ggml : fix Q4_3 cuBLAS

* Unit test for quantization functions

Use the ggml_internal_get_quantize_fn function to loop through all
quantization formats and run a sanity check on the result.

Also add a microbenchmark that times these functions directly without
running the rest of the GGML graph.

* test-quantize-fns: CI fixes

Fix issues uncovered in CI
 - need to use sizes divisible by 32*8 for loop unrolling
 - use intrinsic header that should work on Mac

* test-quantize: remove

Per PR comment, subsumed by test-quantize-fns

* test-quantize: fix for q8_0 intermediates

.github/workflows/build.yml

@@ -12,7 +12,7 @@ on:
       - master
     paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.c', '**/*.cpp']
   pull_request:
-    types: [opened, synchronize, edited, reopened, review_requested, ready_for_review]
+    types: [opened, synchronize, reopened]
     paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.c', '**/*.cpp']
 
 env:
@@ -20,8 +20,6 @@ env:
 
 jobs:
   ubuntu-latest-make:
-    if: github.event.pull_request.draft == false
-
     runs-on: ubuntu-latest
 
     steps:
@@ -41,8 +39,6 @@ jobs:
           make
 
   ubuntu-latest-cmake:
-    if: github.event.pull_request.draft == false
-
     runs-on: ubuntu-latest
 
     steps:
@@ -71,8 +67,6 @@ jobs:
           ctest --verbose
 
   ubuntu-latest-cmake-sanitizer:
-    if: github.event.pull_request.draft == false
-
     runs-on: ubuntu-latest
 
     continue-on-error: true
@@ -108,8 +102,6 @@ jobs:
           ctest --verbose
 
   macOS-latest-make:
-    if: github.event.pull_request.draft == false
-
     runs-on: macos-latest
 
     steps:
@@ -128,8 +120,6 @@ jobs:
           make
 
   macOS-latest-cmake:
-    if: github.event.pull_request.draft == false
-
     runs-on: macOS-latest
 
     steps:
@@ -157,8 +147,6 @@ jobs:
           ctest --verbose
 
   windows-latest-cmake:
-    if: github.event.pull_request.draft == false
-
     runs-on: windows-latest
 
     strategy:
@@ -169,7 +157,7 @@ jobs:
          - build: 'avx'
            defines: '-DLLAMA_AVX2=OFF'
          - build: 'avx512'
-           defines: '-DLLAMA_AVX512=ON'
+           defines: '-DLLAMA_AVX512=ON -DBUILD_SHARED_LIBS=ON'
 
     steps:
       - name: Clone

CMakeLists.txt

@@ -201,6 +201,10 @@ endif()
 
 if (MSVC)
     add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
+
+    if (BUILD_SHARED_LIBS)
+        set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
+    endif()
 endif()
 
 if (LLAMA_LTO)
@@ -308,6 +312,7 @@ add_library(ggml OBJECT
 target_include_directories(ggml PUBLIC .)
 target_compile_features(ggml PUBLIC c_std_11) # don't bump
 target_link_libraries(ggml PUBLIC Threads::Threads ${LLAMA_EXTRA_LIBS})
+
 if (BUILD_SHARED_LIBS)
     set_target_properties(ggml PROPERTIES POSITION_INDEPENDENT_CODE ON)
 endif()
@@ -320,6 +325,7 @@ add_library(llama
 target_include_directories(llama PUBLIC .)
 target_compile_features(llama PUBLIC cxx_std_11) # don't bump
 target_link_libraries(llama PRIVATE ggml ${LLAMA_EXTRA_LIBS})
+
 if (BUILD_SHARED_LIBS)
     set_target_properties(llama PROPERTIES POSITION_INDEPENDENT_CODE ON)
     target_compile_definitions(llama PRIVATE LLAMA_SHARED LLAMA_BUILD)

Makefile

@@ -74,13 +74,17 @@ endif
 #       feel free to update the Makefile for your architecture and send a pull request or issue
 ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686))
 	# Use all CPU extensions that are available:
-	CFLAGS += -march=native -mtune=native
+	CFLAGS   += -march=native -mtune=native
 	CXXFLAGS += -march=native -mtune=native
+
+	# Usage AVX-only
+	#CFLAGS   += -mfma -mf16c -mavx
+	#CXXFLAGS += -mfma -mf16c -mavx
 endif
 ifneq ($(filter ppc64%,$(UNAME_M)),)
 	POWER9_M := $(shell grep "POWER9" /proc/cpuinfo)
 	ifneq (,$(findstring POWER9,$(POWER9_M)))
-		CFLAGS += -mcpu=power9
+		CFLAGS   += -mcpu=power9
 		CXXFLAGS += -mcpu=power9
 	endif
 	# Require c++23's std::byteswap for big-endian support.
@@ -114,7 +118,7 @@ ifdef LLAMA_GPROF
 	CXXFLAGS += -pg
 endif
 ifneq ($(filter aarch64%,$(UNAME_M)),)
-	CFLAGS += -mcpu=native
+	CFLAGS   += -mcpu=native
 	CXXFLAGS += -mcpu=native
 endif
 ifneq ($(filter armv6%,$(UNAME_M)),)

examples/common.h

@@ -20,7 +20,7 @@ struct gpt_params {
     int32_t repeat_last_n = 64;   // last n tokens to penalize
     int32_t n_parts       = -1;   // amount of model parts (-1 = determine from model dimensions)
     int32_t n_ctx         = 512;  // context size
-    int32_t n_batch       = 8;    // batch size for prompt processing
+    int32_t n_batch       = 512;  // batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_keep        = 0;    // number of tokens to keep from initial prompt
 
     // sampling parameters

examples/main/main.cpp

@@ -25,6 +25,7 @@
 #endif
 
 static console_state con_st;
+static llama_context ** g_ctx;
 
 static bool is_interacting = false;
 
@@ -36,6 +37,7 @@ void sigint_handler(int signo) {
         if (!is_interacting) {
             is_interacting=true;
         } else {
+            llama_print_timings(*g_ctx);
             _exit(130);
         }
     }
@@ -94,6 +96,7 @@ int main(int argc, char ** argv) {
 //bool is_prime(int n) {)";
 
     llama_context * ctx;
+    g_ctx = &ctx;
 
     // load the model
     {

ggml.c

@@ -468,6 +468,14 @@ static inline int hsum_i32_8(const __m256i a) {
     return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32));
 }
 
+// horizontally add 4 int32_t
+static inline int hsum_i32_4(const __m128i a) {
+    const __m128i hi64 = _mm_unpackhi_epi64(a, a);
+    const __m128i sum64 = _mm_add_epi32(hi64, a);
+    const __m128i hi32  = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1));
+    return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32));
+}
+
 #if __AVX2__ || __AVX512F__
 // Unpack 32 4-bit fields into 32 bytes
 // The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval
@@ -656,10 +664,11 @@ static_assert(sizeof(block_q4_3) == 2 * sizeof(ggml_fp16_t) + QK4_3 / 2, "wrong
 #define QK8_0 32
 typedef struct {
     float   d;          // delta
-    float   s;          // d * sum(qs[i])
+    float   s0;         // d * sum(qs[i]) low
+    float   s1;         // d * sum(qs[i]) high
     int8_t  qs[QK8_0];  // quants
 } block_q8_0;
-static_assert(sizeof(block_q8_0) == 2*sizeof(float) + QK8_0, "wrong q8_0 block size/padding");
+static_assert(sizeof(block_q8_0) == 3*sizeof(float) + QK8_0, "wrong q8_0 block size/padding");
 
 
 // reference implementation for deterministic creation of model files
@@ -1299,13 +1308,22 @@ static void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * r
 
         y[i].d = d;
 
-        int sum = 0;
-        for (int l = 0; l < QK8_0; ++l) {
-            const float v = x[i*QK8_0 + l]*id;
-            y[i].qs[l] = roundf(v);
-            sum += y[i].qs[l];
+        int sum0 = 0;
+        int sum1 = 0;
+
+        for (int l = 0; l < QK8_0/2; ++l) {
+            const float v0 = x[i*QK8_0           + l]*id;
+            const float v1 = x[i*QK8_0 + QK8_0/2 + l]*id;
+
+            y[i].qs[          l] = roundf(v0);
+            y[i].qs[QK8_0/2 + l] = roundf(v1);
+
+            sum0 += y[i].qs[          l];
+            sum1 += y[i].qs[QK8_0/2 + l];
         }
-        y[i].s = d * sum;
+
+        y[i].s0 = d * sum0;
+        y[i].s1 = d * sum1;
     }
 }
 
@@ -1335,9 +1353,11 @@ static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int
 
         y[i].d = d;
 
-        int32x4_t accv = vdupq_n_s32(0);
+        int32x4_t accv0 = vdupq_n_s32(0);
+        int32x4_t accv1 = vdupq_n_s32(0);
 
-        for (int l = 0; l < 8; l++) {
+        // low half
+        for (int l = 0; l < 4; l++) {
             const float32x4_t v  = vmulq_n_f32(srcv[l], id);
             const int32x4_t   vi = vcvtnq_s32_f32(v);
 
@@ -1346,10 +1366,27 @@ static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int
             y[i].qs[4*l + 2] = vgetq_lane_s32(vi, 2);
             y[i].qs[4*l + 3] = vgetq_lane_s32(vi, 3);
 
-            accv = vaddq_s32(accv, vi);
+            accv0 = vaddq_s32(accv0, vi);
         }
-        int32_t sum = vaddvq_s32(accv);
-        y[i].s = d * sum;
+
+        // high half
+        for (int l = 4; l < 8; l++) {
+            const float32x4_t v  = vmulq_n_f32(srcv[l], id);
+            const int32x4_t   vi = vcvtnq_s32_f32(v);
+
+            y[i].qs[4*l + 0] = vgetq_lane_s32(vi, 0);
+            y[i].qs[4*l + 1] = vgetq_lane_s32(vi, 1);
+            y[i].qs[4*l + 2] = vgetq_lane_s32(vi, 2);
+            y[i].qs[4*l + 3] = vgetq_lane_s32(vi, 3);
+
+            accv1 = vaddq_s32(accv1, vi);
+        }
+
+        const int32_t sum0 = vaddvq_s32(accv0);
+        const int32_t sum1 = vaddvq_s32(accv1);
+
+        y[i].s0 = d * sum0;
+        y[i].s1 = d * sum1;
     }
 #elif defined(__AVX2__) || defined(__AVX__)
     for (int i = 0; i < nb; i++) {
@@ -1398,7 +1435,9 @@ static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int
 
 #if defined(__AVX2__)
         // Compute the sum of the quants and set y[i].s
-        y[i].s = d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3)));
+        //y[i].s = d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3)));
+        y[i].s0 = d * hsum_i32_8(_mm256_add_epi32(i0, i1));
+        y[i].s1 = d * hsum_i32_8(_mm256_add_epi32(i2, i3));
 
         // Convert int32 to int16
         i0 = _mm256_packs_epi32( i0, i1 );	// 0, 1, 2, 3,  8, 9, 10, 11,  4, 5, 6, 7, 12, 13, 14, 15
@@ -1428,7 +1467,8 @@ static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int
         // Compute the sum of the quants and set y[i].s
         const __m128i s0 = _mm_add_epi32(_mm_add_epi32(ni0, ni1), _mm_add_epi32(ni2, ni3));
         const __m128i s1 = _mm_add_epi32(_mm_add_epi32(ni4, ni5), _mm_add_epi32(ni6, ni7));
-        y[i].s = d * hsum_i32_8(_mm256_set_m128i(s1, s0));
+        y[i].s0 = d * hsum_i32_4(s0);
+        y[i].s1 = d * hsum_i32_4(s1);
 
         // Convert int32 to int16
         ni0 = _mm_packs_epi32( ni0, ni1 );
@@ -2395,7 +2435,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
         const block_q8_0 * restrict y0 = &y[i + 0];
         const block_q8_0 * restrict y1 = &y[i + 1];
 
-        sum8 += x0->d * y0->s + x1->d * y1->s;
+        sum8 += x0->d * (y0->s0 + y0->s1) + x1->d * (y1->s0 + y1->s1);
 
         const uint8x16_t m4b   = vdupq_n_u8(0xf);
 
@@ -2562,7 +2602,7 @@ static void ggml_vec_dot_q4_1_q8_0(const int n, float * restrict s, const void *
         const block_q8_0 * restrict y0 = &y[i + 0];
         const block_q8_0 * restrict y1 = &y[i + 1];
 
-        summs += x0->m * y0->s + x1->m * y1->s;
+        summs += x0->m * (y0->s0 + y0->s1) + x1->m * (y1->s0 + y1->s1);
 
         const uint8x16_t m4b = vdupq_n_u8(0xf);
 
@@ -2575,35 +2615,35 @@ static void ggml_vec_dot_q4_1_q8_0(const int n, float * restrict s, const void *
         const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
         const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
 
+        // interleave
+        const int8x16_t v0_0lz = vzip1q_s8(v0_0l, v0_0h);
+        const int8x16_t v0_0hz = vzip2q_s8(v0_0l, v0_0h);
+        const int8x16_t v0_1lz = vzip1q_s8(v0_1l, v0_1h);
+        const int8x16_t v0_1hz = vzip2q_s8(v0_1l, v0_1h);
+
         // load y
         const int8x16_t v1_0l = vld1q_s8(y0->qs);
         const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
         const int8x16_t v1_1l = vld1q_s8(y1->qs);
         const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
 
-        // interleave
-        const int8x16_t v1_0ls = vuzp1q_s8(v1_0l, v1_0h);
-        const int8x16_t v1_0hs = vuzp2q_s8(v1_0l, v1_0h);
-        const int8x16_t v1_1ls = vuzp1q_s8(v1_1l, v1_1h);
-        const int8x16_t v1_1hs = vuzp2q_s8(v1_1l, v1_1h);
-
 #if defined(__ARM_FEATURE_DOTPROD)
         // dot product into int32x4_t
-        const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0ls), v0_0h, v1_0hs);
-        const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1ls), v0_1h, v1_1hs);
+        const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0lz, v1_0l), v0_0hz, v1_0h);
+        const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1lz, v1_1l), v0_1hz, v1_1h);
 
         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_0l), vget_low_s8 (v1_0ls));
-        const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0l), vget_high_s8(v1_0ls));
-        const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0h), vget_low_s8 (v1_0hs));
-        const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0h), vget_high_s8(v1_0hs));
+        const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lz), vget_low_s8 (v1_0l));
+        const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lz), vget_high_s8(v1_0l));
+        const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hz), vget_low_s8 (v1_0h));
+        const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hz), vget_high_s8(v1_0h));
 
-        const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1l), vget_low_s8 (v1_1ls));
-        const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1l), vget_high_s8(v1_1ls));
-        const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1h), vget_low_s8 (v1_1hs));
-        const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1h), vget_high_s8(v1_1hs));
+        const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lz), vget_low_s8 (v1_1l));
+        const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lz), vget_high_s8(v1_1l));
+        const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hz), vget_low_s8 (v1_1h));
+        const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hz), vget_high_s8(v1_1h));
 
         const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h));
         const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h));
@@ -2627,7 +2667,7 @@ static void ggml_vec_dot_q4_1_q8_0(const int n, float * restrict s, const void *
         const float * d0 = &x[i].d;
         const float * d1 = &y[i].d;
 
-        summs += x[i].m * y[i].s;
+        summs += x[i].m * (y[i].s0 + y[i].s1);
 
         const __m256 d0v = _mm256_broadcast_ss( d0 );
         const __m256 d1v = _mm256_broadcast_ss( d1 );
@@ -2845,88 +2885,53 @@ static void ggml_vec_dot_q4_3_q8_0(const int n, float * restrict s, const void *
     float32x4_t sumv0 = vdupq_n_f32(0.0f);
     float32x4_t sumv1 = vdupq_n_f32(0.0f);
 
-    for (int i = 0; i < nb; i += 2) {
+    float summs0 = 0.0f;
+    float summs1 = 0.0f;
+
+    for (int i = 0; i < nb; ++i) {
         const block_q4_3 * restrict x0_0 = &x[2*(i + 0) + 0];
         const block_q4_3 * restrict x0_1 = &x[2*(i + 0) + 1];
-        const block_q4_3 * restrict x1_0 = &x[2*(i + 1) + 0];
-        const block_q4_3 * restrict x1_1 = &x[2*(i + 1) + 1];
 
         const block_q8_0 * restrict y0 = &y[i + 0];
-        const block_q8_0 * restrict y1 = &y[i + 1];
 
-        const uint8x16_t m4b = vdupq_n_u8(0xf);
-
-        const float x0_0d = GGML_FP16_TO_FP32(x0_0->d);
-        const float x0_1d = GGML_FP16_TO_FP32(x0_1->d);
-        const float x1_0d = GGML_FP16_TO_FP32(x1_0->d);
-        const float x1_1d = GGML_FP16_TO_FP32(x1_1->d);
-
-        const float x0_0m = GGML_FP16_TO_FP32(x0_0->m);
-        const float x0_1m = GGML_FP16_TO_FP32(x0_1->m);
-        const float x1_0m = GGML_FP16_TO_FP32(x1_0->m);
-        const float x1_1m = GGML_FP16_TO_FP32(x1_1->m);
+        summs0 += GGML_FP16_TO_FP32(x0_0->m) * y0->s0;
+        summs1 += GGML_FP16_TO_FP32(x0_1->m) * y0->s1;
 
         const uint8x16_t v0_0 = vcombine_u8(vld1_u8(x0_0->qs), vld1_u8(x0_1->qs));
-        const uint8x16_t v0_1 = vcombine_u8(vld1_u8(x1_0->qs), vld1_u8(x1_1->qs));
 
         // 4-bit -> 8-bit
-        const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
+        const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, vdupq_n_u8(0xf)));
         const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
-        const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
-        const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
 
         // interleave
         const int8x16_t v0_0lz = vzip1q_s8(v0_0l, v0_0h);
         const int8x16_t v0_0hz = vzip2q_s8(v0_0l, v0_0h);
-        const int8x16_t v0_1lz = vzip1q_s8(v0_1l, v0_1h);
-        const int8x16_t v0_1hz = vzip2q_s8(v0_1l, v0_1h);
 
         // load y
         const int8x16_t v1_0l = vld1q_s8(y0->qs);
         const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
-        const int8x16_t v1_1l = vld1q_s8(y1->qs);
-        const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
 
-        const int16x8_t sy0_0 = vaddq_s16(vmovl_s8(vget_low_s8(v1_0l)), vmovl_s8(vget_high_s8(v1_0l)));
-        const int16x8_t sy0_1 = vaddq_s16(vmovl_s8(vget_low_s8(v1_0h)), vmovl_s8(vget_high_s8(v1_0h)));
-
-        const int16x8_t sy1_0 = vaddq_s16(vmovl_s8(vget_low_s8(v1_1l)), vmovl_s8(vget_high_s8(v1_1l)));
-        const int16x8_t sy1_1 = vaddq_s16(vmovl_s8(vget_low_s8(v1_1h)), vmovl_s8(vget_high_s8(v1_1h)));
-
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddl_s16(vget_low_s16(sy0_0), vget_high_s16(sy0_0))), x0_0m*y0->d);
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddl_s16(vget_low_s16(sy0_1), vget_high_s16(sy0_1))), x0_1m*y0->d);
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddl_s16(vget_low_s16(sy1_0), vget_high_s16(sy1_0))), x1_0m*y1->d);
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddl_s16(vget_low_s16(sy1_1), vget_high_s16(sy1_1))), x1_1m*y1->d);
+        const float x0_0d = GGML_FP16_TO_FP32(x0_0->d);
+        const float x0_1d = GGML_FP16_TO_FP32(x0_1->d);
 
 #if defined(__ARM_FEATURE_DOTPROD)
         sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_0lz, v1_0l)), x0_0d*y0->d);
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_0hz, v1_0h)), x0_1d*y0->d);
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_1lz, v1_1l)), x1_0d*y1->d);
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_1hz, v1_1h)), x1_1d*y1->d);
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_0hz, v1_0h)), x0_1d*y0->d);
 #else
         const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lz), vget_low_s8 (v1_0l));
         const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lz), vget_high_s8(v1_0l));
         const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hz), vget_low_s8 (v1_0h));
         const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hz), vget_high_s8(v1_0h));
 
-        const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lz), vget_low_s8 (v1_1l));
-        const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lz), vget_high_s8(v1_1l));
-        const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hz), vget_low_s8 (v1_1h));
-        const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hz), vget_high_s8(v1_1h));
-
         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));
 
         sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(pl0), x0_0d*y0->d);
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(ph0), x0_1d*y0->d);
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(pl1), x1_0d*y1->d);
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(ph1), x1_1d*y1->d);
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(ph0), x0_1d*y0->d);
 #endif
     }
 
-    *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
+    *s = vaddvq_f32(vaddq_f32(sumv0, sumv1)) + summs0 + summs1;
 #elif defined(__AVX2__)
     // Initialize accumulator with zeros
     __m256 acc = _mm256_setzero_ps();
@@ -2971,9 +2976,6 @@ static void ggml_vec_dot_q4_3_q8_0(const int n, float * restrict s, const void *
         const float d1 = GGML_FP16_TO_FP32(x[2*i + 1].d);
         const float m1 = GGML_FP16_TO_FP32(x[2*i + 1].m);
 
-        int sy_0 = 0;
-        int sy_1 = 0;
-
         int sxy_0 = 0;
         int sxy_1 = 0;
 
@@ -2993,15 +2995,11 @@ static void ggml_vec_dot_q4_3_q8_0(const int n, float * restrict s, const void *
             const int y0_1 = y0[2*(j + QK8_0/4) + 0];
             const int y1_1 = y0[2*(j + QK8_0/4) + 1];
 
-            sy_0 += y0_0 + y1_0;
-            sy_1 += y0_1 + y1_1;
-
             sxy_0 += x0_0*y0_0 + x1_0*y1_0;
             sxy_1 += x0_1*y0_1 + x1_1*y1_1;
         }
 
-        sumf += (d0*sxy_0 + m0*sy_0)*y[i].d;
-        sumf += (d1*sxy_1 + m1*sy_1)*y[i].d;
+        sumf += (d0*sxy_0 + d1*sxy_1)*y[i].d + m0*y[i].s0 + m1*y[i].s1;
     }
     *s = sumf;
 #endif
@@ -7994,6 +7992,9 @@ static void ggml_compute_forward_mul_mat_q_f32(
         else if (type == GGML_TYPE_Q4_2) {
             dequantize_row_q_cuda = dequantize_row_q4_2_cuda;
         }
+        else if (type == GGML_TYPE_Q4_3) {
+            dequantize_row_q_cuda = dequantize_row_q4_3_cuda;
+        }
         else {
             GGML_ASSERT(false);
         }

llama.cpp

@@ -68,7 +68,7 @@ static const std::map<e_model, size_t> & MEM_REQ_SCRATCH1()
         { MODEL_65B,   512ull * MB },
     };
     return _MEM_REQ_SCRATCH1;
-};
+}
 
 // 2*n_embd*n_ctx*n_layer*sizeof(float16)
 static const std::map<e_model, size_t> & MEM_REQ_KV_SELF()
@@ -80,7 +80,7 @@ static const std::map<e_model, size_t> & MEM_REQ_KV_SELF()
         { MODEL_65B,  5120ull * MB },
     };
     return _MEM_REQ_KV_SELF;
-};
+}
 
 // this is mostly needed for temporary mul_mat buffers to dequantize the data
 // not actually needed if BLAS is disabled
@@ -93,7 +93,7 @@ static const std::map<e_model, size_t> & MEM_REQ_EVAL()
         { MODEL_65B, 1536ull * MB },
     };
     return _MEM_REQ_EVAL;
-};
+}
 
 // default hparams (LLaMA 7B)
 struct llama_hparams {
@@ -2256,7 +2256,6 @@ std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_te
 
 // Returns the size of the state
 size_t llama_get_state_size(struct llama_context * ctx) {
-    const size_t s_bool = sizeof(int32_t);
     // we don't know size of rng until we actually serialize it. so reserve more than enough memory for its serialized state.
     // for reference, std::mt19937(1337) serializes to 6701 bytes.
     const size_t s_rng_size = sizeof(size_t);

tests/CMakeLists.txt

@@ -6,5 +6,6 @@ function(llama_add_test source)
 endfunction()
 
 # llama_add_test(test-double-float.c) # SLOW
-llama_add_test(test-quantize.c)
+llama_add_test(test-quantize-fns.cpp)
+llama_add_test(test-quantize-perf.cpp)
 llama_add_test(test-tokenizer-0.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab.bin)

tests/test-quantize-fns.cpp

@@ -0,0 +1,154 @@
+// Unit tests for quantization specific functions - quantize, dequantize and dot product
+
+#include "ggml.h"
+
+#undef NDEBUG
+#include <assert.h>
+#include <math.h>
+#include <stdio.h>
+#include <string>
+#include <vector>
+
+
+const float MAX_QUANTIZATION_REFERENCE_ERROR = 0.0001;
+const float MAX_QUANTIZATION_TOTAL_ERROR = 0.002;
+const float MAX_DOT_PRODUCT_ERROR = 0.02;
+
+const char* RESULT_STR[] = {"ok", "FAILED"};
+
+
+// Generate synthetic data
+void generate_data(float offset, size_t n, float * dst) {
+    for (size_t i = 0; i < n; i++) {
+        dst[i] = 0.1 + 2*cosf(i + offset);
+    }
+}
+
+// Calculate RMSE between two float arrays
+float array_rmse(const float * a1, const float * a2, size_t n) {
+    double sum = 0;
+    for (size_t i = 0; i < n; i++) {
+        double diff = a1[i] - a2[i];
+        sum += diff * diff;
+    }
+    return sqrtf(sum) / n;
+}
+
+// Total quantization error on test data
+float total_quantization_error(quantize_fns_t & qfns, size_t test_size, const float * test_data) {
+    std::vector<uint8_t> tmp_q(test_size);
+    std::vector<float> tmp_out(test_size);
+
+    qfns.quantize_row_q(test_data, tmp_q.data(), test_size);
+    qfns.dequantize_row_q(tmp_q.data(), tmp_out.data(), test_size);
+    return array_rmse(test_data, tmp_out.data(), test_size);
+}
+
+// Total quantization error on test data
+float reference_quantization_error(quantize_fns_t & qfns, size_t test_size, const float * test_data) {
+    std::vector<uint8_t> tmp_q(test_size);
+    std::vector<float> tmp_out(test_size);
+    std::vector<float> tmp_out_ref(test_size);
+
+    qfns.quantize_row_q(test_data, tmp_q.data(), test_size);
+    qfns.dequantize_row_q(tmp_q.data(), tmp_out.data(), test_size);
+
+    qfns.quantize_row_q_reference(test_data, tmp_q.data(), test_size);
+    qfns.dequantize_row_q(tmp_q.data(), tmp_out_ref.data(), test_size);
+
+    return array_rmse(tmp_out.data(), tmp_out_ref.data(), test_size);
+}
+
+float dot_product(const float * a1, const float * a2, size_t test_size) {
+    double sum = 0;
+    for (size_t i = 0; i < test_size; i++) {
+        sum += a1[i] * a2[i];
+    }
+    return sum;
+}
+
+// Total dot product error
+float dot_product_error(quantize_fns_t & qfns, size_t test_size, const float * test_data1, const float *test_data2) {
+    std::vector<uint8_t> tmp_q1(test_size);
+    std::vector<uint8_t> tmp_q2(test_size*2);
+
+    qfns.quantize_row_q(test_data1, tmp_q1.data(), test_size);
+    qfns.quantize_row_q_dot(test_data2, tmp_q2.data(), test_size);
+
+    float result = INFINITY;
+    qfns.vec_dot_q(test_size, &result, tmp_q1.data(), tmp_q2.data());
+
+    const float dot_ref = dot_product(test_data1, test_data2, test_size);
+
+    return fabsf(result - dot_ref) / test_size;
+}
+
+int main(int argc, char * argv[]) {
+    bool verbose = false;
+    const size_t test_size = 32 * 128;
+
+    std::string arg;
+    for (int i = 1; i < argc; i++) {
+        arg = argv[i];
+
+        if (arg == "-v") {
+            verbose = true;
+        } else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            return 1;
+        }
+    }
+
+    std::vector<float> test_data(test_size);
+    std::vector<float> test_data2(test_size);
+
+    generate_data(0.0, test_data.size(), test_data.data());
+    generate_data(1.0, test_data2.size(), test_data2.data());
+
+    // Initialize GGML, ensures float conversion tables are initialized
+    struct ggml_init_params ggml_params = {
+        /* .mem_size   = */ 1*1024,
+        /* .mem_buffer = */ NULL,
+        /* .no_alloc   = */ true,
+    };
+    struct ggml_context * ctx = ggml_init(ggml_params);
+
+    int num_failed = 0;
+    bool failed = false;
+
+    for (int i = 0; i < GGML_TYPE_COUNT; i++) {
+        ggml_type type = (ggml_type) i;
+        quantize_fns_t qfns = ggml_internal_get_quantize_fn(i);
+
+        if (qfns.quantize_row_q && qfns.dequantize_row_q) {
+            const float total_error = total_quantization_error(qfns, test_size, test_data.data());
+            failed = !(total_error < MAX_QUANTIZATION_TOTAL_ERROR);
+            num_failed += failed;
+            if (failed || verbose) {
+                printf("%5s absolute quantization error: %s (%f)\n", ggml_type_name(type), RESULT_STR[failed], total_error);
+            }
+
+            const float reference_error = reference_quantization_error(qfns, test_size, test_data.data());
+            failed = !(reference_error < MAX_QUANTIZATION_REFERENCE_ERROR);
+            num_failed += failed;
+            if (failed || verbose) {
+                printf("%5s reference implementation error: %s (%f)\n", ggml_type_name(type), RESULT_STR[failed], reference_error);
+            }
+
+            const float vec_dot_error = dot_product_error(qfns, test_size, test_data.data(), test_data2.data());
+            failed = !(vec_dot_error < MAX_DOT_PRODUCT_ERROR);
+            num_failed += failed;
+            if (failed || verbose) {
+                printf("%5s dot product error: %s (%f)\n", ggml_type_name(type), RESULT_STR[failed], vec_dot_error);
+            }
+        }
+    }
+
+    if (num_failed || verbose) {
+        printf("%d tests failed\n", num_failed);
+    }
+
+    ggml_free(ctx);
+
+    return num_failed > 0;
+}

tests/test-quantize-perf.cpp

@@ -0,0 +1,310 @@
+// Benchmark quantization specific functions on synthetic data
+
+#include "ggml.h"
+
+#undef NDEBUG
+#include <algorithm>
+#include <assert.h>
+#include <functional>
+#include <inttypes.h>
+#include <math.h>
+#include <memory>
+#include <stdio.h>
+#include <string>
+#include <vector>
+
+#define MAX_ALIGNMENT 64
+#define QK 32
+#define WARMUP 5
+#define ITERATIONS 10
+
+#define L1_SIZE      32*128
+#define L2_SIZE     32*2048
+#define L3_SIZE    32*20480
+#define MEM_SIZE 32*2048000
+
+struct quantize_perf_params {
+    std::vector<std::string> include_types;
+    std::vector<size_t> test_sizes;
+    size_t alignment_offset = 0;
+    bool op_quantize_row_q_reference = false;
+    bool op_quantize_row_q = false;
+    bool op_dequantize_row_q = false;
+    bool op_quantize_row_q_dot = false;
+    bool op_vec_dot_q = false;
+};
+
+
+#if defined(__x86_64__) || defined(__i386__)
+
+#include <x86intrin.h>
+inline int64_t cpu_cycles() {
+// Rough way to detect new-ish CPUs
+#ifdef __POPCNT__
+    unsigned int dummy;
+    return __rdtscp(&dummy);
+#else
+    return __rdtsc();
+#endif
+}
+
+#else
+
+#define cpu_cycles() 0
+
+#endif
+
+
+// Generate synthetic data
+void generate_data(float offset, size_t n, float * dst) {
+    for (size_t i = 0; i < n; i++) {
+        dst[i] = 0.1 + 2*cosf(i + offset);
+    }
+}
+
+float gigabytes_per_second(size_t bytes, int64_t usecs) {
+    return bytes / (float) usecs * 1000000 / (1024*1024*1024);
+}
+
+void * align_with_offset(void * ptr, int offset) {
+    size_t dummy_size = MAX_ALIGNMENT * 4;
+    return (char *) std::align(MAX_ALIGNMENT, MAX_ALIGNMENT, ptr, dummy_size) + offset;
+}
+
+void benchmark_function(size_t size, size_t q_size, std::function<size_t(void)> function) {
+    int64_t min_time_us = INT64_MAX;
+    int64_t total_time_us = 0;
+    int64_t min_time_cycles = INT64_MAX;
+    int64_t total_time_cycles = 0;
+
+    for (int i = 0; i < WARMUP; i++) {
+        function();
+    }
+
+
+    for (int i = 0; i < ITERATIONS; i++) {
+        const int64_t start_time = ggml_time_us();
+        const int64_t start_cycles = cpu_cycles();
+
+        function();
+
+        const int64_t end_cycles = cpu_cycles();
+        const int64_t end_time = ggml_time_us();
+
+        total_time_cycles += end_cycles - start_cycles;
+        min_time_cycles = std::min(min_time_cycles, end_cycles - start_cycles);
+        total_time_us += end_time - start_time;
+        min_time_us = std::min(min_time_us, end_time - start_time);
+    }
+
+    printf("      min cycles/%d vals   : %9.2f\n",  QK, QK * min_time_cycles / (float) size);
+    printf("      avg cycles/%d vals   : %9.2f\n",  QK, QK * total_time_cycles / (float) (size * ITERATIONS));
+    printf("      float32 throughput   : %9.2f GB/s\n",  gigabytes_per_second(4 * size * ITERATIONS, total_time_us));
+    printf("      quantized throughput : %9.2f GB/s\n",  gigabytes_per_second(q_size * ITERATIONS, total_time_us));
+}
+
+int main(int argc, char * argv[]) {
+    quantize_perf_params params {};
+
+    // read command line
+
+    bool invalid_param = false;
+    std::string arg;
+    for (int i = 1; i < argc; i++) {
+        arg = argv[i];
+
+        if (arg == "--size") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            size_t size = std::stoi(argv[i]);
+            if (size % 32 != 0) {
+                fprintf(stderr, "error: size %zu not divisible by 32\n", size);
+                invalid_param = true;
+                break;
+            }
+            params.test_sizes.push_back(size);
+        } else if (arg == "-3") {
+            // quick select sizes that probably fit in CPU caches
+            params.test_sizes.push_back(L1_SIZE);
+            params.test_sizes.push_back(L2_SIZE);
+            params.test_sizes.push_back(L3_SIZE);
+        } else if (arg == "-4") {
+            // quick select cache sizes + memory
+            params.test_sizes.push_back(L1_SIZE);
+            params.test_sizes.push_back(L2_SIZE);
+            params.test_sizes.push_back(L3_SIZE);
+            params.test_sizes.push_back(MEM_SIZE);
+        } else if (arg == "--op") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            std::string op {argv[i]};
+            if (op == "quantize_row_q_reference") {
+                params.op_quantize_row_q_reference = true;
+            } else if (op == "quantize_row_q") {
+                params.op_quantize_row_q = true;
+            } else if (op == "dequantize_row_q") {
+                params.op_dequantize_row_q = true;
+            } else if (op == "quantize_row_q_dot") {
+                params.op_quantize_row_q_dot = true;
+            } else if (op == "vec_dot_q") {
+                params.op_vec_dot_q = true;
+            } else {
+                invalid_param = true;
+                break;
+            }
+        } else if (arg == "--type") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.include_types.push_back(argv[i]);
+        } else if (arg == "--alignment-offset") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            int alignment = std::stoi(argv[i]);
+            if (alignment < 0 || alignment > MAX_ALIGNMENT) {
+            fprintf(stderr, "error: aligment-offset must be less than %d\n", MAX_ALIGNMENT);
+                invalid_param = true;
+                break;
+            }
+            params.alignment_offset = alignment;
+        } else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            return 1;
+        }
+    }
+    if (invalid_param) {
+        fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
+        return 1;
+    }
+
+    if (params.test_sizes.empty()) {
+        params.test_sizes.push_back(L1_SIZE);
+    }
+    if (!(params.op_quantize_row_q_reference || params.op_quantize_row_q || params.op_dequantize_row_q || params.op_quantize_row_q_dot || params.op_vec_dot_q)) {
+        params.op_quantize_row_q_reference = params.op_quantize_row_q = params.op_dequantize_row_q = params.op_quantize_row_q_dot = params.op_vec_dot_q = true;
+    }
+
+    std::sort(params.test_sizes.begin(), params.test_sizes.end());
+    size_t largest = params.test_sizes.back();
+
+    std::vector<uint8_t> test_data1_v(largest*4 + MAX_ALIGNMENT*2);
+    std::vector<uint8_t> test_data2_v(largest*4 + MAX_ALIGNMENT*2);
+    std::vector<uint8_t> test_q1_v(largest*4 + MAX_ALIGNMENT*2);
+    std::vector<uint8_t> test_q2_v(largest*4 + MAX_ALIGNMENT*2);
+    std::vector<uint8_t> test_out_v(largest*4 + MAX_ALIGNMENT*2);
+
+    float * test_data1 = (float *) align_with_offset(test_data1_v.data(), params.alignment_offset);
+    float * test_data2 = (float *) align_with_offset(test_data2_v.data(), params.alignment_offset);
+    float * test_q1 = (float *) align_with_offset(test_q1_v.data(), params.alignment_offset);
+    float * test_q2 = (float *) align_with_offset(test_q2_v.data(), params.alignment_offset);
+    float * test_out = (float *) align_with_offset(test_out_v.data(), params.alignment_offset);
+
+    generate_data(0, largest, test_data1);
+    generate_data(1, largest, test_data2);
+
+
+    // Initialize GGML, ensures float conversion tables are initialized
+    struct ggml_init_params ggml_params = {
+        /* .mem_size   = */ 1*1024,
+        /* .mem_buffer = */ NULL,
+        /* .no_alloc   = */ true,
+    };
+    struct ggml_context * ctx = ggml_init(ggml_params);
+
+    for (int i = 0; i < GGML_TYPE_COUNT; i++) {
+        ggml_type type = (ggml_type) i;
+        quantize_fns_t qfns = ggml_internal_get_quantize_fn(i);
+        if (!params.include_types.empty() && std::find(params.include_types.begin(), params.include_types.end(), ggml_type_name(type)) == params.include_types.end()) {
+            continue;
+        }
+
+        if (qfns.quantize_row_q && qfns.dequantize_row_q) {
+            printf("%s\n", ggml_type_name(type));
+
+            if (params.op_quantize_row_q_reference) {
+                printf("  quantize_row_q_reference\n");
+                for (size_t size : params.test_sizes) {
+                    printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
+                    auto quantize_fn = [&](void ) {
+                        qfns.quantize_row_q_reference(test_data1, test_q1, size);
+                        return test_q1[0];
+                    };
+                    size_t quantized_size = size / ggml_blck_size(type) * ggml_type_size(type);
+                    benchmark_function(size, quantized_size, quantize_fn);
+                }
+                printf("\n");
+            }
+
+            if (params.op_quantize_row_q) {
+                printf("  quantize_row_q\n");
+                for (size_t size : params.test_sizes) {
+                    printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
+                    auto quantize_fn = [&](void ) {
+                        qfns.quantize_row_q(test_data1, test_q1, size);
+                        return test_q1[0];
+                    };
+                    size_t quantized_size = size / ggml_blck_size(type) * ggml_type_size(type);
+                    benchmark_function(size, quantized_size, quantize_fn);
+                }
+                printf("\n");
+            }
+
+            if (params.op_dequantize_row_q) {
+                printf("  dequantize_row_q\n");
+                qfns.quantize_row_q(test_data1, test_q1, largest);
+                for (size_t size : params.test_sizes) {
+                    printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
+                    auto quantize_fn = [&](void ) {
+                        qfns.dequantize_row_q(test_q1, test_out, size);
+                        return test_out[0];
+                    };
+                    size_t quantized_size = size / ggml_blck_size(type) * ggml_type_size(type);
+                    benchmark_function(size, quantized_size, quantize_fn);
+                }
+                printf("\n");
+            }
+
+            if (params.op_quantize_row_q_dot) {
+                printf("  quantize_row_q_dot\n");
+                for (size_t size : params.test_sizes) {
+                    printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
+                    auto quantize_fn = [&](void ) {
+                        qfns.quantize_row_q_dot(test_data1, test_q1, size);
+                        return test_q1[0];
+                    };
+                    size_t quantized_size = size / ggml_blck_size(type) * ggml_type_size(type);
+                    benchmark_function(size, quantized_size, quantize_fn);
+                }
+                printf("\n");
+            }
+
+            if (params.op_vec_dot_q) {
+                printf("  vec_dot_q\n");
+                qfns.quantize_row_q(test_data1, test_q1, largest);
+                qfns.quantize_row_q(test_data2, test_q2, largest);
+                for (size_t size : params.test_sizes) {
+                    printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
+                    auto quantize_fn = [&](void ) {
+                        float result;
+                        qfns.vec_dot_q(size, &result, test_q1, test_q2);
+                        return result;
+                    };
+                    size_t quantized_size = size / ggml_blck_size(type) * ggml_type_size(type);
+                    benchmark_function(size, quantized_size, quantize_fn);
+                }
+                printf("\n");
+            }
+        }
+    }
+
+    ggml_free(ctx);
+
+    return 0;
+}

tests/test-quantize.c

@@ -1,42 +0,0 @@
-#include "ggml.h"
-#undef NDEBUG
-#include <assert.h>
-#include <math.h>
-
-int main(void) {
-    #define QK 32
-    float src[QK];
-    uint8_t dst[24];
-    int64_t hist[16];
-
-    for (int i = 0; i < QK; i++) {
-        src[i] = (float)(i + 1);
-    }
-
-    size_t size = ggml_quantize_q4_0(src, dst, QK, QK, hist);
-    assert(size == 20);
-    float max_result = ((float *)dst)[0];
-    float max_expected = src[31] / ((1 << 3) - 1);
-    assert(max_result == max_expected);
-    for (int i = 0; i < QK; i++) {
-        uint8_t q4_result = (i % 2) ? (dst[sizeof(float) + i/2] >> 4) : (dst[sizeof(float) + i/2] & 0xF);
-        uint8_t q4_expected = roundf(src[i] / max_expected) + 8;
-        assert(q4_result == q4_expected);
-    }
-
-    size = ggml_quantize_q4_1(src, dst, QK, QK, hist);
-    assert(size == 24);
-    float delta_result = ((float *)dst)[0];
-    float delta_expected = (src[31] - src[0]) / ((1 << 4) - 1);
-    assert(delta_result == delta_expected);
-    float min_result = ((float *)dst)[1];
-    float min_expected = src[0];
-    assert(min_result == min_expected);
-    for (int i = 0; i < QK; i++) {
-        uint8_t q4_result = (i % 2) ? (dst[sizeof(float)*2 + i/2] >> 4) : (dst[sizeof(float)*2 + i/2] & 0xF);
-        uint8_t q4_expected = roundf((src[i] - min_expected) / delta_expected);
-        assert(q4_result == q4_expected);
-    }
-
-    return 0;
-}