|
|
@@ -131,6 +131,8 @@ if (DEFINED CORE AND CMAKE_CROSSCOMPILING AND NOT (${HOST_OS} STREQUAL "WINDOWSS |
|
|
|
set(HAVE_SSE2 1) |
|
|
|
set(HAVE_SSE3 1) |
|
|
|
set(HAVE_SSSE3 1) |
|
|
|
set(SBGEMM_UNROLL_M 8) |
|
|
|
set(SBGEMM_UNROLL_N 4) |
|
|
|
set(SGEMM_UNROLL_M 8) |
|
|
|
set(SGEMM_UNROLL_N 4) |
|
|
|
set(DGEMM_UNROLL_M 4) |
|
|
@@ -143,6 +145,684 @@ if (DEFINED CORE AND CMAKE_CROSSCOMPILING AND NOT (${HOST_OS} STREQUAL "WINDOWSS |
|
|
|
set(CGEMM3M_UNROLL_N 4) |
|
|
|
set(ZGEMM3M_UNROLL_M 4) |
|
|
|
set(ZGEMM3M_UNROLL_N 4) |
|
|
|
elseif ("${TCORE}" STREQUAL "ATOM") |
|
|
|
file(APPEND ${TARGET_CONF_TEMP} |
|
|
|
"#define L1_DATA_SIZE\t24576\n" |
|
|
|
"#define L1_DATA_LINESIZE\t64\n" |
|
|
|
"#define L2_SIZE\t524288\n" |
|
|
|
"#define L2_LINESIZE\t64\n" |
|
|
|
"#define DTB_DEFAULT_ENTRIES\t64\n" |
|
|
|
"#define DTB_SIZE\t4096\n" |
|
|
|
"#define HAVE_CMOV\n" |
|
|
|
"#define HAVE_MMX\n" |
|
|
|
"#define HAVE_SSE\n" |
|
|
|
"#define HAVE_SSE2\n" |
|
|
|
"#define HAVE_SSE3\n" |
|
|
|
"#define HAVE_SSSE3\n" |
|
|
|
"#define SLOCAL_BUFFER_SIZE\t16384\n" |
|
|
|
"#define DLOCAL_BUFFER_SIZE\t8192\n" |
|
|
|
"#define CLOCAL_BUFFER_SIZE\t16384\n" |
|
|
|
"#define ZLOCAL_BUFFER_SIZE\t8192\n") |
|
|
|
set(HAVE_SSE 1) |
|
|
|
set(HAVE_SSE2 1) |
|
|
|
set(HAVE_SSE3 1) |
|
|
|
set(HAVE_SSSE3 1) |
|
|
|
set(SBGEMM_UNROLL_M 8) |
|
|
|
set(SBGEMM_UNROLL_N 4) |
|
|
|
set(SGEMM_UNROLL_M 8) |
|
|
|
set(SGEMM_UNROLL_N 4) |
|
|
|
set(DGEMM_UNROLL_M 4) |
|
|
|
set(DGEMM_UNROLL_N 2) |
|
|
|
set(CGEMM_UNROLL_M 4) |
|
|
|
set(CGEMM_UNROLL_N 2) |
|
|
|
set(ZGEMM_UNROLL_M 2) |
|
|
|
set(ZGEMM_UNROLL_N 1) |
|
|
|
set(CGEMM3M_UNROLL_M 8) |
|
|
|
set(CGEMM3M_UNROLL_N 4) |
|
|
|
set(ZGEMM3M_UNROLL_M 4) |
|
|
|
set(ZGEMM3M_UNROLL_N 4) |
|
|
|
elseif ("${TCORE}" STREQUAL "PRESCOTT") |
|
|
|
file(APPEND ${TARGET_CONF_TEMP} |
|
|
|
"#define L1_DATA_SIZE\t16384\n" |
|
|
|
"#define L1_DATA_LINESIZE\t64\n" |
|
|
|
"#define L2_SIZE\t1048576\n" |
|
|
|
"#define L2_LINESIZE\t64\n" |
|
|
|
"#define DTB_DEFAULT_ENTRIES\t64\n" |
|
|
|
"#define DTB_SIZE\t4096\n" |
|
|
|
"#define HAVE_CMOV\n" |
|
|
|
"#define HAVE_MMX\n" |
|
|
|
"#define HAVE_SSE\n" |
|
|
|
"#define HAVE_SSE2\n" |
|
|
|
"#define HAVE_SSE3\n" |
|
|
|
"#define SLOCAL_BUFFER_SIZE\t8192\n" |
|
|
|
"#define DLOCAL_BUFFER_SIZE\t8192\n" |
|
|
|
"#define CLOCAL_BUFFER_SIZE\t8192\n" |
|
|
|
"#define ZLOCAL_BUFFER_SIZE\t8192\n") |
|
|
|
set(HAVE_SSE 1) |
|
|
|
set(HAVE_SSE2 1) |
|
|
|
set(HAVE_SSE3 1) |
|
|
|
set(SBGEMM_UNROLL_M 8) |
|
|
|
set(SBGEMM_UNROLL_N 4) |
|
|
|
set(SGEMM_UNROLL_M 8) |
|
|
|
set(SGEMM_UNROLL_N 4) |
|
|
|
set(DGEMM_UNROLL_M 4) |
|
|
|
set(DGEMM_UNROLL_N 4) |
|
|
|
set(CGEMM_UNROLL_M 4) |
|
|
|
set(CGEMM_UNROLL_N 2) |
|
|
|
set(ZGEMM_UNROLL_M 2) |
|
|
|
set(ZGEMM_UNROLL_N 2) |
|
|
|
set(CGEMM3M_UNROLL_M 8) |
|
|
|
set(CGEMM3M_UNROLL_N 4) |
|
|
|
set(ZGEMM3M_UNROLL_M 4) |
|
|
|
set(ZGEMM3M_UNROLL_N 4) |
|
|
|
elseif ("${TCORE}" STREQUAL "NEHALEM") |
|
|
|
file(APPEND ${TARGET_CONF_TEMP} |
|
|
|
"#define L1_DATA_SIZE\t32768\n" |
|
|
|
"#define L1_DATA_LINESIZE\t64\n" |
|
|
|
"#define L2_SIZE\t262144\n" |
|
|
|
"#define L2_LINESIZE\t64\n" |
|
|
|
"#define DTB_DEFAULT_ENTRIES\t64\n" |
|
|
|
"#define DTB_SIZE\t4096\n" |
|
|
|
"#define HAVE_CMOV\n" |
|
|
|
"#define HAVE_MMX\n" |
|
|
|
"#define HAVE_SSE\n" |
|
|
|
"#define HAVE_SSE2\n" |
|
|
|
"#define HAVE_SSE3\n" |
|
|
|
"#define HAVE_SSSE3\n" |
|
|
|
"#define HAVE_SSE4_1\n" |
|
|
|
"#define HAVE_SSE4_2\n" |
|
|
|
"#define SLOCAL_BUFFER_SIZE\t65535\n" |
|
|
|
"#define DLOCAL_BUFFER_SIZE\t32768\n" |
|
|
|
"#define CLOCAL_BUFFER_SIZE\t65536\n" |
|
|
|
"#define ZLOCAL_BUFFER_SIZE\t32768\n") |
|
|
|
set(HAVE_SSE 1) |
|
|
|
set(HAVE_SSE2 1) |
|
|
|
set(HAVE_SSE3 1) |
|
|
|
set(HAVE_SSSE3 1) |
|
|
|
set(HAVE_SSE4_1 1) |
|
|
|
set(HAVE_SSE4_2 1) |
|
|
|
set(SBGEMM_UNROLL_M 8) |
|
|
|
set(SBGEMM_UNROLL_N 4) |
|
|
|
set(SGEMM_UNROLL_M 4) |
|
|
|
set(SGEMM_UNROLL_N 8) |
|
|
|
set(DGEMM_UNROLL_M 2) |
|
|
|
set(DGEMM_UNROLL_N 8) |
|
|
|
set(CGEMM_UNROLL_M 2) |
|
|
|
set(CGEMM_UNROLL_N 4) |
|
|
|
set(ZGEMM_UNROLL_M 1) |
|
|
|
set(ZGEMM_UNROLL_N 4) |
|
|
|
set(CGEMM3M_UNROLL_M 4) |
|
|
|
set(CGEMM3M_UNROLL_N 8) |
|
|
|
set(ZGEMM3M_UNROLL_M 2) |
|
|
|
set(ZGEMM3M_UNROLL_N 8) |
|
|
|
elseif ("${TCORE}" STREQUAL "SANDYBRIDGE") |
|
|
|
file(APPEND ${TARGET_CONF_TEMP} |
|
|
|
"#define L1_DATA_SIZE\t32768\n" |
|
|
|
"#define L1_DATA_LINESIZE\t64\n" |
|
|
|
"#define L2_SIZE\t262144\n" |
|
|
|
"#define L2_LINESIZE\t64\n" |
|
|
|
"#define DTB_DEFAULT_ENTRIES\t64\n" |
|
|
|
"#define DTB_SIZE\t4096\n" |
|
|
|
"#define HAVE_CMOV\n" |
|
|
|
"#define HAVE_MMX\n" |
|
|
|
"#define HAVE_SSE\n" |
|
|
|
"#define HAVE_SSE2\n" |
|
|
|
"#define HAVE_SSE3\n" |
|
|
|
"#define HAVE_SSSE3\n" |
|
|
|
"#define HAVE_SSE4_1\n" |
|
|
|
"#define HAVE_SSE4_2\n" |
|
|
|
"#define HAVE_AVX\n" |
|
|
|
"#define SLOCAL_BUFFER_SIZE\t24576\n" |
|
|
|
"#define DLOCAL_BUFFER_SIZE\t16384\n" |
|
|
|
"#define CLOCAL_BUFFER_SIZE\t32768\n" |
|
|
|
"#define ZLOCAL_BUFFER_SIZE\t24576\n") |
|
|
|
set(HAVE_SSE 1) |
|
|
|
set(HAVE_SSE2 1) |
|
|
|
set(HAVE_SSE3 1) |
|
|
|
set(HAVE_SSSE3 1) |
|
|
|
set(HAVE_SSE4_1 1) |
|
|
|
set(HAVE_SSE4_2 1) |
|
|
|
set(HAVE_AVX 1) |
|
|
|
set(SBGEMM_UNROLL_M 8) |
|
|
|
set(SBGEMM_UNROLL_N 4) |
|
|
|
set(SGEMM_UNROLL_M 16) |
|
|
|
set(SGEMM_UNROLL_N 4) |
|
|
|
set(DGEMM_UNROLL_M 8) |
|
|
|
set(DGEMM_UNROLL_N 4) |
|
|
|
set(CGEMM_UNROLL_M 8) |
|
|
|
set(CGEMM_UNROLL_N 2) |
|
|
|
set(ZGEMM_UNROLL_M 1) |
|
|
|
set(ZGEMM_UNROLL_N 4) |
|
|
|
set(CGEMM3M_UNROLL_M 4) |
|
|
|
set(CGEMM3M_UNROLL_N 8) |
|
|
|
set(ZGEMM3M_UNROLL_M 2) |
|
|
|
set(ZGEMM3M_UNROLL_N 8) |
|
|
|
elseif ("${TCORE}" STREQUAL "HASWELL") |
|
|
|
file(APPEND ${TARGET_CONF_TEMP} |
|
|
|
"#define L1_DATA_SIZE\t32768\n" |
|
|
|
"#define L1_DATA_LINESIZE\t64\n" |
|
|
|
"#define L2_SIZE\t262144\n" |
|
|
|
"#define L2_LINESIZE\t64\n" |
|
|
|
"#define DTB_DEFAULT_ENTRIES\t64\n" |
|
|
|
"#define DTB_SIZE\t4096\n" |
|
|
|
"#define HAVE_CMOV\n" |
|
|
|
"#define HAVE_MMX\n" |
|
|
|
"#define HAVE_SSE\n" |
|
|
|
"#define HAVE_SSE2\n" |
|
|
|
"#define HAVE_SSE3\n" |
|
|
|
"#define HAVE_SSSE3\n" |
|
|
|
"#define HAVE_SSE4_1\n" |
|
|
|
"#define HAVE_SSE4_2\n" |
|
|
|
"#define HAVE_AVX\n" |
|
|
|
"#define HAVE_AVX2\n" |
|
|
|
"#define HAVE_FMA3\n" |
|
|
|
"#define SLOCAL_BUFFER_SIZE\t20480\n" |
|
|
|
"#define DLOCAL_BUFFER_SIZE\t32768\n" |
|
|
|
"#define CLOCAL_BUFFER_SIZE\t16384\n" |
|
|
|
"#define ZLOCAL_BUFFER_SIZE\t12288\n") |
|
|
|
set(HAVE_SSE 1) |
|
|
|
set(HAVE_SSE2 1) |
|
|
|
set(HAVE_SSE3 1) |
|
|
|
set(HAVE_SSSE3 1) |
|
|
|
set(HAVE_SSE4_1 1) |
|
|
|
set(HAVE_SSE4_2 1) |
|
|
|
set(HAVE_AVX 1) |
|
|
|
set(HAVE_AVX2 1) |
|
|
|
set(HAVE_FMA3 1) |
|
|
|
set(SBGEMM_UNROLL_M 8) |
|
|
|
set(SBGEMM_UNROLL_N 4) |
|
|
|
set(SGEMM_UNROLL_M 8) |
|
|
|
set(SGEMM_UNROLL_N 4) |
|
|
|
set(DGEMM_UNROLL_M 4) |
|
|
|
set(DGEMM_UNROLL_N 8) |
|
|
|
set(CGEMM_UNROLL_M 8) |
|
|
|
set(CGEMM_UNROLL_N 2) |
|
|
|
set(ZGEMM_UNROLL_M 4) |
|
|
|
set(ZGEMM_UNROLL_N 2) |
|
|
|
set(CGEMM3M_UNROLL_M 8) |
|
|
|
set(CGEMM3M_UNROLL_N 4) |
|
|
|
set(ZGEMM3M_UNROLL_M 4) |
|
|
|
set(ZGEMM3M_UNROLL_N 4) |
|
|
|
elseif ("${TCORE}" STREQUAL "SKYLAKEX") |
|
|
|
file(APPEND ${TARGET_CONF_TEMP} |
|
|
|
"#define L1_DATA_SIZE\t32768\n" |
|
|
|
"#define L1_DATA_LINESIZE\t64\n" |
|
|
|
"#define L2_SIZE\t262144\n" |
|
|
|
"#define L2_LINESIZE\t64\n" |
|
|
|
"#define DTB_DEFAULT_ENTRIES\t64\n" |
|
|
|
"#define DTB_SIZE\t4096\n" |
|
|
|
"#define HAVE_CMOV\n" |
|
|
|
"#define HAVE_MMX\n" |
|
|
|
"#define HAVE_SSE\n" |
|
|
|
"#define HAVE_SSE2\n" |
|
|
|
"#define HAVE_SSE3\n" |
|
|
|
"#define HAVE_SSSE3\n" |
|
|
|
"#define HAVE_SSE4_1\n" |
|
|
|
"#define HAVE_SSE4_2\n" |
|
|
|
"#define HAVE_AVX\n" |
|
|
|
"#define HAVE_AVX2\n" |
|
|
|
"#define HAVE_FMA3\n" |
|
|
|
"#define HAVE_AVX512VL\n" |
|
|
|
"#define SLOCAL_BUFFER_SIZE\t28672\n" |
|
|
|
"#define DLOCAL_BUFFER_SIZE\t12288\n" |
|
|
|
"#define CLOCAL_BUFFER_SIZE\t12288\n" |
|
|
|
"#define ZLOCAL_BUFFER_SIZE\t8192\n") |
|
|
|
set(HAVE_CMOV 1) |
|
|
|
set(HAVE_MMX 1) |
|
|
|
set(HAVE_SSE 1) |
|
|
|
set(HAVE_SSE2 1) |
|
|
|
set(HAVE_SSE3 1) |
|
|
|
set(HAVE_SSSE3 1) |
|
|
|
set(HAVE_SSE4_1 1) |
|
|
|
set(HAVE_SSE4_2 1) |
|
|
|
set(HAVE_AVX 1) |
|
|
|
set(HAVE_AVX2 1) |
|
|
|
set(HAVE_FMA3 1) |
|
|
|
set(HAVE_AVX512VL 1) |
|
|
|
set(SBGEMM_UNROLL_M 8) |
|
|
|
set(SBGEMM_UNROLL_N 4) |
|
|
|
set(SGEMM_UNROLL_M 16) |
|
|
|
set(SGEMM_UNROLL_N 4) |
|
|
|
set(DGEMM_UNROLL_M 16) |
|
|
|
set(DGEMM_UNROLL_N 2) |
|
|
|
set(CGEMM_UNROLL_M 8) |
|
|
|
set(CGEMM_UNROLL_N 2) |
|
|
|
set(ZGEMM_UNROLL_M 4) |
|
|
|
set(ZGEMM_UNROLL_N 2) |
|
|
|
set(CGEMM3M_UNROLL_M 8) |
|
|
|
set(CGEMM3M_UNROLL_N 4) |
|
|
|
set(ZGEMM3M_UNROLL_M 4) |
|
|
|
set(ZGEMM3M_UNROLL_N 4) |
|
|
|
elseif ("${TCORE}" STREQUAL "COOPERLAKE") |
|
|
|
file(APPEND ${TARGET_CONF_TEMP} |
|
|
|
"#define L1_DATA_SIZE\t32768\n" |
|
|
|
"#define L1_DATA_LINESIZE\t64\n" |
|
|
|
"#define L2_SIZE\t262144\n" |
|
|
|
"#define L2_LINESIZE\t64\n" |
|
|
|
"#define DTB_DEFAULT_ENTRIES\t64\n" |
|
|
|
"#define DTB_SIZE\t4096\n" |
|
|
|
"#define HAVE_CMOV\n" |
|
|
|
"#define HAVE_MMX\n" |
|
|
|
"#define HAVE_SSE\n" |
|
|
|
"#define HAVE_SSE2\n" |
|
|
|
"#define HAVE_SSE3\n" |
|
|
|
"#define HAVE_SSSE3\n" |
|
|
|
"#define HAVE_SSE4_1\n" |
|
|
|
"#define HAVE_SSE4_2\n" |
|
|
|
"#define HAVE_AVX\n" |
|
|
|
"#define HAVE_AVX2\n" |
|
|
|
"#define HAVE_FMA3\n" |
|
|
|
"#define HAVE_AVX512VL\n" |
|
|
|
"#define HAVE_AVX512BF16\n" |
|
|
|
"#define SLOCAL_BUFFER_SIZE\t20480\n" |
|
|
|
"#define DLOCAL_BUFFER_SIZE\t12288\n" |
|
|
|
"#define CLOCAL_BUFFER_SIZE\t12288\n" |
|
|
|
"#define ZLOCAL_BUFFER_SIZE\t8192\n") |
|
|
|
set(HAVE_CMOV 1) |
|
|
|
set(HAVE_MMX 1) |
|
|
|
set(HAVE_SSE 1) |
|
|
|
set(HAVE_SSE2 1) |
|
|
|
set(HAVE_SSE3 1) |
|
|
|
set(HAVE_SSSE3 1) |
|
|
|
set(HAVE_SSE4_1 1) |
|
|
|
set(HAVE_SSE4_2 1) |
|
|
|
set(HAVE_AVX 1) |
|
|
|
set(HAVE_AVX2 1) |
|
|
|
set(HAVE_FMA3 1) |
|
|
|
set(HAVE_AVX512VL 1) |
|
|
|
set(HAVE_AVX512BF16 1) |
|
|
|
set(SBGEMM_UNROLL_M 16) |
|
|
|
set(SBGEMM_UNROLL_N 4) |
|
|
|
set(SGEMM_UNROLL_M 16) |
|
|
|
set(SGEMM_UNROLL_N 4) |
|
|
|
set(DGEMM_UNROLL_M 16) |
|
|
|
set(DGEMM_UNROLL_N 2) |
|
|
|
set(CGEMM_UNROLL_M 8) |
|
|
|
set(CGEMM_UNROLL_N 2) |
|
|
|
set(ZGEMM_UNROLL_M 4) |
|
|
|
set(ZGEMM_UNROLL_N 2) |
|
|
|
set(CGEMM3M_UNROLL_M 8) |
|
|
|
set(CGEMM3M_UNROLL_N 4) |
|
|
|
set(ZGEMM3M_UNROLL_M 4) |
|
|
|
set(ZGEMM3M_UNROLL_N 4) |
|
|
|
elseif ("${TCORE}" STREQUAL "SAPPHIRERAPIDS") |
|
|
|
file(APPEND ${TARGET_CONF_TEMP} |
|
|
|
"#define L1_DATA_SIZE\t32768\n" |
|
|
|
"#define L1_DATA_LINESIZE\t64\n" |
|
|
|
"#define L2_SIZE\t262144\n" |
|
|
|
"#define L2_LINESIZE\t64\n" |
|
|
|
"#define DTB_DEFAULT_ENTRIES\t64\n" |
|
|
|
"#define DTB_SIZE\t4096\n" |
|
|
|
"#define HAVE_CMOV\n" |
|
|
|
"#define HAVE_MMX\n" |
|
|
|
"#define HAVE_SSE\n" |
|
|
|
"#define HAVE_SSE2\n" |
|
|
|
"#define HAVE_SSE3\n" |
|
|
|
"#define HAVE_SSSE3\n" |
|
|
|
"#define HAVE_SSE4_1\n" |
|
|
|
"#define HAVE_SSE4_2\n" |
|
|
|
"#define HAVE_AVX\n" |
|
|
|
"#define HAVE_AVX2\n" |
|
|
|
"#define HAVE_FMA3\n" |
|
|
|
"#define HAVE_AVX512VL\n" |
|
|
|
"#define HAVE_AVX512BF16\n" |
|
|
|
"#define SLOCAL_BUFFER_SIZE\t20480\n" |
|
|
|
"#define DLOCAL_BUFFER_SIZE\t12288\n" |
|
|
|
"#define CLOCAL_BUFFER_SIZE\t12288\n" |
|
|
|
"#define ZLOCAL_BUFFER_SIZE\t8192\n") |
|
|
|
set(HAVE_CMOV 1) |
|
|
|
set(HAVE_MMX 1) |
|
|
|
set(HAVE_SSE 1) |
|
|
|
set(HAVE_SSE2 1) |
|
|
|
set(HAVE_SSE3 1) |
|
|
|
set(HAVE_SSSE3 1) |
|
|
|
set(HAVE_SSE4_1 1) |
|
|
|
set(HAVE_SSE4_2 1) |
|
|
|
set(HAVE_AVX 1) |
|
|
|
set(HAVE_AVX2 1) |
|
|
|
set(HAVE_FMA3 1) |
|
|
|
set(HAVE_AVX512VL 1) |
|
|
|
set(HAVE_AVX512BF16 1) |
|
|
|
set(SBGEMM_UNROLL_M 32) |
|
|
|
set(SBGEMM_UNROLL_N 16) |
|
|
|
set(SGEMM_UNROLL_M 16) |
|
|
|
set(SGEMM_UNROLL_N 4) |
|
|
|
set(DGEMM_UNROLL_M 16) |
|
|
|
set(DGEMM_UNROLL_N 2) |
|
|
|
set(CGEMM_UNROLL_M 8) |
|
|
|
set(CGEMM_UNROLL_N 2) |
|
|
|
set(ZGEMM_UNROLL_M 4) |
|
|
|
set(ZGEMM_UNROLL_N 2) |
|
|
|
set(CGEMM3M_UNROLL_M 8) |
|
|
|
set(CGEMM3M_UNROLL_N 4) |
|
|
|
set(ZGEMM3M_UNROLL_M 4) |
|
|
|
set(ZGEMM3M_UNROLL_N 4) |
|
|
|
elseif ("${TCORE}" STREQUAL "OPTERON") |
|
|
|
file(APPEND ${TARGET_CONF_TEMP} |
|
|
|
"#define L1_DATA_SIZE\t65536\n" |
|
|
|
"#define L1_DATA_LINESIZE\t64\n" |
|
|
|
"#define L2_SIZE\t1048576\n" |
|
|
|
"#define L2_LINESIZE\t64\n" |
|
|
|
"#define DTB_DEFAULT_ENTRIES\t32\n" |
|
|
|
"#define DTB_SIZE\t4096\n" |
|
|
|
"#define HAVE_3DNOW\n" |
|
|
|
"#define HAVE_3DNOWEX\n" |
|
|
|
"#define HAVE_MMX\n" |
|
|
|
"#define HAVE_SSE\n" |
|
|
|
"#define HAVE_SSE2\n" |
|
|
|
"#define SLOCAL_BUFFER_SIZE\t15360\n" |
|
|
|
"#define DLOCAL_BUFFER_SIZE\t15360\n" |
|
|
|
"#define CLOCAL_BUFFER_SIZE\t15360\n" |
|
|
|
"#define ZLOCAL_BUFFER_SIZE\t15360\n") |
|
|
|
set(HAVE_3DNOW 1) |
|
|
|
set(HAVE_3DNOWEX 1) |
|
|
|
set(HAVE_MMX 1) |
|
|
|
set(HAVE_SSE 1) |
|
|
|
set(HAVE_SSE2 1) |
|
|
|
set(SBGEMM_UNROLL_M 8) |
|
|
|
set(SBGEMM_UNROLL_N 4) |
|
|
|
set(SGEMM_UNROLL_M 8) |
|
|
|
set(SGEMM_UNROLL_N 4) |
|
|
|
set(DGEMM_UNROLL_M 4) |
|
|
|
set(DGEMM_UNROLL_N 4) |
|
|
|
set(CGEMM_UNROLL_M 4) |
|
|
|
set(CGEMM_UNROLL_N 2) |
|
|
|
set(ZGEMM_UNROLL_M 2) |
|
|
|
set(ZGEMM_UNROLL_N 2) |
|
|
|
set(CGEMM3M_UNROLL_M 8) |
|
|
|
set(CGEMM3M_UNROLL_N 4) |
|
|
|
set(ZGEMM3M_UNROLL_M 4) |
|
|
|
set(ZGEMM3M_UNROLL_N 4) |
|
|
|
elseif ("${TCORE}" STREQUAL "BARCELONA") |
|
|
|
file(APPEND ${TARGET_CONF_TEMP} |
|
|
|
"#define L1_DATA_SIZE\t32768\n" |
|
|
|
"#define L1_DATA_LINESIZE\t64\n" |
|
|
|
"#define L2_SIZE\t524288\n" |
|
|
|
"#define L2_LINESIZE\t64\n" |
|
|
|
"#define DTB_DEFAULT_ENTRIES\t64\n" |
|
|
|
"#define DTB_SIZE\t4096\n" |
|
|
|
"#define HAVE_MMX\n" |
|
|
|
"#define HAVE_SSE\n" |
|
|
|
"#define HAVE_SSE2\n" |
|
|
|
"#define HAVE_SSE3\n" |
|
|
|
"#define HAVE_SSE4A\n" |
|
|
|
"#define HAVE_MISALIGNSSE\n" |
|
|
|
"#define HAVE_128BITFPU\n" |
|
|
|
"#define HAVE_FASTMOVU\n" |
|
|
|
"#define SLOCAL_BUFFER_SIZE\t14336\n" |
|
|
|
"#define DLOCAL_BUFFER_SIZE\t14336\n" |
|
|
|
"#define CLOCAL_BUFFER_SIZE\t14336\n" |
|
|
|
"#define ZLOCAL_BUFFER_SIZE\t14336\n") |
|
|
|
set(HAVE_SSE 1) |
|
|
|
set(HAVE_SSE2 1) |
|
|
|
set(HAVE_SSE3 1) |
|
|
|
set(HAVE_SSE4A 1) |
|
|
|
set(HAVE_MISALIGNSSE 1) |
|
|
|
set(HAVE_128BITFPU 1) |
|
|
|
set(HAVE_FASTMOVU 1) |
|
|
|
set(SBGEMM_UNROLL_M 8) |
|
|
|
set(SBGEMM_UNROLL_N 4) |
|
|
|
set(SGEMM_UNROLL_M 8) |
|
|
|
set(SGEMM_UNROLL_N 4) |
|
|
|
set(DGEMM_UNROLL_M 4) |
|
|
|
set(DGEMM_UNROLL_N 4) |
|
|
|
set(CGEMM_UNROLL_M 4) |
|
|
|
set(CGEMM_UNROLL_N 2) |
|
|
|
set(ZGEMM_UNROLL_M 2) |
|
|
|
set(ZGEMM_UNROLL_N 2) |
|
|
|
set(CGEMM3M_UNROLL_M 8) |
|
|
|
set(CGEMM3M_UNROLL_N 4) |
|
|
|
set(ZGEMM3M_UNROLL_M 4) |
|
|
|
set(ZGEMM3M_UNROLL_N 4) |
|
|
|
elseif ("${TCORE}" STREQUAL "BULLDOZER") |
|
|
|
file(APPEND ${TARGET_CONF_TEMP} |
|
|
|
"#define L1_DATA_SIZE\t49152\n" |
|
|
|
"#define L1_DATA_LINESIZE\t64\n" |
|
|
|
"#define L2_SIZE\t1024000\n" |
|
|
|
"#define L2_LINESIZE\t64\n" |
|
|
|
"#define DTB_DEFAULT_ENTRIES\t32\n" |
|
|
|
"#define DTB_SIZE\t4096\n" |
|
|
|
"#define HAVE_MMX\n" |
|
|
|
"#define HAVE_SSE\n" |
|
|
|
"#define HAVE_SSE2\n" |
|
|
|
"#define HAVE_SSE3\n" |
|
|
|
"#define HAVE_SSE4A\n" |
|
|
|
"#define HAVE_AVX\n" |
|
|
|
"#define HAVE_MISALIGNSSE\n" |
|
|
|
"#define HAVE_128BITFPU\n" |
|
|
|
"#define HAVE_FASTMOVU\n" |
|
|
|
"#define SLOCAL_BUFFER_SIZE\t5376\n" |
|
|
|
"#define DLOCAL_BUFFER_SIZE\t5376\n" |
|
|
|
"#define CLOCAL_BUFFER_SIZE\t14336\n" |
|
|
|
"#define ZLOCAL_BUFFER_SIZE\t14336\n") |
|
|
|
set(HAVE_SSE 1) |
|
|
|
set(HAVE_SSE2 1) |
|
|
|
set(HAVE_SSE3 1) |
|
|
|
set(HAVE_SSE4A 1) |
|
|
|
set(HAVE_AVX 1) |
|
|
|
set(HAVE_MISALIGNSSE 1) |
|
|
|
set(HAVE_128BITFPU 1) |
|
|
|
set(HAVE_FASTMOVU 1) |
|
|
|
set(SBGEMM_UNROLL_M 8) |
|
|
|
set(SBGEMM_UNROLL_N 4) |
|
|
|
set(SGEMM_UNROLL_M 16) |
|
|
|
set(SGEMM_UNROLL_N 2) |
|
|
|
set(DGEMM_UNROLL_M 8) |
|
|
|
set(DGEMM_UNROLL_N 2) |
|
|
|
set(CGEMM_UNROLL_M 2) |
|
|
|
set(CGEMM_UNROLL_N 2) |
|
|
|
set(ZGEMM_UNROLL_M 2) |
|
|
|
set(ZGEMM_UNROLL_N 2) |
|
|
|
set(CGEMM3M_UNROLL_M 8) |
|
|
|
set(CGEMM3M_UNROLL_N 4) |
|
|
|
set(ZGEMM3M_UNROLL_M 4) |
|
|
|
set(ZGEMM3M_UNROLL_N 4) |
|
|
|
elseif ("${TCORE}" STREQUAL "PILEDRIVER") |
|
|
|
file(APPEND ${TARGET_CONF_TEMP} |
|
|
|
"#define L1_DATA_SIZE\t16384\n" |
|
|
|
"#define L1_DATA_LINESIZE\t64\n" |
|
|
|
"#define L2_SIZE\t2097152\n" |
|
|
|
"#define L2_LINESIZE\t64\n" |
|
|
|
"#define DTB_DEFAULT_ENTRIES\t64\n" |
|
|
|
"#define DTB_SIZE\t4096\n" |
|
|
|
"#define HAVE_MMX\n" |
|
|
|
"#define HAVE_SSE\n" |
|
|
|
"#define HAVE_SSE2\n" |
|
|
|
"#define HAVE_SSE3\n" |
|
|
|
"#define HAVE_SSE4_1\n" |
|
|
|
"#define HAVE_SSE4_2\n" |
|
|
|
"#define HAVE_SSE4A\n" |
|
|
|
"#define HAVE_AVX\n" |
|
|
|
"#define HAVE_MISALIGNSSE\n" |
|
|
|
"#define HAVE_128BITFPU\n" |
|
|
|
"#define HAVE_FASTMOVU\n" |
|
|
|
"#define HAVE_CFLUSH\n" |
|
|
|
"#define HAVE_FMA3\n" |
|
|
|
"#define SLOCAL_BUFFER_SIZE\t6144\n" |
|
|
|
"#define DLOCAL_BUFFER_SIZE\t5376\n" |
|
|
|
"#define CLOCAL_BUFFER_SIZE\t10752\n" |
|
|
|
"#define ZLOCAL_BUFFER_SIZE\t10752\n") |
|
|
|
set(HAVE_SSE 1) |
|
|
|
set(HAVE_SSE2 1) |
|
|
|
set(HAVE_SSE3 1) |
|
|
|
set(HAVE_SSE4_1 1) |
|
|
|
set(HAVE_SSE4_2 1) |
|
|
|
set(HAVE_SSE4A 1) |
|
|
|
set(HAVE_AVX 1) |
|
|
|
set(HAVE_FMA3 1) |
|
|
|
set(HAVE_MISALIGNSSE 1) |
|
|
|
set(HAVE_128BITFPU 1) |
|
|
|
set(HAVE_FASTMOVU 1) |
|
|
|
set(HAVE_CFLUSH 1) |
|
|
|
set(SBGEMM_UNROLL_M 8) |
|
|
|
set(SBGEMM_UNROLL_N 4) |
|
|
|
set(SGEMM_UNROLL_M 16) |
|
|
|
set(SGEMM_UNROLL_N 2) |
|
|
|
set(DGEMM_UNROLL_M 8) |
|
|
|
set(DGEMM_UNROLL_N 2) |
|
|
|
set(CGEMM_UNROLL_M 4) |
|
|
|
set(CGEMM_UNROLL_N 2) |
|
|
|
set(ZGEMM_UNROLL_M 2) |
|
|
|
set(ZGEMM_UNROLL_N 2) |
|
|
|
set(CGEMM3M_UNROLL_M 8) |
|
|
|
set(CGEMM3M_UNROLL_N 4) |
|
|
|
set(ZGEMM3M_UNROLL_M 4) |
|
|
|
set(ZGEMM3M_UNROLL_N 4) |
|
|
|
elseif ("${TCORE}" STREQUAL "STEAMROLLER") |
|
|
|
file(APPEND ${TARGET_CONF_TEMP} |
|
|
|
"#define L1_DATA_SIZE\t16384\n" |
|
|
|
"#define L1_DATA_LINESIZE\t64\n" |
|
|
|
"#define L2_SIZE\t2097152\n" |
|
|
|
"#define L2_LINESIZE\t64\n" |
|
|
|
"#define DTB_DEFAULT_ENTRIES\t64\n" |
|
|
|
"#define DTB_SIZE\t4096\n" |
|
|
|
"#define HAVE_MMX\n" |
|
|
|
"#define HAVE_SSE\n" |
|
|
|
"#define HAVE_SSE2\n" |
|
|
|
"#define HAVE_SSE3\n" |
|
|
|
"#define HAVE_SSE4_1\n" |
|
|
|
"#define HAVE_SSE4_2\n" |
|
|
|
"#define HAVE_SSE4A\n" |
|
|
|
"#define HAVE_AVX\n" |
|
|
|
"#define HAVE_MISALIGNSSE\n" |
|
|
|
"#define HAVE_128BITFPU\n" |
|
|
|
"#define HAVE_FASTMOVU\n" |
|
|
|
"#define HAVE_CFLUSH\n" |
|
|
|
"#define HAVE_FMA3\n" |
|
|
|
"#define SLOCAL_BUFFER_SIZE\t6144\n" |
|
|
|
"#define DLOCAL_BUFFER_SIZE\t5120\n" |
|
|
|
"#define CLOCAL_BUFFER_SIZE\t10240\n" |
|
|
|
"#define ZLOCAL_BUFFER_SIZE\t10240\n") |
|
|
|
set(HAVE_SSE 1) |
|
|
|
set(HAVE_SSE2 1) |
|
|
|
set(HAVE_SSE3 1) |
|
|
|
set(HAVE_SSE4_1 1) |
|
|
|
set(HAVE_SSE4_2 1) |
|
|
|
set(HAVE_SSE4A 1) |
|
|
|
set(HAVE_AVX 1) |
|
|
|
set(HAVE_FMA3 1) |
|
|
|
set(HAVE_MISALIGNSSE 1) |
|
|
|
set(HAVE_128BITFPU 1) |
|
|
|
set(HAVE_FASTMOVU 1) |
|
|
|
set(HAVE_CFLUSH 1) |
|
|
|
set(SBGEMM_UNROLL_M 8) |
|
|
|
set(SBGEMM_UNROLL_N 4) |
|
|
|
set(SGEMM_UNROLL_M 16) |
|
|
|
set(SGEMM_UNROLL_N 2) |
|
|
|
set(DGEMM_UNROLL_M 8) |
|
|
|
set(DGEMM_UNROLL_N 2) |
|
|
|
set(CGEMM_UNROLL_M 4) |
|
|
|
set(CGEMM_UNROLL_N 2) |
|
|
|
set(ZGEMM_UNROLL_M 2) |
|
|
|
set(ZGEMM_UNROLL_N 2) |
|
|
|
set(CGEMM3M_UNROLL_M 8) |
|
|
|
set(CGEMM3M_UNROLL_N 4) |
|
|
|
set(ZGEMM3M_UNROLL_M 4) |
|
|
|
set(ZGEMM3M_UNROLL_N 4) |
|
|
|
elseif ("${TCORE}" STREQUAL "EXCAVATOR") |
|
|
|
file(APPEND ${TARGET_CONF_TEMP} |
|
|
|
"#define L1_DATA_SIZE\t16384\n" |
|
|
|
"#define L1_DATA_LINESIZE\t64\n" |
|
|
|
"#define L2_SIZE\t2097152\n" |
|
|
|
"#define L2_LINESIZE\t64\n" |
|
|
|
"#define DTB_DEFAULT_ENTRIES\t64\n" |
|
|
|
"#define DTB_SIZE\t4096\n" |
|
|
|
"#define HAVE_MMX\n" |
|
|
|
"#define HAVE_SSE\n" |
|
|
|
"#define HAVE_SSE2\n" |
|
|
|
"#define HAVE_SSE3\n" |
|
|
|
"#define HAVE_SSE4_1\n" |
|
|
|
"#define HAVE_SSE4_2\n" |
|
|
|
"#define HAVE_SSE4A\n" |
|
|
|
"#define HAVE_AVX\n" |
|
|
|
"#define HAVE_MISALIGNSSE\n" |
|
|
|
"#define HAVE_128BITFPU\n" |
|
|
|
"#define HAVE_FASTMOVU\n" |
|
|
|
"#define HAVE_CFLUSH\n" |
|
|
|
"#define HAVE_FMA3\n" |
|
|
|
"#define SLOCAL_BUFFER_SIZE\t6144\n" |
|
|
|
"#define DLOCAL_BUFFER_SIZE\t5120\n" |
|
|
|
"#define CLOCAL_BUFFER_SIZE\t10240\n" |
|
|
|
"#define ZLOCAL_BUFFER_SIZE\t10240\n") |
|
|
|
set(HAVE_SSE 1) |
|
|
|
set(HAVE_SSE2 1) |
|
|
|
set(HAVE_SSE3 1) |
|
|
|
set(HAVE_SSE4_1 1) |
|
|
|
set(HAVE_SSE4_2 1) |
|
|
|
set(HAVE_SSE4A 1) |
|
|
|
set(HAVE_AVX 1) |
|
|
|
set(HAVE_FMA3 1) |
|
|
|
set(HAVE_MISALIGNSSE 1) |
|
|
|
set(HAVE_128BITFPU 1) |
|
|
|
set(HAVE_FASTMOVU 1) |
|
|
|
set(HAVE_CFLUSH 1) |
|
|
|
set(SBGEMM_UNROLL_M 8) |
|
|
|
set(SBGEMM_UNROLL_N 4) |
|
|
|
set(SGEMM_UNROLL_M 16) |
|
|
|
set(SGEMM_UNROLL_N 2) |
|
|
|
set(DGEMM_UNROLL_M 8) |
|
|
|
set(DGEMM_UNROLL_N 2) |
|
|
|
set(CGEMM_UNROLL_M 4) |
|
|
|
set(CGEMM_UNROLL_N 2) |
|
|
|
set(ZGEMM_UNROLL_M 2) |
|
|
|
set(ZGEMM_UNROLL_N 2) |
|
|
|
set(CGEMM3M_UNROLL_M 8) |
|
|
|
set(CGEMM3M_UNROLL_N 4) |
|
|
|
set(ZGEMM3M_UNROLL_M 4) |
|
|
|
set(ZGEMM3M_UNROLL_N 4) |
|
|
|
elseif ("${TCORE}" STREQUAL "ZEN") |
|
|
|
file(APPEND ${TARGET_CONF_TEMP} |
|
|
|
"#define L1_DATA_SIZE\t32768\n" |
|
|
|
"#define L1_DATA_LINESIZE\t64\n" |
|
|
|
"#define L2_SIZE\t524288\n" |
|
|
|
"#define L2_LINESIZE\t64\n" |
|
|
|
"#define DTB_DEFAULT_ENTRIES\t64\n" |
|
|
|
"#define DTB_SIZE\t4096\n" |
|
|
|
"#define HAVE_MMX\n" |
|
|
|
"#define HAVE_SSE\n" |
|
|
|
"#define HAVE_SSE2\n" |
|
|
|
"#define HAVE_SSE3\n" |
|
|
|
"#define HAVE_SSE4_1\n" |
|
|
|
"#define HAVE_SSE4_2\n" |
|
|
|
"#define HAVE_SSE4A\n" |
|
|
|
"#define HAVE_MISALIGNSSE\n" |
|
|
|
"#define HAVE_128BITFPU\n" |
|
|
|
"#define HAVE_FASTMOVU\n" |
|
|
|
"#define HAVE_CFLUSH\n" |
|
|
|
"#define HAVE_AVX\n" |
|
|
|
"#define HAVE_AVX2\n" |
|
|
|
"#define HAVE_FMA3\n" |
|
|
|
"#define SLOCAL_BUFFER_SIZE\t20480\n" |
|
|
|
"#define DLOCAL_BUFFER_SIZE\t32768\n" |
|
|
|
"#define CLOCAL_BUFFER_SIZE\t16384\n" |
|
|
|
"#define ZLOCAL_BUFFER_SIZE\t12288\n") |
|
|
|
set(HAVE_SSE 1) |
|
|
|
set(HAVE_SSE2 1) |
|
|
|
set(HAVE_SSE3 1) |
|
|
|
set(HAVE_SSE4_1 1) |
|
|
|
set(HAVE_SSE4_2 1) |
|
|
|
set(HAVE_AVX 1) |
|
|
|
set(HAVE_AVX2 1) |
|
|
|
set(HAVE_FMA3 1) |
|
|
|
set(HAVE_SSE4A 1) |
|
|
|
set(HAVE_MISALIGNSSE 1) |
|
|
|
set(HAVE_128BITFPU 1) |
|
|
|
set(HAVE_FASTMOVU 1) |
|
|
|
set(HAVE_CFLUSH 1) |
|
|
|
set(SBGEMM_UNROLL_M 8) |
|
|
|
set(SBGEMM_UNROLL_N 4) |
|
|
|
set(SGEMM_UNROLL_M 8) |
|
|
|
set(SGEMM_UNROLL_N 4) |
|
|
|
set(DGEMM_UNROLL_M 4) |
|
|
|
set(DGEMM_UNROLL_N 8) |
|
|
|
set(CGEMM_UNROLL_M 8) |
|
|
|
set(CGEMM_UNROLL_N 2) |
|
|
|
set(ZGEMM_UNROLL_M 4) |
|
|
|
set(ZGEMM_UNROLL_N 2) |
|
|
|
set(CGEMM3M_UNROLL_M 8) |
|
|
|
set(CGEMM3M_UNROLL_N 4) |
|
|
|
set(ZGEMM3M_UNROLL_M 4) |
|
|
|
set(ZGEMM3M_UNROLL_N 4) |
|
|
|
elseif ("${TCORE}" STREQUAL "ARMV7") |
|
|
|
file(APPEND ${TARGET_CONF_TEMP} |
|
|
|
"#define L1_DATA_SIZE\t65536\n" |
|
|
|