You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
 
 
 
 

247 lines
8.2 KiB

/**********************************************************************
Copyright (c) 2019 Advanced Micro Devices, Inc. All rights reserved.
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in
all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
THE SOFTWARE.
********************************************************************/
// WebGPU has extremely strict uniformity requirements that are incompatible with the current implementation of this shader.
#pragma exclude_renderers webgpu
#define SORT_VALUES
uint g_constants_num_keys;
uint g_constants_num_blocks;
uint g_constants_bit_shift;
uint g_input_keys_offset;
uint g_group_histograms_offset;
uint g_output_keys_offset;
uint g_input_values_offset;
uint g_output_values_offset;
RWStructuredBuffer<uint> g_buffer : register(u0);
#define NUM_BITS_PER_PASS 4
#define NUM_BINS (1 << NUM_BITS_PER_PASS)
#define GROUP_SIZE 256u
#define KEYS_PER_THREAD 4u
//#define HISTOGRAM_TYPE short
#define INT_MAX 0x7fffffff
// Scratch memory for group operations
groupshared int lds_scratch[GROUP_SIZE];
// Temporary storage for the keys
groupshared int lds_keys[GROUP_SIZE];
// Cache for scanned histogram for faster indexing
groupshared int lds_scanned_histogram[NUM_BINS];
// Block histogram
groupshared int lds_histogram[NUM_BINS];
int BlockScan(int key, uint lidx)
{
#ifndef USE_WAVE_INTRINSICS
// Load the key into LDS
lds_keys[lidx] = key;
GroupMemoryBarrierWithGroupSync();
// Calculate reduction
uint stride = 0;
for (stride = 1; stride < GROUP_SIZE; stride <<= 1)
{
if (lidx < GROUP_SIZE / (2 * stride))
{
lds_keys[2 * (lidx + 1) * stride - 1] =
lds_keys[2 * (lidx + 1) * stride - 1] + lds_keys[(2 * lidx + 1) * stride - 1];
}
GroupMemoryBarrierWithGroupSync();
}
// Then put 0 into the root for downsweep
if (lidx == 0)
lds_keys[GROUP_SIZE - 1] = 0;
GroupMemoryBarrierWithGroupSync();
// Perform downsweep
for (stride = GROUP_SIZE >> 1; stride > 0; stride >>= 1)
{
if (lidx < GROUP_SIZE / (2 * stride))
{
int temp = lds_keys[(2 * lidx + 1) * stride - 1];
lds_keys[(2 * lidx + 1) * stride - 1] = lds_keys[2 * (lidx + 1) * stride - 1];
lds_keys[2 * (lidx + 1) * stride - 1] = lds_keys[2 * (lidx + 1) * stride - 1] + temp;
}
GroupMemoryBarrierWithGroupSync();
}
return lds_keys[lidx];
#else
lds_keys[lidx] = 0;
GroupMemoryBarrierWithGroupSync();
// Perform scan within a subgroup
int wave_scanned = WavePrefixSum(key);
uint widx = lidx / WaveGetLaneCount();
uint wlidx = WaveGetLaneIndex();
// Last element in each subgroup writes partial sum into LDS
if (wlidx == WaveGetLaneCount() - 1)
{
lds_keys[widx] = wave_scanned + key;
}
GroupMemoryBarrierWithGroupSync();
// Then first subgroup scannes partial sums
if (widx == 0)
{
lds_keys[lidx] = WavePrefixSum(lds_keys[lidx]);
}
GroupMemoryBarrierWithGroupSync();
// And we add partial sums back to each subgroup-scanned element
wave_scanned += lds_keys[widx];
return wave_scanned;
#endif
}
//[RootSignature(ROOT_SIGNATURE)]
#pragma kernel Scatter
[numthreads(GROUP_SIZE, 1, 1)]
void Scatter(
in uint gidx: SV_DispatchThreadID,
in uint lidx: SV_GroupThreadID,
in uint bidx: SV_GroupID)
{
// Cache scanned histogram in LDS and clear block histogram
if (lidx < NUM_BINS)
{
lds_scanned_histogram[lidx] = g_buffer[g_group_histograms_offset + g_constants_num_blocks * lidx + bidx];
}
// Starting point of our block in global memory
uint block_start_index = bidx * GROUP_SIZE * KEYS_PER_THREAD;
// Each thread handles PP_KEYS_PER_THREAD elements
for (uint i = 0; i < KEYS_PER_THREAD; ++i)
{
// Clear block histogram
if (lidx < NUM_BINS)
{
lds_histogram[lidx] = 0;
}
// Calculate next input element index
uint key_index = block_start_index + i * GROUP_SIZE + lidx;
// Fetch next element and put it in LDS
int key = (key_index < g_constants_num_keys) ? g_buffer[g_input_keys_offset + key_index] : INT_MAX;
#ifdef SORT_VALUES
int value = (key_index < g_constants_num_keys) ? g_buffer[g_input_values_offset + key_index] : 0;
#endif
// Sort keys locally in LDS
for (uint shift = 0; shift < NUM_BITS_PER_PASS; shift += 2)
{
// Detemine bin index for the key
int bin_index = ((key >> g_constants_bit_shift) >> shift) & 0x3;
// Create local packed histogram (0th in 1st byte, 1th in 2nd, etc)
int local_histogram = 1 << (bin_index * 8);
// Scan local histograms
int local_histogram_scanned = BlockScan(local_histogram, lidx);
// Last thread in a block broadcasts total block histogram
if (lidx == (GROUP_SIZE - 1))
{
lds_scratch[0] = local_histogram_scanned + local_histogram;
}
// Make sure broadcast happened
GroupMemoryBarrierWithGroupSync();
// Load broadcast value
local_histogram = lds_scratch[0];
// Scan block histogram in order to add scanned values as offsets
local_histogram = (local_histogram << 8) +
(local_histogram << 16) +
(local_histogram << 24);
// Add offsets from previos bins
local_histogram_scanned += local_histogram;
// Calculate target offset
int offset = (local_histogram_scanned >> (bin_index * 8)) & 0xff;
// Distribute the key
lds_keys[offset] = key;
//
GroupMemoryBarrierWithGroupSync();
// Load key
key = lds_keys[lidx];
#ifdef SORT_VALUES
// Perform value exchange
GroupMemoryBarrierWithGroupSync();
lds_keys[offset] = value;
GroupMemoryBarrierWithGroupSync();
value = lds_keys[lidx];
GroupMemoryBarrierWithGroupSync();
#endif
}
// Reconstruct original 16-bins histogram
int bin_index = (key >> g_constants_bit_shift) & 0xf;
InterlockedAdd(lds_histogram[bin_index], 1);
GroupMemoryBarrierWithGroupSync();
// Scan original histogram
int histogram_value = BlockScan(lidx < NUM_BINS ? lds_histogram[lidx] : 0, lidx);
// Broadcast scanned histogram via LDS
if (lidx < NUM_BINS)
{
lds_scratch[lidx] = histogram_value;
}
// Fetch scanned block histogram index
int global_offset = lds_scanned_histogram[bin_index];
GroupMemoryBarrierWithGroupSync();
// Fetch scanned histogram within a block
int local_offset = int(lidx) - lds_scratch[bin_index];
// Write the element back to global memory
if (global_offset + local_offset < int(g_constants_num_keys))
{
g_buffer[g_output_keys_offset + global_offset + local_offset] = key;
#ifdef SORT_VALUES
g_buffer[g_output_values_offset + global_offset + local_offset] = value;
#endif
}
GroupMemoryBarrierWithGroupSync();
// Update scanned histogram
if (lidx < NUM_BINS)
{
lds_scanned_histogram[lidx] += lds_histogram[lidx];
}
}
}