202504 tmac

common/common.cpp

@@ -10,6 +10,7 @@
 #include "llama.h"
 
 #include <algorithm>
+#include <chrono>
 #include <cinttypes>
 #include <climits>
 #include <cmath>

ggml/CMakeLists.txt

@@ -208,6 +208,8 @@ set   (GGML_OPENCL_TARGET_VERSION "300" CACHE STRING
 # toolchain for vulkan-shaders-gen
 set   (GGML_VULKAN_SHADERS_GEN_TOOLCHAIN "" CACHE FILEPATH "ggml: toolchain file for vulkan-shaders-gen")
 
+option(GGML_TMAC                            "ggml: use TMAC"                                  OFF)
+
 # extra artifacts
 option(GGML_BUILD_TESTS    "ggml: build tests"    ${GGML_STANDALONE})
 option(GGML_BUILD_EXAMPLES "ggml: build examples" ${GGML_STANDALONE})
@@ -217,6 +219,9 @@ option(GGML_BUILD_EXAMPLES "ggml: build examples" ${GGML_STANDALONE})
 #
 
 set(CMAKE_C_STANDARD 11)
+if (GGML_TMAC)
+    set(CMAKE_C_STANDARD 17)
+endif()
 set(CMAKE_C_STANDARD_REQUIRED true)
 
 set(CMAKE_CXX_STANDARD 17)

ggml/include/ggml-cpu.h

@@ -57,6 +57,8 @@ extern "C" {
     GGML_BACKEND_API int                           ggml_threadpool_get_n_threads (struct ggml_threadpool * threadpool);
     GGML_BACKEND_API void                          ggml_threadpool_pause         (struct ggml_threadpool * threadpool);
     GGML_BACKEND_API void                          ggml_threadpool_resume        (struct ggml_threadpool * threadpool);
+    GGML_BACKEND_API void                          ggml_threadpool_atomic_store_explicit(struct ggml_threadpool * threadpool, int value);
+    GGML_BACKEND_API int                           ggml_threadpool_atomic_fetch_add_explicit(struct ggml_threadpool * threadpool, int value);
 
     // ggml_graph_plan() has to be called before ggml_graph_compute()
     // when plan.work_size > 0, caller must allocate memory for plan.work_data
@@ -120,12 +122,12 @@ extern "C" {
 
     GGML_BACKEND_API void ggml_cpu_init(void);
 
+    GGML_BACKEND_API void ggml_cpu_tmac_init(const char * fname);
+
     //
     // CPU backend
     //
 
-    GGML_BACKEND_API ggml_backend_t ggml_backend_cpu_init(void);
-
     GGML_BACKEND_API bool ggml_backend_is_cpu                (ggml_backend_t backend);
     GGML_BACKEND_API void ggml_backend_cpu_set_n_threads     (ggml_backend_t backend_cpu, int n_threads);
     GGML_BACKEND_API void ggml_backend_cpu_set_threadpool    (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);

ggml/include/ggml.h

@@ -388,7 +388,16 @@ extern "C" {
         // GGML_TYPE_IQ4_NL_4_4 = 36,
         // GGML_TYPE_IQ4_NL_4_8 = 37,
         // GGML_TYPE_IQ4_NL_8_8 = 38,
-        GGML_TYPE_COUNT   = 39,
+        GGML_TYPE_TMAC_BN_0         = 39,
+        GGML_TYPE_TMAC_W2G64_0      = 40,
+        GGML_TYPE_TMAC_W2G64_1      = 41,
+        GGML_TYPE_TMAC_W2G128_0     = 42,
+        GGML_TYPE_TMAC_W2G128_1     = 43,
+        GGML_TYPE_TMAC_W4G64_0      = 44,
+        GGML_TYPE_TMAC_W4G64_1      = 45,
+        GGML_TYPE_TMAC_W4G128_0     = 46,
+        GGML_TYPE_TMAC_W4G128_1     = 47,
+        GGML_TYPE_COUNT   = 48,
     };
 
     // precision

ggml/src/CMakeLists.txt

@@ -196,6 +196,7 @@ add_library(ggml-base
             ggml.c
             ggml-alloc.c
             ggml-backend.cpp
+            ggml-common.h
             ggml-opt.cpp
             ggml-threading.cpp
             ggml-threading.h

ggml/src/ggml-cpu/CMakeLists.txt

@@ -22,6 +22,14 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         ggml-cpu/amx/amx.h
         ggml-cpu/amx/mmq.cpp
         ggml-cpu/amx/mmq.h
+        ggml-cpu/tmac/tmac.cpp
+        ggml-cpu/tmac/tmac.h
+        ggml-cpu/tmac/lut_mul_mat.cpp
+        ggml-cpu/tmac/lut_mul_mat.h
+        ggml-cpu/tmac/lut_ctor.cpp
+        ggml-cpu/tmac/lut_ctor.h
+        ggml-cpu/tmac/tbl.cpp
+        ggml-cpu/tmac/tbl.h
         ggml-cpu/ggml-cpu-impl.h
         ggml-cpu/common.h
         ggml-cpu/binary-ops.h
@@ -72,6 +80,22 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                     ggml-cpu/llamafile/sgemm.h)
     endif()
 
+    if (GGML_TMAC)
+        target_compile_definitions(${GGML_CPU_NAME} PUBLIC GGML_USE_TMAC)
+        target_include_directories(${GGML_CPU_NAME} PUBLIC ggml-cpu/tmac)
+        get_target_property(cdefs ${GGML_CPU_NAME} COMPILE_DEFINITIONS)
+        message(STATUS "GGML_CPU_NAME: ${GGML_CPU_NAME} COMPILE_DEFINITIONS: ${cdefs}")
+
+        if ((NOT ${CMAKE_C_COMPILER_ID} MATCHES "Clang") OR
+            (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang"))
+            message(FATAL_ERROR "Clang is required for T-MAC compilation")
+        endif()
+
+        if (GGML_TMAC_RECHUNK)
+            target_compile_definitions(${GGML_CPU_NAME} PRIVATE TMAC_RECHUNK)
+        endif()
+    endif()
+
     if (GGML_CPU_HBM)
         find_library(memkind memkind REQUIRED)
 
@@ -145,6 +169,12 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                     list(APPEND ARCH_FLAGS -march=${GGML_CPU_ARM_ARCH})
                 endif()
             endif()
+            if (GGML_TMAC)
+                # ARM Windows with LLVM clang GNU interface
+                # We need fullfp16 for T-MAC
+                # TODO: check_cxx_source_compiles
+                list(APPEND ARCH_FLAGS -march=armv8.2a+fp16)
+            endif()
 
             # show enabled features
             if (CMAKE_HOST_SYSTEM_NAME STREQUAL "Windows")
@@ -181,7 +211,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             if (GGML_NATIVE)
                 include(ggml-cpu/cmake/FindSIMD.cmake)
             endif ()
-            if (GGML_AVX512)
+            # Can't use GGML_AVX512 with T-MAC and Clang for MSVC
+            # with error: conflicting types for '_m_prefetchw
+            if (GGML_AVX512 AND (NOT ${CMAKE_C_COMPILER_ID} MATCHES "Clang") AND (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang"))
                 list(APPEND ARCH_FLAGS /arch:AVX512)
                 # /arch:AVX512 includes: __AVX512F__, __AVX512CD__, __AVX512BW__, __AVX512DQ__, and __AVX512VL__
                 # MSVC has no compile-time flags enabling specific
@@ -323,6 +355,19 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 list(APPEND ARCH_FLAGS -mcpu=${GGML_CPU_POWERPC_CPUTYPE})
             endif()
         endif()
+    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64" AND GGML_TMAC)
+        # We need fullfp16 for T-MAC
+        # TODO: we need to simplify this logic through check_cxx_source_compiles or Presets?
+        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vmlaq_f32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8)
+        if (GGML_COMPILER_SUPPORT_MATMUL_INT8)
+            # Device with armv8.7a+ cpu, e.g., WSL on Surface Laptop 7
+            # based on arm64-windows-llvm.cmake
+            list(APPEND ARCH_FLAGS -march=armv8.7-a+fp16 -fvectorize -ffp-model=fast -fno-finite-math-only)
+            add_compile_definitions(__ARM_FEATURE_MATMUL_INT8)
+        else ()
+            # Jetson AGX Orin, Raspberry Pi 5
+            list(APPEND ARCH_FLAGS -march=armv8.2a+fp16)
+        endif ()
     elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
         message(STATUS "loongarch64 detected")
 

ggml/src/ggml-cpu/ggml-cpu.c

@@ -50,6 +50,10 @@
 #include "llamafile/sgemm.h"
 #endif
 
+#ifdef GGML_USE_TMAC
+#include "tmac.h"
+#endif
+
 #if defined(_MSC_VER)
 // disable "possible loss of data" to avoid hundreds of casts
 // we should just be careful :)
@@ -373,6 +377,51 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
+    [GGML_TYPE_TMAC_BN_0] = {
+        .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_f32,
+        .vec_dot_type             = GGML_TYPE_F32,
+        .nrows                    = 1,
+    },
+    [GGML_TYPE_TMAC_W2G64_0] = {
+        .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_f32,
+        .vec_dot_type             = GGML_TYPE_F32,
+        .nrows                    = 1,
+    },
+    [GGML_TYPE_TMAC_W2G64_1] = {
+        .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_f32,
+        .vec_dot_type             = GGML_TYPE_F32,
+        .nrows                    = 1,
+    },
+    [GGML_TYPE_TMAC_W2G128_0] = {
+        .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_f32,
+        .vec_dot_type             = GGML_TYPE_F32,
+        .nrows                    = 1,
+    },
+    [GGML_TYPE_TMAC_W2G128_1] = {
+        .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_f32,
+        .vec_dot_type             = GGML_TYPE_F32,
+        .nrows                    = 1,
+    },
+    [GGML_TYPE_TMAC_W4G64_0] = {
+        .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_f32,
+        .vec_dot_type             = GGML_TYPE_F32,
+        .nrows                    = 1,
+    },
+    [GGML_TYPE_TMAC_W4G64_1] = {
+        .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_f32,
+        .vec_dot_type             = GGML_TYPE_F32,
+        .nrows                    = 1,
+    },
+    [GGML_TYPE_TMAC_W4G128_0] = {
+        .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_f32,
+        .vec_dot_type             = GGML_TYPE_F32,
+        .nrows                    = 1,
+    },
+    [GGML_TYPE_TMAC_W4G128_1] = {
+        .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_f32,
+        .vec_dot_type             = GGML_TYPE_F32,
+        .nrows                    = 1,
+    },
 };
 
 const struct ggml_type_traits_cpu * ggml_get_type_traits_cpu(enum ggml_type type) {
@@ -2639,6 +2688,14 @@ void ggml_threadpool_resume(struct ggml_threadpool * threadpool) {
 #endif
 }
 
+void ggml_threadpool_atomic_store_explicit(struct ggml_threadpool * threadpool, int value) {
+    atomic_store_explicit(&threadpool->current_chunk, value, memory_order_relaxed);
+}
+
+int ggml_threadpool_atomic_fetch_add_explicit(struct ggml_threadpool * threadpool, int value) {
+    return (int)atomic_fetch_add_explicit(&threadpool->current_chunk, value, memory_order_relaxed);
+}
+
 struct ggml_cplan ggml_graph_plan(
           const struct ggml_cgraph * cgraph,
                                int   n_threads,
@@ -3406,6 +3463,10 @@ void ggml_cpu_init(void) {
         ggml_init_arm_arch_features();
 #endif
 
+#ifdef GGML_USE_TMAC
+        ggml_tmac_init();
+#endif
+
         is_first_call = false;
     }
 

ggml/src/ggml-cpu/ggml-cpu.cpp

@@ -5,6 +5,7 @@
 #include "ggml-cpu-traits.h"
 #include "ggml-impl.h"
 #include "amx/amx.h"
+#include "tmac/tmac.h"
 
 #include <cctype>
 #include <string>
@@ -43,6 +44,12 @@ std::vector<ggml_backend_buffer_type_t>& ggml_backend_cpu_get_extra_buffers_type
         }
 #endif
 
+#ifdef GGML_USE_TMAC
+        if (ggml_backend_tmac_buffer_type()) {
+            bufts.push_back(ggml_backend_tmac_buffer_type());
+        }
+#endif
+
 #ifdef GGML_USE_CPU_KLEIDIAI
         if (ggml_backend_cpu_kleidiai_buffer_type()) {
             bufts.push_back(ggml_backend_cpu_kleidiai_buffer_type());

ggml/src/ggml-cpu/ops.cpp

@@ -4978,6 +4978,15 @@ void ggml_compute_forward_clamp(
         case GGML_TYPE_I32:
         case GGML_TYPE_I64:
         case GGML_TYPE_F64:
+        case GGML_TYPE_TMAC_BN_0:
+        case GGML_TYPE_TMAC_W2G64_0:
+        case GGML_TYPE_TMAC_W2G64_1:
+        case GGML_TYPE_TMAC_W2G128_0:
+        case GGML_TYPE_TMAC_W2G128_1:
+        case GGML_TYPE_TMAC_W4G64_0:
+        case GGML_TYPE_TMAC_W4G64_1:
+        case GGML_TYPE_TMAC_W4G128_0:
+        case GGML_TYPE_TMAC_W4G128_1:
         case GGML_TYPE_COUNT:
             {
                 GGML_ABORT("fatal error");