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_interleave1VL_fp32_fp32.hpp | 144 ++++++++++----------- 1 file changed, 72 insertions(+), 72 deletions(-) (limited to 'src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_fp32_fp32.hpp') diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_fp32_fp32.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_fp32_fp32.hpp index ab290649fd..b76ec57d22 100644 --- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_fp32_fp32.hpp +++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_fp32_fp32.hpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2022-2023 Arm Limited. + * Copyright (c) 2022-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -32,24 +32,24 @@ void interleave_block<1, 1, VLType::SME, false>( { __asm__ __volatile__( ".inst 0xd503477f // SMSTART ZA\n" - "mov x22, %x[width]\n" - "incw x22\n" - "cntw x10\n" - "sub x22, x22, #0x1\n" - "udiv x22, x22, x10\n" // n_passes = ceildiv(width, VL) "mov x21, %x[width]\n" + "mov x20, %x[width]\n" + "incw x21\n" + "cntw x10\n" + "sub x21, x21, #0x1\n" "sub x9, x10, #0x1\n" - "sub x20, x22, #0x1\n" - "ands x9, x21, x9\n" + "udiv x21, x21, x10\n" // n_passes = ceildiv(width, VL) + "ands x9, x20, x9\n" + "sub x20, x21, #0x1\n" "sub x28, x10, #0x2\n" "mov x27, #0x0\n" "mov x26, %x[in]\n" - "ldr x25, [x26, #0x0]\n" "lsr x20, x20, #0x1\n" // n_loops = (n_passes - 1) / 2 - "and x24, x22, #0x1\n" // odd_tail = bool(n_passes & 0x1) - "ldr x23, [x26, #0x8]\n" + "and x25, x21, #0x1\n" // odd_tail = bool(n_passes & 0x1) + "ldr x24, [x26, #0x0]\n" "csel x9, x9, x10, NE\n" "ptrue p11.s\n" + "ldr x23, [x26, #0x8]\n" "whilelt p10.s, XZR, %x[height]\n" "mov x22, %x[row_offset]\n" "mov x21, %x[out]\n" @@ -59,119 +59,119 @@ void interleave_block<1, 1, VLType::SME, false>( "mov x12, #0x0\n" "cbz x28, 2f\n" "1:" // K loop: Charge: Loop - ".inst 0x25306140 // psel p0.s, p8.s/Z, p10.s[w12]\n" - ".inst 0xe0960320 // ld1w { za0h.s[x12] }, p0/Z, [x25, x22, LSL #2]\n" + ".inst 0x25306141 // psel p1.s, p8.s/Z, p10.s[w12]\n" ".inst 0x25706140 // psel p0.s, p8.s/Z, p10.s[w12, #1]\n" - "ldr x25, [x26, #0x0]\n" + ".inst 0xe0960700 // ld1w { za0h.s[x12] }, p1/Z, [x24, x22, LSL #2]\n" + "ldr x24, [x26, #0x0]\n" ".inst 0xe09602e1 // ld1w { za0h.s[x12, #1] }, p0/Z, [x23, x22, LSL #2]\n" "add x12, x12, #0x2\n" - "cmp x12, x28\n" "ldr x23, [x26, #0x8]\n" "add x26, x26, #0x10\n" + "cmp x12, x28\n" "blt 1b\n" "2:" // K loop: Charge: End - ".inst 0x25306140 // psel p0.s, p8.s/Z, p10.s[w12]\n" - ".inst 0xe0960320 // ld1w { za0h.s[x12] }, p0/Z, [x25, x22, LSL #2]\n" + ".inst 0x25306141 // psel p1.s, p8.s/Z, p10.s[w12]\n" ".inst 0x25706140 // psel p0.s, p8.s/Z, p10.s[w12, #1]\n" "mov x26, %x[in]\n" + "incw x27\n" + ".inst 0xe0960700 // ld1w { za0h.s[x12] }, p1/Z, [x24, x22, LSL #2]\n" + "ldr x24, [x26, #0x0]\n" ".inst 0xe09602e1 // ld1w { za0h.s[x12, #1] }, p0/Z, [x23, x22, LSL #2]\n" - "ldr x25, [x26, #0x0]\n" - "incw x22\n" "ldr x23, [x26, #0x8]\n" "add x26, x26, #0x10\n" - "incw x27\n" + "incw x22\n" "cbz x20, 8f\n" "mov x20, x20\n" "3:" // K loop: Main loop "whilelt p8.s, x27, %x[width]\n" - "mov x12, #0x0\n" + "mov x13, #0x0\n" "cbz x28, 5f\n" "4:" // K loop: Main loop: First: Loop - ".inst 0x25306140 // psel p0.s, p8.s/Z, p10.s[w12]\n" - ".inst 0xe0960328 // ld1w { za2h.s[x12] }, p0/Z, [x25, x22, LSL #2]\n" - ".inst 0x25706141 // psel p1.s, p8.s/Z, p10.s[w12, #1]\n" - "ldr x25, [x26, #0x0]\n" - ".inst 0x25306d20 // psel p0.s, p11.s/Z, p9.s[w12]\n" - ".inst 0xe09606e9 // ld1w { za2h.s[x12, #1] }, p1/Z, [x23, x22, LSL #2]\n" + ".inst 0x25316143 // psel p3.s, p8.s/Z, p10.s[w13]\n" + ".inst 0x25716142 // psel p2.s, p8.s/Z, p10.s[w13, #1]\n" + ".inst 0x25316d21 // psel p1.s, p11.s/Z, p9.s[w13]\n" + ".inst 0x25716d20 // psel p0.s, p11.s/Z, p9.s[w13, #1]\n" + ".inst 0xe0962f08 // ld1w { za2h.s[x13] }, p3/Z, [x24, x22, LSL #2]\n" + "ldr x24, [x26, #0x0]\n" + ".inst 0xe0962ae9 // ld1w { za2h.s[x13, #1] }, p2/Z, [x23, x22, LSL #2]\n" "ldr x23, [x26, #0x8]\n" - ".inst 0xe0bf82a0 // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n" - ".inst 0x25706d20 // psel p0.s, p11.s/Z, p9.s[w12, #1]\n" - ".inst 0xe0aa82a1 // st1w { za0v.s[x12, #1] }, p0/Z, [x21, x10, LSL #2]\n" - "add x12, x12, #0x2\n" - "cmp x12, x28\n" + ".inst 0xe0bfa6a0 // st1w { za0v.s[x13] }, p1/Z, [x21, XZR, LSL #2]\n" "add x26, x26, #0x10\n" + ".inst 0xe0aaa2a1 // st1w { za0v.s[x13, #1] }, p0/Z, [x21, x10, LSL #2]\n" + "add x13, x13, #0x2\n" "addvl x21, x21, #2\n" + "cmp x13, x28\n" "blt 4b\n" "5:" // K loop: Main loop: First: Tail - ".inst 0x25306140 // psel p0.s, p8.s/Z, p10.s[w12]\n" - ".inst 0xe0960328 // ld1w { za2h.s[x12] }, p0/Z, [x25, x22, LSL #2]\n" + ".inst 0x25316143 // psel p3.s, p8.s/Z, p10.s[w13]\n" + ".inst 0x25716142 // psel p2.s, p8.s/Z, p10.s[w13, #1]\n" + ".inst 0x25316d21 // psel p1.s, p11.s/Z, p9.s[w13]\n" + ".inst 0x25716d20 // psel p0.s, p11.s/Z, p9.s[w13, #1]\n" "mov x26, %x[in]\n" - "ldr x25, [x26, #0x0]\n" - ".inst 0x25706141 // psel p1.s, p8.s/Z, p10.s[w12, #1]\n" - ".inst 0x25306d20 // psel p0.s, p11.s/Z, p9.s[w12]\n" - ".inst 0xe09606e9 // ld1w { za2h.s[x12, #1] }, p1/Z, [x23, x22, LSL #2]\n" - "ldr x23, [x26, #0x8]\n" - ".inst 0xe0bf82a0 // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n" - ".inst 0x25706d20 // psel p0.s, p11.s/Z, p9.s[w12, #1]\n" "whilelt p9.s, x27, %x[width]\n" + ".inst 0xe0962f08 // ld1w { za2h.s[x13] }, p3/Z, [x24, x22, LSL #2]\n" + "ldr x24, [x26, #0x0]\n" "incw x27\n" + "mov x12, #0x0\n" + ".inst 0xe0962ae9 // ld1w { za2h.s[x13, #1] }, p2/Z, [x23, x22, LSL #2]\n" + "ldr x23, [x26, #0x8]\n" + ".inst 0xe0bfa6a0 // st1w { za0v.s[x13] }, p1/Z, [x21, XZR, LSL #2]\n" "add x26, x26, #0x10\n" - ".inst 0xe0aa82a1 // st1w { za0v.s[x12, #1] }, p0/Z, [x21, x10, LSL #2]\n" + ".inst 0xe0aaa2a1 // st1w { za0v.s[x13, #1] }, p0/Z, [x21, x10, LSL #2]\n" "addvl x21, x21, #2\n" "incw x22\n" "whilelt p8.s, x27, %x[width]\n" - "mov x12, #0x0\n" "cbz x28, 7f\n" "6:" // K loop: Main loop: Second: Loop - ".inst 0x25306140 // psel p0.s, p8.s/Z, p10.s[w12]\n" - ".inst 0xe0960320 // ld1w { za0h.s[x12] }, p0/Z, [x25, x22, LSL #2]\n" - ".inst 0x25706141 // psel p1.s, p8.s/Z, p10.s[w12, #1]\n" - "ldr x25, [x26, #0x0]\n" - ".inst 0x25306d20 // psel p0.s, p11.s/Z, p9.s[w12]\n" - ".inst 0xe09606e1 // ld1w { za0h.s[x12, #1] }, p1/Z, [x23, x22, LSL #2]\n" - "ldr x23, [x26, #0x8]\n" - ".inst 0xe0bf82a8 // st1w { za2v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n" + ".inst 0x25306143 // psel p3.s, p8.s/Z, p10.s[w12]\n" + ".inst 0x25706142 // psel p2.s, p8.s/Z, p10.s[w12, #1]\n" + ".inst 0x25306d21 // psel p1.s, p11.s/Z, p9.s[w12]\n" ".inst 0x25706d20 // psel p0.s, p11.s/Z, p9.s[w12, #1]\n" + ".inst 0xe0960f00 // ld1w { za0h.s[x12] }, p3/Z, [x24, x22, LSL #2]\n" + "ldr x24, [x26, #0x0]\n" + ".inst 0xe0960ae1 // ld1w { za0h.s[x12, #1] }, p2/Z, [x23, x22, LSL #2]\n" + "ldr x23, [x26, #0x8]\n" + ".inst 0xe0bf86a8 // st1w { za2v.s[x12] }, p1/Z, [x21, XZR, LSL #2]\n" + "add x26, x26, #0x10\n" ".inst 0xe0aa82a9 // st1w { za2v.s[x12, #1] }, p0/Z, [x21, x10, LSL #2]\n" "add x12, x12, #0x2\n" - "cmp x12, x28\n" - "add x26, x26, #0x10\n" "addvl x21, x21, #2\n" + "cmp x12, x28\n" "blt 6b\n" "7:" // K loop: Main loop: Second: Tail - ".inst 0x25306140 // psel p0.s, p8.s/Z, p10.s[w12]\n" - ".inst 0xe0960320 // ld1w { za0h.s[x12] }, p0/Z, [x25, x22, LSL #2]\n" - "mov x26, %x[in]\n" - "ldr x25, [x26, #0x0]\n" - ".inst 0x25706141 // psel p1.s, p8.s/Z, p10.s[w12, #1]\n" - ".inst 0x25306d20 // psel p0.s, p11.s/Z, p9.s[w12]\n" - ".inst 0xe09606e1 // ld1w { za0h.s[x12, #1] }, p1/Z, [x23, x22, LSL #2]\n" - "ldr x23, [x26, #0x8]\n" - ".inst 0xe0bf82a8 // st1w { za2v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n" + ".inst 0x25306143 // psel p3.s, p8.s/Z, p10.s[w12]\n" + ".inst 0x25706142 // psel p2.s, p8.s/Z, p10.s[w12, #1]\n" + ".inst 0x25306d21 // psel p1.s, p11.s/Z, p9.s[w12]\n" ".inst 0x25706d20 // psel p0.s, p11.s/Z, p9.s[w12, #1]\n" + "mov x26, %x[in]\n" "whilelt p9.s, x27, %x[width]\n" + ".inst 0xe0960f00 // ld1w { za0h.s[x12] }, p3/Z, [x24, x22, LSL #2]\n" + "ldr x24, [x26, #0x0]\n" "subs x20, x20, #0x1\n" + "incw x27\n" + ".inst 0xe0960ae1 // ld1w { za0h.s[x12, #1] }, p2/Z, [x23, x22, LSL #2]\n" + "ldr x23, [x26, #0x8]\n" + ".inst 0xe0bf86a8 // st1w { za2v.s[x12] }, p1/Z, [x21, XZR, LSL #2]\n" "add x26, x26, #0x10\n" ".inst 0xe0aa82a9 // st1w { za2v.s[x12, #1] }, p0/Z, [x21, x10, LSL #2]\n" "addvl x21, x21, #2\n" - "incw x27\n" "incw x22\n" "bgt 3b\n" "8:" // K loop: Tails - "cbnz x24, 11f\n" + "cbnz x25, 11f\n" "mov x26, %x[in]\n" "whilelt p8.s, x27, %x[width]\n" "mov x12, #0x0\n" "9:" // K loop: Tails: Even: First - ".inst 0x25306d20 // psel p0.s, p11.s/Z, p9.s[w12]\n" - ".inst 0xe0bf82a0 // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n" - "ldr x20, [x26, #0x0]\n" + ".inst 0x25306d21 // psel p1.s, p11.s/Z, p9.s[w12]\n" ".inst 0x25306140 // psel p0.s, p8.s/Z, p10.s[w12]\n" + ".inst 0xe0bf86a0 // st1w { za0v.s[x12] }, p1/Z, [x21, XZR, LSL #2]\n" + "addvl x21, x21, #1\n" + "ldr x20, [x26, #0x0]\n" + "add x26, x26, #0x8\n" ".inst 0xe0960288 // ld1w { za2h.s[x12] }, p0/Z, [x20, x22, LSL #2]\n" "add x12, x12, #0x1\n" "cmp x12, x10\n" - "add x26, x26, #0x8\n" - "addvl x21, x21, #1\n" "blt 9b\n" "whilelt p9.s, x27, %x[width]\n" "whilelt p8.s, x27, %x[width]\n" @@ -180,8 +180,8 @@ void interleave_block<1, 1, VLType::SME, false>( ".inst 0x25306d20 // psel p0.s, p11.s/Z, p9.s[w12]\n" ".inst 0xe0bf82a8 // st1w { za2v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n" "add x12, x12, #0x1\n" - "cmp x12, x9\n" "addvl x21, x21, #1\n" + "cmp x12, x9\n" "blt 10b\n" "whilelt p8.s, x27, %x[width]\n" "b 13f\n" @@ -191,15 +191,15 @@ void interleave_block<1, 1, VLType::SME, false>( ".inst 0x25306d20 // psel p0.s, p11.s/Z, p9.s[w12]\n" ".inst 0xe0bf82a0 // st1w { za0v.s[x12] }, p0/Z, [x21, XZR, LSL #2]\n" "add x12, x12, #0x1\n" - "cmp x12, x9\n" "addvl x21, x21, #1\n" + "cmp x12, x9\n" "blt 12b\n" "13:" // K loop: End "mov %x[out], x21\n" ".inst 0xd503467f // SMSTOP\n" : [out] "+&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", "x12", "x13", "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