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_block4_s8_s8_summing.hpp | 192 ++++++++++----------- 1 file changed, 96 insertions(+), 96 deletions(-) (limited to 'src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block4_s8_s8_summing.hpp') diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block4_s8_s8_summing.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block4_s8_s8_summing.hpp index f5ee261964..ba8be81ade 100644 --- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block4_s8_s8_summing.hpp +++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme_interleave1VL_block4_s8_s8_summing.hpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2022-2023 Arm Limited. + * Copyright (c) 2022-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -32,18 +32,18 @@ void interleave_block<1, 4, VLType::SME, true>( { __asm__ __volatile__( ".inst 0xd503477f // SMSTART ZA\n" - "cntb x21\n" "mov x23, %x[width]\n" - "mov z18.b, #0x1\n" + "mov x21, %x[width]\n" + "cntb x20\n" "incb x23\n" - "mov x20, %x[width]\n" + "mov z18.b, #0x1\n" "mov z17.s, #0x0\n" - "sub x10, x21, #0x1\n" + "sub x10, x20, #0x1\n" "cntw x9\n" "sub x23, x23, #0x1\n" - "ands x10, x20, x10\n" - "udiv x23, x23, x21\n" // n_passes = ceildiv(width, VL) - "csel x10, x10, x21, NE\n" + "ands x10, x21, x10\n" + "udiv x23, x23, x20\n" // n_passes = ceildiv(width, VL) + "csel x10, x10, x20, NE\n" "lsl x22, %x[height], #0x1\n" // height * 2 "lsl x21, x9, #0x1\n" "sub x20, x23, #0x1\n" @@ -51,7 +51,7 @@ void interleave_block<1, 4, VLType::SME, true>( "whilelt p9.b, XZR, x22\n" "whilelt p8.b, x21, x22\n" "mov x28, #0x0\n" - "ptrue p2.b\n" + "ptrue p4.b\n" "lsr x20, x20, #0x1\n" // n_loops = (n_passes - 1) / 2 "and x27, x23, #0x1\n" // odd_tail = bool(n_passes & 0x1) "lsr x10, x10, #0x2\n" @@ -64,132 +64,132 @@ void interleave_block<1, 4, VLType::SME, true>( "whilelt p8.b, x28, %x[width]\n" "cbnz %x[first], 1f\n" "addvl x24, x24, #-1\n" - "ld1w { z17.s }, p2/Z, [x24]\n" + "ld1w { z17.s }, p4/Z, [x24]\n" "1:" // K loop: Load row sums: End "mov x23, %x[in]\n" - "ldr x22, [x23, #0x0]\n" "mov x12, #0x0\n" + "ldr x22, [x23, #0x0]\n" "ldr x21, [x23, #0x8]\n" "add x23, x23, #0x10\n" "cbz x26, 3f\n" "2:" // K loop: Charge: Loop - ".inst 0x25246140 // psel p0.b, p8.b/Z, p10.b[w12]\n" - ".inst 0xe01902c0 // ld1b { za0h.b[x12] }, p0/Z, [x22, x25]\n" + ".inst 0x25246141 // psel p1.b, p8.b/Z, p10.b[w12]\n" ".inst 0x25646140 // psel p0.b, p8.b/Z, p10.b[w12, #4]\n" + ".inst 0xe01906c0 // ld1b { za0h.b[x12] }, p1/Z, [x22, x25]\n" "ldr x22, [x23, #0x0]\n" ".inst 0xe01902a4 // ld1b { za0h.b[x12, #4] }, p0/Z, [x21, x25]\n" "add x12, x12, #0x8\n" - "cmp x12, x26, LSL #2\n" "ldr x21, [x23, #0x8]\n" "add x23, x23, #0x10\n" + "cmp x12, x26, LSL #2\n" "blt 2b\n" "3:" // K loop: Charge: End - ".inst 0x25246140 // psel p0.b, p8.b/Z, p10.b[w12]\n" - ".inst 0xe01902c0 // ld1b { za0h.b[x12] }, p0/Z, [x22, x25]\n" + ".inst 0x25246141 // psel p1.b, p8.b/Z, p10.b[w12]\n" ".inst 0x25646140 // psel p0.b, p8.b/Z, p10.b[w12, #4]\n" "mov x23, %x[in]\n" - ".inst 0xe01902a4 // ld1b { za0h.b[x12, #4] }, p0/Z, [x21, x25]\n" + "incb x28\n" + ".inst 0xe01906c0 // ld1b { za0h.b[x12] }, p1/Z, [x22, x25]\n" "ldr x22, [x23, #0x0]\n" - "incb x25\n" + ".inst 0xe01902a4 // ld1b { za0h.b[x12, #4] }, p0/Z, [x21, x25]\n" "ldr x21, [x23, #0x8]\n" "add x23, x23, #0x10\n" - "incb x28\n" + "incb x25\n" "cbz x20, 9f\n" "mov x20, x20\n" "4:" // K loop: Main loop "whilelt p8.b, x28, %x[width]\n" - "mov x13, #0x0\n" "mov x12, #0x0\n" + "mov x14, #0x0\n" "cbz x26, 6f\n" "5:" // K loop: Main loop: First: Loop - ".inst 0x25356140 // psel p0.b, p8.b/Z, p10.b[w13, #2]\n" - ".inst 0xe01922c2 // ld1b { za0h.b[x13, #2] }, p0/Z, [x22, x25]\n" - ".inst 0x25756140 // psel p0.b, p8.b/Z, p10.b[w13, #6]\n" + ".inst 0x25346143 // psel p3.b, p8.b/Z, p10.b[w12, #2]\n" + ".inst 0x25746142 // psel p2.b, p8.b/Z, p10.b[w12, #6]\n" + ".inst 0x25266d21 // psel p1.b, p11.b/Z, p9.b[w14]\n" + ".inst 0x252e6d20 // psel p0.b, p11.b/Z, p9.b[w14, #1]\n" + ".inst 0xe0190ec2 // ld1b { za0h.b[x12, #2] }, p3/Z, [x22, x25]\n" "ldr x22, [x23, #0x0]\n" - ".inst 0xe01922a6 // ld1b { za0h.b[x13, #6] }, p0/Z, [x21, x25]\n" - ".inst 0xc0828810 // mova z16.s, p2/M, za0v.s[x12]\n" - ".inst 0x25306d20 // psel p0.s, p11.s/Z, p9.s[w12]\n" - "sdot z17.s, z16.b, z18.b\n" + ".inst 0xe0190aa6 // ld1b { za0h.b[x12, #6] }, p2/Z, [x21, x25]\n" "ldr x21, [x23, #0x8]\n" - ".inst 0xe0bf8300 // st1w { za0v.s[x12] }, p0/Z, [x24, XZR, LSL #2]\n" - ".inst 0x25706d20 // psel p0.s, p11.s/Z, p9.s[w12, #1]\n" - ".inst 0xc0828830 // mova z16.s, p2/M, za0v.s[x12, #1]\n" - ".inst 0xe0a98301 // st1w { za0v.s[x12, #1] }, p0/Z, [x24, x9, LSL #2]\n" - "add x12, x12, #0x2\n" - "cmp x12, x26\n" - "sdot z17.s, z16.b, z18.b\n" "add x23, x23, #0x10\n" + "add x12, x12, #0x8\n" + ".inst 0xc082d010 // mova z16.s, p4/M, za0v.s[x14]\n" + ".inst 0xe0bfc700 // st1w { za0v.s[x14] }, p1/Z, [x24, XZR, LSL #2]\n" + "sdot z17.s, z16.b, z18.b\n" + ".inst 0xe0a9c301 // st1w { za0v.s[x14, #1] }, p0/Z, [x24, x9, LSL #2]\n" + ".inst 0xc082d030 // mova z16.s, p4/M, za0v.s[x14, #1]\n" + "add x14, x14, #0x2\n" + "cmp x14, x26\n" "addvl x24, x24, #2\n" - "add x13, x13, #0x8\n" + "sdot z17.s, z16.b, z18.b\n" "blt 5b\n" "6:" // K loop: Main loop: First: Tail - ".inst 0x25356140 // psel p0.b, p8.b/Z, p10.b[w13, #2]\n" - ".inst 0xe01922c2 // ld1b { za0h.b[x13, #2] }, p0/Z, [x22, x25]\n" - ".inst 0x25756140 // psel p0.b, p8.b/Z, p10.b[w13, #6]\n" - ".inst 0xe01922a6 // ld1b { za0h.b[x13, #6] }, p0/Z, [x21, x25]\n" - ".inst 0xc0828810 // mova z16.s, p2/M, za0v.s[x12]\n" - "sdot z17.s, z16.b, z18.b\n" + ".inst 0x25346143 // psel p3.b, p8.b/Z, p10.b[w12, #2]\n" + ".inst 0x25746142 // psel p2.b, p8.b/Z, p10.b[w12, #6]\n" + ".inst 0x25266d21 // psel p1.b, p11.b/Z, p9.b[w14]\n" + ".inst 0x252e6d20 // psel p0.b, p11.b/Z, p9.b[w14, #1]\n" "mov x23, %x[in]\n" - ".inst 0x25306d21 // psel p1.s, p11.s/Z, p9.s[w12]\n" - "ldr x22, [x23, #0x0]\n" - ".inst 0x25706d20 // psel p0.s, p11.s/Z, p9.s[w12, #1]\n" - ".inst 0xc0828830 // mova z16.s, p2/M, za0v.s[x12, #1]\n" - "ldr x21, [x23, #0x8]\n" - ".inst 0xe0bf8700 // st1w { za0v.s[x12] }, p1/Z, [x24, XZR, LSL #2]\n" "whilelt p9.b, x28, %x[width]\n" + ".inst 0xe0190ec2 // ld1b { za0h.b[x12, #2] }, p3/Z, [x22, x25]\n" + "ldr x22, [x23, #0x0]\n" "incb x28\n" + "mov x13, #0x0\n" + ".inst 0xe0190aa6 // ld1b { za0h.b[x12, #6] }, p2/Z, [x21, x25]\n" + "ldr x21, [x23, #0x8]\n" "add x23, x23, #0x10\n" - ".inst 0xe0a98301 // st1w { za0v.s[x12, #1] }, p0/Z, [x24, x9, LSL #2]\n" - "sdot z17.s, z16.b, z18.b\n" - "addvl x24, x24, #2\n" "incb x25\n" + ".inst 0xc082d010 // mova z16.s, p4/M, za0v.s[x14]\n" + ".inst 0xe0bfc700 // st1w { za0v.s[x14] }, p1/Z, [x24, XZR, LSL #2]\n" "whilelt p8.b, x28, %x[width]\n" - "mov x13, #0x0\n" "mov x12, #0x0\n" + "sdot z17.s, z16.b, z18.b\n" + ".inst 0xc082d030 // mova z16.s, p4/M, za0v.s[x14, #1]\n" + ".inst 0xe0a9c301 // st1w { za0v.s[x14, #1] }, p0/Z, [x24, x9, LSL #2]\n" + "addvl x24, x24, #2\n" + "sdot z17.s, z16.b, z18.b\n" "cbz x26, 8f\n" "7:" // K loop: Main loop: Second: Loop - ".inst 0x25256140 // psel p0.b, p8.b/Z, p10.b[w13]\n" - ".inst 0xe01922c0 // ld1b { za0h.b[x13] }, p0/Z, [x22, x25]\n" - ".inst 0x25656140 // psel p0.b, p8.b/Z, p10.b[w13, #4]\n" + ".inst 0x25256143 // psel p3.b, p8.b/Z, p10.b[w13]\n" + ".inst 0x25656142 // psel p2.b, p8.b/Z, p10.b[w13, #4]\n" + ".inst 0x25246d21 // psel p1.b, p11.b/Z, p9.b[w12]\n" + ".inst 0x252c6d20 // psel p0.b, p11.b/Z, p9.b[w12, #1]\n" + ".inst 0xe0192ec0 // ld1b { za0h.b[x13] }, p3/Z, [x22, x25]\n" "ldr x22, [x23, #0x0]\n" - ".inst 0xe01922a4 // ld1b { za0h.b[x13, #4] }, p0/Z, [x21, x25]\n" - ".inst 0xc0828910 // mova z16.s, p2/M, za2v.s[x12]\n" - ".inst 0x25306d20 // psel p0.s, p11.s/Z, p9.s[w12]\n" - "sdot z17.s, z16.b, z18.b\n" + ".inst 0xe0192aa4 // ld1b { za0h.b[x13, #4] }, p2/Z, [x21, x25]\n" "ldr x21, [x23, #0x8]\n" - ".inst 0xe0bf8308 // st1w { za2v.s[x12] }, p0/Z, [x24, XZR, LSL #2]\n" - ".inst 0x25706d20 // psel p0.s, p11.s/Z, p9.s[w12, #1]\n" - ".inst 0xc0828930 // mova z16.s, p2/M, za2v.s[x12, #1]\n" + "add x23, x23, #0x10\n" + "add x13, x13, #0x8\n" + ".inst 0xc0829110 // mova z16.s, p4/M, za2v.s[x12]\n" + ".inst 0xe0bf8708 // st1w { za2v.s[x12] }, p1/Z, [x24, XZR, LSL #2]\n" + "sdot z17.s, z16.b, z18.b\n" ".inst 0xe0a98309 // st1w { za2v.s[x12, #1] }, p0/Z, [x24, x9, LSL #2]\n" + ".inst 0xc0829130 // mova z16.s, p4/M, za2v.s[x12, #1]\n" "add x12, x12, #0x2\n" "cmp x12, x26\n" - "sdot z17.s, z16.b, z18.b\n" - "add x23, x23, #0x10\n" "addvl x24, x24, #2\n" - "add x13, x13, #0x8\n" + "sdot z17.s, z16.b, z18.b\n" "blt 7b\n" "8:" // K loop: Main loop: Second: Tail - ".inst 0x25256140 // psel p0.b, p8.b/Z, p10.b[w13]\n" - ".inst 0xe01922c0 // ld1b { za0h.b[x13] }, p0/Z, [x22, x25]\n" - ".inst 0x25656140 // psel p0.b, p8.b/Z, p10.b[w13, #4]\n" - ".inst 0xe01922a4 // ld1b { za0h.b[x13, #4] }, p0/Z, [x21, x25]\n" - ".inst 0xc0828910 // mova z16.s, p2/M, za2v.s[x12]\n" - "sdot z17.s, z16.b, z18.b\n" + ".inst 0x25256143 // psel p3.b, p8.b/Z, p10.b[w13]\n" + ".inst 0x25656142 // psel p2.b, p8.b/Z, p10.b[w13, #4]\n" + ".inst 0x25246d21 // psel p1.b, p11.b/Z, p9.b[w12]\n" + ".inst 0x252c6d20 // psel p0.b, p11.b/Z, p9.b[w12, #1]\n" "mov x23, %x[in]\n" - ".inst 0x25306d21 // psel p1.s, p11.s/Z, p9.s[w12]\n" - "ldr x22, [x23, #0x0]\n" - ".inst 0x25706d20 // psel p0.s, p11.s/Z, p9.s[w12, #1]\n" - ".inst 0xc0828930 // mova z16.s, p2/M, za2v.s[x12, #1]\n" - "ldr x21, [x23, #0x8]\n" - ".inst 0xe0bf8708 // st1w { za2v.s[x12] }, p1/Z, [x24, XZR, LSL #2]\n" "whilelt p9.b, x28, %x[width]\n" + ".inst 0xe0192ec0 // ld1b { za0h.b[x13] }, p3/Z, [x22, x25]\n" + "ldr x22, [x23, #0x0]\n" "subs x20, x20, #0x1\n" + "incb x28\n" + ".inst 0xe0192aa4 // ld1b { za0h.b[x13, #4] }, p2/Z, [x21, x25]\n" + "ldr x21, [x23, #0x8]\n" "add x23, x23, #0x10\n" - ".inst 0xe0a98309 // st1w { za2v.s[x12, #1] }, p0/Z, [x24, x9, LSL #2]\n" + "incb x25\n" + ".inst 0xc0829110 // mova z16.s, p4/M, za2v.s[x12]\n" + ".inst 0xe0bf8708 // st1w { za2v.s[x12] }, p1/Z, [x24, XZR, LSL #2]\n" "sdot z17.s, z16.b, z18.b\n" + ".inst 0xc0829130 // mova z16.s, p4/M, za2v.s[x12, #1]\n" + ".inst 0xe0a98309 // st1w { za2v.s[x12, #1] }, p0/Z, [x24, x9, LSL #2]\n" "addvl x24, x24, #2\n" - "incb x28\n" - "incb x25\n" + "sdot z17.s, z16.b, z18.b\n" "bgt 4b\n" "9:" // K loop: Tails "cbnz x27, 12f\n" @@ -198,17 +198,17 @@ void interleave_block<1, 4, VLType::SME, true>( "mov x13, #0x0\n" "mov x12, #0x0\n" "10:" // K loop: Tails: Even: First - ".inst 0x25306d20 // psel p0.s, p11.s/Z, p9.s[w12]\n" - ".inst 0xe0bf8300 // st1w { za0v.s[x12] }, p0/Z, [x24, XZR, LSL #2]\n" - "ldr x20, [x23, #0x0]\n" - ".inst 0xc0828810 // mova z16.s, p2/M, za0v.s[x12]\n" - "add x12, x12, #0x1\n" + ".inst 0x25306d21 // psel p1.s, p11.s/Z, p9.s[w12]\n" ".inst 0x25356140 // psel p0.b, p8.b/Z, p10.b[w13, #2]\n" + ".inst 0xc0829010 // mova z16.s, p4/M, za0v.s[x12]\n" "sdot z17.s, z16.b, z18.b\n" - ".inst 0xe0192282 // ld1b { za0h.b[x13, #2] }, p0/Z, [x20, x25]\n" + ".inst 0xe0bf8700 // st1w { za0v.s[x12] }, p1/Z, [x24, XZR, LSL #2]\n" + "add x12, x12, #0x1\n" + "addvl x24, x24, #1\n" + "ldr x20, [x23, #0x0]\n" "cmp x12, x9\n" "add x23, x23, #0x8\n" - "addvl x24, x24, #1\n" + ".inst 0xe0192282 // ld1b { za0h.b[x13, #2] }, p0/Z, [x20, x25]\n" "add x13, x13, #0x4\n" "blt 10b\n" "whilelt p9.b, x28, %x[width]\n" @@ -217,13 +217,13 @@ void interleave_block<1, 4, VLType::SME, true>( "mov x12, #0x0\n" "11:" // K loop: Tails: Even: Second ".inst 0x25306d20 // psel p0.s, p11.s/Z, p9.s[w12]\n" + ".inst 0xc0829110 // mova z16.s, p4/M, za2v.s[x12]\n" + "add x20, x20, #0x4\n" + "sdot z17.s, z16.b, z18.b\n" ".inst 0xe0bf8308 // st1w { za2v.s[x12] }, p0/Z, [x24, XZR, LSL #2]\n" - ".inst 0xc0828910 // mova z16.s, p2/M, za2v.s[x12]\n" "add x12, x12, #0x1\n" - "cmp x12, x10\n" - "sdot z17.s, z16.b, z18.b\n" "addvl x24, x24, #1\n" - "add x20, x20, #0x4\n" + "cmp x12, x10\n" "blt 11b\n" "whilelt p8.b, x28, %x[width]\n" "b 14f\n" @@ -231,21 +231,21 @@ void interleave_block<1, 4, VLType::SME, true>( "mov x12, #0x0\n" "13:" // K loop: Tails: Odd: Loop ".inst 0x25306d20 // psel p0.s, p11.s/Z, p9.s[w12]\n" + ".inst 0xc0829010 // mova z16.s, p4/M, za0v.s[x12]\n" + "sdot z17.s, z16.b, z18.b\n" ".inst 0xe0bf8300 // st1w { za0v.s[x12] }, p0/Z, [x24, XZR, LSL #2]\n" - ".inst 0xc0828810 // mova z16.s, p2/M, za0v.s[x12]\n" "add x12, x12, #0x1\n" - "cmp x12, x10\n" - "sdot z17.s, z16.b, z18.b\n" "addvl x24, x24, #1\n" + "cmp x12, x10\n" "blt 13b\n" "14:" // K loop: End - "st1w { z17.s }, p2, [x24]\n" + "st1w { z17.s }, p4, [x24]\n" "addvl x24, x24, #1\n" - "mov %x[out], x24\n" ".inst 0xd503467f // SMSTOP\n" + "mov %x[out], x24\n" : [out] "+&r" (out) : [first] "r" (first), [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", "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" + : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "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