diff options
Diffstat (limited to 'recipes/xorg-lib/pixman/over-n-8-0565.patch')
| -rw-r--r-- | recipes/xorg-lib/pixman/over-n-8-0565.patch | 231 |
1 files changed, 231 insertions, 0 deletions
diff --git a/recipes/xorg-lib/pixman/over-n-8-0565.patch b/recipes/xorg-lib/pixman/over-n-8-0565.patch new file mode 100644 index 0000000000..3911068d94 --- /dev/null +++ b/recipes/xorg-lib/pixman/over-n-8-0565.patch @@ -0,0 +1,231 @@ +From de2221a32d0b6628116565563f7b4ccd0a44e8b6 Mon Sep 17 00:00:00 2001 +From: Siarhei Siamashka <siarhei.siamashka@nokia.com> +Date: Thu, 04 Mar 2010 23:20:25 +0000 +Subject: ARM: added 'armv6_composite_over_n_8_0565' fast path + +Provides ~3x performance improvement when working with +data in L1 cache and memory. This fast path is important +for fonts rendering when using 16bpp desktop. + +Microbenchmark from N800 (ARM11 @ 400MHz), measured in MPix/s: + +before: + + over_n_8_0565 = L1: 2.99 M: 2.86 + +after: + + over_n_8_0565 = L1: 9.07 M: 8.05 +--- +diff --git a/pixman/pixman-arm-simd.c b/pixman/pixman-arm-simd.c +index 09a2888..c375c01 100644 +--- a/pixman/pixman-arm-simd.c ++++ b/pixman/pixman-arm-simd.c +@@ -419,6 +419,193 @@ arm_composite_over_n_8_8888 (pixman_implementation_t * impl, + } + } + ++#if defined(__ARM_EABI__) && defined(__linux__) ++/* ++ * ARMv6 assembly optimized version of 'composite_over_n_8_0565'. It is ++ * a bare metal 'naked' function which uses all the available CPU registers ++ * and is compatible with ARM EABI. It might (or might not) break when used ++ * with a different ABI, anyway it is better to be safe than sorry. ++ */ ++static void __attribute__((naked)) armv6_composite_over_n_8_0565_asm ( ++ uint16_t *dst, uint8_t *mask, uint32_t src, int w, ++ int dst_stride_delta, int mask_stride_delta, int h) ++{ ++ asm volatile ( ++ ".macro composite_internal_armv6_asm opaque_flag\n" ++ /* save all registers (8 words) to stack */ ++ "stmdb sp!, {r4-r11, ip, lr}\n" ++ /* some register aliases for better readability */ ++ "DST .req r0\n" ++ "MASK .req r1\n" ++ "S .req r2\n" ++ "W .req r3\n" ++ "A .req r8\n" ++ "D .req r10\n" ++ "C0000FF .req r11\n" ++ "C00001F .req r9\n" ++ "C800080 .req ip\n" ++ "CE000E0 .req lr\n" ++ /* precalculate some stuff and put it on stack */ ++ "mov r6, #0xF8\n" ++ "mov r7, #0xFC\n" ++ ++ "str W, [sp, #-8]!\n" ++ ++ ".if \\opaque_flag\n" ++ /* precalculate and save it to stack for later use: ++ * ((src >> 3) & 0x001F) | ++ * ((src >> 5) & 0x07E0) | ++ * ((src >> 8) & 0xF800) ++ */ ++ "mov A, #0x1F\n" ++ "and D, A, S, lsr #3\n" ++ "and r4, S, #0xF80000\n" ++ "and r5, S, #0xFC00\n" ++ "orr D, r4, lsr #8\n" ++ "orr D, r5, lsr #5\n" ++ "str D, [sp, #4]\n" ++ ".endif\n" ++ ++ "ldr D, [sp, #(8 + 10*4 + 8)]\n" /* h */ ++ "ldr A, =0xFF00FF\n" ++ "ldr C800080, =0x800080\n" ++ "ldr CE000E0, =0xE000E0\n" ++ "ldr C0000FF, =0xFF\n" ++ "ldr C00001F, =0x1F\n" ++ "and r4, A, S\n" /* r4 = src & 0x00FF00FF */ ++ "and r5, A, S, lsr #8\n" /* r5 = (src >> 8) & 0x00FF00FF */ ++ "stmdb sp!, {r4, r5, r6, r7}\n" ++ "0:\n" ++ "subs D, D, #1\n" ++ "blt 6f\n" ++ "1:\n" ++ "subs W, W, #1\n" ++ "blt 5f\n" ++ "2:\n" ++ "ldrb A, [MASK], #1\n" ++ "ldmia sp, {r4, r5, r6, r7}\n" /* load constants from stack */ ++ "add DST, DST, #2\n" ++ "cmp A, #0\n" ++ "beq 1b\n" ++ ++ ".if \\opaque_flag\n" ++ "cmp A, #0xFF\n" ++ "bne 3f\n" ++ "ldr D, [sp, #(4*4 + 4)]\n" /* load precalculated value */ ++ "subs W, #1\n" ++ "strh D, [DST, #-2]\n" ++ "bge 2b\n" ++ ".endif\n" ++ ++ "3:\n" ++ "ldrh D, [DST, #-2]\n" ++ "mla r4, A, r4, C800080\n" ++ "mla r5, A, r5, C800080\n" ++ "and r6, r6, D, lsl #3\n" /* & 0xF8 */ ++ "and r7, r7, D, lsr #3\n" /* & 0xFC */ ++ "and D, D, #0xF800\n" ++ "bic S, r4, #0xFF0000\n" ++ "bic A, r5, #0xFF0000\n" ++ "add r4, r4, S, lsr #8\n" ++ "add r5, r5, A, lsr #8\n" ++ ++ "and S, r7, #0xC0\n" ++ "orr r6, r6, D, lsl #8\n" ++ "and D, r6, CE000E0\n" ++ "eor A, C0000FF, r5, lsr #24\n" ++ "orr r6, D, lsr #5\n" ++ "orr r7, S, lsr #6\n" ++ ++ "mla r6, A, r6, C800080\n" ++ "mla r7, A, r7, C800080\n" ++ "subs W, #1\n" ++ "bic D, r6, #0xFF0000\n" ++ "bic A, r7, #0xFF0000\n" ++ "add r6, r6, D, lsr #8\n" ++ "uqadd8 r4, r4, r6\n" ++ "add r7, r7, A, lsr #8\n" ++ "uqadd8 r5, r5, r7\n" ++ "and D, C00001F, r4, lsr #11\n" ++ "and r4, r4, #0xF8000000\n" ++ "and r5, r5, #0xFC00\n" ++ "orr D, r4, lsr #16\n" ++ "orr D, r5, lsr #5\n" ++ "strh D, [DST, #-2]\n" ++ "bge 2b\n" ++ "5:\n" ++ "ldr r6, [sp, #(4*4 + 8 + 10*4 + 8)]\n" /* h */ ++ "ldr r4, [sp, #(4*4 + 8 + 10*4 + 4)]\n" /* mask stride */ ++ "ldr r5, [sp, #(4*4 + 8 + 10*4 + 0)]\n" /* dst stride */ ++ "ldr W, [sp, #(4*4)]\n" ++ "subs r6, r6, #1\n" /* h */ ++ "str r6, [sp, #(4*4 + 8 + 10*4 + 8)]\n" /* h */ ++ "add MASK, MASK, r4\n" ++ "add DST, DST, r5, lsl #1\n" ++ "bgt 1b\n" ++ "6:\n" ++ "add sp, sp, #(4*4 + 8)\n" ++ /* restore all registers and return */ ++ "ldmia sp!, {r4-r11, ip, pc}\n" ++ ".unreq DST\n" ++ ".unreq MASK\n" ++ ".unreq S\n" ++ ".unreq W\n" ++ ".unreq A\n" ++ ".unreq D\n" ++ ".unreq C0000FF\n" ++ ".unreq C00001F\n" ++ ".unreq C800080\n" ++ ".unreq CE000E0\n" ++ ".endm\n" ++ ++ "mov ip, r2, lsr #24\n" ++ "cmp ip, #0xFF\n" ++ "beq 9f\n" ++ "composite_internal_armv6_asm 0\n" ++ "9:\n" ++ "composite_internal_armv6_asm 1\n" ++ ".ltorg\n" ++ ".purgem composite_internal_armv6_asm\n" ++ ); ++} ++ ++static void ++armv6_composite_over_n_8_0565 (pixman_implementation_t * impl, ++ pixman_op_t op, ++ pixman_image_t * src_image, ++ pixman_image_t * mask_image, ++ pixman_image_t * dst_image, ++ int32_t src_x, ++ int32_t src_y, ++ int32_t mask_x, ++ int32_t mask_y, ++ int32_t dest_x, ++ int32_t dest_y, ++ int32_t width, ++ int32_t height) ++{ ++ uint32_t src; ++ uint16_t *dst; ++ uint8_t *mask; ++ int dst_stride, mask_stride; ++ ++ src = _pixman_image_get_solid (src_image, dst_image->bits.format); ++ ++ /* bail out if fully transparent */ ++ if (src == 0) ++ return; ++ ++ PIXMAN_IMAGE_GET_LINE (dst_image, dest_x, dest_y, uint16_t, ++ dst_stride, dst, 1); ++ PIXMAN_IMAGE_GET_LINE (mask_image, mask_x, mask_y, uint8_t, ++ mask_stride, mask, 1); ++ ++ armv6_composite_over_n_8_0565_asm (dst, mask, src, width, ++ dst_stride - width, mask_stride - width, height); ++} ++ ++#endif ++ + static const pixman_fast_path_t arm_simd_fast_paths[] = + { + PIXMAN_STD_FAST_PATH (OVER, a8r8g8b8, null, a8r8g8b8, arm_composite_over_8888_8888), +@@ -434,7 +621,10 @@ static const pixman_fast_path_t arm_simd_fast_paths[] = + PIXMAN_STD_FAST_PATH (OVER, solid, a8, x8r8g8b8, arm_composite_over_n_8_8888), + PIXMAN_STD_FAST_PATH (OVER, solid, a8, a8b8g8r8, arm_composite_over_n_8_8888), + PIXMAN_STD_FAST_PATH (OVER, solid, a8, x8b8g8r8, arm_composite_over_n_8_8888), +- ++#if defined(__ARM_EABI__) && defined(__linux__) ++ PIXMAN_STD_FAST_PATH (OVER, solid, a8, r5g6b5, armv6_composite_over_n_8_0565), ++ PIXMAN_STD_FAST_PATH (OVER, solid, a8, b5g6r5, armv6_composite_over_n_8_0565), ++#endif + { PIXMAN_OP_NONE }, + }; + +-- +cgit v0.8.3-6-g21f6 |