remove ifdefs

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

@@ -15,48 +15,54 @@ layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     return vec2(data_a[a_offset + ib], data_a[a_offset + ib + 1]);
 }
+vec2 get_dm(uint ib, uint a_offset) {
+    return vec2(0, 0);
+}
 #endif
 
 #if defined(DATA_A_F16)
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     return vec2(data_a[a_offset + ib], data_a[a_offset + ib + 1]);
 }
+vec2 get_dm(uint ib, uint a_offset) {
+    return vec2(0, 0);
+}
 #endif
 
 #if defined(DATA_A_Q4_0)
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
-    const float d = float(data_a[a_offset + ib].d);
     const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
-    return (vec2(vui & 0xF, vui >> 4) - 8.0f) * d;
+    return (vec2(vui & 0xF, vui >> 4) - 8.0f);
 }
 vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
     const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]);
     return (vec4(vui & 0xF, (vui >> 4) & 0xF, (vui >> 8) & 0xF, vui >> 12) - 8.0f);
 }
+vec2 get_dm(uint ib, uint a_offset) {
+    return vec2(float(data_a[a_offset + ib].d), 0);
+}
 #endif
 
 #if defined(DATA_A_Q4_1)
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
-    const float d = float(data_a[a_offset + ib].d);
-    const float m = float(data_a[a_offset + ib].m);
     const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
-    return vec2(vui & 0xF, vui >> 4) * d + m;
+    return vec2(vui & 0xF, vui >> 4);
 }
 vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
-    const float d = float(data_a_packed16[a_offset + ib].d);
-    const float m = float(data_a_packed16[a_offset + ib].m);
     const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]);
-    return vec4(vui & 0xF, (vui >> 4) & 0xF, (vui >> 8) & 0xF, vui >> 12) * d + m;
+    return vec4(vui & 0xF, (vui >> 4) & 0xF, (vui >> 8) & 0xF, vui >> 12);
+}
+vec2 get_dm(uint ib, uint a_offset) {
+    return vec2(float(data_a[a_offset + ib].d), float(data_a[a_offset + ib].m));
 }
 #endif
 
 #if defined(DATA_A_Q5_0)
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
-    const float d = float(data_a[a_offset + ib].d);
     const uint uint_qh = uint(data_a[a_offset + ib].qh[1]) << 16 | data_a[a_offset + ib].qh[0];
     const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
     const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
-    return (vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) - 16.0f) * d;
+    return (vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) - 16.0f);
 }
 vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
     const uint uint_qh = uint(data_a_packed16[a_offset + ib].qh[1]) << 16 | data_a_packed16[a_offset + ib].qh[0];
@@ -65,48 +71,54 @@ vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
     const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]);
     return (vec4(((vui >> 0) & 0xF) | qh0.x, ((vui >> 4) & 0xF) | qh0.y, ((vui >> 8) & 0xF) | qh1.x, (vui >> 12) | qh1.y) - 16.0f);
 }
+vec2 get_dm(uint ib, uint a_offset) {
+    return vec2(float(data_a[a_offset + ib].d), 0);
+}
 #endif
 
 #if defined(DATA_A_Q5_1)
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
-    const float d = float(data_a[a_offset + ib].d);
-    const float m = float(data_a[a_offset + ib].m);
     const uint uint_qh = data_a[a_offset + ib].qh;
     const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
     const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
-    return vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) * d + m;
+    return vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y);
 }
 vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
-    const float d = float(data_a_packed16[a_offset + ib].d);
-    const float m = float(data_a_packed16[a_offset + ib].m);
     const uint uint_qh = data_a_packed16[a_offset + ib].qh;
     const ivec2 qh0 = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
     const ivec2 qh1 = ivec2(((uint_qh >> (iqs + 1)) << 4) & 0x10, (uint_qh >> (iqs + 13)) & 0x10);
     const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]);
-    return vec4(((vui >> 0) & 0xF) | qh0.x, ((vui >> 4) & 0xF) | qh0.y, ((vui >> 8) & 0xF) | qh1.x, (vui >> 12) | qh1.y) * d + m;
+    return vec4(((vui >> 0) & 0xF) | qh0.x, ((vui >> 4) & 0xF) | qh0.y, ((vui >> 8) & 0xF) | qh1.x, (vui >> 12) | qh1.y);
+}
+vec2 get_dm(uint ib, uint a_offset) {
+    return vec2(float(data_a[a_offset + ib].d), float(data_a[a_offset + ib].m));
 }
 #endif
 
 #if defined(DATA_A_Q8_0)
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
-    const float d = float(data_a[a_offset + ib].d);
+    return vec2(int(data_a[a_offset + ib].qs[iqs]), int(data_a[a_offset + ib].qs[iqs + 1]));
-    return vec2(int(data_a[a_offset + ib].qs[iqs]), int(data_a[a_offset + ib].qs[iqs + 1])) * d;
 }
 vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
     uint32_t v0 = data_a_packed16[a_offset + ib].qs[iqs/2];
     uint32_t v1 = data_a_packed16[a_offset + ib].qs[iqs/2 + 1];
     return vec4(int8_t(v0 & 0xFF), int8_t(v0 >> 8), int8_t(v1 & 0xFF), int8_t(v1 >> 8));
 }
+vec2 get_dm(uint ib, uint a_offset) {
+    return vec2(float(data_a[a_offset + ib].d), 0);
+}
 #endif
 
 #if defined(DATA_A_IQ4_NL)
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
-    const float d = float(data_a[a_offset + ib].d);
     const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
-    return vec2(kvalues_iq4nl[vui & 0xF], kvalues_iq4nl[vui >> 4]) * d;
+    return vec2(kvalues_iq4nl[vui & 0xF], kvalues_iq4nl[vui >> 4]);
 }
 vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
     const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]);
     return vec4(kvalues_iq4nl[vui & 0xF], kvalues_iq4nl[(vui >> 4) & 0xF], kvalues_iq4nl[(vui >> 8) & 0xF], kvalues_iq4nl[vui >> 12]);
 }
+vec2 get_dm(uint ib, uint a_offset) {
+    return vec2(float(data_a[a_offset + ib].d), 0);
+}
 #endif

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

@@ -31,6 +31,8 @@ void main() {
     const uint y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;
 
     vec2 v = dequantize(ib, iqs, 0);
+    const vec2 dm = get_dm(ib, 0);
+    v = v * dm.x + dm.y;
 
     data_d[d_offset + iybs + iqs           ] = D_TYPE(v.x);
     data_d[d_offset + iybs + iqs + y_offset] = D_TYPE(v.y);

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

@@ -59,18 +59,22 @@ void iter(inout FLOAT_TYPE temp[NUM_ROWS], const uint first_row, const uint num_
         ibi += p.ncols;
 
 #if K_PER_ITER == 8
-        const vec4 v = dequantize4(ib, iqs, a_offset);
+        vec4 v = dequantize4(ib, iqs, a_offset);
-        const vec4 v2 = dequantize4(ib, iqs+(4/QUANT_R), a_offset);
+        vec4 v2 = dequantize4(ib, iqs+(4/QUANT_R), a_offset);
-        FLOAT_TYPE rowtmp = 0;
+
+        const vec2 dm = get_dm(ib, a_offset);
+        if (dm.y != 0) { // quant has min component
+            v = v * dm.x + dm.y;
+            v2 = v2 * dm.x + dm.y;
+        }
 
         // matrix multiplication
-        rowtmp += dot(bv0, v);
+        FLOAT_TYPE rowtmp = dot(bv0, v);
         rowtmp += dot(bv1, v2);
 
-#if !defined(DATA_A_Q4_1) && !defined(DATA_A_Q5_1)
+        if (dm.y == 0)
-        const float d = float(data_a[a_offset + ib].d);
+            rowtmp *= dm.x;
-        rowtmp *= d;
+
-#endif
         temp[n] += rowtmp;
 #else
         const vec2 v = dequantize(ib, iqs, a_offset);