integrated new version of clblast kernels as a separate file

Makefile

@@ -218,6 +218,8 @@ ggml_clblast.o: ggml.c ggml.h
 	$(CC)  $(CFLAGS) $(BONUSCFLAGS1) $(BONUSCFLAGS2) $(CLBLAST_FLAGS) -c $< -o $@
 ggml-opencl.o: ggml-opencl.c ggml-opencl.h
 	$(CC) $(CFLAGS) -c $< -o $@
+ggml-opencl-legacy.o: ggml-opencl-legacy.c ggml-opencl-legacy.h
+	$(CC) $(CFLAGS) -c $< -o $@
 
 #extreme old version compat
 ggml_v1.o: otherarch/ggml_v1.c otherarch/ggml_v1.h
@@ -262,7 +264,7 @@ koboldcpp_noavx2: ggml_noavx2.o ggml_v1_noavx2.o expose.o common.o gpttype_adapt
 koboldcpp_openblas_noavx2: ggml_openblas_noavx2.o ggml_v1_noavx2.o expose.o common.o gpttype_adapter.o 
 	$(OPENBLAS_NOAVX2_BUILD)
 
-koboldcpp_clblast: ggml_clblast.o ggml_v1.o expose.o common.o gpttype_adapter.o ggml-opencl.o 
+koboldcpp_clblast: ggml_clblast.o ggml_v1.o expose.o common.o gpttype_adapter.o ggml-opencl.o ggml-opencl-legacy.o
 	$(CLBLAST_BUILD)
 		
 quantize_llama: examples/quantize/quantize.cpp ggml.o llama.o

ggml-opencl-legacy.c

@@ -0,0 +1,426 @@
+#include "ggml-opencl-legacy.h"
+
+#define CL_TARGET_OPENCL_VERSION 110
+#include <clblast_c.h>
+
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+
+#include "ggml.h"
+
+#define MULTILINE_QUOTE(...) #__VA_ARGS__
+const char * clblast_dequant_legacy = 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_q4_3
+{
+    ushort d;
+    ushort m;
+    uchar qs[8];
+};
+
+__kernel void dequantize_row_q4_3(__global struct block_q4_3* 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 float m = vload_half(0, (__global half*) &(blocks[i].m));
+
+    const uchar vi = blocks[i].qs[l];
+
+    const uint index = i*16 + l*2;
+    result[index + 0] = (vi & 0xf) * d + m;
+    result[index + 1] = (vi >> 4) * d + m;
+}
+
+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 {                                                                                        \
+        cl_int err_ = (err);                                                                    \
+        if (err_ != CL_SUCCESS) {                                                               \
+            fprintf(stderr, "OpenCL %s error %d at %s:%d\n", name, err_, __FILE__, __LINE__);   \
+            exit(1);                                                                            \
+        }                                                                                       \
+    } 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, 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;
+
+static cl_program build_program_from_source(cl_context ctx, cl_device_id dev, const char* program_buffer) {
+    cl_program p;
+    char *program_log;
+    size_t program_size, log_size;
+    int err;
+
+    program_size = strlen(program_buffer);
+
+    p = clCreateProgramWithSource(ctx, 1, (const char**)&program_buffer, &program_size, &err);
+    if(err < 0) {
+        fprintf(stderr, "OpenCL error creating program");
+        exit(1);
+    }
+
+    err = clBuildProgram(p, 0, NULL, NULL, NULL, NULL);
+    if(err < 0) {
+
+        clGetProgramBuildInfo(p, dev, CL_PROGRAM_BUILD_LOG, 0, NULL, &log_size);
+        program_log = (char*) malloc(log_size + 1);
+        program_log[log_size] = '\0';
+        clGetProgramBuildInfo(p, dev, CL_PROGRAM_BUILD_LOG, log_size + 1, program_log, NULL);
+        printf("%s\n", program_log);
+        free(program_log);
+        exit(1);
+    }
+
+    return p;
+}
+
+void ggml_cl_init_legacy(void) {
+    cl_int err = 0;
+    char * GGML_CLBLAST_PLATFORM = getenv("GGML_CLBLAST_PLATFORM");
+    char * GGML_CLBLAST_DEVICE = getenv("GGML_CLBLAST_DEVICE");
+    int plat_num = (GGML_CLBLAST_PLATFORM == NULL ? 0 : atoi(GGML_CLBLAST_PLATFORM));
+    int dev_num = (GGML_CLBLAST_DEVICE == NULL ? 0 : atoi(GGML_CLBLAST_DEVICE));
+    printf("\nInitializing LEGACY CLBlast (First Run)...");
+    printf("\nAttempting to use: Platform=%d, Device=%d (If invalid, program will crash)\n",plat_num,dev_num);
+    cl_uint num_platforms;
+    clGetPlatformIDs(0, NULL, &num_platforms);
+    cl_platform_id* platforms = (cl_platform_id*)malloc(num_platforms*sizeof(cl_platform_id));
+    clGetPlatformIDs(num_platforms, platforms, NULL);
+    platform = platforms[plat_num];
+    char platform_buffer[1024];
+    clGetPlatformInfo(platform, CL_PLATFORM_NAME, sizeof(platform_buffer), &platform_buffer, NULL);
+    cl_uint num_devices;
+    clGetDeviceIDs(platform, CL_DEVICE_TYPE_ALL, 0, NULL, &num_devices);
+    cl_device_id* devices = (cl_device_id*)malloc(num_devices*sizeof(cl_device_id));
+    clGetDeviceIDs(platform, CL_DEVICE_TYPE_ALL, num_devices, devices, NULL);
+    device = devices[dev_num];
+    char device_buffer[1024];
+    clGetDeviceInfo(device, CL_DEVICE_NAME, sizeof(device_buffer), &device_buffer, NULL);
+    printf("Using Platform: %s Device: %s\n", platform_buffer, device_buffer);
+    context = clCreateContext(NULL, 1, &device, NULL, NULL, &err);
+    CL_CHECK(err, "clCreateContext");
+    queue = clCreateCommandQueue(context, device, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, &err);
+    CL_CHECK(err, "clCreateCommandQueue");
+
+    free(platforms);
+    free(devices);
+
+    program = build_program_from_source(context, device, clblast_dequant_legacy);
+
+    // Prepare dequantize kernels
+    kernel_q4_0 = clCreateKernel(program, "dequantize_row_q4_0", &err);
+    CL_CHECK(err, "clCreateKernel");
+    kernel_q4_1 = clCreateKernel(program, "dequantize_row_q4_1", &err);
+    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);
+    CL_CHECK(err, "clCreateKernel");
+    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");
+}
+
+static void ggml_cl_malloc(size_t req_size, size_t* cur_size, cl_mem_flags flags, cl_mem* buf) {
+    if (req_size <= *cur_size) {
+        return;
+    }
+
+    // Reallocate buffer with enough space
+    if (*cur_size > 0) {
+        clReleaseMemObject(*buf);
+    }
+    cl_int err;
+    *buf = clCreateBuffer(context, flags, req_size, NULL, &err);
+    *cur_size = req_size;
+    CL_CHECK(err, "clCreateBuffer");
+}
+
+void ggml_cl_sgemm_wrapper_legacy(
+        const enum ggml_blas_order order, const enum ggml_blas_op trans_a, const enum ggml_blas_op trans_b,
+        const int m, const int n, const int k,
+        const float alpha, const void *host_a, const int lda,
+        const float *host_b, const int ldb, const float beta,
+        float *host_c, const int ldc, const int btype) {
+    cl_int err = 0;
+
+    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:
+        dequant = false;
+        break;
+    case GGML_TYPE_Q4_0:
+        dequant = true;
+        kernel = kernel_q4_0;
+        local = 16;
+        size_qb = global * (sizeof(float) + local) / 32;
+        break;
+    case GGML_TYPE_Q4_1:
+        dequant = true;
+        kernel = kernel_q4_1;
+        local = 16;
+        size_qb = global * (sizeof(float) * 2 + local) / 32;
+        break;
+    case GGML_TYPE_Q4_2:
+        dequant = true;
+        kernel = kernel_q4_2;
+        local = 8;
+        size_qb = global * (sizeof(ggml_fp16_t) + local) / 16;
+        break;
+    case GGML_TYPE_Q4_3:
+        dequant = true;
+        kernel = kernel_q4_3;
+        local = 8;
+        size_qb = global * (sizeof(short) * 2 + local) / 16;
+        break;
+    case GGML_TYPE_Q5_0:
+        dequant = true;
+        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);
+        abort();
+    }
+
+    const size_t size_a =  m * k * sizeof(float);
+    const size_t size_b =  n * k * sizeof(float);
+    const size_t size_c =  m * n * sizeof(float);
+
+    // Prepare buffers
+    ggml_cl_malloc(size_a, &cl_size_a, CL_MEM_READ_ONLY, &cl_buffer_a);
+    if (dequant) {
+        ggml_cl_malloc(size_qb, &cl_size_qb, CL_MEM_READ_ONLY, &cl_buffer_qb);
+    }
+    ggml_cl_malloc(size_b, &cl_size_b, CL_MEM_READ_WRITE, &cl_buffer_b);
+    ggml_cl_malloc(size_c, &cl_size_c, CL_MEM_WRITE_ONLY, &cl_buffer_c);
+
+    cl_event ev_a, ev_qb, ev_b;
+
+    if (dequant) {
+        err = clSetKernelArg(kernel, 0, sizeof(cl_mem), &cl_buffer_qb);
+        err |= clSetKernelArg(kernel, 1, sizeof(cl_mem), &cl_buffer_b);
+        CL_CHECK(err, "clSetKernelArg");
+        err = clEnqueueWriteBuffer(queue, cl_buffer_qb, CL_FALSE, 0, size_qb, host_b, 0, NULL, &ev_qb);
+        CL_CHECK(err, "clEnqueueWriteBuffer qb");
+    } else {
+        err = clEnqueueWriteBuffer(queue, cl_buffer_b, CL_FALSE, 0, size_b, host_b, 0, NULL, &ev_b);
+        CL_CHECK(err, "clEnqueueWriteBuffer b");
+    }
+
+    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");
+        clReleaseEvent(ev_qb);
+    }
+    clWaitForEvents(1, &ev_a);
+    clWaitForEvents(1, &ev_b);
+    clReleaseEvent(ev_a);
+    clReleaseEvent(ev_b);
+
+    cl_event 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);
+
+    // Wait for completion
+    clWaitForEvents(1, &ev_c);
+    clReleaseEvent(ev_sgemm);
+    clReleaseEvent(ev_c);
+    if (btype == GGML_TYPE_Q5_0) {
+        free((void*) cl_host_b);
+    }
+}

ggml-opencl-legacy.h

@@ -0,0 +1,15 @@
+#pragma once
+
+#include "ggml-opencl.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+void ggml_cl_init_legacy(void);
+
+void ggml_cl_sgemm_wrapper_legacy(const enum ggml_blas_order order, const enum ggml_blas_op trans_a, const enum ggml_blas_op trans_b, const int m, const int n, const int k, const float alpha, const void *host_a, const int lda, const float *host_b, const int ldb, const float beta, float *host_c, const int ldc, const int btype);
+
+#ifdef  __cplusplus
+}
+#endif

ggml-opencl.c

@@ -7,155 +7,134 @@
 #include <stdio.h>
 #include <string.h>
 
-
 #include "ggml.h"
 
 #define MULTILINE_QUOTE(...) #__VA_ARGS__
 const char * clblast_dequant = MULTILINE_QUOTE(
 
+typedef uchar uint8_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+constant uint QK4_0 = 32;
 struct block_q4_0
 {
     float d;
-    uchar qs[16];
+    uint8_t qs[QK4_0 / 2];
 };
 
-__kernel void dequantize_row_q4_0(__global struct block_q4_0* blocks, __global float* result) {
+constant uint QK4_1 = 32;
-    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];
+    uint8_t qs[QK4_1 / 2];
 };
 
-__kernel void dequantize_row_q4_1(__global struct block_q4_1* blocks, __global float* result) {
+constant uint QK5_0 = 32;
-    const uint i = get_global_id(0) / 32;
+struct __attribute__ ((packed)) block_q5_0
-    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;
+    half d;
-    uchar qs[8];
+    uint32_t qh;
+    uint8_t qs[QK5_0 / 2];
 };
 
-__kernel void dequantize_row_q4_2(__global struct block_q4_2* blocks, __global float* result) {
+constant uint QK5_1 = 32;
-    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_q4_3
-{
-    ushort d;
-    ushort m;
-    uchar qs[8];
-};
-
-__kernel void dequantize_row_q4_3(__global struct block_q4_3* 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 float m = vload_half(0, (__global half*) &(blocks[i].m));
-
-    const uchar vi = blocks[i].qs[l];
-
-    const uint index = i*16 + l*2;
-    result[index + 0] = (vi & 0xf) * d + m;
-    result[index + 1] = (vi >> 4) * d + m;
-}
-
-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;
+    half d;
-    ushort m;
+    half m;
-    uint qh;
+    uint32_t qh;
-    uchar qs[16];
+    uint8_t qs[QK5_1 / 2];
 };
 
-__kernel void dequantize_row_q5_1(__global struct block_q5_1* blocks, __global float* result) {
+constant uint QK8_0 = 32;
-    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];
+    uint8_t qs[QK8_0];
 };
 
-__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;
+__kernel void dequantize_row_q4_0(__global struct block_q4_0* x, __global float* y) {
+    constant uint qk = QK4_0;
+
+    const uint i = get_global_id(0) / qk;
+    const uint j = get_local_id(0);
+
+    const float d = x[i].d;
+
+    const int x0 = (x[i].qs[j] & 0xf) - 8;
+    const int x1 = (x[i].qs[j] >>  4) - 8;
+
+    y[i*qk + j + 0   ] = x0*d;
+    y[i*qk + j + qk/2] = x1*d;
+}
+
+__kernel void dequantize_row_q4_1(__global struct block_q4_1* x, __global float* y) {
+    constant uint qk = QK4_1;
+
+    const uint i = get_global_id(0) / qk;
+    const uint j = get_local_id(0);
+
+    const float d = x[i].d;
+    const float m = x[i].m;
+
+    const int x0 = (x[i].qs[j] & 0xf);
+    const int x1 = (x[i].qs[j] >>  4);
+
+    y[i*qk + j + 0   ] = x0*d + m;
+    y[i*qk + j + qk/2] = x1*d + m;
+}
+
+__kernel void dequantize_row_q5_0(__global struct block_q5_0* x, __global float* y) {
+    constant uint qk = QK5_0;
+
+    const uint i = get_global_id(0) / qk;
+    const uint j = get_local_id(0);
+
+    const float d = vload_half(0, (__global half*) &x[i].d);
+
+    uint32_t qh = x[i].qh;
+
+    const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
+    const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
+
+    const int32_t x0 = ((x[i].qs[j] & 0xf) | xh_0) - 16;
+    const int32_t x1 = ((x[i].qs[j] >>  4) | xh_1) - 16;
+
+    y[i*qk + j + 0   ] = x0*d;
+    y[i*qk + j + qk/2] = x1*d;
+}
+
+__kernel void dequantize_row_q5_1(__global struct block_q5_1* x, __global float* y) {
+    constant uint qk = QK5_1;
+
+    const uint i = get_global_id(0) / qk;
+    const uint j = get_local_id(0);
+
+    const float d = vload_half(0, (__global half*) &x[i].d);
+    const float m = vload_half(0, (__global half*) &x[i].m);
+
+    uint32_t qh = x[i].qh;
+
+    const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
+    const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
+
+    const int x0 = (x[i].qs[j] & 0xf) | xh_0;
+    const int x1 = (x[i].qs[j] >>  4) | xh_1;
+
+    y[i*qk + j + 0   ] = x0*d + m;
+    y[i*qk + j + qk/2] = x1*d + m;
+}
+
+__kernel void dequantize_row_q8_0(__global struct block_q8_0* x, __global float* y) {
+    constant uint qk = QK8_0;
+    const uint i = get_global_id(0) / qk;
+    const uint j = get_local_id(0);
+
+    const float d = x[i].d;
+    y[i*qk + j] = x[i].qs[j]*d;
 }
 
 );
@@ -169,26 +148,12 @@ __kernel void dequantize_row_q8_0(__global struct block_q8_0* blocks, __global f
         }                                                                                       \
     } 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, kernel_q5_0, kernel_q5_1, kernel_q8_0;
+static cl_kernel kernel_q4_0, kernel_q4_1, 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;
 
@@ -259,10 +224,6 @@ void ggml_cl_init(void) {
     CL_CHECK(err, "clCreateKernel");
     kernel_q4_1 = clCreateKernel(program, "dequantize_row_q4_1", &err);
     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);
-    CL_CHECK(err, "clCreateKernel");
     kernel_q5_0 = clCreateKernel(program, "dequantize_row_q5_0", &err);
     CL_CHECK(err, "clCreateKernel");
     kernel_q5_1 = clCreateKernel(program, "dequantize_row_q5_1", &err);
@@ -297,7 +258,6 @@ 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:
@@ -315,34 +275,11 @@ void ggml_cl_sgemm_wrapper(
         local = 16;
         size_qb = global * (sizeof(float) * 2 + local) / 32;
         break;
-    case GGML_TYPE_Q4_2:
-        dequant = true;
-        kernel = kernel_q4_2;
-        local = 8;
-        size_qb = global * (sizeof(ggml_fp16_t) + local) / 16;
-        break;
-    case GGML_TYPE_Q4_3:
-        dequant = true;
-        kernel = kernel_q4_3;
-        local = 8;
-        size_qb = global * (sizeof(short) * 2 + local) / 16;
-        break;
     case GGML_TYPE_Q5_0:
         dequant = true;
         kernel = kernel_q5_0;
         local = 16;
-        // For some reason OpenCL seems to be incapable of working with structs of size 22.
+        size_qb = global * (sizeof(ggml_fp16_t) + sizeof(uint32_t) + local) / 32;
-        // 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;
@@ -421,7 +358,4 @@ 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);
-    }
-}

ggml.c

@@ -144,6 +144,7 @@ inline static void* ggml_aligned_malloc(size_t size) {
 #endif
 #if defined(GGML_USE_CLBLAST)
 #include "ggml-opencl.h"
+#include "ggml-opencl-legacy.h"
 #endif
 
 #undef MIN
@@ -3560,7 +3561,14 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
 #if defined(GGML_USE_CUBLAS)
         ggml_init_cublas();
 #elif defined(GGML_USE_CLBLAST)
-        ggml_cl_init();
+        if(quants_unshuffled)
+        {
+            ggml_cl_init();
+        }
+        else
+        {
+            ggml_cl_init_legacy();
+        }
 #endif
 
         is_first_call = false;
@@ -7474,12 +7482,24 @@ static void ggml_compute_forward_mul_mat_f32(
 
 #if defined(GGML_USE_CLBLAST)
                 // zT = y * xT
+                if(quants_unshuffled)
+                {
                 ggml_cl_sgemm_wrapper(GGML_BLAS_ORDER_ROW_MAJOR, GGML_BLAS_OP_N, GGML_BLAS_OP_T,
                         ne11, ne01, ne10,
                         1.0f,    y, ne10,
                                  x, ne10,
                         0.0f,    d, ne01,
                         GGML_TYPE_F32);
+                }
+                else
+                {
+                ggml_cl_sgemm_wrapper_legacy(GGML_BLAS_ORDER_ROW_MAJOR, GGML_BLAS_OP_N, GGML_BLAS_OP_T,
+                    ne11, ne01, ne10,
+                    1.0f,    y, ne10,
+                                x, ne10,
+                    0.0f,    d, ne01,
+                    GGML_TYPE_F32);
+                }
 #else
                 cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
                         ne11, ne01, ne10,
@@ -7664,12 +7684,24 @@ static void ggml_compute_forward_mul_mat_f16_f32(
                 float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
 
                 // zT = y * xT
+                if(quants_unshuffled)
+                {
                 ggml_cl_sgemm_wrapper(GGML_BLAS_ORDER_ROW_MAJOR, GGML_BLAS_OP_N, GGML_BLAS_OP_T,
                         ne11, ne01, ne10,
                         1.0f,    y, ne10,
                                  x, ne10,
                         0.0f,    d, ne01,
                         GGML_TYPE_F32);
+                }
+                else
+                {
+                ggml_cl_sgemm_wrapper_legacy(GGML_BLAS_ORDER_ROW_MAJOR, GGML_BLAS_OP_N, GGML_BLAS_OP_T,
+                        ne11, ne01, ne10,
+                        1.0f,    y, ne10,
+                                 x, ne10,
+                        0.0f,    d, ne01,
+                        GGML_TYPE_F32);
+                }
 #else
                 const float * x = wdata;
                 const float * y = (float *) ((char *) src1->data + i02*nb12 + i03*nb13);
@@ -7888,12 +7920,24 @@ static void ggml_compute_forward_mul_mat_q_f32(
 
 #if defined(GGML_USE_CLBLAST)
                 // zT = y * xT
+                if(quants_unshuffled)
+                {
                 ggml_cl_sgemm_wrapper(GGML_BLAS_ORDER_ROW_MAJOR, GGML_BLAS_OP_N, GGML_BLAS_OP_T,
                         ne11, ne01, ne10,
                         1.0f,    y, ne10,
                                  x, ne10,
                         0.0f,    d, ne01,
                         type);
+                }
+                else
+                {
+                ggml_cl_sgemm_wrapper_legacy(GGML_BLAS_ORDER_ROW_MAJOR, GGML_BLAS_OP_N, GGML_BLAS_OP_T,
+                        ne11, ne01, ne10,
+                        1.0f,    y, ne10,
+                                 x, ne10,
+                        0.0f,    d, ne01,
+                        type);
+                }
 #else
                 cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
                         ne11, ne01, ne10,