// Kernels adapted from llama.cpp ggml-cuda.cu
// https://github.com/ggerganov/llama.cpp/blob/master/ggml-cuda.cu
#include "cuda_fp16.h"
#include "cuda_bf16.h"
#include<stdint.h>
#define GGML_UNUSED(x) (void)(x)
#define GGML_CUDA_ASSUME(x)
#ifdef GGML_QKK_64
#define QK_K 64
#define K_SCALE_SIZE 4
#else
#define QK_K 256
#define K_SCALE_SIZE 12
#endif
#undef GGML_CUDA_F16
#define GGML_CUDA_DMMV_X 32
#define CUDA_QUANTIZE_BLOCK_SIZE 256
#define CUDA_DEQUANTIZE_BLOCK_SIZE 256
#define K_QUANTS_PER_ITERATION 2
typedef uint16_t ggml_fp16_t;
typedef float dfloat; // dequantize float
typedef float2 dfloat2;
typedef void (*dequantize_kernel_t)(const void * vx, const int ib, const int iqs, dfloat2 & v);
static __device__ __forceinline__ float warp_reduce_sum(float x) {
#pragma unroll
for (int mask = 16; mask > 0; mask >>= 1) {
x += __shfl_xor_sync(0xffffffff, x, mask, 32);
}
return x;
}
static __device__ __forceinline__ float warp_reduce_max(float x) {
#pragma unroll
for (int mask = 16; mask > 0; mask >>= 1) {
x = fmaxf(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
}
return x;
}
static __device__ __forceinline__ int get_int_from_int8(const int8_t * x8, const int & i32) {
const uint16_t * x16 = (const uint16_t *) (x8 + sizeof(int) * i32); // assume at least 2 byte alignment
int x32 = 0;
x32 |= x16[0] << 0;
x32 |= x16[1] << 16;
return x32;
}
static __device__ __forceinline__ int get_int_from_uint8(const uint8_t * x8, const int & i32) {
const uint16_t * x16 = (const uint16_t *) (x8 + sizeof(int) * i32); // assume at least 2 byte alignment
int x32 = 0;
x32 |= x16[0] << 0;
x32 |= x16[1] << 16;
return x32;
}
static __device__ __forceinline__ int get_int_from_int8_aligned(const int8_t * x8, const int & i32) {
return *((const int *) (x8 + sizeof(int) * i32)); // assume at least 4 byte alignment
}
static __device__ __forceinline__ int get_int_from_uint8_aligned(const uint8_t * x8, const int & i32) {
return *((const int *) (x8 + sizeof(int) * i32)); // assume at least 4 byte alignment
}
#define WARP_SIZE 32
#define CUDART_HMAX 11070 // CUDA 11.7, min. ver. for which __hmax and __hmax2 are known to work (may be higher than needed)
#define CUDA_CC_PASCAL 600
#define MIN_CC_DP4A 610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products
#define CUDA_CC_VOLTA 700
#define CC_OFFSET_AMD 1000000
#define CC_RDNA1 (CC_OFFSET_AMD + 1010)
#define CC_RDNA2 (CC_OFFSET_AMD + 1030)
#define CC_RDNA3 (CC_OFFSET_AMD + 1100)
static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, int c) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A
return __dp4a(a, b, c);
#else // __CUDA_ARCH__ >= MIN_CC_DP4A
const int8_t * a8 = (const int8_t *) &a;
const int8_t * b8 = (const int8_t *) &b;
return c + a8[0]*b8[0] + a8[1]*b8[1] + a8[2]*b8[2] + a8[3]*b8[3];
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
}
#define MMQ_X_Q4_0_RDNA2 64
#define MMQ_Y_Q4_0_RDNA2 128
#define NWARPS_Q4_0_RDNA2 8
#define MMQ_X_Q4_0_RDNA1 64
#define MMQ_Y_Q4_0_RDNA1 64
#define NWARPS_Q4_0_RDNA1 8
#if defined(CUDA_USE_TENSOR_CORES)
#define MMQ_X_Q4_0_AMPERE 4
#define MMQ_Y_Q4_0_AMPERE 32
#define NWARPS_Q4_0_AMPERE 4
#else
#define MMQ_X_Q4_0_AMPERE 64
#define MMQ_Y_Q4_0_AMPERE 128
#define NWARPS_Q4_0_AMPERE 4
#endif
#define MMQ_X_Q4_0_PASCAL 64
#define MMQ_Y_Q4_0_PASCAL 64
#define NWARPS_Q4_0_PASCAL 8
#define MMQ_X_Q4_1_RDNA2 64
#define MMQ_Y_Q4_1_RDNA2 128
#define NWARPS_Q4_1_RDNA2 8
#define MMQ_X_Q4_1_RDNA1 64
#define MMQ_Y_Q4_1_RDNA1 64
#define NWARPS_Q4_1_RDNA1 8
#if defined(CUDA_USE_TENSOR_CORES)
#define MMQ_X_Q4_1_AMPERE 4
#define MMQ_Y_Q4_1_AMPERE 32
#define NWARPS_Q4_1_AMPERE 4
#else
#define MMQ_X_Q4_1_AMPERE 64
#define MMQ_Y_Q4_1_AMPERE 128
#define NWARPS_Q4_1_AMPERE 4
#endif
#define MMQ_X_Q4_1_PASCAL 64
#define MMQ_Y_Q4_1_PASCAL 64
#define NWARPS_Q4_1_PASCAL 8
#define MMQ_X_Q5_0_RDNA2 64
#define MMQ_Y_Q5_0_RDNA2 128
#define NWARPS_Q5_0_RDNA2 8
#define MMQ_X_Q5_0_RDNA1 64
#define MMQ_Y_Q5_0_RDNA1 64
#define NWARPS_Q5_0_RDNA1 8
#if defined(CUDA_USE_TENSOR_CORES)
#define MMQ_X_Q5_0_AMPERE 4
#define MMQ_Y_Q5_0_AMPERE 32
#define NWARPS_Q5_0_AMPERE 4
#else
#define MMQ_X_Q5_0_AMPERE 128
#define MMQ_Y_Q5_0_AMPERE 64
#define NWARPS_Q5_0_AMPERE 4
#endif
#define MMQ_X_Q5_0_PASCAL 64
#define MMQ_Y_Q5_0_PASCAL 64
#define NWARPS_Q5_0_PASCAL 8
#define MMQ_X_Q5_1_RDNA2 64
#define MMQ_Y_Q5_1_RDNA2 128
#define NWARPS_Q5_1_RDNA2 8
#define MMQ_X_Q5_1_RDNA1 64
#define MMQ_Y_Q5_1_RDNA1 64
#define NWARPS_Q5_1_RDNA1 8
#if defined(CUDA_USE_TENSOR_CORES)
#define MMQ_X_Q5_1_AMPERE 4
#define MMQ_Y_Q5_1_AMPERE 32
#define NWARPS_Q5_1_AMPERE 4
#else
#define MMQ_X_Q5_1_AMPERE 128
#define MMQ_Y_Q5_1_AMPERE 64
#define NWARPS_Q5_1_AMPERE 4
#endif
#define MMQ_X_Q5_1_PASCAL 64
#define MMQ_Y_Q5_1_PASCAL 64
#define NWARPS_Q5_1_PASCAL 8
#define MMQ_X_Q8_0_RDNA2 64
#define MMQ_Y_Q8_0_RDNA2 128
#define NWARPS_Q8_0_RDNA2 8
#define MMQ_X_Q8_0_RDNA1 64
#define MMQ_Y_Q8_0_RDNA1 64
#define NWARPS_Q8_0_RDNA1 8
#if defined(CUDA_USE_TENSOR_CORES)
#define MMQ_X_Q8_0_AMPERE 4
#define MMQ_Y_Q8_0_AMPERE 32
#define NWARPS_Q8_0_AMPERE 4
#else
#define MMQ_X_Q8_0_AMPERE 128
#define MMQ_Y_Q8_0_AMPERE 64
#define NWARPS_Q8_0_AMPERE 4
#endif
#define MMQ_X_Q8_0_PASCAL 64
#define MMQ_Y_Q8_0_PASCAL 64
#define NWARPS_Q8_0_PASCAL 8
#define MMQ_X_Q2_K_RDNA2 64
#define MMQ_Y_Q2_K_RDNA2 128
#define NWARPS_Q2_K_RDNA2 8
#define MMQ_X_Q2_K_RDNA1 128
#define MMQ_Y_Q2_K_RDNA1 32
#define NWARPS_Q2_K_RDNA1 8
#if defined(CUDA_USE_TENSOR_CORES)
#define MMQ_X_Q2_K_AMPERE 4
#define MMQ_Y_Q2_K_AMPERE 32
#define NWARPS_Q2_K_AMPERE 4
#else
#define MMQ_X_Q2_K_AMPERE 64
#define MMQ_Y_Q2_K_AMPERE 128
#define NWARPS_Q2_K_AMPERE 4
#endif
#define MMQ_X_Q2_K_PASCAL 64
#define MMQ_Y_Q2_K_PASCAL 64
#define NWARPS_Q2_K_PASCAL 8
#define MMQ_X_Q3_K_RDNA2 128
#define MMQ_Y_Q3_K_RDNA2 64
#define NWARPS_Q3_K_RDNA2 8
#define MMQ_X_Q3_K_RDNA1 32
#define MMQ_Y_Q3_K_RDNA1 128
#define NWARPS_Q3_K_RDNA1 8
#if defined(CUDA_USE_TENSOR_CORES)
#define MMQ_X_Q3_K_AMPERE 4
#define MMQ_Y_Q3_K_AMPERE 32
#define NWARPS_Q3_K_AMPERE 4
#else
#define MMQ_X_Q3_K_AMPERE 128
#define MMQ_Y_Q3_K_AMPERE 128
#define NWARPS_Q3_K_AMPERE 4
#endif
#define MMQ_X_Q3_K_PASCAL 64
#define MMQ_Y_Q3_K_PASCAL 64
#define NWARPS_Q3_K_PASCAL 8
#define MMQ_X_Q4_K_RDNA2 64
#define MMQ_Y_Q4_K_RDNA2 128
#define NWARPS_Q4_K_RDNA2 8
#define MMQ_X_Q4_K_RDNA1 32
#define MMQ_Y_Q4_K_RDNA1 64
#define NWARPS_Q4_K_RDNA1 8
#if defined(CUDA_USE_TENSOR_CORES)
#define MMQ_X_Q4_K_AMPERE 4
#define MMQ_Y_Q4_K_AMPERE 32
#define NWARPS_Q4_K_AMPERE 4
#else
#define MMQ_X_Q4_K_AMPERE 64
#define MMQ_Y_Q4_K_AMPERE 128
#define NWARPS_Q4_K_AMPERE 4
#endif
#define MMQ_X_Q4_K_PASCAL 64
#define MMQ_Y_Q4_K_PASCAL 64
#define NWARPS_Q4_K_PASCAL 8
#define MMQ_X_Q5_K_RDNA2 64
#define MMQ_Y_Q5_K_RDNA2 128
#define NWARPS_Q5_K_RDNA2 8
#define MMQ_X_Q5_K_RDNA1 32
#define MMQ_Y_Q5_K_RDNA1 64
#define NWARPS_Q5_K_RDNA1 8
#if defined(CUDA_USE_TENSOR_CORES)
#define MMQ_X_Q5_K_AMPERE 4
#define MMQ_Y_Q5_K_AMPERE 32
#define NWARPS_Q5_K_AMPERE 4
#else
#define MMQ_X_Q5_K_AMPERE 64
#define MMQ_Y_Q5_K_AMPERE 128
#define NWARPS_Q5_K_AMPERE 4
#endif
#define MMQ_X_Q5_K_PASCAL 64
#define MMQ_Y_Q5_K_PASCAL 64
#define NWARPS_Q5_K_PASCAL 8
#define MMQ_X_Q6_K_RDNA2 64
#define MMQ_Y_Q6_K_RDNA2 128
#define NWARPS_Q6_K_RDNA2 8
#define MMQ_X_Q6_K_RDNA1 32
#define MMQ_Y_Q6_K_RDNA1 64
#define NWARPS_Q6_K_RDNA1 8
#if defined(CUDA_USE_TENSOR_CORES)
#define MMQ_X_Q6_K_AMPERE 4
#define MMQ_Y_Q6_K_AMPERE 32
#define NWARPS_Q6_K_AMPERE 4
#else
#define MMQ_X_Q6_K_AMPERE 64
#define MMQ_Y_Q6_K_AMPERE 64
#define NWARPS_Q6_K_AMPERE 4
#endif
#define MMQ_X_Q6_K_PASCAL 64
#define MMQ_Y_Q6_K_PASCAL 64
#define NWARPS_Q6_K_PASCAL 8
// QK = number of values after dequantization
// QR = QK / number of values before dequantization
// QI = number of 32 bit integers before dequantization
#define QK4_0 32
#define QR4_0 2
#define QI4_0 (QK4_0 / (4 * QR4_0))
typedef struct {
half d; // delta
uint8_t qs[QK4_0 / 2]; // nibbles / quants
} block_q4_0;
static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");
#define QK4_1 32
#define QR4_1 2
#define QI4_1 (QK4_1 / (4 * QR4_1))
typedef struct {
half2 dm; // dm.x = delta, dm.y = min
uint8_t qs[QK4_1 / 2]; // nibbles / quants
} block_q4_1;
static_assert(sizeof(block_q4_1) == sizeof(ggml_fp16_t) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding");
#define QK5_0 32
#define QR5_0 2
#define QI5_0 (QK5_0 / (4 * QR5_0))
typedef struct {
half d; // delta
uint8_t qh[4]; // 5-th bit of quants
uint8_t qs[QK5_0 / 2]; // nibbles / quants
} block_q5_0;
static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
#define QK5_1 32
#define QR5_1 2
#define QI5_1 (QK5_1 / (4 * QR5_1))
typedef struct {
half2 dm; // dm.x = delta, dm.y = min
uint8_t qh[4]; // 5-th bit of quants
uint8_t qs[QK5_1 / 2]; // nibbles / quants
} block_q5_1;
static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
#define QK8_0 32
#define QR8_0 1
#define QI8_0 (QK8_0 / (4 * QR8_0))
typedef struct {
half d; // delta
int8_t qs[QK8_0]; // quants
} block_q8_0;
static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
#define QK8_1 32
#define QR8_1 1
#define QI8_1 (QK8_1 / (4 * QR8_1))
typedef struct {
half2 ds; // ds.x = delta, ds.y = sum
int8_t qs[QK8_0]; // quants
} block_q8_1;
static_assert(sizeof(block_q8_1) == 2*sizeof(ggml_fp16_t) + QK8_0, "wrong q8_1 block size/padding");
typedef float (*vec_dot_q_cuda_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs);
#define QR2_K 4
#define QI2_K (QK_K / (4*QR2_K))
typedef struct {
uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
uint8_t qs[QK_K/4]; // quants
half2 dm; // super-block scale for quantized scales/mins
} block_q2_K;
static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
#define QR3_K 4
#define QI3_K (QK_K / (4*QR3_K))
typedef struct {
uint8_t hmask[QK_K/8]; // quants - high bit
uint8_t qs[QK_K/4]; // quants - low 2 bits
#ifdef GGML_QKK_64
uint8_t scales[2]; // scales, quantized with 8 bits
#else
uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
#endif
half d; // super-block scale
} block_q3_K;
//static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + K_SCALE_SIZE, "wrong q3_K block size/padding");
#define QR4_K 2
#define QI4_K (QK_K / (4*QR4_K))
#ifdef GGML_QKK_64
typedef struct {
half dm[2]; // super-block scales/mins
uint8_t scales[2]; // 4-bit block scales/mins
uint8_t qs[QK_K/2]; // 4--bit quants
} block_q4_K;
static_assert(sizeof(block_q4_K) == sizeof(half2) + QK_K/2 + 2, "wrong q4_K block size/padding");
#else
typedef struct {
half2 dm; // super-block scale for quantized scales/mins
uint8_t scales[3*QK_K/64]; // scales, quantized with 6 bits
uint8_t qs[QK_K/2]; // 4--bit quants
} block_q4_K;
static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + 3*QK_K/64 + QK_K/2, "wrong q4_K block size/padding");
#endif
#define QR5_K 2
#define QI5_K (QK_K / (4*QR5_K))
#ifdef GGML_QKK_64
typedef struct {
half d; // super-block scale
int8_t scales[QK_K/16]; // block scales
uint8_t qh[QK_K/8]; // quants, high bit
uint8_t qs[QK_K/2]; // quants, low 4 bits
} block_q5_K;
static_assert(sizeof(block_q5_K) == sizeof(ggml_fp16_t) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
#else
typedef struct {
half2 dm; // super-block scale for quantized scales/mins
uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
uint8_t qh[QK_K/8]; // quants, high bit
uint8_t qs[QK_K/2]; // quants, low 4 bits
} block_q5_K;
static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
#endif
#define QR6_K 2
#define QI6_K (QK_K / (4*QR6_K))
typedef struct {
uint8_t ql[QK_K/2]; // quants, lower 4 bits
uint8_t qh[QK_K/4]; // quants, upper 2 bits
int8_t scales[QK_K/16]; // scales
half d; // delta
} block_q6_K;
static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_K block size/padding");
// In llama.cpp this is only used for intermediate quantization and dot products
typedef struct {
float d; // delta
int8_t qs[QK_K]; // quants
int16_t bsums[QK_K/16]; // sum of quants in groups of 16
} block_q8_K;
static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding");
// VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called
// MMVQ = mul_mat_vec_q, MMQ = mul_mat_q
#define VDR_Q4_0_Q8_1_MMVQ 2
#define VDR_Q4_0_Q8_1_MMQ 4
template <int vdr> static __device__ __forceinline__ float vec_dot_q4_0_q8_1_impl(
const int * v, const int * u, const float & d4, const half2 & ds8) {
int sumi = 0;
#pragma unroll
for (int i = 0; i < vdr; ++i) {
const int vi0 = (v[i] >> 0) & 0x0F0F0F0F;
const int vi1 = (v[i] >> 4) & 0x0F0F0F0F;
// SIMD dot product of quantized values
sumi = ggml_cuda_dp4a(vi0, u[2*i+0], sumi);
sumi = ggml_cuda_dp4a(vi1, u[2*i+1], sumi);
}
const float2 ds8f = __half22float2(ds8);
// second part effectively subtracts 8 from each quant value
return d4 * (sumi * ds8f.x - (8*vdr/QI4_0) * ds8f.y);
}
#define VDR_Q4_1_Q8_1_MMVQ 2
#define VDR_Q4_1_Q8_1_MMQ 4
template <int vdr> static __device__ __forceinline__ float vec_dot_q4_1_q8_1_impl(
const int * v, const int * u, const half2 & dm4, const half2 & ds8) {
int sumi = 0;
#pragma unroll
for (int i = 0; i < vdr; ++i) {
const int vi0 = (v[i] >> 0) & 0x0F0F0F0F;
const int vi1 = (v[i] >> 4) & 0x0F0F0F0F;
// SIMD dot product of quantized values
sumi = ggml_cuda_dp4a(vi0, u[2*i+0], sumi);
sumi = ggml_cuda_dp4a(vi1, u[2*i+1], sumi);
}
#ifdef GGML_CUDA_F16
const float2 tmp = __half22float2(__hmul2(dm4, ds8));
const float d4d8 = tmp.x;
const float m4s8 = tmp.y;
#else
const float2 dm4f = __half22float2(dm4);
const float2 ds8f = __half22float2(ds8);
const float d4d8 = dm4f.x * ds8f.x;
const float m4s8 = dm4f.y * ds8f.y;
#endif // GGML_CUDA_F16
// scale second part of sum by QI8_1/(vdr * QR4_1) to compensate for multiple threads adding it
return sumi * d4d8 + m4s8 / (QI8_1 / (vdr * QR4_1));
}
#define VDR_Q5_0_Q8_1_MMVQ 2
#define VDR_Q5_0_Q8_1_MMQ 4
template <int vdr> static __device__ __forceinline__ float vec_dot_q5_0_q8_1_impl(
const int * vl, const int * vh, const int * u, const float & d5, const half2 & ds8) {
int sumi = 0;
#pragma unroll
for (int i = 0; i < vdr; ++i) {
int vi0 = (vl[i] >> 0) & 0x0F0F0F0F; // lower 4 qs bits, still need qh as 5th bits
vi0 |= (vh[i] << 4) & 0x00000010; // 0 -> 4
vi0 |= (vh[i] << 11) & 0x00001000; // 1 -> 12
vi0 |= (vh[i] << 18) & 0x00100000; // 2 -> 20
vi0 |= (vh[i] << 25) & 0x10000000; // 3 -> 28
sumi = ggml_cuda_dp4a(vi0, u[2*i+0], sumi); // SIMD dot product of quantized values
int vi1 = (vl[i] >> 4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh as 5th bits
vi1 |= (vh[i] >> 12) & 0x00000010; // 16 -> 4
vi1 |= (vh[i] >> 5) & 0x00001000; // 17 -> 12
vi1 |= (vh[i] << 2) & 0x00100000; // 18 -> 20
vi1 |= (vh[i] << 9) & 0x10000000; // 19 -> 28
sumi = ggml_cuda_dp4a(vi1, u[2*i+1], sumi); // SIMD dot product of quantized values
}
const float2 ds8f = __half22float2(ds8);
// second part effectively subtracts 16 from each quant value
return d5 * (sumi * ds8f.x - (16*vdr/QI5_0) * ds8f.y);
}
#define VDR_Q5_1_Q8_1_MMVQ 2
#define VDR_Q5_1_Q8_1_MMQ 4
template <int vdr> static __device__ __forceinline__ float vec_dot_q5_1_q8_1_impl(
const int * vl, const int * vh, const int * u, const half2 & dm5, const half2 & ds8) {
int sumi = 0;
#pragma unroll
for (int i = 0; i < vdr; ++i) {
int vi0 = (vl[i] >> 0) & 0x0F0F0F0F; // lower 4 qs bits, still need qh as 5th bits
vi0 |= (vh[i] << 4) & 0x00000010; // 0 -> 4
vi0 |= (vh[i] << 11) & 0x00001000; // 1 -> 12
vi0 |= (vh[i] << 18) & 0x00100000; // 2 -> 20
vi0 |= (vh[i] << 25) & 0x10000000; // 3 -> 28
sumi = ggml_cuda_dp4a(vi0, u[2*i+0], sumi); // SIMD dot product of quantized values
int vi1 = (vl[i] >> 4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh as 5th bits
vi1 |= (vh[i] >> 12) & 0x00000010; // 16 -> 4
vi1 |= (vh[i] >> 5) & 0x00001000; // 17 -> 12
vi1 |= (vh[i] << 2) & 0x00100000; // 18 -> 20
vi1 |= (vh[i] << 9) & 0x10000000; // 19 -> 28
sumi = ggml_cuda_dp4a(vi1, u[2*i+1], sumi); // SIMD dot product of quantized values
}
#ifdef GGML_CUDA_F16
const float2 tmp = __half22float2(__hmul2(dm5, ds8));
const float d5d8 = tmp.x;
const float m5s8 = tmp.y;
#else
const float2 dm5f = __half22float2(dm5);
const float2 ds8f = __half22float2(ds8);
const float d5d8 = dm5f.x * ds8f.x;
const float m5s8 = dm5f.y * ds8f.y;
#endif // GGML_CUDA_F16
// scale second part of sum by QI5_1 / vdr to compensate for multiple threads adding it
return sumi*d5d8 + m5s8 / (QI5_1 / vdr);
}
#define VDR_Q8_0_Q8_1_MMVQ 2
#define VDR_Q8_0_Q8_1_MMQ 8
template <int vdr> static __device__ __forceinline__ float vec_dot_q8_0_q8_1_impl(
const int * v, const int * u, const float & d8_0, const float & d8_1) {
int sumi = 0;
#pragma unroll
for (int i = 0; i < vdr; ++i) {
// SIMD dot product of quantized values
sumi = ggml_cuda_dp4a(v[i], u[i], sumi);
}
return d8_0*d8_1 * sumi;
}
template <int vdr> static __device__ __forceinline__ float vec_dot_q8_1_q8_1_impl(
const int * v, const int * u, const half2 & dm8, const half2 & ds8) {
int sumi = 0;
#pragma unroll
for (int i = 0; i < vdr; ++i) {
// SIMD dot product of quantized values
sumi = ggml_cuda_dp4a(v[i], u[i], sumi);
}
#ifdef GGML_CUDA_F16
const float2 tmp = __half22float2(__hmul2(dm8, ds8));
const float d8d8 = tmp.x;
const float m8s8 = tmp.y;
#else
const float2 dm8f = __half22float2(dm8);
const float2 ds8f = __half22float2(ds8);
const float d8d8 = dm8f.x * ds8f.x;
const float m8s8 = dm8f.y * ds8f.y;
#endif // GGML_CUDA_F16
// scale second part of sum by QI8_1/ vdr to compensate for multiple threads adding it
return sumi*d8d8 + m8s8 / (QI8_1 / vdr);
}
#define VDR_Q2_K_Q8_1_MMVQ 1
#define VDR_Q2_K_Q8_1_MMQ 2
// contiguous v/x values
static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmvq(
const int & v, const int * __restrict__ u, const uint8_t * __restrict__ scales,
const half2 & dm2, const float * __restrict__ d8) {
float sumf_d = 0.0f;
float sumf_m = 0.0f;
#pragma unroll
for (int i = 0; i < QR2_K; ++i) {
const int sc = scales[2*i];
const int vi = (v >> (2*i)) & 0x03030303;
sumf_d += d8[i] * (ggml_cuda_dp4a(vi, u[i], 0) * (sc & 0xF)); // SIMD dot product
// fill int with 4x m
int m = sc >> 4;
m |= m << 8;
m |= m << 16;
sumf_m += d8[i] * ggml_cuda_dp4a(m, u[i], 0); // multiply constant q2_K part with sum of q8_1 values
}
const float2 dm2f = __half22float2(dm2);
return dm2f.x*sumf_d - dm2f.y*sumf_m;
}
// contiguous u/y values
static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmq(
const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ scales,
const half2 & dm2, const float & d8) {
int sumi_d = 0;
int sumi_m = 0;
#pragma unroll
for (int i0 = 0; i0 < QI8_1; i0 += QI8_1/2) {
int sumi_d_sc = 0;
const int sc = scales[i0 / (QI8_1/2)];
// fill int with 4x m
int m = sc >> 4;
m |= m << 8;
m |= m << 16;
#pragma unroll
for (int i = i0; i < i0 + QI8_1/2; ++i) {
sumi_d_sc = ggml_cuda_dp4a(v[i], u[i], sumi_d_sc); // SIMD dot product
sumi_m = ggml_cuda_dp4a(m, u[i], sumi_m); // multiply sum of q8_1 values with m
}
sumi_d += sumi_d_sc * (sc & 0xF);
}
const float2 dm2f = __half22float2(dm2);
return d8 * (dm2f.x*sumi_d - dm2f.y*sumi_m);
}
#define VDR_Q3_K_Q8_1_MMVQ 1
#define VDR_Q3_K_Q8_1_MMQ 2
// contiguous v/x values
static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmvq(
const int & vl, const int & vh, const int * __restrict__ u, const uint8_t * __restrict__ scales,
const int & scale_offset, const float & d3, const float * __restrict__ d8) {
float sumf = 0.0f;
#pragma unroll
for (int i = 0; i < QR3_K; ++i) {
const int isc = scale_offset + 2*i;
const int isc_low = isc % (QK_K/32);
const int sc_shift_low = 4 * (isc / (QK_K/32));
const int sc_low = (scales[isc_low] >> sc_shift_low) & 0xF;
const int isc_high = isc % (QK_K/64);
const int sc_shift_high = 2 * (isc / (QK_K/64));
const int sc_high = ((scales[(QK_K/32) + isc_high] >> sc_shift_high) & 3) << 4;
const int sc = (sc_low | sc_high) - 32;
const int vil = (vl >> (2*i)) & 0x03030303;
const int vih = ((vh >> i) << 2) & 0x04040404;
const int vi = __vsubss4(vil, vih);
sumf += d8[i] * (ggml_cuda_dp4a(vi, u[i], 0) * sc); // SIMD dot product
}
return d3 * sumf;
}
// contiguous u/y values
static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmq(
const int * __restrict__ v, const int * __restrict__ u, const int8_t * __restrict__ scales,
const float & d3, const float & d8) {
int sumi = 0;
#pragma unroll
for (int i0 = 0; i0 < QR3_K*VDR_Q3_K_Q8_1_MMQ; i0 += QI8_1/2) {
int sumi_sc = 0;
for (int i = i0; i < i0 + QI8_1/2; ++i) {
sumi_sc = ggml_cuda_dp4a(v[i], u[i], sumi_sc); // SIMD dot product
}
sumi += sumi_sc * scales[i0 / (QI8_1/2)];
}
return d3*d8 * sumi;
}
#define VDR_Q4_K_Q8_1_MMVQ 2
#define VDR_Q4_K_Q8_1_MMQ 8
// contiguous v/x values
static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_vmmq(
const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ sc,
const uint8_t * __restrict__ m, const half2 & dm4, const float * __restrict__ d8) {
float sumf_d = 0.0f;
float sumf_m = 0.0f;
#pragma unroll
for (int i = 0; i < QR4_K; ++i) {
const int v0i = (v[0] >> (4*i)) & 0x0F0F0F0F;
const int v1i = (v[1] >> (4*i)) & 0x0F0F0F0F;
const int dot1 = ggml_cuda_dp4a(v1i, u[2*i+1], ggml_cuda_dp4a(v0i, u[2*i+0], 0)); // SIMD dot product
const int dot2 = ggml_cuda_dp4a(0x01010101, u[2*i+1], ggml_cuda_dp4a(0x01010101, u[2*i+0], 0)); // sum of u
sumf_d += d8[i] * (dot1 * sc[i]);
sumf_m += d8[i] * (dot2 * m[i]); // multiply constant part of q4_K with sum of q8_1 values
}
const float2 dm4f = __half22float2(dm4);
return dm4f.x*sumf_d - dm4f.y*sumf_m;
}
// contiguous u/y values
static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_mmq(
const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ sc,
const uint8_t * __restrict__ m, const half2 & dm4, const half2 * __restrict__ ds8) {
float sumf_d = 0.0f;
float sumf_m = 0.0f;
#pragma unroll
for (int i = 0; i < QR4_K*VDR_Q4_K_Q8_1_MMQ/QI8_1; ++i) {
int sumi_d = 0;
#pragma unroll
for (int j = 0; j < QI8_1; ++j) {
sumi_d = ggml_cuda_dp4a((v[j] >> (4*i)) & 0x0F0F0F0F, u[i*QI8_1 + j], sumi_d); // SIMD dot product
}
const float2 ds8f = __half22float2(ds8[i]);
sumf_d += ds8f.x * (sc[i] * sumi_d);
sumf_m += ds8f.y * m[i]; // sum of q8_1 block * q4_K min val
}
const float2 dm4f = __half22float2(dm4);
return dm4f.x*sumf_d - dm4f.y*sumf_m;
}
#define VDR_Q5_K_Q8_1_MMVQ 2
#define VDR_Q5_K_Q8_1_MMQ 8
// contiguous v/x values
static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_vmmq(
const int * __restrict__ vl, const int * __restrict__ vh, const int * __restrict__ u, const uint8_t * __restrict__ sc,
const uint8_t * __restrict__ m, const half2 & dm5, const float * __restrict__ d8) {
float sumf_d = 0.0f;
float sumf_m = 0.0f;
#pragma unroll
for (int i = 0; i < QR5_K; ++i) {
const int vl0i = (vl[0] >> (4*i)) & 0x0F0F0F0F;
const int vl1i = (vl[1] >> (4*i)) & 0x0F0F0F0F;
const int vh0i = ((vh[0] >> i) << 4) & 0x10101010;
const int vh1i = ((vh[1] >> i) << 4) & 0x10101010;
const int v0i = vl0i | vh0i;
const int v1i = vl1i | vh1i;
const int dot1 = ggml_cuda_dp4a(v0i, u[2*i+0], ggml_cuda_dp4a(v1i, u[2*i+1], 0)); // SIMD dot product
const int dot2 = ggml_cuda_dp4a(0x01010101, u[2*i+0], ggml_cuda_dp4a(0x01010101, u[2*i+1], 0)); // sum of u
sumf_d += d8[i] * (dot1 * sc[i]);
sumf_m += d8[i] * (dot2 * m[i]);
}
const float2 dm5f = __half22float2(dm5);
return dm5f.x*sumf_d - dm5f.y*sumf_m;
}
// contiguous u/y values
static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_mmq(
const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ sc,
const uint8_t * __restrict__ m, const half2 & dm4, const half2 * __restrict__ ds8) {
float sumf_d = 0.0f;
float sumf_m = 0.0f;
#pragma unroll
for (int i = 0; i < QR5_K*VDR_Q5_K_Q8_1_MMQ/QI8_1; ++i) {
int sumi_d = 0;
#pragma unroll
for (int j = 0; j < QI8_1; ++j) {
sumi_d = ggml_cuda_dp4a(v[i*QI8_1 + j], u[i*QI8_1 + j], sumi_d); // SIMD dot product
}
const float2 ds8f = __half22float2(ds8[i]);
sumf_d += ds8f.x * (sc[i] * sumi_d);
sumf_m += ds8f.y * m[i]; // sum of q8_1 block * q4_K min val
}
const float2 dm4f = __half22float2(dm4);
return dm4f.x*sumf_d - dm4f.y*sumf_m;
}
#define VDR_Q6_K_Q8_1_MMVQ 1
#define VDR_Q6_K_Q8_1_MMQ 8
// contiguous v/x values
static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmvq(
const int & vl, const int & vh, const int * __restrict__ u, const int8_t * __restrict__ scales,
const float & d, const float * __restrict__ d8) {
float sumf = 0.0f;
#pragma unroll
for (int i = 0; i < QR6_K; ++i) {
const int sc = scales[4*i];
const int vil = (vl >> (4*i)) & 0x0F0F0F0F;
const int vih = ((vh >> (4*i)) << 4) & 0x30303030;
const int vi = __vsubss4((vil | vih), 0x20202020); // vi = (vil | vih) - 32
sumf += d8[i] * (ggml_cuda_dp4a(vi, u[i], 0) * sc); // SIMD dot product
}
return d*sumf;
}
// contiguous u/y values
static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmq(
const int * __restrict__ v, const int * __restrict__ u, const int8_t * __restrict__ sc,
const float & d6, const float * __restrict__ d8) {
float sumf_d = 0.0f;
#pragma unroll
for (int i0 = 0; i0 < VDR_Q6_K_Q8_1_MMQ; i0 += 4) {
int2 sumi_d = {0, 0}; // 2 q6_K scales per q8_1 scale
#pragma unroll
for (int i = i0; i < i0 + 2; ++i) {
sumi_d.x = ggml_cuda_dp4a(v[2*i+0], u[2*i+0], sumi_d.x); // SIMD dot product
sumi_d.x = ggml_cuda_dp4a(v[2*i+1], u[2*i+1], sumi_d.x); // SIMD dot product
sumi_d.y = ggml_cuda_dp4a(v[2*i+4], u[2*i+4], sumi_d.y); // SIMD dot product
sumi_d.y = ggml_cuda_dp4a(v[2*i+5], u[2*i+5], sumi_d.y); // SIMD dot product
}
sumf_d += d8[i0/4] * (sc[i0/2+0]*sumi_d.x + sc[i0/2+1]*sumi_d.y);
}
return d6 * sumf_d;
}
static __device__ __forceinline__ float vec_dot_q4_0_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
const block_q4_0 * bq4_0 = (const block_q4_0 *) vbq;
int v[VDR_Q4_0_Q8_1_MMVQ];
int u[2*VDR_Q4_0_Q8_1_MMVQ];
#pragma unroll
for (int i = 0; i < VDR_Q4_0_Q8_1_MMVQ; ++i) {
v[i] = get_int_from_uint8(bq4_0->qs, iqs + i);
u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI4_0);
}
return vec_dot_q4_0_q8_1_impl<VDR_Q4_0_Q8_1_MMVQ>(v, u, bq4_0->d, bq8_1->ds);
}
static __device__ __forceinline__ float vec_dot_q4_1_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
const block_q4_1 * bq4_1 = (const block_q4_1 *) vbq;
int v[VDR_Q4_1_Q8_1_MMVQ];
int u[2*VDR_Q4_1_Q8_1_MMVQ];
#pragma unroll
for (int i = 0; i < VDR_Q4_1_Q8_1_MMVQ; ++i) {
v[i] = get_int_from_uint8_aligned(bq4_1->qs, iqs + i);
u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI4_1);
}
return vec_dot_q4_1_q8_1_impl<VDR_Q4_1_Q8_1_MMVQ>(v, u, bq4_1->dm, bq8_1->ds);
}
static __device__ __forceinline__ float vec_dot_q5_0_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
const block_q5_0 * bq5_0 = (const block_q5_0 *) vbq;
int vl[VDR_Q5_0_Q8_1_MMVQ];
int vh[VDR_Q5_0_Q8_1_MMVQ];
int u[2*VDR_Q5_0_Q8_1_MMVQ];
#pragma unroll
for (int i = 0; i < VDR_Q5_0_Q8_1_MMVQ; ++i) {
vl[i] = get_int_from_uint8(bq5_0->qs, iqs + i);
vh[i] = get_int_from_uint8(bq5_0->qh, 0) >> (4 * (iqs + i));
u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI5_0);
}
return vec_dot_q5_0_q8_1_impl<VDR_Q5_0_Q8_1_MMVQ>(vl, vh, u, bq5_0->d, bq8_1->ds);
}
static __device__ __forceinline__ float vec_dot_q5_1_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
const block_q5_1 * bq5_1 = (const block_q5_1 *) vbq;
int vl[VDR_Q5_1_Q8_1_MMVQ];
int vh[VDR_Q5_1_Q8_1_MMVQ];
int u[2*VDR_Q5_1_Q8_1_MMVQ];
#pragma unroll
for (int i = 0; i < VDR_Q5_1_Q8_1_MMVQ; ++i) {
vl[i] = get_int_from_uint8_aligned(bq5_1->qs, iqs + i);
vh[i] = get_int_from_uint8_aligned(bq5_1->qh, 0) >> (4 * (iqs + i));
u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI5_1);
}
return vec_dot_q5_1_q8_1_impl<VDR_Q5_1_Q8_1_MMVQ>(vl, vh, u, bq5_1->dm, bq8_1->ds);
}
static __device__ __forceinline__ float vec_dot_q8_0_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
const block_q8_0 * bq8_0 = (const block_q8_0 *) vbq;
int v[VDR_Q8_0_Q8_1_MMVQ];
int u[VDR_Q8_0_Q8_1_MMVQ];
#pragma unroll
for (int i = 0; i < VDR_Q8_0_Q8_1_MMVQ; ++i) {
v[i] = get_int_from_int8(bq8_0->qs, iqs + i);
u[i] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
}
return vec_dot_q8_0_q8_1_impl<VDR_Q8_0_Q8_1_MMVQ>(v, u, bq8_0->d, __low2half(bq8_1->ds));
}
static __device__ __forceinline__ float vec_dot_q2_K_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
const block_q2_K * bq2_K = (const block_q2_K *) vbq;
const int bq8_offset = QR2_K * (iqs / QI8_1);
const int scale_offset = iqs - iqs % QI8_1 + (iqs % QI8_1) / (QI8_1/2);
const uint8_t * scales = bq2_K->scales + scale_offset;
const int v = get_int_from_uint8_aligned(bq2_K->qs, iqs);
int u[QR2_K];
float d8[QR2_K];
#pragma unroll
for (int i = 0; i < QR2_K; ++ i) {
u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1);
d8[i] = __low2float(bq8_1[bq8_offset + i].ds);
}
return vec_dot_q2_K_q8_1_impl_mmvq(v, u, scales, bq2_K->dm, d8);
}
static __device__ __forceinline__ float vec_dot_q3_K_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
const block_q3_K * bq3_K = (const block_q3_K *) vbq;
const int bq8_offset = QR3_K * (iqs / (QI3_K/2));
const int scale_offset = iqs - iqs % QI8_1 + (iqs % QI8_1) / (QI8_1/2);
const float d = bq3_K->d;
const int vl = get_int_from_uint8(bq3_K->qs, iqs);
// invert the mask with ~ so that a 0/1 results in 4/0 being subtracted
const int vh = ~get_int_from_uint8(bq3_K->hmask, iqs % (QI3_K/2)) >> bq8_offset;
int u[QR3_K];
float d8[QR3_K];
#pragma unroll
for (int i = 0; i < QR3_K; ++i) {
u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1);
d8[i] = __low2float(bq8_1[bq8_offset + i].ds);
}
return vec_dot_q3_K_q8_1_impl_mmvq(vl, vh, u, bq3_K->scales, scale_offset, d, d8);
}
static __device__ __forceinline__ float vec_dot_q4_K_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
#ifndef GGML_QKK_64
const block_q4_K * bq4_K = (const block_q4_K *) vbq;
int v[2];
int u[2*QR4_K];
float d8[QR4_K];
// iqs is in 0,2..30. bq8_offset = iqs/4 -> bq8_offset = 0, 2, 4, 6
const int bq8_offset = QR4_K * ((iqs/2) / (QI8_1/2));
// iqs = 0....3 -> bq8_offset = 0, want q4_offset = 0, 4, 8, 12
// iqs = 4....7 -> bq8_offset = 2, want q4_offset = 32, 36, 40, 44
// iqs = 8...11 -> bq8_offset = 4, want q4_offset = 64, 68, 72, 76
// iqs = 12..15 -> bq8_offset = 6, want q4_offset = 96, 100, 104, 108
const int * q4 = (const int *)(bq4_K->qs + 16 * bq8_offset + 4 * ((iqs/2)%4));
v[0] = q4[0];
v[1] = q4[4];
const uint16_t * scales = (const uint16_t *)bq4_K->scales;
uint16_t aux[2];
const int j = bq8_offset/2;
if (j < 2) {
aux[0] = scales[j+0] & 0x3f3f;
aux[1] = scales[j+2] & 0x3f3f;
} else {
aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2);
aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2);
}
const uint8_t * sc = (const uint8_t *)aux;
const uint8_t * m = sc + 2;
for (int i = 0; i < QR4_K; ++i) {
const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
d8[i] = __low2float(bq8i->ds);
const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4);
u[2*i+0] = q8[0];
u[2*i+1] = q8[4];
}
return vec_dot_q4_K_q8_1_impl_vmmq(v, u, sc, m, bq4_K->dm, d8);
#else
const block_q4_K * bq4_K = (const block_q4_K *) vbq;
float sumf_d = 0.0f;
float sumf_m = 0.0f;
uint16_t aux16[2];
const uint8_t * s = (const uint8_t *)aux16;
const uint16_t * a = (const uint16_t *)bq4_K->scales;
aux16[0] = a[0] & 0x0f0f;
aux16[1] = (a[0] >> 4) & 0x0f0f;
const float dall = bq4_K->dm[0];
const float dmin = bq4_K->dm[1];
const float d8_1 = __low2float(bq8_1[0].ds);
const float d8_2 = __low2float(bq8_1[1].ds);
const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2));
const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4);
const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2));
const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4);
const int * q4 = (const int *)bq4_K->qs + (iqs/2);
const int v1 = q4[0];
const int v2 = q4[4];
const int dot1 = ggml_cuda_dp4a(ui2, v2 & 0x0f0f0f0f, ggml_cuda_dp4a(ui1, v1 & 0x0f0f0f0f, 0));
const int dot2 = ggml_cuda_dp4a(ui4, (v2 >> 4) & 0x0f0f0f0f, ggml_cuda_dp4a(ui3, (v1 >> 4) & 0x0f0f0f0f, 0));
const int dot3 = ggml_cuda_dp4a(0x01010101, ui2, ggml_cuda_dp4a(0x01010101, ui1, 0));
const int dot4 = ggml_cuda_dp4a(0x01010101, ui4, ggml_cuda_dp4a(0x01010101, ui3, 0));
sumf_d += d8_1 * (dot1 * s[0]) + d8_2 * (dot2 * s[1]);
sumf_m += d8_1 * (dot3 * s[2]) + d8_2 * (dot4 * s[3]);
return dall * sumf_d - dmin * sumf_m;
#endif
}
static __device__ __forceinline__ float vec_dot_q5_K_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
#ifndef GGML_QKK_64
const block_q5_K * bq5_K = (const block_q5_K *) vbq;
int vl[2];
int vh[2];
int u[2*QR5_K];
float d8[QR5_K];
const int bq8_offset = QR5_K * ((iqs/2) / (QI8_1/2));
const int * ql = (const int *)(bq5_K->qs + 16 * bq8_offset + 4 * ((iqs/2)%4));
const int * qh = (const int *)(bq5_K->qh + 4 * ((iqs/2)%4));
vl[0] = ql[0];
vl[1] = ql[4];
vh[0] = qh[0] >> bq8_offset;
vh[1] = qh[4] >> bq8_offset;
const uint16_t * scales = (const uint16_t *)bq5_K->scales;
uint16_t aux[2];
const int j = bq8_offset/2;
if (j < 2) {
aux[0] = scales[j+0] & 0x3f3f;
aux[1] = scales[j+2] & 0x3f3f;
} else {
aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2);
aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2);
}
const uint8_t * sc = (const uint8_t *)aux;
const uint8_t * m = sc + 2;
#pragma unroll
for (int i = 0; i < QR5_K; ++i) {
const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
d8[i] = __low2float(bq8i->ds);
const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4);
u[2*i+0] = q8[0];
u[2*i+1] = q8[4];
}
return vec_dot_q5_K_q8_1_impl_vmmq(vl, vh, u, sc, m, bq5_K->dm, d8);
#else
const block_q5_K * bq5_K = (const block_q5_K *) vbq;
const int8_t * s = bq5_K->scales;
const float d = bq5_K->d;
const float d8_1 = __low2half(bq8_1[0].ds);
const float d8_2 = __low2half(bq8_1[1].ds);
const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2));
const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4);
const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2));
const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4);
const int * ql = (const int *)bq5_K->qs + (iqs/2);
const int vl1 = ql[0];
const int vl2 = ql[4];
const int step = 4 * (iqs/2); // 0, 4, 8, 12
const int im = step/8; // = 0 for iqs = 0, 2, = 1 for iqs = 4, 6
const int in = step%8; // 0, 4, 0, 4
const int vh = (*((const int *)(bq5_K->qh + in))) >> im;
const int v1 = (((vh << 4) & 0x10101010) ^ 0x10101010) | ((vl1 >> 0) & 0x0f0f0f0f);
const int v2 = (((vh << 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 0) & 0x0f0f0f0f);
const int v3 = (((vh >> 0) & 0x10101010) ^ 0x10101010) | ((vl1 >> 4) & 0x0f0f0f0f);
const int v4 = (((vh >> 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 4) & 0x0f0f0f0f);
const float sumf_d = d8_1 * (ggml_cuda_dp4a(ui1, v1, 0) * s[0] + ggml_cuda_dp4a(ui2, v2, 0) * s[1])
+ d8_2 * (ggml_cuda_dp4a(ui3, v3, 0) * s[2] + ggml_cuda_dp4a(ui4, v4, 0) * s[3]);
return d * sumf_d;
#endif
}
static __device__ __forceinline__ float vec_dot_q6_K_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
const block_q6_K * bq6_K = (const block_q6_K *) vbq;
const int bq8_offset = 2 * QR6_K * (iqs / (QI6_K/2)) + (iqs % (QI6_K/2)) / (QI6_K/4);
const int scale_offset = (QI6_K/4) * (iqs / (QI6_K/2)) + (iqs % (QI6_K/2)) / (QI6_K/8);
const int vh_shift = 2 * ((iqs % (QI6_K/2)) / (QI6_K/4));
const int vl = get_int_from_uint8(bq6_K->ql, iqs);
const int vh = get_int_from_uint8(bq6_K->qh, (QI6_K/4) * (iqs / (QI6_K/2)) + iqs % (QI6_K/4)) >> vh_shift;
const int8_t * scales = bq6_K->scales + scale_offset;
int u[QR6_K];
float d8[QR6_K];
#pragma unroll
for (int i = 0; i < QR6_K; ++i) {
u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + 2*i].qs, iqs % QI8_1);
d8[i] = __low2float(bq8_1[bq8_offset + 2*i].ds);
}
return vec_dot_q6_K_q8_1_impl_mmvq(vl, vh, u, scales, bq6_K->d, d8);
}
static __global__ void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy, const int kx, const int kx_padded) {
const int ix = blockDim.x*blockIdx.x + threadIdx.x;
if (ix >= kx_padded) {
return;
}
const int iy = blockDim.y*blockIdx.y + threadIdx.y;
const int i_padded = iy*kx_padded + ix;
block_q8_1 * y = (block_q8_1 *) vy;
const int ib = i_padded / QK8_1; // block index
const int iqs = i_padded % QK8_1; // quant index
const float xi = ix < kx ? x[iy*kx + ix] : 0.0f;
float amax = fabsf(xi);
float sum = xi;
amax = warp_reduce_max(amax);
sum = warp_reduce_sum(sum);
const float d = amax / 127;
const int8_t q = amax == 0.0f ? 0 : roundf(xi / d);
y[ib].qs[iqs] = q;
if (iqs > 0) {
return;
}
reinterpret_cast<half&>(y[ib].ds.x) = d;
reinterpret_cast<half&>(y[ib].ds.y) = sum;
}
template<typename dst_t>
static __device__ __forceinline__ dst_t convert_from_half(half val) {
return val;
}
template<>
__device__ __forceinline__ nv_bfloat16 convert_from_half<nv_bfloat16>(half val) {
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
return __float2bfloat16(__half2float(val));
#else
return __half2float(val);
#endif // defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
}
template<>
__device__ __forceinline__ float convert_from_half<float>(half val) {
return __half2float(val);
}
template<typename dst_t>
inline __device__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
const auto i = 0; //we only need dequant one block in each call
const block_q2_K * x = (const block_q2_K *) vx;
const auto tid = threadIdx.x;
const int n = tid/32;
const int l = tid - 32*n;
const int is = 8*n + l/16;
const uint8_t q = x[i].qs[32*n + l];
dst_t * y = yy + i*QK_K + 128*n;
half dall = __low2half(x[i].dm);
half dmin = __high2half(x[i].dm);
y[l+ 0] = convert_from_half<dst_t>(__hsub(__hmul(dall, __int2half_rn((x[i].scales[is+0] & 0xF) * ((q >> 0) & 3))), __hmul(dmin, __int2half_rn(x[i].scales[is+0] >> 4))));
y[l+32] = convert_from_half<dst_t>(__hsub(__hmul(dall, __int2half_rn((x[i].scales[is+2] & 0xF) * ((q >> 2) & 3))), __hmul(dmin, __int2half_rn(x[i].scales[is+2] >> 4))));
y[l+64] = convert_from_half<dst_t>(__hsub(__hmul(dall, __int2half_rn((x[i].scales[is+4] & 0xF) * ((q >> 4) & 3))), __hmul(dmin, __int2half_rn(x[i].scales[is+4] >> 4))));
y[l+96] = convert_from_half<dst_t>(__hsub(__hmul(dall, __int2half_rn((x[i].scales[is+6] & 0xF) * ((q >> 6) & 3))), __hmul(dmin, __int2half_rn(x[i].scales[is+6] >> 4))));
}
template<typename dst_t>
inline __device__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
const auto i = 0;
const block_q3_K * x = (const block_q3_K *) vx;
const auto r = threadIdx.x/4;
const int tid = r/2;
const int is0 = r%2;
const int l0 = 16*is0 + 4*(threadIdx.x%4);
const int n = tid / 4;
const int j = tid - 4*n;
uint8_t m = 1 << (4*n + j);
int is = 8*n + 2*j + is0;
int shift = 2*j;
int8_t us = is < 4 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+8] >> 0) & 3) << 4) :
is < 8 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+4] >> 2) & 3) << 4) :
is < 12 ? (x[i].scales[is-8] >> 4) | (((x[i].scales[is+0] >> 4) & 3) << 4) :
(x[i].scales[is-8] >> 4) | (((x[i].scales[is-4] >> 6) & 3) << 4);
half d_all = x[i].d;
half dl = __hmul(d_all, __int2half_rn(us - 32));
dst_t * y = yy + i*QK_K + 128*n + 32*j;
const uint8_t * q = x[i].qs + 32*n;
const uint8_t * hm = x[i].hmask;
for (int l = l0; l < l0+4; ++l) {
y[l] = convert_from_half<dst_t>(__hmul(dl, __int2half_rn((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4))));
}
}
static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) {
if (j < 4) {
d = q[j] & 63; m = q[j + 4] & 63;
} else {
d = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4);
m = (q[j+4] >> 4) | ((q[j-0] >> 6) << 4);
}
}
template<typename dst_t>
inline __device__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
const block_q4_K * x = (const block_q4_K *) vx;
const auto i = 0;
// assume 32 threads
const auto tid = threadIdx.x;
const int il = tid/8;
const int ir = tid%8;
const int is = 2*il;
const int n = 4;
dst_t * y = yy + i*QK_K + 64*il + n*ir;
const half dall = __low2half(x[i].dm);
const half dmin = __high2half(x[i].dm);
const uint8_t * q = x[i].qs + 32*il + n*ir;
uint8_t sc, m;
get_scale_min_k4(is + 0, x[i].scales, sc, m);
const half d1 = __hmul(dall, __int2half_rn(sc));
const half m1 = __hmul(dmin, __int2half_rn(m));
get_scale_min_k4(is + 1, x[i].scales, sc, m);
const half d2 = __hmul(dall, __int2half_rn(sc));
const half m2 = __hmul(dmin, __int2half_rn(m));
for (int l = 0; l < n; ++l) {
y[l + 0] = convert_from_half<dst_t>(__hsub(__hmul(d1, __int2half_rn(q[l] & 0xF)), m1));
y[l +32] = convert_from_half<dst_t>(__hsub(__hmul(d2, __int2half_rn(q[l] >> 4)), m2));
}
}
template<typename dst_t>
inline __device__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
const block_q5_K * x = (const block_q5_K *) vx;
const auto i = 0;
// assume 64 threads - this is very slightly better than the one below
const auto tid = threadIdx.x;
const int il = tid/16; // il is in 0...3
const int ir = tid%16; // ir is in 0...15
const int is = 2*il; // is is in 0...6
dst_t * y = yy + i*QK_K + 64*il + 2*ir;
const half dall = __low2half(x[i].dm);
const half dmin = __high2half(x[i].dm);
const uint8_t * ql = x[i].qs + 32*il + 2*ir;
const uint8_t * qh = x[i].qh + 2*ir;
uint8_t sc, m;
get_scale_min_k4(is + 0, x[i].scales, sc, m);
const half d1 = __hmul(dall, __int2half_rn(sc)); const half m1 = __hmul(dmin, __int2half_rn(m));
get_scale_min_k4(is + 1, x[i].scales, sc, m);
const half d2 = __hmul(dall, __int2half_rn(sc)); const half m2 = __hmul(dmin, __int2half_rn(m));
uint8_t hm = 1 << (2*il);
y[ 0] = convert_from_half<dst_t>(__hsub(__hmul(d1, __int2half_rn((ql[0] & 0xF) + (qh[0] & hm ? 16 : 0))), m1));
y[ 1] = convert_from_half<dst_t>(__hsub(__hmul(d1, __int2half_rn((ql[1] & 0xF) + (qh[1] & hm ? 16 : 0))), m1));
hm <<= 1;
y[32] = convert_from_half<dst_t>(__hsub(__hmul(d2, __int2half_rn((ql[0] >> 4) + (qh[0] & hm ? 16 : 0))), m2));
y[33] = convert_from_half<dst_t>(__hsub(__hmul(d2, __int2half_rn((ql[1] >> 4) + (qh[1] & hm ? 16 : 0))), m2));
}
template<typename dst_t>
inline __device__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
const block_q6_K * x = (const block_q6_K *) vx;
const auto i = 0;
// assume 64 threads - this is very slightly better than the one below
const auto tid = threadIdx.x;
const int ip = tid/32; // ip is 0 or 1
const int il = tid - 32*ip; // 0...32
const int is = 8*ip + il/16;
dst_t * y = yy + i*QK_K + 128*ip + il;
const half d = x[i].d;
const uint8_t * ql = x[i].ql + 64*ip + il;
const uint8_t qh = x[i].qh[32*ip + il];
const int8_t * sc = x[i].scales + is;
y[ 0] = convert_from_half<dst_t>(__hmul(d, __int2half_rn(sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32))));
y[32] = convert_from_half<dst_t>(__hmul(d, __int2half_rn(sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32))));
y[64] = convert_from_half<dst_t>(__hmul(d, __int2half_rn(sc[4] * ((int8_t)((ql[ 0] >> 4) | (((qh >> 4) & 3) << 4)) - 32))));
y[96] = convert_from_half<dst_t>(__hmul(d, __int2half_rn(sc[6] * ((int8_t)((ql[32] >> 4) | (((qh >> 6) & 3) << 4)) - 32))));
}