Vulkan: Use improved q4_k and q5_k dequant code in dequant shaders (#10798)

ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_k.comp

@@ -9,8 +9,8 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
 
 void main() {
     [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
-        const uint i = gl_WorkGroupID.x * 256 + wgy;
-        if (i >= p.M * p.K / QUANT_K) {
+        const uint ib = gl_WorkGroupID.x * 256 + wgy;
+        if (ib >= p.M * p.K / QUANT_K) {
             return;
         }
 
@@ -20,37 +20,49 @@ void main() {
         const uint is = 2 * il;
         const uint n = 4;
 
-        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[i].d.x);
-        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[i].d.y);
+        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib].d.x);
+        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib].d.y);
 
-        const uint y_idx = i * QUANT_K + 64 * il + n * ir;
+        const uint y_idx = ib * QUANT_K + 64 * il + n * ir;
         const uint qs_idx = 32*il + n * ir;
 
-        uint8_t sc;
-        uint8_t m;
-        if (is < 4) {
-            sc = uint8_t(data_a[i].scales[is] & 63);
-            m  = uint8_t(data_a[i].scales[is + 4] & 63);
-        } else {
-            sc = uint8_t((data_a[i].scales[is + 4] & 0xF) | ((data_a[i].scales[is - 4] >> 6) << 4));
-            m  = uint8_t((data_a[i].scales[is + 4] >>  4) | ((data_a[i].scales[is    ] >> 6) << 4));
-        }
+        uint scidx0 = (is < 4) ? is : (is + 4);
+        uint scidx1 = (is < 4) ? is : (is - 4);
+        uint scidxmask1 = (is < 4) ? 0x30 : 0xC0;
+        uint scidxshift1 = (is < 4) ? 0 : 2;
+        uint mbidx0 = is + 4;
+        uint mbidx1 = (is < 4) ? is + 4 : is;
+        uint mbidxmask0 = (is < 4) ? 0xF : 0xF0;
+        uint mbidxshift0 = (is < 4) ? 0 : 4;
+        uint mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
+        uint mbidxshift1 = (is < 4) ? 0 : 2;
+
+        uint8_t sc = uint8_t((data_a[ib].scales[scidx0] & 0xF) | ((data_a[ib].scales[scidx1] & scidxmask1) >> scidxshift1));
+        uint8_t mbyte = uint8_t((data_a[ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0 | ((data_a[ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1));
+
         const FLOAT_TYPE d1 = dall * sc;
-        const FLOAT_TYPE m1 = dmin * m;
-
-        if (is < 4) {
-            sc = uint8_t(data_a[i].scales[is + 1] & 63);
-            m  = uint8_t(data_a[i].scales[is + 5] & 63);
-        } else {
-            sc = uint8_t((data_a[i].scales[is + 5] & 0xF) | ((data_a[i].scales[is - 3] >> 6) << 4));
-            m  = uint8_t((data_a[i].scales[is + 5] >>  4) | ((data_a[i].scales[is + 1] >> 6) << 4));
-        }
+        const FLOAT_TYPE m1 = dmin * mbyte;
+
+        scidx0 = (is < 4) ? is + 1 : (is + 5);
+        scidx1 = (is < 4) ? is + 1 : (is - 3);
+        scidxmask1 = (is < 4) ? 0x30 : 0xC0;
+        scidxshift1 = (is < 4) ? 0 : 2;
+        mbidx0 = is + 5;
+        mbidx1 = (is < 4) ? is + 5 : is + 1;
+        mbidxmask0 = (is < 4) ? 0xF : 0xF0;
+        mbidxshift0 = (is < 4) ? 0 : 4;
+        mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
+        mbidxshift1 = (is < 4) ? 0 : 2;
+
+        sc = uint8_t((data_a[ib].scales[scidx0] & 0xF) | ((data_a[ib].scales[scidx1] & scidxmask1) >> scidxshift1));
+        mbyte = uint8_t((data_a[ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0 | ((data_a[ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1));
+
         const FLOAT_TYPE d2 = dall * sc;
-        const FLOAT_TYPE m2 = dmin * m;
+        const FLOAT_TYPE m2 = dmin * mbyte;
 
         [[unroll]] for (uint l = 0; l < n; ++l) {
-            data_b[y_idx + l     ] = D_TYPE(d1 * FLOAT_TYPE(data_a[i].qs[qs_idx + l] & 0xF) - m1);
-            data_b[y_idx + l + 32] = D_TYPE(d2 * FLOAT_TYPE(data_a[i].qs[qs_idx + l] >>  4) - m2);
+            data_b[y_idx + l     ] = D_TYPE(d1 * FLOAT_TYPE(data_a[ib].qs[qs_idx + l] & 0xF) - m1);
+            data_b[y_idx + l + 32] = D_TYPE(d2 * FLOAT_TYPE(data_a[ib].qs[qs_idx + l] >>  4) - m2);
         }
     }
 }

ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_k.comp

@@ -9,8 +9,8 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
 
 void main() {
     [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
-        const uint i = gl_WorkGroupID.x * 256 + wgy;
-        if (i >= p.M * p.K / QUANT_K) {
+        const uint ib = gl_WorkGroupID.x * 256 + wgy;
+        if (ib >= p.M * p.K / QUANT_K) {
             return;
         }
 
@@ -19,40 +19,52 @@ void main() {
         const uint ir = tid % 16;
         const uint is = 2 * il;
 
-        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[i].d.x);
-        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[i].d.y);
+        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib].d.x);
+        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib].d.y);
 
-        const uint y_idx = i * QUANT_K + 64 * il + 2 * ir;
+        const uint y_idx = ib * QUANT_K + 64 * il + 2 * ir;
         const uint qs_idx = 32*il + 2 * ir;
         const uint qh_idx = 2 * ir;
 
-        uint8_t sc;
-        uint8_t m;
-        if (is < 4) {
-            sc = uint8_t(data_a[i].scales[is] & 63);
-            m  = uint8_t(data_a[i].scales[is + 4] & 63);
-        } else {
-            sc = uint8_t((data_a[i].scales[is + 4] & 0xF) | ((data_a[i].scales[is - 4] >> 6) << 4));
-            m  = uint8_t((data_a[i].scales[is + 4] >>  4) | ((data_a[i].scales[is    ] >> 6) << 4));
-        }
+        uint scidx0 = (is < 4) ? is : (is + 4);
+        uint scidx1 = (is < 4) ? is : (is - 4);
+        uint scidxmask1 = (is < 4) ? 0x30 : 0xC0;
+        uint scidxshift1 = (is < 4) ? 0 : 2;
+        uint mbidx0 = is + 4;
+        uint mbidx1 = (is < 4) ? is + 4 : is;
+        uint mbidxmask0 = (is < 4) ? 0xF : 0xF0;
+        uint mbidxshift0 = (is < 4) ? 0 : 4;
+        uint mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
+        uint mbidxshift1 = (is < 4) ? 0 : 2;
+
+        uint8_t sc = uint8_t((data_a[ib].scales[scidx0] & 0xF) | ((data_a[ib].scales[scidx1] & scidxmask1) >> scidxshift1));
+        uint8_t mbyte = uint8_t((data_a[ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0 | ((data_a[ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1));
+
         const FLOAT_TYPE d1 = dall * sc;
-        const FLOAT_TYPE m1 = dmin * m;
-
-        if (is < 4) {
-            sc = uint8_t(data_a[i].scales[is + 1] & 63);
-            m  = uint8_t(data_a[i].scales[is + 5] & 63);
-        } else {
-            sc = uint8_t((data_a[i].scales[is + 5] & 0xF) | ((data_a[i].scales[is - 3] >> 6) << 4));
-            m  = uint8_t((data_a[i].scales[is + 5] >>  4) | ((data_a[i].scales[is + 1] >> 6) << 4));
-        }
+        const FLOAT_TYPE m1 = dmin * mbyte;
+
+        scidx0 = (is < 4) ? is + 1 : (is + 5);
+        scidx1 = (is < 4) ? is + 1 : (is - 3);
+        scidxmask1 = (is < 4) ? 0x30 : 0xC0;
+        scidxshift1 = (is < 4) ? 0 : 2;
+        mbidx0 = is + 5;
+        mbidx1 = (is < 4) ? is + 5 : is + 1;
+        mbidxmask0 = (is < 4) ? 0xF : 0xF0;
+        mbidxshift0 = (is < 4) ? 0 : 4;
+        mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
+        mbidxshift1 = (is < 4) ? 0 : 2;
+
+        sc = uint8_t((data_a[ib].scales[scidx0] & 0xF) | ((data_a[ib].scales[scidx1] & scidxmask1) >> scidxshift1));
+        mbyte = uint8_t((data_a[ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0 | ((data_a[ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1));
+
         const FLOAT_TYPE d2 = dall * sc;
-        const FLOAT_TYPE m2 = dmin * m;
+        const FLOAT_TYPE m2 = dmin * mbyte;
 
         const uint8_t hm1 = uint8_t(1 << (2 * il    ));
         const uint8_t hm2 = uint8_t(1 << (2 * il + 1));
-        data_b[y_idx     ] = D_TYPE(d1 * FLOAT_TYPE((data_a[i].qs[qs_idx    ] & 0xF) + (((data_a[i].qh[qh_idx    ] & hm1) != 0) ? 16 : 0)) - m1);
-        data_b[y_idx +  1] = D_TYPE(d1 * FLOAT_TYPE((data_a[i].qs[qs_idx + 1] & 0xF) + (((data_a[i].qh[qh_idx + 1] & hm1) != 0) ? 16 : 0)) - m1);
-        data_b[y_idx + 32] = D_TYPE(d2 * FLOAT_TYPE((data_a[i].qs[qs_idx    ]  >> 4) + (((data_a[i].qh[qh_idx    ] & hm2) != 0) ? 16 : 0)) - m2);
-        data_b[y_idx + 33] = D_TYPE(d2 * FLOAT_TYPE((data_a[i].qs[qs_idx + 1]  >> 4) + (((data_a[i].qh[qh_idx + 1] & hm2) != 0) ? 16 : 0)) - m2);
+        data_b[y_idx     ] = D_TYPE(d1 * FLOAT_TYPE((data_a[ib].qs[qs_idx    ] & 0xF) + (((data_a[ib].qh[qh_idx    ] & hm1) != 0) ? 16 : 0)) - m1);
+        data_b[y_idx +  1] = D_TYPE(d1 * FLOAT_TYPE((data_a[ib].qs[qs_idx + 1] & 0xF) + (((data_a[ib].qh[qh_idx + 1] & hm1) != 0) ? 16 : 0)) - m1);
+        data_b[y_idx + 32] = D_TYPE(d2 * FLOAT_TYPE((data_a[ib].qs[qs_idx    ]  >> 4) + (((data_a[ib].qh[qh_idx    ] & hm2) != 0) ? 16 : 0)) - m2);
+        data_b[y_idx + 33] = D_TYPE(d2 * FLOAT_TYPE((data_a[ib].qs[qs_idx + 1]  >> 4) + (((data_a[ib].qh[qh_idx + 1] & hm2) != 0) ? 16 : 0)) - m2);
     }
 }