From fa9e69fdd6ee0a2f0e8bb39029121310ccb1e58d Mon Sep 17 00:00:00 2001 From: Iwan Kawrakow Date: Thu, 21 Aug 2025 14:04:18 +0300 Subject: [PATCH] Use bperm trick for iq3_k -> 8% PP performance gain --- .../mmq-instance-iq3_k_r4.cu | 34 ++++++++----------- 1 file changed, 14 insertions(+), 20 deletions(-) diff --git a/ggml/src/ggml-cuda/template-instances/mmq-instance-iq3_k_r4.cu b/ggml/src/ggml-cuda/template-instances/mmq-instance-iq3_k_r4.cu index a588969f..b828055c 100644 --- a/ggml/src/ggml-cuda/template-instances/mmq-instance-iq3_k_r4.cu +++ b/ggml/src/ggml-cuda/template-instances/mmq-instance-iq3_k_r4.cu @@ -16,8 +16,6 @@ template static __device__ __forceinlin const int kqsx = threadIdx.x/4; // 0...7 -> block of 32 - uint32_t aux32[4]; - const uint8_t * aux8 = (const uint8_t *)aux32; #pragma unroll for (int i0 = 0; i0 < mmq_y; i0 += 4*nwarps) { int i = i0 + 4*threadIdx.y + threadIdx.x%4; @@ -37,29 +35,25 @@ template static __device__ __forceinlin #pragma unroll for (int l = 0; l < 2; ++l) { - auto values_l = iq3k_table + (((bxi->extra[ir+4*l] >> kqsx) & 1) << 6); + //auto values_l = iq3k_table + (((bxi->extra[ir+4*l] >> kqsx) & 1) << 6); + uint32_t extra32 = uint32_t((bxi->extra[ir+4*l] >> kqsx) & 1) * 0x88888888; const int ql = get_int_b4(bxi->qs, 8*kqsx + ir + 4*l); - aux32[0] = ((ql >> 0) & 0x03030303) | ((qh << 2) & 0x04040404); - aux32[1] = ((ql >> 2) & 0x03030303) | ((qh << 1) & 0x04040404); - aux32[2] = ((ql >> 4) & 0x03030303) | ((qh >> 0) & 0x04040404); - aux32[3] = ((ql >> 6) & 0x03030303) | ((qh >> 1) & 0x04040404); - - int val0 = int_from_table_2(aux8+ 0, values_l); - int val1 = int_from_table_2(aux8+ 4, values_l); - int val2 = int_from_table_2(aux8+ 8, values_l); - int val3 = int_from_table_2(aux8+12, values_l); + uint32_t val1 = ((ql >> 0) & 0x33333333) | extra32 | ((qh << 2) & 0x04040404) | ((qh << 4) & 0x40404040); + uint32_t val2 = ((ql >> 2) & 0x33333333) | extra32 | ((qh << 1) & 0x04040404) | ((qh << 3) & 0x40404040); + int2 v1 = get_int_from_table_16(val1, iq3nl_values); + int2 v2 = get_int_from_table_16(val2, iq3nl_values); #ifdef INT8_MMA_AVAILABLE - x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + 4*l + 0] = val0; - x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + 4*l + 1] = val1; - x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + 4*l + 2] = val2; - x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + 4*l + 3] = val3; + x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + 4*l + 0] = v1.x; + x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + 4*l + 1] = v2.x; + x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + 4*l + 2] = v1.y; + x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + 4*l + 3] = v2.y; #else - x_qs[i*(2*WARP_SIZE + 1) + 8*kqsx + 4*l + 0] = val0; - x_qs[i*(2*WARP_SIZE + 1) + 8*kqsx + 4*l + 1] = val1; - x_qs[i*(2*WARP_SIZE + 1) + 8*kqsx + 4*l + 2] = val2; - x_qs[i*(2*WARP_SIZE + 1) + 8*kqsx + 4*l + 3] = val3; + x_qs[i*(2*WARP_SIZE + 1) + 8*kqsx + 4*l + 0] = v1.x; + x_qs[i*(2*WARP_SIZE + 1) + 8*kqsx + 4*l + 1] = v2.x; + x_qs[i*(2*WARP_SIZE + 1) + 8*kqsx + 4*l + 2] = v1.y; + x_qs[i*(2*WARP_SIZE + 1) + 8*kqsx + 4*l + 3] = v2.y; #endif // INT8_MMA_AVAILABLE qh >>= 4;