diff options
author | Suren A. Chilingaryan <csa@suren.me> | 2020-03-29 23:13:45 +0200 |
---|---|---|
committer | Suren A. Chilingaryan <csa@suren.me> | 2020-03-29 23:13:45 +0200 |
commit | cd0dd76a48aada43c3066642a8eb327a0374d113 (patch) | |
tree | d5ce8cebda9511c5ac6280df54c29e8b631fa1cb | |
parent | febfe9a6490052d4b8789fd8f7a0342115bfd55e (diff) | |
download | regularization-cd0dd76a48aada43c3066642a8eb327a0374d113.tar.gz regularization-cd0dd76a48aada43c3066642a8eb327a0374d113.tar.bz2 regularization-cd0dd76a48aada43c3066642a8eb327a0374d113.tar.xz regularization-cd0dd76a48aada43c3066642a8eb327a0374d113.zip |
Add optimization steps for reference
-rw-r--r-- | src/Core/performance_CPU/README | 25 | ||||
-rwxr-xr-x | src/Core/performance_CPU/TNV_core.c.v15 | 731 | ||||
-rwxr-xr-x | src/Core/performance_CPU/TNV_core.c.v17 | 690 | ||||
-rwxr-xr-x | src/Core/performance_CPU/TNV_core.c.v18 | 688 | ||||
-rwxr-xr-x | src/Core/performance_CPU/TNV_core.c.v19 | 681 | ||||
-rwxr-xr-x | src/Core/performance_CPU/TNV_core.c.v27 | 650 | ||||
-rwxr-xr-x | src/Core/performance_CPU/TNV_core.c.v32 | 676 | ||||
-rwxr-xr-x | src/Core/performance_CPU/TNV_core.c.v4.stdver | 629 | ||||
-rw-r--r-- | src/Core/performance_CPU/TNV_core_backtrack_loop.h.v19 | 100 | ||||
-rw-r--r-- | src/Core/performance_CPU/TNV_core_loop.h.v32 | 119 |
10 files changed, 4989 insertions, 0 deletions
diff --git a/src/Core/performance_CPU/README b/src/Core/performance_CPU/README new file mode 100644 index 0000000..948c544 --- /dev/null +++ b/src/Core/performance_CPU/README @@ -0,0 +1,25 @@ +TNV_core.c.v4.stdver: Fully equivalent (signle-threads). There is two potential breakage points. + - If OpenMP is enabled, the acces to div_upd will not be serialized and results will breaj + - The results will slightly differ due to different order of summation if loop summing resprimal/resdual organized in a logical way +TNV_core.c.v15: Multi-threads. Works correctly only in the single-threaded mode (if TNV_NEW_STYLE is disabled). In multi-threaded there results slightly differ due to changed order of operation + - TNV_NEW_STYLE slightly disturbs results in both single- and multi-threaded modes + - Resprimal/resdual are summed in groups (not sequentially) if multiple threads. But his actually should improve precision. Use TNV_CHECK_RES to check conformance + - Afterwards, in multi-threaded moded there is a still minor descripancy which first occurs in resprimal (after a few iterations). This is because of changed order of operations while computing + div_upd (only on the first lines of each new sub-block). Normally, we first compute the vertical and, then, add horizontal. On the border rows, instead we first add horizontals... + To check, div/div_upd changed to double. There is no difference then. +TNV_core.c.v17: Computationaly comptabile with v15. + - Padding actually harms performance + - Intel compiler gives about 10% speed-up +TNV_core.c.v18: Blocking helps to boost performance further but only with Intel Compiler. Gcc/Clang is slightly slower here. + - Padding here doesn't harm performance, but is not helpful either + - Difference between icc and gcc is probably due to auto-vectorization. + - Results slightly changed due to different order of operations +TNV_core.c.v19: Eliminate conditionals in the inner loops to help gcc-autovectorisation + - Last version implementing full algrotithm with backtrack in the middle of iterations. + - Again results slightly diverge from v18 due to different order of operations +TNV_core.c.v27: v18 with backtracking only on first iterations (otherwise warning reported) +TNV_core.c.v32: v19 with backtracking only on first iterations (otherwise warning reported) + + +Repo: + - Padding seems to have effect on the newer AVX2 systems. Re-enabled. diff --git a/src/Core/performance_CPU/TNV_core.c.v15 b/src/Core/performance_CPU/TNV_core.c.v15 new file mode 100755 index 0000000..3161cbf --- /dev/null +++ b/src/Core/performance_CPU/TNV_core.c.v15 @@ -0,0 +1,731 @@ +/* + * This work is part of the Core Imaging Library developed by + * Visual Analytics and Imaging System Group of the Science Technology + * Facilities Council, STFC + * + * Copyright 2017 Daniil Kazantsev + * Copyright 2017 Srikanth Nagella, Edoardo Pasca + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * http://www.apache.org/licenses/LICENSE-2.0 + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include "TNV_core.h" + +#define min(a,b) (((a)<(b))?(a):(b)) + +/*inline*/ void coefF(float *t, float M1, float M2, float M3, float sigma, int p, int q, int r) { + int ii, num; + float divsigma = 1.0f / sigma; + float sum, shrinkfactor; + float T,D,det,eig1,eig2,sig1,sig2,V1, V2, V3, V4, v0,v1,v2, mu1,mu2,sig1_upd,sig2_upd; + float proj[2] = {0}; + + // Compute eigenvalues of M + T = M1 + M3; + D = M1 * M3 - M2 * M2; + det = sqrtf(MAX((T * T / 4.0f) - D, 0.0f)); + eig1 = MAX((T / 2.0f) + det, 0.0f); + eig2 = MAX((T / 2.0f) - det, 0.0f); + sig1 = sqrtf(eig1); + sig2 = sqrtf(eig2); + + // Compute normalized eigenvectors + V1 = V2 = V3 = V4 = 0.0f; + + if(M2 != 0.0f) + { + v0 = M2; + v1 = eig1 - M3; + v2 = eig2 - M3; + + mu1 = sqrtf(v0 * v0 + v1 * v1); + mu2 = sqrtf(v0 * v0 + v2 * v2); + + if(mu1 > fTiny) + { + V1 = v1 / mu1; + V3 = v0 / mu1; + } + + if(mu2 > fTiny) + { + V2 = v2 / mu2; + V4 = v0 / mu2; + } + + } else + { + if(M1 > M3) + { + V1 = V4 = 1.0f; + V2 = V3 = 0.0f; + + } else + { + V1 = V4 = 0.0f; + V2 = V3 = 1.0f; + } + } + + // Compute prox_p of the diagonal entries + sig1_upd = sig2_upd = 0.0f; + + if(p == 1) + { + sig1_upd = MAX(sig1 - divsigma, 0.0f); + sig2_upd = MAX(sig2 - divsigma, 0.0f); + + } else if(p == INFNORM) + { + proj[0] = sigma * fabs(sig1); + proj[1] = sigma * fabs(sig2); + + /*l1 projection part */ + sum = fLarge; + num = 0l; + shrinkfactor = 0.0f; + while(sum > 1.0f) + { + sum = 0.0f; + num = 0; + + for(ii = 0; ii < 2; ii++) + { + proj[ii] = MAX(proj[ii] - shrinkfactor, 0.0f); + + sum += fabs(proj[ii]); + if(proj[ii]!= 0.0f) + num++; + } + + if(num > 0) + shrinkfactor = (sum - 1.0f) / num; + else + break; + } + /*l1 proj ends*/ + + sig1_upd = sig1 - divsigma * proj[0]; + sig2_upd = sig2 - divsigma * proj[1]; + } + + // Compute the diagonal entries of $\widehat{\Sigma}\Sigma^{\dagger}_0$ + if(sig1 > fTiny) + sig1_upd /= sig1; + + if(sig2 > fTiny) + sig2_upd /= sig2; + + // Compute solution + t[0] = sig1_upd * V1 * V1 + sig2_upd * V2 * V2; + t[1] = sig1_upd * V1 * V3 + sig2_upd * V2 * V4; + t[2] = sig1_upd * V3 * V3 + sig2_upd * V4 * V4; +} + + +#include "hw_sched.h" +typedef struct { + int offY, stepY, realY, copY; + float *Input, *u, *u_upd, *qx, *qy, *qx_upd, *qy_upd, *gradx, *grady, *gradx_upd, *grady_upd; + double *div, *div_upd; + float resprimal, resdual; + float unorm, qnorm, product; +} tnv_thread_t; + +typedef struct { + int threads; + tnv_thread_t *thr_ctx; + float *InputT, *uT; + int dimX, dimY, dimZ, padZ; + float lambda, sigma, tau, theta; +} tnv_context_t; + +HWSched sched = NULL; +tnv_context_t tnv_ctx; + + +static int tnv_free(HWThread thr, void *hwctx, int device_id, void *data) { + int i,j,k; + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + free(ctx->Input); + free(ctx->u); + free(ctx->u_upd); + free(ctx->qx); + free(ctx->qy); + free(ctx->qx_upd); + free(ctx->qy_upd); + free(ctx->gradx); + free(ctx->grady); + free(ctx->gradx_upd); + free(ctx->grady_upd); + free(ctx->div); + free(ctx->div_upd); + + return 0; +} + +static int tnv_init(HWThread thr, void *hwctx, int device_id, void *data) { + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int offY = ctx->offY; + int stepY = ctx->stepY; + int realY = ctx->realY; + +// printf("%i %p - %i %i %i x %i %i\n", device_id, ctx, dimX, dimY, dimZ, offY, stepY); + + long DimTotal = (long)(dimX*realY*dimZ); + long Dim1Total = (long)(dimX*(stepY+1)*dimZ); + + // Auxiliar vectors + ctx->Input = calloc(Dim1Total, sizeof(float)); + ctx->u = calloc(Dim1Total, sizeof(float)); + ctx->u_upd = calloc(Dim1Total, sizeof(float)); + ctx->qx = calloc(DimTotal, sizeof(float)); + ctx->qy = calloc(DimTotal, sizeof(float)); + ctx->qx_upd = calloc(DimTotal, sizeof(float)); + ctx->qy_upd = calloc(DimTotal, sizeof(float)); + ctx->gradx = calloc(DimTotal, sizeof(float)); + ctx->grady = calloc(DimTotal, sizeof(float)); + ctx->gradx_upd = calloc(DimTotal, sizeof(float)); + ctx->grady_upd = calloc(DimTotal, sizeof(float)); + ctx->div = calloc(Dim1Total, sizeof(double)); + ctx->div_upd = calloc(Dim1Total, sizeof(double)); + + if ((!ctx->Input)||(!ctx->u)||(!ctx->u_upd)||(!ctx->qx)||(!ctx->qy)||(!ctx->qx_upd)||(!ctx->qy_upd)||(!ctx->gradx)||(!ctx->grady)||(!ctx->gradx_upd)||(!ctx->grady_upd)||(!ctx->div)||(!ctx->div_upd)) { + fprintf(stderr, "Error allocating memory\n"); + exit(-1); + } + + return 0; +} + +static int tnv_start(HWThread thr, void *hwctx, int device_id, void *data) { + int i,j,k; + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int offY = ctx->offY; + int stepY = ctx->stepY; + int realY = ctx->realY; + int copY = ctx->copY; + +// printf("%i %p - %i %i %i x %i %i\n", device_id, ctx, dimX, dimY, dimZ, offY, stepY); + + long DimTotal = (long)(dimX*realY*dimZ); + long Dim1Total = (long)(dimX*(stepY+1)*dimZ); + + memset(ctx->u, 0, Dim1Total * sizeof(float)); + memset(ctx->qx, 0, DimTotal * sizeof(float)); + memset(ctx->qy, 0, DimTotal * sizeof(float)); + memset(ctx->gradx, 0, DimTotal * sizeof(float)); + memset(ctx->grady, 0, DimTotal * sizeof(float)); + memset(ctx->div, 0, Dim1Total * sizeof(double)); + memset(ctx->u_upd, 0, Dim1Total * sizeof(float)); + memset(ctx->qx_upd, 0, DimTotal * sizeof(float)); + memset(ctx->qy_upd, 0, DimTotal * sizeof(float)); + memset(ctx->gradx_upd, 0, DimTotal * sizeof(float)); + memset(ctx->grady_upd, 0, DimTotal * sizeof(float)); + memset(ctx->div_upd, 0, Dim1Total * sizeof(double)); + + for(k=0; k<dimZ; k++) { + for(j=0; j<copY; j++) { + for(i=0; i<dimX; i++) { + ctx->Input[j * dimX * dimZ + i * dimZ + k] = tnv_ctx->InputT[k * dimX * dimY + (j + offY) * dimX + i]; + ctx->u[j * dimX * dimZ + i * dimZ + k] = tnv_ctx->uT[k * dimX * dimY + (j + offY) * dimX + i]; + } + } + } + + return 0; +} + +static int tnv_finish(HWThread thr, void *hwctx, int device_id, void *data) { + int i,j,k; + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int offY = ctx->offY; + int stepY = ctx->stepY; + int realY = ctx->realY; + int copY = ctx->copY; + + long DimTotal = (long)(dimX*realY*dimZ); + long Dim1Total = (long)(dimX*(stepY+1)*dimZ); + + for(k=0; k<dimZ; k++) { + for(j=0; j<copY; j++) { + for(i=0; i<dimX; i++) { + tnv_ctx->uT[k * dimX * dimY + (j + offY) * dimX + i] = ctx->u[j * dimX * dimZ + i * dimZ + k]; + } + } + } + + return 0; +} + + +static int tnv_copy(HWThread thr, void *hwctx, int device_id, void *data) { + int i,j,k; + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int stepY = ctx->stepY; + int realY = ctx->realY; + long DimTotal = (long)(dimX*realY*dimZ); + long Dim1Total = (long)(dimX*(stepY+1)*dimZ); + + // Auxiliar vectors + memcpy(ctx->u, ctx->u_upd, Dim1Total * sizeof(float)); + memcpy(ctx->qx, ctx->qx_upd, DimTotal * sizeof(float)); + memcpy(ctx->qy, ctx->qy_upd, DimTotal * sizeof(float)); + memcpy(ctx->gradx, ctx->gradx_upd, DimTotal * sizeof(float)); + memcpy(ctx->grady, ctx->grady_upd, DimTotal * sizeof(float)); + memcpy(ctx->div, ctx->div_upd, Dim1Total * sizeof(double)); + + return 0; +} + +static int tnv_restore(HWThread thr, void *hwctx, int device_id, void *data) { + int i,j,k; + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int stepY = ctx->stepY; + int realY = ctx->realY; + long DimTotal = (long)(dimX*realY*dimZ); + long Dim1Total = (long)(dimX*(stepY+1)*dimZ); + + // Auxiliar vectors + memcpy(ctx->u_upd, ctx->u, Dim1Total * sizeof(float)); + memcpy(ctx->qx_upd, ctx->qx, DimTotal * sizeof(float)); + memcpy(ctx->qy_upd, ctx->qy, DimTotal * sizeof(float)); + memcpy(ctx->gradx_upd, ctx->gradx, DimTotal * sizeof(float)); + memcpy(ctx->grady_upd, ctx->grady, DimTotal * sizeof(float)); + memcpy(ctx->div_upd, ctx->div, Dim1Total * sizeof(double)); + + return 0; +} + + +static int tnv_step(HWThread thr, void *hwctx, int device_id, void *data) { + long i, j, k, l, m; + + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int padZ = tnv_ctx->padZ; + int offY = ctx->offY; + int stepY = ctx->stepY; + int copY = ctx->copY; + + float *Input = ctx->Input; + float *u = ctx->u; + float *u_upd = ctx->u_upd; + float *qx = ctx->qx; + float *qy = ctx->qy; + float *qx_upd = ctx->qx_upd; + float *qy_upd = ctx->qy_upd; + float *gradx = ctx->gradx; + float *grady = ctx->grady; + float *gradx_upd = ctx->gradx_upd; + float *grady_upd = ctx->grady_upd; + double *div = ctx->div; + double *div_upd = ctx->div_upd; + + long p = 1l; + long q = 1l; + long r = 0l; + + float lambda = tnv_ctx->lambda; + float sigma = tnv_ctx->sigma; + float tau = tnv_ctx->tau; + float theta = tnv_ctx->theta; + + float taulambda = tau * lambda; + float divtau = 1.0f / tau; + float divsigma = 1.0f / sigma; + float theta1 = 1.0f + theta; + float constant = 1.0f + taulambda; + + float resprimal = 0.0f; + float resdual = 0.0f; + float product = 0.0f; + float unorm = 0.0f; + float qnorm = 0.0f; + + float udiff[dimZ]; + float qxdiff; + float qydiff; + float divdiff; + float gradxdiff[dimZ]; + float gradydiff[dimZ]; + + for(k=0; k < dimZ * dimX; k++) { + u_upd[k] = (u[k] + tau * div[k] + taulambda * Input[k])/constant; + div_upd[k] = 0; + } + + for(j = 0; j < stepY; j++) { +/* m = j * dimX * dimZ + (dimX - 1) * dimZ; + for(k = 0; k < dimZ; k++) { + u_upd[k + m] = (u[k + m] + tau * div[k + m] + taulambda * Input[k + m]) / constant; + }*/ + + for(i=0; i < (dimX /*- 1*/); i++) { + l = (j * dimX + i) * dimZ; + float t[3]; + float M1 = 0.0f, M2 = 0.0f, M3 = 0.0f; + m = dimX * dimZ; + +//#pragma unroll 64 + for(k = 0; k < dimZ; k++) { + u_upd[l + k + m] = (u[l + k + m] + tau * div[l + k + m] + taulambda * Input[l + k + m]) / constant; + + gradx_upd[l + k] = (i == (dimX - 1))?0:(u_upd[l + k + dimZ] - u_upd[l + k]); + grady_upd[l + k] = (j == (copY - 1))?0:(u_upd[l + k + dimX * dimZ] - u_upd[l + k]); // We need div from the next thread on last iter + + udiff[k] = u[l + k] - u_upd[l + k]; + unorm += (udiff[k] * udiff[k]); +// if ((!k)&&(!i)) printf("%i = %f %f, %f %f\n", offY + j, u[l + k], u_upd[l + k], udiff[k], unorm); + + gradxdiff[k] = gradx[l + k] - gradx_upd[l + k]; + gradydiff[k] = grady[l + k] - grady_upd[l + k]; + + float ubarx = theta1 * gradx_upd[l + k] - theta * gradx[l + k]; + float ubary = theta1 * grady_upd[l + k] - theta * grady[l + k]; +//#define TNV_NEW_STYLE +#ifdef TNV_NEW_STYLE + qx_upd[l + k] = qx[l + k] + sigma * ubarx; + qy_upd[l + k] = qy[l + k] + sigma * ubary; + + float vx = divsigma * qx_upd[l + k]; //+ ubarx + float vy = divsigma * qy_upd[l + k]; //+ ubary +#else + float vx = ubarx + divsigma * qx[l + k]; + float vy = ubary + divsigma * qy[l + k]; +#endif + + M1 += (vx * vx); M2 += (vx * vy); M3 += (vy * vy); + } + + coefF(t, M1, M2, M3, sigma, p, q, r); + +//#pragma unroll 64 + for(k = 0; k < dimZ; k++) { +#ifdef TNV_NEW_STYLE + float vx = divsigma * qx_upd[l + k]; + float vy = divsigma * qy_upd[l + k]; + + float gx_upd = vx * t[0] + vy * t[1]; + float gy_upd = vx * t[1] + vy * t[2]; + + qx_upd[l + k] -= sigma * gx_upd; + qy_upd[l + k] -= sigma * gy_upd; +#else + float ubarx = theta1 * gradx_upd[l + k] - theta * gradx[l + k]; + float ubary = theta1 * grady_upd[l + k] - theta * grady[l + k]; + float vx = ubarx + divsigma * qx[l + k]; + float vy = ubary + divsigma * qy[l + k]; + + float gx_upd = vx * t[0] + vy * t[1]; + float gy_upd = vx * t[1] + vy * t[2]; + + qx_upd[l + k] = qx[l + k] + sigma * (ubarx - gx_upd); + qy_upd[l + k] = qy[l + k] + sigma * (ubary - gy_upd); +#endif + +if(i != (dimX-1)) { + div_upd[l + k] += qx_upd[l + k]; + div_upd[l + k + dimZ] -= qx_upd[l + k]; +} +if(j != (copY-1)) { + div_upd[l + k] += qy_upd[l + k]; + div_upd[l + k + dimX * dimZ] = -qy_upd[l + k]; // We need to update div in the next thread on last iter +} + + qxdiff = qx[l + k] - qx_upd[l + k]; + qydiff = qy[l + k] - qy_upd[l + k]; + qnorm += (qxdiff * qxdiff + qydiff * qydiff); + + resdual += fabs(divsigma * qxdiff - gradxdiff[k]); + resdual += fabs(divsigma * qydiff - gradydiff[k]); + product += (gradxdiff[k] * qxdiff + gradydiff[k] * qydiff); + + if ((offY == 0)||(j > 0)) { + divdiff = div[l + k] - div_upd[l + k]; // Multiple steps... How we compute without history? + resprimal += fabs(divtau * udiff[k] + divdiff); + } + } + + } // i + } + + + ctx->resprimal = resprimal; + ctx->resdual = resdual; + ctx->product = product; + ctx->unorm = unorm; + ctx->qnorm = qnorm; + + return 0; +} + +static void TNV_CPU_init(float *InputT, float *uT, int dimX, int dimY, int dimZ) { + int i, off, size, err; + + if (sched) return; + + tnv_ctx.dimX = dimX; + tnv_ctx.dimY = dimY; + tnv_ctx.dimZ = dimZ; + tnv_ctx.padZ = 64 * ((dimZ / 64) + ((dimZ % 64)?1:0)); + + hw_sched_init(); + + int threads = hw_sched_get_cpu_count(); + if (threads > dimY) threads = dimY/2; + + int step = dimY / threads; + int extra = dimY % threads; + + tnv_ctx.threads = threads; + tnv_ctx.thr_ctx = (tnv_thread_t*)calloc(threads, sizeof(tnv_thread_t)); + for (i = 0, off = 0; i < threads; i++, off += size) { + tnv_thread_t *ctx = tnv_ctx.thr_ctx + i; + size = step + ((i < extra)?1:0); + + ctx->offY = off; + ctx->stepY = size; + ctx->realY = ctx->stepY; + + if (i == (threads-1)) ctx->copY = ctx->stepY; + else ctx->copY = ctx->stepY + 1; + } + + sched = hw_sched_create(threads); + if (!sched) { fprintf(stderr, "Error creating threads\n"); exit(-1); } + + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_init); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling init threads", err); exit(-1); } +} + + + +/* + * C-OMP implementation of Total Nuclear Variation regularisation model (2D + channels) [1] + * The code is modified from the implementation by Joan Duran <joan.duran@uib.es> see + * "denoisingPDHG_ipol.cpp" in Joans Collaborative Total Variation package + * + * Input Parameters: + * 1. Noisy volume of 2D + channel dimension, i.e. 3D volume + * 2. lambda - regularisation parameter + * 3. Number of iterations [OPTIONAL parameter] + * 4. eplsilon - tolerance constant [OPTIONAL parameter] + * 5. print information: 0 (off) or 1 (on) [OPTIONAL parameter] + * + * Output: + * 1. Filtered/regularized image (u) + * + * [1]. Duran, J., Moeller, M., Sbert, C. and Cremers, D., 2016. Collaborative total variation: a general framework for vectorial TV models. SIAM Journal on Imaging Sciences, 9(1), pp.116-151. + */ + +float TNV_CPU_main(float *InputT, float *uT, float lambda, int maxIter, float tol, int dimX, int dimY, int dimZ) +{ + int err; + int iter; + int i,j,k,l; + + lambda = 1.0f/(2.0f*lambda); + tnv_ctx.lambda = lambda; + + // PDHG algorithm parameters + float tau = 0.5f; + float sigma = 0.5f; + float theta = 1.0f; + + // Backtracking parameters + float s = 1.0f; + float gamma = 0.75f; + float beta = 0.95f; + float alpha0 = 0.2f; + float alpha = alpha0; + float delta = 1.5f; + float eta = 0.95f; + + TNV_CPU_init(InputT, uT, dimX, dimY, dimZ); + + tnv_ctx.InputT = InputT; + tnv_ctx.uT = uT; + + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_start); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling start threads", err); exit(-1); } + + + // Apply Primal-Dual Hybrid Gradient scheme + float residual = fLarge; + for(iter = 0; iter < maxIter; iter++) { + float resprimal = 0.0f; + float resdual = 0.0f; + float product = 0.0f; + float unorm = 0.0f; + float qnorm = 0.0f; + + float divtau = 1.0f / tau; + + tnv_ctx.sigma = sigma; + tnv_ctx.tau = tau; + tnv_ctx.theta = theta; + + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_step); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling tnv threads", err); exit(-1); } + + // border regions + for (i = 1; i < tnv_ctx.threads; i++) { + tnv_thread_t *ctx0 = tnv_ctx.thr_ctx + (i - 1); + tnv_thread_t *ctx = tnv_ctx.thr_ctx + i; + + l = ctx0->stepY * dimX * dimZ; + for(k=0; k < dimZ * (dimX /*- 1*/); k++) { + double div_upd_add = ctx0->div_upd[l + k]; + ctx->div_upd[k] += div_upd_add; + ctx0->div_upd[l + k] = ctx->div_upd[k]; + +// ctx0->u_upd[l + k] = ctx->u_upd[k]; + + float divdiff = ctx->div[k] - ctx->div_upd[k]; // Multiple steps... How we compute without history? + float udiff = ctx->u[k] - ctx->u_upd[k]; + resprimal += fabs(divtau * udiff + divdiff); + } + } + +#define TNV_CHECK_RES +#ifndef TNV_CHECK_RES + for (i = 0; i < tnv_ctx.threads; i++) { + tnv_thread_t *ctx = tnv_ctx.thr_ctx + i; + resprimal += ctx->resprimal; + resdual += ctx->resdual; + product += ctx->product; + unorm += ctx->unorm; + qnorm += ctx->qnorm; + } +#else + resprimal = 0; + float divsigma = 1.0f / sigma; + for(j=0; j<dimY; j++) + for(i=0; i<dimX; i++) + for(int l=0; l<dimZ; l++) + { + int step = dimY / tnv_ctx.threads; + int extra = dimY % tnv_ctx.threads; + + int thr, subj = j; + for (thr = 0; thr < tnv_ctx.threads; thr++) { + int size = step; + if (thr < extra) size++; + + if (subj >= size) subj-= size; + else break; + } + + tnv_thread_t *ctx = tnv_ctx.thr_ctx + thr; + + int k = subj * dimX * dimZ + i * dimZ + l; + + float udiff = ctx->u[k] - ctx->u_upd[k]; + float qxdiff = ctx->qx[k] - ctx->qx_upd[k]; + float qydiff = ctx->qy[k] - ctx->qy_upd[k]; + float divdiff = ctx->div[k] - ctx->div_upd[k]; + float gradxdiff = ctx->gradx[k] - ctx->gradx_upd[k]; + float gradydiff = ctx->grady[k] - ctx->grady_upd[k]; + + resprimal += fabs(divtau*udiff + divdiff); + resdual += fabs(divsigma*qxdiff - gradxdiff); + resdual += fabs(divsigma*qydiff - gradydiff); + + unorm += (udiff * udiff); + qnorm += (qxdiff * qxdiff + qydiff * qydiff); + product += (gradxdiff * qxdiff + gradydiff * qydiff); + } +#endif + + + + residual = (resprimal + resdual) / ((float) (dimX*dimY*dimZ)); + float b = (2.0f * tau * sigma * product) / (gamma * sigma * unorm + gamma * tau * qnorm); + float dual_dot_delta = resdual * s * delta; + float dual_div_delta = (resdual * s) / delta; + printf("resprimal: %f, resdual: %f, b: %f (product: %f, unorm: %f, qnorm: %f)\n", resprimal, resdual, b, product, unorm, qnorm); + + + if(b > 1) { + // Decrease step-sizes to fit balancing principle + tau = (beta * tau) / b; + sigma = (beta * sigma) / b; + alpha = alpha0; + + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_restore); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling restore threads", err); exit(-1); } + } else { + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_copy); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling copy threads", err); exit(-1); } + + if(resprimal > dual_dot_delta) { + // Increase primal step-size and decrease dual step-size + tau = tau / (1.0f - alpha); + sigma = sigma * (1.0f - alpha); + alpha = alpha * eta; + } else if(resprimal < dual_div_delta) { + // Decrease primal step-size and increase dual step-size + tau = tau * (1.0f - alpha); + sigma = sigma / (1.0f - alpha); + alpha = alpha * eta; + } + } + + if (residual < tol) break; + } + + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_finish); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling finish threads", err); exit(-1); } + + + printf("Iterations stopped at %i with the residual %f \n", iter, residual); + printf("Return: %f\n", *uT); + + return *uT; +} diff --git a/src/Core/performance_CPU/TNV_core.c.v17 b/src/Core/performance_CPU/TNV_core.c.v17 new file mode 100755 index 0000000..60739cd --- /dev/null +++ b/src/Core/performance_CPU/TNV_core.c.v17 @@ -0,0 +1,690 @@ +/* + * This work is part of the Core Imaging Library developed by + * Visual Analytics and Imaging System Group of the Science Technology + * Facilities Council, STFC + * + * Copyright 2017 Daniil Kazantsev + * Copyright 2017 Srikanth Nagella, Edoardo Pasca + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * http://www.apache.org/licenses/LICENSE-2.0 + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include "TNV_core.h" + +#define min(a,b) (((a)<(b))?(a):(b)) + +inline void coefF(float *t, float M1, float M2, float M3, float sigma, int p, int q, int r) { + int ii, num; + float divsigma = 1.0f / sigma; + float sum, shrinkfactor; + float T,D,det,eig1,eig2,sig1,sig2,V1, V2, V3, V4, v0,v1,v2, mu1,mu2,sig1_upd,sig2_upd; + float proj[2] = {0}; + + // Compute eigenvalues of M + T = M1 + M3; + D = M1 * M3 - M2 * M2; + det = sqrtf(MAX((T * T / 4.0f) - D, 0.0f)); + eig1 = MAX((T / 2.0f) + det, 0.0f); + eig2 = MAX((T / 2.0f) - det, 0.0f); + sig1 = sqrtf(eig1); + sig2 = sqrtf(eig2); + + // Compute normalized eigenvectors + V1 = V2 = V3 = V4 = 0.0f; + + if(M2 != 0.0f) + { + v0 = M2; + v1 = eig1 - M3; + v2 = eig2 - M3; + + mu1 = sqrtf(v0 * v0 + v1 * v1); + mu2 = sqrtf(v0 * v0 + v2 * v2); + + if(mu1 > fTiny) + { + V1 = v1 / mu1; + V3 = v0 / mu1; + } + + if(mu2 > fTiny) + { + V2 = v2 / mu2; + V4 = v0 / mu2; + } + + } else + { + if(M1 > M3) + { + V1 = V4 = 1.0f; + V2 = V3 = 0.0f; + + } else + { + V1 = V4 = 0.0f; + V2 = V3 = 1.0f; + } + } + + // Compute prox_p of the diagonal entries + sig1_upd = sig2_upd = 0.0f; + + if(p == 1) + { + sig1_upd = MAX(sig1 - divsigma, 0.0f); + sig2_upd = MAX(sig2 - divsigma, 0.0f); + + } else if(p == INFNORM) + { + proj[0] = sigma * fabs(sig1); + proj[1] = sigma * fabs(sig2); + + /*l1 projection part */ + sum = fLarge; + num = 0l; + shrinkfactor = 0.0f; + while(sum > 1.0f) + { + sum = 0.0f; + num = 0; + + for(ii = 0; ii < 2; ii++) + { + proj[ii] = MAX(proj[ii] - shrinkfactor, 0.0f); + + sum += fabs(proj[ii]); + if(proj[ii]!= 0.0f) + num++; + } + + if(num > 0) + shrinkfactor = (sum - 1.0f) / num; + else + break; + } + /*l1 proj ends*/ + + sig1_upd = sig1 - divsigma * proj[0]; + sig2_upd = sig2 - divsigma * proj[1]; + } + + // Compute the diagonal entries of $\widehat{\Sigma}\Sigma^{\dagger}_0$ + if(sig1 > fTiny) + sig1_upd /= sig1; + + if(sig2 > fTiny) + sig2_upd /= sig2; + + // Compute solution + t[0] = sig1_upd * V1 * V1 + sig2_upd * V2 * V2; + t[1] = sig1_upd * V1 * V3 + sig2_upd * V2 * V4; + t[2] = sig1_upd * V3 * V3 + sig2_upd * V4 * V4; +} + + +#include "hw_sched.h" +typedef struct { + int offY, stepY, copY; + float *Input, *u, *u_upd, *qx, *qy, *qx_upd, *qy_upd, *gradx, *grady, *gradx_upd, *grady_upd; + float *div, *div_upd; + float resprimal, resdual; + float unorm, qnorm, product; +} tnv_thread_t; + +typedef struct { + int threads; + tnv_thread_t *thr_ctx; + float *InputT, *uT; + int dimX, dimY, dimZ, padZ; + float lambda, sigma, tau, theta; +} tnv_context_t; + +HWSched sched = NULL; +tnv_context_t tnv_ctx; + + +static int tnv_free(HWThread thr, void *hwctx, int device_id, void *data) { + int i,j,k; + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + free(ctx->Input); + free(ctx->u); + free(ctx->u_upd); + free(ctx->qx); + free(ctx->qy); + free(ctx->qx_upd); + free(ctx->qy_upd); + free(ctx->gradx); + free(ctx->grady); + free(ctx->gradx_upd); + free(ctx->grady_upd); + free(ctx->div); + free(ctx->div_upd); + + return 0; +} + +static int tnv_init(HWThread thr, void *hwctx, int device_id, void *data) { + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int padZ = tnv_ctx->padZ; + int offY = ctx->offY; + int stepY = ctx->stepY; + +// printf("%i %p - %i %i %i x %i %i\n", device_id, ctx, dimX, dimY, dimZ, offY, stepY); + + long DimTotal = (long)(dimX*stepY*padZ); + long Dim1Total = (long)(dimX*(stepY+1)*padZ); + + // Auxiliar vectors + ctx->Input = malloc(Dim1Total * sizeof(float)); + ctx->u = malloc(Dim1Total * sizeof(float)); + ctx->u_upd = malloc(Dim1Total * sizeof(float)); + ctx->qx = malloc(DimTotal * sizeof(float)); + ctx->qy = malloc(DimTotal * sizeof(float)); + ctx->qx_upd = malloc(DimTotal * sizeof(float)); + ctx->qy_upd = malloc(DimTotal * sizeof(float)); + ctx->gradx = malloc(DimTotal * sizeof(float)); + ctx->grady = malloc(DimTotal * sizeof(float)); + ctx->gradx_upd = malloc(DimTotal * sizeof(float)); + ctx->grady_upd = malloc(DimTotal * sizeof(float)); + ctx->div = malloc(Dim1Total * sizeof(float)); + ctx->div_upd = malloc(Dim1Total * sizeof(float)); + + if ((!ctx->Input)||(!ctx->u)||(!ctx->u_upd)||(!ctx->qx)||(!ctx->qy)||(!ctx->qx_upd)||(!ctx->qy_upd)||(!ctx->gradx)||(!ctx->grady)||(!ctx->gradx_upd)||(!ctx->grady_upd)||(!ctx->div)||(!ctx->div_upd)) { + fprintf(stderr, "Error allocating memory\n"); + exit(-1); + } + + return 0; +} + +static int tnv_start(HWThread thr, void *hwctx, int device_id, void *data) { + int i,j,k; + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int padZ = tnv_ctx->padZ; + int offY = ctx->offY; + int stepY = ctx->stepY; + int copY = ctx->copY; + +// printf("%i %p - %i %i %i (%i) x %i %i\n", device_id, ctx, dimX, dimY, dimZ, padZ, offY, stepY); + + long DimTotal = (long)(dimX*stepY*padZ); + long Dim1Total = (long)(dimX*copY*padZ); + + memset(ctx->u, 0, Dim1Total * sizeof(float)); + memset(ctx->qx, 0, DimTotal * sizeof(float)); + memset(ctx->qy, 0, DimTotal * sizeof(float)); + memset(ctx->gradx, 0, DimTotal * sizeof(float)); + memset(ctx->grady, 0, DimTotal * sizeof(float)); + memset(ctx->div, 0, Dim1Total * sizeof(float)); + + for(k=0; k<dimZ; k++) { + for(j=0; j<copY; j++) { + for(i=0; i<dimX; i++) { + ctx->Input[j * dimX * padZ + i * padZ + k] = tnv_ctx->InputT[k * dimX * dimY + (j + offY) * dimX + i]; + ctx->u[j * dimX * padZ + i * padZ + k] = tnv_ctx->uT[k * dimX * dimY + (j + offY) * dimX + i]; + } + } + } + + return 0; +} + +static int tnv_finish(HWThread thr, void *hwctx, int device_id, void *data) { + int i,j,k; + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int padZ = tnv_ctx->padZ; + int offY = ctx->offY; + int stepY = ctx->stepY; + int copY = ctx->copY; + + for(k=0; k<dimZ; k++) { + for(j=0; j<stepY; j++) { + for(i=0; i<dimX; i++) { + tnv_ctx->uT[k * dimX * dimY + (j + offY) * dimX + i] = ctx->u[j * dimX * padZ + i * padZ + k]; + } + } + } + + return 0; +} + + +static int tnv_copy(HWThread thr, void *hwctx, int device_id, void *data) { + int i,j,k; + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int stepY = ctx->stepY; + int copY = ctx->copY; + int padZ = tnv_ctx->padZ; + long DimTotal = (long)(dimX*stepY*padZ); + long Dim1Total = (long)(dimX*copY*padZ); + + // Auxiliar vectors + memcpy(ctx->u, ctx->u_upd, Dim1Total * sizeof(float)); + memcpy(ctx->qx, ctx->qx_upd, DimTotal * sizeof(float)); + memcpy(ctx->qy, ctx->qy_upd, DimTotal * sizeof(float)); + memcpy(ctx->gradx, ctx->gradx_upd, DimTotal * sizeof(float)); + memcpy(ctx->grady, ctx->grady_upd, DimTotal * sizeof(float)); + memcpy(ctx->div, ctx->div_upd, Dim1Total * sizeof(float)); + + return 0; +} + +static int tnv_restore(HWThread thr, void *hwctx, int device_id, void *data) { + int i,j,k; + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int stepY = ctx->stepY; + int copY = ctx->copY; + int padZ = tnv_ctx->padZ; + long DimTotal = (long)(dimX*stepY*padZ); + long Dim1Total = (long)(dimX*copY*padZ); + + // Auxiliar vectors + memcpy(ctx->u_upd, ctx->u, Dim1Total * sizeof(float)); + memcpy(ctx->qx_upd, ctx->qx, DimTotal * sizeof(float)); + memcpy(ctx->qy_upd, ctx->qy, DimTotal * sizeof(float)); + memcpy(ctx->gradx_upd, ctx->gradx, DimTotal * sizeof(float)); + memcpy(ctx->grady_upd, ctx->grady, DimTotal * sizeof(float)); + memcpy(ctx->div_upd, ctx->div, Dim1Total * sizeof(float)); + + return 0; +} + + +static int tnv_step(HWThread thr, void *hwctx, int device_id, void *data) { + long i, j, k, l, m; + + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int padZ = tnv_ctx->padZ; + int offY = ctx->offY; + int stepY = ctx->stepY; + int copY = ctx->copY; + + float *Input = ctx->Input; + float *u = ctx->u; + float *u_upd = ctx->u_upd; + float *qx = ctx->qx; + float *qy = ctx->qy; + float *qx_upd = ctx->qx_upd; + float *qy_upd = ctx->qy_upd; + float *gradx = ctx->gradx; + float *grady = ctx->grady; + float *gradx_upd = ctx->gradx_upd; + float *grady_upd = ctx->grady_upd; + float *div = ctx->div; + float *div_upd = ctx->div_upd; + + long p = 1l; + long q = 1l; + long r = 0l; + + float lambda = tnv_ctx->lambda; + float sigma = tnv_ctx->sigma; + float tau = tnv_ctx->tau; + float theta = tnv_ctx->theta; + + float taulambda = tau * lambda; + float divtau = 1.0f / tau; + float divsigma = 1.0f / sigma; + float theta1 = 1.0f + theta; + float constant = 1.0f + taulambda; + + float resprimal = 0.0f; + float resdual = 0.0f; + float product = 0.0f; + float unorm = 0.0f; + float qnorm = 0.0f; + + float udiff[dimZ]; + float qxdiff; + float qydiff; + float divdiff; + float gradxdiff[dimZ]; + float gradydiff[dimZ]; + + for(i=0; i < dimX; i++) { + for(k = 0; k < dimZ; k++) { + int l = i * padZ + k; + u_upd[l] = (u[l] + tau * div[l] + taulambda * Input[l])/constant; + div_upd[l] = 0; + } + } + + for(j = 0; j < stepY; j++) { +/* m = j * dimX * dimZ + (dimX - 1) * dimZ; + for(k = 0; k < dimZ; k++) { + u_upd[k + m] = (u[k + m] + tau * div[k + m] + taulambda * Input[k + m]) / constant; + }*/ + + for(i = 0; i < dimX/* - 1*/; i++) { + float t[3]; + float M1 = 0.0f, M2 = 0.0f, M3 = 0.0f; + l = (j * dimX + i) * padZ; + m = dimX * padZ; + +//#pragma unroll 64 + for(k = 0; k < dimZ; k++) { + u_upd[l + k + m] = (u[l + k + m] + tau * div[l + k + m] + taulambda * Input[l + k + m]) / constant; + + gradx_upd[l + k] = (i == (dimX - 1))?0:(u_upd[l + k + padZ] - u_upd[l + k]); + grady_upd[l + k] = (j == (copY - 1))?0:(u_upd[l + k + m] - u_upd[l + k]); // We need div from the next thread on last iter + + udiff[k] = u[l + k] - u_upd[l + k]; + unorm += (udiff[k] * udiff[k]); +// if ((!k)&&(!i)) printf("%i = %f %f, %f %f\n", offY + j, u[l + k], u_upd[l + k], udiff[k], unorm); + + gradxdiff[k] = gradx[l + k] - gradx_upd[l + k]; + gradydiff[k] = grady[l + k] - grady_upd[l + k]; + + float ubarx = theta1 * gradx_upd[l + k] - theta * gradx[l + k]; + float ubary = theta1 * grady_upd[l + k] - theta * grady[l + k]; +//#define TNV_NEW_STYLE +#ifdef TNV_NEW_STYLE + qx_upd[l + k] = qx[l + k] + sigma * ubarx; + qy_upd[l + k] = qy[l + k] + sigma * ubary; + + float vx = divsigma * qx_upd[l + k]; //+ ubarx + float vy = divsigma * qy_upd[l + k]; //+ ubary +#else + float vx = ubarx + divsigma * qx[l + k]; + float vy = ubary + divsigma * qy[l + k]; +#endif + + M1 += (vx * vx); M2 += (vx * vy); M3 += (vy * vy); + } + + coefF(t, M1, M2, M3, sigma, p, q, r); + +//#pragma unroll 64 + for(k = 0; k < dimZ; k++) { +#ifdef TNV_NEW_STYLE + float vx = divsigma * qx_upd[l + k]; + float vy = divsigma * qy_upd[l + k]; + + float gx_upd = vx * t[0] + vy * t[1]; + float gy_upd = vx * t[1] + vy * t[2]; + + qx_upd[l + k] -= sigma * gx_upd; + qy_upd[l + k] -= sigma * gy_upd; +#else + float ubarx = theta1 * gradx_upd[l + k] - theta * gradx[l + k]; + float ubary = theta1 * grady_upd[l + k] - theta * grady[l + k]; + float vx = ubarx + divsigma * qx[l + k]; + float vy = ubary + divsigma * qy[l + k]; + + float gx_upd = vx * t[0] + vy * t[1]; + float gy_upd = vx * t[1] + vy * t[2]; + + qx_upd[l + k] = qx[l + k] + sigma * (ubarx - gx_upd); + qy_upd[l + k] = qy[l + k] + sigma * (ubary - gy_upd); +#endif + +if(i != (dimX-1)) { + div_upd[l + k] += qx_upd[l + k]; + div_upd[l + k + padZ] -= qx_upd[l + k]; +} +if(j != (copY-1)) { + div_upd[l + k] += qy_upd[l + k]; + div_upd[l + k + m] = -qy_upd[l + k]; // We need to update div in the next thread on last iter +} + + qxdiff = qx[l + k] - qx_upd[l + k]; + qydiff = qy[l + k] - qy_upd[l + k]; + qnorm += (qxdiff * qxdiff + qydiff * qydiff); + + resdual += fabs(divsigma * qxdiff - gradxdiff[k]); + resdual += fabs(divsigma * qydiff - gradydiff[k]); + product += (gradxdiff[k] * qxdiff + gradydiff[k] * qydiff); + + if ((offY == 0)||(j > 0)) { + divdiff = div[l + k] - div_upd[l + k]; // Multiple steps... How we compute without history? + resprimal += fabs(divtau * udiff[k] + divdiff); + } + } + + } // i + } + + + ctx->resprimal = resprimal; + ctx->resdual = resdual; + ctx->product = product; + ctx->unorm = unorm; + ctx->qnorm = qnorm; + + return 0; +} + +static void TNV_CPU_init(float *InputT, float *uT, int dimX, int dimY, int dimZ) { + int i, off, size, err; + + if (sched) return; + + tnv_ctx.dimX = dimX; + tnv_ctx.dimY = dimY; + tnv_ctx.dimZ = dimZ; + // Padding seems actually slower +// tnv_ctx.padZ = 64 * ((dimZ / 64) + ((dimZ % 64)?1:0)); + tnv_ctx.padZ = dimZ; + + hw_sched_init(); + + int threads = hw_sched_get_cpu_count(); + if (threads > dimY) threads = dimY/2; + + int step = dimY / threads; + int extra = dimY % threads; + + tnv_ctx.threads = threads; + tnv_ctx.thr_ctx = (tnv_thread_t*)calloc(threads, sizeof(tnv_thread_t)); + for (i = 0, off = 0; i < threads; i++, off += size) { + tnv_thread_t *ctx = tnv_ctx.thr_ctx + i; + size = step + ((i < extra)?1:0); + + ctx->offY = off; + ctx->stepY = size; + + if (i == (threads-1)) ctx->copY = ctx->stepY; + else ctx->copY = ctx->stepY + 1; + } + + sched = hw_sched_create(threads); + if (!sched) { fprintf(stderr, "Error creating threads\n"); exit(-1); } + + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_init); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling init threads", err); exit(-1); } +} + + + +/* + * C-OMP implementation of Total Nuclear Variation regularisation model (2D + channels) [1] + * The code is modified from the implementation by Joan Duran <joan.duran@uib.es> see + * "denoisingPDHG_ipol.cpp" in Joans Collaborative Total Variation package + * + * Input Parameters: + * 1. Noisy volume of 2D + channel dimension, i.e. 3D volume + * 2. lambda - regularisation parameter + * 3. Number of iterations [OPTIONAL parameter] + * 4. eplsilon - tolerance constant [OPTIONAL parameter] + * 5. print information: 0 (off) or 1 (on) [OPTIONAL parameter] + * + * Output: + * 1. Filtered/regularized image (u) + * + * [1]. Duran, J., Moeller, M., Sbert, C. and Cremers, D., 2016. Collaborative total variation: a general framework for vectorial TV models. SIAM Journal on Imaging Sciences, 9(1), pp.116-151. + */ + +float TNV_CPU_main(float *InputT, float *uT, float lambda, int maxIter, float tol, int dimX, int dimY, int dimZ) +{ + int err; + int iter; + int i,j,k,l,m; + + lambda = 1.0f/(2.0f*lambda); + tnv_ctx.lambda = lambda; + + // PDHG algorithm parameters + float tau = 0.5f; + float sigma = 0.5f; + float theta = 1.0f; + + // Backtracking parameters + float s = 1.0f; + float gamma = 0.75f; + float beta = 0.95f; + float alpha0 = 0.2f; + float alpha = alpha0; + float delta = 1.5f; + float eta = 0.95f; + + TNV_CPU_init(InputT, uT, dimX, dimY, dimZ); + + tnv_ctx.InputT = InputT; + tnv_ctx.uT = uT; + + int padZ = tnv_ctx.padZ; + + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_start); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling start threads", err); exit(-1); } + + + // Apply Primal-Dual Hybrid Gradient scheme + float residual = fLarge; + for(iter = 0; iter < maxIter; iter++) { + float resprimal = 0.0f; + float resdual = 0.0f; + float product = 0.0f; + float unorm = 0.0f; + float qnorm = 0.0f; + + float divtau = 1.0f / tau; + + tnv_ctx.sigma = sigma; + tnv_ctx.tau = tau; + tnv_ctx.theta = theta; + + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_step); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling tnv threads", err); exit(-1); } + + // border regions + for (j = 1; j < tnv_ctx.threads; j++) { + tnv_thread_t *ctx0 = tnv_ctx.thr_ctx + (j - 1); + tnv_thread_t *ctx = tnv_ctx.thr_ctx + j; + + m = ctx0->stepY * dimX * padZ; + for(i = 0; i < dimX; i++) { + for(k = 0; k < dimZ; k++) { + int l = i * padZ + k; + + float div_upd_add = ctx0->div_upd[m + l]; + ctx->div_upd[l] += div_upd_add; + ctx0->div_upd[m + l] = ctx->div_upd[l]; + //ctx0->u_upd[m + l] = ctx->u_upd[l]; + + float divdiff = ctx->div[l] - ctx->div_upd[l]; // Multiple steps... How we compute without history? + float udiff = ctx->u[l] - ctx->u_upd[l]; + resprimal += fabs(divtau * udiff + divdiff); + } + } + } + + for (j = 0; j < tnv_ctx.threads; j++) { + tnv_thread_t *ctx = tnv_ctx.thr_ctx + j; + resprimal += ctx->resprimal; + resdual += ctx->resdual; + product += ctx->product; + unorm += ctx->unorm; + qnorm += ctx->qnorm; + } + + residual = (resprimal + resdual) / ((float) (dimX*dimY*dimZ)); + float b = (2.0f * tau * sigma * product) / (gamma * sigma * unorm + gamma * tau * qnorm); + float dual_dot_delta = resdual * s * delta; + float dual_div_delta = (resdual * s) / delta; + printf("resprimal: %f, resdual: %f, b: %f (product: %f, unorm: %f, qnorm: %f)\n", resprimal, resdual, b, product, unorm, qnorm); + + + if(b > 1) { + // Decrease step-sizes to fit balancing principle + tau = (beta * tau) / b; + sigma = (beta * sigma) / b; + alpha = alpha0; + + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_restore); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling restore threads", err); exit(-1); } + } else { + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_copy); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling copy threads", err); exit(-1); } + + if(resprimal > dual_dot_delta) { + // Increase primal step-size and decrease dual step-size + tau = tau / (1.0f - alpha); + sigma = sigma * (1.0f - alpha); + alpha = alpha * eta; + } else if(resprimal < dual_div_delta) { + // Decrease primal step-size and increase dual step-size + tau = tau * (1.0f - alpha); + sigma = sigma / (1.0f - alpha); + alpha = alpha * eta; + } + } + + if (residual < tol) break; + } + + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_finish); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling finish threads", err); exit(-1); } + + + printf("Iterations stopped at %i with the residual %f \n", iter, residual); + printf("Return: %f\n", *uT); + + return *uT; +} diff --git a/src/Core/performance_CPU/TNV_core.c.v18 b/src/Core/performance_CPU/TNV_core.c.v18 new file mode 100755 index 0000000..7192475 --- /dev/null +++ b/src/Core/performance_CPU/TNV_core.c.v18 @@ -0,0 +1,688 @@ +/* + * This work is part of the Core Imaging Library developed by + * Visual Analytics and Imaging System Group of the Science Technology + * Facilities Council, STFC + * + * Copyright 2017 Daniil Kazantsev + * Copyright 2017 Srikanth Nagella, Edoardo Pasca + * + * Copyriht 2020 Suren A. Chlingaryan + * Optimized version with 1/2 of memory consumption and ~4x performance + * This version is algorithmicly comptable with the original, but slight change in results + * is expected due to different order of floating-point operations. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * http://www.apache.org/licenses/LICENSE-2.0 + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include <malloc.h> +#include "TNV_core.h" + +#define BLOCK 32 +#define min(a,b) (((a)<(b))?(a):(b)) + +inline void coefF(float *t, float M1, float M2, float M3, float sigma, int p, int q, int r) { + int ii, num; + float divsigma = 1.0f / sigma; + float sum, shrinkfactor; + float T,D,det,eig1,eig2,sig1,sig2,V1, V2, V3, V4, v0,v1,v2, mu1,mu2,sig1_upd,sig2_upd; + float proj[2] = {0}; + + // Compute eigenvalues of M + T = M1 + M3; + D = M1 * M3 - M2 * M2; + det = sqrtf(MAX((T * T / 4.0f) - D, 0.0f)); + eig1 = MAX((T / 2.0f) + det, 0.0f); + eig2 = MAX((T / 2.0f) - det, 0.0f); + sig1 = sqrtf(eig1); + sig2 = sqrtf(eig2); + + // Compute normalized eigenvectors + V1 = V2 = V3 = V4 = 0.0f; + + if(M2 != 0.0f) + { + v0 = M2; + v1 = eig1 - M3; + v2 = eig2 - M3; + + mu1 = sqrtf(v0 * v0 + v1 * v1); + mu2 = sqrtf(v0 * v0 + v2 * v2); + + if(mu1 > fTiny) + { + V1 = v1 / mu1; + V3 = v0 / mu1; + } + + if(mu2 > fTiny) + { + V2 = v2 / mu2; + V4 = v0 / mu2; + } + + } else + { + if(M1 > M3) + { + V1 = V4 = 1.0f; + V2 = V3 = 0.0f; + + } else + { + V1 = V4 = 0.0f; + V2 = V3 = 1.0f; + } + } + + // Compute prox_p of the diagonal entries + sig1_upd = sig2_upd = 0.0f; + + if(p == 1) + { + sig1_upd = MAX(sig1 - divsigma, 0.0f); + sig2_upd = MAX(sig2 - divsigma, 0.0f); + + } else if(p == INFNORM) + { + proj[0] = sigma * fabs(sig1); + proj[1] = sigma * fabs(sig2); + + /*l1 projection part */ + sum = fLarge; + num = 0l; + shrinkfactor = 0.0f; + while(sum > 1.0f) + { + sum = 0.0f; + num = 0; + + for(ii = 0; ii < 2; ii++) + { + proj[ii] = MAX(proj[ii] - shrinkfactor, 0.0f); + + sum += fabs(proj[ii]); + if(proj[ii]!= 0.0f) + num++; + } + + if(num > 0) + shrinkfactor = (sum - 1.0f) / num; + else + break; + } + /*l1 proj ends*/ + + sig1_upd = sig1 - divsigma * proj[0]; + sig2_upd = sig2 - divsigma * proj[1]; + } + + // Compute the diagonal entries of $\widehat{\Sigma}\Sigma^{\dagger}_0$ + if(sig1 > fTiny) + sig1_upd /= sig1; + + if(sig2 > fTiny) + sig2_upd /= sig2; + + // Compute solution + t[0] = sig1_upd * V1 * V1 + sig2_upd * V2 * V2; + t[1] = sig1_upd * V1 * V3 + sig2_upd * V2 * V4; + t[2] = sig1_upd * V3 * V3 + sig2_upd * V4 * V4; +} + + +#include "hw_sched.h" +typedef struct { + int offY, stepY, copY; + float *Input, *u, *u_upd, *qx, *qy, *qx_upd, *qy_upd, *gradx, *grady, *gradx_upd, *grady_upd; + float *div, *div_upd; + float resprimal, resdual; + float unorm, qnorm, product; +} tnv_thread_t; + +typedef struct { + int threads; + tnv_thread_t *thr_ctx; + float *InputT, *uT; + int dimX, dimY, dimZ, padZ; + float lambda, sigma, tau, theta; +} tnv_context_t; + +HWSched sched = NULL; +tnv_context_t tnv_ctx; + + +static int tnv_free(HWThread thr, void *hwctx, int device_id, void *data) { + int i,j,k; + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + free(ctx->Input); + free(ctx->u); + free(ctx->u_upd); + free(ctx->qx); + free(ctx->qy); + free(ctx->qx_upd); + free(ctx->qy_upd); + free(ctx->gradx); + free(ctx->grady); + free(ctx->gradx_upd); + free(ctx->grady_upd); + free(ctx->div); + free(ctx->div_upd); + + return 0; +} + +static int tnv_init(HWThread thr, void *hwctx, int device_id, void *data) { + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int padZ = tnv_ctx->padZ; + int offY = ctx->offY; + int stepY = ctx->stepY; + +// printf("%i %p - %i %i %i x %i %i\n", device_id, ctx, dimX, dimY, dimZ, offY, stepY); + + long DimTotal = (long)(dimX*stepY*padZ); + long Dim1Total = (long)(dimX*(stepY+1)*padZ); + + // Auxiliar vectors + ctx->Input = memalign(64, Dim1Total * sizeof(float)); + ctx->u = memalign(64, Dim1Total * sizeof(float)); + ctx->u_upd = memalign(64, Dim1Total * sizeof(float)); + ctx->qx = memalign(64, DimTotal * sizeof(float)); + ctx->qy = memalign(64, DimTotal * sizeof(float)); + ctx->qx_upd = memalign(64, DimTotal * sizeof(float)); + ctx->qy_upd = memalign(64, DimTotal * sizeof(float)); + ctx->gradx = memalign(64, DimTotal * sizeof(float)); + ctx->grady = memalign(64, DimTotal * sizeof(float)); + ctx->gradx_upd = memalign(64, DimTotal * sizeof(float)); + ctx->grady_upd = memalign(64, DimTotal * sizeof(float)); + ctx->div = memalign(64, Dim1Total * sizeof(float)); + ctx->div_upd = malloc(Dim1Total * sizeof(float)); + + if ((!ctx->Input)||(!ctx->u)||(!ctx->u_upd)||(!ctx->qx)||(!ctx->qy)||(!ctx->qx_upd)||(!ctx->qy_upd)||(!ctx->gradx)||(!ctx->grady)||(!ctx->gradx_upd)||(!ctx->grady_upd)||(!ctx->div)||(!ctx->div_upd)) { + fprintf(stderr, "Error allocating memory\n"); + exit(-1); + } + + return 0; +} + +static int tnv_start(HWThread thr, void *hwctx, int device_id, void *data) { + int i,j,k; + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int padZ = tnv_ctx->padZ; + int offY = ctx->offY; + int stepY = ctx->stepY; + int copY = ctx->copY; + +// printf("%i %p - %i %i %i (%i) x %i %i\n", device_id, ctx, dimX, dimY, dimZ, padZ, offY, stepY); + + long DimTotal = (long)(dimX*stepY*padZ); + long Dim1Total = (long)(dimX*copY*padZ); + + memset(ctx->u, 0, Dim1Total * sizeof(float)); + memset(ctx->qx, 0, DimTotal * sizeof(float)); + memset(ctx->qy, 0, DimTotal * sizeof(float)); + memset(ctx->gradx, 0, DimTotal * sizeof(float)); + memset(ctx->grady, 0, DimTotal * sizeof(float)); + memset(ctx->div, 0, Dim1Total * sizeof(float)); + + for(k=0; k<dimZ; k++) { + for(j=0; j<copY; j++) { + for(i=0; i<dimX; i++) { + ctx->Input[j * dimX * padZ + i * padZ + k] = tnv_ctx->InputT[k * dimX * dimY + (j + offY) * dimX + i]; + ctx->u[j * dimX * padZ + i * padZ + k] = tnv_ctx->uT[k * dimX * dimY + (j + offY) * dimX + i]; + } + } + } + + return 0; +} + +static int tnv_finish(HWThread thr, void *hwctx, int device_id, void *data) { + int i,j,k; + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int padZ = tnv_ctx->padZ; + int offY = ctx->offY; + int stepY = ctx->stepY; + int copY = ctx->copY; + + for(k=0; k<dimZ; k++) { + for(j=0; j<stepY; j++) { + for(i=0; i<dimX; i++) { + tnv_ctx->uT[k * dimX * dimY + (j + offY) * dimX + i] = ctx->u[j * dimX * padZ + i * padZ + k]; + } + } + } + + return 0; +} + + +static int tnv_copy(HWThread thr, void *hwctx, int device_id, void *data) { + int i,j,k; + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int stepY = ctx->stepY; + int copY = ctx->copY; + int padZ = tnv_ctx->padZ; + long DimTotal = (long)(dimX*stepY*padZ); + long Dim1Total = (long)(dimX*copY*padZ); + + // Auxiliar vectors + memcpy(ctx->u, ctx->u_upd, Dim1Total * sizeof(float)); + memcpy(ctx->qx, ctx->qx_upd, DimTotal * sizeof(float)); + memcpy(ctx->qy, ctx->qy_upd, DimTotal * sizeof(float)); + memcpy(ctx->gradx, ctx->gradx_upd, DimTotal * sizeof(float)); + memcpy(ctx->grady, ctx->grady_upd, DimTotal * sizeof(float)); + memcpy(ctx->div, ctx->div_upd, Dim1Total * sizeof(float)); + + return 0; +} + +static int tnv_restore(HWThread thr, void *hwctx, int device_id, void *data) { + int i,j,k; + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int stepY = ctx->stepY; + int copY = ctx->copY; + int padZ = tnv_ctx->padZ; + long DimTotal = (long)(dimX*stepY*padZ); + long Dim1Total = (long)(dimX*copY*padZ); + + // Auxiliar vectors + memcpy(ctx->u_upd, ctx->u, Dim1Total * sizeof(float)); + memcpy(ctx->qx_upd, ctx->qx, DimTotal * sizeof(float)); + memcpy(ctx->qy_upd, ctx->qy, DimTotal * sizeof(float)); + memcpy(ctx->gradx_upd, ctx->gradx, DimTotal * sizeof(float)); + memcpy(ctx->grady_upd, ctx->grady, DimTotal * sizeof(float)); + memcpy(ctx->div_upd, ctx->div, Dim1Total * sizeof(float)); + + return 0; +} + + +static int tnv_step(HWThread thr, void *hwctx, int device_id, void *data) { + long i, j, k, l; + + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int padZ = tnv_ctx->padZ; + int offY = ctx->offY; + int stepY = ctx->stepY; + int copY = ctx->copY; + + float *Input = ctx->Input; + float *u = ctx->u; + float *u_upd = ctx->u_upd; + float *qx = ctx->qx; + float *qy = ctx->qy; + float *qx_upd = ctx->qx_upd; + float *qy_upd = ctx->qy_upd; + float *gradx = ctx->gradx; + float *grady = ctx->grady; + float *gradx_upd = ctx->gradx_upd; + float *grady_upd = ctx->grady_upd; + float *div = ctx->div; + float *div_upd = ctx->div_upd; + + long p = 1l; + long q = 1l; + long r = 0l; + + float lambda = tnv_ctx->lambda; + float sigma = tnv_ctx->sigma; + float tau = tnv_ctx->tau; + float theta = tnv_ctx->theta; + + float taulambda = tau * lambda; + float divtau = 1.0f / tau; + float divsigma = 1.0f / sigma; + float theta1 = 1.0f + theta; + float constant = 1.0f + taulambda; + + float resprimal = 0.0; + float resdual = 0.0; + float product = 0.0; + float unorm = 0.0; + float qnorm = 0.0; + + float udiff[dimZ] __attribute__((aligned(64))); + float qxdiff __attribute__((aligned(64))); + float qydiff __attribute__((aligned(64))); + float divdiff __attribute__((aligned(64))); + float gradxdiff[dimZ] __attribute__((aligned(64))); + float gradydiff[dimZ] __attribute__((aligned(64))); + + for(int j1 = 0; j1 < stepY; j1 += BLOCK) { + for(int i1 = 0; i1 < dimX; i1 += BLOCK) { + for(int j2 = 0; j2 < BLOCK; j2++) { + j = j1 + j2; + for(int i2 = 0; i2 < BLOCK; i2++) { + float t[3]; + float M1 = 0.0f, M2 = 0.0f, M3 = 0.0f; + + i = i1 + i2; + if (i == dimX) break; + if (j == stepY) { j2 = BLOCK; break; } + l = (j * dimX + i) * padZ; + +//#pragma vector aligned +#pragma GCC ivdep + for(k = 0; k < dimZ; k++) { + u_upd[l + k] = (u[l + k] + tau * div[l + k] + taulambda * Input[l + k]) / constant; + udiff[k] = u[l + k] - u_upd[l + k]; + unorm += (udiff[k] * udiff[k]); + + gradx_upd[l + k] = (i == (dimX - 1))?0:((u[l + k + padZ] + tau * div[l + k + padZ] + taulambda * Input[l + k + padZ]) / constant - u_upd[l + k]); + grady_upd[l + k] = (j == (copY - 1))?0:((u[l + k + dimX*padZ] + tau * div[l + k + dimX*padZ] + taulambda * Input[l + k + dimX*padZ]) / constant - u_upd[l + k]); + gradxdiff[k] = gradx[l + k] - gradx_upd[l + k]; + gradydiff[k] = grady[l + k] - grady_upd[l + k]; + + float ubarx = theta1 * gradx_upd[l + k] - theta * gradx[l + k]; + float ubary = theta1 * grady_upd[l + k] - theta * grady[l + k]; +//#define TNV_NEW_STYLE +#ifdef TNV_NEW_STYLE + qx_upd[l + k] = qx[l + k] + sigma * ubarx; + qy_upd[l + k] = qy[l + k] + sigma * ubary; + + float vx = divsigma * qx_upd[l + k]; //+ ubarx + float vy = divsigma * qy_upd[l + k]; //+ ubary +#else + float vx = ubarx + divsigma * qx[l + k]; + float vy = ubary + divsigma * qy[l + k]; +#endif + + M1 += (vx * vx); M2 += (vx * vy); M3 += (vy * vy); + } + + coefF(t, M1, M2, M3, sigma, p, q, r); + +//#pragma vector aligned +#pragma GCC ivdep + for(k = 0; k < dimZ; k++) { +#ifdef TNV_NEW_STYLE + float vx = divsigma * qx_upd[l + k]; + float vy = divsigma * qy_upd[l + k]; + + float gx_upd = vx * t[0] + vy * t[1]; + float gy_upd = vx * t[1] + vy * t[2]; + + qx_upd[l + k] -= sigma * gx_upd; + qy_upd[l + k] -= sigma * gy_upd; +#else + float ubarx = theta1 * gradx_upd[l + k] - theta * gradx[l + k]; + float ubary = theta1 * grady_upd[l + k] - theta * grady[l + k]; + float vx = ubarx + divsigma * qx[l + k]; + float vy = ubary + divsigma * qy[l + k]; + + float gx_upd = vx * t[0] + vy * t[1]; + float gy_upd = vx * t[1] + vy * t[2]; + + qx_upd[l + k] = qx[l + k] + sigma * (ubarx - gx_upd); + qy_upd[l + k] = qy[l + k] + sigma * (ubary - gy_upd); +#endif + + float div_upd_val = 0; + div_upd_val -= (i > 0)?qx_upd[l + k - padZ]:0; + div_upd_val -= (j > 0)?qy_upd[l + k - dimX * padZ]:0; + div_upd_val += (i < (dimX-1))?qx_upd[l + k]:0; + div_upd_val += (j < (copY-1))?qy_upd[l + k]:0; + div_upd[l + k] = div_upd_val; + + qxdiff = qx[l + k] - qx_upd[l + k]; + qydiff = qy[l + k] - qy_upd[l + k]; + qnorm += (qxdiff * qxdiff + qydiff * qydiff); + + resdual += fabs(divsigma * qxdiff - gradxdiff[k]); + resdual += fabs(divsigma * qydiff - gradydiff[k]); + product += (gradxdiff[k] * qxdiff + gradydiff[k] * qydiff); + + if ((offY == 0)||(j > 0)) { + divdiff = div[l + k] - div_upd[l + k]; // Multiple steps... How we compute without history? + resprimal += fabs(divtau * udiff[k] + divdiff); + } + } + + } // i + } // j + } // i + } // j + + + ctx->resprimal = resprimal; + ctx->resdual = resdual; + ctx->product = product; + ctx->unorm = unorm; + ctx->qnorm = qnorm; + + return 0; +} + +static void TNV_CPU_init(float *InputT, float *uT, int dimX, int dimY, int dimZ) { + int i, off, size, err; + + if (sched) return; + + tnv_ctx.dimX = dimX; + tnv_ctx.dimY = dimY; + tnv_ctx.dimZ = dimZ; + // Padding seems actually slower + tnv_ctx.padZ = dimZ; +// tnv_ctx.padZ = 16 * ((dimZ / 16) + ((dimZ % 16)?1:0)); + + hw_sched_init(); + + int threads = hw_sched_get_cpu_count(); + if (threads > dimY) threads = dimY/2; + + int step = dimY / threads; + int extra = dimY % threads; + + tnv_ctx.threads = threads; + tnv_ctx.thr_ctx = (tnv_thread_t*)calloc(threads, sizeof(tnv_thread_t)); + for (i = 0, off = 0; i < threads; i++, off += size) { + tnv_thread_t *ctx = tnv_ctx.thr_ctx + i; + size = step + ((i < extra)?1:0); + + ctx->offY = off; + ctx->stepY = size; + + if (i == (threads-1)) ctx->copY = ctx->stepY; + else ctx->copY = ctx->stepY + 1; + } + + sched = hw_sched_create(threads); + if (!sched) { fprintf(stderr, "Error creating threads\n"); exit(-1); } + + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_init); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling init threads", err); exit(-1); } +} + + + +/* + * C-OMP implementation of Total Nuclear Variation regularisation model (2D + channels) [1] + * The code is modified from the implementation by Joan Duran <joan.duran@uib.es> see + * "denoisingPDHG_ipol.cpp" in Joans Collaborative Total Variation package + * + * Input Parameters: + * 1. Noisy volume of 2D + channel dimension, i.e. 3D volume + * 2. lambda - regularisation parameter + * 3. Number of iterations [OPTIONAL parameter] + * 4. eplsilon - tolerance constant [OPTIONAL parameter] + * 5. print information: 0 (off) or 1 (on) [OPTIONAL parameter] + * + * Output: + * 1. Filtered/regularized image (u) + * + * [1]. Duran, J., Moeller, M., Sbert, C. and Cremers, D., 2016. Collaborative total variation: a general framework for vectorial TV models. SIAM Journal on Imaging Sciences, 9(1), pp.116-151. + */ + +float TNV_CPU_main(float *InputT, float *uT, float lambda, int maxIter, float tol, int dimX, int dimY, int dimZ) +{ + int err; + int iter; + int i,j,k,l,m; + + lambda = 1.0f/(2.0f*lambda); + tnv_ctx.lambda = lambda; + + // PDHG algorithm parameters + float tau = 0.5f; + float sigma = 0.5f; + float theta = 1.0f; + + // Backtracking parameters + float s = 1.0f; + float gamma = 0.75f; + float beta = 0.95f; + float alpha0 = 0.2f; + float alpha = alpha0; + float delta = 1.5f; + float eta = 0.95f; + + TNV_CPU_init(InputT, uT, dimX, dimY, dimZ); + + tnv_ctx.InputT = InputT; + tnv_ctx.uT = uT; + + int padZ = tnv_ctx.padZ; + + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_start); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling start threads", err); exit(-1); } + + + // Apply Primal-Dual Hybrid Gradient scheme + float residual = fLarge; + for(iter = 0; iter < maxIter; iter++) { + float resprimal = 0.0f; + float resdual = 0.0f; + float product = 0.0f; + float unorm = 0.0f; + float qnorm = 0.0f; + + float divtau = 1.0f / tau; + + tnv_ctx.sigma = sigma; + tnv_ctx.tau = tau; + tnv_ctx.theta = theta; + + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_step); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling tnv threads", err); exit(-1); } + + // border regions + for (j = 1; j < tnv_ctx.threads; j++) { + tnv_thread_t *ctx0 = tnv_ctx.thr_ctx + (j - 1); + tnv_thread_t *ctx = tnv_ctx.thr_ctx + j; + + m = ctx0->stepY * dimX * padZ; + for(i = 0; i < dimX; i++) { + for(k = 0; k < dimZ; k++) { + int l = i * padZ + k; + + ctx->div_upd[l] -= ctx0->qy_upd[m - dimX * padZ + l]; + ctx0->div_upd[m + l] = ctx->div_upd[l]; + ctx0->u_upd[m + l] = ctx->u_upd[l]; + + float divdiff = ctx->div[l] - ctx->div_upd[l]; + float udiff = ctx->u[l] - ctx->u_upd[l]; + resprimal += fabs(divtau * udiff + divdiff); + } + } + } + + for (j = 0; j < tnv_ctx.threads; j++) { + tnv_thread_t *ctx = tnv_ctx.thr_ctx + j; + resprimal += ctx->resprimal; + resdual += ctx->resdual; + product += ctx->product; + unorm += ctx->unorm; + qnorm += ctx->qnorm; + } + + residual = (resprimal + resdual) / ((float) (dimX*dimY*dimZ)); + float b = (2.0f * tau * sigma * product) / (gamma * sigma * unorm + gamma * tau * qnorm); + float dual_dot_delta = resdual * s * delta; + float dual_div_delta = (resdual * s) / delta; + printf("resprimal: %f, resdual: %f, b: %f (product: %f, unorm: %f, qnorm: %f)\n", resprimal, resdual, b, product, unorm, qnorm); + + + if(b > 1) { + // Decrease step-sizes to fit balancing principle + tau = (beta * tau) / b; + sigma = (beta * sigma) / b; + alpha = alpha0; + + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_restore); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling restore threads", err); exit(-1); } + } else { + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_copy); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling copy threads", err); exit(-1); } + + if(resprimal > dual_dot_delta) { + // Increase primal step-size and decrease dual step-size + tau = tau / (1.0f - alpha); + sigma = sigma * (1.0f - alpha); + alpha = alpha * eta; + } else if(resprimal < dual_div_delta) { + // Decrease primal step-size and increase dual step-size + tau = tau * (1.0f - alpha); + sigma = sigma / (1.0f - alpha); + alpha = alpha * eta; + } + } + + if (residual < tol) break; + } + + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_finish); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling finish threads", err); exit(-1); } + + + printf("Iterations stopped at %i with the residual %f \n", iter, residual); + printf("Return: %f\n", *uT); + + return *uT; +} diff --git a/src/Core/performance_CPU/TNV_core.c.v19 b/src/Core/performance_CPU/TNV_core.c.v19 new file mode 100755 index 0000000..9b19ed5 --- /dev/null +++ b/src/Core/performance_CPU/TNV_core.c.v19 @@ -0,0 +1,681 @@ +/* + * This work is part of the Core Imaging Library developed by + * Visual Analytics and Imaging System Group of the Science Technology + * Facilities Council, STFC + * + * Copyright 2017 Daniil Kazantsev + * Copyright 2017 Srikanth Nagella, Edoardo Pasca + * + * Copyriht 2020 Suren A. Chlingaryan + * Optimized version with 1/2 of memory consumption and ~4x performance + * This version is algorithmicly comptable with the original, but slight change in results + * is expected due to different order of floating-point operations. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * http://www.apache.org/licenses/LICENSE-2.0 + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include <malloc.h> +#include "TNV_core.h" + +#define BLOCK 32 +#define min(a,b) (((a)<(b))?(a):(b)) + +inline void coefF(float *t, float M1, float M2, float M3, float sigma, int p, int q, int r) { + int ii, num; + float divsigma = 1.0f / sigma; + float sum, shrinkfactor; + float T,D,det,eig1,eig2,sig1,sig2,V1, V2, V3, V4, v0,v1,v2, mu1,mu2,sig1_upd,sig2_upd; + float proj[2] = {0}; + + // Compute eigenvalues of M + T = M1 + M3; + D = M1 * M3 - M2 * M2; + det = sqrtf(MAX((T * T / 4.0f) - D, 0.0f)); + eig1 = MAX((T / 2.0f) + det, 0.0f); + eig2 = MAX((T / 2.0f) - det, 0.0f); + sig1 = sqrtf(eig1); + sig2 = sqrtf(eig2); + + // Compute normalized eigenvectors + V1 = V2 = V3 = V4 = 0.0f; + + if(M2 != 0.0f) + { + v0 = M2; + v1 = eig1 - M3; + v2 = eig2 - M3; + + mu1 = sqrtf(v0 * v0 + v1 * v1); + mu2 = sqrtf(v0 * v0 + v2 * v2); + + if(mu1 > fTiny) + { + V1 = v1 / mu1; + V3 = v0 / mu1; + } + + if(mu2 > fTiny) + { + V2 = v2 / mu2; + V4 = v0 / mu2; + } + + } else + { + if(M1 > M3) + { + V1 = V4 = 1.0f; + V2 = V3 = 0.0f; + + } else + { + V1 = V4 = 0.0f; + V2 = V3 = 1.0f; + } + } + + // Compute prox_p of the diagonal entries + sig1_upd = sig2_upd = 0.0f; + + if(p == 1) + { + sig1_upd = MAX(sig1 - divsigma, 0.0f); + sig2_upd = MAX(sig2 - divsigma, 0.0f); + + } else if(p == INFNORM) + { + proj[0] = sigma * fabs(sig1); + proj[1] = sigma * fabs(sig2); + + /*l1 projection part */ + sum = fLarge; + num = 0l; + shrinkfactor = 0.0f; + while(sum > 1.0f) + { + sum = 0.0f; + num = 0; + + for(ii = 0; ii < 2; ii++) + { + proj[ii] = MAX(proj[ii] - shrinkfactor, 0.0f); + + sum += fabs(proj[ii]); + if(proj[ii]!= 0.0f) + num++; + } + + if(num > 0) + shrinkfactor = (sum - 1.0f) / num; + else + break; + } + /*l1 proj ends*/ + + sig1_upd = sig1 - divsigma * proj[0]; + sig2_upd = sig2 - divsigma * proj[1]; + } + + // Compute the diagonal entries of $\widehat{\Sigma}\Sigma^{\dagger}_0$ + if(sig1 > fTiny) + sig1_upd /= sig1; + + if(sig2 > fTiny) + sig2_upd /= sig2; + + // Compute solution + t[0] = sig1_upd * V1 * V1 + sig2_upd * V2 * V2; + t[1] = sig1_upd * V1 * V3 + sig2_upd * V2 * V4; + t[2] = sig1_upd * V3 * V3 + sig2_upd * V4 * V4; +} + + +#include "hw_sched.h" +typedef struct { + int offY, stepY, copY; + float *Input, *u, *u_upd, *qx, *qy, *qx_upd, *qy_upd, *gradx, *grady, *gradx_upd, *grady_upd; + float *div, *div_upd; + float resprimal, resdual; + float unorm, qnorm, product; +} tnv_thread_t; + +typedef struct { + int threads; + tnv_thread_t *thr_ctx; + float *InputT, *uT; + int dimX, dimY, dimZ, padZ; + float lambda, sigma, tau, theta; +} tnv_context_t; + +HWSched sched = NULL; +tnv_context_t tnv_ctx; + + +static int tnv_free(HWThread thr, void *hwctx, int device_id, void *data) { + int i,j,k; + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + free(ctx->Input); + free(ctx->u); + free(ctx->u_upd); + free(ctx->qx); + free(ctx->qy); + free(ctx->qx_upd); + free(ctx->qy_upd); + free(ctx->gradx); + free(ctx->grady); + free(ctx->gradx_upd); + free(ctx->grady_upd); + free(ctx->div); + free(ctx->div_upd); + + return 0; +} + +static int tnv_init(HWThread thr, void *hwctx, int device_id, void *data) { + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int padZ = tnv_ctx->padZ; + int offY = ctx->offY; + int stepY = ctx->stepY; + +// printf("%i %p - %i %i %i x %i %i\n", device_id, ctx, dimX, dimY, dimZ, offY, stepY); + + long DimTotal = (long)(dimX*stepY*padZ); + long Dim1Total = (long)(dimX*(stepY+1)*padZ); + + // Auxiliar vectors + ctx->Input = memalign(64, Dim1Total * sizeof(float)); + ctx->u = memalign(64, Dim1Total * sizeof(float)); + ctx->u_upd = memalign(64, Dim1Total * sizeof(float)); + ctx->qx = memalign(64, DimTotal * sizeof(float)); + ctx->qy = memalign(64, DimTotal * sizeof(float)); + ctx->qx_upd = memalign(64, DimTotal * sizeof(float)); + ctx->qy_upd = memalign(64, DimTotal * sizeof(float)); + ctx->gradx = memalign(64, DimTotal * sizeof(float)); + ctx->grady = memalign(64, DimTotal * sizeof(float)); + ctx->gradx_upd = memalign(64, DimTotal * sizeof(float)); + ctx->grady_upd = memalign(64, DimTotal * sizeof(float)); + ctx->div = memalign(64, Dim1Total * sizeof(float)); + ctx->div_upd = malloc(Dim1Total * sizeof(float)); + + if ((!ctx->Input)||(!ctx->u)||(!ctx->u_upd)||(!ctx->qx)||(!ctx->qy)||(!ctx->qx_upd)||(!ctx->qy_upd)||(!ctx->gradx)||(!ctx->grady)||(!ctx->gradx_upd)||(!ctx->grady_upd)||(!ctx->div)||(!ctx->div_upd)) { + fprintf(stderr, "Error allocating memory\n"); + exit(-1); + } + + return 0; +} + +static int tnv_start(HWThread thr, void *hwctx, int device_id, void *data) { + int i,j,k; + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int padZ = tnv_ctx->padZ; + int offY = ctx->offY; + int stepY = ctx->stepY; + int copY = ctx->copY; + +// printf("%i %p - %i %i %i (%i) x %i %i\n", device_id, ctx, dimX, dimY, dimZ, padZ, offY, stepY); + + long DimTotal = (long)(dimX*stepY*padZ); + long Dim1Total = (long)(dimX*copY*padZ); + + memset(ctx->u, 0, Dim1Total * sizeof(float)); + memset(ctx->qx, 0, DimTotal * sizeof(float)); + memset(ctx->qy, 0, DimTotal * sizeof(float)); + memset(ctx->gradx, 0, DimTotal * sizeof(float)); + memset(ctx->grady, 0, DimTotal * sizeof(float)); + memset(ctx->div, 0, Dim1Total * sizeof(float)); + + for(k=0; k<dimZ; k++) { + for(j=0; j<copY; j++) { + for(i=0; i<dimX; i++) { + ctx->Input[j * dimX * padZ + i * padZ + k] = tnv_ctx->InputT[k * dimX * dimY + (j + offY) * dimX + i]; + ctx->u[j * dimX * padZ + i * padZ + k] = tnv_ctx->uT[k * dimX * dimY + (j + offY) * dimX + i]; + } + } + } + + return 0; +} + +static int tnv_finish(HWThread thr, void *hwctx, int device_id, void *data) { + int i,j,k; + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int padZ = tnv_ctx->padZ; + int offY = ctx->offY; + int stepY = ctx->stepY; + int copY = ctx->copY; + + for(k=0; k<dimZ; k++) { + for(j=0; j<stepY; j++) { + for(i=0; i<dimX; i++) { + tnv_ctx->uT[k * dimX * dimY + (j + offY) * dimX + i] = ctx->u[j * dimX * padZ + i * padZ + k]; + } + } + } + + return 0; +} + + +static int tnv_copy(HWThread thr, void *hwctx, int device_id, void *data) { + int i,j,k; + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int stepY = ctx->stepY; + int copY = ctx->copY; + int padZ = tnv_ctx->padZ; + long DimTotal = (long)(dimX*stepY*padZ); + long Dim1Total = (long)(dimX*copY*padZ); + + // Auxiliar vectors + memcpy(ctx->u, ctx->u_upd, Dim1Total * sizeof(float)); + memcpy(ctx->qx, ctx->qx_upd, DimTotal * sizeof(float)); + memcpy(ctx->qy, ctx->qy_upd, DimTotal * sizeof(float)); + memcpy(ctx->gradx, ctx->gradx_upd, DimTotal * sizeof(float)); + memcpy(ctx->grady, ctx->grady_upd, DimTotal * sizeof(float)); + memcpy(ctx->div, ctx->div_upd, Dim1Total * sizeof(float)); + + return 0; +} + +static int tnv_restore(HWThread thr, void *hwctx, int device_id, void *data) { + int i,j,k; + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int stepY = ctx->stepY; + int copY = ctx->copY; + int padZ = tnv_ctx->padZ; + long DimTotal = (long)(dimX*stepY*padZ); + long Dim1Total = (long)(dimX*copY*padZ); + + // Auxiliar vectors + memcpy(ctx->u_upd, ctx->u, Dim1Total * sizeof(float)); + memcpy(ctx->qx_upd, ctx->qx, DimTotal * sizeof(float)); + memcpy(ctx->qy_upd, ctx->qy, DimTotal * sizeof(float)); + memcpy(ctx->gradx_upd, ctx->gradx, DimTotal * sizeof(float)); + memcpy(ctx->grady_upd, ctx->grady, DimTotal * sizeof(float)); + memcpy(ctx->div_upd, ctx->div, Dim1Total * sizeof(float)); + + return 0; +} + + +static int tnv_step(HWThread thr, void *hwctx, int device_id, void *data) { + long i, j, k, l; + long i1, i2, j1, j2; + + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int padZ = tnv_ctx->padZ; + int offY = ctx->offY; + int stepY = ctx->stepY; + int copY = ctx->copY; + + float *Input = ctx->Input; + float *u = ctx->u; + float *u_upd = ctx->u_upd; + float *qx = ctx->qx; + float *qy = ctx->qy; + float *qx_upd = ctx->qx_upd; + float *qy_upd = ctx->qy_upd; + float *gradx = ctx->gradx; + float *grady = ctx->grady; + float *gradx_upd = ctx->gradx_upd; + float *grady_upd = ctx->grady_upd; + float *div = ctx->div; + float *div_upd = ctx->div_upd; + + long p = 1l; + long q = 1l; + long r = 0l; + + float lambda = tnv_ctx->lambda; + float sigma = tnv_ctx->sigma; + float tau = tnv_ctx->tau; + float theta = tnv_ctx->theta; + + float taulambda = tau * lambda; + float divtau = 1.0f / tau; + float divsigma = 1.0f / sigma; + float theta1 = 1.0f + theta; + float constant = 1.0f + taulambda; + + float resprimal = 0.0f; + float resdual1 = 0.0f; + float resdual2 = 0.0f; + float product = 0.0f; + float unorm = 0.0f; + float qnorm = 0.0f; + + float udiff[dimZ] __attribute__((aligned(64))); + float qxdiff __attribute__((aligned(64))); + float qydiff __attribute__((aligned(64))); + float divdiff __attribute__((aligned(64))); + float gradxdiff[dimZ] __attribute__((aligned(64))); + float gradydiff[dimZ] __attribute__((aligned(64))); + + + j = 0; { +# define TNV_LOOP_FIRST_J + i = 0; { +# define TNV_LOOP_FIRST_I +# include "TNV_core_backtrack_loop.h" +# undef TNV_LOOP_FIRST_I + } + for(i = 1; i < (dimX - 1); i++) { +# include "TNV_core_backtrack_loop.h" + } + i = dimX - 1; { +# define TNV_LOOP_LAST_I +# include "TNV_core_backtrack_loop.h" +# undef TNV_LOOP_LAST_I + } +# undef TNV_LOOP_FIRST_J + } + + + + for(int j = 1; j < (copY - 1); j++) { + i = 0; { +# define TNV_LOOP_FIRST_I +# include "TNV_core_backtrack_loop.h" +# undef TNV_LOOP_FIRST_I + } + } + + for(int j1 = 1; j1 < (copY - 1); j1 += BLOCK) { + for(int i1 = 1; i1 < (dimX - 1); i1 += BLOCK) { + for(int j2 = 0; j2 < BLOCK; j2 ++) { + j = j1 + j2; + for(int i2 = 0; i2 < BLOCK; i2++) { + i = i1 + i2; + + if (i == (dimX - 1)) break; + if (j == (copY - 1)) { j2 = BLOCK; break; } +# include "TNV_core_backtrack_loop.h" + } + } + } // i + + } + + for(int j = 1; j < (copY - 1); j++) { + i = dimX - 1; { +# define TNV_LOOP_LAST_I +# include "TNV_core_backtrack_loop.h" +# undef TNV_LOOP_LAST_I + } + } + + + + for (j = copY - 1; j < stepY; j++) { +# define TNV_LOOP_LAST_J + i = 0; { +# define TNV_LOOP_FIRST_I +# include "TNV_core_backtrack_loop.h" +# undef TNV_LOOP_FIRST_I + } + for(i = 1; i < (dimX - 1); i++) { +# include "TNV_core_backtrack_loop.h" + } + i = dimX - 1; { +# define TNV_LOOP_LAST_I +# include "TNV_core_backtrack_loop.h" +# undef TNV_LOOP_LAST_I + } +# undef TNV_LOOP_LAST_J + } + + + ctx->resprimal = resprimal; + ctx->resdual = resdual1 + resdual2; + ctx->product = product; + ctx->unorm = unorm; + ctx->qnorm = qnorm; + + return 0; +} + +static void TNV_CPU_init(float *InputT, float *uT, int dimX, int dimY, int dimZ) { + int i, off, size, err; + + if (sched) return; + + tnv_ctx.dimX = dimX; + tnv_ctx.dimY = dimY; + tnv_ctx.dimZ = dimZ; + // Padding seems actually slower + tnv_ctx.padZ = dimZ; +// tnv_ctx.padZ = 16 * ((dimZ / 16) + ((dimZ % 16)?1:0)); + + hw_sched_init(); + + int threads = hw_sched_get_cpu_count(); + if (threads > dimY) threads = dimY/2; + + int step = dimY / threads; + int extra = dimY % threads; + + tnv_ctx.threads = threads; + tnv_ctx.thr_ctx = (tnv_thread_t*)calloc(threads, sizeof(tnv_thread_t)); + for (i = 0, off = 0; i < threads; i++, off += size) { + tnv_thread_t *ctx = tnv_ctx.thr_ctx + i; + size = step + ((i < extra)?1:0); + + ctx->offY = off; + ctx->stepY = size; + + if (i == (threads-1)) ctx->copY = ctx->stepY; + else ctx->copY = ctx->stepY + 1; + } + + sched = hw_sched_create(threads); + if (!sched) { fprintf(stderr, "Error creating threads\n"); exit(-1); } + + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_init); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling init threads", err); exit(-1); } +} + + + +/* + * C-OMP implementation of Total Nuclear Variation regularisation model (2D + channels) [1] + * The code is modified from the implementation by Joan Duran <joan.duran@uib.es> see + * "denoisingPDHG_ipol.cpp" in Joans Collaborative Total Variation package + * + * Input Parameters: + * 1. Noisy volume of 2D + channel dimension, i.e. 3D volume + * 2. lambda - regularisation parameter + * 3. Number of iterations [OPTIONAL parameter] + * 4. eplsilon - tolerance constant [OPTIONAL parameter] + * 5. print information: 0 (off) or 1 (on) [OPTIONAL parameter] + * + * Output: + * 1. Filtered/regularized image (u) + * + * [1]. Duran, J., Moeller, M., Sbert, C. and Cremers, D., 2016. Collaborative total variation: a general framework for vectorial TV models. SIAM Journal on Imaging Sciences, 9(1), pp.116-151. + */ + +float TNV_CPU_main(float *InputT, float *uT, float lambda, int maxIter, float tol, int dimX, int dimY, int dimZ) +{ + int err; + int iter; + int i,j,k,l,m; + + lambda = 1.0f/(2.0f*lambda); + tnv_ctx.lambda = lambda; + + // PDHG algorithm parameters + float tau = 0.5f; + float sigma = 0.5f; + float theta = 1.0f; + + // Backtracking parameters + float s = 1.0f; + float gamma = 0.75f; + float beta = 0.95f; + float alpha0 = 0.2f; + float alpha = alpha0; + float delta = 1.5f; + float eta = 0.95f; + + TNV_CPU_init(InputT, uT, dimX, dimY, dimZ); + + tnv_ctx.InputT = InputT; + tnv_ctx.uT = uT; + + int padZ = tnv_ctx.padZ; + + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_start); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling start threads", err); exit(-1); } + + + // Apply Primal-Dual Hybrid Gradient scheme + float residual = fLarge; + for(iter = 0; iter < maxIter; iter++) { + float resprimal = 0.0f; + float resdual = 0.0f; + float product = 0.0f; + float unorm = 0.0f; + float qnorm = 0.0f; + + float divtau = 1.0f / tau; + + tnv_ctx.sigma = sigma; + tnv_ctx.tau = tau; + tnv_ctx.theta = theta; + + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_step); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling tnv threads", err); exit(-1); } + + // border regions + for (j = 1; j < tnv_ctx.threads; j++) { + tnv_thread_t *ctx0 = tnv_ctx.thr_ctx + (j - 1); + tnv_thread_t *ctx = tnv_ctx.thr_ctx + j; + + m = ctx0->stepY * dimX * padZ; + for(i = 0; i < dimX; i++) { + for(k = 0; k < dimZ; k++) { + int l = i * padZ + k; + + ctx->div_upd[l] -= ctx0->qy_upd[m - dimX * padZ + l]; + ctx0->div_upd[m + l] = ctx->div_upd[l]; + ctx0->u_upd[m + l] = ctx->u_upd[l]; + + float divdiff = ctx->div[l] - ctx->div_upd[l]; + float udiff = ctx->u[l] - ctx->u_upd[l]; + resprimal += fabs(divtau * udiff + divdiff); + } + } + } + + { + tnv_thread_t *ctx = tnv_ctx.thr_ctx + 0; + for(i = 0; i < dimX; i++) { + for(k = 0; k < dimZ; k++) { + int l = i * padZ + k; + + float divdiff = ctx->div[l] - ctx->div_upd[l]; + float udiff = ctx->u[l] - ctx->u_upd[l]; + resprimal += fabs(divtau * udiff + divdiff); + } + } + } + + for (j = 0; j < tnv_ctx.threads; j++) { + tnv_thread_t *ctx = tnv_ctx.thr_ctx + j; + resprimal += ctx->resprimal; + resdual += ctx->resdual; + product += ctx->product; + unorm += ctx->unorm; + qnorm += ctx->qnorm; + } + + residual = (resprimal + resdual) / ((float) (dimX*dimY*dimZ)); + float b = (2.0f * tau * sigma * product) / (gamma * sigma * unorm + gamma * tau * qnorm); + float dual_dot_delta = resdual * s * delta; + float dual_div_delta = (resdual * s) / delta; + printf("resprimal: %f, resdual: %f, b: %f (product: %f, unorm: %f, qnorm: %f)\n", resprimal, resdual, b, product, unorm, qnorm); + + + if(b > 1) { + // Decrease step-sizes to fit balancing principle + tau = (beta * tau) / b; + sigma = (beta * sigma) / b; + alpha = alpha0; + + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_restore); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling restore threads", err); exit(-1); } + } else { + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_copy); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling copy threads", err); exit(-1); } + + if(resprimal > dual_dot_delta) { + // Increase primal step-size and decrease dual step-size + tau = tau / (1.0f - alpha); + sigma = sigma * (1.0f - alpha); + alpha = alpha * eta; + } else if(resprimal < dual_div_delta) { + // Decrease primal step-size and increase dual step-size + tau = tau * (1.0f - alpha); + sigma = sigma / (1.0f - alpha); + alpha = alpha * eta; + } + } + + if (residual < tol) break; + } + + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_finish); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling finish threads", err); exit(-1); } + + + printf("Iterations stopped at %i with the residual %f \n", iter, residual); + printf("Return: %f\n", *uT); + + return *uT; +} diff --git a/src/Core/performance_CPU/TNV_core.c.v27 b/src/Core/performance_CPU/TNV_core.c.v27 new file mode 100755 index 0000000..dce414a --- /dev/null +++ b/src/Core/performance_CPU/TNV_core.c.v27 @@ -0,0 +1,650 @@ +/* + * This work is part of the Core Imaging Library developed by + * Visual Analytics and Imaging System Group of the Science Technology + * Facilities Council, STFC + * + * Copyright 2017 Daniil Kazantsev + * Copyright 2017 Srikanth Nagella, Edoardo Pasca + * + * Copyriht 2020 Suren A. Chlingaryan + * Optimized version with 1/3 of memory consumption and ~10x performance + * This version is not able to perform back-track except during first iterations + * But warning would be printed if backtracking is required and slower version (TNV_core_backtrack.c) + * could be executed instead. It still better than original with 1/2 of memory consumption and 4x performance gain + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * http://www.apache.org/licenses/LICENSE-2.0 + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include "TNV_core.h" + +#define BLOCK 32 +#define min(a,b) (((a)<(b))?(a):(b)) + +inline void coefF(float *t, float M1, float M2, float M3, float sigma, int p, int q, int r) { + int ii, num; + float divsigma = 1.0f / sigma; + float sum, shrinkfactor; + float T,D,det,eig1,eig2,sig1,sig2,V1, V2, V3, V4, v0,v1,v2, mu1,mu2,sig1_upd,sig2_upd; + float proj[2] = {0}; + + // Compute eigenvalues of M + T = M1 + M3; + D = M1 * M3 - M2 * M2; + det = sqrtf(MAX((T * T / 4.0f) - D, 0.0f)); + eig1 = MAX((T / 2.0f) + det, 0.0f); + eig2 = MAX((T / 2.0f) - det, 0.0f); + sig1 = sqrtf(eig1); + sig2 = sqrtf(eig2); + + // Compute normalized eigenvectors + V1 = V2 = V3 = V4 = 0.0f; + + if(M2 != 0.0f) + { + v0 = M2; + v1 = eig1 - M3; + v2 = eig2 - M3; + + mu1 = sqrtf(v0 * v0 + v1 * v1); + mu2 = sqrtf(v0 * v0 + v2 * v2); + + if(mu1 > fTiny) + { + V1 = v1 / mu1; + V3 = v0 / mu1; + } + + if(mu2 > fTiny) + { + V2 = v2 / mu2; + V4 = v0 / mu2; + } + + } else + { + if(M1 > M3) + { + V1 = V4 = 1.0f; + V2 = V3 = 0.0f; + + } else + { + V1 = V4 = 0.0f; + V2 = V3 = 1.0f; + } + } + + // Compute prox_p of the diagonal entries + sig1_upd = sig2_upd = 0.0f; + + if(p == 1) + { + sig1_upd = MAX(sig1 - divsigma, 0.0f); + sig2_upd = MAX(sig2 - divsigma, 0.0f); + + } else if(p == INFNORM) + { + proj[0] = sigma * fabs(sig1); + proj[1] = sigma * fabs(sig2); + + /*l1 projection part */ + sum = fLarge; + num = 0l; + shrinkfactor = 0.0f; + while(sum > 1.0f) + { + sum = 0.0f; + num = 0; + + for(ii = 0; ii < 2; ii++) + { + proj[ii] = MAX(proj[ii] - shrinkfactor, 0.0f); + + sum += fabs(proj[ii]); + if(proj[ii]!= 0.0f) + num++; + } + + if(num > 0) + shrinkfactor = (sum - 1.0f) / num; + else + break; + } + /*l1 proj ends*/ + + sig1_upd = sig1 - divsigma * proj[0]; + sig2_upd = sig2 - divsigma * proj[1]; + } + + // Compute the diagonal entries of $\widehat{\Sigma}\Sigma^{\dagger}_0$ + if(sig1 > fTiny) + sig1_upd /= sig1; + + if(sig2 > fTiny) + sig2_upd /= sig2; + + // Compute solution + t[0] = sig1_upd * V1 * V1 + sig2_upd * V2 * V2; + t[1] = sig1_upd * V1 * V3 + sig2_upd * V2 * V4; + t[2] = sig1_upd * V3 * V3 + sig2_upd * V4 * V4; +} + + +#include "hw_sched.h" +typedef struct { + int offY, stepY, copY; + float *Input, *u, *qx, *qy, *gradx, *grady, *div; + float *div0, *udiff0; + float resprimal, resdual; + float unorm, qnorm, product; +} tnv_thread_t; + +typedef struct { + int threads; + tnv_thread_t *thr_ctx; + float *InputT, *uT; + int dimX, dimY, dimZ, padZ; + float lambda, sigma, tau, theta; +} tnv_context_t; + +HWSched sched = NULL; +tnv_context_t tnv_ctx; + + +static int tnv_free(HWThread thr, void *hwctx, int device_id, void *data) { + int i,j,k; + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + free(ctx->Input); + free(ctx->u); + free(ctx->qx); + free(ctx->qy); + free(ctx->gradx); + free(ctx->grady); + free(ctx->div); + + free(ctx->div0); + free(ctx->udiff0); + + return 0; +} + +static int tnv_init(HWThread thr, void *hwctx, int device_id, void *data) { + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int padZ = tnv_ctx->padZ; + int offY = ctx->offY; + int stepY = ctx->stepY; + +// printf("%i %p - %i %i %i x %i %i\n", device_id, ctx, dimX, dimY, dimZ, offY, stepY); + + long DimTotal = (long)(dimX*stepY*padZ); + long Dim1Total = (long)(dimX*(stepY+1)*padZ); + long DimRow = (long)(dimX * padZ); + + // Auxiliar vectors + ctx->Input = memalign(64, Dim1Total * sizeof(float)); + ctx->u = memalign(64, Dim1Total * sizeof(float)); + ctx->qx = memalign(64, DimTotal * sizeof(float)); + ctx->qy = memalign(64, DimTotal * sizeof(float)); + ctx->gradx = memalign(64, DimTotal * sizeof(float)); + ctx->grady = memalign(64, DimTotal * sizeof(float)); + ctx->div = memalign(64, Dim1Total * sizeof(float)); + + ctx->div0 = memalign(64, DimRow * sizeof(float)); + ctx->udiff0 = memalign(64, DimRow * sizeof(float)); + + if ((!ctx->Input)||(!ctx->u)||(!ctx->qx)||(!ctx->qy)||(!ctx->gradx)||(!ctx->grady)||(!ctx->div)||(!ctx->div0)||(!ctx->udiff0)) { + fprintf(stderr, "Error allocating memory\n"); + exit(-1); + } + + return 0; +} + +static int tnv_start(HWThread thr, void *hwctx, int device_id, void *data) { + int i,j,k; + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int padZ = tnv_ctx->padZ; + int offY = ctx->offY; + int stepY = ctx->stepY; + int copY = ctx->copY; + +// printf("%i %p - %i %i %i (%i) x %i %i\n", device_id, ctx, dimX, dimY, dimZ, padZ, offY, stepY); + + long DimTotal = (long)(dimX*stepY*padZ); + long Dim1Total = (long)(dimX*copY*padZ); + + memset(ctx->u, 0, Dim1Total * sizeof(float)); + memset(ctx->qx, 0, DimTotal * sizeof(float)); + memset(ctx->qy, 0, DimTotal * sizeof(float)); + memset(ctx->gradx, 0, DimTotal * sizeof(float)); + memset(ctx->grady, 0, DimTotal * sizeof(float)); + memset(ctx->div, 0, Dim1Total * sizeof(float)); + + for(k=0; k<dimZ; k++) { + for(j=0; j<copY; j++) { + for(i=0; i<dimX; i++) { + ctx->Input[j * dimX * padZ + i * padZ + k] = tnv_ctx->InputT[k * dimX * dimY + (j + offY) * dimX + i]; + ctx->u[j * dimX * padZ + i * padZ + k] = tnv_ctx->uT[k * dimX * dimY + (j + offY) * dimX + i]; + } + } + } + + return 0; +} + +static int tnv_finish(HWThread thr, void *hwctx, int device_id, void *data) { + int i,j,k; + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int padZ = tnv_ctx->padZ; + int offY = ctx->offY; + int stepY = ctx->stepY; + int copY = ctx->copY; + + for(k=0; k<dimZ; k++) { + for(j=0; j<stepY; j++) { + for(i=0; i<dimX; i++) { + tnv_ctx->uT[k * dimX * dimY + (j + offY) * dimX + i] = ctx->u[j * dimX * padZ + i * padZ + k]; + } + } + } + + return 0; +} + + +static int tnv_restore(HWThread thr, void *hwctx, int device_id, void *data) { + int i,j,k; + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int stepY = ctx->stepY; + int copY = ctx->copY; + int padZ = tnv_ctx->padZ; + long DimTotal = (long)(dimX*stepY*padZ); + long Dim1Total = (long)(dimX*copY*padZ); + + memset(ctx->u, 0, Dim1Total * sizeof(float)); + memset(ctx->qx, 0, DimTotal * sizeof(float)); + memset(ctx->qy, 0, DimTotal * sizeof(float)); + memset(ctx->gradx, 0, DimTotal * sizeof(float)); + memset(ctx->grady, 0, DimTotal * sizeof(float)); + memset(ctx->div, 0, Dim1Total * sizeof(float)); + + return 0; +} + + +static int tnv_step(HWThread thr, void *hwctx, int device_id, void *data) { + long i, j, k, l; + + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int padZ = tnv_ctx->padZ; + int offY = ctx->offY; + int stepY = ctx->stepY; + int copY = ctx->copY; + + float *Input = ctx->Input; + float *u = ctx->u; + float *qx = ctx->qx; + float *qy = ctx->qy; + float *gradx = ctx->gradx; + float *grady = ctx->grady; + float *div = ctx->div; + + long p = 1l; + long q = 1l; + long r = 0l; + + float lambda = tnv_ctx->lambda; + float sigma = tnv_ctx->sigma; + float tau = tnv_ctx->tau; + float theta = tnv_ctx->theta; + + float taulambda = tau * lambda; + float divtau = 1.0f / tau; + float divsigma = 1.0f / sigma; + float theta1 = 1.0f + theta; + float constant = 1.0f + taulambda; + + float resprimal = 0.0f; + float resdual = 0.0f; + float product = 0.0f; + float unorm = 0.0f; + float qnorm = 0.0f; + + float qxdiff; + float qydiff; + float divdiff; + float gradxdiff[dimZ] __attribute__((aligned(64))); + float gradydiff[dimZ] __attribute__((aligned(64))); + float ubarx[dimZ] __attribute__((aligned(64))); + float ubary[dimZ] __attribute__((aligned(64))); + float udiff[dimZ] __attribute__((aligned(64))); + + + for(i=0; i < dimX; i++) { + for(k = 0; k < dimZ; k++) { + int l = i * padZ + k; + float u_upd = (u[l] + tau * div[l] + taulambda * Input[l])/constant; + float udiff = u[l] - u_upd; + ctx->udiff0[l] = udiff; + ctx->div0[l] = div[l]; + } + } + + for(int j1 = 0; j1 < stepY; j1 += BLOCK) { + for(int i1 = 0; i1 < dimX; i1 += BLOCK) { + for(int j2 = 0; j2 < BLOCK; j2++) { + j = j1 + j2; + for(int i2 = 0; i2 < BLOCK; i2++) { + float t[3]; + float M1 = 0.0f, M2 = 0.0f, M3 = 0.0f; + + i = i1 + i2; + if (i == dimX) break; + if (j == stepY) { j2 = BLOCK; break; } + l = (j * dimX + i) * padZ; + +#pragma vector aligned +#pragma GCC ivdep + for(k = 0; k < dimZ; k++) { + float u_upd = (u[l + k] + tau * div[l + k] + taulambda * Input[l + k]) / constant; + udiff[k] = u[l + k] - u_upd; + u[l + k] = u_upd; + + float gradx_upd = (i == (dimX - 1))?0:((u[l + k + padZ] + tau * div[l + k + padZ] + taulambda * Input[l + k + padZ]) / constant - u_upd); + float grady_upd = (j == (copY - 1))?0:((u[l + k + dimX*padZ] + tau * div[l + k + dimX*padZ] + taulambda * Input[l + k + dimX*padZ]) / constant - u_upd); + gradxdiff[k] = gradx[l + k] - gradx_upd; + gradydiff[k] = grady[l + k] - grady_upd; + gradx[l + k] = gradx_upd; + grady[l + k] = grady_upd; + + ubarx[k] = gradx_upd - theta * gradxdiff[k]; + ubary[k] = grady_upd - theta * gradydiff[k]; + + float vx = ubarx[k] + divsigma * qx[l + k]; + float vy = ubary[k] + divsigma * qy[l + k]; + + M1 += (vx * vx); M2 += (vx * vy); M3 += (vy * vy); + } + + coefF(t, M1, M2, M3, sigma, p, q, r); + +#pragma vector aligned +#pragma GCC ivdep + for(k = 0; k < dimZ; k++) { + float vx = ubarx[k] + divsigma * qx[l + k]; + float vy = ubary[k] + divsigma * qy[l + k]; + float gx_upd = vx * t[0] + vy * t[1]; + float gy_upd = vx * t[1] + vy * t[2]; + + qxdiff = sigma * (ubarx[k] - gx_upd); + qydiff = sigma * (ubary[k] - gy_upd); + + qx[l + k] += qxdiff; + qy[l + k] += qydiff; + + unorm += (udiff[k] * udiff[k]); + qnorm += (qxdiff * qxdiff + qydiff * qydiff); + + float div_upd = 0; + div_upd -= (i > 0)?qx[l + k - padZ]:0; + div_upd -= (j > 0)?qy[l + k - dimX*padZ]:0; + div_upd += (i < (dimX-1))?qx[l + k]:0; + div_upd += (j < (copY-1))?qy[l + k]:0; + divdiff = div[l + k] - div_upd; + div[l + k] = div_upd; + + resprimal += ((offY == 0)||(j > 0))?fabs(divtau * udiff[k] + divdiff):0; + resdual += fabs(divsigma * qxdiff + gradxdiff[k]); + resdual += fabs(divsigma * qydiff + gradydiff[k]); + product -= (gradxdiff[k] * qxdiff + gradydiff[k] * qydiff); + } + + } // i + } // j + } // i + } // j + + + ctx->resprimal = resprimal; + ctx->resdual = resdual; + ctx->product = product; + ctx->unorm = unorm; + ctx->qnorm = qnorm; + + return 0; +} + +static void TNV_CPU_init(float *InputT, float *uT, int dimX, int dimY, int dimZ) { + int i, off, size, err; + + if (sched) return; + + tnv_ctx.dimX = dimX; + tnv_ctx.dimY = dimY; + tnv_ctx.dimZ = dimZ; + // Padding seems actually slower +// tnv_ctx.padZ = dimZ; +// tnv_ctx.padZ = 4 * ((dimZ / 4) + ((dimZ % 4)?1:0)); + tnv_ctx.padZ = 16 * ((dimZ / 16) + ((dimZ % 16)?1:0)); + + hw_sched_init(); + + int threads = hw_sched_get_cpu_count(); + if (threads > dimY) threads = dimY/2; + + int step = dimY / threads; + int extra = dimY % threads; + + tnv_ctx.threads = threads; + tnv_ctx.thr_ctx = (tnv_thread_t*)calloc(threads, sizeof(tnv_thread_t)); + for (i = 0, off = 0; i < threads; i++, off += size) { + tnv_thread_t *ctx = tnv_ctx.thr_ctx + i; + size = step + ((i < extra)?1:0); + + ctx->offY = off; + ctx->stepY = size; + + if (i == (threads-1)) ctx->copY = ctx->stepY; + else ctx->copY = ctx->stepY + 1; + } + + sched = hw_sched_create(threads); + if (!sched) { fprintf(stderr, "Error creating threads\n"); exit(-1); } + + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_init); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling init threads", err); exit(-1); } +} + + + +/* + * C-OMP implementation of Total Nuclear Variation regularisation model (2D + channels) [1] + * The code is modified from the implementation by Joan Duran <joan.duran@uib.es> see + * "denoisingPDHG_ipol.cpp" in Joans Collaborative Total Variation package + * + * Input Parameters: + * 1. Noisy volume of 2D + channel dimension, i.e. 3D volume + * 2. lambda - regularisation parameter + * 3. Number of iterations [OPTIONAL parameter] + * 4. eplsilon - tolerance constant [OPTIONAL parameter] + * 5. print information: 0 (off) or 1 (on) [OPTIONAL parameter] + * + * Output: + * 1. Filtered/regularized image (u) + * + * [1]. Duran, J., Moeller, M., Sbert, C. and Cremers, D., 2016. Collaborative total variation: a general framework for vectorial TV models. SIAM Journal on Imaging Sciences, 9(1), pp.116-151. + */ + +float TNV_CPU_main(float *InputT, float *uT, float lambda, int maxIter, float tol, int dimX, int dimY, int dimZ) +{ + int err; + int iter; + int i,j,k,l,m; + + lambda = 1.0f/(2.0f*lambda); + tnv_ctx.lambda = lambda; + + // PDHG algorithm parameters + float tau = 0.5f; + float sigma = 0.5f; + float theta = 1.0f; + + // Backtracking parameters + float s = 1.0f; + float gamma = 0.75f; + float beta = 0.95f; + float alpha0 = 0.2f; + float alpha = alpha0; + float delta = 1.5f; + float eta = 0.95f; + + TNV_CPU_init(InputT, uT, dimX, dimY, dimZ); + + tnv_ctx.InputT = InputT; + tnv_ctx.uT = uT; + + int padZ = tnv_ctx.padZ; + + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_start); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling start threads", err); exit(-1); } + + + // Apply Primal-Dual Hybrid Gradient scheme + float residual = fLarge; + int started = 0; + for(iter = 0; iter < maxIter; iter++) { + float resprimal = 0.0f; + float resdual = 0.0f; + float product = 0.0f; + float unorm = 0.0f; + float qnorm = 0.0f; + + float divtau = 1.0f / tau; + + tnv_ctx.sigma = sigma; + tnv_ctx.tau = tau; + tnv_ctx.theta = theta; + + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_step); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling tnv threads", err); exit(-1); } + + // border regions + for (j = 1; j < tnv_ctx.threads; j++) { + tnv_thread_t *ctx0 = tnv_ctx.thr_ctx + (j - 1); + tnv_thread_t *ctx = tnv_ctx.thr_ctx + j; + + m = (ctx0->stepY - 1) * dimX * padZ; + for(i = 0; i < dimX; i++) { + for(k = 0; k < dimZ; k++) { + int l = i * padZ + k; + + float divdiff = ctx->div0[l] - ctx->div[l]; + float udiff = ctx->udiff0[l]; + + ctx->div[l] -= ctx0->qy[l + m]; + ctx0->div[m + l + dimX * padZ] = ctx->div[l]; + ctx0->u[m + l + dimX * padZ] = ctx->u[l]; + + divdiff += ctx0->qy[l + m]; + resprimal += fabs(divtau * udiff + divdiff); + } + } + } + + for (j = 0; j < tnv_ctx.threads; j++) { + tnv_thread_t *ctx = tnv_ctx.thr_ctx + j; + resprimal += ctx->resprimal; + resdual += ctx->resdual; + product += ctx->product; + unorm += ctx->unorm; + qnorm += ctx->qnorm; + } + + residual = (resprimal + resdual) / ((float) (dimX*dimY*dimZ)); + float b = (2.0f * tau * sigma * product) / (gamma * sigma * unorm + gamma * tau * qnorm); + float dual_dot_delta = resdual * s * delta; + float dual_div_delta = (resdual * s) / delta; + printf("resprimal: %f, resdual: %f, b: %f (product: %f, unorm: %f, qnorm: %f)\n", resprimal, resdual, b, product, unorm, qnorm); + + + if(b > 1) { + + // Decrease step-sizes to fit balancing principle + tau = (beta * tau) / b; + sigma = (beta * sigma) / b; + alpha = alpha0; + + if (started) { + fprintf(stderr, "\n\n\nWARNING: Back-tracking is required in the middle of iterative optimization! We CAN'T do it in the fast version. The standard TNV recommended\n\n\n"); + } else { + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_restore); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling restore threads", err); exit(-1); } + } + } else { + started = 1; + if(resprimal > dual_dot_delta) { + // Increase primal step-size and decrease dual step-size + tau = tau / (1.0f - alpha); + sigma = sigma * (1.0f - alpha); + alpha = alpha * eta; + } else if(resprimal < dual_div_delta) { + // Decrease primal step-size and increase dual step-size + tau = tau * (1.0f - alpha); + sigma = sigma / (1.0f - alpha); + alpha = alpha * eta; + } + } + + if (residual < tol) break; + } + + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_finish); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling finish threads", err); exit(-1); } + + + printf("Iterations stopped at %i with the residual %f \n", iter, residual); + printf("Return: %f\n", *uT); + +// exit(-1); + return *uT; +} diff --git a/src/Core/performance_CPU/TNV_core.c.v32 b/src/Core/performance_CPU/TNV_core.c.v32 new file mode 100755 index 0000000..fcb2f87 --- /dev/null +++ b/src/Core/performance_CPU/TNV_core.c.v32 @@ -0,0 +1,676 @@ +/* + * This work is part of the Core Imaging Library developed by + * Visual Analytics and Imaging System Group of the Science Technology + * Facilities Council, STFC + * + * Copyright 2017 Daniil Kazantsev + * Copyright 2017 Srikanth Nagella, Edoardo Pasca + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * http://www.apache.org/licenses/LICENSE-2.0 + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include <malloc.h> +#include "TNV_core.h" + +#define min(a,b) (((a)<(b))?(a):(b)) + +inline void coefF(float *t, float M1, float M2, float M3, float sigma, int p, int q, int r) { + int ii, num; + float divsigma = 1.0f / sigma; + float sum, shrinkfactor; + float T,D,det,eig1,eig2,sig1,sig2,V1, V2, V3, V4, v0,v1,v2, mu1,mu2,sig1_upd,sig2_upd; + float proj[2] = {0}; + + // Compute eigenvalues of M + T = M1 + M3; + D = M1 * M3 - M2 * M2; + det = sqrtf(MAX((T * T / 4.0f) - D, 0.0f)); + eig1 = MAX((T / 2.0f) + det, 0.0f); + eig2 = MAX((T / 2.0f) - det, 0.0f); + sig1 = sqrtf(eig1); + sig2 = sqrtf(eig2); + + // Compute normalized eigenvectors + V1 = V2 = V3 = V4 = 0.0f; + + if(M2 != 0.0f) + { + v0 = M2; + v1 = eig1 - M3; + v2 = eig2 - M3; + + mu1 = sqrtf(v0 * v0 + v1 * v1); + mu2 = sqrtf(v0 * v0 + v2 * v2); + + if(mu1 > fTiny) + { + V1 = v1 / mu1; + V3 = v0 / mu1; + } + + if(mu2 > fTiny) + { + V2 = v2 / mu2; + V4 = v0 / mu2; + } + + } else + { + if(M1 > M3) + { + V1 = V4 = 1.0f; + V2 = V3 = 0.0f; + + } else + { + V1 = V4 = 0.0f; + V2 = V3 = 1.0f; + } + } + + // Compute prox_p of the diagonal entries + sig1_upd = sig2_upd = 0.0f; + + if(p == 1) + { + sig1_upd = MAX(sig1 - divsigma, 0.0f); + sig2_upd = MAX(sig2 - divsigma, 0.0f); + + } else if(p == INFNORM) + { + proj[0] = sigma * fabs(sig1); + proj[1] = sigma * fabs(sig2); + + /*l1 projection part */ + sum = fLarge; + num = 0l; + shrinkfactor = 0.0f; + while(sum > 1.0f) + { + sum = 0.0f; + num = 0; + + for(ii = 0; ii < 2; ii++) + { + proj[ii] = MAX(proj[ii] - shrinkfactor, 0.0f); + + sum += fabs(proj[ii]); + if(proj[ii]!= 0.0f) + num++; + } + + if(num > 0) + shrinkfactor = (sum - 1.0f) / num; + else + break; + } + /*l1 proj ends*/ + + sig1_upd = sig1 - divsigma * proj[0]; + sig2_upd = sig2 - divsigma * proj[1]; + } + + // Compute the diagonal entries of $\widehat{\Sigma}\Sigma^{\dagger}_0$ + if(sig1 > fTiny) + sig1_upd /= sig1; + + if(sig2 > fTiny) + sig2_upd /= sig2; + + // Compute solution + t[0] = sig1_upd * V1 * V1 + sig2_upd * V2 * V2; + t[1] = sig1_upd * V1 * V3 + sig2_upd * V2 * V4; + t[2] = sig1_upd * V3 * V3 + sig2_upd * V4 * V4; +} + + +#include "hw_sched.h" +typedef struct { + int offY, stepY, copY; + float *Input, *u, *qx, *qy, *gradx, *grady, *div; + float *div0, *udiff0; + float *gradxdiff, *gradydiff, *ubarx, *ubary, *udiff; + float resprimal, resdual; + float unorm, qnorm, product; +} tnv_thread_t; + +typedef struct { + int threads; + tnv_thread_t *thr_ctx; + float *InputT, *uT; + int dimX, dimY, dimZ, padZ; + float lambda, sigma, tau, theta; +} tnv_context_t; + +HWSched sched = NULL; +tnv_context_t tnv_ctx; + + +static int tnv_free(HWThread thr, void *hwctx, int device_id, void *data) { + int i,j,k; + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + free(ctx->Input); + free(ctx->u); + free(ctx->qx); + free(ctx->qy); + free(ctx->gradx); + free(ctx->grady); + free(ctx->div); + + free(ctx->div0); + free(ctx->udiff0); + + free(ctx->gradxdiff); + free(ctx->gradydiff); + free(ctx->ubarx); + free(ctx->ubary); + free(ctx->udiff); + + return 0; +} + +static int tnv_init(HWThread thr, void *hwctx, int device_id, void *data) { + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int padZ = tnv_ctx->padZ; + int offY = ctx->offY; + int stepY = ctx->stepY; + +// printf("%i %p - %i %i %i x %i %i\n", device_id, ctx, dimX, dimY, dimZ, offY, stepY); + + long DimTotal = (long)(dimX*stepY*padZ); + long Dim1Total = (long)(dimX*(stepY+1)*padZ); + long DimRow = (long)(dimX * padZ); + long DimCell = (long)(padZ); + + // Auxiliar vectors + ctx->Input = memalign(64, Dim1Total * sizeof(float)); + ctx->u = memalign(64, Dim1Total * sizeof(float)); + ctx->qx = memalign(64, DimTotal * sizeof(float)); + ctx->qy = memalign(64, DimTotal * sizeof(float)); + ctx->gradx = memalign(64, DimTotal * sizeof(float)); + ctx->grady = memalign(64, DimTotal * sizeof(float)); + ctx->div = memalign(64, Dim1Total * sizeof(float)); + + ctx->div0 = memalign(64, DimRow * sizeof(float)); + ctx->udiff0 = memalign(64, DimRow * sizeof(float)); + + ctx->gradxdiff = memalign(64, DimCell * sizeof(float)); + ctx->gradydiff = memalign(64, DimCell * sizeof(float)); + ctx->ubarx = memalign(64, DimCell * sizeof(float)); + ctx->ubary = memalign(64, DimCell * sizeof(float)); + ctx->udiff = memalign(64, DimCell * sizeof(float)); + + if ((!ctx->Input)||(!ctx->u)||(!ctx->qx)||(!ctx->qy)||(!ctx->gradx)||(!ctx->grady)||(!ctx->div)||(!ctx->div0)||(!ctx->udiff)||(!ctx->udiff0)) { + fprintf(stderr, "Error allocating memory\n"); + exit(-1); + } + + return 0; +} + +static int tnv_start(HWThread thr, void *hwctx, int device_id, void *data) { + int i,j,k; + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int padZ = tnv_ctx->padZ; + int offY = ctx->offY; + int stepY = ctx->stepY; + int copY = ctx->copY; + +// printf("%i %p - %i %i %i (%i) x %i %i\n", device_id, ctx, dimX, dimY, dimZ, padZ, offY, stepY); + + long DimTotal = (long)(dimX*stepY*padZ); + long Dim1Total = (long)(dimX*copY*padZ); + + memset(ctx->u, 0, Dim1Total * sizeof(float)); + memset(ctx->qx, 0, DimTotal * sizeof(float)); + memset(ctx->qy, 0, DimTotal * sizeof(float)); + memset(ctx->gradx, 0, DimTotal * sizeof(float)); + memset(ctx->grady, 0, DimTotal * sizeof(float)); + memset(ctx->div, 0, Dim1Total * sizeof(float)); + + for(k=0; k<dimZ; k++) { + for(j=0; j<copY; j++) { + for(i=0; i<dimX; i++) { + ctx->Input[j * dimX * padZ + i * padZ + k] = tnv_ctx->InputT[k * dimX * dimY + (j + offY) * dimX + i]; + ctx->u[j * dimX * padZ + i * padZ + k] = tnv_ctx->uT[k * dimX * dimY + (j + offY) * dimX + i]; + } + } + } + + return 0; +} + +static int tnv_finish(HWThread thr, void *hwctx, int device_id, void *data) { + int i,j,k; + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int padZ = tnv_ctx->padZ; + int offY = ctx->offY; + int stepY = ctx->stepY; + int copY = ctx->copY; + + for(k=0; k<dimZ; k++) { + for(j=0; j<stepY; j++) { + for(i=0; i<dimX; i++) { + tnv_ctx->uT[k * dimX * dimY + (j + offY) * dimX + i] = ctx->u[j * dimX * padZ + i * padZ + k]; + } + } + } + + return 0; +} + + +static int tnv_restore(HWThread thr, void *hwctx, int device_id, void *data) { + int i,j,k; + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int stepY = ctx->stepY; + int copY = ctx->copY; + int padZ = tnv_ctx->padZ; + long DimTotal = (long)(dimX*stepY*padZ); + long Dim1Total = (long)(dimX*copY*padZ); + + memset(ctx->u, 0, Dim1Total * sizeof(float)); + memset(ctx->qx, 0, DimTotal * sizeof(float)); + memset(ctx->qy, 0, DimTotal * sizeof(float)); + memset(ctx->gradx, 0, DimTotal * sizeof(float)); + memset(ctx->grady, 0, DimTotal * sizeof(float)); + memset(ctx->div, 0, Dim1Total * sizeof(float)); + + return 0; +} + + +static int tnv_step(HWThread thr, void *hwctx, int device_id, void *data) { + long i, j, k, l, m; + + tnv_context_t *tnv_ctx = (tnv_context_t*)data; + tnv_thread_t *ctx = tnv_ctx->thr_ctx + device_id; + + int dimX = tnv_ctx->dimX; + int dimY = tnv_ctx->dimY; + int dimZ = tnv_ctx->dimZ; + int padZ = tnv_ctx->padZ; + int offY = ctx->offY; + int stepY = ctx->stepY; + int copY = ctx->copY; + + float *Input = ctx->Input; + float *u = ctx->u; + float *qx = ctx->qx; + float *qy = ctx->qy; + float *gradx = ctx->gradx; + float *grady = ctx->grady; + float *div = ctx->div; + + long p = 1l; + long q = 1l; + long r = 0l; + + float lambda = tnv_ctx->lambda; + float sigma = tnv_ctx->sigma; + float tau = tnv_ctx->tau; + float theta = tnv_ctx->theta; + + float taulambda = tau * lambda; + float divtau = 1.0f / tau; + float divsigma = 1.0f / sigma; + float theta1 = 1.0f + theta; + float constant = 1.0f + taulambda; + + float resprimal = 0.0f; + float resdual1 = 0.0f; + float resdual2 = 0.0f; + float product = 0.0f; + float unorm = 0.0f; + float qnorm = 0.0f; + + float qxdiff; + float qydiff; + float divdiff; + float *gradxdiff = ctx->gradxdiff; + float *gradydiff = ctx->gradydiff; + float *ubarx = ctx->ubarx; + float *ubary = ctx->ubary; + float *udiff = ctx->udiff; + + float *udiff0 = ctx->udiff0; + float *div0 = ctx->div0; + +/* + for(i=0; i < dimX; i++) { + for(k = 0; k < dimZ; k++) { + int l = i * padZ + k; + float u_upd = (u[l] + tau * div[l] + taulambda * Input[l])/constant; + float udiff_val = u[l] - u_upd; + udiff0[l] = udiff_val; + div0[l] = div[l]; + } + } +*/ + + + j = 0; { +# define TNV_LOOP_FIRST_J + i = 0; { +# define TNV_LOOP_FIRST_I +# include "TNV_core_loop.h" +# undef TNV_LOOP_FIRST_I + } + for(i = 1; i < (dimX - 1); i++) { +# include "TNV_core_loop.h" + } + i = dimX - 1; { +# define TNV_LOOP_LAST_I +# include "TNV_core_loop.h" +# undef TNV_LOOP_LAST_I + } +# undef TNV_LOOP_FIRST_J + } + + + + for(int j = 1; j < (copY - 1); j++) { + i = 0; { +# define TNV_LOOP_FIRST_I +# include "TNV_core_loop.h" +# undef TNV_LOOP_FIRST_I + } + } + +#define BLOCK 32 + for(int j1 = 1; j1 < (copY - 1); j1 += BLOCK) { + for(int i1 = 1; i1 < (dimX - 1); i1 += BLOCK) { + for(int j2 = 0; j2 < BLOCK; j2 ++) { + j = j1 + j2; + for(int i2 = 0; i2 < BLOCK; i2++) { + i = i1 + i2; + + if (i == (dimX - 1)) break; + if (j == (copY - 1)) { j2 = BLOCK; break; } +# include "TNV_core_loop.h" + } + } + } // i + + } + + for(int j = 1; j < (copY - 1); j++) { + i = dimX - 1; { +# define TNV_LOOP_LAST_I +# include "TNV_core_loop.h" +# undef TNV_LOOP_LAST_I + } + } + + + + for (j = copY - 1; j < stepY; j++) { +# define TNV_LOOP_LAST_J + i = 0; { +# define TNV_LOOP_FIRST_I +# include "TNV_core_loop.h" +# undef TNV_LOOP_FIRST_I + } + for(i = 1; i < (dimX - 1); i++) { +# include "TNV_core_loop.h" + } + i = dimX - 1; { +# define TNV_LOOP_LAST_I +# include "TNV_core_loop.h" +# undef TNV_LOOP_LAST_I + } +# undef TNV_LOOP_LAST_J + } + + + + ctx->resprimal = resprimal; + ctx->resdual = resdual1 + resdual2; + ctx->product = product; + ctx->unorm = unorm; + ctx->qnorm = qnorm; + + return 0; +} + +static void TNV_CPU_init(float *InputT, float *uT, int dimX, int dimY, int dimZ) { + int i, off, size, err; + + if (sched) return; + + tnv_ctx.dimX = dimX; + tnv_ctx.dimY = dimY; + tnv_ctx.dimZ = dimZ; + // Padding seems actually slower +// tnv_ctx.padZ = dimZ; +// tnv_ctx.padZ = 4 * ((dimZ / 4) + ((dimZ % 4)?1:0)); + tnv_ctx.padZ = 16 * ((dimZ / 16) + ((dimZ % 16)?1:0)); + + hw_sched_init(); + + int threads = hw_sched_get_cpu_count(); + if (threads > dimY) threads = dimY/2; + + int step = dimY / threads; + int extra = dimY % threads; + + tnv_ctx.threads = threads; + tnv_ctx.thr_ctx = (tnv_thread_t*)calloc(threads, sizeof(tnv_thread_t)); + for (i = 0, off = 0; i < threads; i++, off += size) { + tnv_thread_t *ctx = tnv_ctx.thr_ctx + i; + size = step + ((i < extra)?1:0); + + ctx->offY = off; + ctx->stepY = size; + + if (i == (threads-1)) ctx->copY = ctx->stepY; + else ctx->copY = ctx->stepY + 1; + } + + sched = hw_sched_create(threads); + if (!sched) { fprintf(stderr, "Error creating threads\n"); exit(-1); } + + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_init); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling init threads", err); exit(-1); } +} + + + +/* + * C-OMP implementation of Total Nuclear Variation regularisation model (2D + channels) [1] + * The code is modified from the implementation by Joan Duran <joan.duran@uib.es> see + * "denoisingPDHG_ipol.cpp" in Joans Collaborative Total Variation package + * + * Input Parameters: + * 1. Noisy volume of 2D + channel dimension, i.e. 3D volume + * 2. lambda - regularisation parameter + * 3. Number of iterations [OPTIONAL parameter] + * 4. eplsilon - tolerance constant [OPTIONAL parameter] + * 5. print information: 0 (off) or 1 (on) [OPTIONAL parameter] + * + * Output: + * 1. Filtered/regularized image (u) + * + * [1]. Duran, J., Moeller, M., Sbert, C. and Cremers, D., 2016. Collaborative total variation: a general framework for vectorial TV models. SIAM Journal on Imaging Sciences, 9(1), pp.116-151. + */ + +float TNV_CPU_main(float *InputT, float *uT, float lambda, int maxIter, float tol, int dimX, int dimY, int dimZ) +{ + int err; + int iter; + int i,j,k,l,m; + + lambda = 1.0f/(2.0f*lambda); + tnv_ctx.lambda = lambda; + + // PDHG algorithm parameters + float tau = 0.5f; + float sigma = 0.5f; + float theta = 1.0f; + + // Backtracking parameters + float s = 1.0f; + float gamma = 0.75f; + float beta = 0.95f; + float alpha0 = 0.2f; + float alpha = alpha0; + float delta = 1.5f; + float eta = 0.95f; + + TNV_CPU_init(InputT, uT, dimX, dimY, dimZ); + + tnv_ctx.InputT = InputT; + tnv_ctx.uT = uT; + + int padZ = tnv_ctx.padZ; + + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_start); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling start threads", err); exit(-1); } + + + // Apply Primal-Dual Hybrid Gradient scheme + float residual = fLarge; + int started = 0; + for(iter = 0; iter < maxIter; iter++) { + float resprimal = 0.0f; + float resdual = 0.0f; + float product = 0.0f; + float unorm = 0.0f; + float qnorm = 0.0f; + + float divtau = 1.0f / tau; + + tnv_ctx.sigma = sigma; + tnv_ctx.tau = tau; + tnv_ctx.theta = theta; + + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_step); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling tnv threads", err); exit(-1); } + + // border regions + for (j = 1; j < tnv_ctx.threads; j++) { + tnv_thread_t *ctx0 = tnv_ctx.thr_ctx + (j - 1); + tnv_thread_t *ctx = tnv_ctx.thr_ctx + j; + + m = (ctx0->stepY - 1) * dimX * padZ; + for(i = 0; i < dimX; i++) { + for(k = 0; k < dimZ; k++) { + int l = i * padZ + k; + + float divdiff = ctx->div0[l] - ctx->div[l]; + float udiff = ctx->udiff0[l]; + + ctx->div[l] -= ctx0->qy[l + m]; + ctx0->div[m + l + dimX*padZ] = ctx->div[l]; + ctx0->u[m + l + dimX*padZ] = ctx->u[l]; + + divdiff += ctx0->qy[l + m]; + resprimal += fabs(divtau * udiff + divdiff); + } + } + } + + { + tnv_thread_t *ctx = tnv_ctx.thr_ctx + 0; + for(i = 0; i < dimX; i++) { + for(k = 0; k < dimZ; k++) { + int l = i * padZ + k; + + float divdiff = ctx->div0[l] - ctx->div[l]; + float udiff = ctx->udiff0[l]; + resprimal += fabs(divtau * udiff + divdiff); + } + } + } + + for (j = 0; j < tnv_ctx.threads; j++) { + tnv_thread_t *ctx = tnv_ctx.thr_ctx + j; + resprimal += ctx->resprimal; + resdual += ctx->resdual; + product += ctx->product; + unorm += ctx->unorm; + qnorm += ctx->qnorm; + } + + residual = (resprimal + resdual) / ((float) (dimX*dimY*dimZ)); + float b = (2.0f * tau * sigma * product) / (gamma * sigma * unorm + gamma * tau * qnorm); + float dual_dot_delta = resdual * s * delta; + float dual_div_delta = (resdual * s) / delta; + printf("resprimal: %f, resdual: %f, b: %f (product: %f, unorm: %f, qnorm: %f)\n", resprimal, resdual, b, product, unorm, qnorm); + + + if(b > 1) { + + // Decrease step-sizes to fit balancing principle + tau = (beta * tau) / b; + sigma = (beta * sigma) / b; + alpha = alpha0; + + if (started) { + fprintf(stderr, "\n\n\nWARNING: Back-tracking is required in the middle of iterative optimization! We CAN'T do it in the fast version. The standard TNV recommended\n\n\n"); + } else { + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_restore); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling restore threads", err); exit(-1); } + } + } else { + started = 1; + if(resprimal > dual_dot_delta) { + // Increase primal step-size and decrease dual step-size + tau = tau / (1.0f - alpha); + sigma = sigma * (1.0f - alpha); + alpha = alpha * eta; + } else if(resprimal < dual_div_delta) { + // Decrease primal step-size and increase dual step-size + tau = tau * (1.0f - alpha); + sigma = sigma / (1.0f - alpha); + alpha = alpha * eta; + } + } + + if (residual < tol) break; + } + + err = hw_sched_schedule_thread_task(sched, (void*)&tnv_ctx, tnv_finish); + if (!err) err = hw_sched_wait_task(sched); + if (err) { fprintf(stderr, "Error %i scheduling finish threads", err); exit(-1); } + + + printf("Iterations stopped at %i with the residual %f \n", iter, residual); + printf("Return: %f\n", *uT); + + return *uT; +} diff --git a/src/Core/performance_CPU/TNV_core.c.v4.stdver b/src/Core/performance_CPU/TNV_core.c.v4.stdver new file mode 100755 index 0000000..6be60ae --- /dev/null +++ b/src/Core/performance_CPU/TNV_core.c.v4.stdver @@ -0,0 +1,629 @@ +/* + * This work is part of the Core Imaging Library developed by + * Visual Analytics and Imaging System Group of the Science Technology + * Facilities Council, STFC + * + * Copyright 2017 Daniil Kazantsev + * Copyright 2017 Srikanth Nagella, Edoardo Pasca + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * http://www.apache.org/licenses/LICENSE-2.0 + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include "TNV_core.h" + + +inline void coefF(float *t, float M1, float M2, float M3, float sigma, int p, int q, int r) { + int ii, num; + float divsigma = 1.0f / sigma; + float sum, shrinkfactor; + float T,D,det,eig1,eig2,sig1,sig2,V1, V2, V3, V4, v0,v1,v2, mu1,mu2,sig1_upd,sig2_upd; + float proj[2] = {0}; + + // Compute eigenvalues of M + T = M1 + M3; + D = M1 * M3 - M2 * M2; + det = sqrt(MAX((T * T / 4.0f) - D, 0.0f)); + eig1 = MAX((T / 2.0f) + det, 0.0f); + eig2 = MAX((T / 2.0f) - det, 0.0f); + sig1 = sqrt(eig1); + sig2 = sqrt(eig2); + + // Compute normalized eigenvectors + V1 = V2 = V3 = V4 = 0.0f; + + if(M2 != 0.0f) + { + v0 = M2; + v1 = eig1 - M3; + v2 = eig2 - M3; + + mu1 = sqrtf(v0 * v0 + v1 * v1); + mu2 = sqrtf(v0 * v0 + v2 * v2); + + if(mu1 > fTiny) + { + V1 = v1 / mu1; + V3 = v0 / mu1; + } + + if(mu2 > fTiny) + { + V2 = v2 / mu2; + V4 = v0 / mu2; + } + + } else + { + if(M1 > M3) + { + V1 = V4 = 1.0f; + V2 = V3 = 0.0f; + + } else + { + V1 = V4 = 0.0f; + V2 = V3 = 1.0f; + } + } + + // Compute prox_p of the diagonal entries + sig1_upd = sig2_upd = 0.0f; + + if(p == 1) + { + sig1_upd = MAX(sig1 - divsigma, 0.0f); + sig2_upd = MAX(sig2 - divsigma, 0.0f); + + } else if(p == INFNORM) + { + proj[0] = sigma * fabs(sig1); + proj[1] = sigma * fabs(sig2); + + /*l1 projection part */ + sum = fLarge; + num = 0l; + shrinkfactor = 0.0f; + while(sum > 1.0f) + { + sum = 0.0f; + num = 0; + + for(ii = 0; ii < 2; ii++) + { + proj[ii] = MAX(proj[ii] - shrinkfactor, 0.0f); + + sum += fabs(proj[ii]); + if(proj[ii]!= 0.0f) + num++; + } + + if(num > 0) + shrinkfactor = (sum - 1.0f) / num; + else + break; + } + /*l1 proj ends*/ + + sig1_upd = sig1 - divsigma * proj[0]; + sig2_upd = sig2 - divsigma * proj[1]; + } + + // Compute the diagonal entries of $\widehat{\Sigma}\Sigma^{\dagger}_0$ + if(sig1 > fTiny) + sig1_upd /= sig1; + + if(sig2 > fTiny) + sig2_upd /= sig2; + + // Compute solution + t[0] = sig1_upd * V1 * V1 + sig2_upd * V2 * V2; + t[1] = sig1_upd * V1 * V3 + sig2_upd * V2 * V4; + t[2] = sig1_upd * V3 * V3 + sig2_upd * V4 * V4; +} + + +/* + * C-OMP implementation of Total Nuclear Variation regularisation model (2D + channels) [1] + * The code is modified from the implementation by Joan Duran <joan.duran@uib.es> see + * "denoisingPDHG_ipol.cpp" in Joans Collaborative Total Variation package + * + * Input Parameters: + * 1. Noisy volume of 2D + channel dimension, i.e. 3D volume + * 2. lambda - regularisation parameter + * 3. Number of iterations [OPTIONAL parameter] + * 4. eplsilon - tolerance constant [OPTIONAL parameter] + * 5. print information: 0 (off) or 1 (on) [OPTIONAL parameter] + * + * Output: + * 1. Filtered/regularized image (u) + * + * [1]. Duran, J., Moeller, M., Sbert, C. and Cremers, D., 2016. Collaborative total variation: a general framework for vectorial TV models. SIAM Journal on Imaging Sciences, 9(1), pp.116-151. + */ + +float TNV_CPU_main(float *InputT, float *uT, float lambda, int maxIter, float tol, int dimX, int dimY, int dimZ) +{ + long i, j, k, p, q, r, DimTotal; + float taulambda; + float *u_upd, *qx, *qy, *qx_upd, *qy_upd, *gradx, *grady, *gradx_upd, *grady_upd, *div, *div_upd; + + p = 1l; + q = 1l; + r = 0l; + + lambda = 1.0f/(2.0f*lambda); + DimTotal = (long)(dimX*dimY*dimZ); + /* PDHG algorithm parameters*/ + float tau = 0.5f; + float sigma = 0.5f; + float theta = 1.0f; + + // Auxiliar vectors + u_upd = calloc(DimTotal, sizeof(float)); + qx = calloc(DimTotal, sizeof(float)); + qy = calloc(DimTotal, sizeof(float)); + qx_upd = calloc(DimTotal, sizeof(float)); + qy_upd = calloc(DimTotal, sizeof(float)); + gradx = calloc(DimTotal, sizeof(float)); + grady = calloc(DimTotal, sizeof(float)); + gradx_upd = calloc(DimTotal, sizeof(float)); + grady_upd = calloc(DimTotal, sizeof(float)); + div = calloc(DimTotal, sizeof(float)); + div_upd = calloc(DimTotal, sizeof(float)); + + + float *Input = calloc(DimTotal, sizeof(float)); + float *u = calloc(DimTotal, sizeof(float)); + for(k=0; k<dimZ; k++) { + for(j=0; j<dimY; j++) { + for(i=0; i<dimX; i++) { + Input[j * dimX * dimZ + i * dimZ + k] = InputT[k * dimX * dimY + j * dimX + i]; + u[j * dimX * dimZ + i * dimZ + k] = uT[k * dimX * dimY + j * dimX + i]; + } + } + } + + + // Backtracking parameters + float s = 1.0f; + float gamma = 0.75f; + float beta = 0.95f; + float alpha0 = 0.2f; + float alpha = alpha0; + float delta = 1.5f; + float eta = 0.95f; + + // PDHG algorithm parameters + taulambda = tau * lambda; + float divtau = 1.0f / tau; + float divsigma = 1.0f / sigma; + float theta1 = 1.0f + theta; + + /*allocate memory for taulambda */ + //taulambda = (float*) calloc(dimZ, sizeof(float)); + //for(k=0; k < dimZ; k++) {taulambda[k] = tau*lambda[k];} + + // Apply Primal-Dual Hybrid Gradient scheme + int iter = 0; + float residual = fLarge; + + for(iter = 0; iter < maxIter; iter++) { + // Argument of proximal mapping of fidelity term + // Proximal solution of fidelity term + // proxG(u_upd, u, div, Input, tau, taulambda, (long)(dimX), (long)(dimY), (long)(dimZ)); + float constant = 1.0f + taulambda; + #pragma omp parallel for shared(Input, u, u_upd) private(k) + for(k=0; k<dimZ*dimX*dimY; k++) { + float v = u[k] + tau * div[k]; + u_upd[k] = (v + taulambda * Input[k])/constant; + } + + memset(div_upd, 0, dimX*dimY*dimZ*sizeof(float)); + // This changes results, I guess access to div_upd is violated +// #pragma omp parallel for shared (gradx_upd,grady_upd,gradx,grady,qx,qy,qx_upd,qy_upd) private(i,j,k) + for(j=0; j<dimY; j++) { + for(i=0; i<dimX; i++) { + float t[3]; + float M1 = 0.0f, M2 = 0.0f, M3 = 0.0f; + int l = (j * dimX + i) * dimZ; + + for(k = 0; k < dimZ; k++) { + if(i != dimX-1) + gradx_upd[l + k] = u_upd[l + k + dimZ] - u_upd[l + k]; + else + gradx_upd[l + k] = 0.0f; + + if(j != dimY-1) + grady_upd[l + k] = u_upd[l + k + dimX * dimZ] - u_upd[l + k]; + else + grady_upd[l + k] = 0.0f; + + float ubarx = theta1 * gradx_upd[l + k] - theta * gradx[l + k]; + float ubary = theta1 * grady_upd[l + k] - theta * grady[l + k]; + float vx = ubarx + divsigma * qx[l + k]; + float vy = ubary + divsigma * qy[l + k]; + + M1 += (vx * vx); M2 += (vx * vy); M3 += (vy * vy); + } + + coefF(t, M1, M2, M3, sigma, p, q, r); + + for(k = 0; k < dimZ; k++) { + float ubarx = theta1 * gradx_upd[l + k] - theta * gradx[l + k]; + float ubary = theta1 * grady_upd[l + k] - theta * grady[l + k]; + float vx = ubarx + divsigma * qx[l + k]; + float vy = ubary + divsigma * qy[l + k]; + + float gx_upd = vx * t[0] + vy * t[1]; + float gy_upd = vx * t[1] + vy * t[2]; + + qx_upd[l + k] = qx[l + k] + sigma * (ubarx - gx_upd); + qy_upd[l + k] = qy[l + k] + sigma * (ubary - gy_upd); + + if(i != dimX-1) { + div_upd[l + k] += qx_upd[l + k]; + div_upd[l + k + dimZ] -= qx_upd[l + k]; + } + + if(j != dimY-1) { + div_upd[l + k] += qy_upd[l + k]; + div_upd[l + k + dimX * dimZ] -= qy_upd[l + k]; + } + } + } + } + +// Compute primal residual, dual residual, and backtracking condition + float resprimal = 0.0f; + float resdual = 0.0f; + float product = 0.0f; + float unorm = 0.0f; + float qnorm = 0.0f; + + // If this loop is inner, the result slightly changed due to different summation order + for(int l=0; l<dimZ; l++) + for(j=0; j<dimY; j++) + for(i=0; i<dimX; i++) + { +// for(k=0; k<dimX*dimY*dimZ; k++) { + int k = j * dimX * dimZ + i * dimZ + l; + + float udiff = u[k] - u_upd[k]; + float qxdiff = qx[k] - qx_upd[k]; + float qydiff = qy[k] - qy_upd[k]; + float divdiff = div[k] - div_upd[k]; + float gradxdiff = gradx[k] - gradx_upd[k]; + float gradydiff = grady[k] - grady_upd[k]; + + resprimal += fabs(divtau*udiff + divdiff); + resdual += fabs(divsigma*qxdiff - gradxdiff); + resdual += fabs(divsigma*qydiff - gradydiff); + + unorm += (udiff * udiff); + qnorm += (qxdiff * qxdiff + qydiff * qydiff); + product += (gradxdiff * qxdiff + gradydiff * qydiff); + } + + float b = (2.0f * tau * sigma * product) / (gamma * sigma * unorm + + gamma * tau * qnorm); + +// printf("resprimal: %f, resdual: %f, b: %f\n", resprimal, resdual, b); + printf("resprimal: %f, resdual: %f, b: %f (product: %f, unorm: %f, qnorm: %f)\n", resprimal, resdual, b, product, unorm, qnorm); + +// printf("b: %f\n", b); + +// Adapt step-size parameters + float dual_dot_delta = resdual * s * delta; + float dual_div_delta = (resdual * s) / delta; + + if(b > 1) + { + // Decrease step-sizes to fit balancing principle + tau = (beta * tau) / b; + sigma = (beta * sigma) / b; + alpha = alpha0; + + copyIm(u, u_upd, (long)(dimX), (long)(dimY), (long)(dimZ)); + copyIm(qx, qx_upd, (long)(dimX), (long)(dimY), (long)(dimZ)); + copyIm(qy, qy_upd, (long)(dimX), (long)(dimY), (long)(dimZ)); + copyIm(gradx, gradx_upd, (long)(dimX), (long)(dimY), (long)(dimZ)); + copyIm(grady, grady_upd, (long)(dimX), (long)(dimY), (long)(dimZ)); + copyIm(div, div_upd, (long)(dimX), (long)(dimY), (long)(dimZ)); + + } else if(resprimal > dual_dot_delta) + { + // Increase primal step-size and decrease dual step-size + tau = tau / (1.0f - alpha); + sigma = sigma * (1.0f - alpha); + alpha = alpha * eta; + + } else if(resprimal < dual_div_delta) + { + // Decrease primal step-size and increase dual step-size + tau = tau * (1.0f - alpha); + sigma = sigma / (1.0f - alpha); + alpha = alpha * eta; + } + +// Update variables + taulambda = tau * lambda; +//for(k=0; k < dimZ; k++) taulambda[k] = tau*lambda[k]; + + divsigma = 1.0f / sigma; + divtau = 1.0f / tau; + + copyIm(u_upd, u, (long)(dimX), (long)(dimY), (long)(dimZ)); + copyIm(qx_upd, qx, (long)(dimX), (long)(dimY), (long)(dimZ)); + copyIm(qy_upd, qy, (long)(dimX), (long)(dimY), (long)(dimZ)); + copyIm(gradx_upd, gradx, (long)(dimX), (long)(dimY), (long)(dimZ)); + copyIm(grady_upd, grady, (long)(dimX), (long)(dimY), (long)(dimZ)); + copyIm(div_upd, div, (long)(dimX), (long)(dimY), (long)(dimZ)); + +// Compute residual at current iteration + residual = (resprimal + resdual) / ((float) (dimX*dimY*dimZ)); + +// printf("%f \n", residual); + if (residual < tol) { + printf("Iterations stopped at %i with the residual %f \n", iter, residual); + break; + } + + } + printf("Iterations stopped at %i with the residual %f \n", iter, residual); + + free (u_upd); + free(qx); + free(qy); + free(qx_upd); + free(qy_upd); + free(gradx); + free(grady); + free(gradx_upd); + free(grady_upd); + free(div); + free(div_upd); + printf("Return: %f\n", *u); + + for(k=0; k<dimZ; k++) { + for(j=0; j<dimY; j++) { + for(i=0; i<dimX; i++) { + uT[k * dimX * dimY + j * dimX + i] = u[j * dimX * dimZ + i * dimZ + k]; + } + } + } + + + + + return *u; +} + +float proxG(float *u_upd, float *v, float *f, float taulambda, long dimX, long dimY, long dimZ) +{ + float constant; + long k; + constant = 1.0f + taulambda; + #pragma omp parallel for shared(v, f, u_upd) private(k) + for(k=0; k<dimZ*dimX*dimY; k++) { + u_upd[k] = (v[k] + taulambda * f[k])/constant; + //u_upd[(dimX*dimY)*k + l] = (v[(dimX*dimY)*k + l] + taulambda * f[(dimX*dimY)*k + l])/constant; + } + return *u_upd; +} + +float gradient(float *u_upd, float *gradx_upd, float *grady_upd, long dimX, long dimY, long dimZ) +{ + long i, j, k, l; + // Compute discrete gradient using forward differences + #pragma omp parallel for shared(gradx_upd,grady_upd,u_upd) private(i, j, k, l) + for(k = 0; k < dimZ; k++) { + for(j = 0; j < dimY; j++) { + l = j * dimX; + for(i = 0; i < dimX; i++) { + // Derivatives in the x-direction + if(i != dimX-1) + gradx_upd[(dimX*dimY)*k + i+l] = u_upd[(dimX*dimY)*k + i+1+l] - u_upd[(dimX*dimY)*k + i+l]; + else + gradx_upd[(dimX*dimY)*k + i+l] = 0.0f; + + // Derivatives in the y-direction + if(j != dimY-1) + //grady_upd[(dimX*dimY)*k + i+l] = u_upd[(dimX*dimY)*k + i+dimY+l] -u_upd[(dimX*dimY)*k + i+l]; + grady_upd[(dimX*dimY)*k + i+l] = u_upd[(dimX*dimY)*k + i+(j+1)*dimX] -u_upd[(dimX*dimY)*k + i+l]; + else + grady_upd[(dimX*dimY)*k + i+l] = 0.0f; + } + } + } + return 1; +} + +float proxF(float *gx, float *gy, float *vx, float *vy, float sigma, int p, int q, int r, long dimX, long dimY, long dimZ) +{ + // (S^p, \ell^1) norm decouples at each pixel +// Spl1(gx, gy, vx, vy, sigma, p, num_channels, dim); + float divsigma = 1.0f / sigma; + + // $\ell^{1,1,1}$-TV regularization +// int i,j,k; +// #pragma omp parallel for shared (gx,gy,vx,vy) private(i,j,k) +// for(k = 0; k < dimZ; k++) { +// for(i=0; i<dimX; i++) { +// for(j=0; j<dimY; j++) { +// gx[(dimX*dimY)*k + (i)*dimY + (j)] = SIGN(vx[(dimX*dimY)*k + (i)*dimY + (j)]) * MAX(fabs(vx[(dimX*dimY)*k + (i)*dimY + (j)]) - divsigma, 0.0f); +// gy[(dimX*dimY)*k + (i)*dimY + (j)] = SIGN(vy[(dimX*dimY)*k + (i)*dimY + (j)]) * MAX(fabs(vy[(dimX*dimY)*k + (i)*dimY + (j)]) - divsigma, 0.0f); +// }}} + + // Auxiliar vector + float *proj, sum, shrinkfactor ; + float M1,M2,M3,valuex,valuey,T,D,det,eig1,eig2,sig1,sig2,V1, V2, V3, V4, v0,v1,v2, mu1,mu2,sig1_upd,sig2_upd,t1,t2,t3; + long i,j,k, ii, num; + #pragma omp parallel for shared (gx,gy,vx,vy,p) private(i,ii,j,k,proj,num, sum, shrinkfactor, M1,M2,M3,valuex,valuey,T,D,det,eig1,eig2,sig1,sig2,V1, V2, V3, V4,v0,v1,v2,mu1,mu2,sig1_upd,sig2_upd,t1,t2,t3) + for(i=0; i<dimX; i++) { + for(j=0; j<dimY; j++) { + + proj = (float*) calloc (2,sizeof(float)); + // Compute matrix $M\in\R^{2\times 2}$ + M1 = 0.0f; + M2 = 0.0f; + M3 = 0.0f; + + for(k = 0; k < dimZ; k++) + { + valuex = vx[(dimX*dimY)*k + (j)*dimX + (i)]; + valuey = vy[(dimX*dimY)*k + (j)*dimX + (i)]; + + M1 += (valuex * valuex); + M2 += (valuex * valuey); + M3 += (valuey * valuey); + } + + // Compute eigenvalues of M + T = M1 + M3; + D = M1 * M3 - M2 * M2; + det = sqrt(MAX((T * T / 4.0f) - D, 0.0f)); + eig1 = MAX((T / 2.0f) + det, 0.0f); + eig2 = MAX((T / 2.0f) - det, 0.0f); + sig1 = sqrt(eig1); + sig2 = sqrt(eig2); + + // Compute normalized eigenvectors + V1 = V2 = V3 = V4 = 0.0f; + + if(M2 != 0.0f) + { + v0 = M2; + v1 = eig1 - M3; + v2 = eig2 - M3; + + mu1 = sqrtf(v0 * v0 + v1 * v1); + mu2 = sqrtf(v0 * v0 + v2 * v2); + + if(mu1 > fTiny) + { + V1 = v1 / mu1; + V3 = v0 / mu1; + } + + if(mu2 > fTiny) + { + V2 = v2 / mu2; + V4 = v0 / mu2; + } + + } else + { + if(M1 > M3) + { + V1 = V4 = 1.0f; + V2 = V3 = 0.0f; + + } else + { + V1 = V4 = 0.0f; + V2 = V3 = 1.0f; + } + } + + // Compute prox_p of the diagonal entries + sig1_upd = sig2_upd = 0.0f; + + if(p == 1) + { + sig1_upd = MAX(sig1 - divsigma, 0.0f); + sig2_upd = MAX(sig2 - divsigma, 0.0f); + + } else if(p == INFNORM) + { + proj[0] = sigma * fabs(sig1); + proj[1] = sigma * fabs(sig2); + + /*l1 projection part */ + sum = fLarge; + num = 0l; + shrinkfactor = 0.0f; + while(sum > 1.0f) + { + sum = 0.0f; + num = 0; + + for(ii = 0; ii < 2; ii++) + { + proj[ii] = MAX(proj[ii] - shrinkfactor, 0.0f); + + sum += fabs(proj[ii]); + if(proj[ii]!= 0.0f) + num++; + } + + if(num > 0) + shrinkfactor = (sum - 1.0f) / num; + else + break; + } + /*l1 proj ends*/ + + sig1_upd = sig1 - divsigma * proj[0]; + sig2_upd = sig2 - divsigma * proj[1]; + } + + // Compute the diagonal entries of $\widehat{\Sigma}\Sigma^{\dagger}_0$ + if(sig1 > fTiny) + sig1_upd /= sig1; + + if(sig2 > fTiny) + sig2_upd /= sig2; + + // Compute solution + t1 = sig1_upd * V1 * V1 + sig2_upd * V2 * V2; + t2 = sig1_upd * V1 * V3 + sig2_upd * V2 * V4; + t3 = sig1_upd * V3 * V3 + sig2_upd * V4 * V4; + + for(k = 0; k < dimZ; k++) + { + gx[(dimX*dimY)*k + j*dimX + i] = vx[(dimX*dimY)*k + j*dimX + i] * t1 + vy[(dimX*dimY)*k + j*dimX + i] * t2; + gy[(dimX*dimY)*k + j*dimX + i] = vx[(dimX*dimY)*k + j*dimX + i] * t2 + vy[(dimX*dimY)*k + j*dimX + i] * t3; + } + + // Delete allocated memory + free(proj); + } + } + + return 1; +} + +float divergence(float *qx_upd, float *qy_upd, float *div_upd, long dimX, long dimY, long dimZ) +{ + long i, j, k, l; + #pragma omp parallel for shared(qx_upd,qy_upd,div_upd) private(i, j, k, l) + for(k = 0; k < dimZ; k++) { + for(j = 0; j < dimY; j++) { + l = j * dimX; + for(i = 0; i < dimX; i++) { + if(i != dimX-1) + { + // ux[k][i+l] = u[k][i+1+l] - u[k][i+l] + div_upd[(dimX*dimY)*k + i+1+l] -= qx_upd[(dimX*dimY)*k + i+l]; + div_upd[(dimX*dimY)*k + i+l] += qx_upd[(dimX*dimY)*k + i+l]; + } + + if(j != dimY-1) + { + // uy[k][i+l] = u[k][i+width+l] - u[k][i+l] + //div_upd[(dimX*dimY)*k + i+dimY+l] -= qy_upd[(dimX*dimY)*k + i+l]; + div_upd[(dimX*dimY)*k + i+(j+1)*dimX] -= qy_upd[(dimX*dimY)*k + i+l]; + div_upd[(dimX*dimY)*k + i+l] += qy_upd[(dimX*dimY)*k + i+l]; + } + } + } + } + return *div_upd; +} diff --git a/src/Core/performance_CPU/TNV_core_backtrack_loop.h.v19 b/src/Core/performance_CPU/TNV_core_backtrack_loop.h.v19 new file mode 100644 index 0000000..3ec4250 --- /dev/null +++ b/src/Core/performance_CPU/TNV_core_backtrack_loop.h.v19 @@ -0,0 +1,100 @@ + float t[3]; + float M1 = 0.0f, M2 = 0.0f, M3 = 0.0f; + + l = (j * dimX + i) * padZ; + +//#pragma vector aligned +#pragma GCC ivdep + for(k = 0; k < dimZ; k++) { + u_upd[l + k] = (u[l + k] + tau * div[l + k] + taulambda * Input[l + k]) / constant; + udiff[k] = u[l + k] - u_upd[l + k]; + unorm += (udiff[k] * udiff[k]); + +#ifdef TNV_LOOP_LAST_I + gradx_upd[l + k] = 0; +#else + gradx_upd[l + k] = ((u[l + k + padZ] + tau * div[l + k + padZ] + taulambda * Input[l + k + padZ]) / constant - u_upd[l + k]); +#endif + +#ifdef TNV_LOOP_LAST_J + grady_upd[l + k] = 0; +#else + grady_upd[l + k] = ((u[l + k + dimX*padZ] + tau * div[l + k + dimX*padZ] + taulambda * Input[l + k + dimX*padZ]) / constant - u_upd[l + k]); +#endif + + gradxdiff[k] = gradx[l + k] - gradx_upd[l + k]; + gradydiff[k] = grady[l + k] - grady_upd[l + k]; + + float ubarx = theta1 * gradx_upd[l + k] - theta * gradx[l + k]; + float ubary = theta1 * grady_upd[l + k] - theta * grady[l + k]; +//#define TNV_NEW_STYLE +#ifdef TNV_NEW_STYLE + qx_upd[l + k] = qx[l + k] + sigma * ubarx; + qy_upd[l + k] = qy[l + k] + sigma * ubary; + + float vx = divsigma * qx_upd[l + k]; //+ ubarx + float vy = divsigma * qy_upd[l + k]; //+ ubary +#else + float vx = ubarx + divsigma * qx[l + k]; + float vy = ubary + divsigma * qy[l + k]; +#endif + + M1 += (vx * vx); M2 += (vx * vy); M3 += (vy * vy); + } + + coefF(t, M1, M2, M3, sigma, p, q, r); + +//#pragma vector aligned +#pragma GCC ivdep + for(k = 0; k < dimZ; k++) { +#ifdef TNV_NEW_STYLE + float vx = divsigma * qx_upd[l + k]; + float vy = divsigma * qy_upd[l + k]; + + float gx_upd = vx * t[0] + vy * t[1]; + float gy_upd = vx * t[1] + vy * t[2]; + + qx_upd[l + k] -= sigma * gx_upd; + qy_upd[l + k] -= sigma * gy_upd; +#else + float ubarx = theta1 * gradx_upd[l + k] - theta * gradx[l + k]; + float ubary = theta1 * grady_upd[l + k] - theta * grady[l + k]; + float vx = ubarx + divsigma * qx[l + k]; + float vy = ubary + divsigma * qy[l + k]; + + float gx_upd = vx * t[0] + vy * t[1]; + float gy_upd = vx * t[1] + vy * t[2]; + + qx_upd[l + k] = qx[l + k] + sigma * (ubarx - gx_upd); + qy_upd[l + k] = qy[l + k] + sigma * (ubary - gy_upd); +#endif + + float div_upd_val = 0; +#ifndef TNV_LOOP_FIRST_I + div_upd_val -= qx_upd[l + k - padZ]; +#endif + +#ifndef TNV_LOOP_FIRST_J + div_upd_val -= qy_upd[l + k - dimX * padZ]; +#endif +#ifndef TNV_LOOP_LAST_I + div_upd_val += qx_upd[l + k]; +#endif +#ifndef TNV_LOOP_LAST_J + div_upd_val += qy_upd[l + k]; +#endif + div_upd[l + k] = div_upd_val; + + qxdiff = qx[l + k] - qx_upd[l + k]; + qydiff = qy[l + k] - qy_upd[l + k]; + qnorm += (qxdiff * qxdiff + qydiff * qydiff); + + resdual1 += fabs(divsigma * qxdiff - gradxdiff[k]); + resdual2 += fabs(divsigma * qydiff - gradydiff[k]); + product += (gradxdiff[k] * qxdiff + gradydiff[k] * qydiff); + +#ifndef TNV_LOOP_FIRST_J + divdiff = div[l + k] - div_upd[l + k]; // Multiple steps... How we compute without history? + resprimal += fabs(divtau * udiff[k] + divdiff); +#endif + } diff --git a/src/Core/performance_CPU/TNV_core_loop.h.v32 b/src/Core/performance_CPU/TNV_core_loop.h.v32 new file mode 100644 index 0000000..be53156 --- /dev/null +++ b/src/Core/performance_CPU/TNV_core_loop.h.v32 @@ -0,0 +1,119 @@ + { + float t[3]; + float M1 = 0.0f, M2 = 0.0f, M3 = 0.0f; + l = (j * dimX + i) * padZ; + m = dimX * padZ; + + float *__restrict u_next = u + l + padZ; + float *__restrict u_current = u + l; + float *__restrict u_next_row = u + l + m; + + + float *__restrict qx_current = qx + l; + float *__restrict qy_current = qy + l; + float *__restrict qx_prev = qx + l - padZ; + float *__restrict qy_prev = qy + l - m; + + + __assume(padZ%16==0); +/* + __assume_aligned(Input, 64); + __assume_aligned(div, 64); + __assume_aligned(gradx, 64); + __assume_aligned(grady, 64); + __assume_aligned(u, 64); + __assume_aligned(qx, 64); + __assume_aligned(qy, 64); + __assume_aligned(u_current, 64); + __assume_aligned(u_next, 64); + __assume_aligned(u_next_row, 64); +*/ + +#pragma vector aligned +#pragma GCC ivdep + for(k = 0; k < dimZ; k++) { + float u_upd = (u[l + k] + tau * div[l + k] + taulambda * Input[l + k]) / constant; // 3 reads + udiff[k] = u[l + k] - u_upd; // cache 1w + u[l + k] = u_upd; // 1 write + +#ifdef TNV_LOOP_FIRST_J + udiff0[l + k] = udiff[k]; + div0[l + k] = div[l + k]; +#endif + +#ifdef TNV_LOOP_LAST_I + float gradx_upd = 0; +#else + float u_next_upd = (u[l + k + padZ] + tau * div[l + k + padZ] + taulambda * Input[l + k + padZ]) / constant; // 3 reads + float gradx_upd = (u_next_upd - u_upd); // 2 reads +#endif + +#ifdef TNV_LOOP_LAST_J + float grady_upd = 0; +#else + float u_next_row_upd = (u[l + k + m] + tau * div[l + k + m] + taulambda * Input[l + k + m]) / constant; // 3 reads + float grady_upd = (u_next_row_upd - u_upd); // 1 read +#endif + + gradxdiff[k] = gradx[l + k] - gradx_upd; // 1 read, cache 1w + gradydiff[k] = grady[l + k] - grady_upd; // 1 read, cache 1w + gradx[l + k] = gradx_upd; // 1 write + grady[l + k] = grady_upd; // 1 write + + ubarx[k] = gradx_upd - theta * gradxdiff[k]; // cache 1w + ubary[k] = grady_upd - theta * gradydiff[k]; // cache 1w + + float vx = ubarx[k] + divsigma * qx[l + k]; // 1 read + float vy = ubary[k] + divsigma * qy[l + k]; // 1 read + + M1 += (vx * vx); M2 += (vx * vy); M3 += (vy * vy); + } + + coefF(t, M1, M2, M3, sigma, p, q, r); + +#pragma vector aligned +#pragma GCC ivdep + for(k = 0; k < padZ; k++) { + float vx = ubarx[k] + divsigma * qx_current[k]; // cache 2r + float vy = ubary[k] + divsigma * qy_current[k]; // cache 2r + float gx_upd = vx * t[0] + vy * t[1]; + float gy_upd = vx * t[1] + vy * t[2]; + + qxdiff = sigma * (ubarx[k] - gx_upd); + qydiff = sigma * (ubary[k] - gy_upd); + + qx_current[k] += qxdiff; // write 1 + qy_current[k] += qydiff; // write 1 + + unorm += (udiff[k] * udiff[k]); + qnorm += (qxdiff * qxdiff + qydiff * qydiff); + + float div_upd = 0; + +#ifndef TNV_LOOP_FIRST_I + div_upd -= qx_prev[k]; // 1 read +#endif +#ifndef TNV_LOOP_FIRST_J + div_upd -= qy_prev[k]; // 1 read +#endif +#ifndef TNV_LOOP_LAST_I + div_upd += qx_current[k]; +#endif +#ifndef TNV_LOOP_LAST_J + div_upd += qy_current[k]; +#endif + + divdiff = div[l + k] - div_upd; // 1 read + div[l + k] = div_upd; // 1 write + +#ifndef TNV_LOOP_FIRST_J + resprimal += fabs(divtau * udiff[k] + divdiff); +#endif + + // We need to have two independent accumulators to allow gcc-autovectorization + resdual1 += fabs(divsigma * qxdiff + gradxdiff[k]); // cache 1r + resdual2 += fabs(divsigma * qydiff + gradydiff[k]); // cache 1r + product -= (gradxdiff[k] * qxdiff + gradydiff[k] * qydiff); + } + } +
\ No newline at end of file |