cannam@154: /*********************************************************************** cannam@154: Copyright (C) 2014 Vidyo cannam@154: Redistribution and use in source and binary forms, with or without cannam@154: modification, are permitted provided that the following conditions cannam@154: are met: cannam@154: - Redistributions of source code must retain the above copyright notice, cannam@154: this list of conditions and the following disclaimer. cannam@154: - Redistributions in binary form must reproduce the above copyright cannam@154: notice, this list of conditions and the following disclaimer in the cannam@154: documentation and/or other materials provided with the distribution. cannam@154: - Neither the name of Internet Society, IETF or IETF Trust, nor the cannam@154: names of specific contributors, may be used to endorse or promote cannam@154: products derived from this software without specific prior written cannam@154: permission. cannam@154: THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" cannam@154: AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE cannam@154: IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE cannam@154: ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE cannam@154: LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR cannam@154: CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF cannam@154: SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS cannam@154: INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN cannam@154: CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) cannam@154: ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE cannam@154: POSSIBILITY OF SUCH DAMAGE. cannam@154: ***********************************************************************/ cannam@154: #ifdef HAVE_CONFIG_H cannam@154: #include "config.h" cannam@154: #endif cannam@154: cannam@154: #include cannam@154: #include "main.h" cannam@154: #include "stack_alloc.h" cannam@154: #include "NSQ.h" cannam@154: #include "celt/cpu_support.h" cannam@154: #include "celt/arm/armcpu.h" cannam@154: cannam@154: opus_int32 silk_noise_shape_quantizer_short_prediction_neon(const opus_int32 *buf32, const opus_int32 *coef32, opus_int order) cannam@154: { cannam@154: int32x4_t coef0 = vld1q_s32(coef32); cannam@154: int32x4_t coef1 = vld1q_s32(coef32 + 4); cannam@154: int32x4_t coef2 = vld1q_s32(coef32 + 8); cannam@154: int32x4_t coef3 = vld1q_s32(coef32 + 12); cannam@154: cannam@154: int32x4_t a0 = vld1q_s32(buf32 - 15); cannam@154: int32x4_t a1 = vld1q_s32(buf32 - 11); cannam@154: int32x4_t a2 = vld1q_s32(buf32 - 7); cannam@154: int32x4_t a3 = vld1q_s32(buf32 - 3); cannam@154: cannam@154: int32x4_t b0 = vqdmulhq_s32(coef0, a0); cannam@154: int32x4_t b1 = vqdmulhq_s32(coef1, a1); cannam@154: int32x4_t b2 = vqdmulhq_s32(coef2, a2); cannam@154: int32x4_t b3 = vqdmulhq_s32(coef3, a3); cannam@154: cannam@154: int32x4_t c0 = vaddq_s32(b0, b1); cannam@154: int32x4_t c1 = vaddq_s32(b2, b3); cannam@154: cannam@154: int32x4_t d = vaddq_s32(c0, c1); cannam@154: cannam@154: int64x2_t e = vpaddlq_s32(d); cannam@154: cannam@154: int64x1_t f = vadd_s64(vget_low_s64(e), vget_high_s64(e)); cannam@154: cannam@154: opus_int32 out = vget_lane_s32(vreinterpret_s32_s64(f), 0); cannam@154: cannam@154: out += silk_RSHIFT( order, 1 ); cannam@154: cannam@154: return out; cannam@154: } cannam@154: cannam@154: cannam@154: opus_int32 silk_NSQ_noise_shape_feedback_loop_neon(const opus_int32 *data0, opus_int32 *data1, const opus_int16 *coef, opus_int order) cannam@154: { cannam@154: opus_int32 out; cannam@154: if (order == 8) cannam@154: { cannam@154: int32x4_t a00 = vdupq_n_s32(data0[0]); cannam@154: int32x4_t a01 = vld1q_s32(data1); /* data1[0] ... [3] */ cannam@154: cannam@154: int32x4_t a0 = vextq_s32 (a00, a01, 3); /* data0[0] data1[0] ...[2] */ cannam@154: int32x4_t a1 = vld1q_s32(data1 + 3); /* data1[3] ... [6] */ cannam@154: cannam@154: /*TODO: Convert these once in advance instead of once per sample, like cannam@154: silk_noise_shape_quantizer_short_prediction_neon() does.*/ cannam@154: int16x8_t coef16 = vld1q_s16(coef); cannam@154: int32x4_t coef0 = vmovl_s16(vget_low_s16(coef16)); cannam@154: int32x4_t coef1 = vmovl_s16(vget_high_s16(coef16)); cannam@154: cannam@154: /*This is not bit-exact with the C version, since we do not drop the cannam@154: lower 16 bits of each multiply, but wait until the end to truncate cannam@154: precision. This is an encoder-specific calculation (and unlike cannam@154: silk_noise_shape_quantizer_short_prediction_neon(), is not meant to cannam@154: simulate what the decoder will do). We still could use vqdmulhq_s32() cannam@154: like silk_noise_shape_quantizer_short_prediction_neon() and save cannam@154: half the multiplies, but the speed difference is not large, since we cannam@154: then need two extra adds.*/ cannam@154: int64x2_t b0 = vmull_s32(vget_low_s32(a0), vget_low_s32(coef0)); cannam@154: int64x2_t b1 = vmlal_s32(b0, vget_high_s32(a0), vget_high_s32(coef0)); cannam@154: int64x2_t b2 = vmlal_s32(b1, vget_low_s32(a1), vget_low_s32(coef1)); cannam@154: int64x2_t b3 = vmlal_s32(b2, vget_high_s32(a1), vget_high_s32(coef1)); cannam@154: cannam@154: int64x1_t c = vadd_s64(vget_low_s64(b3), vget_high_s64(b3)); cannam@154: int64x1_t cS = vrshr_n_s64(c, 15); cannam@154: int32x2_t d = vreinterpret_s32_s64(cS); cannam@154: cannam@154: out = vget_lane_s32(d, 0); cannam@154: vst1q_s32(data1, a0); cannam@154: vst1q_s32(data1 + 4, a1); cannam@154: return out; cannam@154: } cannam@154: return silk_NSQ_noise_shape_feedback_loop_c(data0, data1, coef, order); cannam@154: }