llama_cpp 0.0.7 → 0.1.0

data/ext/llama_cpp/src/ggml-cuda.h

@@ -1,43 +1,18 @@
-#include <cublas_v2.h>
-#include <cuda_runtime.h>
+#include "ggml.h"
 
 #ifdef  __cplusplus
 extern "C" {
 #endif
 
-#define CUDA_CHECK(err)                                                                 \
-    do {                                                                                \
-        cudaError_t err_ = (err);                                                       \
-        if (err_ != cudaSuccess) {                                                      \
-            fprintf(stderr, "CUDA error %d at %s:%d: %s\n", err_, __FILE__, __LINE__,   \
-                cudaGetErrorString(err_));                                              \
-            exit(1);                                                                    \
-        }                                                                               \
-    } while (0)
-
-#define CUBLAS_CHECK(err)                                                               \
-    do {                                                                                \
-        cublasStatus_t err_ = (err);                                                    \
-        if (err_ != CUBLAS_STATUS_SUCCESS) {                                            \
-            fprintf(stderr, "cuBLAS error %d at %s:%d\n", err_, __FILE__, __LINE__);    \
-            exit(1);                                                                    \
-        }                                                                               \
-    } while (0)
-
-extern cublasHandle_t g_cublasH;
-extern cudaStream_t   g_cudaStream;
-
 void   ggml_init_cublas(void);
-void * ggml_cuda_pool_malloc(size_t size, size_t * actual_size);
-void   ggml_cuda_pool_free(void * ptr, size_t size);
 
-void dequantize_row_q4_0_cuda(const void * vx, float * y, int k, cudaStream_t stream);
-void dequantize_row_q4_1_cuda(const void * vx, float * y, int k, cudaStream_t stream);
-void dequantize_row_q4_2_cuda(const void * vx, float * y, int k, cudaStream_t stream);
-void dequantize_row_q4_3_cuda(const void * vx, float * y, int k, cudaStream_t stream);
-void dequantize_row_q5_0_cuda(const void * vx, float * y, int k, cudaStream_t stream);
-void dequantize_row_q5_1_cuda(const void * vx, float * y, int k, cudaStream_t stream);
-void dequantize_row_q8_0_cuda(const void * vx, float * y, int k, cudaStream_t stream);
+bool   ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+size_t ggml_cuda_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+void   ggml_cuda_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
+
+// TODO: export these with GGML_API
+void * ggml_cuda_host_malloc(size_t size);
+void   ggml_cuda_host_free(void * ptr);
 
 #ifdef  __cplusplus
 }

data/ext/llama_cpp/src/ggml-opencl.c

@@ -3,12 +3,141 @@
 #define CL_TARGET_OPENCL_VERSION 110
 #include <clblast_c.h>
 
+#include <stdlib.h>
 #include <stdio.h>
 #include <string.h>
 
 #include "ggml.h"
 
-#include "ggml-opencl-dequant.cl"
+#define MULTILINE_QUOTE(...) #__VA_ARGS__
+const char * clblast_dequant = MULTILINE_QUOTE(
+
+struct block_q4_0
+{
+    float d;
+    uchar qs[16];
+};
+
+__kernel void dequantize_row_q4_0(__global struct block_q4_0* blocks, __global float* result) {
+    const uint i = get_global_id(0) / 32;
+    const uint l = get_local_id(0);
+
+    const float d = blocks[i].d;
+
+    const uchar vi = blocks[i].qs[l];
+
+    const uint index = i*32 + l*2;
+    result[index + 0] = ((vi & 0xf) - 8)*d;
+    result[index + 1] = ((vi >> 4) - 8)*d;
+}
+
+struct block_q4_1
+{
+    float d;
+    float m;
+    uchar qs[16];
+};
+
+__kernel void dequantize_row_q4_1(__global struct block_q4_1* blocks, __global float* result) {
+    const uint i = get_global_id(0) / 32;
+    const uint l = get_local_id(0);
+
+    const float d = blocks[i].d;
+    const float m = blocks[i].m;
+
+    const uchar vi = blocks[i].qs[l];
+
+    const uint index = i*32 + l*2;
+    result[index + 0] = (vi & 0xf) * d + m;
+    result[index + 1] = (vi >> 4) * d + m;
+}
+
+struct block_q4_2
+{
+    ushort d;
+    uchar qs[8];
+};
+
+__kernel void dequantize_row_q4_2(__global struct block_q4_2* blocks, __global float* result) {
+    const uint i = get_global_id(0) / 16;
+    const uint l = get_local_id(0);
+
+    const float d = vload_half(0, (__global half*) &blocks[i].d);
+
+    const uchar vi = blocks[i].qs[l];
+
+    const uint index = i*16 + l*2;
+    result[index + 0] = ((vi & 0xf) - 8)*d;
+    result[index + 1] = ((vi >> 4) - 8)*d;
+}
+
+
+struct block_q5_0
+{
+    float d;
+    uint qh;
+    uchar qs[16];
+};
+
+__kernel void dequantize_row_q5_0(__global struct block_q5_0* blocks, __global float* result) {
+    const uint i = get_global_id(0) / 32;
+    const uint l = get_local_id(0);
+
+    const float d = blocks[i].d;
+
+    const uchar vi = blocks[i].qs[l];
+
+    const uint l2 = l * 2;
+
+    const uchar vh0 = ((blocks[i].qh & (1 << (l2 + 0))) >> (l2 + 0)) << 4;
+    const uchar vh1 = ((blocks[i].qh & (1 << (l2 + 1))) >> (l2 + 1)) << 4;
+
+    const uint index = i*32 + l2;
+    result[index + 0] = (((vi & 0xf) | vh0) - 16)*d;
+    result[index + 1] = (((vi >>  4) | vh1) - 16)*d;
+}
+
+struct block_q5_1
+{
+    ushort d;
+    ushort m;
+    uint qh;
+    uchar qs[16];
+};
+
+__kernel void dequantize_row_q5_1(__global struct block_q5_1* blocks, __global float* result) {
+    const uint i = get_global_id(0) / 32;
+    const uint l = get_local_id(0);
+
+    const float d = vload_half(0, (__global half*) &blocks[i].d);
+    const float m = vload_half(0, (__global half*) &blocks[i].m);
+
+    const uchar vi = blocks[i].qs[l];
+
+    const uint l2 = l * 2;
+
+    const uchar vh0 = ((blocks[i].qh & (1 << (l2 + 0))) >> (l2 + 0)) << 4;
+    const uchar vh1 = ((blocks[i].qh & (1 << (l2 + 1))) >> (l2 + 1)) << 4;
+
+    const uint index = i*32 + l2;
+    result[index + 0] = ((vi & 0xf) | vh0)*d + m;
+    result[index + 1] = ((vi >>  4) | vh1)*d + m;
+}
+
+struct block_q8_0
+{
+    float d;
+    char qs[32];
+};
+
+__kernel void dequantize_row_q8_0(__global struct block_q8_0* blocks, __global float* result) {
+    const uint i = get_global_id(0) / 32;
+    const uint l = get_local_id(0);
+
+    result[i*32 + l] = blocks[i].qs[l] * blocks[i].d;
+}
+
+);
 
 #define CL_CHECK(err, name)                                                                     \
     do {                                                                                        \
@@ -19,12 +148,26 @@
         }                                                                                       \
     } while (0)
 
+#define QK5_0 32
+typedef struct {
+    ggml_fp16_t d;         // delta
+    uint8_t qh[4];         // 5-th bit of quants
+    uint8_t qs[QK5_0 / 2]; // nibbles / quants
+} block_q5_0;
+
+
+typedef struct {
+    float d;                // delta
+    uint32_t qh;          // 5-th bit of quants
+    uint8_t qs[QK5_0 / 2];  // nibbles / quants
+} cl_block_q5_0;
+
 static cl_platform_id platform;
 static cl_device_id device;
 static cl_context context;
 static cl_command_queue queue;
 static cl_program program;
-static cl_kernel kernel_q4_0, kernel_q4_1, kernel_q4_2, kernel_q4_3;
+static cl_kernel kernel_q4_0, kernel_q4_1, kernel_q4_2, kernel_q5_0, kernel_q5_1, kernel_q8_0;
 static cl_mem cl_buffer_a, cl_buffer_qb, cl_buffer_b, cl_buffer_c;
 static size_t cl_size_a = 0, cl_size_qb = 0, cl_size_b = 0, cl_size_c = 0;
 
@@ -97,7 +240,11 @@ void ggml_cl_init(void) {
     CL_CHECK(err, "clCreateKernel");
     kernel_q4_2 = clCreateKernel(program, "dequantize_row_q4_2", &err);
     CL_CHECK(err, "clCreateKernel");
-    kernel_q4_3 = clCreateKernel(program, "dequantize_row_q4_3", &err);
+    kernel_q5_0 = clCreateKernel(program, "dequantize_row_q5_0", &err);
+    CL_CHECK(err, "clCreateKernel");
+    kernel_q5_1 = clCreateKernel(program, "dequantize_row_q5_1", &err);
+    CL_CHECK(err, "clCreateKernel");
+    kernel_q8_0 = clCreateKernel(program, "dequantize_row_q8_0", &err);
     CL_CHECK(err, "clCreateKernel");
 }
 
@@ -127,6 +274,7 @@ void ggml_cl_sgemm_wrapper(
     cl_kernel kernel;
     size_t global = n * k, local, size_qb;
     bool dequant;
+    cl_block_q5_0* cl_host_b;
 
     switch (btype) {
     case GGML_TYPE_F32:
@@ -148,13 +296,36 @@ void ggml_cl_sgemm_wrapper(
         dequant = true;
         kernel = kernel_q4_2;
         local = 8;
-        size_qb = global * (sizeof(short) + local) / 16;
+        size_qb = global * (sizeof(ggml_fp16_t) + local) / 16;
         break;
-    case GGML_TYPE_Q4_3:
+    case GGML_TYPE_Q5_0:
         dequant = true;
-        kernel = kernel_q4_3;
-        local = 8;
-        size_qb = global * (sizeof(short) * 2 + local) / 16;
+        kernel = kernel_q5_0;
+        local = 16;
+        // For some reason OpenCL seems to be incapable of working with structs of size 22.
+        // 20 and 24 bytes are fine. Workaround to do the fp16 to fp32 step on CPU...
+        // TODO Find the reason, fix and remove workaround.
+        const block_q5_0* b = (const block_q5_0*) host_b;
+        cl_host_b = (cl_block_q5_0*) malloc(sizeof(cl_block_q5_0) * global / 32);
+        for (size_t i = 0; i < global / 32; i++) {
+            cl_host_b[i].d = ggml_fp16_to_fp32(b[i].d);
+            memcpy(&cl_host_b[i].qh, b[i].qh, sizeof(uint32_t));
+            memcpy(&cl_host_b[i].qs, b[i].qs, QK5_0 / 2);
+        }
+        host_b = (const float*) cl_host_b;
+        size_qb = global * (sizeof(float) + sizeof(uint32_t) + local) / 32;
+        break;
+    case GGML_TYPE_Q5_1:
+        dequant = true;
+        kernel = kernel_q5_1;
+        local = 16;
+        size_qb = global * (sizeof(ggml_fp16_t) * 2 + sizeof(uint32_t) + local) / 32;
+        break;
+    case GGML_TYPE_Q8_0:
+        dequant = true;
+        kernel = kernel_q8_0;
+        local = 32;
+        size_qb = global * (sizeof(float) + local) / 32;
         break;
     default:
         fprintf(stderr, "Error: Unsupported OpenCL btype %d\n", btype);
@@ -179,12 +350,15 @@ void ggml_cl_sgemm_wrapper(
         err = clSetKernelArg(kernel, 0, sizeof(cl_mem), &cl_buffer_qb);
         err |= clSetKernelArg(kernel, 1, sizeof(cl_mem), &cl_buffer_b);
         CL_CHECK(err, "clSetKernelArg");
-        clEnqueueWriteBuffer(queue, cl_buffer_qb, CL_FALSE, 0, size_qb, host_b, 0, NULL, &ev_qb);
+        err = clEnqueueWriteBuffer(queue, cl_buffer_qb, CL_FALSE, 0, size_qb, host_b, 0, NULL, &ev_qb);
+        CL_CHECK(err, "clEnqueueWriteBuffer qb");
     } else {
-        clEnqueueWriteBuffer(queue, cl_buffer_b, CL_FALSE, 0, size_b, host_b, 0, NULL, &ev_b);
+        err = clEnqueueWriteBuffer(queue, cl_buffer_b, CL_FALSE, 0, size_b, host_b, 0, NULL, &ev_b);
+        CL_CHECK(err, "clEnqueueWriteBuffer b");
     }
 
-    clEnqueueWriteBuffer(queue, cl_buffer_a, CL_FALSE, 0, size_a, host_a, 0, NULL, &ev_a);
+    err = clEnqueueWriteBuffer(queue, cl_buffer_a, CL_FALSE, 0, size_a, host_a, 0, NULL, &ev_a);
+    CL_CHECK(err, "clEnqueueWriteBuffer a");
     if (dequant) {
         err = clEnqueueNDRangeKernel(queue, kernel, 1, NULL, &global, &local, 1, &ev_qb, &ev_b);
         CL_CHECK(err, "clEnqueueNDRangeKernel");
@@ -196,15 +370,20 @@ void ggml_cl_sgemm_wrapper(
     clReleaseEvent(ev_b);
 
     cl_event ev_sgemm;
-    CLBlastSgemm((CLBlastLayout)order,
-                 (CLBlastTranspose)trans_a, (CLBlastTranspose)trans_b,
-                 m, n, k,
-                 alpha,
-                 cl_buffer_a, 0, lda,
-                 cl_buffer_b, 0, ldb,
-                 beta,
-                 cl_buffer_c, 0, ldc,
-                 &queue, &ev_sgemm);
+    CLBlastStatusCode status = CLBlastSgemm((CLBlastLayout)order,
+                                            (CLBlastTranspose)trans_a, (CLBlastTranspose)trans_b,
+                                            m, n, k,
+                                            alpha,
+                                            cl_buffer_a, 0, lda,
+                                            cl_buffer_b, 0, ldb,
+                                            beta,
+                                            cl_buffer_c, 0, ldc,
+                                            &queue, &ev_sgemm);
+
+    if (status != CLBlastSuccess) {
+        fprintf(stderr, "Error: CLBlast SGEMM %d\n", status);
+        abort();
+    }
 
     cl_event ev_c;
     clEnqueueReadBuffer(queue, cl_buffer_c, CL_TRUE, 0, size_c, host_c, 1, &ev_sgemm, &ev_c);
@@ -213,4 +392,7 @@ void ggml_cl_sgemm_wrapper(
     clWaitForEvents(1, &ev_c);
     clReleaseEvent(ev_sgemm);
     clReleaseEvent(ev_c);
+    if (btype == GGML_TYPE_Q5_0) {
+        free((void*) cl_host_b);
+    }
 }