Optimize SVE implementation of av1_warp_affine In case of beta == 0 and alpha == 0 we know filter values before processing loops so add new logic using Neon USMMLA instructions, keeping the SVE implementation for the remaining cases. By permuting the input samples and the 6-tap filter we can use the Armv8.6 I8MM USMMLA matrix multiply instructions to accelerate horizontal 6-tap convolutions. The 2x8 by 8x2 matrix multiply instruction does twice the work of the USDOT dot product instructions. In case of 8-tap filter we can replace the USDOT instruction with USMMLA, to apply a 7-tap filter, and an extra multiplication. Change-Id: Ia5df8a05512525f1eb4add4da4423c67aff2ca74
diff --git a/av1/common/arm/warp_plane_sve.c b/av1/common/arm/warp_plane_sve.c index 10aee35..455e29d 100644 --- a/av1/common/arm/warp_plane_sve.c +++ b/av1/common/arm/warp_plane_sve.c
@@ -20,6 +20,24 @@ 8, 9, 10, 11, 9, 10, 11, 12, 10, 11, 12, 13, 11, 12, 13, 14 }; +DECLARE_ALIGNED(16, static const uint8_t, kMatMul6PermuteTbl[32]) = { + // clang-format off + 0, 1, 2, 3, 4, 5, 6, 7, 2, 3, 4, 5, 6, 7, 8, 9, + 4, 5, 6, 7, 8, 9, 10, 11, 6, 7, 8, 9, 10, 11, 12, 13 + // clang-format on +}; + +DECLARE_ALIGNED(16, static const uint8_t, kMatMul8PermuteTbl[32]) = { + // clang-format off + 1, 2, 3, 4, 5, 6, 7, 8, 3, 4, 5, 6, 7, 8, 9, 10, + 5, 6, 7, 8, 9, 10, 11, 12, 7, 8, 9, 10, 11, 12, 13, 14 + // clang-format on +}; + +DECLARE_ALIGNED(16, static const uint8_t, kTblIdx0_3[16]) = { + 0, -1, -1, -1, 1, -1, -1, -1, 2, -1, -1, -1, 3, -1, -1, -1, +}; + static AOM_FORCE_INLINE int16x8_t horizontal_filter_4x1_f4(const uint8x16_t in, int sx, int alpha) { // Only put the constant in every other lane to avoid double-counting when @@ -87,6 +105,47 @@ return vreinterpretq_s16_u16(res); } +static AOM_FORCE_INLINE int16x8_t horizontal_filter_4x1_f1_6tap_beta0( + const uint8x16_t in, const int8x16_t filter, const uint8x16_t perm_tbl) { + const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1)); + + // Permute samples ready for matrix multiply. + // { 0, 1, 2, 3, 4, 5, 6, 7, 2, 3, 4, 5, 6, 7, 8, 9 } + const uint8x16_t perm_samples = vqtbl1q_u8(in, perm_tbl); + + // These instructions multiply a 2x8 matrix (samples) by an 8x2 matrix + // (filter), destructively accumulating into the destination register. + int32x4_t sum = vusmmlaq_s32(add_const, perm_samples, filter); + + uint16x8_t res = + vcombine_u16(vqrshrun_n_s32(sum, ROUND0_BITS), vdup_n_u16(0)); + + return vreinterpretq_s16_u16(res); +} + +static AOM_FORCE_INLINE int16x8_t horizontal_filter_4x1_f1_8tap_beta0( + const uint8x16_t in, const int8x16_t filter, const int32x4_t f0, + const uint8x16_t perm_tbl, const uint8x16_t tbl_idx0_3) { + const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1)); + + // Permute samples ready for matrix multiply. + // { 1, 2, 3, 4, 5, 6, 7, 8, 3, 4, 5, 6, 7, 8, 9, 10 } + const uint8x16_t perm_samples = vqtbl1q_u8(in, perm_tbl); + // Get samples 0..3 to apply tap 0 after matrix multiply. + const int32x4_t samples_0_3 = + vreinterpretq_s32_u8(vqtbl1q_u8(in, tbl_idx0_3)); + + // Calculate partial 7-tap convolution. + int32x4_t sum = vusmmlaq_s32(add_const, perm_samples, filter); + // Apply tap 0 and accumulate. + sum = vmlaq_s32(sum, samples_0_3, f0); + + uint16x8_t res = + vcombine_u16(vqrshrun_n_s32(sum, ROUND0_BITS), vdup_n_u16(0)); + + return vreinterpretq_s16_u16(res); +} + static AOM_FORCE_INLINE int16x8_t horizontal_filter_4x1_f1_beta0(const uint8x16_t in, int16x8_t f_s16) { const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1)); @@ -116,6 +175,53 @@ return horizontal_filter_4x1_f1_beta0(in, f_s16); } +static AOM_FORCE_INLINE int16x8_t horizontal_filter_8x1_f1_6tap_beta0( + const uint8x16_t in, const int8x16_t filter, const uint8x16x2_t perm_tbl) { + const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1)); + + // Permute samples ready for matrix multiply. + // { 0, 1, 2, 3, 4, 5, 6, 7, 2, 3, 4, 5, 6, 7, 8, 9 } + // { 4, 5, 6, 7, 8, 9, 10, 11, 6, 7, 8, 9, 10, 11, 12, 13 } + uint8x16_t perm_samples[2] = { vqtbl1q_u8(in, perm_tbl.val[0]), + vqtbl1q_u8(in, perm_tbl.val[1]) }; + + // These instructions multiply a 2x8 matrix (samples) by an 8x2 matrix + // (filter), destructively accumulating into the destination register. + int32x4_t sum0123 = vusmmlaq_s32(add_const, perm_samples[0], filter); + int32x4_t sum4567 = vusmmlaq_s32(add_const, perm_samples[1], filter); + + uint16x8_t res = vcombine_u16(vqrshrun_n_s32(sum0123, ROUND0_BITS), + vqrshrun_n_s32(sum4567, ROUND0_BITS)); + + return vreinterpretq_s16_u16(res); +} + +static AOM_FORCE_INLINE int16x8_t horizontal_filter_8x1_f1_8tap_beta0( + const uint8x16_t in, const int8x16_t filter, const int16x4_t f0, + const uint8x16x2_t perm_tbl) { + const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1)); + + // Permute samples ready for matrix multiply. + // { 1, 2, 3, 4, 5, 6, 7, 8, 3, 4, 5, 6, 7, 8, 9, 10 } + // { 5, 6, 7, 8, 9, 10, 11, 12, 7, 8, 9, 10, 11, 12, 13, 14 } + uint8x16_t perm_samples[2] = { vqtbl1q_u8(in, perm_tbl.val[0]), + vqtbl1q_u8(in, perm_tbl.val[1]) }; + // Get samples 0..7 to apply tap 0 after matrix multiply. + int16x8_t samples_0_7 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(in))); + + // Calculate partial 7-tap convolution. + int32x4_t sum0123 = vusmmlaq_s32(add_const, perm_samples[0], filter); + int32x4_t sum4567 = vusmmlaq_s32(add_const, perm_samples[1], filter); + // Apply tap 0 and accumulate. + sum0123 = vmlal_s16(sum0123, vget_low_s16(samples_0_7), f0); + sum4567 = vmlal_s16(sum4567, vget_high_s16(samples_0_7), f0); + + uint16x8_t res = vcombine_u16(vqrshrun_n_s32(sum0123, ROUND0_BITS), + vqrshrun_n_s32(sum4567, ROUND0_BITS)); + + return vreinterpretq_s16_u16(res); +} + static AOM_FORCE_INLINE int16x8_t horizontal_filter_8x1_f1_beta0(const uint8x16_t in, int16x8_t f_s16) { const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1)); @@ -271,13 +377,209 @@ *res_high = vcombine_s32(vmovn_s64(m45), vmovn_s64(m67)); } +static AOM_FORCE_INLINE void warp_affine_horizontal_sve( + const uint8_t *ref, int width, int height, int stride, int p_width, + int p_height, int16_t alpha, int16_t beta, const int64_t x4, + const int64_t y4, const int i, int16x8_t tmp[]) { + const int bd = 8; + const int reduce_bits_horiz = ROUND0_BITS; + const int height_limit = AOMMIN(8, p_height - i) + 7; + + int32_t ix4 = (int32_t)(x4 >> WARPEDMODEL_PREC_BITS); + int32_t iy4 = (int32_t)(y4 >> WARPEDMODEL_PREC_BITS); + + int32_t sx4 = x4 & ((1 << WARPEDMODEL_PREC_BITS) - 1); + sx4 += alpha * (-4) + beta * (-4) + (1 << (WARPEDDIFF_PREC_BITS - 1)) + + (WARPEDPIXEL_PREC_SHIFTS << WARPEDDIFF_PREC_BITS); + sx4 &= ~((1 << WARP_PARAM_REDUCE_BITS) - 1); + + if (ix4 <= -7) { + for (int k = 0; k < height_limit; ++k) { + int iy = clamp_iy(iy4 + k - 7, height); + int16_t dup_val = + (1 << (bd + FILTER_BITS - reduce_bits_horiz - 1)) + + ref[iy * stride] * (1 << (FILTER_BITS - reduce_bits_horiz)); + tmp[k] = vdupq_n_s16(dup_val); + } + return; + } else if (ix4 >= width + 6) { + for (int k = 0; k < height_limit; ++k) { + int iy = clamp_iy(iy4 + k - 7, height); + int16_t dup_val = (1 << (bd + FILTER_BITS - reduce_bits_horiz - 1)) + + ref[iy * stride + (width - 1)] * + (1 << (FILTER_BITS - reduce_bits_horiz)); + tmp[k] = vdupq_n_s16(dup_val); + } + return; + } + + static const uint8_t kIotaArr[] = { 0, 1, 2, 3, 4, 5, 6, 7, + 8, 9, 10, 11, 12, 13, 14, 15 }; + const uint8x16_t indx = vld1q_u8(kIotaArr); + + const int out_of_boundary_left = -(ix4 - 6); + const int out_of_boundary_right = (ix4 + 8) - width; + + if (p_width == 4) { + if (beta == 0) { + if (alpha == 0) { + int16_t *f_ptr = + (int16_t *)(av1_warped_filter + (sx4 >> WARPEDDIFF_PREC_BITS)); + int16x8_t f_s16 = vld1q_s16(f_ptr); + const int8x8_t x_filter = vmovn_s16(f_s16); + if ((f_ptr[0] | f_ptr[1]) == 0) { + uint8x16_t perm_tbl = vld1q_u8(kMatMul6PermuteTbl); + // Offset the permutation table to match filter layout. + perm_tbl = vaddq_u8(perm_tbl, vdupq_n_u8(2)); + // Stagger filter for use with the matrix multiply instructions. + // { f2, f3, f4, f5, f6, f7, 0, 0, 0, f2, f3, f4, f5, f6, f7, 0 } + const int8x16_t filter = vcombine_s8(vext_s8(x_filter, x_filter, 2), + vext_s8(x_filter, x_filter, 1)); + APPLY_HORIZONTAL_SHIFT(horizontal_filter_4x1_f1_6tap_beta0, filter, + perm_tbl); + } else if ((f_ptr[0] | f_ptr[7]) == 0) { + uint8x16_t perm_tbl = vld1q_u8(kMatMul6PermuteTbl); + // Offset the permutation table to match filter layout. + perm_tbl = vaddq_u8(perm_tbl, vdupq_n_u8(1)); + // Stagger filter for use with the matrix multiply instructions. + // { f1, f2, f3, f4, f5, f6, 0, 0, 0, f1, f2, f3, f4, f5, f6, 0 } + const int8x16_t filter = + vcombine_s8(vext_s8(x_filter, x_filter, 1), x_filter); + APPLY_HORIZONTAL_SHIFT(horizontal_filter_4x1_f1_6tap_beta0, filter, + perm_tbl); + } else if ((f_ptr[6] | f_ptr[7]) == 0) { + const uint8x16_t perm_tbl = vld1q_u8(kMatMul6PermuteTbl); + // Stagger filter for use with the matrix multiply instructions. + // { f0, f1, f2, f3, f4, f5, 0, 0, 0, f0, f1, f2, f3, f4, f5, 0 } + const int8x16_t filter = + vcombine_s8(x_filter, vext_s8(x_filter, x_filter, 7)); + APPLY_HORIZONTAL_SHIFT(horizontal_filter_4x1_f1_6tap_beta0, filter, + perm_tbl); + } else { + const uint8x16_t perm_tbl = vld1q_u8(kMatMul8PermuteTbl); + const uint8x16_t tbl_idx0_3 = vld1q_u8(kTblIdx0_3); + + // Stagger filter for use with the matrix multiply + // instructions. + // { f1, f2, f3, f4, f5, f6, f7, 0, 0, f1, f2, f3, f4, f5, f6, f7 } + const int8x16_t filter = vcombine_s8( + vext_s8(x_filter, vdup_n_s8(0), 1), vset_lane_s8(0, x_filter, 0)); + const int32x4_t f0 = vdupq_n_s32(f_ptr[0]); + + APPLY_HORIZONTAL_SHIFT(horizontal_filter_4x1_f1_8tap_beta0, filter, + f0, perm_tbl, tbl_idx0_3); + } + } else { + APPLY_HORIZONTAL_SHIFT(horizontal_filter_4x1_f4, sx4, alpha); + } + } else { + if (alpha == 0) { + APPLY_HORIZONTAL_SHIFT(horizontal_filter_4x1_f1, + (sx4 + beta * (k - 3))); + } else { + APPLY_HORIZONTAL_SHIFT(horizontal_filter_4x1_f4, (sx4 + beta * (k - 3)), + alpha); + } + } + } else { + if (beta == 0) { + if (alpha == 0) { + int16_t *f_ptr = + (int16_t *)(av1_warped_filter + (sx4 >> WARPEDDIFF_PREC_BITS)); + int16x8_t f_s16 = vld1q_s16(f_ptr); + const int8x8_t x_filter = vmovn_s16(f_s16); + if ((f_ptr[0] | f_ptr[1]) == 0) { + uint8x16x2_t perm_tbl = vld1q_u8_x2(kMatMul6PermuteTbl); + // Offset the permutation table to match filter layout. + perm_tbl.val[0] = vaddq_u8(perm_tbl.val[0], vdupq_n_u8(2)); + perm_tbl.val[1] = vaddq_u8(perm_tbl.val[1], vdupq_n_u8(2)); + // Stagger filter for use with the matrix multiply instructions. + // { f0, f1, f2, f3, f4, f5, 0, 0, 0, f0, f1, f2, f3, f4, f5, 0 } + const int8x16_t filter = vcombine_s8(vext_s8(x_filter, x_filter, 2), + vext_s8(x_filter, x_filter, 1)); + APPLY_HORIZONTAL_SHIFT(horizontal_filter_8x1_f1_6tap_beta0, filter, + perm_tbl); + } else if ((f_ptr[0] | f_ptr[7]) == 0) { + uint8x16x2_t perm_tbl = vld1q_u8_x2(kMatMul6PermuteTbl); + // Offset the permutation table to match filter layout. + perm_tbl.val[0] = vaddq_u8(perm_tbl.val[0], vdupq_n_u8(1)); + perm_tbl.val[1] = vaddq_u8(perm_tbl.val[1], vdupq_n_u8(1)); + // Stagger filter for use with the matrix multiply instructions. + // { f1, f2, f3, f4, f5, f6, 0, 0, 0, f1, f2, f3, f4, f5, f6, 0 } + const int8x16_t filter = + vcombine_s8(vext_s8(x_filter, x_filter, 1), x_filter); + APPLY_HORIZONTAL_SHIFT(horizontal_filter_8x1_f1_6tap_beta0, filter, + perm_tbl); + } else if ((f_ptr[6] | f_ptr[7]) == 0) { + uint8x16x2_t perm_tbl = vld1q_u8_x2(kMatMul6PermuteTbl); + // Stagger filter for use with the matrix multiply instructions. + // { f0, f1, f2, f3, f4, f5, 0, 0, 0, f0, f1, f2, f3, f4, f5, 0 } + const int8x16_t filter = + vcombine_s8(x_filter, vext_s8(x_filter, x_filter, 7)); + APPLY_HORIZONTAL_SHIFT(horizontal_filter_8x1_f1_6tap_beta0, filter, + perm_tbl); + } else { + uint8x16x2_t perm_tbl = vld1q_u8_x2(kMatMul8PermuteTbl); + // Stagger filter for use with the matrix multiply instructions. + // { f1, f2, f3, f4, f5, f6, f7, 0, 0, f1, f2, f3, f4, f5, f6, f7 } + const int8x16_t filter = vcombine_s8( + vext_s8(x_filter, vdup_n_s8(0), 1), vset_lane_s8(0, x_filter, 0)); + + const int16x4_t f0 = vdup_n_s16(f_ptr[0]); + + APPLY_HORIZONTAL_SHIFT(horizontal_filter_8x1_f1_8tap_beta0, filter, + f0, perm_tbl); + } + } else { + APPLY_HORIZONTAL_SHIFT(horizontal_filter_8x1_f8, sx4, alpha); + } + } else { + if (alpha == 0) { + APPLY_HORIZONTAL_SHIFT(horizontal_filter_8x1_f1, + (sx4 + beta * (k - 3))); + } else { + APPLY_HORIZONTAL_SHIFT(horizontal_filter_8x1_f8, (sx4 + beta * (k - 3)), + alpha); + } + } + } +} + void av1_warp_affine_sve(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta) { - av1_warp_affine_common(mat, ref, width, height, stride, pred, p_col, p_row, - p_width, p_height, p_stride, subsampling_x, - subsampling_y, conv_params, alpha, beta, gamma, delta); + const int w0 = conv_params->fwd_offset; + const int w1 = conv_params->bck_offset; + const int is_compound = conv_params->is_compound; + uint16_t *const dst = conv_params->dst; + const int dst_stride = conv_params->dst_stride; + const int do_average = conv_params->do_average; + const int use_dist_wtd_comp_avg = conv_params->use_dist_wtd_comp_avg; + + assert(IMPLIES(is_compound, dst != NULL)); + assert(IMPLIES(do_average, is_compound)); + + for (int i = 0; i < p_height; i += 8) { + for (int j = 0; j < p_width; j += 8) { + const int32_t src_x = (p_col + j + 4) << subsampling_x; + const int32_t src_y = (p_row + i + 4) << subsampling_y; + const int64_t dst_x = + (int64_t)mat[2] * src_x + (int64_t)mat[3] * src_y + (int64_t)mat[0]; + const int64_t dst_y = + (int64_t)mat[4] * src_x + (int64_t)mat[5] * src_y + (int64_t)mat[1]; + + const int64_t x4 = dst_x >> subsampling_x; + const int64_t y4 = dst_y >> subsampling_y; + + int16x8_t tmp[15]; + warp_affine_horizontal_sve(ref, width, height, stride, p_width, p_height, + alpha, beta, x4, y4, i, tmp); + warp_affine_vertical(pred, p_width, p_height, p_stride, is_compound, dst, + dst_stride, do_average, use_dist_wtd_comp_avg, gamma, + delta, y4, i, j, tmp, w0, w1); + } + } }