From 4fde5f6a57b41cb35fe54322d3d6f38defe1f490 Mon Sep 17 00:00:00 2001 From: Henrik Smiding Date: Mon, 9 Jan 2012 16:46:54 +0100 Subject: Add shader for Clamp_SI8_opaque_D32_filter_DX. Adds a NEON optimized shader for SI8_opaque_D32_filter_DX when running with clamped tile mode. Improves performance of browser. Change-Id: I0e1fa03d4104210953b7ec81ce10eba5f77459cc Signed-off-by: Henrik Smiding Signed-off-by: Patrik Ryd --- src/core/SkBitmapProcState.h | 2 + src/opts/SkBitmapProcState_opts_arm.cpp | 215 ++++++++++++++++++++++++++++++++ 2 files changed, 217 insertions(+) (limited to 'src') diff --git a/src/core/SkBitmapProcState.h b/src/core/SkBitmapProcState.h index c04992b..1b42b61 100644 --- a/src/core/SkBitmapProcState.h +++ b/src/core/SkBitmapProcState.h @@ -144,5 +144,7 @@ void ClampX_ClampY_filter_affine(const SkBitmapProcState& s, uint32_t xy[], int count, int x, int y); void ClampX_ClampY_nofilter_affine(const SkBitmapProcState& s, uint32_t xy[], int count, int x, int y); +void SI8_opaque_D32_filter_DX(const SkBitmapProcState& s, const uint32_t xy[], + int count, SkPMColor colors[]); #endif diff --git a/src/opts/SkBitmapProcState_opts_arm.cpp b/src/opts/SkBitmapProcState_opts_arm.cpp index 63854f7..415bd0c 100644 --- a/src/opts/SkBitmapProcState_opts_arm.cpp +++ b/src/opts/SkBitmapProcState_opts_arm.cpp @@ -358,6 +358,214 @@ void Clamp_S32_opaque_D32_filter_DX_shaderproc(const SkBitmapProcState& s, int x } } while (count != 0); } + + +void Clamp_SI8_opaque_D32_filter_DX_shaderproc(const SkBitmapProcState& s, int x, int y, + uint32_t* SK_RESTRICT colors, int count) { + SkASSERT((s.fInvType & ~(SkMatrix::kTranslate_Mask | + SkMatrix::kScale_Mask)) == 0); + SkASSERT(s.fInvKy == 0); + SkASSERT(count > 0 && colors != NULL); + SkASSERT(s.fDoFilter); + SkDEBUGCODE(SkASSERT(s.fBitmap->config() == SkBitmap::kIndex8_Config);) + + const unsigned maxX = s.fBitmap->width() - 1; + const SkFixed oneX = s.fFilterOneX; + const SkFixed dx = s.fInvSx; + SkFixed fx; + const uint8_t* SK_RESTRICT row0; + const uint8_t* SK_RESTRICT row1; + unsigned subY; + + { + SkPoint pt; + s.fInvProc(*s.fInvMatrix, SkIntToScalar(x) + SK_ScalarHalf, + SkIntToScalar(y) + SK_ScalarHalf, &pt); + SkFixed fy = SkScalarToFixed(pt.fY) - (s.fFilterOneY >> 1); + const unsigned maxY = s.fBitmap->height() - 1; + // compute our two Y values up front + subY = ((fy >> 12) & 0xF); + int y0 = SkClampMax(fy >> 16, maxY); + int y1 = SkClampMax((fy + s.fFilterOneY) >> 16, maxY); + + const char* SK_RESTRICT srcAddr = (const char*)s.fBitmap->getPixels(); + unsigned rb = s.fBitmap->rowBytes(); + row0 = (const uint8_t*)(srcAddr + y0 * rb); + row1 = (const uint8_t*)(srcAddr + y1 * rb); + // now initialize fx + fx = SkScalarToFixed(pt.fX) - (oneX >> 1); + } + + const SkPMColor* SK_RESTRICT table = s.fBitmap->getColorTable()->lockColors(); + + do { + // Check if we can do the next four pixels using ARM NEON asm + if ((count >= 4) && + (((dx >= 0) && (fx >= 0) && (((fx + 3 * dx) >> 16) < (const signed)maxX)) || + ((dx < 0) && ((fx >> 16) < (const signed)maxX) && (((fx + 3 * dx) >> 16) >= 0)))) { + int asm_count; + + // How many iterations can we do while still clamped? + if (dx >= 0) { + asm_count = (((((const signed)maxX - 1) << 16) - fx) / dx) >> 2; + } else { + asm_count = ((0 - fx) / dx) >> 2; + } + + if (asm_count <= 0) { + asm_count = 1; + } else if ((asm_count << 2) > count) { + asm_count = count >> 2; + } + + count -= asm_count << 2; + + // We know that oneX is 1.0 since we are running clamped. + // This means that we can load both x0 and x1 offsets in one go. + asm volatile ( + // Setup constants + "rsb r8, %[subY], #16 \n\t" // 16 - subY + "vdup.8 d30, %[subY] \n\t" // Create constant for subY + "vdup.8 d31, r8 \n\t" // Create constant for 16 - subY + "vmov.u16 d29, #16 \n\t" // Create constant for 16 + "1: \n\t" // Loop start + // Pre-load pixel #1 + "asr r7, %[fx], #16 \n\t" // Calculate offset fx >> 16 + "ldrh r8, [%[row0], r7] \n\t" // Fetch row0 color table offsets + "ldrh r7, [%[row1], r7] \n\t" // Fetch row1 color table offsets + "subs %[cnt], %[cnt], #1 \n\t" // Decrement loop counter + "and r6, r8, #0xFF \n\t" // Extract first offset + "lsr r8, r8, #8 \n\t" // Extract second offset + "add r6, %[table], r6, lsl #2 \n\t" // Calculate a00 address for table + "add r8, %[table], r8, lsl #2 \n\t" // Calculate a01 address for table + "vld1.32 {d0[0]}, [r6] \n\t" // Load a00 RGBA pixel from table + "vld1.32 {d0[1]}, [r8] \n\t" // Load a01 RGBA pixel from table + "and r6, r7, #0xFF \n\t" // Extract first offset + "lsr r7, r7, #8 \n\t" // Extract second offset + "add r6, %[table], r6, lsl #2 \n\t" // Calculate a10 address for table + "add r7, %[table], r7, lsl #2 \n\t" // Calculate a11 address for table + "vld1.32 {d1[0]}, [r6] \n\t" // Load a10 RGBA pixel from table + "vld1.32 {d1[1]}, [r7] \n\t" // Load a11 RGBA pixel from table + // Calculate pixel #1 and pre-load #2 + "lsr r8, %[fx], #12 \n\t" // Calculate subX = ((fx >> 12) & 0xF) + "and r8, r8, #0xF \n\t" // + "add %[fx], %[fx], %[dx] \n\t" // Move fx to next position + "vdup.16 d28, r8 \n\t" // subX + "asr r7, %[fx], #16 \n\t" // Calculate offset fx >> 16 + "ldrh r8, [%[row0], r7] \n\t" // Fetch row0 color table offsets + "ldrh r7, [%[row1], r7] \n\t" // Fetch row1 color table offsets + "vmull.u8 q1, d0, d31 \n\t" // q0 = [a00|a01] * (16 - y) + "vmull.u8 q2, d1, d30 \n\t" // q1 = [a10|a11] * y + "and r6, r8, #0xFF \n\t" // Extract first offset + "lsr r8, r8, #8 \n\t" // Extract second offset + "add r6, %[table], r6, lsl #2 \n\t" // Calculate a00 address for table + "add r8, %[table], r8, lsl #2 \n\t" // Calculate a01 address for table + "vld1.32 {d0[0]}, [r6] \n\t" // Load a00 RGBA pixel from table + "vld1.32 {d0[1]}, [r8] \n\t" // Load a01 RGBA pixel from table + "vmul.i16 d16, d3, d28 \n\t" // d16 = a01 * x + "vmla.i16 d16, d5, d28 \n\t" // d16 += a11 * x + "and r6, r7, #0xFF \n\t" // Extract first offset + "lsr r7, r7, #8 \n\t" // Extract second offset + "add r6, %[table], r6, lsl #2 \n\t" // Calculate a10 address for table + "add r7, %[table], r7, lsl #2 \n\t" // Calculate a11 address for table + "vld1.32 {d1[0]}, [r6] \n\t" // Load a10 RGBA pixel from table + "vld1.32 {d1[1]}, [r7] \n\t" // Load a11 RGBA pixel from table + "vsub.i16 d27, d29, d28 \n\t" // 16 - subX + "vmla.i16 d16, d2, d27 \n\t" // d16 += a00 * (16 - x) + "vmla.i16 d16, d4, d27 \n\t" // d16 += a10 * (16 - x) + // Calculate pixel #2 and pre-load #3 + "lsr r8, %[fx], #12 \n\t" // Calculate subX = ((fx >> 12) & 0xF) + "and r8, r8, #0xF \n\t" // + "add %[fx], %[fx], %[dx] \n\t" // Move fx to next position + "vdup.16 d28, r8 \n\t" // subX + "asr r7, %[fx], #16 \n\t" // Calculate offset fx >> 16 + "ldrh r8, [%[row0], r7] \n\t" // Fetch row0 color table offsets + "ldrh r7, [%[row1], r7] \n\t" // Fetch row1 color table offsets + "vmull.u8 q1, d0, d31 \n\t" // q0 = [a00|a01] * (16 - y) + "vmull.u8 q2, d1, d30 \n\t" // q1 = [a10|a11] * y + "and r6, r8, #0xFF \n\t" // Extract first offset + "lsr r8, r8, #8 \n\t" // Extract second offset + "add r6, %[table], r6, lsl #2 \n\t" // Calculate a00 address for table + "add r8, %[table], r8, lsl #2 \n\t" // Calculate a01 address for table + "vld1.32 {d0[0]}, [r6] \n\t" // Load a00 RGBA pixel from table + "vld1.32 {d0[1]}, [r8] \n\t" // Load a01 RGBA pixel from table + "vmul.i16 d17, d3, d28 \n\t" // d17 = a01 * x + "vmla.i16 d17, d5, d28 \n\t" // d17 += a11 * x + "and r6, r7, #0xFF \n\t" // Extract first offset + "lsr r7, r7, #8 \n\t" // Extract second offset + "add r6, %[table], r6, lsl #2 \n\t" // Calculate a10 address for table + "add r7, %[table], r7, lsl #2 \n\t" // Calculate a11 address for table + "vld1.32 {d1[0]}, [r6] \n\t" // Load a10 RGBA pixel from table + "vld1.32 {d1[1]}, [r7] \n\t" // Load a11 RGBA pixel from table + "vsub.i16 d27, d29, d28 \n\t" // 16 - subX + "vmla.i16 d17, d2, d27 \n\t" // d17 += a00 * (16 - x) + "vmla.i16 d17, d4, d27 \n\t" // d17 += a10 * (16 - x) + // Calculate pixel #3 and pre-load #4 + "lsr r8, %[fx], #12 \n\t" // Calculate subX = ((fx >> 12) & 0xF) + "and r8, r8, #0xF \n\t" // + "add %[fx], %[fx], %[dx] \n\t" // Move fx to next position + "vdup.16 d28, r8 \n\t" // subX + "asr r7, %[fx], #16 \n\t" // Calculate offset fx >> 16 + "ldrh r8, [%[row0], r7] \n\t" // Fetch row0 color table offsets + "ldrh r7, [%[row1], r7] \n\t" // Fetch row1 color table offsets + "vmull.u8 q1, d0, d31 \n\t" // q0 = [a00|a01] * (16 - y) + "vmull.u8 q2, d1, d30 \n\t" // q1 = [a10|a11] * y + "and r6, r8, #0xFF \n\t" // Extract first offset + "lsr r8, r8, #8 \n\t" // Extract second offset + "add r6, %[table], r6, lsl #2 \n\t" // Calculate a00 address for table + "add r8, %[table], r8, lsl #2 \n\t" // Calculate a01 address for table + "vld1.32 {d0[0]}, [r6] \n\t" // Load a00 RGBA pixel from table + "vld1.32 {d0[1]}, [r8] \n\t" // Load a01 RGBA pixel from table + "vmul.i16 d18, d3, d28 \n\t" // d18 = a01 * x + "vmla.i16 d18, d5, d28 \n\t" // d18 += a11 * x + "and r6, r7, #0xFF \n\t" // Extract first offset + "lsr r7, r7, #8 \n\t" // Extract second offset + "add r6, %[table], r6, lsl #2 \n\t" // Calculate a10 address for table + "add r7, %[table], r7, lsl #2 \n\t" // Calculate a11 address for table + "vld1.32 {d1[0]}, [r6] \n\t" // Load a10 RGBA pixel from table + "vld1.32 {d1[1]}, [r7] \n\t" // Load a11 RGBA pixel from table + "vsub.i16 d27, d29, d28 \n\t" // 16 - subX + "vmla.i16 d18, d2, d27 \n\t" // d18 += a00 * (16 - x) + "vmla.i16 d18, d4, d27 \n\t" // d18 += a10 * (16 - x) + "vshrn.i16 d16, q8, #8 \n\t" // shift down result by 8 + // Calculate pixel #4 + "vmull.u8 q1, d0, d31 \n\t" // q0 = [a00|a01] * (16 - y) + "vmull.u8 q2, d1, d30 \n\t" // q1 = [a10|a11] * y + "lsr r8, %[fx], #12 \n\t" // Calculate subX = ((fx >> 12) & 0xF) + "and r8, r8, #0xF \n\t" // + "add %[fx], %[fx], %[dx] \n\t" // Move fx to next position + "vdup.16 d28, r8 \n\t" // subX + "vmul.i16 d19, d3, d28 \n\t" // d19 = a01 * x + "vmla.i16 d19, d5, d28 \n\t" // d19 += a11 * x + "vsub.i16 d27, d29, d28 \n\t" // 16 - subX + "vmla.i16 d19, d2, d27 \n\t" // d19 += a00 * (16 - x) + "vmla.i16 d19, d4, d27 \n\t" // d19 += a10 * (16 - x) + "vshrn.i16 d17, q9, #8 \n\t" // shift down result by 8 + "vst1.32 {d16-d17}, [%[colors]]! \n\t" // Write result to memory + "bne 1b \n\t" + : [fx] "+r" (fx), [colors] "+r" (colors), [cnt] "+r" (asm_count) + : [row0] "r" (row0), [row1] "r" (row1), [subY] "r" (subY), [dx] "r" (dx), [table] "r" (table) + : "cc", "memory", "r6", "r7", "r8", "d0", "d1", "d2", "d3", "d4", "d5", "d16", "d17", "d18", "d19", "d27", "d28", "d29", "d30", "d31" + ); + } else { + unsigned subX = ((fx >> 12) & 0xF); + unsigned x0 = SkClampMax(fx >> 16, maxX); + unsigned x1 = SkClampMax((fx + oneX) >> 16, maxX); + + Filter_32_opaque(subX, subY, + table[row0[x0]], + table[row0[x1]], + table[row1[x0]], + table[row1[x1]], + colors); + colors += 1; + fx += dx; + count--; + } + } while (count != 0); + + s.fBitmap->getColorTable()->unlockColors(false); +} #endif @@ -375,6 +583,13 @@ void SkBitmapProcState::platformProcs() { switch (fBitmap->config()) { case SkBitmap::kIndex8_Config: +#if defined(__ARM_HAVE_NEON) && defined(SK_CPU_LENDIAN) + if (SI8_opaque_D32_filter_DX == fSampleProc32) { + if (clamp_clamp) { + fShaderProc32 = Clamp_SI8_opaque_D32_filter_DX_shaderproc; + } + } else +#endif #if __ARM_ARCH__ >= 6 && !defined(SK_CPU_BENDIAN) if (justDx && !doFilter) { #if 0 /* crashing on android device */ -- cgit v1.1