sgemm_tcore: Use old opcodes to match frozen rtl
This commit is contained in:
@@ -6,4 +6,4 @@ VX_SRCS = kernel.cpp
|
|||||||
|
|
||||||
OPTS ?= -n16
|
OPTS ?= -n16
|
||||||
|
|
||||||
include ../common.mk
|
include ../common.mk
|
||||||
|
|||||||
@@ -334,7 +334,7 @@ inline void thread_block_gemm(kernel_arg_t *__UNIFORM__ arg,
|
|||||||
GEMMINI_CISC_CMD_R((dim_n << 16) | (dim_k << 8) | 8);
|
GEMMINI_CISC_CMD_R((dim_n << 16) | (dim_k << 8) | 8);
|
||||||
gemmini_fence();
|
gemmini_fence();
|
||||||
|
|
||||||
GEMMINI_CISC_CMD_I(12);
|
GEMMINI_CISC_CMD_I(10);
|
||||||
gemmini_fence();
|
gemmini_fence();
|
||||||
|
|
||||||
#if 0
|
#if 0
|
||||||
@@ -380,9 +380,9 @@ inline void thread_block_gemm(kernel_arg_t *__UNIFORM__ arg,
|
|||||||
GEMMINI_CISC_CMD_R((dim_n << 16) | (dim_k << 8) | 8);
|
GEMMINI_CISC_CMD_R((dim_n << 16) | (dim_k << 8) | 8);
|
||||||
// gemmini_fence();
|
// gemmini_fence();
|
||||||
|
|
||||||
// block_k is even: opcode 13 (write to local_a_buf)
|
// block_k is even: opcode 11 (write to local_a_buf)
|
||||||
// block_k is odd: opcode 12 (write to local_a)
|
// block_k is odd: opcode 10 (write to local_a)
|
||||||
const uint32_t opcode = 13 - (block_k & 1);
|
const uint32_t opcode = 11 - (block_k & 1);
|
||||||
GEMMINI_CISC_CMD_R(opcode);
|
GEMMINI_CISC_CMD_R(opcode);
|
||||||
// // TODO: branch is probably slow
|
// // TODO: branch is probably slow
|
||||||
// if (block_k & 1) {
|
// if (block_k & 1) {
|
||||||
|
|||||||
Reference in New Issue
Block a user