metal : add Q4_K implementation (ggml-org#1733)

* Metal implementation for Q4_K

Very slow for now:
42 ms / token, Q4_0 runs in 28 ms/token on my
30-core M2 Max GPU.

* Optimizing Q4_K on metal

The first token always takes longer, I guess because
the metal kernel is being jit-compiled.
So, using n = 128 to measure time.

At this point Q4_K takes 29.5 ms / token
compared to 27.2 ms / token for Q4_0.
Quite a bit better than the initial attempt,
but still not good enough.

* Optimizing q4_K metal dot some more

For n = 256 it is now 28.1 ms/token compared to
27 ms/token for q4_0.

* Fix after merge with master

---------

Co-authored-by: Iwan Kawrakow <[email protected]>

Author: ikawrakow

.clang-tidy

@@ -49,9 +49,11 @@
     GGML_METAL_DECL_KERNEL(diag_mask_inf);
     GGML_METAL_DECL_KERNEL(get_rows_f16);
     GGML_METAL_DECL_KERNEL(get_rows_q4_0);
+    GGML_METAL_DECL_KERNEL(get_rows_q4_k);
     GGML_METAL_DECL_KERNEL(rms_norm);
     GGML_METAL_DECL_KERNEL(mul_mat_f16_f32);
     GGML_METAL_DECL_KERNEL(mul_mat_q4_0_f32);
+    GGML_METAL_DECL_KERNEL(mul_mat_q4_k_f32);
     GGML_METAL_DECL_KERNEL(rope);
     GGML_METAL_DECL_KERNEL(cpy_f32_f16);
     GGML_METAL_DECL_KERNEL(cpy_f32_f32);
@@ -133,9 +135,11 @@
         GGML_METAL_ADD_KERNEL(diag_mask_inf);
         GGML_METAL_ADD_KERNEL(get_rows_f16);
         GGML_METAL_ADD_KERNEL(get_rows_q4_0);
+        GGML_METAL_ADD_KERNEL(get_rows_q4_k);
         GGML_METAL_ADD_KERNEL(rms_norm);
         GGML_METAL_ADD_KERNEL(mul_mat_f16_f32);
         GGML_METAL_ADD_KERNEL(mul_mat_q4_0_f32);
+        GGML_METAL_ADD_KERNEL(mul_mat_q4_k_f32);
         GGML_METAL_ADD_KERNEL(rope);
         GGML_METAL_ADD_KERNEL(cpy_f32_f16);
         GGML_METAL_ADD_KERNEL(cpy_f32_f32);
@@ -517,7 +521,20 @@ void ggml_metal_graph_compute(
                                     nth1 = 4;
                                     [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_0_f32];
                                 } break;
-                            default: GGML_ASSERT(false && "not implemented");
+                            case GGML_TYPE_Q4_K:
+                                {
+                                    GGML_ASSERT(ne02 == 1);
+                                    GGML_ASSERT(ne12 == 1);
+
+                                    nth0 = 4;
+                                    nth1 = 16;
+                                    [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_k_f32];
+                                } break;
+                            default:
+                                {
+                                    fprintf(stderr, "Asserting on type %d\n",(int)src0t);
+                                    GGML_ASSERT(false && "not implemented");
+                                }
                         };
 
 
@@ -540,6 +557,9 @@ void ggml_metal_graph_compute(
                         if (src0t == GGML_TYPE_Q4_0) {
                             [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
                             [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                        } else if (src0t == GGML_TYPE_Q4_K) {
+                            [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                         } else {
                             [encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0];
                             [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
@@ -555,6 +575,7 @@ void ggml_metal_graph_compute(
                     switch (src0->type) {
                         case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_get_rows_f16]; break;
                         case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_0]; break;
+                        case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_k]; break;
                         default: GGML_ASSERT(false && "not implemented");
                     }
 

ggml-metal.m

@@ -503,3 +503,165 @@ kernel void kernel_cpy_f32_f32(
         dst_data[i00] = src[0];
     }
 }
+
+//============================================ k-quants ======================================================
+
+#define QK_K 256
+
+typedef struct {
+    half d;             // super-block scale for quantized scales
+    half dmin;          // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+} block_q4_k;
+
+static inline uchar4 get_scale_min_k4(int j, device const uint8_t * q) {
+    uchar4 r;
+    if (j < 4) {
+        r[0] = q[j+0] & 63; r[1] = q[j+4] & 63;
+        r[2] = q[j+1] & 63; r[3] = q[j+5] & 63;
+    } else {
+        r[0] = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4);
+        r[1] = (q[j+4] >>  4) | ((q[j-0] >> 6) << 4);
+        r[2] = (q[j+5] & 0xF) | ((q[j-3] >> 6) << 4);
+        r[3] = (q[j+5] >>  4) | ((q[j+1] >> 6) << 4);
+    }
+    return r;
+}
+
+static void dequantize_row_q4_k(device const block_q4_k * x, device float * y, int k) {
+    assert(k % QK_K == 0);
+    const int nb = k / QK_K;
+
+    for (int i = 0; i < nb; i++) {
+
+        const float d = x[i].d;
+        const float min = x[i].dmin;
+
+        device const uint8_t * q = x[i].qs;
+        device const uint8_t * scales = x[i].scales;
+
+        int is = 0;
+        for (int j = 0; j < QK_K; j += 64) {
+            const uchar4 sc = get_scale_min_k4(is, scales);
+            const float d1 = d * sc[0]; const float m1 = min * sc[1];
+            const float d2 = d * sc[2]; const float m2 = min * sc[3];
+            for (int l = 0; l < 32; ++l) *y++ = d1 * (q[l] & 0xF) - m1;
+            for (int l = 0; l < 32; ++l) *y++ = d2 * (q[l]  >> 4) - m2;
+            q += 32; is += 2;
+        }
+
+    }
+}
+
+kernel void kernel_get_rows_q4_k(
+        device const  void * src0,
+        device const   int * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb1,
+        uint tpig[[thread_position_in_grid]]) {
+    const int i = tpig;
+    const int r = ((device int32_t *) src1)[i];
+
+    dequantize_row_q4_k(
+            (device const block_q4_k *) ((device char *) src0 + r*nb01),
+                       (device float *) ((device char *)  dst + i*nb1), ne00);
+}
+
+kernel void kernel_mul_mat_q4_k_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        threadgroup float  * sum [[threadgroup(0)]],
+        uint2 tgpig[[threadgroup_position_in_grid]],
+        uint2  tpig[[thread_position_in_grid]],               // we don't use this for now
+        uint2 tpitg[[thread_position_in_threadgroup]],
+        uint2  tptg[[threads_per_threadgroup]]) {
+
+    const int nb = ne00/QK_K;
+
+    const int64_t r0 = tgpig.x;
+    const int64_t r1 = tgpig.y;
+
+    device const block_q4_k * x = (device const block_q4_k *) src0 + r0*nb;
+    device const float     * yy = (device const float      *) src1 + r1*ne10;
+
+    const uint nth = tptg.x*tptg.y;
+    const uint ith = tptg.y*tpitg.x + tpitg.y;
+
+    const int tid = tpitg.y;   // 0...16
+    const int il  = tid/4;     // 0...3
+    const int ir  = tid%4;     // 0...3
+    const int n   = 8;
+    const int is  = 2*il;
+
+    sum[ith] = 0.0f;
+
+    float sumf = 0;
+    for (int i = tpitg.x; i < nb; i += tptg.x) {
+
+        device const uint8_t * q = (x + i)->qs + 32*il + n*ir;
+        device const float   * y = yy + i*QK_K + 64*il + n*ir;
+        device const uint8_t * scales = (x + i)->scales;
+
+        const float dall = (float)((x + i)->d);
+        const float dmin = (float)((x + i)->dmin);
+
+        const uchar4 sc = get_scale_min_k4(is, scales);
+
+        float4 s = {0.f, 0.f, 0.f, 0.f};
+        for (int l = 0; l < n; ++l) {
+            s[0] += y[l+ 0] * (q[l] & 0xF); s[1] += y[l+ 0];
+            s[2] += y[l+32] * (q[l] >>  4); s[3] += y[l+32];
+        }
+        sumf += dall * (s[0] * sc[0] + s[2] * sc[2]) - dmin * (s[1] * sc[1] + s[3] * sc[3]);
+
+    }
+    sum[ith] = sumf;
+
+    //
+    // Accumulate the sum from all threads in the threadgroup
+    // This version is slightly faster than the commented out one below,
+    // which I copy-pasted from ggerganov's q4_0 dot product for metal.
+    //
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    if (ith%4 == 0) {
+        for (int i = 1; i < 4; ++i) sum[ith] += sum[ith + i];
+    }
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    if (ith%16 == 0) {
+        for (int i = 4; i < 16; i += 4) sum[ith] += sum[ith + i];
+    }
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    if (ith == 0) {
+        for (int i = 16; i < nth; i += 16) sum[0] += sum[i];
+        dst[r1*ne0 + r0] = sum[0];
+    }
+
+    //// accumulate the sum from all threads in the threadgroup
+    //threadgroup_barrier(mem_flags::mem_threadgroup);
+    //for (uint i = nth/2; i > 0; i /= 2) {
+    //    if (ith < i) {
+    //        sum[ith] += sum[ith + i];
+    //    }
+    //    threadgroup_barrier(mem_flags::mem_threadgroup);
+    //}
+
+    //if (ith == 0) {
+    //    dst[r1*ne0 + r0] = sum[0];
+    //}
+}