iq4_ks: faster dot product on Metal (#90)

TG-128(LLaMA-3.1-8B) goes to 52.5 t/s up from 48.4 t/s. Co-authored-by: Iwan Kawrakow <[email protected]>
Nexesenex · Oct 16, 2024 · 993ca95 · 993ca95
1 parent ff23008
commit 993ca95
Showing 1 changed file with 14 additions and 10 deletions.
diff --git a/ggml/src/ggml-metal.metal b/ggml/src/ggml-metal.metal
@@ -6079,6 +6079,7 @@ void kernel_mul_mv_iq4_ks_f32_impl(
 
     float4 yl[4];
     float2 sumf = 0.f;
+    float  d[2];
 
     device const float * yb = y + ix * QK_K + ib * 32 + il * 8;
 
@@ -6087,22 +6088,25 @@ void kernel_mul_mv_iq4_ks_f32_impl(
 
     float4 qf1, qf2;
 
+    device const float * dptr = (device const float *)cx;
+    d[0] = *dptr;
+    device const block_iq4_ks * x = (device const block_iq4_ks *)(dptr + 1) + ix;
+    dptr += row_size/4;
+    d[1] = *dptr;
+
     for (int ibl = ix; ibl < nb; ibl += 2) {
 
         device const float4 * y4 = (device const float4 *)yb;
         yl[0] = y4[0]; yl[1] = y4[4]; yl[2] = y4[1]; yl[3] = y4[5];
 
-        device const float * dptr = (device const float *)cx;
+        device const uint8_t * scales = x->scales;
 
         for (int row = 0; row < 2; ++row) {
 
-            //device const float * dptr = (device const float *)(cx + row*row_size);
-            const float d = *dptr;
-            device const block_iq4_ks *  x = (device const block_iq4_ks *)(dptr + 1);
-            device const block_iq4_ks & xb = x[ibl];
-            device const uint32_t * q4 = (device const uint32_t *)(xb.qs + 16*ib + 8*il);
+            threadgroup const float * block_values = shared_values + ((scales[ib] & 1) << 4);
+            const float ls = ((scales[ib] & 254) - 127);
 
-            threadgroup const float * block_values = shared_values + ((xb.scales[ib] & 1) << 4);
+            device const uint32_t * q4 = (device const uint32_t *)scales + QK_K/128 + 4*ib + 2*il;
 
             float4 acc1 = {0.f}, acc2 = {0.f};
 
@@ -6122,14 +6126,14 @@ void kernel_mul_mv_iq4_ks_f32_impl(
 
             acc1 += acc2;
 
-            const int ls = (xb.scales[ib] & 254) - 127;
-            sumf[row] += d * ls * (acc1[0] + acc1[1] + acc1[2] + acc1[3]);
+            sumf[row] += d[row] * ls * (acc1[0] + acc1[1] + acc1[2] + acc1[3]);
 
-            dptr += row_size/4;
+            scales += row_size;
 
         }
 
         yb += 2 * QK_K;
+        x  += 2;
     }
 
     sumf = simd_sum(sumf);