[Bf-blender-cvs] [33b5e8daff2] master: Code refactor: replace CUDA array with linear memory for 1D and 2D textures.
Brecht Van Lommel
noreply at git.blender.org
Sat Nov 4 02:23:24 CET 2017
Commit: 33b5e8daff2cfe269bdb527cf0de6dd083b01daf
Author: Brecht Van Lommel
Date: Fri Nov 3 21:28:13 2017 +0100
Branches: master
https://developer.blender.org/rB33b5e8daff2cfe269bdb527cf0de6dd083b01daf
Code refactor: replace CUDA array with linear memory for 1D and 2D textures.
This is a prequisite for getting host memory allocation to work. There appears
to be no support for 3D textures using host memory. The original version of
this code was written by Stefan Werner for D2056.
===================================================================
M intern/cycles/device/device_cuda.cpp
===================================================================
diff --git a/intern/cycles/device/device_cuda.cpp b/intern/cycles/device/device_cuda.cpp
index c951364b53a..278fff02ae1 100644
--- a/intern/cycles/device/device_cuda.cpp
+++ b/intern/cycles/device/device_cuda.cpp
@@ -553,7 +553,7 @@ public:
}
}
- void generic_alloc(device_memory& mem)
+ void generic_alloc(device_memory& mem, size_t padding = 0)
{
CUDAContextScope scope(this);
@@ -565,7 +565,7 @@ public:
CUdeviceptr device_pointer;
size_t size = mem.memory_size();
- cuda_assert(cuMemAlloc(&device_pointer, size));
+ cuda_assert(cuMemAlloc(&device_pointer, size + padding));
mem.device_pointer = (device_ptr)device_pointer;
mem.device_size = size;
stats.mem_alloc(size);
@@ -732,30 +732,8 @@ public:
filter_mode = CU_TR_FILTER_MODE_LINEAR;
}
- /* General variables for Fermi */
- CUtexref texref = NULL;
-
- if(has_fermi_limits && mem.interpolation != INTERPOLATION_NONE) {
- if(mem.data_depth > 1) {
- /* Kernel uses different bind names for 2d and 3d float textures,
- * so we have to adjust couple of things here.
- */
- vector<string> tokens;
- string_split(tokens, mem.name, "_");
- bind_name = string_printf("__tex_image_%s_3d_%s",
- tokens[2].c_str(),
- tokens[3].c_str());
- }
-
- cuda_assert(cuModuleGetTexRef(&texref, cuModule, bind_name.c_str()));
-
- if(!texref) {
- return;
- }
- }
-
+ /* Data Storage */
if(mem.interpolation == INTERPOLATION_NONE) {
- /* Data Storage */
generic_alloc(mem);
generic_copy_to(mem);
@@ -774,179 +752,243 @@ public:
uint32_t ptr = (uint32_t)mem.device_pointer;
cuda_assert(cuMemcpyHtoD(cumem, (void*)&ptr, cubytes));
}
+
+ tex_interp_map[mem.device_pointer] = false;
+ return;
}
- else {
- /* Texture Storage */
- CUarray handle = NULL;
-
- CUarray_format_enum format;
- switch(mem.data_type) {
- case TYPE_UCHAR: format = CU_AD_FORMAT_UNSIGNED_INT8; break;
- case TYPE_UINT: format = CU_AD_FORMAT_UNSIGNED_INT32; break;
- case TYPE_INT: format = CU_AD_FORMAT_SIGNED_INT32; break;
- case TYPE_FLOAT: format = CU_AD_FORMAT_FLOAT; break;
- case TYPE_HALF: format = CU_AD_FORMAT_HALF; break;
- default: assert(0); return;
- }
+ /* Image Texture Storage */
+ CUtexref texref = NULL;
+
+ if(has_fermi_limits) {
if(mem.data_depth > 1) {
- CUDA_ARRAY3D_DESCRIPTOR desc;
+ /* Kernel uses different bind names for 2d and 3d float textures,
+ * so we have to adjust couple of things here.
+ */
+ vector<string> tokens;
+ string_split(tokens, mem.name, "_");
+ bind_name = string_printf("__tex_image_%s_3d_%s",
+ tokens[2].c_str(),
+ tokens[3].c_str());
+ }
- desc.Width = mem.data_width;
- desc.Height = mem.data_height;
- desc.Depth = mem.data_depth;
- desc.Format = format;
- desc.NumChannels = mem.data_elements;
- desc.Flags = 0;
+ cuda_assert(cuModuleGetTexRef(&texref, cuModule, bind_name.c_str()));
- cuda_assert(cuArray3DCreate(&handle, &desc));
+ if(!texref) {
+ return;
}
- else {
- CUDA_ARRAY_DESCRIPTOR desc;
+ }
- desc.Width = mem.data_width;
- desc.Height = mem.data_height;
- desc.Format = format;
- desc.NumChannels = mem.data_elements;
+ CUarray_format_enum format;
+ switch(mem.data_type) {
+ case TYPE_UCHAR: format = CU_AD_FORMAT_UNSIGNED_INT8; break;
+ case TYPE_UINT: format = CU_AD_FORMAT_UNSIGNED_INT32; break;
+ case TYPE_INT: format = CU_AD_FORMAT_SIGNED_INT32; break;
+ case TYPE_FLOAT: format = CU_AD_FORMAT_FLOAT; break;
+ case TYPE_HALF: format = CU_AD_FORMAT_HALF; break;
+ default: assert(0); return;
+ }
- cuda_assert(cuArrayCreate(&handle, &desc));
- }
- if(!handle) {
+ CUarray array_3d = NULL;
+ size_t src_pitch = mem.data_width * dsize * mem.data_elements;
+ size_t dst_pitch = src_pitch;
+
+ if(mem.data_depth > 1) {
+ /* 3D texture using array, there is no API for linear memory. */
+ CUDA_ARRAY3D_DESCRIPTOR desc;
+
+ desc.Width = mem.data_width;
+ desc.Height = mem.data_height;
+ desc.Depth = mem.data_depth;
+ desc.Format = format;
+ desc.NumChannels = mem.data_elements;
+ desc.Flags = 0;
+
+ cuda_assert(cuArray3DCreate(&array_3d, &desc));
+
+ if(!array_3d) {
return;
}
- /* Allocate 3D, 2D or 1D memory */
- if(mem.data_depth > 1) {
- CUDA_MEMCPY3D param;
- memset(¶m, 0, sizeof(param));
- param.dstMemoryType = CU_MEMORYTYPE_ARRAY;
- param.dstArray = handle;
- param.srcMemoryType = CU_MEMORYTYPE_HOST;
- param.srcHost = (void*)mem.data_pointer;
- param.srcPitch = mem.data_width*dsize*mem.data_elements;
- param.WidthInBytes = param.srcPitch;
- param.Height = mem.data_height;
- param.Depth = mem.data_depth;
-
- cuda_assert(cuMemcpy3D(¶m));
- }
- else if(mem.data_height > 1) {
- CUDA_MEMCPY2D param;
- memset(¶m, 0, sizeof(param));
- param.dstMemoryType = CU_MEMORYTYPE_ARRAY;
- param.dstArray = handle;
- param.srcMemoryType = CU_MEMORYTYPE_HOST;
- param.srcHost = (void*)mem.data_pointer;
- param.srcPitch = mem.data_width*dsize*mem.data_elements;
- param.WidthInBytes = param.srcPitch;
- param.Height = mem.data_height;
-
- cuda_assert(cuMemcpy2D(¶m));
- }
- else
- cuda_assert(cuMemcpyHtoA(handle, 0, (void*)mem.data_pointer, size));
+ CUDA_MEMCPY3D param;
+ memset(¶m, 0, sizeof(param));
+ param.dstMemoryType = CU_MEMORYTYPE_ARRAY;
+ param.dstArray = array_3d;
+ param.srcMemoryType = CU_MEMORYTYPE_HOST;
+ param.srcHost = (void*)mem.data_pointer;
+ param.srcPitch = src_pitch;
+ param.WidthInBytes = param.srcPitch;
+ param.Height = mem.data_height;
+ param.Depth = mem.data_depth;
- /* Fermi and Kepler */
- mem.device_pointer = (device_ptr)handle;
- mem.device_size = size;
+ cuda_assert(cuMemcpy3D(¶m));
+ mem.device_pointer = (device_ptr)array_3d;
+ mem.device_size = size;
stats.mem_alloc(size);
+ }
+ else if(mem.data_height > 1) {
+ /* 2D texture, using pitch aligned linear memory. */
+ int alignment = 0;
+ cuda_assert(cuDeviceGetAttribute(&alignment, CU_DEVICE_ATTRIBUTE_TEXTURE_PITCH_ALIGNMENT, cuDevice));
+ dst_pitch = align_up(src_pitch, alignment);
+ size_t dst_size = dst_pitch * mem.data_height;
- if(!has_fermi_limits) {
- /* Bindless Textures - Kepler */
- int flat_slot = 0;
- if(string_startswith(mem.name, "__tex_image")) {
- int pos = string(mem.name).rfind("_");
- flat_slot = atoi(mem.name + pos + 1);
- }
- else {
- assert(0);
- }
+ generic_alloc(mem, dst_size - mem.memory_size());
+
+ CUDA_MEMCPY2D param;
+ memset(¶m, 0, sizeof(param));
+ param.dstMemoryType = CU_MEMORYTYPE_DEVICE;
+ param.dstDevice = mem.device_pointer;
+ param.dstPitch = dst_pitch;
+ param.srcMemoryType = CU_MEMORYTYPE_HOST;
+ param.srcHost = (void*)mem.data_pointer;
+ param.srcPitch = src_pitch;
+ param.WidthInBytes = param.srcPitch;
+ param.Height = mem.data_height;
- CUDA_RESOURCE_DESC resDesc;
- memset(&resDesc, 0, sizeof(resDesc));
+ cuda_assert(cuMemcpy2DUnaligned(¶m));
+ }
+ else {
+ /* 1D texture, using linear memory. */
+ generic_alloc(mem);
+ cuda_assert(cuMemcpyHtoD(mem.device_pointer, (void*)mem.data_pointer, size));
+ }
+
+ if(!has_fermi_limits) {
+ /* Kepler+, bindless textures. */
+ int flat_slot = 0;
+ if(string_startswith(mem.name, "__tex_image")) {
+ int pos = string(mem.name).rfind("_");
+ flat_slot = atoi(mem.name + pos + 1);
+ }
+ else {
+ assert(0);
+ }
+
+ CUDA_RESOURCE_DESC resDesc;
+ memset(&resDesc, 0, sizeof(resDesc));
+
+ if(mem.data_depth > 1) {
resDesc.resType = CU_RESOURCE_TYPE_ARRAY;
- resDesc.res.array.hArray = handle;
+ resDesc.res.array.hArray = array_3d;
resDesc.flags = 0;
+ }
+ else if(mem.data_height > 1) {
+ resDesc.resType = CU_RESOURCE_TYPE_PITCH2D;
+ resDesc.res.pitch2D.devPtr = mem.device_pointer;
+ resDesc.res.pitch2D.format = format;
+ resDesc.res.pitch2D.numChannels = mem.data_elements;
+ resDesc.res.pitch2D.height = mem.data_height;
+ resDesc.res.pitch2D.width = mem.data_width;
+ resDesc.res.pitch2D.pitchInBytes = dst_pitch;
+ }
+ else {
+ resDesc.resType = CU_RESOURCE_TYPE_LINEAR;
+ resDesc.res.linear.devPtr = mem.device_pointer;
+ resDesc.res.linear.format = format;
+ resDesc.res.linear.numChannels = mem.data_elements;
+ resDesc.res.linear.sizeInBytes = mem.device_size;
+ }
- CUDA_TEXTURE_DESC texDesc;
- memset(&texDesc, 0, sizeof(texDesc));
- texDesc.addressMode[0] = address_mode;
- texDesc.addressMode[1] = address_mode;
- texDesc.addressMode[2] = address_mode;
- texDesc.filterMode = filter_mode;
- texDesc.flags = CU_TRSF_NORMALIZED_COORDINATES;
+ CUDA_TEXTURE_DESC texDesc;
+ memset(&texDesc, 0, sizeof(texDesc));
+ texDesc.addressMode[0] = address_mode;
+ texDesc.addressMode[1] = address_mode;
+ texDesc.addressMode[2] = address_mode;
+ texDesc.filterMode = filter_mode;
+ texDesc.flags = CU_TRSF_NORMALIZED_COORDINATES;
- CUtexObject tex = 0;
- cuda_assert(cuTexObjectCreate(&tex, &resDesc, &texDesc, NULL));
+ CUtexObject tex = 0;
+ cuda_assert(cuTexObjectCreate(&tex, &resDesc, &texDesc, NULL));
- /* Safety check */
- if((uint)tex > UINT_MAX) {
- assert(0);
- }
+ /* Safety check */
+ if((uint)tex > UINT_MAX) {
+ assert(0);
+ }
- /* Resize once */
- if(flat_slot >= texture_info.size()) {
- /* Allocate some slots in advance, to reduce amount
- * of re-allocations. */
- texture_info.resize(flat_slot + 128);
- }
+ /* Resize once */
+ if(flat_slot >= texture_info.size()) {
+ /* Allocate some slots in advance, to reduce amount
+ * of re-allocations. */
+ texture_info.resize(flat_slot + 128);
+ }
- /* Set Mapping and tag that we need to (re-)upload to device */
- TextureInfo& info = texture_info[flat_slot];
- info.data = (uint64_t)tex;
- info.cl_buffer = 0;
- info.interpolation = mem.interpolation;
- info.extension = mem.extension;
- info.width = mem.data_width;
- in
@@ Diff output truncated at 10240 characters. @@
More information about the Bf-blender-cvs
mailing list