cmake build update

Djip007 · Djip007 · commit 0723ac5e6562 · 2024-10-26T08:49:17.000+02:00
and correct warning
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -88,6 +88,10 @@ if (NOT DEFINED GGML_LLAMAFILE)
     set(GGML_LLAMAFILE_DEFAULT ON)
 endif()
 
+if (NOT DEFINED GGML_OPENMP_SIMD)
+    set(GGML_OPENMP_SIMD_DEFAULT ON)
+endif()
+
 if (NOT DEFINED GGML_AMX)
     set(GGML_AMX ON)
 endif()
diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp
@@ -1846,9 +1846,9 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {
                 total_seconds = total_seconds % (60*60);
             }
             LOG("%.2f minutes\n", total_seconds / 60.0);
+            LOG("\n");
+            LOG("chunk             PPL               ln(PPL(Q)/PPL(base))          KL Divergence              Δp RMS            Same top p\n");
         }
-        LOG("\n");
-        LOG("chunk             PPL               ln(PPL(Q)/PPL(base))          KL Divergence              Δp RMS            Same top p\n");
 
         const int first = n_ctx/2;
         const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx);
diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt
@@ -61,6 +61,10 @@ if (NOT GGML_LLAMAFILE_DEFAULT)
     set(GGML_LLAMAFILE_DEFAULT OFF)
 endif()
 
+if (NOT GGML_OPENMP_SIMD_DEFAULT)
+    set(GGML_OPENMP_SIMD_DEFAULT OFF)
+endif()
+
 if (NOT GGML_CUDA_GRAPHS_DEFAULT)
     set(GGML_CUDA_GRAPHS_DEFAULT OFF)
 endif()
@@ -109,6 +113,7 @@ endif()
 option(GGML_LASX        "ggml: enable lasx"             ON)
 option(GGML_LSX         "ggml: enable lsx"              ON)
 option(GGML_SVE         "ggml: enable SVE"              OFF)
+option(GGML_OPENMP_SIMD "ggml: enable OPENMP_SIMD"      ${GGML_OPENMP_SIMD_DEFAULT})
 
 if (WIN32)
     set(GGML_WIN_VER "0x602" CACHE STRING "ggml: Windows Version")
@@ -178,11 +183,11 @@ option(GGML_BUILD_EXAMPLES "ggml: build examples" ${GGML_STANDALONE})
 set(CMAKE_C_STANDARD 11)
 set(CMAKE_C_STANDARD_REQUIRED true)
 
-if (GGML_SYCL)
+#if (GGML_SYCL)
     set(CMAKE_CXX_STANDARD 17)
-else()
-    set(CMAKE_CXX_STANDARD 11)
-endif()
+#else()
+#    set(CMAKE_CXX_STANDARD 11)
+#endif()
 set(CMAKE_CXX_STANDARD_REQUIRED true)
 
 set(THREADS_PREFER_PTHREAD_FLAG ON)
diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h
@@ -394,7 +394,7 @@ extern "C" {
         GGML_TYPE_E4M3    = 37,
         GGML_TYPE_E4M3_Q  = 38,
         GGML_TYPE_E3M4_Q  = 39,
-        // E5M6 => 12 bits vs 16 bits for BF16 = E8M7 / FP16 = E5M10  
+        // E5M6 => 12 bits vs 16 bits for BF16 = E8M7 / FP16 = E5M10
         GGML_TYPE_COUNT,
     };
 
diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
@@ -154,6 +154,7 @@ if (GGML_MUSA)
 endif()
 
 if (GGML_OPENMP)
+    set(OpenMP_RUNTIME_MSVC "experimental")
     find_package(OpenMP)
     if (OpenMP_FOUND)
         message(STATUS "OpenMP found")
@@ -171,6 +172,18 @@ if (GGML_OPENMP)
     endif()
 endif()
 
+if (GGML_OPENMP_SIMD)
+    check_cxx_compiler_flag("-fopenmp-simd" SUPPORTS_OPENMP_SIMD)
+    if(SUPPORTS_OPENMP_SIMD)
+        # OpenMP_RUNTIME_MSVC=experimental / if (MSVC)
+        message(STATUS "Using openmp_simd.")
+        add_compile_definitions(GGML_USE_OPENMP_SIMD)
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fopenmp-simd")
+    else()
+        message(FATAL_ERROR, "C++ compiler lacks OPENMP_SIMD support.")
+    endif()
+endif()
+
 if (GGML_BLAS)
     if (GGML_STATIC)
         set(BLA_STATIC ON)
@@ -1360,6 +1373,10 @@ endif()
 # libraries
 #
 
+# FP8
+file(GLOB   GGML_HEADERS_FP8 "ggml-fp8.h")
+file(GLOB   GGML_SOURCES_FP8 "ggml-fp8.cpp")
+
 # ggml
 
 add_library(ggml
@@ -1384,6 +1401,7 @@ add_library(ggml
             ${GGML_SOURCES_AMX}       ${GGML_HEADERS_AMX}
             ${GGML_SOURCES_CANN}      ${GGML_HEADERS_CANN}
             ggml-aarch64.c            ggml-aarch64.h
+            ${GGML_SOURCES_FP8}       ${GGML_HEADERS_FP8}
             )
 
 if (EMSCRIPTEN)
diff --git a/ggml/src/ggml-common.h b/ggml/src/ggml-common.h
@@ -8,6 +8,15 @@ typedef uint32_t ggml_half2;
 
 #define GGML_COMMON_AGGR
 
+#define GGML_COMMON_DECL
+#elif defined(GGML_COMMON_DECL_CPP)
+#include <cstdint>
+
+typedef uint16_t ggml_half;
+typedef uint32_t ggml_half2;
+
+#define GGML_COMMON_AGGR data
+
 #define GGML_COMMON_DECL
 #elif defined(GGML_COMMON_DECL_METAL)
 #include <metal_stdlib>
@@ -419,22 +428,7 @@ typedef struct {
 static_assert(sizeof(block_iq4_xs) == sizeof(ggml_half) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding");
 
 // the fp8 types.
-typedef uint8_t ggml_e5m2;
-typedef uint8_t ggml_e4m3;
-typedef uint8_t ggml_e3m4;
-
-// fp8 with bloc delta => 8.125 bpw
-typedef struct {
-    float d;  // delta
-    ggml_e4m3 qs[QK_K];
-} block_e4m3_q;
-static_assert(sizeof(block_e4m3_q) == sizeof(float) + QK_K, "wrong block_e4m3_q block size/padding");
-
-typedef struct {
-    float d;  // delta
-    ggml_e3m4 qs[QK_K];
-} block_e3m4_q;
-static_assert(sizeof(block_e3m4_q) == sizeof(float) + QK_K, "wrong block_e3m4_q block size/padding");
+#include "ggml-fp8.h"
 
 #endif // GGML_COMMON_DECL
 #endif // GGML_COMMON_DECL
@@ -449,6 +443,13 @@ static_assert(sizeof(block_e3m4_q) == sizeof(float) + QK_K, "wrong block_e3m4_q
 #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-fp8.cpp b/ggml/src/ggml-fp8.cpp
@@ -1,22 +1,26 @@
-#define GGML_COMMON_IMPL_C
-#include "ggml-common.h"
+// #include <iostream>
+#include <cstdint>
+#include <cassert>
+#include <cmath>
 
 #include "ggml-fp8.h"
 
-#include <cassert>
-
 /*
+make clean
+make -j8
 # ./llama-quantize --output-tensor-type fp8_e3m4_q ~/LLM/Mistral-Nemo-Instruct-2407.BF16.gguf ~/LLM/Mistral-Nemo-Instruct-2407.E3M4_Q.gguf E3M4_Q
 ./llama-quantize ~/LLM/Mistral-Nemo-Instruct-2407.BF16.gguf ~/LLM/Mistral-Nemo-Instruct-2407.E3M4_Q.gguf E3M4_Q
 ./llama-cli -c 1024 -m ~/LLM/Mistral-Nemo-Instruct-2407.E3M4_Q.gguf -p "[INST]bonjour a tu un nom. je ne sais pas comment t'appeler. Si tu n'en as pas je peux t'appeler TINTIN[/INST]" -s 42
-# ./llama-perplexity -f ~/LLM/wikitext-2-raw/wiki.test.raw  -s 31337 -m ~/LLM/Mistral-Nemo-Instruct-2407.E3M4_Q.gguf
 ./llama-perplexity --kl-divergence-base ~/LLM/Mistral-Nemo-Instruct-2407.BF16.kld --kl-divergence -s 31337 -m ~/LLM/Mistral-Nemo-Instruct-2407.E3M4_Q.gguf
 
-*/
+rm -rf build
+cmake -B build
+cmake --build build --config Release
+./build/bin/llama-quantize ~/LLM/Mistral-Nemo-Instruct-2407.BF16.gguf ~/LLM/Mistral-Nemo-Instruct-2407.E3M4_Q.gguf E3M4_Q
+./build/bin/llama-cli -c 1024 -m ~/LLM/Mistral-Nemo-Instruct-2407.E3M4_Q.gguf -p "[INST]bonjour a tu un nom. je ne sais pas comment t'appeler. Si tu n'en as pas je peux t'appeler TINTIN[/INST]" -s 42
+./build/bin/llama-perplexity --kl-divergence-base ~/LLM/Mistral-Nemo-Instruct-2407.BF16.kld --kl-divergence -s 31337 -m ~/LLM/Mistral-Nemo-Instruct-2407.E3M4_Q.gguf
 
-#include <iostream>
-#include <cstdint>
-#include <immintrin.h>
+*/
 
 template<int N> constexpr float EXP2() {
     if constexpr (N==0) return 1;
@@ -188,29 +192,29 @@ void ggml_fp32_to_e4m3_row_ref(const float * GGML_RESTRICT x, ggml_e4m3_t * GGML
 }
 
 void dequantize_row_e4m3_q(const block_e4m3_q * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
-    assert(k % QK_K == 0);
-    conv(reinterpret_cast<const bloc_fp8<4, QK_K>*>(x), y, k);
+    assert(k % FP8_QK == 0);
+    conv(reinterpret_cast<const bloc_fp8<4, FP8_QK>*>(x), y, k);
 }
 void quantize_row_e4m3_q(const float * GGML_RESTRICT x, block_e4m3_q * GGML_RESTRICT y, int64_t k) {
-    assert(k % QK_K == 0);
-    conv(x, reinterpret_cast<bloc_fp8<4, QK_K>*>(y), k);
+    assert(k % FP8_QK == 0);
+    conv(x, reinterpret_cast<bloc_fp8<4, FP8_QK>*>(y), k);
 }
 void quantize_row_e4m3_q_ref(const float * GGML_RESTRICT x, block_e4m3_q * GGML_RESTRICT y, int64_t k) {
-    assert(k % QK_K == 0);
-    conv(x, reinterpret_cast<bloc_fp8<4, QK_K>*>(y), k);
+    assert(k % FP8_QK == 0);
+    conv(x, reinterpret_cast<bloc_fp8<4, FP8_QK>*>(y), k);
 }
 
 void dequantize_row_e3m4_q(const block_e3m4_q * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
-    assert(k % QK_K == 0);
-    conv(reinterpret_cast<const bloc_fp8<3, QK_K>*>(x), y, k);
+    assert(k % FP8_QK == 0);
+    conv(reinterpret_cast<const bloc_fp8<3, FP8_QK>*>(x), y, k);
 }
 void quantize_row_e3m4_q(const float * GGML_RESTRICT x, block_e3m4_q * GGML_RESTRICT y, int64_t k) {
-    assert(k % QK_K == 0);
-    conv(x, reinterpret_cast<bloc_fp8<3, QK_K>*>(y), k);
+    assert(k % FP8_QK == 0);
+    conv(x, reinterpret_cast<bloc_fp8<3, FP8_QK>*>(y), k);
 }
 void quantize_row_e3m4_q_ref(const float * GGML_RESTRICT x, block_e3m4_q * GGML_RESTRICT y, int64_t k) {
-    assert(k % QK_K == 0);
-    conv(x, reinterpret_cast<bloc_fp8<3, QK_K>*>(y), k);
+    assert(k % FP8_QK == 0);
+    conv(x, reinterpret_cast<bloc_fp8<3, FP8_QK>*>(y), k);
 }
 
 // the dot product for FP8 weight
@@ -238,7 +242,7 @@ void ggml_vec_dot_e4m3_q(int n, float * GGML_RESTRICT s, size_t bs, const block_
     GGML_UNUSED(bx);
     GGML_UNUSED(by);
     GGML_UNUSED(bs);
-    *s = dot(reinterpret_cast<const bloc_fp8<4, QK_K>*>(vx), vy, n);
+    *s = dot(reinterpret_cast<const bloc_fp8<4, FP8_QK>*>(vx), vy, n);
 }
 
 void ggml_vec_dot_e3m4_q(int n, float * GGML_RESTRICT s, size_t bs, const block_e3m4_q * GGML_RESTRICT vx, size_t bx, const float * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -247,5 +251,5 @@ void ggml_vec_dot_e3m4_q(int n, float * GGML_RESTRICT s, size_t bs, const block_
     GGML_UNUSED(bx);
     GGML_UNUSED(by);
     GGML_UNUSED(bs);
-    *s = dot(reinterpret_cast<const bloc_fp8<3, QK_K>*>(vx), vy, n);
+    *s = dot(reinterpret_cast<const bloc_fp8<3, FP8_QK>*>(vx), vy, n);
 }
diff --git a/ggml/src/ggml-fp8.h b/ggml/src/ggml-fp8.h
@@ -1,41 +1,51 @@
 #pragma once
-
-#define GGML_COMMON_DECL_C
-#include "ggml-common.h"
-
 #include "ggml.h"
 
-// les definitions / converstion FP8 <=> FP32
 #ifdef  __cplusplus
 extern "C" {
 #endif
 
+#define FP8_QK 256
+
     typedef struct { uint8_t bits; } ggml_e5m2_t;
     typedef struct { uint8_t bits; } ggml_e4m3_t;
     typedef struct { uint8_t bits; } ggml_e3m4_t;
 
-    void ggml_e5m2_to_fp32_row(const ggml_e5m2_t * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-    void ggml_fp32_to_e5m2_row(const float * GGML_RESTRICT x, ggml_e5m2_t * GGML_RESTRICT y, int64_t k);
-    void ggml_fp32_to_e5m2_row_ref(const float * GGML_RESTRICT x, ggml_e5m2_t * GGML_RESTRICT y, int64_t k);
+    // fp8 with bloc delta => 8.125 bpw
+    typedef struct {
+        float d;  // delta
+        uint8_t qs[FP8_QK];
+    } block_e4m3_q;
+    static_assert(sizeof(block_e4m3_q) == sizeof(float) + FP8_QK, "wrong block_e4m3_q block size/padding");
+
+    typedef struct {
+        float d;  // delta
+        uint8_t qs[FP8_QK];
+    } block_e3m4_q;
+    static_assert(sizeof(block_e3m4_q) == sizeof(float) + FP8_QK, "wrong block_e3m4_q block size/padding");
+
+    GGML_API void ggml_e5m2_to_fp32_row(const ggml_e5m2_t * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+    GGML_API void ggml_fp32_to_e5m2_row(const float * GGML_RESTRICT x, ggml_e5m2_t * GGML_RESTRICT y, int64_t k);
+    GGML_API void ggml_fp32_to_e5m2_row_ref(const float * GGML_RESTRICT x, ggml_e5m2_t * GGML_RESTRICT y, int64_t k);
 
-    void ggml_e4m3_to_fp32_row(const ggml_e4m3_t * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-    void ggml_fp32_to_e4m3_row(const float * GGML_RESTRICT x, ggml_e4m3_t * GGML_RESTRICT y, int64_t k);
-    void ggml_fp32_to_e4m3_row_ref(const float * GGML_RESTRICT x, ggml_e4m3_t * GGML_RESTRICT y, int64_t k);
+    GGML_API void ggml_e4m3_to_fp32_row(const ggml_e4m3_t * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+    GGML_API void ggml_fp32_to_e4m3_row(const float * GGML_RESTRICT x, ggml_e4m3_t * GGML_RESTRICT y, int64_t k);
+    GGML_API void ggml_fp32_to_e4m3_row_ref(const float * GGML_RESTRICT x, ggml_e4m3_t * GGML_RESTRICT y, int64_t k);
 
-    void dequantize_row_e4m3_q(const block_e4m3_q * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-    void quantize_row_e4m3_q(const float * GGML_RESTRICT x, block_e4m3_q * GGML_RESTRICT y, int64_t k);
-    void quantize_row_e4m3_q_ref(const float * GGML_RESTRICT x, block_e4m3_q * GGML_RESTRICT y, int64_t k);
+    GGML_API void dequantize_row_e4m3_q(const block_e4m3_q * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+    GGML_API void quantize_row_e4m3_q(const float * GGML_RESTRICT x, block_e4m3_q * GGML_RESTRICT y, int64_t k);
+    GGML_API void quantize_row_e4m3_q_ref(const float * GGML_RESTRICT x, block_e4m3_q * GGML_RESTRICT y, int64_t k);
 
-    void dequantize_row_e3m4_q(const block_e3m4_q * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-    void quantize_row_e3m4_q(const float * GGML_RESTRICT x, block_e3m4_q * GGML_RESTRICT y, int64_t k);
-    void quantize_row_e3m4_q_ref(const float * GGML_RESTRICT x, block_e3m4_q * GGML_RESTRICT y, int64_t k);
+    GGML_API void dequantize_row_e3m4_q(const block_e3m4_q * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+    GGML_API void quantize_row_e3m4_q(const float * GGML_RESTRICT x, block_e3m4_q * GGML_RESTRICT y, int64_t k);
+    GGML_API void quantize_row_e3m4_q_ref(const float * GGML_RESTRICT x, block_e3m4_q * GGML_RESTRICT y, int64_t k);
 
     // TODO: the best depend on the CPU fp32 / bf16 / fp16
 #define GGML_FP8_VECT_DOT_TYPE GGML_TYPE_F32
-    void ggml_vec_dot_e5m2(int n, float * GGML_RESTRICT s, size_t bs, const ggml_e5m2_t * GGML_RESTRICT vx, size_t bx, const float * GGML_RESTRICT vy, size_t by, int nrc);
-    void ggml_vec_dot_e4m3(int n, float * GGML_RESTRICT s, size_t bs, const ggml_e4m3_t * GGML_RESTRICT vx, size_t bx, const float * GGML_RESTRICT vy, size_t by, int nrc);
-    void ggml_vec_dot_e4m3_q(int n, float * GGML_RESTRICT s, size_t bs, const block_e4m3_q * GGML_RESTRICT vx, size_t bx, const float * GGML_RESTRICT vy, size_t by, int nrc);
-    void ggml_vec_dot_e3m4_q(int n, float * GGML_RESTRICT s, size_t bs, const block_e3m4_q * GGML_RESTRICT vx, size_t bx, const float * GGML_RESTRICT vy, size_t by, int nrc);
+    GGML_API void ggml_vec_dot_e5m2(int n, float * GGML_RESTRICT s, size_t bs, const ggml_e5m2_t * GGML_RESTRICT vx, size_t bx, const float * GGML_RESTRICT vy, size_t by, int nrc);
+    GGML_API void ggml_vec_dot_e4m3(int n, float * GGML_RESTRICT s, size_t bs, const ggml_e4m3_t * GGML_RESTRICT vx, size_t bx, const float * GGML_RESTRICT vy, size_t by, int nrc);
+    GGML_API void ggml_vec_dot_e4m3_q(int n, float * GGML_RESTRICT s, size_t bs, const block_e4m3_q * GGML_RESTRICT vx, size_t bx, const float * GGML_RESTRICT vy, size_t by, int nrc);
+    GGML_API void ggml_vec_dot_e3m4_q(int n, float * GGML_RESTRICT s, size_t bs, const block_e3m4_q * GGML_RESTRICT vx, size_t bx, const float * GGML_RESTRICT vy, size_t by, int nrc);
 
 #ifdef  __cplusplus
 }

Original file line number	Diff line number	Diff line change
`@@ -1846,9 +1846,9 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {`
`1846`	`1846`	`total_seconds = total_seconds % (60*60);`
`1847`	`1847`	`}`
`1848`	`1848`	`LOG("%.2f minutes\n", total_seconds / 60.0);`
	`1849`	`+ LOG("\n");`
	`1850`	`+ LOG("chunk PPL ln(PPL(Q)/PPL(base)) KL Divergence Δp RMS Same top p\n");`
`1849`	`1851`	`}`
`1850`		`- LOG("\n");`
`1851`		`- LOG("chunk PPL ln(PPL(Q)/PPL(base)) KL Divergence Δp RMS Same top p\n");`
`1852`	`1852`
`1853`	`1853`	`const int first = n_ctx/2;`
`1854`	`1854`	`const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx);`