CUDA: faster non k-quant mul_mat_q kernels (#2483)

This commit is contained in:
Johannes Gäßler 2023-08-02 18:04:04 +02:00 committed by GitHub
parent 4f6b60c776
commit 468ea24fb4
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23

View file

@ -1362,22 +1362,185 @@ static __global__ void dequantize_block(const void * __restrict__ vx, float * __
} }
// VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called // 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 1 #define VDR_Q4_0_Q8_1_MMVQ 2
#define VDR_Q4_0_Q8_1_MMQ 4
static __device__ __forceinline__ float vec_dot_q4_0_q8_1_impl( template <int vdr> static __device__ __forceinline__ float vec_dot_q4_0_q8_1_impl(
const int & vi, const int & ui0, const int & ui1, const half & d4, const half2 & ds8) { const int * v, const int * u, const float & d4, const half2 & ds8) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics #if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
// subtract 8 from each quantized value int sumi = 0;
const int vi0 = (vi >> 0) & 0x0F0F0F0F;
const int vi1 = (vi >> 4) & 0x0F0F0F0F; #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 // SIMD dot product of quantized values
int sumi = __dp4a(vi0, ui0, 0); sumi = __dp4a(vi0, u[2*i+0], sumi);
sumi = __dp4a(vi1, ui1, sumi); sumi = __dp4a(vi1, u[2*i+1], sumi);
}
return __half2float(d4) * (sumi * __half2float(ds8.x) - (8/QI4_0) * __half2float(ds8.y)); // second part effectively subtracts 8 from each quant value
return d4 * (sumi * __half2float(ds8.x) - (8*vdr/QI4_0) * __half2float(ds8.y));
#else
return 0.0f; // only to satisfy the compiler
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
}
#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) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
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 = __dp4a(vi0, u[2*i+0], sumi);
sumi = __dp4a(vi1, u[2*i+1], sumi);
}
#ifdef GGML_CUDA_F16
const half2 tmp = __hmul2(dm4, ds8);
const float d4d8 = __half2float(tmp.x);
const float m4s8 = __half2float(tmp.y);
#else
const float d4d8 = __half2float(dm4.x) * __half2float(ds8.x);
const float m4s8 = __half2float(dm4.y) * __half2float(ds8.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));
#else
return 0.0f; // only to satisfy the compiler
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
}
#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) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
int sumi = 0;
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 = __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 = __dp4a(vi1, u[2*i+1], sumi); // SIMD dot product of quantized values
}
// second part effectively subtracts 16 from each quant value
return d5 * (sumi*__half2float(ds8.x) - (16*vdr/QI5_0) * __half2float(ds8.y));
#else
return 0.0f; // only to satisfy the compiler
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
}
#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) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
int sumi = 0;
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 = __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 = __dp4a(vi1, u[2*i+1], sumi); // SIMD dot product of quantized values
}
#ifdef GGML_CUDA_F16
const half2 tmp = __hmul2(dm5, ds8);
const float d5d8 = __half2float(tmp.x);
const float m5s8 = __half2float(tmp.y);
#else
const float d5d8 = __half2float(dm5.x) * __half2float(ds8.x);
const float m5s8 = __half2float(dm5.y) * __half2float(ds8.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);
#else
return 0.0f; // only to satisfy the compiler
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
}
#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 half2 & ds8_1) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
int sumi = 0;
for (int i = 0; i < vdr; ++i) {
// SIMD dot product of quantized values
sumi = __dp4a(v[i], u[i], sumi);
}
return sumi * d8_0 * __half2float(ds8_1.x);
#else
return 0.0f; // only to satisfy the compiler
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
}
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) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
int sumi = 0;
for (int i = 0; i < vdr; ++i) {
// SIMD dot product of quantized values
sumi = __dp4a(v[i], u[i], sumi);
}
#ifdef GGML_CUDA_F16
const half2 tmp = __hmul2(dm8, ds8);
const float d8d8 = __half2float(tmp.x);
const float m8s8 = __half2float(tmp.y);
#else
const float d8d8 = __half2float(dm8.x) * __half2float(ds8.x);
const float m8s8 = __half2float(dm8.y) * __half2float(ds8.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);
#else #else
return 0.0f; // only to satisfy the compiler return 0.0f; // only to satisfy the compiler
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
@ -1388,20 +1551,26 @@ static __device__ __forceinline__ float vec_dot_q4_0_q8_1(
const block_q4_0 * bq4_0 = (const block_q4_0 *) vbq; const block_q4_0 * bq4_0 = (const block_q4_0 *) vbq;
const int vi = get_int_from_uint8(bq4_0->qs, iqs); int v[VDR_Q4_0_Q8_1_MMVQ];
const int ui0 = get_int_from_int8_aligned(bq8_1->qs, iqs); int u[2*VDR_Q4_0_Q8_1_MMVQ];
const int ui1 = get_int_from_int8_aligned(bq8_1->qs, iqs + QI4_0);
return vec_dot_q4_0_q8_1_impl(vi, ui0, ui1, bq4_0->d, bq8_1->ds); #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__ void allocate_tiles_q4_0(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { static __device__ __forceinline__ void allocate_tiles_q4_0(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
__shared__ int tile_x_qs[GGML_CUDA_MMQ_Y * (WARP_SIZE) + GGML_CUDA_MMQ_Y]; __shared__ int tile_x_qs[GGML_CUDA_MMQ_Y * (WARP_SIZE) + GGML_CUDA_MMQ_Y];
__shared__ half2 tile_x_d[GGML_CUDA_MMQ_Y * (WARP_SIZE/QI4_0) + GGML_CUDA_MMQ_Y/QI4_0]; __shared__ float tile_x_d[GGML_CUDA_MMQ_Y * (WARP_SIZE/QI4_0) + GGML_CUDA_MMQ_Y/QI4_0];
*x_ql = tile_x_qs; *x_ql = tile_x_qs;
*x_dm = tile_x_d; *x_dm = (half2 *) tile_x_d;
} }
template <bool need_check> static __device__ __forceinline__ void load_tiles_q4_0( template <bool need_check> static __device__ __forceinline__ void load_tiles_q4_0(
@ -1418,6 +1587,8 @@ template <bool need_check> static __device__ __forceinline__ void load_tiles_q4_
const block_q4_0 * bx0 = (block_q4_0 *) vx; const block_q4_0 * bx0 = (block_q4_0 *) vx;
float * x_dmf = (float *) x_dm;
#pragma unroll #pragma unroll
for (int i0 = 0; i0 < GGML_CUDA_MMQ_Y; i0 += 8) { for (int i0 = 0; i0 < GGML_CUDA_MMQ_Y; i0 += 8) {
int i = i0 + i_offset; int i = i0 + i_offset;
@ -1429,7 +1600,7 @@ template <bool need_check> static __device__ __forceinline__ void load_tiles_q4_
const block_q4_0 * bxi = bx0 + i*blocks_per_row + kbx; const block_q4_0 * bxi = bx0 + i*blocks_per_row + kbx;
x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8(bxi->qs, kqsx); x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8(bxi->qs, kqsx);
x_dm[i * (WARP_SIZE/QI4_0) + i / QI4_0 + kbx].x = bxi->d; x_dmf[i * (WARP_SIZE/QI4_0) + i / QI4_0 + kbx] = bxi->d;
} }
// const int blocks_per_tile_x_row = WARP_SIZE / QI4_0; // const int blocks_per_tile_x_row = WARP_SIZE / QI4_0;
@ -1462,39 +1633,19 @@ static __device__ __forceinline__ float vec_dot_q4_0_q8_1_mul_mat(
__builtin_assume(k < WARP_SIZE); __builtin_assume(k < WARP_SIZE);
const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2)); const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2));
const float * x_dmf = (float *) x_dm;
return vec_dot_q4_0_q8_1_impl( int u[2*VDR_Q4_0_Q8_1_MMQ];
x_ql[i * (WARP_SIZE + 1) + k], y_qs[j * (2*WARP_SIZE) + kyqs], y_qs[j * (2*WARP_SIZE) + kyqs + (QI8_1/2)],
x_dm[i * (WARP_SIZE/QI4_0) + i/QI4_0 + k/QI4_0].x, y_ds[j * (2*WARP_SIZE/QI8_1) + 2*k/QI8_1]);
}
#define VDR_q4_1_q8_1 1 #pragma unroll
for (int l = 0; l < VDR_Q4_0_Q8_1_MMQ; ++l) {
u[2*l+0] = y_qs[j * (2*WARP_SIZE) + kyqs + l];
u[2*l+1] = y_qs[j * (2*WARP_SIZE) + kyqs + l + QI4_0];
}
static __device__ __forceinline__ float vec_dot_q4_1_q8_1_impl( return vec_dot_q4_0_q8_1_impl<VDR_Q4_0_Q8_1_MMQ>
const int & vi, const int & ui0, const int & ui1, const half2 & dm4, const half2 & ds8) { (&x_ql[i * (WARP_SIZE + 1) + k], u, x_dmf[i * (WARP_SIZE/QI4_0) + i/QI4_0 + k/QI4_0],
y_ds[j * (2*WARP_SIZE/QI8_1) + 2*k/QI8_1]);
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
const int vi0 = (vi >> 0) & 0x0F0F0F0F;
const int vi1 = (vi >> 4) & 0x0F0F0F0F;
// SIMD dot product of quantized values
int sumi = __dp4a(vi0, ui0, 0);
sumi = __dp4a(vi1, ui1, sumi);
#ifdef GGML_CUDA_F16
const half2 tmp = __hmul2(dm4, ds8);
const float d4d8 = __half2float(tmp.x);
const float m4s8 = __half2float(tmp.y);
#else
const float d4d8 = __half2float(dm4.x) * __half2float(ds8.x);
const float m4s8 = __half2float(dm4.y) * __half2float(ds8.y);
#endif // GGML_CUDA_F16
// scale second part of sum by QI8_1/QR4_1 to compensate for multiple threads adding it
return sumi * d4d8 + m4s8 / (QI8_1 / QR4_1);
#else
return 0.0f; // only to satisfy the compiler
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
} }
static __device__ __forceinline__ float vec_dot_q4_1_q8_1( static __device__ __forceinline__ float vec_dot_q4_1_q8_1(
@ -1502,11 +1653,17 @@ static __device__ __forceinline__ float vec_dot_q4_1_q8_1(
const block_q4_1 * bq4_1 = (const block_q4_1 *) vbq; const block_q4_1 * bq4_1 = (const block_q4_1 *) vbq;
const int vi = get_int_from_uint8_aligned(bq4_1->qs, iqs); int v[VDR_Q4_1_Q8_1_MMVQ];
const int ui0 = get_int_from_int8_aligned(bq8_1->qs, iqs); int u[2*VDR_Q4_1_Q8_1_MMVQ];
const int ui1 = get_int_from_int8_aligned(bq8_1->qs, iqs + QI4_1);
return vec_dot_q4_1_q8_1_impl(vi, ui0, ui1, bq4_1->dm, bq8_1->ds); #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__ void allocate_tiles_q4_1(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { static __device__ __forceinline__ void allocate_tiles_q4_1(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
@ -1575,35 +1732,17 @@ static __device__ __forceinline__ float vec_dot_q4_1_q8_1_mul_mat(
const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2)); const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2));
return vec_dot_q4_1_q8_1_impl( int u[2*VDR_Q4_1_Q8_1_MMQ];
x_ql[i * (WARP_SIZE + 1) + k], y_qs[j * (2*WARP_SIZE) + kyqs], y_qs[j * (2*WARP_SIZE) + kyqs + (QI8_1/2)],
x_dm[i * (WARP_SIZE/QI4_1) + i/QI4_1 + k/QI4_1], y_ds[j * (2*WARP_SIZE/QI8_1) + 2*k/QI8_1]);
}
#define VDR_q5_0_q8_1 1 #pragma unroll
for (int l = 0; l < VDR_Q4_1_Q8_1_MMQ; ++l) {
u[2*l+0] = y_qs[j * (2*WARP_SIZE) + kyqs + l];
u[2*l+1] = y_qs[j * (2*WARP_SIZE) + kyqs + l + QI4_1];
}
static __device__ __forceinline__ float vec_dot_q5_0_q8_1_impl( return vec_dot_q4_1_q8_1_impl<VDR_Q4_1_Q8_1_MMQ>
const int & qs, const int & qh, const int & ui0, const int & ui1, const half & d5, const half2 & ds8) { (&x_ql[i * (WARP_SIZE + 1) + k], u, x_dm[i * (WARP_SIZE/QI4_1) + i/QI4_1 + k/QI4_1],
y_ds[j * (2*WARP_SIZE/QI8_1) + 2*k/QI8_1]);
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
int vi0 = (qs >> 0) & 0x0F0F0F0F; // lower 4 qs bits, still need qh as 5th bits
vi0 |= (qh << 4) & 0x00000010; // 0 -> 4
vi0 |= (qh << 11) & 0x00001000; // 1 -> 12
vi0 |= (qh << 18) & 0x00100000; // 2 -> 20
vi0 |= (qh << 25) & 0x10000000; // 3 -> 28
int sumi = __dp4a(vi0, ui0, 0); // SIMD dot product of quantized values
int vi1 = (qs >> 4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh as 5th bits
vi1 |= (qh >> 12) & 0x00000010; // 16 -> 4
vi1 |= (qh >> 5) & 0x00001000; // 17 -> 12
vi1 |= (qh << 2) & 0x00100000; // 18 -> 20
vi1 |= (qh << 9) & 0x10000000; // 19 -> 28
sumi = __dp4a(vi1, ui1, sumi); // SIMD dot product of quantized values
return __half2float(d5) * (sumi*__half2float(ds8.x) - (16/QI5_0) * __half2float(ds8.y));
#else
return 0.0f; // only to satisfy the compiler
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
} }
static __device__ __forceinline__ float vec_dot_q5_0_q8_1( static __device__ __forceinline__ float vec_dot_q5_0_q8_1(
@ -1611,23 +1750,28 @@ static __device__ __forceinline__ float vec_dot_q5_0_q8_1(
const block_q5_0 * bq5_0 = (const block_q5_0 *) vbq; const block_q5_0 * bq5_0 = (const block_q5_0 *) vbq;
const int qs = get_int_from_uint8(bq5_0->qs, iqs); int vl[VDR_Q5_0_Q8_1_MMVQ];
const int qh = get_int_from_uint8(bq5_0->qh, 0) >> (4 * iqs); int vh[VDR_Q5_0_Q8_1_MMVQ];
const int ui0 = get_int_from_int8_aligned(bq8_1->qs, iqs); int u[2*VDR_Q5_0_Q8_1_MMVQ];
const int ui1 = get_int_from_int8_aligned(bq8_1->qs, iqs + QI5_0);
return vec_dot_q5_0_q8_1_impl(qs, qh, ui0, ui1, bq5_0->d, bq8_1->ds); #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__ void allocate_tiles_q5_0(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { static __device__ __forceinline__ void allocate_tiles_q5_0(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
__shared__ int tile_x_ql[GGML_CUDA_MMQ_Y * (WARP_SIZE) + GGML_CUDA_MMQ_Y]; __shared__ int tile_x_ql[GGML_CUDA_MMQ_Y * (2*WARP_SIZE) + GGML_CUDA_MMQ_Y];
__shared__ int tile_x_qh[GGML_CUDA_MMQ_Y * (WARP_SIZE/QI5_0) + GGML_CUDA_MMQ_Y/QI5_0]; __shared__ float tile_x_d[GGML_CUDA_MMQ_Y * (WARP_SIZE/QI5_0) + GGML_CUDA_MMQ_Y/QI5_0];
__shared__ half2 tile_x_d[GGML_CUDA_MMQ_Y * (WARP_SIZE/QI5_0) + GGML_CUDA_MMQ_Y/QI5_0];
*x_ql = tile_x_ql; *x_ql = tile_x_ql;
*x_qh = tile_x_qh; *x_dm = (half2 *) tile_x_d;
*x_dm = tile_x_d;
} }
template <bool need_check> static __device__ __forceinline__ void load_tiles_q5_0( template <bool need_check> static __device__ __forceinline__ void load_tiles_q5_0(
@ -1654,11 +1798,31 @@ template <bool need_check> static __device__ __forceinline__ void load_tiles_q5_
const block_q5_0 * bxi = bx0 + i*blocks_per_row + kbx; const block_q5_0 * bxi = bx0 + i*blocks_per_row + kbx;
x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8(bxi->qs, kqsx); const int ql = get_int_from_uint8(bxi->qs, kqsx);
const int qh = get_int_from_uint8(bxi->qh, 0) >> (4 * (k % QI5_0));
int qs0 = (ql >> 0) & 0x0F0F0F0F;
qs0 |= (qh << 4) & 0x00000010; // 0 -> 4
qs0 |= (qh << 11) & 0x00001000; // 1 -> 12
qs0 |= (qh << 18) & 0x00100000; // 2 -> 20
qs0 |= (qh << 25) & 0x10000000; // 3 -> 28
qs0 = __vsubss4(qs0, 0x10101010); // subtract 16
x_ql[i * (2*WARP_SIZE + 1) + 2*k+0] = qs0;
int qs1 = (ql >> 4) & 0x0F0F0F0F;
qs1 |= (qh >> 12) & 0x00000010; // 16 -> 4
qs1 |= (qh >> 5) & 0x00001000; // 17 -> 12
qs1 |= (qh << 2) & 0x00100000; // 18 -> 20
qs1 |= (qh << 9) & 0x10000000; // 19 -> 28
qs1 = __vsubss4(qs1, 0x10101010); // subtract 16
x_ql[i * (2*WARP_SIZE + 1) + 2*k+1] = qs1;
} }
const int blocks_per_tile_x_row = WARP_SIZE / QI5_0; const int blocks_per_tile_x_row = WARP_SIZE / QI5_0;
const int kbxd = k % blocks_per_tile_x_row; const int kbxd = k % blocks_per_tile_x_row;
float * x_dmf = (float *) x_dm;
#pragma unroll #pragma unroll
for (int i0 = 0; i0 < GGML_CUDA_MMQ_Y; i0 += 8 * QI5_0) { for (int i0 = 0; i0 < GGML_CUDA_MMQ_Y; i0 += 8 * QI5_0) {
@ -1670,8 +1834,7 @@ template <bool need_check> static __device__ __forceinline__ void load_tiles_q5_
const block_q5_0 * bxi = bx0 + i*blocks_per_row + kbxd; const block_q5_0 * bxi = bx0 + i*blocks_per_row + kbxd;
x_qh[i * (WARP_SIZE/QI5_0) + i / QI5_0 + kbxd] = get_int_from_uint8(bxi->qh, 0); x_dmf[i * (WARP_SIZE/QI5_0) + i / QI5_0 + kbxd] = bxi->d;
x_dm[i * (WARP_SIZE/QI5_0) + i / QI5_0 + kbxd].x = bxi->d;
} }
} }
@ -1688,46 +1851,18 @@ static __device__ __forceinline__ float vec_dot_q5_0_q8_1_mul_mat(
const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2)); const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2));
const int index_bx = i * (WARP_SIZE/QI5_0) + i/QI5_0 + k/QI5_0; const int index_bx = i * (WARP_SIZE/QI5_0) + i/QI5_0 + k/QI5_0;
const float * x_dmf = (float *) x_dm;
return vec_dot_q5_0_q8_1_impl( int u[2*VDR_Q5_0_Q8_1_MMQ];
x_ql[i * (WARP_SIZE + 1) + k], x_qh[index_bx] >> (4 * (k % QI5_0)), y_qs[j * (2*WARP_SIZE) + kyqs],
y_qs[j * (2*WARP_SIZE) + kyqs + (QI8_1/2)], x_dm[index_bx].x, y_ds[j * (2*WARP_SIZE/QI8_1) + 2*k/QI8_1]);
}
#define VDR_q5_1_q8_1 1 #pragma unroll
for (int l = 0; l < VDR_Q5_0_Q8_1_MMQ; ++l) {
u[2*l+0] = y_qs[j * (2*WARP_SIZE) + kyqs + l];
u[2*l+1] = y_qs[j * (2*WARP_SIZE) + kyqs + l + QI5_0];
}
static __device__ __forceinline__ float vec_dot_q5_1_q8_1_impl( return vec_dot_q8_0_q8_1_impl<QR5_0*VDR_Q5_0_Q8_1_MMQ>
const int & qs, const int & qh, const int & ui0, const int & ui1, const half2 & dm5, const half2 & ds8) { (&x_ql[i * (2*WARP_SIZE + 1) + 2 * k], u, x_dmf[index_bx], y_ds[j * (2*WARP_SIZE/QI8_1) + 2*k/QI8_1]);
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
int vi0 = (qs >> 0) & 0x0F0F0F0F; // lower 4 qs bits, still need qh0 as 5th bits
vi0 |= (qh << 4) & 0x00000010; // 0 -> 4
vi0 |= (qh << 11) & 0x00001000; // 1 -> 12
vi0 |= (qh << 18) & 0x00100000; // 2 -> 20
vi0 |= (qh << 25) & 0x10000000; // 3 -> 28
int sumi = __dp4a(vi0, ui0, 0); // SIMD dot product of quantized values
int vi1 = (qs >> 4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh1 as 5th bits
vi1 |= (qh >> 12) & 0x00000010; // 16 -> 4
vi1 |= (qh >> 5) & 0x00001000; // 17 -> 12
vi1 |= (qh << 2) & 0x00100000; // 18 -> 20
vi1 |= (qh << 9) & 0x10000000; // 19 -> 28
sumi = __dp4a(vi1, ui1, sumi); // SIMD dot product of quantized values
#ifdef GGML_CUDA_F16
const half2 tmp = __hmul2(dm5, ds8);
const float d5d8 = __half2float(tmp.x);
const float m5s8 = __half2float(tmp.y);
#else
const float d5d8 = __half2float(dm5.x) * __half2float(ds8.x);
const float m5s8 = __half2float(dm5.y) * __half2float(ds8.y);
#endif // GGML_CUDA_F16
return sumi*d5d8 + m5s8/QI5_1; // scale sum by QI5_1 because there are QI5_1 threads working on this block
#else
return 0.0f; // only to satisfy the compiler
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
} }
static __device__ __forceinline__ float vec_dot_q5_1_q8_1( static __device__ __forceinline__ float vec_dot_q5_1_q8_1(
@ -1735,22 +1870,27 @@ static __device__ __forceinline__ float vec_dot_q5_1_q8_1(
const block_q5_1 * bq5_1 = (const block_q5_1 *) vbq; const block_q5_1 * bq5_1 = (const block_q5_1 *) vbq;
const int qs = get_int_from_uint8_aligned(bq5_1->qs, iqs); int vl[VDR_Q5_1_Q8_1_MMVQ];
const int qh = get_int_from_uint8_aligned(bq5_1->qh, 0) >> (4 * iqs); int vh[VDR_Q5_1_Q8_1_MMVQ];
const int ui0 = get_int_from_int8_aligned(bq8_1->qs, iqs); int u[2*VDR_Q5_1_Q8_1_MMVQ];
const int ui1 = get_int_from_int8_aligned(bq8_1->qs, iqs + QI5_1);
return vec_dot_q5_1_q8_1_impl(qs, qh, ui0, ui1, bq5_1->dm, bq8_1->ds); #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__ void allocate_tiles_q5_1(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { static __device__ __forceinline__ void allocate_tiles_q5_1(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
__shared__ int tile_x_ql[GGML_CUDA_MMQ_Y * (WARP_SIZE ) + GGML_CUDA_MMQ_Y]; __shared__ int tile_x_ql[GGML_CUDA_MMQ_Y * (2*WARP_SIZE) + GGML_CUDA_MMQ_Y];
__shared__ int tile_x_qh[GGML_CUDA_MMQ_Y * (WARP_SIZE/QI5_1) + GGML_CUDA_MMQ_Y/QI5_1];
__shared__ half2 tile_x_dm[GGML_CUDA_MMQ_Y * (WARP_SIZE/QI5_1) + GGML_CUDA_MMQ_Y/QI5_1]; __shared__ half2 tile_x_dm[GGML_CUDA_MMQ_Y * (WARP_SIZE/QI5_1) + GGML_CUDA_MMQ_Y/QI5_1];
*x_ql = tile_x_ql; *x_ql = tile_x_ql;
*x_qh = tile_x_qh;
*x_dm = tile_x_dm; *x_dm = tile_x_dm;
} }
@ -1778,7 +1918,24 @@ template <bool need_check> static __device__ __forceinline__ void load_tiles_q5_
const block_q5_1 * bxi = bx0 + i*blocks_per_row + kbx; const block_q5_1 * bxi = bx0 + i*blocks_per_row + kbx;
x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx); const int ql = get_int_from_uint8_aligned(bxi->qs, kqsx);
const int qh = get_int_from_uint8_aligned(bxi->qh, 0) >> (4 * (k % QI5_1));
int qs0 = (ql >> 0) & 0x0F0F0F0F;
qs0 |= (qh << 4) & 0x00000010; // 0 -> 4
qs0 |= (qh << 11) & 0x00001000; // 1 -> 12
qs0 |= (qh << 18) & 0x00100000; // 2 -> 20
qs0 |= (qh << 25) & 0x10000000; // 3 -> 28
x_ql[i * (2*WARP_SIZE + 1) + 2*k+0] = qs0;
int qs1 = (ql >> 4) & 0x0F0F0F0F;
qs1 |= (qh >> 12) & 0x00000010; // 16 -> 4
qs1 |= (qh >> 5) & 0x00001000; // 17 -> 12
qs1 |= (qh << 2) & 0x00100000; // 18 -> 20
qs1 |= (qh << 9) & 0x10000000; // 19 -> 28
x_ql[i * (2*WARP_SIZE + 1) + 2*k+1] = qs1;
} }
const int blocks_per_tile_x_row = WARP_SIZE / QI5_1; const int blocks_per_tile_x_row = WARP_SIZE / QI5_1;
@ -1794,7 +1951,6 @@ template <bool need_check> static __device__ __forceinline__ void load_tiles_q5_
const block_q5_1 * bxi = bx0 + i*blocks_per_row + kbxd; const block_q5_1 * bxi = bx0 + i*blocks_per_row + kbxd;
x_qh[i * (WARP_SIZE/QI5_1) + i / QI5_1 + kbxd] = get_int_from_uint8_aligned(bxi->qh, 0);
x_dm[i * (WARP_SIZE/QI5_1) + i / QI5_1 + kbxd] = bxi->dm; x_dm[i * (WARP_SIZE/QI5_1) + i / QI5_1 + kbxd] = bxi->dm;
} }
} }
@ -1813,24 +1969,16 @@ static __device__ __forceinline__ float vec_dot_q5_1_q8_1_mul_mat(
const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2)); const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2));
const int index_bx = i * (WARP_SIZE/QI5_1) + + i/QI5_1 + k/QI5_1; const int index_bx = i * (WARP_SIZE/QI5_1) + + i/QI5_1 + k/QI5_1;
return vec_dot_q5_1_q8_1_impl( int u[2*VDR_Q5_1_Q8_1_MMQ];
x_ql[i * (WARP_SIZE + 1) + k], x_qh[index_bx] >> (4 * (k % QI5_1)), y_qs[j * (2*WARP_SIZE) + kyqs],
y_qs[j * (2*WARP_SIZE) + kyqs + (QI8_1/2)], x_dm[index_bx], y_ds[j * (2*WARP_SIZE/QI8_1) + 2*k/QI8_1]);
}
#define VDR_q8_0_q8_1 1 #pragma unroll
for (int l = 0; l < VDR_Q5_1_Q8_1_MMQ; ++l) {
u[2*l+0] = y_qs[j * (2*WARP_SIZE) + kyqs + l];
u[2*l+1] = y_qs[j * (2*WARP_SIZE) + kyqs + l + QI5_1];
}
static __device__ __forceinline__ float vec_dot_q8_0_q8_1_impl( return vec_dot_q8_1_q8_1_impl<QR5_1*VDR_Q5_1_Q8_1_MMQ>
const int & vi, const int & ui, const half & d8_0, const half2 & ds8_1) { (&x_ql[i * (2*WARP_SIZE + 1) + 2 * k], u, x_dm[index_bx], y_ds[j * (2*WARP_SIZE/QI8_1) + 2*k/QI8_1]);
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
// SIMD dot product of quantized values
const int sumi = __dp4a(vi, ui, 0);
return sumi * __half2float(d8_0) * __half2float(ds8_1.x);
#else
return 0.0f; // only to satisfy the compiler
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
} }
static __device__ __forceinline__ float vec_dot_q8_0_q8_1( static __device__ __forceinline__ float vec_dot_q8_0_q8_1(
@ -1838,19 +1986,24 @@ static __device__ __forceinline__ float vec_dot_q8_0_q8_1(
const block_q8_0 * bq8_0 = (const block_q8_0 *) vbq; const block_q8_0 * bq8_0 = (const block_q8_0 *) vbq;
const int vi = get_int_from_int8(bq8_0->qs, iqs); int v[VDR_Q8_0_Q8_1_MMVQ];
const int ui = get_int_from_int8_aligned(bq8_1->qs, iqs); int u[VDR_Q8_0_Q8_1_MMVQ];
return vec_dot_q8_0_q8_1_impl(vi, ui, bq8_0->d, bq8_1->ds); 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, bq8_1->ds);
} }
static __device__ __forceinline__ void allocate_tiles_q8_0(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { static __device__ __forceinline__ void allocate_tiles_q8_0(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
__shared__ int tile_x_qs[GGML_CUDA_MMQ_Y * (WARP_SIZE) + GGML_CUDA_MMQ_Y]; __shared__ int tile_x_qs[GGML_CUDA_MMQ_Y * (WARP_SIZE) + GGML_CUDA_MMQ_Y];
__shared__ half2 tile_x_d[GGML_CUDA_MMQ_Y * (WARP_SIZE/QI8_0) + GGML_CUDA_MMQ_Y/QI8_0]; __shared__ float tile_x_d[GGML_CUDA_MMQ_Y * (WARP_SIZE/QI8_0) + GGML_CUDA_MMQ_Y/QI8_0];
*x_ql = tile_x_qs; *x_ql = tile_x_qs;
*x_dm = tile_x_d; *x_dm = (half2 *) tile_x_d;
} }
template <bool need_check> static __device__ __forceinline__ void load_tiles_q8_0( template <bool need_check> static __device__ __forceinline__ void load_tiles_q8_0(
@ -1864,6 +2017,7 @@ template <bool need_check> static __device__ __forceinline__ void load_tiles_q8_
const int kbx = k / QI8_0; const int kbx = k / QI8_0;
const int kqsx = k % QI8_0; const int kqsx = k % QI8_0;
float * x_dmf = (float *) x_dm;
const block_q8_0 * bx0 = (block_q8_0 *) vx; const block_q8_0 * bx0 = (block_q8_0 *) vx;
@ -1878,7 +2032,7 @@ template <bool need_check> static __device__ __forceinline__ void load_tiles_q8_
const block_q8_0 * bxi = bx0 + i*blocks_per_row + kbx; const block_q8_0 * bxi = bx0 + i*blocks_per_row + kbx;
x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_int8(bxi->qs, kqsx); x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_int8(bxi->qs, kqsx);
x_dm[i * (WARP_SIZE/QI8_0) + i / QI8_0 + kbx].x = bxi->d; x_dmf[i * (WARP_SIZE/QI8_0) + i / QI8_0 + kbx] = bxi->d;
} }
// const int blocks_per_tile_x_row = WARP_SIZE / QI8_0; // const int blocks_per_tile_x_row = WARP_SIZE / QI8_0;
@ -1912,9 +2066,11 @@ static __device__ __forceinline__ float vec_dot_q8_0_q8_1_mul_mat(
__builtin_assume(k >= 0); __builtin_assume(k >= 0);
__builtin_assume(k < WARP_SIZE); __builtin_assume(k < WARP_SIZE);
return vec_dot_q8_0_q8_1_impl( const float * x_dmf = (float *) x_dm;
x_ql[i * (WARP_SIZE + 1) + k], y_qs[j*WARP_SIZE + k],
x_dm[i * (WARP_SIZE/QI8_0) + i/QI8_0 + k/QI8_0].x, y_ds[j * (WARP_SIZE/QI8_1) + k/QI8_1]); return vec_dot_q8_0_q8_1_impl<VDR_Q8_0_Q8_1_MMQ>
(&x_ql[i * (WARP_SIZE + 1) + k], &y_qs[j * WARP_SIZE + k], x_dmf[i * (WARP_SIZE/QI8_0) + i/QI8_0 + k/QI8_0],
y_ds[j * (WARP_SIZE/QI8_1) + k/QI8_1]);
} }
#define VDR_q2_K_q8_1 1 #define VDR_q2_K_q8_1 1
@ -2288,15 +2444,15 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1(
int u[2*QR4_K]; int u[2*QR4_K];
float d8[QR4_K]; float d8[QR4_K];
// iqs is in 0...15. bq8_offset = 2 * (iqs/4) -> bq8_offset = 0, 2, 4, 6 // iqs is in 0,2..30. bq8_offset = iqs/4 -> bq8_offset = 0, 2, 4, 6
const int bq8_offset = QR4_K * (iqs / (QI8_1/2)); 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 = 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 = 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 = 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 // 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%4)); const int * q4 = (const int *)(bq4_K->qs + 16 * bq8_offset + 4 * ((iqs/2)%4));
v[0] = q4[0]; v[0] = q4[0];
v[1] = q4[4]; v[1] = q4[4];
@ -2317,7 +2473,7 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1(
const block_q8_1 * bq8i = bq8_1 + bq8_offset + i; const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
d8[i] = bq8i->ds.x; d8[i] = bq8i->ds.x;
const int * q8 = (const int *)bq8i->qs + (iqs%4); const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4);
u[2*i+0] = q8[0]; u[2*i+0] = q8[0];
u[2*i+1] = q8[4]; u[2*i+1] = q8[4];
} }
@ -2345,12 +2501,12 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1(
const float d8_1 = bq8_1[0].ds.x; const float d8_1 = bq8_1[0].ds.x;
const float d8_2 = bq8_1[1].ds.x; const float d8_2 = bq8_1[1].ds.x;
const int ui1 = *((const int *)bq8_1[0].qs + iqs); const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2));
const int ui2 = *((const int *)bq8_1[0].qs + iqs + 4); const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4);
const int ui3 = *((const int *)bq8_1[1].qs + iqs); const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2));
const int ui4 = *((const int *)bq8_1[1].qs + iqs + 4); const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4);
const int * q4 = (const int *)bq4_K->qs + iqs; const int * q4 = (const int *)bq4_K->qs + (iqs/2);
const int v1 = q4[0]; const int v1 = q4[0];
const int v2 = q4[4]; const int v2 = q4[4];
@ -2457,11 +2613,11 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1_mul_mat(
int u[2*QR4_K]; int u[2*QR4_K];
float d8[QR4_K]; float d8[QR4_K];
// iqs is in 0...15. bq8_offset = 2 * (iqs/4) -> bq8_offset = 0, 2, 4, 6 // kqsx is in 0,2...30. bq8_offset = 2 * (kqsx/4) -> bq8_offset = 0, 2, 4, 6
const int bq8_offset = QR4_K * (kqsx / (QI8_1/2)); const int bq8_offset = QR4_K * ((kqsx/2) / (QI8_1/2));
v[0] = x_ql[i * (WARP_SIZE + 1) + 4 * bq8_offset + kqsx % 4 + 0]; v[0] = x_ql[i * (WARP_SIZE + 1) + 4 * bq8_offset + (kqsx/2) % 4 + 0];
v[1] = x_ql[i * (WARP_SIZE + 1) + 4 * bq8_offset + kqsx % 4 + 4]; v[1] = x_ql[i * (WARP_SIZE + 1) + 4 * bq8_offset + (kqsx/2) % 4 + 4];
const uint16_t * scales = (const uint16_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + kbx * 4]; const uint16_t * scales = (const uint16_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + kbx * 4];
uint16_t aux[2]; uint16_t aux[2];
@ -2477,7 +2633,7 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1_mul_mat(
const uint8_t * m = sc + 2; const uint8_t * m = sc + 2;
for (int l = 0; l < QR4_K; ++l) { for (int l = 0; l < QR4_K; ++l) {
const int kqsy = j * (QR4_K*WARP_SIZE) + kbx * (QR4_K*QI4_K) + (bq8_offset + l) * QI8_1 + kqsx % (QI8_1/2); const int kqsy = j * (QR4_K*WARP_SIZE) + kbx * (QR4_K*QI4_K) + (bq8_offset + l) * QI8_1 + (kqsx/2) % (QI8_1/2);
u[2*l+0] = y_qs[kqsy + 0*(QI8_1/2)]; u[2*l+0] = y_qs[kqsy + 0*(QI8_1/2)];
u[2*l+1] = y_qs[kqsy + 1*(QI8_1/2)]; u[2*l+1] = y_qs[kqsy + 1*(QI8_1/2)];
d8[l] = y_ds[kqsy / QI8_1].x; d8[l] = y_ds[kqsy / QI8_1].x;
@ -2532,9 +2688,9 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1(
int u[2*QR5_K]; int u[2*QR5_K];
float d8[QR5_K]; float d8[QR5_K];
const int bq8_offset = QR5_K * (iqs / (QI8_1/2)); const int bq8_offset = QR5_K * ((iqs/2) / (QI8_1/2));
const int * ql = (const int *)(bq5_K->qs + 16 * bq8_offset + 4 * (iqs%4)); 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%4)); const int * qh = (const int *)(bq5_K->qh + 4 * ((iqs/2)%4));
vl[0] = ql[0]; vl[0] = ql[0];
vl[1] = ql[4]; vl[1] = ql[4];
@ -2559,7 +2715,7 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1(
const block_q8_1 * bq8i = bq8_1 + bq8_offset + i; const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
d8[i] = bq8i->ds.x; d8[i] = bq8i->ds.x;
const int * q8 = (const int *)bq8i->qs + (iqs%4); const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4);
u[2*i+0] = q8[0]; u[2*i+0] = q8[0];
u[2*i+1] = q8[4]; u[2*i+1] = q8[4];
} }
@ -2578,17 +2734,17 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1(
const float d8_1 = bq8_1[0].ds.x; const float d8_1 = bq8_1[0].ds.x;
const float d8_2 = bq8_1[1].ds.x; const float d8_2 = bq8_1[1].ds.x;
const int ui1 = *((const int *)bq8_1[0].qs + iqs); const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2));
const int ui2 = *((const int *)bq8_1[0].qs + iqs + 4); const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4);
const int ui3 = *((const int *)bq8_1[1].qs + iqs); const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2));
const int ui4 = *((const int *)bq8_1[1].qs + iqs + 4); const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4);
const int * ql = (const int *)bq5_K->qs + iqs; const int * ql = (const int *)bq5_K->qs + (iqs/2);
const int vl1 = ql[0]; const int vl1 = ql[0];
const int vl2 = ql[4]; const int vl2 = ql[4];
const int step = 4 * iqs; // 0, 4, 8, 12 const int step = 4 * (iqs/2); // 0, 4, 8, 12
const int im = step/8; // = 0 for iqs = 0, 1, = 1 for iqs = 2, 3 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 in = step%8; // 0, 4, 0, 4
const int vh = (*((const int *)(bq5_K->qh + in))) >> im; const int vh = (*((const int *)(bq5_K->qh + in))) >> im;
@ -2711,13 +2867,13 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1_mul_mat(
int u[2*QR4_K]; int u[2*QR4_K];
float d8[QR4_K]; float d8[QR4_K];
const int bq8_offset = QR5_K * (kqsx / (QI8_1/2)); const int bq8_offset = QR5_K * ((kqsx/2) / (QI8_1/2));
vl[0] = x_ql[i * (WARP_SIZE + 1) + 4 * bq8_offset + kqsx % 4 + 0]; vl[0] = x_ql[i * (WARP_SIZE + 1) + 4 * bq8_offset + (kqsx/2) % 4 + 0];
vl[1] = x_ql[i * (WARP_SIZE + 1) + 4 * bq8_offset + kqsx % 4 + 4]; vl[1] = x_ql[i * (WARP_SIZE + 1) + 4 * bq8_offset + (kqsx/2) % 4 + 4];
vh[0] = x_qh[i * (WARP_SIZE/4) + i/4 + kqsx % 4 + 0] >> bq8_offset; vh[0] = x_qh[i * (WARP_SIZE/4) + i/4 + (kqsx/2) % 4 + 0] >> bq8_offset;
vh[1] = x_qh[i * (WARP_SIZE/4) + i/4 + kqsx % 4 + 4] >> bq8_offset; vh[1] = x_qh[i * (WARP_SIZE/4) + i/4 + (kqsx/2) % 4 + 4] >> bq8_offset;
const uint16_t * scales = (const uint16_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + kbx * 4]; const uint16_t * scales = (const uint16_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + kbx * 4];
uint16_t aux[2]; uint16_t aux[2];
@ -2733,7 +2889,7 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1_mul_mat(
const uint8_t * m = sc + 2; const uint8_t * m = sc + 2;
for (int l = 0; l < QR5_K; ++l) { for (int l = 0; l < QR5_K; ++l) {
const int kqsy = j * (QR5_K*WARP_SIZE) + kbx * (QR5_K*QI5_K) + (bq8_offset + l) * QI8_1 + kqsx % (QI8_1/2); const int kqsy = j * (QR5_K*WARP_SIZE) + kbx * (QR5_K*QI5_K) + (bq8_offset + l) * QI8_1 + (kqsx/2) % (QI8_1/2);
u[2*l+0] = y_qs[kqsy + 0*(QI8_1/2)]; u[2*l+0] = y_qs[kqsy + 0*(QI8_1/2)];
u[2*l+1] = y_qs[kqsy + 1*(QI8_1/2)]; u[2*l+1] = y_qs[kqsy + 1*(QI8_1/2)];
d8[l] = y_ds[kqsy / QI8_1].x; d8[l] = y_ds[kqsy / QI8_1].x;
@ -2982,7 +3138,7 @@ static __global__ void mul_mat_q(
#if __CUDA_ARCH__ >= 700 // Unrolling the loop is slower on Pascal #if __CUDA_ARCH__ >= 700 // Unrolling the loop is slower on Pascal
#pragma unroll #pragma unroll
#endif // __CUDA_ARCH__ >= 700 #endif // __CUDA_ARCH__ >= 700
for (int k = 0; k < WARP_SIZE/vdr; ++k) { for (int k = 0; k < WARP_SIZE; k += vdr) {
#pragma unroll #pragma unroll
for (int j = 0; j < WARP_SIZE; j += 8) { for (int j = 0; j < WARP_SIZE; j += 8) {
#pragma unroll #pragma unroll
@ -3034,9 +3190,9 @@ static __global__ void mul_mat_vec_q(const void * __restrict__ vx, const void *
for (int i = 0; i < blocks_per_row; i += blocks_per_warp) { for (int i = 0; i < blocks_per_row; i += blocks_per_warp) {
const int ibx = row*blocks_per_row + i + threadIdx.x / (qi/vdr); // x block index const int ibx = row*blocks_per_row + i + threadIdx.x / (qi/vdr); // x block index
const int iby = (i + threadIdx.x / (qi/vdr)) * qk/QK8_1; // y block index that aligns with ibx const int iby = (i + threadIdx.x / (qi/vdr)) * (qk/QK8_1); // y block index that aligns with ibx
const int iqs = threadIdx.x % (qi/vdr); // x block quant index when casting the quants to int const int iqs = vdr * (threadIdx.x % (qi/vdr)); // x block quant index when casting the quants to int
tmp += vec_dot_q_cuda(&x[ibx], &y[iby], iqs); tmp += vec_dot_q_cuda(&x[ibx], &y[iby], iqs);
} }
@ -3579,7 +3735,7 @@ static void mul_mat_vec_q4_0_q8_1_cuda(const void * vx, const void * vy, float *
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
const dim3 block_nums(1, block_num_y, 1); const dim3 block_nums(1, block_num_y, 1);
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
mul_mat_vec_q<QK4_0, QI4_0, block_q4_0, VDR_q4_0_q8_1, vec_dot_q4_0_q8_1> mul_mat_vec_q<QK4_0, QI4_0, block_q4_0, VDR_Q4_0_Q8_1_MMVQ, vec_dot_q4_0_q8_1>
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows); <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
} }
@ -3588,7 +3744,7 @@ static void mul_mat_vec_q4_1_q8_1_cuda(const void * vx, const void * vy, float *
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
const dim3 block_nums(1, block_num_y, 1); const dim3 block_nums(1, block_num_y, 1);
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
mul_mat_vec_q<QK4_0, QI4_1, block_q4_1, VDR_q4_1_q8_1, vec_dot_q4_1_q8_1> mul_mat_vec_q<QK4_0, QI4_1, block_q4_1, VDR_Q4_1_Q8_1_MMVQ, vec_dot_q4_1_q8_1>
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows); <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
} }
@ -3597,7 +3753,7 @@ static void mul_mat_vec_q5_0_q8_1_cuda(const void * vx, const void * vy, float *
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
const dim3 block_nums(1, block_num_y, 1); const dim3 block_nums(1, block_num_y, 1);
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
mul_mat_vec_q<QK5_0, QI5_0, block_q5_0, VDR_q5_0_q8_1, vec_dot_q5_0_q8_1> mul_mat_vec_q<QK5_0, QI5_0, block_q5_0, VDR_Q5_0_Q8_1_MMVQ, vec_dot_q5_0_q8_1>
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows); <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
} }
@ -3606,7 +3762,7 @@ static void mul_mat_vec_q5_1_q8_1_cuda(const void * vx, const void * vy, float *
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
const dim3 block_nums(1, block_num_y, 1); const dim3 block_nums(1, block_num_y, 1);
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
mul_mat_vec_q<QK5_1, QI5_1, block_q5_1, VDR_q5_1_q8_1, vec_dot_q5_1_q8_1> mul_mat_vec_q<QK5_1, QI5_1, block_q5_1, VDR_Q5_1_Q8_1_MMVQ, vec_dot_q5_1_q8_1>
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows); <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
} }
@ -3615,7 +3771,7 @@ static void mul_mat_vec_q8_0_q8_1_cuda(const void * vx, const void * vy, float *
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
const dim3 block_nums(1, block_num_y, 1); const dim3 block_nums(1, block_num_y, 1);
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
mul_mat_vec_q<QK8_0, QI8_0, block_q8_0, VDR_q8_0_q8_1, vec_dot_q8_0_q8_1> mul_mat_vec_q<QK8_0, QI8_0, block_q8_0, VDR_Q8_0_Q8_1_MMVQ, vec_dot_q8_0_q8_1>
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows); <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
} }
@ -3717,10 +3873,10 @@ static void ggml_mul_mat_q4_0_q8_1_cuda(
const dim3 block_dims(WARP_SIZE, WARP_SIZE/4, 1); const dim3 block_dims(WARP_SIZE, WARP_SIZE/4, 1);
if (nrows_x % GGML_CUDA_MMQ_Y == 0) { if (nrows_x % GGML_CUDA_MMQ_Y == 0) {
mul_mat_q<QK4_0, QR4_0, QI4_0, block_q4_0, allocate_tiles_q4_0, load_tiles_q4_0<false>, VDR_q4_0_q8_1, vec_dot_q4_0_q8_1_mul_mat> mul_mat_q<QK4_0, QR4_0, QI4_0, block_q4_0, allocate_tiles_q4_0, load_tiles_q4_0<false>, VDR_Q4_0_Q8_1_MMQ, vec_dot_q4_0_q8_1_mul_mat>
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
} else { } else {
mul_mat_q<QK4_0, QR4_0, QI4_0, block_q4_0, allocate_tiles_q4_0, load_tiles_q4_0<true>, VDR_q4_0_q8_1, vec_dot_q4_0_q8_1_mul_mat> mul_mat_q<QK4_0, QR4_0, QI4_0, block_q4_0, allocate_tiles_q4_0, load_tiles_q4_0<true>, VDR_Q4_0_Q8_1_MMQ, vec_dot_q4_0_q8_1_mul_mat>
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
} }
} }
@ -3735,10 +3891,10 @@ static void ggml_mul_mat_q4_1_q8_1_cuda(
const dim3 block_dims(WARP_SIZE, WARP_SIZE/4, 1); const dim3 block_dims(WARP_SIZE, WARP_SIZE/4, 1);
if (nrows_x % GGML_CUDA_MMQ_Y == 0) { if (nrows_x % GGML_CUDA_MMQ_Y == 0) {
mul_mat_q<QK4_1, QR4_1, QI4_1, block_q4_1, allocate_tiles_q4_1, load_tiles_q4_1<false>, VDR_q4_1_q8_1, vec_dot_q4_1_q8_1_mul_mat> mul_mat_q<QK4_1, QR4_1, QI4_1, block_q4_1, allocate_tiles_q4_1, load_tiles_q4_1<false>, VDR_Q4_1_Q8_1_MMQ, vec_dot_q4_1_q8_1_mul_mat>
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
} else { } else {
mul_mat_q<QK4_1, QR4_1, QI4_1, block_q4_1, allocate_tiles_q4_1, load_tiles_q4_1<true>, VDR_q4_1_q8_1, vec_dot_q4_1_q8_1_mul_mat> mul_mat_q<QK4_1, QR4_1, QI4_1, block_q4_1, allocate_tiles_q4_1, load_tiles_q4_1<true>, VDR_Q4_1_Q8_1_MMQ, vec_dot_q4_1_q8_1_mul_mat>
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
} }
} }
@ -3753,10 +3909,10 @@ static void ggml_mul_mat_q5_0_q8_1_cuda(
const dim3 block_dims(WARP_SIZE, WARP_SIZE/4, 1); const dim3 block_dims(WARP_SIZE, WARP_SIZE/4, 1);
if (nrows_x % GGML_CUDA_MMQ_Y == 0) { if (nrows_x % GGML_CUDA_MMQ_Y == 0) {
mul_mat_q<QK5_0, QR5_0, QI5_0, block_q5_0, allocate_tiles_q5_0, load_tiles_q5_0<false>, VDR_q5_0_q8_1, vec_dot_q5_0_q8_1_mul_mat> mul_mat_q<QK5_0, QR5_0, QI5_0, block_q5_0, allocate_tiles_q5_0, load_tiles_q5_0<false>, VDR_Q5_0_Q8_1_MMQ, vec_dot_q5_0_q8_1_mul_mat>
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
} else { } else {
mul_mat_q<QK5_0, QR5_0, QI5_0, block_q5_0, allocate_tiles_q5_0, load_tiles_q5_0<true>, VDR_q5_0_q8_1, vec_dot_q5_0_q8_1_mul_mat> mul_mat_q<QK5_0, QR5_0, QI5_0, block_q5_0, allocate_tiles_q5_0, load_tiles_q5_0<true>, VDR_Q5_0_Q8_1_MMQ, vec_dot_q5_0_q8_1_mul_mat>
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
} }
} }
@ -3771,10 +3927,10 @@ static void ggml_mul_mat_q5_1_q8_1_cuda(
const dim3 block_dims(WARP_SIZE, WARP_SIZE/4, 1); const dim3 block_dims(WARP_SIZE, WARP_SIZE/4, 1);
if (nrows_x % GGML_CUDA_MMQ_Y == 0) { if (nrows_x % GGML_CUDA_MMQ_Y == 0) {
mul_mat_q<QK5_1, QR5_1, QI5_1, block_q5_1, allocate_tiles_q5_1, load_tiles_q5_1<false>, VDR_q5_1_q8_1, vec_dot_q5_1_q8_1_mul_mat> mul_mat_q<QK5_1, QR5_1, QI5_1, block_q5_1, allocate_tiles_q5_1, load_tiles_q5_1<false>, VDR_Q5_1_Q8_1_MMQ, vec_dot_q5_1_q8_1_mul_mat>
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
} else { } else {
mul_mat_q<QK5_1, QR5_1, QI5_1, block_q5_1, allocate_tiles_q5_1, load_tiles_q5_1<true>, VDR_q5_1_q8_1, vec_dot_q5_1_q8_1_mul_mat> mul_mat_q<QK5_1, QR5_1, QI5_1, block_q5_1, allocate_tiles_q5_1, load_tiles_q5_1<true>, VDR_Q5_1_Q8_1_MMQ, vec_dot_q5_1_q8_1_mul_mat>
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
} }
} }
@ -3789,10 +3945,10 @@ static void ggml_mul_mat_q8_0_q8_1_cuda(
const dim3 block_dims(WARP_SIZE, WARP_SIZE/4, 1); const dim3 block_dims(WARP_SIZE, WARP_SIZE/4, 1);
if (nrows_x % GGML_CUDA_MMQ_Y == 0) { if (nrows_x % GGML_CUDA_MMQ_Y == 0) {
mul_mat_q<QK8_0, QR8_0, QI8_0, block_q8_0, allocate_tiles_q8_0, load_tiles_q8_0<false>, VDR_q8_0_q8_1, vec_dot_q8_0_q8_1_mul_mat> mul_mat_q<QK8_0, QR8_0, QI8_0, block_q8_0, allocate_tiles_q8_0, load_tiles_q8_0<false>, VDR_Q8_0_Q8_1_MMQ, vec_dot_q8_0_q8_1_mul_mat>
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
} else { } else {
mul_mat_q<QK8_0, QR8_0, QI8_0, block_q8_0, allocate_tiles_q8_0, load_tiles_q8_0<true>, VDR_q8_0_q8_1, vec_dot_q8_0_q8_1_mul_mat> mul_mat_q<QK8_0, QR8_0, QI8_0, block_q8_0, allocate_tiles_q8_0, load_tiles_q8_0<true>, VDR_Q8_0_Q8_1_MMQ, vec_dot_q8_0_q8_1_mul_mat>
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
} }
} }