wu-arch/kernels
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge branch 'ae' into ae-flash-virgo

Browse Source
This commit is contained in:
Richard Yan
2025-01-30 23:42:05 -08:00
parent 692f3dddff f8c51669c1
commit a4bd41392c
2 changed files with 17 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
kernels/sgemm_tcore/sgemm_impl.hpp
View File

@@ -1190,10 +1190,10 @@ inline void thread_block_gemm(const T *A, const T *B, float *C,
(uint64_t)(B + /*block_k:*/ 0 * BK * dim_n + block_n * BN),
k_LOOP_WS_CONFIG_ADDRS_AB)
// GEMMINI_CISC(8) does k_LOOP_WS_CONFIG_STRIDES_AB
GEMMINI_CISC_CMD_R((dim_n << 20) | (dim_k << 8) | 8);
GEMMINI_CISC_CMD_R((dim_n << 20) | (dim_k << 8) | GEMMINI_CISC_SET_AB_STRIDE);
gemmini_fence();
GEMMINI_CISC_CMD_I(10);
GEMMINI_CISC_CMD_R((11 << 16) | (0 << 8) | GEMMINI_CISC_LOAD_TO_HEXADECILES);
gemmini_fence();
#if 0
@@ -1266,8 +1266,9 @@ inline void thread_block_gemm(const T *A, const T *B, float *C,
// the last iteration of the k-loop is prefetching for the first
// iteration of the n-loop. The ping-poing indexing has to match for
// the two loop end to connect.
const uint32_t opcode = 11 - (block_k & 1);
GEMMINI_CISC_CMD_I(opcode);
const uint32_t a_hexadecile = (block_k & 1) * 4;
const uint32_t b_hexadecile = (block_k & 1) * 4 + 11;
GEMMINI_CISC_CMD_R((b_hexadecile << 16) | (a_hexadecile << 8) | GEMMINI_CISC_LOAD_TO_HEXADECILES);
// // TODO: branch is probably slow
// if (block_k & 1) {
// GEMMINI_CISC_CMD_I(12);

16
scripts/toolchain_env.sh
View File

@@ -1,4 +1,4 @@
#!/bin/sh
#!/bin/bash
# Copyright 2023 blaise
#
@@ -14,8 +14,10 @@
# See the License for the specific language governing permissions and
# limitations under the License.
TOOLDIR=${TOOLDIR:=$HOME/build/vortex-toolchain-prebuilt}
export TOOLDIR
ENV_SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
AE_TOOLCHAIN_DIR="$(realpath ${ENV_SCRIPT_DIR}/../../toolchain)"
export TOOLDIR=${AE_TOOLCHAIN_DIR}/vortex-toolchain-prebuilt
export VERILATOR_ROOT=$TOOLDIR/verilator
export PATH=$VERILATOR_ROOT/bin:$PATH
@@ -27,7 +29,7 @@ export YOSYS_PATH=$TOOLDIR/yosys
export PATH=$YOSYS_PATH/bin:$PATH
# LLVM_POCL seems to be only used in tests/opencl
export LLVM_POCL=/home/virgo-ae/build/llvm-vortex2
export LLVM_VORTEX=/home/virgo-ae/build/llvm-vortex2
export POCL_CC_PATH=/home/virgo-ae/build/pocl-vortex2/compiler
export POCL_RT_PATH=/home/virgo-ae/build/pocl-vortex2/runtime
export LLVM_POCL=${AE_TOOLCHAIN_DIR}/llvm-vortex2
export LLVM_VORTEX=${AE_TOOLCHAIN_DIR}/llvm-vortex2
export POCL_CC_PATH=${AE_TOOLCHAIN_DIR}/pocl-vortex2/compiler
export POCL_RT_PATH=${AE_TOOLCHAIN_DIR}/pocl-vortex2/runtime