fix msse3

Makefile.am

@@ -22,8 +22,7 @@ i3lock_CFLAGS = \
 	$(XKBCOMMON_CFLAGS) \
 	$(CAIRO_CFLAGS) \
 	$(CODE_COVERAGE_CFLAGS) \
-	$(X11_CFLAGS) \
-	$(SIMD_CFLAGS)
+	$(X11_CFLAGS)
 
 i3lock_CPPFLAGS = \
 	$(AM_CPPFLAGS) \

blur.c

@@ -1,6 +1,6 @@
 /*
- * Copyright © 2008 Kristian Høgsberg
- * Copyright © 2009 Chris Wilson
+ * Copyright © 2008 Kristian Høgsberg
+ * Copyright © 2009 Chris Wilson
  *
  * Permission to use, copy, modify, distribute, and sell this software and its
  * documentation for any purpose is hereby granted without fee, provided that
@@ -24,12 +24,15 @@
 #include <math.h>
 #include "blur.h"
 
-/* Performs a simple 2D Gaussian blur of standard devation @sigma surface @surface. */
+#define ARRAY_LENGTH(a) (sizeof (a) / sizeof (a)[0])
+
+/* Performs a simple 2D Gaussian blur of radius @radius on surface @surface. */
 void
-blur_image_surface (cairo_surface_t *surface, int sigma)
+blur_image_surface (cairo_surface_t *surface, int radius)
 {
     cairo_surface_t *tmp;
     int width, height;
+//    int src_stride, dst_stride;
     uint32_t *src, *dst;
 
     if (cairo_surface_status (surface))
@@ -61,85 +64,89 @@ blur_image_surface (cairo_surface_t *surface, int sigma)
     return;
 
     src = (uint32_t*)cairo_image_surface_get_data (surface);
+//    src_stride = cairo_image_surface_get_stride (surface);
+
     dst = (uint32_t*)cairo_image_surface_get_data (tmp);
+//    dst_stride = cairo_image_surface_get_stride (tmp);
+
+    //blur_impl_naive(src, dst, width, height, src_stride, dst_stride, 10000);
+    //blur_impl_sse2(src, dst, width, height, 4.5);
+    blur_impl_ssse3(src, dst, width, height, 4.5);
 
-    // according to a paper by Peter Kovesi [1], box filter of width w, equals to Gaussian blur of following sigma:
-    // σ_av = sqrt((w*w-1)/12)
-    // for our 8x8 filter we have σ_av = 2.0.
-    // applying the same Gaussian filter n times results in σ_n = sqrt(n*σ_av*σ_av) [2]
-    // after some trivial math, we arrive at n = ((σ_d)/(σ_av))^2
-    // since it's a box blur filter, n >= 3
-    //
-    // [1]: http://www.peterkovesi.com/papers/FastGaussianSmoothing.pdf
-    // [2]: https://en.wikipedia.org/wiki/Gaussian_blur#Mathematics
-    
-    int n = lrintf((sigma*sigma)/(SIGMA_AV*SIGMA_AV));
-    if (n < 3) n = 3;
-
-    for (int i = 0; i < n; i++)
-    {
-        // horizontal pass includes image transposition:
-        // instead of writing pixel src[x] to dst[x],
-        // we write it to transposed location.
-        // (to be exact: dst[height * current_column + current_row])
-#ifdef __SSE2__
-        blur_impl_horizontal_pass_sse2(src, dst, width, height);
-        blur_impl_horizontal_pass_sse2(dst, src, height, width);
-#else
-        blur_impl_horizontal_pass_generic(src, dst, width, height);
-        blur_impl_horizontal_pass_generic(dst, src, height, width);
-#endif
-    }
- 
     cairo_surface_destroy (tmp);
     cairo_surface_flush (surface);
     cairo_surface_mark_dirty (surface);
 }
 
-void blur_impl_horizontal_pass_generic(uint32_t *src, uint32_t *dst, int width, int height) {
-    for (int row = 0; row < height; row++) {
-        for (int column = 0; column < width; column++, src++) {
-            uint32_t rgbaIn[KERNEL_SIZE];
-
-            // handle borders
-            int leftBorder = column < HALF_KERNEL;
-            int rightBorder = column > width - HALF_KERNEL;
-            int i = 0;
-            if (leftBorder) {
-                // for kernel size 7x7 and column == 0, we have:
-                // x x x P0 P1 P2 P3
-                // first loop mirrors P{0..3} to fill x's,
-                // second one loads P{0..3}
-                for (; i < HALF_KERNEL - column; i++)
-                    rgbaIn[i] = *(src + (HALF_KERNEL - i));
-                for (; i < KERNEL_SIZE; i++)
-                    rgbaIn[i] = *(src - (HALF_KERNEL - i));
-            } else if (rightBorder) {
-                for (; i < width - column; i++)
-                    rgbaIn[i] = *(src + i);
-                for (int k = 0; i < KERNEL_SIZE; i++, k++)
-                    rgbaIn[i] = *(src - k);
-            } else {
-                for (; i < KERNEL_SIZE; i++)
-                    rgbaIn[i] = *(src + i - HALF_KERNEL);
-            }
-
-            uint32_t acc[4] = {0};
-
-            for (i = 0; i < KERNEL_SIZE; i++) {
-                acc[0] += (rgbaIn[i] & 0xFF000000) >> 24;
-                acc[1] += (rgbaIn[i] & 0x00FF0000) >> 16;
-                acc[2] += (rgbaIn[i] & 0x0000FF00) >> 8;
-                acc[3] += (rgbaIn[i] & 0x000000FF) >> 0;
-            }
-
-            for(i = 0; i < 4; i++)
-                acc[i] *= 1.0/KERNEL_SIZE;
-
-            *(dst + height * column + row) = (acc[0] << 24) |
-                                             (acc[1] << 16) |
-                                             (acc[2] << 8 ) |
-                                             (acc[3] << 0);
+void blur_impl_naive(uint32_t* _src, uint32_t* _dst, int width, int height, int src_stride, int dst_stride, int radius)
+{
+    int x, y, z, w;
+    uint32_t *s, *d, a, p;
+    int i, j, k;
+    uint8_t kernel[17];
+    const int size = ARRAY_LENGTH (kernel);
+    const int half = size / 2;
+
+    uint8_t *src = (uint8_t*)_src;
+    uint8_t *dst = (uint8_t*)_dst;
+
+    a = 0;
+    for (i = 0; i < size; i++) {
+    double f = i - half;
+    a += kernel[i] = exp (- f * f / 30.0) * 80;
+    }
+
+    /* Horizontally blur from surface -> tmp */
+    for (i = 0; i < height; i++) {
+    s = (uint32_t *) (src + i * src_stride);
+    d = (uint32_t *) (dst + i * dst_stride);
+    for (j = 0; j < width; j++) {
+        if (radius < j && j < width - radius) {
+        d[j] = s[j];
+        continue;
+        }
+
+        x = y = z = w = 0;
+        for (k = 0; k < size; k++) {
+        if (j - half + k < 0 || j - half + k >= width)
+            continue;
+
+        p = s[j - half + k];
+
+        x += ((p >> 24) & 0xff) * kernel[k];
+        y += ((p >> 16) & 0xff) * kernel[k];
+        z += ((p >>  8) & 0xff) * kernel[k];
+        w += ((p >>  0) & 0xff) * kernel[k];
         }
+        d[j] = (x / a << 24) | (y / a << 16) | (z / a << 8) | w / a;
+    }
+    }
+
+    /* Then vertically blur from tmp -> surface */
+    for (i = 0; i < height; i++) {
+    s = (uint32_t *) (dst + i * dst_stride);
+    d = (uint32_t *) (src + i * src_stride);
+    for (j = 0; j < width; j++) {
+        if (radius <= i && i < height - radius) {
+        d[j] = s[j];
+        continue;
+        }
+
+        x = y = z = w = 0;
+        for (k = 0; k < size; k++) {
+        if (i - half + k < 0 || i - half + k >= height)
+            continue;
+
+        s = (uint32_t *) (dst + (i - half + k) * dst_stride);
+        p = s[j];
+
+        x += ((p >> 24) & 0xff) * kernel[k];
+        y += ((p >> 16) & 0xff) * kernel[k];
+        z += ((p >>  8) & 0xff) * kernel[k];
+        w += ((p >>  0) & 0xff) * kernel[k];
+        }
+        d[j] = (x / a << 24) | (y / a << 16) | (z / a << 8) | w / a;
+    }
     }
 }
+

blur.h

@@ -4,13 +4,12 @@
 #include <stdint.h>
 #include <cairo.h>
 
-#define KERNEL_SIZE 7 
-#define SIGMA_AV 2
-#define HALF_KERNEL KERNEL_SIZE / 2
-
-void blur_image_surface(cairo_surface_t *surface, int sigma);
-void blur_impl_horizontal_pass_sse2(uint32_t *src, uint32_t *dst, int width, int height);
-void blur_impl_horizontal_pass_generic(uint32_t *src, uint32_t *dst, int width, int height);
+void blur_image_surface (cairo_surface_t *surface, int radius);
+void blur_impl_naive(uint32_t* src, uint32_t* dst, int width, int height, int src_stride, int dst_stride, int radius);
+void blur_impl_sse2(uint32_t* src, uint32_t* dst, int width, int height, float sigma);
+void blur_impl_horizontal_pass_sse2(uint32_t *src, uint32_t *dst, float *kernel, int width, int height);
+void blur_impl_ssse3(uint32_t* src, uint32_t* dst, int width, int height, float sigma);
+void blur_impl_horizontal_pass_ssse3(uint32_t *src, uint32_t *dst, int8_t *kernel, int width, int height);
 
 #endif
 

blur_simd.c

@@ -8,76 +8,261 @@
  */
 
 #include "blur.h"
+#include <math.h>
 #include <xmmintrin.h>
+#include <tmmintrin.h>
 
 #include <stdio.h>
 
-// number of xmm registers needed to store input pixels for given kernel size
+#define ALIGN16 __attribute__((aligned(16)))
+#define KERNEL_SIZE 15 
+#define HALF_KERNEL KERNEL_SIZE / 2
+
+// number of xmm registers needed to store 
+// input pixels for given kernel size
 #define REGISTERS_CNT (KERNEL_SIZE + 4/2) / 4
 
-void blur_impl_horizontal_pass_sse2(uint32_t *src, uint32_t *dst, int width, int height) {
-    printf("height %d by width %d\n", height, width);
+// scaling factor for kernel coefficients.
+// higher values cause desaturation.
+// used in SSSE3 implementation.
+#define SCALE_FACTOR 7
+
+void blur_impl_sse2(uint32_t *src, uint32_t *dst, int width, int height, float sigma) {
+    // prepare kernel
+    float kernel[KERNEL_SIZE];
+    float coeff = 1.0 / sqrtf(2 * M_PI * sigma * sigma), sum = 0;
+
+    for (int i = 0; i < KERNEL_SIZE; i++) {
+        float x = HALF_KERNEL - i;
+        kernel[i] = coeff * expf(-x * x / (2.0 * sigma * sigma));
+        sum += kernel[i];
+    }
+
+    // normalize kernel
+    for (int i = 0; i < KERNEL_SIZE; i++)
+        kernel[i] /= sum;
+
+    // horizontal pass includes image transposition:
+    // instead of writing pixel src[x] to dst[x],
+    // we write it to transposed location.
+    // (to be exact: dst[height * current_column + current_row])
+    blur_impl_horizontal_pass_sse2(src, dst, kernel, width, height);
+    blur_impl_horizontal_pass_sse2(dst, src, kernel, height, width);
+}
+
+void blur_impl_horizontal_pass_sse2(uint32_t *src, uint32_t *dst, float *kernel, int width, int height) {
     for (int row = 0; row < height; row++) {
-        printf("row %d\n", row);
         for (int column = 0; column < width; column++, src++) {
-            printf("\tcol %d\n", column);
             __m128i rgbaIn[REGISTERS_CNT];
 
             // handle borders
             int leftBorder = column < HALF_KERNEL;
-            int rightBorder = column > (width - HALF_KERNEL);
-            // +1 to make memory checkers not complain
-            uint32_t _rgbaIn[KERNEL_SIZE + 1] __attribute__((aligned(16)));
-            int i = 0;
-            if (leftBorder) {
-                // for kernel size 8x8 and column == 0, we have:
-                // x x x x P0 P1 P2 P3
-                // first loop mirrors P{0..4} to fill x's,
-                // second one loads P{0..4}
-                for (; i < HALF_KERNEL - column; i++)
-                    _rgbaIn[i] = *(src + (HALF_KERNEL - i));
-                for (; i < KERNEL_SIZE; i++)
-                    _rgbaIn[i] = *(src - (HALF_KERNEL - i));
+            int rightBorder = column > width - HALF_KERNEL;
+            if (leftBorder || rightBorder) {
+                uint32_t _rgbaIn[KERNEL_SIZE] ALIGN16;
+                int i = 0;
+                if (leftBorder) {
+                    // for kernel size 7x7 and column == 0, we have:
+                    // x x x P0 P1 P2 P3
+                    // first loop mirrors P{0..3} to fill x's,
+                    // second one loads P{0..3}
+                    for (; i < HALF_KERNEL - column; i++)
+                        _rgbaIn[i] = *(src + (HALF_KERNEL - i));
+                    for (; i < KERNEL_SIZE; i++)
+                        _rgbaIn[i] = *(src - (HALF_KERNEL - i));
+                } else {
+                    for (; i < width - column; i++)
+                        _rgbaIn[i] = *(src + i);
+                    for (int k = 0; i < KERNEL_SIZE; i++, k++)
+                        _rgbaIn[i] = *(src - k);
+                }
 
                 for (int k = 0; k < REGISTERS_CNT; k++)
                     rgbaIn[k] = _mm_load_si128((__m128i*)(_rgbaIn + 4*k));
-            } else if (rightBorder) {
-                for (; i < width - column; i++)
-                    _rgbaIn[i] = *(src + i);
-                for (int k = 0; i < KERNEL_SIZE; i++, k++)
-                    _rgbaIn[i] = *(src - k);
-
+            } else {
                 for (int k = 0; k < REGISTERS_CNT; k++)
+                    rgbaIn[k] = _mm_loadu_si128((__m128i*)(src + 4*k - HALF_KERNEL));
+            }
+
+            // unpack each pixel, convert to float,
+            // multiply by corresponding kernel value
+            // and add to accumulator
+            __m128i tmp;
+            __m128i zero = _mm_setzero_si128();
+            __m128 rgba_ps;
+            __m128 acc = _mm_setzero_ps();
+            int counter = 0;
+
+            for (int i = 0; i < 3; i++)
+            {
+                tmp = _mm_unpacklo_epi8(rgbaIn[i], zero);
+                rgba_ps = _mm_cvtepi32_ps(_mm_unpacklo_epi16(tmp, zero));
+                acc = _mm_add_ps(acc, _mm_mul_ps(rgba_ps, _mm_set1_ps(kernel[counter++])));
+                rgba_ps = _mm_cvtepi32_ps(_mm_unpackhi_epi16(tmp, zero));
+                acc = _mm_add_ps(acc, _mm_mul_ps(rgba_ps, _mm_set1_ps(kernel[counter++])));
+
+                tmp = _mm_unpackhi_epi8(rgbaIn[i], zero);
+                rgba_ps = _mm_cvtepi32_ps(_mm_unpacklo_epi16(tmp, zero));
+                acc = _mm_add_ps(acc, _mm_mul_ps(rgba_ps, _mm_set1_ps(kernel[counter++])));
+                rgba_ps = _mm_cvtepi32_ps(_mm_unpackhi_epi16(tmp, zero));
+                acc = _mm_add_ps(acc, _mm_mul_ps(rgba_ps, _mm_set1_ps(kernel[counter++])));
+            }
+
+            tmp = _mm_unpacklo_epi8(rgbaIn[3], zero);
+            rgba_ps = _mm_cvtepi32_ps(_mm_unpacklo_epi16(tmp, zero));
+            acc = _mm_add_ps(acc, _mm_mul_ps(rgba_ps, _mm_set1_ps(kernel[counter++])));
+            rgba_ps = _mm_cvtepi32_ps(_mm_unpackhi_epi16(tmp, zero));
+            acc = _mm_add_ps(acc, _mm_mul_ps(rgba_ps, _mm_set1_ps(kernel[counter++])));
+
+            tmp = _mm_unpackhi_epi8(rgbaIn[3], zero);
+            rgba_ps = _mm_cvtepi32_ps(_mm_unpacklo_epi16(tmp, zero));
+            acc = _mm_add_ps(acc, _mm_mul_ps(rgba_ps, _mm_set1_ps(kernel[counter++])));
+
+            __m128i rgbaOut = _mm_cvtps_epi32(acc);
+            rgbaOut = _mm_packs_epi32(rgbaOut, zero);
+            rgbaOut = _mm_packus_epi16(rgbaOut, zero);
+            *(dst + height * column + row) = _mm_cvtsi128_si32(rgbaOut);
+        }
+    }
+}
+
+void blur_impl_ssse3(uint32_t *src, uint32_t *dst, int width, int height, float sigma) {
+    // prepare kernel
+    float kernelf[KERNEL_SIZE];
+    int8_t kernel[KERNEL_SIZE + 1];
+    float coeff = 1.0 / sqrtf(2 * M_PI * sigma * sigma), sum = 0;
+
+    for (int i = 0; i < KERNEL_SIZE; i++) {
+        float x = HALF_KERNEL - i;
+        kernelf[i] = coeff * expf(-x * x / (2.0 * sigma * sigma));
+        sum += kernelf[i];
+    }
+
+    // normalize kernel
+    for (int i = 0; i < KERNEL_SIZE; i++)
+        kernelf[i] /= sum;
+
+    // round to nearest integer and convert to int
+    for (int i = 0; i < KERNEL_SIZE; i++)
+        kernel[i] = (int8_t)rintf(kernelf[i] * (1 << SCALE_FACTOR));
+    kernel[KERNEL_SIZE] = 0;
+
+    // horizontal pass includes image transposition:
+    // instead of writing pixel src[x] to dst[x],
+    // we write it to transposed location.
+    // (to be exact: dst[height * current_column + current_row])
+    blur_impl_horizontal_pass_ssse3(src, dst, kernel, width, height);
+    blur_impl_horizontal_pass_ssse3(dst, src, kernel, height, width);
+}
+
+
+void blur_impl_horizontal_pass_ssse3(uint32_t *src, uint32_t *dst, int8_t *kernel, int width, int height) {
+    uint32_t* o_src = src;
+    __m128i _kern = _mm_loadu_si128((__m128i*)kernel);
+    __m128i rgbaIn[REGISTERS_CNT];
+
+    for (int row = 0; row < height; row++) {
+        for (int column = 0; column < width; column++, src++) {
+            uint32_t _rgbaIn[KERNEL_SIZE + 1] ALIGN16;
+#if 0
+            for (int j = 0; j < KERNEL_SIZE; ++j) {
+                printf("_rgbaIn[%d]: %p->%p\n", j, &_rgbaIn[j], &_rgbaIn[j] + 1);
+            }
+#endif
+            // handle borders
+            int leftBorder = column < HALF_KERNEL;
+            int rightBorder = column > width - HALF_KERNEL;
+            if (leftBorder || rightBorder) {
+                int i = 0;
+                if (leftBorder) {
+                    // for kernel size 7x7 and column == 0, we have:
+                    // x x x P0 P1 P2 P3
+                    // first loop mirrors P{0..3} to fill x's,
+                    // second one loads P{0..3}
+                    for (; i < HALF_KERNEL - column; i++)
+                        _rgbaIn[i] = *(src + (HALF_KERNEL - i));
+                    for (; i < KERNEL_SIZE; i++)
+                        _rgbaIn[i] = *(src - (HALF_KERNEL - i));
+                } else {
+                    for (; i < width - column; i++)
+                        _rgbaIn[i] = *(src + i);
+                    for (int k = 0; i < KERNEL_SIZE; i++, k++)
+                        _rgbaIn[i] = *(src - k);
+                }
+
+                for (int k = 0; k < REGISTERS_CNT; k++) {
+#if 0
+                    printf("K: %d; p: %p, p+4*k: %p, end of p: %p\n", k, _rgbaIn, _rgbaIn+4*k, ((__m128i*) (_rgbaIn +4*k)) + 1);
+#endif
                     rgbaIn[k] = _mm_load_si128((__m128i*)(_rgbaIn + 4*k));
+                }
             } else {
                 for (int k = 0; k < REGISTERS_CNT; k++) {
-                    printf("\t\tk: %d %p\n", k, src);
-                    rgbaIn[k] = _mm_load_si128((__m128i*)(src + 4*k - HALF_KERNEL));
+                    if ((long long)(((__m128i*) src + 4*k - HALF_KERNEL) + 1) > (long long)((o_src + (width * height)))) break;
+#if 0
+                    printf("K: %d; p: %p -> %p\n", k, src+4*k - HALF_KERNEL, ((__m128i*) (src +4*k - HALF_KERNEL)) + 1);
+                    printf("%p->%p, %p->%p (%ld)\n", (__m128i*) src + 4*k - HALF_KERNEL, ((__m128i*) src + 4*k - HALF_KERNEL) + 1, o_src, o_src + (width * height), o_src + (width * height) - src);
+#endif
+                    rgbaIn[k] = _mm_loadu_si128((__m128i*)(src + 4*k - HALF_KERNEL));
                 }
             }
 
+            // basis of this implementation is _mm_maddubs_epi16 (aka pmaddubsw).
+            // 'rgba' holds 16 unsigned bytes, so 4 pixels.
+            // 'kern' holds 16 signed bytes kernel values multiplied by (1 << SCALE_FACTOR).
+            // before multiplication takes place, vectors need to be prepared:
+            // 'rgba' is shuffled from R1B1G1A1...R4B4G4A4 to R1R2R3R4...A1A2A3A4
+            // 'kern' is shuffled from w1w2w3w4...w13w14w15w16 to w1w2w3w4 repeated 4 times
+            // then we call _mm_maddubs_epi16 and we get:
+            // --------------------------------------------------------------------------------------
+            // | R1*w1 + R2*w2 | R3*w3 + R4*w4 | G1*w1 + G2*w2 | G3*w3 + G4*w4 | repeat for B and A |
+            // --------------------------------------------------------------------------------------
+            // each 'rectangle' is a 16-byte signed int.
+            // then we repeat the process for the rest of input pixels,
+            // call _mm_hadds_epi16 to add adjacent ints and shift right to scale by SCALE_FACTOR.
+
+            __m128i rgba, kern;
             __m128i zero = _mm_setzero_si128();
             __m128i acc = _mm_setzero_si128();
 
-            acc = _mm_add_epi16(acc, _mm_unpacklo_epi8(rgbaIn[0], zero));
-            acc = _mm_add_epi16(acc, _mm_unpackhi_epi8(rgbaIn[0], zero));
-            acc = _mm_add_epi16(acc, _mm_unpacklo_epi8(rgbaIn[1], zero));
+            const __m128i rgba_shuf_mask = _mm_setr_epi8(0, 4, 8,  12,
+                                                         1, 5, 9,  13,
+                                                         2, 6, 10, 14,
+                                                         3, 7, 11, 15);
 
-            // kernel size equals to 7, but we can only load multiples of 4 pixels
-            // we have to set 8th pixel to zero
-/*
-            acc = _mm_add_epi16(acc, _mm_andnot_si128(_mm_set_epi32(0xFFFFFFFF, 0xFFFFFFFF, 0, 0), 
-                                                      _mm_unpackhi_epi8(rgbaIn[1], zero)));
-            acc = _mm_add_epi32(_mm_unpacklo_epi16(acc, zero), 
-                                _mm_unpackhi_epi16(acc, zero));
-*/
-
-            // multiplication is significantly faster than division
-            acc = _mm_cvtps_epi32(_mm_mul_ps(_mm_cvtepi32_ps(acc),
-                                             _mm_set1_ps(1.0/KERNEL_SIZE)));
-
-            *(dst + height * column + row) = 
-                _mm_cvtsi128_si32(_mm_packus_epi16(_mm_packs_epi32(acc, zero), zero));
+            const __m128i kern_shuf_mask = _mm_setr_epi8(0, 1, 2, 3,
+                                                         0, 1, 2, 3,
+                                                         0, 1, 2, 3,
+                                                         0, 1, 2, 3);
+
+            rgba = _mm_shuffle_epi8(rgbaIn[0], rgba_shuf_mask);
+            kern = _mm_shuffle_epi8(_kern, kern_shuf_mask);
+            acc = _mm_adds_epi16(acc, _mm_maddubs_epi16(rgba, kern));
+
+            rgba = _mm_shuffle_epi8(rgbaIn[1], rgba_shuf_mask);
+            kern = _mm_shuffle_epi8(_mm_srli_si128(_kern, 4), kern_shuf_mask);
+            acc = _mm_adds_epi16(acc, _mm_maddubs_epi16(rgba, kern));
+
+            rgba = _mm_shuffle_epi8(rgbaIn[2], rgba_shuf_mask);
+            kern = _mm_shuffle_epi8(_mm_srli_si128(_kern, 8), kern_shuf_mask);
+            acc = _mm_adds_epi16(acc, _mm_maddubs_epi16(rgba, kern));
+
+            rgba = _mm_shuffle_epi8(rgbaIn[3], rgba_shuf_mask);
+            kern = _mm_shuffle_epi8(_mm_srli_si128(_kern, 12), kern_shuf_mask);
+            acc = _mm_adds_epi16(acc, _mm_maddubs_epi16(rgba, kern));
+
+            acc = _mm_hadds_epi16(acc, zero);
+            acc = _mm_srai_epi16(acc, SCALE_FACTOR);
+
+            // Cairo sets alpha channel to 255
+            // (or -1, depending how you look at it)
+            // this quickly overflows accumulator,
+            // and alpha is calculated completely wrong.
+            // I assume most people don't use semi-transparent
+            // lock screen images, so no one will mind if we
+            // 'correct it' by setting alpha to 255.
+            *(dst + height * column + row) =
+                _mm_cvtsi128_si32(_mm_packus_epi16(acc, zero)) | 0xFF000000;
         }
     }
 }

configure.ac

@@ -89,7 +89,6 @@ PKG_CHECK_MODULES([XKBCOMMON], [xkbcommon xkbcommon-x11])
 PKG_CHECK_MODULES([CAIRO], [cairo])
 PKG_CHECK_MODULES([X11], [x11])
 
-SIMD_CFLAGS=" -funroll-loops -msse2 -std=c99 -pipe -O2"
 
 # Checks for programs.
 AC_PROG_AWK
@@ -102,6 +101,10 @@ AC_PROG_LN_S
 AM_PROG_AR
 
 AX_FLAGS_WARN_ALL
+AX_APPEND_FLAG([-msse4.1], [AM_CFLAGS])
+AX_APPEND_FLAG([-O2], [AM_CFLAGS])
+AX_APPEND_FLAG([-funroll-loops], [AM_CFLAGS])
+AX_APPEND_FLAG([-std=gnu99], [AM_CFLAGS])
 AX_CHECK_COMPILE_FLAG([-Wunused-value], [AX_APPEND_FLAG([-Wunused-value], [AM_CFLAGS])])
 AC_SUBST(AM_CFLAGS)
 

i3lock.c

@@ -1199,7 +1199,7 @@ int main(int argc, char *argv[]) {
                         arg++;
 
                     if (strlen(arg) != 8 || sscanf(arg, "%08[0-9a-fA-F]", ringwrongcolor) != 1)
-                        errx(1, "ringwrongcolor is invalid, color must be given in r-byte format: rrggbb\n");
+                        errx(1, "ringwrongcolor is invalid, color must be given in 4-byte format: rrggbb\n");
                 }
                 else if (strcmp(longopts[longoptind].name, "ringcolor") == 0) {
                     char *arg = optarg;
@@ -1570,6 +1570,7 @@ int main(int argc, char *argv[]) {
     if (!load_keymap())
         errx(EXIT_FAILURE, "Could not load keymap");
 
+
     const char *locale = getenv("LC_ALL");
     if (!locale || !*locale)
         locale = getenv("LC_CTYPE");
@@ -1587,6 +1588,7 @@ int main(int argc, char *argv[]) {
     load_compose_table(locale);
 #endif
 
+
     screen = xcb_setup_roots_iterator(xcb_get_setup(conn)).data;
 
     randr_init(&randr_base, screen->root);

m4/ax_check_enable_debug.m4

@@ -76,11 +76,11 @@ AC_DEFUN([AX_CHECK_ENABLE_DEBUG],[
     AS_CASE([$enable_debug],
       [yes],[
 	AC_MSG_RESULT(yes)
-	CFLAGS="${CFLAGS} -g -O0"
-	CXXFLAGS="${CXXFLAGS} -g -O0"
-	FFLAGS="${FFLAGS} -g -O0"
-	FCFLAGS="${FCFLAGS} -g -O0"
-	OBJCFLAGS="${OBJCFLAGS} -g -O0"
+	CFLAGS="${CFLAGS} -g -O2"
+	CXXFLAGS="${CXXFLAGS} -g -O2"
+	FFLAGS="${FFLAGS} -g -O2"
+	FCFLAGS="${FCFLAGS} -g -O2"
+	OBJCFLAGS="${OBJCFLAGS} -g -O2"
       ],
       [info],[
 	AC_MSG_RESULT(info)

m4/ax_code_coverage.m4

@@ -141,8 +141,8 @@ AC_DEFUN([AX_CODE_COVERAGE],[
 
 		dnl Build the code coverage flags
 		CODE_COVERAGE_CPPFLAGS="-DNDEBUG"
-		CODE_COVERAGE_CFLAGS="-O0 -g -fprofile-arcs -ftest-coverage"
-		CODE_COVERAGE_CXXFLAGS="-O0 -g -fprofile-arcs -ftest-coverage"
+		CODE_COVERAGE_CFLAGS="-O2 -g -fprofile-arcs -ftest-coverage"
+		CODE_COVERAGE_CXXFLAGS="-O2 -g -fprofile-arcs -ftest-coverage"
 		CODE_COVERAGE_LDFLAGS="-lgcov"
 
 		AC_SUBST([CODE_COVERAGE_CPPFLAGS])

unlock_indicator.c

@@ -830,3 +830,4 @@ void start_time_redraw_tick(struct ev_loop* main_loop) {
         ev_periodic_start(main_loop, time_redraw_tick);
     }
 }
+