re-add lib files for streamer-gemm and streamer-gemm-add-c (#281)

target/snitch_cluster/sw/snax/streamer-gemm-add-c/Makefile

@@ -0,0 +1,78 @@
+# Copyright 2023 KU Leuven.
+# Licensed under the Apache License, Version 2.0, see LICENSE for details.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Xiaoling Yi <[email protected]>
+
+# Usage of absolute paths is required to externally include
+# this Makefile from multiple different locations
+
+MK_DIR := $(dir $(realpath $(lastword $(MAKEFILE_LIST))))
+include $(MK_DIR)/../../toolchain.mk
+
+# Banshee runtime is not supported
+ifeq ($(SELECT_RUNTIME), rtl-generic)
+RUNTIME_DIR := rtl-generic
+else
+RUNTIME_DIR := rtl
+endif
+
+################
+## Directories #
+################
+
+# Fixed paths in repository tree
+ROOT     = $(abspath $(MK_DIR)/../../../../..)
+SNRT_DIR = $(ROOT)/sw/snRuntime
+## Paths relative to the runtime including this Makefile
+BUILDDIR = $(abspath build)
+SRC_DIR  = $(abspath src)
+
+####################
+## Build variables #
+####################
+
+INCDIRS += $(abspath include)
+INCDIRS += $(SNRT_DIR)/src
+INCDIRS += $(SNRT_DIR)/api
+INCDIRS += $(SNRT_DIR)/src/omp
+INCDIRS += $(SNRT_DIR)/api/omp
+INCDIRS += $(SNRT_DIR)/vendor/riscv-opcodes
+INCDIRS += $(ROOT)/target/snitch_cluster/sw/runtime/common
+
+# math.h needed by snRuntime
+
+INCDIRS += $(SNRT_DIR)/../math/arch/riscv64/
+INCDIRS += $(SNRT_DIR)/../math/arch/generic
+INCDIRS += $(SNRT_DIR)/../math/src/include
+INCDIRS += $(SNRT_DIR)/../math/src/internal
+INCDIRS += $(SNRT_DIR)/../math/include/bits
+INCDIRS += $(SNRT_DIR)/../math/include
+
+INCDIRS += $(ROOT)/target/snitch_cluster/sw/runtime/$(RUNTIME_DIR)/src
+INCDIRS += $(ROOT)/target/snitch_cluster/sw/runtime/$(RUNTIME_DIR)/include
+
+############
+## Outputs #
+############
+
+OBJS        = $(BUILDDIR)/snax-streamer-gemm-add-c-lib.o
+ALL_OUTPUTS = $(OBJS)
+
+
+##########
+## Rules #
+##########
+
+.PHONY: all
+all: $(ALL_OUTPUTS)
+
+.PHONY: clean
+clean:
+	rm -rf $(BUILDDIR)
+
+$(BUILDDIR):
+	mkdir -p $@
+
+$(BUILDDIR)/%.o: $(SRC_DIR)/%.c | $(BUILDDIR)
+	$(RISCV_CC) $(RISCV_CFLAGS) -c $< -o $@

target/snitch_cluster/sw/snax/streamer-gemm-add-c/include/snax-streamer-gemm-add-c-lib.h

@@ -0,0 +1,38 @@
+// Copyright 2023 KU Leuven.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+//
+// Xiaoling Yi <[email protected]>
+
+#include <stdbool.h>
+#include "snrt.h"
+#include "stdint.h"
+
+#pragma once
+
+// Set STREAMER configuration CSR
+void set_streamer_csr(int tempLoop0, int tempLoop1, int tempLoop2,
+                      int tempStride0A, int tempStride2A, int spatialStride1A,
+                      int tempStride0B, int tempStride1B, int spatialStride1B,
+                      int tempStride1C, int tempStride2C, int spatialStride1C,
+                      int delta_local_a, int delta_local_b, int delta_local_c,
+                      int delta_local_d);
+
+// Set CSR to start STREAMER
+void set_streamer_start();
+
+// Set GEMM configuration CSR
+void set_block_gemm_csr(int tempLoop0, int tempLoop1, int tempLoop2,
+                        int subtractions);
+
+// Set CSR to start GEMM
+void set_block_gemm_start();
+
+// Poll until Streamer and GEMM accelerator finish
+void wait_streamer_gemm();
+
+// Read perforamcne counter of the Streamer, a read-only CSR
+uint32_t read_gemm_streamer_perf_counter();
+
+// Read perforamcne counter of GEMM, a read-only CSR
+uint32_t read_gemm_perf_counter();

target/snitch_cluster/sw/snax/streamer-gemm-add-c/src/snax-streamer-gemm-add-c-lib.c

@@ -0,0 +1,104 @@
+// Copyright 2023 KU Leuven.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+//
+// Xiaoling Yi <[email protected]>
+
+#include "snax-streamer-gemm-add-c-lib.h"
+
+// Set STREAMER configuration CSR
+void set_streamer_csr(int tempLoop0, int tempLoop1, int tempLoop2,
+                      int tempStride0A, int tempStride2A, int spatialStride1A,
+                      int tempStride0B, int tempStride1B, int spatialStride1B,
+                      int tempStride1C, int tempStride2C, int spatialStride1C,
+                      int delta_local_a, int delta_local_b, int delta_local_c,
+                      int delta_local_d) {
+    // loop bounds, from innermost to outermost, from K to N to M
+    write_csr(960, tempLoop0);
+    write_csr(961, tempLoop1);
+    write_csr(962, tempLoop2);
+
+    // temporal strides for A
+    write_csr(963, tempStride0A);
+    write_csr(964, 0);
+    write_csr(965, tempStride2A);
+
+    // temporal strides for B
+    write_csr(966, tempStride0B);
+    write_csr(967, tempStride1B);
+    write_csr(968, 0);
+
+    // temporal strides for C
+    write_csr(969, 0);
+    write_csr(970, tempStride1C);
+    write_csr(971, tempStride2C);
+
+    // temporal strides for D
+    write_csr(972, 0);
+    write_csr(973, tempStride1C);
+    write_csr(974, tempStride2C);
+
+    // spatial strides for A
+    write_csr(975, 1);
+    write_csr(976, spatialStride1A);
+
+    // spatial strides for B
+    write_csr(977, 1);
+    write_csr(978, spatialStride1B);
+
+    // spatial strides for C
+    write_csr(979, 4);
+    write_csr(980, spatialStride1C);
+
+    // spatial strides for D
+    write_csr(981, 4);
+    write_csr(982, spatialStride1C);
+
+    // base ptr for A
+    write_csr(983, (uint32_t)(delta_local_a + snrt_l1_next()));
+
+    // base ptr for B
+    write_csr(984, (uint32_t)(delta_local_b + snrt_l1_next()));
+
+    // base ptr for C
+    write_csr(985, (uint32_t)(delta_local_c + snrt_l1_next()));
+
+    // base ptr for D
+    write_csr(986, (uint32_t)(delta_local_d + snrt_l1_next()));
+}
+
+// Set CSR to start STREAMER
+void set_streamer_start() { write_csr(987, 1); }
+
+// Set GEMM configuration CSR
+void set_block_gemm_csr(int tempLoop0, int tempLoop1, int tempLoop2,
+                        int subtractions) {
+    // set loop bounds, from innermost to outermost, aka from K to N to M
+    write_csr(990, tempLoop0);
+    write_csr(991, tempLoop1);
+    write_csr(992, tempLoop2);
+
+    // set subtraction a and b
+    write_csr(993, subtractions);
+}
+
+// Set CSR to start GEMM
+void set_block_gemm_start() { write_csr(994, 1); }
+
+// Poll until Streamer and GEMM accelerator finish
+void wait_streamer_gemm() {
+    write_csr(994, 0);
+    write_csr(988, 0);
+}
+
+// Read performance counter of the Streamer, a read-only CSR
+uint32_t read_gemm_streamer_perf_counter() {
+    uint32_t perf_counter = read_csr(989);
+    return perf_counter;
+}
+
+// Read performance counter of GEMM, a read-only CSR
+uint32_t read_gemm_perf_counter() {
+    uint32_t perf_counter = read_csr(996);
+    return perf_counter;
+}

target/snitch_cluster/sw/snax/streamer-gemm/Makefile

@@ -0,0 +1,85 @@
+# Copyright 2023 KU Leuven.
+# Licensed under the Apache License, Version 2.0, see LICENSE for details.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Xiaoling Yi <[email protected]>
+
+# Usage of absolute paths is required to externally include
+# this Makefile from multiple different locations
+
+MK_DIR := $(dir $(realpath $(lastword $(MAKEFILE_LIST))))
+include $(MK_DIR)/../../toolchain.mk
+
+# Banshee runtime is not supported
+ifeq ($(SELECT_RUNTIME), rtl-generic)
+RUNTIME_DIR := rtl-generic
+else
+RUNTIME_DIR := rtl
+endif
+
+################
+## Directories #
+################
+
+# Fixed paths in repository tree
+ROOT     = $(abspath $(MK_DIR)/../../../../..)
+SNRT_DIR = $(ROOT)/sw/snRuntime
+## Paths relative to the runtime including this Makefile
+BUILDDIR = $(abspath build)
+SRC_DIR  = $(abspath src)
+
+####################
+## Build variables #
+####################
+
+INCDIRS += $(abspath include)
+INCDIRS += $(SNRT_DIR)/src
+INCDIRS += $(SNRT_DIR)/api
+INCDIRS += $(SNRT_DIR)/src/omp
+INCDIRS += $(SNRT_DIR)/api/omp
+INCDIRS += $(SNRT_DIR)/vendor/riscv-opcodes
+INCDIRS += $(ROOT)/target/snitch_cluster/sw/runtime/common
+
+# math.h needed by snRuntime
+
+INCDIRS += $(SNRT_DIR)/../math/arch/riscv64/
+INCDIRS += $(SNRT_DIR)/../math/arch/generic
+INCDIRS += $(SNRT_DIR)/../math/src/include
+INCDIRS += $(SNRT_DIR)/../math/src/internal
+INCDIRS += $(SNRT_DIR)/../math/include/bits
+INCDIRS += $(SNRT_DIR)/../math/include
+
+INCDIRS += $(ROOT)/target/snitch_cluster/sw/runtime/$(RUNTIME_DIR)/src
+INCDIRS += $(ROOT)/target/snitch_cluster/sw/runtime/$(RUNTIME_DIR)/include
+
+############
+## Outputs #
+############
+
+OBJS        = $(BUILDDIR)/snax-streamer-gemm-lib.o
+ALL_OUTPUTS = $(OBJS)
+
+
+##########
+## Rules #
+##########
+
+.PHONY: all
+all: $(ALL_OUTPUTS)
+
+.PHONY: clean
+clean:
+	rm -rf $(BUILDDIR)
+
+$(BUILDDIR):
+	mkdir -p $@
+
+$(BUILDDIR)/%.o: $(SRC_DIR)/%.c | $(BUILDDIR)
+	$(RISCV_CC) $(RISCV_CFLAGS) -c $< -o $@
+
+#ifneq ($(MAKECMDGOALS),clean)
+#-include $(DEPS)
+#endif
+
+# mkdir -p /repo/target/snitch_cluster/sw/snax/streamer-gemm/build
+# /tools/riscv-llvm/bin/clang -mcpu=snitch -menable-experimental-extensions -I/repo/target/snitch_cluster/sw/snax/streamer-gemm/include -I/repo/sw/snRuntime/src -I/repo/sw/snRuntime/api -I/repo/sw/snRuntime/src/omp -I/repo/sw/snRuntime/api/omp -I/repo/sw/snRuntime/vendor/riscv-opcodes -I/repo/target/snitch_cluster/sw/runtime/common -I/repo/target/snitch_cluster/sw/runtime/rtl/src -I/repo/target/snitch_cluster/sw/runtime/rtl/include -mabi=ilp32d -mcmodel=medany -ffast-math -fno-builtin-printf -fno-builtin-sqrtf -fno-common -fopenmp -ftls-model=local-exec -O3 -D__DEFINED_uint64_t -c /repo/target/snitch_cluster/sw/snax/streamer-gemm/src/snax-streamer-gemm-lib.c -o /repo/target/snitch_cluster/sw/snax/streamer-gemm/build/streamer-gemm.o

target/snitch_cluster/sw/snax/streamer-gemm/include/snax-streamer-gemm-lib.h

@@ -0,0 +1,37 @@
+// Copyright 2023 KU Leuven.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+//
+// Xiaoling Yi <[email protected]>
+
+#include <stdbool.h>
+#include "snrt.h"
+#include "stdint.h"
+
+#pragma once
+
+// Set STREAMER configuration CSR
+void set_streamer_csr(int tempLoop0, int tempLoop1, int tempLoop2,
+                      int tempStride0A, int tempStride2A, int spatialStride1A,
+                      int tempStride0B, int tempStride1B, int spatialStride1B,
+                      int tempStride1C, int tempStride2C, int spatialStride1C,
+                      int delta_local_a, int delta_local_b, int delta_local_c);
+
+// Set CSR to start STREAMER
+void set_streamer_start();
+
+// Set GEMM configuration CSR
+void set_block_gemm_csr(int tempLoop0, int tempLoop1, int tempLoop2,
+                        int subtractions);
+
+// Set CSR to start GEMM
+void set_block_gemm_start();
+
+// Poll until Streamer and GEMM accelerator finish
+void wait_streamer_gemm();
+
+// Read perforamcne counter of the Streamer, a read-only CSR
+uint32_t read_gemm_streamer_perf_counter();
+
+// Read perforamcne counter of GEMM, a read-only CSR
+uint32_t read_gemm_perf_counter();