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 | 62 +++++++++++----------- 1 file changed, 31 insertions(+), 31 deletions(-) (limited to 'src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp') diff --git a/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp index 647103d3a4..5df848d1dd 100644 --- a/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp @@ -196,38 +196,38 @@ void a64_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst_impl( "add x5, x5, #0x10\n" "cbz x3, 4f\n" "3:" // Oddments - "ldr h6, [x11, x4]\n" - "ldr h5, [x10, x4]\n" - "fadd v17.8h, v6.8h, v5.8h\n" + "ldr h17, [x11, x4]\n" + "ldr h16, [x10, x4]\n" + "fadd v18.8h, v17.8h, v16.8h\n" "subs x3, x3, #0x1\n" - "ldr h4, [x27, x4]\n" - "ldr h3, [x26, x4]\n" - "fadd v16.8h, v4.8h, v3.8h\n" - "fadd v19.8h, v17.8h, v16.8h\n" - "ldr h2, [x15, x4]\n" - "ldr h1, [x14, x4]\n" - "fadd v18.8h, v2.8h, v1.8h\n" - "fadd v21.8h, v18.8h, v19.8h\n" - "ldr h0, [x12, x4]\n" - "ldr h31, [x28, x4]\n" - "fadd v17.8h, v0.8h, v31.8h\n" - "ldr h30, [x9, x4]\n" - "ldr h29, [x25, x4]\n" - "fadd v22.8h, v30.8h, v29.8h\n" - "ldr h28, [x23, x4]\n" - "ldr h27, [x22, x4]\n" - "fadd v16.8h, v28.8h, v27.8h\n" - "fadd v20.8h, v16.8h, v19.8h\n" - "ldr h26, [x16, x4]\n" - "ldr h25, [x13, x4]\n" - "fadd v19.8h, v26.8h, v17.8h\n" - "fadd v18.8h, v25.8h, v22.8h\n" - "ldr h24, [x24, x4]\n" - "ldr h23, [x21, x4]\n" - "fadd v17.8h, v24.8h, v17.8h\n" - "fadd v16.8h, v23.8h, v22.8h\n" - "fadd v19.8h, v21.8h, v19.8h\n" - "fadd v18.8h, v21.8h, v18.8h\n" + "ldr h17, [x27, x4]\n" + "ldr h16, [x26, x4]\n" + "fadd v16.8h, v17.8h, v16.8h\n" + "fadd v18.8h, v18.8h, v16.8h\n" + "ldr h17, [x15, x4]\n" + "ldr h16, [x14, x4]\n" + "fadd v16.8h, v17.8h, v16.8h\n" + "fadd v23.8h, v16.8h, v18.8h\n" + "ldr h17, [x12, x4]\n" + "ldr h16, [x28, x4]\n" + "fadd v22.8h, v17.8h, v16.8h\n" + "ldr h17, [x9, x4]\n" + "ldr h16, [x25, x4]\n" + "fadd v21.8h, v17.8h, v16.8h\n" + "ldr h17, [x23, x4]\n" + "ldr h16, [x22, x4]\n" + "fadd v16.8h, v17.8h, v16.8h\n" + "fadd v20.8h, v16.8h, v18.8h\n" + "ldr h17, [x16, x4]\n" + "ldr h16, [x13, x4]\n" + "fadd v19.8h, v17.8h, v22.8h\n" + "fadd v18.8h, v16.8h, v21.8h\n" + "ldr h17, [x24, x4]\n" + "ldr h16, [x21, x4]\n" + "fadd v17.8h, v17.8h, v22.8h\n" + "fadd v16.8h, v16.8h, v21.8h\n" + "fadd v19.8h, v23.8h, v19.8h\n" + "fadd v18.8h, v23.8h, v18.8h\n" "add x4, x4, #0x2\n" "fadd v17.8h, v17.8h, v20.8h\n" "fadd v16.8h, v16.8h, v20.8h\n" -- cgit v1.2.1