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.cpp | 532 +++++++++++---------- 1 file changed, 267 insertions(+), 265 deletions(-) (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp') diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp index 9fa38c6efe..17afc92e30 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp @@ -22,12 +22,13 @@ * SOFTWARE. */ -#if defined(__aarch64__) #include "arm_gemm.hpp" #include #include +#if defined(__aarch64__) + namespace arm_conv { namespace depthwise { @@ -42,133 +43,133 @@ void a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst_impl __asm__ __volatile__( "ldr q12, [%x[params], #0x0]\n" "ldr q8, [%x[params], #0x10]\n" - "movi v28.16b, #0x1\n" - "movi v18.4s, #0x0\n" + "movi v30.16b, #0x1\n" + "movi v17.4s, #0x0\n" "ldr q9, [%x[params], #0x20]\n" "ldr q10, [%x[params], #0x30]\n" - "movi v31.4s, #0x0\n" - "movi v24.4s, #0x0\n" + "movi v16.4s, #0x0\n" + "movi v25.4s, #0x0\n" "ldr q11, [%x[params], #0x40]\n" "ldr x20, [%x[inptrs], #0x18]\n" - "movi v30.4s, #0x0\n" - "movi v21.4s, #0x0\n" + "movi v24.4s, #0x0\n" + "movi v31.4s, #0x0\n" "ld1 { v3.16b }, [x20]\n" "ldr x20, [%x[inptrs], #0x20]\n" - "mov v16.16b, v3.16b\n" - "ext v16.16b, v16.16b, v16.16b, #0x1\n" + "mov v26.16b, v3.16b\n" + "ext v26.16b, v26.16b, v26.16b, #0x1\n" "ld1 { v4.16b }, [x20]\n" "ldr x20, [%x[inptrs], #0x10]\n" - "mov v15.16b, v4.16b\n" - "ext v15.16b, v15.16b, v15.16b, #0x1\n" + "mov v21.16b, v4.16b\n" + "ext v21.16b, v21.16b, v21.16b, #0x1\n" "ld1 { v2.16b }, [x20]\n" "ldr x20, [%x[inptrs], #0x8]\n" - "mov v20.16b, v2.16b\n" - "ext v20.16b, v20.16b, v20.16b, #0x1\n" + "mov v27.16b, v2.16b\n" + "ext v27.16b, v27.16b, v27.16b, #0x1\n" "ld1 { v1.16b }, [x20]\n" "ldr x20, [%x[inptrs], #0x28]\n" - "zip1 v3.2d, v3.2d, v16.2d\n" - "zip1 v4.2d, v4.2d, v15.2d\n" + "zip1 v3.2d, v3.2d, v26.2d\n" + "zip1 v4.2d, v4.2d, v21.2d\n" "ld1 { v5.16b }, [x20]\n" "ldr x20, [%x[inptrs], #0x30]\n" "mov v26.16b, v1.16b\n" - "mov v13.16b, v5.16b\n" + "mov v22.16b, v5.16b\n" "ld1 { v6.16b }, [x20]\n" "ldr x20, [%x[inptrs], #0x38]\n" "mov v19.16b, v6.16b\n" "ext v26.16b, v26.16b, v26.16b, #0x1\n" "ld1 { v7.16b }, [x20]\n" "ldr x20, [%x[inptrs], #0x0]\n" - "mov v17.16b, v7.16b\n" - "zip1 v2.2d, v2.2d, v20.2d\n" + "mov v21.16b, v7.16b\n" + "zip1 v2.2d, v2.2d, v27.2d\n" "ld1 { v0.16b }, [x20]\n" - "ext v13.16b, v13.16b, v13.16b, #0x1\n" + "ext v22.16b, v22.16b, v22.16b, #0x1\n" "ext v19.16b, v19.16b, v19.16b, #0x1\n" - ".inst 0x4f83e392 // sdot v18.4s, v28.16b, v3.4b[0]\n" - "ext v17.16b, v17.16b, v17.16b, #0x1\n" - ".inst 0x4f83eb9f // sdot v31.4s, v28.16b, v3.4b[2]\n" - ".inst 0x4f84e398 // sdot v24.4s, v28.16b, v4.4b[0]\n" + ".inst 0x4f83e3d1 // sdot v17.4s, v30.16b, v3.4b[0]\n" + "ext v21.16b, v21.16b, v21.16b, #0x1\n" + ".inst 0x4f83ebd0 // sdot v16.4s, v30.16b, v3.4b[2]\n" + ".inst 0x4f84e3d9 // sdot v25.4s, v30.16b, v4.4b[0]\n" "add x20, %x[qp], %[offsetof_Requantize32_b_offset]\n" "ld1r { v23.4s }, [x20]\n" - ".inst 0x4f84eb9e // sdot v30.4s, v28.16b, v4.4b[2]\n" - "mov v16.16b, v0.16b\n" - ".inst 0x4f82e395 // sdot v21.4s, v28.16b, v2.4b[0]\n" - "movi v20.4s, #0x0\n" - "movi v29.4s, #0x1\n" - ".inst 0x4f82eb94 // sdot v20.4s, v28.16b, v2.4b[2]\n" + ".inst 0x4f84ebd8 // sdot v24.4s, v30.16b, v4.4b[2]\n" + "mov v18.16b, v0.16b\n" + ".inst 0x4f82e3df // sdot v31.4s, v30.16b, v2.4b[0]\n" + "movi v29.4s, #0x0\n" + "movi v28.4s, #0x1\n" + ".inst 0x4f82ebdd // sdot v29.4s, v30.16b, v2.4b[2]\n" "add x20, %x[qp], %[offsetof_Requantize32_c_offset]\n" - "ld1r { v14.4s }, [x20]\n" - "ext v16.16b, v16.16b, v16.16b, #0x1\n" + "ld1r { v13.4s }, [x20]\n" + "ext v18.16b, v18.16b, v18.16b, #0x1\n" "zip1 v1.2d, v1.2d, v26.2d\n" - ".inst 0x4fa3e3b2 // sdot v18.4s, v29.16b, v3.4b[1]\n" - "zip1 v5.2d, v5.2d, v13.2d\n" + ".inst 0x4fa3e391 // sdot v17.4s, v28.16b, v3.4b[1]\n" + "zip1 v5.2d, v5.2d, v22.2d\n" "zip1 v6.2d, v6.2d, v19.2d\n" - ".inst 0x4fa3ebbf // sdot v31.4s, v29.16b, v3.4b[3]\n" + ".inst 0x4fa3eb90 // sdot v16.4s, v28.16b, v3.4b[3]\n" "add x20, %x[qp], %[offsetof_Requantize32_minval]\n" - "ld1r { v13.4s }, [x20]\n" - "zip1 v7.2d, v7.2d, v17.2d\n" + "ld1r { v14.4s }, [x20]\n" + "zip1 v7.2d, v7.2d, v21.2d\n" "movi v22.4s, #0x0\n" - ".inst 0x4fa4e3b8 // sdot v24.4s, v29.16b, v4.4b[1]\n" - "movi v26.4s, #0x0\n" - ".inst 0x4fa4ebbe // sdot v30.4s, v29.16b, v4.4b[3]\n" - ".inst 0x4f81e396 // sdot v22.4s, v28.16b, v1.4b[0]\n" + ".inst 0x4fa4e399 // sdot v25.4s, v28.16b, v4.4b[1]\n" + "movi v21.4s, #0x0\n" + ".inst 0x4fa4eb98 // sdot v24.4s, v28.16b, v4.4b[3]\n" + ".inst 0x4f81e3d6 // sdot v22.4s, v30.16b, v1.4b[0]\n" "add x20, %x[qp], %[offsetof_Requantize32_maxval]\n" "ld1r { v15.4s }, [x20]\n" - "movi v25.4s, #0x0\n" + "movi v26.4s, #0x0\n" "movi v27.4s, #0x0\n" - ".inst 0x4f81eb9a // sdot v26.4s, v28.16b, v1.4b[2]\n" - "zip1 v0.2d, v0.2d, v16.2d\n" + ".inst 0x4f81ebd5 // sdot v21.4s, v30.16b, v1.4b[2]\n" + "movi v20.4s, #0x0\n" "movi v19.4s, #0x0\n" - ".inst 0x4f85e399 // sdot v25.4s, v28.16b, v5.4b[0]\n" + ".inst 0x4f85e3da // sdot v26.4s, v30.16b, v5.4b[0]\n" "cmp %x[n_channels], #0x4\n" - ".inst 0x4f85eb9b // sdot v27.4s, v28.16b, v5.4b[2]\n" - ".inst 0x4f86e393 // sdot v19.4s, v28.16b, v6.4b[0]\n" - "add v24.4s, v18.4s, v24.4s\n" - "mov x9, #0x0\n" + "zip1 v0.2d, v0.2d, v18.2d\n" "movi v18.4s, #0x0\n" - ".inst 0x4f86eb92 // sdot v18.4s, v28.16b, v6.4b[2]\n" - ".inst 0x4fa2e3b5 // sdot v21.4s, v29.16b, v2.4b[1]\n" + ".inst 0x4f85ebdb // sdot v27.4s, v30.16b, v5.4b[2]\n" + "mov x9, #0x0\n" + ".inst 0x4f86e3d4 // sdot v20.4s, v30.16b, v6.4b[0]\n" + ".inst 0x4f86ebd3 // sdot v19.4s, v30.16b, v6.4b[2]\n" + "add v17.4s, v17.4s, v25.4s\n" "mov x28, #0x0\n" - ".inst 0x4fa2ebb4 // sdot v20.4s, v29.16b, v2.4b[3]\n" - "add v17.4s, v31.4s, v30.4s\n" - ".inst 0x4fa1e3b6 // sdot v22.4s, v29.16b, v1.4b[1]\n" + "movi v25.4s, #0x0\n" + ".inst 0x4f87e3d2 // sdot v18.4s, v30.16b, v7.4b[0]\n" + ".inst 0x4f87ebd9 // sdot v25.4s, v30.16b, v7.4b[2]\n" "ldp x27, x26, [%x[outptrs], #0x0]\n" - "movi v16.4s, #0x0\n" - ".inst 0x4f87e390 // sdot v16.4s, v28.16b, v7.4b[0]\n" - ".inst 0x4fa1ebba // sdot v26.4s, v29.16b, v1.4b[3]\n" + ".inst 0x4fa2e39f // sdot v31.4s, v28.16b, v2.4b[1]\n" + ".inst 0x4fa2eb9d // sdot v29.4s, v28.16b, v2.4b[3]\n" + "add v16.4s, v16.4s, v24.4s\n" "ldp x25, x24, [%x[outptrs], #0x10]\n" - ".inst 0x4fa5e3b9 // sdot v25.4s, v29.16b, v5.4b[1]\n" - ".inst 0x4fa5ebbb // sdot v27.4s, v29.16b, v5.4b[3]\n" - "add v30.4s, v21.4s, v24.4s\n" + "movi v24.4s, #0x0\n" + ".inst 0x4f80e3d8 // sdot v24.4s, v30.16b, v0.4b[0]\n" + ".inst 0x4fa1e396 // sdot v22.4s, v28.16b, v1.4b[1]\n" "ldp x23, x22, [%x[outptrs], #0x20]\n" - ".inst 0x4fa6e3b3 // sdot v19.4s, v29.16b, v6.4b[1]\n" - ".inst 0x4fa6ebb2 // sdot v18.4s, v29.16b, v6.4b[3]\n" - "add v31.4s, v20.4s, v17.4s\n" + ".inst 0x4fa1eb95 // sdot v21.4s, v28.16b, v1.4b[3]\n" + ".inst 0x4fa5e39a // sdot v26.4s, v28.16b, v5.4b[1]\n" + "add v31.4s, v31.4s, v17.4s\n" "ldp x21, x20, [%x[outptrs], #0x30]\n" - ".inst 0x4fa7e3b0 // sdot v16.4s, v29.16b, v7.4b[1]\n" - "add v22.4s, v22.4s, v30.4s\n" + ".inst 0x4fa5eb9b // sdot v27.4s, v28.16b, v5.4b[3]\n" + ".inst 0x4fa6e394 // sdot v20.4s, v28.16b, v6.4b[1]\n" + "add v29.4s, v29.4s, v16.4s\n" "add %x[params], %x[params], #0x50\n" - "add v21.4s, v26.4s, v31.4s\n" - "add v20.4s, v25.4s, v19.4s\n" - "add v19.4s, v27.4s, v18.4s\n" - "add v18.4s, v16.4s, v24.4s\n" - "movi v16.4s, #0x0\n" - ".inst 0x4f87eb90 // sdot v16.4s, v28.16b, v7.4b[2]\n" - ".inst 0x4fa7ebb0 // sdot v16.4s, v29.16b, v7.4b[3]\n" - "add v17.4s, v16.4s, v17.4s\n" - "movi v16.4s, #0x0\n" - ".inst 0x4f80e390 // sdot v16.4s, v28.16b, v0.4b[0]\n" - ".inst 0x4fa0e3b0 // sdot v16.4s, v29.16b, v0.4b[1]\n" - "add v24.4s, v22.4s, v16.4s\n" - "add v26.4s, v22.4s, v25.4s\n" - "movi v16.4s, #0x0\n" - ".inst 0x4f80eb90 // sdot v16.4s, v28.16b, v0.4b[2]\n" - ".inst 0x4fa0ebb0 // sdot v16.4s, v29.16b, v0.4b[3]\n" - "add v25.4s, v21.4s, v16.4s\n" - "add v27.4s, v21.4s, v27.4s\n" - "add v28.4s, v20.4s, v30.4s\n" - "add v29.4s, v19.4s, v31.4s\n" - "add v30.4s, v18.4s, v20.4s\n" - "add v31.4s, v17.4s, v19.4s\n" + ".inst 0x4fa6eb93 // sdot v19.4s, v28.16b, v6.4b[3]\n" + ".inst 0x4fa7e392 // sdot v18.4s, v28.16b, v7.4b[1]\n" + "add v22.4s, v22.4s, v31.4s\n" + ".inst 0x4fa7eb99 // sdot v25.4s, v28.16b, v7.4b[3]\n" + ".inst 0x4fa0e398 // sdot v24.4s, v28.16b, v0.4b[1]\n" + "add v21.4s, v21.4s, v29.4s\n" + "add v20.4s, v26.4s, v20.4s\n" + "add v19.4s, v27.4s, v19.4s\n" + "add v18.4s, v18.4s, v17.4s\n" + "movi v17.4s, #0x0\n" + ".inst 0x4f80ebd1 // sdot v17.4s, v30.16b, v0.4b[2]\n" + ".inst 0x4fa0eb91 // sdot v17.4s, v28.16b, v0.4b[3]\n" + "add v16.4s, v25.4s, v16.4s\n" + "add v24.4s, v22.4s, v24.4s\n" + "add v25.4s, v21.4s, v17.4s\n" + "add v26.4s, v26.4s, v22.4s\n" + "add v27.4s, v27.4s, v21.4s\n" + "add v28.4s, v20.4s, v31.4s\n" + "add v29.4s, v19.4s, v29.4s\n" + "add v30.4s, v20.4s, v18.4s\n" + "add v31.4s, v19.4s, v16.4s\n" "neg v23.4s, v23.4s\n" "mul v24.4s, v24.4s, v23.4s\n" "mul v25.4s, v25.4s, v23.4s\n" @@ -194,11 +195,11 @@ void a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst_impl "add v31.4s, v31.4s, v12.4s\n" "ble 2f\n" "1:" // Loop - "ldr q21, [%x[params], #0x60]\n" - "ldr q20, [%x[params], #0x70]\n" + "ldr q12, [%x[params], #0x60]\n" + "ldr q21, [%x[params], #0x70]\n" ".inst 0x4f80e118 // sdot v24.4s, v8.16b, v0.4b[0]\n" ".inst 0x4f80e919 // sdot v25.4s, v8.16b, v0.4b[2]\n" - "ldr q12, [%x[params], #0x80]\n" + "ldr q20, [%x[params], #0x80]\n" ".inst 0x4f81e11a // sdot v26.4s, v8.16b, v1.4b[0]\n" ".inst 0x4f81e91b // sdot v27.4s, v8.16b, v1.4b[2]\n" "sub %x[n_channels], %x[n_channels], #0x4\n" @@ -212,7 +213,7 @@ void a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst_impl ".inst 0x4f82e91d // sdot v29.4s, v8.16b, v2.4b[2]\n" ".inst 0x4f83e11e // sdot v30.4s, v8.16b, v3.4b[0]\n" ".inst 0x4f83e91f // sdot v31.4s, v8.16b, v3.4b[2]\n" - "ldr q8, [%x[params], #0x0]\n" + "ldr q17, [%x[params], #0x0]\n" ".inst 0x4f81e158 // sdot v24.4s, v10.16b, v1.4b[0]\n" ".inst 0x4f81e959 // sdot v25.4s, v10.16b, v1.4b[2]\n" ".inst 0x4f82e15a // sdot v26.4s, v10.16b, v2.4b[0]\n" @@ -221,7 +222,7 @@ void a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst_impl ".inst 0x4fa2e93d // sdot v29.4s, v9.16b, v2.4b[3]\n" ".inst 0x4fa3e13e // sdot v30.4s, v9.16b, v3.4b[1]\n" ".inst 0x4fa3e93f // sdot v31.4s, v9.16b, v3.4b[3]\n" - "ldr q9, [%x[params], #0x10]\n" + "ldr q16, [%x[params], #0x10]\n" ".inst 0x4fa1e178 // sdot v24.4s, v11.16b, v1.4b[1]\n" ".inst 0x4fa1e979 // sdot v25.4s, v11.16b, v1.4b[3]\n" ".inst 0x4fa2e17a // sdot v26.4s, v11.16b, v2.4b[1]\n" @@ -230,115 +231,115 @@ void a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst_impl ".inst 0x4f83e95d // sdot v29.4s, v10.16b, v3.4b[2]\n" ".inst 0x4f84e15e // sdot v30.4s, v10.16b, v4.4b[0]\n" ".inst 0x4f84e95f // sdot v31.4s, v10.16b, v4.4b[2]\n" - "ldr q10, [%x[params], #0x20]\n" - ".inst 0x4f82e118 // sdot v24.4s, v8.16b, v2.4b[0]\n" - ".inst 0x4f82e919 // sdot v25.4s, v8.16b, v2.4b[2]\n" - ".inst 0x4f83e11a // sdot v26.4s, v8.16b, v3.4b[0]\n" - ".inst 0x4f83e91b // sdot v27.4s, v8.16b, v3.4b[2]\n" + "ldr q19, [%x[params], #0x20]\n" + ".inst 0x4f82e238 // sdot v24.4s, v17.16b, v2.4b[0]\n" + ".inst 0x4f82ea39 // sdot v25.4s, v17.16b, v2.4b[2]\n" + ".inst 0x4f83e23a // sdot v26.4s, v17.16b, v3.4b[0]\n" + ".inst 0x4f83ea3b // sdot v27.4s, v17.16b, v3.4b[2]\n" ".inst 0x4fa3e17c // sdot v28.4s, v11.16b, v3.4b[1]\n" ".inst 0x4fa3e97d // sdot v29.4s, v11.16b, v3.4b[3]\n" ".inst 0x4fa4e17e // sdot v30.4s, v11.16b, v4.4b[1]\n" ".inst 0x4fa4e97f // sdot v31.4s, v11.16b, v4.4b[3]\n" - "ldr q11, [%x[params], #0x30]\n" - ".inst 0x4fa2e138 // sdot v24.4s, v9.16b, v2.4b[1]\n" - ".inst 0x4fa2e939 // sdot v25.4s, v9.16b, v2.4b[3]\n" - ".inst 0x4fa3e13a // sdot v26.4s, v9.16b, v3.4b[1]\n" - ".inst 0x4fa3e93b // sdot v27.4s, v9.16b, v3.4b[3]\n" - ".inst 0x4f84e11c // sdot v28.4s, v8.16b, v4.4b[0]\n" - ".inst 0x4f84e91d // sdot v29.4s, v8.16b, v4.4b[2]\n" - ".inst 0x4f85e11e // sdot v30.4s, v8.16b, v5.4b[0]\n" - ".inst 0x4f85e91f // sdot v31.4s, v8.16b, v5.4b[2]\n" - "ldr q8, [%x[params], #0x40]\n" - ".inst 0x4f83e158 // sdot v24.4s, v10.16b, v3.4b[0]\n" - ".inst 0x4f83e959 // sdot v25.4s, v10.16b, v3.4b[2]\n" - ".inst 0x4f84e15a // sdot v26.4s, v10.16b, v4.4b[0]\n" - ".inst 0x4f84e95b // sdot v27.4s, v10.16b, v4.4b[2]\n" - ".inst 0x4fa4e13c // sdot v28.4s, v9.16b, v4.4b[1]\n" - ".inst 0x4fa4e93d // sdot v29.4s, v9.16b, v4.4b[3]\n" - ".inst 0x4fa5e13e // sdot v30.4s, v9.16b, v5.4b[1]\n" - ".inst 0x4fa5e93f // sdot v31.4s, v9.16b, v5.4b[3]\n" - "ldr q9, [%x[params], #0x50]\n" - ".inst 0x4fa3e178 // sdot v24.4s, v11.16b, v3.4b[1]\n" - ".inst 0x4fa3e979 // sdot v25.4s, v11.16b, v3.4b[3]\n" - ".inst 0x4fa4e17a // sdot v26.4s, v11.16b, v4.4b[1]\n" - ".inst 0x4fa4e97b // sdot v27.4s, v11.16b, v4.4b[3]\n" - ".inst 0x4f85e15c // sdot v28.4s, v10.16b, v5.4b[0]\n" - ".inst 0x4f85e95d // sdot v29.4s, v10.16b, v5.4b[2]\n" - ".inst 0x4f86e15e // sdot v30.4s, v10.16b, v6.4b[0]\n" - ".inst 0x4f86e95f // sdot v31.4s, v10.16b, v6.4b[2]\n" + "ldr q18, [%x[params], #0x30]\n" + ".inst 0x4fa2e218 // sdot v24.4s, v16.16b, v2.4b[1]\n" + ".inst 0x4fa2ea19 // sdot v25.4s, v16.16b, v2.4b[3]\n" + ".inst 0x4fa3e21a // sdot v26.4s, v16.16b, v3.4b[1]\n" + ".inst 0x4fa3ea1b // sdot v27.4s, v16.16b, v3.4b[3]\n" + ".inst 0x4f84e23c // sdot v28.4s, v17.16b, v4.4b[0]\n" + ".inst 0x4f84ea3d // sdot v29.4s, v17.16b, v4.4b[2]\n" + ".inst 0x4f85e23e // sdot v30.4s, v17.16b, v5.4b[0]\n" + ".inst 0x4f85ea3f // sdot v31.4s, v17.16b, v5.4b[2]\n" + "ldr q17, [%x[params], #0x40]\n" + ".inst 0x4f83e278 // sdot v24.4s, v19.16b, v3.4b[0]\n" + ".inst 0x4f83ea79 // sdot v25.4s, v19.16b, v3.4b[2]\n" + ".inst 0x4f84e27a // sdot v26.4s, v19.16b, v4.4b[0]\n" + ".inst 0x4f84ea7b // sdot v27.4s, v19.16b, v4.4b[2]\n" + ".inst 0x4fa4e21c // sdot v28.4s, v16.16b, v4.4b[1]\n" + ".inst 0x4fa4ea1d // sdot v29.4s, v16.16b, v4.4b[3]\n" + ".inst 0x4fa5e21e // sdot v30.4s, v16.16b, v5.4b[1]\n" + ".inst 0x4fa5ea1f // sdot v31.4s, v16.16b, v5.4b[3]\n" + "ldr q16, [%x[params], #0x50]\n" + ".inst 0x4fa3e258 // sdot v24.4s, v18.16b, v3.4b[1]\n" + ".inst 0x4fa3ea59 // sdot v25.4s, v18.16b, v3.4b[3]\n" + ".inst 0x4fa4e25a // sdot v26.4s, v18.16b, v4.4b[1]\n" + ".inst 0x4fa4ea5b // sdot v27.4s, v18.16b, v4.4b[3]\n" + ".inst 0x4f85e27c // sdot v28.4s, v19.16b, v5.4b[0]\n" + ".inst 0x4f85ea7d // sdot v29.4s, v19.16b, v5.4b[2]\n" + ".inst 0x4f86e27e // sdot v30.4s, v19.16b, v6.4b[0]\n" + ".inst 0x4f86ea7f // sdot v31.4s, v19.16b, v6.4b[2]\n" "ldr q10, [%x[params], #0xb0]\n" - ".inst 0x4f84e118 // sdot v24.4s, v8.16b, v4.4b[0]\n" - ".inst 0x4f84e919 // sdot v25.4s, v8.16b, v4.4b[2]\n" - ".inst 0x4f85e11a // sdot v26.4s, v8.16b, v5.4b[0]\n" - ".inst 0x4f85e91b // sdot v27.4s, v8.16b, v5.4b[2]\n" - ".inst 0x4fa5e17c // sdot v28.4s, v11.16b, v5.4b[1]\n" - ".inst 0x4fa5e97d // sdot v29.4s, v11.16b, v5.4b[3]\n" - ".inst 0x4fa6e17e // sdot v30.4s, v11.16b, v6.4b[1]\n" - ".inst 0x4fa6e97f // sdot v31.4s, v11.16b, v6.4b[3]\n" + ".inst 0x4f84e238 // sdot v24.4s, v17.16b, v4.4b[0]\n" + ".inst 0x4f84ea39 // sdot v25.4s, v17.16b, v4.4b[2]\n" + ".inst 0x4f85e23a // sdot v26.4s, v17.16b, v5.4b[0]\n" + ".inst 0x4f85ea3b // sdot v27.4s, v17.16b, v5.4b[2]\n" + ".inst 0x4fa5e25c // sdot v28.4s, v18.16b, v5.4b[1]\n" + ".inst 0x4fa5ea5d // sdot v29.4s, v18.16b, v5.4b[3]\n" + ".inst 0x4fa6e25e // sdot v30.4s, v18.16b, v6.4b[1]\n" + ".inst 0x4fa6ea5f // sdot v31.4s, v18.16b, v6.4b[3]\n" "ldr q11, [%x[params], #0xc0]\n" - ".inst 0x4fa4e138 // sdot v24.4s, v9.16b, v4.4b[1]\n" - ".inst 0x4fa4e939 // sdot v25.4s, v9.16b, v4.4b[3]\n" - "sqrdmulh v24.4s, v24.4s, v21.4s\n" - ".inst 0x4fa5e13a // sdot v26.4s, v9.16b, v5.4b[1]\n" - ".inst 0x4fa5e93b // sdot v27.4s, v9.16b, v5.4b[3]\n" - "sqrdmulh v25.4s, v25.4s, v21.4s\n" - ".inst 0x4f86e11c // sdot v28.4s, v8.16b, v6.4b[0]\n" - ".inst 0x4f86e91d // sdot v29.4s, v8.16b, v6.4b[2]\n" - "sqrdmulh v26.4s, v26.4s, v21.4s\n" - ".inst 0x4f87e11e // sdot v30.4s, v8.16b, v7.4b[0]\n" - ".inst 0x4f87e91f // sdot v31.4s, v8.16b, v7.4b[2]\n" + ".inst 0x4fa4e218 // sdot v24.4s, v16.16b, v4.4b[1]\n" + ".inst 0x4fa4ea19 // sdot v25.4s, v16.16b, v4.4b[3]\n" + "sqrdmulh v24.4s, v24.4s, v12.4s\n" + ".inst 0x4fa5e21a // sdot v26.4s, v16.16b, v5.4b[1]\n" + ".inst 0x4fa5ea1b // sdot v27.4s, v16.16b, v5.4b[3]\n" + "sqrdmulh v25.4s, v25.4s, v12.4s\n" + ".inst 0x4f86e23c // sdot v28.4s, v17.16b, v6.4b[0]\n" + ".inst 0x4f86ea3d // sdot v29.4s, v17.16b, v6.4b[2]\n" + "sqrdmulh v26.4s, v26.4s, v12.4s\n" + ".inst 0x4f87e23e // sdot v30.4s, v17.16b, v7.4b[0]\n" + ".inst 0x4f87ea3f // sdot v31.4s, v17.16b, v7.4b[2]\n" "ldr q8, [%x[params], #0x90]\n" - "sqrdmulh v27.4s, v27.4s, v21.4s\n" - ".inst 0x4fa6e13c // sdot v28.4s, v9.16b, v6.4b[1]\n" - ".inst 0x4fa6e93d // sdot v29.4s, v9.16b, v6.4b[3]\n" - "and v19.16b, v24.16b, v20.16b\n" - ".inst 0x4fa7e13e // sdot v30.4s, v9.16b, v7.4b[1]\n" - ".inst 0x4fa7e93f // sdot v31.4s, v9.16b, v7.4b[3]\n" + "sqrdmulh v27.4s, v27.4s, v12.4s\n" + ".inst 0x4fa6e21c // sdot v28.4s, v16.16b, v6.4b[1]\n" + ".inst 0x4fa6ea1d // sdot v29.4s, v16.16b, v6.4b[3]\n" + "and v19.16b, v24.16b, v21.16b\n" + ".inst 0x4fa7e21e // sdot v30.4s, v16.16b, v7.4b[1]\n" + ".inst 0x4fa7ea1f // sdot v31.4s, v16.16b, v7.4b[3]\n" "ldr q9, [%x[params], #0xa0]\n" - "and v18.16b, v25.16b, v20.16b\n" + "and v18.16b, v25.16b, v21.16b\n" + "and v17.16b, v26.16b, v21.16b\n" + "and v16.16b, v27.16b, v21.16b\n" + "add %x[params], %x[params], #0xd0\n" "sshr v19.4s, v19.4s, #0x1f\n" "sshr v18.4s, v18.4s, #0x1f\n" - "add %x[params], %x[params], #0xd0\n" - "sqrdmulh v28.4s, v28.4s, v21.4s\n" - "sqrdmulh v29.4s, v29.4s, v21.4s\n" - "sqrdmulh v30.4s, v30.4s, v21.4s\n" - "sqrdmulh v31.4s, v31.4s, v21.4s\n" - "and v17.16b, v26.16b, v20.16b\n" "sshr v17.4s, v17.4s, #0x1f\n" - "sqadd v24.4s, v24.4s, v19.4s\n" - "and v16.16b, v27.16b, v20.16b\n" "sshr v16.4s, v16.4s, #0x1f\n" + "sqrdmulh v28.4s, v28.4s, v12.4s\n" + "sqrdmulh v29.4s, v29.4s, v12.4s\n" + "sqrdmulh v30.4s, v30.4s, v12.4s\n" + "sqrdmulh v31.4s, v31.4s, v12.4s\n" + "sqadd v24.4s, v24.4s, v19.4s\n" "sqadd v25.4s, v25.4s, v18.4s\n" "sqadd v26.4s, v26.4s, v17.4s\n" "sqadd v27.4s, v27.4s, v16.4s\n" - "and v19.16b, v28.16b, v20.16b\n" - "and v18.16b, v29.16b, v20.16b\n" - "and v17.16b, v30.16b, v20.16b\n" + "and v19.16b, v28.16b, v21.16b\n" + "and v18.16b, v29.16b, v21.16b\n" + "and v17.16b, v30.16b, v21.16b\n" + "and v16.16b, v31.16b, v21.16b\n" "sshr v19.4s, v19.4s, #0x1f\n" "sshr v18.4s, v18.4s, #0x1f\n" "sshr v17.4s, v17.4s, #0x1f\n" - "sqadd v28.4s, v28.4s, v19.4s\n" - "and v16.16b, v31.16b, v20.16b\n" "sshr v16.4s, v16.4s, #0x1f\n" + "sqadd v28.4s, v28.4s, v19.4s\n" "sqadd v29.4s, v29.4s, v18.4s\n" "sqadd v30.4s, v30.4s, v17.4s\n" "sqadd v31.4s, v31.4s, v16.4s\n" - "srshl v24.4s, v24.4s, v20.4s\n" - "srshl v25.4s, v25.4s, v20.4s\n" - "srshl v26.4s, v26.4s, v20.4s\n" - "srshl v27.4s, v27.4s, v20.4s\n" - "srshl v28.4s, v28.4s, v20.4s\n" - "srshl v29.4s, v29.4s, v20.4s\n" - "srshl v30.4s, v30.4s, v20.4s\n" - "srshl v31.4s, v31.4s, v20.4s\n" - "add v24.4s, v24.4s, v14.4s\n" - "add v25.4s, v25.4s, v14.4s\n" - "add v26.4s, v26.4s, v14.4s\n" - "add v27.4s, v27.4s, v14.4s\n" - "add v28.4s, v28.4s, v14.4s\n" - "add v29.4s, v29.4s, v14.4s\n" - "add v30.4s, v30.4s, v14.4s\n" - "add v31.4s, v31.4s, v14.4s\n" + "srshl v24.4s, v24.4s, v21.4s\n" + "srshl v25.4s, v25.4s, v21.4s\n" + "srshl v26.4s, v26.4s, v21.4s\n" + "srshl v27.4s, v27.4s, v21.4s\n" + "srshl v28.4s, v28.4s, v21.4s\n" + "srshl v29.4s, v29.4s, v21.4s\n" + "srshl v30.4s, v30.4s, v21.4s\n" + "srshl v31.4s, v31.4s, v21.4s\n" + "add v24.4s, v24.4s, v13.4s\n" + "add v25.4s, v25.4s, v13.4s\n" + "add v26.4s, v26.4s, v13.4s\n" + "add v27.4s, v27.4s, v13.4s\n" + "add v28.4s, v28.4s, v13.4s\n" + "add v29.4s, v29.4s, v13.4s\n" + "add v30.4s, v30.4s, v13.4s\n" + "add v31.4s, v31.4s, v13.4s\n" "smin v24.4s, v24.4s, v15.4s\n" "smin v25.4s, v25.4s, v15.4s\n" "smin v26.4s, v26.4s, v15.4s\n" @@ -347,14 +348,14 @@ void a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst_impl "smin v29.4s, v29.4s, v15.4s\n" "smin v30.4s, v30.4s, v15.4s\n" "smin v31.4s, v31.4s, v15.4s\n" - "smax v24.4s, v24.4s, v13.4s\n" - "smax v25.4s, v25.4s, v13.4s\n" - "smax v26.4s, v26.4s, v13.4s\n" - "smax v27.4s, v27.4s, v13.4s\n" - "smax v28.4s, v28.4s, v13.4s\n" - "smax v29.4s, v29.4s, v13.4s\n" - "smax v30.4s, v30.4s, v13.4s\n" - "smax v31.4s, v31.4s, v13.4s\n" + "smax v24.4s, v24.4s, v14.4s\n" + "smax v25.4s, v25.4s, v14.4s\n" + "smax v26.4s, v26.4s, v14.4s\n" + "smax v27.4s, v27.4s, v14.4s\n" + "smax v28.4s, v28.4s, v14.4s\n" + "smax v29.4s, v29.4s, v14.4s\n" + "smax v30.4s, v30.4s, v14.4s\n" + "smax v31.4s, v31.4s, v14.4s\n" "uzp1 v24.16b, v24.16b, v24.16b\n" "uzp1 v25.16b, v25.16b, v25.16b\n" "uzp1 v26.16b, v26.16b, v26.16b\n" @@ -388,14 +389,14 @@ void a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst_impl "dup v30.4s, v23.s[2]\n" "dup v31.4s, v23.s[3]\n" "add x28, x28, #0x4\n" - "add v24.4s, v24.4s, v12.4s\n" - "add v25.4s, v25.4s, v12.4s\n" - "add v26.4s, v26.4s, v12.4s\n" - "add v27.4s, v27.4s, v12.4s\n" - "add v28.4s, v28.4s, v12.4s\n" - "add v29.4s, v29.4s, v12.4s\n" - "add v30.4s, v30.4s, v12.4s\n" - "add v31.4s, v31.4s, v12.4s\n" + "add v24.4s, v24.4s, v20.4s\n" + "add v25.4s, v25.4s, v20.4s\n" + "add v26.4s, v26.4s, v20.4s\n" + "add v27.4s, v27.4s, v20.4s\n" + "add v28.4s, v28.4s, v20.4s\n" + "add v29.4s, v29.4s, v20.4s\n" + "add v30.4s, v30.4s, v20.4s\n" + "add v31.4s, v31.4s, v20.4s\n" "bgt 1b\n" "2:" // Tail "ldr q21, [%x[params], #0x60]\n" @@ -420,7 +421,7 @@ void a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst_impl "add x21, x21, x28\n" ".inst 0x4f83e11e // sdot v30.4s, v8.16b, v3.4b[0]\n" ".inst 0x4f83e91f // sdot v31.4s, v8.16b, v3.4b[2]\n" - "ldr q8, [%x[params], #0x0]\n" + "ldr q17, [%x[params], #0x0]\n" "add x20, x20, x28\n" ".inst 0x4f81e158 // sdot v24.4s, v10.16b, v1.4b[0]\n" ".inst 0x4f81e959 // sdot v25.4s, v10.16b, v1.4b[2]\n" @@ -430,7 +431,7 @@ void a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst_impl ".inst 0x4fa2e93d // sdot v29.4s, v9.16b, v2.4b[3]\n" ".inst 0x4fa3e13e // sdot v30.4s, v9.16b, v3.4b[1]\n" ".inst 0x4fa3e93f // sdot v31.4s, v9.16b, v3.4b[3]\n" - "ldr q9, [%x[params], #0x10]\n" + "ldr q16, [%x[params], #0x10]\n" ".inst 0x4fa1e178 // sdot v24.4s, v11.16b, v1.4b[1]\n" ".inst 0x4fa1e979 // sdot v25.4s, v11.16b, v1.4b[3]\n" ".inst 0x4fa2e17a // sdot v26.4s, v11.16b, v2.4b[1]\n" @@ -439,68 +440,68 @@ void a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst_impl ".inst 0x4f83e95d // sdot v29.4s, v10.16b, v3.4b[2]\n" ".inst 0x4f84e15e // sdot v30.4s, v10.16b, v4.4b[0]\n" ".inst 0x4f84e95f // sdot v31.4s, v10.16b, v4.4b[2]\n" - "ldr q10, [%x[params], #0x20]\n" - ".inst 0x4f82e118 // sdot v24.4s, v8.16b, v2.4b[0]\n" - ".inst 0x4f82e919 // sdot v25.4s, v8.16b, v2.4b[2]\n" - ".inst 0x4f83e11a // sdot v26.4s, v8.16b, v3.4b[0]\n" - ".inst 0x4f83e91b // sdot v27.4s, v8.16b, v3.4b[2]\n" + "ldr q19, [%x[params], #0x20]\n" + ".inst 0x4f82e238 // sdot v24.4s, v17.16b, v2.4b[0]\n" + ".inst 0x4f82ea39 // sdot v25.4s, v17.16b, v2.4b[2]\n" + ".inst 0x4f83e23a // sdot v26.4s, v17.16b, v3.4b[0]\n" + ".inst 0x4f83ea3b // sdot v27.4s, v17.16b, v3.4b[2]\n" ".inst 0x4fa3e17c // sdot v28.4s, v11.16b, v3.4b[1]\n" ".inst 0x4fa3e97d // sdot v29.4s, v11.16b, v3.4b[3]\n" ".inst 0x4fa4e17e // sdot v30.4s, v11.16b, v4.4b[1]\n" ".inst 0x4fa4e97f // sdot v31.4s, v11.16b, v4.4b[3]\n" - "ldr q11, [%x[params], #0x30]\n" - ".inst 0x4fa2e138 // sdot v24.4s, v9.16b, v2.4b[1]\n" - ".inst 0x4fa2e939 // sdot v25.4s, v9.16b, v2.4b[3]\n" - ".inst 0x4fa3e13a // sdot v26.4s, v9.16b, v3.4b[1]\n" - ".inst 0x4fa3e93b // sdot v27.4s, v9.16b, v3.4b[3]\n" - ".inst 0x4f84e11c // sdot v28.4s, v8.16b, v4.4b[0]\n" - ".inst 0x4f84e91d // sdot v29.4s, v8.16b, v4.4b[2]\n" - ".inst 0x4f85e11e // sdot v30.4s, v8.16b, v5.4b[0]\n" - ".inst 0x4f85e91f // sdot v31.4s, v8.16b, v5.4b[2]\n" - "ldr q8, [%x[params], #0x40]\n" - ".inst 0x4f83e158 // sdot v24.4s, v10.16b, v3.4b[0]\n" - ".inst 0x4f83e959 // sdot v25.4s, v10.16b, v3.4b[2]\n" - ".inst 0x4f84e15a // sdot v26.4s, v10.16b, v4.4b[0]\n" - ".inst 0x4f84e95b // sdot v27.4s, v10.16b, v4.4b[2]\n" - ".inst 0x4fa4e13c // sdot v28.4s, v9.16b, v4.4b[1]\n" - ".inst 0x4fa4e93d // sdot v29.4s, v9.16b, v4.4b[3]\n" - ".inst 0x4fa5e13e // sdot v30.4s, v9.16b, v5.4b[1]\n" - ".inst 0x4fa5e93f // sdot v31.4s, v9.16b, v5.4b[3]\n" - "ldr q9, [%x[params], #0x50]\n" + "ldr q18, [%x[params], #0x30]\n" + ".inst 0x4fa2e218 // sdot v24.4s, v16.16b, v2.4b[1]\n" + ".inst 0x4fa2ea19 // sdot v25.4s, v16.16b, v2.4b[3]\n" + ".inst 0x4fa3e21a // sdot v26.4s, v16.16b, v3.4b[1]\n" + ".inst 0x4fa3ea1b // sdot v27.4s, v16.16b, v3.4b[3]\n" + ".inst 0x4f84e23c // sdot v28.4s, v17.16b, v4.4b[0]\n" + ".inst 0x4f84ea3d // sdot v29.4s, v17.16b, v4.4b[2]\n" + ".inst 0x4f85e23e // sdot v30.4s, v17.16b, v5.4b[0]\n" + ".inst 0x4f85ea3f // sdot v31.4s, v17.16b, v5.4b[2]\n" + "ldr q17, [%x[params], #0x40]\n" + ".inst 0x4f83e278 // sdot v24.4s, v19.16b, v3.4b[0]\n" + ".inst 0x4f83ea79 // sdot v25.4s, v19.16b, v3.4b[2]\n" + ".inst 0x4f84e27a // sdot v26.4s, v19.16b, v4.4b[0]\n" + ".inst 0x4f84ea7b // sdot v27.4s, v19.16b, v4.4b[2]\n" + ".inst 0x4fa4e21c // sdot v28.4s, v16.16b, v4.4b[1]\n" + ".inst 0x4fa4ea1d // sdot v29.4s, v16.16b, v4.4b[3]\n" + ".inst 0x4fa5e21e // sdot v30.4s, v16.16b, v5.4b[1]\n" + ".inst 0x4fa5ea1f // sdot v31.4s, v16.16b, v5.4b[3]\n" + "ldr q16, [%x[params], #0x50]\n" "add %x[params], %x[params], #0x80\n" - ".inst 0x4fa3e178 // sdot v24.4s, v11.16b, v3.4b[1]\n" - ".inst 0x4fa3e979 // sdot v25.4s, v11.16b, v3.4b[3]\n" - ".inst 0x4fa4e17a // sdot v26.4s, v11.16b, v4.4b[1]\n" - ".inst 0x4fa4e97b // sdot v27.4s, v11.16b, v4.4b[3]\n" - ".inst 0x4f85e15c // sdot v28.4s, v10.16b, v5.4b[0]\n" - ".inst 0x4f85e95d // sdot v29.4s, v10.16b, v5.4b[2]\n" - ".inst 0x4f86e15e // sdot v30.4s, v10.16b, v6.4b[0]\n" - ".inst 0x4f86e95f // sdot v31.4s, v10.16b, v6.4b[2]\n" - ".inst 0x4f84e118 // sdot v24.4s, v8.16b, v4.4b[0]\n" - ".inst 0x4f84e919 // sdot v25.4s, v8.16b, v4.4b[2]\n" - ".inst 0x4f85e11a // sdot v26.4s, v8.16b, v5.4b[0]\n" - ".inst 0x4f85e91b // sdot v27.4s, v8.16b, v5.4b[2]\n" - ".inst 0x4fa5e17c // sdot v28.4s, v11.16b, v5.4b[1]\n" - ".inst 0x4fa5e97d // sdot v29.4s, v11.16b, v5.4b[3]\n" - ".inst 0x4fa6e17e // sdot v30.4s, v11.16b, v6.4b[1]\n" - ".inst 0x4fa6e97f // sdot v31.4s, v11.16b, v6.4b[3]\n" - ".inst 0x4fa4e138 // sdot v24.4s, v9.16b, v4.4b[1]\n" - ".inst 0x4fa4e939 // sdot v25.4s, v9.16b, v4.4b[3]\n" + ".inst 0x4fa3e258 // sdot v24.4s, v18.16b, v3.4b[1]\n" + ".inst 0x4fa3ea59 // sdot v25.4s, v18.16b, v3.4b[3]\n" + ".inst 0x4fa4e25a // sdot v26.4s, v18.16b, v4.4b[1]\n" + ".inst 0x4fa4ea5b // sdot v27.4s, v18.16b, v4.4b[3]\n" + ".inst 0x4f85e27c // sdot v28.4s, v19.16b, v5.4b[0]\n" + ".inst 0x4f85ea7d // sdot v29.4s, v19.16b, v5.4b[2]\n" + ".inst 0x4f86e27e // sdot v30.4s, v19.16b, v6.4b[0]\n" + ".inst 0x4f86ea7f // sdot v31.4s, v19.16b, v6.4b[2]\n" + ".inst 0x4f84e238 // sdot v24.4s, v17.16b, v4.4b[0]\n" + ".inst 0x4f84ea39 // sdot v25.4s, v17.16b, v4.4b[2]\n" + ".inst 0x4f85e23a // sdot v26.4s, v17.16b, v5.4b[0]\n" + ".inst 0x4f85ea3b // sdot v27.4s, v17.16b, v5.4b[2]\n" + ".inst 0x4fa5e25c // sdot v28.4s, v18.16b, v5.4b[1]\n" + ".inst 0x4fa5ea5d // sdot v29.4s, v18.16b, v5.4b[3]\n" + ".inst 0x4fa6e25e // sdot v30.4s, v18.16b, v6.4b[1]\n" + ".inst 0x4fa6ea5f // sdot v31.4s, v18.16b, v6.4b[3]\n" + ".inst 0x4fa4e218 // sdot v24.4s, v16.16b, v4.4b[1]\n" + ".inst 0x4fa4ea19 // sdot v25.4s, v16.16b, v4.4b[3]\n" "sqrdmulh v24.4s, v24.4s, v21.4s\n" - ".inst 0x4fa5e13a // sdot v26.4s, v9.16b, v5.4b[1]\n" - ".inst 0x4fa5e93b // sdot v27.4s, v9.16b, v5.4b[3]\n" + ".inst 0x4fa5e21a // sdot v26.4s, v16.16b, v5.4b[1]\n" + ".inst 0x4fa5ea1b // sdot v27.4s, v16.16b, v5.4b[3]\n" "sqrdmulh v25.4s, v25.4s, v21.4s\n" - ".inst 0x4f86e11c // sdot v28.4s, v8.16b, v6.4b[0]\n" - ".inst 0x4f86e91d // sdot v29.4s, v8.16b, v6.4b[2]\n" + ".inst 0x4f86e23c // sdot v28.4s, v17.16b, v6.4b[0]\n" + ".inst 0x4f86ea3d // sdot v29.4s, v17.16b, v6.4b[2]\n" "sqrdmulh v26.4s, v26.4s, v21.4s\n" - ".inst 0x4f87e11e // sdot v30.4s, v8.16b, v7.4b[0]\n" - ".inst 0x4f87e91f // sdot v31.4s, v8.16b, v7.4b[2]\n" + ".inst 0x4f87e23e // sdot v30.4s, v17.16b, v7.4b[0]\n" + ".inst 0x4f87ea3f // sdot v31.4s, v17.16b, v7.4b[2]\n" "sqrdmulh v27.4s, v27.4s, v21.4s\n" - ".inst 0x4fa6e13c // sdot v28.4s, v9.16b, v6.4b[1]\n" - ".inst 0x4fa6e93d // sdot v29.4s, v9.16b, v6.4b[3]\n" + ".inst 0x4fa6e21c // sdot v28.4s, v16.16b, v6.4b[1]\n" + ".inst 0x4fa6ea1d // sdot v29.4s, v16.16b, v6.4b[3]\n" "and v19.16b, v24.16b, v20.16b\n" - ".inst 0x4fa7e13e // sdot v30.4s, v9.16b, v7.4b[1]\n" - ".inst 0x4fa7e93f // sdot v31.4s, v9.16b, v7.4b[3]\n" + ".inst 0x4fa7e21e // sdot v30.4s, v16.16b, v7.4b[1]\n" + ".inst 0x4fa7ea1f // sdot v31.4s, v16.16b, v7.4b[3]\n" "and v18.16b, v25.16b, v20.16b\n" "and v17.16b, v26.16b, v20.16b\n" "and v16.16b, v27.16b, v20.16b\n" @@ -536,14 +537,14 @@ void a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst_impl "srshl v29.4s, v29.4s, v20.4s\n" "srshl v30.4s, v30.4s, v20.4s\n" "srshl v31.4s, v31.4s, v20.4s\n" - "add v24.4s, v24.4s, v14.4s\n" - "add v25.4s, v25.4s, v14.4s\n" - "add v26.4s, v26.4s, v14.4s\n" - "add v27.4s, v27.4s, v14.4s\n" - "add v28.4s, v28.4s, v14.4s\n" - "add v29.4s, v29.4s, v14.4s\n" - "add v30.4s, v30.4s, v14.4s\n" - "add v31.4s, v31.4s, v14.4s\n" + "add v24.4s, v24.4s, v13.4s\n" + "add v25.4s, v25.4s, v13.4s\n" + "add v26.4s, v26.4s, v13.4s\n" + "add v27.4s, v27.4s, v13.4s\n" + "add v28.4s, v28.4s, v13.4s\n" + "add v29.4s, v29.4s, v13.4s\n" + "add v30.4s, v30.4s, v13.4s\n" + "add v31.4s, v31.4s, v13.4s\n" "smin v24.4s, v24.4s, v15.4s\n" "smin v25.4s, v25.4s, v15.4s\n" "smin v26.4s, v26.4s, v15.4s\n" @@ -552,14 +553,14 @@ void a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst_impl "smin v29.4s, v29.4s, v15.4s\n" "smin v30.4s, v30.4s, v15.4s\n" "smin v31.4s, v31.4s, v15.4s\n" - "smax v24.4s, v24.4s, v13.4s\n" - "smax v25.4s, v25.4s, v13.4s\n" - "smax v26.4s, v26.4s, v13.4s\n" - "smax v27.4s, v27.4s, v13.4s\n" - "smax v28.4s, v28.4s, v13.4s\n" - "smax v29.4s, v29.4s, v13.4s\n" - "smax v30.4s, v30.4s, v13.4s\n" - "smax v31.4s, v31.4s, v13.4s\n" + "smax v24.4s, v24.4s, v14.4s\n" + "smax v25.4s, v25.4s, v14.4s\n" + "smax v26.4s, v26.4s, v14.4s\n" + "smax v27.4s, v27.4s, v14.4s\n" + "smax v28.4s, v28.4s, v14.4s\n" + "smax v29.4s, v29.4s, v14.4s\n" + "smax v30.4s, v30.4s, v14.4s\n" + "smax v31.4s, v31.4s, v14.4s\n" "uzp1 v24.16b, v24.16b, v24.16b\n" "uzp1 v25.16b, v25.16b, v25.16b\n" "uzp1 v26.16b, v26.16b, v26.16b\n" @@ -635,4 +636,5 @@ void a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst_impl } // namespace depthwise } // namespace arm_conv -#endif // defined(__aarch64__) + +#endif // defined(__aarch64__) -- cgit v1.2.1