local-llm-rn 1.0.0

package/cpp/ggml/src/ggml-opencl/kernels/mean.cl

@@ -0,0 +1,140 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+
+// Most devices have max workgroup size of 1024, so this is enough for subgroup
+// sizes of 16, 32, 64 and 128. Increase this value for smaller subgroups sizes
+#define MAX_SUBGROUPS 64
+kernel void kernel_mean_f32(
+    global char *  src0,
+    ulong           offset0,
+    global char *  dst,
+    ulong           offsetd,
+    int             ne00,
+    int             ne01,
+    int             ne02,
+    int             ne03,
+    ulong           nb01,
+    ulong           nb02,
+    ulong           nb03,
+    ulong           nb1,
+    ulong           nb2,
+    ulong           nb3
+) {
+    src0 = src0 + offset0;
+    dst  = dst  + offsetd;
+
+    const int i3 = get_group_id(2);
+    const int i2 = get_group_id(1);
+    const int i1 = get_group_id(0);
+
+    const int lid = get_local_id(0);
+    const int lsize = get_local_size(0);
+
+    const uint sg_size = get_sub_group_size();
+    const uint sg_id = get_sub_group_id();
+    const uint sg_lid = get_sub_group_local_id();
+
+    __local float lmem[MAX_SUBGROUPS];
+
+    if (i3 >= ne03 || i2 >= ne02 || i1 >= ne01) {
+        return;
+    }
+
+    if(sg_id == 0){
+        lmem[sg_lid] = 0.0f;
+    }
+
+    global float * src_row = (global float *) (src0 + i1*nb01 + i2*nb02 + i3*nb03);
+    global float * dst_row = (global float *) (dst  + i1*nb1  + i2*nb2  + i3*nb3);
+
+    float sumf = 0.0f;
+
+    for (int i0 = lid; i0 < ne00; i0 += lsize) {
+        sumf += src_row[i0];
+    }
+
+    sumf = sub_group_reduce_add(sumf);
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    if(sg_lid == 0){
+        lmem[sg_id] = sumf;
+    }
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    sumf = lmem[sg_lid];
+    sumf = sub_group_reduce_add(sumf);
+
+    if (lid == 0) {
+        dst_row[0] = sumf / ne00;
+    }
+}
+
+kernel void kernel_mean_f32_4(
+    global char *  src0,
+    ulong           offset0,
+    global char *  dst,
+    ulong           offsetd,
+    int             ne00,
+    int             ne01,
+    int             ne02,
+    int             ne03,
+    ulong           nb01,
+    ulong           nb02,
+    ulong           nb03,
+    ulong           nb1,
+    ulong           nb2,
+    ulong           nb3
+) {
+    src0 = src0 + offset0;
+    dst  = dst  + offsetd;
+
+    const int i3 = get_group_id(2);
+    const int i2 = get_group_id(1);
+    const int i1 = get_group_id(0);
+
+    const int lid = get_local_id(0);
+    const int lsize = get_local_size(0);
+
+    const uint sg_size = get_sub_group_size();
+    const uint sg_id = get_sub_group_id();
+    const uint sg_lid = get_sub_group_local_id();
+
+    __local float lmem[MAX_SUBGROUPS];
+
+    if (i3 >= ne03 || i2 >= ne02 || i1 >= ne01) {
+        return;
+    }
+
+    if(sg_id == 0){
+        lmem[sg_lid] = 0.0f;
+    }
+
+    global float4 * src_row = (global float4 *) (src0 + i1*nb01 + i2*nb02 + i3*nb03);
+    global float  * dst_row = (global float  *) (dst  + i1*nb1  + i2*nb2  + i3*nb3);
+
+    float4 sum_vec = (float4)0.0f;
+
+    for (int i0 = lid; i0 < ne00 / 4; i0 += lsize) {
+        sum_vec += src_row[i0];
+    }
+
+    float sumf = dot(sum_vec, (float4)(1.0f));
+    sumf = sub_group_reduce_add(sumf);
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    if(sg_lid == 0){
+        lmem[sg_id] = sumf;
+    }
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    sumf = lmem[sg_lid];
+    sumf = sub_group_reduce_add(sumf);
+
+    if (lid == 0) {
+        dst_row[0] = sumf / ne00;
+    }
+}

package/cpp/ggml/src/ggml-opencl/kernels/mul.cl

@@ -0,0 +1,152 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// mul
+//------------------------------------------------------------------------------
+kernel void kernel_mul(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        int ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int i13 = i03 % ne13;
+    int i12 = i02 % ne12;
+    int i11 = i01 % ne11;
+
+    global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
+    global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
+    global char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const int i10 = i0 % ne10;
+        *((global float *)(dst_ptr + i0*nb0)) = *((global float *)(src0_ptr + i0*nb00)) * *((global float *)(src1_ptr + i10*nb10));
+    }
+}
+
+// assumption: src1 is a row
+// broadcast src1 into src0
+kernel void kernel_mul_row(
+        global float4 * src0,
+        ulong offset0,
+        global float4 * src1,
+        ulong offset1,
+        global float4 * dst,
+        ulong offsetd,
+        int ne
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    src1 = (global float4*)((global char*)src1 + offset1);
+    dst = (global float4*)((global char*)dst + offsetd);
+
+    // This performs better than using %.
+    uint gid = get_global_id(0);
+    uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
+    dst[gid] = src0[gid] * src1[idx1];
+}
+
+kernel void kernel_mul_f16(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        int ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int i13 = i03 % ne13;
+    int i12 = i02 % ne12;
+    int i11 = i01 % ne11;
+
+    global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
+    global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
+    global char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const int i10 = i0 % ne10;
+        *((global half *)(dst_ptr + i0*nb0)) = *((global half *)(src0_ptr + i0*nb00)) * *((global half *)(src1_ptr + i10*nb10));
+    }
+}
+
+kernel void kernel_mul_row_f16(
+        global half4 * src0,
+        ulong offset0,
+        global half4 * src1,
+        ulong offset1,
+        global half4 * dst,
+        ulong offsetd,
+        int ne
+) {
+    src0 = (global half4*)((global char*)src0 + offset0);
+    src1 = (global half4*)((global char*)src1 + offset1);
+    dst = (global half4*)((global char*)dst + offsetd);
+
+    // This performs better than using %.
+    uint gid = get_global_id(0);
+    uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
+    dst[gid] = src0[gid] * src1[idx1];
+}

package/cpp/ggml/src/ggml-opencl/kernels/mul_mat_Ab_Bi_8x4.cl

@@ -0,0 +1,139 @@
+// src0_q, src0_d, src1 are transposed as a preprocessing step
+// 4-bit weights are transposed in groups of 4 (unsigned short int)
+// consider weights originally "next to each other", now "on top of each other"
+// each fiber computes a 8x4 tile of output elements
+// using unshuffled weights
+
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_128
+#endif
+
+kernel void kernel_mul_mat_Ab_Bi_8x4(
+        global const ushort * src0_q,       // quantized A
+        global const half  * src0_d,        // A scales
+        __read_only image1d_buffer_t src1,  // B (1d image)
+        global float * dst,                 // C
+        int m,                              // M
+        int n,                              // N with padding
+        int k,                              // K
+        int n_no_padding                    // N without padding
+) {
+
+    int m_4 = m >> 2;
+    int n_4 = n >> 2;
+
+    int gy = get_global_id(0);
+    int gx = get_global_id(1);
+    int gx_2 = gx << 2;
+
+    half8 c0 = 0, c1 = 0, c2 = 0, c3 = 0; // 8x4 output elements
+    half8 B; // registers for activations
+    half4 dequantized_weights; // registers for dequantized weights
+    __global const ushort* weight_ptr = src0_q + gx_2; // pointer for weights
+    __global const half* scale_ptr = src0_d + gx_2; // pointer for scales
+
+    for(int i=0; i<k; i+=4){ //loop through K dimension
+
+        B.s0123 = read_imageh(src1, gy*2 + (i)*(n_4));
+        B.s4567 = read_imageh(src1, gy*2 + (i)*(n_4)+1);
+
+        // keep (i/4) and (i/32) in parenthesis, rounds down
+        // load 4 consecutive groups of 4 weights
+        ushort4 bits4 = vload4(0, weight_ptr + (i/4)*(m)); // (i/4) because weights grouped in 4s
+
+        // load 4 consecutive scales
+        half4 scale = vload4(0, scale_ptr + (i/32)*(m));// (i/32) because 1 scale per 32 elements
+
+        // j=0
+        dequantized_weights.s0 = ((bits4.s0 & (0x000F)) - 8) * scale.s0; // dequantize a row of the 16 weights
+        dequantized_weights.s1 = ((bits4.s1 & (0x000F)) - 8) * scale.s1;
+        dequantized_weights.s2 = ((bits4.s2 & (0x000F)) - 8) * scale.s2;
+        dequantized_weights.s3 = ((bits4.s3 & (0x000F)) - 8) * scale.s3;
+        c0 += B * dequantized_weights.s0; // vector-scalar multiplication to accumulate
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+
+        // j=1
+        B.s0123 = read_imageh(src1, gy*2 + (i+1)*(n_4));
+        B.s4567 = read_imageh(src1, gy*2 + (i+1)*(n_4)+1);
+        dequantized_weights.s0 = (((bits4.s0 & (0x00F0)) >> 4) - 8) * scale.s0; // dequantize a row of the 16 weights
+        dequantized_weights.s1 = (((bits4.s1 & (0x00F0)) >> 4) - 8) * scale.s1;
+        dequantized_weights.s2 = (((bits4.s2 & (0x00F0)) >> 4) - 8) * scale.s2;
+        dequantized_weights.s3 = (((bits4.s3 & (0x00F0)) >> 4) - 8) * scale.s3;
+        c0 += B * dequantized_weights.s0; //vector-scalar multiplication to accumulate
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+
+        // j=2
+        B.s0123 = read_imageh(src1, gy*2 + (i+2)*(n_4));
+        B.s4567 = read_imageh(src1, gy*2 + (i+2)*(n_4)+1);
+        dequantized_weights.s0 = (((bits4.s0 & (0x0F00)) >> 8) - 8) * scale.s0; // dequantize a row of the 16 weights
+        dequantized_weights.s1 = (((bits4.s1 & (0x0F00)) >> 8) - 8) * scale.s1;
+        dequantized_weights.s2 = (((bits4.s2 & (0x0F00)) >> 8) - 8) * scale.s2;
+        dequantized_weights.s3 = (((bits4.s3 & (0x0F00)) >> 8) - 8) * scale.s3;
+        c0 += B * dequantized_weights.s0; // vector-scalar multiplication to accumulate
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+
+        // j=3
+        B.s0123 = read_imageh(src1, gy*2 + (i+3)*(n_4));
+        B.s4567 = read_imageh(src1, gy*2 + (i+3)*(n_4)+1);
+        dequantized_weights.s0 = (((bits4.s0 & (0xF000)) >> 12) - 8) * scale.s0; // dequantize a row of the 16 weights
+        dequantized_weights.s1 = (((bits4.s1 & (0xF000)) >> 12) - 8) * scale.s1;
+        dequantized_weights.s2 = (((bits4.s2 & (0xF000)) >> 12) - 8) * scale.s2;
+        dequantized_weights.s3 = (((bits4.s3 & (0xF000)) >> 12) - 8) * scale.s3;
+        c0 += B * dequantized_weights.s0; // vector-scalar multiplication to accumulate
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+    }
+
+    int idx = (gy<<3)*m + (gx<<2); // vectorized store 16 elements
+
+    // conditional check if store is to a valid location. Required when N is not a multiple of 8
+    // if statements allow registers to be reused for each store
+    // provides a performance boost due to reduced register footprint, which increases number of concurrent waves
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s0, c1.s0, c2.s0, c3.s0), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s1, c1.s1, c2.s1, c3.s1), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s2, c1.s2, c2.s2, c3.s2), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s3, c1.s3, c2.s3, c3.s3), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s4, c1.s4, c2.s4, c3.s4), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s5, c1.s5, c2.s5, c3.s5), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s6, c1.s6, c2.s6, c3.s6), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s7, c1.s7, c2.s7, c3.s7), 0, dst + idx);
+    }
+}

package/cpp/ggml/src/ggml-opencl/kernels/mul_mat_f16_f32.cl

@@ -0,0 +1,130 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#if defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#else
+#define REQD_SUBGROUP_SIZE_128
+#endif
+
+#define OPWM 64
+#define OPWN 64
+#define CPWK 8
+#define OPTM 4
+#define OPTN 8
+
+#define WG_M (OPWM / OPTM)
+#define WG_N (OPWN / OPTN)
+#define VEC_K (CPWK / 4)
+
+REQD_SUBGROUP_SIZE_128
+__kernel void mul_mat_f16_f32(
+    const int M, const int N, const int K,
+    __global const void* A_void, ulong A_offset,
+    __global const void* B_void, ulong B_offset,
+    __global       void* C_void, ulong C_offset) {
+
+    __global const half*  A = (__global const half* )((__global const char*)A_void + A_offset);
+    __global const float* B = (__global const float*)((__global const char*)B_void + B_offset);
+    __global       float* C = (__global       float*)((__global       char*)C_void + C_offset);
+
+    const int lidm = get_local_id(0);
+    const int lidn = get_local_id(1);
+    const int lid = lidn * WG_M + lidm;
+
+    const int offsetM = get_group_id(0) * OPWM;
+    const int offsetN = get_group_id(1) * OPWN;
+
+    __local half4  Alocal[OPWM][VEC_K];
+    __local float4 Blocal[OPWN][VEC_K];
+
+    float sum[OPTM][OPTN];
+
+    for (int wm = 0; wm < OPTM; wm++) {
+        for (int wn = 0; wn < OPTN; wn++) {
+            sum[wm][wn] = 0.0f;
+        }
+    }
+
+    const int numTiles = (K + CPWK - 1) / CPWK;
+
+    const int load_row_a = lid % OPWM;
+    const int load_vec_k_a = lid / OPWM;
+    const int global_row_a = offsetM + load_row_a;
+
+    const int load_row_b = lid % OPWN;
+    const int load_vec_k_b = lid / OPWN;
+    const int global_row_b = offsetN + load_row_b;
+
+    for (int t = 0; t < numTiles; t++) {
+        const int k_start = t * CPWK;
+        const int k_vec_start_a = k_start + load_vec_k_a * 4;
+        const int k_vec_start_b = k_start + load_vec_k_b * 4;
+
+        if (global_row_a < M && k_vec_start_a < K) {
+            if (k_vec_start_a + 3 < K) {
+                Alocal[load_row_a][load_vec_k_a] = vload4(0, A + global_row_a * K + k_vec_start_a);
+            } else {
+                half4 tempA = (half4)(0.0h);
+                if (k_vec_start_a < K) tempA.s0 = A[global_row_a * K + k_vec_start_a];
+                if (k_vec_start_a + 1 < K) tempA.s1 = A[global_row_a * K + k_vec_start_a + 1];
+                if (k_vec_start_a + 2 < K) tempA.s2 = A[global_row_a * K + k_vec_start_a + 2];
+                Alocal[load_row_a][load_vec_k_a] = tempA;
+            }
+        } else {
+            Alocal[load_row_a][load_vec_k_a] = (half4)(0.0h);
+        }
+
+        if (global_row_b < N && k_vec_start_b < K) {
+            if (k_vec_start_b + 3 < K) {
+                Blocal[load_row_b][load_vec_k_b] = vload4(0, B + global_row_b * K + k_vec_start_b);
+            } else {
+                float4 tempB = (float4)(0.0f);
+                if (k_vec_start_b < K) tempB.s0 = B[global_row_b * K + k_vec_start_b];
+                if (k_vec_start_b + 1 < K) tempB.s1 = B[global_row_b * K + k_vec_start_b + 1];
+                if (k_vec_start_b + 2 < K) tempB.s2 = B[global_row_b * K + k_vec_start_b + 2];
+                Blocal[load_row_b][load_vec_k_b] = tempB;
+            }
+        } else {
+            Blocal[load_row_b][load_vec_k_b] = (float4)(0.0f);
+        }
+
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        #pragma unroll
+        for (int k_vec = 0; k_vec < VEC_K; k_vec++) {
+            float4 a_fvecs[OPTM];
+            int current_row_a = lidm;
+            for (int wm = 0; wm < OPTM; wm++) {
+                a_fvecs[wm] = convert_float4(Alocal[current_row_a][k_vec]);
+                current_row_a += WG_M;
+            }
+
+            float4 b_fvecs[OPTN];
+            int current_row_b = lidn;
+            for (int wn = 0; wn < OPTN; wn++) {
+                b_fvecs[wn] = Blocal[current_row_b][k_vec];
+                current_row_b += WG_N;
+            }
+
+            for (int wm = 0; wm < OPTM; wm++) {
+                for (int wn = 0; wn < OPTN; wn++) {
+                    sum[wm][wn] += dot(a_fvecs[wm], b_fvecs[wn]);
+                }
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    for (int wm = 0; wm < OPTM; wm++) {
+        int globalRow = offsetM + lidm + wm * WG_M;
+        if (globalRow < M) {
+            for (int wn = 0; wn < OPTN; wn++) {
+                int globalCol = offsetN + lidn + wn * WG_N;
+                if (globalCol < N) {
+                    C[globalCol * M + globalRow] = sum[wm][wn];
+                }
+            }
+        }
+    }
+}