metal : Q6_K implementation (#1752)

* 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

* Metal implementation for Q6_K

Similar to the CUDA implementation.
No idea if this is the optimum for Metal, but the few
alternative variants I tried all had a lower performance.

We get 36.5 ms / token on M2 Max with 30 GPU cores.
This corresponds to ~200 GB/second throughput.

* clang-tidy : add config back

* Much better Q6_K implementation for metal

28.3 ms / token for 7B. Subtracting ~9 ms that is spent in
other compute graph operations, we are left with ~19 ms
for the matrix multiplications. The model is ~5.5 GB,
so we are getting 1000 / 19 * 5.5 = 290 GB/s!

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>

ggml-metal.m

@@ -50,10 +50,12 @@ struct ggml_metal_context {
     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(get_rows_q6_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(mul_mat_q6_k_f32);
     GGML_METAL_DECL_KERNEL(rope);
     GGML_METAL_DECL_KERNEL(cpy_f32_f16);
     GGML_METAL_DECL_KERNEL(cpy_f32_f32);
@@ -136,10 +138,12 @@ struct ggml_metal_context * ggml_metal_init(void) {
         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(get_rows_q6_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(mul_mat_q6_k_f32);
         GGML_METAL_ADD_KERNEL(rope);
         GGML_METAL_ADD_KERNEL(cpy_f32_f16);
         GGML_METAL_ADD_KERNEL(cpy_f32_f32);
@@ -530,6 +534,15 @@ void ggml_metal_graph_compute(
                                     nth1 = 16;
                                     [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_k_f32];
                                 } break;
+                            case GGML_TYPE_Q6_K:
+                                {
+                                    GGML_ASSERT(ne02 == 1);
+                                    GGML_ASSERT(ne12 == 1);
+
+                                    nth0 = 4;
+                                    nth1 = 16;
+                                    [encoder setComputePipelineState:ctx->pipeline_mul_mat_q6_k_f32];
+                                } break;
                             default:
                                 {
                                     fprintf(stderr, "Asserting on type %d\n",(int)src0t);
@@ -560,6 +573,9 @@ void ggml_metal_graph_compute(
                         } 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 if (src0t == GGML_TYPE_Q6_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)];
@@ -576,6 +592,7 @@ void ggml_metal_graph_compute(
                         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;
+                        case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q6_k]; break;
                         default: GGML_ASSERT(false && "not implemented");
                     }
 

ggml-metal.metal

@@ -303,18 +303,37 @@ kernel void kernel_mul_mat_q4_0_f32(
         sum[ith] += acc*d;
     }
 
-    // accumulate the sum from all threads in the threadgroup
+    //
+    // 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);
-    for (uint i = nth/2; i > 0; i /= 2) {
+    if (ith%4 == 0) {
-        if (ith < i) {
+        for (int i = 1; i < 4; ++i) sum[ith] += sum[ith + 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];
+    //}
 }
 
 kernel void kernel_mul_mat_f16_f32(
@@ -515,6 +534,13 @@ typedef struct {
     uint8_t qs[QK_K/2];        // 4--bit quants
 } block_q4_k;
 
+typedef struct {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    half d;                  // super-block scale
+} block_q6_k;
+
 static inline uchar4 get_scale_min_k4(int j, device const uint8_t * q) {
     uchar4 r;
     if (j < 4) {
@@ -554,6 +580,38 @@ static void dequantize_row_q4_k(device const block_q4_k * x, device float * y, i
     }
 }
 
+static void dequantize_row_q6_k(device const block_q6_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;
+
+        device const uint8_t * ql = x[i].ql;
+        device const uint8_t * qh = x[i].qh;
+        device const int8_t  * sc = x[i].scales;
+
+        for (int n = 0; n < QK_K; n += 128) {
+            for (int l = 0; l < 32; ++l) {
+                int is = l/16;
+                const int8_t q1 = (int8_t)((ql[l +  0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
+                const int8_t q2 = (int8_t)((ql[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
+                const int8_t q3 = (int8_t)((ql[l +  0]  >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32;
+                const int8_t q4 = (int8_t)((ql[l + 32]  >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32;
+                y[l +  0] = d * sc[is + 0] * q1;
+                y[l + 32] = d * sc[is + 2] * q2;
+                y[l + 64] = d * sc[is + 4] * q3;
+                y[l + 96] = d * sc[is + 6] * q4;
+            }
+            y  += 128;
+            ql += 64;
+            qh += 32;
+            sc += 8;
+        }
+    }
+}
+
 kernel void kernel_get_rows_q4_k(
         device const  void * src0,
         device const   int * src1,
@@ -665,3 +723,108 @@ kernel void kernel_mul_mat_q4_k_f32(
     //    dst[r1*ne0 + r0] = sum[0];
     //}
 }
+
+kernel void kernel_get_rows_q6_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_q6_k(
+            (device const block_q6_k *) ((device char *) src0 + r*nb01),
+                       (device float *) ((device char *)  dst + i*nb1), ne00);
+}
+
+kernel void kernel_mul_mat_q6_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 uint8_t kmask1 = 0x03;
+    const uint8_t kmask2 = 0x0C;
+    const uint8_t kmask3 = 0x30;
+    const uint8_t kmask4 = 0xC0;
+
+    const int nb = ne00/QK_K;
+
+    const int64_t r0 = tgpig.x;
+    const int64_t r1 = tgpig.y;
+
+    device const block_q6_k * x = (device const block_q6_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 step = QK_K / tptg.y;     // we expect this to be 16
+    const int iqs  = step * tpitg.y;    // 0...240 in steps of 16
+    const int ip   = iqs / 128;         // 0 or 1
+    const int il   = (iqs - 128*ip)/16; // 0...7
+    const int n    = 4;
+    const int is   = 8*ip + (n*il)/16;
+
+    float sumf = 0;
+    for (int i = tpitg.x; i < nb; i += tptg.x) {
+
+        device const uint8_t * ql = x[i].ql + 64*ip + n*il;
+        device const uint8_t * qh = x[i].qh + 32*ip + n*il;
+        device const int8_t  * sc = x[i].scales + is;
+
+        device const float * y = yy + i * QK_K + 128*ip + n*il;
+
+        const float dall = x[i].d;
+
+        float4 sums = {0.f, 0.f, 0.f, 0.f};
+        for (int l = 0; l < n; ++l) {
+            sums[0] += y[l+ 0] * ((int8_t)((ql[l+ 0] & 0xF) | ((qh[l] & kmask1) << 4)) - 32);
+            sums[1] += y[l+32] * ((int8_t)((ql[l+32] & 0xF) | ((qh[l] & kmask2) << 2)) - 32);
+            sums[2] += y[l+64] * ((int8_t)((ql[l+ 0]  >> 4) | ((qh[l] & kmask3) << 0)) - 32);
+            sums[3] += y[l+96] * ((int8_t)((ql[l+32]  >> 4) | ((qh[l] & kmask4) >> 2)) - 32);
+        }
+
+        sumf += dall * (sums[0] * sc[0] + sums[1] * sc[2] + sums[2] * sc[4] + sums[3] * sc[6]);
+
+    }
+
+    sum[ith] = sumf;
+
+    //
+    // Accumulate the sum from all threads in the threadgroup
+    //
+    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];
+    }
+
+}