michael@0: /* -*- Mode: C++; tab-width: 20; indent-tabs-mode: nil; c-basic-offset: 2 -*-
michael@0:  * This Source Code Form is subject to the terms of the Mozilla Public
michael@0:  * License, v. 2.0. If a copy of the MPL was not distributed with this
michael@0:  * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
michael@0: 
michael@0: #include "ImageScaling.h"
michael@0: #include "mozilla/Attributes.h"
michael@0: 
michael@0: #include "SSEHelpers.h"
michael@0: 
michael@0: /* The functions below use the following system for averaging 4 pixels:
michael@0:  *
michael@0:  * The first observation is that a half-adder is implemented as follows:
michael@0:  * R = S + 2C or in the case of a and b (a ^ b) + ((a & b) << 1);
michael@0:  *
michael@0:  * This can be trivially extended to three pixels by observaring that when
michael@0:  * doing (a ^ b ^ c) as the sum, the carry is simply the bitwise-or of the
michael@0:  * carries of the individual numbers, since the sum of 3 bits can only ever
michael@0:  * have a carry of one.
michael@0:  *
michael@0:  * We then observe that the average is then ((carry << 1) + sum) >> 1, or,
michael@0:  * assuming eliminating overflows and underflows, carry + (sum >> 1).
michael@0:  *
michael@0:  * We now average our existing sum with the fourth number, so we get:
michael@0:  * sum2 = (sum + d) >> 1 or (sum >> 1) + (d >> 1).
michael@0:  *
michael@0:  * We now observe that our sum has been moved into place relative to the
michael@0:  * carry, so we can now average with the carry to get the final 4 input
michael@0:  * average: avg = (sum2 + carry) >> 1;
michael@0:  *
michael@0:  * Or to reverse the proof:
michael@0:  * avg = ((sum >> 1) + carry + d >> 1) >> 1
michael@0:  * avg = ((a + b + c) >> 1 + d >> 1) >> 1
michael@0:  * avg = ((a + b + c + d) >> 2)
michael@0:  *
michael@0:  * An additional fact used in the SSE versions is the concept that we can
michael@0:  * trivially convert a rounded average to a truncated average:
michael@0:  *
michael@0:  * We have:
michael@0:  * f(a, b) = (a + b + 1) >> 1
michael@0:  *
michael@0:  * And want:
michael@0:  * g(a, b) = (a + b) >> 1
michael@0:  *
michael@0:  * Observe:
michael@0:  * ~f(~a, ~b) == ~((~a + ~b + 1) >> 1)
michael@0:  *            == ~((-a - 1 + -b - 1 + 1) >> 1)
michael@0:  *            == ~((-a - 1 + -b) >> 1)
michael@0:  *            == ~((-(a + b) - 1) >> 1)
michael@0:  *            == ~((~(a + b)) >> 1)
michael@0:  *            == (a + b) >> 1
michael@0:  *            == g(a, b)
michael@0:  */
michael@0: 
michael@0: MOZ_ALWAYS_INLINE __m128i _mm_not_si128(__m128i arg)
michael@0: {
michael@0:   __m128i minusone = _mm_set1_epi32(0xffffffff);
michael@0:   return _mm_xor_si128(arg, minusone);
michael@0: }
michael@0: 
michael@0: /* We have to pass pointers here, MSVC does not allow passing more than 3
michael@0:  * __m128i arguments on the stack. And it does not allow 16-byte aligned
michael@0:  * stack variables. This inlines properly on MSVC 2010. It does -not- inline
michael@0:  * with just the inline directive.
michael@0:  */
michael@0: MOZ_ALWAYS_INLINE __m128i avg_sse2_8x2(__m128i *a, __m128i *b, __m128i *c, __m128i *d)
michael@0: {
michael@0: #define shuf1 _MM_SHUFFLE(2, 0, 2, 0)
michael@0: #define shuf2 _MM_SHUFFLE(3, 1, 3, 1)
michael@0: 
michael@0: // This cannot be an inline function as the __Imm argument to _mm_shuffle_ps
michael@0: // needs to be a compile time constant.
michael@0: #define shuffle_si128(arga, argb, imm) \
michael@0:   _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps((arga)), _mm_castsi128_ps((argb)), (imm)));
michael@0: 
michael@0:   __m128i t = shuffle_si128(*a, *b, shuf1);
michael@0:   *b = shuffle_si128(*a, *b, shuf2);
michael@0:   *a = t;
michael@0:   t = shuffle_si128(*c, *d, shuf1);
michael@0:   *d = shuffle_si128(*c, *d, shuf2);
michael@0:   *c = t;
michael@0: 
michael@0: #undef shuf1
michael@0: #undef shuf2
michael@0: #undef shuffle_si128
michael@0: 
michael@0:   __m128i sum = _mm_xor_si128(*a, _mm_xor_si128(*b, *c));
michael@0: 
michael@0:   __m128i carry = _mm_or_si128(_mm_and_si128(*a, *b), _mm_or_si128(_mm_and_si128(*a, *c), _mm_and_si128(*b, *c)));
michael@0: 
michael@0:   sum = _mm_avg_epu8(_mm_not_si128(sum), _mm_not_si128(*d));
michael@0: 
michael@0:   return _mm_not_si128(_mm_avg_epu8(sum, _mm_not_si128(carry)));
michael@0: }
michael@0: 
michael@0: MOZ_ALWAYS_INLINE __m128i avg_sse2_4x2_4x1(__m128i a, __m128i b)
michael@0: {
michael@0:   return _mm_not_si128(_mm_avg_epu8(_mm_not_si128(a), _mm_not_si128(b)));
michael@0: }
michael@0: 
michael@0: MOZ_ALWAYS_INLINE __m128i avg_sse2_8x1_4x1(__m128i a, __m128i b)
michael@0: {
michael@0:   __m128i t = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b), _MM_SHUFFLE(3, 1, 3, 1)));
michael@0:   b = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b), _MM_SHUFFLE(2, 0, 2, 0)));
michael@0:   a = t;
michael@0: 
michael@0:   return _mm_not_si128(_mm_avg_epu8(_mm_not_si128(a), _mm_not_si128(b)));
michael@0: }
michael@0: 
michael@0: MOZ_ALWAYS_INLINE uint32_t Avg2x2(uint32_t a, uint32_t b, uint32_t c, uint32_t d)
michael@0: {
michael@0:   uint32_t sum = a ^ b ^ c;
michael@0:   uint32_t carry = (a & b) | (a & c) | (b & c);
michael@0: 
michael@0:   uint32_t mask = 0xfefefefe;
michael@0: 
michael@0:   // Not having a byte based average instruction means we should mask to avoid
michael@0:   // underflow.
michael@0:   sum = (((sum ^ d) & mask) >> 1) + (sum & d);
michael@0: 
michael@0:   return (((sum ^ carry) & mask) >> 1) + (sum & carry);
michael@0: }
michael@0: 
michael@0: // Simple 2 pixel average version of the function above.
michael@0: MOZ_ALWAYS_INLINE uint32_t Avg2(uint32_t a, uint32_t b)
michael@0: {
michael@0:   uint32_t sum = a ^ b;
michael@0:   uint32_t carry = (a & b);
michael@0: 
michael@0:   uint32_t mask = 0xfefefefe;
michael@0: 
michael@0:   return ((sum & mask) >> 1) + carry;
michael@0: }
michael@0: 
michael@0: namespace mozilla {
michael@0: namespace gfx {
michael@0: 
michael@0: void
michael@0: ImageHalfScaler::HalfImage2D_SSE2(uint8_t *aSource, int32_t aSourceStride,
michael@0:                                   const IntSize &aSourceSize, uint8_t *aDest,
michael@0:                                   uint32_t aDestStride)
michael@0: {
michael@0:   const int Bpp = 4;
michael@0: 
michael@0:   for (int y = 0; y < aSourceSize.height; y += 2) {
michael@0:     __m128i *storage = (__m128i*)(aDest + (y / 2) * aDestStride);
michael@0:     int x = 0;
michael@0:     // Run a loop depending on alignment.
michael@0:     if (!(uintptr_t(aSource + (y * aSourceStride)) % 16) &&
michael@0:         !(uintptr_t(aSource + ((y + 1) * aSourceStride)) % 16)) {
michael@0:       for (; x < (aSourceSize.width - 7); x += 8) {
michael@0:         __m128i *upperRow = (__m128i*)(aSource + (y * aSourceStride + x * Bpp));
michael@0:         __m128i *lowerRow = (__m128i*)(aSource + ((y + 1) * aSourceStride + x * Bpp));
michael@0: 
michael@0:         __m128i a = _mm_load_si128(upperRow);
michael@0:         __m128i b = _mm_load_si128(upperRow + 1);
michael@0:         __m128i c = _mm_load_si128(lowerRow);
michael@0:         __m128i d = _mm_load_si128(lowerRow + 1);
michael@0: 
michael@0:         *storage++ = avg_sse2_8x2(&a, &b, &c, &d);
michael@0:       }
michael@0:     } else if (!(uintptr_t(aSource + (y * aSourceStride)) % 16)) {
michael@0:       for (; x < (aSourceSize.width - 7); x += 8) {
michael@0:         __m128i *upperRow = (__m128i*)(aSource + (y * aSourceStride + x * Bpp));
michael@0:         __m128i *lowerRow = (__m128i*)(aSource + ((y + 1) * aSourceStride + x * Bpp));
michael@0: 
michael@0:         __m128i a = _mm_load_si128(upperRow);
michael@0:         __m128i b = _mm_load_si128(upperRow + 1);
michael@0:         __m128i c = loadUnaligned128(lowerRow);
michael@0:         __m128i d = loadUnaligned128(lowerRow + 1);
michael@0: 
michael@0:         *storage++ = avg_sse2_8x2(&a, &b, &c, &d);
michael@0:       }
michael@0:     } else if (!(uintptr_t(aSource + ((y + 1) * aSourceStride)) % 16)) {
michael@0:       for (; x < (aSourceSize.width - 7); x += 8) {
michael@0:         __m128i *upperRow = (__m128i*)(aSource + (y * aSourceStride + x * Bpp));
michael@0:         __m128i *lowerRow = (__m128i*)(aSource + ((y + 1) * aSourceStride + x * Bpp));
michael@0: 
michael@0:         __m128i a = loadUnaligned128((__m128i*)upperRow);
michael@0:         __m128i b = loadUnaligned128((__m128i*)upperRow + 1);
michael@0:         __m128i c = _mm_load_si128((__m128i*)lowerRow);
michael@0:         __m128i d = _mm_load_si128((__m128i*)lowerRow + 1);
michael@0: 
michael@0:         *storage++ = avg_sse2_8x2(&a, &b, &c, &d);
michael@0:       }
michael@0:     } else {
michael@0:       for (; x < (aSourceSize.width - 7); x += 8) {
michael@0:         __m128i *upperRow = (__m128i*)(aSource + (y * aSourceStride + x * Bpp));
michael@0:         __m128i *lowerRow = (__m128i*)(aSource + ((y + 1) * aSourceStride + x * Bpp));
michael@0: 
michael@0:         __m128i a = loadUnaligned128(upperRow);
michael@0:         __m128i b = loadUnaligned128(upperRow + 1);
michael@0:         __m128i c = loadUnaligned128(lowerRow);
michael@0:         __m128i d = loadUnaligned128(lowerRow + 1);
michael@0: 
michael@0:         *storage++ = avg_sse2_8x2(&a, &b, &c, &d);
michael@0:       }
michael@0:     }
michael@0: 
michael@0:     uint32_t *unalignedStorage = (uint32_t*)storage;
michael@0:     // Take care of the final pixels, we know there's an even number of pixels
michael@0:     // in the source rectangle. We use a 2x2 'simd' implementation for this.
michael@0:     //
michael@0:     // Potentially we only have to do this in the last row since overflowing 
michael@0:     // 8 pixels in an earlier row would appear to be harmless as it doesn't
michael@0:     // touch invalid memory. Even when reading and writing to the same surface.
michael@0:     // in practice we only do this when doing an additional downscale pass, and
michael@0:     // in this situation we have unused stride to write into harmlessly.
michael@0:     // I do not believe the additional code complexity would be worth it though.
michael@0:     for (; x < aSourceSize.width; x += 2) {
michael@0:       uint8_t *upperRow = aSource + (y * aSourceStride + x * Bpp);
michael@0:       uint8_t *lowerRow = aSource + ((y + 1) * aSourceStride + x * Bpp);
michael@0: 
michael@0:       *unalignedStorage++ = Avg2x2(*(uint32_t*)upperRow, *((uint32_t*)upperRow + 1),
michael@0:                                    *(uint32_t*)lowerRow, *((uint32_t*)lowerRow + 1));
michael@0:     }
michael@0:   }
michael@0: }
michael@0: 
michael@0: void
michael@0: ImageHalfScaler::HalfImageVertical_SSE2(uint8_t *aSource, int32_t aSourceStride,
michael@0:                                         const IntSize &aSourceSize, uint8_t *aDest,
michael@0:                                         uint32_t aDestStride)
michael@0: {
michael@0:   for (int y = 0; y < aSourceSize.height; y += 2) {
michael@0:     __m128i *storage = (__m128i*)(aDest + (y / 2) * aDestStride);
michael@0:     int x = 0;
michael@0:     // Run a loop depending on alignment.
michael@0:     if (!(uintptr_t(aSource + (y * aSourceStride)) % 16) &&
michael@0:         !(uintptr_t(aSource + ((y + 1) * aSourceStride)) % 16)) {
michael@0:       for (; x < (aSourceSize.width - 3); x += 4) {
michael@0:         uint8_t *upperRow = aSource + (y * aSourceStride + x * 4);
michael@0:         uint8_t *lowerRow = aSource + ((y + 1) * aSourceStride + x * 4);
michael@0: 
michael@0:         __m128i a = _mm_load_si128((__m128i*)upperRow);
michael@0:         __m128i b = _mm_load_si128((__m128i*)lowerRow);
michael@0: 
michael@0:         *storage++ = avg_sse2_4x2_4x1(a, b);
michael@0:       }
michael@0:     } else if (!(uintptr_t(aSource + (y * aSourceStride)) % 16)) {
michael@0:       // This line doesn't align well.
michael@0:       for (; x < (aSourceSize.width - 3); x += 4) {
michael@0:         uint8_t *upperRow = aSource + (y * aSourceStride + x * 4);
michael@0:         uint8_t *lowerRow = aSource + ((y + 1) * aSourceStride + x * 4);
michael@0: 
michael@0:         __m128i a = _mm_load_si128((__m128i*)upperRow);
michael@0:         __m128i b = loadUnaligned128((__m128i*)lowerRow);
michael@0: 
michael@0:         *storage++ = avg_sse2_4x2_4x1(a, b);
michael@0:       }
michael@0:     } else if (!(uintptr_t(aSource + (y * aSourceStride)) % 16)) {
michael@0:       for (; x < (aSourceSize.width - 3); x += 4) {
michael@0:         uint8_t *upperRow = aSource + (y * aSourceStride + x * 4);
michael@0:         uint8_t *lowerRow = aSource + ((y + 1) * aSourceStride + x * 4);
michael@0: 
michael@0:         __m128i a = loadUnaligned128((__m128i*)upperRow);
michael@0:         __m128i b = _mm_load_si128((__m128i*)lowerRow);
michael@0: 
michael@0:         *storage++ = avg_sse2_4x2_4x1(a, b);
michael@0:       }
michael@0:     } else {
michael@0:       for (; x < (aSourceSize.width - 3); x += 4) {
michael@0:         uint8_t *upperRow = aSource + (y * aSourceStride + x * 4);
michael@0:         uint8_t *lowerRow = aSource + ((y + 1) * aSourceStride + x * 4);
michael@0: 
michael@0:         __m128i a = loadUnaligned128((__m128i*)upperRow);
michael@0:         __m128i b = loadUnaligned128((__m128i*)lowerRow);
michael@0: 
michael@0:         *storage++ = avg_sse2_4x2_4x1(a, b);
michael@0:       }
michael@0:     }
michael@0: 
michael@0:     uint32_t *unalignedStorage = (uint32_t*)storage;
michael@0:     // Take care of the final pixels, we know there's an even number of pixels
michael@0:     // in the source rectangle.
michael@0:     //
michael@0:     // Similar overflow considerations are valid as in the previous function.
michael@0:     for (; x < aSourceSize.width; x++) {
michael@0:       uint8_t *upperRow = aSource + (y * aSourceStride + x * 4);
michael@0:       uint8_t *lowerRow = aSource + ((y + 1) * aSourceStride + x * 4);
michael@0: 
michael@0:       *unalignedStorage++ = Avg2(*(uint32_t*)upperRow, *(uint32_t*)lowerRow);
michael@0:     }
michael@0:   }
michael@0: }
michael@0: 
michael@0: void
michael@0: ImageHalfScaler::HalfImageHorizontal_SSE2(uint8_t *aSource, int32_t aSourceStride,
michael@0:                                           const IntSize &aSourceSize, uint8_t *aDest,
michael@0:                                           uint32_t aDestStride)
michael@0: {
michael@0:   for (int y = 0; y < aSourceSize.height; y++) {
michael@0:     __m128i *storage = (__m128i*)(aDest + (y * aDestStride));
michael@0:     int x = 0;
michael@0:     // Run a loop depending on alignment.
michael@0:     if (!(uintptr_t(aSource + (y * aSourceStride)) % 16)) {
michael@0:       for (; x < (aSourceSize.width - 7); x += 8) {
michael@0:         __m128i* pixels = (__m128i*)(aSource + (y * aSourceStride + x * 4));
michael@0: 
michael@0:         __m128i a = _mm_load_si128(pixels);
michael@0:         __m128i b = _mm_load_si128(pixels + 1);
michael@0: 
michael@0:         *storage++ = avg_sse2_8x1_4x1(a, b);
michael@0:       }
michael@0:     } else {
michael@0:       for (; x < (aSourceSize.width - 7); x += 8) {
michael@0:         __m128i* pixels = (__m128i*)(aSource + (y * aSourceStride + x * 4));
michael@0: 
michael@0:         __m128i a = loadUnaligned128(pixels);
michael@0:         __m128i b = loadUnaligned128(pixels + 1);
michael@0: 
michael@0:         *storage++ = avg_sse2_8x1_4x1(a, b);
michael@0:       }
michael@0:     }
michael@0: 
michael@0:     uint32_t *unalignedStorage = (uint32_t*)storage;
michael@0:     // Take care of the final pixels, we know there's an even number of pixels
michael@0:     // in the source rectangle.
michael@0:     //
michael@0:     // Similar overflow considerations are valid as in the previous function.
michael@0:     for (; x < aSourceSize.width; x += 2) {
michael@0:       uint32_t *pixels = (uint32_t*)(aSource + (y * aSourceStride + x * 4));
michael@0: 
michael@0:       *unalignedStorage++ = Avg2(*pixels, *(pixels + 1));
michael@0:     }
michael@0:   }
michael@0: }
michael@0: 
michael@0: }
michael@0: }