some cleanup with tinyblas backend

Djip007 · Djip007 · commit fd08ab83ca99 · 2024-11-16T22:41:43.000+01:00
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -84,8 +84,8 @@ set(GGML_ALL_WARNINGS       ${LLAMA_ALL_WARNINGS})
 set(GGML_FATAL_WARNINGS     ${LLAMA_FATAL_WARNINGS})
 
 # change the default for these ggml options
-if (NOT DEFINED GGML_LLAMAFILE)
-    set(GGML_LLAMAFILE_DEFAULT ON)
+if (NOT DEFINED GGML_TINYBLAS)
+    set(GGML_TINYBLAS ON)
 endif()
 
 if (NOT DEFINED GGML_AMX)
diff --git a/docs/android.md b/docs/android.md
@@ -45,7 +45,7 @@ $ cmake \
   -DCMAKE_C_FLAGS="-march=armv8.7a" \
   -DCMAKE_CXX_FLAGS="-march=armv8.7a" \
   -DGGML_OPENMP=OFF \
-  -DGGML_LLAMAFILE=OFF \
+  -DGGML_TINYBLAS=OFF \
   -B build-android
 ```
 
diff --git a/docs/build.md b/docs/build.md
@@ -42,7 +42,7 @@ In order to build llama.cpp you have four different options.
 
   **Notes**:
 
-    - For `Q4_0_4_4` quantization type build, add the `-DGGML_LLAMAFILE=OFF` cmake option. For example, use `cmake -B build -DGGML_LLAMAFILE=OFF`.
+    - For `Q4_0_4_4` quantization type build, add the `-DGGML_TINYBLAS=OFF` cmake option. For example, use `cmake -B build -DGGML_TINYBLAS=OFF`.
     - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `cmake --build build --config Release -j 8` will run 8 jobs in parallel.
     - For faster repeated compilation, install [ccache](https://ccache.dev/).
     - For debug builds, there are two cases:
@@ -393,4 +393,4 @@ To read documentation for how to build on Android, [click here](./android.md)
 
 Llama.cpp includes a set of optimized mulmat kernels for the Arm architecture, leveraging Arm® Neon™, int8mm and SVE instructions. These kernels are enabled at build time through the appropriate compiler cpu-type flags, such as `-DCMAKE_C_FLAGS=-march=armv8.2a+i8mm+sve`. Note that these optimized kernels require the model to be quantized into one of the formats: `Q4_0_4_4` (Arm Neon), `Q4_0_4_8` (int8mm) or `Q4_0_8_8` (SVE). The SVE mulmat kernel specifically requires a vector width of 256 bits. When running on devices with a different vector width, it is recommended to use the `Q4_0_4_8` (int8mm) or `Q4_0_4_4` (Arm Neon) formats for better performance. Refer to [examples/quantize/README.md](../examples/quantize/README.md) for more information on the quantization formats.
 
-To support `Q4_0_4_4`, you must build with `GGML_NO_LLAMAFILE=1` (`make`) or `-DGGML_LLAMAFILE=OFF` (`cmake`).
+To support `Q4_0_4_4`, you must build with `GGML_NO_LLAMAFILE=1` (`make`) or `-DGGML_TINYBLAS=OFF` (`cmake`).
diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt
@@ -57,8 +57,8 @@ else()
 endif()
 
 # defaults
-if (NOT GGML_LLAMAFILE_DEFAULT)
-    set(GGML_LLAMAFILE_DEFAULT OFF)
+if (NOT GGML_TINYBLAS_DEFAULT)
+    set(GGML_TINYBLAS_DEFAULT OFF)
 endif()
 
 if (NOT GGML_CUDA_GRAPHS_DEFAULT)
@@ -124,8 +124,7 @@ option(GGML_ACCELERATE                      "ggml: enable Accelerate framework"
 option(GGML_BLAS                            "ggml: use BLAS"                                  ${GGML_BLAS_DEFAULT})
 set(GGML_BLAS_VENDOR ${GGML_BLAS_VENDOR_DEFAULT} CACHE STRING
                                             "ggml: BLAS library vendor")
-option(GGML_LLAMAFILE                       "ggml: use LLAMAFILE"                             ${GGML_LLAMAFILE_DEFAULT})
-
+option(GGML_TINYBLAS                        "ggml: use TINYBLAS"                              OFF)
 option(GGML_CUDA                            "ggml: use CUDA"                                  OFF)
 option(GGML_MUSA                            "ggml: use MUSA"                                  OFF)
 option(GGML_CUDA_FORCE_DMMV                 "ggml: use dmmv instead of mmvq CUDA kernels"     OFF)
@@ -231,6 +230,7 @@ set(GGML_PUBLIC_HEADERS
     include/ggml-metal.h
     include/ggml-rpc.h
     include/ggml-sycl.h
+    include/ggml-tinyblas.h
     include/ggml-vulkan.h)
 
 set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
@@ -256,6 +256,7 @@ ggml_add_backend(Kompute)
 ggml_add_backend(METAL)
 ggml_add_backend(RPC)
 ggml_add_backend(SYCL)
+ggml_add_backend(TINYBLAS)
 ggml_add_backend(Vulkan)
 ggml_add_backend(MUSA)
 
diff --git a/ggml/src/ggml-backend-reg.cpp b/ggml/src/ggml-backend-reg.cpp
@@ -91,10 +91,12 @@ struct ggml_backend_registry {
             return;
         }
 
-#ifndef NDEBUG
-        GGML_LOG_DEBUG("%s: registered backend %s (%zu devices)\n",
+        GGML_LOG_INFO("%s: registered backend %s (%zu devices)\n",
             __func__, ggml_backend_reg_name(reg), ggml_backend_reg_dev_count(reg));
-#endif
+//#ifndef NDEBUG
+//        GGML_LOG_DEBUG("%s: registered backend %s (%zu devices)\n",
+//            __func__, ggml_backend_reg_name(reg), ggml_backend_reg_dev_count(reg));
+//#endif
         backends.push_back(reg);
         for (size_t i = 0; i < ggml_backend_reg_dev_count(reg); i++) {
             register_device(ggml_backend_reg_dev_get(reg, i));
diff --git a/ggml/src/ggml-common.h b/ggml/src/ggml-common.h
@@ -6,7 +6,20 @@
 typedef uint16_t ggml_half;
 typedef uint32_t ggml_half2;
 
-#define GGML_COMMON_AGGR
+#define GGML_COMMON_AGGR_U
+#define GGML_COMMON_AGGR_S
+
+#define GGML_COMMON_DECL
+#elif defined(GGML_COMMON_DECL_CPP)
+#include <cstdint>
+
+typedef uint16_t ggml_half;
+typedef uint32_t ggml_half2;
+
+// std-c++ allow anonymous unions but some compiler warn on it
+#define GGML_COMMON_AGGR_U data
+// std-c++ do not allow it.
+#define GGML_COMMON_AGGR_S data
 
 #define GGML_COMMON_DECL
 #elif defined(GGML_COMMON_DECL_METAL)
@@ -15,7 +28,8 @@ typedef uint32_t ggml_half2;
 typedef half  ggml_half;
 typedef half2 ggml_half2;
 
-#define GGML_COMMON_AGGR
+#define GGML_COMMON_AGGR_U
+#define GGML_COMMON_AGGR_S
 
 #define GGML_COMMON_DECL
 #elif defined(GGML_COMMON_DECL_CUDA)
@@ -29,7 +43,8 @@ typedef half2 ggml_half2;
 typedef half  ggml_half;
 typedef half2 ggml_half2;
 
-#define GGML_COMMON_AGGR data
+#define GGML_COMMON_AGGR_U
+#define GGML_COMMON_AGGR_S data
 
 #define GGML_COMMON_DECL
 #elif defined(GGML_COMMON_DECL_HIP)
@@ -39,7 +54,8 @@ typedef half2 ggml_half2;
 typedef half  ggml_half;
 typedef half2 ggml_half2;
 
-#define GGML_COMMON_AGGR data
+#define GGML_COMMON_AGGR_U
+#define GGML_COMMON_AGGR_S data
 
 #define GGML_COMMON_DECL
 #elif defined(GGML_COMMON_DECL_SYCL)
@@ -49,7 +65,8 @@ typedef half2 ggml_half2;
 typedef sycl::half  ggml_half;
 typedef sycl::half2 ggml_half2;
 
-#define GGML_COMMON_AGGR data
+#define GGML_COMMON_AGGR_U
+#define GGML_COMMON_AGGR_S data
 
 #define GGML_COMMON_DECL
 #endif
@@ -154,9 +171,9 @@ typedef struct {
         struct {
             ggml_half d; // delta
             ggml_half m; // min
-        } GGML_COMMON_AGGR;
+        } GGML_COMMON_AGGR_S;
         ggml_half2 dm;
-    };
+    } GGML_COMMON_AGGR_U;
     uint8_t qs[QK4_1 / 2]; // nibbles / quants
 } block_q4_1;
 static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_half) + QK4_1 / 2, "wrong q4_1 block size/padding");
@@ -175,9 +192,9 @@ typedef struct {
         struct {
             ggml_half d; // delta
             ggml_half m; // min
-        } GGML_COMMON_AGGR;
+        } GGML_COMMON_AGGR_S;
         ggml_half2 dm;
-    };
+    } GGML_COMMON_AGGR_U;
     uint8_t qh[4];         // 5-th bit of quants
     uint8_t qs[QK5_1 / 2]; // nibbles / quants
 } block_q5_1;
@@ -196,9 +213,9 @@ typedef struct {
         struct {
             ggml_half d; // delta
             ggml_half s; // d * sum(qs[i])
-        } GGML_COMMON_AGGR;
+        } GGML_COMMON_AGGR_S;
         ggml_half2 ds;
-    };
+    } GGML_COMMON_AGGR_U;
     int8_t qs[QK8_1]; // quants
 } block_q8_1;
 static_assert(sizeof(block_q8_1) == 2*sizeof(ggml_half) + QK8_1, "wrong q8_1 block size/padding");
@@ -261,9 +278,9 @@ typedef struct {
         struct {
             ggml_half d;    // super-block scale for quantized scales
             ggml_half dmin; // super-block scale for quantized mins
-        } GGML_COMMON_AGGR;
+        } GGML_COMMON_AGGR_S;
         ggml_half2 dm;
-    };
+    } GGML_COMMON_AGGR_U;
 } block_q2_K;
 static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_half) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
 
@@ -288,9 +305,9 @@ typedef struct {
         struct {
             ggml_half d;    // super-block scale for quantized scales
             ggml_half dmin; // super-block scale for quantized mins
-        } GGML_COMMON_AGGR;
+        } GGML_COMMON_AGGR_S;
         ggml_half2 dm;
-    };
+    } GGML_COMMON_AGGR_U;
     uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
     uint8_t qs[QK_K/2];           // 4--bit quants
 } block_q4_K;
@@ -305,9 +322,9 @@ typedef struct {
         struct {
             ggml_half d;    // super-block scale for quantized scales
             ggml_half dmin; // super-block scale for quantized mins
-        } GGML_COMMON_AGGR;
+        } GGML_COMMON_AGGR_S;
         ggml_half2 dm;
-    };
+    } GGML_COMMON_AGGR_U;
     uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
     uint8_t qh[QK_K/8];           // quants, high bit
     uint8_t qs[QK_K/2];           // quants, low 4 bits
@@ -431,6 +448,13 @@ static_assert(sizeof(block_iq4_xs) == sizeof(ggml_half) + sizeof(uint16_t) + QK_
 #define GGML_TABLE_BEGIN(type, name, size) static const type name[size] = {
 #define GGML_TABLE_END() };
 
+#define GGML_COMMON_IMPL
+#elif defined(GGML_COMMON_IMPL_CPP)
+#include <cstdint>
+
+#define GGML_TABLE_BEGIN(type, name, size) static const type name[size] = {
+#define GGML_TABLE_END() };
+
 #define GGML_COMMON_IMPL
 #elif defined(GGML_COMMON_IMPL_METAL)
 #include <metal_stdlib>
diff --git a/ggml/src/ggml-cpu/CMakeLists.txt b/ggml/src/ggml-cpu/CMakeLists.txt
@@ -44,16 +44,6 @@ if (GGML_OPENMP)
     endif()
 endif()
 
-if (GGML_LLAMAFILE)
-    message(STATUS "Using llamafile")
-
-    add_compile_definitions(GGML_USE_LLAMAFILE)
-
-    target_sources(ggml-cpu PRIVATE
-                    llamafile/sgemm.cpp
-                    llamafile/sgemm.h)
-endif()
-
 if (GGML_CPU_HBM)
     find_library(memkind memkind REQUIRED)
 
diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c
@@ -39,14 +39,6 @@
 #include <omp.h>
 #endif
 
-#if defined(__ARM_FEATURE_SVE) || defined(__ARM_FEATURE_MATMUL_INT8)
-#undef GGML_USE_LLAMAFILE
-#endif
-
-#ifdef GGML_USE_LLAMAFILE
-#include "llamafile/sgemm.h"
-#endif
-
 #if defined(_MSC_VER)
 // disable "possible loss of data" to avoid hundreds of casts
 // we should just be careful :)
@@ -7466,33 +7458,6 @@ static void ggml_compute_forward_mul_mat(
     // nb01 >= nb00 - src0 is not transposed
     //   compute by src0 rows
 
-#if GGML_USE_LLAMAFILE
-    // broadcast factors
-    const int64_t r2 = ne12 / ne02;
-    const int64_t r3 = ne13 / ne03;
-
-    const bool src1_cont = ggml_is_contiguous(src1);
-
-    if (src1_cont) {
-        for (int64_t i13 = 0; i13 < ne13; i13++)
-            for (int64_t i12 = 0; i12 < ne12; i12++)
-                if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(type),
-                                     (const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
-                                     nb01/ggml_type_size(type),
-                                     (const char *)src1->data + i12*nb12 + i13*nb13,
-                                     nb11/ggml_type_size(src1->type),
-                                     (char *)dst->data + i12*nb2 + i13*nb3,
-                                     nb1/ggml_type_size(dst->type),
-                                     ith, nth,
-                                     type,
-                                     src1->type,
-                                     dst->type))
-                    goto UseGgmlGemm1;
-        return;
-    }
-UseGgmlGemm1:;
-#endif
-
     if (src1->type != vec_dot_type) {
         char * wdata = params->wdata;
 
@@ -7530,30 +7495,6 @@ UseGgmlGemm1:;
 
     ggml_barrier(params->threadpool);
 
-#if GGML_USE_LLAMAFILE
-    if (src1->type != vec_dot_type) {
-        const void* wdata = (src1->type == vec_dot_type) ? src1->data : params->wdata;
-        const size_t row_size = ggml_row_size(vec_dot_type, ne10);
-
-        for (int64_t i13 = 0; i13 < ne13; i13++)
-            for (int64_t i12 = 0; i12 < ne12; i12++)
-                if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(type),
-                                     (const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
-                                     nb01/ggml_type_size(type),
-                                     (const char *)wdata + (i12*ne11 + i13*ne12*ne11)*row_size,
-                                     row_size/ggml_type_size(vec_dot_type),
-                                     (char *)dst->data + i12*nb2 + i13*nb3,
-                                     nb1/ggml_type_size(dst->type),
-                                     ith, nth,
-                                     type,
-                                     vec_dot_type,
-                                     dst->type))
-                    goto UseGgmlGemm2;
-        return;
-    }
-UseGgmlGemm2:;
-#endif
-
     // This is the size of the first dimension of the result, so we can iterate that way. (see the ASSERT above, these are the same numbers)
     const int64_t nr0 = ne0;
 
diff --git a/ggml/src/ggml-tinyblas/CMakeLists.txt b/ggml/src/ggml-tinyblas/CMakeLists.txt
@@ -1,3 +1,5 @@
+message(STATUS "Using TINYBLAS")
+
 add_library(ggml-tinyblas
             ggml-tinyblas.cpp
             )
@@ -225,6 +227,10 @@ endif()
 target_compile_options(ggml-tinyblas PRIVATE "$<$<COMPILE_LANGUAGE:CXX>:${ARCH_FLAGS}>")
 target_compile_options(ggml-tinyblas PRIVATE "$<$<COMPILE_LANGUAGE:C>:${ARCH_FLAGS}>")
 
+#set_source_files_properties( ${GGML_SOURCES_FP8} PROPERTIES CXX_STANDARD 17)
+#set_source_files_properties( ${GGML_SOURCES_FP8} PROPERTIES COMPILE_FLAGS "-std=c++17")
+target_compile_features   (ggml-tinyblas PRIVATE cxx_std_17)
+
 if (EMSCRIPTEN)
     set_target_properties(ggml-tinyblas PROPERTIES COMPILE_FLAGS "-msimd128")
 endif()
diff --git a/ggml/src/ggml-tinyblas/ggml-tinyblas.cpp b/ggml/src/ggml-tinyblas/ggml-tinyblas.cpp
diff --git a/ggml/src/ggml-tinyblas/sgemm.cpp b/ggml/src/ggml-tinyblas/sgemm.cpp
diff --git a/ggml/src/ggml-tinyblas/sgemm.h b/ggml/src/ggml-tinyblas/sgemm.h
