wine/gecko
0
0
Fork 0
mirror of https://gitlab.winehq.org/wine/wine-gecko.git synced 2024-09-13 09:24:08 -07:00
Code Issues Packages Projects Releases Wiki Activity

Bug 1054510 - Turned on SSE for image resizing. r=jrmuizel

Browse Source
This commit is contained in:
Artem Sobolev 2014-09-03 11:20:18 -07:00
parent 94927c5089
commit 910938cba0
6 changed files with 612 additions and 494 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

527
gfx/2d/convolver.cpp
View File

@ -26,18 +26,20 @@
// OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
// SUCH DAMAGE.
#include "2D.h"
#include "convolver.h"
#include <algorithm>
#include "skia/SkTypes.h"
// note: SIMD_SSE2 is not enabled because of bugs, apparently
#if defined(SIMD_SSE2)
#include <emmintrin.h> // ARCH_CPU_X86_FAMILY was defined in build/config.h
#if defined(USE_SSE2)
#include "convolverSSE2.h"
#endif
using mozilla::gfx::Factory;
#if defined(SK_CPU_LENDIAN)
#define R_OFFSET_IDX 0
#define G_OFFSET_IDX 1
@ -152,7 +154,7 @@ class CircularRowBuffer {
};
// Convolves horizontally along a single row. The row data is given in
// |src_data| and continues for the num_values() of the filter.
// |src_data| and continues for the [begin, end) of the filter.
template<bool has_alpha>
// This function is miscompiled with gcc 4.5 with pgo. See bug 827946.
#if defined(__GNUC__) && defined(MOZ_GCC_VERSION_AT_LEAST)
@ -161,11 +163,11 @@ __attribute__((optimize("-O1")))
#endif
#endif
void ConvolveHorizontally(const unsigned char* src_data,
int begin, int end,
const ConvolutionFilter1D& filter,
unsigned char* out_row) {
// Loop over each pixel on this row in the output image.
int num_values = filter.num_values();
for (int out_x = 0; out_x < num_values; out_x++) {
for (int out_x = begin; out_x < end; out_x++) {
// Get the filter that determines the current output pixel.
int filter_offset, filter_length;
const ConvolutionFilter1D::Fixed* filter_values =
@ -206,18 +208,17 @@ void ConvolveHorizontally(const unsigned char* src_data,
// Does vertical convolution to produce one output row. The filter values and
// length are given in the first two parameters. These are applied to each
// of the rows pointed to in the |source_data_rows| array, with each row
// being |pixel_width| wide.
// being |end - begin| wide.
//
// The output must have room for |pixel_width * 4| bytes.
// The output must have room for |(end - begin) * 4| bytes.
template<bool has_alpha>
void ConvolveVertically(const ConvolutionFilter1D::Fixed* filter_values,
int filter_length,
unsigned char* const* source_data_rows,
int pixel_width,
unsigned char* out_row) {
int begin, int end, unsigned char* out_row) {
// We go through each column in the output and do a vertical convolution,
// generating one output pixel each time.
for (int out_x = 0; out_x < pixel_width; out_x++) {
for (int out_x = begin; out_x < end; out_x++) {
// Compute the number of bytes over in each row that the current column
// we're convolving starts at. The pixel will cover the next 4 bytes.
int byte_offset = out_x * 4;
@ -272,433 +273,6 @@ void ConvolveVertically(const ConvolutionFilter1D::Fixed* filter_values,
}
}
// Convolves horizontally along a single row. The row data is given in
// |src_data| and continues for the num_values() of the filter.
void ConvolveHorizontally_SSE2(const unsigned char* src_data,
const ConvolutionFilter1D& filter,
unsigned char* out_row) {
#if defined(SIMD_SSE2)
int num_values = filter.num_values();
int filter_offset, filter_length;
__m128i zero = _mm_setzero_si128();
__m128i mask[4];
// |mask| will be used to decimate all extra filter coefficients that are
// loaded by SIMD when |filter_length| is not divisible by 4.
// mask[0] is not used in following algorithm.
mask[1] = _mm_set_epi16(0, 0, 0, 0, 0, 0, 0, -1);
mask[2] = _mm_set_epi16(0, 0, 0, 0, 0, 0, -1, -1);
mask[3] = _mm_set_epi16(0, 0, 0, 0, 0, -1, -1, -1);
// Output one pixel each iteration, calculating all channels (RGBA) together.
for (int out_x = 0; out_x < num_values; out_x++) {
const ConvolutionFilter1D::Fixed* filter_values =
filter.FilterForValue(out_x, &filter_offset, &filter_length);
__m128i accum = _mm_setzero_si128();
// Compute the first pixel in this row that the filter affects. It will
// touch |filter_length| pixels (4 bytes each) after this.
const __m128i* row_to_filter =
reinterpret_cast<const __m128i*>(&src_data[filter_offset << 2]);
// We will load and accumulate with four coefficients per iteration.
for (int filter_x = 0; filter_x < filter_length >> 2; filter_x++) {
// Load 4 coefficients => duplicate 1st and 2nd of them for all channels.
__m128i coeff, coeff16;
// [16] xx xx xx xx c3 c2 c1 c0
coeff = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(filter_values));
// [16] xx xx xx xx c1 c1 c0 c0
coeff16 = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(1, 1, 0, 0));
// [16] c1 c1 c1 c1 c0 c0 c0 c0
coeff16 = _mm_unpacklo_epi16(coeff16, coeff16);
// Load four pixels => unpack the first two pixels to 16 bits =>
// multiply with coefficients => accumulate the convolution result.
// [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0
__m128i src8 = _mm_loadu_si128(row_to_filter);
// [16] a1 b1 g1 r1 a0 b0 g0 r0
__m128i src16 = _mm_unpacklo_epi8(src8, zero);
__m128i mul_hi = _mm_mulhi_epi16(src16, coeff16);
__m128i mul_lo = _mm_mullo_epi16(src16, coeff16);
// [32] a0*c0 b0*c0 g0*c0 r0*c0
__m128i t = _mm_unpacklo_epi16(mul_lo, mul_hi);
accum = _mm_add_epi32(accum, t);
// [32] a1*c1 b1*c1 g1*c1 r1*c1
t = _mm_unpackhi_epi16(mul_lo, mul_hi);
accum = _mm_add_epi32(accum, t);
// Duplicate 3rd and 4th coefficients for all channels =>
// unpack the 3rd and 4th pixels to 16 bits => multiply with coefficients
// => accumulate the convolution results.
// [16] xx xx xx xx c3 c3 c2 c2
coeff16 = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(3, 3, 2, 2));
// [16] c3 c3 c3 c3 c2 c2 c2 c2
coeff16 = _mm_unpacklo_epi16(coeff16, coeff16);
// [16] a3 g3 b3 r3 a2 g2 b2 r2
src16 = _mm_unpackhi_epi8(src8, zero);
mul_hi = _mm_mulhi_epi16(src16, coeff16);
mul_lo = _mm_mullo_epi16(src16, coeff16);
// [32] a2*c2 b2*c2 g2*c2 r2*c2
t = _mm_unpacklo_epi16(mul_lo, mul_hi);
accum = _mm_add_epi32(accum, t);
// [32] a3*c3 b3*c3 g3*c3 r3*c3
t = _mm_unpackhi_epi16(mul_lo, mul_hi);
accum = _mm_add_epi32(accum, t);
// Advance the pixel and coefficients pointers.
row_to_filter += 1;
filter_values += 4;
}
// When |filter_length| is not divisible by 4, we need to decimate some of
// the filter coefficient that was loaded incorrectly to zero; Other than
// that the algorithm is same with above, exceot that the 4th pixel will be
// always absent.
int r = filter_length&3;
if (r) {
// Note: filter_values must be padded to align_up(filter_offset, 8).
__m128i coeff, coeff16;
coeff = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(filter_values));
// Mask out extra filter taps.
coeff = _mm_and_si128(coeff, mask[r]);
coeff16 = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(1, 1, 0, 0));
coeff16 = _mm_unpacklo_epi16(coeff16, coeff16);
// Note: line buffer must be padded to align_up(filter_offset, 16).
// We resolve this by use C-version for the last horizontal line.
__m128i src8 = _mm_loadu_si128(row_to_filter);
__m128i src16 = _mm_unpacklo_epi8(src8, zero);
__m128i mul_hi = _mm_mulhi_epi16(src16, coeff16);
__m128i mul_lo = _mm_mullo_epi16(src16, coeff16);
__m128i t = _mm_unpacklo_epi16(mul_lo, mul_hi);
accum = _mm_add_epi32(accum, t);
t = _mm_unpackhi_epi16(mul_lo, mul_hi);
accum = _mm_add_epi32(accum, t);
src16 = _mm_unpackhi_epi8(src8, zero);
coeff16 = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(3, 3, 2, 2));
coeff16 = _mm_unpacklo_epi16(coeff16, coeff16);
mul_hi = _mm_mulhi_epi16(src16, coeff16);
mul_lo = _mm_mullo_epi16(src16, coeff16);
t = _mm_unpacklo_epi16(mul_lo, mul_hi);
accum = _mm_add_epi32(accum, t);
}
// Shift right for fixed point implementation.
accum = _mm_srai_epi32(accum, ConvolutionFilter1D::kShiftBits);
// Packing 32 bits |accum| to 16 bits per channel (signed saturation).
accum = _mm_packs_epi32(accum, zero);
// Packing 16 bits |accum| to 8 bits per channel (unsigned saturation).
accum = _mm_packus_epi16(accum, zero);
// Store the pixel value of 32 bits.
*(reinterpret_cast<int*>(out_row)) = _mm_cvtsi128_si32(accum);
out_row += 4;
}
#endif
}
// Convolves horizontally along four rows. The row data is given in
// |src_data| and continues for the num_values() of the filter.
// The algorithm is almost same as |ConvolveHorizontally_SSE2|. Please
// refer to that function for detailed comments.
void ConvolveHorizontally4_SSE2(const unsigned char* src_data[4],
const ConvolutionFilter1D& filter,
unsigned char* out_row[4]) {
#if defined(SIMD_SSE2)
int num_values = filter.num_values();
int filter_offset, filter_length;
__m128i zero = _mm_setzero_si128();
__m128i mask[4];
// |mask| will be used to decimate all extra filter coefficients that are
// loaded by SIMD when |filter_length| is not divisible by 4.
// mask[0] is not used in following algorithm.
mask[1] = _mm_set_epi16(0, 0, 0, 0, 0, 0, 0, -1);
mask[2] = _mm_set_epi16(0, 0, 0, 0, 0, 0, -1, -1);
mask[3] = _mm_set_epi16(0, 0, 0, 0, 0, -1, -1, -1);
// Output one pixel each iteration, calculating all channels (RGBA) together.
for (int out_x = 0; out_x < num_values; out_x++) {
const ConvolutionFilter1D::Fixed* filter_values =
filter.FilterForValue(out_x, &filter_offset, &filter_length);
// four pixels in a column per iteration.
__m128i accum0 = _mm_setzero_si128();
__m128i accum1 = _mm_setzero_si128();
__m128i accum2 = _mm_setzero_si128();
__m128i accum3 = _mm_setzero_si128();
int start = (filter_offset<<2);
// We will load and accumulate with four coefficients per iteration.
for (int filter_x = 0; filter_x < (filter_length >> 2); filter_x++) {
__m128i coeff, coeff16lo, coeff16hi;
// [16] xx xx xx xx c3 c2 c1 c0
coeff = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(filter_values));
// [16] xx xx xx xx c1 c1 c0 c0
coeff16lo = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(1, 1, 0, 0));
// [16] c1 c1 c1 c1 c0 c0 c0 c0
coeff16lo = _mm_unpacklo_epi16(coeff16lo, coeff16lo);
// [16] xx xx xx xx c3 c3 c2 c2
coeff16hi = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(3, 3, 2, 2));
// [16] c3 c3 c3 c3 c2 c2 c2 c2
coeff16hi = _mm_unpacklo_epi16(coeff16hi, coeff16hi);
__m128i src8, src16, mul_hi, mul_lo, t;
#define ITERATION(src, accum) \
src8 = _mm_loadu_si128(reinterpret_cast<const __m128i*>(src)); \
src16 = _mm_unpacklo_epi8(src8, zero); \
mul_hi = _mm_mulhi_epi16(src16, coeff16lo); \
mul_lo = _mm_mullo_epi16(src16, coeff16lo); \
t = _mm_unpacklo_epi16(mul_lo, mul_hi); \
accum = _mm_add_epi32(accum, t); \
t = _mm_unpackhi_epi16(mul_lo, mul_hi); \
accum = _mm_add_epi32(accum, t); \
src16 = _mm_unpackhi_epi8(src8, zero); \
mul_hi = _mm_mulhi_epi16(src16, coeff16hi); \
mul_lo = _mm_mullo_epi16(src16, coeff16hi); \
t = _mm_unpacklo_epi16(mul_lo, mul_hi); \
accum = _mm_add_epi32(accum, t); \
t = _mm_unpackhi_epi16(mul_lo, mul_hi); \
accum = _mm_add_epi32(accum, t)
ITERATION(src_data[0] + start, accum0);
ITERATION(src_data[1] + start, accum1);
ITERATION(src_data[2] + start, accum2);
ITERATION(src_data[3] + start, accum3);
start += 16;
filter_values += 4;
}
int r = filter_length & 3;
if (r) {
// Note: filter_values must be padded to align_up(filter_offset, 8);
__m128i coeff;
coeff = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(filter_values));
// Mask out extra filter taps.
coeff = _mm_and_si128(coeff, mask[r]);
__m128i coeff16lo = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(1, 1, 0, 0));
/* c1 c1 c1 c1 c0 c0 c0 c0 */
coeff16lo = _mm_unpacklo_epi16(coeff16lo, coeff16lo);
__m128i coeff16hi = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(3, 3, 2, 2));
coeff16hi = _mm_unpacklo_epi16(coeff16hi, coeff16hi);
__m128i src8, src16, mul_hi, mul_lo, t;
ITERATION(src_data[0] + start, accum0);
ITERATION(src_data[1] + start, accum1);
ITERATION(src_data[2] + start, accum2);
ITERATION(src_data[3] + start, accum3);
}
accum0 = _mm_srai_epi32(accum0, ConvolutionFilter1D::kShiftBits);
accum0 = _mm_packs_epi32(accum0, zero);
accum0 = _mm_packus_epi16(accum0, zero);
accum1 = _mm_srai_epi32(accum1, ConvolutionFilter1D::kShiftBits);
accum1 = _mm_packs_epi32(accum1, zero);
accum1 = _mm_packus_epi16(accum1, zero);
accum2 = _mm_srai_epi32(accum2, ConvolutionFilter1D::kShiftBits);
accum2 = _mm_packs_epi32(accum2, zero);
accum2 = _mm_packus_epi16(accum2, zero);
accum3 = _mm_srai_epi32(accum3, ConvolutionFilter1D::kShiftBits);
accum3 = _mm_packs_epi32(accum3, zero);
accum3 = _mm_packus_epi16(accum3, zero);
*(reinterpret_cast<int*>(out_row[0])) = _mm_cvtsi128_si32(accum0);
*(reinterpret_cast<int*>(out_row[1])) = _mm_cvtsi128_si32(accum1);
*(reinterpret_cast<int*>(out_row[2])) = _mm_cvtsi128_si32(accum2);
*(reinterpret_cast<int*>(out_row[3])) = _mm_cvtsi128_si32(accum3);
out_row[0] += 4;
out_row[1] += 4;
out_row[2] += 4;
out_row[3] += 4;
}
#endif
}
// Does vertical convolution to produce one output row. The filter values and
// length are given in the first two parameters. These are applied to each
// of the rows pointed to in the |source_data_rows| array, with each row
// being |pixel_width| wide.
//
// The output must have room for |pixel_width * 4| bytes.
template<bool has_alpha>
void ConvolveVertically_SSE2(const ConvolutionFilter1D::Fixed* filter_values,
int filter_length,
unsigned char* const* source_data_rows,
int pixel_width,
unsigned char* out_row) {
#if defined(SIMD_SSE2)
int width = pixel_width & ~3;
__m128i zero = _mm_setzero_si128();
__m128i accum0, accum1, accum2, accum3, coeff16;
const __m128i* src;
// Output four pixels per iteration (16 bytes).
for (int out_x = 0; out_x < width; out_x += 4) {
// Accumulated result for each pixel. 32 bits per RGBA channel.
accum0 = _mm_setzero_si128();
accum1 = _mm_setzero_si128();
accum2 = _mm_setzero_si128();
accum3 = _mm_setzero_si128();
// Convolve with one filter coefficient per iteration.
for (int filter_y = 0; filter_y < filter_length; filter_y++) {
// Duplicate the filter coefficient 8 times.
// [16] cj cj cj cj cj cj cj cj
coeff16 = _mm_set1_epi16(filter_values[filter_y]);
// Load four pixels (16 bytes) together.
// [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0
src = reinterpret_cast<const __m128i*>(
&source_data_rows[filter_y][out_x << 2]);
__m128i src8 = _mm_loadu_si128(src);
// Unpack 1st and 2nd pixels from 8 bits to 16 bits for each channels =>
// multiply with current coefficient => accumulate the result.
// [16] a1 b1 g1 r1 a0 b0 g0 r0
__m128i src16 = _mm_unpacklo_epi8(src8, zero);
__m128i mul_hi = _mm_mulhi_epi16(src16, coeff16);
__m128i mul_lo = _mm_mullo_epi16(src16, coeff16);
// [32] a0 b0 g0 r0
__m128i t = _mm_unpacklo_epi16(mul_lo, mul_hi);
accum0 = _mm_add_epi32(accum0, t);
// [32] a1 b1 g1 r1
t = _mm_unpackhi_epi16(mul_lo, mul_hi);
accum1 = _mm_add_epi32(accum1, t);
// Unpack 3rd and 4th pixels from 8 bits to 16 bits for each channels =>
// multiply with current coefficient => accumulate the result.
// [16] a3 b3 g3 r3 a2 b2 g2 r2
src16 = _mm_unpackhi_epi8(src8, zero);
mul_hi = _mm_mulhi_epi16(src16, coeff16);
mul_lo = _mm_mullo_epi16(src16, coeff16);
// [32] a2 b2 g2 r2
t = _mm_unpacklo_epi16(mul_lo, mul_hi);
accum2 = _mm_add_epi32(accum2, t);
// [32] a3 b3 g3 r3
t = _mm_unpackhi_epi16(mul_lo, mul_hi);
accum3 = _mm_add_epi32(accum3, t);
}
// Shift right for fixed point implementation.
accum0 = _mm_srai_epi32(accum0, ConvolutionFilter1D::kShiftBits);
accum1 = _mm_srai_epi32(accum1, ConvolutionFilter1D::kShiftBits);
accum2 = _mm_srai_epi32(accum2, ConvolutionFilter1D::kShiftBits);
accum3 = _mm_srai_epi32(accum3, ConvolutionFilter1D::kShiftBits);
// Packing 32 bits |accum| to 16 bits per channel (signed saturation).
// [16] a1 b1 g1 r1 a0 b0 g0 r0
accum0 = _mm_packs_epi32(accum0, accum1);
// [16] a3 b3 g3 r3 a2 b2 g2 r2
accum2 = _mm_packs_epi32(accum2, accum3);
// Packing 16 bits |accum| to 8 bits per channel (unsigned saturation).
// [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0
accum0 = _mm_packus_epi16(accum0, accum2);
if (has_alpha) {
// Compute the max(ri, gi, bi) for each pixel.
// [8] xx a3 b3 g3 xx a2 b2 g2 xx a1 b1 g1 xx a0 b0 g0
__m128i a = _mm_srli_epi32(accum0, 8);
// [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0
__m128i b = _mm_max_epu8(a, accum0); // Max of r and g.
// [8] xx xx a3 b3 xx xx a2 b2 xx xx a1 b1 xx xx a0 b0
a = _mm_srli_epi32(accum0, 16);
// [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0
b = _mm_max_epu8(a, b); // Max of r and g and b.
// [8] max3 00 00 00 max2 00 00 00 max1 00 00 00 max0 00 00 00
b = _mm_slli_epi32(b, 24);
// Make sure the value of alpha channel is always larger than maximum
// value of color channels.
accum0 = _mm_max_epu8(b, accum0);
} else {
// Set value of alpha channels to 0xFF.
__m128i mask = _mm_set1_epi32(0xff000000);
accum0 = _mm_or_si128(accum0, mask);
}
// Store the convolution result (16 bytes) and advance the pixel pointers.
_mm_storeu_si128(reinterpret_cast<__m128i*>(out_row), accum0);
out_row += 16;
}
// When the width of the output is not divisible by 4, We need to save one
// pixel (4 bytes) each time. And also the fourth pixel is always absent.
if (pixel_width & 3) {
accum0 = _mm_setzero_si128();
accum1 = _mm_setzero_si128();
accum2 = _mm_setzero_si128();
for (int filter_y = 0; filter_y < filter_length; ++filter_y) {
coeff16 = _mm_set1_epi16(filter_values[filter_y]);
// [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0
src = reinterpret_cast<const __m128i*>(
&source_data_rows[filter_y][width<<2]);
__m128i src8 = _mm_loadu_si128(src);
// [16] a1 b1 g1 r1 a0 b0 g0 r0
__m128i src16 = _mm_unpacklo_epi8(src8, zero);
__m128i mul_hi = _mm_mulhi_epi16(src16, coeff16);
__m128i mul_lo = _mm_mullo_epi16(src16, coeff16);
// [32] a0 b0 g0 r0
__m128i t = _mm_unpacklo_epi16(mul_lo, mul_hi);
accum0 = _mm_add_epi32(accum0, t);
// [32] a1 b1 g1 r1
t = _mm_unpackhi_epi16(mul_lo, mul_hi);
accum1 = _mm_add_epi32(accum1, t);
// [16] a3 b3 g3 r3 a2 b2 g2 r2
src16 = _mm_unpackhi_epi8(src8, zero);
mul_hi = _mm_mulhi_epi16(src16, coeff16);
mul_lo = _mm_mullo_epi16(src16, coeff16);
// [32] a2 b2 g2 r2
t = _mm_unpacklo_epi16(mul_lo, mul_hi);
accum2 = _mm_add_epi32(accum2, t);
}
accum0 = _mm_srai_epi32(accum0, ConvolutionFilter1D::kShiftBits);
accum1 = _mm_srai_epi32(accum1, ConvolutionFilter1D::kShiftBits);
accum2 = _mm_srai_epi32(accum2, ConvolutionFilter1D::kShiftBits);
// [16] a1 b1 g1 r1 a0 b0 g0 r0
accum0 = _mm_packs_epi32(accum0, accum1);
// [16] a3 b3 g3 r3 a2 b2 g2 r2
accum2 = _mm_packs_epi32(accum2, zero);
// [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0
accum0 = _mm_packus_epi16(accum0, accum2);
if (has_alpha) {
// [8] xx a3 b3 g3 xx a2 b2 g2 xx a1 b1 g1 xx a0 b0 g0
__m128i a = _mm_srli_epi32(accum0, 8);
// [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0
__m128i b = _mm_max_epu8(a, accum0); // Max of r and g.
// [8] xx xx a3 b3 xx xx a2 b2 xx xx a1 b1 xx xx a0 b0
a = _mm_srli_epi32(accum0, 16);
// [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0
b = _mm_max_epu8(a, b); // Max of r and g and b.
// [8] max3 00 00 00 max2 00 00 00 max1 00 00 00 max0 00 00 00
b = _mm_slli_epi32(b, 24);
accum0 = _mm_max_epu8(b, accum0);
} else {
__m128i mask = _mm_set1_epi32(0xff000000);
accum0 = _mm_or_si128(accum0, mask);
}
for (int out_x = width; out_x < pixel_width; out_x++) {
*(reinterpret_cast<int*>(out_row)) = _mm_cvtsi128_si32(accum0);
accum0 = _mm_srli_si128(accum0, 4);
out_row += 4;
}
}
#endif
}
} // namespace
// ConvolutionFilter1D ---------------------------------------------------------
@ -770,14 +344,16 @@ void BGRAConvolve2D(const unsigned char* source_data,
const ConvolutionFilter1D& filter_x,
const ConvolutionFilter1D& filter_y,
int output_byte_row_stride,
unsigned char* output,
bool use_sse2) {
#if !defined(SIMD_SSE2)
unsigned char* output) {
bool use_sse2 = Factory::HasSSE2();
#if !defined(USE_SSE2)
// Even we have runtime support for SSE2 instructions, since the binary
// was not built with SSE2 support, we had to fallback to C version.
use_sse2 = false;
#endif
int max_y_filter_size = filter_y.max_filter();
// The next row in the input that we will generate a horizontally
@ -810,6 +386,7 @@ void BGRAConvolve2D(const unsigned char* source_data,
// convolutions to run each subsequent vertical convolution.
SkASSERT(output_byte_row_stride >= filter_x.num_values() * 4);
int num_output_rows = filter_y.num_values();
int pixel_width = filter_x.num_values();
// We need to check which is the last line to convolve before we advance 4
// lines in one iteration.
@ -823,47 +400,56 @@ void BGRAConvolve2D(const unsigned char* source_data,
// Generate output rows until we have enough to run the current filter.
if (use_sse2) {
#if defined(USE_SSE2)
// We don't want to process too much rows in batches of 4 because
// we can go out-of-bounds at the end
while (next_x_row < filter_offset + filter_length) {
if (next_x_row + 3 < last_filter_offset + last_filter_length - 1) {
if (next_x_row + 3 < last_filter_offset + last_filter_length - 3) {
const unsigned char* src[4];
unsigned char* out_row[4];
for (int i = 0; i < 4; ++i) {
src[i] = &source_data[(next_x_row + i) * source_byte_row_stride];
out_row[i] = row_buffer.AdvanceRow();
}
ConvolveHorizontally4_SSE2(src, filter_x, out_row);
ConvolveHorizontally4_SSE2(src, 0, pixel_width, filter_x, out_row);
next_x_row += 4;
} else {
// For the last row, SSE2 load possibly to access data beyond the
// image area. therefore we use C version here.
if (next_x_row == last_filter_offset + last_filter_length - 1) {
unsigned char* buffer = row_buffer.AdvanceRow();
// For last rows, SSE2 load possibly to access data beyond the
// image area. therefore we use cobined C+SSE version here
int simd_width = pixel_width & ~3;
if (simd_width) {
ConvolveHorizontally_SSE2(
&source_data[next_x_row * source_byte_row_stride],
0, simd_width, filter_x, buffer);
}
if (pixel_width > simd_width) {
if (source_has_alpha) {
ConvolveHorizontally<true>(
&source_data[next_x_row * source_byte_row_stride],
filter_x, row_buffer.AdvanceRow());
simd_width, pixel_width, filter_x, buffer);
} else {
ConvolveHorizontally<false>(
&source_data[next_x_row * source_byte_row_stride],
filter_x, row_buffer.AdvanceRow());
simd_width, pixel_width, filter_x, buffer);
}
} else {
ConvolveHorizontally_SSE2(
&source_data[next_x_row * source_byte_row_stride],
filter_x, row_buffer.AdvanceRow());
}
next_x_row++;
}
}
#endif
} else {
while (next_x_row < filter_offset + filter_length) {
if (source_has_alpha) {
ConvolveHorizontally<true>(
&source_data[next_x_row * source_byte_row_stride],
filter_x, row_buffer.AdvanceRow());
0, pixel_width, filter_x, row_buffer.AdvanceRow());
} else {
ConvolveHorizontally<false>(
&source_data[next_x_row * source_byte_row_stride],
filter_x, row_buffer.AdvanceRow());
0, pixel_width, filter_x, row_buffer.AdvanceRow());
}
next_x_row++;
}
@ -882,26 +468,23 @@ void BGRAConvolve2D(const unsigned char* source_data,
unsigned char* const* first_row_for_filter =
&rows_to_convolve[filter_offset - first_row_in_circular_buffer];
int processed = 0;
#if defined(USE_SSE2)
int simd_width = pixel_width & ~3;
if (use_sse2 && simd_width) {
ConvolveVertically_SSE2(filter_values, filter_length, first_row_for_filter,
0, simd_width, cur_output_row, source_has_alpha);
processed = simd_width;
}
#endif
if (source_has_alpha) {
if (use_sse2) {
ConvolveVertically_SSE2<true>(filter_values, filter_length,
first_row_for_filter,
filter_x.num_values(), cur_output_row);
} else {
ConvolveVertically<true>(filter_values, filter_length,
first_row_for_filter,
filter_x.num_values(), cur_output_row);
}
ConvolveVertically<true>(filter_values, filter_length,
first_row_for_filter,
processed, pixel_width, cur_output_row);
} else {
if (use_sse2) {
ConvolveVertically_SSE2<false>(filter_values, filter_length,
first_row_for_filter,
filter_x.num_values(), cur_output_row);
} else {
ConvolveVertically<false>(filter_values, filter_length,
first_row_for_filter,
filter_x.num_values(), cur_output_row);
}
ConvolveVertically<false>(filter_values, filter_length,
first_row_for_filter,
processed, pixel_width, cur_output_row);
}
}
}

4
gfx/2d/convolver.h
View File

@ -184,8 +184,8 @@ void BGRAConvolve2D(const unsigned char* source_data,
const ConvolutionFilter1D& xfilter,
const ConvolutionFilter1D& yfilter,
int output_byte_row_stride,
unsigned char* output,
bool use_sse2);
unsigned char* output);
} // namespace skia
#endif // SKIA_EXT_CONVOLVER_H_

476
gfx/2d/convolverSSE2.cpp Normal file
View File

@ -0,0 +1,476 @@
// Copyright (c) 2006-2011 The Chromium Authors. All rights reserved.
//
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions
// are met:
// * Redistributions of source code must retain the above copyright
// notice, this list of conditions and the following disclaimer.
// * Redistributions in binary form must reproduce the above copyright
// notice, this list of conditions and the following disclaimer in
// the documentation and/or other materials provided with the
// distribution.
// * Neither the name of Google, Inc. nor the names of its contributors
// may be used to endorse or promote products derived from this
// software without specific prior written permission.
//
// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
// FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
// COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
// INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
// BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
// OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
// AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
// OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
// OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
// SUCH DAMAGE.
#include "convolver.h"
#include <algorithm>
#include "skia/SkTypes.h"
#include <emmintrin.h> // ARCH_CPU_X86_FAMILY was defined in build/config.h
namespace skia {
// Convolves horizontally along a single row. The row data is given in
// |src_data| and continues for the [begin, end) of the filter.
void ConvolveHorizontally_SSE2(const unsigned char* src_data,
int begin, int end,
const ConvolutionFilter1D& filter,
unsigned char* out_row) {
int filter_offset, filter_length;
__m128i zero = _mm_setzero_si128();
__m128i mask[3];
// |mask| will be used to decimate all extra filter coefficients that are
// loaded by SIMD when |filter_length| is not divisible by 4.
mask[0] = _mm_set_epi16(0, 0, 0, 0, 0, 0, 0, -1);
mask[1] = _mm_set_epi16(0, 0, 0, 0, 0, 0, -1, -1);
mask[2] = _mm_set_epi16(0, 0, 0, 0, 0, -1, -1, -1);
// This buffer is used for tails
__m128i buffer;
// Output one pixel each iteration, calculating all channels (RGBA) together.
for (int out_x = begin; out_x < end; out_x++) {
const ConvolutionFilter1D::Fixed* filter_values =
filter.FilterForValue(out_x, &filter_offset, &filter_length);
__m128i accum = _mm_setzero_si128();
// Compute the first pixel in this row that the filter affects. It will
// touch |filter_length| pixels (4 bytes each) after this.
const __m128i* row_to_filter =
reinterpret_cast<const __m128i*>(&src_data[filter_offset << 2]);
// We will load and accumulate with four coefficients per iteration.
for (int filter_x = 0; filter_x < filter_length >> 2; filter_x++) {
// Load 4 coefficients => duplicate 1st and 2nd of them for all channels.
__m128i coeff, coeff16;
// [16] xx xx xx xx c3 c2 c1 c0
coeff = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(filter_values));
// [16] xx xx xx xx c1 c1 c0 c0
coeff16 = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(1, 1, 0, 0));
// [16] c1 c1 c1 c1 c0 c0 c0 c0
coeff16 = _mm_unpacklo_epi16(coeff16, coeff16);
// Load four pixels => unpack the first two pixels to 16 bits =>
// multiply with coefficients => accumulate the convolution result.
// [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0
__m128i src8 = _mm_loadu_si128(row_to_filter);
// [16] a1 b1 g1 r1 a0 b0 g0 r0
__m128i src16 = _mm_unpacklo_epi8(src8, zero);
__m128i mul_hi = _mm_mulhi_epi16(src16, coeff16);
__m128i mul_lo = _mm_mullo_epi16(src16, coeff16);
// [32] a0*c0 b0*c0 g0*c0 r0*c0
__m128i t = _mm_unpacklo_epi16(mul_lo, mul_hi);
accum = _mm_add_epi32(accum, t);
// [32] a1*c1 b1*c1 g1*c1 r1*c1
t = _mm_unpackhi_epi16(mul_lo, mul_hi);
accum = _mm_add_epi32(accum, t);
// Duplicate 3rd and 4th coefficients for all channels =>
// unpack the 3rd and 4th pixels to 16 bits => multiply with coefficients
// => accumulate the convolution results.
// [16] xx xx xx xx c3 c3 c2 c2
coeff16 = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(3, 3, 2, 2));
// [16] c3 c3 c3 c3 c2 c2 c2 c2
coeff16 = _mm_unpacklo_epi16(coeff16, coeff16);
// [16] a3 g3 b3 r3 a2 g2 b2 r2
src16 = _mm_unpackhi_epi8(src8, zero);
mul_hi = _mm_mulhi_epi16(src16, coeff16);
mul_lo = _mm_mullo_epi16(src16, coeff16);
// [32] a2*c2 b2*c2 g2*c2 r2*c2
t = _mm_unpacklo_epi16(mul_lo, mul_hi);
accum = _mm_add_epi32(accum, t);
// [32] a3*c3 b3*c3 g3*c3 r3*c3
t = _mm_unpackhi_epi16(mul_lo, mul_hi);
accum = _mm_add_epi32(accum, t);
// Advance the pixel and coefficients pointers.
row_to_filter += 1;
filter_values += 4;
}
// When |filter_length| is not divisible by 4, we need to decimate some of
// the filter coefficient that was loaded incorrectly to zero; Other than
// that the algorithm is same with above, except that the 4th pixel will be
// always absent.
int r = filter_length & 3;
if (r) {
memcpy(&buffer, row_to_filter, r * 4);
// Note: filter_values must be padded to align_up(filter_offset, 8).
__m128i coeff, coeff16;
coeff = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(filter_values));
// Mask out extra filter taps.
coeff = _mm_and_si128(coeff, mask[r-1]);
coeff16 = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(1, 1, 0, 0));
coeff16 = _mm_unpacklo_epi16(coeff16, coeff16);
// Note: line buffer must be padded to align_up(filter_offset, 16).
// We resolve this by temporary buffer
__m128i src8 = _mm_loadu_si128(&buffer);
__m128i src16 = _mm_unpacklo_epi8(src8, zero);
__m128i mul_hi = _mm_mulhi_epi16(src16, coeff16);
__m128i mul_lo = _mm_mullo_epi16(src16, coeff16);
__m128i t = _mm_unpacklo_epi16(mul_lo, mul_hi);
accum = _mm_add_epi32(accum, t);
t = _mm_unpackhi_epi16(mul_lo, mul_hi);
accum = _mm_add_epi32(accum, t);
src16 = _mm_unpackhi_epi8(src8, zero);
coeff16 = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(3, 3, 2, 2));
coeff16 = _mm_unpacklo_epi16(coeff16, coeff16);
mul_hi = _mm_mulhi_epi16(src16, coeff16);
mul_lo = _mm_mullo_epi16(src16, coeff16);
t = _mm_unpacklo_epi16(mul_lo, mul_hi);
accum = _mm_add_epi32(accum, t);
}
// Shift right for fixed point implementation.
accum = _mm_srai_epi32(accum, ConvolutionFilter1D::kShiftBits);
// Packing 32 bits |accum| to 16 bits per channel (signed saturation).
accum = _mm_packs_epi32(accum, zero);
// Packing 16 bits |accum| to 8 bits per channel (unsigned saturation).
accum = _mm_packus_epi16(accum, zero);
// Store the pixel value of 32 bits.
*(reinterpret_cast<int*>(out_row)) = _mm_cvtsi128_si32(accum);
out_row += 4;
}
}
// Convolves horizontally along four rows. The row data is given in
// |src_data| and continues for the [begin, end) of the filter.
// The algorithm is almost same as |ConvolveHorizontally_SSE2|. Please
// refer to that function for detailed comments.
void ConvolveHorizontally4_SSE2(const unsigned char* src_data[4],
int begin, int end,
const ConvolutionFilter1D& filter,
unsigned char* out_row[4]) {
int filter_offset, filter_length;
__m128i zero = _mm_setzero_si128();
__m128i mask[3];
// |mask| will be used to decimate all extra filter coefficients that are
// loaded by SIMD when |filter_length| is not divisible by 4.
mask[0] = _mm_set_epi16(0, 0, 0, 0, 0, 0, 0, -1);
mask[1] = _mm_set_epi16(0, 0, 0, 0, 0, 0, -1, -1);
mask[2] = _mm_set_epi16(0, 0, 0, 0, 0, -1, -1, -1);
// Output one pixel each iteration, calculating all channels (RGBA) together.
for (int out_x = begin; out_x < end; out_x++) {
const ConvolutionFilter1D::Fixed* filter_values =
filter.FilterForValue(out_x, &filter_offset, &filter_length);
// four pixels in a column per iteration.
__m128i accum0 = _mm_setzero_si128();
__m128i accum1 = _mm_setzero_si128();
__m128i accum2 = _mm_setzero_si128();
__m128i accum3 = _mm_setzero_si128();
int start = (filter_offset<<2);
// We will load and accumulate with four coefficients per iteration.
for (int filter_x = 0; filter_x < (filter_length >> 2); filter_x++) {
__m128i coeff, coeff16lo, coeff16hi;
// [16] xx xx xx xx c3 c2 c1 c0
coeff = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(filter_values));
// [16] xx xx xx xx c1 c1 c0 c0
coeff16lo = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(1, 1, 0, 0));
// [16] c1 c1 c1 c1 c0 c0 c0 c0
coeff16lo = _mm_unpacklo_epi16(coeff16lo, coeff16lo);
// [16] xx xx xx xx c3 c3 c2 c2
coeff16hi = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(3, 3, 2, 2));
// [16] c3 c3 c3 c3 c2 c2 c2 c2
coeff16hi = _mm_unpacklo_epi16(coeff16hi, coeff16hi);
__m128i src8, src16, mul_hi, mul_lo, t;
#define ITERATION(src, accum) \
src8 = _mm_loadu_si128(reinterpret_cast<const __m128i*>(src)); \
src16 = _mm_unpacklo_epi8(src8, zero); \
mul_hi = _mm_mulhi_epi16(src16, coeff16lo); \
mul_lo = _mm_mullo_epi16(src16, coeff16lo); \
t = _mm_unpacklo_epi16(mul_lo, mul_hi); \
accum = _mm_add_epi32(accum, t); \
t = _mm_unpackhi_epi16(mul_lo, mul_hi); \
accum = _mm_add_epi32(accum, t); \
src16 = _mm_unpackhi_epi8(src8, zero); \
mul_hi = _mm_mulhi_epi16(src16, coeff16hi); \
mul_lo = _mm_mullo_epi16(src16, coeff16hi); \
t = _mm_unpacklo_epi16(mul_lo, mul_hi); \
accum = _mm_add_epi32(accum, t); \
t = _mm_unpackhi_epi16(mul_lo, mul_hi); \
accum = _mm_add_epi32(accum, t)
ITERATION(src_data[0] + start, accum0);
ITERATION(src_data[1] + start, accum1);
ITERATION(src_data[2] + start, accum2);
ITERATION(src_data[3] + start, accum3);
start += 16;
filter_values += 4;
}
int r = filter_length & 3;
if (r) {
// Note: filter_values must be padded to align_up(filter_offset, 8);
__m128i coeff;
coeff = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(filter_values));
// Mask out extra filter taps.
coeff = _mm_and_si128(coeff, mask[r-1]);
__m128i coeff16lo = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(1, 1, 0, 0));
/* c1 c1 c1 c1 c0 c0 c0 c0 */
coeff16lo = _mm_unpacklo_epi16(coeff16lo, coeff16lo);
__m128i coeff16hi = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(3, 3, 2, 2));
coeff16hi = _mm_unpacklo_epi16(coeff16hi, coeff16hi);
__m128i src8, src16, mul_hi, mul_lo, t;
ITERATION(src_data[0] + start, accum0);
ITERATION(src_data[1] + start, accum1);
ITERATION(src_data[2] + start, accum2);
ITERATION(src_data[3] + start, accum3);
}
accum0 = _mm_srai_epi32(accum0, ConvolutionFilter1D::kShiftBits);
accum0 = _mm_packs_epi32(accum0, zero);
accum0 = _mm_packus_epi16(accum0, zero);
accum1 = _mm_srai_epi32(accum1, ConvolutionFilter1D::kShiftBits);
accum1 = _mm_packs_epi32(accum1, zero);
accum1 = _mm_packus_epi16(accum1, zero);
accum2 = _mm_srai_epi32(accum2, ConvolutionFilter1D::kShiftBits);
accum2 = _mm_packs_epi32(accum2, zero);
accum2 = _mm_packus_epi16(accum2, zero);
accum3 = _mm_srai_epi32(accum3, ConvolutionFilter1D::kShiftBits);
accum3 = _mm_packs_epi32(accum3, zero);
accum3 = _mm_packus_epi16(accum3, zero);
*(reinterpret_cast<int*>(out_row[0])) = _mm_cvtsi128_si32(accum0);
*(reinterpret_cast<int*>(out_row[1])) = _mm_cvtsi128_si32(accum1);
*(reinterpret_cast<int*>(out_row[2])) = _mm_cvtsi128_si32(accum2);
*(reinterpret_cast<int*>(out_row[3])) = _mm_cvtsi128_si32(accum3);
out_row[0] += 4;
out_row[1] += 4;
out_row[2] += 4;
out_row[3] += 4;
}
}
// Does vertical convolution to produce one output row. The filter values and
// length are given in the first two parameters. These are applied to each
// of the rows pointed to in the |source_data_rows| array, with each row
// being |end - begin| wide.
//
// The output must have room for |(end - begin) * 4| bytes.
template<bool has_alpha>
void ConvolveVertically_SSE2_impl(const ConvolutionFilter1D::Fixed* filter_values,
int filter_length,
unsigned char* const* source_data_rows,
int begin, int end,
unsigned char* out_row) {
__m128i zero = _mm_setzero_si128();
__m128i accum0, accum1, accum2, accum3, coeff16;
const __m128i* src;
int out_x;
// Output four pixels per iteration (16 bytes).
for (out_x = begin; out_x + 3 < end; out_x += 4) {
// Accumulated result for each pixel. 32 bits per RGBA channel.
accum0 = _mm_setzero_si128();
accum1 = _mm_setzero_si128();
accum2 = _mm_setzero_si128();
accum3 = _mm_setzero_si128();
// Convolve with one filter coefficient per iteration.
for (int filter_y = 0; filter_y < filter_length; filter_y++) {
// Duplicate the filter coefficient 8 times.
// [16] cj cj cj cj cj cj cj cj
coeff16 = _mm_set1_epi16(filter_values[filter_y]);
// Load four pixels (16 bytes) together.
// [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0
src = reinterpret_cast<const __m128i*>(
&source_data_rows[filter_y][out_x << 2]);
__m128i src8 = _mm_loadu_si128(src);
// Unpack 1st and 2nd pixels from 8 bits to 16 bits for each channels =>
// multiply with current coefficient => accumulate the result.
// [16] a1 b1 g1 r1 a0 b0 g0 r0
__m128i src16 = _mm_unpacklo_epi8(src8, zero);
__m128i mul_hi = _mm_mulhi_epi16(src16, coeff16);
__m128i mul_lo = _mm_mullo_epi16(src16, coeff16);
// [32] a0 b0 g0 r0
__m128i t = _mm_unpacklo_epi16(mul_lo, mul_hi);
accum0 = _mm_add_epi32(accum0, t);
// [32] a1 b1 g1 r1
t = _mm_unpackhi_epi16(mul_lo, mul_hi);
accum1 = _mm_add_epi32(accum1, t);
// Unpack 3rd and 4th pixels from 8 bits to 16 bits for each channels =>
// multiply with current coefficient => accumulate the result.
// [16] a3 b3 g3 r3 a2 b2 g2 r2
src16 = _mm_unpackhi_epi8(src8, zero);
mul_hi = _mm_mulhi_epi16(src16, coeff16);
mul_lo = _mm_mullo_epi16(src16, coeff16);
// [32] a2 b2 g2 r2
t = _mm_unpacklo_epi16(mul_lo, mul_hi);
accum2 = _mm_add_epi32(accum2, t);
// [32] a3 b3 g3 r3
t = _mm_unpackhi_epi16(mul_lo, mul_hi);
accum3 = _mm_add_epi32(accum3, t);
}
// Shift right for fixed point implementation.
accum0 = _mm_srai_epi32(accum0, ConvolutionFilter1D::kShiftBits);
accum1 = _mm_srai_epi32(accum1, ConvolutionFilter1D::kShiftBits);
accum2 = _mm_srai_epi32(accum2, ConvolutionFilter1D::kShiftBits);
accum3 = _mm_srai_epi32(accum3, ConvolutionFilter1D::kShiftBits);
// Packing 32 bits |accum| to 16 bits per channel (signed saturation).
// [16] a1 b1 g1 r1 a0 b0 g0 r0
accum0 = _mm_packs_epi32(accum0, accum1);
// [16] a3 b3 g3 r3 a2 b2 g2 r2
accum2 = _mm_packs_epi32(accum2, accum3);
// Packing 16 bits |accum| to 8 bits per channel (unsigned saturation).
// [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0
accum0 = _mm_packus_epi16(accum0, accum2);
if (has_alpha) {
// Compute the max(ri, gi, bi) for each pixel.
// [8] xx a3 b3 g3 xx a2 b2 g2 xx a1 b1 g1 xx a0 b0 g0
__m128i a = _mm_srli_epi32(accum0, 8);
// [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0
__m128i b = _mm_max_epu8(a, accum0); // Max of r and g.
// [8] xx xx a3 b3 xx xx a2 b2 xx xx a1 b1 xx xx a0 b0
a = _mm_srli_epi32(accum0, 16);
// [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0
b = _mm_max_epu8(a, b); // Max of r and g and b.
// [8] max3 00 00 00 max2 00 00 00 max1 00 00 00 max0 00 00 00
b = _mm_slli_epi32(b, 24);
// Make sure the value of alpha channel is always larger than maximum
// value of color channels.
accum0 = _mm_max_epu8(b, accum0);
} else {
// Set value of alpha channels to 0xFF.
__m128i mask = _mm_set1_epi32(0xff000000);
accum0 = _mm_or_si128(accum0, mask);
}
// Store the convolution result (16 bytes) and advance the pixel pointers.
_mm_storeu_si128(reinterpret_cast<__m128i*>(out_row), accum0);
out_row += 16;
}
// When the width of the output is not divisible by 4, We need to save one
// pixel (4 bytes) each time. And also the fourth pixel is always absent.
int r = end - out_x;
if (r > 0) {
// Since accum3 is never used here, we'll use it as a buffer
__m128i *buffer = &accum3;
accum0 = _mm_setzero_si128();
accum1 = _mm_setzero_si128();
accum2 = _mm_setzero_si128();
for (int filter_y = 0; filter_y < filter_length; ++filter_y) {
coeff16 = _mm_set1_epi16(filter_values[filter_y]);
// [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0
src = reinterpret_cast<const __m128i*>(
&source_data_rows[filter_y][out_x * 4]);
memcpy(buffer, src, r * 4);
__m128i src8 = _mm_loadu_si128(buffer);
// [16] a1 b1 g1 r1 a0 b0 g0 r0
__m128i src16 = _mm_unpacklo_epi8(src8, zero);
__m128i mul_hi = _mm_mulhi_epi16(src16, coeff16);
__m128i mul_lo = _mm_mullo_epi16(src16, coeff16);
// [32] a0 b0 g0 r0
__m128i t = _mm_unpacklo_epi16(mul_lo, mul_hi);
accum0 = _mm_add_epi32(accum0, t);
// [32] a1 b1 g1 r1
t = _mm_unpackhi_epi16(mul_lo, mul_hi);
accum1 = _mm_add_epi32(accum1, t);
// [16] a3 b3 g3 r3 a2 b2 g2 r2
src16 = _mm_unpackhi_epi8(src8, zero);
mul_hi = _mm_mulhi_epi16(src16, coeff16);
mul_lo = _mm_mullo_epi16(src16, coeff16);
// [32] a2 b2 g2 r2
t = _mm_unpacklo_epi16(mul_lo, mul_hi);
accum2 = _mm_add_epi32(accum2, t);
}
accum0 = _mm_srai_epi32(accum0, ConvolutionFilter1D::kShiftBits);
accum1 = _mm_srai_epi32(accum1, ConvolutionFilter1D::kShiftBits);
accum2 = _mm_srai_epi32(accum2, ConvolutionFilter1D::kShiftBits);
// [16] a1 b1 g1 r1 a0 b0 g0 r0
accum0 = _mm_packs_epi32(accum0, accum1);
// [16] a3 b3 g3 r3 a2 b2 g2 r2
accum2 = _mm_packs_epi32(accum2, zero);
// [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0
accum0 = _mm_packus_epi16(accum0, accum2);
if (has_alpha) {
// [8] xx a3 b3 g3 xx a2 b2 g2 xx a1 b1 g1 xx a0 b0 g0
__m128i a = _mm_srli_epi32(accum0, 8);
// [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0
__m128i b = _mm_max_epu8(a, accum0); // Max of r and g.
// [8] xx xx a3 b3 xx xx a2 b2 xx xx a1 b1 xx xx a0 b0
a = _mm_srli_epi32(accum0, 16);
// [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0
b = _mm_max_epu8(a, b); // Max of r and g and b.
// [8] max3 00 00 00 max2 00 00 00 max1 00 00 00 max0 00 00 00
b = _mm_slli_epi32(b, 24);
accum0 = _mm_max_epu8(b, accum0);
} else {
__m128i mask = _mm_set1_epi32(0xff000000);
accum0 = _mm_or_si128(accum0, mask);
}
for (; out_x < end; out_x++) {
*(reinterpret_cast<int*>(out_row)) = _mm_cvtsi128_si32(accum0);
accum0 = _mm_srli_si128(accum0, 4);
out_row += 4;
}
}
}
void ConvolveVertically_SSE2(const ConvolutionFilter1D::Fixed* filter_values,
int filter_length,
unsigned char* const* source_data_rows,
int begin, int end,
unsigned char* out_row, bool has_alpha) {
if (has_alpha) {
ConvolveVertically_SSE2_impl<true>(filter_values, filter_length,
source_data_rows, begin, end, out_row);
} else {
ConvolveVertically_SSE2_impl<false>(filter_values, filter_length,
source_data_rows, begin, end, out_row);
}
}
} // namespace skia

70
gfx/2d/convolverSSE2.h Normal file
View File

@ -0,0 +1,70 @@
// Copyright (c) 2006-2011 The Chromium Authors. All rights reserved.
//
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions
// are met:
// * Redistributions of source code must retain the above copyright
// notice, this list of conditions and the following disclaimer.
// * Redistributions in binary form must reproduce the above copyright
// notice, this list of conditions and the following disclaimer in
// the documentation and/or other materials provided with the
// distribution.
// * Neither the name of Google, Inc. nor the names of its contributors
// may be used to endorse or promote products derived from this
// software without specific prior written permission.
//
// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
// FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
// COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
// INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
// BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
// OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
// AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
// OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
// OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
// SUCH DAMAGE.
#ifndef SKIA_EXT_CONVOLVER_SSE_H_
#define SKIA_EXT_CONVOLVER_SSE_H_
#include "convolver.h"
#include <algorithm>
#include "skia/SkTypes.h"
namespace skia {
// Convolves horizontally along a single row. The row data is given in
// |src_data| and continues for the [begin, end) of the filter.
void ConvolveHorizontally_SSE2(const unsigned char* src_data,
int begin, int end,
const ConvolutionFilter1D& filter,
unsigned char* out_row);
// Convolves horizontally along four rows. The row data is given in
// |src_data| and continues for the [begin, end) of the filter.
// The algorithm is almost same as |ConvolveHorizontally_SSE2|. Please
// refer to that function for detailed comments.
void ConvolveHorizontally4_SSE2(const unsigned char* src_data[4],
int begin, int end,
const ConvolutionFilter1D& filter,
unsigned char* out_row[4]);
// Does vertical convolution to produce one output row. The filter values and
// length are given in the first two parameters. These are applied to each
// of the rows pointed to in the |source_data_rows| array, with each row
// being |pixel_width| wide.
//
// The output must have room for |pixel_width * 4| bytes.
void ConvolveVertically_SSE2(const ConvolutionFilter1D::Fixed* filter_values,
int filter_length,
unsigned char* const* source_data_rows,
int begin, int end,
unsigned char* out_row, bool has_alpha);
} // namespace skia
#endif // SKIA_EXT_CONVOLVER_SSE_H_

27
gfx/2d/image_operations.cpp
View File

@ -160,8 +160,7 @@ class ResizeFilter {
// for the transform is also specified.
void ComputeFilters(int src_size,
int dest_subset_lo, int dest_subset_size,
float scale, float src_support,
ConvolutionFilter1D* output);
float scale, ConvolutionFilter1D* output);
// Computes the filter value given the coordinate in filter space.
inline float ComputeFilter(float pos) {
@ -181,11 +180,6 @@ class ResizeFilter {
ImageOperations::ResizeMethod method_;
// Size of the filter support on one side only in the destination space.
// See GetFilterSupport.
float x_filter_support_;
float y_filter_support_;
// Subset of scaled destination bitmap to compute.
SkIRect out_bounds_;
@ -210,17 +204,10 @@ ResizeFilter::ResizeFilter(ImageOperations::ResizeMethod method,
float scale_y = static_cast<float>(dest_height) /
static_cast<float>(src_full_height);
x_filter_support_ = GetFilterSupport(scale_x);
y_filter_support_ = GetFilterSupport(scale_y);
// Support of the filter in source space.
float src_x_support = x_filter_support_ / scale_x;
float src_y_support = y_filter_support_ / scale_y;
ComputeFilters(src_full_width, dest_subset.fLeft, dest_subset.width(),
scale_x, src_x_support, &x_filter_);
scale_x, &x_filter_);
ComputeFilters(src_full_height, dest_subset.fTop, dest_subset.height(),
scale_y, src_y_support, &y_filter_);
scale_y, &y_filter_);
}
// TODO(egouriou): Take advantage of periods in the convolution.
@ -236,8 +223,7 @@ ResizeFilter::ResizeFilter(ImageOperations::ResizeMethod method,
// loading the factors only once outside the borders.
void ResizeFilter::ComputeFilters(int src_size,
int dest_subset_lo, int dest_subset_size,
float scale, float src_support,
ConvolutionFilter1D* output) {
float scale, ConvolutionFilter1D* output) {
int dest_subset_hi = dest_subset_lo + dest_subset_size; // [lo, hi)
// When we're doing a magnification, the scale will be larger than one. This
@ -247,6 +233,8 @@ void ResizeFilter::ComputeFilters(int src_size,
// some computations.
float clamped_scale = std::min(1.0f, scale);
float src_support = GetFilterSupport(clamped_scale) / clamped_scale;
// Speed up the divisions below by turning them into multiplies.
float inv_scale = 1.0f / scale;
@ -536,8 +524,7 @@ SkBitmap ImageOperations::ResizeBasic(const SkBitmap& source,
BGRAConvolve2D(source_subset, static_cast<int>(source.rowBytes()),
!source.isOpaque(), filter.x_filter(), filter.y_filter(),
static_cast<int>(result.rowBytes()),
static_cast<unsigned char*>(result.getPixels()),
/* sse = */ false);
static_cast<unsigned char*>(result.getPixels()));
// Preserve the "opaque" flag for use as an optimization later.
result.setAlphaType(source.alphaType());

2
gfx/2d/moz.build
View File

@ -88,6 +88,7 @@ if CONFIG['INTEL_ARCHITECTURE']:
if CONFIG['_MSC_VER'] != '1400':
SOURCES += [
'BlurSSE2.cpp',
'convolverSSE2.cpp',
'FilterProcessingSSE2.cpp',
'ImageScalingSSE2.cpp',
]
@ -97,6 +98,7 @@ if CONFIG['INTEL_ARCHITECTURE']:
SOURCES['BlurSSE2.cpp'].flags += CONFIG['SSE2_FLAGS']
SOURCES['FilterProcessingSSE2.cpp'].flags += CONFIG['SSE2_FLAGS']
SOURCES['ImageScalingSSE2.cpp'].flags += CONFIG['SSE2_FLAGS']
SOURCES['convolverSSE2.cpp'].flags += CONFIG['SSE2_FLAGS']
UNIFIED_SOURCES += [
'Blur.cpp',