Skip to content

Commit 68b4dfe

Browse files
ikawrakowIwan Kawrakow
authored and
Nexesenex
committed
Faster prompt processing for IQ2_KS, IQ2_K, IQ2_K_R4 (LostRuins#593)
* cuda: faster MMQ for iq2_ks, iq2_k, iq2_k_r4 * Lookup is still beter for MMQ if we get 4 values at once * Minor --------- And some corrections + non activated IKL LostRuins#559 Co-Authored-By: Iwan Kawrakow <[email protected]>
1 parent 646c331 commit 68b4dfe

File tree

5 files changed

+206
-230
lines changed

5 files changed

+206
-230
lines changed

‎ggml/src/ggml-cuda/iqk_cuda_common.h‎

Lines changed: 75 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,75 @@
1+
#pragma once
2+
3+
#include "common.cuh"
4+
5+
static const __device__ uint32_t iq2k_table[512] = {
6+
0xe1e1e1e1, 0xe1e1e1f3, 0xe1e1e101, 0xe1e1e111, 0xe1e1f3e1, 0xe1e1f3f3, 0xe1e1f301, 0xe1e1f311,
7+
0xe1e101e1, 0xe1e101f3, 0xe1e10101, 0xe1e10111, 0xe1e111e1, 0xe1e111f3, 0xe1e11101, 0xe1e11111,
8+
0xe1f3e1e1, 0xe1f3e1f3, 0xe1f3e101, 0xe1f3e111, 0xe1f3f3e1, 0xe1f3f3f3, 0xe1f3f301, 0xe1f3f311,
9+
0xe1f301e1, 0xe1f301f3, 0xe1f30101, 0xe1f30111, 0xe1f311e1, 0xe1f311f3, 0xe1f31101, 0xe1f31111,
10+
0xe101e1e1, 0xe101e1f3, 0xe101e101, 0xe101e111, 0xe101f3e1, 0xe101f3f3, 0xe101f301, 0xe101f311,
11+
0xe10101e1, 0xe10101f3, 0xe1010101, 0xe1010111, 0xe10111e1, 0xe10111f3, 0xe1011101, 0xe1011111,
12+
0xe111e1e1, 0xe111e1f3, 0xe111e101, 0xe111e111, 0xe111f3e1, 0xe111f3f3, 0xe111f301, 0xe111f311,
13+
0xe11101e1, 0xe11101f3, 0xe1110101, 0xe1110111, 0xe11111e1, 0xe11111f3, 0xe1111101, 0xe1111111,
14+
0xf3e1e1e1, 0xf3e1e1f3, 0xf3e1e101, 0xf3e1e111, 0xf3e1f3e1, 0xf3e1f3f3, 0xf3e1f301, 0xf3e1f311,
15+
0xf3e101e1, 0xf3e101f3, 0xf3e10101, 0xf3e10111, 0xf3e111e1, 0xf3e111f3, 0xf3e11101, 0xf3e11111,
16+
0xf3f3e1e1, 0xf3f3e1f3, 0xf3f3e101, 0xf3f3e111, 0xf3f3f3e1, 0xf3f3f3f3, 0xf3f3f301, 0xf3f3f311,
17+
0xf3f301e1, 0xf3f301f3, 0xf3f30101, 0xf3f30111, 0xf3f311e1, 0xf3f311f3, 0xf3f31101, 0xf3f31111,
18+
0xf301e1e1, 0xf301e1f3, 0xf301e101, 0xf301e111, 0xf301f3e1, 0xf301f3f3, 0xf301f301, 0xf301f311,
19+
0xf30101e1, 0xf30101f3, 0xf3010101, 0xf3010111, 0xf30111e1, 0xf30111f3, 0xf3011101, 0xf3011111,
20+
0xf311e1e1, 0xf311e1f3, 0xf311e101, 0xf311e111, 0xf311f3e1, 0xf311f3f3, 0xf311f301, 0xf311f311,
21+
0xf31101e1, 0xf31101f3, 0xf3110101, 0xf3110111, 0xf31111e1, 0xf31111f3, 0xf3111101, 0xf3111111,
22+
0x01e1e1e1, 0x01e1e1f3, 0x01e1e101, 0x01e1e111, 0x01e1f3e1, 0x01e1f3f3, 0x01e1f301, 0x01e1f311,
23+
0x01e101e1, 0x01e101f3, 0x01e10101, 0x01e10111, 0x01e111e1, 0x01e111f3, 0x01e11101, 0x01e11111,
24+
0x01f3e1e1, 0x01f3e1f3, 0x01f3e101, 0x01f3e111, 0x01f3f3e1, 0x01f3f3f3, 0x01f3f301, 0x01f3f311,
25+
0x01f301e1, 0x01f301f3, 0x01f30101, 0x01f30111, 0x01f311e1, 0x01f311f3, 0x01f31101, 0x01f31111,
26+
0x0101e1e1, 0x0101e1f3, 0x0101e101, 0x0101e111, 0x0101f3e1, 0x0101f3f3, 0x0101f301, 0x0101f311,
27+
0x010101e1, 0x010101f3, 0x01010101, 0x01010111, 0x010111e1, 0x010111f3, 0x01011101, 0x01011111,
28+
0x0111e1e1, 0x0111e1f3, 0x0111e101, 0x0111e111, 0x0111f3e1, 0x0111f3f3, 0x0111f301, 0x0111f311,
29+
0x011101e1, 0x011101f3, 0x01110101, 0x01110111, 0x011111e1, 0x011111f3, 0x01111101, 0x01111111,
30+
0x11e1e1e1, 0x11e1e1f3, 0x11e1e101, 0x11e1e111, 0x11e1f3e1, 0x11e1f3f3, 0x11e1f301, 0x11e1f311,
31+
0x11e101e1, 0x11e101f3, 0x11e10101, 0x11e10111, 0x11e111e1, 0x11e111f3, 0x11e11101, 0x11e11111,
32+
0x11f3e1e1, 0x11f3e1f3, 0x11f3e101, 0x11f3e111, 0x11f3f3e1, 0x11f3f3f3, 0x11f3f301, 0x11f3f311,
33+
0x11f301e1, 0x11f301f3, 0x11f30101, 0x11f30111, 0x11f311e1, 0x11f311f3, 0x11f31101, 0x11f31111,
34+
0x1101e1e1, 0x1101e1f3, 0x1101e101, 0x1101e111, 0x1101f3e1, 0x1101f3f3, 0x1101f301, 0x1101f311,
35+
0x110101e1, 0x110101f3, 0x11010101, 0x11010111, 0x110111e1, 0x110111f3, 0x11011101, 0x11011111,
36+
0x1111e1e1, 0x1111e1f3, 0x1111e101, 0x1111e111, 0x1111f3e1, 0x1111f3f3, 0x1111f301, 0x1111f311,
37+
0x111101e1, 0x111101f3, 0x11110101, 0x11110111, 0x111111e1, 0x111111f3, 0x11111101, 0x11111111,
38+
0xe6e6e6e6, 0xe6e6e6f8, 0xe6e6e606, 0xe6e6e616, 0xe6e6f8e6, 0xe6e6f8f8, 0xe6e6f806, 0xe6e6f816,
39+
0xe6e606e6, 0xe6e606f8, 0xe6e60606, 0xe6e60616, 0xe6e616e6, 0xe6e616f8, 0xe6e61606, 0xe6e61616,
40+
0xe6f8e6e6, 0xe6f8e6f8, 0xe6f8e606, 0xe6f8e616, 0xe6f8f8e6, 0xe6f8f8f8, 0xe6f8f806, 0xe6f8f816,
41+
0xe6f806e6, 0xe6f806f8, 0xe6f80606, 0xe6f80616, 0xe6f816e6, 0xe6f816f8, 0xe6f81606, 0xe6f81616,
42+
0xe606e6e6, 0xe606e6f8, 0xe606e606, 0xe606e616, 0xe606f8e6, 0xe606f8f8, 0xe606f806, 0xe606f816,
43+
0xe60606e6, 0xe60606f8, 0xe6060606, 0xe6060616, 0xe60616e6, 0xe60616f8, 0xe6061606, 0xe6061616,
44+
0xe616e6e6, 0xe616e6f8, 0xe616e606, 0xe616e616, 0xe616f8e6, 0xe616f8f8, 0xe616f806, 0xe616f816,
45+
0xe61606e6, 0xe61606f8, 0xe6160606, 0xe6160616, 0xe61616e6, 0xe61616f8, 0xe6161606, 0xe6161616,
46+
0xf8e6e6e6, 0xf8e6e6f8, 0xf8e6e606, 0xf8e6e616, 0xf8e6f8e6, 0xf8e6f8f8, 0xf8e6f806, 0xf8e6f816,
47+
0xf8e606e6, 0xf8e606f8, 0xf8e60606, 0xf8e60616, 0xf8e616e6, 0xf8e616f8, 0xf8e61606, 0xf8e61616,
48+
0xf8f8e6e6, 0xf8f8e6f8, 0xf8f8e606, 0xf8f8e616, 0xf8f8f8e6, 0xf8f8f8f8, 0xf8f8f806, 0xf8f8f816,
49+
0xf8f806e6, 0xf8f806f8, 0xf8f80606, 0xf8f80616, 0xf8f816e6, 0xf8f816f8, 0xf8f81606, 0xf8f81616,
50+
0xf806e6e6, 0xf806e6f8, 0xf806e606, 0xf806e616, 0xf806f8e6, 0xf806f8f8, 0xf806f806, 0xf806f816,
51+
0xf80606e6, 0xf80606f8, 0xf8060606, 0xf8060616, 0xf80616e6, 0xf80616f8, 0xf8061606, 0xf8061616,
52+
0xf816e6e6, 0xf816e6f8, 0xf816e606, 0xf816e616, 0xf816f8e6, 0xf816f8f8, 0xf816f806, 0xf816f816,
53+
0xf81606e6, 0xf81606f8, 0xf8160606, 0xf8160616, 0xf81616e6, 0xf81616f8, 0xf8161606, 0xf8161616,
54+
0x06e6e6e6, 0x06e6e6f8, 0x06e6e606, 0x06e6e616, 0x06e6f8e6, 0x06e6f8f8, 0x06e6f806, 0x06e6f816,
55+
0x06e606e6, 0x06e606f8, 0x06e60606, 0x06e60616, 0x06e616e6, 0x06e616f8, 0x06e61606, 0x06e61616,
56+
0x06f8e6e6, 0x06f8e6f8, 0x06f8e606, 0x06f8e616, 0x06f8f8e6, 0x06f8f8f8, 0x06f8f806, 0x06f8f816,
57+
0x06f806e6, 0x06f806f8, 0x06f80606, 0x06f80616, 0x06f816e6, 0x06f816f8, 0x06f81606, 0x06f81616,
58+
0x0606e6e6, 0x0606e6f8, 0x0606e606, 0x0606e616, 0x0606f8e6, 0x0606f8f8, 0x0606f806, 0x0606f816,
59+
0x060606e6, 0x060606f8, 0x06060606, 0x06060616, 0x060616e6, 0x060616f8, 0x06061606, 0x06061616,
60+
0x0616e6e6, 0x0616e6f8, 0x0616e606, 0x0616e616, 0x0616f8e6, 0x0616f8f8, 0x0616f806, 0x0616f816,
61+
0x061606e6, 0x061606f8, 0x06160606, 0x06160616, 0x061616e6, 0x061616f8, 0x06161606, 0x06161616,
62+
0x16e6e6e6, 0x16e6e6f8, 0x16e6e606, 0x16e6e616, 0x16e6f8e6, 0x16e6f8f8, 0x16e6f806, 0x16e6f816,
63+
0x16e606e6, 0x16e606f8, 0x16e60606, 0x16e60616, 0x16e616e6, 0x16e616f8, 0x16e61606, 0x16e61616,
64+
0x16f8e6e6, 0x16f8e6f8, 0x16f8e606, 0x16f8e616, 0x16f8f8e6, 0x16f8f8f8, 0x16f8f806, 0x16f8f816,
65+
0x16f806e6, 0x16f806f8, 0x16f80606, 0x16f80616, 0x16f816e6, 0x16f816f8, 0x16f81606, 0x16f81616,
66+
0x1606e6e6, 0x1606e6f8, 0x1606e606, 0x1606e616, 0x1606f8e6, 0x1606f8f8, 0x1606f806, 0x1606f816,
67+
0x160606e6, 0x160606f8, 0x16060606, 0x16060616, 0x160616e6, 0x160616f8, 0x16061606, 0x16061616,
68+
0x1616e6e6, 0x1616e6f8, 0x1616e606, 0x1616e616, 0x1616f8e6, 0x1616f8f8, 0x1616f806, 0x1616f816,
69+
0x161606e6, 0x161606f8, 0x16160606, 0x16160616, 0x161616e6, 0x161616f8, 0x16161606, 0x16161616,
70+
};
71+
72+
__device__ __forceinline__ int int_from_table_4(const uint32_t idx, const int * values) {
73+
return values[ggml_cuda_dp4a(idx, 0x40100401, 0)];
74+
}
75+

‎ggml/src/ggml-cuda/iqk_mmvq.cu‎

Lines changed: 21 additions & 95 deletions
Original file line numberDiff line numberDiff line change
@@ -5,6 +5,7 @@
55
//
66

77
#include "iqk_mmvq.cuh"
8+
#include "iqk_cuda_common.h"
89

910
// typedef float (*vec_dot_q_cuda_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs);
1011

@@ -787,77 +788,6 @@ __device__ __forceinline__ void vec_dot_iq6_k_q8_1(
787788
*result += d6 * (__low2float(bq8_1[2*(i4/2)+0].ds) * sumi1 * bq6->scales[4*(i4/2)+(i4%2)] + __low2float(bq8_1[2*(i4/2)+1].ds) * sumi2 * bq6->scales[4*(i4/2)+(i4%2)+2]);
788789
}
789790

790-
static const __device__ uint32_t iq2k_table[512] = {
791-
0xe1e1e1e1, 0xe1e1e1f3, 0xe1e1e101, 0xe1e1e111, 0xe1e1f3e1, 0xe1e1f3f3, 0xe1e1f301, 0xe1e1f311,
792-
0xe1e101e1, 0xe1e101f3, 0xe1e10101, 0xe1e10111, 0xe1e111e1, 0xe1e111f3, 0xe1e11101, 0xe1e11111,
793-
0xe1f3e1e1, 0xe1f3e1f3, 0xe1f3e101, 0xe1f3e111, 0xe1f3f3e1, 0xe1f3f3f3, 0xe1f3f301, 0xe1f3f311,
794-
0xe1f301e1, 0xe1f301f3, 0xe1f30101, 0xe1f30111, 0xe1f311e1, 0xe1f311f3, 0xe1f31101, 0xe1f31111,
795-
0xe101e1e1, 0xe101e1f3, 0xe101e101, 0xe101e111, 0xe101f3e1, 0xe101f3f3, 0xe101f301, 0xe101f311,
796-
0xe10101e1, 0xe10101f3, 0xe1010101, 0xe1010111, 0xe10111e1, 0xe10111f3, 0xe1011101, 0xe1011111,
797-
0xe111e1e1, 0xe111e1f3, 0xe111e101, 0xe111e111, 0xe111f3e1, 0xe111f3f3, 0xe111f301, 0xe111f311,
798-
0xe11101e1, 0xe11101f3, 0xe1110101, 0xe1110111, 0xe11111e1, 0xe11111f3, 0xe1111101, 0xe1111111,
799-
0xf3e1e1e1, 0xf3e1e1f3, 0xf3e1e101, 0xf3e1e111, 0xf3e1f3e1, 0xf3e1f3f3, 0xf3e1f301, 0xf3e1f311,
800-
0xf3e101e1, 0xf3e101f3, 0xf3e10101, 0xf3e10111, 0xf3e111e1, 0xf3e111f3, 0xf3e11101, 0xf3e11111,
801-
0xf3f3e1e1, 0xf3f3e1f3, 0xf3f3e101, 0xf3f3e111, 0xf3f3f3e1, 0xf3f3f3f3, 0xf3f3f301, 0xf3f3f311,
802-
0xf3f301e1, 0xf3f301f3, 0xf3f30101, 0xf3f30111, 0xf3f311e1, 0xf3f311f3, 0xf3f31101, 0xf3f31111,
803-
0xf301e1e1, 0xf301e1f3, 0xf301e101, 0xf301e111, 0xf301f3e1, 0xf301f3f3, 0xf301f301, 0xf301f311,
804-
0xf30101e1, 0xf30101f3, 0xf3010101, 0xf3010111, 0xf30111e1, 0xf30111f3, 0xf3011101, 0xf3011111,
805-
0xf311e1e1, 0xf311e1f3, 0xf311e101, 0xf311e111, 0xf311f3e1, 0xf311f3f3, 0xf311f301, 0xf311f311,
806-
0xf31101e1, 0xf31101f3, 0xf3110101, 0xf3110111, 0xf31111e1, 0xf31111f3, 0xf3111101, 0xf3111111,
807-
0x01e1e1e1, 0x01e1e1f3, 0x01e1e101, 0x01e1e111, 0x01e1f3e1, 0x01e1f3f3, 0x01e1f301, 0x01e1f311,
808-
0x01e101e1, 0x01e101f3, 0x01e10101, 0x01e10111, 0x01e111e1, 0x01e111f3, 0x01e11101, 0x01e11111,
809-
0x01f3e1e1, 0x01f3e1f3, 0x01f3e101, 0x01f3e111, 0x01f3f3e1, 0x01f3f3f3, 0x01f3f301, 0x01f3f311,
810-
0x01f301e1, 0x01f301f3, 0x01f30101, 0x01f30111, 0x01f311e1, 0x01f311f3, 0x01f31101, 0x01f31111,
811-
0x0101e1e1, 0x0101e1f3, 0x0101e101, 0x0101e111, 0x0101f3e1, 0x0101f3f3, 0x0101f301, 0x0101f311,
812-
0x010101e1, 0x010101f3, 0x01010101, 0x01010111, 0x010111e1, 0x010111f3, 0x01011101, 0x01011111,
813-
0x0111e1e1, 0x0111e1f3, 0x0111e101, 0x0111e111, 0x0111f3e1, 0x0111f3f3, 0x0111f301, 0x0111f311,
814-
0x011101e1, 0x011101f3, 0x01110101, 0x01110111, 0x011111e1, 0x011111f3, 0x01111101, 0x01111111,
815-
0x11e1e1e1, 0x11e1e1f3, 0x11e1e101, 0x11e1e111, 0x11e1f3e1, 0x11e1f3f3, 0x11e1f301, 0x11e1f311,
816-
0x11e101e1, 0x11e101f3, 0x11e10101, 0x11e10111, 0x11e111e1, 0x11e111f3, 0x11e11101, 0x11e11111,
817-
0x11f3e1e1, 0x11f3e1f3, 0x11f3e101, 0x11f3e111, 0x11f3f3e1, 0x11f3f3f3, 0x11f3f301, 0x11f3f311,
818-
0x11f301e1, 0x11f301f3, 0x11f30101, 0x11f30111, 0x11f311e1, 0x11f311f3, 0x11f31101, 0x11f31111,
819-
0x1101e1e1, 0x1101e1f3, 0x1101e101, 0x1101e111, 0x1101f3e1, 0x1101f3f3, 0x1101f301, 0x1101f311,
820-
0x110101e1, 0x110101f3, 0x11010101, 0x11010111, 0x110111e1, 0x110111f3, 0x11011101, 0x11011111,
821-
0x1111e1e1, 0x1111e1f3, 0x1111e101, 0x1111e111, 0x1111f3e1, 0x1111f3f3, 0x1111f301, 0x1111f311,
822-
0x111101e1, 0x111101f3, 0x11110101, 0x11110111, 0x111111e1, 0x111111f3, 0x11111101, 0x11111111,
823-
0xe6e6e6e6, 0xe6e6e6f8, 0xe6e6e606, 0xe6e6e616, 0xe6e6f8e6, 0xe6e6f8f8, 0xe6e6f806, 0xe6e6f816,
824-
0xe6e606e6, 0xe6e606f8, 0xe6e60606, 0xe6e60616, 0xe6e616e6, 0xe6e616f8, 0xe6e61606, 0xe6e61616,
825-
0xe6f8e6e6, 0xe6f8e6f8, 0xe6f8e606, 0xe6f8e616, 0xe6f8f8e6, 0xe6f8f8f8, 0xe6f8f806, 0xe6f8f816,
826-
0xe6f806e6, 0xe6f806f8, 0xe6f80606, 0xe6f80616, 0xe6f816e6, 0xe6f816f8, 0xe6f81606, 0xe6f81616,
827-
0xe606e6e6, 0xe606e6f8, 0xe606e606, 0xe606e616, 0xe606f8e6, 0xe606f8f8, 0xe606f806, 0xe606f816,
828-
0xe60606e6, 0xe60606f8, 0xe6060606, 0xe6060616, 0xe60616e6, 0xe60616f8, 0xe6061606, 0xe6061616,
829-
0xe616e6e6, 0xe616e6f8, 0xe616e606, 0xe616e616, 0xe616f8e6, 0xe616f8f8, 0xe616f806, 0xe616f816,
830-
0xe61606e6, 0xe61606f8, 0xe6160606, 0xe6160616, 0xe61616e6, 0xe61616f8, 0xe6161606, 0xe6161616,
831-
0xf8e6e6e6, 0xf8e6e6f8, 0xf8e6e606, 0xf8e6e616, 0xf8e6f8e6, 0xf8e6f8f8, 0xf8e6f806, 0xf8e6f816,
832-
0xf8e606e6, 0xf8e606f8, 0xf8e60606, 0xf8e60616, 0xf8e616e6, 0xf8e616f8, 0xf8e61606, 0xf8e61616,
833-
0xf8f8e6e6, 0xf8f8e6f8, 0xf8f8e606, 0xf8f8e616, 0xf8f8f8e6, 0xf8f8f8f8, 0xf8f8f806, 0xf8f8f816,
834-
0xf8f806e6, 0xf8f806f8, 0xf8f80606, 0xf8f80616, 0xf8f816e6, 0xf8f816f8, 0xf8f81606, 0xf8f81616,
835-
0xf806e6e6, 0xf806e6f8, 0xf806e606, 0xf806e616, 0xf806f8e6, 0xf806f8f8, 0xf806f806, 0xf806f816,
836-
0xf80606e6, 0xf80606f8, 0xf8060606, 0xf8060616, 0xf80616e6, 0xf80616f8, 0xf8061606, 0xf8061616,
837-
0xf816e6e6, 0xf816e6f8, 0xf816e606, 0xf816e616, 0xf816f8e6, 0xf816f8f8, 0xf816f806, 0xf816f816,
838-
0xf81606e6, 0xf81606f8, 0xf8160606, 0xf8160616, 0xf81616e6, 0xf81616f8, 0xf8161606, 0xf8161616,
839-
0x06e6e6e6, 0x06e6e6f8, 0x06e6e606, 0x06e6e616, 0x06e6f8e6, 0x06e6f8f8, 0x06e6f806, 0x06e6f816,
840-
0x06e606e6, 0x06e606f8, 0x06e60606, 0x06e60616, 0x06e616e6, 0x06e616f8, 0x06e61606, 0x06e61616,
841-
0x06f8e6e6, 0x06f8e6f8, 0x06f8e606, 0x06f8e616, 0x06f8f8e6, 0x06f8f8f8, 0x06f8f806, 0x06f8f816,
842-
0x06f806e6, 0x06f806f8, 0x06f80606, 0x06f80616, 0x06f816e6, 0x06f816f8, 0x06f81606, 0x06f81616,
843-
0x0606e6e6, 0x0606e6f8, 0x0606e606, 0x0606e616, 0x0606f8e6, 0x0606f8f8, 0x0606f806, 0x0606f816,
844-
0x060606e6, 0x060606f8, 0x06060606, 0x06060616, 0x060616e6, 0x060616f8, 0x06061606, 0x06061616,
845-
0x0616e6e6, 0x0616e6f8, 0x0616e606, 0x0616e616, 0x0616f8e6, 0x0616f8f8, 0x0616f806, 0x0616f816,
846-
0x061606e6, 0x061606f8, 0x06160606, 0x06160616, 0x061616e6, 0x061616f8, 0x06161606, 0x06161616,
847-
0x16e6e6e6, 0x16e6e6f8, 0x16e6e606, 0x16e6e616, 0x16e6f8e6, 0x16e6f8f8, 0x16e6f806, 0x16e6f816,
848-
0x16e606e6, 0x16e606f8, 0x16e60606, 0x16e60616, 0x16e616e6, 0x16e616f8, 0x16e61606, 0x16e61616,
849-
0x16f8e6e6, 0x16f8e6f8, 0x16f8e606, 0x16f8e616, 0x16f8f8e6, 0x16f8f8f8, 0x16f8f806, 0x16f8f816,
850-
0x16f806e6, 0x16f806f8, 0x16f80606, 0x16f80616, 0x16f816e6, 0x16f816f8, 0x16f81606, 0x16f81616,
851-
0x1606e6e6, 0x1606e6f8, 0x1606e606, 0x1606e616, 0x1606f8e6, 0x1606f8f8, 0x1606f806, 0x1606f816,
852-
0x160606e6, 0x160606f8, 0x16060606, 0x16060616, 0x160616e6, 0x160616f8, 0x16061606, 0x16061616,
853-
0x1616e6e6, 0x1616e6f8, 0x1616e606, 0x1616e616, 0x1616f8e6, 0x1616f8f8, 0x1616f806, 0x1616f816,
854-
0x161606e6, 0x161606f8, 0x16160606, 0x16160616, 0x161616e6, 0x161616f8, 0x16161606, 0x16161616,
855-
};
856-
857-
__device__ __forceinline__ int int_from_table_4(const uint8_t * a8, const int * values) {
858-
return values[a8[0] | (a8[1] << 2) | (a8[2] << 4) | (a8[3] << 6)];
859-
}
860-
861791
#define VDR_IQ2_K_Q8_1_MMVQ 4
862792
#define VDR_IQ2_K_Q8_1_MMQ 4
863793

@@ -883,7 +813,6 @@ __device__ __forceinline__ void vec_dot_iq2_k_q8_1(
883813
uint32_t val1 = q2[0], val2 = q2[1];
884814

885815
uint32_t aux32[2];
886-
const uint8_t * a8 = (const uint8_t *)&aux32;
887816
int v1, v2;
888817

889818
// Block of 16: (32*(4*(i4/4)+k)+8*(i4%4))/16 = 8*(i4/4) + 2*k + (i4%4)/2
@@ -894,23 +823,23 @@ __device__ __forceinline__ void vec_dot_iq2_k_q8_1(
894823
const int8_t * s8 = (const int8_t *)&s32;
895824

896825
aux32[0] = ((val1 >> 0) & 0x03030303); aux32[1] = ((val2 >> 0) & 0x03030303); values = all_values + ((extra & 0x01) << 8);
897-
v1 = int_from_table_4(a8 + 0, values);
898-
v2 = int_from_table_4(a8 + 4, values);
826+
v1 = int_from_table_4(aux32[0], values);
827+
v2 = int_from_table_4(aux32[1], values);
899828
int sumi1 = ggml_cuda_dp4a(v2, q8_1[1], ggml_cuda_dp4a(v1, q8_1[0], 0)) * s8[0];
900829

901830
aux32[0] = ((val1 >> 2) & 0x03030303); aux32[1] = ((val2 >> 2) & 0x03030303); values = all_values + ((extra & 0x04) << 6);
902-
v1 = int_from_table_4(a8 + 0, values);
903-
v2 = int_from_table_4(a8 + 4, values);
831+
v1 = int_from_table_4(aux32[0], values);
832+
v2 = int_from_table_4(aux32[1], values);
904833
int sumi2 = ggml_cuda_dp4a(v2, q8_2[1], ggml_cuda_dp4a(v1, q8_2[0], 0)) * s8[1];
905834

906835
aux32[0] = ((val1 >> 4) & 0x03030303); aux32[1] = ((val2 >> 4) & 0x03030303); values = all_values + ((extra & 0x10) << 4);
907-
v1 = int_from_table_4(a8 + 0, values);
908-
v2 = int_from_table_4(a8 + 4, values);
836+
v1 = int_from_table_4(aux32[0], values);
837+
v2 = int_from_table_4(aux32[1], values);
909838
int sumi3 = ggml_cuda_dp4a(v2, q8_3[1], ggml_cuda_dp4a(v1, q8_3[0], 0)) * s8[2];
910839

911840
aux32[0] = ((val1 >> 6) & 0x03030303); aux32[1] = ((val2 >> 6) & 0x03030303); values = all_values + ((extra & 0x40) << 2);
912-
v1 = int_from_table_4(a8 + 0, values);
913-
v2 = int_from_table_4(a8 + 4, values);
841+
v1 = int_from_table_4(aux32[0], values);
842+
v2 = int_from_table_4(aux32[1], values);
914843
int sumi4 = ggml_cuda_dp4a(v2, q8_4[1], ggml_cuda_dp4a(v1, q8_4[0], 0)) * s8[3];
915844

916845
*result += __half2float(bq2->d) * (__low2float(bq8_1[4*(i4/4)+0].ds) * sumi1
@@ -943,7 +872,6 @@ __device__ __forceinline__ void vec_dot_iq2_ks_q8_1(
943872
uint32_t val1 = q2[0] | (q2[1] << 16), val2 = q2[2] | (q2[3] << 16);
944873

945874
uint32_t aux32[2];
946-
const uint8_t * a8 = (const uint8_t *)&aux32;
947875
int v1, v2;
948876

949877
int32_t scales32;
@@ -956,23 +884,23 @@ __device__ __forceinline__ void vec_dot_iq2_ks_q8_1(
956884
s8[3] += ((extra >> 7) & 0x10);
957885

958886
aux32[0] = ((val1 >> 0) & 0x03030303); aux32[1] = ((val2 >> 0) & 0x03030303); values = all_values + ((extra & 0x01) << 8);
959-
v1 = int_from_table_4(a8 + 0, values);
960-
v2 = int_from_table_4(a8 + 4, values);
887+
v1 = int_from_table_4(aux32[0], values);
888+
v2 = int_from_table_4(aux32[1], values);
961889
int sumi1 = ggml_cuda_dp4a(v2, q8_1[1], ggml_cuda_dp4a(v1, q8_1[0], 0)) * s8[0];
962890

963891
aux32[0] = ((val1 >> 2) & 0x03030303); aux32[1] = ((val2 >> 2) & 0x03030303); values = all_values + ((extra & 0x02) << 7);
964-
v1 = int_from_table_4(a8 + 0, values);
965-
v2 = int_from_table_4(a8 + 4, values);
892+
v1 = int_from_table_4(aux32[0], values);
893+
v2 = int_from_table_4(aux32[1], values);
966894
int sumi2 = ggml_cuda_dp4a(v2, q8_2[1], ggml_cuda_dp4a(v1, q8_2[0], 0)) * s8[2];
967895

968896
aux32[0] = ((val1 >> 4) & 0x03030303); aux32[1] = ((val2 >> 4) & 0x03030303); values = all_values + ((extra & 0x04) << 6);
969-
v1 = int_from_table_4(a8 + 0, values);
970-
v2 = int_from_table_4(a8 + 4, values);
897+
v1 = int_from_table_4(aux32[0], values);
898+
v2 = int_from_table_4(aux32[1], values);
971899
int sumi3 = ggml_cuda_dp4a(v2, q8_3[1], ggml_cuda_dp4a(v1, q8_3[0], 0)) * s8[1];
972900

973901
aux32[0] = ((val1 >> 6) & 0x03030303); aux32[1] = ((val2 >> 6) & 0x03030303); values = all_values + ((extra & 0x08) << 5);
974-
v1 = int_from_table_4(a8 + 0, values);
975-
v2 = int_from_table_4(a8 + 4, values);
902+
v1 = int_from_table_4(aux32[0], values);
903+
v2 = int_from_table_4(aux32[1], values);
976904
int sumi4 = ggml_cuda_dp4a(v2, q8_4[1], ggml_cuda_dp4a(v1, q8_4[0], 0)) * s8[3];
977905

978906
*result += scale * (__low2float(bq8_1[4*(i4/4)+0].ds) * sumi1
@@ -1002,20 +930,19 @@ __device__ __forceinline__ void vec_dot_iq2_k_r4_q8_1(
1002930
int2 val1;
1003931
const int * q2 = (const int *)bq2->qs + 8*ib32 + 4*is;
1004932
int aux32[2];
1005-
const uint8_t * aux8 = (const uint8_t *)aux32;
1006933
#pragma unroll
1007934
for (int i = 0; i < 4; ++i) {
1008935
auto values1 = all_values + (((bq2->extra[i+4*is] >> ib32) & 1) << 8);
1009936
int sumi1 = 0;
1010937
aux32[0] = ((q2[i] >> 0) & 0x03030303);
1011938
aux32[1] = ((q2[i] >> 2) & 0x03030303);
1012-
val1.x = int_from_table_4(aux8+0, values1);
1013-
val1.y = int_from_table_4(aux8+4, values1);
939+
val1.x = int_from_table_4(aux32[0], values1);
940+
val1.y = int_from_table_4(aux32[1], values1);
1014941
sumi1 = ggml_cuda_dp4a(val1.x, q8[0], ggml_cuda_dp4a(val1.y, q8[1], sumi1));
1015942
aux32[0] = ((q2[i] >> 4) & 0x03030303);
1016943
aux32[1] = ((q2[i] >> 6) & 0x03030303);
1017-
val1.x = int_from_table_4(aux8+0, values1);
1018-
val1.y = int_from_table_4(aux8+4, values1);
944+
val1.x = int_from_table_4(aux32[0], values1);
945+
val1.y = int_from_table_4(aux32[1], values1);
1019946
sumi1 = ggml_cuda_dp4a(val1.x, q8[2], ggml_cuda_dp4a(val1.y, q8[3], sumi1));
1020947
const float d = __half2float(bq2->d[i]) * d8;
1021948
result[i] += d * sumi1 * s8[i];
@@ -1177,7 +1104,6 @@ __device__ __forceinline__ void vec_dot_iq3_ks_q8_1(
11771104
const int ib128 = iqs/4; // 0 or 1. 0 works on quants 0...127, 1 on quants 128...255
11781105
// Each thread processes 8 quants in each of the 4 32-blocks
11791106
const int il8 = iqs%4; // 0...3. 0 works on quants 0...7, 1 on quants 8...15, 2 on 16...23, 3 on 24...31
1180-
const int shift = 4*(il8/2);
11811107

11821108
const uint16_t * ql = (const uint16_t *)bq3->qs + 16*ib128 + 4*il8;
11831109
const uint16_t * qh = (const uint16_t *)bq3->qh + 4*il8;

0 commit comments

Comments
 (0)