From 74921eee924625426429044decefe3673561b174 Mon Sep 17 00:00:00 2001 From: Michael Tyler Date: Wed, 12 Apr 2023 17:43:17 +0100 Subject: Update CPU kernel implementations and guard directives Resolves COMPMID-6023 Change-Id: I868975d14c4f98af6716726feda22405a6a4c891 Signed-off-by: Michael Tyler Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9686 Tested-by: Arm Jenkins Reviewed-by: Viet-Hoa Do Comments-Addressed: Arm Jenkins Benchmark: Arm Jenkins --- .../generic_direct.cpp | 273 ++++++++++----------- 1 file changed, 136 insertions(+), 137 deletions(-) (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp') diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp index f4027df375..d8ca3d7437 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp @@ -116,9 +116,9 @@ void a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst_direct_impl( "add x28, x9, x25, LSL #1\n" "add x12, x12, x20, LSL #1\n" // outptrs[0] += offset * sizeof(__fp16) "add x20, %x[params_struct], %[offsetof_args_min]\n" - "ld1r { v18.8h }, [x20]\n" + "ld1r { v27.8h }, [x20]\n" "add x20, %x[params_struct], %[offsetof_args_max]\n" - "ld1r { v17.8h }, [x20]\n" + "ld1r { v26.8h }, [x20]\n" "add x27, x28, x25, LSL #1\n" "add x26, x11, x15\n" "add x25, x12, x24, LSL #1\n" @@ -126,7 +126,7 @@ void a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst_direct_impl( "mov x21, #0x0\n" "sub x20, XZR, x23\n" "cbz x22, 4f\n" - "ldr q16, [x10, #0x0]\n" + "ldr q25, [x10, #0x0]\n" "ldr q0, [x10, #0x10]\n" "cmp x23, x22, LSL #4\n" "ldr q1, [x10, #0x20]\n" @@ -145,162 +145,162 @@ void a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst_direct_impl( "ldr q13, [x28, x15]\n" "bge 3f\n" "2:" // Tile loop: Channel loop - "mov v28.16b, v16.16b\n fmla v28.8h, v4.8h, v9.8h\n" - "mov v29.16b, v16.16b\n fmla v29.8h, v3.8h, v9.8h\n" + "mov v24.16b, v25.16b\n fmla v24.8h, v4.8h, v9.8h\n" + "mov v23.16b, v25.16b\n fmla v23.8h, v3.8h, v9.8h\n" "add x23, x23, #0x10\n" "cmp x23, x22, LSL #4\n" - "mov v30.16b, v16.16b\n fmla v30.8h, v1.8h, v9.8h\n" - "mov v31.16b, v16.16b\n fmla v31.8h, v0.8h, v9.8h\n" - "ld1 { v9.8h }, [x27]\n" - "ldr q16, [x10, #0x0]\n" - "fmla v28.8h, v0.8h, v10.8h\n" - "ldr q10, [x28, x11]\n" - "fmla v29.8h, v2.8h, v11.8h\n" - "ldr q11, [x27, x26]\n" - "fmla v30.8h, v2.8h, v12.8h\n" - "fmla v31.8h, v1.8h, v12.8h\n" + "mov v22.16b, v25.16b\n fmla v22.8h, v1.8h, v9.8h\n" + "mov v21.16b, v25.16b\n fmla v21.8h, v0.8h, v9.8h\n" + "ld1 { v18.8h }, [x27]\n" + "ldr q25, [x10, #0x0]\n" + "fmla v24.8h, v0.8h, v10.8h\n" + "ldr q20, [x28, x11]\n" + "fmla v23.8h, v2.8h, v11.8h\n" + "ldr q17, [x27, x26]\n" + "fmla v22.8h, v2.8h, v12.8h\n" + "fmla v21.8h, v1.8h, v12.8h\n" "add x20, x20, #0x10\n" "add x21, x21, #0x10\n" - "fmla v28.8h, v5.8h, v12.8h\n" - "fmla v29.8h, v4.8h, v12.8h\n" - "ldr q12, [x13, x15]\n" - "fmla v30.8h, v6.8h, v9.8h\n" - "ldr q9, [x13, x11]\n" - "fmla v31.8h, v3.8h, v13.8h\n" + "fmla v24.8h, v5.8h, v12.8h\n" + "fmla v23.8h, v4.8h, v12.8h\n" + "ldr q16, [x13, x15]\n" + "fmla v22.8h, v6.8h, v18.8h\n" + "ldr q18, [x13, x11]\n" + "fmla v21.8h, v3.8h, v13.8h\n" "add x13, x13, #0x10\n" - "fmla v28.8h, v7.8h, v13.8h\n" - "fmla v29.8h, v6.8h, v13.8h\n" - "fmla v30.8h, v4.8h, v13.8h\n" - "fmla v31.8h, v8.8h, v11.8h\n" - "ld1 { v11.8h }, [x9]\n" - "fmla v28.8h, v1.8h, v12.8h\n" - "fmla v29.8h, v0.8h, v12.8h\n" - "ldr q12, [x9, x26]\n" + "fmla v24.8h, v7.8h, v13.8h\n" + "fmla v23.8h, v6.8h, v13.8h\n" + "fmla v22.8h, v4.8h, v13.8h\n" + "fmla v21.8h, v8.8h, v17.8h\n" + "ld1 { v17.8h }, [x9]\n" + "fmla v24.8h, v1.8h, v16.8h\n" + "fmla v23.8h, v0.8h, v16.8h\n" + "ldr q16, [x9, x26]\n" "add x9, x9, #0x10\n" - "fmla v30.8h, v5.8h, v10.8h\n" - "fmla v31.8h, v4.8h, v10.8h\n" + "fmla v22.8h, v5.8h, v20.8h\n" + "fmla v21.8h, v4.8h, v20.8h\n" "ldr q4, [x10, #0x50]\n" - "fmla v28.8h, v2.8h, v9.8h\n" - "fmla v29.8h, v1.8h, v9.8h\n" - "ld1 { v9.8h }, [x28]\n" + "fmla v24.8h, v2.8h, v18.8h\n" + "fmla v23.8h, v1.8h, v18.8h\n" + "ld1 { v19.8h }, [x28]\n" "ldr q1, [x10, #0x20]\n" - "fmla v30.8h, v0.8h, v11.8h\n" + "fmla v22.8h, v0.8h, v17.8h\n" "ldr q0, [x10, #0x10]\n" - "fmla v31.8h, v2.8h, v12.8h\n" + "fmla v21.8h, v2.8h, v16.8h\n" "ldr q2, [x10, #0x30]\n" - "fmla v28.8h, v8.8h, v10.8h\n" - "fmla v29.8h, v7.8h, v10.8h\n" - "ldr q10, [x28, x26]\n" + "fmla v24.8h, v8.8h, v20.8h\n" + "fmla v23.8h, v7.8h, v20.8h\n" + "ldr q18, [x28, x26]\n" "add x28, x28, #0x10\n" "ldr q13, [x28, x15]\n" - "fmla v30.8h, v3.8h, v9.8h\n" - "fmla v31.8h, v5.8h, v10.8h\n" - "fmla v28.8h, v3.8h, v11.8h\n" - "ldr q11, [x27, x15]\n" + "fmla v22.8h, v3.8h, v19.8h\n" + "fmla v21.8h, v5.8h, v18.8h\n" + "fmla v24.8h, v3.8h, v17.8h\n" + "ldr q17, [x27, x15]\n" "ldr q3, [x10, #0x40]\n" - "fmla v29.8h, v5.8h, v12.8h\n" - "ldr q12, [x27, x11]\n" + "fmla v23.8h, v5.8h, v16.8h\n" + "ldr q16, [x27, x11]\n" "ldr q5, [x10, #0x60]\n" - "fmla v30.8h, v7.8h, v11.8h\n" - "fmla v31.8h, v6.8h, v11.8h\n" + "fmla v22.8h, v7.8h, v17.8h\n" + "fmla v21.8h, v6.8h, v17.8h\n" "ldr q11, [x13, x26]\n" - "fmla v28.8h, v6.8h, v9.8h\n" + "fmla v24.8h, v6.8h, v19.8h\n" "ldr q9, [x9, x15]\n" - "fmla v29.8h, v8.8h, v10.8h\n" + "fmla v23.8h, v8.8h, v18.8h\n" "ld1 { v10.8h }, [x13]\n" "ldr q6, [x10, #0x70]\n" - "fmla v30.8h, v8.8h, v12.8h\n" - "fmla v31.8h, v7.8h, v12.8h\n" + "fmla v22.8h, v8.8h, v16.8h\n" + "fmla v21.8h, v7.8h, v16.8h\n" "ldr q12, [x9, x11]\n" "ldr q7, [x10, #0x80]\n" - "fmax v28.8h, v28.8h, v18.8h\n" - "fmax v29.8h, v29.8h, v18.8h\n" + "fmax v24.8h, v24.8h, v27.8h\n" + "fmax v23.8h, v23.8h, v27.8h\n" "ldr q8, [x10, #0x90]\n" - "fmax v30.8h, v30.8h, v18.8h\n" - "fmax v31.8h, v31.8h, v18.8h\n" + "fmax v22.8h, v22.8h, v27.8h\n" + "fmax v21.8h, v21.8h, v27.8h\n" "add x27, x27, #0x10\n" - "fmin v28.8h, v28.8h, v17.8h\n" - "fmin v29.8h, v29.8h, v17.8h\n" - "st1 { v28.8h }, [x12]\n" + "fmin v24.8h, v24.8h, v26.8h\n" + "fmin v23.8h, v23.8h, v26.8h\n" + "st1 { v24.8h }, [x12]\n" "add x10, x10, #0xa0\n" - "fmin v30.8h, v30.8h, v17.8h\n" - "fmin v31.8h, v31.8h, v17.8h\n" - "str q29, [x12, x14]\n" + "fmin v22.8h, v22.8h, v26.8h\n" + "fmin v21.8h, v21.8h, v26.8h\n" + "str q23, [x12, x14]\n" "add x12, x12, #0x10\n" - "st1 { v30.8h }, [x25]\n" - "str q31, [x25, x14]\n" + "st1 { v22.8h }, [x25]\n" + "str q21, [x25, x14]\n" "add x25, x25, #0x10\n" "blt 2b\n" "3:" // Tile loop: Channel tail - "mov v28.16b, v16.16b\n fmla v28.8h, v4.8h, v9.8h\n" - "mov v29.16b, v16.16b\n fmla v29.8h, v3.8h, v9.8h\n" - "mov v30.16b, v16.16b\n fmla v30.8h, v1.8h, v9.8h\n" - "mov v31.16b, v16.16b\n fmla v31.8h, v0.8h, v9.8h\n" - "ld1 { v9.8h }, [x27]\n" - "fmla v28.8h, v0.8h, v10.8h\n" - "ldr q10, [x28, x11]\n" - "fmla v29.8h, v2.8h, v11.8h\n" - "ldr q11, [x27, x26]\n" - "fmla v30.8h, v2.8h, v12.8h\n" - "fmla v31.8h, v1.8h, v12.8h\n" - "fmla v28.8h, v5.8h, v12.8h\n" - "fmla v29.8h, v4.8h, v12.8h\n" - "ldr q12, [x13, x15]\n" - "fmla v30.8h, v6.8h, v9.8h\n" - "ldr q9, [x13, x11]\n" - "fmla v31.8h, v3.8h, v13.8h\n" + "mov v24.16b, v25.16b\n fmla v24.8h, v4.8h, v9.8h\n" + "mov v23.16b, v25.16b\n fmla v23.8h, v3.8h, v9.8h\n" + "mov v22.16b, v25.16b\n fmla v22.8h, v1.8h, v9.8h\n" + "mov v21.16b, v25.16b\n fmla v21.8h, v0.8h, v9.8h\n" + "ld1 { v18.8h }, [x27]\n" + "fmla v24.8h, v0.8h, v10.8h\n" + "ldr q20, [x28, x11]\n" + "fmla v23.8h, v2.8h, v11.8h\n" + "ldr q17, [x27, x26]\n" + "fmla v22.8h, v2.8h, v12.8h\n" + "fmla v21.8h, v1.8h, v12.8h\n" + "fmla v24.8h, v5.8h, v12.8h\n" + "fmla v23.8h, v4.8h, v12.8h\n" + "ldr q16, [x13, x15]\n" + "fmla v22.8h, v6.8h, v18.8h\n" + "ldr q18, [x13, x11]\n" + "fmla v21.8h, v3.8h, v13.8h\n" "add x13, x13, #0x10\n" - "fmla v28.8h, v7.8h, v13.8h\n" - "fmla v29.8h, v6.8h, v13.8h\n" - "fmla v30.8h, v4.8h, v13.8h\n" - "fmla v31.8h, v8.8h, v11.8h\n" - "ld1 { v11.8h }, [x9]\n" - "fmla v28.8h, v1.8h, v12.8h\n" - "fmla v29.8h, v0.8h, v12.8h\n" - "ldr q12, [x9, x26]\n" + "fmla v24.8h, v7.8h, v13.8h\n" + "fmla v23.8h, v6.8h, v13.8h\n" + "fmla v22.8h, v4.8h, v13.8h\n" + "fmla v21.8h, v8.8h, v17.8h\n" + "ld1 { v17.8h }, [x9]\n" + "fmla v24.8h, v1.8h, v16.8h\n" + "fmla v23.8h, v0.8h, v16.8h\n" + "ldr q16, [x9, x26]\n" "add x9, x9, #0x10\n" - "fmla v30.8h, v5.8h, v10.8h\n" - "fmla v31.8h, v4.8h, v10.8h\n" - "fmla v28.8h, v2.8h, v9.8h\n" - "fmla v29.8h, v1.8h, v9.8h\n" - "ld1 { v9.8h }, [x28]\n" - "fmla v30.8h, v0.8h, v11.8h\n" - "fmla v31.8h, v2.8h, v12.8h\n" - "fmla v28.8h, v8.8h, v10.8h\n" - "fmla v29.8h, v7.8h, v10.8h\n" - "ldr q10, [x28, x26]\n" + "fmla v22.8h, v5.8h, v20.8h\n" + "fmla v21.8h, v4.8h, v20.8h\n" + "fmla v24.8h, v2.8h, v18.8h\n" + "fmla v23.8h, v1.8h, v18.8h\n" + "ld1 { v19.8h }, [x28]\n" + "fmla v22.8h, v0.8h, v17.8h\n" + "fmla v21.8h, v2.8h, v16.8h\n" + "fmla v24.8h, v8.8h, v20.8h\n" + "fmla v23.8h, v7.8h, v20.8h\n" + "ldr q18, [x28, x26]\n" "add x28, x28, #0x10\n" - "fmla v30.8h, v3.8h, v9.8h\n" - "fmla v31.8h, v5.8h, v10.8h\n" - "fmla v28.8h, v3.8h, v11.8h\n" - "ldr q11, [x27, x15]\n" - "fmla v29.8h, v5.8h, v12.8h\n" - "ldr q12, [x27, x11]\n" - "fmla v30.8h, v7.8h, v11.8h\n" - "fmla v31.8h, v6.8h, v11.8h\n" + "fmla v22.8h, v3.8h, v19.8h\n" + "fmla v21.8h, v5.8h, v18.8h\n" + "fmla v24.8h, v3.8h, v17.8h\n" + "ldr q17, [x27, x15]\n" + "fmla v23.8h, v5.8h, v16.8h\n" + "ldr q16, [x27, x11]\n" + "fmla v22.8h, v7.8h, v17.8h\n" + "fmla v21.8h, v6.8h, v17.8h\n" "add x27, x27, #0x10\n" - "fmla v28.8h, v6.8h, v9.8h\n" - "fmla v29.8h, v8.8h, v10.8h\n" - "fmax v28.8h, v28.8h, v18.8h\n" - "fmla v30.8h, v8.8h, v12.8h\n" - "fmla v31.8h, v7.8h, v12.8h\n" - "fmax v29.8h, v29.8h, v18.8h\n" - "fmax v30.8h, v30.8h, v18.8h\n" - "fmax v31.8h, v31.8h, v18.8h\n" - "fmin v28.8h, v28.8h, v17.8h\n" - "fmin v29.8h, v29.8h, v17.8h\n" - "st1 { v28.8h }, [x12]\n" - "fmin v30.8h, v30.8h, v17.8h\n" - "fmin v31.8h, v31.8h, v17.8h\n" - "str q29, [x12, x14]\n" + "fmla v24.8h, v6.8h, v19.8h\n" + "fmla v23.8h, v8.8h, v18.8h\n" + "fmax v24.8h, v24.8h, v27.8h\n" + "fmla v22.8h, v8.8h, v16.8h\n" + "fmla v21.8h, v7.8h, v16.8h\n" + "fmax v23.8h, v23.8h, v27.8h\n" + "fmax v22.8h, v22.8h, v27.8h\n" + "fmax v21.8h, v21.8h, v27.8h\n" + "fmin v24.8h, v24.8h, v26.8h\n" + "fmin v23.8h, v23.8h, v26.8h\n" + "st1 { v24.8h }, [x12]\n" + "fmin v22.8h, v22.8h, v26.8h\n" + "fmin v21.8h, v21.8h, v26.8h\n" + "str q23, [x12, x14]\n" "add x12, x12, #0x10\n" - "st1 { v30.8h }, [x25]\n" - "str q31, [x25, x14]\n" + "st1 { v22.8h }, [x25]\n" + "str q21, [x25, x14]\n" "add x25, x25, #0x10\n" "4:" // Tile loop: Oddments "tst %x[n_channels], #0x7\n" "beq 57f\n" - "ldr q16, [x10, #0x0]\n" + "ldr q25, [x10, #0x0]\n" "ldr q0, [x10, #0x10]\n" "add x24, x9, x15\n" "add x23, x13, XZR\n" @@ -363,11 +363,11 @@ void a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst_direct_impl( "ldr h12, [x21, #0x0]\n" "ldr h13, [x20, #0x0]\n" "8:" // Tile loop: Oddments: Load inputs: (1, 1), (0, 0), (0, 3), (1, 2), (2, 1): Bit 2: End - "mov v28.16b, v16.16b\n fmla v28.8h, v4.8h, v9.8h\n" - "mov v29.16b, v16.16b\n fmla v29.8h, v3.8h, v9.8h\n" + "mov v28.16b, v25.16b\n fmla v28.8h, v4.8h, v9.8h\n" + "mov v29.16b, v25.16b\n fmla v29.8h, v3.8h, v9.8h\n" "add x20, x27, XZR\n" - "mov v30.16b, v16.16b\n fmla v30.8h, v1.8h, v9.8h\n" - "mov v31.16b, v16.16b\n fmla v31.8h, v0.8h, v9.8h\n" + "mov v30.16b, v25.16b\n fmla v30.8h, v1.8h, v9.8h\n" + "mov v31.16b, v25.16b\n fmla v31.8h, v0.8h, v9.8h\n" "fmla v28.8h, v0.8h, v10.8h\n" "fmla v29.8h, v2.8h, v11.8h\n" "fmla v28.8h, v5.8h, v12.8h\n" @@ -630,14 +630,14 @@ void a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst_direct_impl( "52:" // Tile loop: Oddments: Load inputs: (3, 2): Bit 2: End "fmla v30.8h, v8.8h, v12.8h\n" "fmla v31.8h, v7.8h, v12.8h\n" - "fmax v28.8h, v28.8h, v18.8h\n" - "fmax v29.8h, v29.8h, v18.8h\n" - "fmax v30.8h, v30.8h, v18.8h\n" - "fmax v31.8h, v31.8h, v18.8h\n" - "fmin v28.8h, v28.8h, v17.8h\n" - "fmin v29.8h, v29.8h, v17.8h\n" - "fmin v30.8h, v30.8h, v17.8h\n" - "fmin v31.8h, v31.8h, v17.8h\n" + "fmax v28.8h, v28.8h, v27.8h\n" + "fmax v29.8h, v29.8h, v27.8h\n" + "fmax v30.8h, v30.8h, v27.8h\n" + "fmax v31.8h, v31.8h, v27.8h\n" + "fmin v28.8h, v28.8h, v26.8h\n" + "fmin v29.8h, v29.8h, v26.8h\n" + "fmin v30.8h, v30.8h, v26.8h\n" + "fmin v31.8h, v31.8h, v26.8h\n" "tbz %x[n_channels], #2, 54f\n" "mov x21, x12\n" "mov x20, x25\n" @@ -699,7 +699,6 @@ void a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst_direct_impl( "st1 { v29.h }[0], [x21]\n" "st1 { v31.h }[0], [x20]\n" "56:" // Tile loop: Oddments: Store: Bit 2: End - "57:" // Tile loop: End "ldr x22, [%x[params_struct], %[offsetof_args_tile_j]]\n" "ldr x23, [%x[params_struct], %[offsetof_args_tile_i]]\n" @@ -714,7 +713,7 @@ void a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst_direct_impl( "blt 1b\n" : : [n_channels] "r" ((unsigned long) n_channels), [offsetof_args_inptr] "I" (offsetof(Args, inptr)), [offsetof_args_ld_input_col] "I" (offsetof(Args, ld_input_col)), [offsetof_args_ld_input_row] "I" (offsetof(Args, ld_input_row)), [offsetof_args_ld_output_col] "I" (offsetof(Args, ld_output_col)), [offsetof_args_ld_output_row] "I" (offsetof(Args, ld_output_row)), [offsetof_args_max] "I" (offsetof(Args, max)), [offsetof_args_min] "I" (offsetof(Args, min)), [offsetof_args_n_tile_cols] "I" (offsetof(Args, n_tile_cols)), [offsetof_args_n_tile_rows] "I" (offsetof(Args, n_tile_rows)), [offsetof_args_outptr] "I" (offsetof(Args, outptr)), [offsetof_args_params] "I" (offsetof(Args, params)), [offsetof_args_tile_i] "I" (offsetof(Args, tile_i)), [offsetof_args_tile_j] "I" (offsetof(Args, tile_j)), [params_struct] "r" (¶ms_struct) - : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v16", "v17", "v18", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" ); } -- cgit v1.2.1