Merge branch 'master' into q4_1xq8_0

Author: ggerganov

.gitignore

@@ -24,6 +24,7 @@ models/*
 /perplexity
 /embedding
 /benchmark-q4_0-matmult
+/vdot
 /Pipfile
 
 arm_neon.h

CMakeLists.txt

@@ -305,4 +305,5 @@ endif ()
 
 if (LLAMA_BUILD_EXAMPLES)
     add_subdirectory(examples)
+    add_subdirectory(pocs)
 endif()

Makefile

@@ -133,7 +133,7 @@ $(info I CC:       $(CCV))
 $(info I CXX:      $(CXXV))
 $(info )
 
-default: main quantize quantize-stats perplexity embedding
+default: main quantize quantize-stats perplexity embedding vdot
 
 #
 # Build library
@@ -169,6 +169,9 @@ perplexity: examples/perplexity/perplexity.cpp ggml.o llama.o common.o
 embedding: examples/embedding/embedding.cpp ggml.o llama.o common.o
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
+vdot: pocs/vdot/vdot.cpp ggml.o
+	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
+
 libllama.so: llama.o ggml.o
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 

ggml.c

@@ -2249,8 +2249,8 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
     float sumf = 0.0;
 
 #if defined(__ARM_NEON)
-    float sum0 = 0.0f;
-    float sum1 = 0.0f;
+    float32x4_t sumv0 = vdupq_n_f32(0.0f);
+    float32x4_t sumv1 = vdupq_n_f32(0.0f);
 
     for (int i = 0; i < nb; i += 2) {
         const block_q4_0 * restrict x0 = &x[i + 0];
@@ -2290,14 +2290,11 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
 
 #if defined(__ARM_FEATURE_DOTPROD)
         // dot product into int32x4_t
-        int32x4_t p_0 = vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0ls);
-        int32x4_t p_1 = vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1ls);
+        const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0ls), v0_0hs, v1_0hs);
+        const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1ls), v0_1hs, v1_1hs);
 
-        p_0 = vdotq_s32(p_0, v0_0hs, v1_0hs);
-        p_1 = vdotq_s32(p_1, v0_1hs, v1_1hs);
-
-        sum0 += x0->d*y0->d*vaddvq_s32(p_0);
-        sum1 += x1->d*y1->d*vaddvq_s32(p_1);
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), x0->d*y0->d);
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), x1->d*y1->d);
 #else
         const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0ls), vget_low_s8 (v1_0ls));
         const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0ls), vget_high_s8(v1_0ls));
@@ -2309,21 +2306,17 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
         const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hs), vget_low_s8 (v1_1hs));
         const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hs), vget_high_s8(v1_1hs));
 
-        const int16x8_t pl_0 = vaddq_s16(pl0l, pl0h);
-        const int16x8_t ph_0 = vaddq_s16(ph0l, ph0h);
-
-        const int16x8_t pl_1 = vaddq_s16(pl1l, pl1h);
-        const int16x8_t ph_1 = vaddq_s16(ph1l, ph1h);
-
-        const int16x8_t p_0 = vaddq_s16(pl_0, ph_0);
-        const int16x8_t p_1 = vaddq_s16(pl_1, ph_1);
+        const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h));
+        const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h));
+        const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h));
+        const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h));
 
-        sum0 += x0->d*y0->d*vaddvq_s16(p_0);
-        sum1 += x1->d*y1->d*vaddvq_s16(p_1);
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0->d*y0->d);
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), x1->d*y1->d);
 #endif
     }
 
-    sumf = sum0 + sum1;
+    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
 #elif defined(__AVX2__)
     // Initialize accumulator with zeros
     __m256 acc = _mm256_setzero_ps();

pocs/CMakeLists.txt

@@ -0,0 +1,12 @@
+# dependencies
+
+find_package(Threads REQUIRED)
+
+# third-party
+
+include_directories(${CMAKE_CURRENT_SOURCE_DIR})
+
+if (EMSCRIPTEN)
+else()
+    add_subdirectory(vdot)
+endif()

pocs/vdot/CMakeLists.txt

@@ -0,0 +1,4 @@
+set(TARGET vdot)
+add_executable(${TARGET} vdot.cpp)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)