\n" "mov x20, %x[width]\n" - "add x24, x26, x10, LSL #3\n" + "add x9, x13, x11, LSL #3\n" "mov x12, #0x0\n" - ".inst 0x25b44733 // whilelt pn11.s, x25, x20, VLx2\n" - "add x21, x24, x10, LSL #3\n" + "add x27, x9, x11, LSL #3\n" + ".inst 0x25b44794 // whilelt pn12.s, x28, x20, VLx2\n" + "add x24, x27, x11, LSL #3\n" "4:" // Width loop: Load: Loop - "ldr x20, [x27], #0x8\n" - ".inst 0x25306c28 // psel p8.s, p11.s/Z, p1.s[w12]\n" - ".inst 0x25306dca // psel p10.s, p11.s/Z, p14.s[w12]\n" - ".inst 0xa016428c // ld1w { z12.s-z13.s }, pn8.s/Z, [x20, x22, LSL #2]\n" - "ldr x20, [x26], #0x8\n" - ".inst 0x25306da9 // psel p9.s, p11.s/Z, p13.s[w12]\n" - ".inst 0x25306d88 // psel p8.s, p11.s/Z, p12.s[w12]\n" - ".inst 0xa0164a8e // ld1w { z14.s-z15.s }, pn10.s/Z, [x20, x22, LSL #2]\n" + "ldr x23, [x13], #0x8\n" + ".inst 0x2530704b // psel p11.s, p12.s/Z, p2.s[w12]\n" + ".inst 0x2530702a // psel p10.s, p12.s/Z, p1.s[w12]\n" + "ldr x22, [x9], #0x8\n" + ".inst 0x253071c9 // psel p9.s, p12.s/Z, p14.s[w12]\n" + ".inst 0x253071a8 // psel p8.s, p12.s/Z, p13.s[w12]\n" + "ldr x21, [x27], #0x8\n" "ldr x20, [x24], #0x8\n" - ".inst 0xa0164692 // ld1w { z18.s-z19.s }, pn9.s/Z, [x20, x22, LSL #2]\n" + ".inst 0xa0194eec // ld1w { z12.s-z13.s }, pn11.s/Z, [x23, x25, LSL #2]\n" + ".inst 0xa0194ace // ld1w { z14.s-z15.s }, pn10.s/Z, [x22, x25, LSL #2]\n" + ".inst 0xa01946b2 // ld1w { z18.s-z19.s }, pn9.s/Z, [x21, x25, LSL #2]\n" + ".inst 0xa019429e // ld1w { z30.s-z31.s }, pn8.s/Z, [x20, x25, LSL #2]\n" ".inst 0xc160e18c // bfcvt z12.h, { z12.s-z13.s }\n" ".inst 0xc160e1ce // bfcvt z14.h, { z14.s-z15.s }\n" - "ldr x20, [x21], #0x8\n" - ".inst 0xa016429e // ld1w { z30.s-z31.s }, pn8.s/Z, [x20, x22, LSL #2]\n" - ".inst 0xc160e252 // bfcvt z18.h, { z18.s-z19.s }\n" - ".inst 0xc160e3de // bfcvt z30.h, { z30.s-z31.s }\n" ".inst 0xc0800180 // mova za0h.s[x12], p0/M, z12.s\n" + ".inst 0xc160e252 // bfcvt z18.h, { z18.s-z19.s }\n" ".inst 0xc08001c4 // mova za1h.s[x12], p0/M, z14.s\n" + ".inst 0xc160e3de // bfcvt z30.h, { z30.s-z31.s }\n" ".inst 0xc0800248 // mova za2h.s[x12], p0/M, z18.s\n" ".inst 0xc08003cc // mova za3h.s[x12], p0/M, z30.s\n" "add x12, x12, #0x1\n" - "cmp x12, x10\n" + "cmp x12, x11\n" "blt 4b\n" - "subs x9, x9, #0x1\n" - "incw x22, ALL, MUL #2\n" + "subs x14, x14, #0x1\n" "incw x25, ALL, MUL #2\n" + "incw x28, ALL, MUL #2\n" "bgt 2b\n" "5:" // Width loop: Tails "mov x12, #0x0\n" @@ -144,16 +144,16 @@ void interleave_block<4, 2, VLType::SME, false>( ".inst 0xc0828119 // mova z25.s, p0/M, za2v.s[x12]\n" ".inst 0xc082819d // mova z29.s, p0/M, za3v.s[x12]\n" "add x12, x12, #0x1\n" - "cmp x12, x28\n" - ".inst 0xa160def1 // st1w { z17.s, z21.s, z25.s, z29.s }, pn15.b, [x23]\n" - "addvl x23, x23, #4\n" + "cmp x12, x10\n" + ".inst 0xa160df51 // st1w { z17.s, z21.s, z25.s, z29.s }, pn15.b, [x26]\n" + "addvl x26, x26, #4\n" "blt 6b\n" "7:" // End - "mov %x[outptr_raw], x23\n" + "mov %x[outptr_raw], x26\n" ".inst 0xd503467f // SMSTOP\n" : [outptr_raw] "+&r" (out) : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset), [width] "r" (width) - : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" + : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" ); } -- cgit v1.2.1