# Copyright (C) 2026 Free Software Foundation, Inc.
# Copyright (C) 2026 Advanced Micro Devices, Inc. All rights reserved.

# This file is part of GDB.

# This program is free software; you can redistribute it and/or modify
# it under the terms of the GNU General Public License as published by
# the Free Software Foundation; either version 3 of the License, or
# (at your option) any later version.
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
# GNU General Public License for more details.
#
# You should have received a copy of the GNU General Public License
# along with this program.  If not, see <http://www.gnu.org/licenses/>.

# Tests for debugging a single-device cooperative kernel that uses
# cooperative_groups::this_grid ().sync (), implemented on AMD GPUs
# using the Global Wave Sync (GWS) hardware mechanism.
#
# The test verifies that GDB can:
#   - hit a breakpoint placed after grid.sync () in a cooperatively
#     launched kernel (proving GWS-protected code is debuggable across
#     the barrier),
#   - list the cooperative dispatch via "info dispatches",
#   - observe waves from multiple workgroups,
#   - inspect correct per-wave and per-lane state, and
#   - resume the kernel to normal completion.

load_lib rocm.exp

require allow_hip_tests
require supports_cooperative_groups

standard_testfile .cpp

if {[build_executable "failed to prepare" $testfile $srcfile {debug hip}]} {
    return
}

# Test debugging a single-device cooperative kernel.  Stop at a
# breakpoint after grid.sync () -- which fires once all waves have
# crossed the GWS barrier, proving GWS-protected code can be debugged
# across the barrier -- and, with the co-resident waves parked there,
# check the dispatch and GDB's per-wave / per-lane state reporting, all
# in a single debug session.

proc_with_prefix test_coop_grid_sync {} {
    clean_restart $::testfile

    with_rocm_gpu_lock {
	if {![runto_main]} {
	    return
	}

	# Place a breakpoint after grid.sync ().  Use "allow-pending"
	# because the kernel is loaded as device code at dispatch time.
	gdb_breakpoint \
	    [gdb_get_line_number "after-sync line"] allow-pending temporary

	# Continue into the kernel.  If the program self-skips (no device
	# advertises cooperativeLaunch) it exits cleanly; treat that as
	# UNSUPPORTED rather than a FAIL.
	set reached 0
	gdb_test_multiple "continue" "stop after grid.sync" {
	    -re -wrap "Temporary breakpoint $::decimal,\[^\r\n\]*coop_grid_sync_kernel .*" {
		set reached 1
		pass $gdb_test_name
	    }
	    -re -wrap "$::inferior_exited_re normally.*" {
		unsupported "cooperative launch not available"
	    }
	}
	if { !$reached } {
	    return
	}

	# Verify that waves from multiple workgroups are stopped.
	# Counting waves alone is wave-size dependent (1 wave
	# per workgroup on wave64 vs 2 on wave32) and would let the test
	# pass on wave32 even if all visible waves came from a single
	# workgroup.  Instead, collect the distinct workgroup (block)
	# coordinates from the AMDGPU Wave entries in "info threads" and
	# require at least two distinct workgroups.
	set blocks {}
	gdb_test_multiple "info threads" "blocks present" -lbl {
	    -re "AMDGPU Wave \[^(\r\n\]*\\((\[^()\r\n\]+)\\)" {
		if {[lsearch -exact $blocks $expect_out(1,string)] == -1} {
		    lappend blocks $expect_out(1,string)
		}
		exp_continue
	    }
	    -re -wrap "" {
		verbose -log "blocks: $blocks"
		gdb_assert {[llength $blocks] >= 2} \
		    "waves from at least 2 distinct workgroups are present"
	    }
	}

	# "info dispatches" should show the cooperative dispatch.
	gdb_test "info dispatches" \
	    ".*AMDGPU Dispatch\[^\r\n\]*coop_grid_sync_kernel.*" \
	    "info dispatches lists cooperative dispatch"

	# Check that GDB reports correct, distinct per-wave state: switching
	# to each wave and reading blockIdx.x must yield that wave's owning
	# workgroup, and across all waves we must observe more than one
	# distinct workgroup.
	set waves [info_thread_get_wave_list]
	gdb_assert {[llength $waves] >= 2} "multiple GPU waves listed"

	set wave_blocks {}
	foreach wave $waves {
	    with_test_prefix "wave $wave" {
		gdb_test "thread $wave" "Switching to thread $wave.*" \
		    "switch to wave"
		set bidx [get_integer_valueof "blockIdx.x" -1 "blockIdx.x"]
		if {[lsearch -exact $wave_blocks $bidx] == -1} {
		    lappend wave_blocks $bidx
		}
	    }
	}
	gdb_assert {[llength $wave_blocks] >= 2} \
	    "distinct workgroups observed across waves"

	# Per-lane (SIMT) state: within a single wave, different lanes map
	# to different threads, so threadIdx.x must differ between lanes.
	gdb_test "thread [lindex $waves 0]" "Switching to thread .*" \
	    "switch to first wave for lane inspection"
	gdb_test "lane 0" "\\\[Switching to thread $::decimal, lane 0 .*" \
	    "switch to lane 0"
	set tid0 [get_integer_valueof "threadIdx.x" -1 "threadIdx.x at lane 0"]
	gdb_test "lane 1" "\\\[Switching to thread $::decimal, lane 1 .*" \
	    "switch to lane 1"
	set tid1 [get_integer_valueof "threadIdx.x" -1 "threadIdx.x at lane 1"]
	gdb_assert {$tid0 != $tid1} "distinct threadIdx.x across lanes"

	# The kernel should run to completion afterwards.
	gdb_continue_to_end "continue to program end" "continue" 1
    }
}

test_coop_grid_sync