Add 4-tap path for av1_convolve_2d_horiz_sr_neon_i8mm Add 4-tap specialization for the horizontal pass of av1_convolve_2d_sr_neon_i8mm. This gives up to 10% uplift over using the 8-tap path. Change-Id: I6d67190d424a51e077f9cfe8edf8c89b446780a3
diff --git a/av1/common/arm/convolve_neon_i8mm.c b/av1/common/arm/convolve_neon_i8mm.c index f8b11eb..b6a2a41 100644 --- a/av1/common/arm/convolve_neon_i8mm.c +++ b/av1/common/arm/convolve_neon_i8mm.c
@@ -1060,21 +1060,6 @@ } } -static INLINE int16x4_t convolve4_4_2d_h(uint8x16_t samples, - const int8x8_t filters, - const uint8x16_t permute_tbl, - const int32x4_t horiz_const) { - // Permute samples ready for dot product. - // { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6 } - uint8x16_t permuted_samples = vqtbl1q_u8(samples, permute_tbl); - - // First 4 output values. - int32x4_t sum = vusdotq_lane_s32(horiz_const, permuted_samples, filters, 0); - - // We halved the convolution filter values so -1 from the right shift. - return vshrn_n_s32(sum, ROUND0_BITS - 1); -} - static INLINE int16x8_t convolve8_8_2d_h(uint8x16_t samples, const int8x8_t filters, const uint8x16x3_t permute_tbl, @@ -1103,7 +1088,7 @@ vshrn_n_s32(sum[1], ROUND0_BITS - 1)); } -static INLINE void convolve_2d_sr_horiz_neon_i8mm( +static INLINE void convolve_2d_sr_horiz_8tap_neon_i8mm( const uint8_t *src, int src_stride, int16_t *im_block, int im_stride, int w, int im_h, const int16_t *x_filter_ptr) { const int bd = 8; @@ -1118,85 +1103,173 @@ int dst_stride = im_stride; int height = im_h; - if (w <= 4) { - const uint8x16_t permute_tbl = vld1q_u8(kDotProdPermuteTbl); - // 4-tap filters are used for blocks having width <= 4. - // Filter values are even, so halve to reduce intermediate precision reqs. - const int8x8_t x_filter = - vshrn_n_s16(vcombine_s16(vld1_s16(x_filter_ptr + 2), vdup_n_s16(0)), 1); + const uint8x16x3_t permute_tbl = vld1q_u8_x3(kDotProdPermuteTbl); + // Filter values are even, so halve to reduce intermediate precision reqs. + const int8x8_t x_filter = vshrn_n_s16(vld1q_s16(x_filter_ptr), 1); - src_ptr += 2; + do { + const uint8_t *s = src_ptr; + int16_t *d = dst_ptr; + int width = w; do { uint8x16_t s0, s1, s2, s3; - load_u8_16x4(src_ptr, src_stride, &s0, &s1, &s2, &s3); + load_u8_16x4(s, src_stride, &s0, &s1, &s2, &s3); - int16x4_t d0 = convolve4_4_2d_h(s0, x_filter, permute_tbl, horiz_const); - int16x4_t d1 = convolve4_4_2d_h(s1, x_filter, permute_tbl, horiz_const); - int16x4_t d2 = convolve4_4_2d_h(s2, x_filter, permute_tbl, horiz_const); - int16x4_t d3 = convolve4_4_2d_h(s3, x_filter, permute_tbl, horiz_const); + int16x8_t d0 = convolve8_8_2d_h(s0, x_filter, permute_tbl, horiz_const); + int16x8_t d1 = convolve8_8_2d_h(s1, x_filter, permute_tbl, horiz_const); + int16x8_t d2 = convolve8_8_2d_h(s2, x_filter, permute_tbl, horiz_const); + int16x8_t d3 = convolve8_8_2d_h(s3, x_filter, permute_tbl, horiz_const); - store_s16_4x4(dst_ptr, dst_stride, d0, d1, d2, d3); + store_s16_8x4(d, dst_stride, d0, d1, d2, d3); - src_ptr += 4 * src_stride; - dst_ptr += 4 * dst_stride; + s += 8; + d += 8; + width -= 8; + } while (width != 0); + src_ptr += 4 * src_stride; + dst_ptr += 4 * dst_stride; + height -= 4; + } while (height > 4); + + do { + const uint8_t *s = src_ptr; + int16_t *d = dst_ptr; + int width = w; + + do { + uint8x16_t s0 = vld1q_u8(s); + int16x8_t d0 = convolve8_8_2d_h(s0, x_filter, permute_tbl, horiz_const); + vst1q_s16(d, d0); + + s += 8; + d += 8; + width -= 8; + } while (width != 0); + src_ptr += src_stride; + dst_ptr += dst_stride; + } while (--height != 0); +} + +static INLINE int16x4_t convolve4_4_2d_h(const uint8x16_t samples, + const int8x8_t filters, + const uint8x16_t permute_tbl, + const int32x4_t horiz_const) { + // Permute samples ready for dot product. + // { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6 } + uint8x16_t permuted_samples = vqtbl1q_u8(samples, permute_tbl); + + int32x4_t sum = vusdotq_lane_s32(horiz_const, permuted_samples, filters, 0); + + // We halved the convolution filter values so -1 from the right shift. + return vshrn_n_s32(sum, ROUND0_BITS - 1); +} + +static INLINE int16x8_t convolve4_8_2d_h(const uint8x16_t samples, + const int8x8_t filters, + const uint8x16x2_t permute_tbl, + const int32x4_t horiz_const) { + // Permute samples ready for dot product. + // { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6 } + // { 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10 } + uint8x16_t permuted_samples[2] = { vqtbl1q_u8(samples, permute_tbl.val[0]), + vqtbl1q_u8(samples, permute_tbl.val[1]) }; + + int32x4_t sum0123 = + vusdotq_lane_s32(horiz_const, permuted_samples[0], filters, 0); + int32x4_t sum4567 = + vusdotq_lane_s32(horiz_const, permuted_samples[1], filters, 0); + + // Narrow and re-pack. + // We halved the filter values so -1 from right shift. + return vcombine_s16(vshrn_n_s32(sum0123, ROUND0_BITS - 1), + vshrn_n_s32(sum4567, ROUND0_BITS - 1)); +} + +static INLINE void convolve_2d_sr_horiz_4tap_neon_i8mm( + const uint8_t *src, int src_stride, int16_t *dst, int dst_stride, int width, + int height, const int16_t *filter_x) { + const int bd = 8; + const int16x4_t x_filter = vld1_s16(filter_x + 2); + // All 4-tap and bilinear filter values are even, so halve them to reduce + // intermediate precision requirements. + const int8x8_t filter = vshrn_n_s16(vcombine_s16(x_filter, vdup_n_s16(0)), 1); + + // Adding a shim of 1 << (ROUND0_BITS - 1) enables us to use non-rounding + // shifts - which are generally faster than rounding shifts on modern CPUs. + // Halve the total because we will halve the filter values. + const int32x4_t horiz_const = vdupq_n_s32( + (((1 << (bd + FILTER_BITS - 1)) + (1 << (ROUND0_BITS - 1))) / 2)); + + if (width == 4) { + const uint8x16_t perm_tbl = vld1q_u8(kDotProdPermuteTbl); + do { + uint8x16_t s0, s1, s2, s3; + load_u8_16x4(src, src_stride, &s0, &s1, &s2, &s3); + + int16x4_t d0 = convolve4_4_2d_h(s0, filter, perm_tbl, horiz_const); + int16x4_t d1 = convolve4_4_2d_h(s1, filter, perm_tbl, horiz_const); + int16x4_t d2 = convolve4_4_2d_h(s2, filter, perm_tbl, horiz_const); + int16x4_t d3 = convolve4_4_2d_h(s3, filter, perm_tbl, horiz_const); + + store_s16_4x4(dst, dst_stride, d0, d1, d2, d3); + + src += 4 * src_stride; + dst += 4 * dst_stride; height -= 4; } while (height > 4); do { - uint8x16_t s0 = vld1q_u8(src_ptr); - int16x4_t d0 = convolve4_4_2d_h(s0, x_filter, permute_tbl, horiz_const); - vst1_s16(dst_ptr, d0); + uint8x16_t s0 = vld1q_u8(src); + int16x4_t d0 = convolve4_4_2d_h(s0, filter, perm_tbl, horiz_const); + vst1_s16(dst, d0); - src_ptr += src_stride; - dst_ptr += dst_stride; + src += src_stride; + dst += dst_stride; } while (--height != 0); } else { - const uint8x16x3_t permute_tbl = vld1q_u8_x3(kDotProdPermuteTbl); - // Filter values are even, so halve to reduce intermediate precision reqs. - const int8x8_t x_filter = vshrn_n_s16(vld1q_s16(x_filter_ptr), 1); - + const uint8x16x2_t perm_tbl = vld1q_u8_x2(kDotProdPermuteTbl); do { - const uint8_t *s = src_ptr; - int16_t *d = dst_ptr; - int width = w; + int w = width; + const uint8_t *s = src; + int16_t *d = dst; do { uint8x16_t s0, s1, s2, s3; load_u8_16x4(s, src_stride, &s0, &s1, &s2, &s3); - int16x8_t d0 = convolve8_8_2d_h(s0, x_filter, permute_tbl, horiz_const); - int16x8_t d1 = convolve8_8_2d_h(s1, x_filter, permute_tbl, horiz_const); - int16x8_t d2 = convolve8_8_2d_h(s2, x_filter, permute_tbl, horiz_const); - int16x8_t d3 = convolve8_8_2d_h(s3, x_filter, permute_tbl, horiz_const); + int16x8_t d0 = convolve4_8_2d_h(s0, filter, perm_tbl, horiz_const); + int16x8_t d1 = convolve4_8_2d_h(s1, filter, perm_tbl, horiz_const); + int16x8_t d2 = convolve4_8_2d_h(s2, filter, perm_tbl, horiz_const); + int16x8_t d3 = convolve4_8_2d_h(s3, filter, perm_tbl, horiz_const); store_s16_8x4(d, dst_stride, d0, d1, d2, d3); s += 8; d += 8; - width -= 8; - } while (width != 0); - src_ptr += 4 * src_stride; - dst_ptr += 4 * dst_stride; + w -= 8; + } while (w != 0); + src += 4 * src_stride; + dst += 4 * dst_stride; height -= 4; } while (height > 4); do { - const uint8_t *s = src_ptr; - int16_t *d = dst_ptr; - int width = w; + const uint8_t *s = src; + int16_t *d = dst; + int w = width; do { uint8x16_t s0 = vld1q_u8(s); - int16x8_t d0 = convolve8_8_2d_h(s0, x_filter, permute_tbl, horiz_const); + int16x8_t d0 = convolve4_8_2d_h(s0, filter, perm_tbl, horiz_const); vst1q_s16(d, d0); s += 8; d += 8; - width -= 8; - } while (width != 0); - src_ptr += src_stride; - dst_ptr += dst_stride; + w -= 8; + } while (w != 0); + src += src_stride; + dst += dst_stride; } while (--height != 0); } } @@ -1215,6 +1288,7 @@ } const int y_filter_taps = get_filter_tap(filter_params_y, subpel_y_qn); + const int x_filter_taps = get_filter_tap(filter_params_x, subpel_x_qn); const int clamped_y_taps = y_filter_taps < 6 ? 6 : y_filter_taps; const int im_h = h + clamped_y_taps - 1; const int im_stride = MAX_SB_SIZE; @@ -1246,8 +1320,13 @@ DECLARE_ALIGNED(16, int16_t, im_block[(MAX_SB_SIZE + SUBPEL_TAPS - 1) * MAX_SB_SIZE]); - convolve_2d_sr_horiz_neon_i8mm(src_ptr, src_stride, im_block, im_stride, w, - im_h, x_filter_ptr); + if (x_filter_taps <= 4) { + convolve_2d_sr_horiz_4tap_neon_i8mm(src_ptr + 2, src_stride, im_block, + im_stride, w, im_h, x_filter_ptr); + } else { + convolve_2d_sr_horiz_8tap_neon_i8mm(src_ptr, src_stride, im_block, + im_stride, w, im_h, x_filter_ptr); + } const int16x8_t y_filter = vld1q_s16(y_filter_ptr);