vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

fixup! More GPU threads for CUDA kernels

Browse source
This commit is contained in:
JohannesGaessler 2023-05-08 12:56:32 +02:00
parent 8d8de07a4e
commit d0199b3bc3
1 changed files with 145 additions and 101 deletions
Download patch file Download diff file Expand all files Collapse all files

102
ggml-cuda.cu
View file

@ -34,6 +34,8 @@ static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size");
typedef void (*to_fp32_cuda_t)(const void * x, float * y, int k, cudaStream_t stream);
#define GGML_CUDA_MAX_BLOCK_SIZE 256
#define QK4_0 32
typedef struct {
float d; // delta
@ -85,7 +87,10 @@ static __global__ void dequantize_block_q4_0(const void * vx, float * y, int k)
const int i = blockIdx.x*blockDim.x + threadIdx.x;
if (i < k) {
if (i >= k) {
return;
}
const float d = x[i].d;
const uint8_t * pp = x[i].qs;
@ -103,14 +108,16 @@ static __global__ void dequantize_block_q4_0(const void * vx, float * y, int k)
y[i*QK4_0 + l + 1] = v1;
}
}
}
static __global__ void dequantize_block_q4_1(const void * vx, float * y, int k) {
const block_q4_1 * x = (const block_q4_1 *) vx;
const int i = blockIdx.x*blockDim.x + threadIdx.x;
if (i < k) {
if (i >= k) {
return;
}
const float d = x[i].d;
const float m = x[i].m;
@ -129,14 +136,16 @@ static __global__ void dequantize_block_q4_1(const void * vx, float * y, int k)
y[i*QK4_1 + l + 1] = v1;
}
}
}
static __global__ void dequantize_block_q4_2(const void * vx, float * y, int k) {
const block_q4_2 * x = (const block_q4_2 *) vx;
const int i = blockIdx.x*blockDim.x + threadIdx.x;
if (i < k) {
if (i >= k) {
return;
}
const float d = x[i].d;
const uint8_t * pp = x[i].qs;
@ -154,14 +163,16 @@ static __global__ void dequantize_block_q4_2(const void * vx, float * y, int k)
y[i*QK4_2 + l + 1] = v1;
}
}
}
static __global__ void dequantize_block_q5_0(const void * vx, float * y, int k) {
const block_q5_0 * x = (const block_q5_0 *) vx;
const int i = blockIdx.x*blockDim.x + threadIdx.x;
if (i < k) {
if (i >= k) {
return;
}
const float d = x[i].d;
const uint8_t * pp = x[i].qs;
@ -185,14 +196,16 @@ static __global__ void dequantize_block_q5_0(const void * vx, float * y, int k)
y[i*QK5_0 + l + 1] = v1;
}
}
}
static __global__ void dequantize_block_q5_1(const void * vx, float * y, int k) {
const block_q5_1 * x = (const block_q5_1 *) vx;
const int i = blockIdx.x*blockDim.x + threadIdx.x;
if (i < k) {
if (i >= k) {
return;
}
const float d = x[i].d;
const float m = x[i].m;
@ -217,14 +230,16 @@ static __global__ void dequantize_block_q5_1(const void * vx, float * y, int k)
y[i*QK5_1 + l + 1] = v1;
}
}
}
static __global__ void dequantize_block_q8_0(const void * vx, float * y, int k) {
const block_q8_0 * x = (const block_q8_0 *) vx;
const int i = blockIdx.x*blockDim.x + threadIdx.x;
if (i < k) {
if (i >= k) {
return;
}
const float d = x[i].d;
const int8_t * pp = x[i].qs;
@ -235,53 +250,76 @@ static __global__ void dequantize_block_q8_0(const void * vx, float * y, int k)
y[i*QK8_0 + l] = vi*d;
}
}
}
static void dequantize_row_q4_0_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
const int nb = k / QK4_0;
int min_grid_size, block_size = 1; // Initialize to suppress compiler warning.
static int grid_size, block_size = -1;
if (block_size == -1) {
int min_grid_size;
CUDA_CHECK(cudaOccupancyMaxPotentialBlockSize(&min_grid_size, &block_size, dequantize_block_q4_0, 0, 0));
int grid_size = (nb + block_size - 1) / block_size; // Round up.
block_size = min(block_size, GGML_CUDA_MAX_BLOCK_SIZE);
grid_size = (nb + block_size - 1) / block_size; // Round up.
}
dequantize_block_q4_0<<<grid_size, block_size, 0, stream>>>(vx, y, nb);
}
static void dequantize_row_q4_1_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
const int nb = k / QK4_1;
int min_grid_size, block_size = 1; // Initialize to suppress compiler warning.
static int grid_size, block_size = -1;
if (block_size == -1) {
int min_grid_size;
CUDA_CHECK(cudaOccupancyMaxPotentialBlockSize(&min_grid_size, &block_size, dequantize_block_q4_1, 0, 0));
int grid_size = (nb + block_size - 1) / block_size; // Round up.
block_size = min(block_size, GGML_CUDA_MAX_BLOCK_SIZE);
grid_size = (nb + block_size - 1) / block_size; // Round up.
}
dequantize_block_q4_1<<<grid_size, block_size, 0, stream>>>(vx, y, nb);
}
static void dequantize_row_q4_2_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
const int nb = k / QK4_2;
int min_grid_size, block_size = 1; // Initialize to suppress compiler warning.
static int grid_size, block_size = -1;
if (block_size == -1) {
int min_grid_size;
CUDA_CHECK(cudaOccupancyMaxPotentialBlockSize(&min_grid_size, &block_size, dequantize_block_q4_2, 0, 0));
int grid_size = (nb + block_size - 1) / block_size; // Round up.
block_size = min(block_size, GGML_CUDA_MAX_BLOCK_SIZE);
grid_size = (nb + block_size - 1) / block_size; // Round up.
}
dequantize_block_q4_2<<<grid_size, block_size, 0, stream>>>(vx, y, nb);
}
static void dequantize_row_q5_0_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
const int nb = k / QK5_0;
int min_grid_size, block_size = 1; // Initialize to suppress compiler warning.
static int grid_size, block_size = -1;
if (block_size == -1) {
int min_grid_size;
CUDA_CHECK(cudaOccupancyMaxPotentialBlockSize(&min_grid_size, &block_size, dequantize_block_q5_0, 0, 0));
int grid_size = (nb + block_size - 1) / block_size; // Round up.
block_size = min(block_size, GGML_CUDA_MAX_BLOCK_SIZE);
grid_size = (nb + block_size - 1) / block_size; // Round up.
}
dequantize_block_q5_0<<<grid_size, block_size, 0, stream>>>(vx, y, nb);
}
static void dequantize_row_q5_1_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
const int nb = k / QK5_1;
int min_grid_size, block_size = 1; // Initialize to suppress compiler warning.
static int grid_size, block_size = -1;
if (block_size == -1) {
int min_grid_size;
CUDA_CHECK(cudaOccupancyMaxPotentialBlockSize(&min_grid_size, &block_size, dequantize_block_q5_1, 0, 0));
int grid_size = (nb + block_size - 1) / block_size; // Round up.
block_size = min(block_size, GGML_CUDA_MAX_BLOCK_SIZE);
grid_size = (nb + block_size - 1) / block_size; // Round up.
}
dequantize_block_q5_1<<<grid_size, block_size, 0, stream>>>(vx, y, nb);
}
static void dequantize_row_q8_0_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
const int nb = k / QK8_0;
int min_grid_size, block_size = 1; // Initialize to suppress compiler warning.
static int grid_size, block_size = -1;
if (block_size == -1) {
int min_grid_size;
CUDA_CHECK(cudaOccupancyMaxPotentialBlockSize(&min_grid_size, &block_size, dequantize_block_q8_0, 0, 0));
int grid_size = (nb + block_size - 1) / block_size; // Round up.
block_size = min(block_size, GGML_CUDA_MAX_BLOCK_SIZE);
grid_size = (nb + block_size - 1) / block_size; // Round up.
}
dequantize_block_q8_0<<<grid_size, block_size, 0, stream>>>(vx, y, nb);
}
@ -289,17 +327,23 @@ static void dequantize_row_q8_0_cuda(const void * vx, float * y, int k, cudaStre
static __global__ void convert_fp16_to_fp32(const void * vx, float * y, int k) {
const half * x = (const half *) vx;
const int i = blockIdx.x;
const int i = blockIdx.x*blockDim.x + threadIdx.x;
if (i < k) {
y[i] = __half2float(x[i]);
if (i >= k) {
return;
}
y[i] = __half2float(x[i]);
}
static void convert_fp16_to_fp32_cuda(const void * x, float * y, int k, cudaStream_t stream) {
int min_grid_size, block_size = 1; // Initialize to suppress compiler warning.
static int grid_size, block_size = -1;
if (block_size == -1) {
int min_grid_size;
CUDA_CHECK(cudaOccupancyMaxPotentialBlockSize(&min_grid_size, &block_size, convert_fp16_to_fp32, 0, 0));
int grid_size = (k + block_size - 1) / block_size; // Round up.
block_size = min(block_size, GGML_CUDA_MAX_BLOCK_SIZE);
grid_size = (k + block_size - 1) / block_size; // Round up.
}
convert_fp16_to_fp32<<<grid_size, block_size, 0, stream>>>(x, y, k);
}