# SPDX-License-Identifier: AGPL-3.0-or-later
# Copyright (C) 2026 David (Corey) Herrera
# Foundry Blackhole backend — Phase 6 item 6: FLASH-DECODE ATTENTION.
#
# THREE BINARIES, TWO CONFIGURATIONS, AND THE FENCE BETWEEN THEM IS THE POINT.
#
#   make -f Makefile.attention all              -> build-attn/test_attention_ops
#     The OFF-HARDWARE gate. Built WITHOUT -DFT_TT_PHASE1B_NOC_ENABLED, so every device
#     call in the objects it links returns FT_TT_EPERM before touching the NoC. That is not
#     a formality: it means this binary cannot reach the card even if it is wrong, which is
#     what makes it safe to run while somebody else owns the silicon.
#
#   make -f Makefile.attention PHASE1B=1 card   -> build-attn-card/test_phase6_attention
#                                               -> build-attn-card/test_attention_transpose_probe
#     The ON-CARD acceptance, and the transpose probe that has to answer first.
#     PHASE1B=1 is mandatory and deliberate (spec §7.3 — a wrong NoC coordinate hangs the
#     chip and costs a physical power cycle).
#
# Program directories (RELOCATE_KERNEL_TEXT=1 builds; pass kernel sources by ABSOLUTE path,
# because the generated chlkc_*.cpp #includes the path it was given and resolves it
# relative to the OUTPUT directory):
#
#   probe — the transpose question, on the UNMODIFIED matmul reader and writer:
#     RELOCATE_KERNEL_TEXT=1 READER_CTARGS=0,1,2048,2048,1 COMPUTE_CTARGS=0,1,16,1 \
#       WRITER_CTARGS=16,2048,1 \
#       CB_TABLE=$PWD/kernels/cb_tables/matmul_float16b_fp32acc.cbtable \
#       build_tensix_program.sh $PWD/kernels/resident_matmul_reader.c \
#           $PWD/kernels/resident_attention_transpose_probe.c \
#           $PWD/kernels/resident_matmul_writer.c programs/probe
#
#   attention — the real triad:
#     RELOCATE_KERNEL_TEXT=1 READER_CTARGS=0,1,2,3,13,2048,1 \
#       COMPUTE_CTARGS=0,1,2,3,4,5,6,7,8,9,10,11,12,16,14,1 WRITER_CTARGS=16,2048,1 \
#       CB_TABLE=$PWD/kernels/cb_tables/attention_float16b.cbtable \
#       build_tensix_program.sh $PWD/kernels/resident_attention_reader.c \
#           $PWD/kernels/resident_attention_compute.c \
#           $PWD/kernels/resident_attention_writer.c programs/attention
#
# Both must be built against ONE firmware build: a kernel is linked against its own
# firmware's exported symbols, so mixing directories is a wedged core, not a configuration.

CC      ?= cc
SRC_DIR := ..

# The on-card harness opens the card through the product transport seam.
include hardware_session_sources.mk
CONVERT := ../convert
BUILD   := build-attn
CARD_BUILD := build-attn-card

WARNINGS := -Wall -Wextra -Werror -pedantic \
            -Wshadow -Wconversion -Wstrict-prototypes -Wformat=2 \
            -fstack-protector-strong

CFLAGS  := -std=c99 -O1 -g $(WARNINGS) -I$(SRC_DIR)
LDLIBS  := -lm

# `card` and `all` are separate invocations on purpose: the PHASE1B flag changes what the
# backend objects DO, so one command line cannot honestly produce both.
ifneq ($(filter card ornith,$(MAKECMDGOALS)),)
ifneq ($(filter all,$(MAKECMDGOALS)),)
$(error build `all` and `card` in separate invocations — they are different configurations)
endif
ifndef PHASE1B
$(error the on-card acceptance drives the card: build with PHASE1B=1, or every device call \
is refused before it reaches the NoC)
endif
CFLAGS += -DFT_TT_PHASE1B_NOC_ENABLED
endif

# Header dependency tracking. Without it, editing a HEADER rebuilds nothing and the link
# silently mixes objects compiled against two different struct layouts — not a build error,
# a wrong offset.
DEPFLAGS := -MMD -MP

BACKEND := $(SRC_DIR)/ft_tt_kmd.c \
           $(SRC_DIR)/ft_tt_arch.c \
           $(SRC_DIR)/ft_tt_arch_blackhole.c \
           $(SRC_DIR)/ft_tt_device.c \
           $(SRC_DIR)/ft_tt_tlb.c \
           $(SRC_DIR)/ft_tt_arc_telemetry.c \
           $(SRC_DIR)/ft_tt_coord.c \
           $(SRC_DIR)/ft_tt_noc.c \
           $(SRC_DIR)/ft_tt_buffer.c \
           $(SRC_DIR)/ft_tt_elf.c \
           $(SRC_DIR)/ft_tt_kernel_blob.c \
           $(SRC_DIR)/ft_tt_udrv_claim.c \
           $(SRC_DIR)/ft_tt_tensix_claim.c \
           $(SRC_DIR)/ft_tt_launch.c \
           $(SRC_DIR)/ft_tt_tilize.c \
           $(SRC_DIR)/ft_tt_op_dispatch.c \
           $(SRC_DIR)/ft_tt_op_attention.c \
           $(SRC_DIR)/ft_tt_op_memory.c \
           $(SRC_DIR)/ft_tt_queue.c \
           $(SRC_DIR)/ft_tt_group.c \
           $(SRC_DIR)/ft_tt_watchdog.c

# ft_float_convert is READ, not modified: the reference needs the same bf16 rounding the
# card will apply, and compiling it in rather than reimplementing it is what keeps the
# reference honest about rounding while staying independent about the OP.
CONVERT_SRC := $(CONVERT)/ft_float_convert.c

SUPPORT := ft_tt_test_report.c ft_tt_attention_reference.c
CARD_SUPPORT := $(HARDWARE_SESSION_SOURCES) ft_tt_program_directory.c

TARGET        := $(BUILD)/test_attention_ops
CARD_TARGET   := $(CARD_BUILD)/test_phase6_attention
PROBE_TARGET  := $(CARD_BUILD)/test_attention_transpose_probe
ORNITH_TARGET := $(CARD_BUILD)/test_phase6_attention_ornith

# ORNITH'S SHAPE, OUT OF THE SAME SOURCE. 16 Q heads / 4 KV groups / head_dim 256 over 8
# attention layers is q_rows 4 and dim_tiles 8 — a different RESIDENCY, not a different
# descriptor, so it is a second BINARY rather than an environment variable. MAX_KV=32768 is
# what makes the SLOPE measurable where this phase's value actually is; it costs 32 MiB of
# static K and V on the host and 16 MiB of card DRAM a region.
#
#   make -f Makefile.attention PHASE1B=1 ornith
#   FT_TT_FIRMWARE_DIR=../programs/attention FT_TT_TENSIX_INDEX=84 \
#     FT_TT_ATTN_BLOCK_TILES=16 FT_TT_ATTN_CONTEXTS=1024,2048,4096,8192,16384,32768 \
#     ./build-attn-card/test_phase6_attention_ornith ../programs/attention
ORNITH_SHAPE := -DMODEL_Q_HEADS=16u -DMODEL_KV_HEADS=4u -DMODEL_HEAD_DIM=256u \
                -DMODEL_LAYERS=8u -DMAX_KV=32768u

.PHONY: all card ornith red clean

all: $(TARGET)

card: $(CARD_TARGET) $(PROBE_TARGET)

ornith: $(ORNITH_TARGET)

$(BUILD) $(CARD_BUILD):
	mkdir -p $@

# Whole-program links, not incremental objects. `red` deletes the tree first, because
# Apple's GNU make 3.81 compares mtimes at one-second granularity and will silently skip a
# rebuild of a source edited in the same second as its object (best-practices §9).
$(TARGET): test_attention_ops.c $(BACKEND) $(CONVERT_SRC) $(SUPPORT) | $(BUILD)
	$(CC) $(CFLAGS) $(DEPFLAGS) -o $@ $< $(BACKEND) $(CONVERT_SRC) $(SUPPORT) $(LDLIBS)

$(CARD_TARGET): test_phase6_attention.c $(BACKEND) $(CONVERT_SRC) $(SUPPORT) $(CARD_SUPPORT) \
                | $(CARD_BUILD)
	$(CC) $(CFLAGS) $(DEPFLAGS) -o $@ $< $(BACKEND) $(CONVERT_SRC) $(SUPPORT) $(CARD_SUPPORT) \
	    $(LDLIBS)

$(ORNITH_TARGET): test_phase6_attention.c $(BACKEND) $(CONVERT_SRC) $(SUPPORT) $(CARD_SUPPORT) \
                  | $(CARD_BUILD)
	$(CC) $(CFLAGS) $(ORNITH_SHAPE) $(DEPFLAGS) -o $@ $< $(BACKEND) $(CONVERT_SRC) \
	    $(SUPPORT) $(CARD_SUPPORT) $(LDLIBS)

$(PROBE_TARGET): test_attention_transpose_probe.c $(BACKEND) $(CONVERT_SRC) ft_tt_test_report.c \
                 $(CARD_SUPPORT) | $(CARD_BUILD)
	$(CC) $(CFLAGS) $(DEPFLAGS) -o $@ $< $(BACKEND) $(CONVERT_SRC) ft_tt_test_report.c \
	    $(CARD_SUPPORT) $(LDLIBS)

red: clean all

clean:
	rm -rf $(BUILD) $(CARD_BUILD)

-include $(wildcard $(BUILD)/*.d) $(wildcard $(CARD_BUILD)/*.d)