diff options
| author | Alan Kelly <me@alankelly.dev> | 2019-04-10 21:37:05 +0200 |
|---|---|---|
| committer | Georgios Pinitas <georgios.pinitas@arm.com> | 2019-05-17 16:25:00 +0000 |
| commit | efcefeadd0b3ccb78dea1912cfa792cd87088c10 (patch) | |
| tree | 64b233772928cefc71aff8e840926f6974650555 | |
| parent | a400334f60c28a464ba44170350ec475af73a94a (diff) | |
| download | ComputeLibrary-efcefeadd0b3ccb78dea1912cfa792cd87088c10.tar.gz | |
Optimizes NEBox3x3FP16Kernel by removing unnecessary int to float conversions
Signed-off-by: Alan Kelly <me@alankelly.dev>
Change-Id: Icf54d5f37455fc5ba0fbc5f15d6e481a7a751d74
Reviewed-on: https://review.mlplatform.org/c/1044
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
| -rw-r--r-- | src/core/NEON/kernels/NEBox3x3Kernel.cpp | 41 |
1 files changed, 21 insertions, 20 deletions
diff --git a/src/core/NEON/kernels/NEBox3x3Kernel.cpp b/src/core/NEON/kernels/NEBox3x3Kernel.cpp index a401aa7e79..126b62b21e 100644 --- a/src/core/NEON/kernels/NEBox3x3Kernel.cpp +++ b/src/core/NEON/kernels/NEBox3x3Kernel.cpp @@ -55,52 +55,53 @@ void NEBox3x3FP16Kernel::run(const Window &window, const ThreadInfo &info) const uint8x16_t mid_data = vld1q_u8(input_mid_ptr + input.offset()); const uint8x16_t bot_data = vld1q_u8(input_bot_ptr + input.offset()); - const float16x8x2_t top_f16 = + const uint16x8x2_t top_f16 = { { - vcvtq_f16_u16(vmovl_u8(vget_low_u8(top_data))), - vcvtq_f16_u16(vmovl_u8(vget_high_u8(top_data))) + vmovl_u8(vget_low_u8(top_data)), + vmovl_u8(vget_high_u8(top_data)) } }; - const float16x8x2_t mid_f16 = + const uint16x8x2_t mid_f16 = { { - vcvtq_f16_u16(vmovl_u8(vget_low_u8(mid_data))), - vcvtq_f16_u16(vmovl_u8(vget_high_u8(mid_data))) + vmovl_u8(vget_low_u8(mid_data)), + vmovl_u8(vget_high_u8(mid_data)) } }; - const float16x8x2_t bot_f16 = + const uint16x8x2_t bot_f16 = { { - vcvtq_f16_u16(vmovl_u8(vget_low_u8(bot_data))), - vcvtq_f16_u16(vmovl_u8(vget_high_u8(bot_data))) + vmovl_u8(vget_low_u8(bot_data)), + vmovl_u8(vget_high_u8(bot_data)) } }; //top left - float16x8_t out = top_f16.val[0]; + uint16x8_t out = top_f16.val[0]; //top mid - out = vaddq_f16(out, vextq_f16(top_f16.val[0], top_f16.val[1], 1)); + out = vaddq_u16(out, vextq_u16(top_f16.val[0], top_f16.val[1], 1)); //top right - out = vaddq_f16(out, vextq_f16(top_f16.val[0], top_f16.val[1], 2)); + out = vaddq_u16(out, vextq_u16(top_f16.val[0], top_f16.val[1], 2)); //mid left - out = vaddq_f16(out, mid_f16.val[0]); + out = vaddq_u16(out, mid_f16.val[0]); //mid mid - out = vaddq_f16(out, vextq_f16(mid_f16.val[0], mid_f16.val[1], 1)); + out = vaddq_u16(out, vextq_u16(mid_f16.val[0], mid_f16.val[1], 1)); //mid right - out = vaddq_f16(out, vextq_f16(mid_f16.val[0], mid_f16.val[1], 2)); + out = vaddq_u16(out, vextq_u16(mid_f16.val[0], mid_f16.val[1], 2)); //bot left - out = vaddq_f16(out, bot_f16.val[0]); + out = vaddq_u16(out, bot_f16.val[0]); //bot mid - out = vaddq_f16(out, vextq_f16(bot_f16.val[0], bot_f16.val[1], 1)); + out = vaddq_u16(out, vextq_u16(bot_f16.val[0], bot_f16.val[1], 1)); //bot right - out = vaddq_f16(out, vextq_f16(bot_f16.val[0], bot_f16.val[1], 2)); + out = vaddq_u16(out, vextq_u16(bot_f16.val[0], bot_f16.val[1], 2)); - out = vmulq_f16(out, oneovernine); + float16x8_t outfloat = vcvtq_f16_u16(out); + outfloat = vmulq_f16(outfloat, oneovernine); - vst1_u8(output.ptr(), vqmovun_s16(vcvtq_s16_f16(out))); + vst1_u8(output.ptr(), vqmovun_s16(vcvtq_s16_f16(outfloat))); }, input, output); } |