From: Georgi Gerganov <redacted>
Date: Sun, 12 Mar 2023 23:28:08 +0000 (+0200)
Subject: Revert "10% performance boost on ARM"
X-Git-Tag: gguf-v0.4.0~1260
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=c80e2a8f2adeda202cbffe76ef800f134e51f03f;p=pkg%2Fggml%2Fsources%2Fllama.cpp

Revert "10% performance boost on ARM"

This reverts commit 113a9e83ebc0f788f861394437087bf3ca0e019b.

There are some reports for illegal instruction.
Moved this stuff to vdotq_s32 branch until resolve
---

diff --git a/ggml.c b/ggml.c
index e664e718..fbd7b933 100644
--- a/ggml.c
+++ b/ggml.c
@@ -1360,22 +1360,6 @@ inline static void ggml_vec_dot_q4_0(const int n, float * restrict s, const void
         const int8x16_t v1_1hs = vsubq_s8(v1_1h, s8b);
 
         // dot product into int16x8_t
-#if defined(__ARM_FEATURE_DOTPROD)
-        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);
-
-        p_0 = vdotq_s32(p_0, v0_0hs, v1_0hs);
-        p_1 = vdotq_s32(p_1, v0_1hs, v1_1hs);
-
-        // scalar
-#if defined(__ARM_FEATURE_QRDMX)
-        sum0 += d0_0*d1_0*vaddvq_s32(p_0);
-        sum1 += d0_1*d1_1*vaddvq_s32(p_1);
-#else
-        sum0 += d0_0*d1_0*(vgetq_lane_s32(p_0, 0) + vgetq_lane_s32(p_0, 1) + vgetq_lane_s32(p_0, 2) + vgetq_lane_s32(p_0, 3));
-        sum1 += d0_1*d1_1*(vgetq_lane_s32(p_1, 0) + vgetq_lane_s32(p_1, 1) + vgetq_lane_s32(p_1, 2) + vgetq_lane_s32(p_1, 3));
-#endif
-#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));
 
@@ -1404,7 +1388,6 @@ inline static void ggml_vec_dot_q4_0(const int n, float * restrict s, const void
 #else
         sum0 += d0_0*d1_0*(vgetq_lane_s16(p_0, 0) + vgetq_lane_s16(p_0, 1) + vgetq_lane_s16(p_0, 2) + vgetq_lane_s16(p_0, 3) + vgetq_lane_s16(p_0, 4) + vgetq_lane_s16(p_0, 5) + vgetq_lane_s16(p_0, 6) + vgetq_lane_s16(p_0, 7));
         sum1 += d0_1*d1_1*(vgetq_lane_s16(p_1, 0) + vgetq_lane_s16(p_1, 1) + vgetq_lane_s16(p_1, 2) + vgetq_lane_s16(p_1, 3) + vgetq_lane_s16(p_1, 4) + vgetq_lane_s16(p_1, 5) + vgetq_lane_s16(p_1, 6) + vgetq_lane_s16(p_1, 7));
-#endif
 #endif
     }