102 lines
4.1 KiB
Makefile
102 lines
4.1 KiB
Makefile
# SPDX-License-Identifier: AGPL-3.0-or-later
|
|
# Copyright (C) 2026 David (Corey) Herrera
|
|
# Foundry Blackhole backend — DRAM read capacity.
|
|
#
|
|
# ON CARD ONLY. This is the measurement that settles the per-bank contradiction between
|
|
# the cooperative-matmul round (~42 GB/s a bank) and the earlier DRAM round (>= 537 GB/s
|
|
# across 14 cores). See test_dram_capacity.c's header for the design.
|
|
#
|
|
# make -f Makefile.dramcap PHASE1B=1 all
|
|
#
|
|
# PHASE1B=1 is mandatory: without it every device call returns FT_TT_EPERM before touching
|
|
# the NoC, which is the whole point of the fence (spec §7.3 — a wrong coordinate hangs the
|
|
# chip).
|
|
#
|
|
# Run through run_dram_capacity_suite.sh, never directly: the last rule is that the cores
|
|
# it used must be proven to still BOOT *and* still COMPUTE, and the COMPUTE half is
|
|
# `ft_tt_core probe`, a different program.
|
|
#
|
|
# ---------------------------------------------------------------------------
|
|
# THE PROGRAM DIRECTORY
|
|
# ---------------------------------------------------------------------------
|
|
# Like the multicast bench, this is NOT a build_tensix_program.sh triad output: the bench
|
|
# needs one BRISC kernel and a compute no-op rather than a reader/compute/writer triad, so
|
|
# the per-class builders are driven directly into one directory. dram_bandwidth_reader.c
|
|
# takes NO compile-time args — everything about the call is ft_tensix_runtime_arg(0..12),
|
|
# deliberately, because one binary for every arm is the measurement's method. So `CTARGS=0`
|
|
# is a placeholder the builders require, not a contract.
|
|
#
|
|
# On the card host, ABSOLUTE paths (K=$PWD/../kernels, D=/tmp/ttk_dramcap):
|
|
#
|
|
# CTARGS=0 NOC=0 ../build_tensix_dataflow_kernel.sh \
|
|
# $K/dram_bandwidth_reader.c brisc $D
|
|
# CTARGS=0 ../build_tensix_kernel.sh $K/noop_compute.c $D
|
|
# CTARGS=16,2048 NOC=1 ../build_tensix_dataflow_kernel.sh \
|
|
# $K/vector_add_writer.c ncrisc $D
|
|
#
|
|
# All three into the SAME directory and in that order, because each builder writes the
|
|
# firmware it linked against and they must agree; the test loads the mask 0x1d (DM0 + all
|
|
# three TRISCs, NOT DM1), so the ncrisc image is built for completeness and never loaded.
|
|
#
|
|
# `make red` rebuilds from scratch. Apple ships GNU make 3.81, whose one-second mtime
|
|
# granularity silently skips a rebuild of a source edited in the same second as its object,
|
|
# so an incremental mutation loop can be running the PREVIOUS binary (DEV_NOTES.md §3).
|
|
|
|
CC ?= cc
|
|
SRC_DIR := ..
|
|
|
|
# The on-card harness opens the card through the product transport seam.
|
|
include hardware_session_sources.mk
|
|
BUILD := build-dramcap
|
|
|
|
CFLAGS := -std=c99 -O1 -g \
|
|
-Wall -Wextra -Werror -pedantic \
|
|
-Wshadow -Wconversion -Wstrict-prototypes -Wformat=2 \
|
|
-fstack-protector-strong \
|
|
-I$(SRC_DIR)
|
|
|
|
DEPFLAGS := -MMD -MP
|
|
|
|
ifdef PHASE1B
|
|
CFLAGS += -DFT_TT_PHASE1B_NOC_ENABLED
|
|
endif
|
|
|
|
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
|
|
|
|
SUPPORT := ft_tt_test_report.c $(HARDWARE_SESSION_SOURCES) ft_tt_program_directory.c \
|
|
ft_tt_phase2_l1_layout.c
|
|
|
|
.PHONY: all red clean
|
|
|
|
all: $(BUILD)/test_dram_capacity
|
|
|
|
$(BUILD):
|
|
mkdir -p $(BUILD)
|
|
|
|
# `sort` DEDUPLICATES: hardware_session_sources.mk gained ft_tt_udrv_claim.c, which BACKEND
|
|
# also names, and a whole-program link of one file twice is a multiple-definition error.
|
|
# A whole-program link, not incremental objects: this is built once and run against the
|
|
# card, and a stale object there is a wrong answer on silicon.
|
|
$(BUILD)/test_dram_capacity: test_dram_capacity.c $(BACKEND) $(SUPPORT) | $(BUILD)
|
|
$(CC) $(CFLAGS) $(DEPFLAGS) -o $@ $< $(sort $(BACKEND) $(SUPPORT))
|
|
|
|
red: clean all
|
|
|
|
clean:
|
|
rm -rf $(BUILD)
|
|
|
|
-include $(wildcard $(BUILD)/*.d)
|