From b4ae7005e66cb03d8de601dd271b10a1127970c4 Mon Sep 17 00:00:00 2001 From: Eve <139727413+netrunnereve@users.noreply.github.com> Date: Sun, 5 Jan 2025 21:59:43 -0500 Subject: [PATCH] unpack should be u16, add vim swap to gitignore (about time) --- .gitignore | 1 + .../ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp | 12 ++++++------ .../ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp | 12 ++++++------ 3 files changed, 13 insertions(+), 12 deletions(-) diff --git a/.gitignore b/.gitignore index 1df7cf4a1..694f36e04 100644 --- a/.gitignore +++ b/.gitignore @@ -18,6 +18,7 @@ *.metallib *.o *.so +*.swp *.tmp # IDE / OS diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp index 8a3644e96..940ba3f51 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp @@ -47,12 +47,12 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { const FLOAT_TYPE dall = FLOAT_TYPE(d.x); const FLOAT_TYPE dmin = FLOAT_TYPE(d.y); - uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im ]; - uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2]; - uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4]; - uvec2 scale0 = uvec2(unpack8(scale0_u32)); - uvec2 scale4 = uvec2(unpack8(scale4_u32)); - uvec2 scale8 = uvec2(unpack8(scale8_u32)); + uint16_t scale0_u16 = data_a_packed16[ib0 + i].scales[v_im ]; + uint16_t scale4_u16 = data_a_packed16[ib0 + i].scales[v_im + 2]; + uint16_t scale8_u16 = data_a_packed16[ib0 + i].scales[v_im + 4]; + uvec2 scale0 = uvec2(unpack8(scale0_u16)); + uvec2 scale4 = uvec2(unpack8(scale4_u16)); + uvec2 scale8 = uvec2(unpack8(scale8_u16)); const uint32_t sc0 = ( scale0.x & 0x3f); const uint32_t sc1 = ( scale0.y & 0x3f); diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp index bea1dc860..c4aaf9fea 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp @@ -46,12 +46,12 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { const FLOAT_TYPE dall = FLOAT_TYPE(d.x); const FLOAT_TYPE dmin = FLOAT_TYPE(d.y); - uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im ]; - uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2]; - uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4]; - uvec2 scale0 = uvec2(unpack8(scale0_u32)); - uvec2 scale4 = uvec2(unpack8(scale4_u32)); - uvec2 scale8 = uvec2(unpack8(scale8_u32)); + uint16_t scale0_u16 = data_a_packed16[ib0 + i].scales[v_im ]; + uint16_t scale4_u16 = data_a_packed16[ib0 + i].scales[v_im + 2]; + uint16_t scale8_u16 = data_a_packed16[ib0 + i].scales[v_im + 4]; + uvec2 scale0 = uvec2(unpack8(scale0_u16)); + uvec2 scale4 = uvec2(unpack8(scale4_u16)); + uvec2 scale8 = uvec2(unpack8(scale8_u16)); const uint32_t sc0 = ( scale0.x & 0x3f); const uint32_t sc1 = ( scale0.y & 0x3f);