From 7c7836d9d4062d6858e3fb337b135c417ccee6ce Mon Sep 17 00:00:00 2001 From: 0cc4m Date: Sun, 16 Jun 2024 07:17:31 +0200 Subject: Vulkan Shader Refactor, Memory Debugging Option (#7947) * Refactor shaders, extract GLSL code from ggml_vk_generate_shaders.py into vulkan-shaders directory * Improve debug log code * Add memory debug output option * Fix flake8 * Fix unnecessary high llama-3 VRAM use --- vulkan-shaders/dequant_q5_0.comp | 34 ++++++++++++++++++++++++++++++++++ 1 file changed, 34 insertions(+) create mode 100644 vulkan-shaders/dequant_q5_0.comp (limited to 'vulkan-shaders/dequant_q5_0.comp') diff --git a/vulkan-shaders/dequant_q5_0.comp b/vulkan-shaders/dequant_q5_0.comp new file mode 100644 index 00000000..b20b8052 --- /dev/null +++ b/vulkan-shaders/dequant_q5_0.comp @@ -0,0 +1,34 @@ +#version 450 + +#include "dequant_head.comp" + +layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer A {block_q5_0 data_a[];}; +layout (binding = 1) writeonly buffer D {D_TYPE data_b[];}; + +void main() { + const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64; + + const uint tid = gl_LocalInvocationID.x % 64; + const uint il = tid/32; + const uint ir = tid%32; + const uint ib = 32*i + ir; + if (ib >= p.nel / 32) { + return; + } + + const uint b_idx = 1024*i + 32*ir + 8*il; + + const float d = float(data_a[ib].d); + const uint qh = uint(data_a[ib].qh[1]) << 16 | data_a[ib].qh[0]; + + const uint q_idx = 8*il; + + [[unroll]] for (uint l = 0; l < 8; ++l) { + const uint iqs = q_idx + l; + const uint vui = uint(data_a[ib].qs[iqs]); + data_b[b_idx + l + 0] = D_TYPE(d * (((vui & 0xF) | (((qh >> iqs) << 4) & 0x10)) - 16.0f)); + data_b[b_idx + l + 16] = D_TYPE(d * (((vui >> 4) | ((qh >> (iqs + 12)) & 0x10)) - 16.0f)); + } +} -- cgit v1.2.3