ggml : optimize llamafile cpu matrix multiplication for ppc64le (llama/10156)

This change upstreams llamafile's cpu matrix
multiplication kernels for ppc64le using MMA
builtins for FP32 datatype.

This change results in a consistent 90%
improvement in input processing time, and 20%
to 80% improvement in output processing time,
across various batch sizes.

The patch is tested with Meta-Lllama-3-8B,
Mistral-7B, Llama-2-7B-chat-hf models on a
IBM POWER10 machine.

Signed-off-by: Amrita H S <redacted>

diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
index 6c5b816d2f5e7d18fb4c972a7fbe5f3ef70aa925..a05f8c505c49211d61509470ce4b86144438e9ef 100644 (file)
--- a/ggml/src/CMakeLists.txt
+++ b/ggml/src/CMakeLists.txt
@@ -1265,8 +1265,13 @@ elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LW
     endif()
 elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
     message(STATUS "PowerPC detected")
-    if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
-        list(APPEND ARCH_FLAGS -mcpu=powerpc64le)
+    execute_process(COMMAND bash -c "grep POWER10 /proc/cpuinfo | head -n 1"
+                   OUTPUT_VARIABLE POWER10_M)
+    string(FIND ${POWER10_M} "POWER10" substring_index)
+    if(${substring_index} GREATER_EQUAL 0)
+       list(APPEND ARCH_FLAGS -mcpu=power10)
+    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
+       list(APPEND ARCH_FLAGS -mcpu=powerpc64le)
     else()
         list(APPEND ARCH_FLAGS -mcpu=native -mtune=native)
         #TODO: Add  targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be)