void Waifu2x_Process_Base::process_core_gray()
{
FLType *dstYd = nullptr, *srcYd = nullptr; // *refYd = nullptr
// Get write/read pointer
auto dstY = reinterpret_cast<_Ty *>(vsapi->getWritePtr(dst, 0));
auto srcY = reinterpret_cast<const _Ty *>(vsapi->getReadPtr(src, 0));
// Allocate memory for floating point Y data
AlignedMalloc(dstYd, dst_pcount[0]);
AlignedMalloc(srcYd, src_pcount[0]);
// Convert src and ref from integer Y data to floating point Y data
Int2Float(srcYd, srcY, src_height[0], src_width[0], src_stride[0], src_stride[0], false, d.para.full, false);
// Execute kernel
Kernel(dstYd, srcYd);
// Convert dst from floating point Y data to integer Y data
Float2Int(dstY, dstYd, dst_height[0], dst_width[0], dst_stride[0], dst_stride[0], false, d.para.full, !isFloat(_Ty));
// Free memory for floating point Y data
AlignedFree(dstYd);
AlignedFree(srcYd);
}
void TriangleMesh::ReallocVertexBuffer(int numTris, int vertexSizeBytes_)
{
AlignedFree(data);
vertexSizeBytes = vertexSizeBytes_;
data = (float*)AlignedMalloc(numTris * 3 * vertexSizeBytes, 32);
numTriangles = numTris;
}
HOTTILE* HotTileMgr::GetHotTileNoLoad(SWR_CONTEXT* pContext,
DRAW_CONTEXT* pDC,
uint32_t macroID,
SWR_RENDERTARGET_ATTACHMENT attachment,
bool create,
uint32_t numSamples)
{
uint32_t x, y;
MacroTileMgr::getTileIndices(macroID, x, y);
SWR_ASSERT(x < KNOB_NUM_HOT_TILES_X);
SWR_ASSERT(y < KNOB_NUM_HOT_TILES_Y);
HotTileSet& tile = mHotTiles[x][y];
HOTTILE& hotTile = tile.Attachment[attachment];
if (hotTile.pBuffer == NULL)
{
if (create)
{
uint32_t size = numSamples * mHotTileSize[attachment];
hotTile.pBuffer = (uint8_t*)AlignedMalloc(size, 64);
hotTile.state = HOTTILE_INVALID;
hotTile.numSamples = numSamples;
hotTile.renderTargetArrayIndex = 0;
}
else
{
return NULL;
}
}
return &hotTile;
}
void Waifu2x_Process_Base::process_core_rgb()
{
FLType *dstYd = nullptr, *dstUd = nullptr, *dstVd = nullptr;
FLType *srcYd = nullptr, *srcUd = nullptr, *srcVd = nullptr;
// Get write/read pointer
auto dstR = reinterpret_cast<_Ty *>(vsapi->getWritePtr(dst, 0));
auto dstG = reinterpret_cast<_Ty *>(vsapi->getWritePtr(dst, 1));
auto dstB = reinterpret_cast<_Ty *>(vsapi->getWritePtr(dst, 2));
auto srcR = reinterpret_cast<const _Ty *>(vsapi->getReadPtr(src, 0));
auto srcG = reinterpret_cast<const _Ty *>(vsapi->getReadPtr(src, 1));
auto srcB = reinterpret_cast<const _Ty *>(vsapi->getReadPtr(src, 2));
// Allocate memory for floating point YUV data
AlignedMalloc(dstYd, dst_pcount[0]);
if (d.chroma) AlignedMalloc(dstUd, dst_pcount[1]);
if (d.chroma) AlignedMalloc(dstVd, dst_pcount[2]);
AlignedMalloc(srcYd, src_pcount[0]);
AlignedMalloc(srcUd, src_pcount[1]);
AlignedMalloc(srcVd, src_pcount[2]);
// Convert src and ref from RGB data to floating point YUV data
RGB2FloatYUV(srcYd, srcUd, srcVd, srcR, srcG, srcB,
src_height[0], src_width[0], src_stride[0], src_stride[0],
d.para.matrix, d.para.full, false);
// Execute kernel
if (d.chroma)
{
Kernel(dstYd, dstUd, dstVd, srcYd, srcUd, srcVd);
}
else
{
Kernel(dstYd, srcYd);
dstUd = srcUd;
dstVd = srcVd;
}
// Convert dst from floating point YUV data to RGB data
FloatYUV2RGB(dstR, dstG, dstB, dstYd, dstUd, dstVd,
dst_height[0], dst_width[0], dst_stride[0], dst_stride[0],
d.para.matrix, d.para.full, !isFloat(_Ty));
// Free memory for floating point YUV data
AlignedFree(dstYd);
if (d.chroma) AlignedFree(dstUd);
if (d.chroma) AlignedFree(dstVd);
AlignedFree(srcYd);
AlignedFree(srcUd);
AlignedFree(srcVd);
}
bool MultilayerPerceptron::AddLayer( Layer* in_layer)
{
if(_layers.size() > 0)
{
if(_layers.back()->outputs != in_layer->inputs)
{
return false;
}
size_t input_alloc_size = sizeof(float) * BlockCount(in_layer->inputs) * 4;
float* buffer = (float*)AlignedMalloc(input_alloc_size, 16);
memset(buffer, 0x00, input_alloc_size);
_activations.back() = buffer;
_activations.push_back(nullptr);
}
_layers.push_back(in_layer);
return true;
}
void Waifu2x_Process_Base::process_core_yuv()
{
FLType *dstYd = nullptr, *dstUd = nullptr, *dstVd = nullptr;
FLType *srcYd = nullptr, *srcUd = nullptr, *srcVd = nullptr;
// Get write/read pointer
auto dstY = reinterpret_cast<_Ty *>(vsapi->getWritePtr(dst, 0));
auto dstU = reinterpret_cast<_Ty *>(vsapi->getWritePtr(dst, 1));
auto dstV = reinterpret_cast<_Ty *>(vsapi->getWritePtr(dst, 2));
auto srcY = reinterpret_cast<const _Ty *>(vsapi->getReadPtr(src, 0));
auto srcU = reinterpret_cast<const _Ty *>(vsapi->getReadPtr(src, 1));
auto srcV = reinterpret_cast<const _Ty *>(vsapi->getReadPtr(src, 2));
// Allocate memory for floating point YUV data
AlignedMalloc(dstYd, dst_pcount[0]);
if (d.process[1]) AlignedMalloc(dstUd, dst_pcount[1]);
if (d.process[2]) AlignedMalloc(dstVd, dst_pcount[2]);
AlignedMalloc(srcYd, src_pcount[0]);
if (d.process[1]) AlignedMalloc(srcUd, src_pcount[1]);
if (d.process[2]) AlignedMalloc(srcVd, src_pcount[2]);
// Convert src and ref from integer YUV data to floating point YUV data
Int2Float(srcYd, srcY, src_height[0], src_width[0], src_stride[0], src_stride[0], false, d.para.full, false);
if (d.process[1]) Int2Float(srcUd, srcU, src_height[1], src_width[1], src_stride[1], src_stride[1], true, d.para.full, false);
if (d.process[2]) Int2Float(srcVd, srcV, src_height[2], src_width[2], src_stride[2], src_stride[2], true, d.para.full, false);
// Execute kernel
if (d.chroma) Kernel(dstYd, dstUd, dstVd, srcYd, srcUd, srcVd);
else Kernel(dstYd, srcYd);
// Convert dst from floating point YUV data to integer YUV data
Float2Int(dstY, dstYd, dst_height[0], dst_width[0], dst_stride[0], dst_stride[0], false, d.para.full, !isFloat(_Ty));
if (d.process[1]) Float2Int(dstU, dstUd, dst_height[1], dst_width[1], dst_stride[1], dst_stride[1], true, d.para.full, !isFloat(_Ty));
if (d.process[2]) Float2Int(dstV, dstVd, dst_height[2], dst_width[2], dst_stride[2], dst_stride[2], true, d.para.full, !isFloat(_Ty));
// Free memory for floating point YUV data
AlignedFree(dstYd);
if (d.process[1]) AlignedFree(dstUd);
if (d.process[2]) AlignedFree(dstVd);
AlignedFree(srcYd);
if (d.process[1]) AlignedFree(srcUd);
if (d.process[2]) AlignedFree(srcVd);
}
static boolean
swr_texture_layout(struct swr_screen *screen,
struct swr_resource *res,
boolean allocate)
{
struct pipe_resource *pt = &res->base;
pipe_format fmt = pt->format;
const struct util_format_description *desc = util_format_description(fmt);
res->has_depth = util_format_has_depth(desc);
res->has_stencil = util_format_has_stencil(desc);
if (res->has_stencil && !res->has_depth)
fmt = PIPE_FORMAT_R8_UINT;
res->swr.width = pt->width0;
res->swr.height = pt->height0;
res->swr.depth = pt->depth0;
res->swr.type = swr_convert_target_type(pt->target);
res->swr.tileMode = SWR_TILE_NONE;
res->swr.format = mesa_to_swr_format(fmt);
res->swr.numSamples = (1 << pt->nr_samples);
SWR_FORMAT_INFO finfo = GetFormatInfo(res->swr.format);
unsigned total_size = 0;
unsigned width = pt->width0;
unsigned height = pt->height0;
unsigned depth = pt->depth0;
unsigned layers = pt->array_size;
for (int level = 0; level <= pt->last_level; level++) {
unsigned alignedWidth, alignedHeight;
unsigned num_slices;
if (pt->bind & (PIPE_BIND_RENDER_TARGET | PIPE_BIND_DEPTH_STENCIL)) {
alignedWidth = align(width, KNOB_MACROTILE_X_DIM);
alignedHeight = align(height, KNOB_MACROTILE_Y_DIM);
} else {
alignedWidth = width;
alignedHeight = height;
}
if (level == 0) {
res->alignedWidth = alignedWidth;
res->alignedHeight = alignedHeight;
}
res->row_stride[level] = alignedWidth * finfo.Bpp;
res->img_stride[level] = res->row_stride[level] * alignedHeight;
res->mip_offsets[level] = total_size;
if (pt->target == PIPE_TEXTURE_3D)
num_slices = depth;
else if (pt->target == PIPE_TEXTURE_1D_ARRAY
|| pt->target == PIPE_TEXTURE_2D_ARRAY
|| pt->target == PIPE_TEXTURE_CUBE
|| pt->target == PIPE_TEXTURE_CUBE_ARRAY)
num_slices = layers;
else
num_slices = 1;
total_size += res->img_stride[level] * num_slices;
if (total_size > SWR_MAX_TEXTURE_SIZE)
return FALSE;
width = u_minify(width, 1);
height = u_minify(height, 1);
depth = u_minify(depth, 1);
}
res->swr.halign = res->alignedWidth;
res->swr.valign = res->alignedHeight;
res->swr.pitch = res->row_stride[0];
if (allocate) {
res->swr.pBaseAddress = (uint8_t *)AlignedMalloc(total_size, 64);
if (res->has_depth && res->has_stencil) {
SWR_FORMAT_INFO finfo = GetFormatInfo(res->secondary.format);
res->secondary.width = pt->width0;
res->secondary.height = pt->height0;
res->secondary.depth = pt->depth0;
res->secondary.type = SURFACE_2D;
res->secondary.tileMode = SWR_TILE_NONE;
res->secondary.format = R8_UINT;
res->secondary.numSamples = (1 << pt->nr_samples);
res->secondary.pitch = res->alignedWidth * finfo.Bpp;
res->secondary.pBaseAddress = (uint8_t *)AlignedMalloc(
res->alignedHeight * res->secondary.pitch, 64);
}
}
return TRUE;
}
struct pipe_context *
swr_create_context(struct pipe_screen *p_screen, void *priv, unsigned flags)
{
struct swr_context *ctx = (struct swr_context *)
AlignedMalloc(sizeof(struct swr_context), KNOB_SIMD_BYTES);
memset(ctx, 0, sizeof(struct swr_context));
swr_screen(p_screen)->pfnSwrGetInterface(ctx->api);
ctx->swrDC.pAPI = &ctx->api;
ctx->blendJIT =
new std::unordered_map<BLEND_COMPILE_STATE, PFN_BLEND_JIT_FUNC>;
ctx->max_draws_in_flight = KNOB_MAX_DRAWS_IN_FLIGHT;
SWR_CREATECONTEXT_INFO createInfo;
memset(&createInfo, 0, sizeof(createInfo));
createInfo.privateStateSize = sizeof(swr_draw_context);
createInfo.pfnLoadTile = swr_LoadHotTile;
createInfo.pfnStoreTile = swr_StoreHotTile;
createInfo.pfnClearTile = swr_StoreHotTileClear;
createInfo.pfnUpdateStats = swr_UpdateStats;
createInfo.pfnUpdateStatsFE = swr_UpdateStatsFE;
SWR_THREADING_INFO threadingInfo {0};
threadingInfo.MAX_WORKER_THREADS = KNOB_MAX_WORKER_THREADS;
threadingInfo.MAX_NUMA_NODES = KNOB_MAX_NUMA_NODES;
threadingInfo.MAX_CORES_PER_NUMA_NODE = KNOB_MAX_CORES_PER_NUMA_NODE;
threadingInfo.MAX_THREADS_PER_CORE = KNOB_MAX_THREADS_PER_CORE;
threadingInfo.SINGLE_THREADED = KNOB_SINGLE_THREADED;
// Use non-standard settings for KNL
if (swr_screen(p_screen)->is_knl)
{
if (nullptr == getenv("KNOB_MAX_THREADS_PER_CORE"))
threadingInfo.MAX_THREADS_PER_CORE = 2;
if (nullptr == getenv("KNOB_MAX_DRAWS_IN_FLIGHT"))
{
ctx->max_draws_in_flight = 2048;
createInfo.MAX_DRAWS_IN_FLIGHT = ctx->max_draws_in_flight;
}
}
createInfo.pThreadInfo = &threadingInfo;
ctx->swrContext = ctx->api.pfnSwrCreateContext(&createInfo);
ctx->api.pfnSwrInit();
if (ctx->swrContext == NULL)
goto fail;
ctx->pipe.screen = p_screen;
ctx->pipe.destroy = swr_destroy;
ctx->pipe.priv = priv;
ctx->pipe.create_surface = swr_create_surface;
ctx->pipe.surface_destroy = swr_surface_destroy;
ctx->pipe.transfer_map = swr_transfer_map;
ctx->pipe.transfer_unmap = swr_transfer_unmap;
ctx->pipe.transfer_flush_region = swr_transfer_flush_region;
ctx->pipe.buffer_subdata = u_default_buffer_subdata;
ctx->pipe.texture_subdata = u_default_texture_subdata;
ctx->pipe.clear_texture = util_clear_texture;
ctx->pipe.resource_copy_region = swr_resource_copy;
ctx->pipe.render_condition = swr_render_condition;
swr_state_init(&ctx->pipe);
swr_clear_init(&ctx->pipe);
swr_draw_init(&ctx->pipe);
swr_query_init(&ctx->pipe);
ctx->pipe.stream_uploader = u_upload_create_default(&ctx->pipe);
if (!ctx->pipe.stream_uploader)
goto fail;
ctx->pipe.const_uploader = ctx->pipe.stream_uploader;
ctx->pipe.blit = swr_blit;
ctx->blitter = util_blitter_create(&ctx->pipe);
if (!ctx->blitter)
goto fail;
swr_init_scratch_buffers(ctx);
return &ctx->pipe;
fail:
/* Should really validate the init steps and fail gracefully */
swr_destroy(&ctx->pipe);
return NULL;
}
T *alloc(size_t size) const
{
T *p = static_cast<T*>(AlignedMalloc(size * sizeof(T), N));
if (p == 0) throw std::bad_alloc();
return p;
}
static bool
swr_texture_layout(struct swr_screen *screen,
struct swr_resource *res,
boolean allocate)
{
struct pipe_resource *pt = &res->base;
pipe_format fmt = pt->format;
const struct util_format_description *desc = util_format_description(fmt);
res->has_depth = util_format_has_depth(desc);
res->has_stencil = util_format_has_stencil(desc);
if (res->has_stencil && !res->has_depth)
fmt = PIPE_FORMAT_R8_UINT;
/* We always use the SWR layout. For 2D and 3D textures this looks like:
*
* |<------- pitch ------->|
* +=======================+-------
* |Array 0 | ^
* | | |
* | Level 0 | |
* | | |
* | | qpitch
* +-----------+-----------+ |
* | | L2L2L2L2 | |
* | Level 1 | L3L3 | |
* | | L4 | v
* +===========+===========+-------
* |Array 1 |
* | |
* | Level 0 |
* | |
* | |
* +-----------+-----------+
* | | L2L2L2L2 |
* | Level 1 | L3L3 |
* | | L4 |
* +===========+===========+
*
* The overall width in bytes is known as the pitch, while the overall
* height in rows is the qpitch. Array slices are laid out logically below
* one another, qpitch rows apart. For 3D surfaces, the "level" values are
* just invalid for the higher array numbers (since depth is also
* minified). 1D and 1D array surfaces are stored effectively the same way,
* except that pitch never plays into it. All the levels are logically
* adjacent to each other on the X axis. The qpitch becomes the number of
* elements between array slices, while the pitch is unused.
*
* Each level's sizes are subject to the valign and halign settings of the
* surface. For compressed formats that swr is unaware of, we will use an
* appropriately-sized uncompressed format, and scale the widths/heights.
*
* This surface is stored inside res->swr. For depth/stencil textures,
* res->secondary will have an identically-laid-out but R8_UINT-formatted
* stencil tree. In the Z32F_S8 case, the primary surface still has 64-bpp
* texels, to simplify map/unmap logic which copies the stencil values
* in/out.
*/
res->swr.width = pt->width0;
res->swr.height = pt->height0;
res->swr.type = swr_convert_target_type(pt->target);
res->swr.tileMode = SWR_TILE_NONE;
res->swr.format = mesa_to_swr_format(fmt);
res->swr.numSamples = std::max(1u, pt->nr_samples);
if (pt->bind & (PIPE_BIND_RENDER_TARGET | PIPE_BIND_DEPTH_STENCIL)) {
res->swr.halign = KNOB_MACROTILE_X_DIM;
res->swr.valign = KNOB_MACROTILE_Y_DIM;
} else {
res->swr.halign = 1;
res->swr.valign = 1;
}
unsigned halign = res->swr.halign * util_format_get_blockwidth(fmt);
unsigned width = align(pt->width0, halign);
if (pt->target == PIPE_TEXTURE_1D || pt->target == PIPE_TEXTURE_1D_ARRAY) {
for (int level = 1; level <= pt->last_level; level++)
width += align(u_minify(pt->width0, level), halign);
res->swr.pitch = util_format_get_blocksize(fmt);
res->swr.qpitch = util_format_get_nblocksx(fmt, width);
} else {
// The pitch is the overall width of the texture in bytes. Most of the
// time this is the pitch of level 0 since all the other levels fit
// underneath it. However in some degenerate situations, the width of
// level1 + level2 may be larger. In that case, we use those
// widths. This can happen if, e.g. halign is 32, and the width of level
// 0 is 32 or less. In that case, the aligned levels 1 and 2 will also
// be 32 each, adding up to 64.
unsigned valign = res->swr.valign * util_format_get_blockheight(fmt);
if (pt->last_level > 1) {
width = std::max<uint32_t>(
width,
align(u_minify(pt->width0, 1), halign) +
align(u_minify(pt->width0, 2), halign));
}
res->swr.pitch = util_format_get_stride(fmt, width);
// The qpitch is controlled by either the height of the second LOD, or
// the combination of all the later LODs.
unsigned height = align(pt->height0, valign);
if (pt->last_level == 1) {
height += align(u_minify(pt->height0, 1), valign);
} else if (pt->last_level > 1) {
unsigned level1 = align(u_minify(pt->height0, 1), valign);
unsigned level2 = 0;
for (int level = 2; level <= pt->last_level; level++) {
level2 += align(u_minify(pt->height0, level), valign);
}
height += std::max(level1, level2);
}
res->swr.qpitch = util_format_get_nblocksy(fmt, height);
}
if (pt->target == PIPE_TEXTURE_3D)
res->swr.depth = pt->depth0;
else
res->swr.depth = pt->array_size;
// Fix up swr format if necessary so that LOD offset computation works
if (res->swr.format == (SWR_FORMAT)-1) {
switch (util_format_get_blocksize(fmt)) {
default:
unreachable("Unexpected format block size");
case 1: res->swr.format = R8_UINT; break;
case 2: res->swr.format = R16_UINT; break;
case 4: res->swr.format = R32_UINT; break;
case 8:
if (util_format_is_compressed(fmt))
res->swr.format = BC4_UNORM;
else
res->swr.format = R32G32_UINT;
break;
case 16:
if (util_format_is_compressed(fmt))
res->swr.format = BC5_UNORM;
else
res->swr.format = R32G32B32A32_UINT;
break;
}
}
for (int level = 0; level <= pt->last_level; level++) {
res->mip_offsets[level] =
ComputeSurfaceOffset<false>(0, 0, 0, 0, 0, level, &res->swr);
}
size_t total_size =
(size_t)res->swr.depth * res->swr.qpitch * res->swr.pitch;
if (total_size > SWR_MAX_TEXTURE_SIZE)
return false;
if (allocate) {
res->swr.pBaseAddress = (uint8_t *)AlignedMalloc(total_size, 64);
if (res->has_depth && res->has_stencil) {
res->secondary = res->swr;
res->secondary.format = R8_UINT;
res->secondary.pitch = res->swr.pitch / util_format_get_blocksize(fmt);
for (int level = 0; level <= pt->last_level; level++) {
res->secondary_mip_offsets[level] =
ComputeSurfaceOffset<false>(0, 0, 0, 0, 0, level, &res->secondary);
}
res->secondary.pBaseAddress = (uint8_t *)AlignedMalloc(
res->secondary.depth * res->secondary.qpitch *
res->secondary.pitch, 64);
}
}
return true;
}
void TriangleMesh::ReallocVertexBuffer(int numTris)
{
AlignedFree(data);
data = (float*)AlignedMalloc(numTris*3*3*sizeof(float), 32);
numTriangles = numTris;
}
struct pipe_context *
swr_create_context(struct pipe_screen *p_screen, void *priv, unsigned flags)
{
struct swr_context *ctx = (struct swr_context *)
AlignedMalloc(sizeof(struct swr_context), KNOB_SIMD_BYTES);
memset(ctx, 0, sizeof(struct swr_context));
swr_screen(p_screen)->pfnSwrGetInterface(ctx->api);
ctx->swrDC.pAPI = &ctx->api;
ctx->blendJIT =
new std::unordered_map<BLEND_COMPILE_STATE, PFN_BLEND_JIT_FUNC>;
SWR_CREATECONTEXT_INFO createInfo;
memset(&createInfo, 0, sizeof(createInfo));
createInfo.privateStateSize = sizeof(swr_draw_context);
createInfo.pfnLoadTile = swr_LoadHotTile;
createInfo.pfnStoreTile = swr_StoreHotTile;
createInfo.pfnClearTile = swr_StoreHotTileClear;
createInfo.pfnUpdateStats = swr_UpdateStats;
createInfo.pfnUpdateStatsFE = swr_UpdateStatsFE;
ctx->swrContext = ctx->api.pfnSwrCreateContext(&createInfo);
ctx->api.pfnSwrInit();
if (ctx->swrContext == NULL)
goto fail;
ctx->pipe.screen = p_screen;
ctx->pipe.destroy = swr_destroy;
ctx->pipe.priv = priv;
ctx->pipe.create_surface = swr_create_surface;
ctx->pipe.surface_destroy = swr_surface_destroy;
ctx->pipe.transfer_map = swr_transfer_map;
ctx->pipe.transfer_unmap = swr_transfer_unmap;
ctx->pipe.transfer_flush_region = swr_transfer_flush_region;
ctx->pipe.buffer_subdata = u_default_buffer_subdata;
ctx->pipe.texture_subdata = u_default_texture_subdata;
ctx->pipe.clear_texture = util_clear_texture;
ctx->pipe.resource_copy_region = swr_resource_copy;
ctx->pipe.render_condition = swr_render_condition;
swr_state_init(&ctx->pipe);
swr_clear_init(&ctx->pipe);
swr_draw_init(&ctx->pipe);
swr_query_init(&ctx->pipe);
ctx->pipe.stream_uploader = u_upload_create_default(&ctx->pipe);
if (!ctx->pipe.stream_uploader)
goto fail;
ctx->pipe.const_uploader = ctx->pipe.stream_uploader;
ctx->pipe.blit = swr_blit;
ctx->blitter = util_blitter_create(&ctx->pipe);
if (!ctx->blitter)
goto fail;
swr_init_scratch_buffers(ctx);
return &ctx->pipe;
fail:
/* Should really validate the init steps and fail gracefully */
swr_destroy(&ctx->pipe);
return NULL;
}
BM3D_FilterData::BM3D_FilterData(bool wiener, double sigma, PCType GroupSize, PCType BlockSize, double lambda)
: fp(GroupSize), bp(GroupSize), finalAMP(GroupSize), thrTable(wiener ? 0 : GroupSize),
wienerSigmaSqr(wiener ? GroupSize : 0)
{
const unsigned int flags = FFTW_PATIENT;
const fftw::r2r_kind fkind = FFTW_REDFT10;
const fftw::r2r_kind bkind = FFTW_REDFT01;
FLType *temp = nullptr;
for (PCType i = 1; i <= GroupSize; ++i)
{
AlignedMalloc(temp, i * BlockSize * BlockSize);
fp[i - 1].r2r_3d(i, BlockSize, BlockSize, temp, temp, fkind, fkind, fkind, flags);
bp[i - 1].r2r_3d(i, BlockSize, BlockSize, temp, temp, bkind, bkind, bkind, flags);
AlignedFree(temp);
finalAMP[i - 1] = 2 * i * 2 * BlockSize * 2 * BlockSize;
double forwardAMP = sqrt(finalAMP[i - 1]);
if (wiener)
{
wienerSigmaSqr[i - 1] = static_cast<FLType>(sigma * forwardAMP * sigma * forwardAMP);
}
else
{
double thrBase = sigma * lambda * forwardAMP;
std::vector<double> thr(4);
thr[0] = thrBase;
thr[1] = thrBase * sqrt(double(2));
thr[2] = thrBase * double(2);
thr[3] = thrBase * sqrt(double(8));
FLType *thrp = nullptr;
AlignedMalloc(thrp, i * BlockSize * BlockSize);
thrTable[i - 1].reset(thrp, [](FLType *memory)
{
AlignedFree(memory);
});
for (PCType z = 0; z < i; ++z)
{
for (PCType y = 0; y < BlockSize; ++y)
{
for (PCType x = 0; x < BlockSize; ++x, ++thrp)
{
int flag = 0;
if (x == 0)
{
++flag;
}
if (y == 0)
{
++flag;
}
if (z == 0)
{
++flag;
}
*thrp = static_cast<FLType>(thr[flag]);
}
}
}
}
}
}