michael@0: 
michael@0: /*
michael@0:  * Copyright 2012 Google Inc.
michael@0:  *
michael@0:  * Use of this source code is governed by a BSD-style license that can be
michael@0:  * found in the LICENSE file.
michael@0:  */
michael@0: #include "SkBitmapProcState.h"
michael@0: #include "SkBitmapProcState_filter.h"
michael@0: #include "SkColorPriv.h"
michael@0: #include "SkFilterProc.h"
michael@0: #include "SkPaint.h"
michael@0: #include "SkShader.h"   // for tilemodes
michael@0: #include "SkUtilsArm.h"
michael@0: 
michael@0: // Required to ensure the table is part of the final binary.
michael@0: extern const SkBitmapProcState::SampleProc32 gSkBitmapProcStateSample32_neon[];
michael@0: extern const SkBitmapProcState::SampleProc16 gSkBitmapProcStateSample16_neon[];
michael@0: 
michael@0: #define   NAME_WRAP(x)  x ## _neon
michael@0: #include "SkBitmapProcState_filter_neon.h"
michael@0: #include "SkBitmapProcState_procs.h"
michael@0: 
michael@0: const SkBitmapProcState::SampleProc32 gSkBitmapProcStateSample32_neon[] = {
michael@0:     S32_opaque_D32_nofilter_DXDY_neon,
michael@0:     S32_alpha_D32_nofilter_DXDY_neon,
michael@0:     S32_opaque_D32_nofilter_DX_neon,
michael@0:     S32_alpha_D32_nofilter_DX_neon,
michael@0:     S32_opaque_D32_filter_DXDY_neon,
michael@0:     S32_alpha_D32_filter_DXDY_neon,
michael@0:     S32_opaque_D32_filter_DX_neon,
michael@0:     S32_alpha_D32_filter_DX_neon,
michael@0: 
michael@0:     S16_opaque_D32_nofilter_DXDY_neon,
michael@0:     S16_alpha_D32_nofilter_DXDY_neon,
michael@0:     S16_opaque_D32_nofilter_DX_neon,
michael@0:     S16_alpha_D32_nofilter_DX_neon,
michael@0:     S16_opaque_D32_filter_DXDY_neon,
michael@0:     S16_alpha_D32_filter_DXDY_neon,
michael@0:     S16_opaque_D32_filter_DX_neon,
michael@0:     S16_alpha_D32_filter_DX_neon,
michael@0: 
michael@0:     SI8_opaque_D32_nofilter_DXDY_neon,
michael@0:     SI8_alpha_D32_nofilter_DXDY_neon,
michael@0:     SI8_opaque_D32_nofilter_DX_neon,
michael@0:     SI8_alpha_D32_nofilter_DX_neon,
michael@0:     SI8_opaque_D32_filter_DXDY_neon,
michael@0:     SI8_alpha_D32_filter_DXDY_neon,
michael@0:     SI8_opaque_D32_filter_DX_neon,
michael@0:     SI8_alpha_D32_filter_DX_neon,
michael@0: 
michael@0:     S4444_opaque_D32_nofilter_DXDY_neon,
michael@0:     S4444_alpha_D32_nofilter_DXDY_neon,
michael@0:     S4444_opaque_D32_nofilter_DX_neon,
michael@0:     S4444_alpha_D32_nofilter_DX_neon,
michael@0:     S4444_opaque_D32_filter_DXDY_neon,
michael@0:     S4444_alpha_D32_filter_DXDY_neon,
michael@0:     S4444_opaque_D32_filter_DX_neon,
michael@0:     S4444_alpha_D32_filter_DX_neon,
michael@0: 
michael@0:     // A8 treats alpha/opauqe the same (equally efficient)
michael@0:     SA8_alpha_D32_nofilter_DXDY_neon,
michael@0:     SA8_alpha_D32_nofilter_DXDY_neon,
michael@0:     SA8_alpha_D32_nofilter_DX_neon,
michael@0:     SA8_alpha_D32_nofilter_DX_neon,
michael@0:     SA8_alpha_D32_filter_DXDY_neon,
michael@0:     SA8_alpha_D32_filter_DXDY_neon,
michael@0:     SA8_alpha_D32_filter_DX_neon,
michael@0:     SA8_alpha_D32_filter_DX_neon
michael@0: };
michael@0: 
michael@0: const SkBitmapProcState::SampleProc16 gSkBitmapProcStateSample16_neon[] = {
michael@0:     S32_D16_nofilter_DXDY_neon,
michael@0:     S32_D16_nofilter_DX_neon,
michael@0:     S32_D16_filter_DXDY_neon,
michael@0:     S32_D16_filter_DX_neon,
michael@0: 
michael@0:     S16_D16_nofilter_DXDY_neon,
michael@0:     S16_D16_nofilter_DX_neon,
michael@0:     S16_D16_filter_DXDY_neon,
michael@0:     S16_D16_filter_DX_neon,
michael@0: 
michael@0:     SI8_D16_nofilter_DXDY_neon,
michael@0:     SI8_D16_nofilter_DX_neon,
michael@0:     SI8_D16_filter_DXDY_neon,
michael@0:     SI8_D16_filter_DX_neon,
michael@0: 
michael@0:     // Don't support 4444 -> 565
michael@0:     NULL, NULL, NULL, NULL,
michael@0:     // Don't support A8 -> 565
michael@0:     NULL, NULL, NULL, NULL
michael@0: };
michael@0: 
michael@0: ///////////////////////////////////////////////////////////////////////////////
michael@0: 
michael@0: #include <arm_neon.h>
michael@0: #include "SkConvolver.h"
michael@0: 
michael@0: // Convolves horizontally along a single row. The row data is given in
michael@0: // |srcData| and continues for the numValues() of the filter.
michael@0: void convolveHorizontally_neon(const unsigned char* srcData,
michael@0:                                const SkConvolutionFilter1D& filter,
michael@0:                                unsigned char* outRow,
michael@0:                                bool hasAlpha) {
michael@0:     // Loop over each pixel on this row in the output image.
michael@0:     int numValues = filter.numValues();
michael@0:     for (int outX = 0; outX < numValues; outX++) {
michael@0:         uint8x8_t coeff_mask0 = vcreate_u8(0x0100010001000100);
michael@0:         uint8x8_t coeff_mask1 = vcreate_u8(0x0302030203020302);
michael@0:         uint8x8_t coeff_mask2 = vcreate_u8(0x0504050405040504);
michael@0:         uint8x8_t coeff_mask3 = vcreate_u8(0x0706070607060706);
michael@0:         // Get the filter that determines the current output pixel.
michael@0:         int filterOffset, filterLength;
michael@0:         const SkConvolutionFilter1D::ConvolutionFixed* filterValues =
michael@0:             filter.FilterForValue(outX, &filterOffset, &filterLength);
michael@0: 
michael@0:         // Compute the first pixel in this row that the filter affects. It will
michael@0:         // touch |filterLength| pixels (4 bytes each) after this.
michael@0:         const unsigned char* rowToFilter = &srcData[filterOffset * 4];
michael@0: 
michael@0:         // Apply the filter to the row to get the destination pixel in |accum|.
michael@0:         int32x4_t accum = vdupq_n_s32(0);
michael@0:         for (int filterX = 0; filterX < filterLength >> 2; filterX++) {
michael@0:             // Load 4 coefficients
michael@0:             int16x4_t coeffs, coeff0, coeff1, coeff2, coeff3;
michael@0:             coeffs = vld1_s16(filterValues);
michael@0:             coeff0 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask0));
michael@0:             coeff1 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask1));
michael@0:             coeff2 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask2));
michael@0:             coeff3 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask3));
michael@0: 
michael@0:             // Load pixels and calc
michael@0:             uint8x16_t pixels = vld1q_u8(rowToFilter);
michael@0:             int16x8_t p01_16 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(pixels)));
michael@0:             int16x8_t p23_16 = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(pixels)));
michael@0: 
michael@0:             int16x4_t p0_src = vget_low_s16(p01_16);
michael@0:             int16x4_t p1_src = vget_high_s16(p01_16);
michael@0:             int16x4_t p2_src = vget_low_s16(p23_16);
michael@0:             int16x4_t p3_src = vget_high_s16(p23_16);
michael@0: 
michael@0:             int32x4_t p0 = vmull_s16(p0_src, coeff0);
michael@0:             int32x4_t p1 = vmull_s16(p1_src, coeff1);
michael@0:             int32x4_t p2 = vmull_s16(p2_src, coeff2);
michael@0:             int32x4_t p3 = vmull_s16(p3_src, coeff3);
michael@0: 
michael@0:             accum += p0;
michael@0:             accum += p1;
michael@0:             accum += p2;
michael@0:             accum += p3;
michael@0: 
michael@0:             // Advance the pointers
michael@0:             rowToFilter += 16;
michael@0:             filterValues += 4;
michael@0:         }
michael@0:         int r = filterLength & 3;
michael@0:         if (r) {
michael@0:             const uint16_t mask[4][4] = {
michael@0:                 {0, 0, 0, 0},
michael@0:                 {0xFFFF, 0, 0, 0},
michael@0:                 {0xFFFF, 0xFFFF, 0, 0},
michael@0:                 {0xFFFF, 0xFFFF, 0xFFFF, 0}
michael@0:             };
michael@0:             uint16x4_t coeffs;
michael@0:             int16x4_t coeff0, coeff1, coeff2;
michael@0:             coeffs = vld1_u16(reinterpret_cast<const uint16_t*>(filterValues));
michael@0:             coeffs &= vld1_u16(&mask[r][0]);
michael@0:             coeff0 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_u16(coeffs), coeff_mask0));
michael@0:             coeff1 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_u16(coeffs), coeff_mask1));
michael@0:             coeff2 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_u16(coeffs), coeff_mask2));
michael@0: 
michael@0:             // Load pixels and calc
michael@0:             uint8x16_t pixels = vld1q_u8(rowToFilter);
michael@0:             int16x8_t p01_16 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(pixels)));
michael@0:             int16x8_t p23_16 = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(pixels)));
michael@0:             int32x4_t p0 = vmull_s16(vget_low_s16(p01_16), coeff0);
michael@0:             int32x4_t p1 = vmull_s16(vget_high_s16(p01_16), coeff1);
michael@0:             int32x4_t p2 = vmull_s16(vget_low_s16(p23_16), coeff2);
michael@0: 
michael@0:             accum += p0;
michael@0:             accum += p1;
michael@0:             accum += p2;
michael@0:         }
michael@0: 
michael@0:         // Bring this value back in range. All of the filter scaling factors
michael@0:         // are in fixed point with kShiftBits bits of fractional part.
michael@0:         accum = vshrq_n_s32(accum, SkConvolutionFilter1D::kShiftBits);
michael@0: 
michael@0:         // Pack and store the new pixel.
michael@0:         int16x4_t accum16 = vqmovn_s32(accum);
michael@0:         uint8x8_t accum8 = vqmovun_s16(vcombine_s16(accum16, accum16));
michael@0:         vst1_lane_u32(reinterpret_cast<uint32_t*>(outRow), vreinterpret_u32_u8(accum8), 0);
michael@0:         outRow += 4;
michael@0:     }
michael@0: }
michael@0: 
michael@0: // Does vertical convolution to produce one output row. The filter values and
michael@0: // length are given in the first two parameters. These are applied to each
michael@0: // of the rows pointed to in the |sourceDataRows| array, with each row
michael@0: // being |pixelWidth| wide.
michael@0: //
michael@0: // The output must have room for |pixelWidth * 4| bytes.
michael@0: template<bool hasAlpha>
michael@0: void convolveVertically_neon(const SkConvolutionFilter1D::ConvolutionFixed* filterValues,
michael@0:                              int filterLength,
michael@0:                              unsigned char* const* sourceDataRows,
michael@0:                              int pixelWidth,
michael@0:                              unsigned char* outRow) {
michael@0:     int width = pixelWidth & ~3;
michael@0: 
michael@0:     int32x4_t accum0, accum1, accum2, accum3;
michael@0:     int16x4_t coeff16;
michael@0: 
michael@0:     // Output four pixels per iteration (16 bytes).
michael@0:     for (int outX = 0; outX < width; outX += 4) {
michael@0: 
michael@0:         // Accumulated result for each pixel. 32 bits per RGBA channel.
michael@0:         accum0 = accum1 = accum2 = accum3 = vdupq_n_s32(0);
michael@0: 
michael@0:         // Convolve with one filter coefficient per iteration.
michael@0:         for (int filterY = 0; filterY < filterLength; filterY++) {
michael@0: 
michael@0:             // Duplicate the filter coefficient 4 times.
michael@0:             // [16] cj cj cj cj
michael@0:             coeff16 = vdup_n_s16(filterValues[filterY]);
michael@0: 
michael@0:             // Load four pixels (16 bytes) together.
michael@0:             // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0
michael@0:             uint8x16_t src8 = vld1q_u8(&sourceDataRows[filterY][outX << 2]);
michael@0: 
michael@0:             int16x8_t src16_01 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(src8)));
michael@0:             int16x8_t src16_23 = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(src8)));
michael@0:             int16x4_t src16_0 = vget_low_s16(src16_01);
michael@0:             int16x4_t src16_1 = vget_high_s16(src16_01);
michael@0:             int16x4_t src16_2 = vget_low_s16(src16_23);
michael@0:             int16x4_t src16_3 = vget_high_s16(src16_23);
michael@0: 
michael@0:             accum0 += vmull_s16(src16_0, coeff16);
michael@0:             accum1 += vmull_s16(src16_1, coeff16);
michael@0:             accum2 += vmull_s16(src16_2, coeff16);
michael@0:             accum3 += vmull_s16(src16_3, coeff16);
michael@0:         }
michael@0: 
michael@0:         // Shift right for fixed point implementation.
michael@0:         accum0 = vshrq_n_s32(accum0, SkConvolutionFilter1D::kShiftBits);
michael@0:         accum1 = vshrq_n_s32(accum1, SkConvolutionFilter1D::kShiftBits);
michael@0:         accum2 = vshrq_n_s32(accum2, SkConvolutionFilter1D::kShiftBits);
michael@0:         accum3 = vshrq_n_s32(accum3, SkConvolutionFilter1D::kShiftBits);
michael@0: 
michael@0:         // Packing 32 bits |accum| to 16 bits per channel (signed saturation).
michael@0:         // [16] a1 b1 g1 r1 a0 b0 g0 r0
michael@0:         int16x8_t accum16_0 = vcombine_s16(vqmovn_s32(accum0), vqmovn_s32(accum1));
michael@0:         // [16] a3 b3 g3 r3 a2 b2 g2 r2
michael@0:         int16x8_t accum16_1 = vcombine_s16(vqmovn_s32(accum2), vqmovn_s32(accum3));
michael@0: 
michael@0:         // Packing 16 bits |accum| to 8 bits per channel (unsigned saturation).
michael@0:         // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0
michael@0:         uint8x16_t accum8 = vcombine_u8(vqmovun_s16(accum16_0), vqmovun_s16(accum16_1));
michael@0: 
michael@0:         if (hasAlpha) {
michael@0:             // Compute the max(ri, gi, bi) for each pixel.
michael@0:             // [8] xx a3 b3 g3 xx a2 b2 g2 xx a1 b1 g1 xx a0 b0 g0
michael@0:             uint8x16_t a = vreinterpretq_u8_u32(vshrq_n_u32(vreinterpretq_u32_u8(accum8), 8));
michael@0:             // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0
michael@0:             uint8x16_t b = vmaxq_u8(a, accum8); // Max of r and g
michael@0:             // [8] xx xx a3 b3 xx xx a2 b2 xx xx a1 b1 xx xx a0 b0
michael@0:             a = vreinterpretq_u8_u32(vshrq_n_u32(vreinterpretq_u32_u8(accum8), 16));
michael@0:             // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0
michael@0:             b = vmaxq_u8(a, b); // Max of r and g and b.
michael@0:             // [8] max3 00 00 00 max2 00 00 00 max1 00 00 00 max0 00 00 00
michael@0:             b = vreinterpretq_u8_u32(vshlq_n_u32(vreinterpretq_u32_u8(b), 24));
michael@0: 
michael@0:             // Make sure the value of alpha channel is always larger than maximum
michael@0:             // value of color channels.
michael@0:             accum8 = vmaxq_u8(b, accum8);
michael@0:         } else {
michael@0:             // Set value of alpha channels to 0xFF.
michael@0:             accum8 = vreinterpretq_u8_u32(vreinterpretq_u32_u8(accum8) | vdupq_n_u32(0xFF000000));
michael@0:         }
michael@0: 
michael@0:         // Store the convolution result (16 bytes) and advance the pixel pointers.
michael@0:         vst1q_u8(outRow, accum8);
michael@0:         outRow += 16;
michael@0:     }
michael@0: 
michael@0:     // Process the leftovers when the width of the output is not divisible
michael@0:     // by 4, that is at most 3 pixels.
michael@0:     int r = pixelWidth & 3;
michael@0:     if (r) {
michael@0: 
michael@0:         accum0 = accum1 = accum2 = vdupq_n_s32(0);
michael@0: 
michael@0:         for (int filterY = 0; filterY < filterLength; ++filterY) {
michael@0:             coeff16 = vdup_n_s16(filterValues[filterY]);
michael@0: 
michael@0:             // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0
michael@0:             uint8x16_t src8 = vld1q_u8(&sourceDataRows[filterY][width << 2]);
michael@0: 
michael@0:             int16x8_t src16_01 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(src8)));
michael@0:             int16x8_t src16_23 = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(src8)));
michael@0:             int16x4_t src16_0 = vget_low_s16(src16_01);
michael@0:             int16x4_t src16_1 = vget_high_s16(src16_01);
michael@0:             int16x4_t src16_2 = vget_low_s16(src16_23);
michael@0: 
michael@0:             accum0 += vmull_s16(src16_0, coeff16);
michael@0:             accum1 += vmull_s16(src16_1, coeff16);
michael@0:             accum2 += vmull_s16(src16_2, coeff16);
michael@0:         }
michael@0: 
michael@0:         accum0 = vshrq_n_s32(accum0, SkConvolutionFilter1D::kShiftBits);
michael@0:         accum1 = vshrq_n_s32(accum1, SkConvolutionFilter1D::kShiftBits);
michael@0:         accum2 = vshrq_n_s32(accum2, SkConvolutionFilter1D::kShiftBits);
michael@0: 
michael@0:         int16x8_t accum16_0 = vcombine_s16(vqmovn_s32(accum0), vqmovn_s32(accum1));
michael@0:         int16x8_t accum16_1 = vcombine_s16(vqmovn_s32(accum2), vqmovn_s32(accum2));
michael@0: 
michael@0:         uint8x16_t accum8 = vcombine_u8(vqmovun_s16(accum16_0), vqmovun_s16(accum16_1));
michael@0: 
michael@0:         if (hasAlpha) {
michael@0:             // Compute the max(ri, gi, bi) for each pixel.
michael@0:             // [8] xx a3 b3 g3 xx a2 b2 g2 xx a1 b1 g1 xx a0 b0 g0
michael@0:             uint8x16_t a = vreinterpretq_u8_u32(vshrq_n_u32(vreinterpretq_u32_u8(accum8), 8));
michael@0:             // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0
michael@0:             uint8x16_t b = vmaxq_u8(a, accum8); // Max of r and g
michael@0:             // [8] xx xx a3 b3 xx xx a2 b2 xx xx a1 b1 xx xx a0 b0
michael@0:             a = vreinterpretq_u8_u32(vshrq_n_u32(vreinterpretq_u32_u8(accum8), 16));
michael@0:             // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0
michael@0:             b = vmaxq_u8(a, b); // Max of r and g and b.
michael@0:             // [8] max3 00 00 00 max2 00 00 00 max1 00 00 00 max0 00 00 00
michael@0:             b = vreinterpretq_u8_u32(vshlq_n_u32(vreinterpretq_u32_u8(b), 24));
michael@0: 
michael@0:             // Make sure the value of alpha channel is always larger than maximum
michael@0:             // value of color channels.
michael@0:             accum8 = vmaxq_u8(b, accum8);
michael@0:         } else {
michael@0:             // Set value of alpha channels to 0xFF.
michael@0:             accum8 = vreinterpretq_u8_u32(vreinterpretq_u32_u8(accum8) | vdupq_n_u32(0xFF000000));
michael@0:         }
michael@0: 
michael@0:         switch(r) {
michael@0:         case 1:
michael@0:             vst1q_lane_u32(reinterpret_cast<uint32_t*>(outRow), vreinterpretq_u32_u8(accum8), 0);
michael@0:             break;
michael@0:         case 2:
michael@0:             vst1_u32(reinterpret_cast<uint32_t*>(outRow),
michael@0:                      vreinterpret_u32_u8(vget_low_u8(accum8)));
michael@0:             break;
michael@0:         case 3:
michael@0:             vst1_u32(reinterpret_cast<uint32_t*>(outRow),
michael@0:                      vreinterpret_u32_u8(vget_low_u8(accum8)));
michael@0:             vst1q_lane_u32(reinterpret_cast<uint32_t*>(outRow+8), vreinterpretq_u32_u8(accum8), 2);
michael@0:             break;
michael@0:         }
michael@0:     }
michael@0: }
michael@0: 
michael@0: void convolveVertically_neon(const SkConvolutionFilter1D::ConvolutionFixed* filterValues,
michael@0:                              int filterLength,
michael@0:                              unsigned char* const* sourceDataRows,
michael@0:                              int pixelWidth,
michael@0:                              unsigned char* outRow,
michael@0:                              bool sourceHasAlpha) {
michael@0:     if (sourceHasAlpha) {
michael@0:         convolveVertically_neon<true>(filterValues, filterLength,
michael@0:                                       sourceDataRows, pixelWidth,
michael@0:                                       outRow);
michael@0:     } else {
michael@0:         convolveVertically_neon<false>(filterValues, filterLength,
michael@0:                                        sourceDataRows, pixelWidth,
michael@0:                                        outRow);
michael@0:     }
michael@0: }
michael@0: 
michael@0: // Convolves horizontally along four rows. The row data is given in
michael@0: // |src_data| and continues for the num_values() of the filter.
michael@0: // The algorithm is almost same as |ConvolveHorizontally_SSE2|. Please
michael@0: // refer to that function for detailed comments.
michael@0: void convolve4RowsHorizontally_neon(const unsigned char* srcData[4],
michael@0:                                     const SkConvolutionFilter1D& filter,
michael@0:                                     unsigned char* outRow[4]) {
michael@0: 
michael@0:     uint8x8_t coeff_mask0 = vcreate_u8(0x0100010001000100);
michael@0:     uint8x8_t coeff_mask1 = vcreate_u8(0x0302030203020302);
michael@0:     uint8x8_t coeff_mask2 = vcreate_u8(0x0504050405040504);
michael@0:     uint8x8_t coeff_mask3 = vcreate_u8(0x0706070607060706);
michael@0:     int num_values = filter.numValues();
michael@0: 
michael@0:     int filterOffset, filterLength;
michael@0:     // |mask| will be used to decimate all extra filter coefficients that are
michael@0:     // loaded by SIMD when |filter_length| is not divisible by 4.
michael@0:     // mask[0] is not used in following algorithm.
michael@0:     const uint16_t mask[4][4] = {
michael@0:         {0, 0, 0, 0},
michael@0:         {0xFFFF, 0, 0, 0},
michael@0:         {0xFFFF, 0xFFFF, 0, 0},
michael@0:         {0xFFFF, 0xFFFF, 0xFFFF, 0}
michael@0:     };
michael@0: 
michael@0:     // Output one pixel each iteration, calculating all channels (RGBA) together.
michael@0:     for (int outX = 0; outX < num_values; outX++) {
michael@0: 
michael@0:         const SkConvolutionFilter1D::ConvolutionFixed* filterValues =
michael@0:         filter.FilterForValue(outX, &filterOffset, &filterLength);
michael@0: 
michael@0:         // four pixels in a column per iteration.
michael@0:         int32x4_t accum0 = vdupq_n_s32(0);
michael@0:         int32x4_t accum1 = vdupq_n_s32(0);
michael@0:         int32x4_t accum2 = vdupq_n_s32(0);
michael@0:         int32x4_t accum3 = vdupq_n_s32(0);
michael@0: 
michael@0:         int start = (filterOffset<<2);
michael@0: 
michael@0:         // We will load and accumulate with four coefficients per iteration.
michael@0:         for (int filter_x = 0; filter_x < (filterLength >> 2); filter_x++) {
michael@0:             int16x4_t coeffs, coeff0, coeff1, coeff2, coeff3;
michael@0: 
michael@0:             coeffs = vld1_s16(filterValues);
michael@0:             coeff0 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask0));
michael@0:             coeff1 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask1));
michael@0:             coeff2 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask2));
michael@0:             coeff3 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask3));
michael@0: 
michael@0:             uint8x16_t pixels;
michael@0:             int16x8_t p01_16, p23_16;
michael@0:             int32x4_t p0, p1, p2, p3;
michael@0: 
michael@0: 
michael@0: #define ITERATION(src, accum)                                       \
michael@0:     pixels = vld1q_u8(src);                                         \
michael@0:     p01_16 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(pixels)));  \
michael@0:     p23_16 = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(pixels))); \
michael@0:     p0 = vmull_s16(vget_low_s16(p01_16), coeff0);                   \
michael@0:     p1 = vmull_s16(vget_high_s16(p01_16), coeff1);                  \
michael@0:     p2 = vmull_s16(vget_low_s16(p23_16), coeff2);                   \
michael@0:     p3 = vmull_s16(vget_high_s16(p23_16), coeff3);                  \
michael@0:     accum += p0;                                                    \
michael@0:     accum += p1;                                                    \
michael@0:     accum += p2;                                                    \
michael@0:     accum += p3
michael@0: 
michael@0:             ITERATION(srcData[0] + start, accum0);
michael@0:             ITERATION(srcData[1] + start, accum1);
michael@0:             ITERATION(srcData[2] + start, accum2);
michael@0:             ITERATION(srcData[3] + start, accum3);
michael@0: 
michael@0:             start += 16;
michael@0:             filterValues += 4;
michael@0:         }
michael@0: 
michael@0:         int r = filterLength & 3;
michael@0:         if (r) {
michael@0:             int16x4_t coeffs, coeff0, coeff1, coeff2, coeff3;
michael@0:             coeffs = vld1_s16(filterValues);
michael@0:             coeffs &= vreinterpret_s16_u16(vld1_u16(&mask[r][0]));
michael@0:             coeff0 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask0));
michael@0:             coeff1 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask1));
michael@0:             coeff2 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask2));
michael@0:             coeff3 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask3));
michael@0: 
michael@0:             uint8x16_t pixels;
michael@0:             int16x8_t p01_16, p23_16;
michael@0:             int32x4_t p0, p1, p2, p3;
michael@0: 
michael@0:             ITERATION(srcData[0] + start, accum0);
michael@0:             ITERATION(srcData[1] + start, accum1);
michael@0:             ITERATION(srcData[2] + start, accum2);
michael@0:             ITERATION(srcData[3] + start, accum3);
michael@0:         }
michael@0: 
michael@0:         int16x4_t accum16;
michael@0:         uint8x8_t res0, res1, res2, res3;
michael@0: 
michael@0: #define PACK_RESULT(accum, res)                                         \
michael@0:         accum = vshrq_n_s32(accum, SkConvolutionFilter1D::kShiftBits);  \
michael@0:         accum16 = vqmovn_s32(accum);                                    \
michael@0:         res = vqmovun_s16(vcombine_s16(accum16, accum16));
michael@0: 
michael@0:         PACK_RESULT(accum0, res0);
michael@0:         PACK_RESULT(accum1, res1);
michael@0:         PACK_RESULT(accum2, res2);
michael@0:         PACK_RESULT(accum3, res3);
michael@0: 
michael@0:         vst1_lane_u32(reinterpret_cast<uint32_t*>(outRow[0]), vreinterpret_u32_u8(res0), 0);
michael@0:         vst1_lane_u32(reinterpret_cast<uint32_t*>(outRow[1]), vreinterpret_u32_u8(res1), 0);
michael@0:         vst1_lane_u32(reinterpret_cast<uint32_t*>(outRow[2]), vreinterpret_u32_u8(res2), 0);
michael@0:         vst1_lane_u32(reinterpret_cast<uint32_t*>(outRow[3]), vreinterpret_u32_u8(res3), 0);
michael@0:         outRow[0] += 4;
michael@0:         outRow[1] += 4;
michael@0:         outRow[2] += 4;
michael@0:         outRow[3] += 4;
michael@0:     }
michael@0: }
michael@0: 
michael@0: void applySIMDPadding_neon(SkConvolutionFilter1D *filter) {
michael@0:     // Padding |paddingCount| of more dummy coefficients after the coefficients
michael@0:     // of last filter to prevent SIMD instructions which load 8 or 16 bytes
michael@0:     // together to access invalid memory areas. We are not trying to align the
michael@0:     // coefficients right now due to the opaqueness of <vector> implementation.
michael@0:     // This has to be done after all |AddFilter| calls.
michael@0:     for (int i = 0; i < 8; ++i) {
michael@0:         filter->addFilterValue(static_cast<SkConvolutionFilter1D::ConvolutionFixed>(0));
michael@0:     }
michael@0: }
michael@0: 
michael@0: void platformConvolutionProcs_arm_neon(SkConvolutionProcs* procs) {
michael@0:     procs->fExtraHorizontalReads = 3;
michael@0:     procs->fConvolveVertically = &convolveVertically_neon;
michael@0:     procs->fConvolve4RowsHorizontally = &convolve4RowsHorizontally_neon;
michael@0:     procs->fConvolveHorizontally = &convolveHorizontally_neon;
michael@0:     procs->fApplySIMDPadding = &applySIMDPadding_neon;
michael@0: }