iq2_k: better CUDA dot product

Almost on par with iq2_xs (168 t/s vs 172 t/s).

1 files changed, 85 insertions, 15 deletions

diff --git a/ggml/src/ggml-cuda/iqk_mmvq.cu b/ggml/src/ggml-cuda/iqk_mmvq.cu
index b3603697..9ed51c88 100644
--- a/ggml/src/ggml-cuda/iqk_mmvq.cu
+++ b/ggml/src/ggml-cuda/iqk_mmvq.cu
@@ -248,7 +248,77 @@ __device__ __forceinline__ float vec_dot_iq5_k_q8_1(
 #define VDR_IQ2_K_Q8_1_MMVQ 4
 #define VDR_IQ2_K_Q8_1_MMQ  4
 
-// TODO
+static const __device__ uint32_t iq2k_table[512] = {
+    0xe1e1e1e1, 0xe1e1e1f3, 0xe1e1e101, 0xe1e1e111, 0xe1e1f3e1, 0xe1e1f3f3, 0xe1e1f301, 0xe1e1f311,
+    0xe1e101e1, 0xe1e101f3, 0xe1e10101, 0xe1e10111, 0xe1e111e1, 0xe1e111f3, 0xe1e11101, 0xe1e11111,
+    0xe1f3e1e1, 0xe1f3e1f3, 0xe1f3e101, 0xe1f3e111, 0xe1f3f3e1, 0xe1f3f3f3, 0xe1f3f301, 0xe1f3f311,
+    0xe1f301e1, 0xe1f301f3, 0xe1f30101, 0xe1f30111, 0xe1f311e1, 0xe1f311f3, 0xe1f31101, 0xe1f31111,
+    0xe101e1e1, 0xe101e1f3, 0xe101e101, 0xe101e111, 0xe101f3e1, 0xe101f3f3, 0xe101f301, 0xe101f311,
+    0xe10101e1, 0xe10101f3, 0xe1010101, 0xe1010111, 0xe10111e1, 0xe10111f3, 0xe1011101, 0xe1011111,
+    0xe111e1e1, 0xe111e1f3, 0xe111e101, 0xe111e111, 0xe111f3e1, 0xe111f3f3, 0xe111f301, 0xe111f311,
+    0xe11101e1, 0xe11101f3, 0xe1110101, 0xe1110111, 0xe11111e1, 0xe11111f3, 0xe1111101, 0xe1111111,
+    0xf3e1e1e1, 0xf3e1e1f3, 0xf3e1e101, 0xf3e1e111, 0xf3e1f3e1, 0xf3e1f3f3, 0xf3e1f301, 0xf3e1f311,
+    0xf3e101e1, 0xf3e101f3, 0xf3e10101, 0xf3e10111, 0xf3e111e1, 0xf3e111f3, 0xf3e11101, 0xf3e11111,
+    0xf3f3e1e1, 0xf3f3e1f3, 0xf3f3e101, 0xf3f3e111, 0xf3f3f3e1, 0xf3f3f3f3, 0xf3f3f301, 0xf3f3f311,
+    0xf3f301e1, 0xf3f301f3, 0xf3f30101, 0xf3f30111, 0xf3f311e1, 0xf3f311f3, 0xf3f31101, 0xf3f31111,
+    0xf301e1e1, 0xf301e1f3, 0xf301e101, 0xf301e111, 0xf301f3e1, 0xf301f3f3, 0xf301f301, 0xf301f311,
+    0xf30101e1, 0xf30101f3, 0xf3010101, 0xf3010111, 0xf30111e1, 0xf30111f3, 0xf3011101, 0xf3011111,
+    0xf311e1e1, 0xf311e1f3, 0xf311e101, 0xf311e111, 0xf311f3e1, 0xf311f3f3, 0xf311f301, 0xf311f311,
+    0xf31101e1, 0xf31101f3, 0xf3110101, 0xf3110111, 0xf31111e1, 0xf31111f3, 0xf3111101, 0xf3111111,
+    0x01e1e1e1, 0x01e1e1f3, 0x01e1e101, 0x01e1e111, 0x01e1f3e1, 0x01e1f3f3, 0x01e1f301, 0x01e1f311,
+    0x01e101e1, 0x01e101f3, 0x01e10101, 0x01e10111, 0x01e111e1, 0x01e111f3, 0x01e11101, 0x01e11111,
+    0x01f3e1e1, 0x01f3e1f3, 0x01f3e101, 0x01f3e111, 0x01f3f3e1, 0x01f3f3f3, 0x01f3f301, 0x01f3f311,
+    0x01f301e1, 0x01f301f3, 0x01f30101, 0x01f30111, 0x01f311e1, 0x01f311f3, 0x01f31101, 0x01f31111,
+    0x0101e1e1, 0x0101e1f3, 0x0101e101, 0x0101e111, 0x0101f3e1, 0x0101f3f3, 0x0101f301, 0x0101f311,
+    0x010101e1, 0x010101f3, 0x01010101, 0x01010111, 0x010111e1, 0x010111f3, 0x01011101, 0x01011111,
+    0x0111e1e1, 0x0111e1f3, 0x0111e101, 0x0111e111, 0x0111f3e1, 0x0111f3f3, 0x0111f301, 0x0111f311,
+    0x011101e1, 0x011101f3, 0x01110101, 0x01110111, 0x011111e1, 0x011111f3, 0x01111101, 0x01111111,
+    0x11e1e1e1, 0x11e1e1f3, 0x11e1e101, 0x11e1e111, 0x11e1f3e1, 0x11e1f3f3, 0x11e1f301, 0x11e1f311,
+    0x11e101e1, 0x11e101f3, 0x11e10101, 0x11e10111, 0x11e111e1, 0x11e111f3, 0x11e11101, 0x11e11111,
+    0x11f3e1e1, 0x11f3e1f3, 0x11f3e101, 0x11f3e111, 0x11f3f3e1, 0x11f3f3f3, 0x11f3f301, 0x11f3f311,
+    0x11f301e1, 0x11f301f3, 0x11f30101, 0x11f30111, 0x11f311e1, 0x11f311f3, 0x11f31101, 0x11f31111,
+    0x1101e1e1, 0x1101e1f3, 0x1101e101, 0x1101e111, 0x1101f3e1, 0x1101f3f3, 0x1101f301, 0x1101f311,
+    0x110101e1, 0x110101f3, 0x11010101, 0x11010111, 0x110111e1, 0x110111f3, 0x11011101, 0x11011111,
+    0x1111e1e1, 0x1111e1f3, 0x1111e101, 0x1111e111, 0x1111f3e1, 0x1111f3f3, 0x1111f301, 0x1111f311,
+    0x111101e1, 0x111101f3, 0x11110101, 0x11110111, 0x111111e1, 0x111111f3, 0x11111101, 0x11111111,
+    0xe6e6e6e6, 0xe6e6e6f8, 0xe6e6e606, 0xe6e6e616, 0xe6e6f8e6, 0xe6e6f8f8, 0xe6e6f806, 0xe6e6f816,
+    0xe6e606e6, 0xe6e606f8, 0xe6e60606, 0xe6e60616, 0xe6e616e6, 0xe6e616f8, 0xe6e61606, 0xe6e61616,
+    0xe6f8e6e6, 0xe6f8e6f8, 0xe6f8e606, 0xe6f8e616, 0xe6f8f8e6, 0xe6f8f8f8, 0xe6f8f806, 0xe6f8f816,
+    0xe6f806e6, 0xe6f806f8, 0xe6f80606, 0xe6f80616, 0xe6f816e6, 0xe6f816f8, 0xe6f81606, 0xe6f81616,
+    0xe606e6e6, 0xe606e6f8, 0xe606e606, 0xe606e616, 0xe606f8e6, 0xe606f8f8, 0xe606f806, 0xe606f816,
+    0xe60606e6, 0xe60606f8, 0xe6060606, 0xe6060616, 0xe60616e6, 0xe60616f8, 0xe6061606, 0xe6061616,
+    0xe616e6e6, 0xe616e6f8, 0xe616e606, 0xe616e616, 0xe616f8e6, 0xe616f8f8, 0xe616f806, 0xe616f816,
+    0xe61606e6, 0xe61606f8, 0xe6160606, 0xe6160616, 0xe61616e6, 0xe61616f8, 0xe6161606, 0xe6161616,
+    0xf8e6e6e6, 0xf8e6e6f8, 0xf8e6e606, 0xf8e6e616, 0xf8e6f8e6, 0xf8e6f8f8, 0xf8e6f806, 0xf8e6f816,
+    0xf8e606e6, 0xf8e606f8, 0xf8e60606, 0xf8e60616, 0xf8e616e6, 0xf8e616f8, 0xf8e61606, 0xf8e61616,
+    0xf8f8e6e6, 0xf8f8e6f8, 0xf8f8e606, 0xf8f8e616, 0xf8f8f8e6, 0xf8f8f8f8, 0xf8f8f806, 0xf8f8f816,
+    0xf8f806e6, 0xf8f806f8, 0xf8f80606, 0xf8f80616, 0xf8f816e6, 0xf8f816f8, 0xf8f81606, 0xf8f81616,
+    0xf806e6e6, 0xf806e6f8, 0xf806e606, 0xf806e616, 0xf806f8e6, 0xf806f8f8, 0xf806f806, 0xf806f816,
+    0xf80606e6, 0xf80606f8, 0xf8060606, 0xf8060616, 0xf80616e6, 0xf80616f8, 0xf8061606, 0xf8061616,
+    0xf816e6e6, 0xf816e6f8, 0xf816e606, 0xf816e616, 0xf816f8e6, 0xf816f8f8, 0xf816f806, 0xf816f816,
+    0xf81606e6, 0xf81606f8, 0xf8160606, 0xf8160616, 0xf81616e6, 0xf81616f8, 0xf8161606, 0xf8161616,
+    0x06e6e6e6, 0x06e6e6f8, 0x06e6e606, 0x06e6e616, 0x06e6f8e6, 0x06e6f8f8, 0x06e6f806, 0x06e6f816,
+    0x06e606e6, 0x06e606f8, 0x06e60606, 0x06e60616, 0x06e616e6, 0x06e616f8, 0x06e61606, 0x06e61616,
+    0x06f8e6e6, 0x06f8e6f8, 0x06f8e606, 0x06f8e616, 0x06f8f8e6, 0x06f8f8f8, 0x06f8f806, 0x06f8f816,
+    0x06f806e6, 0x06f806f8, 0x06f80606, 0x06f80616, 0x06f816e6, 0x06f816f8, 0x06f81606, 0x06f81616,
+    0x0606e6e6, 0x0606e6f8, 0x0606e606, 0x0606e616, 0x0606f8e6, 0x0606f8f8, 0x0606f806, 0x0606f816,
+    0x060606e6, 0x060606f8, 0x06060606, 0x06060616, 0x060616e6, 0x060616f8, 0x06061606, 0x06061616,
+    0x0616e6e6, 0x0616e6f8, 0x0616e606, 0x0616e616, 0x0616f8e6, 0x0616f8f8, 0x0616f806, 0x0616f816,
+    0x061606e6, 0x061606f8, 0x06160606, 0x06160616, 0x061616e6, 0x061616f8, 0x06161606, 0x06161616,
+    0x16e6e6e6, 0x16e6e6f8, 0x16e6e606, 0x16e6e616, 0x16e6f8e6, 0x16e6f8f8, 0x16e6f806, 0x16e6f816,
+    0x16e606e6, 0x16e606f8, 0x16e60606, 0x16e60616, 0x16e616e6, 0x16e616f8, 0x16e61606, 0x16e61616,
+    0x16f8e6e6, 0x16f8e6f8, 0x16f8e606, 0x16f8e616, 0x16f8f8e6, 0x16f8f8f8, 0x16f8f806, 0x16f8f816,
+    0x16f806e6, 0x16f806f8, 0x16f80606, 0x16f80616, 0x16f816e6, 0x16f816f8, 0x16f81606, 0x16f81616,
+    0x1606e6e6, 0x1606e6f8, 0x1606e606, 0x1606e616, 0x1606f8e6, 0x1606f8f8, 0x1606f806, 0x1606f816,
+    0x160606e6, 0x160606f8, 0x16060606, 0x16060616, 0x160616e6, 0x160616f8, 0x16061606, 0x16061616,
+    0x1616e6e6, 0x1616e6f8, 0x1616e606, 0x1616e616, 0x1616f8e6, 0x1616f8f8, 0x1616f806, 0x1616f816,
+    0x161606e6, 0x161606f8, 0x16160606, 0x16160616, 0x161616e6, 0x161616f8, 0x16161606, 0x16161616,
+};
+
+__device__ __forceinline__ int int_from_table_4(const uint8_t * a8, const int * values) {
+    return values[a8[0] | (a8[1] << 2) | (a8[2] << 4) | (a8[3] << 6)];
+}
+
 __device__ __forceinline__ float vec_dot_iq2_k_q8_1(
     const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
 
@@ -265,8 +335,8 @@ __device__ __forceinline__ float vec_dot_iq2_k_q8_1(
     const uint32_t * q2 = (const uint32_t *)bq2->qs + 8*(i4/4) + 2*(i4%4);
     const uint16_t extra = bq2->extra >> (8*(i4/4) + (i4%4)/2);
 
-    const uint8_t * all_values = (const uint8_t *)iq2nl_values;
-    const uint8_t * values;
+    const int * all_values = (const int *)iq2k_table;
+    const int * values;
 
     uint32_t val1 = q2[0], val2 = q2[1];
 
@@ -278,27 +348,27 @@ __device__ __forceinline__ float vec_dot_iq2_k_q8_1(
     // -> scales_l[4*(i4/4) + k] >> 4*(((i4%4)/2)%2)
 
     ls = (bq2->scales[4*(i4/4) + 0] >> 4*(((i4%4)/2)%2)) & 0xf;
-    aux32[0] = ((val1 >> 0) & 0x03030303); aux32[1] = ((val2 >> 0) & 0x03030303); values = all_values + ((extra & 0x01) << 2);
-    v1 = int_from_table(a8 + 0, values);
-    v2 = int_from_table(a8 + 4, values);
+    aux32[0] = ((val1 >> 0) & 0x03030303); aux32[1] = ((val2 >> 0) & 0x03030303); values = all_values + ((extra & 0x01) << 8);
+    v1 = int_from_table_4(a8 + 0, values);
+    v2 = int_from_table_4(a8 + 4, values);
     int sumi1 = ggml_cuda_dp4a(v2, q8_1[1], ggml_cuda_dp4a(v1, q8_1[0], 0)) * (2*ls - 15);
 
     ls = (bq2->scales[4*(i4/4) + 1] >> 4*(((i4%4)/2)%2)) & 0xf;
-    aux32[0] = ((val1 >> 2) & 0x03030303); aux32[1] = ((val2 >> 2) & 0x03030303); values = all_values + ((extra & 0x04) << 0);
-    v1 = int_from_table(a8 + 0, values);
-    v2 = int_from_table(a8 + 4, values);
+    aux32[0] = ((val1 >> 2) & 0x03030303); aux32[1] = ((val2 >> 2) & 0x03030303); values = all_values + ((extra & 0x04) << 6);
+    v1 = int_from_table_4(a8 + 0, values);
+    v2 = int_from_table_4(a8 + 4, values);
     int sumi2 = ggml_cuda_dp4a(v2, q8_2[1], ggml_cuda_dp4a(v1, q8_2[0], 0)) * (2*ls - 15);
 
     ls = (bq2->scales[4*(i4/4) + 2] >> 4*(((i4%4)/2)%2)) & 0xf;
-    aux32[0] = ((val1 >> 4) & 0x03030303); aux32[1] = ((val2 >> 4) & 0x03030303); values = all_values + ((extra & 0x10) >> 2);
-    v1 = int_from_table(a8 + 0, values);
-    v2 = int_from_table(a8 + 4, values);
+    aux32[0] = ((val1 >> 4) & 0x03030303); aux32[1] = ((val2 >> 4) & 0x03030303); values = all_values + ((extra & 0x10) << 4);
+    v1 = int_from_table_4(a8 + 0, values);
+    v2 = int_from_table_4(a8 + 4, values);
     int sumi3 = ggml_cuda_dp4a(v2, q8_3[1], ggml_cuda_dp4a(v1, q8_3[0], 0)) * (2*ls - 15);
 
     ls = (bq2->scales[4*(i4/4) + 3] >> 4*(((i4%4)/2)%2)) & 0xf;
-    aux32[0] = ((val1 >> 6) & 0x03030303); aux32[1] = ((val2 >> 6) & 0x03030303); values = all_values + ((extra & 0x40) >> 4);
-    v1 = int_from_table(a8 + 0, values);
-    v2 = int_from_table(a8 + 4, values);
+    aux32[0] = ((val1 >> 6) & 0x03030303); aux32[1] = ((val2 >> 6) & 0x03030303); values = all_values + ((extra & 0x40) << 2);
+    v1 = int_from_table_4(a8 + 0, values);
+    v2 = int_from_table_4(a8 + 4, values);
     int sumi4 = ggml_cuda_dp4a(v2, q8_4[1], ggml_cuda_dp4a(v1, q8_4[0], 0)) * (2*ls - 15);
 
     return __half2float(bq2->d) * (__low2float(bq8_1[4*(i4/4)+0].ds) * sumi1