Chris@69: /***********************************************************************
Chris@69: Copyright (c) 2017 Google Inc.
Chris@69: Redistribution and use in source and binary forms, with or without
Chris@69: modification, are permitted provided that the following conditions
Chris@69: are met:
Chris@69: - Redistributions of source code must retain the above copyright notice,
Chris@69: this list of conditions and the following disclaimer.
Chris@69: - Redistributions in binary form must reproduce the above copyright
Chris@69: notice, this list of conditions and the following disclaimer in the
Chris@69: documentation and/or other materials provided with the distribution.
Chris@69: - Neither the name of Internet Society, IETF or IETF Trust, nor the
Chris@69: names of specific contributors, may be used to endorse or promote
Chris@69: products derived from this software without specific prior written
Chris@69: permission.
Chris@69: THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
Chris@69: AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
Chris@69: IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
Chris@69: ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
Chris@69: LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
Chris@69: CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
Chris@69: SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
Chris@69: INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
Chris@69: CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
Chris@69: ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
Chris@69: POSSIBILITY OF SUCH DAMAGE.
Chris@69: ***********************************************************************/
Chris@69: 
Chris@69: #ifdef HAVE_CONFIG_H
Chris@69: #include "config.h"
Chris@69: #endif
Chris@69: 
Chris@69: #include <arm_neon.h>
Chris@69: #include "pitch.h"
Chris@69: 
Chris@69: #ifdef FIXED_POINT
Chris@69: 
Chris@69: opus_val32 celt_inner_prod_neon(const opus_val16 *x, const opus_val16 *y, int N)
Chris@69: {
Chris@69:     int i;
Chris@69:     opus_val32 xy;
Chris@69:     int16x8_t x_s16x8, y_s16x8;
Chris@69:     int32x4_t xy_s32x4 = vdupq_n_s32(0);
Chris@69:     int64x2_t xy_s64x2;
Chris@69:     int64x1_t xy_s64x1;
Chris@69: 
Chris@69:     for (i = 0; i < N - 7; i += 8) {
Chris@69:         x_s16x8  = vld1q_s16(&x[i]);
Chris@69:         y_s16x8  = vld1q_s16(&y[i]);
Chris@69:         xy_s32x4 = vmlal_s16(xy_s32x4, vget_low_s16 (x_s16x8), vget_low_s16 (y_s16x8));
Chris@69:         xy_s32x4 = vmlal_s16(xy_s32x4, vget_high_s16(x_s16x8), vget_high_s16(y_s16x8));
Chris@69:     }
Chris@69: 
Chris@69:     if (N - i >= 4) {
Chris@69:         const int16x4_t x_s16x4 = vld1_s16(&x[i]);
Chris@69:         const int16x4_t y_s16x4 = vld1_s16(&y[i]);
Chris@69:         xy_s32x4 = vmlal_s16(xy_s32x4, x_s16x4, y_s16x4);
Chris@69:         i += 4;
Chris@69:     }
Chris@69: 
Chris@69:     xy_s64x2 = vpaddlq_s32(xy_s32x4);
Chris@69:     xy_s64x1 = vadd_s64(vget_low_s64(xy_s64x2), vget_high_s64(xy_s64x2));
Chris@69:     xy       = vget_lane_s32(vreinterpret_s32_s64(xy_s64x1), 0);
Chris@69: 
Chris@69:     for (; i < N; i++) {
Chris@69:         xy = MAC16_16(xy, x[i], y[i]);
Chris@69:     }
Chris@69: 
Chris@69: #ifdef OPUS_CHECK_ASM
Chris@69:     celt_assert(celt_inner_prod_c(x, y, N) == xy);
Chris@69: #endif
Chris@69: 
Chris@69:     return xy;
Chris@69: }
Chris@69: 
Chris@69: void dual_inner_prod_neon(const opus_val16 *x, const opus_val16 *y01, const opus_val16 *y02,
Chris@69:         int N, opus_val32 *xy1, opus_val32 *xy2)
Chris@69: {
Chris@69:     int i;
Chris@69:     opus_val32 xy01, xy02;
Chris@69:     int16x8_t x_s16x8, y01_s16x8, y02_s16x8;
Chris@69:     int32x4_t xy01_s32x4 = vdupq_n_s32(0);
Chris@69:     int32x4_t xy02_s32x4 = vdupq_n_s32(0);
Chris@69:     int64x2_t xy01_s64x2, xy02_s64x2;
Chris@69:     int64x1_t xy01_s64x1, xy02_s64x1;
Chris@69: 
Chris@69:     for (i = 0; i < N - 7; i += 8) {
Chris@69:         x_s16x8    = vld1q_s16(&x[i]);
Chris@69:         y01_s16x8  = vld1q_s16(&y01[i]);
Chris@69:         y02_s16x8  = vld1q_s16(&y02[i]);
Chris@69:         xy01_s32x4 = vmlal_s16(xy01_s32x4, vget_low_s16 (x_s16x8), vget_low_s16 (y01_s16x8));
Chris@69:         xy02_s32x4 = vmlal_s16(xy02_s32x4, vget_low_s16 (x_s16x8), vget_low_s16 (y02_s16x8));
Chris@69:         xy01_s32x4 = vmlal_s16(xy01_s32x4, vget_high_s16(x_s16x8), vget_high_s16(y01_s16x8));
Chris@69:         xy02_s32x4 = vmlal_s16(xy02_s32x4, vget_high_s16(x_s16x8), vget_high_s16(y02_s16x8));
Chris@69:     }
Chris@69: 
Chris@69:     if (N - i >= 4) {
Chris@69:         const int16x4_t x_s16x4   = vld1_s16(&x[i]);
Chris@69:         const int16x4_t y01_s16x4 = vld1_s16(&y01[i]);
Chris@69:         const int16x4_t y02_s16x4 = vld1_s16(&y02[i]);
Chris@69:         xy01_s32x4 = vmlal_s16(xy01_s32x4, x_s16x4, y01_s16x4);
Chris@69:         xy02_s32x4 = vmlal_s16(xy02_s32x4, x_s16x4, y02_s16x4);
Chris@69:         i += 4;
Chris@69:     }
Chris@69: 
Chris@69:     xy01_s64x2 = vpaddlq_s32(xy01_s32x4);
Chris@69:     xy02_s64x2 = vpaddlq_s32(xy02_s32x4);
Chris@69:     xy01_s64x1 = vadd_s64(vget_low_s64(xy01_s64x2), vget_high_s64(xy01_s64x2));
Chris@69:     xy02_s64x1 = vadd_s64(vget_low_s64(xy02_s64x2), vget_high_s64(xy02_s64x2));
Chris@69:     xy01       = vget_lane_s32(vreinterpret_s32_s64(xy01_s64x1), 0);
Chris@69:     xy02       = vget_lane_s32(vreinterpret_s32_s64(xy02_s64x1), 0);
Chris@69: 
Chris@69:     for (; i < N; i++) {
Chris@69:         xy01 = MAC16_16(xy01, x[i], y01[i]);
Chris@69:         xy02 = MAC16_16(xy02, x[i], y02[i]);
Chris@69:     }
Chris@69:     *xy1 = xy01;
Chris@69:     *xy2 = xy02;
Chris@69: 
Chris@69: #ifdef OPUS_CHECK_ASM
Chris@69:     {
Chris@69:         opus_val32 xy1_c, xy2_c;
Chris@69:         dual_inner_prod_c(x, y01, y02, N, &xy1_c, &xy2_c);
Chris@69:         celt_assert(xy1_c == *xy1);
Chris@69:         celt_assert(xy2_c == *xy2);
Chris@69:     }
Chris@69: #endif
Chris@69: }
Chris@69: 
Chris@69: #else /* !FIXED_POINT */
Chris@69: 
Chris@69: /* ========================================================================== */
Chris@69: 
Chris@69: #ifdef OPUS_CHECK_ASM
Chris@69: 
Chris@69: /* This part of code simulates floating-point NEON operations. */
Chris@69: 
Chris@69: /* celt_inner_prod_neon_float_c_simulation() simulates the floating-point   */
Chris@69: /* operations of celt_inner_prod_neon(), and both functions should have bit */
Chris@69: /* exact output.                                                            */
Chris@69: static opus_val32 celt_inner_prod_neon_float_c_simulation(const opus_val16 *x, const opus_val16 *y, int N)
Chris@69: {
Chris@69:    int i;
Chris@69:    opus_val32 xy, xy0 = 0, xy1 = 0, xy2 = 0, xy3 = 0;
Chris@69:    for (i = 0; i < N - 3; i += 4) {
Chris@69:       xy0 = MAC16_16(xy0, x[i + 0], y[i + 0]);
Chris@69:       xy1 = MAC16_16(xy1, x[i + 1], y[i + 1]);
Chris@69:       xy2 = MAC16_16(xy2, x[i + 2], y[i + 2]);
Chris@69:       xy3 = MAC16_16(xy3, x[i + 3], y[i + 3]);
Chris@69:    }
Chris@69:    xy0 += xy2;
Chris@69:    xy1 += xy3;
Chris@69:    xy = xy0 + xy1;
Chris@69:    for (; i < N; i++) {
Chris@69:       xy = MAC16_16(xy, x[i], y[i]);
Chris@69:    }
Chris@69:    return xy;
Chris@69: }
Chris@69: 
Chris@69: /* dual_inner_prod_neon_float_c_simulation() simulates the floating-point   */
Chris@69: /* operations of dual_inner_prod_neon(), and both functions should have bit */
Chris@69: /* exact output.                                                            */
Chris@69: static void dual_inner_prod_neon_float_c_simulation(const opus_val16 *x, const opus_val16 *y01, const opus_val16 *y02,
Chris@69:       int N, opus_val32 *xy1, opus_val32 *xy2)
Chris@69: {
Chris@69:    int i;
Chris@69:    opus_val32 xy01, xy02, xy01_0 = 0, xy01_1 = 0, xy01_2 = 0, xy01_3 = 0, xy02_0 = 0, xy02_1 = 0, xy02_2 = 0, xy02_3 = 0;
Chris@69:    for (i = 0; i < N - 3; i += 4) {
Chris@69:       xy01_0 = MAC16_16(xy01_0, x[i + 0], y01[i + 0]);
Chris@69:       xy01_1 = MAC16_16(xy01_1, x[i + 1], y01[i + 1]);
Chris@69:       xy01_2 = MAC16_16(xy01_2, x[i + 2], y01[i + 2]);
Chris@69:       xy01_3 = MAC16_16(xy01_3, x[i + 3], y01[i + 3]);
Chris@69:       xy02_0 = MAC16_16(xy02_0, x[i + 0], y02[i + 0]);
Chris@69:       xy02_1 = MAC16_16(xy02_1, x[i + 1], y02[i + 1]);
Chris@69:       xy02_2 = MAC16_16(xy02_2, x[i + 2], y02[i + 2]);
Chris@69:       xy02_3 = MAC16_16(xy02_3, x[i + 3], y02[i + 3]);
Chris@69:    }
Chris@69:    xy01_0 += xy01_2;
Chris@69:    xy02_0 += xy02_2;
Chris@69:    xy01_1 += xy01_3;
Chris@69:    xy02_1 += xy02_3;
Chris@69:    xy01 = xy01_0 + xy01_1;
Chris@69:    xy02 = xy02_0 + xy02_1;
Chris@69:    for (; i < N; i++) {
Chris@69:       xy01 = MAC16_16(xy01, x[i], y01[i]);
Chris@69:       xy02 = MAC16_16(xy02, x[i], y02[i]);
Chris@69:    }
Chris@69:    *xy1 = xy01;
Chris@69:    *xy2 = xy02;
Chris@69: }
Chris@69: 
Chris@69: #endif /* OPUS_CHECK_ASM */
Chris@69: 
Chris@69: /* ========================================================================== */
Chris@69: 
Chris@69: opus_val32 celt_inner_prod_neon(const opus_val16 *x, const opus_val16 *y, int N)
Chris@69: {
Chris@69:     int i;
Chris@69:     opus_val32 xy;
Chris@69:     float32x4_t xy_f32x4 = vdupq_n_f32(0);
Chris@69:     float32x2_t xy_f32x2;
Chris@69: 
Chris@69:     for (i = 0; i < N - 7; i += 8) {
Chris@69:         float32x4_t x_f32x4, y_f32x4;
Chris@69:         x_f32x4  = vld1q_f32(&x[i]);
Chris@69:         y_f32x4  = vld1q_f32(&y[i]);
Chris@69:         xy_f32x4 = vmlaq_f32(xy_f32x4, x_f32x4, y_f32x4);
Chris@69:         x_f32x4  = vld1q_f32(&x[i + 4]);
Chris@69:         y_f32x4  = vld1q_f32(&y[i + 4]);
Chris@69:         xy_f32x4 = vmlaq_f32(xy_f32x4, x_f32x4, y_f32x4);
Chris@69:     }
Chris@69: 
Chris@69:     if (N - i >= 4) {
Chris@69:         const float32x4_t x_f32x4 = vld1q_f32(&x[i]);
Chris@69:         const float32x4_t y_f32x4 = vld1q_f32(&y[i]);
Chris@69:         xy_f32x4 = vmlaq_f32(xy_f32x4, x_f32x4, y_f32x4);
Chris@69:         i += 4;
Chris@69:     }
Chris@69: 
Chris@69:     xy_f32x2 = vadd_f32(vget_low_f32(xy_f32x4), vget_high_f32(xy_f32x4));
Chris@69:     xy_f32x2 = vpadd_f32(xy_f32x2, xy_f32x2);
Chris@69:     xy       = vget_lane_f32(xy_f32x2, 0);
Chris@69: 
Chris@69:     for (; i < N; i++) {
Chris@69:         xy = MAC16_16(xy, x[i], y[i]);
Chris@69:     }
Chris@69: 
Chris@69: #ifdef OPUS_CHECK_ASM
Chris@69:     celt_assert(ABS32(celt_inner_prod_neon_float_c_simulation(x, y, N) - xy) <= VERY_SMALL);
Chris@69: #endif
Chris@69: 
Chris@69:     return xy;
Chris@69: }
Chris@69: 
Chris@69: void dual_inner_prod_neon(const opus_val16 *x, const opus_val16 *y01, const opus_val16 *y02,
Chris@69:         int N, opus_val32 *xy1, opus_val32 *xy2)
Chris@69: {
Chris@69:     int i;
Chris@69:     opus_val32 xy01, xy02;
Chris@69:     float32x4_t xy01_f32x4 = vdupq_n_f32(0);
Chris@69:     float32x4_t xy02_f32x4 = vdupq_n_f32(0);
Chris@69:     float32x2_t xy01_f32x2, xy02_f32x2;
Chris@69: 
Chris@69:     for (i = 0; i < N - 7; i += 8) {
Chris@69:         float32x4_t x_f32x4, y01_f32x4, y02_f32x4;
Chris@69:         x_f32x4    = vld1q_f32(&x[i]);
Chris@69:         y01_f32x4  = vld1q_f32(&y01[i]);
Chris@69:         y02_f32x4  = vld1q_f32(&y02[i]);
Chris@69:         xy01_f32x4 = vmlaq_f32(xy01_f32x4, x_f32x4, y01_f32x4);
Chris@69:         xy02_f32x4 = vmlaq_f32(xy02_f32x4, x_f32x4, y02_f32x4);
Chris@69:         x_f32x4    = vld1q_f32(&x[i + 4]);
Chris@69:         y01_f32x4  = vld1q_f32(&y01[i + 4]);
Chris@69:         y02_f32x4  = vld1q_f32(&y02[i + 4]);
Chris@69:         xy01_f32x4 = vmlaq_f32(xy01_f32x4, x_f32x4, y01_f32x4);
Chris@69:         xy02_f32x4 = vmlaq_f32(xy02_f32x4, x_f32x4, y02_f32x4);
Chris@69:     }
Chris@69: 
Chris@69:     if (N - i >= 4) {
Chris@69:         const float32x4_t x_f32x4   = vld1q_f32(&x[i]);
Chris@69:         const float32x4_t y01_f32x4 = vld1q_f32(&y01[i]);
Chris@69:         const float32x4_t y02_f32x4 = vld1q_f32(&y02[i]);
Chris@69:         xy01_f32x4 = vmlaq_f32(xy01_f32x4, x_f32x4, y01_f32x4);
Chris@69:         xy02_f32x4 = vmlaq_f32(xy02_f32x4, x_f32x4, y02_f32x4);
Chris@69:         i += 4;
Chris@69:     }
Chris@69: 
Chris@69:     xy01_f32x2 = vadd_f32(vget_low_f32(xy01_f32x4), vget_high_f32(xy01_f32x4));
Chris@69:     xy02_f32x2 = vadd_f32(vget_low_f32(xy02_f32x4), vget_high_f32(xy02_f32x4));
Chris@69:     xy01_f32x2 = vpadd_f32(xy01_f32x2, xy01_f32x2);
Chris@69:     xy02_f32x2 = vpadd_f32(xy02_f32x2, xy02_f32x2);
Chris@69:     xy01       = vget_lane_f32(xy01_f32x2, 0);
Chris@69:     xy02       = vget_lane_f32(xy02_f32x2, 0);
Chris@69: 
Chris@69:     for (; i < N; i++) {
Chris@69:         xy01 = MAC16_16(xy01, x[i], y01[i]);
Chris@69:         xy02 = MAC16_16(xy02, x[i], y02[i]);
Chris@69:     }
Chris@69:     *xy1 = xy01;
Chris@69:     *xy2 = xy02;
Chris@69: 
Chris@69: #ifdef OPUS_CHECK_ASM
Chris@69:     {
Chris@69:         opus_val32 xy1_c, xy2_c;
Chris@69:         dual_inner_prod_neon_float_c_simulation(x, y01, y02, N, &xy1_c, &xy2_c);
Chris@69:         celt_assert(ABS32(xy1_c - *xy1) <= VERY_SMALL);
Chris@69:         celt_assert(ABS32(xy2_c - *xy2) <= VERY_SMALL);
Chris@69:     }
Chris@69: #endif
Chris@69: }
Chris@69: 
Chris@69: #endif /* FIXED_POINT */