From 07d0d58bef57366233b52c421632cfb2e54c76d4 Mon Sep 17 00:00:00 2001 From: Eve <139727413+netrunnereve@users.noreply.github.com> Date: Sat, 4 Jan 2025 16:24:29 -0500 Subject: [PATCH] q3_k --- .../vulkan-shaders/mul_mat_vec_q3_k.comp | 36 ++++++++----------- 1 file changed, 14 insertions(+), 22 deletions(-) diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp index b9a87f13a..0266417a8 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp @@ -5,6 +5,9 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; +shared uint32_t scu8[BLOCK_SIZE/16][12]; +shared FLOAT_TYPE sccache[BLOCK_SIZE/16][2][12]; + void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { uint a_offset, b_offset, d_offset; get_offsets(a_offset, b_offset, d_offset); @@ -16,6 +19,7 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { const uint tid = gl_LocalInvocationID.x; const uint itid = tid%16; // 0...15 const uint ix = tid/16; + const uint itid8 = itid%8; const uint v_im = itid/8; // 0 or 1. 0 computes 0..., 1 computes 128... const uint v_in = itid - 8*v_im; // 0...15 or 0...7 @@ -34,8 +38,6 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { } } - const uint s_shift = 4 * v_im; - [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) { const uint y_idx = i * QUANT_K + y_offset; @@ -43,18 +45,8 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row; const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d); - uint16_t s0_16 = data_a_packed16[ib0 + i].scales[0]; - uint16_t s2_16 = data_a_packed16[ib0 + i].scales[1]; - uint16_t s4_16 = data_a_packed16[ib0 + i].scales[2]; - uint16_t s6_16 = data_a_packed16[ib0 + i].scales[3]; - uint16_t s8_16 = data_a_packed16[ib0 + i].scales[4]; - uint16_t s10_16 = data_a_packed16[ib0 + i].scales[5]; - u8vec2 s0 = unpack8(s0_16); - u8vec2 s2 = unpack8(s2_16); - u8vec2 s4 = unpack8(s4_16); - u8vec2 s6 = unpack8(s6_16); - u8vec2 s8 = unpack8(s8_16); - u8vec2 s10 = unpack8(s10_16); + sccache[ix][v_im][itid8] = FLOAT_TYPE(int8_t(((int8_t(data_a[ib0+i].scales[itid8]) >> 4*v_im) & 0xF) | ((int8_t(data_a[ib0+i].scales[itid%4+8]) >> (4*v_im + 2*(itid8/4)) & 0x3) << 4)) - 32); + barrier(); [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { @@ -69,14 +61,14 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { FLOAT_TYPE sum = FLOAT_TYPE(0.0); [[unroll]] for (int l = 0; l < 2; ++l) { - sum = fma(FLOAT_TYPE(b0[l]) * FLOAT_TYPE(int8_t(((s0[0] >> s_shift) & 0xF) | ((s8[0] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] ) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 0)) != 0) ? 0 : 4)), - fma(FLOAT_TYPE(b32[l]) * FLOAT_TYPE(int8_t(((s2[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 1)) != 0) ? 0 : 4)), - fma(FLOAT_TYPE(b64[l]) * FLOAT_TYPE(int8_t(((s4[0] >> s_shift) & 0xF) | ((s8[0] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 2)) != 0) ? 0 : 4)), - fma(FLOAT_TYPE(b96[l]) * FLOAT_TYPE(int8_t(((s6[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 3)) != 0) ? 0 : 4)), - fma(FLOAT_TYPE(b16[l]) * FLOAT_TYPE(int8_t(((s0[1] >> s_shift) & 0xF) | ((s8[1] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4)), - fma(FLOAT_TYPE(b48[l]) * FLOAT_TYPE(int8_t(((s2[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4)), - fma(FLOAT_TYPE(b80[l]) * FLOAT_TYPE(int8_t(((s4[1] >> s_shift) & 0xF) | ((s8[1] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4)), - fma(FLOAT_TYPE(b112[l]) * FLOAT_TYPE(int8_t(((s6[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4)), sum)))))))); + sum = fma(FLOAT_TYPE( b0[l]) * sccache[ix][v_im][0], FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] ) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m )) != 0) ? 0 : 4)), + fma(FLOAT_TYPE( b16[l]) * sccache[ix][v_im][1], FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m )) != 0) ? 0 : 4)), + fma(FLOAT_TYPE( b32[l]) * sccache[ix][v_im][2], FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 1)) != 0) ? 0 : 4)), + fma(FLOAT_TYPE( b48[l]) * sccache[ix][v_im][3], FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4)), + fma(FLOAT_TYPE( b64[l]) * sccache[ix][v_im][4], FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 2)) != 0) ? 0 : 4)), + fma(FLOAT_TYPE( b80[l]) * sccache[ix][v_im][5], FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4)), + fma(FLOAT_TYPE( b96[l]) * sccache[ix][v_im][6], FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 3)) != 0) ? 0 : 4)), + fma(FLOAT_TYPE(b112[l]) * sccache[ix][v_im][7], FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4)), sum)))))))); } temp[j][n] = fma(d, sum, temp[j][n]); }