k-quants : fix build on armv7 (android only) (#2920)

* k-quants : fix build on armv7

* ggml : cleanup unused arm32 specific impl

* k-quants : avoid some unused vzero / mzero define

* ggml-alloc : use 4g for MEASURE_MAX_SIZE in 32-bit arm

ggml-alloc.c

@@ -284,7 +284,14 @@ struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment)
 // address and size of the buffer when measuring
 // it needs to be large enough to fit all the tensors, but it cannot overlap with other existing buffers
 static void * const MEASURE_BASE_ADDR = (void *) 0x1000;
+#if defined(__ARM_NEON) && !defined(__aarch64__)
+// 32-bit
+// TODO: Use for 32-bit x86 as well
+static const size_t MEASURE_MAX_SIZE  = (1ULL<<32) - 1; // 4 GB
+#else
+// 64-bit
 static const size_t MEASURE_MAX_SIZE  = 1ULL<<40; // 1 TB
+#endif
 
 struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) {
     struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */);

ggml.c

@@ -817,46 +817,6 @@ static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128
 
 #if !defined(__aarch64__)
 
-inline static uint16_t vaddvq_u8(uint8x16_t v) {
-    return
-        (uint16_t)vgetq_lane_u8(v, 0)  + (uint16_t)vgetq_lane_u8(v, 1)  +
-        (uint16_t)vgetq_lane_u8(v, 2)  + (uint16_t)vgetq_lane_u8(v, 3)  +
-        (uint16_t)vgetq_lane_u8(v, 4)  + (uint16_t)vgetq_lane_u8(v, 5)  +
-        (uint16_t)vgetq_lane_u8(v, 6)  + (uint16_t)vgetq_lane_u8(v, 7)  +
-        (uint16_t)vgetq_lane_u8(v, 8)  + (uint16_t)vgetq_lane_u8(v, 9)  +
-        (uint16_t)vgetq_lane_u8(v, 10) + (uint16_t)vgetq_lane_u8(v, 11) +
-        (uint16_t)vgetq_lane_u8(v, 12) + (uint16_t)vgetq_lane_u8(v, 13) +
-        (uint16_t)vgetq_lane_u8(v, 14) + (uint16_t)vgetq_lane_u8(v, 15);
-}
-
-inline static int16_t vaddvq_s8(int8x16_t v) {
-    return
-        (int16_t)vgetq_lane_s8(v, 0)  + (int16_t)vgetq_lane_s8(v, 1)  +
-        (int16_t)vgetq_lane_s8(v, 2)  + (int16_t)vgetq_lane_s8(v, 3)  +
-        (int16_t)vgetq_lane_s8(v, 4)  + (int16_t)vgetq_lane_s8(v, 5)  +
-        (int16_t)vgetq_lane_s8(v, 6)  + (int16_t)vgetq_lane_s8(v, 7)  +
-        (int16_t)vgetq_lane_s8(v, 8)  + (int16_t)vgetq_lane_s8(v, 9)  +
-        (int16_t)vgetq_lane_s8(v, 10) + (int16_t)vgetq_lane_s8(v, 11) +
-        (int16_t)vgetq_lane_s8(v, 12) + (int16_t)vgetq_lane_s8(v, 13) +
-        (int16_t)vgetq_lane_s8(v, 14) + (int16_t)vgetq_lane_s8(v, 15);
-}
-
-inline static int32_t vaddvq_s16(int16x8_t v) {
-    return
-        (int32_t)vgetq_lane_s16(v, 0) + (int32_t)vgetq_lane_s16(v, 1) +
-        (int32_t)vgetq_lane_s16(v, 2) + (int32_t)vgetq_lane_s16(v, 3) +
-        (int32_t)vgetq_lane_s16(v, 4) + (int32_t)vgetq_lane_s16(v, 5) +
-        (int32_t)vgetq_lane_s16(v, 6) + (int32_t)vgetq_lane_s16(v, 7);
-}
-
-inline static uint32_t vaddvq_u16(uint16x8_t v) {
-    return
-        (uint32_t)vgetq_lane_u16(v, 0) + (uint32_t)vgetq_lane_u16(v, 1) +
-        (uint32_t)vgetq_lane_u16(v, 2) + (uint32_t)vgetq_lane_u16(v, 3) +
-        (uint32_t)vgetq_lane_u16(v, 4) + (uint32_t)vgetq_lane_u16(v, 5) +
-        (uint32_t)vgetq_lane_u16(v, 6) + (uint32_t)vgetq_lane_u16(v, 7);
-}
-
 inline static int32_t vaddvq_s32(int32x4_t v) {
     return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
 }
@@ -865,12 +825,6 @@ inline static float vaddvq_f32(float32x4_t v) {
     return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
 }
 
-inline static float vminvq_f32(float32x4_t v) {
-    return
-        MIN(MIN(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
-            MIN(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
-}
-
 inline static float vmaxvq_f32(float32x4_t v) {
     return
         MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),

k_quants.c

@@ -13,6 +13,26 @@
 //
 #include <arm_neon.h>
 
+#if !defined(__aarch64__)
+inline static int32_t vaddvq_s16(int16x8_t v) {
+    return
+        (int32_t)vgetq_lane_s16(v, 0) + (int32_t)vgetq_lane_s16(v, 1) +
+        (int32_t)vgetq_lane_s16(v, 2) + (int32_t)vgetq_lane_s16(v, 3) +
+        (int32_t)vgetq_lane_s16(v, 4) + (int32_t)vgetq_lane_s16(v, 5) +
+        (int32_t)vgetq_lane_s16(v, 6) + (int32_t)vgetq_lane_s16(v, 7);
+}
+
+inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
+    int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
+    int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
+    return vcombine_s16(a0, b0);
+}
+
+inline static int32_t vaddvq_s32(int32x4_t v) {
+    return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
+}
+#endif
+
 #else
 
 #ifdef __wasm_simd128__
@@ -1302,7 +1322,9 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
 
     const uint8x16_t m3 = vdupq_n_u8(0x3);
     const uint8x16_t m4 = vdupq_n_u8(0xF);
+#if defined(__ARM_FEATURE_DOTPROD)
     const int32x4_t  vzero = vdupq_n_s32(0);
+#endif
 
     int8x16x2_t q2bytes;
     uint8_t aux[16];
@@ -1608,7 +1630,9 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
 #ifdef __ARM_NEON
 
     const uint8x16_t m3 = vdupq_n_u8(0x3);
+#if defined(__ARM_FEATURE_DOTPROD)
     const int32x4_t  vzero = vdupq_n_s32(0);
+#endif
 
     int8x16x4_t q2bytes;
 
@@ -2592,8 +2616,6 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
         const uint8_t * restrict q4 = x[i].qs;
         const int8_t  * restrict q8 = y[i].qs;
 
-        //int32x4_t isum = mzero;
-
         int32_t sumi1 = 0;
         int32_t sumi2 = 0;
 
@@ -3092,9 +3114,11 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
 #ifdef __ARM_NEON
 
     const uint8x16_t m4b = vdupq_n_u8(0xf);
-    const int32x4_t mzero = vdupq_n_s32(0);
     const uint8x16_t mone = vdupq_n_u8(1);
     const uint8x16_t mtwo = vdupq_n_u8(2);
+#if defined(__ARM_FEATURE_DOTPROD)
+    const int32x4_t mzero = vdupq_n_s32(0);
+#endif
 
     int8x16x4_t q5bytes;
 
@@ -3437,8 +3461,10 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
 #ifdef __ARM_NEON
 
     const uint8x16_t m4b = vdupq_n_u8(0xf);
-    const int32x4_t mzero = vdupq_n_s32(0);
     const uint8x16_t mh = vdupq_n_u8(16);
+#if defined(__ARM_FEATURE_DOTPROD)
+    const int32x4_t mzero = vdupq_n_s32(0);
+#endif
 
     int8x16x4_t q5bytes;
     uint8x16x4_t q5h;
@@ -3656,7 +3682,9 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
     float sum = 0;
 
     const uint8x16_t m4b = vdupq_n_u8(0xF);
+#if defined(__ARM_FEATURE_DOTPROD)
     const int32x4_t  vzero = vdupq_n_s32(0);
+#endif
     //const int8x16_t  m32s = vdupq_n_s8(32);
 
     const uint8x16_t mone = vdupq_n_u8(3);
@@ -4045,8 +4073,10 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
     float sum = 0;
 
     const uint8x16_t m4b = vdupq_n_u8(0xF);
-    const int32x4_t  vzero = vdupq_n_s32(0);
     const int8x16_t  m32s = vdupq_n_s8(32);
+#if defined(__ARM_FEATURE_DOTPROD)
+    const int32x4_t  vzero = vdupq_n_s32(0);
+#endif
 
     const uint8x16_t mone = vdupq_n_u8(3);