From 6878d573e50cca6ad917b67d596e5bb38c93b42a Mon Sep 17 00:00:00 2001 From: Tom Smeding Date: Tue, 5 May 2020 16:47:42 +0200 Subject: Subsampling for faster image --- mandel.cu | 102 +++++++++++++++++++++++++++++++++++++++++++++++++++++--------- 1 file changed, 87 insertions(+), 15 deletions(-) (limited to 'mandel.cu') diff --git a/mandel.cu b/mandel.cu index 480daed..386f962 100644 --- a/mandel.cu +++ b/mandel.cu @@ -60,34 +60,77 @@ void mandel_free(Mandel *ctx) { delete ctx; } +// These macros may assume that the parameters are simple variable names. +#if 1 +#define MANDEL_FUNCTION_REAL(a, b, a2, b2, x, y) (a2 - b2 + x) +#define MANDEL_FUNCTION_IMAG(a, b, a2, b2, x, y) (2 * a * b + y) +#endif +#if 0 +#define MANDEL_FUNCTION_REAL(a, b, a2, b2, x, y) (a2 - b2 + x + 1) +#define MANDEL_FUNCTION_IMAG(a, b, a2, b2, x, y) (2 * a * b + y) +#endif +#if 0 +#define MANDEL_FUNCTION_REAL(a, b, a2, b2, x, y) (a2 * a - 3 * a * b * b + x) +#define MANDEL_FUNCTION_IMAG(a, b, a2, b2, x, y) (3 * a * a * b - b2 * b + y) +#endif + #define MANDEL_GENERIC(dst, ctx, par, ix, iy, idx) { \ const double x = (par).cx - (par).imgw / 2 + (par).imgw * (ix) / ((ctx).w-1); \ const double y = (par).cy - (ctx).imgh / 2 + (ctx).imgh * (iy) / ((ctx).h-1); \ double a = x, b = y, a2 = a * a, b2 = b * b; \ int16_t iter, maxiter = (par).maxit; \ for (iter = 0; iter < maxiter && a2 + b2 < 4; iter++) { \ - b = 2 * a * b + y; a = a2 - b2 + x; \ + double newa = MANDEL_FUNCTION_REAL(a, b, a2, b2, x, y); \ + double newb = MANDEL_FUNCTION_IMAG(a, b, a2, b2, x, y); \ + a = newa; b = newb; \ a2 = a * a; b2 = b * b; \ } \ (dst)[idx] = iter; \ } -__global__ void mandel_gpu(int16_t *dst, const Mandel *ctx, const Params *par) { - const int idx = blockDim.x * blockIdx.x + threadIdx.x; - const int ix = idx % (int)ctx->w, iy = ctx->h - 1 - idx / (int)ctx->w; +__global__ void mandel_gpu(int16_t *dst, const Mandel *ctx, const Params *par, int subsample) { + const int subidx = blockDim.x * blockIdx.x + threadIdx.x; + const int subw = (ctx->w + subsample - 1) / subsample; + const int subix = subidx % subw, subiy = subidx / subw; + const int ix = subix * subsample, iy = subiy * subsample; + const int idx = ctx->w * iy + ix; + const int riy = ctx->h - 1 - iy; + if (iy >= ctx->h) return; - MANDEL_GENERIC(dst, *ctx, *par, ix, iy, idx); + MANDEL_GENERIC(dst, *ctx, *par, ix, riy, idx); +} + +__global__ void unsubsample_gpu(int16_t *dst, const Mandel *ctx, int subsample) { + const int idx = blockDim.x * blockIdx.x + threadIdx.x; + const int ix = idx % ctx->w, iy = idx / ctx->w; + const int cellix = ix / subsample * subsample, celliy = iy / subsample * subsample; + + dst[ctx->w * iy + ix] = dst[ctx->w * celliy + cellix]; } // Unused in GPU mode __attribute__((unused)) -static inline void mandel_cpu(int16_t *dst, const Mandel *ctx, const Params *par) { - int idx = 0; - for (int iy = 0; iy < ctx->h; iy++) { - for (int ix = 0; ix < ctx->w; ix++) { - MANDEL_GENERIC(dst, *ctx, *par, ix, iy, idx); - idx++; +static inline void mandel_cpu(int16_t *dst, const Mandel *ctx, const Params *par, int subsample) { + for (int iy = 0; iy < ctx->h; iy += subsample) { + for (int ix = 0; ix < ctx->w; ix += subsample) { + int idx = ctx->w * iy + ix; + int riy = ctx->h - 1 - iy; + MANDEL_GENERIC(dst, *ctx, *par, ix, riy, idx); + } + } + + if (subsample == 1) return; + + for (int iy = 0; iy < ctx->h; iy += subsample) { + for (int ix = 0; ix < ctx->w; ix += subsample) { + const int16_t value = dst[ctx->w * iy + ix]; + + for (int subiy = 0; subiy < subsample; subiy++) { + for (int subix = 0; subix < subsample; subix++) { + dst[ctx->w * (iy + subiy) + ix + subix] = value; + } + } } } } @@ -126,20 +169,28 @@ double mandel_imgh(const Mandel *ctx, const Params *par) { return par->imgw * ctx->h / ctx->w; } -void mandel_render(uint8_t *dst, Mandel *ctx, const Params *par) { +void mandel_render(uint8_t *dst, Mandel *ctx, const Params *par, size_t subsample) { ctx->imgh = mandel_imgh(ctx, par); int64_t t1 = gettimestamp(); #ifdef USE_GPU + assert((size_t)(int)subsample == subsample); + + const int subfactor = subsample * subsample; + CUDA_CHECK(cudaMemcpy(ctx->devPar, par, sizeof(Params), cudaMemcpyHostToDevice)); CUDA_CHECK(cudaMemcpy(ctx->devCtx, ctx, sizeof(Mandel), cudaMemcpyHostToDevice)); const int nblocks = (ctx->w * ctx->h + 1023) / 1024; - mandel_gpu<<>>(ctx->devImg, ctx->devCtx, ctx->devPar); + mandel_gpu<<>>(ctx->devImg, ctx->devCtx, ctx->devPar, subsample); + + if (subsample > 1) { + unsubsample_gpu<<>>(ctx->devImg, ctx->devCtx, subsample); + } CUDA_CHECK(cudaMemcpy(ctx->img, ctx->devImg, ctx->w * ctx->h * sizeof(int16_t), cudaMemcpyDeviceToHost)); #else - mandel_cpu(ctx->img, ctx, par); + mandel_cpu(ctx->img, ctx, par, subsample); #endif int64_t t2 = gettimestamp(); @@ -155,20 +206,41 @@ void mandel_render(uint8_t *dst, Mandel *ctx, const Params *par) { int64_t t3 = gettimestamp(); cout << "gpu part: " << (t2 - t1) / 1000000.0 << " sec " - << "cpu part: " << (t3 - t2) / 1000000 << " sec" << endl; + << "cpu part: " << (t3 - t2) / 1000000.0 << " sec" << endl; } #if 0 int main() { Mandel *ctx = mandel_init(1920, 1080); Params par = mandel_default_params(); + +#if 0 par.cx = -0.73844331961137488; par.cy = -0.20921562151741355; par.imgw = 9.6624448855068765e-08; par.maxit = 20000; + uint8_t *img = new uint8_t[3 * 1920 * 1080]; mandel_render(img, ctx, &par); mandel_free(ctx); bmp_rgb_encode_file("out.bmp", img, 1920, 1080); +#else + par.cx = -0.23845305789504126; + par.cy = 0.87200567648858607; + par.imgw = 3.5; // zoom to 7.5815186948616694e-14 + par.maxit = 6144; + + uint8_t *img = new uint8_t[3 * 1920 * 1080]; + + for (int i = 0; i <= 141; i++) { + mandel_render(img, ctx, &par, 1); + char fname[64]; + sprintf(fname, "outdir/out%03d.bmp", i); + bmp_rgb_encode_file(fname, img, 1920, 1080); + par.imgw *= 0.8; + } + + mandel_free(ctx); +#endif } #endif -- cgit v1.2.3-54-g00ecf