Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-26 06:10:29 +01:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: define all quant data structures in types.comp (#10440)

Browse Source
This commit is contained in:
Jeff Bolz 2024-11-27 01:32:54 -06:00 committed by GitHub
parent 5b3466bedf
commit c31ed2abfc
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 76 additions and 71 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

147
ggml/src/ggml-vulkan/vulkan-shaders/types.comp
View File

@ -30,10 +30,8 @@
#endif
#endif
#if defined(DATA_A_Q4_0)
#extension GL_EXT_shader_16bit_storage : require
#define QUANT_K 32
#define QUANT_R 2
#define QUANT_K_Q4_0 32
#define QUANT_R_Q4_0 2
struct block_q4_0
{
@ -46,14 +44,15 @@ struct block_q4_0_packed16
uint16_t qs[16/2];
};
#if defined(DATA_A_Q4_0)
#define QUANT_K QUANT_K_Q4_0
#define QUANT_R QUANT_R_Q4_0
#define A_TYPE block_q4_0
#define A_TYPE_PACKED16 block_q4_0_packed16
#endif
#if defined(DATA_A_Q4_1)
#extension GL_EXT_shader_16bit_storage : require
#define QUANT_K 32
#define QUANT_R 2
#define QUANT_K_Q4_1 32
#define QUANT_R_Q4_1 2
struct block_q4_1
{
@ -69,15 +68,15 @@ struct block_q4_1_packed16
uint16_t qs[16/2];
};
#if defined(DATA_A_Q4_1)
#define QUANT_K QUANT_K_Q4_1
#define QUANT_R QUANT_R_Q4_1
#define A_TYPE block_q4_1
#define A_TYPE_PACKED16 block_q4_1_packed16
#endif
#if defined(DATA_A_Q5_0)
#extension GL_EXT_shader_16bit_storage : require
#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
#define QUANT_K 32
#define QUANT_R 2
#define QUANT_K_Q5_0 32
#define QUANT_R_Q5_0 2
struct block_q5_0
{
@ -93,15 +92,15 @@ struct block_q5_0_packed16
uint16_t qs[16/2];
};
#if defined(DATA_A_Q5_0)
#define QUANT_K QUANT_K_Q5_0
#define QUANT_R QUANT_R_Q5_0
#define A_TYPE block_q5_0
#define A_TYPE_PACKED16 block_q5_0_packed16
#endif
#if defined(DATA_A_Q5_1)
#extension GL_EXT_shader_16bit_storage : require
#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
#define QUANT_K 32
#define QUANT_R 2
#define QUANT_K_Q5_1 32
#define QUANT_R_Q5_1 2
struct block_q5_1
{
@ -119,14 +118,15 @@ struct block_q5_1_packed16
uint16_t qs[16/2];
};
#if defined(DATA_A_Q5_1)
#define QUANT_K QUANT_K_Q5_1
#define QUANT_R QUANT_R_Q5_1
#define A_TYPE block_q5_1
#define A_TYPE_PACKED16 block_q5_1_packed16
#endif
#if defined(DATA_A_Q8_0)
#extension GL_EXT_shader_16bit_storage : require
#define QUANT_K 32
#define QUANT_R 1
#define QUANT_K_Q8_0 32
#define QUANT_R_Q8_0 1
struct block_q8_0
{
@ -139,164 +139,164 @@ struct block_q8_0_packed16
uint16_t qs[32/2];
};
#if defined(DATA_A_Q8_0)
#define QUANT_K QUANT_K_Q8_0
#define QUANT_R QUANT_R_Q8_0
#define A_TYPE block_q8_0
#define A_TYPE_PACKED16 block_q8_0_packed16
#endif
// K-quants
#if defined(DATA_A_Q2_K)
#extension GL_EXT_shader_16bit_storage : require
#define QUANT_K 256
#define QUANT_K_Q2_K 256
struct block_q2_K
{
uint8_t scales[QUANT_K/16];
uint8_t qs[QUANT_K/4];
uint8_t scales[QUANT_K_Q2_K/16];
uint8_t qs[QUANT_K_Q2_K/4];
f16vec2 d;
};
struct block_q2_K_packed16
{
uint16_t scales[QUANT_K/16/2];
uint16_t qs[QUANT_K/4/2];
uint16_t scales[QUANT_K_Q2_K/16/2];
uint16_t qs[QUANT_K_Q2_K/4/2];
f16vec2 d;
};
struct block_q2_K_packed32
{
uint32_t scales[QUANT_K/16/4];
uint32_t qs[QUANT_K/4/4];
uint32_t scales[QUANT_K_Q2_K/16/4];
uint32_t qs[QUANT_K_Q2_K/4/4];
f16vec2 d;
};
#if defined(DATA_A_Q2_K)
#define QUANT_K QUANT_K_Q2_K
#define A_TYPE block_q2_K
#define A_TYPE_PACKED16 block_q2_K_packed16
#define A_TYPE_PACKED32 block_q2_K_packed32
#endif
#if defined(DATA_A_Q3_K)
#extension GL_EXT_shader_16bit_storage : require
#define QUANT_K 256
#define QUANT_K_Q3_K 256
struct block_q3_K
{
uint8_t hmask[QUANT_K/8];
uint8_t qs[QUANT_K/4];
uint8_t hmask[QUANT_K_Q3_K/8];
uint8_t qs[QUANT_K_Q3_K/4];
uint8_t scales[12];
float16_t d;
};
struct block_q3_K_packed16
{
uint16_t hmask[QUANT_K/8/2];
uint16_t qs[QUANT_K/4/2];
uint16_t hmask[QUANT_K_Q3_K/8/2];
uint16_t qs[QUANT_K_Q3_K/4/2];
uint16_t scales[12/2];
float16_t d;
};
#if defined(DATA_A_Q3_K)
#define QUANT_K QUANT_K_Q3_K
#define A_TYPE block_q3_K
#define A_TYPE_PACKED16 block_q3_K_packed16
#endif
#if defined(DATA_A_Q4_K)
#extension GL_EXT_shader_16bit_storage : require
#define QUANT_K 256
#define QUANT_K_Q4_K 256
struct block_q4_K
{
f16vec2 d;
uint8_t scales[3*QUANT_K/64];
uint8_t qs[QUANT_K/2];
uint8_t scales[3*QUANT_K_Q4_K/64];
uint8_t qs[QUANT_K_Q4_K/2];
};
struct block_q4_K_packed16
{
f16vec2 d;
uint16_t scales[3*QUANT_K/64/2];
uint16_t qs[QUANT_K/2/2];
uint16_t scales[3*QUANT_K_Q4_K/64/2];
uint16_t qs[QUANT_K_Q4_K/2/2];
};
struct block_q4_K_packed32
{
f16vec2 d;
uint32_t scales[3*QUANT_K/64/4];
uint32_t qs[QUANT_K/2/4];
uint32_t scales[3*QUANT_K_Q4_K/64/4];
uint32_t qs[QUANT_K_Q4_K/2/4];
};
#if defined(DATA_A_Q4_K)
#define QUANT_K QUANT_K_Q4_K
#define A_TYPE block_q4_K
#define A_TYPE_PACKED16 block_q4_K_packed16
#define A_TYPE_PACKED32 block_q4_K_packed32
#endif
#if defined(DATA_A_Q5_K)
#extension GL_EXT_shader_16bit_storage : require
#define QUANT_K 256
#define QUANT_K_Q5_K 256
struct block_q5_K
{
f16vec2 d;
uint8_t scales[12];
uint8_t qh[QUANT_K/8];
uint8_t qs[QUANT_K/2];
uint8_t qh[QUANT_K_Q5_K/8];
uint8_t qs[QUANT_K_Q5_K/2];
};
struct block_q5_K_packed16
{
f16vec2 d;
uint16_t scales[12/2];
uint16_t qh[QUANT_K/8/2];
uint16_t qs[QUANT_K/2/2];
uint16_t qh[QUANT_K_Q5_K/8/2];
uint16_t qs[QUANT_K_Q5_K/2/2];
};
#if defined(DATA_A_Q5_K)
#define QUANT_K QUANT_K_Q5_K
#define A_TYPE block_q5_K
#define A_TYPE_PACKED16 block_q5_K_packed16
#endif
#if defined(DATA_A_Q6_K)
#extension GL_EXT_shader_16bit_storage : require
#define QUANT_K 256
#define QUANT_K_Q6_K 256
struct block_q6_K
{
uint8_t ql[QUANT_K/2];
uint8_t qh[QUANT_K/4];
int8_t scales[QUANT_K/16];
uint8_t ql[QUANT_K_Q6_K/2];
uint8_t qh[QUANT_K_Q6_K/4];
int8_t scales[QUANT_K_Q6_K/16];
float16_t d;
};
struct block_q6_K_packed16
{
uint16_t ql[QUANT_K/2/2];
uint16_t qh[QUANT_K/4/2];
int8_t scales[QUANT_K/16];
uint16_t ql[QUANT_K_Q6_K/2/2];
uint16_t qh[QUANT_K_Q6_K/4/2];
int8_t scales[QUANT_K_Q6_K/16];
float16_t d;
};
#if defined(DATA_A_Q6_K)
#define QUANT_K QUANT_K_Q6_K
#define A_TYPE block_q6_K
#define A_TYPE_PACKED16 block_q6_K_packed16
#endif
// IQuants
#if defined(DATA_A_IQ4_NL)
#extension GL_EXT_shader_16bit_storage : require
#define QUANT_K 32
#define QUANT_R 2
#define QUANT_K_IQ4_NL 32
#define QUANT_R_IQ4_NL 2
struct block_iq4_nl
{
float16_t d;
uint8_t qs[QUANT_K/2];
uint8_t qs[QUANT_K_IQ4_NL/2];
};
struct block_iq4_nl_packed16
{
float16_t d;
uint16_t qs[QUANT_K/2/2];
uint16_t qs[QUANT_K_IQ4_NL/2/2];
};
#define A_TYPE block_iq4_nl
#define A_TYPE_PACKED16 block_iq4_nl_packed16
#if defined(DATA_A_IQ4_NL)
const int8_t kvalues_iq4nl_const[16] = {
int8_t(-127), int8_t(-104), int8_t(-83), int8_t(-65), int8_t(-49), int8_t(-35), int8_t(-22), int8_t(-10),
@ -313,6 +313,11 @@ void init_iq4nl_shmem()
}
barrier();
}
#define QUANT_K QUANT_K_IQ4_NL
#define QUANT_R QUANT_R_IQ4_NL
#define A_TYPE block_iq4_nl
#define A_TYPE_PACKED16 block_iq4_nl_packed16
#endif
#endif // !defined(GGML_TYPES_COMP)