From 6c0d7c1f6cf344ada3ef366e8ffea8da695373f5 Mon Sep 17 00:00:00 2001
From: Hadrien Grasland <hadrien.grasland@ijclab.in2p3.fr>
Date: Wed, 24 Jul 2024 18:33:54 +0200
Subject: [PATCH] Allow work-group reuse after completing a tile
---
exercises/src/gpu/grayscott.comp | 49 ++++++++++++++++++++++----------
exercises/src/gpu/interface.comp | 10 ++++++-
exercises/src/gpu/mod.rs | 8 ++----
exercises/src/gpu/pipeline.rs | 2 ++
exercises/src/gpu/shared.comp | 8 +++++-
exercises/src/gpu/switch.comp | 7 +++--
exercises/src/gpu/tile_map.comp | 48 +++++++++++++++++++++++++++++++
exercises/src/gpu/wait.comp | 2 +-
8 files changed, 109 insertions(+), 25 deletions(-)
diff --git a/exercises/src/gpu/grayscott.comp b/exercises/src/gpu/grayscott.comp
index da65477..53245c4 100644
--- a/exercises/src/gpu/grayscott.comp
+++ b/exercises/src/gpu/grayscott.comp
@@ -14,6 +14,7 @@
#include "wait.comp"
#include "switch.comp"
+
// Initialize the global debug/profiling state
#if DEBUG || PROFILE
void init_global() {
@@ -86,6 +87,12 @@
}
#endif
+// Entry point that sets up global state, preparing for a simulation run
+void init_simulation() {
+ global_init_remaining_tiles();
+ clear_tile_map();
+}
+
// Truth that this work-group must switch to a different tile
shared bool s_must_switch;
@@ -190,13 +197,19 @@ void init_shared() {
}
}
-// Main entry point of the Gray-Scott reaction simulation
+// Entry point that executes the simulation
void run_simulation() {
- // Initialize global, shared and local state
+ // Early exit if the work-group is not needed
#if DEBUG || PROFILE
- init_global();
if(is_leader()) atomicAdd(g_metadata.entry, 1);
#endif
+ if (work_group_finished()) return;
+
+ // Set up global, shared and local state
+ #if DEBUG || PROFILE
+ init_global();
+ if(is_leader()) atomicAdd(g_metadata.has_work, 1);
+ #endif
init_shared();
vec2 uv_cache;
barrier();
@@ -211,7 +224,7 @@ void run_simulation() {
// simulation step (same or greater input_buffer_idx).
// - uv_cache contains the UV data from step `s_input_buffer_idx` or is
// marked invalid by setting `s_uv_cache_valid` to false.
- while(g_metadata.error == ERROR_NONE) {
+ while((work_group_finished() == false) && (g_metadata.error == ERROR_NONE)) {
#if DEBUG || PROFILE
if (is_leader()) {
s_iteration += 1;
@@ -224,8 +237,8 @@ void run_simulation() {
// Switch to a different tile if we have to
//
- // This should always succeed, failure to do so is a fatal error that
- // will lead the entire dispatch to abord ASAP.
+ // This almost always succeeds, if it means it should mean that the
+ // work-group has no hope of ever finding work again and should exit.
//
// This will clear s_acquired_tile, and thus execute the next block too.
if (s_must_switch && !try_switch_tile()) break;
@@ -282,11 +295,9 @@ void run_simulation() {
// Terminate work-group once it finishes a tile
if (s_input_buffer_idx == LAST_BUFFER_IDX) {
- #if DEBUG || PROFILE
- if (is_leader()) atomicAdd(g_metadata.finish_final_step, 1);
- barrier();
- #endif
- break;
+ if (is_leader()) leader_notify_tile_finished();
+ barrier();
+ continue;
}
#if DEBUG
if (is_leader()) s_checkpoint = 5;
@@ -306,6 +317,14 @@ void run_simulation() {
#endif
}
+ // Liberate any acquired tile
+ if (s_acquired_tile && is_leader()) {
+ leader_release_tile();
+ #if DEBUG || PROFILE
+ atomicAdd(g_metadata.final_release, 1);
+ #endif
+ }
+
// Aggregate local execution statistics into global execution statistics
#if PROFILE
#define MERGE_HIST(histogram) \
@@ -348,13 +367,13 @@ void run_simulation() {
#endif
}
-// Main entry point that dispatches to the tile status clear or simulation
+// Main entry point that dispatches to simulation initialization or execution
void main() {
switch (g_push.mode) {
- case 0:
- clear_tile_map();
+ case MODE_INIT:
+ init_simulation();
break;
- case 1:
+ case MODE_RUN:
run_simulation();
break;
default:
diff --git a/exercises/src/gpu/interface.comp b/exercises/src/gpu/interface.comp
index 8383c35..15bc105 100644
--- a/exercises/src/gpu/interface.comp
+++ b/exercises/src/gpu/interface.comp
@@ -12,8 +12,10 @@ const uint NUM_BUFFERS = MAX_STEPS + 1;
const uint LAST_BUFFER_IDX = NUM_BUFFERS - 1;
layout(push_constant) uniform PushConstants {
uint num_steps; // Number of computation steps to be performed
- uint mode; // 0 for clearing the tile status, 1 for running the simulation
+ uint mode; // See MODE_ constants below
} g_push;
+const uint MODE_INIT = 0;
+const uint MODE_RUN = 1;
// Last input buffer index
uint last_input_buffer_idx() {
@@ -76,6 +78,9 @@ layout(set = 1, binding = 0) buffer restrict coherent Metadata {
uvec2 work_group_size;
#endif
+ // Number of simulation tiles that are not fully completed yet
+ uint remaining_tiles;
+
// Detailed error context reporting in debug builds
#if DEBUG
// Work-group and work-item that encountered the first error
@@ -111,6 +116,7 @@ layout(set = 1, binding = 0) buffer restrict coherent Metadata {
// Dispatch-wide work-group control flow tracking
#if DEBUG || PROFILE
uint entry;
+ uint has_work;
uint iteration;
uint try_switch_tile; // See also hist_search_distance and hist_switch_distance
uint try_switch_tile_release;
@@ -128,6 +134,7 @@ layout(set = 1, binding = 0) buffer restrict coherent Metadata {
uint try_await_neighbors;
uint try_await_neighbors_failure;
uint try_await_neighbors_success; // See also hist_log2_polls_to_success
+ uint final_release;
#endif
// Optional detailed performance monitoring
@@ -187,6 +194,7 @@ const uint ERROR_3D_WORK_GROUP = 20;
const uint ERROR_INVALID_UV_LEN = 21;
const uint ERROR_INVALID_TILE_STATUS_LEN = 22;
const uint ERROR_INVALID_MODE = 23;
+const uint ERROR_INVALID_REMAINING_TILES = 24;
// Set the metadata error
//
diff --git a/exercises/src/gpu/mod.rs b/exercises/src/gpu/mod.rs
index f61e4d3..f91da4a 100644
--- a/exercises/src/gpu/mod.rs
+++ b/exercises/src/gpu/mod.rs
@@ -2,7 +2,7 @@
use self::{
context::{CommandBufferAllocator, VulkanContext},
- pipeline::{create_pipeline, IMAGES_SET, MODE, NUM_STEPS, STATIC_SET},
+ pipeline::{create_pipeline, IMAGES_SET, MODE, MODE_INIT, MODE_RUN, NUM_STEPS, STATIC_SET},
resources::{Concentrations, StaticComputeState},
};
use crate::{
@@ -70,11 +70,9 @@ fn add_update(
IMAGES_SET,
concentrations,
)?
- // FIXME: Investigate ways to merge clear into main shader
- .push_constants(pipeline_layout.clone(), MODE, 0)?
+ .push_constants(pipeline_layout.clone(), MODE, MODE_INIT)?
.dispatch(static_state.tile_status_clear_dispatch())?
- .push_constants(pipeline_layout.clone(), MODE, 1)?
- //
+ .push_constants(pipeline_layout.clone(), MODE, MODE_RUN)?
.dispatch(dispatch_size)?;
Ok(())
}
diff --git a/exercises/src/gpu/pipeline.rs b/exercises/src/gpu/pipeline.rs
index c3864ac..ffb4cbd 100644
--- a/exercises/src/gpu/pipeline.rs
+++ b/exercises/src/gpu/pipeline.rs
@@ -85,6 +85,8 @@ pub const NUM_STEPS: u32 = 0;
/// Integer push constant used to indicate if we should set up the simulation
/// state by clearing the tile status (0) or run the simulation (1)
pub const MODE: u32 = 4;
+pub const MODE_INIT: u32 = 0;
+pub const MODE_RUN: u32 = 1;
/// Create the compute pipeline
pub fn create_pipeline(
diff --git a/exercises/src/gpu/shared.comp b/exercises/src/gpu/shared.comp
index f91b27d..0695ba6 100644
--- a/exercises/src/gpu/shared.comp
+++ b/exercises/src/gpu/shared.comp
@@ -7,7 +7,13 @@
//!
//! After this, the work-group will attempt to move the simulation to the next
//! step again and again, hopefully until the last simulation step is reached.
-//! At this point, the work-group would stop executing.
+//!
+//! At this point, the work-group will take note that there is one less tile to
+//! be processed, which will cause one work-group to exit. We priorize killing
+//! workgroups with higher IDs, as these tend to be auxilliary work-groups that
+//! are spawned by the GPU but do not do anything until other workgroups have
+//! exited. These workgroups are not useful to us, as we are trying to make the
+//! most of each workgroup here.
//!
//! If the work-group manages to process several simulation steps in a row,
//! work-items will be able to go faster by keeping the previous (U, V) data
diff --git a/exercises/src/gpu/switch.comp b/exercises/src/gpu/switch.comp
index 1290f52..2162076 100644
--- a/exercises/src/gpu/switch.comp
+++ b/exercises/src/gpu/switch.comp
@@ -425,6 +425,9 @@ bool find_tile(in ivec2 center_block_idx) {
//
// This will call leader_release_tile() for you if needed, so at the end of this
// function it is guaranteed that `s_acquired_tile` is false.
+//
+// This function will almost always return true. If it returns false, it means
+// that this work-group cannot ever expect to find work again and should exit.
bool try_switch_tile() {
#if DEBUG || PROFILE
if (is_leader()) s_try_switch_tile += 1;
@@ -463,8 +466,8 @@ bool try_switch_tile() {
uint retries;
for (retries = 0; retries < MAX_SWITCH_ATTEMPTS; ++retries) {
if (find_tile(center_block_idx)) break;
- // Maximize odds of a tile_status cache flush => Up-to-date view
- memoryBarrierBuffer();
+ if (work_group_finished()) return false;
+ memoryBarrierBuffer(); // Maximize odds of fresh find_tile reads
}
// Handle search success or failure
diff --git a/exercises/src/gpu/tile_map.comp b/exercises/src/gpu/tile_map.comp
index 6279696..3f7b205 100644
--- a/exercises/src/gpu/tile_map.comp
+++ b/exercises/src/gpu/tile_map.comp
@@ -18,6 +18,7 @@
// Number of simulation tiles (and thus dispatch size)
const uvec2 OUTPUT_SIZE_IN_TILES = DIV_CEIL_UVEC2(OUTPUT_SIZE, LOCAL_OUTPUT_SIZE);
+const uint NUM_ACTIVE_TILES = OUTPUT_SIZE_IN_TILES.x * OUTPUT_SIZE_IN_TILES.y;
// The g_tile_status map is actually padded such that the number of tiles is a
// multiple of the local output size, and there is an extra strip of tile status
@@ -27,6 +28,53 @@ const uvec2 OUTPUT_TILES = NEXT_MULTIPLE_UVEC2(OUTPUT_SIZE_IN_TILES,
const uvec2 INPUT_TILES = OUTPUT_TILES + uvec2(2);
+// === Tracking of completed tiles ===
+
+// Set up tracking of fully processed tiles
+//
+// Must be done in a separate initialization dispatch before the main simulation
+// dispatch so that every work-group sees it.
+void global_init_remaining_tiles() {
+ if (gl_GlobalInvocationID == uvec3(0)) {
+ g_metadata.remaining_tiles = NUM_ACTIVE_TILES;
+ }
+}
+
+// (Leader-only) Notify other work-groups that one more tile is fully processed
+void leader_notify_tile_finished() {
+ #if DEBUG
+ if (!is_leader()) {
+ set_error(ERROR_NOT_LEADER);
+ return;
+ }
+ #endif
+ const uint prev_remaining = atomicAdd(g_metadata.remaining_tiles,
+ 0xffffffff /* -1 */);
+ #if DEBUG
+ if ((prev_remaining == 0) || (prev_remaining > NUM_ACTIVE_TILES)) {
+ if (set_error(ERROR_INVALID_REMAINING_TILES)) {
+ g_metadata.invalid_uint = prev_remaining;
+ }
+ }
+ #endif
+ #if DEBUG || PROFILE
+ atomicAdd(g_metadata.finish_final_step, 1);
+ #endif
+}
+
+// Truth that this work-group is not needed anymore
+//
+// Each time a simulation tile is fully processed, one less work-group is
+// necessary to process remaining tiles. We eliminate work-groups with larger
+// dispatch indices first because these are more likely to be extraneous
+// work-groups that were spawned by the GPU but will not run until the main
+// work-groups have completed their work.
+bool work_group_finished() {
+ const uint work_group_idx = gl_WorkGroupID.x + gl_WorkGroupID.y * gl_WorkGroupSize.x;
+ return (work_group_idx >= g_metadata.remaining_tiles);
+}
+
+
// === Basic reads and writes ===
// Check that a tile index is correct
diff --git a/exercises/src/gpu/wait.comp b/exercises/src/gpu/wait.comp
index 1a5e599..9e62d02 100644
--- a/exercises/src/gpu/wait.comp
+++ b/exercises/src/gpu/wait.comp
@@ -133,7 +133,7 @@ bool try_wait_for_neighbors() {
// ...wait a bit more...
for (num_polls = 2; num_polls <= s_max_polls; ++num_polls) {
- memoryBarrierBuffer();
+ memoryBarrierBuffer(); // Maximize odds of fresh tile_status reads
neighbor_status = read_tile_status(neighbor_idx);
decode_tile_status(neighbor_status,
neighbor_input_idx,
--
GitLab