llama_cpp 0.4.0 → 0.5.1

data/ext/llama_cpp/src/ggml-metal.metal

@@ -25,9 +25,9 @@ typedef struct {
 } block_q8_0;
 
 kernel void kernel_add(
-        device const float * src0,
-        device const float * src1,
-        device       float * dst,
+        device const float4 * src0,
+        device const float4 * src1,
+        device       float4 * dst,
         uint tpig[[thread_position_in_grid]]) {
     dst[tpig] = src0[tpig] + src1[tpig];
 }
@@ -35,18 +35,18 @@ kernel void kernel_add(
 // assumption: src1 is a row
 // broadcast src1 into src0
 kernel void kernel_add_row(
-        device const float * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
+        device const float4 * src0,
+        device const float4 * src1,
+        device       float4 * dst,
+        constant   int64_t & nb,
         uint tpig[[thread_position_in_grid]]) {
-    dst[tpig] = src0[tpig] + src1[tpig % ne00];
+    dst[tpig] = src0[tpig] + src1[tpig % nb];
 }
 
 kernel void kernel_mul(
-        device const float * src0,
-        device const float * src1,
-        device       float * dst,
+        device const float4 * src0,
+        device const float4 * src1,
+        device       float4 * dst,
         uint tpig[[thread_position_in_grid]]) {
     dst[tpig] = src0[tpig] * src1[tpig];
 }
@@ -54,12 +54,12 @@ kernel void kernel_mul(
 // assumption: src1 is a row
 // broadcast src1 into src0
 kernel void kernel_mul_row(
-        device const float * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
+        device const float4 * src0,
+        device const float4 * src1,
+        device       float4 * dst,
+        constant    int64_t & nb,
         uint tpig[[thread_position_in_grid]]) {
-    dst[tpig] = src0[tpig] * src1[tpig % ne00];
+    dst[tpig] = src0[tpig] * src1[tpig % nb];
 }
 
 kernel void kernel_scale(
@@ -133,19 +133,24 @@ kernel void kernel_soft_max(
         threadgroup_barrier(mem_flags::mem_threadgroup);
     }
 
-    // broadcast
-    if (tpitg[0] == 0) {
-        buf[0] = buf[0];
-    }
+    //// broadcast - not needed. There is a threadgroup barrier above in the last iteration of
+    //               the loop, and when that is done, buf[0] has the correct (synchronized) value
+    //if (tpitg[0] == 0) {
+    //    buf[0] = buf[0];
+    //}
 
-    threadgroup_barrier(mem_flags::mem_threadgroup);
+    //threadgroup_barrier(mem_flags::mem_threadgroup);
 
     const float max = buf[0];
 
     // parallel sum
     buf[tpitg[0]] = 0.0f;
     for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) {
-        buf[tpitg[0]] += exp(psrc0[i00] - max);
+        const float exp_psrc0 = exp(psrc0[i00] - max);
+        buf[tpitg[0]] += exp_psrc0;
+        // Remember the result of exp here. exp is expensive, so we really do not
+        // whish to compute it twice.
+        pdst[i00] = exp_psrc0;
     }
 
     // reduce
@@ -157,17 +162,18 @@ kernel void kernel_soft_max(
         threadgroup_barrier(mem_flags::mem_threadgroup);
     }
 
-    // broadcast
-    if (tpitg[0] == 0) {
-        buf[0] = buf[0];
-    }
+    // broadcast - not needed, see above
+    //// broadcast
+    //if (tpitg[0] == 0) {
+    //    buf[0] = buf[0];
+    //}
 
-    threadgroup_barrier(mem_flags::mem_threadgroup);
+    //threadgroup_barrier(mem_flags::mem_threadgroup);
 
     const float sum = buf[0];
 
     for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) {
-        pdst[i00] = exp(psrc0[i00] - max) / sum;
+        pdst[i00] /= sum;
     }
 }
 
@@ -214,25 +220,17 @@ kernel void kernel_norm(
         }
         threadgroup_barrier(mem_flags::mem_threadgroup);
     }
-    // broadcast
-    if (tpitg == 0) {
-        sum[0] /= ne00;
-    }
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    const float mean  = sum[0];
+    const float mean  = sum[0] / ne00;
 
-    // recenter
+    // recenter and VARIANCE
+    threadgroup_barrier(mem_flags::mem_threadgroup);
     device float * y = dst + tgpig*ne00;
-    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
-        y[i00] = x[i00] - mean;
-    }
-
-    // VARIANCE
-    // parallel sum
     sum[tpitg] = 0.0f;
     for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
+        y[i00] = x[i00] - mean;
         sum[tpitg] += y[i00] * y[i00];
     }
+
     // reduce
     threadgroup_barrier(mem_flags::mem_threadgroup);
     for (uint i = ntg/2; i > 0; i /= 2) {
@@ -241,12 +239,7 @@ kernel void kernel_norm(
         }
         threadgroup_barrier(mem_flags::mem_threadgroup);
     }
-    // broadcast
-    if (tpitg == 0) {
-        sum[0] /= ne00;
-    }
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    const float variance = sum[0];
+    const float variance = sum[0] / ne00;
 
     const float scale = 1.0f/sqrt(variance + eps);
     for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
@@ -254,7 +247,6 @@ kernel void kernel_norm(
     }
 }
 
-
 kernel void kernel_rms_norm(
         device const  void * src0,
         device       float * dst,
@@ -435,6 +427,8 @@ kernel void kernel_mul_mat_q4_1_f32(
      mul_vec_q_n_f32<block_q4_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,gqa,tgpig,tiisg,sgitg);
 }
 
+#define NB_Q8_0 8
+
 kernel void kernel_mul_mat_q8_0_f32(
         device const  void * src0,
         device const float * src1,
@@ -463,30 +457,30 @@ kernel void kernel_mul_mat_q8_0_f32(
     device const block_q8_0 * x = (device const block_q8_0 *) src0 + offset0;
     device const float      * y = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
 
-    float yl[16];
+    float yl[NB_Q8_0];
     float sumf[nr]={0.f};
 
-    const int ix = tiisg/2;
-    const int il = tiisg%2;
+    const int ix = tiisg/4;
+    const int il = tiisg%4;
 
-    device const float * yb = y + ix * QK8_0 + 16*il;
+    device const float * yb = y + ix * QK8_0 + NB_Q8_0*il;
 
-    // each thread in a SIMD group deals with half a block.
-    for (int ib = ix; ib < nb; ib += nw/2) {
-        for (int i = 0; i < 16; ++i) {
+    // each thread in a SIMD group deals with NB_Q8_0 quants at a time
+    for (int ib = ix; ib < nb; ib += nw/4) {
+        for (int i = 0; i < NB_Q8_0; ++i) {
             yl[i] = yb[i];
         }
 
         for (int row = 0; row < nr; row++) {
-            device const int8_t * qs = x[ib+row*nb].qs + 16*il;
+            device const int8_t * qs = x[ib+row*nb].qs + NB_Q8_0*il;
             float sumq = 0.f;
-            for (int iq = 0; iq < 16; ++iq) {
+            for (int iq = 0; iq < NB_Q8_0; ++iq) {
                 sumq += qs[iq] * yl[iq];
             }
             sumf[row] += sumq*x[ib+row*nb].d;
         }
 
-        yb += QK8_0 * 16;
+        yb += NB_Q8_0 * nw;
     }
 
     for (int row = 0; row < nr; ++row) {
@@ -497,7 +491,7 @@ kernel void kernel_mul_mat_q8_0_f32(
     }
 }
 
-kernel void kernel_mul_mat_f16_f32(
+kernel void kernel_mul_mat_f16_f32_1row(
         device const  char * src0,
         device const  char * src1,
         device       float * dst,
@@ -515,11 +509,8 @@ kernel void kernel_mul_mat_f16_f32(
         constant  uint64_t & nb12,
         constant   int64_t & ne0,
         constant   int64_t & ne1,
-        threadgroup float  * sum [[threadgroup(0)]],
         uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3  tpig[[thread_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3  tptg[[threads_per_threadgroup]]) {
+        uint tiisg[[thread_index_in_simdgroup]]) {
 
     const int64_t r0 = tgpig.x;
     const int64_t r1 = tgpig.y;
@@ -528,23 +519,100 @@ kernel void kernel_mul_mat_f16_f32(
     device const half  * x = (device const half  *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02);
     device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
 
-    sum[tpitg.x] = 0.0f;
-
-    for (int i = tpitg.x; i < ne00; i += tptg.x) {
-        sum[tpitg.x] += (float) x[i] * (float) y[i];
+    float sumf = 0;
+    if (ne00 < 128) {
+        for (int i = tiisg; i < ne00; i += 32) {
+            sumf += (float) x[i] * (float) y[i];
+        }
+        float all_sum = simd_sum(sumf);
+        if (tiisg == 0) {
+            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+        }
+    } else {
+        device const half4  * x4 = (device const half4  *) x;
+        device const float4 * y4 = (device const float4 *) y;
+        for (int i = tiisg; i < ne00/4; i += 32) {
+            for (int k = 0; k < 4; ++k) sumf += (float)x4[i][k] * y4[i][k];
+        }
+        float all_sum = simd_sum(sumf);
+        if (tiisg == 0) {
+            for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (float) x[i] * y[i];
+            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+        }
     }
 
-    // accumulate the sum from all threads in the threadgroup
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    for (uint i = tptg.x/2; i > 0; i /= 2) {
-        if (tpitg.x < i) {
-            sum[tpitg.x] += sum[tpitg.x + i];
+}
+
+#define N_F16_F32 4
+
+kernel void kernel_mul_mat_f16_f32(
+        device const  char * src0,
+        device const  char * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint tiisg[[thread_index_in_simdgroup]]) {
+
+    const int64_t r0 = tgpig.x;
+    const int64_t rb = tgpig.y*N_F16_F32;
+    const int64_t im = tgpig.z;
+
+    device const half * x = (device const half *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02);
+
+    if (ne00 < 128) {
+        for (int row = 0; row < N_F16_F32; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
+
+            float sumf = 0;
+            for (int i = tiisg; i < ne00; i += 32) {
+                sumf += (float) x[i] * (float) y[i];
+            }
+
+            float all_sum = simd_sum(sumf);
+            if (tiisg == 0) {
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
         }
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-    }
+    } else {
+        device const half4 * x4 = (device const half4 *)x;
+        for (int row = 0; row < N_F16_F32; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            device const float  * y  = (device const float  *) (src1 + r1*nb11 + im*nb12);
+            device const float4 * y4 = (device const float4 *) y;
+
+            float sumf = 0;
+            for (int i = tiisg; i < ne00/4; i += 32) {
+                for (int k = 0; k < 4; ++k) sumf += (float) x4[i][k] * y4[i][k];
+            }
 
-    if (tpitg.x == 0) {
-        dst[im*ne1*ne0 + r1*ne0 + r0] = sum[0];
+            float all_sum = simd_sum(sumf);
+            if (tiisg == 0) {
+                for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (float) x[i] * y[i];
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
     }
 }
 
@@ -614,25 +682,27 @@ kernel void kernel_rope(
         constant       int & mode,
         constant     float & freq_base,
         constant     float & freq_scale,
-        uint3 tpig[[thread_position_in_grid]]) {
-    const int64_t i3 = tpig[2];
-    const int64_t i2 = tpig[1];
-    const int64_t i1 = tpig[0];
+        uint  tiitg[[thread_index_in_threadgroup]],
+        uint3 tptg[[threads_per_threadgroup]],
+        uint3 tgpig[[threadgroup_position_in_grid]]) {
+    const int64_t i3 = tgpig[2];
+    const int64_t i2 = tgpig[1];
+    const int64_t i1 = tgpig[0];
 
     const bool is_neox = mode & 2;
-    const float theta_scale = pow(freq_base, -2.0f/n_dims);
 
     const int64_t p = ((mode & 1) == 0 ? n_past + i2 : i2);
 
-    float theta = freq_scale * (float)p;
+    const float theta_0 = freq_scale * (float)p;
+    const float inv_ndims = -1.f/n_dims;
 
     if (!is_neox) {
-        for (int64_t i0 = 0; i0 < ne0; i0 += 2) {
+        for (int64_t i0 = 2*tiitg; i0 < ne0; i0 += 2*tptg.x) {
+
+            const float theta = theta_0 * pow(freq_base, inv_ndims*i0);
             const float cos_theta = cos(theta);
             const float sin_theta = sin(theta);
 
-            theta *= theta_scale;
-
             device const float * const src = (device float *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
             device       float * dst_data  = (device float *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
 
@@ -644,12 +714,12 @@ kernel void kernel_rope(
         }
     } else {
         for (int64_t ib = 0; ib < ne0/n_dims; ++ib) {
-            for (int64_t ic = 0; ic < n_dims; ic += 2) {
+            for (int64_t ic = 2*tiitg; ic < n_dims; ic += 2*tptg.x) {
+
+                const float theta = theta_0 * pow(freq_base, inv_ndims*ic - ib);
                 const float cos_theta = cos(theta);
                 const float sin_theta = sin(theta);
 
-                theta *= theta_scale;
-
                 const int64_t i0 = ib*n_dims + ic/2;
 
                 device const float * const src = (device float *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
@@ -1244,7 +1314,8 @@ kernel void kernel_mul_mat_q4_K_f32(
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int r2 = tgpig.z;
-    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
+    //const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
+    const int first_row = r0 * N_DST;
     const int ib_row = first_row * nb;
     const uint offset0 = r2/gqa*(nb*ne0);
     device const block_q4_K * x = (device const block_q4_K *) src0 + ib_row + offset0;

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

@@ -1334,7 +1334,7 @@ void ggml_cl_free_data(const struct ggml_tensor* tensor) {
         return;
     }
 
-    cl_mem mem = (cl_mem)tensor->data;
+    cl_mem mem = (cl_mem)tensor->extra;
     clReleaseMemObject(mem);
 }
 
@@ -1393,7 +1393,7 @@ static void ggml_cl_mul_f32(const ggml_tensor * src0, const ggml_tensor * src1,
     size_t d_size;
 
     cl_mem d_X = ggml_cl_pool_malloc(ne0 * sizeof(float), &x_size); // src0
-    cl_mem d_Y = (cl_mem) src1->data; // src1 is already on device, broadcasted.
+    cl_mem d_Y = (cl_mem) src1->extra; // src1 is already on device, broadcasted.
     cl_mem d_D = ggml_cl_pool_malloc(ne0 * sizeof(float), &d_size); // dst
 
 
@@ -1491,9 +1491,9 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
     size_t d_size;
     cl_mem d_X;
     if (src0->backend == GGML_BACKEND_GPU) { // NOLINT
-        d_X = (cl_mem) src0->data;
+        d_X = (cl_mem) src0->extra;
     } else {
-        d_X = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * x_ne, &x_size);
+        d_X = ggml_cl_pool_malloc(sizeof(float) * x_ne, &x_size);
     }
     cl_mem d_Y = ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
     cl_mem d_D = ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
@@ -1567,7 +1567,7 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
     size_t d_size;
     cl_mem d_X;
     if (src0->backend == GGML_BACKEND_GPU) { // NOLINT
-        d_X = (cl_mem) src0->data;
+        d_X = (cl_mem) src0->extra;
     } else {
         d_X = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * x_ne, &x_size);
     }
@@ -1697,7 +1697,7 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
                 events.emplace_back();
                 CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Q, 0, src0, i03, i02, events.data() + ev_idx++));
             } else if (src0->backend == GGML_BACKEND_GPU) {
-                d_Q = (cl_mem) src0->data;
+                d_Q = (cl_mem) src0->extra;
             } else {
                 GGML_ASSERT(false);
             }
@@ -1860,6 +1860,6 @@ void ggml_cl_transform_tensor(void * data, ggml_tensor * tensor) {
 
     CL_CHECK(clFinish(queue));
 
-    tensor->data = dst;
+    tensor->extra = dst;
     GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
 }