vulkan: optimize Q2_K and Q3_K mul_mat_vec (llama/10459)

diff --git a/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp b/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp
index e2625d32b902d50b11d72f264ec4b368bebfa218..fcf02210ea6c7ebe4ca68174e8217dfbc8a6d52d 100644 (file)
--- a/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp
+++ b/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp
@@ -1,4 +1,5 @@
 #version 450
+#extension GL_EXT_shader_explicit_arithmetic_types : require
 
 #include "mul_mat_vec_base.comp"
 
@@ -32,38 +33,67 @@ void main() {
     const uint s_offset = 8*v_im;
     const uint y_offset = 128*v_im + l0;
 
-    tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
+    FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
 
     [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
         const uint y_idx = i * QUANT_K + y_offset;
 
-        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib0 + i].d.x);
-        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib0 + i].d.y);
+        f16vec2 d = data_a[ib0 + i].d;
+        const FLOAT_TYPE dall = d.x;
+        const FLOAT_TYPE dmin = d.y;
+
+        B_TYPE_VEC2 b0 = data_b_v2[(b_offset + y_idx) / 2 + 0];
+        B_TYPE_VEC2 b16 = data_b_v2[(b_offset + y_idx) / 2 + 8];
+        B_TYPE_VEC2 b32 = data_b_v2[(b_offset + y_idx) / 2 + 16];
+        B_TYPE_VEC2 b48 = data_b_v2[(b_offset + y_idx) / 2 + 24];
+        B_TYPE_VEC2 b64 = data_b_v2[(b_offset + y_idx) / 2 + 32];
+        B_TYPE_VEC2 b80 = data_b_v2[(b_offset + y_idx) / 2 + 40];
+        B_TYPE_VEC2 b96 = data_b_v2[(b_offset + y_idx) / 2 + 48];
+        B_TYPE_VEC2 b112 = data_b_v2[(b_offset + y_idx) / 2 + 56];
+
+        uint32_t s0_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 0];
+        uint32_t s4_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 1];
+
+        uint32_t s0_lo4_u32 = s0_u32 & 0x0F0F0F0F;
+        uint32_t s0_hi4_u32 = (s0_u32 >> 4) & 0x0F0F0F0F;
+        uint32_t s4_lo4_u32 = s4_u32 & 0x0F0F0F0F;
+        uint32_t s4_hi4_u32 = (s4_u32 >> 4) & 0x0F0F0F0F;
+
+        uvec4 s0_lo4 = uvec4(unpack8(s0_lo4_u32));
+        uvec4 s4_lo4 = uvec4(unpack8(s4_lo4_u32));
+        uvec4 s0_hi4 = uvec4(unpack8(s0_hi4_u32));
+        uvec4 s4_hi4 = uvec4(unpack8(s4_hi4_u32));
+
+        uint16_t qs0_u16 = data_a_packed16[ib0 + i].qs[q_offset / 2 + 0];
+        uint16_t qs16_u16 = data_a_packed16[ib0 + i].qs[q_offset / 2 + 8];
+        uvec2 qs0 =  uvec2(unpack8(qs0_u16));
+        uvec2 qs16 = uvec2(unpack8(qs16_u16));
 
         FLOAT_TYPE sum1 = FLOAT_TYPE(0.0);
         FLOAT_TYPE sum2 = FLOAT_TYPE(0.0);
-        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
-            sum1 = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +  0]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 0) & 3),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 1] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 0) & 3),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 2) & 3),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 3] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 2) & 3),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 4) & 3),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 5] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 4) & 3),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 6) & 3),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +112]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 7] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 6) & 3), sum1))))))));
-            sum2 = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +  0]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 0] >> 4) & 0xF),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 1] >> 4) & 0xF),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 2] >> 4) & 0xF),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 3] >> 4) & 0xF),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 4] >> 4) & 0xF),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 5] >> 4) & 0xF),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 6] >> 4) & 0xF),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +112]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 7] >> 4) & 0xF), sum2))))))));
+        [[unroll]] for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
+            sum1 = fma(FLOAT_TYPE(b0[l]),   FLOAT_TYPE(s0_lo4[0]) * FLOAT_TYPE((qs0[l]  >> 0) & 3),
+                   fma(FLOAT_TYPE(b16[l]),  FLOAT_TYPE(s0_lo4[1]) * FLOAT_TYPE((qs16[l] >> 0) & 3),
+                   fma(FLOAT_TYPE(b32[l]),  FLOAT_TYPE(s0_lo4[2]) * FLOAT_TYPE((qs0[l]  >> 2) & 3),
+                   fma(FLOAT_TYPE(b48[l]),  FLOAT_TYPE(s0_lo4[3]) * FLOAT_TYPE((qs16[l] >> 2) & 3),
+                   fma(FLOAT_TYPE(b64[l]),  FLOAT_TYPE(s4_lo4[0]) * FLOAT_TYPE((qs0[l]  >> 4) & 3),
+                   fma(FLOAT_TYPE(b80[l]),  FLOAT_TYPE(s4_lo4[1]) * FLOAT_TYPE((qs16[l] >> 4) & 3),
+                   fma(FLOAT_TYPE(b96[l]),  FLOAT_TYPE(s4_lo4[2]) * FLOAT_TYPE((qs0[l]  >> 6) & 3),
+                   fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_lo4[3]) * FLOAT_TYPE((qs16[l] >> 6) & 3), sum1))))))));
+            sum2 = fma(FLOAT_TYPE(b0[l]),   FLOAT_TYPE(s0_hi4[0]),
+                   fma(FLOAT_TYPE(b16[l]),  FLOAT_TYPE(s0_hi4[1]),
+                   fma(FLOAT_TYPE(b32[l]),  FLOAT_TYPE(s0_hi4[2]),
+                   fma(FLOAT_TYPE(b48[l]),  FLOAT_TYPE(s0_hi4[3]),
+                   fma(FLOAT_TYPE(b64[l]),  FLOAT_TYPE(s4_hi4[0]),
+                   fma(FLOAT_TYPE(b80[l]),  FLOAT_TYPE(s4_hi4[1]),
+                   fma(FLOAT_TYPE(b96[l]),  FLOAT_TYPE(s4_hi4[2]),
+                   fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_hi4[3]), sum2))))))));
         }
-        const uint tmp_idx = 16 * ix + tid;
-        tmp[tmp_idx] = fma(dall, sum1, fma(-dmin, sum2, tmp[tmp_idx]));
+        temp = fma(dall, sum1, fma(-dmin, sum2, temp));
     }
 
+    tmp[gl_LocalInvocationID.x] = temp;
+
     // sum up partial sums and write back result
     barrier();
     [[unroll]] for (uint s = 16; s > 0; s >>= 1) {

diff --git a/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp b/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp
index a28804533c3205b0a07bcfedca5fcceddf4b0235..723fadde0d52b6033bd2e88ec578140af9d4ea60 100644 (file)
--- a/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp
+++ b/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp
@@ -1,4 +1,5 @@
 #version 450
+#extension GL_EXT_shader_explicit_arithmetic_types : require
 
 #include "mul_mat_vec_base.comp"
 
@@ -33,7 +34,7 @@ void main() {
     const uint q_offset = 32*v_im + l0;
     const uint y_offset = 128*v_im + l0;
 
-    tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
+    FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
 
     const uint s_shift = 4 * v_im;
 
@@ -42,21 +43,44 @@ void main() {
 
         const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
 
+        B_TYPE_VEC2 b0 = data_b_v2[(b_offset + y_idx) / 2 + 0];
+        B_TYPE_VEC2 b16 = data_b_v2[(b_offset + y_idx) / 2 + 8];
+        B_TYPE_VEC2 b32 = data_b_v2[(b_offset + y_idx) / 2 + 16];
+        B_TYPE_VEC2 b48 = data_b_v2[(b_offset + y_idx) / 2 + 24];
+        B_TYPE_VEC2 b64 = data_b_v2[(b_offset + y_idx) / 2 + 32];
+        B_TYPE_VEC2 b80 = data_b_v2[(b_offset + y_idx) / 2 + 40];
+        B_TYPE_VEC2 b96 = data_b_v2[(b_offset + y_idx) / 2 + 48];
+        B_TYPE_VEC2 b112 = data_b_v2[(b_offset + y_idx) / 2 + 56];
+
+        uint16_t s0_16 = data_a_packed16[ib0 + i].scales[0];
+        uint16_t s2_16 = data_a_packed16[ib0 + i].scales[1];
+        uint16_t s4_16 = data_a_packed16[ib0 + i].scales[2];
+        uint16_t s6_16 = data_a_packed16[ib0 + i].scales[3];
+        uint16_t s8_16 = data_a_packed16[ib0 + i].scales[4];
+        uint16_t s10_16 = data_a_packed16[ib0 + i].scales[5];
+        u8vec2 s0 = unpack8(s0_16);
+        u8vec2 s2 = unpack8(s2_16);
+        u8vec2 s4 = unpack8(s4_16);
+        u8vec2 s6 = unpack8(s6_16);
+        u8vec2 s8 = unpack8(s8_16);
+        u8vec2 s10 = unpack8(s10_16);
+
         FLOAT_TYPE sum = FLOAT_TYPE(0.0);
         for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
-            sum = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +  0]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[0] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 8] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 0)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[2] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[10] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 1)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[4] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 8] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 2)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[6] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[10] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 3)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[1] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 9] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[3] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[11] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[5] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 9] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +112]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[7] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[11] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4)), sum))))))));
+            sum = fma(FLOAT_TYPE(b0[l])   * FLOAT_TYPE(int8_t(((s0[0] >> s_shift) & 0xF) | ((s8[0]  >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 0)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(b32[l])  * FLOAT_TYPE(int8_t(((s2[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 1)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(b64[l])  * FLOAT_TYPE(int8_t(((s4[0] >> s_shift) & 0xF) | ((s8[0]  >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 2)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(b96[l])  * FLOAT_TYPE(int8_t(((s6[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 3)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(b16[l])  * FLOAT_TYPE(int8_t(((s0[1] >> s_shift) & 0xF) | ((s8[1]  >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(b48[l])  * FLOAT_TYPE(int8_t(((s2[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(b80[l])  * FLOAT_TYPE(int8_t(((s4[1] >> s_shift) & 0xF) | ((s8[1]  >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(b112[l]) * FLOAT_TYPE(int8_t(((s6[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4)), sum))))))));
         }
-        const uint tmp_idx = 16 * ix + tid;
-        tmp[tmp_idx] = fma(d, sum, tmp[tmp_idx]);
+        temp = fma(d, sum, temp);
     }
 
+    tmp[gl_LocalInvocationID.x] = temp;
+
     // sum up partial sums and write back result
     barrier();
     [[unroll]] for (uint s = 16; s > 0; s >>= 1) {