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 | 542 ++++++++++----------- 1 file changed, 271 insertions(+), 271 deletions(-) (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst/generic.cpp') diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst/generic.cpp index 69b3865a65..9cafd23fb8 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst/generic.cpp @@ -22,10 +22,11 @@ * SOFTWARE. */ -#if defined(__aarch64__) #include #include +#if defined(__aarch64__) + namespace arm_conv { namespace depthwise { @@ -41,7 +42,7 @@ void a64_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst_imp const float minmax_vals[2] = { activation_min, activation_max }; __asm__ __volatile__( - "ld1r { v24.4s }, [%x[clamps]]\n" + "ld1r { v27.4s }, [%x[clamps]]\n" "ldr x21, [%x[inptrs], #0x0]\n" "lsr x22, %x[channel_multiplier], #0x2\n" "add x20, %x[clamps], #0x4\n" @@ -49,7 +50,7 @@ void a64_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst_imp "ldr q1, [x21, #0x10]\n" "mov x21, #0x0\n" "mov x14, #0x0\n" - "ld1r { v23.4s }, [x20]\n" + "ld1r { v26.4s }, [x20]\n" "ldr x20, [%x[inptrs], #0x8]\n" "ldr q2, [x20, #0x0]\n" "ldr q3, [x20, #0x10]\n" @@ -101,7 +102,7 @@ void a64_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst_imp "fmla v20.4s, v31.4s, v8.s[0]\n" "fmla v21.4s, v31.4s, v8.s[2]\n" "fmla v22.4s, v31.4s, v9.s[0]\n" - "ldr q31, [%x[params], #0x0]\n" + "ldr q25, [%x[params], #0x0]\n" "fmla v14.4s, v30.4s, v0.s[1]\n" "fmla v15.4s, v30.4s, v0.s[3]\n" "fmla v16.4s, v30.4s, v1.s[1]\n" @@ -111,7 +112,7 @@ void a64_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst_imp "fmla v20.4s, v30.4s, v8.s[1]\n" "fmla v21.4s, v30.4s, v8.s[3]\n" "fmla v22.4s, v30.4s, v9.s[1]\n" - "ldr q30, [%x[params], #0x10]\n" + "ldr q24, [%x[params], #0x10]\n" "fmla v14.4s, v29.4s, v0.s[2]\n" "fmla v15.4s, v29.4s, v1.s[0]\n" "fmla v16.4s, v29.4s, v1.s[2]\n" @@ -121,92 +122,92 @@ void a64_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst_imp "fmla v20.4s, v29.4s, v8.s[2]\n" "fmla v21.4s, v29.4s, v9.s[0]\n" "fmla v22.4s, v29.4s, v9.s[2]\n" - "ldr q29, [%x[params], #0x20]\n" - "fmla v14.4s, v31.4s, v2.s[0]\n" - "fmla v15.4s, v31.4s, v2.s[2]\n" - "fmla v16.4s, v31.4s, v3.s[0]\n" - "fmla v17.4s, v31.4s, v6.s[0]\n" - "fmla v18.4s, v31.4s, v6.s[2]\n" - "fmla v19.4s, v31.4s, v7.s[0]\n" - "fmla v20.4s, v31.4s, v10.s[0]\n" - "fmla v21.4s, v31.4s, v10.s[2]\n" - "fmla v22.4s, v31.4s, v11.s[0]\n" - "ldr q31, [%x[params], #0x30]\n" - "fmla v14.4s, v30.4s, v2.s[1]\n" - "fmla v15.4s, v30.4s, v2.s[3]\n" - "fmla v16.4s, v30.4s, v3.s[1]\n" - "fmla v17.4s, v30.4s, v6.s[1]\n" - "fmla v18.4s, v30.4s, v6.s[3]\n" - "fmla v19.4s, v30.4s, v7.s[1]\n" - "fmla v20.4s, v30.4s, v10.s[1]\n" - "fmla v21.4s, v30.4s, v10.s[3]\n" - "fmla v22.4s, v30.4s, v11.s[1]\n" - "ldr q30, [%x[params], #0x40]\n" - "fmla v14.4s, v29.4s, v2.s[2]\n" - "fmla v15.4s, v29.4s, v3.s[0]\n" - "fmla v16.4s, v29.4s, v3.s[2]\n" - "fmla v17.4s, v29.4s, v6.s[2]\n" - "fmla v18.4s, v29.4s, v7.s[0]\n" - "fmla v19.4s, v29.4s, v7.s[2]\n" - "fmla v20.4s, v29.4s, v10.s[2]\n" - "fmla v21.4s, v29.4s, v11.s[0]\n" - "fmla v22.4s, v29.4s, v11.s[2]\n" - "ldr q29, [%x[params], #0x50]\n" - "fmla v14.4s, v31.4s, v4.s[0]\n" - "fmla v15.4s, v31.4s, v4.s[2]\n" - "fmla v16.4s, v31.4s, v5.s[0]\n" - "fmla v17.4s, v31.4s, v8.s[0]\n" - "fmla v18.4s, v31.4s, v8.s[2]\n" - "fmla v19.4s, v31.4s, v9.s[0]\n" - "fmla v20.4s, v31.4s, v12.s[0]\n" - "fmla v21.4s, v31.4s, v12.s[2]\n" - "fmla v22.4s, v31.4s, v13.s[0]\n" + "ldr q23, [%x[params], #0x20]\n" + "fmla v14.4s, v25.4s, v2.s[0]\n" + "fmla v15.4s, v25.4s, v2.s[2]\n" + "fmla v16.4s, v25.4s, v3.s[0]\n" + "fmla v17.4s, v25.4s, v6.s[0]\n" + "fmla v18.4s, v25.4s, v6.s[2]\n" + "fmla v19.4s, v25.4s, v7.s[0]\n" + "fmla v20.4s, v25.4s, v10.s[0]\n" + "fmla v21.4s, v25.4s, v10.s[2]\n" + "fmla v22.4s, v25.4s, v11.s[0]\n" + "ldr q25, [%x[params], #0x30]\n" + "fmla v14.4s, v24.4s, v2.s[1]\n" + "fmla v15.4s, v24.4s, v2.s[3]\n" + "fmla v16.4s, v24.4s, v3.s[1]\n" + "fmla v17.4s, v24.4s, v6.s[1]\n" + "fmla v18.4s, v24.4s, v6.s[3]\n" + "fmla v19.4s, v24.4s, v7.s[1]\n" + "fmla v20.4s, v24.4s, v10.s[1]\n" + "fmla v21.4s, v24.4s, v10.s[3]\n" + "fmla v22.4s, v24.4s, v11.s[1]\n" + "ldr q24, [%x[params], #0x40]\n" + "fmla v14.4s, v23.4s, v2.s[2]\n" + "fmla v15.4s, v23.4s, v3.s[0]\n" + "fmla v16.4s, v23.4s, v3.s[2]\n" + "fmla v17.4s, v23.4s, v6.s[2]\n" + "fmla v18.4s, v23.4s, v7.s[0]\n" + "fmla v19.4s, v23.4s, v7.s[2]\n" + "fmla v20.4s, v23.4s, v10.s[2]\n" + "fmla v21.4s, v23.4s, v11.s[0]\n" + "fmla v22.4s, v23.4s, v11.s[2]\n" + "ldr q23, [%x[params], #0x50]\n" + "fmla v14.4s, v25.4s, v4.s[0]\n" + "fmla v15.4s, v25.4s, v4.s[2]\n" + "fmla v16.4s, v25.4s, v5.s[0]\n" + "fmla v17.4s, v25.4s, v8.s[0]\n" + "fmla v18.4s, v25.4s, v8.s[2]\n" + "fmla v19.4s, v25.4s, v9.s[0]\n" + "fmla v20.4s, v25.4s, v12.s[0]\n" + "fmla v21.4s, v25.4s, v12.s[2]\n" + "fmla v22.4s, v25.4s, v13.s[0]\n" "ldr q31, [%x[params], #0x70]\n" - "fmla v14.4s, v30.4s, v4.s[1]\n" - "fmla v15.4s, v30.4s, v4.s[3]\n" - "fmla v16.4s, v30.4s, v5.s[1]\n" - "fmla v17.4s, v30.4s, v8.s[1]\n" - "fmla v18.4s, v30.4s, v8.s[3]\n" - "fmla v19.4s, v30.4s, v9.s[1]\n" - "fmla v20.4s, v30.4s, v12.s[1]\n" - "fmla v21.4s, v30.4s, v12.s[3]\n" - "fmla v22.4s, v30.4s, v13.s[1]\n" + "fmla v14.4s, v24.4s, v4.s[1]\n" + "fmla v15.4s, v24.4s, v4.s[3]\n" + "fmla v16.4s, v24.4s, v5.s[1]\n" + "fmla v17.4s, v24.4s, v8.s[1]\n" + "fmla v18.4s, v24.4s, v8.s[3]\n" + "fmla v19.4s, v24.4s, v9.s[1]\n" + "fmla v20.4s, v24.4s, v12.s[1]\n" + "fmla v21.4s, v24.4s, v12.s[3]\n" + "fmla v22.4s, v24.4s, v13.s[1]\n" "ldr q30, [%x[params], #0x80]\n" - "fmla v14.4s, v29.4s, v4.s[2]\n" - "fmla v15.4s, v29.4s, v5.s[0]\n" - "fmin v14.4s, v14.4s, v23.4s\n" - "fmla v16.4s, v29.4s, v5.s[2]\n" - "fmla v17.4s, v29.4s, v8.s[2]\n" - "fmax v14.4s, v14.4s, v24.4s\n" + "fmla v14.4s, v23.4s, v4.s[2]\n" + "fmla v15.4s, v23.4s, v5.s[0]\n" + "fmin v14.4s, v14.4s, v26.4s\n" + "fmla v16.4s, v23.4s, v5.s[2]\n" + "fmla v17.4s, v23.4s, v8.s[2]\n" + "fmax v14.4s, v14.4s, v27.4s\n" "str q14, [x13, x14]\n" "ldr q14, [%x[params], #0x60]\n" - "fmla v18.4s, v29.4s, v9.s[0]\n" - "fmla v19.4s, v29.4s, v9.s[2]\n" - "fmin v15.4s, v15.4s, v23.4s\n" - "fmla v20.4s, v29.4s, v12.s[2]\n" - "fmla v21.4s, v29.4s, v13.s[0]\n" - "fmin v16.4s, v16.4s, v23.4s\n" - "fmla v22.4s, v29.4s, v13.s[2]\n" + "fmla v18.4s, v23.4s, v9.s[0]\n" + "fmla v19.4s, v23.4s, v9.s[2]\n" + "fmin v15.4s, v15.4s, v26.4s\n" + "fmla v20.4s, v23.4s, v12.s[2]\n" + "fmla v21.4s, v23.4s, v13.s[0]\n" + "fmin v16.4s, v16.4s, v26.4s\n" + "fmla v22.4s, v23.4s, v13.s[2]\n" "ldr q29, [%x[params], #0x90]\n" - "fmin v17.4s, v17.4s, v23.4s\n" + "fmin v17.4s, v17.4s, v26.4s\n" "add %x[params], %x[params], #0xa0\n" - "fmin v18.4s, v18.4s, v23.4s\n" - "fmin v19.4s, v19.4s, v23.4s\n" - "fmin v20.4s, v20.4s, v23.4s\n" - "fmin v21.4s, v21.4s, v23.4s\n" - "fmin v22.4s, v22.4s, v23.4s\n" - "fmax v15.4s, v15.4s, v24.4s\n" + "fmin v18.4s, v18.4s, v26.4s\n" + "fmin v19.4s, v19.4s, v26.4s\n" + "fmin v20.4s, v20.4s, v26.4s\n" + "fmin v21.4s, v21.4s, v26.4s\n" + "fmin v22.4s, v22.4s, v26.4s\n" + "fmax v15.4s, v15.4s, v27.4s\n" "str q15, [x12, x14]\n" - "fmax v16.4s, v16.4s, v24.4s\n" - "fmax v17.4s, v17.4s, v24.4s\n" + "fmax v16.4s, v16.4s, v27.4s\n" + "fmax v17.4s, v17.4s, v27.4s\n" "str q16, [x11, x14]\n" - "fmax v18.4s, v18.4s, v24.4s\n" - "fmax v19.4s, v19.4s, v24.4s\n" + "fmax v18.4s, v18.4s, v27.4s\n" + "fmax v19.4s, v19.4s, v27.4s\n" "str q17, [x10, x14]\n" - "fmax v20.4s, v20.4s, v24.4s\n" - "fmax v21.4s, v21.4s, v24.4s\n" + "fmax v20.4s, v20.4s, v27.4s\n" + "fmax v21.4s, v21.4s, v27.4s\n" "str q18, [x9, x14]\n" - "fmax v22.4s, v22.4s, v24.4s\n" + "fmax v22.4s, v22.4s, v27.4s\n" "str q19, [x28, x14]\n" "mov v15.16b, v14.16b\n" "str q20, [x27, x14]\n" @@ -231,7 +232,7 @@ void a64_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst_imp "fmla v20.4s, v31.4s, v8.s[0]\n" "fmla v21.4s, v31.4s, v8.s[2]\n" "fmla v22.4s, v31.4s, v9.s[0]\n" - "ldr q31, [%x[params], #0x0]\n" + "ldr q25, [%x[params], #0x0]\n" "fmla v14.4s, v30.4s, v0.s[1]\n" "fmla v15.4s, v30.4s, v0.s[3]\n" "fmla v16.4s, v30.4s, v1.s[1]\n" @@ -241,7 +242,7 @@ void a64_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst_imp "fmla v20.4s, v30.4s, v8.s[1]\n" "fmla v21.4s, v30.4s, v8.s[3]\n" "fmla v22.4s, v30.4s, v9.s[1]\n" - "ldr q30, [%x[params], #0x10]\n" + "ldr q24, [%x[params], #0x10]\n" "fmla v14.4s, v29.4s, v0.s[2]\n" "fmla v15.4s, v29.4s, v1.s[0]\n" "fmla v16.4s, v29.4s, v1.s[2]\n" @@ -251,87 +252,87 @@ void a64_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst_imp "fmla v20.4s, v29.4s, v8.s[2]\n" "fmla v21.4s, v29.4s, v9.s[0]\n" "fmla v22.4s, v29.4s, v9.s[2]\n" - "ldr q29, [%x[params], #0x20]\n" - "fmla v14.4s, v31.4s, v2.s[0]\n" - "fmla v15.4s, v31.4s, v2.s[2]\n" - "fmla v16.4s, v31.4s, v3.s[0]\n" - "fmla v17.4s, v31.4s, v6.s[0]\n" - "fmla v18.4s, v31.4s, v6.s[2]\n" - "fmla v19.4s, v31.4s, v7.s[0]\n" - "fmla v20.4s, v31.4s, v10.s[0]\n" - "fmla v21.4s, v31.4s, v10.s[2]\n" - "fmla v22.4s, v31.4s, v11.s[0]\n" - "ldr q31, [%x[params], #0x30]\n" - "fmla v14.4s, v30.4s, v2.s[1]\n" - "fmla v15.4s, v30.4s, v2.s[3]\n" - "fmla v16.4s, v30.4s, v3.s[1]\n" - "fmla v17.4s, v30.4s, v6.s[1]\n" - "fmla v18.4s, v30.4s, v6.s[3]\n" - "fmla v19.4s, v30.4s, v7.s[1]\n" - "fmla v20.4s, v30.4s, v10.s[1]\n" - "fmla v21.4s, v30.4s, v10.s[3]\n" - "fmla v22.4s, v30.4s, v11.s[1]\n" - "ldr q30, [%x[params], #0x40]\n" - "fmla v14.4s, v29.4s, v2.s[2]\n" - "fmla v15.4s, v29.4s, v3.s[0]\n" - "fmla v16.4s, v29.4s, v3.s[2]\n" - "fmla v17.4s, v29.4s, v6.s[2]\n" - "fmla v18.4s, v29.4s, v7.s[0]\n" - "fmla v19.4s, v29.4s, v7.s[2]\n" - "fmla v20.4s, v29.4s, v10.s[2]\n" - "fmla v21.4s, v29.4s, v11.s[0]\n" - "fmla v22.4s, v29.4s, v11.s[2]\n" - "ldr q29, [%x[params], #0x50]\n" + "ldr q23, [%x[params], #0x20]\n" + "fmla v14.4s, v25.4s, v2.s[0]\n" + "fmla v15.4s, v25.4s, v2.s[2]\n" + "fmla v16.4s, v25.4s, v3.s[0]\n" + "fmla v17.4s, v25.4s, v6.s[0]\n" + "fmla v18.4s, v25.4s, v6.s[2]\n" + "fmla v19.4s, v25.4s, v7.s[0]\n" + "fmla v20.4s, v25.4s, v10.s[0]\n" + "fmla v21.4s, v25.4s, v10.s[2]\n" + "fmla v22.4s, v25.4s, v11.s[0]\n" + "ldr q25, [%x[params], #0x30]\n" + "fmla v14.4s, v24.4s, v2.s[1]\n" + "fmla v15.4s, v24.4s, v2.s[3]\n" + "fmla v16.4s, v24.4s, v3.s[1]\n" + "fmla v17.4s, v24.4s, v6.s[1]\n" + "fmla v18.4s, v24.4s, v6.s[3]\n" + "fmla v19.4s, v24.4s, v7.s[1]\n" + "fmla v20.4s, v24.4s, v10.s[1]\n" + "fmla v21.4s, v24.4s, v10.s[3]\n" + "fmla v22.4s, v24.4s, v11.s[1]\n" + "ldr q24, [%x[params], #0x40]\n" + "fmla v14.4s, v23.4s, v2.s[2]\n" + "fmla v15.4s, v23.4s, v3.s[0]\n" + "fmla v16.4s, v23.4s, v3.s[2]\n" + "fmla v17.4s, v23.4s, v6.s[2]\n" + "fmla v18.4s, v23.4s, v7.s[0]\n" + "fmla v19.4s, v23.4s, v7.s[2]\n" + "fmla v20.4s, v23.4s, v10.s[2]\n" + "fmla v21.4s, v23.4s, v11.s[0]\n" + "fmla v22.4s, v23.4s, v11.s[2]\n" + "ldr q23, [%x[params], #0x50]\n" "add %x[params], %x[params], #0x60\n" - "fmla v14.4s, v31.4s, v4.s[0]\n" - "fmla v15.4s, v31.4s, v4.s[2]\n" - "fmla v16.4s, v31.4s, v5.s[0]\n" - "fmla v17.4s, v31.4s, v8.s[0]\n" - "fmla v18.4s, v31.4s, v8.s[2]\n" - "fmla v19.4s, v31.4s, v9.s[0]\n" - "fmla v20.4s, v31.4s, v12.s[0]\n" - "fmla v21.4s, v31.4s, v12.s[2]\n" - "fmla v22.4s, v31.4s, v13.s[0]\n" - "fmla v14.4s, v30.4s, v4.s[1]\n" - "fmla v15.4s, v30.4s, v4.s[3]\n" - "fmla v16.4s, v30.4s, v5.s[1]\n" - "fmla v17.4s, v30.4s, v8.s[1]\n" - "fmla v18.4s, v30.4s, v8.s[3]\n" - "fmla v19.4s, v30.4s, v9.s[1]\n" - "fmla v20.4s, v30.4s, v12.s[1]\n" - "fmla v21.4s, v30.4s, v12.s[3]\n" - "fmla v22.4s, v30.4s, v13.s[1]\n" - "fmla v14.4s, v29.4s, v4.s[2]\n" - "fmla v15.4s, v29.4s, v5.s[0]\n" - "fmin v14.4s, v14.4s, v23.4s\n" - "fmla v16.4s, v29.4s, v5.s[2]\n" - "fmla v17.4s, v29.4s, v8.s[2]\n" - "fmin v15.4s, v15.4s, v23.4s\n" - "fmla v18.4s, v29.4s, v9.s[0]\n" - "fmla v19.4s, v29.4s, v9.s[2]\n" - "fmin v16.4s, v16.4s, v23.4s\n" - "fmla v20.4s, v29.4s, v12.s[2]\n" - "fmla v21.4s, v29.4s, v13.s[0]\n" - "fmin v17.4s, v17.4s, v23.4s\n" - "fmla v22.4s, v29.4s, v13.s[2]\n" - "fmin v18.4s, v18.4s, v23.4s\n" - "fmin v19.4s, v19.4s, v23.4s\n" - "fmin v20.4s, v20.4s, v23.4s\n" - "fmin v21.4s, v21.4s, v23.4s\n" - "fmin v22.4s, v22.4s, v23.4s\n" - "fmax v14.4s, v14.4s, v24.4s\n" - "fmax v15.4s, v15.4s, v24.4s\n" + "fmla v14.4s, v25.4s, v4.s[0]\n" + "fmla v15.4s, v25.4s, v4.s[2]\n" + "fmla v16.4s, v25.4s, v5.s[0]\n" + "fmla v17.4s, v25.4s, v8.s[0]\n" + "fmla v18.4s, v25.4s, v8.s[2]\n" + "fmla v19.4s, v25.4s, v9.s[0]\n" + "fmla v20.4s, v25.4s, v12.s[0]\n" + "fmla v21.4s, v25.4s, v12.s[2]\n" + "fmla v22.4s, v25.4s, v13.s[0]\n" + "fmla v14.4s, v24.4s, v4.s[1]\n" + "fmla v15.4s, v24.4s, v4.s[3]\n" + "fmla v16.4s, v24.4s, v5.s[1]\n" + "fmla v17.4s, v24.4s, v8.s[1]\n" + "fmla v18.4s, v24.4s, v8.s[3]\n" + "fmla v19.4s, v24.4s, v9.s[1]\n" + "fmla v20.4s, v24.4s, v12.s[1]\n" + "fmla v21.4s, v24.4s, v12.s[3]\n" + "fmla v22.4s, v24.4s, v13.s[1]\n" + "fmla v14.4s, v23.4s, v4.s[2]\n" + "fmla v15.4s, v23.4s, v5.s[0]\n" + "fmin v14.4s, v14.4s, v26.4s\n" + "fmla v16.4s, v23.4s, v5.s[2]\n" + "fmla v17.4s, v23.4s, v8.s[2]\n" + "fmin v15.4s, v15.4s, v26.4s\n" + "fmla v18.4s, v23.4s, v9.s[0]\n" + "fmla v19.4s, v23.4s, v9.s[2]\n" + "fmin v16.4s, v16.4s, v26.4s\n" + "fmla v20.4s, v23.4s, v12.s[2]\n" + "fmla v21.4s, v23.4s, v13.s[0]\n" + "fmin v17.4s, v17.4s, v26.4s\n" + "fmla v22.4s, v23.4s, v13.s[2]\n" + "fmin v18.4s, v18.4s, v26.4s\n" + "fmin v19.4s, v19.4s, v26.4s\n" + "fmin v20.4s, v20.4s, v26.4s\n" + "fmin v21.4s, v21.4s, v26.4s\n" + "fmin v22.4s, v22.4s, v26.4s\n" + "fmax v14.4s, v14.4s, v27.4s\n" + "fmax v15.4s, v15.4s, v27.4s\n" "str q14, [x13, x14]\n" - "fmax v16.4s, v16.4s, v24.4s\n" - "fmax v17.4s, v17.4s, v24.4s\n" + "fmax v16.4s, v16.4s, v27.4s\n" + "fmax v17.4s, v17.4s, v27.4s\n" "str q15, [x12, x14]\n" - "fmax v18.4s, v18.4s, v24.4s\n" - "fmax v19.4s, v19.4s, v24.4s\n" + "fmax v18.4s, v18.4s, v27.4s\n" + "fmax v19.4s, v19.4s, v27.4s\n" "str q16, [x11, x14]\n" - "fmax v20.4s, v20.4s, v24.4s\n" - "fmax v21.4s, v21.4s, v24.4s\n" + "fmax v20.4s, v20.4s, v27.4s\n" + "fmax v21.4s, v21.4s, v27.4s\n" "str q17, [x10, x14]\n" - "fmax v22.4s, v22.4s, v24.4s\n" + "fmax v22.4s, v22.4s, v27.4s\n" "str q18, [x9, x14]\n" "str q19, [x28, x14]\n" "str q20, [x27, x14]\n" @@ -342,123 +343,123 @@ void a64_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst_imp "tst %x[channel_multiplier], #0x3\n" "beq 6f\n" "ldr q14, [%x[params], #0x0]\n" - "ldr q31, [%x[params], #0x10]\n" + "ldr q25, [%x[params], #0x10]\n" "mov v15.16b, v14.16b\n" "mov v16.16b, v14.16b\n" - "ldr q30, [%x[params], #0x20]\n" - "ldr q29, [%x[params], #0x30]\n" + "ldr q24, [%x[params], #0x20]\n" + "ldr q23, [%x[params], #0x30]\n" "mov v17.16b, v14.16b\n" "mov v18.16b, v14.16b\n" "mov v19.16b, v14.16b\n" "mov v20.16b, v14.16b\n" - "fmla v15.4s, v31.4s, v0.s[2]\n" + "fmla v15.4s, v25.4s, v0.s[2]\n" "mov v21.16b, v14.16b\n" "mov v22.16b, v14.16b\n" - "fmla v14.4s, v31.4s, v0.s[0]\n" - "fmla v16.4s, v31.4s, v1.s[0]\n" - "fmla v17.4s, v31.4s, v4.s[0]\n" - "fmla v18.4s, v31.4s, v4.s[2]\n" - "fmla v19.4s, v31.4s, v5.s[0]\n" - "fmla v20.4s, v31.4s, v8.s[0]\n" - "fmla v21.4s, v31.4s, v8.s[2]\n" - "fmla v22.4s, v31.4s, v9.s[0]\n" - "ldr q31, [%x[params], #0x40]\n" - "fmla v14.4s, v30.4s, v0.s[1]\n" - "fmla v15.4s, v30.4s, v0.s[3]\n" - "fmla v16.4s, v30.4s, v1.s[1]\n" - "fmla v17.4s, v30.4s, v4.s[1]\n" - "fmla v18.4s, v30.4s, v4.s[3]\n" - "fmla v19.4s, v30.4s, v5.s[1]\n" - "fmla v20.4s, v30.4s, v8.s[1]\n" - "fmla v21.4s, v30.4s, v8.s[3]\n" - "fmla v22.4s, v30.4s, v9.s[1]\n" - "ldr q30, [%x[params], #0x50]\n" - "fmla v14.4s, v29.4s, v0.s[2]\n" - "fmla v15.4s, v29.4s, v1.s[0]\n" - "fmla v16.4s, v29.4s, v1.s[2]\n" - "fmla v17.4s, v29.4s, v4.s[2]\n" - "fmla v18.4s, v29.4s, v5.s[0]\n" - "fmla v19.4s, v29.4s, v5.s[2]\n" - "fmla v20.4s, v29.4s, v8.s[2]\n" - "fmla v21.4s, v29.4s, v9.s[0]\n" - "fmla v22.4s, v29.4s, v9.s[2]\n" - "ldr q29, [%x[params], #0x60]\n" - "fmla v14.4s, v31.4s, v2.s[0]\n" - "fmla v15.4s, v31.4s, v2.s[2]\n" - "fmla v16.4s, v31.4s, v3.s[0]\n" - "fmla v17.4s, v31.4s, v6.s[0]\n" - "fmla v18.4s, v31.4s, v6.s[2]\n" - "fmla v19.4s, v31.4s, v7.s[0]\n" - "fmla v20.4s, v31.4s, v10.s[0]\n" - "fmla v21.4s, v31.4s, v10.s[2]\n" - "fmla v22.4s, v31.4s, v11.s[0]\n" - "ldr q31, [%x[params], #0x70]\n" - "fmla v14.4s, v30.4s, v2.s[1]\n" - "fmla v15.4s, v30.4s, v2.s[3]\n" - "fmla v16.4s, v30.4s, v3.s[1]\n" - "fmla v17.4s, v30.4s, v6.s[1]\n" - "fmla v18.4s, v30.4s, v6.s[3]\n" - "fmla v19.4s, v30.4s, v7.s[1]\n" - "fmla v20.4s, v30.4s, v10.s[1]\n" - "fmla v21.4s, v30.4s, v10.s[3]\n" - "fmla v22.4s, v30.4s, v11.s[1]\n" - "ldr q30, [%x[params], #0x80]\n" - "fmla v14.4s, v29.4s, v2.s[2]\n" - "fmla v15.4s, v29.4s, v3.s[0]\n" - "fmla v16.4s, v29.4s, v3.s[2]\n" - "fmla v17.4s, v29.4s, v6.s[2]\n" - "fmla v18.4s, v29.4s, v7.s[0]\n" - "fmla v19.4s, v29.4s, v7.s[2]\n" - "fmla v20.4s, v29.4s, v10.s[2]\n" - "fmla v21.4s, v29.4s, v11.s[0]\n" - "fmla v22.4s, v29.4s, v11.s[2]\n" - "ldr q29, [%x[params], #0x90]\n" + "fmla v14.4s, v25.4s, v0.s[0]\n" + "fmla v16.4s, v25.4s, v1.s[0]\n" + "fmla v17.4s, v25.4s, v4.s[0]\n" + "fmla v18.4s, v25.4s, v4.s[2]\n" + "fmla v19.4s, v25.4s, v5.s[0]\n" + "fmla v20.4s, v25.4s, v8.s[0]\n" + "fmla v21.4s, v25.4s, v8.s[2]\n" + "fmla v22.4s, v25.4s, v9.s[0]\n" + "ldr q25, [%x[params], #0x40]\n" + "fmla v14.4s, v24.4s, v0.s[1]\n" + "fmla v15.4s, v24.4s, v0.s[3]\n" + "fmla v16.4s, v24.4s, v1.s[1]\n" + "fmla v17.4s, v24.4s, v4.s[1]\n" + "fmla v18.4s, v24.4s, v4.s[3]\n" + "fmla v19.4s, v24.4s, v5.s[1]\n" + "fmla v20.4s, v24.4s, v8.s[1]\n" + "fmla v21.4s, v24.4s, v8.s[3]\n" + "fmla v22.4s, v24.4s, v9.s[1]\n" + "ldr q24, [%x[params], #0x50]\n" + "fmla v14.4s, v23.4s, v0.s[2]\n" + "fmla v15.4s, v23.4s, v1.s[0]\n" + "fmla v16.4s, v23.4s, v1.s[2]\n" + "fmla v17.4s, v23.4s, v4.s[2]\n" + "fmla v18.4s, v23.4s, v5.s[0]\n" + "fmla v19.4s, v23.4s, v5.s[2]\n" + "fmla v20.4s, v23.4s, v8.s[2]\n" + "fmla v21.4s, v23.4s, v9.s[0]\n" + "fmla v22.4s, v23.4s, v9.s[2]\n" + "ldr q23, [%x[params], #0x60]\n" + "fmla v14.4s, v25.4s, v2.s[0]\n" + "fmla v15.4s, v25.4s, v2.s[2]\n" + "fmla v16.4s, v25.4s, v3.s[0]\n" + "fmla v17.4s, v25.4s, v6.s[0]\n" + "fmla v18.4s, v25.4s, v6.s[2]\n" + "fmla v19.4s, v25.4s, v7.s[0]\n" + "fmla v20.4s, v25.4s, v10.s[0]\n" + "fmla v21.4s, v25.4s, v10.s[2]\n" + "fmla v22.4s, v25.4s, v11.s[0]\n" + "ldr q25, [%x[params], #0x70]\n" + "fmla v14.4s, v24.4s, v2.s[1]\n" + "fmla v15.4s, v24.4s, v2.s[3]\n" + "fmla v16.4s, v24.4s, v3.s[1]\n" + "fmla v17.4s, v24.4s, v6.s[1]\n" + "fmla v18.4s, v24.4s, v6.s[3]\n" + "fmla v19.4s, v24.4s, v7.s[1]\n" + "fmla v20.4s, v24.4s, v10.s[1]\n" + "fmla v21.4s, v24.4s, v10.s[3]\n" + "fmla v22.4s, v24.4s, v11.s[1]\n" + "ldr q24, [%x[params], #0x80]\n" + "fmla v14.4s, v23.4s, v2.s[2]\n" + "fmla v15.4s, v23.4s, v3.s[0]\n" + "fmla v16.4s, v23.4s, v3.s[2]\n" + "fmla v17.4s, v23.4s, v6.s[2]\n" + "fmla v18.4s, v23.4s, v7.s[0]\n" + "fmla v19.4s, v23.4s, v7.s[2]\n" + "fmla v20.4s, v23.4s, v10.s[2]\n" + "fmla v21.4s, v23.4s, v11.s[0]\n" + "fmla v22.4s, v23.4s, v11.s[2]\n" + "ldr q23, [%x[params], #0x90]\n" "add %x[params], %x[params], #0xa0\n" - "fmla v14.4s, v31.4s, v4.s[0]\n" - "fmla v15.4s, v31.4s, v4.s[2]\n" - "fmla v16.4s, v31.4s, v5.s[0]\n" - "fmla v17.4s, v31.4s, v8.s[0]\n" - "fmla v18.4s, v31.4s, v8.s[2]\n" - "fmla v19.4s, v31.4s, v9.s[0]\n" - "fmla v20.4s, v31.4s, v12.s[0]\n" - "fmla v21.4s, v31.4s, v12.s[2]\n" - "fmla v22.4s, v31.4s, v13.s[0]\n" - "fmla v14.4s, v30.4s, v4.s[1]\n" - "fmla v15.4s, v30.4s, v4.s[3]\n" - "fmla v16.4s, v30.4s, v5.s[1]\n" - "fmla v17.4s, v30.4s, v8.s[1]\n" - "fmla v18.4s, v30.4s, v8.s[3]\n" - "fmla v19.4s, v30.4s, v9.s[1]\n" - "fmla v20.4s, v30.4s, v12.s[1]\n" - "fmla v21.4s, v30.4s, v12.s[3]\n" - "fmla v22.4s, v30.4s, v13.s[1]\n" - "fmla v14.4s, v29.4s, v4.s[2]\n" - "fmla v15.4s, v29.4s, v5.s[0]\n" - "fmin v14.4s, v14.4s, v23.4s\n" - "fmla v16.4s, v29.4s, v5.s[2]\n" - "fmla v17.4s, v29.4s, v8.s[2]\n" - "fmin v15.4s, v15.4s, v23.4s\n" - "fmla v18.4s, v29.4s, v9.s[0]\n" - "fmla v19.4s, v29.4s, v9.s[2]\n" - "fmin v16.4s, v16.4s, v23.4s\n" - "fmla v20.4s, v29.4s, v12.s[2]\n" - "fmla v21.4s, v29.4s, v13.s[0]\n" - "fmin v17.4s, v17.4s, v23.4s\n" - "fmla v22.4s, v29.4s, v13.s[2]\n" - "fmin v18.4s, v18.4s, v23.4s\n" - "fmin v19.4s, v19.4s, v23.4s\n" - "fmin v20.4s, v20.4s, v23.4s\n" - "fmin v21.4s, v21.4s, v23.4s\n" - "fmin v22.4s, v22.4s, v23.4s\n" - "fmax v14.4s, v14.4s, v24.4s\n" - "fmax v15.4s, v15.4s, v24.4s\n" - "fmax v16.4s, v16.4s, v24.4s\n" - "fmax v17.4s, v17.4s, v24.4s\n" - "fmax v18.4s, v18.4s, v24.4s\n" - "fmax v19.4s, v19.4s, v24.4s\n" - "fmax v20.4s, v20.4s, v24.4s\n" - "fmax v21.4s, v21.4s, v24.4s\n" - "fmax v22.4s, v22.4s, v24.4s\n" + "fmla v14.4s, v25.4s, v4.s[0]\n" + "fmla v15.4s, v25.4s, v4.s[2]\n" + "fmla v16.4s, v25.4s, v5.s[0]\n" + "fmla v17.4s, v25.4s, v8.s[0]\n" + "fmla v18.4s, v25.4s, v8.s[2]\n" + "fmla v19.4s, v25.4s, v9.s[0]\n" + "fmla v20.4s, v25.4s, v12.s[0]\n" + "fmla v21.4s, v25.4s, v12.s[2]\n" + "fmla v22.4s, v25.4s, v13.s[0]\n" + "fmla v14.4s, v24.4s, v4.s[1]\n" + "fmla v15.4s, v24.4s, v4.s[3]\n" + "fmla v16.4s, v24.4s, v5.s[1]\n" + "fmla v17.4s, v24.4s, v8.s[1]\n" + "fmla v18.4s, v24.4s, v8.s[3]\n" + "fmla v19.4s, v24.4s, v9.s[1]\n" + "fmla v20.4s, v24.4s, v12.s[1]\n" + "fmla v21.4s, v24.4s, v12.s[3]\n" + "fmla v22.4s, v24.4s, v13.s[1]\n" + "fmla v14.4s, v23.4s, v4.s[2]\n" + "fmla v15.4s, v23.4s, v5.s[0]\n" + "fmin v14.4s, v14.4s, v26.4s\n" + "fmla v16.4s, v23.4s, v5.s[2]\n" + "fmla v17.4s, v23.4s, v8.s[2]\n" + "fmin v15.4s, v15.4s, v26.4s\n" + "fmla v18.4s, v23.4s, v9.s[0]\n" + "fmla v19.4s, v23.4s, v9.s[2]\n" + "fmin v16.4s, v16.4s, v26.4s\n" + "fmla v20.4s, v23.4s, v12.s[2]\n" + "fmla v21.4s, v23.4s, v13.s[0]\n" + "fmin v17.4s, v17.4s, v26.4s\n" + "fmla v22.4s, v23.4s, v13.s[2]\n" + "fmin v18.4s, v18.4s, v26.4s\n" + "fmin v19.4s, v19.4s, v26.4s\n" + "fmin v20.4s, v20.4s, v26.4s\n" + "fmin v21.4s, v21.4s, v26.4s\n" + "fmin v22.4s, v22.4s, v26.4s\n" + "fmax v14.4s, v14.4s, v27.4s\n" + "fmax v15.4s, v15.4s, v27.4s\n" + "fmax v16.4s, v16.4s, v27.4s\n" + "fmax v17.4s, v17.4s, v27.4s\n" + "fmax v18.4s, v18.4s, v27.4s\n" + "fmax v19.4s, v19.4s, v27.4s\n" + "fmax v20.4s, v20.4s, v27.4s\n" + "fmax v21.4s, v21.4s, v27.4s\n" + "fmax v22.4s, v22.4s, v27.4s\n" "tbz %x[channel_multiplier], #1, 4f\n" "add x20, x13, x14\n" "add x22, x12, x14\n" @@ -519,15 +520,14 @@ void a64_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst_imp "st1 { v21.s }[0], [x21]\n" "st1 { v22.s }[0], [x20]\n" "5:" // Output channel oddments: Store: Bit 1: End - "6:" // End - : [params] "+&r" (params) : [channel_multiplier] "r" (n_output_channels), [clamps] "r" (minmax_vals), [inptrs] "r" (inptrs), [outptrs] "r" (outptrs) - : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "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", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" ); } } // namespace depthwise } // namespace arm_conv -#endif // defined(__aarch64__) + +#endif // defined(__aarch64__) -- cgit v1.2.1