From 251642b374e1e01334d2a92bd31545dc5d5383f3 Mon Sep 17 00:00:00 2001 From: lovedheart <6277001+lovedheart@users.noreply.github.com> Date: Sun, 14 Dec 2025 14:30:32 +0100 Subject: [PATCH 1/2] Some improvement on mul_mat_iq2_xs Refactor calculations for db values and grid data to optimize performance and reduce redundancy. --- .../vulkan-shaders/mul_mat_vec_iq2_xs.comp | 66 ++++++++++++------- 1 file changed, 42 insertions(+), 24 deletions(-) diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xs.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xs.comp index 0cd906dbbf..5fbec7d6b9 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xs.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xs.comp @@ -11,36 +11,54 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint y_idx = i * QUANT_K + 16 * itid; const uint nibble_shift = 4 * (itid & 1); const uint ib32 = itid / 2; // 0..7 - uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i; + // Precompute db multiplication factors + float db_vals[NUM_ROWS]; [[unroll]] for (uint n = 0; n < num_rows; ++n) { const float d = float(data_a[ibi].d); - const uint scale = (data_a[ibi].scales[ib32] >> nibble_shift) & 0xF; - const float db = d * (0.5 + scale) * 0.25; - + const uint scale_raw = data_a[ibi].scales[ib32]; + const uint scale = (scale_raw >> nibble_shift) & 0xF; + // Merge constant calculations d * (0.5 + scale) * 0.25 = d*0.125 + d*scale*0.25 + db_vals[n] = d * (0.125f + float(scale) * 0.25f); + ibi += num_blocks_per_row; + } + ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i; + [[unroll]] for (uint n = 0; n < num_rows; ++n) { + // Preload grid and sign data for all l values + vec4 grid0_vals[2], grid1_vals[2]; + uint sign_vals[2], sign7_vals[2]; [[unroll]] for (uint l = 0; l < 2; ++l) { const uint qs = data_a[ibi].qs[2 * itid + l]; - const uint sign = qs >> 9; - const uint sign7 = bitCount(sign); - const vec4 grid0 = vec4(unpack8(iq2xs_grid[qs & 511].x)); - const vec4 grid1 = vec4(unpack8(iq2xs_grid[qs & 511].y)); - - [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { - vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]); - vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]); - - FLOAT_TYPE sum = - fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign & 1) != 0 ? -grid0.x : grid0.x), - fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign & 2) != 0 ? -grid0.y : grid0.y), - fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign & 4) != 0 ? -grid0.z : grid0.z), - fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign & 8) != 0 ? -grid0.w : grid0.w), - fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign & 16) != 0 ? -grid1.x : grid1.x), - fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign & 32) != 0 ? -grid1.y : grid1.y), - fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign & 64) != 0 ? -grid1.z : grid1.z), - fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign7 & 1) != 0 ? -grid1.w : grid1.w), - FLOAT_TYPE(0.0))))))))); - temp[j][n] = fma(db, sum, temp[j][n]); + sign_vals[l] = qs >> 9; + sign7_vals[l] = bitCount(sign_vals[l]); + const uvec2 grid_data = iq2xs_grid[qs & 511]; + grid0_vals[l] = vec4(unpack8(grid_data.x)); + grid1_vals[l] = vec4(unpack8(grid_data.y)); + } + // Preload B data for all j columns (reduce repeated index calculations) + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + FLOAT_TYPE sum = FLOAT_TYPE(0.0); + [[unroll]] for (uint l = 0; l < 2; ++l) { + const uint sign = sign_vals[l]; + const uint sign7 = sign7_vals[l]; + const vec4 grid0 = grid0_vals[l]; + const vec4 grid1 = grid1_vals[l]; + // Precompute indices + const uint b_idx = (j * p.batch_stride_b + b_offset + y_idx) / 4 + 2 * l; + const vec4 b0 = vec4(data_b_v4[b_idx + 0]); + const vec4 b4 = vec4(data_b_v4[b_idx + 1]); + sum += + fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign & 1) != 0 ? -grid0.x : grid0.x), + fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign & 2) != 0 ? -grid0.y : grid0.y), + fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign & 4) != 0 ? -grid0.z : grid0.z), + fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign & 8) != 0 ? -grid0.w : grid0.w), + fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign & 16) != 0 ? -grid1.x : grid1.x), + fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign & 32) != 0 ? -grid1.y : grid1.y), + fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign & 64) != 0 ? -grid1.z : grid1.z), + fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign7 & 1) != 0 ? -grid1.w : grid1.w), + FLOAT_TYPE(0.0))))))))); } + temp[j][n] = fma(FLOAT_TYPE(db_vals[n]), sum, temp[j][n]); } ibi += num_blocks_per_row; } From 31342c22d6b8c7d631bf6dca36e53c9d4289b16f Mon Sep 17 00:00:00 2001 From: lovedheart <6277001+lovedheart@users.noreply.github.com> Date: Mon, 15 Dec 2025 10:38:42 +0100 Subject: [PATCH 2/2] Fix trailing whitespace --- ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xs.comp | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xs.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xs.comp index 5fbec7d6b9..7ec2e04f58 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xs.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xs.comp @@ -34,15 +34,15 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uvec2 grid_data = iq2xs_grid[qs & 511]; grid0_vals[l] = vec4(unpack8(grid_data.x)); grid1_vals[l] = vec4(unpack8(grid_data.y)); - } + } // Preload B data for all j columns (reduce repeated index calculations) [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { - FLOAT_TYPE sum = FLOAT_TYPE(0.0); + FLOAT_TYPE sum = FLOAT_TYPE(0.0); [[unroll]] for (uint l = 0; l < 2; ++l) { const uint sign = sign_vals[l]; const uint sign7 = sign7_vals[l]; const vec4 grid0 = grid0_vals[l]; - const vec4 grid1 = grid1_vals[l]; + const vec4 grid1 = grid1_vals[l]; // Precompute indices const uint b_idx = (j * p.batch_stride_b + b_offset + y_idx) / 4 + 2 * l; const vec4 b0 = vec4(data_b_v4[b_idx + 0]);