diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp index 098771493..1f00e442d 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp @@ -25,7 +25,6 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { const uint l0 = 2*v_in; // 0...15 const uint q_offset = 32*v_im + l0; - const uint s_offset = 8*v_im; const uint y_offset = 128*v_im + l0; FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; 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 940ba3f51..6761f3d32 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 @@ -43,55 +43,57 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { [[unroll]] for (uint n = 0; n < num_rows; ++n) { const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row; - f16vec2 d = data_a[ib0 + i].d; + const f16vec2 d = data_a[ib0 + i].d; const FLOAT_TYPE dall = FLOAT_TYPE(d.x); const FLOAT_TYPE dmin = FLOAT_TYPE(d.y); - 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 scale0_u32 = data_a_packed16[ib0 + i].scales[v_im ]; + const uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2]; + const uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4]; - const uint32_t sc0 = ( scale0.x & 0x3f); - const uint32_t sc1 = ( scale0.y & 0x3f); - const uint32_t sc2 = ( scale4.x & 0x3f); - const uint32_t sc3 = ( scale4.y & 0x3f); - const uint32_t sc4 = (( scale8.x & 0x0f) | ((scale0.x & 0xc0) >> 2)); - const uint32_t sc5 = (( scale8.y & 0x0f) | ((scale0.y & 0xc0) >> 2)); - const uint32_t sc6 = (((scale8.x >> 4) & 0x0f) | ((scale4.x & 0xc0) >> 2)); - const uint32_t sc7 = (((scale8.y >> 4) & 0x0f) | ((scale4.y & 0xc0) >> 2)); + const uint32_t scale_0_4_l = (scale4_u32 << 16) | scale0_u32; + const uint32_t scale_0_4_h = (scale_0_4_l & 0xc0c0c0c0) >> 2; + const vec4 scale_0_4_l_f = vec4(unpack8(scale_0_4_l & 0x3f3f3f3f)); + const vec4 scale8_f = vec4(unpack8(((((scale8_u32 >> 4) << 16) | scale8_u32) & 0x0f0f0f0f) | scale_0_4_h)); - uint32_t qs0_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4]; - uint32_t qs64_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4 + 16]; + const FLOAT_TYPE sc0 = scale_0_4_l_f.x; + const FLOAT_TYPE sc1 = scale_0_4_l_f.y; + const FLOAT_TYPE sc2 = scale_0_4_l_f.z; + const FLOAT_TYPE sc3 = scale_0_4_l_f.w; + const FLOAT_TYPE sc4 = scale8_f.x; + const FLOAT_TYPE sc5 = scale8_f.y; + const FLOAT_TYPE sc6 = scale8_f.z; + const FLOAT_TYPE sc7 = scale8_f.w; - uint32_t qs0_u32_lo4 = qs0_u32 & 0x0F0F0F0F; - uint32_t qs0_u32_hi4 = (qs0_u32 >> 4) & 0x0F0F0F0F; - uint32_t qs64_u32_lo4 = qs64_u32 & 0x0F0F0F0F; - uint32_t qs64_u32_hi4 = (qs64_u32 >> 4) & 0x0F0F0F0F; + const uint32_t qs0_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4]; + const uint32_t qs64_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4 + 16]; - uvec4 qs0_lo4 = uvec4(unpack8(qs0_u32_lo4)); - uvec4 qs64_lo4 = uvec4(unpack8(qs64_u32_lo4)); - uvec4 qs0_hi4 = uvec4(unpack8(qs0_u32_hi4)); - uvec4 qs64_hi4 = uvec4(unpack8(qs64_u32_hi4)); + const uint32_t qs0_u32_lo4 = qs0_u32 & 0x0F0F0F0F; + const uint32_t qs0_u32_hi4 = (qs0_u32 >> 4) & 0x0F0F0F0F; + const uint32_t qs64_u32_lo4 = qs64_u32 & 0x0F0F0F0F; + const uint32_t qs64_u32_hi4 = (qs64_u32 >> 4) & 0x0F0F0F0F; - const uint32_t q4_0 = qs0_lo4.x; - const uint32_t q4_1 = qs0_lo4.y; - const uint32_t q4_2 = qs0_lo4.z; - const uint32_t q4_3 = qs0_lo4.w; - const uint32_t q4_4 = qs0_hi4.x; - const uint32_t q4_5 = qs0_hi4.y; - const uint32_t q4_6 = qs0_hi4.z; - const uint32_t q4_7 = qs0_hi4.w; - const uint32_t q4_8 = qs64_lo4.x; - const uint32_t q4_9 = qs64_lo4.y; - const uint32_t q4_10 = qs64_lo4.z; - const uint32_t q4_11 = qs64_lo4.w; - const uint32_t q4_12 = qs64_hi4.x; - const uint32_t q4_13 = qs64_hi4.y; - const uint32_t q4_14 = qs64_hi4.z; - const uint32_t q4_15 = qs64_hi4.w; + const vec4 qs0_lo4 = vec4(unpack8(qs0_u32_lo4)); + const vec4 qs64_lo4 = vec4(unpack8(qs64_u32_lo4)); + const vec4 qs0_hi4 = vec4(unpack8(qs0_u32_hi4)); + const vec4 qs64_hi4 = vec4(unpack8(qs64_u32_hi4)); + + const FLOAT_TYPE q4_0 = qs0_lo4.x; + const FLOAT_TYPE q4_1 = qs0_lo4.y; + const FLOAT_TYPE q4_2 = qs0_lo4.z; + const FLOAT_TYPE q4_3 = qs0_lo4.w; + const FLOAT_TYPE q4_4 = qs0_hi4.x; + const FLOAT_TYPE q4_5 = qs0_hi4.y; + const FLOAT_TYPE q4_6 = qs0_hi4.z; + const FLOAT_TYPE q4_7 = qs0_hi4.w; + const FLOAT_TYPE q4_8 = qs64_lo4.x; + const FLOAT_TYPE q4_9 = qs64_lo4.y; + const FLOAT_TYPE q4_10 = qs64_lo4.z; + const FLOAT_TYPE q4_11 = qs64_lo4.w; + const FLOAT_TYPE q4_12 = qs64_hi4.x; + const FLOAT_TYPE q4_13 = qs64_hi4.y; + const FLOAT_TYPE q4_14 = qs64_hi4.z; + const FLOAT_TYPE q4_15 = qs64_hi4.w; [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { B_TYPE_VEC4 by10 = data_b_v4[(j*p.batch_stride_b + b_offset + y1_idx) / 4];